https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116367
Bug ID: 116367
Summary: Handle vector shuffles better
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Apologies for the broad summary, feel free to make it more targeted.
Testcase:
void test(short * restrict x, short * restrict y) {
const char table[] = {7, 6, 5, 4, 3, 2, 1, 0};
for (int i = 0; i < 8; i++) {
y[i] = x[table[i]];
}
}
Compiled with -Ofast on aarch64 gives:
test:
ldr q31, [x0]
adrp x2, .LC0
ldr q29, [x2, #:lo12:.LC0]
mov v30.16b, v31.16b
tbl v30.16b, {v30.16b - v31.16b}, v29.16b
str q30, [x1]
ret
.LC0:
.byte 14
.byte 15
.byte 12
.byte 13
.byte 10
.byte 11
.byte 8
.byte 9
.byte 6
.byte 7
.byte 4
.byte 5
.byte 2
.byte 3
.byte 0
.byte 1
LLVM does something better:
test:
ldr q0, [x0]
rev64 v0.8h, v0.8h
ext v0.16b, v0.16b, v0.16b, #8
str q0, [x1]
ret
