This is mostly my code_gen_buffer cleanup, plus a few other random changes thrown in. Including a fix for a recent float32_exp2 bug. r~ The following changes since commit 894fc4fd670aaf04a67dc7507739f914ff4bacf2: Merge remote-tracking branch 'remotes/jasowang/tags/net-pull-request' into staging (2021-06-11 09:21:48 +0100) are available in the Git repository at: https://gitlab.com/rth7680/qemu.git tags/pull-tcg-20210611 for you to fetch changes up to 60afaddc208d34f6dc86dd974f6e02724fba6eb6: docs/devel: Explain in more detail the TB chaining mechanisms (2021-06-11 09:41:25 -0700) ---------------------------------------------------------------- Clean up code_gen_buffer allocation. Add tcg_remove_ops_after. Fix tcg_constant_* documentation. Improve TB chaining documentation. Fix float32_exp2. ---------------------------------------------------------------- Jose R. Ziviani (1): tcg/arm: Fix tcg_out_op function signature Luis Pires (1): docs/devel: Explain in more detail the TB chaining mechanisms Richard Henderson (32): meson: Split out tcg/meson.build meson: Split out fpu/meson.build tcg: Re-order tcg_region_init vs tcg_prologue_init tcg: Remove error return from tcg_region_initial_alloc__locked tcg: Split out tcg_region_initial_alloc tcg: Split out tcg_region_prologue_set tcg: Split out region.c accel/tcg: Inline cpu_gen_init accel/tcg: Move alloc_code_gen_buffer to tcg/region.c accel/tcg: Rename tcg_init to tcg_init_machine tcg: Create tcg_init accel/tcg: Merge tcg_exec_init into tcg_init_machine accel/tcg: Use MiB in tcg_init_machine accel/tcg: Pass down max_cpus to tcg_init tcg: Introduce tcg_max_ctxs tcg: Move MAX_CODE_GEN_BUFFER_SIZE to tcg-target.h tcg: Replace region.end with region.total_size tcg: Rename region.start to region.after_prologue tcg: Tidy tcg_n_regions tcg: Tidy split_cross_256mb tcg: Move in_code_gen_buffer and tests to region.c tcg: Allocate code_gen_buffer into struct tcg_region_state tcg: Return the map protection from alloc_code_gen_buffer tcg: Sink qemu_madvise call to common code util/osdep: Add qemu_mprotect_rw tcg: Round the tb_size default from qemu_get_host_physmem tcg: Merge buffer protection and guard page protection tcg: When allocating for !splitwx, begin with PROT_NONE tcg: Move tcg_init_ctx and tcg_ctx from accel/tcg/ tcg: Introduce tcg_remove_ops_after tcg: Fix documentation for tcg_constant_* vs tcg_temp_free_* softfloat: Fix tp init in float32_exp2 docs/devel/tcg.rst | 101 ++++- meson.build | 12 +- accel/tcg/internal.h | 2 + include/qemu/osdep.h | 1 + include/sysemu/tcg.h | 2 - include/tcg/tcg.h | 28 +- tcg/aarch64/tcg-target.h | 1 + tcg/arm/tcg-target.h | 1 + tcg/i386/tcg-target.h | 2 + tcg/mips/tcg-target.h | 6 + tcg/ppc/tcg-target.h | 2 + tcg/riscv/tcg-target.h | 1 + tcg/s390/tcg-target.h | 3 + tcg/sparc/tcg-target.h | 1 + tcg/tcg-internal.h | 40 ++ tcg/tci/tcg-target.h | 1 + accel/tcg/tcg-all.c | 32 +- accel/tcg/translate-all.c | 439 +------------------- bsd-user/main.c | 3 +- fpu/softfloat.c | 2 +- linux-user/main.c | 1 - tcg/region.c | 999 ++++++++++++++++++++++++++++++++++++++++++++++ tcg/tcg.c | 649 +++--------------------------- util/osdep.c | 9 + tcg/arm/tcg-target.c.inc | 3 +- fpu/meson.build | 1 + tcg/meson.build | 14 + 27 files changed, 1266 insertions(+), 1090 deletions(-) create mode 100644 tcg/tcg-internal.h create mode 100644 tcg/region.c create mode 100644 fpu/meson.build create mode 100644 tcg/meson.build Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- meson.build | 8 +------- tcg/meson.build | 13 +++++++++++++ 2 files changed, 14 insertions(+), 7 deletions(-) create mode 100644 tcg/meson.build diff --git a/meson.build b/meson.build index XXXXXXX..XXXXXXX 100644 --- a/meson.build +++ b/meson.build @@ -XXX,XX +XXX,XX @@ common_ss.add(capstone) specific_ss.add(files('cpu.c', 'disas.c', 'gdbstub.c'), capstone) specific_ss.add(when: 'CONFIG_TCG', if_true: files( 'fpu/softfloat.c', - 'tcg/optimize.c', - 'tcg/tcg-common.c', - 'tcg/tcg-op-gvec.c', - 'tcg/tcg-op-vec.c', - 'tcg/tcg-op.c', - 'tcg/tcg.c', )) -specific_ss.add(when: 'CONFIG_TCG_INTERPRETER', if_true: files('tcg/tci.c')) # Work around a gcc bug/misfeature wherein constant propagation looks # through an alias: @@ -XXX,XX +XXX,XX @@ subdir('net') subdir('replay') subdir('semihosting') subdir('hw') +subdir('tcg') subdir('accel') subdir('plugins') subdir('bsd-user') diff --git a/tcg/meson.build b/tcg/meson.build new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/tcg/meson.build @@ -XXX,XX +XXX,XX @@ +tcg_ss = ss.source_set() + +tcg_ss.add(files( + 'optimize.c', + 'tcg.c', + 'tcg-common.c', + 'tcg-op.c', + 'tcg-op-gvec.c', + 'tcg-op-vec.c', +)) +tcg_ss.add(when: 'CONFIG_TCG_INTERPRETER', if_true: files('tci.c')) + +specific_ss.add_all(when: 'CONFIG_TCG', if_true: tcg_ss) -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- meson.build | 4 +--- fpu/meson.build | 1 + 2 files changed, 2 insertions(+), 3 deletions(-) create mode 100644 fpu/meson.build diff --git a/meson.build b/meson.build index XXXXXXX..XXXXXXX 100644 --- a/meson.build +++ b/meson.build @@ -XXX,XX +XXX,XX @@ subdir('softmmu') common_ss.add(capstone) specific_ss.add(files('cpu.c', 'disas.c', 'gdbstub.c'), capstone) -specific_ss.add(when: 'CONFIG_TCG', if_true: files( - 'fpu/softfloat.c', -)) # Work around a gcc bug/misfeature wherein constant propagation looks # through an alias: @@ -XXX,XX +XXX,XX @@ subdir('replay') subdir('semihosting') subdir('hw') subdir('tcg') +subdir('fpu') subdir('accel') subdir('plugins') subdir('bsd-user') diff --git a/fpu/meson.build b/fpu/meson.build new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/fpu/meson.build @@ -0,0 +1 @@ +specific_ss.add(when: 'CONFIG_TCG', if_true: files('softfloat.c')) -- 2.25.1 Instead of delaying tcg_region_init until after tcg_prologue_init is complete, do tcg_region_init first and let tcg_prologue_init shrink the first region by the size of the generated prologue. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/tcg-all.c | 11 --------- accel/tcg/translate-all.c | 3 +++ bsd-user/main.c | 1 - linux-user/main.c | 1 - tcg/tcg.c | 52 ++++++++++++++------------------------- 5 files changed, 22 insertions(+), 46 deletions(-) diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/tcg-all.c +++ b/accel/tcg/tcg-all.c @@ -XXX,XX +XXX,XX @@ static int tcg_init(MachineState *ms) tcg_exec_init(s->tb_size * 1024 * 1024, s->splitwx_enabled); mttcg_enabled = s->mttcg_enabled; - - /* - * Initialize TCG regions only for softmmu. - * - * This needs to be done later for user mode, because the prologue - * generation needs to be delayed so that GUEST_BASE is already set. - */ -#ifndef CONFIG_USER_ONLY - tcg_region_init(); -#endif /* !CONFIG_USER_ONLY */ - return 0; } diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ void tcg_exec_init(unsigned long tb_size, int splitwx) splitwx, &error_fatal); assert(ok); + /* TODO: allocating regions is hand-in-glove with code_gen_buffer. */ + tcg_region_init(); + #if defined(CONFIG_SOFTMMU) /* There's no guest base to take into account, so go ahead and initialize the prologue now. */ diff --git a/bsd-user/main.c b/bsd-user/main.c index XXXXXXX..XXXXXXX 100644 --- a/bsd-user/main.c +++ b/bsd-user/main.c @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv) * the real value of GUEST_BASE into account. */ tcg_prologue_init(tcg_ctx); - tcg_region_init(); /* build Task State */ memset(ts, 0, sizeof(TaskState)); diff --git a/linux-user/main.c b/linux-user/main.c index XXXXXXX..XXXXXXX 100644 --- a/linux-user/main.c +++ b/linux-user/main.c @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv, char **envp) generating the prologue until now so that the prologue can take the real value of GUEST_BASE into account. */ tcg_prologue_init(tcg_ctx); - tcg_region_init(); target_cpu_copy_regs(env, regs); diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ TranslationBlock *tcg_tb_alloc(TCGContext *s) void tcg_prologue_init(TCGContext *s) { - size_t prologue_size, total_size; - void *buf0, *buf1; + size_t prologue_size; /* Put the prologue at the beginning of code_gen_buffer. */ - buf0 = s->code_gen_buffer; - total_size = s->code_gen_buffer_size; - s->code_ptr = buf0; - s->code_buf = buf0; + tcg_region_assign(s, 0); + s->code_ptr = s->code_gen_ptr; + s->code_buf = s->code_gen_ptr; s->data_gen_ptr = NULL; - /* - * The region trees are not yet configured, but tcg_splitwx_to_rx - * needs the bounds for an assert. - */ - region.start = buf0; - region.end = buf0 + total_size; - #ifndef CONFIG_TCG_INTERPRETER - tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(buf0); + tcg_qemu_tb_exec = (tcg_prologue_fn *)tcg_splitwx_to_rx(s->code_ptr); #endif - /* Compute a high-water mark, at which we voluntarily flush the buffer - and start over. The size here is arbitrary, significantly larger - than we expect the code generation for any one opcode to require. */ - s->code_gen_highwater = s->code_gen_buffer + (total_size - TCG_HIGHWATER); - #ifdef TCG_TARGET_NEED_POOL_LABELS s->pool_labels = NULL; #endif @@ -XXX,XX +XXX,XX @@ void tcg_prologue_init(TCGContext *s) } #endif - buf1 = s->code_ptr; + prologue_size = tcg_current_code_size(s); + #ifndef CONFIG_TCG_INTERPRETER - flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(buf0), (uintptr_t)buf0, - tcg_ptr_byte_diff(buf1, buf0)); + flush_idcache_range((uintptr_t)tcg_splitwx_to_rx(s->code_buf), + (uintptr_t)s->code_buf, prologue_size); #endif - /* Deduct the prologue from the buffer. */ - prologue_size = tcg_current_code_size(s); - s->code_gen_ptr = buf1; - s->code_gen_buffer = buf1; - s->code_buf = buf1; - total_size -= prologue_size; - s->code_gen_buffer_size = total_size; + /* Deduct the prologue from the first region. */ + region.start = s->code_ptr; - tcg_register_jit(tcg_splitwx_to_rx(s->code_gen_buffer), total_size); + /* Recompute boundaries of the first region. */ + tcg_region_assign(s, 0); + + tcg_register_jit(tcg_splitwx_to_rx(region.start), + region.end - region.start); #ifdef DEBUG_DISAS if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { FILE *logfile = qemu_log_lock(); qemu_log("PROLOGUE: [size=%zu]\n", prologue_size); if (s->data_gen_ptr) { - size_t code_size = s->data_gen_ptr - buf0; + size_t code_size = s->data_gen_ptr - s->code_gen_ptr; size_t data_size = prologue_size - code_size; size_t i; - log_disas(buf0, code_size); + log_disas(s->code_gen_ptr, code_size); for (i = 0; i < data_size; i += sizeof(tcg_target_ulong)) { if (sizeof(tcg_target_ulong) == 8) { @@ -XXX,XX +XXX,XX @@ void tcg_prologue_init(TCGContext *s) } } } else { - log_disas(buf0, prologue_size); + log_disas(s->code_gen_ptr, prologue_size); } qemu_log("\n"); qemu_log_flush(); -- 2.25.1 All callers immediately assert on error, so move the assert into the function itself. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg.c | 19 ++++++------------- 1 file changed, 6 insertions(+), 13 deletions(-) diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static bool tcg_region_alloc(TCGContext *s) * Perform a context's first region allocation. * This function does _not_ increment region.agg_size_full. */ -static inline bool tcg_region_initial_alloc__locked(TCGContext *s) +static void tcg_region_initial_alloc__locked(TCGContext *s) { - return tcg_region_alloc__locked(s); + bool err = tcg_region_alloc__locked(s); + g_assert(!err); } /* Call from a safe-work context */ @@ -XXX,XX +XXX,XX @@ void tcg_region_reset_all(void) for (i = 0; i < n_ctxs; i++) { TCGContext *s = qatomic_read(&tcg_ctxs[i]); - bool err = tcg_region_initial_alloc__locked(s); - - g_assert(!err); + tcg_region_initial_alloc__locked(s); } qemu_mutex_unlock(&region.lock); @@ -XXX,XX +XXX,XX @@ void tcg_region_init(void) /* In user-mode we support only one ctx, so do the initial allocation now */ #ifdef CONFIG_USER_ONLY - { - bool err = tcg_region_initial_alloc__locked(tcg_ctx); - - g_assert(!err); - } + tcg_region_initial_alloc__locked(tcg_ctx); #endif } @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) MachineState *ms = MACHINE(qdev_get_machine()); TCGContext *s = g_malloc(sizeof(*s)); unsigned int i, n; - bool err; *s = tcg_init_ctx; @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) tcg_ctx = s; qemu_mutex_lock(&region.lock); - err = tcg_region_initial_alloc__locked(tcg_ctx); - g_assert(!err); + tcg_region_initial_alloc__locked(s); qemu_mutex_unlock(&region.lock); } #endif /* !CONFIG_USER_ONLY */ -- 2.25.1 This has only one user, and currently needs an ifdef, but will make more sense after some code motion. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg.c | 13 ++++++++++--- 1 file changed, 10 insertions(+), 3 deletions(-) diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static void tcg_region_initial_alloc__locked(TCGContext *s) g_assert(!err); } +#ifndef CONFIG_USER_ONLY +static void tcg_region_initial_alloc(TCGContext *s) +{ + qemu_mutex_lock(&region.lock); + tcg_region_initial_alloc__locked(s); + qemu_mutex_unlock(&region.lock); +} +#endif + /* Call from a safe-work context */ void tcg_region_reset_all(void) { @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) } tcg_ctx = s; - qemu_mutex_lock(&region.lock); - tcg_region_initial_alloc__locked(s); - qemu_mutex_unlock(&region.lock); + tcg_region_initial_alloc(s); } #endif /* !CONFIG_USER_ONLY */ -- 2.25.1 This has only one user, but will make more sense after some code motion. Always leave the tcg_init_ctx initialized to the first region, in preparation for tcg_prologue_init(). This also requires that we don't re-allocate the region for the first cpu, lest we hit the assertion for total number of regions allocated . Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg.c | 37 ++++++++++++++++++++++--------------- 1 file changed, 22 insertions(+), 15 deletions(-) diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ void tcg_region_init(void) tcg_region_trees_init(); - /* In user-mode we support only one ctx, so do the initial allocation now */ -#ifdef CONFIG_USER_ONLY - tcg_region_initial_alloc__locked(tcg_ctx); -#endif + /* + * Leave the initial context initialized to the first region. + * This will be the context into which we generate the prologue. + * It is also the only context for CONFIG_USER_ONLY. + */ + tcg_region_initial_alloc__locked(&tcg_init_ctx); +} + +static void tcg_region_prologue_set(TCGContext *s) +{ + /* Deduct the prologue from the first region. */ + g_assert(region.start == s->code_gen_buffer); + region.start = s->code_ptr; + + /* Recompute boundaries of the first region. */ + tcg_region_assign(s, 0); + + /* Register the balance of the buffer with gdb. */ + tcg_register_jit(tcg_splitwx_to_rx(region.start), + region.end - region.start); } #ifdef CONFIG_DEBUG_TCG @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) if (n > 0) { alloc_tcg_plugin_context(s); + tcg_region_initial_alloc(s); } tcg_ctx = s; - tcg_region_initial_alloc(s); } #endif /* !CONFIG_USER_ONLY */ @@ -XXX,XX +XXX,XX @@ void tcg_prologue_init(TCGContext *s) { size_t prologue_size; - /* Put the prologue at the beginning of code_gen_buffer. */ - tcg_region_assign(s, 0); s->code_ptr = s->code_gen_ptr; s->code_buf = s->code_gen_ptr; s->data_gen_ptr = NULL; @@ -XXX,XX +XXX,XX @@ void tcg_prologue_init(TCGContext *s) (uintptr_t)s->code_buf, prologue_size); #endif - /* Deduct the prologue from the first region. */ - region.start = s->code_ptr; - - /* Recompute boundaries of the first region. */ - tcg_region_assign(s, 0); - - tcg_register_jit(tcg_splitwx_to_rx(region.start), - region.end - region.start); + tcg_region_prologue_set(s); #ifdef DEBUG_DISAS if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM)) { -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg-internal.h | 37 +++ tcg/region.c | 572 +++++++++++++++++++++++++++++++++++++++++++++ tcg/tcg.c | 547 +------------------------------------------ tcg/meson.build | 1 + 4 files changed, 613 insertions(+), 544 deletions(-) create mode 100644 tcg/tcg-internal.h create mode 100644 tcg/region.c diff --git a/tcg/tcg-internal.h b/tcg/tcg-internal.h new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/tcg/tcg-internal.h @@ -XXX,XX +XXX,XX @@ +/* + * Internal declarations for Tiny Code Generator for QEMU + * + * Copyright (c) 2008 Fabrice Bellard + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +#ifndef TCG_INTERNAL_H +#define TCG_INTERNAL_H 1 + +#define TCG_HIGHWATER 1024 + +extern TCGContext **tcg_ctxs; +extern unsigned int n_tcg_ctxs; + +bool tcg_region_alloc(TCGContext *s); +void tcg_region_initial_alloc(TCGContext *s); +void tcg_region_prologue_set(TCGContext *s); + +#endif /* TCG_INTERNAL_H */ diff --git a/tcg/region.c b/tcg/region.c new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ +/* + * Memory region management for Tiny Code Generator for QEMU + * + * Copyright (c) 2008 Fabrice Bellard + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to deal + * in the Software without restriction, including without limitation the rights + * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell + * copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in + * all copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL + * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN + * THE SOFTWARE. + */ + +#include "qemu/osdep.h" +#include "exec/exec-all.h" +#include "tcg/tcg.h" +#if !defined(CONFIG_USER_ONLY) +#include "hw/boards.h" +#endif +#include "tcg-internal.h" + + +struct tcg_region_tree { + QemuMutex lock; + GTree *tree; + /* padding to avoid false sharing is computed at run-time */ +}; + +/* + * We divide code_gen_buffer into equally-sized "regions" that TCG threads + * dynamically allocate from as demand dictates. Given appropriate region + * sizing, this minimizes flushes even when some TCG threads generate a lot + * more code than others. + */ +struct tcg_region_state { + QemuMutex lock; + + /* fields set at init time */ + void *start; + void *start_aligned; + void *end; + size_t n; + size_t size; /* size of one region */ + size_t stride; /* .size + guard size */ + + /* fields protected by the lock */ + size_t current; /* current region index */ + size_t agg_size_full; /* aggregate size of full regions */ +}; + +static struct tcg_region_state region; + +/* + * This is an array of struct tcg_region_tree's, with padding. + * We use void * to simplify the computation of region_trees[i]; each + * struct is found every tree_size bytes. + */ +static void *region_trees; +static size_t tree_size; + +/* compare a pointer @ptr and a tb_tc @s */ +static int ptr_cmp_tb_tc(const void *ptr, const struct tb_tc *s) +{ + if (ptr >= s->ptr + s->size) { + return 1; + } else if (ptr < s->ptr) { + return -1; + } + return 0; +} + +static gint tb_tc_cmp(gconstpointer ap, gconstpointer bp) +{ + const struct tb_tc *a = ap; + const struct tb_tc *b = bp; + + /* + * When both sizes are set, we know this isn't a lookup. + * This is the most likely case: every TB must be inserted; lookups + * are a lot less frequent. + */ + if (likely(a->size && b->size)) { + if (a->ptr > b->ptr) { + return 1; + } else if (a->ptr < b->ptr) { + return -1; + } + /* a->ptr == b->ptr should happen only on deletions */ + g_assert(a->size == b->size); + return 0; + } + /* + * All lookups have either .size field set to 0. + * From the glib sources we see that @ap is always the lookup key. However + * the docs provide no guarantee, so we just mark this case as likely. + */ + if (likely(a->size == 0)) { + return ptr_cmp_tb_tc(a->ptr, b); + } + return ptr_cmp_tb_tc(b->ptr, a); +} + +static void tcg_region_trees_init(void) +{ + size_t i; + + tree_size = ROUND_UP(sizeof(struct tcg_region_tree), qemu_dcache_linesize); + region_trees = qemu_memalign(qemu_dcache_linesize, region.n * tree_size); + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + qemu_mutex_init(&rt->lock); + rt->tree = g_tree_new(tb_tc_cmp); + } +} + +static struct tcg_region_tree *tc_ptr_to_region_tree(const void *p) +{ + size_t region_idx; + + /* + * Like tcg_splitwx_to_rw, with no assert. The pc may come from + * a signal handler over which the caller has no control. + */ + if (!in_code_gen_buffer(p)) { + p -= tcg_splitwx_diff; + if (!in_code_gen_buffer(p)) { + return NULL; + } + } + + if (p < region.start_aligned) { + region_idx = 0; + } else { + ptrdiff_t offset = p - region.start_aligned; + + if (offset > region.stride * (region.n - 1)) { + region_idx = region.n - 1; + } else { + region_idx = offset / region.stride; + } + } + return region_trees + region_idx * tree_size; +} + +void tcg_tb_insert(TranslationBlock *tb) +{ + struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); + + g_assert(rt != NULL); + qemu_mutex_lock(&rt->lock); + g_tree_insert(rt->tree, &tb->tc, tb); + qemu_mutex_unlock(&rt->lock); +} + +void tcg_tb_remove(TranslationBlock *tb) +{ + struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); + + g_assert(rt != NULL); + qemu_mutex_lock(&rt->lock); + g_tree_remove(rt->tree, &tb->tc); + qemu_mutex_unlock(&rt->lock); +} + +/* + * Find the TB 'tb' such that + * tb->tc.ptr <= tc_ptr < tb->tc.ptr + tb->tc.size + * Return NULL if not found. + */ +TranslationBlock *tcg_tb_lookup(uintptr_t tc_ptr) +{ + struct tcg_region_tree *rt = tc_ptr_to_region_tree((void *)tc_ptr); + TranslationBlock *tb; + struct tb_tc s = { .ptr = (void *)tc_ptr }; + + if (rt == NULL) { + return NULL; + } + + qemu_mutex_lock(&rt->lock); + tb = g_tree_lookup(rt->tree, &s); + qemu_mutex_unlock(&rt->lock); + return tb; +} + +static void tcg_region_tree_lock_all(void) +{ + size_t i; + + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + qemu_mutex_lock(&rt->lock); + } +} + +static void tcg_region_tree_unlock_all(void) +{ + size_t i; + + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + qemu_mutex_unlock(&rt->lock); + } +} + +void tcg_tb_foreach(GTraverseFunc func, gpointer user_data) +{ + size_t i; + + tcg_region_tree_lock_all(); + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + g_tree_foreach(rt->tree, func, user_data); + } + tcg_region_tree_unlock_all(); +} + +size_t tcg_nb_tbs(void) +{ + size_t nb_tbs = 0; + size_t i; + + tcg_region_tree_lock_all(); + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + nb_tbs += g_tree_nnodes(rt->tree); + } + tcg_region_tree_unlock_all(); + return nb_tbs; +} + +static gboolean tcg_region_tree_traverse(gpointer k, gpointer v, gpointer data) +{ + TranslationBlock *tb = v; + + tb_destroy(tb); + return FALSE; +} + +static void tcg_region_tree_reset_all(void) +{ + size_t i; + + tcg_region_tree_lock_all(); + for (i = 0; i < region.n; i++) { + struct tcg_region_tree *rt = region_trees + i * tree_size; + + g_tree_foreach(rt->tree, tcg_region_tree_traverse, NULL); + /* Increment the refcount first so that destroy acts as a reset */ + g_tree_ref(rt->tree); + g_tree_destroy(rt->tree); + } + tcg_region_tree_unlock_all(); +} + +static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend) +{ + void *start, *end; + + start = region.start_aligned + curr_region * region.stride; + end = start + region.size; + + if (curr_region == 0) { + start = region.start; + } + if (curr_region == region.n - 1) { + end = region.end; + } + + *pstart = start; + *pend = end; +} + +static void tcg_region_assign(TCGContext *s, size_t curr_region) +{ + void *start, *end; + + tcg_region_bounds(curr_region, &start, &end); + + s->code_gen_buffer = start; + s->code_gen_ptr = start; + s->code_gen_buffer_size = end - start; + s->code_gen_highwater = end - TCG_HIGHWATER; +} + +static bool tcg_region_alloc__locked(TCGContext *s) +{ + if (region.current == region.n) { + return true; + } + tcg_region_assign(s, region.current); + region.current++; + return false; +} + +/* + * Request a new region once the one in use has filled up. + * Returns true on error. + */ +bool tcg_region_alloc(TCGContext *s) +{ + bool err; + /* read the region size now; alloc__locked will overwrite it on success */ + size_t size_full = s->code_gen_buffer_size; + + qemu_mutex_lock(&region.lock); + err = tcg_region_alloc__locked(s); + if (!err) { + region.agg_size_full += size_full - TCG_HIGHWATER; + } + qemu_mutex_unlock(&region.lock); + return err; +} + +/* + * Perform a context's first region allocation. + * This function does _not_ increment region.agg_size_full. + */ +static void tcg_region_initial_alloc__locked(TCGContext *s) +{ + bool err = tcg_region_alloc__locked(s); + g_assert(!err); +} + +void tcg_region_initial_alloc(TCGContext *s) +{ + qemu_mutex_lock(&region.lock); + tcg_region_initial_alloc__locked(s); + qemu_mutex_unlock(&region.lock); +} + +/* Call from a safe-work context */ +void tcg_region_reset_all(void) +{ + unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int i; + + qemu_mutex_lock(&region.lock); + region.current = 0; + region.agg_size_full = 0; + + for (i = 0; i < n_ctxs; i++) { + TCGContext *s = qatomic_read(&tcg_ctxs[i]); + tcg_region_initial_alloc__locked(s); + } + qemu_mutex_unlock(&region.lock); + + tcg_region_tree_reset_all(); +} + +#ifdef CONFIG_USER_ONLY +static size_t tcg_n_regions(void) +{ + return 1; +} +#else +/* + * It is likely that some vCPUs will translate more code than others, so we + * first try to set more regions than max_cpus, with those regions being of + * reasonable size. If that's not possible we make do by evenly dividing + * the code_gen_buffer among the vCPUs. + */ +static size_t tcg_n_regions(void) +{ + size_t i; + + /* Use a single region if all we have is one vCPU thread */ +#if !defined(CONFIG_USER_ONLY) + MachineState *ms = MACHINE(qdev_get_machine()); + unsigned int max_cpus = ms->smp.max_cpus; +#endif + if (max_cpus == 1 || !qemu_tcg_mttcg_enabled()) { + return 1; + } + + /* Try to have more regions than max_cpus, with each region being >= 2 MB */ + for (i = 8; i > 0; i--) { + size_t regions_per_thread = i; + size_t region_size; + + region_size = tcg_init_ctx.code_gen_buffer_size; + region_size /= max_cpus * regions_per_thread; + + if (region_size >= 2 * 1024u * 1024) { + return max_cpus * regions_per_thread; + } + } + /* If we can't, then just allocate one region per vCPU thread */ + return max_cpus; +} +#endif + +/* + * Initializes region partitioning. + * + * Called at init time from the parent thread (i.e. the one calling + * tcg_context_init), after the target's TCG globals have been set. + * + * Region partitioning works by splitting code_gen_buffer into separate regions, + * and then assigning regions to TCG threads so that the threads can translate + * code in parallel without synchronization. + * + * In softmmu the number of TCG threads is bounded by max_cpus, so we use at + * least max_cpus regions in MTTCG. In !MTTCG we use a single region. + * Note that the TCG options from the command-line (i.e. -accel accel=tcg,[...]) + * must have been parsed before calling this function, since it calls + * qemu_tcg_mttcg_enabled(). + * + * In user-mode we use a single region. Having multiple regions in user-mode + * is not supported, because the number of vCPU threads (recall that each thread + * spawned by the guest corresponds to a vCPU thread) is only bounded by the + * OS, and usually this number is huge (tens of thousands is not uncommon). + * Thus, given this large bound on the number of vCPU threads and the fact + * that code_gen_buffer is allocated at compile-time, we cannot guarantee + * that the availability of at least one region per vCPU thread. + * + * However, this user-mode limitation is unlikely to be a significant problem + * in practice. Multi-threaded guests share most if not all of their translated + * code, which makes parallel code generation less appealing than in softmmu. + */ +void tcg_region_init(void) +{ + void *buf = tcg_init_ctx.code_gen_buffer; + void *aligned; + size_t size = tcg_init_ctx.code_gen_buffer_size; + size_t page_size = qemu_real_host_page_size; + size_t region_size; + size_t n_regions; + size_t i; + + n_regions = tcg_n_regions(); + + /* The first region will be 'aligned - buf' bytes larger than the others */ + aligned = QEMU_ALIGN_PTR_UP(buf, page_size); + g_assert(aligned < tcg_init_ctx.code_gen_buffer + size); + /* + * Make region_size a multiple of page_size, using aligned as the start. + * As a result of this we might end up with a few extra pages at the end of + * the buffer; we will assign those to the last region. + */ + region_size = (size - (aligned - buf)) / n_regions; + region_size = QEMU_ALIGN_DOWN(region_size, page_size); + + /* A region must have at least 2 pages; one code, one guard */ + g_assert(region_size >= 2 * page_size); + + /* init the region struct */ + qemu_mutex_init(&region.lock); + region.n = n_regions; + region.size = region_size - page_size; + region.stride = region_size; + region.start = buf; + region.start_aligned = aligned; + /* page-align the end, since its last page will be a guard page */ + region.end = QEMU_ALIGN_PTR_DOWN(buf + size, page_size); + /* account for that last guard page */ + region.end -= page_size; + + /* + * Set guard pages in the rw buffer, as that's the one into which + * buffer overruns could occur. Do not set guard pages in the rx + * buffer -- let that one use hugepages throughout. + */ + for (i = 0; i < region.n; i++) { + void *start, *end; + + tcg_region_bounds(i, &start, &end); + + /* + * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect + * rejects a permission change from RWX -> NONE. Guard pages are + * nice for bug detection but are not essential; ignore any failure. + */ + (void)qemu_mprotect_none(end, page_size); + } + + tcg_region_trees_init(); + + /* + * Leave the initial context initialized to the first region. + * This will be the context into which we generate the prologue. + * It is also the only context for CONFIG_USER_ONLY. + */ + tcg_region_initial_alloc__locked(&tcg_init_ctx); +} + +void tcg_region_prologue_set(TCGContext *s) +{ + /* Deduct the prologue from the first region. */ + g_assert(region.start == s->code_gen_buffer); + region.start = s->code_ptr; + + /* Recompute boundaries of the first region. */ + tcg_region_assign(s, 0); + + /* Register the balance of the buffer with gdb. */ + tcg_register_jit(tcg_splitwx_to_rx(region.start), + region.end - region.start); +} + +/* + * Returns the size (in bytes) of all translated code (i.e. from all regions) + * currently in the cache. + * See also: tcg_code_capacity() + * Do not confuse with tcg_current_code_size(); that one applies to a single + * TCG context. + */ +size_t tcg_code_size(void) +{ + unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int i; + size_t total; + + qemu_mutex_lock(&region.lock); + total = region.agg_size_full; + for (i = 0; i < n_ctxs; i++) { + const TCGContext *s = qatomic_read(&tcg_ctxs[i]); + size_t size; + + size = qatomic_read(&s->code_gen_ptr) - s->code_gen_buffer; + g_assert(size <= s->code_gen_buffer_size); + total += size; + } + qemu_mutex_unlock(&region.lock); + return total; +} + +/* + * Returns the code capacity (in bytes) of the entire cache, i.e. including all + * regions. + * See also: tcg_code_size() + */ +size_t tcg_code_capacity(void) +{ + size_t guard_size, capacity; + + /* no need for synchronization; these variables are set at init time */ + guard_size = region.stride - region.size; + capacity = region.end + guard_size - region.start; + capacity -= region.n * (guard_size + TCG_HIGHWATER); + return capacity; +} + +size_t tcg_tb_phys_invalidate_count(void) +{ + unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int i; + size_t total = 0; + + for (i = 0; i < n_ctxs; i++) { + const TCGContext *s = qatomic_read(&tcg_ctxs[i]); + + total += qatomic_read(&s->tb_phys_invalidate_count); + } + return total; +} diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ #include "elf.h" #include "exec/log.h" +#include "tcg-internal.h" /* Forward declarations for functions declared in tcg-target.c.inc and used here. */ @@ -XXX,XX +XXX,XX @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct); static int tcg_out_ldst_finalize(TCGContext *s); #endif -#define TCG_HIGHWATER 1024 - -static TCGContext **tcg_ctxs; -static unsigned int n_tcg_ctxs; +TCGContext **tcg_ctxs; +unsigned int n_tcg_ctxs; TCGv_env cpu_env = 0; const void *tcg_code_gen_epilogue; uintptr_t tcg_splitwx_diff; @@ -XXX,XX +XXX,XX @@ uintptr_t tcg_splitwx_diff; tcg_prologue_fn *tcg_qemu_tb_exec; #endif -struct tcg_region_tree { - QemuMutex lock; - GTree *tree; - /* padding to avoid false sharing is computed at run-time */ -}; - -/* - * We divide code_gen_buffer into equally-sized "regions" that TCG threads - * dynamically allocate from as demand dictates. Given appropriate region - * sizing, this minimizes flushes even when some TCG threads generate a lot - * more code than others. - */ -struct tcg_region_state { - QemuMutex lock; - - /* fields set at init time */ - void *start; - void *start_aligned; - void *end; - size_t n; - size_t size; /* size of one region */ - size_t stride; /* .size + guard size */ - - /* fields protected by the lock */ - size_t current; /* current region index */ - size_t agg_size_full; /* aggregate size of full regions */ -}; - -static struct tcg_region_state region; -/* - * This is an array of struct tcg_region_tree's, with padding. - * We use void * to simplify the computation of region_trees[i]; each - * struct is found every tree_size bytes. - */ -static void *region_trees; -static size_t tree_size; static TCGRegSet tcg_target_available_regs[TCG_TYPE_COUNT]; static TCGRegSet tcg_target_call_clobber_regs; @@ -XXX,XX +XXX,XX @@ static const TCGTargetOpDef constraint_sets[] = { #include "tcg-target.c.inc" -/* compare a pointer @ptr and a tb_tc @s */ -static int ptr_cmp_tb_tc(const void *ptr, const struct tb_tc *s) -{ - if (ptr >= s->ptr + s->size) { - return 1; - } else if (ptr < s->ptr) { - return -1; - } - return 0; -} - -static gint tb_tc_cmp(gconstpointer ap, gconstpointer bp) -{ - const struct tb_tc *a = ap; - const struct tb_tc *b = bp; - - /* - * When both sizes are set, we know this isn't a lookup. - * This is the most likely case: every TB must be inserted; lookups - * are a lot less frequent. - */ - if (likely(a->size && b->size)) { - if (a->ptr > b->ptr) { - return 1; - } else if (a->ptr < b->ptr) { - return -1; - } - /* a->ptr == b->ptr should happen only on deletions */ - g_assert(a->size == b->size); - return 0; - } - /* - * All lookups have either .size field set to 0. - * From the glib sources we see that @ap is always the lookup key. However - * the docs provide no guarantee, so we just mark this case as likely. - */ - if (likely(a->size == 0)) { - return ptr_cmp_tb_tc(a->ptr, b); - } - return ptr_cmp_tb_tc(b->ptr, a); -} - -static void tcg_region_trees_init(void) -{ - size_t i; - - tree_size = ROUND_UP(sizeof(struct tcg_region_tree), qemu_dcache_linesize); - region_trees = qemu_memalign(qemu_dcache_linesize, region.n * tree_size); - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - qemu_mutex_init(&rt->lock); - rt->tree = g_tree_new(tb_tc_cmp); - } -} - -static struct tcg_region_tree *tc_ptr_to_region_tree(const void *p) -{ - size_t region_idx; - - /* - * Like tcg_splitwx_to_rw, with no assert. The pc may come from - * a signal handler over which the caller has no control. - */ - if (!in_code_gen_buffer(p)) { - p -= tcg_splitwx_diff; - if (!in_code_gen_buffer(p)) { - return NULL; - } - } - - if (p < region.start_aligned) { - region_idx = 0; - } else { - ptrdiff_t offset = p - region.start_aligned; - - if (offset > region.stride * (region.n - 1)) { - region_idx = region.n - 1; - } else { - region_idx = offset / region.stride; - } - } - return region_trees + region_idx * tree_size; -} - -void tcg_tb_insert(TranslationBlock *tb) -{ - struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); - - g_assert(rt != NULL); - qemu_mutex_lock(&rt->lock); - g_tree_insert(rt->tree, &tb->tc, tb); - qemu_mutex_unlock(&rt->lock); -} - -void tcg_tb_remove(TranslationBlock *tb) -{ - struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); - - g_assert(rt != NULL); - qemu_mutex_lock(&rt->lock); - g_tree_remove(rt->tree, &tb->tc); - qemu_mutex_unlock(&rt->lock); -} - -/* - * Find the TB 'tb' such that - * tb->tc.ptr <= tc_ptr < tb->tc.ptr + tb->tc.size - * Return NULL if not found. - */ -TranslationBlock *tcg_tb_lookup(uintptr_t tc_ptr) -{ - struct tcg_region_tree *rt = tc_ptr_to_region_tree((void *)tc_ptr); - TranslationBlock *tb; - struct tb_tc s = { .ptr = (void *)tc_ptr }; - - if (rt == NULL) { - return NULL; - } - - qemu_mutex_lock(&rt->lock); - tb = g_tree_lookup(rt->tree, &s); - qemu_mutex_unlock(&rt->lock); - return tb; -} - -static void tcg_region_tree_lock_all(void) -{ - size_t i; - - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - qemu_mutex_lock(&rt->lock); - } -} - -static void tcg_region_tree_unlock_all(void) -{ - size_t i; - - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - qemu_mutex_unlock(&rt->lock); - } -} - -void tcg_tb_foreach(GTraverseFunc func, gpointer user_data) -{ - size_t i; - - tcg_region_tree_lock_all(); - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - g_tree_foreach(rt->tree, func, user_data); - } - tcg_region_tree_unlock_all(); -} - -size_t tcg_nb_tbs(void) -{ - size_t nb_tbs = 0; - size_t i; - - tcg_region_tree_lock_all(); - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - nb_tbs += g_tree_nnodes(rt->tree); - } - tcg_region_tree_unlock_all(); - return nb_tbs; -} - -static gboolean tcg_region_tree_traverse(gpointer k, gpointer v, gpointer data) -{ - TranslationBlock *tb = v; - - tb_destroy(tb); - return FALSE; -} - -static void tcg_region_tree_reset_all(void) -{ - size_t i; - - tcg_region_tree_lock_all(); - for (i = 0; i < region.n; i++) { - struct tcg_region_tree *rt = region_trees + i * tree_size; - - g_tree_foreach(rt->tree, tcg_region_tree_traverse, NULL); - /* Increment the refcount first so that destroy acts as a reset */ - g_tree_ref(rt->tree); - g_tree_destroy(rt->tree); - } - tcg_region_tree_unlock_all(); -} - -static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend) -{ - void *start, *end; - - start = region.start_aligned + curr_region * region.stride; - end = start + region.size; - - if (curr_region == 0) { - start = region.start; - } - if (curr_region == region.n - 1) { - end = region.end; - } - - *pstart = start; - *pend = end; -} - -static void tcg_region_assign(TCGContext *s, size_t curr_region) -{ - void *start, *end; - - tcg_region_bounds(curr_region, &start, &end); - - s->code_gen_buffer = start; - s->code_gen_ptr = start; - s->code_gen_buffer_size = end - start; - s->code_gen_highwater = end - TCG_HIGHWATER; -} - -static bool tcg_region_alloc__locked(TCGContext *s) -{ - if (region.current == region.n) { - return true; - } - tcg_region_assign(s, region.current); - region.current++; - return false; -} - -/* - * Request a new region once the one in use has filled up. - * Returns true on error. - */ -static bool tcg_region_alloc(TCGContext *s) -{ - bool err; - /* read the region size now; alloc__locked will overwrite it on success */ - size_t size_full = s->code_gen_buffer_size; - - qemu_mutex_lock(&region.lock); - err = tcg_region_alloc__locked(s); - if (!err) { - region.agg_size_full += size_full - TCG_HIGHWATER; - } - qemu_mutex_unlock(&region.lock); - return err; -} - -/* - * Perform a context's first region allocation. - * This function does _not_ increment region.agg_size_full. - */ -static void tcg_region_initial_alloc__locked(TCGContext *s) -{ - bool err = tcg_region_alloc__locked(s); - g_assert(!err); -} - -#ifndef CONFIG_USER_ONLY -static void tcg_region_initial_alloc(TCGContext *s) -{ - qemu_mutex_lock(&region.lock); - tcg_region_initial_alloc__locked(s); - qemu_mutex_unlock(&region.lock); -} -#endif - -/* Call from a safe-work context */ -void tcg_region_reset_all(void) -{ - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); - unsigned int i; - - qemu_mutex_lock(&region.lock); - region.current = 0; - region.agg_size_full = 0; - - for (i = 0; i < n_ctxs; i++) { - TCGContext *s = qatomic_read(&tcg_ctxs[i]); - tcg_region_initial_alloc__locked(s); - } - qemu_mutex_unlock(&region.lock); - - tcg_region_tree_reset_all(); -} - -#ifdef CONFIG_USER_ONLY -static size_t tcg_n_regions(void) -{ - return 1; -} -#else -/* - * It is likely that some vCPUs will translate more code than others, so we - * first try to set more regions than max_cpus, with those regions being of - * reasonable size. If that's not possible we make do by evenly dividing - * the code_gen_buffer among the vCPUs. - */ -static size_t tcg_n_regions(void) -{ - size_t i; - - /* Use a single region if all we have is one vCPU thread */ -#if !defined(CONFIG_USER_ONLY) - MachineState *ms = MACHINE(qdev_get_machine()); - unsigned int max_cpus = ms->smp.max_cpus; -#endif - if (max_cpus == 1 || !qemu_tcg_mttcg_enabled()) { - return 1; - } - - /* Try to have more regions than max_cpus, with each region being >= 2 MB */ - for (i = 8; i > 0; i--) { - size_t regions_per_thread = i; - size_t region_size; - - region_size = tcg_init_ctx.code_gen_buffer_size; - region_size /= max_cpus * regions_per_thread; - - if (region_size >= 2 * 1024u * 1024) { - return max_cpus * regions_per_thread; - } - } - /* If we can't, then just allocate one region per vCPU thread */ - return max_cpus; -} -#endif - -/* - * Initializes region partitioning. - * - * Called at init time from the parent thread (i.e. the one calling - * tcg_context_init), after the target's TCG globals have been set. - * - * Region partitioning works by splitting code_gen_buffer into separate regions, - * and then assigning regions to TCG threads so that the threads can translate - * code in parallel without synchronization. - * - * In softmmu the number of TCG threads is bounded by max_cpus, so we use at - * least max_cpus regions in MTTCG. In !MTTCG we use a single region. - * Note that the TCG options from the command-line (i.e. -accel accel=tcg,[...]) - * must have been parsed before calling this function, since it calls - * qemu_tcg_mttcg_enabled(). - * - * In user-mode we use a single region. Having multiple regions in user-mode - * is not supported, because the number of vCPU threads (recall that each thread - * spawned by the guest corresponds to a vCPU thread) is only bounded by the - * OS, and usually this number is huge (tens of thousands is not uncommon). - * Thus, given this large bound on the number of vCPU threads and the fact - * that code_gen_buffer is allocated at compile-time, we cannot guarantee - * that the availability of at least one region per vCPU thread. - * - * However, this user-mode limitation is unlikely to be a significant problem - * in practice. Multi-threaded guests share most if not all of their translated - * code, which makes parallel code generation less appealing than in softmmu. - */ -void tcg_region_init(void) -{ - void *buf = tcg_init_ctx.code_gen_buffer; - void *aligned; - size_t size = tcg_init_ctx.code_gen_buffer_size; - size_t page_size = qemu_real_host_page_size; - size_t region_size; - size_t n_regions; - size_t i; - - n_regions = tcg_n_regions(); - - /* The first region will be 'aligned - buf' bytes larger than the others */ - aligned = QEMU_ALIGN_PTR_UP(buf, page_size); - g_assert(aligned < tcg_init_ctx.code_gen_buffer + size); - /* - * Make region_size a multiple of page_size, using aligned as the start. - * As a result of this we might end up with a few extra pages at the end of - * the buffer; we will assign those to the last region. - */ - region_size = (size - (aligned - buf)) / n_regions; - region_size = QEMU_ALIGN_DOWN(region_size, page_size); - - /* A region must have at least 2 pages; one code, one guard */ - g_assert(region_size >= 2 * page_size); - - /* init the region struct */ - qemu_mutex_init(&region.lock); - region.n = n_regions; - region.size = region_size - page_size; - region.stride = region_size; - region.start = buf; - region.start_aligned = aligned; - /* page-align the end, since its last page will be a guard page */ - region.end = QEMU_ALIGN_PTR_DOWN(buf + size, page_size); - /* account for that last guard page */ - region.end -= page_size; - - /* - * Set guard pages in the rw buffer, as that's the one into which - * buffer overruns could occur. Do not set guard pages in the rx - * buffer -- let that one use hugepages throughout. - */ - for (i = 0; i < region.n; i++) { - void *start, *end; - - tcg_region_bounds(i, &start, &end); - - /* - * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect - * rejects a permission change from RWX -> NONE. Guard pages are - * nice for bug detection but are not essential; ignore any failure. - */ - (void)qemu_mprotect_none(end, page_size); - } - - tcg_region_trees_init(); - - /* - * Leave the initial context initialized to the first region. - * This will be the context into which we generate the prologue. - * It is also the only context for CONFIG_USER_ONLY. - */ - tcg_region_initial_alloc__locked(&tcg_init_ctx); -} - -static void tcg_region_prologue_set(TCGContext *s) -{ - /* Deduct the prologue from the first region. */ - g_assert(region.start == s->code_gen_buffer); - region.start = s->code_ptr; - - /* Recompute boundaries of the first region. */ - tcg_region_assign(s, 0); - - /* Register the balance of the buffer with gdb. */ - tcg_register_jit(tcg_splitwx_to_rx(region.start), - region.end - region.start); -} - #ifdef CONFIG_DEBUG_TCG const void *tcg_splitwx_to_rx(void *rw) { @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) } #endif /* !CONFIG_USER_ONLY */ -/* - * Returns the size (in bytes) of all translated code (i.e. from all regions) - * currently in the cache. - * See also: tcg_code_capacity() - * Do not confuse with tcg_current_code_size(); that one applies to a single - * TCG context. - */ -size_t tcg_code_size(void) -{ - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); - unsigned int i; - size_t total; - - qemu_mutex_lock(&region.lock); - total = region.agg_size_full; - for (i = 0; i < n_ctxs; i++) { - const TCGContext *s = qatomic_read(&tcg_ctxs[i]); - size_t size; - - size = qatomic_read(&s->code_gen_ptr) - s->code_gen_buffer; - g_assert(size <= s->code_gen_buffer_size); - total += size; - } - qemu_mutex_unlock(&region.lock); - return total; -} - -/* - * Returns the code capacity (in bytes) of the entire cache, i.e. including all - * regions. - * See also: tcg_code_size() - */ -size_t tcg_code_capacity(void) -{ - size_t guard_size, capacity; - - /* no need for synchronization; these variables are set at init time */ - guard_size = region.stride - region.size; - capacity = region.end + guard_size - region.start; - capacity -= region.n * (guard_size + TCG_HIGHWATER); - return capacity; -} - -size_t tcg_tb_phys_invalidate_count(void) -{ - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); - unsigned int i; - size_t total = 0; - - for (i = 0; i < n_ctxs; i++) { - const TCGContext *s = qatomic_read(&tcg_ctxs[i]); - - total += qatomic_read(&s->tb_phys_invalidate_count); - } - return total; -} - /* pool based memory allocation */ void *tcg_malloc_internal(TCGContext *s, int size) { diff --git a/tcg/meson.build b/tcg/meson.build index XXXXXXX..XXXXXXX 100644 --- a/tcg/meson.build +++ b/tcg/meson.build @@ -XXX,XX +XXX,XX @@ tcg_ss = ss.source_set() tcg_ss.add(files( 'optimize.c', + 'region.c', 'tcg.c', 'tcg-common.c', 'tcg-op.c', -- 2.25.1 It consists of one function call and has only one caller. