GCC vs. LLVM Clang vs. AOCC Compilers On AMD Threadripper

[phoronix]

Phoronix: GCC vs. LLVM Clang vs. AOCC Compilers On AMD Threadripper

Given recent improvements to AMD Zen (znver1) with LLVM, the new AMD AOCC 1.2 compiler release, and GCC 8.1 having premiered just weeks ago, here is a fresh look at the performance of six different C/C++ code compilers when testing the performance of the resulting binaries on an AMD Threadripper 1950X system.

http://www.phoronix.com/vr.php?view=26336

[tildearrow]

Possible typo:

to the CLang-based compilers.

[schmidtbag]

Just to clarify: are these optimizations supposed to improve the performance of the compilation process, or the performance of the compiled product? Because considering this is supposed to be an architecture-specific optimization, these results are overall very unimpressive. But, if the output performs better then I could see this being worth using.

[reavertm]

Compiled binary performance ("more is better" gives it away). Results look pretty homeopathic, compared to baseline clang. Or baseline clang inherited some AMD optimisations already?

[Meteorhead]

I too think that the results seem very unimpressive. It looks to me that the effort put into AOCC up until now would've been more useful placed elsewhere. GCC and Clang seem to be doing fine. If there is compiler dev FTE to spare, it should be put into mainline SPIR-V back-end for LLVM. That would have far greater use to AMD as well.

[carewolf]

Possible typo:

Looks right to me, if a little old fashioned. Clang comes from C-Lang frontend.

[oleid]

I too think that the results seem very unimpressive. It looks to me that the effort put into AOCC up until now would've been more useful placed elsewhere. .

AFAIK AMD's patches are going to be upstreamed to vanilla clang, however, for the time being, they're distributing pre-compiled binaries for faster distribution.

[oleid]

What's wrong with clang in "Timed PHP Compilation v7.1.9" benchmark? Isn't clang supposed to compile things faster than gcc? Or is this test doing something else than compiling php?

[mlau]

gcc's scimark2 scores can be improved (~10-15%) with "-flto -funroll-loops". MonteCarlo score almost triples with lto, and the LU benchmark also improves by almost 10% with funroll-loops; clang has a better vectorizer as shown by the sparse matmult test (although it has regressed significatly since clang-4 there as well).

[hubicka]

gcc's scimark2 scores can be improved (~10-15%) with "-flto -funroll-loops". MonteCarlo score almost triples with lto, and the LU benchmark also improves by almost 10% with funroll-loops; clang has a better vectorizer as shown by the sparse matmult test (although it has regressed significatly since clang-4 there as well).

On my skylake notebook I get improvement for matmul when vectorizer is disabled (i.e. -fno-tree-vectorize).
Clang 5 does not vectorize the benchmark. So it is misbehaving heuristics perhaps hoping for matrix to be denser