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

tree-optimization/115843 - fix wrong-code with fully-masked loop and peeling

Browse source 
When AVX512 uses a fully masked loop and peeling we fail to create the
correct initial loop mask when the mask is composed of multiple
components in some cases.  The following fixes this by properly applying
the bias for the component to the shift amount.

	PR tree-optimization/115843
	* tree-vect-loop-manip.cc
	(vect_set_loop_condition_partial_vectors_avx512): Properly
	bias the shift of the initial mask for alignment peeling.

	* gcc.dg/vect/pr115843.c: New testcase.

(cherry picked from commit a177be05f6952c3f7e62186d2e138d96c475b81a)
This commit is contained in:
Richard Biener 2024-07-15 13:50:58 +02:00 • committed by Richard Biener
parent e01012c459
commit 06829e593d
2 changed files with 47 additions and 2 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

41
gcc/testsuite/gcc.dg/vect/pr115843.c Normal file
View file

@ -0,0 +1,41 @@
/* { dg-additional-options "-mavx512f --param vect-partial-vector-usage=2" { target avx512f_runtime } } */
#include "tree-vect.h"
typedef __UINT64_TYPE__ BITBOARD;
BITBOARD KingPressureMask1[64], KingSafetyMask1[64];
void __attribute__((noinline))
foo()
{
for (int i = 0; i < 64; i++)
{
if ((i & 7) == 0)
KingPressureMask1[i] = KingSafetyMask1[i + 1];
else if ((i & 7) == 7)
KingPressureMask1[i] = KingSafetyMask1[i - 1];
else
KingPressureMask1[i] = KingSafetyMask1[i];
}
}
BITBOARD verify[64]
= {1, 1, 2, 3, 4, 5, 6, 6, 9, 9, 10, 11, 12, 13, 14, 14, 17, 17, 18, 19,
20, 21, 22, 22, 25, 25, 26, 27, 28, 29, 30, 30, 33, 33, 34, 35, 36, 37, 38,
38, 41, 41, 42, 43, 44, 45, 46, 46, 49, 49, 50, 51, 52, 53, 54, 54, 57, 57,
58, 59, 60, 61, 62, 62};
int main()
{
check_vect ();
#pragma GCC novector
for (int i = 0; i < 64; ++i)
KingSafetyMask1[i] = i;
foo ();
#pragma GCC novector
for (int i = 0; i < 64; ++i)
if (KingPressureMask1[i] != verify[i])
__builtin_abort ();
return 0;
}

6
gcc/tree-vect-loop-manip.cc
View file

@ -1149,8 +1149,12 @@ vect_set_loop_condition_partial_vectors_avx512 (class loop *loop,
/* ??? But when the shift amount isn't constant this requires
a round-trip to GRPs. We could apply the bias to either
side of the compare instead. */
tree shift = gimple_build (&preheader_seq, MULT_EXPR,
tree shift = gimple_build (&preheader_seq, MINUS_EXPR,
TREE_TYPE (niters_skip), niters_skip,
build_int_cst (TREE_TYPE (niters_skip),
bias));
shift = gimple_build (&preheader_seq, MULT_EXPR,
TREE_TYPE (niters_skip), shift,
build_int_cst (TREE_TYPE (niters_skip),
rgc.max_nscalars_per_iter));
init_ctrl = gimple_build (&preheader_seq, LSHIFT_EXPR,