https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128872

            Bug ID: 128872
           Summary: [15/16/17 Regression] wrong code at -O2 and above:
                    unsigned saturating subtraction (a >= b ? a - b : 0)
                    on unsigned char/short is sign-extended when widened
           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,

$cat repro.c

[[gnu::noipa]] short f (unsigned char b, unsigned char c)
{
  return (c >= b) * (unsigned char) (c - b);
}

int main (void)
{
  /* c >= b, c - b == 200: the unsigned char product is 200, so f must return
200.  */
  if (f (0, 200) != 200)
    __builtin_abort ();
  return 0;
}


$ gcc -O2 repro.c && ./a.out      # also -O3, -Os;  -O0 / -O1 fine
Aborted


(c >= b) * (unsigned char)(c - b) is an unsigned char saturating subtraction;
its value is in [0, 255], so
f (0, 200) must be 200. At -O2 the result is sign-extended to short (f returns
-56):


## Checked Versions
- gcc 13: correct
- gcc 15.2, gcc 16, trunk 9571dd0b721 (2026-11-17): wrong at -O2/-O3/-Os
- clang: correct

Reproduced in Godbolt: https://godbolt.org/z/6YPKefc7Y

Thanks.

Reply via email to