https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117939
Bug ID: 117939
Summary: [15 Regression] nvptx: CRC test cases execution test
FAILs
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: testsuite-fail, wrong-code
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: tschwinge at gcc dot gnu.org
CC: law at gcc dot gnu.org
Target Milestone: ---
Target: nvptx
For '--target=nvptx-none', there's one existing test case where a recent
regression at first wasn't obviously related to the CRC optimization work:
PASS: gcc.c-torture/execute/pr84524.c -O0 (test for excess errors)
PASS: gcc.c-torture/execute/pr84524.c -O0 execution test
PASS: gcc.c-torture/execute/pr84524.c -O1 (test for excess errors)
PASS: gcc.c-torture/execute/pr84524.c -O1 execution test
PASS: gcc.c-torture/execute/pr84524.c -O2 (test for excess errors)
[-PASS:-]{+FAIL:+} gcc.c-torture/execute/pr84524.c -O2 execution test
PASS: gcc.c-torture/execute/pr84524.c -O3 -fomit-frame-pointer
-funroll-loops -fpeel-loops -ftracer -finline-functions (test for excess
errors)
PASS: gcc.c-torture/execute/pr84524.c -O3 -fomit-frame-pointer
-funroll-loops -fpeel-loops -ftracer -finline-functions execution test
PASS: gcc.c-torture/execute/pr84524.c -O3 -g (test for excess errors)
PASS: gcc.c-torture/execute/pr84524.c -O3 -g execution test
PASS: gcc.c-torture/execute/pr84524.c -Os (test for excess errors)
[-PASS:-]{+FAIL:+} gcc.c-torture/execute/pr84524.c -Os execution test
For '-O2', '-Os' (but not '-O3') that now FAILs with:
nvptx-run: error getting kernel result: an illegal memory access was
encountered (CUDA_ERROR_ILLEGAL_ADDRESS, 700)
I have to have a look what's going on. (For now, let's assume it's specific to
nvptx target.) I can get '-O2', '-Os' back to PASS with '-fno-optimize-crc',
thus "[15 Regression]".
However, a different observation: with plain '-O3' (without
'-fno-optimize-crc'), there are earlier code transformations (to be determined)
that apparently preclude CRC optimization. I can't tell if that's profitable
or not. Jeff: would Mariam et al. maybe want to have a look at that test case
-- assuming that this '-O3' behavior also reproduces on the CRC work's usual
architectures, and unless the behavior is obvious to explain anyway? See the
'diff' of '-O2' vs. '-O3' for nvptx:
--- O2/pr84524.c.169t.crc 2024-12-06 19:30:08.307780439 +0100
+++ O3/pr84524.c.169t.crc 2024-12-06 19:37:03.142273827 +0100
@@ -1,83 +1,213 @@
;; Function foo (foo, funcdef_no=0, decl_uid=1853, cgraph_uid=1,
symbol_order=0)
-
-foo function maybe contains CRC calculation.
-Loop iteration number is 7.
-Bit forward.
-The loop with 4 header BB index calculates CRC!
__attribute__((noipa, noinline, noclone, no_icf))
void foo (short unsigned int * x)
{
[...]
I see big GIMPLE IR changes starting with 'pr84524.c.112t.cunrolli', maybe
that's responsible, maybe not.
Otherwise, a few dozens of the new CRC test cases likewise FAIL their execution
test, as above: 'CUDA_ERROR_ILLEGAL_ADDRESS'. For now I assume it's all the
same underlying issue.
