Intel Xeon Skylake Compilers: Clang Showing Strong Performance Against GCC

[mayankleoboy1]

This stupid race "Clang compiles faster than ..." reminds of the eighties. (The eighties in the last millenium) There were two companies selling C compilers. (that was for the beloved Amiga) Whenever they had a new release they usually went into "boasting mode" about the compile time of their new product. And then, out of nothing, came a single guy that created a compiler that did not compete with the compile speed. He asked a single question: "Did you have a look at the code quality of the produced code of the compilers?" Then there were long faces and a long silence after that.

more details with names please.

[name99]

The apache benchmark was the last one, and everything was tied. The only benchmarks Clang did better than gcc 6 in was the two compilation time results and 7-zip performance.

Btw, Gcc 6 looks really good in this test. Didn't the last gcc 6 benchmark show almost no improvements? What changed?

For Rodinia, the change is presumably that the benchmark makes good use of some OpenMP 4.5 features, and while these have been added to GCC, they are not yet part of LLVM (though on their way).
I'm guessing for Himeno the difference is that gcc uses graphite while LLVM is not yet using Polly as default; and that C-Ray is some combination of both of these.

In other words apart from these two known missing features, the important take-away is really that LLVM and gcc are at good enough parity. More important going forward for performance will be how easily they integrate out-of-compiler optimizations, in particular LTO and PGO (including the real-world, as opposed to toy benchmark, issues of how well they handle these when trying to integrate with dynamic libraries and third party/OS-supplied libraries). THAT's what tribal warrior wanna-be's (and in future Michael) should be focussing on, not 0.2% differences in code generation.

[smitty3268]

and that C-Ray is some combination of both of these.

The C-Ray benchmark is heavily dependent on how much the compiler chooses to inline code. I believe GCC is more aggressive about doing that by default, which is what leads to the better score.