|
 |
190885 |
From 18a0eba6e84524e3a1cbb544ed1f53dce9531b46 Mon Sep 17 00:00:00 2001
|
|
|
190885 |
From: "H.J. Lu" <hjl.tools@gmail.com>
|
|
|
190885 |
Date: Wed, 2 Mar 2022 12:52:23 -0800
|
|
|
190885 |
Subject: [PATCH] x86: Set Prefer_No_VZEROUPPER and add Prefer_AVX2_STRCMP
|
|
|
190885 |
|
|
|
190885 |
1. Set Prefer_No_VZEROUPPER if RTM is usable to avoid RTM abort triggered
|
|
|
190885 |
by VZEROUPPER inside a transactionally executing RTM region.
|
|
|
190885 |
2. Since to compare 2 32-byte strings, 256-bit EVEX strcmp requires 2
|
|
|
190885 |
loads, 3 VPCMPs and 2 KORDs while AVX2 strcmp requires 1 load, 2 VPCMPEQs,
|
|
|
190885 |
1 VPMINU and 1 VPMOVMSKB, AVX2 strcmp is faster than EVEX strcmp. Add
|
|
|
190885 |
Prefer_AVX2_STRCMP to prefer AVX2 strcmp family functions.
|
|
|
190885 |
|
|
|
190885 |
(cherry picked from commit 1da50d4bda07f04135dca39f40e79fc9eabed1f8)
|
|
|
190885 |
---
|
|
|
190885 |
sysdeps/x86/cpu-features.c | 20 +++++++++++++++++--
|
|
|
190885 |
sysdeps/x86/cpu-tunables.c | 2 ++
|
|
|
190885 |
...cpu-features-preferred_feature_index_1.def | 1 +
|
|
|
190885 |
3 files changed, 21 insertions(+), 2 deletions(-)
|
|
|
190885 |
|
|
|
190885 |
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
|
|
|
190885 |
index 91042505..3610ee5c 100644
|
|
|
190885 |
--- a/sysdeps/x86/cpu-features.c
|
|
|
190885 |
+++ b/sysdeps/x86/cpu-features.c
|
|
|
190885 |
@@ -524,8 +524,24 @@ init_cpu_features (struct cpu_features *cpu_features)
|
|
|
190885 |
cpu_features->preferred[index_arch_Prefer_No_VZEROUPPER]
|
|
|
190885 |
|= bit_arch_Prefer_No_VZEROUPPER;
|
|
|
190885 |
else
|
|
|
190885 |
- cpu_features->preferred[index_arch_Prefer_No_AVX512]
|
|
|
190885 |
- |= bit_arch_Prefer_No_AVX512;
|
|
|
190885 |
+ {
|
|
|
190885 |
+ cpu_features->preferred[index_arch_Prefer_No_AVX512]
|
|
|
190885 |
+ |= bit_arch_Prefer_No_AVX512;
|
|
|
190885 |
+
|
|
|
190885 |
+ /* Avoid RTM abort triggered by VZEROUPPER inside a
|
|
|
190885 |
+ transactionally executing RTM region. */
|
|
|
190885 |
+ if (CPU_FEATURE_USABLE_P (cpu_features, RTM))
|
|
|
190885 |
+ cpu_features->preferred[index_arch_Prefer_No_VZEROUPPER]
|
|
|
190885 |
+ |= bit_arch_Prefer_No_VZEROUPPER;
|
|
|
190885 |
+
|
|
|
190885 |
+ /* Since to compare 2 32-byte strings, 256-bit EVEX strcmp
|
|
|
190885 |
+ requires 2 loads, 3 VPCMPs and 2 KORDs while AVX2 strcmp
|
|
|
190885 |
+ requires 1 load, 2 VPCMPEQs, 1 VPMINU and 1 VPMOVMSKB,
|
|
|
190885 |
+ AVX2 strcmp is faster than EVEX strcmp. */
|
|
|
190885 |
+ if (CPU_FEATURE_USABLE_P (cpu_features, AVX2))
|
|
|
190885 |
+ cpu_features->preferred[index_arch_Prefer_AVX2_STRCMP]
|
|
|
190885 |
+ |= bit_arch_Prefer_AVX2_STRCMP;
|
|
|
190885 |
+ }
|
|
|
190885 |
}
|
|
|
190885 |
/* This spells out "AuthenticAMD". */
|
|
|
190885 |
else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
|
|
|
190885 |
diff --git a/sysdeps/x86/cpu-tunables.c b/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
index 3173b2b9..73adbaba 100644
|
|
|
190885 |
--- a/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
+++ b/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
@@ -239,6 +239,8 @@ TUNABLE_CALLBACK (set_hwcaps) (tunable_val_t *valp)
|
|
|
190885 |
CHECK_GLIBC_IFUNC_PREFERRED_BOTH (n, cpu_features,
|
|
|
190885 |
Fast_Copy_Backward,
|
|
|
190885 |
disable, 18);
|
|
|
190885 |
+ CHECK_GLIBC_IFUNC_PREFERRED_NEED_BOTH
|
|
|
190885 |
+ (n, cpu_features, Prefer_AVX2_STRCMP, AVX2, disable, 18);
|
|
|
190885 |
}
|
|
|
190885 |
break;
|
|
|
190885 |
case 19:
|
|
|
190885 |
diff --git a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
|
|
|
190885 |
index 17a5cc42..4ca70b40 100644
|
|
|
190885 |
--- a/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
|
|
|
190885 |
+++ b/sysdeps/x86/include/cpu-features-preferred_feature_index_1.def
|
|
|
190885 |
@@ -32,3 +32,4 @@ BIT (Prefer_ERMS)
|
|
|
190885 |
BIT (Prefer_FSRM)
|
|
|
190885 |
BIT (Prefer_No_AVX512)
|
|
|
190885 |
BIT (MathVec_Prefer_No_AVX512)
|
|
|
190885 |
+BIT (Prefer_AVX2_STRCMP)
|
|
|
190885 |
--
|
|
|
190885 |
GitLab
|
|
|
190885 |
|