procyberian/gcc
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

i386.c (ix86_simd_clone_compute_vecsize_and_simdlen): Bump the upper SIMDLEN limits...

Browse source 
* config/i386/i386.c (ix86_simd_clone_compute_vecsize_and_simdlen):
	Bump the upper SIMDLEN limits, so that if the return type or
	characteristic type if the return type is void can be passed in
	all available SSE2/AVX/AVX2/AVX512-F registers, the SIMDLEN is
	allowed.

	* c-c++-common/cilk-plus/SE/ef_error2.c (func2): Use vectorlength
	128 instead of 32.

From-SVN: r234913
This commit is contained in:
Jakub Jelinek 2016-04-12 21:37:21 +02:00 • committed by Jakub Jelinek
parent 13bd7c2bc5
commit 6512fa6dc0
4 changed files with 36 additions and 18 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

8
gcc/ChangeLog
View file

@ -1,3 +1,11 @@
2016-04-12 Jakub Jelinek <jakub@redhat.com>
* config/i386/i386.c (ix86_simd_clone_compute_vecsize_and_simdlen):
Bump the upper SIMDLEN limits, so that if the return type or
characteristic type if the return type is void can be passed in
all available SSE2/AVX/AVX2/AVX512-F registers, the SIMDLEN is
allowed.
2016-04-12 Michael Meissner <meissner@linux.vnet.ibm.com>
PR target/70680

39
gcc/config/i386/i386.c
View file

@ -53761,7 +53761,7 @@ ix86_simd_clone_compute_vecsize_and_simdlen (struct cgraph_node *node,
if (clonei->simdlen
&& (clonei->simdlen < 2
|| clonei->simdlen > 128
|| clonei->simdlen > 1024
|| (clonei->simdlen & (clonei->simdlen - 1)) != 0))
{
warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
@ -53867,21 +53867,28 @@ ix86_simd_clone_compute_vecsize_and_simdlen (struct cgraph_node *node,
clonei->simdlen /= GET_MODE_BITSIZE (TYPE_MODE (base_type));
}
else if (clonei->simdlen > 16)
switch (clonei->vecsize_int)
{
case 512:
/* For AVX512-F, support VLEN up to 128. */
break;
case 256:
/* For AVX2, support VLEN up to 32. */
if (clonei->simdlen <= 32)
break;
/* FALLTHRU */
default:
/* Otherwise, support VLEN up to 16. */
warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
"unsupported simdlen %d", clonei->simdlen);
return 0;
{
/* For compatibility with ICC, use the same upper bounds
for simdlen. In particular, for CTYPE below, use the return type,
unless the function returns void, in that case use the characteristic
type. If it is possible for given SIMDLEN to pass CTYPE value
in registers (8 [XYZ]MM* regs for 32-bit code, 16 [XYZ]MM* regs
for 64-bit code), accept that SIMDLEN, otherwise warn and don't
emit corresponding clone. */
tree ctype = ret_type;
if (TREE_CODE (ret_type) == VOID_TYPE)
ctype = base_type;
int cnt = GET_MODE_BITSIZE (TYPE_MODE (ctype)) * clonei->simdlen;
if (SCALAR_INT_MODE_P (TYPE_MODE (ctype)))
cnt /= clonei->vecsize_int;
else
cnt /= clonei->vecsize_float;
if (cnt > (TARGET_64BIT ? 16 : 8))
{
warning_at (DECL_SOURCE_LOCATION (node->decl), 0,
"unsupported simdlen %d", clonei->simdlen);
return 0;
}
}
return ret;
}

3
gcc/testsuite/ChangeLog
View file

@ -1,5 +1,8 @@
2016-04-12 Jakub Jelinek <jakub@redhat.com>
* c-c++-common/cilk-plus/SE/ef_error2.c (func2): Use vectorlength
128 instead of 32.
PR c++/70571
* g++.dg/ext/pr70571.C: New test.

4
gcc/testsuite/c-c++-common/cilk-plus/SE/ef_error2.c
View file

@ -1,8 +1,8 @@
/* { dg-do compile { target { i?86-*-* x86_64-*-* } } } */
/* { dg-options "-fcilkplus -Wall" } */
__attribute__((vector (vectorlength(32))))
//#pragma omp simd simdlen (32)
__attribute__((vector (vectorlength(128))))
//#pragma omp simd simdlen (128)
int func2 (int x, int y) /* { dg-warning "unsupported simdlen" } */
{
return (x+y);