The previous attempt was #4654 and it ended up being just memcmp removal; see #4654 (comment)

Benchmark results on my 5950X, split into separate tables for 1 and 2 bytes versus 4 and 8 bytes:

Aside from c_strstr which is of course unchanged, I'm also seeing across-the-board massive improvements for 1 and 2 bytes, so this is great.

I am mildly confused as to why performance for search_default_searcher seems to vary for 4 bytes (better) and 8 bytes (worse) for this PR, when it shouldn't have been altered at all - the if constexpr should be completely vanishing. Codegen gremlins? I don't think it should block merging though.

I am mildly confused as to why performance for search_default_searcher seems to vary for 4 bytes (better) and 8 bytes (worse) for this PR, when it shouldn't have been altered at all - the if constexpr should be completely vanishing. Codegen gremlins? I don't think it should block merging though.

I guess the biggest of codegen gremlin is exact loop alignment. The compiler only align functions to 16-byte boundary, whereas apparently like 32 or 64 bytes boundary in important. You may try /QIntel-jcc-erratum, (yes, even despite you run on AMD!) for both main and changed code, build whole import lib and the benchmark executable with it, and see if this variation disappears.

I've seen this happening even when changing unrelated functions. That's why it doesn't worth hunting for -- eventually we will add or change even more unrelated functions, and alignment would change again.

Thanks, makes sense!

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

🔍 🕵️ 🔎