https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117709
Bug ID: 117709
Summary: [15 regression] maskload else case generating wrong
code
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ams at gcc dot gnu.org
CC: rdapp at gcc dot gnu.org
Target Milestone: ---
The following testcase aborts on amdgcn since the maskload else patches were
added (https://patchwork.sourceware.org/project/gcc/list/?series=40395, plus
the patch for pr117657).
-----------8<-----------
int i, j;
int k[11][101];
__attribute__((noipa)) void
doit (void) {
int err = 0;
for (j = -11; j >= -41; j -= 15) {
k[0][-j] = 1;
}
#pragma omp simd collapse(2) lastprivate (i, j) reduction(|:err)
for (i = 4; i < 8; i += 12)
for (j = -8 * i - 9; j < i * -3 + 6; j += 15)
err |= (k[0][-j] != 1);
if (err)
__builtin_abort ();
}
int main () {
doit ();
}
-----------8<-----------
The testcase is simplified from gcc.dg/vect/vect-simd-17.c.
If I revert most of the gcn patch, but keep the new operand with the
"maskload_else_operand" predicate, so that the backend generates the
zero-initializer, but the middle-end assumes the lanes are undefined, then the
testcase still fails.
I think this shows that the problem in in the additional code generated by the
middle end, not that the vectors are actually corrupt, but I've not identified
exactly how.
