This week I have benchmarked some of the hooks that have been templated a while ago by Teo, using LDC. I have published the results in an upcoming PR [1].

While performing this benchmarks, I have noticed some regressions with _d_arrayctor and _d_newarrayT. The cause of the regression for _d_arrayctor was an unoptimal branching condition in the hook logic, which prevented it from using memcpy in some instances [2]. For _d_newarrayT, on the other hand, the regression is a problem only with LDC, as it still makes use of the old TypeInfo-based hook instead of the templated one. The solution is not as straightforward as I initially thought, because of the extra LLVM GC2Stack pass it performs and so I will need to investigate further [3].

Additionally, I have added some missing exception handling in _d_arrayappendT for postblit exceptions [4], revealing at the same time a bug in the postorder traversal of CatAssignEp and CatExp [5].

Finally, I have moved the lowering of ConstructExp to a dedicated lowering field, eliminating the need for rewriting the original expression in dinterpret.d and overall making it consisten with the other lowered expressions [6].

[1] https://github.com/teodutu/druntime-hooks-benchmarks/pull/9
[2] https://github.com/dlang/dmd/pull/21675
[3] https://github.com/ldc-developers/ldc/pull/4965
[4] https://github.com/dlang/dmd/pull/21701
[5] https://github.com/dlang/dmd/pull/21699
[6] https://github.com/dlang/dmd/pull/21706

On Tuesday, 12 August 2025 at 09:03:55 UTC, AlbertG wrote:

I appreciate the focus put on maintaining the lowest run-time cost as possible.

On Wednesday, 13 August 2025 at 14:16:33 UTC, user1234 wrote:

Thank you!