| Issue |
90936
|
| Summary |
miscompile related to coalescing stores in AArch64 SDAG backend
|
| Labels |
backend:AArch64,
llvm:codegen,
miscompilation,
llvm:SelectionDAG
|
| Assignees |
|
| Reporter |
regehr
|
here's a function:
```llvm
define void @f(i16 %0, ptr %1) {
%3 = trunc i16 %0 to i8
%4 = trunc i16 %0 to i14
%new0 = lshr i14 %4, 8
store i8 %3, ptr %1, align 1
%5 = getelementptr i8, ptr %1, i64 1
%6 = trunc i14 %new0 to i8
store i8 %6, ptr %5, align 1
ret void
}
```
during codegen, the two 8-bit stores are getting merged:
```
_f:
strh w0, [x1]
ret
```
but the backend seems to miss that the trunc changed the value, it dropped those 2 high bits.
I also have versions of this that improperly merge 16-bit stores and 32-bit stores.
we can demonstrate using this driver:
```c
#include <stdio.h>
void f(unsigned short, unsigned char *);
int main(void) {
unsigned char mem[8];
for (int i=0; i<8; ++i)
mem[i] = 0xee;
f(0xFFFF, mem);
for (int i=0; i<8; ++i)
printf("%x\n", mem[i]);
}
```
```
Johns-MacBook-Pro:tmp regehr$ ~/llvm-project/for-alive/bin/llc foo.ll
Johns-MacBook-Pro:tmp regehr$ clang driver.c foo.s && ./a.out
ff
ff
ee
ee
ee
ee
ee
ee
Johns-MacBook-Pro:tmp regehr$
```
in contrast, global isel properly notices that those high bits were chopped off:
```
Johns-MacBook-Pro:tmp regehr$ ~/llvm-project/for-alive/bin/llc foo.ll -global-isel
Johns-MacBook-Pro:tmp regehr$ clang driver.c foo.s && ./a.out
ff
3f
ee
ee
ee
ee
ee
ee
```
cc @hatsunespica
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
