Process five symbols per stream per iteration on AArch64.

[DivineOb]

Decode five symbols per stream per iteration in X1 huffman decompression on AArch64 rather than the default 4. The x86 assembly version already implements this change. Doing so gives a modest decompression speedup on Neoverse N1. Because the portion of runtime used by huffman compression is small this represents a significant speedup to those functions.

gcc: 11.2.0
clang: 14.0.6-2
Tests: silesia.tar
Platform: Neoverse N1

Decompression speedup for different compression levels.

Level         | Clang | gcc
2             | 0.44% | 0.32%
3             | 0.80% | 0.37%
9             | 0.64% | 0.28%
10            | 0.59% | 0.38%
11            | 0.71% | 0.38%

[terrelln]

Sorry for the delay!

We have to retain support for the Huffman table log = 12. Which means that we can't blindly decode 5 symbols per loop (5 * 12 = 60, but we are only guaranteed to have 57 bits in our bitstream). The ASM implementation is only called when tableLog == 11, so it is allowed to make that assumption.

The Zstandard format doesn't actually allow tableLog=12, but we had a bug in our dictionary builder in an early version that could potentially emit tableLog=12. So we want to retain that support.

Rather than just adding a 5th symbol to the loops, I'd likely re-write the decoding loop to use a similar approach to the assembly, somewhat like #3155.

I am going to close this PR in favor of writing an optimized C version of the Huffman decoder in Issue #3425.

Thanks for the PR!

[terrelln]

This is handled in PR #3449. I'd be happy to accept any patches to the fast C decoder that improve aarch64 performance.