Where do I start if I would like help improve GHC compilation times?

Alfredo Di Napoli alfredo.dinapoli at gmail.com
Sun Apr 9 09:37:09 UTC 2017

Hey Ben,

as promised I’m back to you with something more articulated and hopefully
meaningful. I do hear you perfectly — probably trying to dive head-first
into this without at least a rough understanding of the performance
hotspots or the GHC overall architecture is going to do me more harm than
good (I get the overall picture and I’m aware of the different stages of
the GHC compilation pipeline, but it’s far from saying I’m proficient with
the architecture as whole). I have also read a couple of years ago the GHC
chapter on the “Architeture of Open Source Applications” book, but I don’t
know how much that is still relevant. If it is, I guess I should refresh my

I’m currently trying to move on 2 fronts — please advice if I’m a fool
flogging a dead horse or if I have any hope of getting anything done ;)

1. I’m trying to treat indeed the compiler as a black block (as you
adviced) trying to build a sufficiently large program where GHC is not “as
fast as I would like” (I know that’s a very lame definition of “slow”,
hehe). In particular, I have built the stage2 compiler with the “prof”
flavour as you suggested, and I have chosen 2 examples as a reference
“benchmark” for performance; DynFlags.hs (which seems to have been
mentioned multiple times as a GHC perf killer) and the highlighting-kate
package as posted here: https://ghc.haskell.org/trac/ghc/ticket/9221 . The
idea would be to compile those with -v +RTS -p -hc -RTS enabled, look at
the output from the .prof file AND the `-v` flag, find any hotspot, try to
change something, recompile, observe diff, rinse and repeat. Do you think I
have any hope of making progress this way? In particular, I think compiling
DynFlags.hs is a bit of a dead-end; I whipped up this buggy script which
escalated into a Behemoth which is compiling pretty much half of the
compiler once again :D

#!/usr/bin/env bash

../ghc/inplace/bin/ghc-stage2 --make -j8 -v +RTS -A256M -qb0 -p -h \
-RTS -DSTAGE=2 -I../ghc/includes -I../ghc/compiler -I../ghc/compiler/stage2
-I../ghc/compiler/stage2/build \
-fforce-recomp -c $@

I’m running it with `./dynflags.sh ../ghc/compiler/main/DynFlags.hs` but
it’s taking a lot to compile (20+ mins on my 2014 mac Pro) because it’s
pulling in half of the compiler anyway :D I tried to reuse the .hi files
from my stage2 compilation but I failed (GHC was complaining about
interface file mismatch). Short story short, I don’t think it will be a
very agile way to proceed. Am I right? Do you have any recommendation in
such sense? Do I have any hope to compile DynFlags.hs in a way which would
make this perf investigation feasible?

The second example (the highlighting-kate package) seems much more
promising. It takes maybe 1-2 mins on my machine, which is enough to take a
look at the perf output. Do you think I should follow this second lead? In
principle any 50+ modules package I think would do (better if with a lot of
TH ;) ) but this seems like a low-entry barrier start.

2. The second path I’m exploring is simply to take a less holistic approach
and try to dive in into a performance ticket like the ones listed here:
Maybe some are very specific, but it seems like fixing small things and
move forward could help giving me understanding of different sub-parts of
GHC, which seems less intimidating than the black-box approach.

In conclusion, what do you think is the best approach, 1 or 2, both or
none? ;)

Thank you!


