| Issue |
149821
|
| Summary |
LLVM doesn't keep nullability information around when inlining
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
philnik777
|
Consider
```
struct S {
int* i;
};
[[gnu::returns_nonnull]] int* get(S& s) { return s.i; }
auto test1(S& s) {
auto val = get(s);
if (!val)
__builtin_trap();
return *val;
}
auto deref(int* val) {
if (!val)
__builtin_trap();
return *val;
}
auto test2(S& s) {
auto val = get(s);
return deref(val);
}
```
`test1` gets optimized better than `test2`, even though the exact same information should be available to the compiler. It looks like LLVM doesn't keep the information that `s.i` can never be null when inlining `get`. This can be relevant for classes like `string`, where `data()` can never return a null pointer, but annotating the function doesn't help to remove redundant checks, since the annotation is lost as soon as the function is inlined.
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
