This is v4 of my notdirty + rom patch set with two suggested name changes (qemu_build_not_reached, TLB_DISCARD_WRITE) from David and Alex. r~ The following changes since commit 240ab11fb72049d6373cbbec8d788f8e411a00bc: Merge remote-tracking branch 'remotes/aperard/tags/pull-xen-20190924' into staging (2019-09-24 15:36:31 +0100) are available in the Git repository at: https://github.com/rth7680/qemu.git tags/pull-tcg-20190925 for you to fetch changes up to ae57db63acf5a0399232f852acc5c1d83ef63400: cputlb: Pass retaddr to tb_check_watchpoint (2019-09-25 10:56:28 -0700) ---------------------------------------------------------------- Fixes for TLB_BSWAP Coversion of NOTDIRTY and ROM handling to cputlb Followup cleanups to cputlb ---------------------------------------------------------------- Richard Henderson (16): exec: Use TARGET_PAGE_BITS_MIN for TLB flags cputlb: Disable __always_inline__ without optimization qemu/compiler.h: Add qemu_build_not_reached cputlb: Use qemu_build_not_reached in load/store_helpers cputlb: Split out load/store_memop cputlb: Introduce TLB_BSWAP exec: Adjust notdirty tracing cputlb: Move ROM handling from I/O path to TLB path cputlb: Move NOTDIRTY handling from I/O path to TLB path cputlb: Partially inline memory_region_section_get_iotlb cputlb: Merge and move memory_notdirty_write_{prepare,complete} cputlb: Handle TLB_NOTDIRTY in probe_access cputlb: Remove cpu->mem_io_vaddr cputlb: Remove tb_invalidate_phys_page_range is_cpu_write_access cputlb: Pass retaddr to tb_invalidate_phys_page_fast cputlb: Pass retaddr to tb_check_watchpoint accel/tcg/translate-all.h | 8 +- include/exec/cpu-all.h | 23 ++- include/exec/cpu-common.h | 3 - include/exec/exec-all.h | 6 +- include/exec/memory-internal.h | 65 -------- include/hw/core/cpu.h | 2 - include/qemu/compiler.h | 26 +++ accel/tcg/cputlb.c | 348 +++++++++++++++++++++++++---------------- accel/tcg/translate-all.c | 51 +++--- exec.c | 158 +------------------ hw/core/cpu.c | 1 - memory.c | 20 --- trace-events | 4 +- 13 files changed, 288 insertions(+), 427 deletions(-) These bits do not need to vary with the actual page size used by the guest. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Reviewed-by: Paolo Bonzini <pbonzini@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/cpu-all.h | 16 ++++++++++------ 1 file changed, 10 insertions(+), 6 deletions(-) diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/cpu-all.h +++ b/include/exec/cpu-all.h @@ -XXX,XX +XXX,XX @@ CPUArchState *cpu_copy(CPUArchState *env); #if !defined(CONFIG_USER_ONLY) -/* Flags stored in the low bits of the TLB virtual address. These are - * defined so that fast path ram access is all zeros. +/* + * Flags stored in the low bits of the TLB virtual address. + * These are defined so that fast path ram access is all zeros. * The flags all must be between TARGET_PAGE_BITS and * maximum address alignment bit. + * + * Use TARGET_PAGE_BITS_MIN so that these bits are constant + * when TARGET_PAGE_BITS_VARY is in effect. */ /* Zero if TLB entry is valid. */ -#define TLB_INVALID_MASK (1 << (TARGET_PAGE_BITS - 1)) +#define TLB_INVALID_MASK (1 << (TARGET_PAGE_BITS_MIN - 1)) /* Set if TLB entry references a clean RAM page. The iotlb entry will contain the page physical address. */ -#define TLB_NOTDIRTY (1 << (TARGET_PAGE_BITS - 2)) +#define TLB_NOTDIRTY (1 << (TARGET_PAGE_BITS_MIN - 2)) /* Set if TLB entry is an IO callback. */ -#define TLB_MMIO (1 << (TARGET_PAGE_BITS - 3)) +#define TLB_MMIO (1 << (TARGET_PAGE_BITS_MIN - 3)) /* Set if TLB entry contains a watchpoint. */ -#define TLB_WATCHPOINT (1 << (TARGET_PAGE_BITS - 4)) +#define TLB_WATCHPOINT (1 << (TARGET_PAGE_BITS_MIN - 4)) /* Use this mask to check interception with an alignment mask * in a TCG backend. -- 2.17.1 This forced inlining can result in missing symbols, which makes a debugging build harder to follow. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com> Reviewed-by: David Hildenbrand <david@redhat.com> Reported-by: Peter Maydell <peter.maydell@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/qemu/compiler.h | 11 +++++++++++ accel/tcg/cputlb.c | 4 ++-- 2 files changed, 13 insertions(+), 2 deletions(-) diff --git a/include/qemu/compiler.h b/include/qemu/compiler.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/compiler.h +++ b/include/qemu/compiler.h @@ -XXX,XX +XXX,XX @@ # define QEMU_NONSTRING #endif +/* + * Forced inlining may be desired to encourage constant propagation + * of function parameters. However, it can also make debugging harder, + * so disable it for a non-optimizing build. + */ +#if defined(__OPTIMIZE__) +#define QEMU_ALWAYS_INLINE __attribute__((always_inline)) +#else +#define QEMU_ALWAYS_INLINE +#endif + /* Implement C11 _Generic via GCC builtins. Example: * * QEMU_GENERIC(x, (float, sinf), (long double, sinl), sin) (x) diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr, typedef uint64_t FullLoadHelper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, uintptr_t retaddr); -static inline uint64_t __attribute__((always_inline)) +static inline uint64_t QEMU_ALWAYS_INLINE load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, uintptr_t retaddr, MemOp op, bool code_read, FullLoadHelper *full_load) @@ -XXX,XX +XXX,XX @@ tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr, * Store Helpers */ -static inline void __attribute__((always_inline)) +static inline void QEMU_ALWAYS_INLINE store_helper(CPUArchState *env, target_ulong addr, uint64_t val, TCGMemOpIdx oi, uintptr_t retaddr, MemOp op) { -- 2.17.1 Use this as a compile-time assert that a particular code path is not reachable. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/qemu/compiler.h | 15 +++++++++++++++ 1 file changed, 15 insertions(+) diff --git a/include/qemu/compiler.h b/include/qemu/compiler.h index XXXXXXX..XXXXXXX 100644 --- a/include/qemu/compiler.h +++ b/include/qemu/compiler.h @@ -XXX,XX +XXX,XX @@ #define QEMU_GENERIC9(x, a0, ...) QEMU_GENERIC_IF(x, a0, QEMU_GENERIC8(x, __VA_ARGS__)) #define QEMU_GENERIC10(x, a0, ...) QEMU_GENERIC_IF(x, a0, QEMU_GENERIC9(x, __VA_ARGS__)) +/** + * qemu_build_not_reached() + * + * The compiler, during optimization, is expected to prove that a call + * to this function cannot be reached and remove it. If the compiler + * supports QEMU_ERROR, this will be reported at compile time; otherwise + * this will be reported at link time due to the missing symbol. + */ +#ifdef __OPTIMIZE__ +extern void QEMU_NORETURN QEMU_ERROR("code path is reachable") + qemu_build_not_reached(void); +#else +#define qemu_build_not_reached() g_assert_not_reached() +#endif + #endif /* COMPILER_H */ -- 2.17.1 Increase the current runtime assert to a compile-time assert. Reviewed-by: David Hildenbrand <david@redhat.com> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/cputlb.c | 5 ++--- 1 file changed, 2 insertions(+), 3 deletions(-) diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, res = ldq_le_p(haddr); break; default: - g_assert_not_reached(); + qemu_build_not_reached(); } return res; @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, stq_le_p(haddr, val); break; default: - g_assert_not_reached(); - break; + qemu_build_not_reached(); } } -- 2.17.1 We will shortly be using these more than once. