| Issue |
155584
|
| Summary |
(u)comiss instructions not merged together in comparison chains
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
purplesyringa
|
https://godbolt.org/z/a1f4hrPzT
```c
enum Ordering {
LESS = -1,
EQUAL = 0,
GREATER = 1,
UNORDERED = 2
};
enum Ordering compare(float a, float b) {
if (a < b) {
return LESS;
} else if (a == b) {
return EQUAL;
} else if (a > b) {
return GREATER;
} else {
return UNORDERED;
}
}
int another_example(float a, float b) {
if (a < b) {
return 1;
} else if (a <= b) {
return 2;
} else {
return 3;
}
}
```
```x86asm
compare:
mov eax, -1
ucomiss xmm1, xmm0
ja .LBB0_3
xor eax, eax
ucomiss xmm0, xmm1
jne .LBB0_2
jp .LBB0_2
.LBB0_3:
ret
.LBB0_2:
setbe al
movzx eax, al
inc eax
ret
another_example:
xor ecx, ecx
ucomiss xmm1, xmm0
adc ecx, 2
ucomiss xmm1, xmm0
mov eax, 1
cmovbe eax, ecx
ret
```
I see no good reason why the partial ordering implementation cannot emit `ucomiss` just once, e.g.
```x86asm
compare:
ucomiss xmm0, xmm1
mov eax, 2
jp .exit
mov eax, -1
jb .exit
seta al
.exit:
ret
```
`another_example` looks even more straight-forward, with operands not even swapped.
`ucomiss` is not exactly slow, but it's not fast either, so surely this is a missed optimization?
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
