https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110192
Bug ID: 110192
Summary: Missed GIMPLE optimization with calls
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: jakub at gcc dot gnu.org
Target Milestone: ---
int *t;
long b;
void *
foo ()
{
void *c;
if (t)
{
c = operator new (b);
return c;
}
void *d = operator new (b);
return d;
}
has still in optimized dump
t.0_1 = t;
pretmp_13 = b;
_14 = (long unsigned int) pretmp_13;
if (t.0_1 != 0B)
goto <bb 3>; [54.59%]
else
goto <bb 4>; [45.41%]
<bb 3> [local count: 586155665]:
_12 = operator new (_14); [tail call]
goto <bb 5>; [100.00%]
<bb 4> [local count: 487586160]:
_10 = operator new (_14); [tail call]
<bb 5> [local count: 1073741824]:
# _6 = PHI <_12(3), _10(4)>
return _6;
and only RTL optimizations manage (jump2) to just a single unconditional call.
In GCC 8 this causes ICE when t is thread_local (testcase reduced from skia) on
powerpc64le because it isn't prepared to have all TLSLD references disappear
during RTL optimizations (I think even trunk has a theoretical such problem
where get_some_local_dynamic_name will return NULL if all such references are
optimized out and it might misbehave in the callers).
