CMake dependencies and building LLVM

Project Infrastructure LLVM Dev List Archives
1

Hi everyone,

I have felt for quite some time that during my regular LLVM workflow, CMake recompiles a lot of files after I update to the latest ToT (which I do more than once per week). Of course, this wasn’t really based on any real data, just felt like every time I update (even if it’s a day or two later), we end up recompiling everything. For a while I assumed that patches touch a few headers that are used all over the place or something like that, but this was happening with what seems like unexpected frequency.

So what I’ve done is do a few update/rebuild cycles for revisions that are 200 apart, then 100, then 50, then 25. And it seems that quite commonly, we recompile more than 50% of the source with just a few landed revisions. I’ve included the numbers below (and I do realize that this is a fairly small sample).

If this seems reasonable to most of the community, so be it - I work on a large server that chews through this stuff pretty quickly so I can certainly live with it. On the other hand, if this seems way higher than what you expect, it could very well be that something is wrong in my setup and I’d appreciate some pointers about how to go about investigating this. Finally, if others are seeing this and we as a community do feel that something can/should be done about this (although I don’t know what that might be), perhaps we can start that discussion.

Setup: CMake with -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_ASSERTIONS=ON -G Ninja

The numbers here are the total numbers of targets ninja reports as having rebuilt. Ninja is invoked only -j <N> for parameters (i.e. not check-all or any specific targets).

RevFrm, RevTo, NumNinjaObjsRebuilt
300000, 300200, 2957
300200, 300400, 2692
300400, 300600, 3001
300600, 300800, 2877
300800, 301000, 3003
301000, 301200, 2750
301200, 301400, 2579
301400, 301600, 2709
301600, 301700, 2325
301700, 301800, 2357
301800, 301900, 2667
301900, 302000, 2466
302000, 302100, 2499
302100, 302200, 2063
302200, 302250, 2884
302250, 302300, 2007
302300, 302350, 2209
302350, 302400, 737
302400, 302450, 2398
302450, 302500, 2856
302500, 302550, 2090
302550, 302600, 2866
302600, 302625, 1786
302625, 302650, 1966
302650, 302675, 1583
302675, 302700, 307
302700, 302725, 2246

2

I’m /guessing/ this is probably pretty normal - but likely the best solution would be to get down to granularity of 1 commit in each of those ranges, to see if it’s really justified that the change causes so much recompilation (ie: do all the things that got compiled really depend on the modified file)

Some of this might be tablegen recompiling - but I /think/ that ninja has some support/seems to do the right thing when the resulting tablegen binary isn’t different from previously (or maybe when it doesn’t generate different output - this would be better) & doesn’t cause everything to rebuild (you’ll notice Ninja sometimes starts with a very high estimate (1000+) of targets to rebuild, but the estimate is adjusted down to only a couple of hundred in many cases probably due to this.

3

I’m /guessing/ this is probably pretty normal - but likely the best solution would be to get down to granularity of 1 commit in each of those ranges, to see if it’s really justified that the change causes so much recompilation (ie: do all the things that got compiled really depend on the modified file)

Some of this might be tablegen recompiling - but I /think/ that ninja has some support/seems to do the right thing when the resulting tablegen binary isn’t different from previously (or maybe when it doesn’t generate different output - this would be better) & doesn’t cause everything to rebuild (you’ll notice Ninja sometimes starts with a very high estimate (1000+) of targets to rebuild, but the estimate is adjusted down to only a couple of hundred in many cases probably due to this.

I can confirm that tablegen-erated files do the right thing and that it’s based on the output of tablegen. Our cmake scripts do it by having llvm-tblgen write to a temporary file and then use ‘cmake -E copy_if_different’ to update the real one when necessary. Once ninja has run the copy_if_different command it re-checks the dependencies and culls the ones it doesn’t need.

4

Yes, I have seen TableGen creating tempoaries and doing the right thing. Keep in mind that the stated numbers are final numbers (i.e. When ninja is done). I’ll try with some single changesets to try to narrow down what it is that causes a lot of recompilation.

5

For what it’s worth:

  • I’m usually seeing similar behavior even in shorter timeframes like updating after half a day
  • From all I could tell the buildsystem was doing the right thing. There’s just a lot of headers in Support and ADT that are included pretty much everywhere and people change them (legitimately).

Sure you may find the occasional instance where doing some class XXX; forward declaration instead of #include “XXX.h” would improve dependencies. But overall LLVM code is already amazingly clean in this regard so I don’t see a lot of room for improvement there either.

  • If you haven’t seen it look at Tillmanns excellent blogpost about all the ways to optimize an LLVM build: https://blogs.s-osg.org/an-introduction-to-accelerating-your-build-with-clang/

  • The biggest savings for me personally are disabling clang builds and restricting myself to the targets I am developing for

  • Either way it’s never fast and I end up drinking way too much coffee because it’s so easy to go down the hall to grab another one while the build is running…

  • Matthias