[PATCH] D26855: New unsafe-fp-math implementation for X86 target

[Chandler Carruth via Phabricator via llvm-commits]

chandlerc added a reviewer: scanon.
chandlerc added a comment.

In https://reviews.llvm.org/D26855#647135, @avt77 wrote:

> I fixed everything except one comment (see below).  And I collected new perf numbers.  Now I used the following command for bootstrap building:
>
> time make -j 1
>
> As result the reproducing is very well from my point of view. In addition I tried to get numbers accordingly to description in  http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20161017/398831.html . The reproducing is almost the same but the required time is even longer (about 2 hours for every test). Because of that I kept my numbers. The test itself is very simple: I created 2 versions of compiler: the first one was built directly from trunk and the second one was built after applying of my patch. Then with help of every compiler I created 4 bootstraps. The results are below:


So, I'm mostly lurking, but I want to point out a serious issue here: Clang and LLVM have as little floating point as we could manage in them. So I would expect them to be quite uninteresting in testing the compile time impact of a patch that is only concerned with floating point code....

And there still are no numbers around the improvement here...

I suspect you'll need to provide benchmark data from at least SPEC and/or the LLVM test suite that shows this is an improvement and compile time numbers to show that the improvement doesn't cost too much... At least, that would be my expectation. Those at least do include some floating point code. You might also try running benchmarks from the Eigen project which has a very large amount of floating point code. However, they usually don't build with any unsafe math flags, so correctness issues may dominate.

It'd also be great to hear from others invested in LLVM's FP lowering like Hal, Steve, etc...


https://reviews.llvm.org/D26855