[chiranmoyf]

Motivation

Decompression on AArch64 (e.g., AWS Graviton3) currently doesn't take advantage of SVE (Scalable Vector Extension) instructions, leaving performance on the table in the Huffman decode and sequence decompression hot paths.

Changes

• Introduces a DYNAMIC_SVE build macro, auto-enabled for Clang/GCC on aarch64, allowing SVE support to be detected and used at runtime.
• Adds SVE_TARGET_ATTRIBUTE (target("+sve")) so the compiler can generate SVE-enabled variants of the Huffman fast-decode loops and ZSTD_decompressSequences, alongside the existing default implementations.
• At decompress time, the SVE variant is selected automatically when supported, falling back to the default path otherwise.

Result

~2–3% faster decompression on Graviton3 (fullbench), with no change to compression speed.
Benchmarked using fullbench.
Instance: m7g.8xlarge (Graviton 3)

#	Benchmark	Baseline (MB/s)	SVE (MB/s)	Diff (%)
1	decompress	1216.9	1248.4	+2.59%
3	decompressDCtx	1215.3	1250.0	+2.86%
6	decompressContinue	1217.5	1243.8	+2.16%
9	decompressStream	1218.7	1243.7	+2.05%

Related

#4564

This PR is a joint contribution by:

Abhishek Jain(@abhijain1204fujitsu)
Mohak Goel(@mohakgoel1)

[chiranmoyf]

cc @iksaif @terrelln @Cyan4973 would appreciate your feedback when you get a chance.