huf_decompress: enable 4-way fast loop on riscv64

[Polaris-911]

Summary

• In lib/decompress/huf_decompress.c, enable HUF_4X2_4WAY for riscv64 (defined(__riscv) && __riscv_xlen == 64).
• Keep existing logic unchanged for all other architectures.
• No functional/format changes: only changes fast-loop scheduling policy in the hot decode path.

Why this change is reasonable

• The HUF_4X2_4WAY switch controls whether the fast loop decodes all 4 streams in the main unrolled body, or uses a 3-way fallback to reduce register pressure.
• This loop is already enabled as 4-way on aarch64, which indicates the 4-way variant is considered stable and beneficial on modern 64-bit RISC-style cores with larger register files.
• riscv64 has similar 64-bit execution characteristics for this scalar path (64-bit container ops, stream-parallel decode pattern), so selecting 4-way is a reasonable architecture-level default.
• Scope is intentionally conservative: only riscv64 is opted in; rv32 and all non-RISC-V targets keep existing behavior.
• If performance differs by compiler/micro-architecture, this change is still safe functionally because both paths implement the same decode semantics.

[terrelln]

Do you have benchmarks for this?

[Polaris-911]

Benchmark Results (silesia.tar, level 1)

Hi @terrelln Here are the benchmark results comparing the baseline zstd against zstd_huf with #define HUF_4X2_4WAY 1 enabled across 5 iterations:

Run	zstd Comp (MB/s)	zstd Decomp (MB/s)	zstd_huf Comp (MB/s)	zstd_huf Decomp (MB/s)
1	91.3	235.3	91.8	235.8
2	92.1	227.3	90.6	231.6
3	91.2	236.4	91.4	235.5
4	92.2	233.0	91.7	236.8
5	91.5	234.7	91.5	237.6
Average	91.66	233.34	91.40	235.46
Difference	-	-	-0.28%	+0.91%

Summary:
After setting #define HUF_4X2_4WAY 1, the decompression speed shows a consistent improvement (+0.91%). The slight variance in compression speed (-0.28%) is within the normal margin of error.