| Issue |
79315
|
| Summary |
Redundant equality check when value is known to be 0 or 1
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
Sp00ph
|
Example:
```ll
declare void @llvm.assume(i1)
define i1 @foo(i8 %a) {
%lt = icmp ult i8 %a, 2
tail call void @llvm.assume(i1 %lt)
%ret = icmp ne i8 %a, 0
ret i1 %ret
}
```
The `assume` call constrains the value of `%a` to be either 0 or 1, so the function should get compiled to a single `mov al, dil` instruction. Instead, clang produces this redundant instruction sequence:
```asm
foo:
test dil, dil
setne al
ret
```
The same thing happens on aarch64, and replacing `icmp ne i8 %a, 0` with `trunc i8 %a to i1` doesn't change the output either.
I discovered this when playing around with inline asm in Rust: https://rust.godbolt.org/z/oafj35dn5 . In this case, the value is set to either 0 or 1 by the `sete` instruction in the asm block, but LLVM inserts two redundant instructions afterwards.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
