| Issue |
140997
|
| Summary |
DSEPass incorrectly eliminates store to captured (and later read) pointer
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
mhjacobson
|
llvm v20.1.3
macOS "Big Sur" 11.7.10 (20G1427)
I have this code:
```
declare i32 @printf(i8* noalias readonly, ...)
@fmt = constant [6 x i8] c"%hhd\0a\00"
define ptr @makeValue(ptr %p) noinline optnone memory(none) {
ret ptr %p
}
define void @useValue(ptr %value) noinline optnone memory(argmem: read) {
%val = load i8, ptr %value
call i32 @printf(ptr @fmt, i8 %val)
ret void
}
define i32 @main() {
entry:
%1 = alloca i8, align 4
store i8 42, ptr %1
%value = call i64 @makeValue(ptr %1)
call void @useValue(i64 %value)
ret i32 0
}
```
`opt --passes=dse` transforms `@main()` to (<https://godbolt.org/z/rhx1z9K4W>):
```
define i32 @main() {
entry:
%0 = alloca i8, align 4
%value = call i64 @makeValue(ptr %0)
call void @useValue(i64 %value)
ret i32 0
}
```
i.e., it eliminates the store to `%1`, even though `%1` is captured by the call to `@makeValue()` and ultimately read in `@useValue()`.
I believe the `memory` annotations on `@makeValue()` and `@useValue()` are legal. Note that the `noinline optnone` are just to prevent the optimizer from seeing into the contents of the function, equivalent to if the functions were simply `declare`d and implemented in another module.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
