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

tree-optimization/104010 - fix SLP scalar costing with patterns

Browse source 
When doing BB vectorization the scalar cost compute is derailed
by patterns, causing lanes to be considered live and thus not
costed on the scalar side.  For the testcase in PR104010 this
prevents vectorization which was done by GCC 11.  PR103941
shows similar cases of missed optimizations that are fixed by
this patch.

2022-04-13  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/104010
	PR tree-optimization/103941
	* tree-vect-slp.cc (vect_bb_slp_scalar_cost): When
	we run into stmts in patterns continue walking those
	for uses outside of the vectorized region instead of
	marking the lane live.

	* gcc.target/i386/pr103941-1.c: New testcase.
	* gcc.target/i386/pr103941-2.c: Likewise.
This commit is contained in:
Richard Biener 2022-04-13 13:49:45 +02:00
parent cd3964ebd3
commit 353434b65e
3 changed files with 62 additions and 12 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
gcc/testsuite/gcc.target/i386/pr103941-1.c Normal file
View file

@ -0,0 +1,14 @@
/* { dg-do compile } */
/* { dg-options "-O2 -msse2" } */
unsigned char ur[16], ua[16], ub[16];
void avgu_v2qi (void)
{
int i;
for (i = 0; i < 2; i++)
ur[i] = (ua[i] + ub[i] + 1) >> 1;
}
/* { dg-final { scan-assembler "pavgb" } } */

12
gcc/testsuite/gcc.target/i386/pr103941-2.c Normal file
View file

@ -0,0 +1,12 @@
/* { dg-do compile } */
/* { dg-options "-O2 -msse2" } */
void foo (int *c, float *x, float *y)
{
c[0] = x[0] < y[0];
c[1] = x[1] < y[1];
c[2] = x[2] < y[2];
c[3] = x[3] < y[3];
}
/* { dg-final { scan-assembler "cmpltps" } } */

48
gcc/tree-vect-slp.cc
View file

@ -5185,22 +5185,46 @@ vect_bb_slp_scalar_cost (vec_info *vinfo,
the scalar cost. */
if (!STMT_VINFO_LIVE_P (stmt_info))
{
FOR_EACH_PHI_OR_STMT_DEF (def_p, orig_stmt, op_iter, SSA_OP_DEF)
auto_vec<gimple *, 8> worklist;
hash_set<gimple *> *worklist_visited = NULL;
worklist.quick_push (orig_stmt);
do
{
imm_use_iterator use_iter;
gimple *use_stmt;
FOR_EACH_IMM_USE_STMT (use_stmt, use_iter, DEF_FROM_PTR (def_p))
if (!is_gimple_debug (use_stmt))
{
stmt_vec_info use_stmt_info = vinfo->lookup_stmt (use_stmt);
if (!use_stmt_info
|| !vectorized_scalar_stmts.contains (use_stmt_info))
gimple *work_stmt = worklist.pop ();
FOR_EACH_PHI_OR_STMT_DEF (def_p, work_stmt, op_iter, SSA_OP_DEF)
{
imm_use_iterator use_iter;
gimple *use_stmt;
FOR_EACH_IMM_USE_STMT (use_stmt, use_iter,
DEF_FROM_PTR (def_p))
if (!is_gimple_debug (use_stmt))
{
(*life)[i] = true;
break;
stmt_vec_info use_stmt_info
= vinfo->lookup_stmt (use_stmt);
if (!use_stmt_info
|| !vectorized_scalar_stmts.contains (use_stmt_info))
{
if (use_stmt_info
&& STMT_VINFO_IN_PATTERN_P (use_stmt_info))
{
/* For stmts participating in patterns we have
to check its uses recursively. */
if (!worklist_visited)
worklist_visited = new hash_set<gimple *> ();
if (!worklist_visited->add (use_stmt))
worklist.safe_push (use_stmt);
continue;
}
(*life)[i] = true;
goto next_lane;
}
}
}
}
}
while (!worklist.is_empty ());
next_lane:
if (worklist_visited)
delete worklist_visited;
if ((*life)[i])
continue;
}