https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69442
--- Comment #4 from ktkachov at gcc dot gnu.org ---
The differences between various -mtune options come out during expand phase,
where the expansion for the mod operation is different.
For example, for -mtune=cortex-a9 the code doesn't abort and the assembly for
foo is:
foo:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
push {r4, r5}
mvn ip, r0, lsr #16
mvn ip, ip, lsl #16
subs r4, r2, #255
sbc r5, r3, #0
mvn r2, #24
mov r3, #0
movt r3, 65535
cmp r5, r3
cmpeq r4, r2
movcs r3, #1
movcc r3, #0
mov r0, r3
mov r1, #0
vmov d16, r0, r1 @ int
vshl.u64 d17, d16, #16
vsub.i64 d17, d16, d17
vshl.u64 d17, d17, #27
vsub.i64 d17, d17, d16
vshl.u64 d17, d17, #2
vadd.i64 d17, d17, d16
vshl.u64 d16, d17, #3
vmov r2, r3, d16 @ int
subs r2, r2, r0
sbc r3, r3, r1
mov r0, r2
mov r1, r3
subs r0, r4, r0
sbc r1, r5, r1
adds r0, r0, ip
adc r1, r1, #0
pop {r4, r5}
bx lr
but for -mtune=cortex-a8 we have a completely different expansion of mod that
seems to be misoptimised or something:
foo:
@ args = 0, pretend = 0, frame = 0
@ frame_needed = 0, uses_anonymous_args = 0
@ link register save eliminated.
mvn ip, r0, lsr #16
mvn ip, ip, lsl #16
subs r2, r2, #255
sbc r3, r3, #0
mvn r0, #24
mov r1, #0
movt r1, 65535
cmp r3, r1
cmpeq r2, r0
movcs r1, #1
movcc r1, #0
mvn r0, #24
umull r0, r1, r1, r0
subs r0, r2, r0
sbc r1, r3, r1
adds r0, r0, ip
adc r1, r1, #0
bx lr