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/translate-all.c | 7 +------ 1 file changed, 1 insertion(+), 6 deletions(-) diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ static void page_table_config_init(void) assert(v_l2_levels >= 0); } -static void cpu_gen_init(void) -{ - tcg_context_init(&tcg_init_ctx); -} - /* Encode VAL as a signed leb128 sequence at P. Return P incremented past the encoded value. */ static uint8_t *encode_sleb128(uint8_t *p, target_long val) @@ -XXX,XX +XXX,XX @@ void tcg_exec_init(unsigned long tb_size, int splitwx) bool ok; tcg_allowed = true; - cpu_gen_init(); + tcg_context_init(&tcg_init_ctx); page_init(); tb_htable_init(); -- 2.25.1 Buffer management is integral to tcg. Do not leave the allocation to code outside of tcg/. This is code movement, with further cleanups to follow. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 2 +- accel/tcg/translate-all.c | 414 +----------------------------------- tcg/region.c | 431 +++++++++++++++++++++++++++++++++++++- 3 files changed, 428 insertions(+), 419 deletions(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ void *tcg_malloc_internal(TCGContext *s, int size); void tcg_pool_reset(TCGContext *s); TranslationBlock *tcg_tb_alloc(TCGContext *s); -void tcg_region_init(void); +void tcg_region_init(size_t tb_size, int splitwx); void tb_destroy(TranslationBlock *tb); void tcg_region_reset_all(void); diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ */ #include "qemu/osdep.h" -#include "qemu/units.h" #include "qemu-common.h" #define NO_CPU_IO_DEFS @@ -XXX,XX +XXX,XX @@ #include "exec/cputlb.h" #include "exec/translate-all.h" #include "qemu/bitmap.h" -#include "qemu/error-report.h" #include "qemu/qemu-print.h" #include "qemu/timer.h" #include "qemu/main-loop.h" @@ -XXX,XX +XXX,XX @@ static void page_lock_pair(PageDesc **ret_p1, tb_page_addr_t phys1, } } -/* Minimum size of the code gen buffer. This number is randomly chosen, - but not so small that we can't have a fair number of TB's live. */ -#define MIN_CODE_GEN_BUFFER_SIZE (1 * MiB) - -/* Maximum size of the code gen buffer we'd like to use. Unless otherwise - indicated, this is constrained by the range of direct branches on the - host cpu, as used by the TCG implementation of goto_tb. */ -#if defined(__x86_64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__sparc__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__powerpc64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__powerpc__) -# define MAX_CODE_GEN_BUFFER_SIZE (32 * MiB) -#elif defined(__aarch64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__s390x__) - /* We have a +- 4GB range on the branches; leave some slop. */ -# define MAX_CODE_GEN_BUFFER_SIZE (3 * GiB) -#elif defined(__mips__) - /* We have a 256MB branch region, but leave room to make sure the - main executable is also within that region. */ -# define MAX_CODE_GEN_BUFFER_SIZE (128 * MiB) -#else -# define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1) -#endif - -#if TCG_TARGET_REG_BITS == 32 -#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (32 * MiB) -#ifdef CONFIG_USER_ONLY -/* - * For user mode on smaller 32 bit systems we may run into trouble - * allocating big chunks of data in the right place. On these systems - * we utilise a static code generation buffer directly in the binary. - */ -#define USE_STATIC_CODE_GEN_BUFFER -#endif -#else /* TCG_TARGET_REG_BITS == 64 */ -#ifdef CONFIG_USER_ONLY -/* - * As user-mode emulation typically means running multiple instances - * of the translator don't go too nuts with our default code gen - * buffer lest we make things too hard for the OS. - */ -#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (128 * MiB) -#else -/* - * We expect most system emulation to run one or two guests per host. - * Users running large scale system emulation may want to tweak their - * runtime setup via the tb-size control on the command line. - */ -#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (1 * GiB) -#endif -#endif - -#define DEFAULT_CODE_GEN_BUFFER_SIZE \ - (DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \ - ? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE) - -static size_t size_code_gen_buffer(size_t tb_size) -{ - /* Size the buffer. */ - if (tb_size == 0) { - size_t phys_mem = qemu_get_host_physmem(); - if (phys_mem == 0) { - tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE; - } else { - tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, phys_mem / 8); - } - } - if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) { - tb_size = MIN_CODE_GEN_BUFFER_SIZE; - } - if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) { - tb_size = MAX_CODE_GEN_BUFFER_SIZE; - } - return tb_size; -} - -#ifdef __mips__ -/* In order to use J and JAL within the code_gen_buffer, we require - that the buffer not cross a 256MB boundary. */ -static inline bool cross_256mb(void *addr, size_t size) -{ - return ((uintptr_t)addr ^ ((uintptr_t)addr + size)) & ~0x0ffffffful; -} - -/* We weren't able to allocate a buffer without crossing that boundary, - so make do with the larger portion of the buffer that doesn't cross. - Returns the new base of the buffer, and adjusts code_gen_buffer_size. */ -static inline void *split_cross_256mb(void *buf1, size_t size1) -{ - void *buf2 = (void *)(((uintptr_t)buf1 + size1) & ~0x0ffffffful); - size_t size2 = buf1 + size1 - buf2; - - size1 = buf2 - buf1; - if (size1 < size2) { - size1 = size2; - buf1 = buf2; - } - - tcg_ctx->code_gen_buffer_size = size1; - return buf1; -} -#endif - -#ifdef USE_STATIC_CODE_GEN_BUFFER -static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE] - __attribute__((aligned(CODE_GEN_ALIGN))); - -static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) -{ - void *buf, *end; - size_t size; - - if (splitwx > 0) { - error_setg(errp, "jit split-wx not supported"); - return false; - } - - /* page-align the beginning and end of the buffer */ - buf = static_code_gen_buffer; - end = static_code_gen_buffer + sizeof(static_code_gen_buffer); - buf = QEMU_ALIGN_PTR_UP(buf, qemu_real_host_page_size); - end = QEMU_ALIGN_PTR_DOWN(end, qemu_real_host_page_size); - - size = end - buf; - - /* Honor a command-line option limiting the size of the buffer. */ - if (size > tb_size) { - size = QEMU_ALIGN_DOWN(tb_size, qemu_real_host_page_size); - } - tcg_ctx->code_gen_buffer_size = size; - -#ifdef __mips__ - if (cross_256mb(buf, size)) { - buf = split_cross_256mb(buf, size); - size = tcg_ctx->code_gen_buffer_size; - } -#endif - - if (qemu_mprotect_rwx(buf, size)) { - error_setg_errno(errp, errno, "mprotect of jit buffer"); - return false; - } - qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); - - tcg_ctx->code_gen_buffer = buf; - return true; -} -#elif defined(_WIN32) -static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) -{ - void *buf; - - if (splitwx > 0) { - error_setg(errp, "jit split-wx not supported"); - return false; - } - - buf = VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT, - PAGE_EXECUTE_READWRITE); - if (buf == NULL) { - error_setg_win32(errp, GetLastError(), - "allocate %zu bytes for jit buffer", size); - return false; - } - - tcg_ctx->code_gen_buffer = buf; - tcg_ctx->code_gen_buffer_size = size; - return true; -} -#else -static bool alloc_code_gen_buffer_anon(size_t size, int prot, - int flags, Error **errp) -{ - void *buf; - - buf = mmap(NULL, size, prot, flags, -1, 0); - if (buf == MAP_FAILED) { - error_setg_errno(errp, errno, - "allocate %zu bytes for jit buffer", size); - return false; - } - tcg_ctx->code_gen_buffer_size = size; - -#ifdef __mips__ - if (cross_256mb(buf, size)) { - /* - * Try again, with the original still mapped, to avoid re-acquiring - * the same 256mb crossing. - */ - size_t size2; - void *buf2 = mmap(NULL, size, prot, flags, -1, 0); - switch ((int)(buf2 != MAP_FAILED)) { - case 1: - if (!cross_256mb(buf2, size)) { - /* Success! Use the new buffer. */ - munmap(buf, size); - break; - } - /* Failure. Work with what we had. */ - munmap(buf2, size); - /* fallthru */ - default: - /* Split the original buffer. Free the smaller half. */ - buf2 = split_cross_256mb(buf, size); - size2 = tcg_ctx->code_gen_buffer_size; - if (buf == buf2) { - munmap(buf + size2, size - size2); - } else { - munmap(buf, size - size2); - } - size = size2; - break; - } - buf = buf2; - } -#endif - - /* Request large pages for the buffer. */ - qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); - - tcg_ctx->code_gen_buffer = buf; - return true; -} - -#ifndef CONFIG_TCG_INTERPRETER -#ifdef CONFIG_POSIX -#include "qemu/memfd.h" - -static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) -{ - void *buf_rw = NULL, *buf_rx = MAP_FAILED; - int fd = -1; - -#ifdef __mips__ - /* Find space for the RX mapping, vs the 256MiB regions. */ - if (!alloc_code_gen_buffer_anon(size, PROT_NONE, - MAP_PRIVATE | MAP_ANONYMOUS | - MAP_NORESERVE, errp)) { - return false; - } - /* The size of the mapping may have been adjusted. */ - size = tcg_ctx->code_gen_buffer_size; - buf_rx = tcg_ctx->code_gen_buffer; -#endif - - buf_rw = qemu_memfd_alloc("tcg-jit", size, 0, &fd, errp); - if (buf_rw == NULL) { - goto fail; - } - -#ifdef __mips__ - void *tmp = mmap(buf_rx, size, PROT_READ | PROT_EXEC, - MAP_SHARED | MAP_FIXED, fd, 0); - if (tmp != buf_rx) { - goto fail_rx; - } -#else - buf_rx = mmap(NULL, size, PROT_READ | PROT_EXEC, MAP_SHARED, fd, 0); - if (buf_rx == MAP_FAILED) { - goto fail_rx; - } -#endif - - close(fd); - tcg_ctx->code_gen_buffer = buf_rw; - tcg_ctx->code_gen_buffer_size = size; - tcg_splitwx_diff = buf_rx - buf_rw; - - /* Request large pages for the buffer and the splitwx. */ - qemu_madvise(buf_rw, size, QEMU_MADV_HUGEPAGE); - qemu_madvise(buf_rx, size, QEMU_MADV_HUGEPAGE); - return true; - - fail_rx: - error_setg_errno(errp, errno, "failed to map shared memory for execute"); - fail: - if (buf_rx != MAP_FAILED) { - munmap(buf_rx, size); - } - if (buf_rw) { - munmap(buf_rw, size); - } - if (fd >= 0) { - close(fd); - } - return false; -} -#endif /* CONFIG_POSIX */ - -#ifdef CONFIG_DARWIN -#include <mach/mach.h> - -extern kern_return_t mach_vm_remap(vm_map_t target_task, - mach_vm_address_t *target_address, - mach_vm_size_t size, - mach_vm_offset_t mask, - int flags, - vm_map_t src_task, - mach_vm_address_t src_address, - boolean_t copy, - vm_prot_t *cur_protection, - vm_prot_t *max_protection, - vm_inherit_t inheritance); - -static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) -{ - kern_return_t ret; - mach_vm_address_t buf_rw, buf_rx; - vm_prot_t cur_prot, max_prot; - - /* Map the read-write portion via normal anon memory. */ - if (!alloc_code_gen_buffer_anon(size, PROT_READ | PROT_WRITE, - MAP_PRIVATE | MAP_ANONYMOUS, errp)) { - return false; - } - - buf_rw = (mach_vm_address_t)tcg_ctx->code_gen_buffer; - buf_rx = 0; - ret = mach_vm_remap(mach_task_self(), - &buf_rx, - size, - 0, - VM_FLAGS_ANYWHERE, - mach_task_self(), - buf_rw, - false, - &cur_prot, - &max_prot, - VM_INHERIT_NONE); - if (ret != KERN_SUCCESS) { - /* TODO: Convert "ret" to a human readable error message. */ - error_setg(errp, "vm_remap for jit splitwx failed"); - munmap((void *)buf_rw, size); - return false; - } - - if (mprotect((void *)buf_rx, size, PROT_READ | PROT_EXEC) != 0) { - error_setg_errno(errp, errno, "mprotect for jit splitwx"); - munmap((void *)buf_rx, size); - munmap((void *)buf_rw, size); - return false; - } - - tcg_splitwx_diff = buf_rx - buf_rw; - return true; -} -#endif /* CONFIG_DARWIN */ -#endif /* CONFIG_TCG_INTERPRETER */ - -static bool alloc_code_gen_buffer_splitwx(size_t size, Error **errp) -{ -#ifndef CONFIG_TCG_INTERPRETER -# ifdef CONFIG_DARWIN - return alloc_code_gen_buffer_splitwx_vmremap(size, errp); -# endif -# ifdef CONFIG_POSIX - return alloc_code_gen_buffer_splitwx_memfd(size, errp); -# endif -#endif - error_setg(errp, "jit split-wx not supported"); - return false; -} - -static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) -{ - ERRP_GUARD(); - int prot, flags; - - if (splitwx) { - if (alloc_code_gen_buffer_splitwx(size, errp)) { - return true; - } - /* - * If splitwx force-on (1), fail; - * if splitwx default-on (-1), fall through to splitwx off. - */ - if (splitwx > 0) { - return false; - } - error_free_or_abort(errp); - } - - prot = PROT_READ | PROT_WRITE | PROT_EXEC; - flags = MAP_PRIVATE | MAP_ANONYMOUS; -#ifdef CONFIG_TCG_INTERPRETER - /* The tcg interpreter does not need execute permission. */ - prot = PROT_READ | PROT_WRITE; -#elif defined(CONFIG_DARWIN) - /* Applicable to both iOS and macOS (Apple Silicon). */ - if (!splitwx) { - flags |= MAP_JIT; - } -#endif - - return alloc_code_gen_buffer_anon(size, prot, flags, errp); -} -#endif /* USE_STATIC_CODE_GEN_BUFFER, WIN32, POSIX */ - static bool tb_cmp(const void *ap, const void *bp) { const TranslationBlock *a = ap; @@ -XXX,XX +XXX,XX @@ static void tb_htable_init(void) size. */ void tcg_exec_init(unsigned long tb_size, int splitwx) { - bool ok; - tcg_allowed = true; tcg_context_init(&tcg_init_ctx); page_init(); tb_htable_init(); - - ok = alloc_code_gen_buffer(size_code_gen_buffer(tb_size), - splitwx, &error_fatal); - assert(ok); - - /* TODO: allocating regions is hand-in-glove with code_gen_buffer. */ - tcg_region_init(); + tcg_region_init(tb_size, splitwx); #if defined(CONFIG_SOFTMMU) /* There's no guest base to take into account, so go ahead and diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ */ #include "qemu/osdep.h" +#include "qemu/units.h" +#include "qapi/error.h" #include "exec/exec-all.h" #include "tcg/tcg.h" #if !defined(CONFIG_USER_ONLY) @@ -XXX,XX +XXX,XX @@ static size_t tcg_n_regions(void) } #endif +/* + * Minimum size of the code gen buffer. This number is randomly chosen, + * but not so small that we can't have a fair number of TB's live. + */ +#define MIN_CODE_GEN_BUFFER_SIZE (1 * MiB) + +/* + * Maximum size of the code gen buffer we'd like to use. Unless otherwise + * indicated, this is constrained by the range of direct branches on the + * host cpu, as used by the TCG implementation of goto_tb. + */ +#if defined(__x86_64__) +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) +#elif defined(__sparc__) +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) +#elif defined(__powerpc64__) +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) +#elif defined(__powerpc__) +# define MAX_CODE_GEN_BUFFER_SIZE (32 * MiB) +#elif defined(__aarch64__) +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) +#elif defined(__s390x__) + /* We have a +- 4GB range on the branches; leave some slop. */ +# define MAX_CODE_GEN_BUFFER_SIZE (3 * GiB) +#elif defined(__mips__) + /* + * We have a 256MB branch region, but leave room to make sure the + * main executable is also within that region. + */ +# define MAX_CODE_GEN_BUFFER_SIZE (128 * MiB) +#else +# define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1) +#endif + +#if TCG_TARGET_REG_BITS == 32 +#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (32 * MiB) +#ifdef CONFIG_USER_ONLY +/* + * For user mode on smaller 32 bit systems we may run into trouble + * allocating big chunks of data in the right place. On these systems + * we utilise a static code generation buffer directly in the binary. + */ +#define USE_STATIC_CODE_GEN_BUFFER +#endif +#else /* TCG_TARGET_REG_BITS == 64 */ +#ifdef CONFIG_USER_ONLY +/* + * As user-mode emulation typically means running multiple instances + * of the translator don't go too nuts with our default code gen + * buffer lest we make things too hard for the OS. + */ +#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (128 * MiB) +#else +/* + * We expect most system emulation to run one or two guests per host. + * Users running large scale system emulation may want to tweak their + * runtime setup via the tb-size control on the command line. + */ +#define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (1 * GiB) +#endif +#endif + +#define DEFAULT_CODE_GEN_BUFFER_SIZE \ + (DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \ + ? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE) + +static size_t size_code_gen_buffer(size_t tb_size) +{ + /* Size the buffer. */ + if (tb_size == 0) { + size_t phys_mem = qemu_get_host_physmem(); + if (phys_mem == 0) { + tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE; + } else { + tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, phys_mem / 8); + } + } + if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) { + tb_size = MIN_CODE_GEN_BUFFER_SIZE; + } + if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) { + tb_size = MAX_CODE_GEN_BUFFER_SIZE; + } + return tb_size; +} + +#ifdef __mips__ +/* + * In order to use J and JAL within the code_gen_buffer, we require + * that the buffer not cross a 256MB boundary. + */ +static inline bool cross_256mb(void *addr, size_t size) +{ + return ((uintptr_t)addr ^ ((uintptr_t)addr + size)) & ~0x0ffffffful; +} + +/* + * We weren't able to allocate a buffer without crossing that boundary, + * so make do with the larger portion of the buffer that doesn't cross. + * Returns the new base of the buffer, and adjusts code_gen_buffer_size. + */ +static inline void *split_cross_256mb(void *buf1, size_t size1) +{ + void *buf2 = (void *)(((uintptr_t)buf1 + size1) & ~0x0ffffffful); + size_t size2 = buf1 + size1 - buf2; + + size1 = buf2 - buf1; + if (size1 < size2) { + size1 = size2; + buf1 = buf2; + } + + tcg_ctx->code_gen_buffer_size = size1; + return buf1; +} +#endif + +#ifdef USE_STATIC_CODE_GEN_BUFFER +static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE] + __attribute__((aligned(CODE_GEN_ALIGN))); + +static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) +{ + void *buf, *end; + size_t size; + + if (splitwx > 0) { + error_setg(errp, "jit split-wx not supported"); + return false; + } + + /* page-align the beginning and end of the buffer */ + buf = static_code_gen_buffer; + end = static_code_gen_buffer + sizeof(static_code_gen_buffer); + buf = QEMU_ALIGN_PTR_UP(buf, qemu_real_host_page_size); + end = QEMU_ALIGN_PTR_DOWN(end, qemu_real_host_page_size); + + size = end - buf; + + /* Honor a command-line option limiting the size of the buffer. */ + if (size > tb_size) { + size = QEMU_ALIGN_DOWN(tb_size, qemu_real_host_page_size); + } + tcg_ctx->code_gen_buffer_size = size; + +#ifdef __mips__ + if (cross_256mb(buf, size)) { + buf = split_cross_256mb(buf, size); + size = tcg_ctx->code_gen_buffer_size; + } +#endif + + if (qemu_mprotect_rwx(buf, size)) { + error_setg_errno(errp, errno, "mprotect of jit buffer"); + return false; + } + qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); + + tcg_ctx->code_gen_buffer = buf; + return true; +} +#elif defined(_WIN32) +static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) +{ + void *buf; + + if (splitwx > 0) { + error_setg(errp, "jit split-wx not supported"); + return false; + } + + buf = VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT, + PAGE_EXECUTE_READWRITE); + if (buf == NULL) { + error_setg_win32(errp, GetLastError(), + "allocate %zu bytes for jit buffer", size); + return false; + } + + tcg_ctx->code_gen_buffer = buf; + tcg_ctx->code_gen_buffer_size = size; + return true; +} +#else +static bool alloc_code_gen_buffer_anon(size_t size, int prot, + int flags, Error **errp) +{ + void *buf; + + buf = mmap(NULL, size, prot, flags, -1, 0); + if (buf == MAP_FAILED) { + error_setg_errno(errp, errno, + "allocate %zu bytes for jit buffer", size); + return false; + } + tcg_ctx->code_gen_buffer_size = size; + +#ifdef __mips__ + if (cross_256mb(buf, size)) { + /* + * Try again, with the original still mapped, to avoid re-acquiring + * the same 256mb crossing. + */ + size_t size2; + void *buf2 = mmap(NULL, size, prot, flags, -1, 0); + switch ((int)(buf2 != MAP_FAILED)) { + case 1: + if (!cross_256mb(buf2, size)) { + /* Success! Use the new buffer. */ + munmap(buf, size); + break; + } + /* Failure. Work with what we had. */ + munmap(buf2, size); + /* fallthru */ + default: + /* Split the original buffer. Free the smaller half. */ + buf2 = split_cross_256mb(buf, size); + size2 = tcg_ctx->code_gen_buffer_size; + if (buf == buf2) { + munmap(buf + size2, size - size2); + } else { + munmap(buf, size - size2); + } + size = size2; + break; + } + buf = buf2; + } +#endif + + /* Request large pages for the buffer. */ + qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); + + tcg_ctx->code_gen_buffer = buf; + return true; +} + +#ifndef CONFIG_TCG_INTERPRETER +#ifdef CONFIG_POSIX +#include "qemu/memfd.h" + +static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) +{ + void *buf_rw = NULL, *buf_rx = MAP_FAILED; + int fd = -1; + +#ifdef __mips__ + /* Find space for the RX mapping, vs the 256MiB regions. */ + if (!alloc_code_gen_buffer_anon(size, PROT_NONE, + MAP_PRIVATE | MAP_ANONYMOUS | + MAP_NORESERVE, errp)) { + return false; + } + /* The size of the mapping may have been adjusted. */ + size = tcg_ctx->code_gen_buffer_size; + buf_rx = tcg_ctx->code_gen_buffer; +#endif + + buf_rw = qemu_memfd_alloc("tcg-jit", size, 0, &fd, errp); + if (buf_rw == NULL) { + goto fail; + } + +#ifdef __mips__ + void *tmp = mmap(buf_rx, size, PROT_READ | PROT_EXEC, + MAP_SHARED | MAP_FIXED, fd, 0); + if (tmp != buf_rx) { + goto fail_rx; + } +#else + buf_rx = mmap(NULL, size, PROT_READ | PROT_EXEC, MAP_SHARED, fd, 0); + if (buf_rx == MAP_FAILED) { + goto fail_rx; + } +#endif + + close(fd); + tcg_ctx->code_gen_buffer = buf_rw; + tcg_ctx->code_gen_buffer_size = size; + tcg_splitwx_diff = buf_rx - buf_rw; + + /* Request large pages for the buffer and the splitwx. */ + qemu_madvise(buf_rw, size, QEMU_MADV_HUGEPAGE); + qemu_madvise(buf_rx, size, QEMU_MADV_HUGEPAGE); + return true; + + fail_rx: + error_setg_errno(errp, errno, "failed to map shared memory for execute"); + fail: + if (buf_rx != MAP_FAILED) { + munmap(buf_rx, size); + } + if (buf_rw) { + munmap(buf_rw, size); + } + if (fd >= 0) { + close(fd); + } + return false; +} +#endif /* CONFIG_POSIX */ + +#ifdef CONFIG_DARWIN +#include <mach/mach.h> + +extern kern_return_t mach_vm_remap(vm_map_t target_task, + mach_vm_address_t *target_address, + mach_vm_size_t size, + mach_vm_offset_t mask, + int flags, + vm_map_t src_task, + mach_vm_address_t src_address, + boolean_t copy, + vm_prot_t *cur_protection, + vm_prot_t *max_protection, + vm_inherit_t inheritance); + +static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) +{ + kern_return_t ret; + mach_vm_address_t buf_rw, buf_rx; + vm_prot_t cur_prot, max_prot; + + /* Map the read-write portion via normal anon memory. */ + if (!alloc_code_gen_buffer_anon(size, PROT_READ | PROT_WRITE, + MAP_PRIVATE | MAP_ANONYMOUS, errp)) { + return false; + } + + buf_rw = (mach_vm_address_t)tcg_ctx->code_gen_buffer; + buf_rx = 0; + ret = mach_vm_remap(mach_task_self(), + &buf_rx, + size, + 0, + VM_FLAGS_ANYWHERE, + mach_task_self(), + buf_rw, + false, + &cur_prot, + &max_prot, + VM_INHERIT_NONE); + if (ret != KERN_SUCCESS) { + /* TODO: Convert "ret" to a human readable error message. */ + error_setg(errp, "vm_remap for jit splitwx failed"); + munmap((void *)buf_rw, size); + return false; + } + + if (mprotect((void *)buf_rx, size, PROT_READ | PROT_EXEC) != 0) { + error_setg_errno(errp, errno, "mprotect for jit splitwx"); + munmap((void *)buf_rx, size); + munmap((void *)buf_rw, size); + return false; + } + + tcg_splitwx_diff = buf_rx - buf_rw; + return true; +} +#endif /* CONFIG_DARWIN */ +#endif /* CONFIG_TCG_INTERPRETER */ + +static bool alloc_code_gen_buffer_splitwx(size_t size, Error **errp) +{ +#ifndef CONFIG_TCG_INTERPRETER +# ifdef CONFIG_DARWIN + return alloc_code_gen_buffer_splitwx_vmremap(size, errp); +# endif +# ifdef CONFIG_POSIX + return alloc_code_gen_buffer_splitwx_memfd(size, errp); +# endif +#endif + error_setg(errp, "jit split-wx not supported"); + return false; +} + +static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) +{ + ERRP_GUARD(); + int prot, flags; + + if (splitwx) { + if (alloc_code_gen_buffer_splitwx(size, errp)) { + return true; + } + /* + * If splitwx force-on (1), fail; + * if splitwx default-on (-1), fall through to splitwx off. + */ + if (splitwx > 0) { + return false; + } + error_free_or_abort(errp); + } + + prot = PROT_READ | PROT_WRITE | PROT_EXEC; + flags = MAP_PRIVATE | MAP_ANONYMOUS; +#ifdef CONFIG_TCG_INTERPRETER + /* The tcg interpreter does not need execute permission. */ + prot = PROT_READ | PROT_WRITE; +#elif defined(CONFIG_DARWIN) + /* Applicable to both iOS and macOS (Apple Silicon). */ + if (!splitwx) { + flags |= MAP_JIT; + } +#endif + + return alloc_code_gen_buffer_anon(size, prot, flags, errp); +} +#endif /* USE_STATIC_CODE_GEN_BUFFER, WIN32, POSIX */ + /* * Initializes region partitioning. * @@ -XXX,XX +XXX,XX @@ static size_t tcg_n_regions(void) * in practice. Multi-threaded guests share most if not all of their translated * code, which makes parallel code generation less appealing than in softmmu. */ -void tcg_region_init(void) +void tcg_region_init(size_t tb_size, int splitwx) { - void *buf = tcg_init_ctx.code_gen_buffer; - void *aligned; - size_t size = tcg_init_ctx.code_gen_buffer_size; - size_t page_size = qemu_real_host_page_size; + void *buf, *aligned; + size_t size; + size_t page_size; size_t region_size; size_t n_regions; size_t i; + bool ok; + ok = alloc_code_gen_buffer(size_code_gen_buffer(tb_size), + splitwx, &error_fatal); + assert(ok); + + buf = tcg_init_ctx.code_gen_buffer; + size = tcg_init_ctx.code_gen_buffer_size; + page_size = qemu_real_host_page_size; n_regions = tcg_n_regions(); /* The first region will be 'aligned - buf' bytes larger than the others */ -- 2.25.1 We shortly want to use tcg_init for something else. Since the hook is called init_machine, match that. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/tcg-all.c | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/tcg-all.c +++ b/accel/tcg/tcg-all.c @@ -XXX,XX +XXX,XX @@ static void tcg_accel_instance_init(Object *obj) bool mttcg_enabled; -static int tcg_init(MachineState *ms) +static int tcg_init_machine(MachineState *ms) { TCGState *s = TCG_STATE(current_accel()); @@ -XXX,XX +XXX,XX @@ static void tcg_accel_class_init(ObjectClass *oc, void *data) { AccelClass *ac = ACCEL_CLASS(oc); ac->name = "tcg"; - ac->init_machine = tcg_init; + ac->init_machine = tcg_init_machine; ac->allowed = &tcg_allowed; object_class_property_add_str(oc, "thread", -- 2.25.1 Perform both tcg_context_init and tcg_region_init. Do not leave this split to the caller. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 3 +-- tcg/tcg-internal.h | 1 + accel/tcg/translate-all.c | 3 +-- tcg/tcg.c | 9 ++++++++- 4 files changed, 11 insertions(+), 5 deletions(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ void *tcg_malloc_internal(TCGContext *s, int size); void tcg_pool_reset(TCGContext *s); TranslationBlock *tcg_tb_alloc(TCGContext *s); -void tcg_region_init(size_t tb_size, int splitwx); void tb_destroy(TranslationBlock *tb); void tcg_region_reset_all(void); @@ -XXX,XX +XXX,XX @@ static inline void *tcg_malloc(int size) } } -void tcg_context_init(TCGContext *s); +void tcg_init(size_t tb_size, int splitwx); void tcg_register_thread(void); void tcg_prologue_init(TCGContext *s); void tcg_func_start(TCGContext *s); diff --git a/tcg/tcg-internal.h b/tcg/tcg-internal.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg-internal.h +++ b/tcg/tcg-internal.h @@ -XXX,XX +XXX,XX @@ extern TCGContext **tcg_ctxs; extern unsigned int n_tcg_ctxs; +void tcg_region_init(size_t tb_size, int splitwx); bool tcg_region_alloc(TCGContext *s); void tcg_region_initial_alloc(TCGContext *s); void tcg_region_prologue_set(TCGContext *s); diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ static void tb_htable_init(void) void tcg_exec_init(unsigned long tb_size, int splitwx) { tcg_allowed = true; - tcg_context_init(&tcg_init_ctx); page_init(); tb_htable_init(); - tcg_region_init(tb_size, splitwx); + tcg_init(tb_size, splitwx); #if defined(CONFIG_SOFTMMU) /* There's no guest base to take into account, so go ahead and diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static void process_op_defs(TCGContext *s); static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, TCGReg reg, const char *name); -void tcg_context_init(TCGContext *s) +static void tcg_context_init(void) { + TCGContext *s = &tcg_init_ctx; int op, total_args, n, i; TCGOpDef *def; TCGArgConstraint *args_ct; @@ -XXX,XX +XXX,XX @@ void tcg_context_init(TCGContext *s) cpu_env = temp_tcgv_ptr(ts); } +void tcg_init(size_t tb_size, int splitwx) +{ + tcg_context_init(); + tcg_region_init(tb_size, splitwx); +} + /* * Allocate TBs right before their corresponding translated code, making * sure that TBs and code are on different cache lines. -- 2.25.1 There is only one caller, and shortly we will need access to the MachineState, which tcg_init_machine already has. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/internal.h | 2 ++ include/sysemu/tcg.h | 2 -- accel/tcg/tcg-all.c | 16 +++++++++++++++- accel/tcg/translate-all.c | 21 ++------------------- bsd-user/main.c | 2 +- 5 files changed, 20 insertions(+), 23 deletions(-) diff --git a/accel/tcg/internal.h b/accel/tcg/internal.h index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/internal.h +++ b/accel/tcg/internal.h @@ -XXX,XX +XXX,XX @@ TranslationBlock *tb_gen_code(CPUState *cpu, target_ulong pc, int cflags); void QEMU_NORETURN cpu_io_recompile(CPUState *cpu, uintptr_t retaddr); +void page_init(void); +void tb_htable_init(void); #endif /* ACCEL_TCG_INTERNAL_H */ diff --git a/include/sysemu/tcg.h b/include/sysemu/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/sysemu/tcg.h +++ b/include/sysemu/tcg.h @@ -XXX,XX +XXX,XX @@ #ifndef SYSEMU_TCG_H #define SYSEMU_TCG_H -void tcg_exec_init(unsigned long tb_size, int splitwx); - #ifdef CONFIG_TCG extern bool tcg_allowed; #define tcg_enabled() (tcg_allowed) diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/tcg-all.c +++ b/accel/tcg/tcg-all.c @@ -XXX,XX +XXX,XX @@ #include "qemu/error-report.h" #include "qemu/accel.h" #include "qapi/qapi-builtin-visit.h" +#include "internal.h" struct TCGState { AccelState parent_obj; @@ -XXX,XX +XXX,XX @@ static int tcg_init_machine(MachineState *ms) { TCGState *s = TCG_STATE(current_accel()); - tcg_exec_init(s->tb_size * 1024 * 1024, s->splitwx_enabled); + tcg_allowed = true; mttcg_enabled = s->mttcg_enabled; + + page_init(); + tb_htable_init(); + tcg_init(s->tb_size * 1024 * 1024, s->splitwx_enabled); + +#if defined(CONFIG_SOFTMMU) + /* + * There's no guest base to take into account, so go ahead and + * initialize the prologue now. + */ + tcg_prologue_init(tcg_ctx); +#endif + return 0; } diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ bool cpu_restore_state(CPUState *cpu, uintptr_t host_pc, bool will_exit) return false; } -static void page_init(void) +void page_init(void) { page_size_init(); page_table_config_init(); @@ -XXX,XX +XXX,XX @@ static bool tb_cmp(const void *ap, const void *bp) a->page_addr[1] == b->page_addr[1]; } -static void tb_htable_init(void) +void tb_htable_init(void) { unsigned int mode = QHT_MODE_AUTO_RESIZE; qht_init(&tb_ctx.htable, tb_cmp, CODE_GEN_HTABLE_SIZE, mode); } -/* Must be called before using the QEMU cpus. 'tb_size' is the size - (in bytes) allocated to the translation buffer. Zero means default - size. */ -void tcg_exec_init(unsigned long tb_size, int splitwx) -{ - tcg_allowed = true; - page_init(); - tb_htable_init(); - tcg_init(tb_size, splitwx); - -#if defined(CONFIG_SOFTMMU) - /* There's no guest base to take into account, so go ahead and - initialize the prologue now. */ - tcg_prologue_init(tcg_ctx); -#endif -} - /* call with @p->lock held */ static inline void invalidate_page_bitmap(PageDesc *p) { diff --git a/bsd-user/main.c b/bsd-user/main.c index XXXXXXX..XXXXXXX 100644 --- a/bsd-user/main.c +++ b/bsd-user/main.c @@ -XXX,XX +XXX,XX @@ int main(int argc, char **argv) envlist_free(envlist); /* - * Now that page sizes are configured in tcg_exec_init() we can do + * Now that page sizes are configured we can do * proper page alignment for guest_base. */ guest_base = HOST_PAGE_ALIGN(guest_base); -- 2.25.1 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/tcg-all.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/tcg-all.c +++ b/accel/tcg/tcg-all.c @@ -XXX,XX +XXX,XX @@ #include "qemu/error-report.h" #include "qemu/accel.h" #include "qapi/qapi-builtin-visit.h" +#include "qemu/units.h" #include "internal.h" struct TCGState { @@ -XXX,XX +XXX,XX @@ static int tcg_init_machine(MachineState *ms) page_init(); tb_htable_init(); - tcg_init(s->tb_size * 1024 * 1024, s->splitwx_enabled); + tcg_init(s->tb_size * MiB, s->splitwx_enabled); #if defined(CONFIG_SOFTMMU) /* -- 2.25.1 Start removing the include of hw/boards.h from tcg/. Pass down the max_cpus value from tcg_init_machine, where we have the MachineState already. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 2 +- tcg/tcg-internal.h | 2 +- accel/tcg/tcg-all.c | 10 +++++++++- tcg/region.c | 32 +++++++++++--------------------- tcg/tcg.c | 10 ++++------ 5 files changed, 26 insertions(+), 30 deletions(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ static inline void *tcg_malloc(int size) } } -void tcg_init(size_t tb_size, int splitwx); +void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus); void tcg_register_thread(void); void tcg_prologue_init(TCGContext *s); void tcg_func_start(TCGContext *s); diff --git a/tcg/tcg-internal.h b/tcg/tcg-internal.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg-internal.h +++ b/tcg/tcg-internal.h @@ -XXX,XX +XXX,XX @@ extern TCGContext **tcg_ctxs; extern unsigned int n_tcg_ctxs; -void tcg_region_init(size_t tb_size, int splitwx); +void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus); bool tcg_region_alloc(TCGContext *s); void tcg_region_initial_alloc(TCGContext *s); void tcg_region_prologue_set(TCGContext *s); diff --git a/accel/tcg/tcg-all.c b/accel/tcg/tcg-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/tcg-all.c +++ b/accel/tcg/tcg-all.c @@ -XXX,XX +XXX,XX @@ #include "qemu/accel.h" #include "qapi/qapi-builtin-visit.h" #include "qemu/units.h" +#if !defined(CONFIG_USER_ONLY) +#include "hw/boards.h" +#endif #include "internal.h" struct TCGState { @@ -XXX,XX +XXX,XX @@ bool mttcg_enabled; static int tcg_init_machine(MachineState *ms) { TCGState *s = TCG_STATE(current_accel()); +#ifdef CONFIG_USER_ONLY + unsigned max_cpus = 1; +#else + unsigned max_cpus = ms->smp.max_cpus; +#endif tcg_allowed = true; mttcg_enabled = s->mttcg_enabled; page_init(); tb_htable_init(); - tcg_init(s->tb_size * MiB, s->splitwx_enabled); + tcg_init(s->tb_size * MiB, s->splitwx_enabled, max_cpus); #if defined(CONFIG_SOFTMMU) /* diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ #include "qapi/error.h" #include "exec/exec-all.h" #include "tcg/tcg.h" -#if !defined(CONFIG_USER_ONLY) -#include "hw/boards.h" -#endif #include "tcg-internal.h" @@ -XXX,XX +XXX,XX @@ void tcg_region_reset_all(void) tcg_region_tree_reset_all(); } +static size_t tcg_n_regions(unsigned max_cpus) +{ #ifdef CONFIG_USER_ONLY -static size_t tcg_n_regions(void) -{ return 1; -} #else -/* - * It is likely that some vCPUs will translate more code than others, so we - * first try to set more regions than max_cpus, with those regions being of - * reasonable size. If that's not possible we make do by evenly dividing - * the code_gen_buffer among the vCPUs. - */ -static size_t tcg_n_regions(void) -{ + /* + * It is likely that some vCPUs will translate more code than others, + * so we first try to set more regions than max_cpus, with those regions + * being of reasonable size. If that's not possible we make do by evenly + * dividing the code_gen_buffer among the vCPUs. + */ size_t i; /* Use a single region if all we have is one vCPU thread */ -#if !defined(CONFIG_USER_ONLY) - MachineState *ms = MACHINE(qdev_get_machine()); - unsigned int max_cpus = ms->smp.max_cpus; -#endif if (max_cpus == 1 || !qemu_tcg_mttcg_enabled()) { return 1; } @@ -XXX,XX +XXX,XX @@ static size_t tcg_n_regions(void) } /* If we can't, then just allocate one region per vCPU thread */ return max_cpus; -} #endif +} /* * Minimum size of the code gen buffer. This number is randomly chosen, @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) * in practice. Multi-threaded guests share most if not all of their translated * code, which makes parallel code generation less appealing than in softmmu. */ -void tcg_region_init(size_t tb_size, int splitwx) +void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) { void *buf, *aligned; size_t size; @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx) buf = tcg_init_ctx.code_gen_buffer; size = tcg_init_ctx.code_gen_buffer_size; page_size = qemu_real_host_page_size; - n_regions = tcg_n_regions(); + n_regions = tcg_n_regions(max_cpus); /* The first region will be 'aligned - buf' bytes larger than the others */ aligned = QEMU_ALIGN_PTR_UP(buf, page_size); diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static void process_op_defs(TCGContext *s); static TCGTemp *tcg_global_reg_new_internal(TCGContext *s, TCGType type, TCGReg reg, const char *name); -static void tcg_context_init(void) +static void tcg_context_init(unsigned max_cpus) { TCGContext *s = &tcg_init_ctx; int op, total_args, n, i; @@ -XXX,XX +XXX,XX @@ static void tcg_context_init(void) tcg_ctxs = &tcg_ctx; n_tcg_ctxs = 1; #else - MachineState *ms = MACHINE(qdev_get_machine()); - unsigned int max_cpus = ms->smp.max_cpus; tcg_ctxs = g_new(TCGContext *, max_cpus); #endif @@ -XXX,XX +XXX,XX @@ static void tcg_context_init(void) cpu_env = temp_tcgv_ptr(ts); } -void tcg_init(size_t tb_size, int splitwx) +void tcg_init(size_t tb_size, int splitwx, unsigned max_cpus) { - tcg_context_init(); - tcg_region_init(tb_size, splitwx); + tcg_context_init(max_cpus); + tcg_region_init(tb_size, splitwx, max_cpus); } /* -- 2.25.1 Finish the divorce of tcg/ from hw/, and do not take the max cpu value from MachineState; just remember what we were passed in tcg_init. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/tcg-internal.h | 3 ++- tcg/region.c | 6 +++--- tcg/tcg.c | 23 ++++++++++------------- 3 files changed, 15 insertions(+), 17 deletions(-) diff --git a/tcg/tcg-internal.h b/tcg/tcg-internal.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg-internal.h +++ b/tcg/tcg-internal.h @@ -XXX,XX +XXX,XX @@ #define TCG_HIGHWATER 1024 extern TCGContext **tcg_ctxs; -extern unsigned int n_tcg_ctxs; +extern unsigned int tcg_cur_ctxs; +extern unsigned int tcg_max_ctxs; void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus); bool tcg_region_alloc(TCGContext *s); diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ void tcg_region_initial_alloc(TCGContext *s) /* Call from a safe-work context */ void tcg_region_reset_all(void) { - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); unsigned int i; qemu_mutex_lock(&region.lock); @@ -XXX,XX +XXX,XX @@ void tcg_region_prologue_set(TCGContext *s) */ size_t tcg_code_size(void) { - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); unsigned int i; size_t total; @@ -XXX,XX +XXX,XX @@ size_t tcg_code_capacity(void) size_t tcg_tb_phys_invalidate_count(void) { - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); unsigned int i; size_t total = 0; diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ #define NO_CPU_IO_DEFS #include "exec/exec-all.h" - -#if !defined(CONFIG_USER_ONLY) -#include "hw/boards.h" -#endif - #include "tcg/tcg-op.h" #if UINTPTR_MAX == UINT32_MAX @@ -XXX,XX +XXX,XX @@ static int tcg_out_ldst_finalize(TCGContext *s); #endif TCGContext **tcg_ctxs; -unsigned int n_tcg_ctxs; +unsigned int tcg_cur_ctxs; +unsigned int tcg_max_ctxs; TCGv_env cpu_env = 0; const void *tcg_code_gen_epilogue; uintptr_t tcg_splitwx_diff; @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) #else void tcg_register_thread(void) { - MachineState *ms = MACHINE(qdev_get_machine()); TCGContext *s = g_malloc(sizeof(*s)); unsigned int i, n; @@ -XXX,XX +XXX,XX @@ void tcg_register_thread(void) } /* Claim an entry in tcg_ctxs */ - n = qatomic_fetch_inc(&n_tcg_ctxs); - g_assert(n < ms->smp.max_cpus); + n = qatomic_fetch_inc(&tcg_cur_ctxs); + g_assert(n < tcg_max_ctxs); qatomic_set(&tcg_ctxs[n], s); if (n > 0) { @@ -XXX,XX +XXX,XX @@ static void tcg_context_init(unsigned max_cpus) */ #ifdef CONFIG_USER_ONLY tcg_ctxs = &tcg_ctx; - n_tcg_ctxs = 1; + tcg_cur_ctxs = 1; + tcg_max_ctxs = 1; #else - tcg_ctxs = g_new(TCGContext *, max_cpus); + tcg_max_ctxs = max_cpus; + tcg_ctxs = g_new0(TCGContext *, max_cpus); #endif tcg_debug_assert(!tcg_regset_test_reg(s->reserved_regs, TCG_AREG0)); @@ -XXX,XX +XXX,XX @@ static void tcg_reg_alloc_call(TCGContext *s, TCGOp *op) static inline void tcg_profile_snapshot(TCGProfile *prof, bool counters, bool table) { - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); unsigned int i; for (i = 0; i < n_ctxs; i++) { @@ -XXX,XX +XXX,XX @@ void tcg_dump_op_count(void) int64_t tcg_cpu_exec_time(void) { - unsigned int n_ctxs = qatomic_read(&n_tcg_ctxs); + unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); unsigned int i; int64_t ret = 0; -- 2.25.1 Remove the ifdef ladder and move each define into the appropriate header file. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/aarch64/tcg-target.h | 1 + tcg/arm/tcg-target.h | 1 + tcg/i386/tcg-target.h | 2 ++ tcg/mips/tcg-target.h | 6 ++++++ tcg/ppc/tcg-target.h | 2 ++ tcg/riscv/tcg-target.h | 1 + tcg/s390/tcg-target.h | 3 +++ tcg/sparc/tcg-target.h | 1 + tcg/tci/tcg-target.h | 1 + tcg/region.c | 33 +++++---------------------------- 10 files changed, 23 insertions(+), 28 deletions(-) diff --git a/tcg/aarch64/tcg-target.h b/tcg/aarch64/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/aarch64/tcg-target.h +++ b/tcg/aarch64/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_INSN_UNIT_SIZE 4 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 24 +#define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) #undef TCG_TARGET_STACK_GROWSUP typedef enum { diff --git a/tcg/arm/tcg-target.h b/tcg/arm/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/arm/tcg-target.h +++ b/tcg/arm/tcg-target.h @@ -XXX,XX +XXX,XX @@ extern int arm_arch; #undef TCG_TARGET_STACK_GROWSUP #define TCG_TARGET_INSN_UNIT_SIZE 4 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 16 +#define MAX_CODE_GEN_BUFFER_SIZE UINT32_MAX typedef enum { TCG_REG_R0 = 0, diff --git a/tcg/i386/tcg-target.h b/tcg/i386/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/i386/tcg-target.h +++ b/tcg/i386/tcg-target.h @@ -XXX,XX +XXX,XX @@ #ifdef __x86_64__ # define TCG_TARGET_REG_BITS 64 # define TCG_TARGET_NB_REGS 32 +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) #else # define TCG_TARGET_REG_BITS 32 # define TCG_TARGET_NB_REGS 24 +# define MAX_CODE_GEN_BUFFER_SIZE UINT32_MAX #endif typedef enum { diff --git a/tcg/mips/tcg-target.h b/tcg/mips/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/mips/tcg-target.h +++ b/tcg/mips/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_TLB_DISPLACEMENT_BITS 16 #define TCG_TARGET_NB_REGS 32 +/* + * We have a 256MB branch region, but leave room to make sure the + * main executable is also within that region. + */ +#define MAX_CODE_GEN_BUFFER_SIZE (128 * MiB) + typedef enum { TCG_REG_ZERO = 0, TCG_REG_AT, diff --git a/tcg/ppc/tcg-target.h b/tcg/ppc/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/ppc/tcg-target.h +++ b/tcg/ppc/tcg-target.h @@ -XXX,XX +XXX,XX @@ #ifdef _ARCH_PPC64 # define TCG_TARGET_REG_BITS 64 +# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) #else # define TCG_TARGET_REG_BITS 32 +# define MAX_CODE_GEN_BUFFER_SIZE (32 * MiB) #endif #define TCG_TARGET_NB_REGS 64 diff --git a/tcg/riscv/tcg-target.h b/tcg/riscv/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/riscv/tcg-target.h +++ b/tcg/riscv/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_INSN_UNIT_SIZE 4 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 20 #define TCG_TARGET_NB_REGS 32 +#define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1) typedef enum { TCG_REG_ZERO, diff --git a/tcg/s390/tcg-target.h b/tcg/s390/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/s390/tcg-target.h +++ b/tcg/s390/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_INSN_UNIT_SIZE 2 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 19 +/* We have a +- 4GB range on the branches; leave some slop. */ +#define MAX_CODE_GEN_BUFFER_SIZE (3 * GiB) + typedef enum TCGReg { TCG_REG_R0 = 0, TCG_REG_R1, diff --git a/tcg/sparc/tcg-target.h b/tcg/sparc/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/sparc/tcg-target.h +++ b/tcg/sparc/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_INSN_UNIT_SIZE 4 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 32 #define TCG_TARGET_NB_REGS 32 +#define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) typedef enum { TCG_REG_G0 = 0, diff --git a/tcg/tci/tcg-target.h b/tcg/tci/tcg-target.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/tci/tcg-target.h +++ b/tcg/tci/tcg-target.h @@ -XXX,XX +XXX,XX @@ #define TCG_TARGET_INTERPRETER 1 #define TCG_TARGET_INSN_UNIT_SIZE 1 #define TCG_TARGET_TLB_DISPLACEMENT_BITS 32 +#define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1) #if UINTPTR_MAX == UINT32_MAX # define TCG_TARGET_REG_BITS 32 diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static size_t tcg_n_regions(unsigned max_cpus) /* * Minimum size of the code gen buffer. This number is randomly chosen, * but not so small that we can't have a fair number of TB's live. + * + * Maximum size, MAX_CODE_GEN_BUFFER_SIZE, is defined in tcg-target.h. + * Unless otherwise indicated, this is constrained by the range of + * direct branches on the host cpu, as used by the TCG implementation + * of goto_tb. */ #define MIN_CODE_GEN_BUFFER_SIZE (1 * MiB) -/* - * Maximum size of the code gen buffer we'd like to use. Unless otherwise - * indicated, this is constrained by the range of direct branches on the - * host cpu, as used by the TCG implementation of goto_tb. - */ -#if defined(__x86_64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__sparc__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__powerpc64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__powerpc__) -# define MAX_CODE_GEN_BUFFER_SIZE (32 * MiB) -#elif defined(__aarch64__) -# define MAX_CODE_GEN_BUFFER_SIZE (2 * GiB) -#elif defined(__s390x__) - /* We have a +- 4GB range on the branches; leave some slop. */ -# define MAX_CODE_GEN_BUFFER_SIZE (3 * GiB) -#elif defined(__mips__) - /* - * We have a 256MB branch region, but leave room to make sure the - * main executable is also within that region. - */ -# define MAX_CODE_GEN_BUFFER_SIZE (128 * MiB) -#else -# define MAX_CODE_GEN_BUFFER_SIZE ((size_t)-1) -#endif - #if TCG_TARGET_REG_BITS == 32 #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (32 * MiB) #ifdef CONFIG_USER_ONLY -- 2.25.1 A size is easier to work with than an end point, particularly during initial buffer allocation. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 30 ++++++++++++++++++------------ 1 file changed, 18 insertions(+), 12 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ struct tcg_region_state { /* fields set at init time */ void *start; void *start_aligned; - void *end; size_t n; size_t size; /* size of one region */ size_t stride; /* .size + guard size */ + size_t total_size; /* size of entire buffer, >= n * stride */ /* fields protected by the lock */ size_t current; /* current region index */ @@ -XXX,XX +XXX,XX @@ static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend) if (curr_region == 0) { start = region.start; } + /* The final region may have a few extra pages due to earlier rounding. */ if (curr_region == region.n - 1) { - end = region.end; + end = region.start_aligned + region.total_size; } *pstart = start; @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) */ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) { - void *buf, *aligned; - size_t size; + void *buf, *aligned, *end; + size_t total_size; size_t page_size; size_t region_size; size_t n_regions; @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) assert(ok); buf = tcg_init_ctx.code_gen_buffer; - size = tcg_init_ctx.code_gen_buffer_size; + total_size = tcg_init_ctx.code_gen_buffer_size; page_size = qemu_real_host_page_size; n_regions = tcg_n_regions(max_cpus); /* The first region will be 'aligned - buf' bytes larger than the others */ aligned = QEMU_ALIGN_PTR_UP(buf, page_size); - g_assert(aligned < tcg_init_ctx.code_gen_buffer + size); + g_assert(aligned < tcg_init_ctx.code_gen_buffer + total_size); + /* * Make region_size a multiple of page_size, using aligned as the start. * As a result of this we might end up with a few extra pages at the end of * the buffer; we will assign those to the last region. */ - region_size = (size - (aligned - buf)) / n_regions; + region_size = (total_size - (aligned - buf)) / n_regions; region_size = QEMU_ALIGN_DOWN(region_size, page_size); /* A region must have at least 2 pages; one code, one guard */ @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) region.start = buf; region.start_aligned = aligned; /* page-align the end, since its last page will be a guard page */ - region.end = QEMU_ALIGN_PTR_DOWN(buf + size, page_size); + end = QEMU_ALIGN_PTR_DOWN(buf + total_size, page_size); /* account for that last guard page */ - region.end -= page_size; + end -= page_size; + total_size = end - aligned; + region.total_size = total_size; /* * Set guard pages in the rw buffer, as that's the one into which @@ -XXX,XX +XXX,XX @@ void tcg_region_prologue_set(TCGContext *s) /* Register the balance of the buffer with gdb. */ tcg_register_jit(tcg_splitwx_to_rx(region.start), - region.end - region.start); + region.start_aligned + region.total_size - region.start); } /* @@ -XXX,XX +XXX,XX @@ size_t tcg_code_capacity(void) /* no need for synchronization; these variables are set at init time */ guard_size = region.stride - region.size; - capacity = region.end + guard_size - region.start; - capacity -= region.n * (guard_size + TCG_HIGHWATER); + capacity = region.total_size; + capacity -= (region.n - 1) * guard_size; + capacity -= region.n * TCG_HIGHWATER; + return capacity; } -- 2.25.1 Give the field a name reflecting its actual meaning. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 15 ++++++++------- 1 file changed, 8 insertions(+), 7 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ struct tcg_region_state { QemuMutex lock; /* fields set at init time */ - void *start; void *start_aligned; + void *after_prologue; size_t n; size_t size; /* size of one region */ size_t stride; /* .size + guard size */ @@ -XXX,XX +XXX,XX @@ static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend) end = start + region.size; if (curr_region == 0) { - start = region.start; + start = region.after_prologue; } /* The final region may have a few extra pages due to earlier rounding. */ if (curr_region == region.n - 1) { @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) region.n = n_regions; region.size = region_size - page_size; region.stride = region_size; - region.start = buf; + region.after_prologue = buf; region.start_aligned = aligned; /* page-align the end, since its last page will be a guard page */ end = QEMU_ALIGN_PTR_DOWN(buf + total_size, page_size); @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) void tcg_region_prologue_set(TCGContext *s) { /* Deduct the prologue from the first region. */ - g_assert(region.start == s->code_gen_buffer); - region.start = s->code_ptr; + g_assert(region.start_aligned == s->code_gen_buffer); + region.after_prologue = s->code_ptr; /* Recompute boundaries of the first region. */ tcg_region_assign(s, 0); /* Register the balance of the buffer with gdb. */ - tcg_register_jit(tcg_splitwx_to_rx(region.start), - region.start_aligned + region.total_size - region.start); + tcg_register_jit(tcg_splitwx_to_rx(region.after_prologue), + region.start_aligned + region.total_size - + region.after_prologue); } /* -- 2.25.1 Compute the value using straight division and bounds, rather than a loop. Pass in tb_size rather than reading from tcg_init_ctx.code_gen_buffer_size, Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 29 ++++++++++++----------------- 1 file changed, 12 insertions(+), 17 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ void tcg_region_reset_all(void) tcg_region_tree_reset_all(); } -static size_t tcg_n_regions(unsigned max_cpus) +static size_t tcg_n_regions(size_t tb_size, unsigned max_cpus) { #ifdef CONFIG_USER_ONLY return 1; #else + size_t n_regions; + /* * It is likely that some vCPUs will translate more code than others, * so we first try to set more regions than max_cpus, with those regions * being of reasonable size. If that's not possible we make do by evenly * dividing the code_gen_buffer among the vCPUs. */ - size_t i; - /* Use a single region if all we have is one vCPU thread */ if (max_cpus == 1 || !qemu_tcg_mttcg_enabled()) { return 1; } - /* Try to have more regions than max_cpus, with each region being >= 2 MB */ - for (i = 8; i > 0; i--) { - size_t regions_per_thread = i; - size_t region_size; - - region_size = tcg_init_ctx.code_gen_buffer_size; - region_size /= max_cpus * regions_per_thread; - - if (region_size >= 2 * 1024u * 1024) { - return max_cpus * regions_per_thread; - } + /* + * Try to have more regions than max_cpus, with each region being >= 2 MB. + * If we can't, then just allocate one region per vCPU thread. + */ + n_regions = tb_size / (2 * MiB); + if (n_regions <= max_cpus) { + return max_cpus; } - /* If we can't, then just allocate one region per vCPU thread */ - return max_cpus; + return MIN(n_regions, max_cpus * 8); #endif } @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) buf = tcg_init_ctx.code_gen_buffer; total_size = tcg_init_ctx.code_gen_buffer_size; page_size = qemu_real_host_page_size; - n_regions = tcg_n_regions(max_cpus); + n_regions = tcg_n_regions(total_size, max_cpus); /* The first region will be 'aligned - buf' bytes larger than the others */ aligned = QEMU_ALIGN_PTR_UP(buf, page_size); -- 2.25.1 Return output buffer and size via output pointer arguments, rather than returning size via tcg_ctx->code_gen_buffer_size. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static inline bool cross_256mb(void *addr, size_t size) /* * We weren't able to allocate a buffer without crossing that boundary, * so make do with the larger portion of the buffer that doesn't cross. - * Returns the new base of the buffer, and adjusts code_gen_buffer_size. + * Returns the new base and size of the buffer in *obuf and *osize. */ -static inline void *split_cross_256mb(void *buf1, size_t size1) +static inline void split_cross_256mb(void **obuf, size_t *osize, + void *buf1, size_t size1) { void *buf2 = (void *)(((uintptr_t)buf1 + size1) & ~0x0ffffffful); size_t size2 = buf1 + size1 - buf2; @@ -XXX,XX +XXX,XX @@ static inline void *split_cross_256mb(void *buf1, size_t size1) buf1 = buf2; } - tcg_ctx->code_gen_buffer_size = size1; - return buf1; + *obuf = buf1; + *osize = size1; } #endif @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) if (size > tb_size) { size = QEMU_ALIGN_DOWN(tb_size, qemu_real_host_page_size); } - tcg_ctx->code_gen_buffer_size = size; #ifdef __mips__ if (cross_256mb(buf, size)) { - buf = split_cross_256mb(buf, size); - size = tcg_ctx->code_gen_buffer_size; + split_cross_256mb(&buf, &size, buf, size); } #endif @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); tcg_ctx->code_gen_buffer = buf; + tcg_ctx->code_gen_buffer_size = size; return true; } #elif defined(_WIN32) @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, "allocate %zu bytes for jit buffer", size); return false; } - tcg_ctx->code_gen_buffer_size = size; #ifdef __mips__ if (cross_256mb(buf, size)) { @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, /* fallthru */ default: /* Split the original buffer. Free the smaller half. */ - buf2 = split_cross_256mb(buf, size); - size2 = tcg_ctx->code_gen_buffer_size; + split_cross_256mb(&buf2, &size2, buf, size); if (buf == buf2) { munmap(buf + size2, size - size2); } else { @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); tcg_ctx->code_gen_buffer = buf; + tcg_ctx->code_gen_buffer_size = size; return true; } -- 2.25.1 Shortly, the full code_gen_buffer will only be visible to region.c, so move in_code_gen_buffer out-of-line. Move the debugging versions of tcg_splitwx_to_{rx,rw} to region.c as well, so that the compiler gets to see the implementation of in_code_gen_buffer. This leaves exactly one use of in_code_gen_buffer outside of region.c, in cpu_restore_state. Which, being on the exception path, is not performance critical. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 11 +---------- tcg/region.c | 34 ++++++++++++++++++++++++++++++++++ tcg/tcg.c | 23 ----------------------- 3 files changed, 35 insertions(+), 33 deletions(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ extern const void *tcg_code_gen_epilogue; extern uintptr_t tcg_splitwx_diff; extern TCGv_env cpu_env; -static inline bool in_code_gen_buffer(const void *p) -{ - const TCGContext *s = &tcg_init_ctx; - /* - * Much like it is valid to have a pointer to the byte past the - * end of an array (so long as you don't dereference it), allow - * a pointer to the byte past the end of the code gen buffer. - */ - return (size_t)(p - s->code_gen_buffer) <= s->code_gen_buffer_size; -} +bool in_code_gen_buffer(const void *p); #ifdef CONFIG_DEBUG_TCG const void *tcg_splitwx_to_rx(void *rw); diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static struct tcg_region_state region; static void *region_trees; static size_t tree_size; +bool in_code_gen_buffer(const void *p) +{ + const TCGContext *s = &tcg_init_ctx; + /* + * Much like it is valid to have a pointer to the byte past the + * end of an array (so long as you don't dereference it), allow + * a pointer to the byte past the end of the code gen buffer. + */ + return (size_t)(p - s->code_gen_buffer) <= s->code_gen_buffer_size; +} + +#ifdef CONFIG_DEBUG_TCG +const void *tcg_splitwx_to_rx(void *rw) +{ + /* Pass NULL pointers unchanged. */ + if (rw) { + g_assert(in_code_gen_buffer(rw)); + rw += tcg_splitwx_diff; + } + return rw; +} + +void *tcg_splitwx_to_rw(const void *rx) +{ + /* Pass NULL pointers unchanged. */ + if (rx) { + rx -= tcg_splitwx_diff; + /* Assert that we end with a pointer in the rw region. */ + g_assert(in_code_gen_buffer(rx)); + } + return (void *)rx; +} +#endif /* CONFIG_DEBUG_TCG */ + /* compare a pointer @ptr and a tb_tc @s */ static int ptr_cmp_tb_tc(const void *ptr, const struct tb_tc *s) { diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static const TCGTargetOpDef constraint_sets[] = { #include "tcg-target.c.inc" -#ifdef CONFIG_DEBUG_TCG -const void *tcg_splitwx_to_rx(void *rw) -{ - /* Pass NULL pointers unchanged. */ - if (rw) { - g_assert(in_code_gen_buffer(rw)); - rw += tcg_splitwx_diff; - } - return rw; -} - -void *tcg_splitwx_to_rw(const void *rx) -{ - /* Pass NULL pointers unchanged. */ - if (rx) { - rx -= tcg_splitwx_diff; - /* Assert that we end with a pointer in the rw region. */ - g_assert(in_code_gen_buffer(rx)); - } - return (void *)rx; -} -#endif /* CONFIG_DEBUG_TCG */ - static void alloc_tcg_plugin_context(TCGContext *s) { #ifdef CONFIG_PLUGIN -- 2.25.1 Do not mess around with setting values within tcg_init_ctx. Put the values into 'region' directly, which is where they will live for the lifetime of the program. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 64 ++++++++++++++++++++++------------------------------ 1 file changed, 27 insertions(+), 37 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static size_t tree_size; bool in_code_gen_buffer(const void *p) { - const TCGContext *s = &tcg_init_ctx; /* * Much like it is valid to have a pointer to the byte past the * end of an array (so long as you don't dereference it), allow * a pointer to the byte past the end of the code gen buffer. */ - return (size_t)(p - s->code_gen_buffer) <= s->code_gen_buffer_size; + return (size_t)(p - region.start_aligned) <= region.total_size; } #ifdef CONFIG_DEBUG_TCG @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) } qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); - tcg_ctx->code_gen_buffer = buf; - tcg_ctx->code_gen_buffer_size = size; + region.start_aligned = buf; + region.total_size = size; return true; } #elif defined(_WIN32) @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) return false; } - tcg_ctx->code_gen_buffer = buf; - tcg_ctx->code_gen_buffer_size = size; + region.start_aligned = buf; + region.total_size = size; return true; } #else @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, /* Request large pages for the buffer. */ qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); - tcg_ctx->code_gen_buffer = buf; - tcg_ctx->code_gen_buffer_size = size; + region.start_aligned = buf; + region.total_size = size; return true; } @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) return false; } /* The size of the mapping may have been adjusted. */ - size = tcg_ctx->code_gen_buffer_size; - buf_rx = tcg_ctx->code_gen_buffer; + buf_rx = region.start_aligned; + size = region.total_size; #endif buf_rw = qemu_memfd_alloc("tcg-jit", size, 0, &fd, errp); @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) #endif close(fd); - tcg_ctx->code_gen_buffer = buf_rw; - tcg_ctx->code_gen_buffer_size = size; + region.start_aligned = buf_rw; + region.total_size = size; tcg_splitwx_diff = buf_rx - buf_rw; /* Request large pages for the buffer and the splitwx. */ @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) return false; } - buf_rw = (mach_vm_address_t)tcg_ctx->code_gen_buffer; + buf_rw = region.start_aligned; buf_rx = 0; ret = mach_vm_remap(mach_task_self(), &buf_rx, @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) */ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) { - void *buf, *aligned, *end; - size_t total_size; size_t page_size; size_t region_size; - size_t n_regions; size_t i; bool ok; @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) splitwx, &error_fatal); assert(ok); - buf = tcg_init_ctx.code_gen_buffer; - total_size = tcg_init_ctx.code_gen_buffer_size; - page_size = qemu_real_host_page_size; - n_regions = tcg_n_regions(total_size, max_cpus); - - /* The first region will be 'aligned - buf' bytes larger than the others */ - aligned = QEMU_ALIGN_PTR_UP(buf, page_size); - g_assert(aligned < tcg_init_ctx.code_gen_buffer + total_size); - /* * Make region_size a multiple of page_size, using aligned as the start. * As a result of this we might end up with a few extra pages at the end of * the buffer; we will assign those to the last region. */ - region_size = (total_size - (aligned - buf)) / n_regions; + region.n = tcg_n_regions(region.total_size, max_cpus); + page_size = qemu_real_host_page_size; + region_size = region.total_size / region.n; region_size = QEMU_ALIGN_DOWN(region_size, page_size); /* A region must have at least 2 pages; one code, one guard */ g_assert(region_size >= 2 * page_size); + region.stride = region_size; + + /* Reserve space for guard pages. */ + region.size = region_size - page_size; + region.total_size -= page_size; + + /* + * The first region will be smaller than the others, via the prologue, + * which has yet to be allocated. For now, the first region begins at + * the page boundary. + */ + region.after_prologue = region.start_aligned; /* init the region struct */ qemu_mutex_init(&region.lock); - region.n = n_regions; - region.size = region_size - page_size; - region.stride = region_size; - region.after_prologue = buf; - region.start_aligned = aligned; - /* page-align the end, since its last page will be a guard page */ - end = QEMU_ALIGN_PTR_DOWN(buf + total_size, page_size); - /* account for that last guard page */ - end -= page_size; - total_size = end - aligned; - region.total_size = total_size; /* * Set guard pages in the rw buffer, as that's the one into which -- 2.25.1 Change the interface from a boolean error indication to a negative error vs a non-negative protection. For the moment this is only interface change, not making use of the new data. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 63 +++++++++++++++++++++++++++------------------------- 1 file changed, 33 insertions(+), 30 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static inline void split_cross_256mb(void **obuf, size_t *osize, static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE] __attribute__((aligned(CODE_GEN_ALIGN))); -static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) +static int alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) { void *buf, *end; size_t size; if (splitwx > 0) { error_setg(errp, "jit split-wx not supported"); - return false; + return -1; } /* page-align the beginning and end of the buffer */ @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) region.start_aligned = buf; region.total_size = size; - return true; + + return PROT_READ | PROT_WRITE; } #elif defined(_WIN32) -static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) +static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) { void *buf; if (splitwx > 0) { error_setg(errp, "jit split-wx not supported"); - return false; + return -1; } buf = VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT, @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) region.start_aligned = buf; region.total_size = size; - return true; + + return PAGE_READ | PAGE_WRITE | PAGE_EXEC; } #else -static bool alloc_code_gen_buffer_anon(size_t size, int prot, - int flags, Error **errp) +static int alloc_code_gen_buffer_anon(size_t size, int prot, + int flags, Error **errp) { void *buf; @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, if (buf == MAP_FAILED) { error_setg_errno(errp, errno, "allocate %zu bytes for jit buffer", size); - return false; + return -1; } #ifdef __mips__ @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_anon(size_t size, int prot, region.start_aligned = buf; region.total_size = size; - return true; + return prot; } #ifndef CONFIG_TCG_INTERPRETER @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) #ifdef __mips__ /* Find space for the RX mapping, vs the 256MiB regions. */ - if (!alloc_code_gen_buffer_anon(size, PROT_NONE, - MAP_PRIVATE | MAP_ANONYMOUS | - MAP_NORESERVE, errp)) { + if (alloc_code_gen_buffer_anon(size, PROT_NONE, + MAP_PRIVATE | MAP_ANONYMOUS | + MAP_NORESERVE, errp) < 0) { return false; } /* The size of the mapping may have been adjusted. */ @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) /* Request large pages for the buffer and the splitwx. */ qemu_madvise(buf_rw, size, QEMU_MADV_HUGEPAGE); qemu_madvise(buf_rx, size, QEMU_MADV_HUGEPAGE); - return true; + return PROT_READ | PROT_WRITE; fail_rx: error_setg_errno(errp, errno, "failed to map shared memory for execute"); @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) if (fd >= 0) { close(fd); } - return false; + return -1; } #endif /* CONFIG_POSIX */ @@ -XXX,XX +XXX,XX @@ extern kern_return_t mach_vm_remap(vm_map_t target_task, vm_prot_t *max_protection, vm_inherit_t inheritance); -static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) +static int alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) { kern_return_t ret; mach_vm_address_t buf_rw, buf_rx; @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) /* Map the read-write portion via normal anon memory. */ if (!alloc_code_gen_buffer_anon(size, PROT_READ | PROT_WRITE, MAP_PRIVATE | MAP_ANONYMOUS, errp)) { - return false; + return -1; } buf_rw = region.start_aligned; @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) /* TODO: Convert "ret" to a human readable error message. */ error_setg(errp, "vm_remap for jit splitwx failed"); munmap((void *)buf_rw, size); - return false; + return -1; } if (mprotect((void *)buf_rx, size, PROT_READ | PROT_EXEC) != 0) { error_setg_errno(errp, errno, "mprotect for jit splitwx"); munmap((void *)buf_rx, size); munmap((void *)buf_rw, size); - return false; + return -1; } tcg_splitwx_diff = buf_rx - buf_rw; - return true; + return PROT_READ | PROT_WRITE; } #endif /* CONFIG_DARWIN */ #endif /* CONFIG_TCG_INTERPRETER */ -static bool alloc_code_gen_buffer_splitwx(size_t size, Error **errp) +static int alloc_code_gen_buffer_splitwx(size_t size, Error **errp) { #ifndef CONFIG_TCG_INTERPRETER # ifdef CONFIG_DARWIN @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx(size_t size, Error **errp) # endif #endif error_setg(errp, "jit split-wx not supported"); - return false; + return -1; } -static bool alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) +static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) { ERRP_GUARD(); int prot, flags; if (splitwx) { - if (alloc_code_gen_buffer_splitwx(size, errp)) { - return true; + prot = alloc_code_gen_buffer_splitwx(size, errp); + if (prot >= 0) { + return prot; } /* * If splitwx force-on (1), fail; * if splitwx default-on (-1), fall through to splitwx off. */ if (splitwx > 0) { - return false; + return -1; } error_free_or_abort(errp); } @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) size_t page_size; size_t region_size; size_t i; - bool ok; + int have_prot; - ok = alloc_code_gen_buffer(size_code_gen_buffer(tb_size), - splitwx, &error_fatal); - assert(ok); + have_prot = alloc_code_gen_buffer(size_code_gen_buffer(tb_size), + splitwx, &error_fatal); + assert(have_prot >= 0); /* * Make region_size a multiple of page_size, using aligned as the start. -- 2.25.1 Move the call out of the N versions of alloc_code_gen_buffer and into tcg_region_init. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static int alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) error_setg_errno(errp, errno, "mprotect of jit buffer"); return false; } - qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); region.start_aligned = buf; region.total_size = size; @@ -XXX,XX +XXX,XX @@ static int alloc_code_gen_buffer_anon(size_t size, int prot, } #endif - /* Request large pages for the buffer. */ - qemu_madvise(buf, size, QEMU_MADV_HUGEPAGE); - region.start_aligned = buf; region.total_size = size; return prot; @@ -XXX,XX +XXX,XX @@ static bool alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) region.total_size = size; tcg_splitwx_diff = buf_rx - buf_rw; - /* Request large pages for the buffer and the splitwx. */ - qemu_madvise(buf_rw, size, QEMU_MADV_HUGEPAGE); - qemu_madvise(buf_rx, size, QEMU_MADV_HUGEPAGE); return PROT_READ | PROT_WRITE; fail_rx: @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) splitwx, &error_fatal); assert(have_prot >= 0); + /* Request large pages for the buffer and the splitwx. */ + qemu_madvise(region.start_aligned, region.total_size, QEMU_MADV_HUGEPAGE); + if (tcg_splitwx_diff) { + qemu_madvise(region.start_aligned + tcg_splitwx_diff, + region.total_size, QEMU_MADV_HUGEPAGE); + } + /* * Make region_size a multiple of page_size, using aligned as the start. * As a result of this we might end up with a few extra pages at the end of -- 2.25.1 For --enable-tcg-interpreter on Windows, we will need this. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/qemu/osdep.h | 1 + util/osdep.c | 9 +++++++++ 2 files changed, 10 insertions(+) diff --git a/include/qemu/osdep.h b/include/qemu/osdep.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/osdep.h +++ b/include/qemu/osdep.h @@ -XXX,XX +XXX,XX @@ void sigaction_invoke(struct sigaction *action, #endif int qemu_madvise(void *addr, size_t len, int advice); +int qemu_mprotect_rw(void *addr, size_t size); int qemu_mprotect_rwx(void *addr, size_t size); int qemu_mprotect_none(void *addr, size_t size); diff --git a/util/osdep.c b/util/osdep.c index XXXXXXX..XXXXXXX 100644 --- a/util/osdep.c +++ b/util/osdep.c @@ -XXX,XX +XXX,XX @@ static int qemu_mprotect__osdep(void *addr, size_t size, int prot) #endif } +int qemu_mprotect_rw(void *addr, size_t size) +{ +#ifdef _WIN32 + return qemu_mprotect__osdep(addr, size, PAGE_READWRITE); +#else + return qemu_mprotect__osdep(addr, size, PROT_READ | PROT_WRITE); +#endif +} + int qemu_mprotect_rwx(void *addr, size_t size) { #ifdef _WIN32 -- 2.25.1 If qemu_get_host_physmem returns an odd number of pages, then physmem / 8 will not be a multiple of the page size. The following was observed on a gitlab runner: ERROR qtest-arm/boot-serial-test - Bail out! ERROR:../util/osdep.c:80:qemu_mprotect__osdep: \ assertion failed: (!(size & ~qemu_real_host_page_mask)) Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 47 +++++++++++++++++++++-------------------------- 1 file changed, 21 insertions(+), 26 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static size_t tcg_n_regions(size_t tb_size, unsigned max_cpus) (DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \ ? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE) -static size_t size_code_gen_buffer(size_t tb_size) -{ - /* Size the buffer. */ - if (tb_size == 0) { - size_t phys_mem = qemu_get_host_physmem(); - if (phys_mem == 0) { - tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE; - } else { - tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, phys_mem / 8); - } - } - if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) { - tb_size = MIN_CODE_GEN_BUFFER_SIZE; - } - if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) { - tb_size = MAX_CODE_GEN_BUFFER_SIZE; - } - return tb_size; -} - #ifdef __mips__ /* * In order to use J and JAL within the code_gen_buffer, we require @@ -XXX,XX +XXX,XX @@ static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) */ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) { - size_t page_size; + const size_t page_size = qemu_real_host_page_size; size_t region_size; size_t i; int have_prot; - have_prot = alloc_code_gen_buffer(size_code_gen_buffer(tb_size), - splitwx, &error_fatal); + /* Size the buffer. */ + if (tb_size == 0) { + size_t phys_mem = qemu_get_host_physmem(); + if (phys_mem == 0) { + tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE; + } else { + tb_size = QEMU_ALIGN_DOWN(phys_mem / 8, page_size); + tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, tb_size); + } + } + if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) { + tb_size = MIN_CODE_GEN_BUFFER_SIZE; + } + if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) { + tb_size = MAX_CODE_GEN_BUFFER_SIZE; + } + + have_prot = alloc_code_gen_buffer(tb_size, splitwx, &error_fatal); assert(have_prot >= 0); /* Request large pages for the buffer and the splitwx. */ @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) * As a result of this we might end up with a few extra pages at the end of * the buffer; we will assign those to the last region. */ - region.n = tcg_n_regions(region.total_size, max_cpus); - page_size = qemu_real_host_page_size; - region_size = region.total_size / region.n; + region.n = tcg_n_regions(tb_size, max_cpus); + region_size = tb_size / region.n; region_size = QEMU_ALIGN_DOWN(region_size, page_size); /* A region must have at least 2 pages; one code, one guard */ -- 2.25.1 Do not handle protections on a case-by-case basis in the various alloc_code_gen_buffer instances; do it within a single loop in tcg_region_init. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 45 +++++++++++++++++++++++++++++++-------------- 1 file changed, 31 insertions(+), 14 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static int alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) } #endif - if (qemu_mprotect_rwx(buf, size)) { - error_setg_errno(errp, errno, "mprotect of jit buffer"); - return false; - } - region.start_aligned = buf; region.total_size = size; @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) { const size_t page_size = qemu_real_host_page_size; size_t region_size; - size_t i; - int have_prot; + int have_prot, need_prot; /* Size the buffer. */ if (tb_size == 0) { @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) * Set guard pages in the rw buffer, as that's the one into which * buffer overruns could occur. Do not set guard pages in the rx * buffer -- let that one use hugepages throughout. + * Work with the page protections set up with the initial mapping. */ - for (i = 0; i < region.n; i++) { + need_prot = PAGE_READ | PAGE_WRITE; +#ifndef CONFIG_TCG_INTERPRETER + if (tcg_splitwx_diff == 0) { + need_prot |= PAGE_EXEC; + } +#endif + for (size_t i = 0, n = region.n; i < n; i++) { void *start, *end; tcg_region_bounds(i, &start, &end); + if (have_prot != need_prot) { + int rc; - /* - * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect - * rejects a permission change from RWX -> NONE. Guard pages are - * nice for bug detection but are not essential; ignore any failure. - */ - (void)qemu_mprotect_none(end, page_size); + if (need_prot == (PAGE_READ | PAGE_WRITE | PAGE_EXEC)) { + rc = qemu_mprotect_rwx(start, end - start); + } else if (need_prot == (PAGE_READ | PAGE_WRITE)) { + rc = qemu_mprotect_rw(start, end - start); + } else { + g_assert_not_reached(); + } + if (rc) { + error_setg_errno(&error_fatal, errno, + "mprotect of jit buffer"); + } + } + if (have_prot != 0) { + /* + * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect + * rejects a permission change from RWX -> NONE. Guard pages are + * nice for bug detection but are not essential; ignore any failure. + */ + (void)qemu_mprotect_none(end, page_size); + } } tcg_region_trees_init(); -- 2.25.1 There's a change in mprotect() behaviour [1] in the latest macOS on M1 and it's not yet clear if it's going to be fixed by Apple. In this case, instead of changing permissions of N guard pages, we change permissions of N rwx regions. The same number of syscalls are required either way. [1] https://gist.github.com/hikalium/75ae822466ee4da13cbbe486498a191f Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/region.c | 19 +++++++++---------- 1 file changed, 9 insertions(+), 10 deletions(-) diff --git a/tcg/region.c b/tcg/region.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/region.c +++ b/tcg/region.c @@ -XXX,XX +XXX,XX @@ static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) error_free_or_abort(errp); } - prot = PROT_READ | PROT_WRITE | PROT_EXEC; + /* + * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect + * rejects a permission change from RWX -> NONE when reserving the + * guard pages later. We can go the other way with the same number + * of syscalls, so always begin with PROT_NONE. + */ + prot = PROT_NONE; flags = MAP_PRIVATE | MAP_ANONYMOUS; -#ifdef CONFIG_TCG_INTERPRETER - /* The tcg interpreter does not need execute permission. */ - prot = PROT_READ | PROT_WRITE; -#elif defined(CONFIG_DARWIN) +#ifdef CONFIG_DARWIN /* Applicable to both iOS and macOS (Apple Silicon). */ if (!splitwx) { flags |= MAP_JIT; @@ -XXX,XX +XXX,XX @@ void tcg_region_init(size_t tb_size, int splitwx, unsigned max_cpus) } } if (have_prot != 0) { - /* - * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect - * rejects a permission change from RWX -> NONE. Guard pages are - * nice for bug detection but are not essential; ignore any failure. - */ + /* Guard pages are nice for bug detection but are not essential. */ (void)qemu_mprotect_none(end, page_size); } } -- 2.25.1 These variables belong to the jit side, not the user side. Since tcg_init_ctx is no longer used outside of tcg/, move the declaration to tcg-internal.h. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Suggested-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 1 - tcg/tcg-internal.h | 1 + accel/tcg/translate-all.c | 3 --- tcg/tcg.c | 3 +++ 4 files changed, 4 insertions(+), 4 deletions(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ static inline bool temp_readonly(TCGTemp *ts) return ts->kind >= TEMP_FIXED; } -extern TCGContext tcg_init_ctx; extern __thread TCGContext *tcg_ctx; extern const void *tcg_code_gen_epilogue; extern uintptr_t tcg_splitwx_diff; diff --git a/tcg/tcg-internal.h b/tcg/tcg-internal.h index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg-internal.h +++ b/tcg/tcg-internal.h @@ -XXX,XX +XXX,XX @@ #define TCG_HIGHWATER 1024 +extern TCGContext tcg_init_ctx; extern TCGContext **tcg_ctxs; extern unsigned int tcg_cur_ctxs; extern unsigned int tcg_max_ctxs; diff --git a/accel/tcg/translate-all.c b/accel/tcg/translate-all.