[PATCH] D154812: [Support] Add llvm::xxh3_64bits

Tue Jul 18 12:45:37 PDT 2023

MaskRay added a comment.

In D154812#4511112 <https://reviews.llvm.org/D154812#4511112>, @andrewng wrote:

> I haven't looked at the code but I have built `ld.lld` with this patch and the patch from D154813 <https://reviews.llvm.org/D154813> on Windows with both VS 2022 `cl` 19.36.32537 and `clang-cl` 16.0.6. I have run the usual benchmarks (clang, chrome, mozilla, scylla) and also a UE5 link and I don't see any difference in performance on my Windows 10 PC, i.e. any delta is in the measurement "noise".

Thank you for benchmarking this on Windows. Do these programs have large `.debug_str` sections? (Windows builds normally use CodeView/PDB, not DWARF).
I think only any effect is likely only observed with large `.debug_str` sections. To observe the effect of replacing the hash function, I need to compute the time ratio of hashing

> The remaining part of lld remains unchanged. Consequently, a lower ratio indicates that hashing is faster.

I have observed ~2% speedup on an x86-64 machine, but the result on an aarch64 machine (Cavium ThunderX2) is so-so.
In PiotrZSL's setup the speedup seems larger?

I think xxh3 is still worth it and we can probably drop the xxh64 implementation once internal implementations have migrated.

  % hyperfine --warmup 1 --min-runs 16 "numactl -C 16-23 "{/tmp/lld0,/tmp/lld1}" -flavor gnu @response.txt --threads=8"
  Benchmark 1: numactl -C 16-23 /tmp/lld0 -flavor gnu @response.txt --threads=8
    Time (mean ± σ):      7.393 s ±  0.089 s    [User: 17.077 s, System: 4.274 s]
    Range (min … max):    7.208 s …  7.561 s    16 runs

  Benchmark 2: numactl -C 16-23 /tmp/lld1 -flavor gnu @response.txt --threads=8
    Time (mean ± σ):      7.351 s ±  0.093 s    [User: 16.757 s, System: 4.236 s]
    Range (min … max):    7.165 s …  7.519 s    16 runs

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D154812/new/

https://reviews.llvm.org/D154812