Misc. bug: Incorrect usage of __riscv_v_intrinsic macro - should use __riscv_v for RVV separation

[arniechangsifive]

Name and Version

version: 8683 (d0a6dfe)
built with Clang 20.9.0 for Linux riscv64

Operating systems

Linux

Which llama.cpp modules do you know to be affected?

llama-cli

Command line

Problem description & steps to reproduce

Summary
The llama.cpp codebase incorrectly uses __riscv_v_intrinsic for conditional compilation to separate RISC-V Vector (RVV) implementation from scalar code. According to the official RISC-V "V" Vector Extension Intrinsic specification, __riscv_v should be used instead.

The Issue
According to the RISC-V "V" Vector Extension Intrinsic Document, section on predefined macros:
__riscv_v_intrinsic: Version number macro for the intrinsic API
This macro should be defined even if the vector extension is not enabled.
__riscv_v: RVV extension detection macro
Only defined when the vector extension is actually enabled. This is the correct macro for RVV/scalar code separation

Current State in llama.cpp
The codebase has inconsistent usage:

1. Correct usage (older code):
   

   llama.cpp/ggml/src/ggml-cpu/arch/riscv/quants.c

   Line 39 in 81df3f7

   #if defined(__riscv_v)

2. Incorrect usage (newer code)
   

   llama.cpp/ggml/src/ggml-cpu/arch/riscv/quants.c

   Line 2061 in 81df3f7

   #if defined __riscv_v_intrinsic

Impact
Using __riscv_v_intrinsic instead of __riscv_v for RVV separation can lead to runtime crashes, vector instructions executed on hardware without vector support.

First Bad Commit

No response

Relevant log output

Logs