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.c +++ b/accel/tcg/translate-all.c @@ -XXX,XX +XXX,XX @@ static int v_l2_levels; static void *l1_map[V_L1_MAX_SIZE]; -/* code generation context */ -TCGContext tcg_init_ctx; -__thread TCGContext *tcg_ctx; TBContext tb_ctx; static void page_table_config_init(void) diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ static bool tcg_target_const_match(int64_t val, TCGType type, int ct); static int tcg_out_ldst_finalize(TCGContext *s); #endif +TCGContext tcg_init_ctx; +__thread TCGContext *tcg_ctx; + TCGContext **tcg_ctxs; unsigned int tcg_cur_ctxs; unsigned int tcg_max_ctxs; -- 2.25.1 Introduce a function to remove everything emitted since a given point. Reviewed-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 10 ++++++++++ tcg/tcg.c | 13 +++++++++++++ 2 files changed, 23 insertions(+) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ void tcg_op_remove(TCGContext *s, TCGOp *op); TCGOp *tcg_op_insert_before(TCGContext *s, TCGOp *op, TCGOpcode opc); TCGOp *tcg_op_insert_after(TCGContext *s, TCGOp *op, TCGOpcode opc); +/** + * tcg_remove_ops_after: + * @op: target operation + * + * Discard any opcodes emitted since @op. Expected usage is to save + * a starting point with tcg_last_op(), speculatively emit opcodes, + * then decide whether or not to keep those opcodes after the fact. + */ +void tcg_remove_ops_after(TCGOp *op); + void tcg_optimize(TCGContext *s); /* Allocate a new temporary and initialize it with a constant. */ diff --git a/tcg/tcg.c b/tcg/tcg.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/tcg.c +++ b/tcg/tcg.c @@ -XXX,XX +XXX,XX @@ void tcg_op_remove(TCGContext *s, TCGOp *op) #endif } +void tcg_remove_ops_after(TCGOp *op) +{ + TCGContext *s = tcg_ctx; + + while (true) { + TCGOp *last = tcg_last_op(); + if (last == op) { + return; + } + tcg_op_remove(s, last); + } +} + static TCGOp *tcg_op_alloc(TCGOpcode opc) { TCGContext *s = tcg_ctx; -- 2.25.1 At some point during the development of tcg_constant_*, I changed my mind about whether such temps should be able to be passed to tcg_temp_free_*. The final version committed allows this, but the commentary was not updated to match. Fixes: c0522136adf Reported-by: Peter Maydell <peter.maydell@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/tcg/tcg.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/include/tcg/tcg.h b/include/tcg/tcg.h index XXXXXXX..XXXXXXX 100644 --- a/include/tcg/tcg.h +++ b/include/tcg/tcg.h @@ -XXX,XX +XXX,XX @@ TCGv_vec tcg_const_ones_vec_matching(TCGv_vec); /* * Locate or create a read-only temporary that is a constant. - * This kind of temporary need not and should not be freed. + * This kind of temporary need not be freed, but for convenience + * will be silently ignored by tcg_temp_free_*. */ TCGTemp *tcg_constant_internal(TCGType type, int64_t val); -- 2.25.1 From: "Jose R. Ziviani" <jziviani@suse.de> Commit 5e8892db93 fixed several function signatures but tcg_out_op for arm is missing. This patch fixes it as well. Signed-off-by: Jose R. Ziviani <jziviani@suse.de> Message-Id: <20210610224450.23425-1-jziviani@suse.de> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/arm/tcg-target.c.inc | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tcg/arm/tcg-target.c.inc b/tcg/arm/tcg-target.c.inc index XXXXXXX..XXXXXXX 100644 --- a/tcg/arm/tcg-target.c.inc +++ b/tcg/arm/tcg-target.c.inc @@ -XXX,XX +XXX,XX @@ static void tcg_out_qemu_st(TCGContext *s, const TCGArg *args, bool is64) static void tcg_out_epilogue(TCGContext *s); static inline void tcg_out_op(TCGContext *s, TCGOpcode opc, - const TCGArg *args, const int *const_args) + const TCGArg args[TCG_MAX_OP_ARGS], + const int const_args[TCG_MAX_OP_ARGS]) { TCGArg a0, a1, a2, a3, a4, a5; int c; -- 2.25.1 Typo in the conversion to FloatParts64. Fixes: 572c4d862ff2 Fixes: Coverity CID 1457457 Signed-off-by: Richard Henderson <richard.henderson@linaro.org> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Message-Id: <20210607223812.110596-1-richard.henderson@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- fpu/softfloat.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/fpu/softfloat.c b/fpu/softfloat.c index XXXXXXX..XXXXXXX 100644 --- a/fpu/softfloat.c +++ b/fpu/softfloat.c @@ -XXX,XX +XXX,XX @@ float32 float32_exp2(float32 a, float_status *status) float_raise(float_flag_inexact, status); - float64_unpack_canonical(&xnp, float64_ln2, status); + float64_unpack_canonical(&tp, float64_ln2, status); xp = *parts_mul(&xp, &tp, status); xnp = xp; -- 2.25.1 From: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Message-Id: <20210601125143.191165-1-luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- docs/devel/tcg.rst | 101 ++++++++++++++++++++++++++++++++++++++++----- 1 file changed, 90 insertions(+), 11 deletions(-) diff --git a/docs/devel/tcg.rst b/docs/devel/tcg.rst index XXXXXXX..XXXXXXX 100644 --- a/docs/devel/tcg.rst +++ b/docs/devel/tcg.rst @@ -XXX,XX +XXX,XX @@ performances. QEMU's dynamic translation backend is called TCG, for "Tiny Code Generator". For more information, please take a look at ``tcg/README``. -Some notable features of QEMU's dynamic translator are: +The following sections outline some notable features and implementation +details of QEMU's dynamic translator. CPU state optimisations ----------------------- -The target CPUs have many internal states which change the way it -evaluates instructions. In order to achieve a good speed, the +The target CPUs have many internal states which change the way they +evaluate instructions. In order to achieve a good speed, the translation phase considers that some state information of the virtual CPU cannot change in it. The state is recorded in the Translation Block (TB). If the state changes (e.g. privilege level), a new TB will @@ -XXX,XX +XXX,XX @@ Direct block chaining --------------------- After each translated basic block is executed, QEMU uses the simulated -Program Counter (PC) and other cpu state information (such as the CS +Program Counter (PC) and other CPU state information (such as the CS segment base value) to find the next basic block. -In order to accelerate the most common cases where the new simulated PC -is known, QEMU can patch a basic block so that it jumps directly to the -next one. +In its simplest, less optimized form, this is done by exiting from the +current TB, going through the TB epilogue, and then back to the +main loop. That’s where QEMU looks for the next TB to execute, +translating it from the guest architecture if it isn’t already available +in memory. Then QEMU proceeds to execute this next TB, starting at the +prologue and then moving on to the translated instructions. -The most portable code uses an indirect jump. An indirect jump makes -it easier to make the jump target modification atomic. On some host -architectures (such as x86 or PowerPC), the ``JUMP`` opcode is -directly patched so that the block chaining has no overhead. +Exiting from the TB this way will cause the ``cpu_exec_interrupt()`` +callback to be re-evaluated before executing additional instructions. +It is mandatory to exit this way after any CPU state changes that may +unmask interrupts. + +In order to accelerate the cases where the TB for the new +simulated PC is already available, QEMU has mechanisms that allow +multiple TBs to be chained directly, without having to go back to the +main loop as described above. These mechanisms are: + +``lookup_and_goto_ptr`` +^^^^^^^^^^^^^^^^^^^^^^^ + +Calling ``tcg_gen_lookup_and_goto_ptr()`` will emit a call to +``helper_lookup_tb_ptr``. This helper will look for an existing TB that +matches the current CPU state. If the destination TB is available its +code address is returned, otherwise the address of the JIT epilogue is +returned. The call to the helper is always followed by the tcg ``goto_ptr`` +opcode, which branches to the returned address. In this way, we either +branch to the next TB or return to the main loop. + +``goto_tb + exit_tb`` +^^^^^^^^^^^^^^^^^^^^^ + +The translation code usually implements branching by performing the +following steps: + +1. Call ``tcg_gen_goto_tb()`` passing a jump slot index (either 0 or 1) + as a parameter. + +2. Emit TCG instructions to update the CPU state with any information + that has been assumed constant and is required by the main loop to + correctly locate and execute the next TB. For most guests, this is + just the PC of the branch destination, but others may store additional + data. The information updated in this step must be inferable from both + ``cpu_get_tb_cpu_state()`` and ``cpu_restore_state()``. + +3. Call ``tcg_gen_exit_tb()`` passing the address of the current TB and + the jump slot index again. + +Step 1, ``tcg_gen_goto_tb()``, will emit a ``goto_tb`` TCG +instruction that later on gets translated to a jump to an address +associated with the specified jump slot. Initially, this is the address +of step 2's instructions, which update the CPU state information. Step 3, +``tcg_gen_exit_tb()``, exits from the current TB returning a tagged +pointer composed of the last executed TB’s address and the jump slot +index. + +The first time this whole sequence is executed, step 1 simply jumps +to step 2. Then the CPU state information gets updated and we exit from +the current TB. As a result, the behavior is very similar to the less +optimized form described earlier in this section. + +Next, the main loop looks for the next TB to execute using the +current CPU state information (creating the TB if it wasn’t already +available) and, before starting to execute the new TB’s instructions, +patches the previously executed TB by associating one of its jump +slots (the one specified in the call to ``tcg_gen_exit_tb()``) with the +address of the new TB. + +The next time this previous TB is executed and we get to that same +``goto_tb`` step, it will already be patched (assuming the destination TB +is still in memory) and will jump directly to the first instruction of +the destination TB, without going back to the main loop. + +For the ``goto_tb + exit_tb`` mechanism to be used, the following +conditions need to be satisfied: + +* The change in CPU state must be constant, e.g., a direct branch and + not an indirect branch. + +* The direct branch cannot cross a page boundary. Memory mappings + may change, causing the code at the destination address to change. + +Note that, on step 3 (``tcg_gen_exit_tb()``), in addition to the +jump slot index, the address of the TB just executed is also returned. +This address corresponds to the TB that will be patched; it may be +different than the one that was directly executed from the main loop +if the latter had already been chained to other TBs. Self-modifying code and translated code invalidation ---------------------------------------------------- -- 2.25.1 The following changes since commit c52d69e7dbaaed0ffdef8125e79218672c30161d: Merge remote-tracking branch 'remotes/cschoenebeck/tags/pull-9p-20211027' into staging (2021-10-27 11:45:18 -0700) are available in the Git repository at: https://gitlab.com/rth7680/qemu.git tags/pull-tcg-20211027 for you to fetch changes up to 820c025f0dcacf2f3c12735b1f162893fbfa7bc6: tcg/optimize: Propagate sign info for shifting (2021-10-27 17:11:23 -0700) ---------------------------------------------------------------- Improvements to qemu/int128 Fixes for 128/64 division. Cleanup tcg/optimize.c Optimize redundant sign extensions ---------------------------------------------------------------- Frédéric Pétrot (1): qemu/int128: Add int128_{not,xor} Luis Pires (4): host-utils: move checks out of divu128/divs128 host-utils: move udiv_qrnnd() to host-utils host-utils: add 128-bit quotient support to divu128/divs128 host-utils: add unit tests for divu128/divs128 Richard Henderson (51): tcg/optimize: Rename "mask" to "z_mask" tcg/optimize: Split out OptContext tcg/optimize: Remove do_default label tcg/optimize: Change tcg_opt_gen_{mov,movi} interface tcg/optimize: Move prev_mb into OptContext tcg/optimize: Split out init_arguments tcg/optimize: Split out copy_propagate tcg/optimize: Split out fold_call tcg/optimize: Drop nb_oargs, nb_iargs locals tcg/optimize: Change fail return for do_constant_folding_cond* tcg/optimize: Return true from tcg_opt_gen_{mov,movi} tcg/optimize: Split out finish_folding tcg/optimize: Use a boolean to avoid a mass of continues tcg/optimize: Split out fold_mb, fold_qemu_{ld,st} tcg/optimize: Split out fold_const{1,2} tcg/optimize: Split out fold_setcond2 tcg/optimize: Split out fold_brcond2 tcg/optimize: Split out fold_brcond tcg/optimize: Split out fold_setcond tcg/optimize: Split out fold_mulu2_i32 tcg/optimize: Split out fold_addsub2_i32 tcg/optimize: Split out fold_movcond tcg/optimize: Split out fold_extract2 tcg/optimize: Split out fold_extract, fold_sextract tcg/optimize: Split out fold_deposit tcg/optimize: Split out fold_count_zeros tcg/optimize: Split out fold_bswap tcg/optimize: Split out fold_dup, fold_dup2 tcg/optimize: Split out fold_mov tcg/optimize: Split out fold_xx_to_i tcg/optimize: Split out fold_xx_to_x tcg/optimize: Split out fold_xi_to_i tcg/optimize: Add type to OptContext tcg/optimize: Split out fold_to_not tcg/optimize: Split out fold_sub_to_neg tcg/optimize: Split out fold_xi_to_x tcg/optimize: Split out fold_ix_to_i tcg/optimize: Split out fold_masks tcg/optimize: Expand fold_mulu2_i32 to all 4-arg multiplies tcg/optimize: Expand fold_addsub2_i32 to 64-bit ops tcg/optimize: Sink commutative operand swapping into fold functions tcg/optimize: Stop forcing z_mask to "garbage" for 32-bit values tcg/optimize: Use fold_xx_to_i for orc tcg/optimize: Use fold_xi_to_x for mul tcg/optimize: Use fold_xi_to_x for div tcg/optimize: Use fold_xx_to_i for rem tcg/optimize: Optimize sign extensions tcg/optimize: Propagate sign info for logical operations tcg/optimize: Propagate sign info for setcond tcg/optimize: Propagate sign info for bit counting tcg/optimize: Propagate sign info for shifting include/fpu/softfloat-macros.h | 82 -- include/hw/clock.h | 5 +- include/qemu/host-utils.h | 121 +- include/qemu/int128.h | 20 + target/ppc/int_helper.c | 23 +- tcg/optimize.c | 2644 ++++++++++++++++++++++++---------------- tests/unit/test-div128.c | 197 +++ util/host-utils.c | 147 ++- tests/unit/meson.build | 1 + 9 files changed, 2053 insertions(+), 1187 deletions(-) create mode 100644 tests/unit/test-div128.c From: Frédéric Pétrot <frederic.petrot@univ-grenoble-alpes.fr> Addition of not and xor on 128-bit integers. Signed-off-by: Frédéric Pétrot <frederic.petrot@univ-grenoble-alpes.fr> Co-authored-by: Fabien Portas <fabien.portas@grenoble-inp.org> Message-Id: <20211025122818.168890-3-frederic.petrot@univ-grenoble-alpes.fr> [rth: Split out logical operations.] Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/qemu/int128.h | 20 ++++++++++++++++++++ 1 file changed, 20 insertions(+) diff --git a/include/qemu/int128.h b/include/qemu/int128.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/int128.h +++ b/include/qemu/int128.h @@ -XXX,XX +XXX,XX @@ static inline Int128 int128_exts64(int64_t a) return a; } +static inline Int128 int128_not(Int128 a) +{ + return ~a; +} + static inline Int128 int128_and(Int128 a, Int128 b) { return a & b; @@ -XXX,XX +XXX,XX @@ static inline Int128 int128_or(Int128 a, Int128 b) return a | b; } +static inline Int128 int128_xor(Int128 a, Int128 b) +{ + return a ^ b; +} + static inline Int128 int128_rshift(Int128 a, int n) { return a >> n; @@ -XXX,XX +XXX,XX @@ static inline Int128 int128_exts64(int64_t a) return int128_make128(a, (a < 0) ? -1 : 0); } +static inline Int128 int128_not(Int128 a) +{ + return int128_make128(~a.lo, ~a.hi); +} + static inline Int128 int128_and(Int128 a, Int128 b) { return int128_make128(a.lo & b.lo, a.hi & b.hi); @@ -XXX,XX +XXX,XX @@ static inline Int128 int128_or(Int128 a, Int128 b) return int128_make128(a.lo | b.lo, a.hi | b.hi); } +static inline Int128 int128_xor(Int128 a, Int128 b) +{ + return int128_make128(a.lo ^ b.lo, a.hi ^ b.hi); +} + static inline Int128 int128_rshift(Int128 a, int n) { int64_t h; -- 2.25.1 From: Luis Pires <luis.pires@eldorado.org.br> In preparation for changing the divu128/divs128 implementations to allow for quotients larger than 64 bits, move the div-by-zero and overflow checks to the callers. Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <20211025191154.350831-2-luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/hw/clock.h | 5 +++-- include/qemu/host-utils.h | 34 ++++++++++++--------------------- target/ppc/int_helper.c | 14 +++++++++----- util/host-utils.c | 40 ++++++++++++++++++--------------------- 4 files changed, 42 insertions(+), 51 deletions(-) diff --git a/include/hw/clock.h b/include/hw/clock.h index XXXXXXX..XXXXXXX 100644 --- a/include/hw/clock.h +++ b/include/hw/clock.h @@ -XXX,XX +XXX,XX @@ static inline uint64_t clock_ns_to_ticks(const Clock *clk, uint64_t ns) return 0; } /* - * Ignore divu128() return value as we've caught div-by-zero and don't - * need different behaviour for overflow. + * BUG: when CONFIG_INT128 is not defined, the current implementation of + * divu128 does not return a valid truncated quotient, so the result will + * be wrong. */ divu128(&lo, &hi, clk->period); return lo; diff --git a/include/qemu/host-utils.h b/include/qemu/host-utils.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/host-utils.h +++ b/include/qemu/host-utils.h @@ -XXX,XX +XXX,XX @@ static inline uint64_t muldiv64(uint64_t a, uint32_t b, uint32_t c) return (__int128_t)a * b / c; } -static inline int divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) +static inline void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) { - if (divisor == 0) { - return 1; - } else { - __uint128_t dividend = ((__uint128_t)*phigh << 64) | *plow; - __uint128_t result = dividend / divisor; - *plow = result; - *phigh = dividend % divisor; - return result > UINT64_MAX; - } + __uint128_t dividend = ((__uint128_t)*phigh << 64) | *plow; + __uint128_t result = dividend / divisor; + *plow = result; + *phigh = dividend % divisor; } -static inline int divs128(int64_t *plow, int64_t *phigh, int64_t divisor) +static inline void divs128(int64_t *plow, int64_t *phigh, int64_t divisor) { - if (divisor == 0) { - return 1; - } else { - __int128_t dividend = ((__int128_t)*phigh << 64) | (uint64_t)*plow; - __int128_t result = dividend / divisor; - *plow = result; - *phigh = dividend % divisor; - return result != *plow; - } + __int128_t dividend = ((__int128_t)*phigh << 64) | (uint64_t)*plow; + __int128_t result = dividend / divisor; + *plow = result; + *phigh = dividend % divisor; } #else void muls64(uint64_t *plow, uint64_t *phigh, int64_t a, int64_t b); void mulu64(uint64_t *plow, uint64_t *phigh, uint64_t a, uint64_t b); -int divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor); -int divs128(int64_t *plow, int64_t *phigh, int64_t divisor); +void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor); +void divs128(int64_t *plow, int64_t *phigh, int64_t divisor); static inline uint64_t muldiv64(uint64_t a, uint32_t b, uint32_t c) { diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/int_helper.c +++ b/target/ppc/int_helper.c @@ -XXX,XX +XXX,XX @@ uint64_t helper_divdeu(CPUPPCState *env, uint64_t ra, uint64_t rb, uint32_t oe) uint64_t rt = 0; int overflow = 0; - overflow = divu128(&rt, &ra, rb); - - if (unlikely(overflow)) { + if (unlikely(rb == 0 || ra >= rb)) { + overflow = 1; rt = 0; /* Undefined */ + } else { + divu128(&rt, &ra, rb); } if (oe) { @@ -XXX,XX +XXX,XX @@ uint64_t helper_divde(CPUPPCState *env, uint64_t rau, uint64_t rbu, uint32_t oe) int64_t rt = 0; int64_t ra = (int64_t)rau; int64_t rb = (int64_t)rbu; - int overflow = divs128(&rt, &ra, rb); + int overflow = 0; - if (unlikely(overflow)) { + if (unlikely(rb == 0 || uabs64(ra) >= uabs64(rb))) { + overflow = 1; rt = 0; /* Undefined */ + } else { + divs128(&rt, &ra, rb); } if (oe) { diff --git a/util/host-utils.c b/util/host-utils.c index XXXXXXX..XXXXXXX 100644 --- a/util/host-utils.c +++ b/util/host-utils.c @@ -XXX,XX +XXX,XX @@ void muls64 (uint64_t *plow, uint64_t *phigh, int64_t a, int64_t b) *phigh = rh; } -/* Unsigned 128x64 division. Returns 1 if overflow (divide by zero or */ -/* quotient exceeds 64 bits). Otherwise returns quotient via plow and */ -/* remainder via phigh. */ -int divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) +/* + * Unsigned 128-by-64 division. Returns quotient via plow and + * remainder via phigh. + * The result must fit in 64 bits (plow) - otherwise, the result + * is undefined. + * This function will cause a division by zero if passed a zero divisor. + */ +void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) { uint64_t dhi = *phigh; uint64_t dlo = *plow; unsigned i; uint64_t carry = 0; - if (divisor == 0) { - return 1; - } else if (dhi == 0) { + if (divisor == 0 || dhi == 0) { *plow = dlo / divisor; *phigh = dlo % divisor; - return 0; - } else if (dhi >= divisor) { - return 1; } else { for (i = 0; i < 64; i++) { @@ -XXX,XX +XXX,XX @@ int divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) *plow = dlo; *phigh = dhi; - return 0; } } -int divs128(int64_t *plow, int64_t *phigh, int64_t divisor) +/* + * Signed 128-by-64 division. Returns quotient via plow and + * remainder via phigh. + * The result must fit in 64 bits (plow) - otherwise, the result + * is undefined. + * This function will cause a division by zero if passed a zero divisor. + */ +void divs128(int64_t *plow, int64_t *phigh, int64_t divisor) { int sgn_dvdnd = *phigh < 0; int sgn_divsr = divisor < 0; - int overflow = 0; if (sgn_dvdnd) { *plow = ~(*plow); @@ -XXX,XX +XXX,XX @@ int divs128(int64_t *plow, int64_t *phigh, int64_t divisor) divisor = 0 - divisor; } - overflow = divu128((uint64_t *)plow, (uint64_t *)phigh, (uint64_t)divisor); + divu128((uint64_t *)plow, (uint64_t *)phigh, (uint64_t)divisor); if (sgn_dvdnd ^ sgn_divsr) { *plow = 0 - *plow; } - - if (!overflow) { - if ((*plow < 0) ^ (sgn_dvdnd ^ sgn_divsr)) { - overflow = 1; - } - } - - return overflow; } #endif -- 2.25.1 From: Luis Pires <luis.pires@eldorado.org.br> Move udiv_qrnnd() from include/fpu/softfloat-macros.h to host-utils, so it can be reused by divu128(). Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <20211025191154.350831-3-luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/fpu/softfloat-macros.h | 82 ---------------------------------- include/qemu/host-utils.h | 81 +++++++++++++++++++++++++++++++++ 2 files changed, 81 insertions(+), 82 deletions(-) diff --git a/include/fpu/softfloat-macros.h b/include/fpu/softfloat-macros.h index XXXXXXX..XXXXXXX 100644 --- a/include/fpu/softfloat-macros.h +++ b/include/fpu/softfloat-macros.h @@ -XXX,XX +XXX,XX @@ * so some portions are provided under: * the SoftFloat-2a license * the BSD license - * GPL-v2-or-later * * Any future contributions to this file after December 1st 2014 will be * taken to be licensed under the Softfloat-2a license unless specifically @@ -XXX,XX +XXX,XX @@ this code that are retained. * THE POSSIBILITY OF SUCH DAMAGE. */ -/* Portions of this work are licensed under the terms of the GNU GPL, - * version 2 or later. See the COPYING file in the top-level directory. - */ - #ifndef FPU_SOFTFLOAT_MACROS_H #define FPU_SOFTFLOAT_MACROS_H @@ -XXX,XX +XXX,XX @@ static inline uint64_t estimateDiv128To64(uint64_t a0, uint64_t a1, uint64_t b) } -/* From the GNU Multi Precision Library - longlong.h __udiv_qrnnd - * (https://gmplib.org/repo/gmp/file/tip/longlong.h) - * - * Licensed under the GPLv2/LGPLv3 - */ -static inline uint64_t udiv_qrnnd(uint64_t *r, uint64_t n1, - uint64_t n0, uint64_t d) -{ -#if defined(__x86_64__) - uint64_t q; - asm("divq %4" : "=a"(q), "=d"(*r) : "0"(n0), "1"(n1), "rm"(d)); - return q; -#elif defined(__s390x__) && !defined(__clang__) - /* Need to use a TImode type to get an even register pair for DLGR. */ - unsigned __int128 n = (unsigned __int128)n1 << 64 | n0; - asm("dlgr %0, %1" : "+r"(n) : "r"(d)); - *r = n >> 64; - return n; -#elif defined(_ARCH_PPC64) && defined(_ARCH_PWR7) - /* From Power ISA 2.06, programming note for divdeu. */ - uint64_t q1, q2, Q, r1, r2, R; - asm("divdeu %0,%2,%4; divdu %1,%3,%4" - : "=&r"(q1), "=r"(q2) - : "r"(n1), "r"(n0), "r"(d)); - r1 = -(q1 * d); /* low part of (n1<<64) - (q1 * d) */ - r2 = n0 - (q2 * d); - Q = q1 + q2; - R = r1 + r2; - if (R >= d || R < r2) { /* overflow implies R > d */ - Q += 1; - R -= d; - } - *r = R; - return Q; -#else - uint64_t d0, d1, q0, q1, r1, r0, m; - - d0 = (uint32_t)d; - d1 = d >> 32; - - r1 = n1 % d1; - q1 = n1 / d1; - m = q1 * d0; - r1 = (r1 << 32) | (n0 >> 32); - if (r1 < m) { - q1 -= 1; - r1 += d; - if (r1 >= d) { - if (r1 < m) { - q1 -= 1; - r1 += d; - } - } - } - r1 -= m; - - r0 = r1 % d1; - q0 = r1 / d1; - m = q0 * d0; - r0 = (r0 << 32) | (uint32_t)n0; - if (r0 < m) { - q0 -= 1; - r0 += d; - if (r0 >= d) { - if (r0 < m) { - q0 -= 1; - r0 += d; - } - } - } - r0 -= m; - - *r = r0; - return (q1 << 32) | q0; -#endif -} - /*---------------------------------------------------------------------------- | Returns an approximation to the square root of the 32-bit significand given | by `a'. Considered as an integer, `a' must be at least 2^31. If bit 0 of diff --git a/include/qemu/host-utils.h b/include/qemu/host-utils.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/host-utils.h +++ b/include/qemu/host-utils.h @@ -XXX,XX +XXX,XX @@ * THE SOFTWARE. */ +/* Portions of this work are licensed under the terms of the GNU GPL, + * version 2 or later. See the COPYING file in the top-level directory. + */ + #ifndef HOST_UTILS_H #define HOST_UTILS_H @@ -XXX,XX +XXX,XX @@ void urshift(uint64_t *plow, uint64_t *phigh, int32_t shift); */ void ulshift(uint64_t *plow, uint64_t *phigh, int32_t shift, bool *overflow); +/* From the GNU Multi Precision Library - longlong.h __udiv_qrnnd + * (https://gmplib.org/repo/gmp/file/tip/longlong.h) + * + * Licensed under the GPLv2/LGPLv3 + */ +static inline uint64_t udiv_qrnnd(uint64_t *r, uint64_t n1, + uint64_t n0, uint64_t d) +{ +#if defined(__x86_64__) + uint64_t q; + asm("divq %4" : "=a"(q), "=d"(*r) : "0"(n0), "1"(n1), "rm"(d)); + return q; +#elif defined(__s390x__) && !defined(__clang__) + /* Need to use a TImode type to get an even register pair for DLGR. */ + unsigned __int128 n = (unsigned __int128)n1 << 64 | n0; + asm("dlgr %0, %1" : "+r"(n) : "r"(d)); + *r = n >> 64; + return n; +#elif defined(_ARCH_PPC64) && defined(_ARCH_PWR7) + /* From Power ISA 2.06, programming note for divdeu. */ + uint64_t q1, q2, Q, r1, r2, R; + asm("divdeu %0,%2,%4; divdu %1,%3,%4" + : "=&r"(q1), "=r"(q2) + : "r"(n1), "r"(n0), "r"(d)); + r1 = -(q1 * d); /* low part of (n1<<64) - (q1 * d) */ + r2 = n0 - (q2 * d); + Q = q1 + q2; + R = r1 + r2; + if (R >= d || R < r2) { /* overflow implies R > d */ + Q += 1; + R -= d; + } + *r = R; + return Q; +#else + uint64_t d0, d1, q0, q1, r1, r0, m; + + d0 = (uint32_t)d; + d1 = d >> 32; + + r1 = n1 % d1; + q1 = n1 / d1; + m = q1 * d0; + r1 = (r1 << 32) | (n0 >> 32); + if (r1 < m) { + q1 -= 1; + r1 += d; + if (r1 >= d) { + if (r1 < m) { + q1 -= 1; + r1 += d; + } + } + } + r1 -= m; + + r0 = r1 % d1; + q0 = r1 / d1; + m = q0 * d0; + r0 = (r0 << 32) | (uint32_t)n0; + if (r0 < m) { + q0 -= 1; + r0 += d; + if (r0 >= d) { + if (r0 < m) { + q0 -= 1; + r0 += d; + } + } + } + r0 -= m; + + *r = r0; + return (q1 << 32) | q0; +#endif +} + #endif -- 2.25.1 From: Luis Pires <luis.pires@eldorado.org.br> These will be used to implement new decimal floating point instructions from Power ISA 3.1. The remainder is now returned directly by divu128/divs128, freeing up phigh to receive the high 64 bits of the quotient. Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <20211025191154.350831-4-luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/hw/clock.h | 6 +- include/qemu/host-utils.h | 20 ++++-- target/ppc/int_helper.c | 9 +-- util/host-utils.c | 133 +++++++++++++++++++++++++------------- 4 files changed, 108 insertions(+), 60 deletions(-) diff --git a/include/hw/clock.h b/include/hw/clock.h index XXXXXXX..XXXXXXX 100644 --- a/include/hw/clock.h +++ b/include/hw/clock.h @@ -XXX,XX +XXX,XX @@ static inline uint64_t clock_ns_to_ticks(const Clock *clk, uint64_t ns) if (clk->period == 0) { return 0; } - /* - * BUG: when CONFIG_INT128 is not defined, the current implementation of - * divu128 does not return a valid truncated quotient, so the result will - * be wrong. - */ + divu128(&lo, &hi, clk->period); return lo; } diff --git a/include/qemu/host-utils.h b/include/qemu/host-utils.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/host-utils.h +++ b/include/qemu/host-utils.h @@ -XXX,XX +XXX,XX @@ static inline uint64_t muldiv64(uint64_t a, uint32_t b, uint32_t c) return (__int128_t)a * b / c; } -static inline void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) +static inline uint64_t divu128(uint64_t *plow, uint64_t *phigh, + uint64_t divisor) { __uint128_t dividend = ((__uint128_t)*phigh << 64) | *plow; __uint128_t result = dividend / divisor; + *plow = result; - *phigh = dividend % divisor; + *phigh = result >> 64; + return dividend % divisor; } -static inline void divs128(int64_t *plow, int64_t *phigh, int64_t divisor) +static inline int64_t divs128(uint64_t *plow, int64_t *phigh, + int64_t divisor) { - __int128_t dividend = ((__int128_t)*phigh << 64) | (uint64_t)*plow; + __int128_t dividend = ((__int128_t)*phigh << 64) | *plow; __int128_t result = dividend / divisor; + *plow = result; - *phigh = dividend % divisor; + *phigh = result >> 64; + return dividend % divisor; } #else void muls64(uint64_t *plow, uint64_t *phigh, int64_t a, int64_t b); void mulu64(uint64_t *plow, uint64_t *phigh, uint64_t a, uint64_t b); -void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor); -void divs128(int64_t *plow, int64_t *phigh, int64_t divisor); +uint64_t divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor); +int64_t divs128(uint64_t *plow, int64_t *phigh, int64_t divisor); static inline uint64_t muldiv64(uint64_t a, uint32_t b, uint32_t c) { diff --git a/target/ppc/int_helper.c b/target/ppc/int_helper.c index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/int_helper.c +++ b/target/ppc/int_helper.c @@ -XXX,XX +XXX,XX @@ uint64_t helper_divdeu(CPUPPCState *env, uint64_t ra, uint64_t rb, uint32_t oe) uint64_t helper_divde(CPUPPCState *env, uint64_t rau, uint64_t rbu, uint32_t oe) { - int64_t rt = 0; + uint64_t rt = 0; int64_t ra = (int64_t)rau; int64_t rb = (int64_t)rbu; int overflow = 0; @@ -XXX,XX +XXX,XX @@ uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) int cr; uint64_t lo_value; uint64_t hi_value; + uint64_t rem; ppc_avr_t ret = { .u64 = { 0, 0 } }; if (b->VsrSD(0) < 0) { @@ -XXX,XX +XXX,XX @@ uint32_t helper_bcdcfsq(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps) * In that case, we leave r unchanged. */ } else { - divu128(&lo_value, &hi_value, 1000000000000000ULL); + rem = divu128(&lo_value, &hi_value, 1000000000000000ULL); - for (i = 1; i < 16; hi_value /= 10, i++) { - bcd_put_digit(&ret, hi_value % 10, i); + for (i = 1; i < 16; rem /= 10, i++) { + bcd_put_digit(&ret, rem % 10, i); } for (; i < 32; lo_value /= 10, i++) { diff --git a/util/host-utils.c b/util/host-utils.c index XXXXXXX..XXXXXXX 100644 --- a/util/host-utils.c +++ b/util/host-utils.c @@ -XXX,XX +XXX,XX @@ void muls64 (uint64_t *plow, uint64_t *phigh, int64_t a, int64_t b) } /* - * Unsigned 128-by-64 division. Returns quotient via plow and - * remainder via phigh. - * The result must fit in 64 bits (plow) - otherwise, the result - * is undefined. - * This function will cause a division by zero if passed a zero divisor. + * Unsigned 128-by-64 division. + * Returns the remainder. + * Returns quotient via plow and phigh. + * Also returns the remainder via the function return value. */ -void divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) +uint64_t divu128(uint64_t *plow, uint64_t *phigh, uint64_t divisor) { uint64_t dhi = *phigh; uint64_t dlo = *plow; - unsigned i; - uint64_t carry = 0; + uint64_t rem, dhighest; + int sh; if (divisor == 0 || dhi == 0) { *plow = dlo / divisor; - *phigh = dlo % divisor; + *phigh = 0; + return dlo % divisor; } else { + sh = clz64(divisor); - for (i = 0; i < 64; i++) { - carry = dhi >> 63; - dhi = (dhi << 1) | (dlo >> 63); - if (carry || (dhi >= divisor)) { - dhi -= divisor; - carry = 1; - } else { - carry = 0; + if (dhi < divisor) { + if (sh != 0) { + /* normalize the divisor, shifting the dividend accordingly */ + divisor <<= sh; + dhi = (dhi << sh) | (dlo >> (64 - sh)); + dlo <<= sh; } - dlo = (dlo << 1) | carry; + + *phigh = 0; + *plow = udiv_qrnnd(&rem, dhi, dlo, divisor); + } else { + if (sh != 0) { + /* normalize the divisor, shifting the dividend accordingly */ + divisor <<= sh; + dhighest = dhi >> (64 - sh); + dhi = (dhi << sh) | (dlo >> (64 - sh)); + dlo <<= sh; + + *phigh = udiv_qrnnd(&dhi, dhighest, dhi, divisor); + } else { + /** + * dhi >= divisor + * Since the MSB of divisor is set (sh == 0), + * (dhi - divisor) < divisor + * + * Thus, the high part of the quotient is 1, and we can + * calculate the low part with a single call to udiv_qrnnd + * after subtracting divisor from dhi + */ + dhi -= divisor; + *phigh = 1; + } + + *plow = udiv_qrnnd(&rem, dhi, dlo, divisor); } - *plow = dlo; - *phigh = dhi; + /* + * since the dividend/divisor might have been normalized, + * the remainder might also have to be shifted back + */ + return rem >> sh; } } /* - * Signed 128-by-64 division. Returns quotient via plow and - * remainder via phigh. - * The result must fit in 64 bits (plow) - otherwise, the result - * is undefined. - * This function will cause a division by zero if passed a zero divisor. + * Signed 128-by-64 division. + * Returns quotient via plow and phigh. + * Also returns the remainder via the function return value. */ -void divs128(int64_t *plow, int64_t *phigh, int64_t divisor) +int64_t divs128(uint64_t *plow, int64_t *phigh, int64_t divisor) { - int sgn_dvdnd = *phigh < 0; - int sgn_divsr = divisor < 0; + bool neg_quotient = false, neg_remainder = false; + uint64_t unsig_hi = *phigh, unsig_lo = *plow; + uint64_t rem; - if (sgn_dvdnd) { - *plow = ~(*plow); - *phigh = ~(*phigh); - if (*plow == (int64_t)-1) { + if (*phigh < 0) { + neg_quotient = !neg_quotient; + neg_remainder = !neg_remainder; + + if (unsig_lo == 0) { + unsig_hi = -unsig_hi; + } else { + unsig_hi = ~unsig_hi; + unsig_lo = -unsig_lo; + } + } + + if (divisor < 0) { + neg_quotient = !neg_quotient; + + divisor = -divisor; + } + + rem = divu128(&unsig_lo, &unsig_hi, (uint64_t)divisor); + + if (neg_quotient) { + if (unsig_lo == 0) { + *phigh = -unsig_hi; *plow = 0; - (*phigh)++; - } else { - (*plow)++; - } + } else { + *phigh = ~unsig_hi; + *plow = -unsig_lo; + } + } else { + *phigh = unsig_hi; + *plow = unsig_lo; } - if (sgn_divsr) { - divisor = 0 - divisor; - } - - divu128((uint64_t *)plow, (uint64_t *)phigh, (uint64_t)divisor); - - if (sgn_dvdnd ^ sgn_divsr) { - *plow = 0 - *plow; + if (neg_remainder) { + return -rem; + } else { + return rem; } } #endif -- 2.25.1 From: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Richard Henderson <richard.henderson@linaro.org> Message-Id: <20211025191154.350831-5-luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tests/unit/test-div128.c | 197 +++++++++++++++++++++++++++++++++++++++ tests/unit/meson.build | 1 + 2 files changed, 198 insertions(+) create mode 100644 tests/unit/test-div128.c diff --git a/tests/unit/test-div128.c b/tests/unit/test-div128.c new file mode 100644 index XXXXXXX..XXXXXXX --- /dev/null +++ b/tests/unit/test-div128.c @@ -XXX,XX +XXX,XX @@ +/* + * Test 128-bit division functions + * + * Copyright (c) 2021 Instituto de Pesquisas Eldorado (eldorado.org.br) + * + * This library is free software; you can redistribute it and/or + * modify it under the terms of the GNU Lesser General Public + * License as published by the Free Software Foundation; either + * version 2.1 of the License, or (at your option) any later version. + * + * This library is distributed in the hope that it will be useful, + * but WITHOUT ANY WARRANTY; without even the implied warranty of + * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + * Lesser General Public License for more details. + * + * You should have received a copy of the GNU Lesser General Public + * License along with this library; if not, see <http://www.gnu.org/licenses/>. + */ + +#include "qemu/osdep.h" +#include "qemu/host-utils.h" + +typedef struct { + uint64_t high; + uint64_t low; + uint64_t rhigh; + uint64_t rlow; + uint64_t divisor; + uint64_t remainder; +} test_data_unsigned; + +typedef struct { + int64_t high; + uint64_t low; + int64_t rhigh; + uint64_t rlow; + int64_t divisor; + int64_t remainder; +} test_data_signed; + +static const test_data_unsigned test_table_unsigned[] = { + /* Dividend fits in 64 bits */ + { 0x0000000000000000ULL, 0x0000000000000000ULL, + 0x0000000000000000ULL, 0x0000000000000000ULL, + 0x0000000000000001ULL, 0x0000000000000000ULL}, + { 0x0000000000000000ULL, 0x0000000000000001ULL, + 0x0000000000000000ULL, 0x0000000000000001ULL, + 0x0000000000000001ULL, 0x0000000000000000ULL}, + { 0x0000000000000000ULL, 0x0000000000000003ULL, + 0x0000000000000000ULL, 0x0000000000000001ULL, + 0x0000000000000002ULL, 0x0000000000000001ULL}, + { 0x0000000000000000ULL, 0x8000000000000000ULL, + 0x0000000000000000ULL, 0x8000000000000000ULL, + 0x0000000000000001ULL, 0x0000000000000000ULL}, + { 0x0000000000000000ULL, 0xa000000000000000ULL, + 0x0000000000000000ULL, 0x0000000000000002ULL, + 0x4000000000000000ULL, 0x2000000000000000ULL}, + { 0x0000000000000000ULL, 0x8000000000000000ULL, + 0x0000000000000000ULL, 0x0000000000000001ULL, + 0x8000000000000000ULL, 0x0000000000000000ULL}, + + /* Dividend > 64 bits, with MSB 0 */ + { 0x123456789abcdefeULL, 0xefedcba987654321ULL, + 0x123456789abcdefeULL, 0xefedcba987654321ULL, + 0x0000000000000001ULL, 0x0000000000000000ULL}, + { 0x123456789abcdefeULL, 0xefedcba987654321ULL, + 0x0000000000000001ULL, 0x000000000000000dULL, + 0x123456789abcdefeULL, 0x03456789abcdf03bULL}, + { 0x123456789abcdefeULL, 0xefedcba987654321ULL, + 0x0123456789abcdefULL, 0xeefedcba98765432ULL, + 0x0000000000000010ULL, 0x0000000000000001ULL}, + + /* Dividend > 64 bits, with MSB 1 */ + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x0000000000000001ULL, 0x0000000000000000ULL}, + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x0000000000000001ULL, 0x0000000000000000ULL, + 0xfeeddccbbaa99887ULL, 0x766554433221100fULL}, + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x0feeddccbbaa9988ULL, 0x7766554433221100ULL, + 0x0000000000000010ULL, 0x000000000000000fULL}, + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x000000000000000eULL, 0x00f0f0f0f0f0f35aULL, + 0x123456789abcdefeULL, 0x0f8922bc55ef90c3ULL}, + + /** + * Divisor == 64 bits, with MSB 1 + * and high 64 bits of dividend >= divisor + * (for testing normalization) + */ + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x0000000000000001ULL, 0x0000000000000000ULL, + 0xfeeddccbbaa99887ULL, 0x766554433221100fULL}, + { 0xfeeddccbbaa99887ULL, 0x766554433221100fULL, + 0x0000000000000001ULL, 0xfddbb9977553310aULL, + 0x8000000000000001ULL, 0x78899aabbccddf05ULL}, + + /* Dividend > 64 bits, divisor almost as big */ + { 0x0000000000000001ULL, 0x23456789abcdef01ULL, + 0x0000000000000000ULL, 0x000000000000000fULL, + 0x123456789abcdefeULL, 0x123456789abcde1fULL}, +}; + +static const test_data_signed test_table_signed[] = { + /* Positive dividend, positive/negative divisors */ + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0x0000000000000000LL, 0x0000000000bc614eULL, + 0x0000000000000001LL, 0x0000000000000000LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0xffffffffffffffffLL, 0x0000000000000000LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0x0000000000000000LL, 0x00000000005e30a7ULL, + 0x0000000000000002LL, 0x0000000000000000LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0xffffffffffffffffLL, 0xffffffffffa1cf59ULL, + 0xfffffffffffffffeLL, 0x0000000000000000LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0x0000000000000000LL, 0x0000000000178c29ULL, + 0x0000000000000008LL, 0x0000000000000006LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0xffffffffffffffffLL, 0xffffffffffe873d7ULL, + 0xfffffffffffffff8LL, 0x0000000000000006LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0x0000000000000000LL, 0x000000000000550dULL, + 0x0000000000000237LL, 0x0000000000000183LL}, + { 0x0000000000000000LL, 0x0000000000bc614eULL, + 0xffffffffffffffffLL, 0xffffffffffffaaf3ULL, + 0xfffffffffffffdc9LL, 0x0000000000000183LL}, + + /* Negative dividend, positive/negative divisors */ + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0x0000000000000001LL, 0x0000000000000000LL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0x0000000000000000LL, 0x0000000000bc614eULL, + 0xffffffffffffffffLL, 0x0000000000000000LL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0xffffffffffffffffLL, 0xffffffffffa1cf59ULL, + 0x0000000000000002LL, 0x0000000000000000LL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0x0000000000000000LL, 0x00000000005e30a7ULL, + 0xfffffffffffffffeLL, 0x0000000000000000LL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0xffffffffffffffffLL, 0xffffffffffe873d7ULL, + 0x0000000000000008LL, 0xfffffffffffffffaLL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0x0000000000000000LL, 0x0000000000178c29ULL, + 0xfffffffffffffff8LL, 0xfffffffffffffffaLL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0xffffffffffffffffLL, 0xffffffffffffaaf3ULL, + 0x0000000000000237LL, 0xfffffffffffffe7dLL}, + { 0xffffffffffffffffLL, 0xffffffffff439eb2ULL, + 0x0000000000000000LL, 0x000000000000550dULL, + 0xfffffffffffffdc9LL, 0xfffffffffffffe7dLL}, +}; + +static void test_divu128(void) +{ + int i; + uint64_t rem; + test_data_unsigned tmp; + + for (i = 0; i < ARRAY_SIZE(test_table_unsigned); ++i) { + tmp = test_table_unsigned[i]; + + rem = divu128(&tmp.low, &tmp.high, tmp.divisor); + g_assert_cmpuint(tmp.low, ==, tmp.rlow); + g_assert_cmpuint(tmp.high, ==, tmp.rhigh); + g_assert_cmpuint(rem, ==, tmp.remainder); + } +} + +static void test_divs128(void) +{ + int i; + int64_t rem; + test_data_signed tmp; + + for (i = 0; i < ARRAY_SIZE(test_table_signed); ++i) { + tmp = test_table_signed[i]; + + rem = divs128(&tmp.low, &tmp.high, tmp.divisor); + g_assert_cmpuint(tmp.low, ==, tmp.rlow); + g_assert_cmpuint(tmp.high, ==, tmp.rhigh); + g_assert_cmpuint(rem, ==, tmp.remainder); + } +} + +int main(int argc, char **argv) +{ + g_test_init(&argc, &argv, NULL); + g_test_add_func("/host-utils/test_divu128", test_divu128); + g_test_add_func("/host-utils/test_divs128", test_divs128); + return g_test_run(); +} diff --git a/tests/unit/meson.build b/tests/unit/meson.build index XXXXXXX..XXXXXXX 100644 --- a/tests/unit/meson.build +++ b/tests/unit/meson.build @@ -XXX,XX +XXX,XX @@ tests = { # all code tested by test-x86-cpuid is inside topology.h 'test-x86-cpuid': [], 'test-cutils': [], + 'test-div128': [], 'test-shift128': [], 'test-mul64': [], # all code tested by test-int128 is inside int128.h -- 2.25.1 Prepare for tracking different masks by renaming this one. