tree-optimization/106081 - elide redundant permute

The following patch makes sure to elide a redundant permute that
can be merged with existing splats represented as load permutations
as we now do for non-grouped SLP loads.  This is the last bit
missing to fix this PR where the main fix was already done by
r14-2117-gdd86a5a69cbda4

	PR tree-optimization/106081
	* tree-vect-slp.cc (vect_optimize_slp_pass::start_choosing_layouts):
	Assign layout -1 to splats.

	* gcc.dg/vect/pr106081.c: New testcase.
This commit is contained in:
Richard Biener 2023-07-26 13:31:16 +02:00
parent 386df7ce7b
commit 5d09fb683a
2 changed files with 37 additions and 1 deletions

View file

@ -0,0 +1,33 @@
/* { dg-do compile } */
/* { dg-additional-options "-ffast-math -fdump-tree-optimized" } */
/* { dg-additional-options "-mavx2" { target x86_64-*-* i?86-*-* } } */
/* { dg-require-effective-target vect_double } */
/* { dg-require-effective-target vect_unpack } */
/* { dg-require-effective-target vect_intdouble_cvt } */
/* { dg-require-effective-target vect_perm } */
struct pixels
{
short a,b,c,d;
} *pixels;
struct dpixels
{
double a,b,c,d;
};
double
test(double *k)
{
struct dpixels results={};
for (int u=0; u<1000*16;u++,k--)
{
results.a += *k*pixels[u].a;
results.b += *k*pixels[u].b;
results.c += *k*pixels[u].c;
results.d += *k*pixels[u].d;
}
return results.a+results.b*2+results.c*3+results.d*4;
}
/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } } */
/* { dg-final { scan-tree-dump-times "VEC_PERM" 4 "optimized" { target x86_64-*-* i?86-*-* } } } */

View file

@ -4605,7 +4605,10 @@ vect_optimize_slp_pass::start_choosing_layouts ()
IFN_MASK_LOADs). */
gcc_assert (partition.layout == 0 && !m_slpg->vertices[node_i].succ);
if (!STMT_VINFO_GROUPED_ACCESS (dr_stmt))
continue;
{
partition.layout = -1;
continue;
}
dr_stmt = DR_GROUP_FIRST_ELEMENT (dr_stmt);
imin = DR_GROUP_SIZE (dr_stmt) + 1;
tmp_perm.safe_splice (SLP_TREE_LOAD_PERMUTATION (node));