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/cputlb.c | 107 +++++++++++++++++++++++---------------------- 1 file changed, 55 insertions(+), 52 deletions(-) diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr, typedef uint64_t FullLoadHelper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, uintptr_t retaddr); +static inline uint64_t QEMU_ALWAYS_INLINE +load_memop(const void *haddr, MemOp op) +{ + switch (op) { + case MO_UB: + return ldub_p(haddr); + case MO_BEUW: + return lduw_be_p(haddr); + case MO_LEUW: + return lduw_le_p(haddr); + case MO_BEUL: + return (uint32_t)ldl_be_p(haddr); + case MO_LEUL: + return (uint32_t)ldl_le_p(haddr); + case MO_BEQ: + return ldq_be_p(haddr); + case MO_LEQ: + return ldq_le_p(haddr); + default: + qemu_build_not_reached(); + } +} + static inline uint64_t QEMU_ALWAYS_INLINE load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, uintptr_t retaddr, MemOp op, bool code_read, @@ -XXX,XX +XXX,XX @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, do_aligned_access: haddr = (void *)((uintptr_t)addr + entry->addend); - switch (op) { - case MO_UB: - res = ldub_p(haddr); - break; - case MO_BEUW: - res = lduw_be_p(haddr); - break; - case MO_LEUW: - res = lduw_le_p(haddr); - break; - case MO_BEUL: - res = (uint32_t)ldl_be_p(haddr); - break; - case MO_LEUL: - res = (uint32_t)ldl_le_p(haddr); - break; - case MO_BEQ: - res = ldq_be_p(haddr); - break; - case MO_LEQ: - res = ldq_le_p(haddr); - break; - default: - qemu_build_not_reached(); - } - - return res; + return load_memop(haddr, op); } /* @@ -XXX,XX +XXX,XX @@ tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr, * Store Helpers */ +static inline void QEMU_ALWAYS_INLINE +store_memop(void *haddr, uint64_t val, MemOp op) +{ + switch (op) { + case MO_UB: + stb_p(haddr, val); + break; + case MO_BEUW: + stw_be_p(haddr, val); + break; + case MO_LEUW: + stw_le_p(haddr, val); + break; + case MO_BEUL: + stl_be_p(haddr, val); + break; + case MO_LEUL: + stl_le_p(haddr, val); + break; + case MO_BEQ: + stq_be_p(haddr, val); + break; + case MO_LEQ: + stq_le_p(haddr, val); + break; + default: + qemu_build_not_reached(); + } +} + static inline void QEMU_ALWAYS_INLINE store_helper(CPUArchState *env, target_ulong addr, uint64_t val, TCGMemOpIdx oi, uintptr_t retaddr, MemOp op) @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, do_aligned_access: haddr = (void *)((uintptr_t)addr + entry->addend); - switch (op) { - case MO_UB: - stb_p(haddr, val); - break; - case MO_BEUW: - stw_be_p(haddr, val); - break; - case MO_LEUW: - stw_le_p(haddr, val); - break; - case MO_BEUL: - stl_be_p(haddr, val); - break; - case MO_LEUL: - stl_le_p(haddr, val); - break; - case MO_BEQ: - stq_be_p(haddr, val); - break; - case MO_LEQ: - stq_le_p(haddr, val); - break; - default: - qemu_build_not_reached(); - } + store_memop(haddr, val, op); } void helper_ret_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val, -- 2.17.1 Handle bswap on ram directly in load/store_helper. This fixes a bug with the previous implementation in that one cannot use the I/O path for RAM. Fixes: a26fc6f5152b47f1 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/cpu-all.h | 4 ++- accel/tcg/cputlb.c | 72 +++++++++++++++++++++++++----------------- 2 files changed, 46 insertions(+), 30 deletions(-) diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/cpu-all.h +++ b/include/exec/cpu-all.h @@ -XXX,XX +XXX,XX @@ CPUArchState *cpu_copy(CPUArchState *env); #define TLB_MMIO (1 << (TARGET_PAGE_BITS_MIN - 3)) /* Set if TLB entry contains a watchpoint. */ #define TLB_WATCHPOINT (1 << (TARGET_PAGE_BITS_MIN - 4)) +/* Set if TLB entry requires byte swap. */ +#define TLB_BSWAP (1 << (TARGET_PAGE_BITS_MIN - 5)) /* Use this mask to check interception with an alignment mask * in a TCG backend. */ #define TLB_FLAGS_MASK \ - (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO | TLB_WATCHPOINT) + (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO | TLB_WATCHPOINT | TLB_BSWAP) /** * tlb_hit_page: return true if page aligned @addr is a hit against the diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, address |= TLB_INVALID_MASK; } if (attrs.byte_swap) { - /* Force the access through the I/O slow path. */ - address |= TLB_MMIO; + address |= TLB_BSWAP; } if (!memory_region_is_ram(section->mr) && !memory_region_is_romd(section->mr)) { @@ -XXX,XX +XXX,XX @@ static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry, bool locked = false; MemTxResult r; - if (iotlbentry->attrs.byte_swap) { - op ^= MO_BSWAP; - } - section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs); mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; @@ -XXX,XX +XXX,XX @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry, bool locked = false; MemTxResult r; - if (iotlbentry->attrs.byte_swap) { - op ^= MO_BSWAP; - } - section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs); mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; @@ -XXX,XX +XXX,XX @@ void *probe_access(CPUArchState *env, target_ulong addr, int size, wp_access, retaddr); } - if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO)) { - /* I/O access */ + /* Reject I/O access, or other required slow-path. */ + if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO | TLB_BSWAP)) { return NULL; } @@ -XXX,XX +XXX,XX @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, /* Handle anything that isn't just a straight memory access. */ if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) { CPUIOTLBEntry *iotlbentry; + bool need_swap; /* For anything that is unaligned, recurse through full_load. */ if ((addr & (size - 1)) != 0) { @@ -XXX,XX +XXX,XX @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, /* On watchpoint hit, this will longjmp out. */ cpu_check_watchpoint(env_cpu(env), addr, size, iotlbentry->attrs, BP_MEM_READ, retaddr); - - /* The backing page may or may not require I/O. */ - tlb_addr &= ~TLB_WATCHPOINT; - if ((tlb_addr & ~TARGET_PAGE_MASK) == 0) { - goto do_aligned_access; - } } + need_swap = size > 1 && (tlb_addr & TLB_BSWAP); + /* Handle I/O access. */ - return io_readx(env, iotlbentry, mmu_idx, addr, - retaddr, access_type, op); + if (likely(tlb_addr & TLB_MMIO)) { + return io_readx(env, iotlbentry, mmu_idx, addr, retaddr, + access_type, op ^ (need_swap * MO_BSWAP)); + } + + haddr = (void *)((uintptr_t)addr + entry->addend); + + /* + * Keep these two load_memop separate to ensure that the compiler + * is able to fold the entire function to a single instruction. + * There is a build-time assert inside to remind you of this. ;-) + */ + if (unlikely(need_swap)) { + return load_memop(haddr, op ^ MO_BSWAP); + } + return load_memop(haddr, op); } /* Handle slow unaligned access (it spans two pages or IO). */ @@ -XXX,XX +XXX,XX @@ load_helper(CPUArchState *env, target_ulong addr, TCGMemOpIdx oi, return res & MAKE_64BIT_MASK(0, size * 8); } - do_aligned_access: haddr = (void *)((uintptr_t)addr + entry->addend); return load_memop(haddr, op); } @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, /* Handle anything that isn't just a straight memory access. */ if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) { CPUIOTLBEntry *iotlbentry; + bool need_swap; /* For anything that is unaligned, recurse through byte stores. */ if ((addr & (size - 1)) != 0) { @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, /* On watchpoint hit, this will longjmp out. */ cpu_check_watchpoint(env_cpu(env), addr, size, iotlbentry->attrs, BP_MEM_WRITE, retaddr); - - /* The backing page may or may not require I/O. */ - tlb_addr &= ~TLB_WATCHPOINT; - if ((tlb_addr & ~TARGET_PAGE_MASK) == 0) { - goto do_aligned_access; - } } + need_swap = size > 1 && (tlb_addr & TLB_BSWAP); + /* Handle I/O access. */ - io_writex(env, iotlbentry, mmu_idx, val, addr, retaddr, op); + if (likely(tlb_addr & (TLB_MMIO | TLB_NOTDIRTY))) { + io_writex(env, iotlbentry, mmu_idx, val, addr, retaddr, + op ^ (need_swap * MO_BSWAP)); + return; + } + + haddr = (void *)((uintptr_t)addr + entry->addend); + + /* + * Keep these two store_memop separate to ensure that the compiler + * is able to fold the entire function to a single instruction. + * There is a build-time assert inside to remind you of this. ;-) + */ + if (unlikely(need_swap)) { + store_memop(haddr, val, op ^ MO_BSWAP); + } else { + store_memop(haddr, val, op); + } return; } @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, return; } - do_aligned_access: haddr = (void *)((uintptr_t)addr + entry->addend); store_memop(haddr, val, op); } -- 2.17.1 The memory_region_tb_read tracepoint is unreachable, since notdirty is supposed to apply only to writes. The memory_region_tb_write tracepoint is mis-named, because notdirty is not only used for TB invalidation. It is also used for e.g. VGA RAM updates and migration. Replace memory_region_tb_write with memory_notdirty_write_access, and place it in memory_notdirty_write_prepare where it can catch all of the instances. Add memory_notdirty_set_dirty to log when we no longer intercept writes to a page. