https://gcc.gnu.org/bugzilla/show_bug.cgi?id=127086
--- Comment #2 from Anton Blanchard <anton at ozlabs dot org> ---
In this case it took -flto to trigger, but I'm not sure if it is -flto
specific.
When trying to understand the problem I ended up with the following test case.
The call to tick() after the loop causes gcc to write VXRM every row, remove it
and we hoist it before the loop.
#include <stdint.h>
void tick(void);
void avg_then_call(uint8_t *dst, const uint8_t *s1, const uint8_t *s2,
int w, int h, int ds, int ss1, int ss2)
{
for (int y = 0; y < h; y++) {
for (int x = 0; x < w; x++)
dst[x] = (s1[x] + s2[x] + 1) >> 1;
dst += ds; s1 += ss1; s2 += ss2;
}
tick();
}
bad:
.L3:
minu t1,a2,t2
bleu t1,a7,.L18
csrwi vxrm,0
mv t4,s1
mv t3,a0
mv t6,a5
mv t5,a1
.L4:
vsetvli t1,t4,e8,m1,ta,ma
vle8.v v1,0(t5)
vle8.v v2,0(t6)
sub t4,t4,t1
add t5,t5,t1
add t6,t6,t1
vaaddu.vv v1,v1,v2
vse8.v v1,0(t3)
add t3,t3,t1
bne t4,zero,.L4
addiw a4,a4,1
add a0,a0,a3
add a1,a1,a6
add a5,a5,s2
add a2,a2,s4
add t2,t2,s3
add t0,t0,a6
bne s0,a4,.L3
good (remove call to tick()):
.L3:
minu t1,a2,t2
bleu t1,a7,.L19
mv t4,s1
mv t3,a0
mv t6,a5
mv t5,a1
.L4:
vsetvli t1,t4,e8,m1,ta,ma
vle8.v v1,0(t5)
vle8.v v2,0(t6)
sub t4,t4,t1
add t5,t5,t1
add t6,t6,t1
vaaddu.vv v1,v1,v2
vse8.v v1,0(t3)
add t3,t3,t1
bne t4,zero,.L4
addiw a4,a4,1
add a0,a0,a3
add a1,a1,a6
add a5,a5,s2
add a2,a2,s4
add t2,t2,s3
add t0,t0,a6
bne s0,a4,.L3
