[Bug middle-end/128138] New: [17 Regression] wrong code with -fno-math-errno: pow(-Inf, 3.5) is folded to -Inf instead of +Inf

[yuki.tanaka at defaultcase dot de via Gcc-bugs]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128138

            Bug ID: 128138
           Summary: [17 Regression] wrong code with -fno-math-errno:
                    pow(-Inf, 3.5) is folded to -Inf instead of +Inf
           Product: gcc
           Version: 17.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: yuki.tanaka at defaultcase dot de
  Target Milestone: ---

double a(void)
{
  return __builtin_pow(-__builtin_inf(), 3.5);
}

int main(void)
{
  if (a() != __builtin_inf())
    __builtin_abort();
  return 0;
}

Compiler Explorer (GCC 16.1 vs. 17):
https://godbolt.org/z/cj1c8sEqs
-O2 -fno-math-errno -lm

x86-64 GCC 17:
"a":
        movsd   xmm0, QWORD PTR .LC0[rip]
        ret
"main":
        push    rax
        call    "abort"
.LC0:
        .long   0
        .long   -1048576

3.5 is not an odd integer, so pow(-Inf, 3.5) is +Inf, but trunk folds it to
-Inf. GCC 16.1 does not fold it and calls pow, which returns +Inf. Without
-fno-math-errno trunk is fine.

With 1e300 instead of 3.5, trunk crashes in real_isoddinteger (internal
compiler error: Segmentation fault).

Possibly related to r17-5455-ga327f4cf837 ("fold-const: Constant fold
pow(-Inf,0.5) at compile-time").