mikaelholmen wrote:
I think this patch causes miscompiles. Reproduce with
```opt bbi-88690.ll -passes=instcombine -S -o -```
So with this patch instcombine turns
```
@v_936 = global i16 -3276, align 1
@v_937 = global i24 0, align 1
define i16 @main() {
entry:
%0 = load i16, ptr @v_936, align 1
%unsclear = and i16 %0, 32767
%resize = zext i16 %unsclear to i24
%unsclear1 = and i24 %resize, 8388607
store i24 %unsclear1, ptr @v_937, align 1
ret i16 0
}
```
into
```
@v_936 = global i16 -3276, align 1
@v_937 = global i24 0, align 1
define i16 @main() {
entry:
%0 = load i16, ptr @v_936, align 1
%resize = zext nneg i16 %0 to i24
store i24 %resize, ptr @v_937, align 1
ret i16 0
}
```
I.e the and with 32767 (0x7fff) is gone and instead the zext got "nneg"?
But the value in v_936 can be, and actually _is_ negative.
[bbi-88690.ll.gz](https://github.com/llvm/llvm-project/files/13306009/bbi-88690.ll.gz)
https://github.com/llvm/llvm-project/pull/71534
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