On 7 April 2017 at 18:30, Alfredo Di Napoli <alfredo.dinapoli at gmail.com>

> Hey Ben,
> thanks for the quite exhaustive reply! I’m on the go right now, but I
> promise to get back to you with a meaningful reply later this weekend ;)
> Alfredo
> On 7 April 2017 at 18:22, Ben Gamari <ben at smart-cactus.org> wrote:
>> Alfredo Di Napoli <alfredo.dinapoli at gmail.com> writes:
>> > Hey folks,
>> >
>> Hi Alfredo!
>> First, thanks for writing. More eyes looking at GHC's compiler
>> performance is badly needed.
>> > maybe I’m setting up for something too ambitious for me, but I would
>> like
>> > to take an active stance to the overlasting “GHC compilation times are
>> > terrible” matter, instead of simply stare at the screen with despair
>> > whenever GHC compiles a sufficiently large Haskell program ;)
>> >
>> > To make this even more interesting, I have never contributed to GHC
>> either!
>> > The max I have pushed myself into was 2 years ago when I successfully
>> built
>> > GHC head from source and tried to fix an Haddock “easy” ticket I don’t
>> even
>> > recall (full disclosure, eventually I didn’t :D ).
>> >
>> > Specifically, I would love community recommendations & guidance about:
>> >
>> > 1. Is this simply too daunting for somebody like me? Maybe is better to
>> > first start contributing more regularly, take confidence with the code
>> base
>> > AND then move forward?
>> >
>> As with any software project, it is possible to treat the compiler as a
>> black box, throw a profiler at it and see what hotspots show up. This
>> gives you a place to focus your effort, allowing you to learn a small
>> area and broaden your knowledge as necessary.
>> However, I think it's fair to say that you will be significantly more
>> productive if you first develop a basic understanding of the compilation
>> pipeline. I'd recommend having a look at the GHC Commentary [1] for a
>> start.
>> I think it also helps to have a rough idea of what "slow" means to you.
>> I find it is quite helpful if you have a particular program which you
>> feel compiles more slowly than you would like (especially if it even
>> compiles slowly with -O0, since then much less of the compiler is
>> involved in compilation). Another approach is to look for programs whose
>> compilation time has regressed over the course of GHC releases. It is
>> not hard to find these examples and it is often possible to bisect your
>> way back to the regressing commit.
>> Also, note that I have collected some notes pertaining to compiler
>> performance on the Wiki [2]. Here you will find a number of tickets of
>> interest (as well a some rough themes which I've noticed), some nofib
>> results which might guide your efforts, as well as a list of some
>> fixes which have been committed in the past.
>> [1] https://ghc.haskell.org/trac/ghc/wiki/Commentary/Compiler
>> [2] https://ghc.haskell.org/trac/ghc/wiki/Performance/Compiler
>> > 2. Are compilation times largely dependant from the target platform
>> (I’m on
>> > Darwin) or there is something which can be done “globally” so that the
>> > benefits can be experienced by everybody?
>> >
>> There are some external considerations (e.g. the platform's compiler and
>> linking toolchain) which contribute to GHC's runtime. For instance, it
>> is known that the BFD ld linker implementation that many Linux
>> distributions use by default is a great deal slower than it could be.
>> This particular issue has come up recently and I'm currently working on
>> allowing us to use the more performant gold linker when available.
>> However, I think it's fair to say that for most programs GHC's runtime
>> is largely independent of platform. I would invite you to try compiling
>> a package which you consider GHC to compile "slowly" with GHC's -v flag
>> (and GHC 8.0.1 or newer). This will give you a rough breakdown of where
>> time is spent. For many packages you will find that the simplifier
>> and/or typechecker dominate, followed (often distantly) by native code
>> generation. Of these steps native code generation is the only one with a
>> strong platform dependence.
>> > 3. Is there any recommended workflow to profile GHC compilation times?
>> Is
>> > there any build flavour one should prefer when doing so? (Maybe the
>> full,
>> > slowest one?)
>> >
>> There are a few options here:
>>  * As of GHC 8.0 the compiler will output timing and allocation
>>    information for its various stages if run with -v. This can be
>>    extremely helpful to get a high-level picture of where the compiler
>>    is spending its time while compiling your program. This is almost
>>    always the right place to start.
>>  * As with any Haskell program, the cost centre profiler can be used to
>>    characterize the memory and CPU behavior of various parts of the
>>    compiler.
>>    GHC's source tree includes a "prof" build flavour which builds the
>>    compiler with profiling enabled. However it only includes a handful
>>    of cost-centres and is best used when you already have a rough idea
>>    where you are looking and can add further cost-centres to drill down
>>    to your hotspot.
>>    Simply enabling -fprof-exported across the entire tree just doesn't
>>    work in my experience: not only is the resulting compiler quite slow,
>>    but the profile you get is far too unwieldy to learn from.
>>  * Occassionally the ticky-ticky profiler can be helpful in identifying
>>    allocation hotspots without the full overhead of the cost-centre
>>    profiler.
>>  * In principle our newly-stable DWARF debug information can be used for
>>    profiling, although this is still a work in progress and requires a
>>    patched GHC for best results. It's probably best to stick to the more
>>    traditional profiling mechanisms for now.
>> Anyways, I hope this helps. Always feel free to get in touch with me
>> personally (IRC and email are both great) if you would like to discuss
>> particular issues. Thanks again for your interest!
>> Cheers,
>> - Ben
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.haskell.org/pipermail/ghc-devs/attachments/20170409/55c81643/attachment-0001.html>

More information about the ghc-devs mailing list