diff options
| author | H.J. Lu <hjl.tools@gmail.com> | 2021-02-26 05:36:59 -0800 |
|---|---|---|
| committer | H.J. Lu <hjl.tools@gmail.com> | 2022-01-27 12:47:19 -0800 |
| commit | b2e59f87ae10e5874bb150bd87f3b92b47b1090e (patch) | |
| tree | d241c960302351ce88e7577da5bea9b4c2e212b4 | |
| parent | 43c27a754bd417726032006d0a564d946994554e (diff) | |
| download | glibc-b2e59f87ae10e5874bb150bd87f3b92b47b1090e.zip glibc-b2e59f87ae10e5874bb150bd87f3b92b47b1090e.tar.gz glibc-b2e59f87ae10e5874bb150bd87f3b92b47b1090e.tar.bz2 | |
x86: Set Prefer_No_VZEROUPPER and add Prefer_AVX2_STRCMP
1. Set Prefer_No_VZEROUPPER if RTM is usable to avoid RTM abort triggered
by VZEROUPPER inside a transactionally executing RTM region.
2. Since to compare 2 32-byte strings, 256-bit EVEX strcmp requires 2
loads, 3 VPCMPs and 2 KORDs while AVX2 strcmp requires 1 load, 2 VPCMPEQs,
1 VPMINU and 1 VPMOVMSKB, AVX2 strcmp is faster than EVEX strcmp. Add
Prefer_AVX2_STRCMP to prefer AVX2 strcmp family functions.
(cherry picked from commit 1da50d4bda07f04135dca39f40e79fc9eabed1f8)
| -rw-r--r-- | sysdeps/x86/cpu-features.c | 20 | ||||
| -rw-r--r-- | sysdeps/x86/cpu-features.h | 2 | ||||
| -rw-r--r-- | sysdeps/x86/cpu-tunables.c | 3 |
3 files changed, 23 insertions, 2 deletions
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c index 4695ac8..5ddc8ee 100644 --- a/sysdeps/x86/cpu-features.c +++ b/sysdeps/x86/cpu-features.c @@ -347,8 +347,24 @@ init_cpu_features (struct cpu_features *cpu_features) cpu_features->feature[index_arch_Prefer_No_VZEROUPPER] |= bit_arch_Prefer_No_VZEROUPPER; else - cpu_features->feature[index_arch_Prefer_No_AVX512] - |= bit_arch_Prefer_No_AVX512; + { + cpu_features->feature[index_arch_Prefer_No_AVX512] + |= bit_arch_Prefer_No_AVX512; + + /* Avoid RTM abort triggered by VZEROUPPER inside a + transactionally executing RTM region. */ + if (CPU_FEATURES_CPU_P (cpu_features, RTM)) + cpu_features->feature[index_arch_Prefer_No_VZEROUPPER] + |= bit_arch_Prefer_No_VZEROUPPER; + + /* Since to compare 2 32-byte strings, 256-bit EVEX strcmp + requires 2 loads, 3 VPCMPs and 2 KORDs while AVX2 strcmp + requires 1 load, 2 VPCMPEQs, 1 VPMINU and 1 VPMOVMSKB, + AVX2 strcmp is faster than EVEX strcmp. */ + if (CPU_FEATURES_ARCH_P (cpu_features, AVX2_Usable)) + cpu_features->feature[index_arch_Prefer_AVX2_STRCMP] + |= bit_arch_Prefer_AVX2_STRCMP; + } } /* This spells out "AuthenticAMD". */ else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65) diff --git a/sysdeps/x86/cpu-features.h b/sysdeps/x86/cpu-features.h index 347a4b1..1f60296 100644 --- a/sysdeps/x86/cpu-features.h +++ b/sysdeps/x86/cpu-features.h @@ -41,6 +41,7 @@ #define bit_arch_MathVec_Prefer_No_AVX512 (1 << 21) #define bit_arch_XSAVEC_Usable (1 << 22) #define bit_arch_Prefer_FSRM (1 << 23) +#define bit_arch_Prefer_AVX2_STRCMP (1 << 24) /* CPUID Feature flags. */ @@ -266,6 +267,7 @@ extern const struct cpu_features *__get_cpu_features (void) # define index_arch_MathVec_Prefer_No_AVX512 FEATURE_INDEX_1 # define index_arch_XSAVEC_Usable FEATURE_INDEX_1 # define index_arch_Prefer_FSRM FEATURE_INDEX_1 +# define index_arch_Prefer_AVX2_STRCMP FEATURE_INDEX_1 #endif /* !__ASSEMBLER__ */ diff --git a/sysdeps/x86/cpu-tunables.c b/sysdeps/x86/cpu-tunables.c index 69155a8..824ee76 100644 --- a/sysdeps/x86/cpu-tunables.c +++ b/sysdeps/x86/cpu-tunables.c @@ -282,6 +282,9 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp) CHECK_GLIBC_IFUNC_ARCH_BOTH (n, cpu_features, Fast_Copy_Backward, disable, 18); + CHECK_GLIBC_IFUNC_ARCH_NEED_ARCH_BOTH + (n, cpu_features, Prefer_AVX2_STRCMP, AVX2_Usable, + disable, 18); } break; case 19: |