Instead of passing a bool and select a value within dcbz_common() let the callers pass in the right value to avoid this conditional statement. On PPC dcbz is often used to zero memory and some code uses it a lot. This change improves the run time of a test case that copies memory with a dcbz call in every iteration from 6.23 to 5.83 seconds. Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> --- This is just a small optimisation removing some of the overhead but dcbz still seems to be the biggest issue with this test. Removing the dcbz call it runs in 2 seconds. In a profile I see: Children Self Command Shared Object Symbol - 55.01% 11.44% qemu-ppc qemu-ppc [.] dcbz_common.constprop.0 - 43.57% dcbz_common.constprop.0 - probe_access - page_get_flags interval_tree_iter_first - 11.44% helper_raise_exception_err cpu_loop_exit_restore cpu_loop cpu_exec cpu_exec_setjmp.isra.0 cpu_exec_loop.constprop.0 cpu_tb_exec 0x7f262403636e helper_raise_exception_err cpu_loop_exit_restore cpu_loop cpu_exec cpu_exec_setjmp.isra.0 cpu_exec_loop.constprop.0 cpu_tb_exec - 0x7f26240386a4 11.20% helper_dcbz + 43.81% 12.28% qemu-ppc qemu-ppc [.] probe_access + 39.31% 0.00% qemu-ppc [JIT] tid 9969 [.] 0x00007f2624000000 + 32.45% 4.51% qemu-ppc qemu-ppc [.] page_get_flags + 25.50% 2.10% qemu-ppc qemu-ppc [.] interval_tree_iter_first + 24.67% 24.67% qemu-ppc qemu-ppc [.] interval_tree_subtree_search + 16.75% 1.19% qemu-ppc qemu-ppc [.] helper_dcbz + 4.78% 4.78% qemu-ppc [JIT] tid 9969 [.] 0x00007f26240386be + 3.46% 3.46% qemu-ppc libc-2.32.so [.] __memset_avx2_unaligned_erms Any idea how this could be optimised further? (This is running with qemu-ppc user mode emulation but I think with system it might be even worse.) Could an inline implementation with TCG vector ops work to avoid the helper and let it compile to efficient host code? Even if that could work I don't know how to do that so I'd need some further advice on this. target/ppc/mem_helper.c | 7 +++---- 1 file changed, 3 insertions(+), 4 deletions(-) diff --git a/target/ppc/mem_helper.c b/target/ppc/mem_helper.c index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/mem_helper.c +++ b/target/ppc/mem_helper.c @@ -XXX,XX +XXX,XX @@ void helper_stsw(CPUPPCState *env, target_ulong addr, uint32_t nb, } static void dcbz_common(CPUPPCState *env, target_ulong addr, - uint32_t opcode, bool epid, uintptr_t retaddr) + uint32_t opcode, int mmu_idx, uintptr_t retaddr) { target_ulong mask, dcbz_size = env->dcache_line_size; uint32_t i; void *haddr; - int mmu_idx = epid ? PPC_TLB_EPID_STORE : ppc_env_mmu_index(env, false); #if defined(TARGET_PPC64) /* Check for dcbz vs dcbzl on 970 */ @@ -XXX,XX +XXX,XX @@ static void dcbz_common(CPUPPCState *env, target_ulong addr, void helper_dcbz(CPUPPCState *env, target_ulong addr, uint32_t opcode) { - dcbz_common(env, addr, opcode, false, GETPC()); + dcbz_common(env, addr, opcode, ppc_env_mmu_index(env, false), GETPC()); } void helper_dcbzep(CPUPPCState *env, target_ulong addr, uint32_t opcode) { - dcbz_common(env, addr, opcode, true, GETPC()); + dcbz_common(env, addr, opcode, PPC_TLB_EPID_STORE, GETPC()); } void helper_icbi(CPUPPCState *env, target_ulong addr) -- 2.30.9 This is an updated version of this patch as suggested by Richard but it runs slower and only gets 5.9 seconds instead of 5.83-5.81 with v1 so this is not for merge, only for reference