| Issue |
128488
|
| Summary |
LLVM fails to recognize the same condition sometimes
|
| Labels |
missed-optimization
|
| Assignees |
|
| Reporter |
philnik777
|
In libc++ we have the following string representation:
```c++
namespace std {
using size_t = __SIZE_TYPE__;
struct string {
union {
struct {
size_t __is_long_ : 1;
size_t __long_cap_ : 63;
size_t __long_size_;
char* __long_data_;
};
struct {
unsigned char : 1;
unsigned char __short_size_ : 7;
char __short_data_[23];
};
};
void __set_size(size_t __s) {
if (__is_long_) {
__long_size_ = __s;
} else {
__short_size_ = __s;
__is_long_ = false;
}
}
char* __get_pointer() {
return __is_long_ ? __long_data_ : &__short_data_[0];
}
};
}
auto test1(std::string& str, std::size_t size) {
str.__set_size(size);
return str.__get_pointer();
}
auto test2(std::string& str, std::size_t size) {
auto __ptr = str.__get_pointer();
str.__set_size(size);
return __ptr;
}
```
I've noticed recently that the code generated for the two functions above are significantly different, even though they are semantically equivalent. For some reason `test1` generates significantly worse code than `test2`. https://godbolt.org/z/15czrP69d
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
