[PATCH] D23755: cmake: Add an ordering dep between HTML & man Sphinx targets

Chris Bieneman via llvm-commits llvm-commits at lists.llvm.org
Mon Oct 3 10:22:29 PDT 2016


beanz added a comment.

In https://reviews.llvm.org/D23755#558390, @mgorny wrote:

> Tell that to the person who is running Gentoo on old x86. Because most of binary distros don't even work on CPU that old ;-).


I care very little about minor impacts on downstream consumers of source distributions. I care way more about people who are building LLVM regularly because it impacts us every day. Further, I very strongly believe that we need to build our software thinking about the future, and MP is the present, and the future. Single core CPUs are becoming less and less common every day. Optimizing for them doesn't seem prudent to me.

> I disagree here. We're talking about losing a major cache opportunity. We lose the opportunity to run twice as many Sphinx instances in parallel but the free cores are better utilized compiling objects at the time. If I read the timing right, I would guess around 75% of the CPU time is consumed on building the doctree.
> 
> Besides, Sphinx already does parallel jobs internally.

This is just an argument for why my earlier suggestion of USES_TERMINAL would not be a bad fit. If sphinx significantly used multi-threading you wouldn't want more than one instance of it running at a time. While I know it has multiple threads, based on your single invocation or parallel time tests, I suspect they aren't CPU bound.

> That would be nice but I should point out that the 'dummy' builder is not available as separate builder in sphinx-1.3.5 (which is the Gentoo current for some reason). It seems that upstream split it out 3 months ago, so this would mean restricting LLVM to very fresh Sphinx versions.

I didn't realize the dummy builder was that new. Requiring a very new sphinx is not reasonable.


https://reviews.llvm.org/D23755





More information about the llvm-commits mailing list