enable MEM_FORCE_MEMORY_ACCESS=2 for RISC-V targets with zicclsm

[Polaris-911]

Summary

This PR updates lib/common/mem.h to select MEM_FORCE_MEMORY_ACCESS=2 for RISC-V builds when both __riscv and __riscv_zicclsm are defined.

Motivation

Zicclsm indicates support for misaligned loads/stores in main memory.
For these targets, using method 2 enables direct unaligned memory access in mem.h.
References
GCC Zicclsm
RVA20U64 specification

Changes

• Updated the default MEM_FORCE_MEMORY_ACCESS selection logic:
  • MEM_FORCE_MEMORY_ACCESS=2 when defined(__riscv) && defined(__riscv_zicclsm)
  • Keep existing GCC fallback: MEM_FORCE_MEMORY_ACCESS=1 for other GCC targets
• No changes to behavior when MEM_FORCE_MEMORY_ACCESS is explicitly set by the build system.

Validation

• Verified the updated preprocessor branch compiles cleanly in mem.h.
• No linter issues reported for the modified file.

Benchmark Data (Compression Speed)

Method	zstd -1 Speed	vs Method 0
Method 2	74.2 MB/s	+74%
Method 1	59.2 MB/s	+39%
Method 0	42.5 MB/s	-

Data screenshot

[terrelln]

This looks good to me. Ultimately, it would be better if the compiler was able to handle optimizing memcpy() to use unaligned loads & stores on RISC-V with zicclsm. Once they can do that we can turn this off.

CC @Cyan4973 in case you have an opinion, before I merge

[terrelln]

@Polaris-911 I don't see the same improvement to decompression speed as we see in PR #4584.

Do you know what function the difference in decompression speed is coming from between these two PRs? Other than these functions, the decompressor really shouldn't be doing any unaligned accesses, so I don't know where the difference would be coming from.

[Polaris-911]

@terrelln Thanks for pointing that out! I suspect the difference in decompression speed might be related to ZSTD_wildcopy rather than the MEM_read/MEM_write macros.

While the decompressor doesn't directly use MEM_read* for data manipulation, ZSTD_execSequence relies heavily on ZSTD_wildcopy (which eventually calls __builtin_memcpy via ZSTD_copy8/ZSTD_copy16) to copy both literals and matches.

[Polaris-911]

@terrelln @Cyan4973 Hi , I submitted a similar PR( #4524) before, and at that time @Cyan4973 was concerned about undefined behavior (UB).

I completely understand @Cyan4973's concerns about Undefined Behavior and that the compiler should handle memcpy optimizations natively (Method 0). However, given the current state of RISC-V compilers, they still fall short in optimizing memcpy for unaligned accesses, leaving a significant performance gap on hardware that supports it (like with zicclsm).

Would you be open to accepting either this MEM_FORCE_MEMORY_ACCESS=2 approach or explicitly adding -mno-strict-align to the RISC-V build flags(#4584) as a temporary workaround? We can absolutely revert this once the upstream compilers improve their memcpy optimizations for RISC-V. This would allow users to benefit from the immediate performance gains in the meantime.