|
 |
190885 |
From 69ce2769234e36e62c2db82e1e638124f076e82e Mon Sep 17 00:00:00 2001
|
|
|
190885 |
From: "H.J. Lu" <hjl.tools@gmail.com>
|
|
|
190885 |
Date: Wed, 2 Mar 2022 15:42:47 -0800
|
|
|
190885 |
Subject: [PATCH] x86-64: Remove Prefer_AVX2_STRCMP
|
|
|
190885 |
|
|
|
190885 |
Remove Prefer_AVX2_STRCMP to enable EVEX strcmp. When comparing 2 32-byte
|
|
|
190885 |
strings, EVEX strcmp has been improved to require 1 load, 1 VPTESTM, 1
|
|
|
190885 |
VPCMP, 1 KMOVD and 1 INCL instead of 2 loads, 3 VPCMPs, 2 KORDs, 1 KMOVD
|
|
|
190885 |
and 1 TESTL while AVX2 strcmp requires 1 load, 2 VPCMPEQs, 1 VPMINU, 1
|
|
|
190885 |
VPMOVMSKB and 1 TESTL. EVEX strcmp is now faster than AVX2 strcmp by up
|
|
|
190885 |
to 40% on Tiger Lake and Ice Lake.
|
|
|
190885 |
|
|
|
190885 |
(cherry picked from commit 14dbbf46a007ae5df36646b51ad0c9e5f5259f30)
|
|
|
190885 |
---
|
|
|
190885 |
sysdeps/x86/cpu-features.c | 8 --------
|
|
|
190885 |
sysdeps/x86/cpu-tunables.c | 2 --
|
|
|
190885 |
.../include/cpu-features-preferred_feature_index_1.def | 1 -
|
|
|
190885 |
sysdeps/x86_64/multiarch/strcmp.c | 3 +--
|
|
|
190885 |
sysdeps/x86_64/multiarch/strncmp.c | 3 +--
|
|
|
190885 |
5 files changed, 2 insertions(+), 15 deletions(-)
|
|
|
190885 |
|
|
|
190885 |
diff --git a/sysdeps/x86/cpu-features.c b/sysdeps/x86/cpu-features.c
|
|
|
190885 |
index 8885b48e..956bfb4f 100644
|
|
|
190885 |
--- a/sysdeps/x86/cpu-features.c
|
|
|
190885 |
+++ b/sysdeps/x86/cpu-features.c
|
|
|
190885 |
@@ -533,14 +533,6 @@ init_cpu_features (struct cpu_features *cpu_features)
|
|
|
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 |
diff --git a/sysdeps/x86/cpu-tunables.c b/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
index 73adbaba..3173b2b9 100644
|
|
|
190885 |
--- a/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
+++ b/sysdeps/x86/cpu-tunables.c
|
|
|
190885 |
@@ -239,8 +239,6 @@ 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 f2340624..e7277b33 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 |
@@ -32,5 +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 |
BIT (Avoid_Short_Distance_REP_MOVSB)
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strcmp.c b/sysdeps/x86_64/multiarch/strcmp.c
|
|
|
190885 |
index 11bbea2b..f8a7220e 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/strcmp.c
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strcmp.c
|
|
|
190885 |
@@ -43,8 +43,7 @@ IFUNC_SELECTOR (void)
|
|
|
190885 |
{
|
|
|
190885 |
if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
|
|
|
190885 |
&& CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
|
|
|
190885 |
- && CPU_FEATURE_USABLE_P (cpu_features, BMI2)
|
|
|
190885 |
- && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_AVX2_STRCMP))
|
|
|
190885 |
+ && CPU_FEATURE_USABLE_P (cpu_features, BMI2))
|
|
|
190885 |
return OPTIMIZE (evex);
|
|
|
190885 |
|
|
|
190885 |
if (CPU_FEATURE_USABLE_P (cpu_features, RTM))
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/strncmp.c b/sysdeps/x86_64/multiarch/strncmp.c
|
|
|
190885 |
index 44c85116..7903f807 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/strncmp.c
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/strncmp.c
|
|
|
190885 |
@@ -43,8 +43,7 @@ IFUNC_SELECTOR (void)
|
|
|
190885 |
{
|
|
|
190885 |
if (CPU_FEATURE_USABLE_P (cpu_features, AVX512VL)
|
|
|
190885 |
&& CPU_FEATURE_USABLE_P (cpu_features, AVX512BW)
|
|
|
190885 |
- && CPU_FEATURE_USABLE_P (cpu_features, BMI2)
|
|
|
190885 |
- && !CPU_FEATURES_ARCH_P (cpu_features, Prefer_AVX2_STRCMP))
|
|
|
190885 |
+ && CPU_FEATURE_USABLE_P (cpu_features, BMI2))
|
|
|
190885 |
return OPTIMIZE (evex);
|
|
|
190885 |
|
|
|
190885 |
if (CPU_FEATURE_USABLE_P (cpu_features, RTM))
|
|
|
190885 |
--
|
|
|
190885 |
GitLab
|
|
|
190885 |
|