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- exec.c | 3 +++ memory.c | 4 ---- trace-events | 4 ++-- 3 files changed, 5 insertions(+), 6 deletions(-) diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ void memory_notdirty_write_prepare(NotDirtyInfo *ndi, ndi->size = size; ndi->pages = NULL; + trace_memory_notdirty_write_access(mem_vaddr, ram_addr, size); + assert(tcg_enabled()); if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) { ndi->pages = page_collection_lock(ram_addr, ram_addr + size); @@ -XXX,XX +XXX,XX @@ void memory_notdirty_write_complete(NotDirtyInfo *ndi) /* we remove the notdirty callback only if the code has been flushed */ if (!cpu_physical_memory_is_clean(ndi->ram_addr)) { + trace_memory_notdirty_set_dirty(ndi->mem_vaddr); tlb_set_dirty(ndi->cpu, ndi->mem_vaddr); } } diff --git a/memory.c b/memory.c index XXXXXXX..XXXXXXX 100644 --- a/memory.c +++ b/memory.c @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_read_accessor(MemoryRegion *mr, /* Accesses to code which has previously been translated into a TB show * up in the MMIO path, as accesses to the io_mem_notdirty * MemoryRegion. */ - trace_memory_region_tb_read(get_cpu_index(), addr, tmp, size); } else if (TRACE_MEMORY_REGION_OPS_READ_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_read(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_read_with_attrs_accessor(MemoryRegion *mr, /* Accesses to code which has previously been translated into a TB show * up in the MMIO path, as accesses to the io_mem_notdirty * MemoryRegion. */ - trace_memory_region_tb_read(get_cpu_index(), addr, tmp, size); } else if (TRACE_MEMORY_REGION_OPS_READ_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_read(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_write_accessor(MemoryRegion *mr, /* Accesses to code which has previously been translated into a TB show * up in the MMIO path, as accesses to the io_mem_notdirty * MemoryRegion. */ - trace_memory_region_tb_write(get_cpu_index(), addr, tmp, size); } else if (TRACE_MEMORY_REGION_OPS_WRITE_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_write(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_write_with_attrs_accessor(MemoryRegion *mr, /* Accesses to code which has previously been translated into a TB show * up in the MMIO path, as accesses to the io_mem_notdirty * MemoryRegion. */ - trace_memory_region_tb_write(get_cpu_index(), addr, tmp, size); } else if (TRACE_MEMORY_REGION_OPS_WRITE_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_write(get_cpu_index(), mr, abs_addr, tmp, size); diff --git a/trace-events b/trace-events index XXXXXXX..XXXXXXX 100644 --- a/trace-events +++ b/trace-events @@ -XXX,XX +XXX,XX @@ dma_map_wait(void *dbs) "dbs=%p" find_ram_offset(uint64_t size, uint64_t offset) "size: 0x%" PRIx64 " @ 0x%" PRIx64 find_ram_offset_loop(uint64_t size, uint64_t candidate, uint64_t offset, uint64_t next, uint64_t mingap) "trying size: 0x%" PRIx64 " @ 0x%" PRIx64 ", offset: 0x%" PRIx64" next: 0x%" PRIx64 " mingap: 0x%" PRIx64 ram_block_discard_range(const char *rbname, void *hva, size_t length, bool need_madvise, bool need_fallocate, int ret) "%s@%p + 0x%zx: madvise: %d fallocate: %d ret: %d" +memory_notdirty_write_access(uint64_t vaddr, uint64_t ram_addr, unsigned size) "0x%" PRIx64 " ram_addr 0x%" PRIx64 " size %u" +memory_notdirty_set_dirty(uint64_t vaddr) "0x%" PRIx64 # memory.c memory_region_ops_read(int cpu_index, void *mr, uint64_t addr, uint64_t value, unsigned size) "cpu %d mr %p addr 0x%"PRIx64" value 0x%"PRIx64" size %u" memory_region_ops_write(int cpu_index, void *mr, uint64_t addr, uint64_t value, unsigned size) "cpu %d mr %p addr 0x%"PRIx64" value 0x%"PRIx64" size %u" memory_region_subpage_read(int cpu_index, void *mr, uint64_t offset, uint64_t value, unsigned size) "cpu %d mr %p offset 0x%"PRIx64" value 0x%"PRIx64" size %u" memory_region_subpage_write(int cpu_index, void *mr, uint64_t offset, uint64_t value, unsigned size) "cpu %d mr %p offset 0x%"PRIx64" value 0x%"PRIx64" size %u" -memory_region_tb_read(int cpu_index, uint64_t addr, uint64_t value, unsigned size) "cpu %d addr 0x%"PRIx64" value 0x%"PRIx64" size %u" -memory_region_tb_write(int cpu_index, uint64_t addr, uint64_t value, unsigned size) "cpu %d addr 0x%"PRIx64" value 0x%"PRIx64" size %u" memory_region_ram_device_read(int cpu_index, void *mr, uint64_t addr, uint64_t value, unsigned size) "cpu %d mr %p addr 0x%"PRIx64" value 0x%"PRIx64" size %u" memory_region_ram_device_write(int cpu_index, void *mr, uint64_t addr, uint64_t value, unsigned size) "cpu %d mr %p addr 0x%"PRIx64" value 0x%"PRIx64" size %u" flatview_new(void *view, void *root) "%p (root %p)" -- 2.17.1 It does not require going through the whole I/O path in order to discard a write. Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/cpu-all.h | 5 ++++- include/exec/cpu-common.h | 1 - accel/tcg/cputlb.c | 36 ++++++++++++++++++++-------------- exec.c | 41 +-------------------------------------- 4 files changed, 26 insertions(+), 57 deletions(-) diff --git a/include/exec/cpu-all.h b/include/exec/cpu-all.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/cpu-all.h +++ b/include/exec/cpu-all.h @@ -XXX,XX +XXX,XX @@ CPUArchState *cpu_copy(CPUArchState *env); #define TLB_WATCHPOINT (1 << (TARGET_PAGE_BITS_MIN - 4)) /* Set if TLB entry requires byte swap. */ #define TLB_BSWAP (1 << (TARGET_PAGE_BITS_MIN - 5)) +/* Set if TLB entry writes ignored. */ +#define TLB_DISCARD_WRITE (1 << (TARGET_PAGE_BITS_MIN - 6)) /* Use this mask to check interception with an alignment mask * in a TCG backend. */ #define TLB_FLAGS_MASK \ - (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO | TLB_WATCHPOINT | TLB_BSWAP) + (TLB_INVALID_MASK | TLB_NOTDIRTY | TLB_MMIO \ + | TLB_WATCHPOINT | TLB_BSWAP | TLB_DISCARD_WRITE) /** * tlb_hit_page: return true if page aligned @addr is a hit against the diff --git a/include/exec/cpu-common.h b/include/exec/cpu-common.