[llvm] [RISCV][CostModel] Refine Arithmetic reduction costs (PR #79103)

[Shih-Po Hung via llvm-commits]

arcbbb wrote:

> LGTM. Assuming that the spec tests were built with -O2/-O3 it makes sense that there's not any changes, seeing as most of the diffs here are for the size cost.

Yes, the command I use is
```
cmake -DTEST_SUITE_RUN_BENCHMARKS=Off -DARCH=RISCV \
      -DTEST_SUITE_COLLECT_CODE_SIZE=OFF \
      -DTEST_SUITE_COLLECT_STATS=ON \
      -DOPTFLAGS="-march=rv64gcv_zfh_zba_zbb_zvfh_zvl512b -menable-experimental-extensions" \
      -DTEST_SUITE_SPEC2017_ROOT=$SPEC2017 \
      -DTEST_SUITE_SPEC2006_ROOT=$SPEC2006 \
      -DCMAKE_C_COMPILER=$CLANG \
      -DCMAKE_CXX_COMPILER="$CLANG++" \
        -C../test-suite/cmake/caches/O3.cmake \
        ../test-suite
make -j
```
and the result CXXFLAGS are
`CXX_FLAGS = -march=rv64gcv_zfh_zba_zbb_zvfh_zvl512b -menable-experimental-extensions -O3   -w -Werror=date-time -save-stats=obj -Wno-implicit-function-declaration -DSPEC_CPU -DSPEC_CPU_LINUX -DSPEC_CPU_LITTLEENDIAN -DSPEC_CPU_LITTLE_ENDIAN -DSPEC_CPU_LP64 -std=c++98 -save-stats=obj`

https://github.com/llvm/llvm-project/pull/79103