| Message ID | 20190214040652.4811-5-richard.henderson@linaro.org |
|---|---|
| State | New |
| Headers | show |
| Series | target/arm: Reduce overhead of cpu_get_tb_cpu_state | expand |
Richard Henderson <richard.henderson@linaro.org> writes: > This is the payoff. \o/ Running my own silly pigz benchmark: Before: Time (mean ± σ): 66.035 s ± 1.425 s After: Time (mean ± σ): 57.191 s ± 0.675 s > > From perf record -g data of ubuntu 18 boot and shutdown: > > BEFORE: > > - 23.02% 2.82% qemu-system-aar [.] helper_lookup_tb_ptr > - 20.22% helper_lookup_tb_ptr > + 10.05% tb_htable_lookup > - 9.13% cpu_get_tb_cpu_state > 3.20% aa64_va_parameters_both > 0.55% fp_exception_el > > - 11.66% 4.74% qemu-system-aar [.] cpu_get_tb_cpu_state > - 6.96% cpu_get_tb_cpu_state > 3.63% aa64_va_parameters_both > 0.60% fp_exception_el > 0.53% sve_exception_el > > AFTER: > > - 16.40% 3.40% qemu-system-aar [.] helper_lookup_tb_ptr > - 13.03% helper_lookup_tb_ptr > + 11.19% tb_htable_lookup > 0.55% cpu_get_tb_cpu_state > > 0.98% 0.71% qemu-system-aar [.] cpu_get_tb_cpu_state > > 0.87% 0.24% qemu-system-aar [.] rebuild_hflags_a64 > > Before, helper_lookup_tb_ptr is the second hottest function in the > application, consuming almost a quarter of the runtime. Within the > entire execution, cpu_get_tb_cpu_state consumes about 12%. > > After, helper_lookup_tb_ptr has dropped to the fourth hottest function, > with consumption dropping to a sixth of the runtime. Within the > entire execution, cpu_get_tb_cpu_state has dropped below 1%, and the > supporting function to rebuild hflags also consumes about 1%. > > Signed-off-by: Richard Henderson <richard.henderson@linaro.org> > --- > target/arm/helper.c | 9 ++------- > 1 file changed, 2 insertions(+), 7 deletions(-) > > diff --git a/target/arm/helper.c b/target/arm/helper.c > index 3c8724883d..1bdb87267e 100644 > --- a/target/arm/helper.c > +++ b/target/arm/helper.c > @@ -13894,21 +13894,16 @@ void HELPER(rebuild_hflags_a64)(CPUARMState *env, uint32_t el) > void cpu_get_tb_cpu_state(CPUARMState *env, target_ulong *pc, > target_ulong *cs_base, uint32_t *pflags) > { > - int current_el = arm_current_el(env); > - uint32_t flags; > + uint32_t flags = env->hflags; > uint32_t pstate_for_ss; > > *cs_base = 0; > - if (is_a64(env)) { > + if (FIELD_EX32(flags, TBFLAG_ANY, AARCH64_STATE)) { > *pc = env->pc; > - flags = rebuild_hflags_a64(env, current_el); > - assert(flags == env->hflags); While debugging I came up with this monstrosity: if (FIELD_EX32(flags, TBFLAG_ANY, AARCH64_STATE)) { #ifdef CONFIG_DEBUG_TCG static uint32_t tb_state = 0; uint32_t recalc_flags = rebuild_hflags_a64(env, arm_current_el(env)); tb_state++; if (flags != recalc_flags) { fprintf(stderr, "%s: flags %#x, should be %#x (%#x/%d)\n", __func__, flags, recalc_flags, flags ^ recalc_flags, tb_state); abort(); } #endif *pc = env->pc; flags = FIELD_DP32(flags, TBFLAG_A64, BTYPE, env->btype); pstate_for_ss = env->pstate; } else { I suspect given cached flags are common about the translator it might be worth coming up with a nicer macro to encapsulate the check for DEBUG_TCG builds. > flags = FIELD_DP32(flags, TBFLAG_A64, BTYPE, env->btype); > pstate_for_ss = env->pstate; > } else { > *pc = env->regs[15]; > - flags = rebuild_hflags_a32(env, current_el); > - assert(flags == env->hflags); > flags = FIELD_DP32(flags, TBFLAG_A32, THUMB, env->thumb); > flags = FIELD_DP32(flags, TBFLAG_A32, CONDEXEC, env->condexec_bits); > pstate_for_ss = env->uncached_cpsr; -- Alex Bennée
