| Issue |
174933
|
| Summary |
Missed optimization x86-64 clang 21.1.0: lock xadd not replaced with lock add
|
| Labels |
clang
|
| Assignees |
|
| Reporter |
piotr-topnotch
|
```
#include <atomic>
bool fn_sub(std::atomic_uint64_t& a, std::uint64_t n) noexcept {
return a.fetch_sub(n, std::memory_order_relaxed) == n;
}
bool fn_add(std::atomic_uint64_t& a, std::uint64_t n) noexcept {
return a.fetch_add(-n, std::memory_order_relaxed) == n;
}
```
The first function correctly undergoes strength reduction replacing xadd with sub:
```
fn_sub(std::atomic<unsigned long>&, unsigned long):
lock sub qword ptr [rdi], rsi
sete al
ret
```
Whereas the second pattern is not recognized:
```
fn_add(std::atomic<unsigned long>&, unsigned long):
mov rax, rsi
neg rax
lock xadd qword ptr [rdi], rax
cmp rax, rsi
sete al
ret
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
