https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106467
Bug ID: 106467
Summary: [OpenMP] Wrong code with collapse – tree sharing
issue.
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: ice-on-valid-code, openmp
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: burnus at gcc dot gnu.org
CC: burnus at gcc dot gnu.org, jakub at gcc dot gnu.org,
sandra at gcc dot gnu.org
Depends on: 106449
Target Milestone: ---
Created attachment 53375
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=53375&action=edit
Testcase, modified version of libgomp.c-c++-common/pr106449.c from attachment
53370 to PR 106449
+++ This bug was initially created as a clone of Bug #106449 +++
(In reply to Jakub Jelinek from bug 106449 comment #9)
> Created attachment 53370 [details]
> gcc13-pr106449-2.patch
>
> Updated fix for this PR which fixes even the runtime case (it was caused by
> missing unsharing and regimplification clobbering a shared tree).
My potentially wrong impression is that the unshare_tree issue also exists
with a bare collapse.
At least, when converting the testcase from the patch to one which uses 'FOR'
instead of '(FOR) SIMD', it fails for BAR in a similar way as the SIMD version
with the initial version of the patch.
Namely, the result it the __builtin_abort call for i = 24 in bar (n=64, m=128):
33 for (i = 0; i < 32768; i++)
34 if (b[2 * i] != &a[i / 64] || b[2 * i + 1] != &a[(i / 64) + 64 + (i
% 64)])
35 __builtin_abort ();
The LHS of the is false (sub-LHS/sub-RHS are idential) but the RHS, i.e.
b[2 * i + 1] != &a[(i / 64) + 64 + (i % 64)]
fails as
b[2 * i + 1] = 0x7fffffffd588
while
&a[(i / 64) + 64 + (i % 64)] = 0x7fffffffd490
Referenced Bugs:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106449
[Bug 106449] ICE in #pragma omp parallel for simd since
r6-4544-ge01d41e553aae245