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 142 +++++++++++++++++++++++++------------------------ 1 file changed, 72 insertions(+), 70 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct TempOptInfo { TCGTemp *prev_copy; TCGTemp *next_copy; uint64_t val; - uint64_t mask; + uint64_t z_mask; /* mask bit is 0 if and only if value bit is 0 */ } TempOptInfo; static inline TempOptInfo *ts_info(TCGTemp *ts) @@ -XXX,XX +XXX,XX @@ static void reset_ts(TCGTemp *ts) ti->next_copy = ts; ti->prev_copy = ts; ti->is_const = false; - ti->mask = -1; + ti->z_mask = -1; } static void reset_temp(TCGArg arg) @@ -XXX,XX +XXX,XX @@ static void init_ts_info(TCGTempSet *temps_used, TCGTemp *ts) if (ts->kind == TEMP_CONST) { ti->is_const = true; ti->val = ts->val; - ti->mask = ts->val; + ti->z_mask = ts->val; if (TCG_TARGET_REG_BITS > 32 && ts->type == TCG_TYPE_I32) { /* High bits of a 32-bit quantity are garbage. */ - ti->mask |= ~0xffffffffull; + ti->z_mask |= ~0xffffffffull; } } else { ti->is_const = false; - ti->mask = -1; + ti->z_mask = -1; } } @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) const TCGOpDef *def; TempOptInfo *di; TempOptInfo *si; - uint64_t mask; + uint64_t z_mask; TCGOpcode new_op; if (ts_are_copies(dst_ts, src_ts)) { @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) op->args[0] = dst; op->args[1] = src; - mask = si->mask; + z_mask = si->z_mask; if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) { /* High bits of the destination are now garbage. */ - mask |= ~0xffffffffull; + z_mask |= ~0xffffffffull; } - di->mask = mask; + di->z_mask = z_mask; if (src_ts->type == dst_ts->type) { TempOptInfo *ni = ts_info(si->next_copy); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { - uint64_t mask, partmask, affected, tmp; + uint64_t z_mask, partmask, affected, tmp; int nb_oargs, nb_iargs; TCGOpcode opc = op->opc; const TCGOpDef *def = &tcg_op_defs[opc]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* Simplify using known-zero bits. Currently only ops with a single output argument is supported. */ - mask = -1; + z_mask = -1; affected = -1; switch (opc) { CASE_OP_32_64(ext8s): - if ((arg_info(op->args[1])->mask & 0x80) != 0) { + if ((arg_info(op->args[1])->z_mask & 0x80) != 0) { break; } QEMU_FALLTHROUGH; CASE_OP_32_64(ext8u): - mask = 0xff; + z_mask = 0xff; goto and_const; CASE_OP_32_64(ext16s): - if ((arg_info(op->args[1])->mask & 0x8000) != 0) { + if ((arg_info(op->args[1])->z_mask & 0x8000) != 0) { break; } QEMU_FALLTHROUGH; CASE_OP_32_64(ext16u): - mask = 0xffff; + z_mask = 0xffff; goto and_const; case INDEX_op_ext32s_i64: - if ((arg_info(op->args[1])->mask & 0x80000000) != 0) { + if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) { break; } QEMU_FALLTHROUGH; case INDEX_op_ext32u_i64: - mask = 0xffffffffU; + z_mask = 0xffffffffU; goto and_const; CASE_OP_32_64(and): - mask = arg_info(op->args[2])->mask; + z_mask = arg_info(op->args[2])->z_mask; if (arg_is_const(op->args[2])) { and_const: - affected = arg_info(op->args[1])->mask & ~mask; + affected = arg_info(op->args[1])->z_mask & ~z_mask; } - mask = arg_info(op->args[1])->mask & mask; + z_mask = arg_info(op->args[1])->z_mask & z_mask; break; case INDEX_op_ext_i32_i64: - if ((arg_info(op->args[1])->mask & 0x80000000) != 0) { + if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) { break; } QEMU_FALLTHROUGH; case INDEX_op_extu_i32_i64: /* We do not compute affected as it is a size changing op. */ - mask = (uint32_t)arg_info(op->args[1])->mask; + z_mask = (uint32_t)arg_info(op->args[1])->z_mask; break; CASE_OP_32_64(andc): /* Known-zeros does not imply known-ones. Therefore unless op->args[2] is constant, we can't infer anything from it. */ if (arg_is_const(op->args[2])) { - mask = ~arg_info(op->args[2])->mask; + z_mask = ~arg_info(op->args[2])->z_mask; goto and_const; } /* But we certainly know nothing outside args[1] may be set. */ - mask = arg_info(op->args[1])->mask; + z_mask = arg_info(op->args[1])->z_mask; break; case INDEX_op_sar_i32: if (arg_is_const(op->args[2])) { tmp = arg_info(op->args[2])->val & 31; - mask = (int32_t)arg_info(op->args[1])->mask >> tmp; + z_mask = (int32_t)arg_info(op->args[1])->z_mask >> tmp; } break; case INDEX_op_sar_i64: if (arg_is_const(op->args[2])) { tmp = arg_info(op->args[2])->val & 63; - mask = (int64_t)arg_info(op->args[1])->mask >> tmp; + z_mask = (int64_t)arg_info(op->args[1])->z_mask >> tmp; } break; case INDEX_op_shr_i32: if (arg_is_const(op->args[2])) { tmp = arg_info(op->args[2])->val & 31; - mask = (uint32_t)arg_info(op->args[1])->mask >> tmp; + z_mask = (uint32_t)arg_info(op->args[1])->z_mask >> tmp; } break; case INDEX_op_shr_i64: if (arg_is_const(op->args[2])) { tmp = arg_info(op->args[2])->val & 63; - mask = (uint64_t)arg_info(op->args[1])->mask >> tmp; + z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> tmp; } break; case INDEX_op_extrl_i64_i32: - mask = (uint32_t)arg_info(op->args[1])->mask; + z_mask = (uint32_t)arg_info(op->args[1])->z_mask; break; case INDEX_op_extrh_i64_i32: - mask = (uint64_t)arg_info(op->args[1])->mask >> 32; + z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> 32; break; CASE_OP_32_64(shl): if (arg_is_const(op->args[2])) { tmp = arg_info(op->args[2])->val & (TCG_TARGET_REG_BITS - 1); - mask = arg_info(op->args[1])->mask << tmp; + z_mask = arg_info(op->args[1])->z_mask << tmp; } break; CASE_OP_32_64(neg): /* Set to 1 all bits to the left of the rightmost. */ - mask = -(arg_info(op->args[1])->mask - & -arg_info(op->args[1])->mask); + z_mask = -(arg_info(op->args[1])->z_mask + & -arg_info(op->args[1])->z_mask); break; CASE_OP_32_64(deposit): - mask = deposit64(arg_info(op->args[1])->mask, - op->args[3], op->args[4], - arg_info(op->args[2])->mask); + z_mask = deposit64(arg_info(op->args[1])->z_mask, + op->args[3], op->args[4], + arg_info(op->args[2])->z_mask); break; CASE_OP_32_64(extract): - mask = extract64(arg_info(op->args[1])->mask, - op->args[2], op->args[3]); + z_mask = extract64(arg_info(op->args[1])->z_mask, + op->args[2], op->args[3]); if (op->args[2] == 0) { - affected = arg_info(op->args[1])->mask & ~mask; + affected = arg_info(op->args[1])->z_mask & ~z_mask; } break; CASE_OP_32_64(sextract): - mask = sextract64(arg_info(op->args[1])->mask, - op->args[2], op->args[3]); - if (op->args[2] == 0 && (tcg_target_long)mask >= 0) { - affected = arg_info(op->args[1])->mask & ~mask; + z_mask = sextract64(arg_info(op->args[1])->z_mask, + op->args[2], op->args[3]); + if (op->args[2] == 0 && (tcg_target_long)z_mask >= 0) { + affected = arg_info(op->args[1])->z_mask & ~z_mask; } break; CASE_OP_32_64(or): CASE_OP_32_64(xor): - mask = arg_info(op->args[1])->mask | arg_info(op->args[2])->mask; + z_mask = arg_info(op->args[1])->z_mask + | arg_info(op->args[2])->z_mask; break; case INDEX_op_clz_i32: case INDEX_op_ctz_i32: - mask = arg_info(op->args[2])->mask | 31; + z_mask = arg_info(op->args[2])->z_mask | 31; break; case INDEX_op_clz_i64: case INDEX_op_ctz_i64: - mask = arg_info(op->args[2])->mask | 63; + z_mask = arg_info(op->args[2])->z_mask | 63; break; case INDEX_op_ctpop_i32: - mask = 32 | 31; + z_mask = 32 | 31; break; case INDEX_op_ctpop_i64: - mask = 64 | 63; + z_mask = 64 | 63; break; CASE_OP_32_64(setcond): case INDEX_op_setcond2_i32: - mask = 1; + z_mask = 1; break; CASE_OP_32_64(movcond): - mask = arg_info(op->args[3])->mask | arg_info(op->args[4])->mask; + z_mask = arg_info(op->args[3])->z_mask + | arg_info(op->args[4])->z_mask; break; CASE_OP_32_64(ld8u): - mask = 0xff; + z_mask = 0xff; break; CASE_OP_32_64(ld16u): - mask = 0xffff; + z_mask = 0xffff; break; case INDEX_op_ld32u_i64: - mask = 0xffffffffu; + z_mask = 0xffffffffu; break; CASE_OP_32_64(qemu_ld): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) MemOpIdx oi = op->args[nb_oargs + nb_iargs]; MemOp mop = get_memop(oi); if (!(mop & MO_SIGN)) { - mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1; + z_mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1; } } break; CASE_OP_32_64(bswap16): - mask = arg_info(op->args[1])->mask; - if (mask <= 0xffff) { + z_mask = arg_info(op->args[1])->z_mask; + if (z_mask <= 0xffff) { op->args[2] |= TCG_BSWAP_IZ; } - mask = bswap16(mask); + z_mask = bswap16(z_mask); switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { case TCG_BSWAP_OZ: break; case TCG_BSWAP_OS: - mask = (int16_t)mask; + z_mask = (int16_t)z_mask; break; default: /* undefined high bits */ - mask |= MAKE_64BIT_MASK(16, 48); + z_mask |= MAKE_64BIT_MASK(16, 48); break; } break; case INDEX_op_bswap32_i64: - mask = arg_info(op->args[1])->mask; - if (mask <= 0xffffffffu) { + z_mask = arg_info(op->args[1])->z_mask; + if (z_mask <= 0xffffffffu) { op->args[2] |= TCG_BSWAP_IZ; } - mask = bswap32(mask); + z_mask = bswap32(z_mask); switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { case TCG_BSWAP_OZ: break; case TCG_BSWAP_OS: - mask = (int32_t)mask; + z_mask = (int32_t)z_mask; break; default: /* undefined high bits */ - mask |= MAKE_64BIT_MASK(32, 32); + z_mask |= MAKE_64BIT_MASK(32, 32); break; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* 32-bit ops generate 32-bit results. For the result is zero test below, we can ignore high bits, but for further optimizations we need to record that the high bits contain garbage. */ - partmask = mask; + partmask = z_mask; if (!(def->flags & TCG_OPF_64BIT)) { - mask |= ~(tcg_target_ulong)0xffffffffu; + z_mask |= ~(tcg_target_ulong)0xffffffffu; partmask &= 0xffffffffu; affected &= 0xffffffffu; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) vs the high word of the input. */ do_setcond_high: reset_temp(op->args[0]); - arg_info(op->args[0])->mask = 1; + arg_info(op->args[0])->z_mask = 1; op->opc = INDEX_op_setcond_i32; op->args[1] = op->args[2]; op->args[2] = op->args[4]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } do_setcond_low: reset_temp(op->args[0]); - arg_info(op->args[0])->mask = 1; + arg_info(op->args[0])->z_mask = 1; op->opc = INDEX_op_setcond_i32; op->args[2] = op->args[3]; op->args[3] = op->args[5]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* Default case: we know nothing about operation (or were unable to compute the operation result) so no propagation is done. We trash everything if the operation is the end of a basic - block, otherwise we only trash the output args. "mask" is + block, otherwise we only trash the output args. "z_mask" is the non-zero bits mask for the first output arg. */ if (def->flags & TCG_OPF_BB_END) { memset(&temps_used, 0, sizeof(temps_used)); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* Save the corresponding known-zero bits mask for the first output argument (only one supported so far). */ if (i == 0) { - arg_info(op->args[i])->mask = mask; + arg_info(op->args[i])->z_mask = z_mask; } } } -- 2.25.1 Provide what will become a larger context for splitting the very large tcg_optimize function. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 77 ++++++++++++++++++++++++++------------------------ 1 file changed, 40 insertions(+), 37 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct TempOptInfo { uint64_t z_mask; /* mask bit is 0 if and only if value bit is 0 */ } TempOptInfo; +typedef struct OptContext { + TCGTempSet temps_used; +} OptContext; + static inline TempOptInfo *ts_info(TCGTemp *ts) { return ts->state_ptr; @@ -XXX,XX +XXX,XX @@ static void reset_temp(TCGArg arg) } /* Initialize and activate a temporary. */ -static void init_ts_info(TCGTempSet *temps_used, TCGTemp *ts) +static void init_ts_info(OptContext *ctx, TCGTemp *ts) { size_t idx = temp_idx(ts); TempOptInfo *ti; - if (test_bit(idx, temps_used->l)) { + if (test_bit(idx, ctx->temps_used.l)) { return; } - set_bit(idx, temps_used->l); + set_bit(idx, ctx->temps_used.l); ti = ts->state_ptr; if (ti == NULL) { @@ -XXX,XX +XXX,XX @@ static void init_ts_info(TCGTempSet *temps_used, TCGTemp *ts) } } -static void init_arg_info(TCGTempSet *temps_used, TCGArg arg) +static void init_arg_info(OptContext *ctx, TCGArg arg) { - init_ts_info(temps_used, arg_temp(arg)); + init_ts_info(ctx, arg_temp(arg)); } static TCGTemp *find_better_copy(TCGContext *s, TCGTemp *ts) @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) } } -static void tcg_opt_gen_movi(TCGContext *s, TCGTempSet *temps_used, +static void tcg_opt_gen_movi(TCGContext *s, OptContext *ctx, TCGOp *op, TCGArg dst, uint64_t val) { const TCGOpDef *def = &tcg_op_defs[op->opc]; @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_movi(TCGContext *s, TCGTempSet *temps_used, /* Convert movi to mov with constant temp. */ tv = tcg_constant_internal(type, val); - init_ts_info(temps_used, tv); + init_ts_info(ctx, tv); tcg_opt_gen_mov(s, op, dst, temp_arg(tv)); } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) { int nb_temps, nb_globals, i; TCGOp *op, *op_next, *prev_mb = NULL; - TCGTempSet temps_used; + OptContext ctx = {}; /* Array VALS has an element for each temp. If this temp holds a constant then its value is kept in VALS' element. @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) nb_temps = s->nb_temps; nb_globals = s->nb_globals; - memset(&temps_used, 0, sizeof(temps_used)); for (i = 0; i < nb_temps; ++i) { s->temps[i].state_ptr = NULL; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) for (i = 0; i < nb_oargs + nb_iargs; i++) { TCGTemp *ts = arg_temp(op->args[i]); if (ts) { - init_ts_info(&temps_used, ts); + init_ts_info(&ctx, ts); } } } else { nb_oargs = def->nb_oargs; nb_iargs = def->nb_iargs; for (i = 0; i < nb_oargs + nb_iargs; i++) { - init_arg_info(&temps_used, op->args[i]); + init_arg_info(&ctx, op->args[i]); } } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(rotr): if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == 0) { - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], 0); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (partmask == 0) { tcg_debug_assert(nb_oargs == 1); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], 0); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); continue; } if (affected == 0) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(mulsh): if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0) { - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], 0); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(sub): CASE_OP_32_64_VEC(xor): if (args_are_copies(op->args[1], op->args[2])) { - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], 0); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = arg_info(op->args[1])->val; tmp = dup_const(TCGOP_VECE(op), tmp); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_dup2_vec: assert(TCG_TARGET_REG_BITS == 32); if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], + tcg_opt_gen_movi(s, &ctx, op, op->args[0], deposit64(arg_info(op->args[1])->val, 32, 32, arg_info(op->args[2])->val)); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_extrh_i64_i32: if (arg_is_const(op->args[1])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, 0); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, op->args[2]); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, arg_info(op->args[2])->val); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) TCGArg v = arg_info(op->args[1])->val; if (v != 0) { tmp = do_constant_folding(opc, v, 0); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); } else { tcg_opt_gen_mov(s, op, op->args[0], op->args[2]); } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = deposit64(arg_info(op->args[1])->val, op->args[3], op->args[4], arg_info(op->args[2])->val); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = extract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = sextract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = (int32_t)(((uint32_t)v1 >> shr) | ((uint32_t)v2 << (32 - shr))); } - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[3]); if (tmp != 2) { - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[1], op->args[2]); if (tmp != 2) { if (tmp) { - memset(&temps_used, 0, sizeof(temps_used)); + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); op->opc = INDEX_op_br; op->args[0] = op->args[3]; } else { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(s, &temps_used, op, rl, (int32_t)a); - tcg_opt_gen_movi(s, &temps_used, op2, rh, (int32_t)(a >> 32)); + tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)a); + tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(a >> 32)); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(s, &temps_used, op, rl, (int32_t)r); - tcg_opt_gen_movi(s, &temps_used, op2, rh, (int32_t)(r >> 32)); + tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)r); + tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(r >> 32)); break; } goto do_default; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (tmp != 2) { if (tmp) { do_brcond_true: - memset(&temps_used, 0, sizeof(temps_used)); + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); op->opc = INDEX_op_br; op->args[0] = op->args[5]; } else { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* Simplify LT/GE comparisons vs zero to a single compare vs the high word of the input. */ do_brcond_high: - memset(&temps_used, 0, sizeof(temps_used)); + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); op->opc = INDEX_op_brcond_i32; op->args[0] = op->args[1]; op->args[1] = op->args[3]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) goto do_default; } do_brcond_low: - memset(&temps_used, 0, sizeof(temps_used)); + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); op->opc = INDEX_op_brcond_i32; op->args[1] = op->args[2]; op->args[2] = op->args[4]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[5]); if (tmp != 2) { do_setcond_const: - tcg_opt_gen_movi(s, &temps_used, op, op->args[0], tmp); + tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); } else if ((op->args[5] == TCG_COND_LT || op->args[5] == TCG_COND_GE) && arg_is_const(op->args[3]) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (!(tcg_call_flags(op) & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { for (i = 0; i < nb_globals; i++) { - if (test_bit(i, temps_used.l)) { + if (test_bit(i, ctx.temps_used.l)) { reset_ts(&s->temps[i]); } } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) block, otherwise we only trash the output args. "z_mask" is the non-zero bits mask for the first output arg. */ if (def->flags & TCG_OPF_BB_END) { - memset(&temps_used, 0, sizeof(temps_used)); + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); } else { do_reset_output: for (i = 0; i < nb_oargs; i++) { -- 2.25.1 Break the final cleanup clause out of the main switch statement. When fully folding an opcode to mov/movi, use "continue" to process the next opcode, else break to fall into the final cleanup. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 190 ++++++++++++++++++++++++------------------------- 1 file changed, 94 insertions(+), 96 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) switch (opc) { CASE_OP_32_64_VEC(mov): tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); - break; + continue; case INDEX_op_dup_vec: if (arg_is_const(op->args[1])) { tmp = arg_info(op->args[1])->val; tmp = dup_const(TCGOP_VECE(op), tmp); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; case INDEX_op_dup2_vec: assert(TCG_TARGET_REG_BITS == 32); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tcg_opt_gen_movi(s, &ctx, op, op->args[0], deposit64(arg_info(op->args[1])->val, 32, 32, arg_info(op->args[2])->val)); - break; + continue; } else if (args_are_copies(op->args[1], op->args[2])) { op->opc = INDEX_op_dup_vec; TCGOP_VECE(op) = MO_32; nb_iargs = 1; } - goto do_default; + break; CASE_OP_32_64(not): CASE_OP_32_64(neg): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, 0); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(bswap16): CASE_OP_32_64(bswap32): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = do_constant_folding(opc, arg_info(op->args[1])->val, op->args[2]); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(add): CASE_OP_32_64(sub): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = do_constant_folding(opc, arg_info(op->args[1])->val, arg_info(op->args[2])->val); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(clz): CASE_OP_32_64(ctz): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } else { tcg_opt_gen_mov(s, op, op->args[0], op->args[2]); } - break; + continue; } - goto do_default; + break; CASE_OP_32_64(deposit): if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[3], op->args[4], arg_info(op->args[2])->val); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(extract): if (arg_is_const(op->args[1])) { tmp = extract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(sextract): if (arg_is_const(op->args[1])) { tmp = sextract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(extract2): if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) ((uint32_t)v2 << (32 - shr))); } tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(setcond): tmp = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[3]); if (tmp != 2) { tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - break; + continue; } - goto do_default; + break; CASE_OP_32_64(brcond): tmp = do_constant_folding_cond(opc, op->args[0], op->args[1], op->args[2]); - if (tmp != 2) { - if (tmp) { - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - op->opc = INDEX_op_br; - op->args[0] = op->args[3]; - } else { - tcg_op_remove(s, op); - } + switch (tmp) { + case 0: + tcg_op_remove(s, op); + continue; + case 1: + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); + op->opc = opc = INDEX_op_br; + op->args[0] = op->args[3]; break; } - goto do_default; + break; CASE_OP_32_64(movcond): tmp = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[5]); if (tmp != 2) { tcg_opt_gen_mov(s, op, op->args[0], op->args[4-tmp]); - break; + continue; } if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { uint64_t tv = arg_info(op->args[3])->val; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (fv == 1 && tv == 0) { cond = tcg_invert_cond(cond); } else if (!(tv == 1 && fv == 0)) { - goto do_default; + break; } op->args[3] = cond; op->opc = opc = (opc == INDEX_op_movcond_i32 @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) : INDEX_op_setcond_i64); nb_iargs = 2; } - goto do_default; + break; case INDEX_op_add2_i32: case INDEX_op_sub2_i32: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rh = op->args[1]; tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)a); tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(a >> 32)); - break; + continue; } - goto do_default; + break; case INDEX_op_mulu2_i32: if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rh = op->args[1]; tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)r); tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(r >> 32)); - break; + continue; } - goto do_default; + break; case INDEX_op_brcond2_i32: tmp = do_constant_folding_cond2(&op->args[0], &op->args[2], op->args[4]); - if (tmp != 2) { - if (tmp) { - do_brcond_true: - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - op->opc = INDEX_op_br; - op->args[0] = op->args[5]; - } else { + if (tmp == 0) { do_brcond_false: - tcg_op_remove(s, op); - } - } else if ((op->args[4] == TCG_COND_LT - || op->args[4] == TCG_COND_GE) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == 0 - && arg_is_const(op->args[3]) - && arg_info(op->args[3])->val == 0) { + tcg_op_remove(s, op); + continue; + } + if (tmp == 1) { + do_brcond_true: + op->opc = opc = INDEX_op_br; + op->args[0] = op->args[5]; + break; + } + if ((op->args[4] == TCG_COND_LT || op->args[4] == TCG_COND_GE) + && arg_is_const(op->args[2]) + && arg_info(op->args[2])->val == 0 + && arg_is_const(op->args[3]) + && arg_info(op->args[3])->val == 0) { /* Simplify LT/GE comparisons vs zero to a single compare vs the high word of the input. */ do_brcond_high: - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - op->opc = INDEX_op_brcond_i32; + op->opc = opc = INDEX_op_brcond_i32; op->args[0] = op->args[1]; op->args[1] = op->args[3]; op->args[2] = op->args[4]; op->args[3] = op->args[5]; - } else if (op->args[4] == TCG_COND_EQ) { + break; + } + if (op->args[4] == TCG_COND_EQ) { /* Simplify EQ comparisons where one of the pairs can be simplified. */ tmp = do_constant_folding_cond(INDEX_op_brcond_i32, @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (tmp == 0) { goto do_brcond_false; } else if (tmp != 1) { - goto do_default; + break; } do_brcond_low: memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[1] = op->args[2]; op->args[2] = op->args[4]; op->args[3] = op->args[5]; - } else if (op->args[4] == TCG_COND_NE) { + break; + } + if (op->args[4] == TCG_COND_NE) { /* Simplify NE comparisons where one of the pairs can be simplified. */ tmp = do_constant_folding_cond(INDEX_op_brcond_i32, @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } else if (tmp == 1) { goto do_brcond_true; } - goto do_default; - } else { - goto do_default; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (tmp != 2) { do_setcond_const: tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); - } else if ((op->args[5] == TCG_COND_LT - || op->args[5] == TCG_COND_GE) - && arg_is_const(op->args[3]) - && arg_info(op->args[3])->val == 0 - && arg_is_const(op->args[4]) - && arg_info(op->args[4])->val == 0) { + continue; + } + if ((op->args[5] == TCG_COND_LT || op->args[5] == TCG_COND_GE) + && arg_is_const(op->args[3]) + && arg_info(op->args[3])->val == 0 + && arg_is_const(op->args[4]) + && arg_info(op->args[4])->val == 0) { /* Simplify LT/GE comparisons vs zero to a single compare vs the high word of the input. */ do_setcond_high: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[1] = op->args[2]; op->args[2] = op->args[4]; op->args[3] = op->args[5]; - } else if (op->args[5] == TCG_COND_EQ) { + break; + } + if (op->args[5] == TCG_COND_EQ) { /* Simplify EQ comparisons where one of the pairs can be simplified. */ tmp = do_constant_folding_cond(INDEX_op_setcond_i32, @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (tmp == 0) { goto do_setcond_high; } else if (tmp != 1) { - goto do_default; + break; } do_setcond_low: reset_temp(op->args[0]); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->opc = INDEX_op_setcond_i32; op->args[2] = op->args[3]; op->args[3] = op->args[5]; - } else if (op->args[5] == TCG_COND_NE) { + break; + } + if (op->args[5] == TCG_COND_NE) { /* Simplify NE comparisons where one of the pairs can be simplified. */ tmp = do_constant_folding_cond(INDEX_op_setcond_i32, @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } else if (tmp == 1) { goto do_setcond_const; } - goto do_default; - } else { - goto do_default; } break; - case INDEX_op_call: - if (!(tcg_call_flags(op) + default: + break; + } + + /* Some of the folding above can change opc. */ + opc = op->opc; + def = &tcg_op_defs[opc]; + if (def->flags & TCG_OPF_BB_END) { + memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); + } else { + if (opc == INDEX_op_call && + !(tcg_call_flags(op) & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { for (i = 0; i < nb_globals; i++) { if (test_bit(i, ctx.temps_used.l)) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } } } - goto do_reset_output; - default: - do_default: - /* Default case: we know nothing about operation (or were unable - to compute the operation result) so no propagation is done. - We trash everything if the operation is the end of a basic - block, otherwise we only trash the output args. "z_mask" is - the non-zero bits mask for the first output arg. */ - if (def->flags & TCG_OPF_BB_END) { - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - } else { - do_reset_output: - for (i = 0; i < nb_oargs; i++) { - reset_temp(op->args[i]); - /* Save the corresponding known-zero bits mask for the - first output argument (only one supported so far). */ - if (i == 0) { - arg_info(op->args[i])->z_mask = z_mask; - } + for (i = 0; i < nb_oargs; i++) { + reset_temp(op->args[i]); + /* Save the corresponding known-zero bits mask for the + first output argument (only one supported so far). */ + if (i == 0) { + arg_info(op->args[i])->z_mask = z_mask; } } - break; } /* Eliminate duplicate and redundant fence instructions. */ -- 2.25.1 Adjust the interface to take the OptContext parameter instead of TCGContext or both. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 67 +++++++++++++++++++++++++------------------------- 1 file changed, 34 insertions(+), 33 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct TempOptInfo { } TempOptInfo; typedef struct OptContext { + TCGContext *tcg; TCGTempSet temps_used; } OptContext; @@ -XXX,XX +XXX,XX @@ static bool args_are_copies(TCGArg arg1, TCGArg arg2) return ts_are_copies(arg_temp(arg1), arg_temp(arg2)); } -static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) +static void tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) { TCGTemp *dst_ts = arg_temp(dst); TCGTemp *src_ts = arg_temp(src); @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) TCGOpcode new_op; if (ts_are_copies(dst_ts, src_ts)) { - tcg_op_remove(s, op); + tcg_op_remove(ctx->tcg, op); return; } @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(TCGContext *s, TCGOp *op, TCGArg dst, TCGArg src) } } -static void tcg_opt_gen_movi(TCGContext *s, OptContext *ctx, - TCGOp *op, TCGArg dst, uint64_t val) +static void tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, + TCGArg dst, uint64_t val) { const TCGOpDef *def = &tcg_op_defs[op->opc]; TCGType type; @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_movi(TCGContext *s, OptContext *ctx, /* Convert movi to mov with constant temp. */ tv = tcg_constant_internal(type, val); init_ts_info(ctx, tv); - tcg_opt_gen_mov(s, op, dst, temp_arg(tv)); + tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv)); } static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) { int nb_temps, nb_globals, i; TCGOp *op, *op_next, *prev_mb = NULL; - OptContext ctx = {}; + OptContext ctx = { .tcg = s }; /* Array VALS has an element for each temp. If this temp holds a constant then its value is kept in VALS' element. @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(rotr): if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == 0) { - tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); + tcg_opt_gen_movi(&ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (!arg_is_const(op->args[1]) && arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0) { - tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (!arg_is_const(op->args[1]) && arg_is_const(op->args[2]) && arg_info(op->args[2])->val == -1) { - tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (partmask == 0) { tcg_debug_assert(nb_oargs == 1); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); + tcg_opt_gen_movi(&ctx, op, op->args[0], 0); continue; } if (affected == 0) { tcg_debug_assert(nb_oargs == 1); - tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(mulsh): if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0) { - tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); + tcg_opt_gen_movi(&ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(or): CASE_OP_32_64_VEC(and): if (args_are_copies(op->args[1], op->args[2])) { - tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(sub): CASE_OP_32_64_VEC(xor): if (args_are_copies(op->args[1], op->args[2])) { - tcg_opt_gen_movi(s, &ctx, op, op->args[0], 0); + tcg_opt_gen_movi(&ctx, op, op->args[0], 0); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) allocator where needed and possible. Also detect copies. */ switch (opc) { CASE_OP_32_64_VEC(mov): - tcg_opt_gen_mov(s, op, op->args[0], op->args[1]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; case INDEX_op_dup_vec: if (arg_is_const(op->args[1])) { tmp = arg_info(op->args[1])->val; tmp = dup_const(TCGOP_VECE(op), tmp); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_dup2_vec: assert(TCG_TARGET_REG_BITS == 32); if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - tcg_opt_gen_movi(s, &ctx, op, op->args[0], + tcg_opt_gen_movi(&ctx, op, op->args[0], deposit64(arg_info(op->args[1])->val, 32, 32, arg_info(op->args[2])->val)); continue; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_extrh_i64_i32: if (arg_is_const(op->args[1])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, 0); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, op->args[2]); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { tmp = do_constant_folding(opc, arg_info(op->args[1])->val, arg_info(op->args[2])->val); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) TCGArg v = arg_info(op->args[1])->val; if (v != 0) { tmp = do_constant_folding(opc, v, 0); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); } else { - tcg_opt_gen_mov(s, op, op->args[0], op->args[2]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[2]); } continue; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = deposit64(arg_info(op->args[1])->val, op->args[3], op->args[4], arg_info(op->args[2])->val); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = extract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (arg_is_const(op->args[1])) { tmp = sextract64(arg_info(op->args[1])->val, op->args[2], op->args[3]); - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = (int32_t)(((uint32_t)v1 >> shr) | ((uint32_t)v2 << (32 - shr))); } - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[3]); if (tmp != 2) { - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) tmp = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[5]); if (tmp != 2) { - tcg_opt_gen_mov(s, op, op->args[0], op->args[4-tmp]); + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[4-tmp]); continue; } if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)a); - tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(a >> 32)); + tcg_opt_gen_movi(&ctx, op, rl, (int32_t)a); + tcg_opt_gen_movi(&ctx, op2, rh, (int32_t)(a >> 32)); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(s, &ctx, op, rl, (int32_t)r); - tcg_opt_gen_movi(s, &ctx, op2, rh, (int32_t)(r >> 32)); + tcg_opt_gen_movi(&ctx, op, rl, (int32_t)r); + tcg_opt_gen_movi(&ctx, op2, rh, (int32_t)(r >> 32)); continue; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->args[5]); if (tmp != 2) { do_setcond_const: - tcg_opt_gen_movi(s, &ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); continue; } if ((op->args[5] == TCG_COND_LT || op->args[5] == TCG_COND_GE) -- 2.25.1 This will expose the variable to subroutines that will be broken out of tcg_optimize. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 11 ++++++----- 1 file changed, 6 insertions(+), 5 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct TempOptInfo { typedef struct OptContext { TCGContext *tcg; + TCGOp *prev_mb; TCGTempSet temps_used; } OptContext; @@ -XXX,XX +XXX,XX @@ static bool swap_commutative2(TCGArg *p1, TCGArg *p2) void tcg_optimize(TCGContext *s) { int nb_temps, nb_globals, i; - TCGOp *op, *op_next, *prev_mb = NULL; + TCGOp *op, *op_next; OptContext ctx = { .tcg = s }; /* Array VALS has an element for each temp. @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } /* Eliminate duplicate and redundant fence instructions. */ - if (prev_mb) { + if (ctx.prev_mb) { switch (opc) { case INDEX_op_mb: /* Merge two barriers of the same type into one, @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) * barrier. This is stricter than specified but for * the purposes of TCG is better than not optimizing. */ - prev_mb->args[0] |= op->args[0]; + ctx.prev_mb->args[0] |= op->args[0]; tcg_op_remove(s, op); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_qemu_st_i64: case INDEX_op_call: /* Opcodes that touch guest memory stop the optimization. */ - prev_mb = NULL; + ctx.prev_mb = NULL; break; } } else if (opc == INDEX_op_mb) { - prev_mb = op; + ctx.prev_mb = op; } } } -- 2.25.1 There was no real reason for calls to have separate code here. Unify init for calls vs non-calls using the call path, which handles TCG_CALL_DUMMY_ARG. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 25 +++++++++++-------------- 1 file changed, 11 insertions(+), 14 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static void init_ts_info(OptContext *ctx, TCGTemp *ts) } } -static void init_arg_info(OptContext *ctx, TCGArg arg) -{ - init_ts_info(ctx, arg_temp(arg)); -} - static TCGTemp *find_better_copy(TCGContext *s, TCGTemp *ts) { TCGTemp *i, *g, *l; @@ -XXX,XX +XXX,XX @@ static bool swap_commutative2(TCGArg *p1, TCGArg *p2) return false; } +static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args) +{ + for (int i = 0; i < nb_args; i++) { + TCGTemp *ts = arg_temp(op->args[i]); + if (ts) { + init_ts_info(ctx, ts); + } + } +} + /* Propagate constants and copies, fold constant expressions. */ void tcg_optimize(TCGContext *s) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (opc == INDEX_op_call) { nb_oargs = TCGOP_CALLO(op); nb_iargs = TCGOP_CALLI(op); - for (i = 0; i < nb_oargs + nb_iargs; i++) { - TCGTemp *ts = arg_temp(op->args[i]); - if (ts) { - init_ts_info(&ctx, ts); - } - } } else { nb_oargs = def->nb_oargs; nb_iargs = def->nb_iargs; - for (i = 0; i < nb_oargs + nb_iargs; i++) { - init_arg_info(&ctx, op->args[i]); - } } + init_arguments(&ctx, op, nb_oargs + nb_iargs); /* Do copy propagation */ for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { -- 2.25.1 Continue splitting tcg_optimize. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 22 ++++++++++++++-------- 1 file changed, 14 insertions(+), 8 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static void init_arguments(OptContext *ctx, TCGOp *op, int nb_args) } } +static void copy_propagate(OptContext *ctx, TCGOp *op, + int nb_oargs, int nb_iargs) +{ + TCGContext *s = ctx->tcg; + + for (int i = nb_oargs; i < nb_oargs + nb_iargs; i++) { + TCGTemp *ts = arg_temp(op->args[i]); + if (ts && ts_is_copy(ts)) { + op->args[i] = temp_arg(find_better_copy(s, ts)); + } + } +} + /* Propagate constants and copies, fold constant expressions. */ void tcg_optimize(TCGContext *s) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) nb_iargs = def->nb_iargs; } init_arguments(&ctx, op, nb_oargs + nb_iargs); - - /* Do copy propagation */ - for (i = nb_oargs; i < nb_oargs + nb_iargs; i++) { - TCGTemp *ts = arg_temp(op->args[i]); - if (ts && ts_is_copy(ts)) { - op->args[i] = temp_arg(find_better_copy(s, ts)); - } - } + copy_propagate(&ctx, op, nb_oargs, nb_iargs); /* For commutative operations make constant second argument */ switch (opc) { -- 2.25.1 Calls are special in that they have a variable number of arguments, and need to be able to clobber globals. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 63 ++++++++++++++++++++++++++++++++------------------ 1 file changed, 41 insertions(+), 22 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static void copy_propagate(OptContext *ctx, TCGOp *op, } } +static bool fold_call(OptContext *ctx, TCGOp *op) +{ + TCGContext *s = ctx->tcg; + int nb_oargs = TCGOP_CALLO(op); + int nb_iargs = TCGOP_CALLI(op); + int flags, i; + + init_arguments(ctx, op, nb_oargs + nb_iargs); + copy_propagate(ctx, op, nb_oargs, nb_iargs); + + /* If the function reads or writes globals, reset temp data. */ + flags = tcg_call_flags(op); + if (!(flags & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { + int nb_globals = s->nb_globals; + + for (i = 0; i < nb_globals; i++) { + if (test_bit(i, ctx->temps_used.l)) { + reset_ts(&ctx->tcg->temps[i]); + } + } + } + + /* Reset temp data for outputs. */ + for (i = 0; i < nb_oargs; i++) { + reset_temp(op->args[i]); + } + + /* Stop optimizing MB across calls. */ + ctx->prev_mb = NULL; + return true; +} + /* Propagate constants and copies, fold constant expressions. */ void tcg_optimize(TCGContext *s) { - int nb_temps, nb_globals, i; + int nb_temps, i; TCGOp *op, *op_next; OptContext ctx = { .tcg = s }; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) available through the doubly linked circular list. */ nb_temps = s->nb_temps; - nb_globals = s->nb_globals; - for (i = 0; i < nb_temps; ++i) { s->temps[i].state_ptr = NULL; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) uint64_t z_mask, partmask, affected, tmp; int nb_oargs, nb_iargs; TCGOpcode opc = op->opc; - const TCGOpDef *def = &tcg_op_defs[opc]; + const TCGOpDef *def; - /* Count the arguments, and initialize the temps that are - going to be used */ + /* Calls are special. */ if (opc == INDEX_op_call) { - nb_oargs = TCGOP_CALLO(op); - nb_iargs = TCGOP_CALLI(op); - } else { - nb_oargs = def->nb_oargs; - nb_iargs = def->nb_iargs; + fold_call(&ctx, op); + continue; } + + def = &tcg_op_defs[opc]; + nb_oargs = def->nb_oargs; + nb_iargs = def->nb_iargs; init_arguments(&ctx, op, nb_oargs + nb_iargs); copy_propagate(&ctx, op, nb_oargs, nb_iargs); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (def->flags & TCG_OPF_BB_END) { memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); } else { - if (opc == INDEX_op_call && - !(tcg_call_flags(op) - & (TCG_CALL_NO_READ_GLOBALS | TCG_CALL_NO_WRITE_GLOBALS))) { - for (i = 0; i < nb_globals; i++) { - if (test_bit(i, ctx.temps_used.l)) { - reset_ts(&s->temps[i]); - } - } - } - for (i = 0; i < nb_oargs; i++) { reset_temp(op->args[i]); /* Save the corresponding known-zero bits mask for the @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_qemu_st_i32: case INDEX_op_qemu_st8_i32: case INDEX_op_qemu_st_i64: - case INDEX_op_call: /* Opcodes that touch guest memory stop the optimization. */ ctx.prev_mb = NULL; break; -- 2.25.1 Rather than try to keep these up-to-date across folding, re-read nb_oargs at the end, after re-reading the opcode. A couple of asserts need dropping, but that will take care of itself as we split the function further. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 14 ++++---------- 1 file changed, 4 insertions(+), 10 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { uint64_t z_mask, partmask, affected, tmp; - int nb_oargs, nb_iargs; TCGOpcode opc = op->opc; const TCGOpDef *def; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } def = &tcg_op_defs[opc]; - nb_oargs = def->nb_oargs; - nb_iargs = def->nb_iargs; - init_arguments(&ctx, op, nb_oargs + nb_iargs); - copy_propagate(&ctx, op, nb_oargs, nb_iargs); + init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs); + copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs); /* For commutative operations make constant second argument */ switch (opc) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(qemu_ld): { - MemOpIdx oi = op->args[nb_oargs + nb_iargs]; + MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs]; MemOp mop = get_memop(oi); if (!(mop & MO_SIGN)) { z_mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } if (partmask == 0) { - tcg_debug_assert(nb_oargs == 1); tcg_opt_gen_movi(&ctx, op, op->args[0], 0); continue; } if (affected == 0) { - tcg_debug_assert(nb_oargs == 1); tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); continue; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } else if (args_are_copies(op->args[1], op->args[2])) { op->opc = INDEX_op_dup_vec; TCGOP_VECE(op) = MO_32; - nb_iargs = 1; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) op->opc = opc = (opc == INDEX_op_movcond_i32 ? INDEX_op_setcond_i32 : INDEX_op_setcond_i64); - nb_iargs = 2; } break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (def->flags & TCG_OPF_BB_END) { memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); } else { + int nb_oargs = def->nb_oargs; for (i = 0; i < nb_oargs; i++) { reset_temp(op->args[i]); /* Save the corresponding known-zero bits mask for the -- 2.25.1 Return -1 instead of 2 for failure, so that we can use comparisons against 0 for all cases. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 145 +++++++++++++++++++++++++------------------------ 1 file changed, 74 insertions(+), 71 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool do_constant_folding_cond_eq(TCGCond c) } } -/* Return 2 if the condition can't be simplified, and the result - of the condition (0 or 1) if it can */ -static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x, - TCGArg y, TCGCond c) +/* + * Return -1 if the condition can't be simplified, + * and the result of the condition (0 or 1) if it can. + */ +static int do_constant_folding_cond(TCGOpcode op, TCGArg x, + TCGArg y, TCGCond c) { uint64_t xv = arg_info(x)->val; uint64_t yv = arg_info(y)->val; @@ -XXX,XX +XXX,XX @@ static TCGArg do_constant_folding_cond(TCGOpcode op, TCGArg x, case TCG_COND_GEU: return 1; default: - return 2; + return -1; } } - return 2; + return -1; } -/* Return 2 if the condition can't be simplified, and the result - of the condition (0 or 1) if it can */ -static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) +/* + * Return -1 if the condition can't be simplified, + * and the result of the condition (0 or 1) if it can. + */ +static int do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) { TCGArg al = p1[0], ah = p1[1]; TCGArg bl = p2[0], bh = p2[1]; @@ -XXX,XX +XXX,XX @@ static TCGArg do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) if (args_are_copies(al, bl) && args_are_copies(ah, bh)) { return do_constant_folding_cond_eq(c); } - return 2; + return -1; } static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; CASE_OP_32_64(setcond): - tmp = do_constant_folding_cond(opc, op->args[1], - op->args[2], op->args[3]); - if (tmp != 2) { - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); + i = do_constant_folding_cond(opc, op->args[1], + op->args[2], op->args[3]); + if (i >= 0) { + tcg_opt_gen_movi(&ctx, op, op->args[0], i); continue; } break; CASE_OP_32_64(brcond): - tmp = do_constant_folding_cond(opc, op->args[0], - op->args[1], op->args[2]); - switch (tmp) { - case 0: + i = do_constant_folding_cond(opc, op->args[0], + op->args[1], op->args[2]); + if (i == 0) { tcg_op_remove(s, op); continue; - case 1: + } else if (i > 0) { memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); op->opc = opc = INDEX_op_br; op->args[0] = op->args[3]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; CASE_OP_32_64(movcond): - tmp = do_constant_folding_cond(opc, op->args[1], - op->args[2], op->args[5]); - if (tmp != 2) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[4-tmp]); + i = do_constant_folding_cond(opc, op->args[1], + op->args[2], op->args[5]); + if (i >= 0) { + tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[4 - i]); continue; } if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; case INDEX_op_brcond2_i32: - tmp = do_constant_folding_cond2(&op->args[0], &op->args[2], - op->args[4]); - if (tmp == 0) { + i = do_constant_folding_cond2(&op->args[0], &op->args[2], + op->args[4]); + if (i == 0) { do_brcond_false: tcg_op_remove(s, op); continue; } - if (tmp == 1) { + if (i > 0) { do_brcond_true: op->opc = opc = INDEX_op_br; op->args[0] = op->args[5]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (op->args[4] == TCG_COND_EQ) { /* Simplify EQ comparisons where one of the pairs can be simplified. */ - tmp = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[0], op->args[2], - TCG_COND_EQ); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_brcond_i32, + op->args[0], op->args[2], + TCG_COND_EQ); + if (i == 0) { goto do_brcond_false; - } else if (tmp == 1) { + } else if (i > 0) { goto do_brcond_high; } - tmp = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[1], op->args[3], - TCG_COND_EQ); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_brcond_i32, + op->args[1], op->args[3], + TCG_COND_EQ); + if (i == 0) { goto do_brcond_false; - } else if (tmp != 1) { + } else if (i < 0) { break; } do_brcond_low: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (op->args[4] == TCG_COND_NE) { /* Simplify NE comparisons where one of the pairs can be simplified. */ - tmp = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[0], op->args[2], - TCG_COND_NE); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_brcond_i32, + op->args[0], op->args[2], + TCG_COND_NE); + if (i == 0) { goto do_brcond_high; - } else if (tmp == 1) { + } else if (i > 0) { goto do_brcond_true; } - tmp = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[1], op->args[3], - TCG_COND_NE); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_brcond_i32, + op->args[1], op->args[3], + TCG_COND_NE); + if (i == 0) { goto do_brcond_low; - } else if (tmp == 1) { + } else if (i > 0) { goto do_brcond_true; } } break; case INDEX_op_setcond2_i32: - tmp = do_constant_folding_cond2(&op->args[1], &op->args[3], - op->args[5]); - if (tmp != 2) { + i = do_constant_folding_cond2(&op->args[1], &op->args[3], + op->args[5]); + if (i >= 0) { do_setcond_const: - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); + tcg_opt_gen_movi(&ctx, op, op->args[0], i); continue; } if ((op->args[5] == TCG_COND_LT || op->args[5] == TCG_COND_GE) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (op->args[5] == TCG_COND_EQ) { /* Simplify EQ comparisons where one of the pairs can be simplified. */ - tmp = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[1], op->args[3], - TCG_COND_EQ); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_setcond_i32, + op->args[1], op->args[3], + TCG_COND_EQ); + if (i == 0) { goto do_setcond_const; - } else if (tmp == 1) { + } else if (i > 0) { goto do_setcond_high; } - tmp = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[2], op->args[4], - TCG_COND_EQ); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_setcond_i32, + op->args[2], op->args[4], + TCG_COND_EQ); + if (i == 0) { goto do_setcond_high; - } else if (tmp != 1) { + } else if (i < 0) { break; } do_setcond_low: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (op->args[5] == TCG_COND_NE) { /* Simplify NE comparisons where one of the pairs can be simplified. */ - tmp = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[1], op->args[3], - TCG_COND_NE); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_setcond_i32, + op->args[1], op->args[3], + TCG_COND_NE); + if (i == 0) { goto do_setcond_high; - } else if (tmp == 1) { + } else if (i > 0) { goto do_setcond_const; } - tmp = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[2], op->args[4], - TCG_COND_NE); - if (tmp == 0) { + i = do_constant_folding_cond(INDEX_op_setcond_i32, + op->args[2], op->args[4], + TCG_COND_NE); + if (i == 0) { goto do_setcond_low; - } else if (tmp == 1) { + } else if (i > 0) { goto do_setcond_const; } } -- 2.25.1 This will allow callers to tail call to these functions and return true indicating processing complete. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool args_are_copies(TCGArg arg1, TCGArg arg2) return ts_are_copies(arg_temp(arg1), arg_temp(arg2)); } -static void tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) +static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) { TCGTemp *dst_ts = arg_temp(dst); TCGTemp *src_ts = arg_temp(src); @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) if (ts_are_copies(dst_ts, src_ts)) { tcg_op_remove(ctx->tcg, op); - return; + return true; } reset_ts(dst_ts); @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) di->is_const = si->is_const; di->val = si->val; } + return true; } -static void tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, +static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, TCGArg dst, uint64_t val) { const TCGOpDef *def = &tcg_op_defs[op->opc]; @@ -XXX,XX +XXX,XX @@ static void tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, /* Convert movi to mov with constant temp. */ tv = tcg_constant_internal(type, val); init_ts_info(ctx, tv); - tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv)); + return tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv)); } static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y) -- 2.25.1 Copy z_mask into OptContext, for writeback to the first output within the new function. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 49 +++++++++++++++++++++++++++++++++---------------- 1 file changed, 33 insertions(+), 16 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct OptContext { TCGContext *tcg; TCGOp *prev_mb; TCGTempSet temps_used; + + /* In flight values from optimization. */ + uint64_t z_mask; } OptContext; static inline TempOptInfo *ts_info(TCGTemp *ts) @@ -XXX,XX +XXX,XX @@ static void copy_propagate(OptContext *ctx, TCGOp *op, } } +static void finish_folding(OptContext *ctx, TCGOp *op) +{ + const TCGOpDef *def = &tcg_op_defs[op->opc]; + int i, nb_oargs; + + /* + * For an opcode that ends a BB, reset all temp data. + * We do no cross-BB optimization. + */ + if (def->flags & TCG_OPF_BB_END) { + memset(&ctx->temps_used, 0, sizeof(ctx->temps_used)); + ctx->prev_mb = NULL; + return; + } + + nb_oargs = def->nb_oargs; + for (i = 0; i < nb_oargs; i++) { + reset_temp(op->args[i]); + /* + * Save the corresponding known-zero bits mask for the + * first output argument (only one supported so far). + */ + if (i == 0) { + arg_info(op->args[i])->z_mask = ctx->z_mask; + } + } +} + static bool fold_call(OptContext *ctx, TCGOp *op) { TCGContext *s = ctx->tcg; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) partmask &= 0xffffffffu; affected &= 0xffffffffu; } + ctx.z_mask = z_mask; if (partmask == 0) { tcg_opt_gen_movi(&ctx, op, op->args[0], 0); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Some of the folding above can change opc. */ - opc = op->opc; - def = &tcg_op_defs[opc]; - if (def->flags & TCG_OPF_BB_END) { - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - } else { - int nb_oargs = def->nb_oargs; - for (i = 0; i < nb_oargs; i++) { - reset_temp(op->args[i]); - /* Save the corresponding known-zero bits mask for the - first output argument (only one supported so far). */ - if (i == 0) { - arg_info(op->args[i])->z_mask = z_mask; - } - } - } + finish_folding(&ctx, op); /* Eliminate duplicate and redundant fence instructions. */ if (ctx.prev_mb) { -- 2.25.1 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) uint64_t z_mask, partmask, affected, tmp; TCGOpcode opc = op->opc; const TCGOpDef *def; + bool done = false; /* Calls are special. */ if (opc == INDEX_op_call) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) allocator where needed and possible. Also detect copies. */ switch (opc) { CASE_OP_32_64_VEC(mov): - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - continue; + done = tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); + break; case INDEX_op_dup_vec: if (arg_is_const(op->args[1])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - finish_folding(&ctx, op); + if (!done) { + finish_folding(&ctx, op); + } /* Eliminate duplicate and redundant fence instructions. */ if (ctx.prev_mb) { -- 2.25.1 This puts the separate mb optimization into the same framework as the others. While fold_qemu_{ld,st} are currently identical, that won't last as more code gets moved. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 89 +++++++++++++++++++++++++++++--------------------- 1 file changed, 51 insertions(+), 38 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_call(OptContext *ctx, TCGOp *op) return true; } +static bool fold_mb(OptContext *ctx, TCGOp *op) +{ + /* Eliminate duplicate and redundant fence instructions. */ + if (ctx->prev_mb) { + /* + * Merge two barriers of the same type into one, + * or a weaker barrier into a stronger one, + * or two weaker barriers into a stronger one. + * mb X; mb Y => mb X|Y + * mb; strl => mb; st + * ldaq; mb => ld; mb + * ldaq; strl => ld; mb; st + * Other combinations are also merged into a strong + * barrier. This is stricter than specified but for + * the purposes of TCG is better than not optimizing. + */ + ctx->prev_mb->args[0] |= op->args[0]; + tcg_op_remove(ctx->tcg, op); + } else { + ctx->prev_mb = op; + } + return true; +} + +static bool fold_qemu_ld(OptContext *ctx, TCGOp *op) +{ + /* Opcodes that touch guest memory stop the mb optimization. */ + ctx->prev_mb = NULL; + return false; +} + +static bool fold_qemu_st(OptContext *ctx, TCGOp *op) +{ + /* Opcodes that touch guest memory stop the mb optimization. */ + ctx->prev_mb = NULL; + return false; +} + /* Propagate constants and copies, fold constant expressions. */ void tcg_optimize(TCGContext *s) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; + case INDEX_op_mb: + done = fold_mb(&ctx, op); + break; + case INDEX_op_qemu_ld_i32: + case INDEX_op_qemu_ld_i64: + done = fold_qemu_ld(&ctx, op); + break; + case INDEX_op_qemu_st_i32: + case INDEX_op_qemu_st8_i32: + case INDEX_op_qemu_st_i64: + done = fold_qemu_st(&ctx, op); + break; + default: break; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) if (!done) { finish_folding(&ctx, op); } - - /* Eliminate duplicate and redundant fence instructions. */ - if (ctx.prev_mb) { - switch (opc) { - case INDEX_op_mb: - /* Merge two barriers of the same type into one, - * or a weaker barrier into a stronger one, - * or two weaker barriers into a stronger one. - * mb X; mb Y => mb X|Y - * mb; strl => mb; st - * ldaq; mb => ld; mb - * ldaq; strl => ld; mb; st - * Other combinations are also merged into a strong - * barrier. This is stricter than specified but for - * the purposes of TCG is better than not optimizing. - */ - ctx.prev_mb->args[0] |= op->args[0]; - tcg_op_remove(s, op); - break; - - default: - /* Opcodes that end the block stop the optimization. */ - if ((def->flags & TCG_OPF_BB_END) == 0) { - break; - } - /* fallthru */ - case INDEX_op_qemu_ld_i32: - case INDEX_op_qemu_ld_i64: - case INDEX_op_qemu_st_i32: - case INDEX_op_qemu_st8_i32: - case INDEX_op_qemu_st_i64: - /* Opcodes that touch guest memory stop the optimization. */ - ctx.prev_mb = NULL; - break; - } - } else if (opc == INDEX_op_mb) { - ctx.prev_mb = op; - } } } -- 2.25.1 Split out a whole bunch of placeholder functions, which are currently identical. That won't last as more code gets moved. Use CASE_32_64_VEC for some logical operators that previously missed the addition of vectors. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 271 +++++++++++++++++++++++++++++++++++++++---------- 1 file changed, 219 insertions(+), 52 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static void finish_folding(OptContext *ctx, TCGOp *op) } } +/* + * The fold_* functions return true when processing is complete, + * usually by folding the operation to a constant or to a copy, + * and calling tcg_opt_gen_{mov,movi}. They may do other things, + * like collect information about the value produced, for use in + * optimizing a subsequent operation. + * + * These first fold_* functions are all helpers, used by other + * folders for more specific operations. + */ + +static bool fold_const1(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t; + + t = arg_info(op->args[1])->val; + t = do_constant_folding(op->opc, t, 0); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return false; +} + +static bool fold_const2(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { + uint64_t t1 = arg_info(op->args[1])->val; + uint64_t t2 = arg_info(op->args[2])->val; + + t1 = do_constant_folding(op->opc, t1, t2); + return tcg_opt_gen_movi(ctx, op, op->args[0], t1); + } + return false; +} + +/* + * These outermost fold_<op> functions are sorted alphabetically. + */ + +static bool fold_add(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_and(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_andc(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + static bool fold_call(OptContext *ctx, TCGOp *op) { TCGContext *s = ctx->tcg; @@ -XXX,XX +XXX,XX @@ static bool fold_call(OptContext *ctx, TCGOp *op) return true; } +static bool fold_ctpop(OptContext *ctx, TCGOp *op) +{ + return fold_const1(ctx, op); +} + +static bool fold_divide(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_eqv(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_exts(OptContext *ctx, TCGOp *op) +{ + return fold_const1(ctx, op); +} + +static bool fold_extu(OptContext *ctx, TCGOp *op) +{ + return fold_const1(ctx, op); +} + static bool fold_mb(OptContext *ctx, TCGOp *op) { /* Eliminate duplicate and redundant fence instructions. */ @@ -XXX,XX +XXX,XX @@ static bool fold_mb(OptContext *ctx, TCGOp *op) return true; } +static bool fold_mul(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_nand(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_neg(OptContext *ctx, TCGOp *op) +{ + return fold_const1(ctx, op); +} + +static bool fold_nor(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_not(OptContext *ctx, TCGOp *op) +{ + return fold_const1(ctx, op); +} + +static bool fold_or(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_orc(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + static bool fold_qemu_ld(OptContext *ctx, TCGOp *op) { /* Opcodes that touch guest memory stop the mb optimization. */ @@ -XXX,XX +XXX,XX @@ static bool fold_qemu_st(OptContext *ctx, TCGOp *op) return false; } +static bool fold_remainder(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_shift(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_sub(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + +static bool fold_xor(OptContext *ctx, TCGOp *op) +{ + return fold_const2(ctx, op); +} + /* Propagate constants and copies, fold constant expressions. */ void tcg_optimize(TCGContext *s) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(not): - CASE_OP_32_64(neg): - CASE_OP_32_64(ext8s): - CASE_OP_32_64(ext8u): - CASE_OP_32_64(ext16s): - CASE_OP_32_64(ext16u): - CASE_OP_32_64(ctpop): - case INDEX_op_ext32s_i64: - case INDEX_op_ext32u_i64: - case INDEX_op_ext_i32_i64: - case INDEX_op_extu_i32_i64: - case INDEX_op_extrl_i64_i32: - case INDEX_op_extrh_i64_i32: - if (arg_is_const(op->args[1])) { - tmp = do_constant_folding(opc, arg_info(op->args[1])->val, 0); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - CASE_OP_32_64(bswap16): CASE_OP_32_64(bswap32): case INDEX_op_bswap64_i64: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(add): - CASE_OP_32_64(sub): - CASE_OP_32_64(mul): - CASE_OP_32_64(or): - CASE_OP_32_64(and): - CASE_OP_32_64(xor): - CASE_OP_32_64(shl): - CASE_OP_32_64(shr): - CASE_OP_32_64(sar): - CASE_OP_32_64(rotl): - CASE_OP_32_64(rotr): - CASE_OP_32_64(andc): - CASE_OP_32_64(orc): - CASE_OP_32_64(eqv): - CASE_OP_32_64(nand): - CASE_OP_32_64(nor): - CASE_OP_32_64(muluh): - CASE_OP_32_64(mulsh): - CASE_OP_32_64(div): - CASE_OP_32_64(divu): - CASE_OP_32_64(rem): - CASE_OP_32_64(remu): - if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - tmp = do_constant_folding(opc, arg_info(op->args[1])->val, - arg_info(op->args[2])->val); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - CASE_OP_32_64(clz): CASE_OP_32_64(ctz): if (arg_is_const(op->args[1])) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; + default: + break; + + /* ---------------------------------------------------------- */ + /* Sorted alphabetically by opcode as much as possible. */ + + CASE_OP_32_64_VEC(add): + done = fold_add(&ctx, op); + break; + CASE_OP_32_64_VEC(and): + done = fold_and(&ctx, op); + break; + CASE_OP_32_64_VEC(andc): + done = fold_andc(&ctx, op); + break; + CASE_OP_32_64(ctpop): + done = fold_ctpop(&ctx, op); + break; + CASE_OP_32_64(div): + CASE_OP_32_64(divu): + done = fold_divide(&ctx, op); + break; + CASE_OP_32_64(eqv): + done = fold_eqv(&ctx, op); + break; + CASE_OP_32_64(ext8s): + CASE_OP_32_64(ext16s): + case INDEX_op_ext32s_i64: + case INDEX_op_ext_i32_i64: + done = fold_exts(&ctx, op); + break; + CASE_OP_32_64(ext8u): + CASE_OP_32_64(ext16u): + case INDEX_op_ext32u_i64: + case INDEX_op_extu_i32_i64: + case INDEX_op_extrl_i64_i32: + case INDEX_op_extrh_i64_i32: + done = fold_extu(&ctx, op); + break; case INDEX_op_mb: done = fold_mb(&ctx, op); break; + CASE_OP_32_64(mul): + done = fold_mul(&ctx, op); + break; + CASE_OP_32_64(mulsh): + CASE_OP_32_64(muluh): + done = fold_mul_highpart(&ctx, op); + break; + CASE_OP_32_64(nand): + done = fold_nand(&ctx, op); + break; + CASE_OP_32_64(neg): + done = fold_neg(&ctx, op); + break; + CASE_OP_32_64(nor): + done = fold_nor(&ctx, op); + break; + CASE_OP_32_64_VEC(not): + done = fold_not(&ctx, op); + break; + CASE_OP_32_64_VEC(or): + done = fold_or(&ctx, op); + break; + CASE_OP_32_64_VEC(orc): + done = fold_orc(&ctx, op); + break; case INDEX_op_qemu_ld_i32: case INDEX_op_qemu_ld_i64: done = fold_qemu_ld(&ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_qemu_st_i64: done = fold_qemu_st(&ctx, op); break; - - default: + CASE_OP_32_64(rem): + CASE_OP_32_64(remu): + done = fold_remainder(&ctx, op); + break; + CASE_OP_32_64(rotl): + CASE_OP_32_64(rotr): + CASE_OP_32_64(sar): + CASE_OP_32_64(shl): + CASE_OP_32_64(shr): + done = fold_shift(&ctx, op); + break; + CASE_OP_32_64_VEC(sub): + done = fold_sub(&ctx, op); + break; + CASE_OP_32_64_VEC(xor): + done = fold_xor(&ctx, op); break; } -- 2.25.1 Reduce some code duplication by folding the NE and EQ cases. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 145 ++++++++++++++++++++++++------------------------- 1 file changed, 72 insertions(+), 73 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_remainder(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_setcond2(OptContext *ctx, TCGOp *op) +{ + TCGCond cond = op->args[5]; + int i = do_constant_folding_cond2(&op->args[1], &op->args[3], cond); + int inv = 0; + + if (i >= 0) { + goto do_setcond_const; + } + + switch (cond) { + case TCG_COND_LT: + case TCG_COND_GE: + /* + * Simplify LT/GE comparisons vs zero to a single compare + * vs the high word of the input. + */ + if (arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0 && + arg_is_const(op->args[4]) && arg_info(op->args[4])->val == 0) { + goto do_setcond_high; + } + break; + + case TCG_COND_NE: + inv = 1; + QEMU_FALLTHROUGH; + case TCG_COND_EQ: + /* + * Simplify EQ/NE comparisons where one of the pairs + * can be simplified. + */ + i = do_constant_folding_cond(INDEX_op_setcond_i32, op->args[1], + op->args[3], cond); + switch (i ^ inv) { + case 0: + goto do_setcond_const; + case 1: + goto do_setcond_high; + } + + i = do_constant_folding_cond(INDEX_op_setcond_i32, op->args[2], + op->args[4], cond); + switch (i ^ inv) { + case 0: + goto do_setcond_const; + case 1: + op->args[2] = op->args[3]; + op->args[3] = cond; + op->opc = INDEX_op_setcond_i32; + break; + } + break; + + default: + break; + + do_setcond_high: + op->args[1] = op->args[2]; + op->args[2] = op->args[4]; + op->args[3] = cond; + op->opc = INDEX_op_setcond_i32; + break; + } + return false; + + do_setcond_const: + return tcg_opt_gen_movi(ctx, op, op->args[0], i); +} + static bool fold_shift(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - case INDEX_op_setcond2_i32: - i = do_constant_folding_cond2(&op->args[1], &op->args[3], - op->args[5]); - if (i >= 0) { - do_setcond_const: - tcg_opt_gen_movi(&ctx, op, op->args[0], i); - continue; - } - if ((op->args[5] == TCG_COND_LT || op->args[5] == TCG_COND_GE) - && arg_is_const(op->args[3]) - && arg_info(op->args[3])->val == 0 - && arg_is_const(op->args[4]) - && arg_info(op->args[4])->val == 0) { - /* Simplify LT/GE comparisons vs zero to a single compare - vs the high word of the input. */ - do_setcond_high: - reset_temp(op->args[0]); - arg_info(op->args[0])->z_mask = 1; - op->opc = INDEX_op_setcond_i32; - op->args[1] = op->args[2]; - op->args[2] = op->args[4]; - op->args[3] = op->args[5]; - break; - } - if (op->args[5] == TCG_COND_EQ) { - /* Simplify EQ comparisons where one of the pairs - can be simplified. */ - i = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[1], op->args[3], - TCG_COND_EQ); - if (i == 0) { - goto do_setcond_const; - } else if (i > 0) { - goto do_setcond_high; - } - i = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[2], op->args[4], - TCG_COND_EQ); - if (i == 0) { - goto do_setcond_high; - } else if (i < 0) { - break; - } - do_setcond_low: - reset_temp(op->args[0]); - arg_info(op->args[0])->z_mask = 1; - op->opc = INDEX_op_setcond_i32; - op->args[2] = op->args[3]; - op->args[3] = op->args[5]; - break; - } - if (op->args[5] == TCG_COND_NE) { - /* Simplify NE comparisons where one of the pairs - can be simplified. */ - i = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[1], op->args[3], - TCG_COND_NE); - if (i == 0) { - goto do_setcond_high; - } else if (i > 0) { - goto do_setcond_const; - } - i = do_constant_folding_cond(INDEX_op_setcond_i32, - op->args[2], op->args[4], - TCG_COND_NE); - if (i == 0) { - goto do_setcond_low; - } else if (i > 0) { - goto do_setcond_const; - } - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(shr): done = fold_shift(&ctx, op); break; + case INDEX_op_setcond2_i32: + done = fold_setcond2(&ctx, op); + break; CASE_OP_32_64_VEC(sub): done = fold_sub(&ctx, op); break; -- 2.25.1 Reduce some code duplication by folding the NE and EQ cases. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 159 +++++++++++++++++++++++++------------------------ 1 file changed, 81 insertions(+), 78 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_brcond2(OptContext *ctx, TCGOp *op) +{ + TCGCond cond = op->args[4]; + int i = do_constant_folding_cond2(&op->args[0], &op->args[2], cond); + TCGArg label = op->args[5]; + int inv = 0; + + if (i >= 0) { + goto do_brcond_const; + } + + switch (cond) { + case TCG_COND_LT: + case TCG_COND_GE: + /* + * Simplify LT/GE comparisons vs zero to a single compare + * vs the high word of the input. + */ + if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == 0 && + arg_is_const(op->args[3]) && arg_info(op->args[3])->val == 0) { + goto do_brcond_high; + } + break; + + case TCG_COND_NE: + inv = 1; + QEMU_FALLTHROUGH; + case TCG_COND_EQ: + /* + * Simplify EQ/NE comparisons where one of the pairs + * can be simplified. + */ + i = do_constant_folding_cond(INDEX_op_brcond_i32, op->args[0], + op->args[2], cond); + switch (i ^ inv) { + case 0: + goto do_brcond_const; + case 1: + goto do_brcond_high; + } + + i = do_constant_folding_cond(INDEX_op_brcond_i32, op->args[1], + op->args[3], cond); + switch (i ^ inv) { + case 0: + goto do_brcond_const; + case 1: + op->opc = INDEX_op_brcond_i32; + op->args[1] = op->args[2]; + op->args[2] = cond; + op->args[3] = label; + break; + } + break; + + default: + break; + + do_brcond_high: + op->opc = INDEX_op_brcond_i32; + op->args[0] = op->args[1]; + op->args[1] = op->args[3]; + op->args[2] = cond; + op->args[3] = label; + break; + + do_brcond_const: + if (i == 0) { + tcg_op_remove(ctx->tcg, op); + return true; + } + op->opc = INDEX_op_br; + op->args[0] = label; + break; + } + return false; +} + static bool fold_call(OptContext *ctx, TCGOp *op) { TCGContext *s = ctx->tcg; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - case INDEX_op_brcond2_i32: - i = do_constant_folding_cond2(&op->args[0], &op->args[2], - op->args[4]); - if (i == 0) { - do_brcond_false: - tcg_op_remove(s, op); - continue; - } - if (i > 0) { - do_brcond_true: - op->opc = opc = INDEX_op_br; - op->args[0] = op->args[5]; - break; - } - if ((op->args[4] == TCG_COND_LT || op->args[4] == TCG_COND_GE) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == 0 - && arg_is_const(op->args[3]) - && arg_info(op->args[3])->val == 0) { - /* Simplify LT/GE comparisons vs zero to a single compare - vs the high word of the input. */ - do_brcond_high: - op->opc = opc = INDEX_op_brcond_i32; - op->args[0] = op->args[1]; - op->args[1] = op->args[3]; - op->args[2] = op->args[4]; - op->args[3] = op->args[5]; - break; - } - if (op->args[4] == TCG_COND_EQ) { - /* Simplify EQ comparisons where one of the pairs - can be simplified. */ - i = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[0], op->args[2], - TCG_COND_EQ); - if (i == 0) { - goto do_brcond_false; - } else if (i > 0) { - goto do_brcond_high; - } - i = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[1], op->args[3], - TCG_COND_EQ); - if (i == 0) { - goto do_brcond_false; - } else if (i < 0) { - break; - } - do_brcond_low: - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - op->opc = INDEX_op_brcond_i32; - op->args[1] = op->args[2]; - op->args[2] = op->args[4]; - op->args[3] = op->args[5]; - break; - } - if (op->args[4] == TCG_COND_NE) { - /* Simplify NE comparisons where one of the pairs - can be simplified. */ - i = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[0], op->args[2], - TCG_COND_NE); - if (i == 0) { - goto do_brcond_high; - } else if (i > 0) { - goto do_brcond_true; - } - i = do_constant_folding_cond(INDEX_op_brcond_i32, - op->args[1], op->args[3], - TCG_COND_NE); - if (i == 0) { - goto do_brcond_low; - } else if (i > 0) { - goto do_brcond_true; - } - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(andc): done = fold_andc(&ctx, op); break; + case INDEX_op_brcond2_i32: + done = fold_brcond2(&ctx, op); + break; CASE_OP_32_64(ctpop): done = fold_ctpop(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 33 +++++++++++++++++++-------------- 1 file changed, 19 insertions(+), 14 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_brcond(OptContext *ctx, TCGOp *op) +{ + TCGCond cond = op->args[2]; + int i = do_constant_folding_cond(op->opc, op->args[0], op->args[1], cond); + + if (i == 0) { + tcg_op_remove(ctx->tcg, op); + return true; + } + if (i > 0) { + op->opc = INDEX_op_br; + op->args[0] = op->args[3]; + } + return false; +} + static bool fold_brcond2(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[4]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(brcond): - i = do_constant_folding_cond(opc, op->args[0], - op->args[1], op->args[2]); - if (i == 0) { - tcg_op_remove(s, op); - continue; - } else if (i > 0) { - memset(&ctx.temps_used, 0, sizeof(ctx.temps_used)); - op->opc = opc = INDEX_op_br; - op->args[0] = op->args[3]; - break; - } - break; - CASE_OP_32_64(movcond): i = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[5]); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(andc): done = fold_andc(&ctx, op); break; + CASE_OP_32_64(brcond): + done = fold_brcond(&ctx, op); + break; case INDEX_op_brcond2_i32: done = fold_brcond2(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 23 ++++++++++++++--------- 1 file changed, 14 insertions(+), 9 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_remainder(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_setcond(OptContext *ctx, TCGOp *op) +{ + TCGCond cond = op->args[3]; + int i = do_constant_folding_cond(op->opc, op->args[1], op->args[2], cond); + + if (i >= 0) { + return tcg_opt_gen_movi(ctx, op, op->args[0], i); + } + return false; +} + static bool fold_setcond2(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[5]; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(setcond): - i = do_constant_folding_cond(opc, op->args[1], - op->args[2], op->args[3]); - if (i >= 0) { - tcg_opt_gen_movi(&ctx, op, op->args[0], i); - continue; - } - break; - CASE_OP_32_64(movcond): i = do_constant_folding_cond(opc, op->args[1], op->args[2], op->args[5]); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(shr): done = fold_shift(&ctx, op); break; + CASE_OP_32_64(setcond): + done = fold_setcond(&ctx, op); + break; case INDEX_op_setcond2_i32: done = fold_setcond2(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 37 +++++++++++++++++++++---------------- 1 file changed, 21 insertions(+), 16 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_mulu2_i32(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { + uint32_t a = arg_info(op->args[2])->val; + uint32_t b = arg_info(op->args[3])->val; + uint64_t r = (uint64_t)a * b; + TCGArg rl, rh; + TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_mov_i32); + + rl = op->args[0]; + rh = op->args[1]; + tcg_opt_gen_movi(ctx, op, rl, (int32_t)r); + tcg_opt_gen_movi(ctx, op2, rh, (int32_t)(r >> 32)); + return true; + } + return false; +} + static bool fold_nand(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - case INDEX_op_mulu2_i32: - if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { - uint32_t a = arg_info(op->args[2])->val; - uint32_t b = arg_info(op->args[3])->val; - uint64_t r = (uint64_t)a * b; - TCGArg rl, rh; - TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_mov_i32); - - rl = op->args[0]; - rh = op->args[1]; - tcg_opt_gen_movi(&ctx, op, rl, (int32_t)r); - tcg_opt_gen_movi(&ctx, op2, rh, (int32_t)(r >> 32)); - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(muluh): done = fold_mul_highpart(&ctx, op); break; + case INDEX_op_mulu2_i32: + done = fold_mulu2_i32(&ctx, op); + break; CASE_OP_32_64(nand): done = fold_nand(&ctx, op); break; -- 2.25.1 Add two additional helpers, fold_add2_i32 and fold_sub2_i32 which will not be simple wrappers forever. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 70 +++++++++++++++++++++++++++++++------------------- 1 file changed, 44 insertions(+), 26 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_add(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_addsub2_i32(OptContext *ctx, TCGOp *op, bool add) +{ + if (arg_is_const(op->args[2]) && arg_is_const(op->args[3]) && + arg_is_const(op->args[4]) && arg_is_const(op->args[5])) { + uint32_t al = arg_info(op->args[2])->val; + uint32_t ah = arg_info(op->args[3])->val; + uint32_t bl = arg_info(op->args[4])->val; + uint32_t bh = arg_info(op->args[5])->val; + uint64_t a = ((uint64_t)ah << 32) | al; + uint64_t b = ((uint64_t)bh << 32) | bl; + TCGArg rl, rh; + TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_mov_i32); + + if (add) { + a += b; + } else { + a -= b; + } + + rl = op->args[0]; + rh = op->args[1]; + tcg_opt_gen_movi(ctx, op, rl, (int32_t)a); + tcg_opt_gen_movi(ctx, op2, rh, (int32_t)(a >> 32)); + return true; + } + return false; +} + +static bool fold_add2_i32(OptContext *ctx, TCGOp *op) +{ + return fold_addsub2_i32(ctx, op, true); +} + static bool fold_and(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ static bool fold_sub(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) +{ + return fold_addsub2_i32(ctx, op, false); +} + static bool fold_xor(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - case INDEX_op_add2_i32: - case INDEX_op_sub2_i32: - if (arg_is_const(op->args[2]) && arg_is_const(op->args[3]) - && arg_is_const(op->args[4]) && arg_is_const(op->args[5])) { - uint32_t al = arg_info(op->args[2])->val; - uint32_t ah = arg_info(op->args[3])->val; - uint32_t bl = arg_info(op->args[4])->val; - uint32_t bh = arg_info(op->args[5])->val; - uint64_t a = ((uint64_t)ah << 32) | al; - uint64_t b = ((uint64_t)bh << 32) | bl; - TCGArg rl, rh; - TCGOp *op2 = tcg_op_insert_before(s, op, INDEX_op_mov_i32); - - if (opc == INDEX_op_add2_i32) { - a += b; - } else { - a -= b; - } - - rl = op->args[0]; - rh = op->args[1]; - tcg_opt_gen_movi(&ctx, op, rl, (int32_t)a); - tcg_opt_gen_movi(&ctx, op2, rh, (int32_t)(a >> 32)); - continue; - } - break; default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(add): done = fold_add(&ctx, op); break; + case INDEX_op_add2_i32: + done = fold_add2_i32(&ctx, op); + break; CASE_OP_32_64_VEC(and): done = fold_and(&ctx, op); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(sub): done = fold_sub(&ctx, op); break; + case INDEX_op_sub2_i32: + done = fold_sub2_i32(&ctx, op); + break; CASE_OP_32_64_VEC(xor): done = fold_xor(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 56 ++++++++++++++++++++++++++++---------------------- 1 file changed, 31 insertions(+), 25 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_mb(OptContext *ctx, TCGOp *op) return true; } +static bool fold_movcond(OptContext *ctx, TCGOp *op) +{ + TCGOpcode opc = op->opc; + TCGCond cond = op->args[5]; + int i = do_constant_folding_cond(opc, op->args[1], op->args[2], cond); + + if (i >= 0) { + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); + } + + if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { + uint64_t tv = arg_info(op->args[3])->val; + uint64_t fv = arg_info(op->args[4])->val; + + opc = (opc == INDEX_op_movcond_i32 + ? INDEX_op_setcond_i32 : INDEX_op_setcond_i64); + + if (tv == 1 && fv == 0) { + op->opc = opc; + op->args[3] = cond; + } else if (fv == 1 && tv == 0) { + op->opc = opc; + op->args[3] = tcg_invert_cond(cond); + } + } + return false; +} + static bool fold_mul(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(movcond): - i = do_constant_folding_cond(opc, op->args[1], - op->args[2], op->args[5]); - if (i >= 0) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[4 - i]); - continue; - } - if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { - uint64_t tv = arg_info(op->args[3])->val; - uint64_t fv = arg_info(op->args[4])->val; - TCGCond cond = op->args[5]; - - if (fv == 1 && tv == 0) { - cond = tcg_invert_cond(cond); - } else if (!