in case it can be useful for further optimisation or can be fixed in some way. Signed-off-by: BALATON Zoltan <balaton@eik.bme.hu> --- target/ppc/helper.h | 3 +-- target/ppc/mem_helper.c | 16 +++------------- target/ppc/translate.c | 18 ++++++++---------- 3 files changed, 12 insertions(+), 25 deletions(-) diff --git a/target/ppc/helper.h b/target/ppc/helper.h index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/helper.h +++ b/target/ppc/helper.h @@ -XXX,XX +XXX,XX @@ DEF_HELPER_FLAGS_3(stmw, TCG_CALL_NO_WG, void, env, tl, i32) DEF_HELPER_4(lsw, void, env, tl, i32, i32) DEF_HELPER_5(lswx, void, env, tl, i32, i32, i32) DEF_HELPER_FLAGS_4(stsw, TCG_CALL_NO_WG, void, env, tl, i32, i32) -DEF_HELPER_FLAGS_3(dcbz, TCG_CALL_NO_WG, void, env, tl, i32) -DEF_HELPER_FLAGS_3(dcbzep, TCG_CALL_NO_WG, void, env, tl, i32) +DEF_HELPER_FLAGS_4(dcbz, TCG_CALL_NO_WG, void, env, tl, i32, i32) DEF_HELPER_FLAGS_2(icbi, TCG_CALL_NO_WG, void, env, tl) DEF_HELPER_FLAGS_2(icbiep, TCG_CALL_NO_WG, void, env, tl) DEF_HELPER_5(lscbx, tl, env, tl, i32, i32, i32) diff --git a/target/ppc/mem_helper.c b/target/ppc/mem_helper.c index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/mem_helper.c +++ b/target/ppc/mem_helper.c @@ -XXX,XX +XXX,XX @@ void helper_stsw(CPUPPCState *env, target_ulong addr, uint32_t nb, } } -static void dcbz_common(CPUPPCState *env, target_ulong addr, - uint32_t opcode, bool epid, uintptr_t retaddr) +void helper_dcbz(CPUPPCState *env, target_ulong addr, uint32_t opcode, + uint32_t mmu_idx) { target_ulong mask, dcbz_size = env->dcache_line_size; uint32_t i; void *haddr; - int mmu_idx = epid ? PPC_TLB_EPID_STORE : ppc_env_mmu_index(env, false); + uintptr_t retaddr = GETPC(); #if defined(TARGET_PPC64) /* Check for dcbz vs dcbzl on 970 */ @@ -XXX,XX +XXX,XX @@ static void dcbz_common(CPUPPCState *env, target_ulong addr, } } -void helper_dcbz(CPUPPCState *env, target_ulong addr, uint32_t opcode) -{ - dcbz_common(env, addr, opcode, false, GETPC()); -} - -void helper_dcbzep(CPUPPCState *env, target_ulong addr, uint32_t opcode) -{ - dcbz_common(env, addr, opcode, true, GETPC()); -} - void helper_icbi(CPUPPCState *env, target_ulong addr) { addr &= ~(env->dcache_line_size - 1); diff --git a/target/ppc/translate.c b/target/ppc/translate.c index XXXXXXX..XXXXXXX 100644 --- a/target/ppc/translate.c +++ b/target/ppc/translate.c @@ -XXX,XX +XXX,XX @@ static void gen_dcblc(DisasContext *ctx) /* dcbz */ static void gen_dcbz(DisasContext *ctx) { - TCGv tcgv_addr; - TCGv_i32 tcgv_op; + TCGv tcgv_addr = tcg_temp_new(); gen_set_access_type(ctx, ACCESS_CACHE); - tcgv_addr = tcg_temp_new(); - tcgv_op = tcg_constant_i32(ctx->opcode & 0x03FF000); gen_addr_reg_index(ctx, tcgv_addr); - gen_helper_dcbz(tcg_env, tcgv_addr, tcgv_op); + gen_helper_dcbz(tcg_env, tcgv_addr, + tcg_constant_i32(ctx->opcode & 0x03FF000), + tcg_constant_i32(ctx->mem_idx)); } /* dcbzep */ static void gen_dcbzep(DisasContext *ctx) { - TCGv tcgv_addr; - TCGv_i32 tcgv_op; + TCGv tcgv_addr = tcg_temp_new(); gen_set_access_type(ctx, ACCESS_CACHE); - tcgv_addr = tcg_temp_new(); - tcgv_op = tcg_constant_i32(ctx->opcode & 0x03FF000); gen_addr_reg_index(ctx, tcgv_addr); - gen_helper_dcbzep(tcg_env, tcgv_addr, tcgv_op); + gen_helper_dcbz(tcg_env, tcgv_addr, + tcg_constant_i32(ctx->opcode & 0x03FF000), + tcg_constant_i32(PPC_TLB_EPID_STORE)); } /* dst / dstt */ -- 2.30.9