| Issue |
58129
|
| Summary |
[Thumb-1] inefficient select in pre-v6 thumb
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
easyaspi314
|
LLVM makes questionable decisions on Thumb-1 with selects pre-v6 (where `tMOVr` only works on hi registers).
```llvm
define i32 @foo(i32 %x, i32 %y, i32 %z) {
%cmp = icmp ne i32 %x, 0
%sel = select i1 %cmp, i32 %y, i32 %z
; breaks return x ? y : z pattern
%sum = add i32 %sel, %z
ret i32 %sum
}
```
`clang --target=thumbv4t-none-eabi -O2 -S foo.ll`
```asm
foo:
cmp r0, #0
push {r2}
pop {r0}
beq .LBB0_2
movs r0, r1
.LBB0_2:
adds r0, r0, r2
bx lr
```
The ideal solution would just be something like this, which is similar to what it would emit with the `return x ? y : z` pattern.
```asm
foo:
cmp r0, #0
beq .LBB0_2
movs r1, r2
.LBB0_2:
adds r0, r1, r2
bx lr
```
However, even putting the optimal code aside, it would be better to try and clobber an available hi register:
```asm
foo:
cmp r0, #0
mov r12, r2
mov r0, r12
beq .LBB0_2
movs r0, r1
.LBB0_2:
adds r0, r0, r2
bx lr
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
