| Issue |
172888
|
| Summary |
[InstCombine] Missed icmp_ne(and(x,1),0) -> trunc(x) fold
|
| Labels |
llvm:instcombine,
missed-optimization
|
| Assignees |
|
| Reporter |
RKSimon
|
https://alive2.llvm.org/ce/z/U5bXFr
```ll
----------------------------------------
define i1 @src(i32 %a2) {
%s2 = and i32 %a2, 1
%t2 = icmp ne i32 %s2, 0
ret i1 %t2
}
=>
define i1 @tgt(i32 %a2) {
%t2 = trunc i32 %a2 to i1
ret i1 %t2
}
Transformation seems to be correct!
```
Noticed while working on vector code trying to fold SSE/AVX blendv intrinsics to generic select ops:
```ll
define <8 x float> @src(<8 x float> %a0, <8 x float> %a1, <8 x i32> %a2) {
%s2 = shl <8 x i32> %a2, splat (i32 31)
%c2 = icmp slt <8 x i32> %s2, zeroinitializer
%r = select <8 x i1> %c2, <8 x float> %a0, <8 x float> %a1
ret <8 x float> %r
}
define <8 x float> @tgt(<8 x float> %a0, <8 x float> %a1, <8 x i32> %a2) {
%t2 = trunc <8 x i32> %a2 to <8 x i1>
%r = select <8 x i1> %t2, <8 x float> %a0, <8 x float> %a1
ret <8 x float> %r
}
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
