| Issue |
179008
|
| Summary |
[AVX-512] Emit `vmovdqu` over `vpexpand` for dense bit sets
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
Rexicon226
|
Given some IR like:
```llvm
define dso_local <8 x i64> @foo(ptr nocapture nonnull readonly align 1 %0) local_unnamed_addr {
Entry:
%1 = load <8 x i64>, ptr %0, align 1
%2 = shufflevector <8 x i64> %1, <8 x i64> <i64 poison, i64 poison, i64 poison, i64 poison, i64 poison, i64 0, i64 0, i64 0>, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 13, i32 14, i32 15>
ret <8 x i64> %2
}
```
LLVM emits:
```asm
foo:
mov al, 31
kmovd k1, eax
vpexpandq zmm0 {k1} {z}, zmmword ptr [rdi]
ret
```
https://zig.godbolt.org/z/aMer3W9n7
when it should instead emit the better instruction:
```asm
foo:
mov al, 31
kmovd k1, eax
vmovdqu64 zmm0 {k1} {z}, zmmword ptr [rdi]
ret
```
https://c.godbolt.org/z/Ya8EYajTE
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