(tv == 1 && fv == 0)) { - break; - } - op->args[3] = cond; - op->opc = opc = (opc == INDEX_op_movcond_i32 - ? INDEX_op_setcond_i32 - : INDEX_op_setcond_i64); - } - break; - - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_mb: done = fold_mb(&ctx, op); break; + CASE_OP_32_64(movcond): + done = fold_movcond(&ctx, op); + break; CASE_OP_32_64(mul): done = fold_mul(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 39 ++++++++++++++++++++++----------------- 1 file changed, 22 insertions(+), 17 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_eqv(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_extract2(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { + uint64_t v1 = arg_info(op->args[1])->val; + uint64_t v2 = arg_info(op->args[2])->val; + int shr = op->args[3]; + + if (op->opc == INDEX_op_extract2_i64) { + v1 >>= shr; + v2 <<= 64 - shr; + } else { + v1 = (uint32_t)v1 >> shr; + v2 = (int32_t)v2 << (32 - shr); + } + return tcg_opt_gen_movi(ctx, op, op->args[0], v1 | v2); + } + return false; +} + static bool fold_exts(OptContext *ctx, TCGOp *op) { return fold_const1(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(extract2): - if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - uint64_t v1 = arg_info(op->args[1])->val; - uint64_t v2 = arg_info(op->args[2])->val; - int shr = op->args[3]; - - if (opc == INDEX_op_extract2_i64) { - tmp = (v1 >> shr) | (v2 << (64 - shr)); - } else { - tmp = (int32_t)(((uint32_t)v1 >> shr) | - ((uint32_t)v2 << (32 - shr))); - } - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(eqv): done = fold_eqv(&ctx, op); break; + CASE_OP_32_64(extract2): + done = fold_extract2(&ctx, op); + break; CASE_OP_32_64(ext8s): CASE_OP_32_64(ext16s): case INDEX_op_ext32s_i64: -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 48 ++++++++++++++++++++++++++++++------------------ 1 file changed, 30 insertions(+), 18 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_eqv(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_extract(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t; + + t = arg_info(op->args[1])->val; + t = extract64(t, op->args[2], op->args[3]); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return false; +} + static bool fold_extract2(OptContext *ctx, TCGOp *op) { if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) return tcg_opt_gen_movi(ctx, op, op->args[0], i); } +static bool fold_sextract(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t; + + t = arg_info(op->args[1])->val; + t = sextract64(t, op->args[2], op->args[3]); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return false; +} + static bool fold_shift(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(extract): - if (arg_is_const(op->args[1])) { - tmp = extract64(arg_info(op->args[1])->val, - op->args[2], op->args[3]); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - - CASE_OP_32_64(sextract): - if (arg_is_const(op->args[1])) { - tmp = sextract64(arg_info(op->args[1])->val, - op->args[2], op->args[3]); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(eqv): done = fold_eqv(&ctx, op); break; + CASE_OP_32_64(extract): + done = fold_extract(&ctx, op); + break; CASE_OP_32_64(extract2): done = fold_extract2(&ctx, op); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_setcond2_i32: done = fold_setcond2(&ctx, op); break; + CASE_OP_32_64(sextract): + done = fold_sextract(&ctx, op); + break; CASE_OP_32_64_VEC(sub): done = fold_sub(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 25 +++++++++++++++---------- 1 file changed, 15 insertions(+), 10 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_ctpop(OptContext *ctx, TCGOp *op) return fold_const1(ctx, op); } +static bool fold_deposit(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { + uint64_t t1 = arg_info(op->args[1])->val; + uint64_t t2 = arg_info(op->args[2])->val; + + t1 = deposit64(t1, op->args[3], op->args[4], t2); + return tcg_opt_gen_movi(ctx, op, op->args[0], t1); + } + return false; +} + static bool fold_divide(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(deposit): - if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - tmp = deposit64(arg_info(op->args[1])->val, - op->args[3], op->args[4], - arg_info(op->args[2])->val); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(ctpop): done = fold_ctpop(&ctx, op); break; + CASE_OP_32_64(deposit): + done = fold_deposit(&ctx, op); + break; CASE_OP_32_64(div): CASE_OP_32_64(divu): done = fold_divide(&ctx, op); -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 32 ++++++++++++++++++-------------- 1 file changed, 18 insertions(+), 14 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_call(OptContext *ctx, TCGOp *op) return true; } +static bool fold_count_zeros(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t = arg_info(op->args[1])->val; + + if (t != 0) { + t = do_constant_folding(op->opc, t, 0); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); + } + return false; +} + static bool fold_ctpop(OptContext *ctx, TCGOp *op) { return fold_const1(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(clz): - CASE_OP_32_64(ctz): - if (arg_is_const(op->args[1])) { - TCGArg v = arg_info(op->args[1])->val; - if (v != 0) { - tmp = do_constant_folding(opc, v, 0); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - } else { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[2]); - } - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_brcond2_i32: done = fold_brcond2(&ctx, op); break; + CASE_OP_32_64(clz): + CASE_OP_32_64(ctz): + done = fold_count_zeros(&ctx, op); + break; CASE_OP_32_64(ctpop): done = fold_ctpop(&ctx, op); break; -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 27 ++++++++++++++++----------- 1 file changed, 16 insertions(+), 11 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op) return false; } +static bool fold_bswap(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t = arg_info(op->args[1])->val; + + t = do_constant_folding(op->opc, t, op->args[2]); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return false; +} + static bool fold_call(OptContext *ctx, TCGOp *op) { TCGContext *s = ctx->tcg; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } break; - CASE_OP_32_64(bswap16): - CASE_OP_32_64(bswap32): - case INDEX_op_bswap64_i64: - if (arg_is_const(op->args[1])) { - tmp = do_constant_folding(opc, arg_info(op->args[1])->val, - op->args[2]); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_brcond2_i32: done = fold_brcond2(&ctx, op); break; + CASE_OP_32_64(bswap16): + CASE_OP_32_64(bswap32): + case INDEX_op_bswap64_i64: + done = fold_bswap(&ctx, op); + break; CASE_OP_32_64(clz): CASE_OP_32_64(ctz): done = fold_count_zeros(&ctx, op); -- 2.25.1 Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 53 +++++++++++++++++++++++++++++--------------------- 1 file changed, 31 insertions(+), 22 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_divide(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_dup(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1])) { + uint64_t t = arg_info(op->args[1])->val; + t = dup_const(TCGOP_VECE(op), t); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + return false; +} + +static bool fold_dup2(OptContext *ctx, TCGOp *op) +{ + if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { + uint64_t t = deposit64(arg_info(op->args[1])->val, 32, 32, + arg_info(op->args[2])->val); + return tcg_opt_gen_movi(ctx, op, op->args[0], t); + } + + if (args_are_copies(op->args[1], op->args[2])) { + op->opc = INDEX_op_dup_vec; + TCGOP_VECE(op) = MO_32; + } + return false; +} + static bool fold_eqv(OptContext *ctx, TCGOp *op) { return fold_const2(ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) done = tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); break; - case INDEX_op_dup_vec: - if (arg_is_const(op->args[1])) { - tmp = arg_info(op->args[1])->val; - tmp = dup_const(TCGOP_VECE(op), tmp); - tcg_opt_gen_movi(&ctx, op, op->args[0], tmp); - continue; - } - break; - - case INDEX_op_dup2_vec: - assert(TCG_TARGET_REG_BITS == 32); - if (arg_is_const(op->args[1]) && arg_is_const(op->args[2])) { - tcg_opt_gen_movi(&ctx, op, op->args[0], - deposit64(arg_info(op->args[1])->val, 32, 32, - arg_info(op->args[2])->val)); - continue; - } else if (args_are_copies(op->args[1], op->args[2])) { - op->opc = INDEX_op_dup_vec; - TCGOP_VECE(op) = MO_32; - } - break; - default: break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(divu): done = fold_divide(&ctx, op); break; + case INDEX_op_dup_vec: + done = fold_dup(&ctx, op); + break; + case INDEX_op_dup2_vec: + done = fold_dup2(&ctx, op); + break; CASE_OP_32_64(eqv): done = fold_eqv(&ctx, op); break; -- 2.25.1 This is the final entry in the main switch that was in a different form. After this, we have the option to convert the switch into a function dispatch table. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 27 ++++++++++++++------------- 1 file changed, 14 insertions(+), 13 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_mb(OptContext *ctx, TCGOp *op) return true; } +static bool fold_mov(OptContext *ctx, TCGOp *op) +{ + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); +} + static bool fold_movcond(OptContext *ctx, TCGOp *op) { TCGOpcode opc = op->opc; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Propagate constants through copy operations and do constant - folding. Constants will be substituted to arguments by register - allocator where needed and possible. Also detect copies. */ + /* + * Process each opcode. + * Sorted alphabetically by opcode as much as possible. + */ switch (opc) { - CASE_OP_32_64_VEC(mov): - done = tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - break; - - default: - break; - - /* ---------------------------------------------------------- */ - /* Sorted alphabetically by opcode as much as possible. */ - CASE_OP_32_64_VEC(add): done = fold_add(&ctx, op); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_mb: done = fold_mb(&ctx, op); break; + CASE_OP_32_64_VEC(mov): + done = fold_mov(&ctx, op); + break; CASE_OP_32_64(movcond): done = fold_movcond(&ctx, op); break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(xor): done = fold_xor(&ctx, op); break; + default: + break; } if (!done) { -- 2.25.1 Pull the "op r, a, a => movi r, 0" optimization into a function, and use it in the outer opcode fold functions. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 41 ++++++++++++++++++++++++----------------- 1 file changed, 24 insertions(+), 17 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) return false; } +/* If the binary operation has both arguments equal, fold to @i. */ +static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (args_are_copies(op->args[1], op->args[2])) { + return tcg_opt_gen_movi(ctx, op, op->args[0], i); + } + return false; +} + /* * These outermost fold_<op> functions are sorted alphabetically. */ @@ -XXX,XX +XXX,XX @@ static bool fold_and(OptContext *ctx, TCGOp *op) static bool fold_andc(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_i(ctx, op, 0)) { + return true; + } + return false; } static bool fold_brcond(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_shift(OptContext *ctx, TCGOp *op) static bool fold_sub(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_i(ctx, op, 0)) { + return true; + } + return false; } static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) static bool fold_xor(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_i(ctx, op, 0)) { + return true; + } + return false; } /* Propagate constants and copies, fold constant expressions. */ @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Simplify expression for "op r, a, a => movi r, 0" cases */ - switch (opc) { - CASE_OP_32_64_VEC(andc): - CASE_OP_32_64_VEC(sub): - CASE_OP_32_64_VEC(xor): - if (args_are_copies(op->args[1], op->args[2])) { - tcg_opt_gen_movi(&ctx, op, op->args[0], 0); - continue; - } - break; - default: - break; - } - /* * Process each opcode. * Sorted alphabetically by opcode as much as possible. -- 2.25.1 Pull the "op r, a, a => mov r, a" optimization into a function, and use it in the outer opcode fold functions. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 39 ++++++++++++++++++++++++--------------- 1 file changed, 24 insertions(+), 15 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) return false; } +/* If the binary operation has both arguments equal, fold to identity. */ +static bool fold_xx_to_x(OptContext *ctx, TCGOp *op) +{ + if (args_are_copies(op->args[1], op->args[2])) { + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); + } + return false; +} + /* * These outermost fold_<op> functions are sorted alphabetically. + * + * The ordering of the transformations should be: + * 1) those that produce a constant + * 2) those that produce a copy + * 3) those that produce information about the result value. */ static bool fold_add(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_add2_i32(OptContext *ctx, TCGOp *op) static bool fold_and(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_x(ctx, op)) { + return true; + } + return false; } static bool fold_andc(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_not(OptContext *ctx, TCGOp *op) static bool fold_or(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_x(ctx, op)) { + return true; + } + return false; } static bool fold_orc(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Simplify expression for "op r, a, a => mov r, a" cases */ - switch (opc) { - CASE_OP_32_64_VEC(or): - CASE_OP_32_64_VEC(and): - if (args_are_copies(op->args[1], op->args[2])) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - continue; - } - break; - default: - break; - } - /* * Process each opcode. * Sorted alphabetically by opcode as much as possible. -- 2.25.1 Pull the "op r, a, 0 => movi r, 0" optimization into a function, and use it in the outer opcode fold functions. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 38 ++++++++++++++++++++------------------ 1 file changed, 20 insertions(+), 18 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) return false; } +/* If the binary operation has second argument @i, fold to @i. */ +static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) { + return tcg_opt_gen_movi(ctx, op, op->args[0], i); + } + return false; +} + /* If the binary operation has both arguments equal, fold to @i. */ static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) { @@ -XXX,XX +XXX,XX @@ static bool fold_add2_i32(OptContext *ctx, TCGOp *op) static bool fold_and(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_xi_to_i(ctx, op, 0) || fold_xx_to_x(ctx, op)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_movcond(OptContext *ctx, TCGOp *op) static bool fold_mul(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_i(ctx, op, 0)) { + return true; + } + return false; } static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_i(ctx, op, 0)) { + return true; + } + return false; } static bool fold_mulu2_i32(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) continue; } - /* Simplify expression for "op r, a, 0 => movi r, 0" cases */ - switch (opc) { - CASE_OP_32_64_VEC(and): - CASE_OP_32_64_VEC(mul): - CASE_OP_32_64(muluh): - CASE_OP_32_64(mulsh): - if (arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == 0) { - tcg_opt_gen_movi(&ctx, op, op->args[0], 0); - continue; - } - break; - default: - break; - } - /* * Process each opcode. * Sorted alphabetically by opcode as much as possible. -- 2.25.1 Compute the type of the operation early. There are at least 4 places that used a def->flags ladder to determine the type of the operation being optimized. There were two places that assumed !TCG_OPF_64BIT means TCG_TYPE_I32, and so could potentially compute incorrect results for vector operations. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 149 +++++++++++++++++++++++++++++-------------------- 1 file changed, 89 insertions(+), 60 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct OptContext { /* In flight values from optimization. */ uint64_t z_mask; + TCGType type; } OptContext; static inline TempOptInfo *ts_info(TCGTemp *ts) @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) { TCGTemp *dst_ts = arg_temp(dst); TCGTemp *src_ts = arg_temp(src); - const TCGOpDef *def; TempOptInfo *di; TempOptInfo *si; uint64_t z_mask; @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) reset_ts(dst_ts); di = ts_info(dst_ts); si = ts_info(src_ts); - def = &tcg_op_defs[op->opc]; - if (def->flags & TCG_OPF_VECTOR) { - new_op = INDEX_op_mov_vec; - } else if (def->flags & TCG_OPF_64BIT) { - new_op = INDEX_op_mov_i64; - } else { + + switch (ctx->type) { + case TCG_TYPE_I32: new_op = INDEX_op_mov_i32; + break; + case TCG_TYPE_I64: + new_op = INDEX_op_mov_i64; + break; + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + /* TCGOP_VECL and TCGOP_VECE remain unchanged. */ + new_op = INDEX_op_mov_vec; + break; + default: + g_assert_not_reached(); } op->opc = new_op; - /* TCGOP_VECL and TCGOP_VECE remain unchanged. */ op->args[0] = dst; op->args[1] = src; @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, TCGArg dst, uint64_t val) { - const TCGOpDef *def = &tcg_op_defs[op->opc]; - TCGType type; - TCGTemp *tv; - - if (def->flags & TCG_OPF_VECTOR) { - type = TCGOP_VECL(op) + TCG_TYPE_V64; - } else if (def->flags & TCG_OPF_64BIT) { - type = TCG_TYPE_I64; - } else { - type = TCG_TYPE_I32; - } - /* Convert movi to mov with constant temp. */ - tv = tcg_constant_internal(type, val); + TCGTemp *tv = tcg_constant_internal(ctx->type, val); + init_ts_info(ctx, tv); return tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv)); } @@ -XXX,XX +XXX,XX @@ static uint64_t do_constant_folding_2(TCGOpcode op, uint64_t x, uint64_t y) } } -static uint64_t do_constant_folding(TCGOpcode op, uint64_t x, uint64_t y) +static uint64_t do_constant_folding(TCGOpcode op, TCGType type, + uint64_t x, uint64_t y) { - const TCGOpDef *def = &tcg_op_defs[op]; uint64_t res = do_constant_folding_2(op, x, y); - if (!(def->flags & TCG_OPF_64BIT)) { + if (type == TCG_TYPE_I32) { res = (int32_t)res; } return res; @@ -XXX,XX +XXX,XX @@ static bool do_constant_folding_cond_eq(TCGCond c) * Return -1 if the condition can't be simplified, * and the result of the condition (0 or 1) if it can. */ -static int do_constant_folding_cond(TCGOpcode op, TCGArg x, +static int do_constant_folding_cond(TCGType type, TCGArg x, TCGArg y, TCGCond c) { uint64_t xv = arg_info(x)->val; uint64_t yv = arg_info(y)->val; if (arg_is_const(x) && arg_is_const(y)) { - const TCGOpDef *def = &tcg_op_defs[op]; - tcg_debug_assert(!(def->flags & TCG_OPF_VECTOR)); - if (def->flags & TCG_OPF_64BIT) { - return do_constant_folding_cond_64(xv, yv, c); - } else { + switch (type) { + case TCG_TYPE_I32: return do_constant_folding_cond_32(xv, yv, c); + case TCG_TYPE_I64: + return do_constant_folding_cond_64(xv, yv, c); + default: + /* Only scalar comparisons are optimizable */ + return -1; } } else if (args_are_copies(x, y)) { return do_constant_folding_cond_eq(c); @@ -XXX,XX +XXX,XX @@ static bool fold_const1(OptContext *ctx, TCGOp *op) uint64_t t; t = arg_info(op->args[1])->val; - t = do_constant_folding(op->opc, t, 0); + t = do_constant_folding(op->opc, ctx->type, t, 0); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } return false; @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) uint64_t t1 = arg_info(op->args[1])->val; uint64_t t2 = arg_info(op->args[2])->val; - t1 = do_constant_folding(op->opc, t1, t2); + t1 = do_constant_folding(op->opc, ctx->type, t1, t2); return tcg_opt_gen_movi(ctx, op, op->args[0], t1); } return false; @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) static bool fold_brcond(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[2]; - int i = do_constant_folding_cond(op->opc, op->args[0], op->args[1], cond); + int i = do_constant_folding_cond(ctx->type, op->args[0], op->args[1], cond); if (i == 0) { tcg_op_remove(ctx->tcg, op); @@ -XXX,XX +XXX,XX @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op) * Simplify EQ/NE comparisons where one of the pairs * can be simplified. */ - i = do_constant_folding_cond(INDEX_op_brcond_i32, op->args[0], + i = do_constant_folding_cond(TCG_TYPE_I32, op->args[0], op->args[2], cond); switch (i ^ inv) { case 0: @@ -XXX,XX +XXX,XX @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op) goto do_brcond_high; } - i = do_constant_folding_cond(INDEX_op_brcond_i32, op->args[1], + i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], op->args[3], cond); switch (i ^ inv) { case 0: @@ -XXX,XX +XXX,XX @@ static bool fold_bswap(OptContext *ctx, TCGOp *op) if (arg_is_const(op->args[1])) { uint64_t t = arg_info(op->args[1])->val; - t = do_constant_folding(op->opc, t, op->args[2]); + t = do_constant_folding(op->opc, ctx->type, t, op->args[2]); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } return false; @@ -XXX,XX +XXX,XX @@ static bool fold_count_zeros(OptContext *ctx, TCGOp *op) uint64_t t = arg_info(op->args[1])->val; if (t != 0) { - t = do_constant_folding(op->opc, t, 0); + t = do_constant_folding(op->opc, ctx->type, t, 0); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); @@ -XXX,XX +XXX,XX @@ static bool fold_mov(OptContext *ctx, TCGOp *op) static bool fold_movcond(OptContext *ctx, TCGOp *op) { - TCGOpcode opc = op->opc; TCGCond cond = op->args[5]; - int i = do_constant_folding_cond(opc, op->args[1], op->args[2], cond); + int i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); if (i >= 0) { return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); @@ -XXX,XX +XXX,XX @@ static bool fold_movcond(OptContext *ctx, TCGOp *op) if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { uint64_t tv = arg_info(op->args[3])->val; uint64_t fv = arg_info(op->args[4])->val; + TCGOpcode opc; - opc = (opc == INDEX_op_movcond_i32 - ? INDEX_op_setcond_i32 : INDEX_op_setcond_i64); + switch (ctx->type) { + case TCG_TYPE_I32: + opc = INDEX_op_setcond_i32; + break; + case TCG_TYPE_I64: + opc = INDEX_op_setcond_i64; + break; + default: + g_assert_not_reached(); + } if (tv == 1 && fv == 0) { op->opc = opc; @@ -XXX,XX +XXX,XX @@ static bool fold_remainder(OptContext *ctx, TCGOp *op) static bool fold_setcond(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[3]; - int i = do_constant_folding_cond(op->opc, op->args[1], op->args[2], cond); + int i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); if (i >= 0) { return tcg_opt_gen_movi(ctx, op, op->args[0], i); @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) * Simplify EQ/NE comparisons where one of the pairs * can be simplified. */ - i = do_constant_folding_cond(INDEX_op_setcond_i32, op->args[1], + i = do_constant_folding_cond(TCG_TYPE_I32, op->args[1], op->args[3], cond); switch (i ^ inv) { case 0: @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) goto do_setcond_high; } - i = do_constant_folding_cond(INDEX_op_setcond_i32, op->args[2], + i = do_constant_folding_cond(TCG_TYPE_I32, op->args[2], op->args[4], cond); switch (i ^ inv) { case 0: @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) init_arguments(&ctx, op, def->nb_oargs + def->nb_iargs); copy_propagate(&ctx, op, def->nb_oargs, def->nb_iargs); + /* Pre-compute the type of the operation. */ + if (def->flags & TCG_OPF_VECTOR) { + ctx.type = TCG_TYPE_V64 + TCGOP_VECL(op); + } else if (def->flags & TCG_OPF_64BIT) { + ctx.type = TCG_TYPE_I64; + } else { + ctx.type = TCG_TYPE_I32; + } + /* For commutative operations make constant second argument */ switch (opc) { CASE_OP_32_64_VEC(add): @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) /* Proceed with possible constant folding. */ break; } - if (opc == INDEX_op_sub_i32) { + switch (ctx.type) { + case TCG_TYPE_I32: neg_op = INDEX_op_neg_i32; have_neg = TCG_TARGET_HAS_neg_i32; - } else if (opc == INDEX_op_sub_i64) { + break; + case TCG_TYPE_I64: neg_op = INDEX_op_neg_i64; have_neg = TCG_TARGET_HAS_neg_i64; - } else if (TCG_TARGET_HAS_neg_vec) { - TCGType type = TCGOP_VECL(op) + TCG_TYPE_V64; - unsigned vece = TCGOP_VECE(op); - neg_op = INDEX_op_neg_vec; - have_neg = tcg_can_emit_vec_op(neg_op, type, vece) > 0; - } else { break; + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + neg_op = INDEX_op_neg_vec; + have_neg = tcg_can_emit_vec_op(neg_op, ctx.type, + TCGOP_VECE(op)) > 0; + break; + default: + g_assert_not_reached(); } if (!have_neg) { break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) TCGOpcode not_op; bool have_not; - if (def->flags & TCG_OPF_VECTOR) { - not_op = INDEX_op_not_vec; - have_not = TCG_TARGET_HAS_not_vec; - } else if (def->flags & TCG_OPF_64BIT) { - not_op = INDEX_op_not_i64; - have_not = TCG_TARGET_HAS_not_i64; - } else { + switch (ctx.type) { + case TCG_TYPE_I32: not_op = INDEX_op_not_i32; have_not = TCG_TARGET_HAS_not_i32; + break; + case TCG_TYPE_I64: + not_op = INDEX_op_not_i64; + have_not = TCG_TARGET_HAS_not_i64; + break; + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + not_op = INDEX_op_not_vec; + have_not = TCG_TARGET_HAS_not_vec; + break; + default: + g_assert_not_reached(); } if (!have_not) { break; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) below, we can ignore high bits, but for further optimizations we need to record that the high bits contain garbage. */ partmask = z_mask; - if (!(def->flags & TCG_OPF_64BIT)) { + if (ctx.type == TCG_TYPE_I32) { z_mask |= ~(tcg_target_ulong)0xffffffffu; partmask &= 0xffffffffu; affected &= 0xffffffffu; -- 2.25.1 Split out the conditional conversion from a more complex logical operation to a simple NOT. Create a couple more helpers to make this easy for the outer-most logical operations. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 158 +++++++++++++++++++++++++++---------------------- 1 file changed, 86 insertions(+), 72 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) return false; } +/* + * Convert @op to NOT, if NOT is supported by the host. + * Return true f the conversion is successful, which will still + * indicate that the processing is complete. + */ +static bool fold_not(OptContext *ctx, TCGOp *op); +static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx) +{ + TCGOpcode not_op; + bool have_not; + + switch (ctx->type) { + case TCG_TYPE_I32: + not_op = INDEX_op_not_i32; + have_not = TCG_TARGET_HAS_not_i32; + break; + case TCG_TYPE_I64: + not_op = INDEX_op_not_i64; + have_not = TCG_TARGET_HAS_not_i64; + break; + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + not_op = INDEX_op_not_vec; + have_not = TCG_TARGET_HAS_not_vec; + break; + default: + g_assert_not_reached(); + } + if (have_not) { + op->opc = not_op; + op->args[1] = op->args[idx]; + return fold_not(ctx, op); + } + return false; +} + +/* If the binary operation has first argument @i, fold to NOT. */ +static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) { + return fold_to_not(ctx, op, 2); + } + return false; +} + /* If the binary operation has second argument @i, fold to @i. */ static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) { @@ -XXX,XX +XXX,XX @@ static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) return false; } +/* If the binary operation has second argument @i, fold to NOT. */ +static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) { + return fold_to_not(ctx, op, 1); + } + return false; +} + /* If the binary operation has both arguments equal, fold to @i. */ static bool fold_xx_to_i(OptContext *ctx, TCGOp *op, uint64_t i) { @@ -XXX,XX +XXX,XX @@ static bool fold_and(OptContext *ctx, TCGOp *op) static bool fold_andc(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || - fold_xx_to_i(ctx, op, 0)) { + fold_xx_to_i(ctx, op, 0) || + fold_ix_to_not(ctx, op, -1)) { return true; } return false; @@ -XXX,XX +XXX,XX @@ static bool fold_dup2(OptContext *ctx, TCGOp *op) static bool fold_eqv(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_not(ctx, op, 0)) { + return true; + } + return false; } static bool fold_extract(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_mulu2_i32(OptContext *ctx, TCGOp *op) static bool fold_nand(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_not(ctx, op, -1)) { + return true; + } + return false; } static bool fold_neg(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_neg(OptContext *ctx, TCGOp *op) static bool fold_nor(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_not(ctx, op, 0)) { + return true; + } + return false; } static bool fold_not(OptContext *ctx, TCGOp *op) { - return fold_const1(ctx, op); + if (fold_const1(ctx, op)) { + return true; + } + + /* Because of fold_to_not, we want to always return true, via finish. */ + finish_folding(ctx, op); + return true; } static bool fold_or(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_or(OptContext *ctx, TCGOp *op) static bool fold_orc(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_ix_to_not(ctx, op, 0)) { + return true; + } + return false; } static bool fold_qemu_ld(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) static bool fold_xor(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || - fold_xx_to_i(ctx, op, 0)) { + fold_xx_to_i(ctx, op, 0) || + fold_xi_to_not(ctx, op, -1)) { return true; } return false; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } } break; - CASE_OP_32_64_VEC(xor): - CASE_OP_32_64(nand): - if (!arg_is_const(op->args[1]) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == -1) { - i = 1; - goto try_not; - } - break; - CASE_OP_32_64(nor): - if (!arg_is_const(op->args[1]) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == 0) { - i = 1; - goto try_not; - } - break; - CASE_OP_32_64_VEC(andc): - if (!arg_is_const(op->args[2]) - && arg_is_const(op->args[1]) - && arg_info(op->args[1])->val == -1) { - i = 2; - goto try_not; - } - break; - CASE_OP_32_64_VEC(orc): - CASE_OP_32_64(eqv): - if (!arg_is_const(op->args[2]) - && arg_is_const(op->args[1]) - && arg_info(op->args[1])->val == 0) { - i = 2; - goto try_not; - } - break; - try_not: - { - TCGOpcode not_op; - bool have_not; - - switch (ctx.type) { - case TCG_TYPE_I32: - not_op = INDEX_op_not_i32; - have_not = TCG_TARGET_HAS_not_i32; - break; - case TCG_TYPE_I64: - not_op = INDEX_op_not_i64; - have_not = TCG_TARGET_HAS_not_i64; - break; - case TCG_TYPE_V64: - case TCG_TYPE_V128: - case TCG_TYPE_V256: - not_op = INDEX_op_not_vec; - have_not = TCG_TARGET_HAS_not_vec; - break; - default: - g_assert_not_reached(); - } - if (!have_not) { - break; - } - op->opc = not_op; - reset_temp(op->args[0]); - op->args[1] = op->args[i]; - continue; - } default: break; } -- 2.25.1 Even though there is only one user, place this more complex conversion into its own helper. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 89 ++++++++++++++++++++++++++------------------------ 1 file changed, 47 insertions(+), 42 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_nand(OptContext *ctx, TCGOp *op) static bool fold_neg(OptContext *ctx, TCGOp *op) { - return fold_const1(ctx, op); + if (fold_const1(ctx, op)) { + return true; + } + /* + * Because of fold_sub_to_neg, we want to always return true, + * via finish_folding. + */ + finish_folding(ctx, op); + return true; } static bool fold_nor(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_shift(OptContext *ctx, TCGOp *op) return fold_const2(ctx, op); } +static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op) +{ + TCGOpcode neg_op; + bool have_neg; + + if (!arg_is_const(op->args[1]) || arg_info(op->args[1])->val != 0) { + return false; + } + + switch (ctx->type) { + case TCG_TYPE_I32: + neg_op = INDEX_op_neg_i32; + have_neg = TCG_TARGET_HAS_neg_i32; + break; + case TCG_TYPE_I64: + neg_op = INDEX_op_neg_i64; + have_neg = TCG_TARGET_HAS_neg_i64; + break; + case TCG_TYPE_V64: + case TCG_TYPE_V128: + case TCG_TYPE_V256: + neg_op = INDEX_op_neg_vec; + have_neg = (TCG_TARGET_HAS_neg_vec && + tcg_can_emit_vec_op(neg_op, ctx->type, TCGOP_VECE(op)) > 0); + break; + default: + g_assert_not_reached(); + } + if (have_neg) { + op->opc = neg_op; + op->args[1] = op->args[2]; + return fold_neg(ctx, op); + } + return false; +} + static bool fold_sub(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || - fold_xx_to_i(ctx, op, 0)) { + fold_xx_to_i(ctx, op, 0) || + fold_sub_to_neg(ctx, op)) { return true; } return false; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) continue; } break; - CASE_OP_32_64_VEC(sub): - { - TCGOpcode neg_op; - bool have_neg; - - if (arg_is_const(op->args[2])) { - /* Proceed with possible constant folding. */ - break; - } - switch (ctx.type) { - case TCG_TYPE_I32: - neg_op = INDEX_op_neg_i32; - have_neg = TCG_TARGET_HAS_neg_i32; - break; - case TCG_TYPE_I64: - neg_op = INDEX_op_neg_i64; - have_neg = TCG_TARGET_HAS_neg_i64; - break; - case TCG_TYPE_V64: - case TCG_TYPE_V128: - case TCG_TYPE_V256: - neg_op = INDEX_op_neg_vec; - have_neg = tcg_can_emit_vec_op(neg_op, ctx.type, - TCGOP_VECE(op)) > 0; - break; - default: - g_assert_not_reached(); - } - if (!have_neg) { - break; - } - if (arg_is_const(op->args[1]) - && arg_info(op->args[1])->val == 0) { - op->opc = neg_op; - reset_temp(op->args[0]); - op->args[1] = op->args[2]; - continue; - } - } - break; default: break; } -- 2.25.1 Pull the "op r, a, i => mov r, a" optimization into a function, and use them in the outer-most logical operations. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 61 +++++++++++++++++++++----------------------------- 1 file changed, 26 insertions(+), 35 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_xi_to_i(OptContext *ctx, TCGOp *op, uint64_t i) return false; } +/* If the binary operation has second argument @i, fold to identity. */ +static bool fold_xi_to_x(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (arg_is_const(op->args[2]) && arg_info(op->args[2])->val == i) { + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); + } + return false; +} + /* If the binary operation has second argument @i, fold to NOT. */ static bool fold_xi_to_not(OptContext *ctx, TCGOp *op, uint64_t i) { @@ -XXX,XX +XXX,XX @@ static bool fold_xx_to_x(OptContext *ctx, TCGOp *op) static bool fold_add(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, 0)) { + return true; + } + return false; } static bool fold_addsub2_i32(OptContext *ctx, TCGOp *op, bool add) @@ -XXX,XX +XXX,XX @@ static bool fold_and(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || fold_xi_to_i(ctx, op, 0) || + fold_xi_to_x(ctx, op, -1) || fold_xx_to_x(ctx, op)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || fold_xx_to_i(ctx, op, 0) || + fold_xi_to_x(ctx, op, 0) || fold_ix_to_not(ctx, op, -1)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_dup2(OptContext *ctx, TCGOp *op) static bool fold_eqv(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, -1) || fold_xi_to_not(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_not(OptContext *ctx, TCGOp *op) static bool fold_or(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, 0) || fold_xx_to_x(ctx, op)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_or(OptContext *ctx, TCGOp *op) static bool fold_orc(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, -1) || fold_ix_to_not(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_sextract(OptContext *ctx, TCGOp *op) static bool fold_shift(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, 0)) { + return true; + } + return false; } static bool fold_sub_to_neg(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_sub(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || fold_xx_to_i(ctx, op, 0) || + fold_xi_to_x(ctx, op, 0) || fold_sub_to_neg(ctx, op)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_xor(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || fold_xx_to_i(ctx, op, 0) || + fold_xi_to_x(ctx, op, 0) || fold_xi_to_not(ctx, op, -1)) { return true; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Simplify expression for "op r, a, const => mov r, a" cases */ - switch (opc) { - CASE_OP_32_64_VEC(add): - CASE_OP_32_64_VEC(sub): - CASE_OP_32_64_VEC(or): - CASE_OP_32_64_VEC(xor): - CASE_OP_32_64_VEC(andc): - CASE_OP_32_64(shl): - CASE_OP_32_64(shr): - CASE_OP_32_64(sar): - CASE_OP_32_64(rotl): - CASE_OP_32_64(rotr): - if (!arg_is_const(op->args[1]) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == 0) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - continue; - } - break; - CASE_OP_32_64_VEC(and): - CASE_OP_32_64_VEC(orc): - CASE_OP_32_64(eqv): - if (!arg_is_const(op->args[1]) - && arg_is_const(op->args[2]) - && arg_info(op->args[2])->val == -1) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - continue; - } - break; - default: - break; - } - /* Simplify using known-zero bits. Currently only ops with a single output argument is supported. */ z_mask = -1; -- 2.25.1 Pull the "op r, 0, b => movi r, 0" optimization into a function, and use it in fold_shift. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 28 ++++++++++------------------ 1 file changed, 10 insertions(+), 18 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_to_not(OptContext *ctx, TCGOp *op, int idx) return false; } +/* If the binary operation has first argument @i, fold to @i. */ +static bool fold_ix_to_i(OptContext *ctx, TCGOp *op, uint64_t i) +{ + if (arg_is_const(op->args[1]) && arg_info(op->args[1])->val == i) { + return tcg_opt_gen_movi(ctx, op, op->args[0], i); + } + return false; +} + /* If the binary operation has first argument @i, fold to NOT. */ static bool fold_ix_to_not(OptContext *ctx, TCGOp *op, uint64_t i) { @@ -XXX,XX +XXX,XX @@ static bool fold_sextract(OptContext *ctx, TCGOp *op) static bool fold_shift(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_ix_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Simplify expressions for "shift/rot r, 0, a => movi r, 0", - and "sub r, 0, a => neg r, a" case. */ - switch (opc) { - CASE_OP_32_64(shl): - CASE_OP_32_64(shr): - CASE_OP_32_64(sar): - CASE_OP_32_64(rotl): - CASE_OP_32_64(rotr): - if (arg_is_const(op->args[1]) - && arg_info(op->args[1])->val == 0) { - tcg_opt_gen_movi(&ctx, op, op->args[0], 0); - continue; - } - break; - default: - break; - } - /* Simplify using known-zero bits. Currently only ops with a single output argument is supported. */ z_mask = -1; -- 2.25.1 Move all of the known-zero optimizations into the per-opcode functions. Use fold_masks when there is a possibility of the result being determined, and simply set ctx->z_mask otherwise. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 545 ++++++++++++++++++++++++++----------------------- 1 file changed, 294 insertions(+), 251 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct OptContext { TCGTempSet temps_used; /* In flight values from optimization. */ - uint64_t z_mask; + uint64_t a_mask; /* mask bit is 0 iff value identical to first input */ + uint64_t z_mask; /* mask bit is 0 iff value bit is 0 */ TCGType type; } OptContext; @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) return false; } +static bool fold_masks(OptContext *ctx, TCGOp *op) +{ + uint64_t a_mask = ctx->a_mask; + uint64_t z_mask = ctx->z_mask; + + /* + * 32-bit ops generate 32-bit results. For the result is zero test + * below, we can ignore high bits, but for further optimizations we + * need to record that the high bits contain garbage. + */ + if (ctx->type == TCG_TYPE_I32) { + ctx->z_mask |= MAKE_64BIT_MASK(32, 32); + a_mask &= MAKE_64BIT_MASK(0, 32); + z_mask &= MAKE_64BIT_MASK(0, 32); + } + + if (z_mask == 0) { + return tcg_opt_gen_movi(ctx, op, op->args[0], 0); + } + if (a_mask == 0) { + return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[1]); + } + return false; +} + /* * Convert @op to NOT, if NOT is supported by the host. * Return true f the conversion is successful, which will still @@ -XXX,XX +XXX,XX @@ static bool fold_add2_i32(OptContext *ctx, TCGOp *op) static bool fold_and(OptContext *ctx, TCGOp *op) { + uint64_t z1, z2; + if (fold_const2(ctx, op) || fold_xi_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, -1) || fold_xx_to_x(ctx, op)) { return true; } - return false; + + z1 = arg_info(op->args[1])->z_mask; + z2 = arg_info(op->args[2])->z_mask; + ctx->z_mask = z1 & z2; + + /* + * Known-zeros does not imply known-ones. Therefore unless + * arg2 is constant, we can't infer affected bits from it. + */ + if (arg_is_const(op->args[2])) { + ctx->a_mask = z1 & ~z2; + } + + return fold_masks(ctx, op); } static bool fold_andc(OptContext *ctx, TCGOp *op) { + uint64_t z1; + if (fold_const2(ctx, op) || fold_xx_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, 0) || fold_ix_to_not(ctx, op, -1)) { return true; } - return false; + + z1 = arg_info(op->args[1])->z_mask; + + /* + * Known-zeros does not imply known-ones. Therefore unless + * arg2 is constant, we can't infer anything from it. + */ + if (arg_is_const(op->args[2])) { + uint64_t z2 = ~arg_info(op->args[2])->z_mask; + ctx->a_mask = z1 & ~z2; + z1 &= z2; + } + ctx->z_mask = z1; + + return fold_masks(ctx, op); } static bool fold_brcond(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op) static bool fold_bswap(OptContext *ctx, TCGOp *op) { + uint64_t z_mask, sign; + if (arg_is_const(op->args[1])) { uint64_t t = arg_info(op->args[1])->val; t = do_constant_folding(op->opc, ctx->type, t, op->args[2]); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } - return false; + + z_mask = arg_info(op->args[1])->z_mask; + switch (op->opc) { + case INDEX_op_bswap16_i32: + case INDEX_op_bswap16_i64: + z_mask = bswap16(z_mask); + sign = INT16_MIN; + break; + case INDEX_op_bswap32_i32: + case INDEX_op_bswap32_i64: + z_mask = bswap32(z_mask); + sign = INT32_MIN; + break; + case INDEX_op_bswap64_i64: + z_mask = bswap64(z_mask); + sign = INT64_MIN; + break; + default: + g_assert_not_reached(); + } + + switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { + case TCG_BSWAP_OZ: + break; + case TCG_BSWAP_OS: + /* If the sign bit may be 1, force all the bits above to 1. */ + if (z_mask & sign) { + z_mask |= sign; + } + break; + default: + /* The high bits are undefined: force all bits above the sign to 1. */ + z_mask |= sign << 1; + break; + } + ctx->z_mask = z_mask; + + return fold_masks(ctx, op); } static bool fold_call(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_call(OptContext *ctx, TCGOp *op) static bool fold_count_zeros(OptContext *ctx, TCGOp *op) { + uint64_t z_mask; + if (arg_is_const(op->args[1])) { uint64_t t = arg_info(op->args[1])->val; @@ -XXX,XX +XXX,XX @@ static bool fold_count_zeros(OptContext *ctx, TCGOp *op) } return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[2]); } + + switch (ctx->type) { + case TCG_TYPE_I32: + z_mask = 31; + break; + case TCG_TYPE_I64: + z_mask = 63; + break; + default: + g_assert_not_reached(); + } + ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask; + return false; } static bool fold_ctpop(OptContext *ctx, TCGOp *op) { - return fold_const1(ctx, op); + if (fold_const1(ctx, op)) { + return true; + } + + switch (ctx->type) { + case TCG_TYPE_I32: + ctx->z_mask = 32 | 31; + break; + case TCG_TYPE_I64: + ctx->z_mask = 64 | 63; + break; + default: + g_assert_not_reached(); + } + return false; } static bool fold_deposit(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_deposit(OptContext *ctx, TCGOp *op) t1 = deposit64(t1, op->args[3], op->args[4], t2); return tcg_opt_gen_movi(ctx, op, op->args[0], t1); } + + ctx->z_mask = deposit64(arg_info(op->args[1])->z_mask, + op->args[3], op->args[4], + arg_info(op->args[2])->z_mask); return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_eqv(OptContext *ctx, TCGOp *op) static bool fold_extract(OptContext *ctx, TCGOp *op) { + uint64_t z_mask_old, z_mask; + if (arg_is_const(op->args[1])) { uint64_t t; @@ -XXX,XX +XXX,XX @@ static bool fold_extract(OptContext *ctx, TCGOp *op) t = extract64(t, op->args[2], op->args[3]); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } - return false; + + z_mask_old = arg_info(op->args[1])->z_mask; + z_mask = extract64(z_mask_old, op->args[2], op->args[3]); + if (op->args[2] == 0) { + ctx->a_mask = z_mask_old ^ z_mask; + } + ctx->z_mask = z_mask; + + return fold_masks(ctx, op); } static bool fold_extract2(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_extract2(OptContext *ctx, TCGOp *op) static bool fold_exts(OptContext *ctx, TCGOp *op) { - return fold_const1(ctx, op); + uint64_t z_mask_old, z_mask, sign; + bool type_change = false; + + if (fold_const1(ctx, op)) { + return true; + } + + z_mask_old = z_mask = arg_info(op->args[1])->z_mask; + + switch (op->opc) { + CASE_OP_32_64(ext8s): + sign = INT8_MIN; + z_mask = (uint8_t)z_mask; + break; + CASE_OP_32_64(ext16s): + sign = INT16_MIN; + z_mask = (uint16_t)z_mask; + break; + case INDEX_op_ext_i32_i64: + type_change = true; + QEMU_FALLTHROUGH; + case INDEX_op_ext32s_i64: + sign = INT32_MIN; + z_mask = (uint32_t)z_mask; + break; + default: + g_assert_not_reached(); + } + + if (z_mask & sign) { + z_mask |= sign; + } else if (!type_change) { + ctx->a_mask = z_mask_old ^ z_mask; + } + ctx->z_mask = z_mask; + + return fold_masks(ctx, op); } static bool fold_extu(OptContext *ctx, TCGOp *op) { - return fold_const1(ctx, op); + uint64_t z_mask_old, z_mask; + bool type_change = false; + + if (fold_const1(ctx, op)) { + return true; + } + + z_mask_old = z_mask = arg_info(op->args[1])->z_mask; + + switch (op->opc) { + CASE_OP_32_64(ext8u): + z_mask = (uint8_t)z_mask; + break; + CASE_OP_32_64(ext16u): + z_mask = (uint16_t)z_mask; + break; + case INDEX_op_extrl_i64_i32: + case INDEX_op_extu_i32_i64: + type_change = true; + QEMU_FALLTHROUGH; + case INDEX_op_ext32u_i64: + z_mask = (uint32_t)z_mask; + break; + case INDEX_op_extrh_i64_i32: + type_change = true; + z_mask >>= 32; + break; + default: + g_assert_not_reached(); + } + + ctx->z_mask = z_mask; + if (!type_change) { + ctx->a_mask = z_mask_old ^ z_mask; + } + return fold_masks(ctx, op); } static bool fold_mb(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_movcond(OptContext *ctx, TCGOp *op) return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); } + ctx->z_mask = arg_info(op->args[3])->z_mask + | arg_info(op->args[4])->z_mask; + if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { uint64_t tv = arg_info(op->args[3])->val; uint64_t fv = arg_info(op->args[4])->val; @@ -XXX,XX +XXX,XX @@ static bool fold_nand(OptContext *ctx, TCGOp *op) static bool fold_neg(OptContext *ctx, TCGOp *op) { + uint64_t z_mask; + if (fold_const1(ctx, op)) { return true; } + + /* Set to 1 all bits to the left of the rightmost. */ + z_mask = arg_info(op->args[1])->z_mask; + ctx->z_mask = -(z_mask & -z_mask); + /* * Because of fold_sub_to_neg, we want to always return true, * via finish_folding. @@ -XXX,XX +XXX,XX @@ static bool fold_or(OptContext *ctx, TCGOp *op) fold_xx_to_x(ctx, op)) { return true; } - return false; + + ctx->z_mask = arg_info(op->args[1])->z_mask + | arg_info(op->args[2])->z_mask; + return fold_masks(ctx, op); } static bool fold_orc(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_orc(OptContext *ctx, TCGOp *op) static bool fold_qemu_ld(OptContext *ctx, TCGOp *op) { + const TCGOpDef *def = &tcg_op_defs[op->opc]; + MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs]; + MemOp mop = get_memop(oi); + int width = 8 * memop_size(mop); + + if (!