[libcxx-commits] [libcxx] [ASan][libc++] Optimize `__annotate_delete` for the default allocator (PR #76176)

[via libcxx-commits]

AdvenamTacet wrote:

I just ran a small benchmark with flags `-stdlib=libc++ -fsanitize=address -O3`.

```cpp
#ifdef VECTOR_FAST
#include "vector-optimized.hpp"
#endif

#ifdef VECTOR_MAIN
#include "vector-main.hpp"
#endif

int main() {
    uint32_t sum = 0;
    for(uint32_t i = 2; i < 100'000; ++i) {
        std::vector<uint64_t> v(1'000'000, i);
        sum += v[i/2];
        v.clear();
    }

    return sum != 0 ? 0 : -1;
}
```

Example results (quite representative):
```bash
# Optimized times:
real    3m6.008s
user    0m25.815s
sys     2m39.491s


# Times from upstream mian:
real    3m32.434s
user    0m27.063s
sys     3m4.495s
```

It gives around 13% of speedup on that benchmark.
```python
In [1]: 1 - (3*60+3.008) / (3*60+32.434)
Out[1]: 0.13851831627705535
```

I don't see many potential issues, maybe problems after replacing default allocator with some problematic allocator?
It only affects strictly ASan related functions, so readability is arguably untouched.
Performance is consistently better.

What do you think?

https://github.com/llvm/llvm-project/pull/76176