https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128882
Bug ID: 128882
Summary: [15/16/17 Regression] wrong code at -O2: subtraction
in a narrower type recognized as .SAT_SUB of the wider
operands
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: liuhongt at gcc dot gnu.org
Target Milestone: ---
Target: x86_64-pc-linux-gnu
$ cat t.c
[[gnu::noipa]] unsigned
f (unsigned a, unsigned b)
{
unsigned short d = (unsigned short) a - (unsigned short) b;
return a >= b ? d : 0;
}
int
main ()
{
if (f (0x10005, 3) != 2)
__builtin_abort ();
return 0;
}
$ gcc -O2 t.c && ./a.out
Aborted
GCC 14 is fine at all optimization levels. GCC 15, 16 and trunk fail at
-O2 and -O3 (-O0, -O1 and -Os are fine). The optimized dump has
prephitmp_10 = .SAT_SUB (a_4(D), b_5(D));
The difference is computed in unsigned short, so it wraps at 16 bits and
is not a saturating subtraction of a and b: f (0x10005, 3) must be 2,
but .SAT_SUB gives 0x10002.
r15-1671 changed the X >= Y ? X - Y : 0 pattern (now in match-sat-alu.pd)
to (convert? (minus (convert1? @0) (convert1? @1))) so that a truncated
.SAT_SUB is recognized. Nothing stops convert1 from narrowing @0 and @1.
The matcher is shared by widening_mul and the vectorizer. The
X <= Y ? 0 : X - Y form added in r17-4219 has the same hole, though I
couldn't get the reproducer above into that shape.
This is separate from PR128872, where the match is valid but widening_mul
drops the conversion to the type of the lhs.