https://gcc.gnu.org/bugzilla/show_bug.cgi?id=123748
Bug ID: 123748
Summary: Conversion from long to float is not vectorized for
aarch64 target
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: pzheng at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
void foo(float *restrict f, long *restrict l) {
for (int i = 0; i < 128; i++)
f[i] = l[i];
}
With -O2, GCC generates:
foo:
mov x2, 0
.L2:
ldr x3, [x1, x2, lsl 3]
scvtf s31, x3
str s31, [x0, x2, lsl 2]
add x2, x2, 1
cmp x2, 128
bne .L2
ret
While, LLVM generates:
foo:
mov x8, xzr
.LBB0_1:
ldp x9, x10, [x1]
scvtf s0, x10
scvtf s1, x9
ldp x9, x10, [x1, #16]
add x1, x1, #32
mov v1.s[1], v0.s[0]
scvtf s0, x9
mov v1.s[2], v0.s[0]
scvtf s0, x10
mov v1.s[3], v0.s[0]
str q1, [x0, x8]
add x8, x8, #16
cmp x8, #512
b.ne .LBB0_1
ret