On 2/19/19 12:17 PM, Alex Bennée wrote: > While debugging I came up with this monstrosity: > > if (FIELD_EX32(flags, TBFLAG_ANY, AARCH64_STATE)) { > #ifdef CONFIG_DEBUG_TCG > static uint32_t tb_state = 0; > uint32_t recalc_flags = rebuild_hflags_a64(env, arm_current_el(env)); > tb_state++; > if (flags != recalc_flags) { > fprintf(stderr, "%s: flags %#x, should be %#x (%#x/%d)\n", __func__, > flags, recalc_flags, flags ^ recalc_flags, tb_state); > abort(); > } > #endif > *pc = env->pc; > flags = FIELD_DP32(flags, TBFLAG_A64, BTYPE, env->btype); > pstate_for_ss = env->pstate; > } else { I have now included +#ifdef CONFIG_DEBUG_TCG + { + int el = arm_current_el(env); + uint32_t check_flags; + if (is_a64(env)) { + check_flags = rebuild_hflags_a64(env, el); + } else { + check_flags = rebuild_hflags_a32(env, el); + } + g_assert_cmphex(flags, ==, check_flags); + } +#endif r~
diff --git a/target/arm/helper.c b/target/arm/helper.c index 3c8724883d..1bdb87267e 100644 --- a/target/arm/helper.c +++ b/target/arm/helper.c @@ -13894,21 +13894,16 @@ void HELPER(rebuild_hflags_a64)(CPUARMState *env, uint32_t el) void cpu_get_tb_cpu_state(CPUARMState *env, target_ulong *pc, target_ulong *cs_base, uint32_t *pflags) { - int current_el = arm_current_el(env); - uint32_t flags; + uint32_t flags = env->hflags; uint32_t pstate_for_ss; *cs_base = 0; - if (is_a64(env)) { + if (FIELD_EX32(flags, TBFLAG_ANY, AARCH64_STATE)) { *pc = env->pc; - flags = rebuild_hflags_a64(env, current_el); - assert(flags == env->hflags); flags = FIELD_DP32(flags, TBFLAG_A64, BTYPE, env->btype); pstate_for_ss = env->pstate; } else { *pc = env->regs[15]; - flags = rebuild_hflags_a32(env, current_el); - assert(flags == env->hflags); flags = FIELD_DP32(flags, TBFLAG_A32, THUMB, env->thumb); flags = FIELD_DP32(flags, TBFLAG_A32, CONDEXEC, env->condexec_bits); pstate_for_ss = env->uncached_cpsr;
This is the payoff. From perf record -g data of ubuntu 18 boot and shutdown: BEFORE: - 23.02% 2.82% qemu-system-aar [.] helper_lookup_tb_ptr - 20.22% helper_lookup_tb_ptr + 10.05% tb_htable_lookup - 9.13% cpu_get_tb_cpu_state 3.20% aa64_va_parameters_both 0.55% fp_exception_el - 11.66% 4.74% qemu-system-aar [.] cpu_get_tb_cpu_state - 6.96% cpu_get_tb_cpu_state 3.63% aa64_va_parameters_both 0.60% fp_exception_el 0.53% sve_exception_el AFTER: - 16.40% 3.40% qemu-system-aar [.] helper_lookup_tb_ptr - 13.03% helper_lookup_tb_ptr + 11.19% tb_htable_lookup 0.55% cpu_get_tb_cpu_state 0.98% 0.71% qemu-system-aar [.] cpu_get_tb_cpu_state 0.87% 0.24% qemu-system-aar [.] rebuild_hflags_a64 Before, helper_lookup_tb_ptr is the second hottest function in the application, consuming almost a quarter of the runtime. Within the entire execution, cpu_get_tb_cpu_state consumes about 12%. After, helper_lookup_tb_ptr has dropped to the fourth hottest function, with consumption dropping to a sixth of the runtime. Within the entire execution, cpu_get_tb_cpu_state has dropped below 1%, and the supporting function to rebuild hflags also consumes about 1%. Signed-off-by: Richard Henderson <richard.henderson@linaro.org> --- target/arm/helper.c | 9 ++------- 1 file changed, 2 insertions(+), 7 deletions(-) -- 2.17.1