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> | 2021-03-29 07:40:17 -0700 |
| commit | 1da50d4bda07f04135dca39f40e79fc9eabed1f8 (patch) | |
| tree | bf3a3ffac8850fffcb22dfb889374349f66136b1 | |
| parent | 934d88d862627a0ded8955985b43f5f12aa38046 (diff) | |
| download | glibc-1da50d4bda07f04135dca39f40e79fc9eabed1f8.zip glibc-1da50d4bda07f04135dca39f40e79fc9eabed1f8.tar.gz glibc-1da50d4bda07f04135dca39f40e79fc9eabed1f8.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.
| -rw-r--r-- | sysdeps/x86/cpu-features.c | 20 | ||||
| -rw-r--r-- | sysdeps/x86/cpu-tunables.c | 2 | ||||
| -rw-r--r-- | sysdeps/x86/include/cpu-features-preferred_feature_index_1.def | 1 |
3 files changed, 21 insertions, 2 deletions
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c index e99e67f..81275db 100644 --- a/sysdeps/x86/cpu-features.c +++ b/sysdeps/x86/cpu-features.c @@ -531,8 +531,24 @@ init_cpu_features (struct cpu_features *cpu_features) cpu_features->preferred[index_arch_Prefer_No_VZEROUPPER] |= bit_arch_Prefer_No_VZEROUPPER; else - cpu_features->preferred[index_arch_Prefer_No_AVX512] - |= bit_arch_Prefer_No_AVX512; + { + cpu_features->preferred[index_arch_Prefer_No_AVX512] + |= bit_arch_Prefer_No_AVX512; + + /* Avoid RTM abort triggered by VZEROUPPER inside a + transactionally executing RTM region. */ + if (CPU_FEATURE_USABLE_P (cpu_features, RTM)) + cpu_features->preferred[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_FEATURE_USABLE_P (cpu_features, AVX2)) + cpu_features->preferred[index_arch_Prefer_AVX2_STRCMP] + |= bit_arch_Prefer_AVX2_STRCMP; + } } /* This spells out "AuthenticAMD" or "HygonGenuine". */ else if ((ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65) diff --git a/sysdeps/x86/cpu-tunables.c b/sysdeps/x86/cpu-tunables.c index 61b05e5..00fe504 100644 --- a/sysdeps/x86/cpu-tunables.c +++ b/sysdeps/x86/cpu-tunables.c @@ -239,6 +239,8 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp) CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features, Fast_Copy_Backward, disable, 18); + CHECK_GLIBC_IFUNC_PREFERRED_NEED_BOTH + (n, cpu_features, Prefer_AVX2_STRCMP, AVX2, disable, 18); } break; case 19: diff --git a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def index 06af1a8..133aab1 100644 --- a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def +++ b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def @@ -32,3 +32,4 @@ BIT (Prefer_ERMS) BIT (Prefer_No_AVX512) BIT (MathVec_Prefer_No_AVX512) BIT (Prefer_FSRM) +BIT (Prefer_AVX2_STRCMP) |