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/cpu-common.h +++ b/include/exec/cpu-common.h @@ -XXX,XX +XXX,XX @@ void qemu_flush_coalesced_mmio_buffer(void); void cpu_flush_icache_range(hwaddr start, hwaddr len); -extern struct MemoryRegion io_mem_rom; extern struct MemoryRegion io_mem_notdirty; typedef int (RAMBlockIterFunc)(RAMBlock *rb, void *opaque); diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static void tlb_reset_dirty_range_locked(CPUTLBEntry *tlb_entry, { uintptr_t addr = tlb_entry->addr_write; - if ((addr & (TLB_INVALID_MASK | TLB_MMIO | TLB_NOTDIRTY)) == 0) { + if ((addr & (TLB_INVALID_MASK | TLB_MMIO | + TLB_DISCARD_WRITE | TLB_NOTDIRTY)) == 0) { addr &= TARGET_PAGE_MASK; addr += tlb_entry->addend; if ((addr - start) < length) { @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, address |= TLB_MMIO; addend = 0; } else { - /* TLB_MMIO for rom/romd handled below */ addend = (uintptr_t)memory_region_get_ram_ptr(section->mr) + xlat; } @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, tn.addr_write = -1; if (prot & PAGE_WRITE) { - if ((memory_region_is_ram(section->mr) && section->readonly) - || memory_region_is_romd(section->mr)) { - /* Write access calls the I/O callback. */ - tn.addr_write = address | TLB_MMIO; - } else if (memory_region_is_ram(section->mr) - && cpu_physical_memory_is_clean( - memory_region_get_ram_addr(section->mr) + xlat)) { - tn.addr_write = address | TLB_NOTDIRTY; - } else { - tn.addr_write = address; + tn.addr_write = address; + if (memory_region_is_romd(section->mr)) { + /* Use the MMIO path so that the device can switch states. */ + tn.addr_write |= TLB_MMIO; + } else if (memory_region_is_ram(section->mr)) { + if (section->readonly) { + tn.addr_write |= TLB_DISCARD_WRITE; + } else if (cpu_physical_memory_is_clean( + memory_region_get_ram_addr(section->mr) + xlat)) { + tn.addr_write |= TLB_NOTDIRTY; + } } if (prot & PAGE_WRITE_INV) { tn.addr_write |= TLB_INVALID_MASK; @@ -XXX,XX +XXX,XX @@ static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry, mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; cpu->mem_io_pc = retaddr; - if (mr != &io_mem_rom && mr != &io_mem_notdirty && !cpu->can_do_io) { + if (mr != &io_mem_notdirty && !cpu->can_do_io) { cpu_io_recompile(cpu, retaddr); } @@ -XXX,XX +XXX,XX @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry, section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs); mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; - if (mr != &io_mem_rom && mr != &io_mem_notdirty && !cpu->can_do_io) { + if (mr != &io_mem_notdirty && !cpu->can_do_io) { cpu_io_recompile(cpu, retaddr); } cpu->mem_io_vaddr = addr; @@ -XXX,XX +XXX,XX @@ void *probe_access(CPUArchState *env, target_ulong addr, int size, } /* Reject I/O access, or other required slow-path. */ - if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO | TLB_BSWAP)) { + if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO | TLB_BSWAP | TLB_DISCARD_WRITE)) { return NULL; } @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, return; } + /* Ignore writes to ROM. */ + if (unlikely(tlb_addr & TLB_DISCARD_WRITE)) { + return; + } + haddr = (void *)((uintptr_t)addr + entry->addend); /* diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ static MemoryRegion *system_io; AddressSpace address_space_io; AddressSpace address_space_memory; -MemoryRegion io_mem_rom, io_mem_notdirty; +MemoryRegion io_mem_notdirty; static MemoryRegion io_mem_unassigned; #endif @@ -XXX,XX +XXX,XX @@ typedef struct subpage_t { #define PHYS_SECTION_UNASSIGNED 0 #define PHYS_SECTION_NOTDIRTY 1 -#define PHYS_SECTION_ROM 2 static void io_mem_init(void); static void memory_map_init(void); @@ -XXX,XX +XXX,XX @@ hwaddr memory_region_section_get_iotlb(CPUState *cpu, iotlb = memory_region_get_ram_addr(section->mr) + xlat; if (!section->readonly) { iotlb |= PHYS_SECTION_NOTDIRTY; - } else { - iotlb |= PHYS_SECTION_ROM; } } else { AddressSpaceDispatch *d; @@ -XXX,XX +XXX,XX @@ static uint16_t dummy_section(PhysPageMap *map, FlatView *fv, MemoryRegion *mr) return phys_section_add(map, &section); } -static void readonly_mem_write(void *opaque, hwaddr addr, - uint64_t val, unsigned size) -{ - /* Ignore any write to ROM. */ -} - -static bool readonly_mem_accepts(void *opaque, hwaddr addr, - unsigned size, bool is_write, - MemTxAttrs attrs) -{ - return is_write; -} - -/* This will only be used for writes, because reads are special cased - * to directly access the underlying host ram. - */ -static const MemoryRegionOps readonly_mem_ops = { - .write = readonly_mem_write, - .valid.accepts = readonly_mem_accepts, - .endianness = DEVICE_NATIVE_ENDIAN, - .valid = { - .min_access_size = 1, - .max_access_size = 8, - .unaligned = false, - }, - .impl = { - .min_access_size = 1, - .max_access_size = 8, - .unaligned = false, - }, -}; - MemoryRegionSection *iotlb_to_section(CPUState *cpu, hwaddr index, MemTxAttrs attrs) { @@ -XXX,XX +XXX,XX @@ MemoryRegionSection *iotlb_to_section(CPUState *cpu, static void io_mem_init(void) { - memory_region_init_io(&io_mem_rom, NULL, &readonly_mem_ops, - NULL, NULL, UINT64_MAX); memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL, NULL, UINT64_MAX); @@ -XXX,XX +XXX,XX @@ AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv) assert(n == PHYS_SECTION_UNASSIGNED); n = dummy_section(&d->map, fv, &io_mem_notdirty); assert(n == PHYS_SECTION_NOTDIRTY); - n = dummy_section(&d->map, fv, &io_mem_rom); - assert(n == PHYS_SECTION_ROM); d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 }; -- 2.17.1 Pages that we want to track for NOTDIRTY are RAM. We do not really need to go through the I/O path to handle them. Acked-by: David Hildenbrand <david@redhat.com> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: Philippe Mathieu-Daudé <philmd@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/cpu-common.h | 2 -- accel/tcg/cputlb.c | 26 +++++++++++++++++--- exec.c | 50 --------------------------------------- memory.c | 16 ------------- 4 files changed, 23 insertions(+), 71 deletions(-) diff --git a/include/exec/cpu-common.h b/include/exec/cpu-common.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/cpu-common.h +++ b/include/exec/cpu-common.h @@ -XXX,XX +XXX,XX @@ void qemu_flush_coalesced_mmio_buffer(void); void cpu_flush_icache_range(hwaddr start, hwaddr len); -extern struct MemoryRegion io_mem_notdirty; - typedef int (RAMBlockIterFunc)(RAMBlock *rb, void *opaque); int qemu_ram_foreach_block(RAMBlockIterFunc func, void *opaque); diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry, mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; cpu->mem_io_pc = retaddr; - if (mr != &io_mem_notdirty && !cpu->can_do_io) { + if (!cpu->can_do_io) { cpu_io_recompile(cpu, retaddr); } @@ -XXX,XX +XXX,XX @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry, section = iotlb_to_section(cpu, iotlbentry->addr, iotlbentry->attrs); mr = section->mr; mr_offset = (iotlbentry->addr & TARGET_PAGE_MASK) + addr; - if (mr != &io_mem_notdirty && !cpu->can_do_io) { + if (!cpu->can_do_io) { cpu_io_recompile(cpu, retaddr); } cpu->mem_io_vaddr = addr; @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, need_swap = size > 1 && (tlb_addr & TLB_BSWAP); /* Handle I/O access. */ - if (likely(tlb_addr & (TLB_MMIO | TLB_NOTDIRTY))) { + if (tlb_addr & TLB_MMIO) { io_writex(env, iotlbentry, mmu_idx, val, addr, retaddr, op ^ (need_swap * MO_BSWAP)); return; @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, haddr = (void *)((uintptr_t)addr + entry->addend); + /* Handle clean RAM pages. */ + if (tlb_addr & TLB_NOTDIRTY) { + NotDirtyInfo ndi; + + /* We require mem_io_pc in tb_invalidate_phys_page_range. */ + env_cpu(env)->mem_io_pc = retaddr; + + memory_notdirty_write_prepare(&ndi, env_cpu(env), addr, + addr + iotlbentry->addr, size); + + if (unlikely(need_swap)) { + store_memop(haddr, val, op ^ MO_BSWAP); + } else { + store_memop(haddr, val, op); + } + + memory_notdirty_write_complete(&ndi); + return; + } + /* * Keep these two store_memop separate to ensure that the compiler * is able to fold the entire function to a single instruction. diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ static MemoryRegion *system_io; AddressSpace address_space_io; AddressSpace address_space_memory; -MemoryRegion io_mem_notdirty; static MemoryRegion io_mem_unassigned; #endif @@ -XXX,XX +XXX,XX @@ typedef struct subpage_t { } subpage_t; #define PHYS_SECTION_UNASSIGNED 0 -#define PHYS_SECTION_NOTDIRTY 1 static void io_mem_init(void); static void memory_map_init(void); @@ -XXX,XX +XXX,XX @@ hwaddr memory_region_section_get_iotlb(CPUState *cpu, if (memory_region_is_ram(section->mr)) { /* Normal RAM. */ iotlb = memory_region_get_ram_addr(section->mr) + xlat; - if (!section->readonly) { - iotlb |= PHYS_SECTION_NOTDIRTY; - } } else { AddressSpaceDispatch *d; @@ -XXX,XX +XXX,XX @@ void memory_notdirty_write_complete(NotDirtyInfo *ndi) } } -/* Called within RCU critical section. */ -static void notdirty_mem_write(void *opaque, hwaddr ram_addr, - uint64_t val, unsigned size) -{ - NotDirtyInfo ndi; - - memory_notdirty_write_prepare(&ndi, current_cpu, current_cpu->mem_io_vaddr, - ram_addr, size); - - stn_p(qemu_map_ram_ptr(NULL, ram_addr), size, val); - memory_notdirty_write_complete(&ndi); -} - -static bool notdirty_mem_accepts(void *opaque, hwaddr addr, - unsigned size, bool is_write, - MemTxAttrs attrs) -{ - return is_write; -} - -static const MemoryRegionOps notdirty_mem_ops = { - .write = notdirty_mem_write, - .valid.accepts = notdirty_mem_accepts, - .endianness = DEVICE_NATIVE_ENDIAN, - .valid = { - .min_access_size = 1, - .max_access_size = 8, - .unaligned = false, - }, - .impl = { - .min_access_size = 1, - .max_access_size = 8, - .unaligned = false, - }, -}; - /* Generate a debug exception if a watchpoint has been hit. */ void cpu_check_watchpoint(CPUState *cpu, vaddr addr, vaddr len, MemTxAttrs attrs, int flags, uintptr_t ra) @@ -XXX,XX +XXX,XX @@ static void io_mem_init(void) { memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL, NULL, UINT64_MAX); - - /* io_mem_notdirty calls tb_invalidate_phys_page_fast, - * which can be called without the iothread mutex. - */ - memory_region_init_io(&io_mem_notdirty, NULL, &notdirty_mem_ops, NULL, - NULL, UINT64_MAX); - memory_region_clear_global_locking(&io_mem_notdirty); } AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv) @@ -XXX,XX +XXX,XX @@ AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv) n = dummy_section(&d->map, fv, &io_mem_unassigned); assert(n == PHYS_SECTION_UNASSIGNED); - n = dummy_section(&d->map, fv, &io_mem_notdirty); - assert(n == PHYS_SECTION_NOTDIRTY); d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 }; diff --git a/memory.c b/memory.c index XXXXXXX..XXXXXXX 100644 --- a/memory.c +++ b/memory.c @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_read_accessor(MemoryRegion *mr, tmp = mr->ops->read(mr->opaque, addr, size); if (mr->subpage) { trace_memory_region_subpage_read(get_cpu_index(), mr, addr, tmp, size); - } else if (mr == &io_mem_notdirty) { - /* Accesses to code which has previously been translated into a TB show - * up in the MMIO path, as accesses to the io_mem_notdirty - * MemoryRegion. */ } else if (TRACE_MEMORY_REGION_OPS_READ_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_read(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_read_with_attrs_accessor(MemoryRegion *mr, r = mr->ops->read_with_attrs(mr->opaque, addr, &tmp, size, attrs); if (mr->subpage) { trace_memory_region_subpage_read(get_cpu_index(), mr, addr, tmp, size); - } else if (mr == &io_mem_notdirty) { - /* Accesses to code which has previously been translated into a TB show - * up in the MMIO path, as accesses to the io_mem_notdirty - * MemoryRegion. */ } else if (TRACE_MEMORY_REGION_OPS_READ_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_read(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_write_accessor(MemoryRegion *mr, if (mr->subpage) { trace_memory_region_subpage_write(get_cpu_index(), mr, addr, tmp, size); - } else if (mr == &io_mem_notdirty) { - /* Accesses to code which has previously been translated into a TB show - * up in the MMIO path, as accesses to the io_mem_notdirty - * MemoryRegion. */ } else if (TRACE_MEMORY_REGION_OPS_WRITE_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_write(get_cpu_index(), mr, abs_addr, tmp, size); @@ -XXX,XX +XXX,XX @@ static MemTxResult memory_region_write_with_attrs_accessor(MemoryRegion *mr, if (mr->subpage) { trace_memory_region_subpage_write(get_cpu_index(), mr, addr, tmp, size); - } else if (mr == &io_mem_notdirty) { - /* Accesses to code which has previously been translated into a TB show - * up in the MMIO path, as accesses to the io_mem_notdirty - * MemoryRegion. */ } else if (TRACE_MEMORY_REGION_OPS_WRITE_ENABLED) { hwaddr abs_addr = memory_region_to_absolute_addr(mr, addr); trace_memory_region_ops_write(get_cpu_index(), mr, abs_addr, tmp, size); -- 2.17.1 There is only one caller, tlb_set_page_with_attrs. We cannot inline the entire function because the AddressSpaceDispatch structure is private to exec.c, and cannot easily be moved to include/exec/memory-internal.h. Compute is_ram and is_romd once within tlb_set_page_with_attrs. Fold the number of tests against these predicates. Compute cpu_physical_memory_is_clean outside of the tlb lock region. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/exec-all.h | 6 +--- accel/tcg/cputlb.c | 68 ++++++++++++++++++++++++++--------------- exec.c | 22 ++----------- 3 files changed, 47 insertions(+), 49 deletions(-) diff --git a/include/exec/exec-all.h b/include/exec/exec-all.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/exec-all.h +++ b/include/exec/exec-all.h @@ -XXX,XX +XXX,XX @@ address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr addr, hwaddr *xlat, hwaddr *plen, MemTxAttrs attrs, int *prot); hwaddr memory_region_section_get_iotlb(CPUState *cpu, - MemoryRegionSection *section, - target_ulong vaddr, - hwaddr paddr, hwaddr xlat, - int prot, - target_ulong *address); + MemoryRegionSection *section); #endif /* vl.c */ diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, MemoryRegionSection *section; unsigned int index; target_ulong address; - target_ulong code_address; + target_ulong write_address; uintptr_t addend; CPUTLBEntry *te, tn; hwaddr iotlb, xlat, sz, paddr_page; target_ulong vaddr_page; int asidx = cpu_asidx_from_attrs(cpu, attrs); int wp_flags; + bool is_ram, is_romd; assert_cpu_is_self(cpu); @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, if (attrs.byte_swap) { address |= TLB_BSWAP; } - if (!memory_region_is_ram(section->mr) && - !memory_region_is_romd(section->mr)) { - /* IO memory case */ - address |= TLB_MMIO; - addend = 0; - } else { + + is_ram = memory_region_is_ram(section->mr); + is_romd = memory_region_is_romd(section->mr); + + if (is_ram || is_romd) { + /* RAM and ROMD both have associated host memory. */ addend = (uintptr_t)memory_region_get_ram_ptr(section->mr) + xlat; + } else { + /* I/O does not; force the host address to NULL. */ + addend = 0; + } + + write_address = address; + if (is_ram) { + iotlb = memory_region_get_ram_addr(section->mr) + xlat; + /* + * Computing is_clean is expensive; avoid all that unless + * the page is actually writable. + */ + if (prot & PAGE_WRITE) { + if (section->readonly) { + write_address |= TLB_DISCARD_WRITE; + } else if (cpu_physical_memory_is_clean(iotlb)) { + write_address |= TLB_NOTDIRTY; + } + } + } else { + /* I/O or ROMD */ + iotlb = memory_region_section_get_iotlb(cpu, section) + xlat; + /* + * Writes to romd devices must go through MMIO to enable write. + * Reads to romd devices go through the ram_ptr found above, + * but of course reads to I/O must go through MMIO. + */ + write_address |= TLB_MMIO; + if (!is_romd) { + address = write_address; + } } - code_address = address; - iotlb = memory_region_section_get_iotlb(cpu, section, vaddr_page, - paddr_page, xlat, prot, &address); wp_flags = cpu_watchpoint_address_matches(cpu, vaddr_page, TARGET_PAGE_SIZE); @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, /* * At this point iotlb contains a physical section number in the lower * TARGET_PAGE_BITS, and either - * + the ram_addr_t of the page base of the target RAM (if NOTDIRTY or ROM) - * + the offset within section->mr of the page base (otherwise) + * + the ram_addr_t of the page base of the target RAM (RAM) + * + the offset within section->mr of the page base (I/O, ROMD) * We subtract the vaddr_page (which is page aligned and thus