| Issue |
178403
|
| Summary |
[SLP] Handle zext(i1 -> i32) and select(i1, i32 1, i32 0) equivalency
|
| Labels |
llvm:SLPVectorizer,
missed-optimization
|
| Assignees |
|
| Reporter |
RKSimon
|
Pulled out of #96395
```rust
pub unsafe fn ascii_prefix(input: &[u8]) -> usize {
let mut mask = 0_u16;
for i in 0..16 {
mask |= ((*input.get_unchecked(i) < 128) as u16) << i;
}
mask.trailing_ones() as usize
}
```
By the time it gets to SLP we see this pattern:
```ll
%0 = load i8, ptr %input, align 1
%cmp3 = icmp sgt i8 %0, -1
%1 = zext i1 %cmp3 to i16
%arrayidx.1 = getelementptr inbounds nuw i8, ptr %input, i64 1
%2 = load i8, ptr %arrayidx.1, align 1
%cmp3.1 = icmp sgt i8 %2, -1
%3 = select i1 %cmp3.1, i16 2, i16 0
%conv9.1 = or disjoint i16 %3, %1
%arrayidx.2 = getelementptr inbounds nuw i8, ptr %input, i64 2
%4 = load i8, ptr %arrayidx.2, align 1
%cmp3.2 = icmp sgt i8 %4, -1
%5 = select i1 %cmp3.2, i16 4, i16 0
%conv9.2 = or disjoint i16 %conv9.1, %5
.....
```
Loop unrolling has constant folded the shifts entirely and replaced them with icmp+select pairs, except for i == 0 which has become a icmp+zext - preventing SLP copyable / sameopcode patterns from cleaning everything up.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
