OpenE2K/gcc
2
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

tree-optimization/100778 - fix placement of trapping vectorized ops

Browse Source 
This avoids placing possibly trapping vectorized operations where
the corresponding scalar operation was possibly not executed.

2021-01-07  Richard Biener  <rguenther@suse.de>

	PR tree-optimization/100778
	* tree-vect-slp.c (vect_schedule_slp_node): Do not place trapping
	vectorized ops ahead of their scalar BB.

	* gcc.dg/torture/pr100778.c: New testcase.
This commit is contained in:
Richard Biener 2021-07-01 10:35:38 +02:00
parent 9f6aeb85ee
commit a3aaba6840
2 changed files with 46 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

31
gcc/testsuite/gcc.dg/torture/pr100778.c Normal file
View File

@ -0,0 +1,31 @@
/* { dg-do run { target *-*-*gnu* } } */
/* { dg-additional-options "-fno-tree-sink -fno-math-errno -ftree-vectorize -D_GNU_SOURCE" } */
/* { dg-require-effective-target fenv_exceptions } */
#include <fenv.h>
double a[2];
void __attribute__((noipa)) foo ()
{
double x = a[0];
double y = a[1];
double norm = __builtin_sqrt (x*x + y*y);
if (norm > 1.)
{
x = x / norm;
y = y / norm;
}
a[0] = x;
a[1] = y;
}
int main()
{
feenableexcept (FE_INVALID);
a[0] = 0.;
a[1] = 0.;
foo ();
if (a[0] != 0. || a[1] != 0.)
__builtin_abort ();
return 0;
}

15
gcc/tree-vect-slp.c
View File

@ -7100,6 +7100,21 @@ vect_schedule_slp_node (vec_info *vinfo,
gcc_assert (seen_vector_def);
si = gsi_after_labels (as_a <bb_vec_info> (vinfo)->bbs[0]);
}
else if (is_a <bb_vec_info> (vinfo)
&& gimple_bb (last_stmt) != gimple_bb (stmt_info->stmt)
&& gimple_could_trap_p (stmt_info->stmt))
{
/* We've constrained possibly trapping operations to all come
from the same basic-block, if vectorized defs would allow earlier
scheduling still force vectorized stmts to the original block.
This is only necessary for BB vectorization since for loop vect
all operations are in a single BB and scalar stmt based
placement doesn't play well with epilogue vectorization. */
gcc_assert (dominated_by_p (CDI_DOMINATORS,
gimple_bb (stmt_info->stmt),
gimple_bb (last_stmt)));
si = gsi_after_labels (gimple_bb (stmt_info->stmt));
}
else if (is_a <gphi *> (last_stmt))
si = gsi_after_labels (gimple_bb (last_stmt));
else