won't * disturb the low bits) to give an offset which can be added to the * (non-page-aligned) vaddr of the eventual memory access to get @@ -XXX,XX +XXX,XX @@ void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr, } if (prot & PAGE_EXEC) { - tn.addr_code = code_address; + tn.addr_code = address; } else { tn.addr_code = -1; } tn.addr_write = -1; if (prot & PAGE_WRITE) { - tn.addr_write = address; - if (memory_region_is_romd(section->mr)) { - /* Use the MMIO path so that the device can switch states. */ - tn.addr_write |= TLB_MMIO; - } else if (memory_region_is_ram(section->mr)) { - if (section->readonly) { - tn.addr_write |= TLB_DISCARD_WRITE; - } else if (cpu_physical_memory_is_clean( - memory_region_get_ram_addr(section->mr) + xlat)) { - tn.addr_write |= TLB_NOTDIRTY; - } - } + tn.addr_write = write_address; if (prot & PAGE_WRITE_INV) { tn.addr_write |= TLB_INVALID_MASK; } diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap, /* Called from RCU critical section */ hwaddr memory_region_section_get_iotlb(CPUState *cpu, - MemoryRegionSection *section, - target_ulong vaddr, - hwaddr paddr, hwaddr xlat, - int prot, - target_ulong *address) + MemoryRegionSection *section) { - hwaddr iotlb; - - if (memory_region_is_ram(section->mr)) { - /* Normal RAM. */ - iotlb = memory_region_get_ram_addr(section->mr) + xlat; - } else { - AddressSpaceDispatch *d; - - d = flatview_to_dispatch(section->fv); - iotlb = section - d->map.sections; - iotlb += xlat; - } - - return iotlb; + AddressSpaceDispatch *d = flatview_to_dispatch(section->fv); + return section - d->map.sections; } #endif /* defined(CONFIG_USER_ONLY) */ -- 2.17.1 Since 9458a9a1df1a, all readers of the dirty bitmaps wait for the rcu lock, which means that they wait until the end of any executing TranslationBlock. As a consequence, there is no need for the actual access to happen in between the _prepare and _complete. Therefore, we can improve things by merging the two functions into notdirty_write and dropping the NotDirtyInfo structure. In addition, the only users of notdirty_write are in cputlb.c, so move the merged function there. Pass in the CPUIOTLBEntry from which the ram_addr_t may be computed. Reviewed-by: David Hildenbrand <david@redhat.com> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/exec/memory-internal.h | 65 ----------------------------- accel/tcg/cputlb.c | 76 +++++++++++++++++++--------------- exec.c | 44 -------------------- 3 files changed, 42 insertions(+), 143 deletions(-) diff --git a/include/exec/memory-internal.h b/include/exec/memory-internal.h index XXXXXXX..XXXXXXX 100644 --- a/include/exec/memory-internal.h +++ b/include/exec/memory-internal.h @@ -XXX,XX +XXX,XX @@ void address_space_dispatch_free(AddressSpaceDispatch *d); void mtree_print_dispatch(struct AddressSpaceDispatch *d, MemoryRegion *root); - -struct page_collection; - -/* Opaque struct for passing info from memory_notdirty_write_prepare() - * to memory_notdirty_write_complete(). Callers should treat all fields - * as private, with the exception of @active. - * - * @active is a field which is not touched by either the prepare or - * complete functions, but which the caller can use if it wishes to - * track whether it has called prepare for this struct and so needs - * to later call the complete function. - */ -typedef struct { - CPUState *cpu; - struct page_collection *pages; - ram_addr_t ram_addr; - vaddr mem_vaddr; - unsigned size; - bool active; -} NotDirtyInfo; - -/** - * memory_notdirty_write_prepare: call before writing to non-dirty memory - * @ndi: pointer to opaque NotDirtyInfo struct - * @cpu: CPU doing the write - * @mem_vaddr: virtual address of write - * @ram_addr: the ram address of the write - * @size: size of write in bytes - * - * Any code which writes to the host memory corresponding to - * guest RAM which has been marked as NOTDIRTY must wrap those - * writes in calls to memory_notdirty_write_prepare() and - * memory_notdirty_write_complete(): - * - * NotDirtyInfo ndi; - * memory_notdirty_write_prepare(&ndi, ....); - * ... perform write here ... - * memory_notdirty_write_complete(&ndi); - * - * These calls will ensure that we flush any TCG translated code for - * the memory being written, update the dirty bits and (if possible) - * remove the slowpath callback for writing to the memory. - * - * This must only be called if we are using TCG; it will assert otherwise. - * - * We may take locks in the prepare call, so callers must ensure that - * they don't exit (via longjump or otherwise) without calling complete. - * - * This call must only be made inside an RCU critical section. - * (Note that while we're executing a TCG TB we're always in an - * RCU critical section, which is likely to be the case for callers - * of these functions.) - */ -void memory_notdirty_write_prepare(NotDirtyInfo *ndi, - CPUState *cpu, - vaddr mem_vaddr, - ram_addr_t ram_addr, - unsigned size); -/** - * memory_notdirty_write_complete: finish write to non-dirty memory - * @ndi: pointer to the opaque NotDirtyInfo struct which was initialized - * by memory_not_dirty_write_prepare(). - */ -void memory_notdirty_write_complete(NotDirtyInfo *ndi); - #endif #endif diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ #include "exec/helper-proto.h" #include "qemu/atomic.h" #include "qemu/atomic128.h" +#include "translate-all.h" /* DEBUG defines, enable DEBUG_TLB_LOG to log to the CPU_LOG_MMU target */ /* #define DEBUG_TLB */ @@ -XXX,XX +XXX,XX @@ tb_page_addr_t get_page_addr_code(CPUArchState *env, target_ulong addr) return qemu_ram_addr_from_host_nofail(p); } +static void notdirty_write(CPUState *cpu, vaddr mem_vaddr, unsigned size, + CPUIOTLBEntry *iotlbentry, uintptr_t retaddr) +{ + ram_addr_t ram_addr = mem_vaddr + iotlbentry->addr; + + trace_memory_notdirty_write_access(mem_vaddr, ram_addr, size); + + if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) { + struct page_collection *pages + = page_collection_lock(ram_addr, ram_addr + size); + + /* We require mem_io_pc in tb_invalidate_phys_page_range. */ + cpu->mem_io_pc = retaddr; + + tb_invalidate_phys_page_fast(pages, ram_addr, size); + page_collection_unlock(pages); + } + + /* + * Set both VGA and migration bits for simplicity and to remove + * the notdirty callback faster. + */ + cpu_physical_memory_set_dirty_range(ram_addr, size, DIRTY_CLIENTS_NOCODE); + + /* We remove the notdirty callback only if the code has been flushed. */ + if (!cpu_physical_memory_is_clean(ram_addr)) { + trace_memory_notdirty_set_dirty(mem_vaddr); + tlb_set_dirty(cpu, mem_vaddr); + } +} + /* * Probe for whether the specified guest access is permitted. If it is not * permitted then an exception will be taken in the same way as if this @@ -XXX,XX +XXX,XX @@ void *tlb_vaddr_to_host(CPUArchState *env, abi_ptr addr, /* Probe for a read-modify-write atomic operation. Do not allow unaligned * operations, or io operations to proceed. Return the host address. */ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr, - TCGMemOpIdx oi, uintptr_t retaddr, - NotDirtyInfo *ndi) + TCGMemOpIdx oi, uintptr_t retaddr) { size_t mmu_idx = get_mmuidx(oi); uintptr_t index = tlb_index(env, mmu_idx, addr); @@ -XXX,XX +XXX,XX @@ static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr, hostaddr = (void *)((uintptr_t)addr + tlbe->addend); - ndi->active = false; if (unlikely(tlb_addr & TLB_NOTDIRTY)) { - ndi->active = true; - memory_notdirty_write_prepare(ndi, env_cpu(env), addr, - qemu_ram_addr_from_host_nofail(hostaddr), - 1 << s_bits); + notdirty_write(env_cpu(env), addr, 1 << s_bits, + &env_tlb(env)->d[mmu_idx].iotlb[index], retaddr); } return hostaddr; @@ -XXX,XX +XXX,XX @@ store_helper(CPUArchState *env, target_ulong addr, uint64_t val, return; } - haddr = (void *)((uintptr_t)addr + entry->addend); - /* Handle clean RAM pages. */ if (tlb_addr & TLB_NOTDIRTY) { - NotDirtyInfo ndi; - - /* We require mem_io_pc in tb_invalidate_phys_page_range. */ - env_cpu(env)->mem_io_pc = retaddr; - - memory_notdirty_write_prepare(&ndi, env_cpu(env), addr, - addr + iotlbentry->addr, size); - - if (unlikely(need_swap)) { - store_memop(haddr, val, op ^ MO_BSWAP); - } else { - store_memop(haddr, val, op); - } - - memory_notdirty_write_complete(&ndi); - return; + notdirty_write(env_cpu(env), addr, size, iotlbentry, retaddr); } + haddr = (void *)((uintptr_t)addr + entry->addend); + /* * Keep these two store_memop separate to ensure that the compiler * is able to fold the entire function to a single instruction. @@ -XXX,XX +XXX,XX @@ void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val, #define EXTRA_ARGS , TCGMemOpIdx oi, uintptr_t retaddr #define ATOMIC_NAME(X) \ HELPER(glue(glue(glue(atomic_ ## X, SUFFIX), END), _mmu)) -#define ATOMIC_MMU_DECLS NotDirtyInfo ndi -#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, retaddr, &ndi) -#define ATOMIC_MMU_CLEANUP \ - do { \ - if (unlikely(ndi.active)) { \ - memory_notdirty_write_complete(&ndi); \ - } \ - } while (0) +#define ATOMIC_MMU_DECLS +#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, retaddr) +#define ATOMIC_MMU_CLEANUP #define DATA_SIZE 1 #include "atomic_template.h" @@ -XXX,XX +XXX,XX @@ void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val, #undef ATOMIC_MMU_LOOKUP #define EXTRA_ARGS , TCGMemOpIdx oi #define ATOMIC_NAME(X) HELPER(glue(glue(atomic_ ## X, SUFFIX), END)) -#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, GETPC(), &ndi) +#define ATOMIC_MMU_LOOKUP atomic_mmu_lookup(env, addr, oi, GETPC()) #define DATA_SIZE 1 #include "atomic_template.h" diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ ram_addr_t qemu_ram_addr_from_host(void *ptr) return block->offset + offset; } -/* Called within RCU critical section. */ -void memory_notdirty_write_prepare(NotDirtyInfo *ndi, - CPUState *cpu, - vaddr mem_vaddr, - ram_addr_t ram_addr, - unsigned size) -{ - ndi->cpu = cpu; - ndi->ram_addr = ram_addr; - ndi->mem_vaddr = mem_vaddr; - ndi->size = size; - ndi->pages = NULL; - - trace_memory_notdirty_write_access(mem_vaddr, ram_addr, size); - - assert(tcg_enabled()); - if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) { - ndi->pages = page_collection_lock(ram_addr, ram_addr + size); - tb_invalidate_phys_page_fast(ndi->pages, ram_addr, size); - } -} - -/* Called within RCU critical section. */ -void memory_notdirty_write_complete(NotDirtyInfo *ndi) -{ - if (ndi->pages) { - assert(tcg_enabled()); - page_collection_unlock(ndi->pages); - ndi->pages = NULL; - } - - /* Set both VGA and migration bits for simplicity and to remove - * the notdirty callback faster. - */ - cpu_physical_memory_set_dirty_range(ndi->ram_addr, ndi->size, - DIRTY_CLIENTS_NOCODE); - /* we remove the notdirty callback only if the code has been - flushed */ - if (!cpu_physical_memory_is_clean(ndi->ram_addr)) { - trace_memory_notdirty_set_dirty(ndi->mem_vaddr); - tlb_set_dirty(ndi->cpu, ndi->mem_vaddr); - } -} - /* Generate a debug exception if a watchpoint has been hit. */ void cpu_check_watchpoint(CPUState *cpu, vaddr addr, vaddr len, MemTxAttrs attrs, int flags, uintptr_t ra) -- 2.17.1 We can use notdirty_write for the write and return a valid host pointer for this case. Reviewed-by: David Hildenbrand <david@redhat.com> Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/cputlb.c | 26 +++++++++++++++++--------- 1 file changed, 17 insertions(+), 9 deletions(-) diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ void *probe_access(CPUArchState *env, target_ulong addr, int size, return NULL; } - /* Handle watchpoints. */ - if (tlb_addr & TLB_WATCHPOINT) { - cpu_check_watchpoint(env_cpu(env), addr, size, - env_tlb(env)->d[mmu_idx].iotlb[index].attrs, - wp_access, retaddr); - } + if (unlikely(tlb_addr & TLB_FLAGS_MASK)) { + CPUIOTLBEntry *iotlbentry = &env_tlb(env)->d[mmu_idx].iotlb[index]; - /* Reject I/O access, or other required slow-path. */ - if (tlb_addr & (TLB_NOTDIRTY | TLB_MMIO | TLB_BSWAP | TLB_DISCARD_WRITE)) { - return NULL; + /* Reject I/O access, or other required slow-path. */ + if (tlb_addr & (TLB_MMIO | TLB_BSWAP | TLB_DISCARD_WRITE)) { + return NULL; + } + + /* Handle watchpoints. */ + if (tlb_addr & TLB_WATCHPOINT) { + cpu_check_watchpoint(env_cpu(env), addr, size, + iotlbentry->attrs, wp_access, retaddr); + } + + /* Handle clean RAM pages. */ + if (tlb_addr & TLB_NOTDIRTY) { + notdirty_write(env_cpu(env), addr, size, iotlbentry, retaddr); + } } return (void *)((uintptr_t)addr + entry->addend); -- 2.17.1 With the merge of notdirty handling into store_helper, the last user of cpu->mem_io_vaddr was removed. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- include/hw/core/cpu.h | 2 -- accel/tcg/cputlb.c | 2 -- hw/core/cpu.c | 1 - 3 files changed, 5 deletions(-) diff --git a/include/hw/core/cpu.h b/include/hw/core/cpu.h index XXXXXXX..XXXXXXX 100644 --- a/include/hw/core/cpu.h +++ b/include/hw/core/cpu.h @@ -XXX,XX +XXX,XX @@ struct qemu_work_item; * @next_cpu: Next CPU sharing TB cache. * @opaque: User data. * @mem_io_pc: Host Program Counter at which the memory was accessed. - * @mem_io_vaddr: Target virtual address at which the memory was accessed. * @kvm_fd: vCPU file descriptor for KVM. * @work_mutex: Lock to prevent multiple access to queued_work_*. * @queued_work_first: First asynchronous work pending. @@ -XXX,XX +XXX,XX @@ struct CPUState { * we store some rarely used information in the CPU context. */ uintptr_t mem_io_pc; - vaddr mem_io_vaddr; /* * This is only needed for the legacy cpu_unassigned_access() hook; * when all targets using it have been converted to use diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static uint64_t io_readx(CPUArchState *env, CPUIOTLBEntry *iotlbentry, cpu_io_recompile(cpu, retaddr); } - cpu->mem_io_vaddr = addr; cpu->mem_io_access_type = access_type; if (mr->global_locking && !qemu_mutex_iothread_locked()) { @@ -XXX,XX +XXX,XX @@ static void io_writex(CPUArchState *env, CPUIOTLBEntry *iotlbentry, if (!cpu->can_do_io) { cpu_io_recompile(cpu, retaddr); } - cpu->mem_io_vaddr = addr; cpu->mem_io_pc = retaddr; if (mr->global_locking && !qemu_mutex_iothread_locked()) { diff --git a/hw/core/cpu.c b/hw/core/cpu.c index XXXXXXX..XXXXXXX 100644 --- a/hw/core/cpu.c +++ b/hw/core/cpu.c @@ -XXX,XX +XXX,XX @@ static void cpu_common_reset(CPUState *cpu) cpu->interrupt_request = 0; cpu->halted = 0; cpu->mem_io_pc = 0; - cpu->mem_io_vaddr = 0; cpu->icount_extra = 0; atomic_set(&cpu->icount_decr_ptr->u32, 0); cpu->can_do_io = 1; -- 2.17.1 All callers pass false to this argument. Remove it and pass the constant on to tb_invalidate_phys_page_range__locked. