| Issue |
56672
|
| Summary |
[LoopRotate][LoopVectorize] Suspected broken analyzes after LoopRotate
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
max-quazan
|
I am not sure if this particular thing is a bug, but looks *extremely* fishy.
Link 1: https://godbolt.org/z/1bfKdadx4
Link 2: https://godbolt.org/z/4hq1931ex
Test:
```
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128-ni:1-p2:32:8:8:32-ni:2"
target triple = "x86_64-unknown-linux-gnu"
define void @test() {
bb:
br label %bb4
bb1: ; preds = %bb4
br label %bb2
bb2: ; preds = %bb1
store i32 1, i32 addrspace(1)* %tmp7, align 4
%tmp = add nuw i64 %tmp5, 1
%tmp3 = icmp ult i64 %tmp, undef
br i1 %tmp3, label %bb4, label %bb11
bb4: ; preds = %bb2, %bb
%tmp5 = phi i64 [ %tmp, %bb2 ], [ 16, %bb ]
%tmp6 = phi i64 [ %tmp5, %bb2 ], [ 15, %bb ]
%tmp7 = getelementptr inbounds i32, i32 addrspace(1)* null, i64 %tmp5
%tmp8 = load i32, i32 addrspace(1)* %tmp7, align 4
%tmp9 = add i32 %tmp8, -5
store i32 %tmp9, i32 addrspace(1)* %tmp7, align 4
br i1 false, label %bb10, label %bb1
bb10: ; preds = %bb4
unreachable
bb11: ; preds = %bb2
ret void
}
```
Two run commands:
```
opt -passes='loop(loop-rotate),loop-vectorize' -S
opt -passes='loop(loop-rotate),invalidate<loops>,loop-vectorize' -S
```
In the former case, loop does vectorize. In the later case, it doesn't. For us, the vectorization is buggy (miscompile), but I could not figure out how to reduce test to demonstrate the miscompile yet.
The suspicion is that some analysis is in broken state after loop-rotate, and LV takes wrong decision basing on it.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
