Revert "Use LLVM intrinsics for madd intrinsics"

[folkertdev]

This reverts commit 3214671.

The commit is from #1985, which itself reverted this original change.

Use intrinsics::simd for madd (again), now that llvm/llvm-project#174149 should make this optimize properly.

[folkertdev]

See also https://godbolt.org/z/8z53f6WPE

r? sayantn

[sayantn]

Please just append this to the test_mm256_madd_epi16 function. Or if you really want to keep it as a separate function, just move it to the tests module

[folkertdev]

The test checks that the instruction is emitted, so it can't just be added to that existing test. Locally it didn't run when I up it with the other tests, but on CI it does, so I've moved it now.

[usamoi]

This seems like it would break the usage of base64. https://godbolt.org/z/q6e1ne568 ¹²

I have some concerns about this kind of change. It might introduce reliance on hard-to-predict optimizer behavior and lead to performance issues that are difficult to detect.

Footnotes

1. https://github.com/ozgb/bs64/blob/ef033b373cb08d76808da9242fe57e89b2eff002/src/avx2/mod.rs#L136 ↩

2. https://github.com/aki-akaguma/basexx/blob/536976ff8aefb8299e0dba335a34198c6ed62d79/src/base64/base64_avx2.rs#L158 ↩

[folkertdev]

Interesting, I think it uses a shift instead of a multiply and that fools the pattern match. Yeah that needs a fix.

Do you have any other cases that don't optimize?

I'll revert this, and then add separate tests for the adler32 and base64 patterns.

[usamoi]

Do you have any other cases that don't optimize?

Can't find more on GitHub. https://github.com/search?q=_mm_madd_epi16+NOT+is%3Afork+language%3ARust&type=code

If not limited to real-world cases, there are also cases like _mm256_madd_epi16(v, _mm256_set1_epi16(-1)) that might make sense.