https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118174
Bug ID: 118174
Summary: AArch64: Miscompilation at -O3
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: kristerw at gcc dot gnu.org
Target Milestone: ---
The following function is miscompiled for AArch64 when compiled with -O3:
int
foo (signed char *p1, signed char *p2)
{
int sum = 0;
for (int i = 0; i < 32; i++)
sum += __builtin_abs (p1[i] - p2[i]);
return sum;
}
The GIMPLE passes have transformed this to SAD_EXPR of 16-element vectors,
followed by a .REDUC_PLUS. This is then generated as the following assembly:
foo:
ldp q3, q28, [x0]
ldp q2, q31, [x1]
uabdl2 v29.8h, v3.16b, v2.16b
uabdl2 v1.8h, v28.16b, v31.16b
uabal v29.8h, v3.8b, v2.8b
uabal v1.8h, v28.8b, v31.8b
uaddlp v29.4s, v29.8h
uadalp v29.4s, v1.8h
uadalp v29.4s, v1.8h
uadalp v29.4s, v1.8h
addv s31, v29.4s
fmov w0, s31
ret
The reduction has been generated incorrectly as
uaddlp v29.4s, v29.8h
uadalp v29.4s, v1.8h
uadalp v29.4s, v1.8h
uadalp v29.4s, v1.8h
addv s31, v29.4s
which accumulates the register v1 multiple times.
