https://gcc.gnu.org/bugzilla/show_bug.cgi?id=89057
Bug ID: 89057
Summary: GCC 7->8 regression: ARM(64) ld3 st4 less optimized
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: linux at carewolf dot com
Target Milestone: ---
When using the vld3_u8 and vst4_u8 instrinsics, the code generated with gcc8 is
less efficient than the code generated with gcc7. One has 3 moves, and the
other 9 moves.
The code in question is:
#include <stdint.h>
#include <arm_neon.h>
void qt_convert_rgb888_to_rgb32_neon(unsigned *dst, const unsigned char *src,
int len)
{
if (!len)
return;
const unsigned *const end = dst + len;
// align dst on 64 bits
const int offsetToAlignOn8Bytes = (reinterpret_cast<uintptr_t>(dst) >> 2) &
0x1;
for (int i = 0; i < offsetToAlignOn8Bytes; ++i) {
*dst++ = 0xff000000 | (src[0] << 16) | (src[1] << 8) | src[2];
src += 3;
}
if ((len - offsetToAlignOn8Bytes) >= 8) {
const unsigned *const simdEnd = end - 7;
// non-inline asm version (uses more moves)
uint8x8x4_t dstVector;
dstVector.val[3] = vdup_n_u8(0xff);
do {
uint8x8x3_t srcVector = vld3_u8(src);
src += 3 * 8;
dstVector.val[0] = srcVector.val[2];
dstVector.val[1] = srcVector.val[1];
dstVector.val[2] = srcVector.val[0];
vst4_u8((uint8_t*)dst, dstVector);
dst += 8;
} while (dst < simdEnd);
}
while (dst != end) {
*dst++ = 0xff000000 | (src[0] << 16) | (src[1] << 8) | src[2];
src += 3;
}
}
With gcc 7.3 the inner loop is:
.L5:
ld3 {v4.8b - v6.8b}, [x1]
add x1, x1, 24
orr v3.16b, v7.16b, v7.16b
mov v0.8b, v6.8b
mov v1.8b, v5.8b
mov v2.8b, v4.8b
st4 {v0.8b - v3.8b}, [x0]
add x0, x0, 32
cmp x3, x0
bhi .L5
With gcc 8.2 the inner loop is:
.L5:
ld3 {v4.8b - v6.8b}, [x1]
adrp x3, .LC1
add x1, x1, 24
ldr q3, [x3, #:lo12:.LC1]
mov v16.8b, v6.8b
mov v7.8b, v5.8b
mov v4.8b, v4.8b
ins v16.d[1], v17.d[0]
ins v7.d[1], v17.d[0]
ins v4.d[1], v17.d[0]
mov v0.16b, v16.16b
mov v1.16b, v7.16b
mov v2.16b, v4.16b
st4 {v0.8b - v3.8b}, [x0]
add x0, x0, 32
cmp x2, x0
bhi .L5