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/translate-all.h | 3 +-- accel/tcg/translate-all.c | 6 ++---- exec.c | 4 ++-- 3 files changed, 5 insertions(+), 8 deletions(-) diff --git a/accel/tcg/translate-all.h b/accel/tcg/translate-all.h index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.h +++ b/accel/tcg/translate-all.h @@ -XXX,XX +XXX,XX @@ struct page_collection *page_collection_lock(tb_page_addr_t start, void page_collection_unlock(struct page_collection *set); void tb_invalidate_phys_page_fast(struct page_collection *pages, tb_page_addr_t start, int len); -void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end, - int is_cpu_write_access); +void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end); void tb_check_watchpoint(CPUState *cpu); #ifdef CONFIG_USER_ONLY 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 @@ tb_invalidate_phys_page_range__locked(struct page_collection *pages, * * Called with mmap_lock held for user-mode emulation */ -void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end, - int is_cpu_write_access) +void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end) { struct page_collection *pages; PageDesc *p; @@ -XXX,XX +XXX,XX @@ void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end, return; } pages = page_collection_lock(start, end); - tb_invalidate_phys_page_range__locked(pages, p, start, end, - is_cpu_write_access); + tb_invalidate_phys_page_range__locked(pages, p, start, end, 0); page_collection_unlock(pages); } diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ const char *parse_cpu_option(const char *cpu_option) void tb_invalidate_phys_addr(target_ulong addr) { mmap_lock(); - tb_invalidate_phys_page_range(addr, addr + 1, 0); + tb_invalidate_phys_page_range(addr, addr + 1); mmap_unlock(); } @@ -XXX,XX +XXX,XX @@ void tb_invalidate_phys_addr(AddressSpace *as, hwaddr addr, MemTxAttrs attrs) return; } ram_addr = memory_region_get_ram_addr(mr) + addr; - tb_invalidate_phys_page_range(ram_addr, ram_addr + 1, 0); + tb_invalidate_phys_page_range(ram_addr, ram_addr + 1); rcu_read_unlock(); } -- 2.17.1 Rather than rely on cpu->mem_io_pc, pass retaddr down directly. Within tb_invalidate_phys_page_range__locked, the is_cpu_write_access parameter is non-zero exactly when retaddr would be non-zero, so that is a simple replacement. Recognize that current_tb_not_found is true only when mem_io_pc (and now retaddr) are also non-zero, so remove a redundant test. Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/translate-all.h | 3 ++- accel/tcg/cputlb.c | 6 +----- accel/tcg/translate-all.c | 39 +++++++++++++++++++-------------------- 3 files changed, 22 insertions(+), 26 deletions(-) diff --git a/accel/tcg/translate-all.h b/accel/tcg/translate-all.h index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.h +++ b/accel/tcg/translate-all.h @@ -XXX,XX +XXX,XX @@ struct page_collection *page_collection_lock(tb_page_addr_t start, tb_page_addr_t end); void page_collection_unlock(struct page_collection *set); void tb_invalidate_phys_page_fast(struct page_collection *pages, - tb_page_addr_t start, int len); + tb_page_addr_t start, int len, + uintptr_t retaddr); void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end); void tb_check_watchpoint(CPUState *cpu); diff --git a/accel/tcg/cputlb.c b/accel/tcg/cputlb.c index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/cputlb.c +++ b/accel/tcg/cputlb.c @@ -XXX,XX +XXX,XX @@ static void notdirty_write(CPUState *cpu, vaddr mem_vaddr, unsigned size, if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) { struct page_collection *pages = page_collection_lock(ram_addr, ram_addr + size); - - /* We require mem_io_pc in tb_invalidate_phys_page_range. */ - cpu->mem_io_pc = retaddr; - - tb_invalidate_phys_page_fast(pages, ram_addr, size); + tb_invalidate_phys_page_fast(pages, ram_addr, size, retaddr); page_collection_unlock(pages); } 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_invalidate_phys_page_range__locked(struct page_collection *pages, PageDesc *p, tb_page_addr_t start, tb_page_addr_t end, - int is_cpu_write_access) + uintptr_t retaddr) { TranslationBlock *tb; tb_page_addr_t tb_start, tb_end; @@ -XXX,XX +XXX,XX @@ tb_invalidate_phys_page_range__locked(struct page_collection *pages, #ifdef TARGET_HAS_PRECISE_SMC CPUState *cpu = current_cpu; CPUArchState *env = NULL; - int current_tb_not_found = is_cpu_write_access; + bool current_tb_not_found = retaddr != 0; + bool current_tb_modified = false; TranslationBlock *current_tb = NULL; - int current_tb_modified = 0; target_ulong current_pc = 0; target_ulong current_cs_base = 0; uint32_t current_flags = 0; @@ -XXX,XX +XXX,XX @@ tb_invalidate_phys_page_range__locked(struct page_collection *pages, if (!(tb_end <= start || tb_start >= end)) { #ifdef TARGET_HAS_PRECISE_SMC if (current_tb_not_found) { - current_tb_not_found = 0; - current_tb = NULL; - if (cpu->mem_io_pc) { - /* now we have a real cpu fault */ - current_tb = tcg_tb_lookup(cpu->mem_io_pc); - } + current_tb_not_found = false; + /* now we have a real cpu fault */ + current_tb = tcg_tb_lookup(retaddr); } if (current_tb == tb && (tb_cflags(current_tb) & CF_COUNT_MASK) != 1) { - /* If we are modifying the current TB, we must stop - its execution. We could be more precise by checking - that the modification is after the current PC, but it - would require a specialized function to partially - restore the CPU state */ - - current_tb_modified = 1; - cpu_restore_state_from_tb(cpu, current_tb, - cpu->mem_io_pc, true); + /* + * If we are modifying the current TB, we must stop + * its execution. We could be more precise by checking + * that the modification is after the current PC, but it + * would require a specialized function to partially + * restore the CPU state. + */ + current_tb_modified = true; + cpu_restore_state_from_tb(cpu, current_tb, retaddr, true); cpu_get_tb_cpu_state(env, &current_pc, &current_cs_base, &current_flags); } @@ -XXX,XX +XXX,XX @@ void tb_invalidate_phys_range(target_ulong start, target_ulong end) * Call with all @pages in the range [@start, @start + len[ locked. */ void tb_invalidate_phys_page_fast(struct page_collection *pages, - tb_page_addr_t start, int len) + tb_page_addr_t start, int len, + uintptr_t retaddr) { PageDesc *p; @@ -XXX,XX +XXX,XX @@ void tb_invalidate_phys_page_fast(struct page_collection *pages, } } else { do_invalidate: - tb_invalidate_phys_page_range__locked(pages, p, start, start + len, 1); + tb_invalidate_phys_page_range__locked(pages, p, start, start + len, + retaddr); } } #else -- 2.17.1 Fixes the previous TLB_WATCHPOINT patches because we are currently failing to set cpu->mem_io_pc with the call to cpu_check_watchpoint. Pass down the retaddr directly because it's readily available. Fixes: 50b107c5d61 Reviewed-by: Alex Bennée <alex.bennee@linaro.org> Reviewed-by: David Hildenbrand <david@redhat.com> Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- accel/tcg/translate-all.h | 2 +- accel/tcg/translate-all.c | 6 +++--- exec.c | 2 +- 3 files changed, 5 insertions(+), 5 deletions(-) diff --git a/accel/tcg/translate-all.h b/accel/tcg/translate-all.h index XXXXXXX..XXXXXXX 100644 --- a/accel/tcg/translate-all.h +++ b/accel/tcg/translate-all.h @@ -XXX,XX +XXX,XX @@ void tb_invalidate_phys_page_fast(struct page_collection *pages, tb_page_addr_t start, int len, uintptr_t retaddr); void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end); -void tb_check_watchpoint(CPUState *cpu); +void tb_check_watchpoint(CPUState *cpu, uintptr_t retaddr); #ifdef CONFIG_USER_ONLY int page_unprotect(target_ulong address, uintptr_t pc); 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 bool tb_invalidate_phys_page(tb_page_addr_t addr, uintptr_t pc) #endif /* user-mode: call with mmap_lock held */ -void tb_check_watchpoint(CPUState *cpu) +void tb_check_watchpoint(CPUState *cpu, uintptr_t retaddr) { TranslationBlock *tb; assert_memory_lock(); - tb = tcg_tb_lookup(cpu->mem_io_pc); + tb = tcg_tb_lookup(retaddr); if (tb) { /* We can use retranslation to find the PC. */ - cpu_restore_state_from_tb(cpu, tb, cpu->mem_io_pc, true); + cpu_restore_state_from_tb(cpu, tb, retaddr, true); tb_phys_invalidate(tb, -1); } else { /* The exception probably happened in a helper. The CPU state should diff --git a/exec.c b/exec.c index XXXXXXX..XXXXXXX 100644 --- a/exec.c +++ b/exec.c @@ -XXX,XX +XXX,XX @@ void cpu_check_watchpoint(CPUState *cpu, vaddr addr, vaddr len, cpu->watchpoint_hit = wp; mmap_lock(); - tb_check_watchpoint(cpu); + tb_check_watchpoint(cpu, ra); if (wp->flags & BP_STOP_BEFORE_ACCESS) { cpu->exception_index = EXCP_DEBUG; mmap_unlock(); -- 2.17.1 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