[PATCH] Fix up -funsigned-char behavior of _mm256_cmpgt_epi8 (PR target/91704)

[Jakub Jelinek <jakub at redhat dot com>]

Hi!

This PR is a repetition of PR87853, just for avx2 instead of sse2.
See https://gcc.gnu.org/ml/gcc-patches/2018-11/msg00195.html
for the previous patch.

This time there is just one intrinsic with the problem (note, the previous
patch didn't have to change _mm_cmpeq_epi8, as __v16qi vs. __v16qs doesn't
make any difference for equality comparisons).

I've grepped for similar issues in other headers, including 512-byte
vectors, but couldn't find any, in those cases we use a builtin with a mask
argument.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and
release branches?  I've also noticed PR87853 didn't come up with a testcase,
so that is also attached, ok to commit too?

As for tests, I chose not to do a dg-do run test with -funsigned-char,
because that option is an ABI change and when including some headers that
also include system headers one is never sure what will become of that.

2019-09-09  Jakub Jelinek  <jakub@redhat.com>

	PR target/91704
	* config/i386/avxintrin.h (__v32qs): New typedef.
	* config/i386/avx2intrin.h (_mm256_cmpgt_epi8): Use casts to __v32qs
	instead of __v32qi.

	* gcc.target/i386/pr91704.c: New test.

--- gcc/config/i386/avxintrin.h.jj	2019-08-12 17:55:19.039139772 +0200
+++ gcc/config/i386/avxintrin.h	2019-09-08 23:22:11.829573162 +0200
@@ -47,6 +47,7 @@ typedef unsigned int __v8su __attribute_
 typedef short __v16hi __attribute__ ((__vector_size__ (32)));
 typedef unsigned short __v16hu __attribute__ ((__vector_size__ (32)));
 typedef char __v32qi __attribute__ ((__vector_size__ (32)));
+typedef signed char __v32qs __attribute__ ((__vector_size__ (32)));
 typedef unsigned char __v32qu __attribute__ ((__vector_size__ (32)));
 
 /* The Intel API is flexible enough that we must allow aliasing with other
--- gcc/config/i386/avx2intrin.h.jj	2019-01-01 12:37:32.000731417 +0100
+++ gcc/config/i386/avx2intrin.h	2019-09-08 23:24:23.391560853 +0200
@@ -258,7 +258,7 @@ extern __inline __m256i
 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 _mm256_cmpgt_epi8 (__m256i __A, __m256i __B)
 {
-  return (__m256i) ((__v32qi)__A > (__v32qi)__B);
+  return (__m256i) ((__v32qs)__A > (__v32qs)__B);
 }
 
 extern __inline __m256i
--- gcc/testsuite/gcc.target/i386/pr91704.c.jj	2019-09-09 11:01:14.588282654 +0200
+++ gcc/testsuite/gcc.target/i386/pr91704.c	2019-09-09 11:09:55.659355290 +0200
@@ -0,0 +1,14 @@
+/* PR target/91704 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -funsigned-char -mavx2 -mavx512f -masm=att" } */
+/* { dg-final { scan-assembler-times "\tvpcmpgtb\t%ymm" 1 } } */
+/* { dg-final { scan-assembler-not "\tvpsubusb\t" } } */
+/* { dg-final { scan-assembler-not "\tvpcmpeqb\t" } } */
+
+#include <x86intrin.h>
+
+__m256i
+foo (__m256i x, __m256i y)
+{
+  return _mm256_cmpgt_epi8 (x, y);
+}

	Jakub

Attachment: T942
Description: Text document