[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 --- Comment #23 from Sam James --- Thanks Martin.
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 Martin Jambor changed: What|Removed |Added Resolution|--- |DUPLICATE Status|NEW |RESOLVED --- Comment #22 from Martin Jambor --- (In reply to mcccs from comment #21) > Okay I'm so sorry for the misinterpretation. I thought DCE should never turn > an invalid program to a valid one because DCE is always an optimization and > optimizations should never modify the behavior of the programs. But since > you're the code owner and someone familiar with GCC internals your word is > worth 100 times my word so this can be closed. Your thinking about DCE, meaning the DCE pass, is actually the one that prevailed. But that is in fact why I was made to write the patch committed as r12-1848. It also adds a special DCE-like feature to IPA-SRA pass but it is there specifically for the pass to clean up after itself, not as a standalone optimization. Before that the pass did rely on the fact that when a parameter is used as a divisor or when it is dereferenced, but then the result is never used, DCE would take care of that operation and remove it. And so IPA-SRA was happy to keep such operations around even if they used uninitialized values from what remained of removed formal parameters of a function. Since it was decided that DCE can be switched off safely, IPA-SRA had to find all such operations it "created" and remove it - or rather not copy it the new clone - itself. The line you removed by sed is an initial part of that process. *** This bug has been marked as a duplicate of bug 99912 ***
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 --- Comment #21 from mcccs at gmx dot com --- Okay I'm so sorry for the misinterpretation. I thought DCE should never turn an invalid program to a valid one because DCE is always an optimization and optimizations should never modify the behavior of the programs. But since you're the code owner and someone familiar with GCC internals your word is worth 100 times my word so this can be closed.
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 --- Comment #20 from Martin Jambor --- (In reply to mcccs from comment #17) > [...] > > The problem is that the parameter pointer is not set to the argument so > dereferencing it causes a segfault. Dumping all the passes, this problem is > first present in the output of the pass "x.c.100t.fixup_cfg3.c" as compiled > with the modified GCC with -O3 -fno-dce -fno-ipa-cp -fno-tree-dce -fno-dse > -fno-tree-dse > This looks pretty much like a bug that... (In reply to mcccs from comment #15) > There's indeed a miscompilation and I've confirmed it's still present in the > current trunk. With -fno-dce -fno-ipa-cp -fno-tree-dce the issue was visible > until r12-248 which made the issue latent. So I added -fno-tree-dse and then > it was made latent by r12-1848. So I added: ...the commit r12-1848 aimed to fix. What makes you think it is "latent?" > [...] > > Download the minimized testcase from > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282#c14 > > -O3 -fno-dce -fno-ipa-cp -fno-tree-dce -fno-dse -fno-tree-dse > > sed -i -e 's/mark_dead_statements (m_oparms\[i\]/(void)3;\/\//g' > ./gcc/gcc/ipa-param-manipulation.cc > This looks like undoing a fix for the problem you described in comment#17, I'm afraid.
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 Sam James changed: What|Removed |Added Ever confirmed|0 |1 Last reconfirmed||2025-05-19 Status|UNCONFIRMED |NEW --- Comment #19 from Sam James --- Tentatively confirmed then.
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 --- Comment #18 from mcccs at gmx dot com --- > this problem is first present in the output of the pass > "x.c.100t.fixup_cfg3.c" Sorry, there are non-tree dumps as well. the first wrong dump is (of course) the "inline" ipa dump
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282
--- Comment #17 from mcccs at gmx dot com ---
Actually my test file in comment 16 is invalid. This means the failure happens
before the optimized pass and it needs the DCE-disabled GCC I described in
comment 15 to reproduce the faulty optimized dump (which is then propagated in
later passes in any GCC).
The problem is that the parameter pointer is not set to the argument so
dereferencing it causes a segfault. Dumping all the passes, this problem is
first present in the output of the pass "x.c.100t.fixup_cfg3.c" as compiled
with the modified GCC with -O3 -fno-dce -fno-ipa-cp -fno-tree-dce -fno-dse
-fno-tree-dse
The previous dump, "x.058t.local-fnsummary2", is sane and looks like this:
int * __GIMPLE (ssa,guessed_local(1073741824))
n (int * o)
{
...
_3 = __MEM (o_17(D));
...
main ()
{
__BB(2):
n (&a);
whereas x.c.100t.fixup_cfg3.c looks like this:
int __GIMPLE (ssa)
main ()
{
int * o;
...(no assignment to o)...
_6 = __MEM (o_5(D));
__attribute__((noclone)) to the function fixes the segmentation fault,
__attribute__((noinline)) doesn't.
[Bug ipa/110282] [12/13/14/15 regression] Segmentation fault with specific optimizations since r10-3311-gff6686d2e5f797
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110282 Sam James changed: What|Removed |Added Target Milestone|--- |12.5 Summary|Segmentation fault with |[12/13/14/15 regression] |specific optimizations |Segmentation fault with ||specific optimizations ||since ||r10-3311-gff6686d2e5f797 Keywords|needs-bisection,| |needs-reduction | Component|middle-end |ipa CC||jamborm at gcc dot gnu.org