(mop & MO_SIGN) && width < 64) { + ctx->z_mask = MAKE_64BIT_MASK(0, width); + } + /* Opcodes that touch guest memory stop the mb optimization. */ ctx->prev_mb = NULL; return false; @@ -XXX,XX +XXX,XX @@ static bool fold_setcond(OptContext *ctx, TCGOp *op) if (i >= 0) { return tcg_opt_gen_movi(ctx, op, op->args[0], i); } + + ctx->z_mask = 1; return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) op->opc = INDEX_op_setcond_i32; break; } + + ctx->z_mask = 1; return false; do_setcond_const: @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) static bool fold_sextract(OptContext *ctx, TCGOp *op) { + int64_t z_mask_old, z_mask; + if (arg_is_const(op->args[1])) { uint64_t t; @@ -XXX,XX +XXX,XX @@ static bool fold_sextract(OptContext *ctx, TCGOp *op) t = sextract64(t, op->args[2], op->args[3]); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } - return false; + + z_mask_old = arg_info(op->args[1])->z_mask; + z_mask = sextract64(z_mask_old, op->args[2], op->args[3]); + if (op->args[2] == 0 && z_mask >= 0) { + ctx->a_mask = z_mask_old ^ z_mask; + } + ctx->z_mask = z_mask; + + return fold_masks(ctx, op); } static bool fold_shift(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_shift(OptContext *ctx, TCGOp *op) fold_xi_to_x(ctx, op, 0)) { return true; } + + if (arg_is_const(op->args[2])) { + ctx->z_mask = do_constant_folding(op->opc, ctx->type, + arg_info(op->args[1])->z_mask, + arg_info(op->args[2])->val); + return fold_masks(ctx, op); + } return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) return fold_addsub2_i32(ctx, op, false); } +static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) +{ + /* We can't do any folding with a load, but we can record bits. */ + switch (op->opc) { + CASE_OP_32_64(ld8u): + ctx->z_mask = MAKE_64BIT_MASK(0, 8); + break; + CASE_OP_32_64(ld16u): + ctx->z_mask = MAKE_64BIT_MASK(0, 16); + break; + case INDEX_op_ld32u_i64: + ctx->z_mask = MAKE_64BIT_MASK(0, 32); + break; + default: + g_assert_not_reached(); + } + return false; +} + static bool fold_xor(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || @@ -XXX,XX +XXX,XX @@ static bool fold_xor(OptContext *ctx, TCGOp *op) fold_xi_to_not(ctx, op, -1)) { return true; } - return false; + + ctx->z_mask = arg_info(op->args[1])->z_mask + | arg_info(op->args[2])->z_mask; + return fold_masks(ctx, op); } /* Propagate constants and copies, fold constant expressions. */ @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) } QTAILQ_FOREACH_SAFE(op, &s->ops, link, op_next) { - uint64_t z_mask, partmask, affected, tmp; TCGOpcode opc = op->opc; const TCGOpDef *def; bool done = false; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) break; } - /* Simplify using known-zero bits. Currently only ops with a single - output argument is supported. */ - z_mask = -1; - affected = -1; - switch (opc) { - CASE_OP_32_64(ext8s): - if ((arg_info(op->args[1])->z_mask & 0x80) != 0) { - break; - } - QEMU_FALLTHROUGH; - CASE_OP_32_64(ext8u): - z_mask = 0xff; - goto and_const; - CASE_OP_32_64(ext16s): - if ((arg_info(op->args[1])->z_mask & 0x8000) != 0) { - break; - } - QEMU_FALLTHROUGH; - CASE_OP_32_64(ext16u): - z_mask = 0xffff; - goto and_const; - case INDEX_op_ext32s_i64: - if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) { - break; - } - QEMU_FALLTHROUGH; - case INDEX_op_ext32u_i64: - z_mask = 0xffffffffU; - goto and_const; - - CASE_OP_32_64(and): - z_mask = arg_info(op->args[2])->z_mask; - if (arg_is_const(op->args[2])) { - and_const: - affected = arg_info(op->args[1])->z_mask & ~z_mask; - } - z_mask = arg_info(op->args[1])->z_mask & z_mask; - break; - - case INDEX_op_ext_i32_i64: - if ((arg_info(op->args[1])->z_mask & 0x80000000) != 0) { - break; - } - QEMU_FALLTHROUGH; - case INDEX_op_extu_i32_i64: - /* We do not compute affected as it is a size changing op. */ - z_mask = (uint32_t)arg_info(op->args[1])->z_mask; - break; - - CASE_OP_32_64(andc): - /* Known-zeros does not imply known-ones. Therefore unless - op->args[2] is constant, we can't infer anything from it. */ - if (arg_is_const(op->args[2])) { - z_mask = ~arg_info(op->args[2])->z_mask; - goto and_const; - } - /* But we certainly know nothing outside args[1] may be set. */ - z_mask = arg_info(op->args[1])->z_mask; - break; - - case INDEX_op_sar_i32: - if (arg_is_const(op->args[2])) { - tmp = arg_info(op->args[2])->val & 31; - z_mask = (int32_t)arg_info(op->args[1])->z_mask >> tmp; - } - break; - case INDEX_op_sar_i64: - if (arg_is_const(op->args[2])) { - tmp = arg_info(op->args[2])->val & 63; - z_mask = (int64_t)arg_info(op->args[1])->z_mask >> tmp; - } - break; - - case INDEX_op_shr_i32: - if (arg_is_const(op->args[2])) { - tmp = arg_info(op->args[2])->val & 31; - z_mask = (uint32_t)arg_info(op->args[1])->z_mask >> tmp; - } - break; - case INDEX_op_shr_i64: - if (arg_is_const(op->args[2])) { - tmp = arg_info(op->args[2])->val & 63; - z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> tmp; - } - break; - - case INDEX_op_extrl_i64_i32: - z_mask = (uint32_t)arg_info(op->args[1])->z_mask; - break; - case INDEX_op_extrh_i64_i32: - z_mask = (uint64_t)arg_info(op->args[1])->z_mask >> 32; - break; - - CASE_OP_32_64(shl): - if (arg_is_const(op->args[2])) { - tmp = arg_info(op->args[2])->val & (TCG_TARGET_REG_BITS - 1); - z_mask = arg_info(op->args[1])->z_mask << tmp; - } - break; - - CASE_OP_32_64(neg): - /* Set to 1 all bits to the left of the rightmost. */ - z_mask = -(arg_info(op->args[1])->z_mask - & -arg_info(op->args[1])->z_mask); - break; - - CASE_OP_32_64(deposit): - z_mask = deposit64(arg_info(op->args[1])->z_mask, - op->args[3], op->args[4], - arg_info(op->args[2])->z_mask); - break; - - CASE_OP_32_64(extract): - z_mask = extract64(arg_info(op->args[1])->z_mask, - op->args[2], op->args[3]); - if (op->args[2] == 0) { - affected = arg_info(op->args[1])->z_mask & ~z_mask; - } - break; - CASE_OP_32_64(sextract): - z_mask = sextract64(arg_info(op->args[1])->z_mask, - op->args[2], op->args[3]); - if (op->args[2] == 0 && (tcg_target_long)z_mask >= 0) { - affected = arg_info(op->args[1])->z_mask & ~z_mask; - } - break; - - CASE_OP_32_64(or): - CASE_OP_32_64(xor): - z_mask = arg_info(op->args[1])->z_mask - | arg_info(op->args[2])->z_mask; - break; - - case INDEX_op_clz_i32: - case INDEX_op_ctz_i32: - z_mask = arg_info(op->args[2])->z_mask | 31; - break; - - case INDEX_op_clz_i64: - case INDEX_op_ctz_i64: - z_mask = arg_info(op->args[2])->z_mask | 63; - break; - - case INDEX_op_ctpop_i32: - z_mask = 32 | 31; - break; - case INDEX_op_ctpop_i64: - z_mask = 64 | 63; - break; - - CASE_OP_32_64(setcond): - case INDEX_op_setcond2_i32: - z_mask = 1; - break; - - CASE_OP_32_64(movcond): - z_mask = arg_info(op->args[3])->z_mask - | arg_info(op->args[4])->z_mask; - break; - - CASE_OP_32_64(ld8u): - z_mask = 0xff; - break; - CASE_OP_32_64(ld16u): - z_mask = 0xffff; - break; - case INDEX_op_ld32u_i64: - z_mask = 0xffffffffu; - break; - - CASE_OP_32_64(qemu_ld): - { - MemOpIdx oi = op->args[def->nb_oargs + def->nb_iargs]; - MemOp mop = get_memop(oi); - if (!(mop & MO_SIGN)) { - z_mask = (2ULL << ((8 << (mop & MO_SIZE)) - 1)) - 1; - } - } - break; - - CASE_OP_32_64(bswap16): - z_mask = arg_info(op->args[1])->z_mask; - if (z_mask <= 0xffff) { - op->args[2] |= TCG_BSWAP_IZ; - } - z_mask = bswap16(z_mask); - switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { - case TCG_BSWAP_OZ: - break; - case TCG_BSWAP_OS: - z_mask = (int16_t)z_mask; - break; - default: /* undefined high bits */ - z_mask |= MAKE_64BIT_MASK(16, 48); - break; - } - break; - - case INDEX_op_bswap32_i64: - z_mask = arg_info(op->args[1])->z_mask; - if (z_mask <= 0xffffffffu) { - op->args[2] |= TCG_BSWAP_IZ; - } - z_mask = bswap32(z_mask); - switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { - case TCG_BSWAP_OZ: - break; - case TCG_BSWAP_OS: - z_mask = (int32_t)z_mask; - break; - default: /* undefined high bits */ - z_mask |= MAKE_64BIT_MASK(32, 32); - break; - } - break; - - default: - break; - } - - /* 32-bit ops generate 32-bit results. For the result is zero test - below, we can ignore high bits, but for further optimizations we - need to record that the high bits contain garbage. */ - partmask = z_mask; - if (ctx.type == TCG_TYPE_I32) { - z_mask |= ~(tcg_target_ulong)0xffffffffu; - partmask &= 0xffffffffu; - affected &= 0xffffffffu; - } - ctx.z_mask = z_mask; - - if (partmask == 0) { - tcg_opt_gen_movi(&ctx, op, op->args[0], 0); - continue; - } - if (affected == 0) { - tcg_opt_gen_mov(&ctx, op, op->args[0], op->args[1]); - continue; - } + /* Assume all bits affected, and no bits known zero. */ + ctx.a_mask = -1; + ctx.z_mask = -1; /* * Process each opcode. @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_extrh_i64_i32: done = fold_extu(&ctx, op); break; + CASE_OP_32_64(ld8u): + CASE_OP_32_64(ld16u): + case INDEX_op_ld32u_i64: + done = fold_tcg_ld(&ctx, op); + break; case INDEX_op_mb: done = fold_mb(&ctx, op); break; -- 2.25.1 Rename to fold_multiply2, and handle muls2_i32, mulu2_i64, and muls2_i64. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 44 +++++++++++++++++++++++++++++++++++--------- 1 file changed, 35 insertions(+), 9 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) return false; } -static bool fold_mulu2_i32(OptContext *ctx, TCGOp *op) +static bool fold_multiply2(OptContext *ctx, TCGOp *op) { if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { - uint32_t a = arg_info(op->args[2])->val; - uint32_t b = arg_info(op->args[3])->val; - uint64_t r = (uint64_t)a * b; + uint64_t a = arg_info(op->args[2])->val; + uint64_t b = arg_info(op->args[3])->val; + uint64_t h, l; TCGArg rl, rh; - TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_mov_i32); + TCGOp *op2; + + switch (op->opc) { + case INDEX_op_mulu2_i32: + l = (uint64_t)(uint32_t)a * (uint32_t)b; + h = (int32_t)(l >> 32); + l = (int32_t)l; + break; + case INDEX_op_muls2_i32: + l = (int64_t)(int32_t)a * (int32_t)b; + h = l >> 32; + l = (int32_t)l; + break; + case INDEX_op_mulu2_i64: + mulu64(&l, &h, a, b); + break; + case INDEX_op_muls2_i64: + muls64(&l, &h, a, b); + break; + default: + g_assert_not_reached(); + } rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(ctx, op, rl, (int32_t)r); - tcg_opt_gen_movi(ctx, op2, rh, (int32_t)(r >> 32)); + + /* The proper opcode is supplied by tcg_opt_gen_mov. */ + op2 = tcg_op_insert_before(ctx->tcg, op, 0); + + tcg_opt_gen_movi(ctx, op, rl, l); + tcg_opt_gen_movi(ctx, op2, rh, h); return true; } return false; @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64(muluh): done = fold_mul_highpart(&ctx, op); break; - case INDEX_op_mulu2_i32: - done = fold_mulu2_i32(&ctx, op); + CASE_OP_32_64(muls2): + CASE_OP_32_64(mulu2): + done = fold_multiply2(&ctx, op); break; CASE_OP_32_64(nand): done = fold_nand(&ctx, op); -- 2.25.1 Rename to fold_addsub2. Use Int128 to implement the wider operation. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 65 ++++++++++++++++++++++++++++++++++---------------- 1 file changed, 44 insertions(+), 21 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ */ #include "qemu/osdep.h" +#include "qemu/int128.h" #include "tcg/tcg-op.h" #include "tcg-internal.h" @@ -XXX,XX +XXX,XX @@ static bool fold_add(OptContext *ctx, TCGOp *op) return false; } -static bool fold_addsub2_i32(OptContext *ctx, TCGOp *op, bool add) +static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add) { if (arg_is_const(op->args[2]) && arg_is_const(op->args[3]) && arg_is_const(op->args[4]) && arg_is_const(op->args[5])) { - uint32_t al = arg_info(op->args[2])->val; - uint32_t ah = arg_info(op->args[3])->val; - uint32_t bl = arg_info(op->args[4])->val; - uint32_t bh = arg_info(op->args[5])->val; - uint64_t a = ((uint64_t)ah << 32) | al; - uint64_t b = ((uint64_t)bh << 32) | bl; + uint64_t al = arg_info(op->args[2])->val; + uint64_t ah = arg_info(op->args[3])->val; + uint64_t bl = arg_info(op->args[4])->val; + uint64_t bh = arg_info(op->args[5])->val; TCGArg rl, rh; - TCGOp *op2 = tcg_op_insert_before(ctx->tcg, op, INDEX_op_mov_i32); + TCGOp *op2; - if (add) { - a += b; + if (ctx->type == TCG_TYPE_I32) { + uint64_t a = deposit64(al, 32, 32, ah); + uint64_t b = deposit64(bl, 32, 32, bh); + + if (add) { + a += b; + } else { + a -= b; + } + + al = sextract64(a, 0, 32); + ah = sextract64(a, 32, 32); } else { - a -= b; + Int128 a = int128_make128(al, ah); + Int128 b = int128_make128(bl, bh); + + if (add) { + a = int128_add(a, b); + } else { + a = int128_sub(a, b); + } + + al = int128_getlo(a); + ah = int128_gethi(a); } rl = op->args[0]; rh = op->args[1]; - tcg_opt_gen_movi(ctx, op, rl, (int32_t)a); - tcg_opt_gen_movi(ctx, op2, rh, (int32_t)(a >> 32)); + + /* The proper opcode is supplied by tcg_opt_gen_mov. */ + op2 = tcg_op_insert_before(ctx->tcg, op, 0); + + tcg_opt_gen_movi(ctx, op, rl, al); + tcg_opt_gen_movi(ctx, op2, rh, ah); return true; } return false; } -static bool fold_add2_i32(OptContext *ctx, TCGOp *op) +static bool fold_add2(OptContext *ctx, TCGOp *op) { - return fold_addsub2_i32(ctx, op, true); + return fold_addsub2(ctx, op, true); } static bool fold_and(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ static bool fold_sub(OptContext *ctx, TCGOp *op) return false; } -static bool fold_sub2_i32(OptContext *ctx, TCGOp *op) +static bool fold_sub2(OptContext *ctx, TCGOp *op) { - return fold_addsub2_i32(ctx, op, false); + return fold_addsub2(ctx, op, false); } static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(add): done = fold_add(&ctx, op); break; - case INDEX_op_add2_i32: - done = fold_add2_i32(&ctx, op); + CASE_OP_32_64(add2): + done = fold_add2(&ctx, op); break; CASE_OP_32_64_VEC(and): done = fold_and(&ctx, op); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) CASE_OP_32_64_VEC(sub): done = fold_sub(&ctx, op); break; - case INDEX_op_sub2_i32: - done = fold_sub2_i32(&ctx, op); + CASE_OP_32_64(sub2): + done = fold_sub2(&ctx, op); break; CASE_OP_32_64_VEC(xor): done = fold_xor(&ctx, op); -- 2.25.1 Most of these are handled by creating a fold_const2_commutative to handle all of the binary operators. The rest were already handled on a case-by-case basis in the switch, and have their own fold function in which to place the call. We now have only one major switch on TCGOpcode. Introduce NO_DEST and a block comment for swap_commutative in order to make the handling of brcond and movcond opcodes cleaner. Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 142 ++++++++++++++++++++++++------------------------- 1 file changed, 70 insertions(+), 72 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static int do_constant_folding_cond2(TCGArg *p1, TCGArg *p2, TCGCond c) return -1; } +/** + * swap_commutative: + * @dest: TCGArg of the destination argument, or NO_DEST. + * @p1: first paired argument + * @p2: second paired argument + * + * If *@p1 is a constant and *@p2 is not, swap. + * If *@p2 matches @dest, swap. + * Return true if a swap was performed. + */ + +#define NO_DEST temp_arg(NULL) + static bool swap_commutative(TCGArg dest, TCGArg *p1, TCGArg *p2) { TCGArg a1 = *p1, a2 = *p2; @@ -XXX,XX +XXX,XX @@ static bool fold_const2(OptContext *ctx, TCGOp *op) return false; } +static bool fold_const2_commutative(OptContext *ctx, TCGOp *op) +{ + swap_commutative(op->args[0], &op->args[1], &op->args[2]); + return fold_const2(ctx, op); +} + static bool fold_masks(OptContext *ctx, TCGOp *op) { uint64_t a_mask = ctx->a_mask; @@ -XXX,XX +XXX,XX @@ static bool fold_xx_to_x(OptContext *ctx, TCGOp *op) static bool fold_add(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_x(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_addsub2(OptContext *ctx, TCGOp *op, bool add) static bool fold_add2(OptContext *ctx, TCGOp *op) { + /* Note that the high and low parts may be independently swapped. */ + swap_commutative(op->args[0], &op->args[2], &op->args[4]); + swap_commutative(op->args[1], &op->args[3], &op->args[5]); + return fold_addsub2(ctx, op, true); } @@ -XXX,XX +XXX,XX @@ static bool fold_and(OptContext *ctx, TCGOp *op) { uint64_t z1, z2; - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, -1) || fold_xx_to_x(ctx, op)) { @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) static bool fold_brcond(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[2]; - int i = do_constant_folding_cond(ctx->type, op->args[0], op->args[1], cond); + int i; + if (swap_commutative(NO_DEST, &op->args[0], &op->args[1])) { + op->args[2] = cond = tcg_swap_cond(cond); + } + + i = do_constant_folding_cond(ctx->type, op->args[0], op->args[1], cond); if (i == 0) { tcg_op_remove(ctx->tcg, op); return true; @@ -XXX,XX +XXX,XX @@ static bool fold_brcond(OptContext *ctx, TCGOp *op) static bool fold_brcond2(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[4]; - int i = do_constant_folding_cond2(&op->args[0], &op->args[2], cond); TCGArg label = op->args[5]; - int inv = 0; + int i, inv = 0; + if (swap_commutative2(&op->args[0], &op->args[2])) { + op->args[4] = cond = tcg_swap_cond(cond); + } + + i = do_constant_folding_cond2(&op->args[0], &op->args[2], cond); if (i >= 0) { goto do_brcond_const; } @@ -XXX,XX +XXX,XX @@ static bool fold_dup2(OptContext *ctx, TCGOp *op) static bool fold_eqv(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_x(ctx, op, -1) || fold_xi_to_not(ctx, op, 0)) { return true; @@ -XXX,XX +XXX,XX @@ static bool fold_mov(OptContext *ctx, TCGOp *op) static bool fold_movcond(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[5]; - int i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); + int i; + if (swap_commutative(NO_DEST, &op->args[1], &op->args[2])) { + op->args[5] = cond = tcg_swap_cond(cond); + } + /* + * Canonicalize the "false" input reg to match the destination reg so + * that the tcg backend can implement a "move if true" operation. + */ + if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { + op->args[5] = cond = tcg_invert_cond(cond); + } + + i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); if (i >= 0) { return tcg_opt_gen_mov(ctx, op, op->args[0], op->args[4 - i]); } @@ -XXX,XX +XXX,XX @@ static bool fold_mul(OptContext *ctx, TCGOp *op) static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_i(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_mul_highpart(OptContext *ctx, TCGOp *op) static bool fold_multiply2(OptContext *ctx, TCGOp *op) { + swap_commutative(op->args[0], &op->args[2], &op->args[3]); + if (arg_is_const(op->args[2]) && arg_is_const(op->args[3])) { uint64_t a = arg_info(op->args[2])->val; uint64_t b = arg_info(op->args[3])->val; @@ -XXX,XX +XXX,XX @@ static bool fold_multiply2(OptContext *ctx, TCGOp *op) static bool fold_nand(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_not(ctx, op, -1)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_neg(OptContext *ctx, TCGOp *op) static bool fold_nor(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_not(ctx, op, 0)) { return true; } @@ -XXX,XX +XXX,XX @@ static bool fold_not(OptContext *ctx, TCGOp *op) static bool fold_or(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xi_to_x(ctx, op, 0) || fold_xx_to_x(ctx, op)) { return true; @@ -XXX,XX +XXX,XX @@ static bool fold_remainder(OptContext *ctx, TCGOp *op) static bool fold_setcond(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[3]; - int i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); + int i; + if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) { + op->args[3] = cond = tcg_swap_cond(cond); + } + + i = do_constant_folding_cond(ctx->type, op->args[1], op->args[2], cond); if (i >= 0) { return tcg_opt_gen_movi(ctx, op, op->args[0], i); } @@ -XXX,XX +XXX,XX @@ static bool fold_setcond(OptContext *ctx, TCGOp *op) static bool fold_setcond2(OptContext *ctx, TCGOp *op) { TCGCond cond = op->args[5]; - int i = do_constant_folding_cond2(&op->args[1], &op->args[3], cond); - int inv = 0; + int i, inv = 0; + if (swap_commutative2(&op->args[1], &op->args[3])) { + op->args[5] = cond = tcg_swap_cond(cond); + } + + i = do_constant_folding_cond2(&op->args[1], &op->args[3], cond); if (i >= 0) { goto do_setcond_const; } @@ -XXX,XX +XXX,XX @@ static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) static bool fold_xor(OptContext *ctx, TCGOp *op) { - if (fold_const2(ctx, op) || + if (fold_const2_commutative(ctx, op) || fold_xx_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, 0) || fold_xi_to_not(ctx, op, -1)) { @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) ctx.type = TCG_TYPE_I32; } - /* For commutative operations make constant second argument */ - switch (opc) { - CASE_OP_32_64_VEC(add): - CASE_OP_32_64_VEC(mul): - CASE_OP_32_64_VEC(and): - CASE_OP_32_64_VEC(or): - CASE_OP_32_64_VEC(xor): - CASE_OP_32_64(eqv): - CASE_OP_32_64(nand): - CASE_OP_32_64(nor): - CASE_OP_32_64(muluh): - CASE_OP_32_64(mulsh): - swap_commutative(op->args[0], &op->args[1], &op->args[2]); - break; - CASE_OP_32_64(brcond): - if (swap_commutative(-1, &op->args[0], &op->args[1])) { - op->args[2] = tcg_swap_cond(op->args[2]); - } - break; - CASE_OP_32_64(setcond): - if (swap_commutative(op->args[0], &op->args[1], &op->args[2])) { - op->args[3] = tcg_swap_cond(op->args[3]); - } - break; - CASE_OP_32_64(movcond): - if (swap_commutative(-1, &op->args[1], &op->args[2])) { - op->args[5] = tcg_swap_cond(op->args[5]); - } - /* For movcond, we canonicalize the "false" input reg to match - the destination reg so that the tcg backend can implement - a "move if true" operation. */ - if (swap_commutative(op->args[0], &op->args[4], &op->args[3])) { - op->args[5] = tcg_invert_cond(op->args[5]); - } - break; - CASE_OP_32_64(add2): - swap_commutative(op->args[0], &op->args[2], &op->args[4]); - swap_commutative(op->args[1], &op->args[3], &op->args[5]); - break; - CASE_OP_32_64(mulu2): - CASE_OP_32_64(muls2): - swap_commutative(op->args[0], &op->args[2], &op->args[3]); - break; - case INDEX_op_brcond2_i32: - if (swap_commutative2(&op->args[0], &op->args[2])) { - op->args[4] = tcg_swap_cond(op->args[4]); - } - break; - case INDEX_op_setcond2_i32: - if (swap_commutative2(&op->args[1], &op->args[3])) { - op->args[5] = tcg_swap_cond(op->args[5]); - } - break; - default: - break; - } - /* Assume all bits affected, and no bits known zero. */ ctx.a_mask = -1; ctx.z_mask = -1; -- 2.25.1 This "garbage" setting pre-dates the addition of the type changing opcodes INDEX_op_ext_i32_i64, INDEX_op_extu_i32_i64, and INDEX_op_extr{l,h}_i64_i32. So now we have a definitive points at which to adjust z_mask to eliminate such bits from the 32-bit operands. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 35 ++++++++++++++++------------------- 1 file changed, 16 insertions(+), 19 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static void init_ts_info(OptContext *ctx, TCGTemp *ts) ti->is_const = true; ti->val = ts->val; ti->z_mask = ts->val; - if (TCG_TARGET_REG_BITS > 32 && ts->type == TCG_TYPE_I32) { - /* High bits of a 32-bit quantity are garbage. */ - ti->z_mask |= ~0xffffffffull; - } } else { ti->is_const = false; ti->z_mask = -1; @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) TCGTemp *src_ts = arg_temp(src); TempOptInfo *di; TempOptInfo *si; - uint64_t z_mask; TCGOpcode new_op; if (ts_are_copies(dst_ts, src_ts)) { @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) op->args[0] = dst; op->args[1] = src; - z_mask = si->z_mask; - if (TCG_TARGET_REG_BITS > 32 && new_op == INDEX_op_mov_i32) { - /* High bits of the destination are now garbage. */ - z_mask |= ~0xffffffffull; - } - di->z_mask = z_mask; + di->z_mask = si->z_mask; if (src_ts->type == dst_ts->type) { TempOptInfo *ni = ts_info(si->next_copy); @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) static bool tcg_opt_gen_movi(OptContext *ctx, TCGOp *op, TCGArg dst, uint64_t val) { - /* Convert movi to mov with constant temp. */ - TCGTemp *tv = tcg_constant_internal(ctx->type, val); + TCGTemp *tv; + if (ctx->type == TCG_TYPE_I32) { + val = (int32_t)val; + } + + /* Convert movi to mov with constant temp. */ + tv = tcg_constant_internal(ctx->type, val); init_ts_info(ctx, tv); return tcg_opt_gen_mov(ctx, op, dst, temp_arg(tv)); } @@ -XXX,XX +XXX,XX @@ static bool fold_masks(OptContext *ctx, TCGOp *op) uint64_t z_mask = ctx->z_mask; /* - * 32-bit ops generate 32-bit results. For the result is zero test - * below, we can ignore high bits, but for further optimizations we - * need to record that the high bits contain garbage. + * 32-bit ops generate 32-bit results, which for the purpose of + * simplifying tcg are sign-extended. Certainly that's how we + * represent our constants elsewhere. Note that the bits will + * be reset properly for a 64-bit value when encountering the + * type changing opcodes. */ if (ctx->type == TCG_TYPE_I32) { - ctx->z_mask |= MAKE_64BIT_MASK(32, 32); - a_mask &= MAKE_64BIT_MASK(0, 32); - z_mask &= MAKE_64BIT_MASK(0, 32); + a_mask = (int32_t)a_mask; + z_mask = (int32_t)z_mask; + ctx->z_mask = z_mask; } if (z_mask == 0) { -- 2.25.1 Recognize the constant function for or-complement. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 1 + 1 file changed, 1 insertion(+) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_or(OptContext *ctx, TCGOp *op) static bool fold_orc(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || + fold_xx_to_i(ctx, op, -1) || fold_xi_to_x(ctx, op, -1) || fold_ix_to_not(ctx, op, 0)) { return true; -- 2.25.1 Recognize the identity function for low-part multiply. Suggested-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_movcond(OptContext *ctx, TCGOp *op) static bool fold_mul(OptContext *ctx, TCGOp *op) { if (fold_const2(ctx, op) || - fold_xi_to_i(ctx, op, 0)) { + fold_xi_to_i(ctx, op, 0) || + fold_xi_to_x(ctx, op, 1)) { return true; } return false; -- 2.25.1 Recognize the identity function for division. Suggested-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_deposit(OptContext *ctx, TCGOp *op) static bool fold_divide(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xi_to_x(ctx, op, 1)) { + return true; + } + return false; } static bool fold_dup(OptContext *ctx, TCGOp *op) -- 2.25.1 Recognize the constant function for remainder. Suggested-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_qemu_st(OptContext *ctx, TCGOp *op) static bool fold_remainder(OptContext *ctx, TCGOp *op) { - return fold_const2(ctx, op); + if (fold_const2(ctx, op) || + fold_xx_to_i(ctx, op, 0)) { + return true; + } + return false; } static bool fold_setcond(OptContext *ctx, TCGOp *op) -- 2.25.1 Certain targets, like riscv, produce signed 32-bit results. This can lead to lots of redundant extensions as values are manipulated. Begin by tracking only the obvious sign-extensions, and converting them to simple copies when possible. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 123 ++++++++++++++++++++++++++++++++++++++++--------- 1 file changed, 102 insertions(+), 21 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ typedef struct TempOptInfo { TCGTemp *next_copy; uint64_t val; uint64_t z_mask; /* mask bit is 0 if and only if value bit is 0 */ + uint64_t s_mask; /* a left-aligned mask of clrsb(value) bits. */ } TempOptInfo; typedef struct OptContext { @@ -XXX,XX +XXX,XX @@ typedef struct OptContext { /* In flight values from optimization. */ uint64_t a_mask; /* mask bit is 0 iff value identical to first input */ uint64_t z_mask; /* mask bit is 0 iff value bit is 0 */ + uint64_t s_mask; /* mask of clrsb(value) bits */ TCGType type; } OptContext; +/* Calculate the smask for a specific value. */ +static uint64_t smask_from_value(uint64_t value) +{ + int rep = clrsb64(value); + return ~(~0ull >> rep); +} + +/* + * Calculate the smask for a given set of known-zeros. + * If there are lots of zeros on the left, we can consider the remainder + * an unsigned field, and thus the corresponding signed field is one bit + * larger. + */ +static uint64_t smask_from_zmask(uint64_t zmask) +{ + /* + * Only the 0 bits are significant for zmask, thus the msb itself + * must be zero, else we have no sign information. + */ + int rep = clz64(zmask); + if (rep == 0) { + return 0; + } + rep -= 1; + return ~(~0ull >> rep); +} + static inline TempOptInfo *ts_info(TCGTemp *ts) { return ts->state_ptr; @@ -XXX,XX +XXX,XX @@ static void reset_ts(TCGTemp *ts) ti->prev_copy = ts; ti->is_const = false; ti->z_mask = -1; + ti->s_mask = 0; } static void reset_temp(TCGArg arg) @@ -XXX,XX +XXX,XX @@ static void init_ts_info(OptContext *ctx, TCGTemp *ts) ti->is_const = true; ti->val = ts->val; ti->z_mask = ts->val; + ti->s_mask = smask_from_value(ts->val); } else { ti->is_const = false; ti->z_mask = -1; + ti->s_mask = 0; } } @@ -XXX,XX +XXX,XX @@ static bool tcg_opt_gen_mov(OptContext *ctx, TCGOp *op, TCGArg dst, TCGArg src) op->args[1] = src; di->z_mask = si->z_mask; + di->s_mask = si->s_mask; if (src_ts->type == dst_ts->type) { TempOptInfo *ni = ts_info(si->next_copy); @@ -XXX,XX +XXX,XX @@ static void finish_folding(OptContext *ctx, TCGOp *op) nb_oargs = def->nb_oargs; for (i = 0; i < nb_oargs; i++) { - reset_temp(op->args[i]); + TCGTemp *ts = arg_temp(op->args[i]); + reset_ts(ts); /* - * Save the corresponding known-zero bits mask for the + * Save the corresponding known-zero/sign bits mask for the * first output argument (only one supported so far). */ if (i == 0) { - arg_info(op->args[i])->z_mask = ctx->z_mask; + ts_info(ts)->z_mask = ctx->z_mask; + ts_info(ts)->s_mask = ctx->s_mask; } } } @@ -XXX,XX +XXX,XX @@ static bool fold_masks(OptContext *ctx, TCGOp *op) { uint64_t a_mask = ctx->a_mask; uint64_t z_mask = ctx->z_mask; + uint64_t s_mask = ctx->s_mask; /* * 32-bit ops generate 32-bit results, which for the purpose of @@ -XXX,XX +XXX,XX @@ static bool fold_masks(OptContext *ctx, TCGOp *op) if (ctx->type == TCG_TYPE_I32) { a_mask = (int32_t)a_mask; z_mask = (int32_t)z_mask; + s_mask |= MAKE_64BIT_MASK(32, 32); ctx->z_mask = z_mask; + ctx->s_mask = s_mask; } if (z_mask == 0) { @@ -XXX,XX +XXX,XX @@ static bool fold_brcond2(OptContext *ctx, TCGOp *op) static bool fold_bswap(OptContext *ctx, TCGOp *op) { - uint64_t z_mask, sign; + uint64_t z_mask, s_mask, sign; if (arg_is_const(op->args[1])) { uint64_t t = arg_info(op->args[1])->val; @@ -XXX,XX +XXX,XX @@ static bool fold_bswap(OptContext *ctx, TCGOp *op) } z_mask = arg_info(op->args[1])->z_mask; + switch (op->opc) { case INDEX_op_bswap16_i32: case INDEX_op_bswap16_i64: @@ -XXX,XX +XXX,XX @@ static bool fold_bswap(OptContext *ctx, TCGOp *op) default: g_assert_not_reached(); } + s_mask = smask_from_zmask(z_mask); switch (op->args[2] & (TCG_BSWAP_OZ | TCG_BSWAP_OS)) { case TCG_BSWAP_OZ: @@ -XXX,XX +XXX,XX @@ static bool fold_bswap(OptContext *ctx, TCGOp *op) /* If the sign bit may be 1, force all the bits above to 1. */ if (z_mask & sign) { z_mask |= sign; + s_mask = sign << 1; } break; default: /* The high bits are undefined: force all bits above the sign to 1. */ z_mask |= sign << 1; + s_mask = 0; break; } ctx->z_mask = z_mask; + ctx->s_mask = s_mask; return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_eqv(OptContext *ctx, TCGOp *op) static bool fold_extract(OptContext *ctx, TCGOp *op) { uint64_t z_mask_old, z_mask; + int pos = op->args[2]; + int len = op->args[3]; if (arg_is_const(op->args[1])) { uint64_t t; t = arg_info(op->args[1])->val; - t = extract64(t, op->args[2], op->args[3]); + t = extract64(t, pos, len); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } z_mask_old = arg_info(op->args[1])->z_mask; - z_mask = extract64(z_mask_old, op->args[2], op->args[3]); - if (op->args[2] == 0) { + z_mask = extract64(z_mask_old, pos, len); + if (pos == 0) { ctx->a_mask = z_mask_old ^ z_mask; } ctx->z_mask = z_mask; + ctx->s_mask = smask_from_zmask(z_mask); return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_extract2(OptContext *ctx, TCGOp *op) static bool fold_exts(OptContext *ctx, TCGOp *op) { - uint64_t z_mask_old, z_mask, sign; + uint64_t s_mask_old, s_mask, z_mask, sign; bool type_change = false; if (fold_const1(ctx, op)) { return true; } - z_mask_old = z_mask = arg_info(op->args[1])->z_mask; + z_mask = arg_info(op->args[1])->z_mask; + s_mask = arg_info(op->args[1])->s_mask; + s_mask_old = s_mask; switch (op->opc) { CASE_OP_32_64(ext8s): @@ -XXX,XX +XXX,XX @@ static bool fold_exts(OptContext *ctx, TCGOp *op) if (z_mask & sign) { z_mask |= sign; - } else if (!type_change) { - ctx->a_mask = z_mask_old ^ z_mask; } + s_mask |= sign << 1; + ctx->z_mask = z_mask; + ctx->s_mask = s_mask; + if (!type_change) { + ctx->a_mask = s_mask & ~s_mask_old; + } return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_extu(OptContext *ctx, TCGOp *op) } ctx->z_mask = z_mask; + ctx->s_mask = smask_from_zmask(z_mask); if (!type_change) { ctx->a_mask = z_mask_old ^ z_mask; } @@ -XXX,XX +XXX,XX @@ static bool fold_qemu_ld(OptContext *ctx, TCGOp *op) MemOp mop = get_memop(oi); int width = 8 * memop_size(mop); - if (!(mop & MO_SIGN) && width < 64) { - ctx->z_mask = MAKE_64BIT_MASK(0, width); + if (width < 64) { + ctx->s_mask = MAKE_64BIT_MASK(width, 64 - width); + if (!(mop & MO_SIGN)) { + ctx->z_mask = MAKE_64BIT_MASK(0, width); + ctx->s_mask <<= 1; + } } /* Opcodes that touch guest memory stop the mb optimization. */ @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) static bool fold_sextract(OptContext *ctx, TCGOp *op) { - int64_t z_mask_old, z_mask; + uint64_t z_mask, s_mask, s_mask_old; + int pos = op->args[2]; + int len = op->args[3]; if (arg_is_const(op->args[1])) { uint64_t t; t = arg_info(op->args[1])->val; - t = sextract64(t, op->args[2], op->args[3]); + t = sextract64(t, pos, len); return tcg_opt_gen_movi(ctx, op, op->args[0], t); } - z_mask_old = arg_info(op->args[1])->z_mask; - z_mask = sextract64(z_mask_old, op->args[2], op->args[3]); - if (op->args[2] == 0 && z_mask >= 0) { - ctx->a_mask = z_mask_old ^ z_mask; - } + z_mask = arg_info(op->args[1])->z_mask; + z_mask = sextract64(z_mask, pos, len); ctx->z_mask = z_mask; + s_mask_old = arg_info(op->args[1])->s_mask; + s_mask = sextract64(s_mask_old, pos, len); + s_mask |= MAKE_64BIT_MASK(len, 64 - len); + ctx->s_mask = s_mask; + + if (pos == 0) { + ctx->a_mask = s_mask & ~s_mask_old; + } + return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_tcg_ld(OptContext *ctx, TCGOp *op) { /* We can't do any folding with a load, but we can record bits. */ switch (op->opc) { + CASE_OP_32_64(ld8s): + ctx->s_mask = MAKE_64BIT_MASK(8, 56); + break; CASE_OP_32_64(ld8u): ctx->z_mask = MAKE_64BIT_MASK(0, 8); + ctx->s_mask = MAKE_64BIT_MASK(9, 55); + break; + CASE_OP_32_64(ld16s): + ctx->s_mask = MAKE_64BIT_MASK(16, 48); break; CASE_OP_32_64(ld16u): ctx->z_mask = MAKE_64BIT_MASK(0, 16); + ctx->s_mask = MAKE_64BIT_MASK(17, 47); + break; + case INDEX_op_ld32s_i64: + ctx->s_mask = MAKE_64BIT_MASK(32, 32); break; case INDEX_op_ld32u_i64: ctx->z_mask = MAKE_64BIT_MASK(0, 32); + ctx->s_mask = MAKE_64BIT_MASK(33, 31); break; default: g_assert_not_reached(); @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) ctx.type = TCG_TYPE_I32; } - /* Assume all bits affected, and no bits known zero. */ + /* Assume all bits affected, no bits known zero, no sign reps. */ ctx.a_mask = -1; ctx.z_mask = -1; + ctx.s_mask = 0; /* * Process each opcode. @@ -XXX,XX +XXX,XX @@ void tcg_optimize(TCGContext *s) case INDEX_op_extrh_i64_i32: done = fold_extu(&ctx, op); break; + CASE_OP_32_64(ld8s): CASE_OP_32_64(ld8u): + CASE_OP_32_64(ld16s): CASE_OP_32_64(ld16u): + case INDEX_op_ld32s_i64: case INDEX_op_ld32u_i64: done = fold_tcg_ld(&ctx, op); break; -- 2.25.1 Sign repetitions are perforce all identical, whether they are 1 or 0. Bitwise operations preserve the relative quantity of the repetitions. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Reviewed-by: Philippe Mathieu-Daudé <f4bug@amsat.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 29 +++++++++++++++++++++++++++++ 1 file changed, 29 insertions(+) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_and(OptContext *ctx, TCGOp *op) z2 = arg_info(op->args[2])->z_mask; ctx->z_mask = z1 & z2; + /* + * Sign repetitions are perforce all identical, whether they are 1 or 0. + * Bitwise operations preserve the relative quantity of the repetitions. + */ + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; + /* * Known-zeros does not imply known-ones. Therefore unless * arg2 is constant, we can't infer affected bits from it. @@ -XXX,XX +XXX,XX @@ static bool fold_andc(OptContext *ctx, TCGOp *op) } ctx->z_mask = z1; + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_eqv(OptContext *ctx, TCGOp *op) fold_xi_to_not(ctx, op, 0)) { return true; } + + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_movcond(OptContext *ctx, TCGOp *op) ctx->z_mask = arg_info(op->args[3])->z_mask | arg_info(op->args[4])->z_mask; + ctx->s_mask = arg_info(op->args[3])->s_mask + & arg_info(op->args[4])->s_mask; if (arg_is_const(op->args[3]) && arg_is_const(op->args[4])) { uint64_t tv = arg_info(op->args[3])->val; @@ -XXX,XX +XXX,XX @@ static bool fold_nand(OptContext *ctx, TCGOp *op) fold_xi_to_not(ctx, op, -1)) { return true; } + + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_nor(OptContext *ctx, TCGOp *op) fold_xi_to_not(ctx, op, 0)) { return true; } + + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_not(OptContext *ctx, TCGOp *op) return true; } + ctx->s_mask = arg_info(op->args[1])->s_mask; + /* Because of fold_to_not, we want to always return true, via finish. */ finish_folding(ctx, op); return true; @@ -XXX,XX +XXX,XX @@ static bool fold_or(OptContext *ctx, TCGOp *op) ctx->z_mask = arg_info(op->args[1])->z_mask | arg_info(op->args[2])->z_mask; + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return fold_masks(ctx, op); } @@ -XXX,XX +XXX,XX @@ static bool fold_orc(OptContext *ctx, TCGOp *op) fold_ix_to_not(ctx, op, 0)) { return true; } + + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_xor(OptContext *ctx, TCGOp *op) ctx->z_mask = arg_info(op->args[1])->z_mask | arg_info(op->args[2])->z_mask; + ctx->s_mask = arg_info(op->args[1])->s_mask + & arg_info(op->args[2])->s_mask; return fold_masks(ctx, op); } -- 2.25.1 The result is either 0 or 1, which means that we have a 2 bit signed result, and thus 62 bits of sign. For clarity, use the smask_from_zmask function. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 2 ++ 1 file changed, 2 insertions(+) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_setcond(OptContext *ctx, TCGOp *op) } ctx->z_mask = 1; + ctx->s_mask = smask_from_zmask(1); return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_setcond2(OptContext *ctx, TCGOp *op) } ctx->z_mask = 1; + ctx->s_mask = smask_from_zmask(1); return false; do_setcond_const: -- 2.25.1 The results are generally 6 bit unsigned values, though the count leading and trailing bits may produce any value for a zero input. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static bool fold_count_zeros(OptContext *ctx, TCGOp *op) g_assert_not_reached(); } ctx->z_mask = arg_info(op->args[2])->z_mask | z_mask; - + ctx->s_mask = smask_from_zmask(ctx->z_mask); return false; } @@ -XXX,XX +XXX,XX @@ static bool fold_ctpop(OptContext *ctx, TCGOp *op) default: g_assert_not_reached(); } + ctx->s_mask = smask_from_zmask(ctx->z_mask); return false; } -- 2.25.1 For constant shifts, we can simply shift the s_mask. For variable shifts, we know that sar does not reduce the s_mask, which helps for sequences like ext32s_i64 t, in sar_i64 t, t, v ext32s_i64 out, t allowing the final extend to be eliminated. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Luis Pires <luis.pires@eldorado.org.br> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- tcg/optimize.c | 50 +++++++++++++++++++++++++++++++++++++++++++++++--- 1 file changed, 47 insertions(+), 3 deletions(-) diff --git a/tcg/optimize.c b/tcg/optimize.c index XXXXXXX..XXXXXXX 100644 --- a/tcg/optimize.c +++ b/tcg/optimize.c @@ -XXX,XX +XXX,XX @@ static uint64_t smask_from_zmask(uint64_t zmask) return ~(~0ull >> rep); } +/* + * Recreate a properly left-aligned smask after manipulation. + * Some bit-shuffling, particularly shifts and rotates, may + * retain sign bits on the left, but may scatter disconnected + * sign bits on the right. Retain only what remains to the left. + */ +static uint64_t smask_from_smask(int64_t smask) +{ + /* Only the 1 bits are significant for smask */ + return smask_from_zmask(~smask); +} + static inline TempOptInfo *ts_info(TCGTemp *ts) { return ts->state_ptr; @@ -XXX,XX +XXX,XX @@ static bool fold_sextract(OptContext *ctx, TCGOp *op) static bool fold_shift(OptContext *ctx, TCGOp *op) { + uint64_t s_mask, z_mask, sign; + if (fold_const2(ctx, op) || fold_ix_to_i(ctx, op, 0) || fold_xi_to_x(ctx, op, 0)) { return true; } + s_mask = arg_info(op->args[1])->s_mask; + z_mask = arg_info(op->args[1])->z_mask; + if (arg_is_const(op->args[2])) { - ctx->z_mask = do_constant_folding(op->opc, ctx->type, - arg_info(op->args[1])->z_mask, - arg_info(op->args[2])->val); + int sh = arg_info(op->args[2])->val; + + ctx->z_mask = do_constant_folding(op->opc, ctx->type, z_mask, sh); + + s_mask = do_constant_folding(op->opc, ctx->type, s_mask, sh); + ctx->s_mask = smask_from_smask(s_mask); + return fold_masks(ctx, op); } + + switch (op->opc) { + CASE_OP_32_64(sar): + /* + * Arithmetic right shift will not reduce the number of + * input sign repetitions. + */ + ctx->s_mask = s_mask; + break; + CASE_OP_32_64(shr): + /* + * If the sign bit is known zero, then logical right shift + * will not reduced the number of input sign repetitions. + */ + sign = (s_mask & -s_mask) >> 1; + if (!(z_mask & sign)) { + ctx->s_mask = s_mask; + } + break; + default: + break; + } + return false; } -- 2.25.1