[libcxx-commits] [libcxx] [libc++] Optimize `std::exception_ptr` (PR #162773)

[Adrian Vogelsgesang via libcxx-commits]

vogelsgesang wrote:

> I'm not convinced 2 and 3 are a good idea so far, so I'd rather do them later. I'm happy with (1) though and I think (4) can be implemented separately from moving any code (you seem to have come to the same conclusion below?)

I created
* for "(1) introduce benchmark": #164278
* for "(4) introduce move ctor/assign": #164281

For PR #164281, we already get measurable performance improvements:

```
Benchmark                          Baseline    Candidate    Difference    % Difference
-------------------------------  ----------  -----------  ------------  --------------
bm_nonnull_exception_ptr              52.22        40.92        -11.31          -21.65
bm_null_exception_ptr                 31.41        23.29         -8.12          -25.85
bm_optimized_null_exception_ptr       28.69        20.50         -8.19          -28.55
```

With this PR (#162733), we get even better numbers (measured relative to #164281)

```
Benchmark                          Baseline    Candidate    Difference    % Difference
-------------------------------  ----------  -----------  ------------  --------------
bm_nonnull_exception_ptr              40.92        33.69         -7.23          -17.66
bm_null_exception_ptr                 23.29         1.21        -22.07          -94.79
bm_optimized_null_exception_ptr       20.50         0.70        -19.80          -96.61
```

I hope that we can ship PR #164278 (introducing the benchmark) pretty much as is.

I am not 100% happy with "(4) introduce move ctor/assign" because I had to
1. the glibcxx implementation of `__do_decrement_refcount` relies on `_M_add_ref`. But this is inconsistent now with other functions, as I did not do "(2) change `src/exception_pointer_glibcxx.ipp` to use `_M_{add,release}ref`" first
2. introduce a new `__do_decrement_refcount` function to the library's ABI. I wasn't able to keep the move assignment completely contained to the header. Doing so would have required me to first do "(3) expose the selected ABI via `__config_site.in`"

https://github.com/llvm/llvm-project/pull/162773