https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126566
Bug ID: 126566
Summary: [16/17 Regression] ICE in vec.h in phiopt
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: ice-on-valid-code
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
/* ICE / out-of-bounds read. gcc/tree-ssa-phiopt.cc:2444-2449,
spaceship_replacement.
cond3_bb = single_pred (cond2_bb);
if (EDGE_COUNT (cond2_bb->succs) != 2) <- should be cond3_bb
return false;
if (EDGE_SUCC (cond3_bb, 0)->dest == cond2_bb)
{
if (EDGE_SUCC (cond3_bb, 1)->dest != phi_bb)
cond2_bb was already proved to have 2 successors at line 2316, so the
guard is a no-op and cond3_bb is never constrained. Here cond3_bb is the
block holding "gp = &&lab;", which has a single successor, so
EDGE_SUCC (cond3_bb, 1) reads one past the end of the edge vector.
vec::operator[] only has a gcc_checking_assert, so a checking compiler
ICEs and a release compiler uses an uninitialised edge pointer.
The forced label keeps cond3_bb from being merged into cond2_bb; the
"if (k)" arm supplies the fourth predecessor of phi_bb that selects the
4-predecessor path.
gcc -O1 / -O2 / -O3 / -Os
internal compiler error: in operator[], at vec.h:911
during GIMPLE pass: phiopt, from spaceship_replacement
gcc -O2 -fno-ssa-phiopt compiles clean */
void *gp;
__attribute__((noipa)) int
f (int x, int y, int k)
{
int c;
if (k) { c = 0; goto out; }
gp = &&lab;
lab:
if (x < y) c = -1;
else if (x > y) c = 1;
else c = -128;
out:
return c > 0;
}
int
main (void)
{
if (f (1, 2, 0) != 0) __builtin_abort ();
if (f (2, 1, 0) != 1) __builtin_abort ();
return 0;
}