https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128867
Bug ID: 128867
Summary: [17 Regression] wrong code at -O1 with
__builtin_islessgreater on a NaN operand (tree DOM)
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: wrong-code
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: haoxintu at gmail dot com
Target Milestone: ---
Hi,
The following program aborts when compiled with -O1 (passes at -O0, -O2, -O3,
-Os, and with clang):
$cat repro.c
[[gnu::noipa]] double f (double a, float b, double c, long long ib)
{
double L = __builtin_isless ((double)__builtin_islessgreater ((double)b /
(double)b, -c), (double)__builtin_isgreater (0.0, (double)__builtin_isinf_sign
((double)b)));
double R = (a + (6.0e-5 + c));
return L < R ? (double)(float)ib : 5.0;
}
int main (void)
{
/* b / b is NaN, so islessgreater (NaN, -inf) == 0; isinf_sign (-inf) == -1,
so isgreater (0.0, -1.0) == 1;
L = isless (0.0, 1.0) = 1.0; R = 0.0 + (6e-5 + inf) = inf; L < R, so the
result is (double)(float)ib. */
if (f (0.0, -__builtin_inff (), __builtin_inf (), 9ll << 40) != 0x1.2p+43)
__builtin_abort ();
return 0;
}
$ gcc -O1 repro.c && ./a.out
Aborted
At -O1, f returns 5.0 instead of 0x1.2p+43. Expected value (no UB; IEEE
semantics):
b / b is NaN, so islessgreater (NaN, -inf) == 0; isinf_sign (-inf) == -1, so
isgreater (0.0, -1.0) == 1;
L = isless (0.0, 1.0) = 1.0; R = 0.0 + (6e-5 + inf) = +inf; L < R holds, so f
returns (double)(float)ib.
-O1 -fno-tree-dominator-opts gives the correct result, so the problem seems to
be in DOM, which was reworked in
October 2026 (e.g. 58294ebfb40 "DOM: replace simplify_binary_operation with the
ranger",
8f91f1288ad "DOM: stop rewriting comparisons into equalities"). Replacing
islessgreater (x, y) by !(x == y) hides it.
Affected Versions
- gcc 12, 13, 15.2, 16: correct
- trunk 9571dd0b721 (2026-11-17): wrong at -O1
Reproduced in Godbolt: https://godbolt.org/z/xrzPaoGa7
Thanks.