Further investigate performance differences vs riscv-gcc

Nobody has claimed this yet.

Using benchmarks from https://github.com/rv8-io/rv8-bench/ I see the current dynamic instruction counts (note: this isn't a perfect proxy for performance, but reasonable for initial comparisons):

Benchmark aes with -march=rv32im and -O3
gcc instret: 283330319
clang instret: 271796365
clang/gcc: 95.00%
Benchmark dhrystone with -march=rv32im and -O3
gcc instret: 274275911
clang instret: 336272568
clang/gcc: 122.00%
Benchmark norx with -march=rv32im and -O3
gcc instret: 129801139
clang instret: 134429519
clang/gcc: 103.00%
Benchmark qsort with -march=rv32im and -O3
gcc instret: 814554722
clang instret: 692674566
clang/gcc: 85.00%
Benchmark sha512 with -march=rv32im and -O3
gcc instret: 227057803
clang instret: 274505822
clang/gcc: 120.00%

clang/gcc < 100% means there are fewer dynamic instructions for the clang-generated binary, and > 100% means fewer dynamic instructions for the gcc-generated binary.

Taken with the current version of this clang+llvm patchset and a recent gcc 8.0.0 build. We did better than gcc 7.1.1 for dhrystone, but gcc has since improved.

Start with the benchmarks in https://github.com/rv8-io/rv8-bench/ and reproduce the reported rv32im -O3 instruction counts for clang and riscv-gcc. Compare the generated results to determine which benchmark differences need further investigation; the issue does not specify a particular file, test, or acceptance criterion.

Written by the indexing model from the issue text.