|
 |
190885 |
From e7e0ac928b21e5f47e5b648723851c5270db24f2 Mon Sep 17 00:00:00 2001
|
|
|
190885 |
From: Noah Goldstein <goldstein.w.n@gmail.com>
|
|
|
190885 |
Date: Sat, 23 Oct 2021 01:26:47 -0400
|
|
|
190885 |
Subject: [PATCH] x86: Replace sse2 instructions with avx in
|
|
|
190885 |
memcmp-evex-movbe.S
|
|
|
190885 |
|
|
|
190885 |
This commit replaces two usages of SSE2 'movups' with AVX 'vmovdqu'.
|
|
|
190885 |
|
|
|
190885 |
it could potentially be dangerous to use SSE2 if this function is ever
|
|
|
190885 |
called without using 'vzeroupper' beforehand. While compilers appear
|
|
|
190885 |
to use 'vzeroupper' before function calls if AVX2 has been used, using
|
|
|
190885 |
SSE2 here is more brittle. Since it is not absolutely necessary it
|
|
|
190885 |
should be avoided.
|
|
|
190885 |
|
|
|
190885 |
It costs 2-extra bytes but the extra bytes should only eat into
|
|
|
190885 |
alignment padding.
|
|
|
190885 |
Reviewed-by: H.J. Lu <hjl.tools@gmail.com>
|
|
|
190885 |
|
|
|
190885 |
(cherry picked from commit bad852b61b79503fcb3c5fc379c70f768df3e1fb)
|
|
|
190885 |
---
|
|
|
190885 |
sysdeps/x86_64/multiarch/memcmp-evex-movbe.S | 4 ++--
|
|
|
190885 |
1 file changed, 2 insertions(+), 2 deletions(-)
|
|
|
190885 |
|
|
|
190885 |
diff --git a/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
|
|
|
190885 |
index 2761b54f..640f6757 100644
|
|
|
190885 |
--- a/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
|
|
|
190885 |
+++ b/sysdeps/x86_64/multiarch/memcmp-evex-movbe.S
|
|
|
190885 |
@@ -561,13 +561,13 @@ L(between_16_31):
|
|
|
190885 |
/* From 16 to 31 bytes. No branch when size == 16. */
|
|
|
190885 |
|
|
|
190885 |
/* Use movups to save code size. */
|
|
|
190885 |
- movups (%rsi), %xmm2
|
|
|
190885 |
+ vmovdqu (%rsi), %xmm2
|
|
|
190885 |
VPCMP $4, (%rdi), %xmm2, %k1
|
|
|
190885 |
kmovd %k1, %eax
|
|
|
190885 |
testl %eax, %eax
|
|
|
190885 |
jnz L(return_vec_0_lv)
|
|
|
190885 |
/* Use overlapping loads to avoid branches. */
|
|
|
190885 |
- movups -16(%rsi, %rdx, CHAR_SIZE), %xmm2
|
|
|
190885 |
+ vmovdqu -16(%rsi, %rdx, CHAR_SIZE), %xmm2
|
|
|
190885 |
VPCMP $4, -16(%rdi, %rdx, CHAR_SIZE), %xmm2, %k1
|
|
|
190885 |
addl $(CHAR_PER_VEC - (16 / CHAR_SIZE)), %edx
|
|
|
190885 |
kmovd %k1, %eax
|
|
|
190885 |
--
|
|
|
190885 |
GitLab
|
|
|
190885 |
|