BLAKE3 SIMD performance on MSVC vs GCC

[Coralesoft]

Hi,
I've added BLAKE3 to https://github.com/cryptopp-modern/cryptopp-modern
with SSE4.1 (4-way) and AVX2 (8-way) parallel chunk hashing. All test vectors pass, thanks for the algorithm!

Benchmarks (Intel Core Ultra 7 155H, Windows 11, 64-bit Release, 16 KiB buffer):

Compiler	Flags	AVX2 speed
MinGW GCC 14.2	-O3 -msse4.1 -mavx2	2591 MiB/s
MSVC 2022	/O2 /arch:AVX2	1829 MiB/s

Same code, but MSVC is ~29% slower. I also tried /Ob3, but it made things worse (1766 MiB/s ~32% slower then gcc).

Have you seen similar compiler differences with your implementations?

Any known MSVC pitfalls or recommended flags/idioms for BLAKE3-style SIMD code?

Thanks.

[BurningEnlightenment]

The CMake buildsystem of this repository uses the intrinsic-based SIMD code only as a fallback for the handwritten assembler SIMD code. Therefore I wouldn't expect there to be much experience with or tuning for a MSVC build. Is there a reason to not use the ASM based code?

[Coralesoft]

Thanks for the explanation.

While cryptopp-modern has some existing asm, for BLAKE3 I went with a pure intrinsics implementation to keep the new code in portable C++ across compilers and platforms. Pulling in the BLAKE3 asm is an option, but it would add more per-architecture code and build logic to maintain.

I opened this mainly to sanity check whether a roughly 30% GCC vs MSVC gap on AVX2/SSE4.1 intrinsics-heavy code matches what you’ve seen elsewhere, and whether there are any MSVC-specific pitfalls or flags you’d recommend.

Thanks again for clarifying how this repo prefers the asm path over the intrinsics fallback.

[BurningEnlightenment]

AFAIK MSVC has a bad register allocator, i.e. it usually needlessly spills register content to the stack only to read it back a few instructions later. I don't think there's anything we can do about that, though.