[llvm-dev] llvm (the middle-end) is getting slower, December edition

Davide Italiano via llvm-dev llvm-dev at lists.llvm.org
Sat Dec 17 13:35:12 PST 2016


First of all, sorry for the long mail.
Inspired by the excellent analysis Rui did for lld, I decided to do
the same for llvm.
I'm personally very interested in build-time for LTO configuration,
with particular attention to the time spent in the optimizer.
Rafael did something similar back in March, so this can be considered
as an update. This tries to include a more accurate high-level
analysis of where llvm is spending CPU cycles.
Here I present 2 cases: clang building itself with `-flto` (Full), and
clang building an internal codebase which I'm going to refer as
`game7`.
It's a mid-sized program (it's actually a game), more or less of the
size of clang, which we use internally as benchmark to track
compile-time/runtime improvements/regression.
I picked two random revisions of llvm: trunk (December 16th 2016) and
trunk (June 2nd 2016), so, roughly, 6 months period.
My setup is a Mac Pro running Linux (NixOS).
These are the numbers I collected (including the output of -mllvm -time-passes).
For clang:

June 2nd:
real    22m9.278s
user    21m30.410s
sys     0m38.834s
  Total Execution Time: 1270.4795 seconds (1269.1330 wall clock)
  289.8102 ( 23.5%)  18.8891 ( 53.7%)  308.6993 ( 24.3%)  308.6906 (
24.3%)  X86 DAG->DAG Instruction Selection
  97.2730 (  7.9%)   0.7656 (  2.2%)  98.0386 (  7.7%)  98.0010 (
7.7%)  Global Value Numbering
  62.4091 (  5.1%)   0.4779 (  1.4%)  62.8870 (  4.9%)  62.8665 (
5.0%)  Function Integration/Inlining
  58.6923 (  4.8%)   0.4767 (  1.4%)  59.1690 (  4.7%)  59.1323 (
4.7%)  Combine redundant instructions
  53.9602 (  4.4%)   0.6163 (  1.8%)  54.5765 (  4.3%)  54.5409 (
4.3%)  Combine redundant instructions
  51.0470 (  4.1%)   0.5703 (  1.6%)  51.6173 (  4.1%)  51.5425 (
4.1%)  Loop Strength Reduction
  47.4067 (  3.8%)   1.3040 (  3.7%)  48.7106 (  3.8%)  48.7034 (
3.8%)  Greedy Register Allocator
  36.7463 (  3.0%)   0.8133 (  2.3%)  37.5597 (  3.0%)  37.4612 (
3.0%)  Induction Variable Simplification
  37.0125 (  3.0%)   0.2699 (  0.8%)  37.2824 (  2.9%)  37.2478 (
2.9%)  Combine redundant instructions
  34.2071 (  2.8%)   0.2737 (  0.8%)  34.4808 (  2.7%)  34.4487 (
2.7%)  Combine redundant instructions
  25.6627 (  2.1%)   0.3215 (  0.9%)  25.9842 (  2.0%)  25.9509 (
2.0%)  Combine redundant instructions

Dec 16th:
real    27m34.922s
user    26m53.489s
sys     0m41.533s

  287.5683 ( 18.5%)  19.7048 ( 52.3%)  307.2731 ( 19.3%)  307.2648 (
19.3%)  X86 DAG->DAG Instruction Selection
  197.9211 ( 12.7%)   0.5104 (  1.4%)  198.4314 ( 12.5%)  198.4091 (
12.5%)  Function Integration/Inlining
  106.9669 (  6.9%)   0.8316 (  2.2%)  107.7984 (  6.8%)  107.7633 (
6.8%)  Global Value Numbering
  89.7571 (  5.8%)   0.4840 (  1.3%)  90.2411 (  5.7%)  90.2067 (
5.7%)  Combine redundant instructions
  79.0456 (  5.1%)   0.7534 (  2.0%)  79.7990 (  5.0%)  79.7630 (
5.0%)  Combine redundant instructions
  55.6393 (  3.6%)   0.3116 (  0.8%)  55.9509 (  3.5%)  55.9187 (
3.5%)  Combine redundant instructions
  51.8663 (  3.3%)   1.4090 (  3.7%)  53.2754 (  3.3%)  53.2684 (
3.3%)  Greedy Register Allocator
  52.5721 (  3.4%)   0.3021 (  0.8%)  52.8743 (  3.3%)  52.8416 (
3.3%)  Combine redundant instructions
  49.0593 (  3.2%)   0.6101 (  1.6%)  49.6694 (  3.1%)  49.5904 (
3.1%)  Loop Strength Reduction
  41.2602 (  2.7%)   0.9608 (  2.5%)  42.2209 (  2.7%)  42.1122 (
2.6%)  Induction Variable Simplification
  38.1438 (  2.5%)   0.3486 (  0.9%)  38.4923 (  2.4%)  38.4603 (
2.4%)  Combine redundant instructions

so, llvm is around 20% slower than it used to be.

For our internal codebase the situation seems slightly worse:

`game7`

June 2nd:

Total Execution Time: 464.3920 seconds (463.8986 wall clock)

  88.0204 ( 20.3%)   6.0310 ( 20.0%)  94.0514 ( 20.3%)  94.0473 (
20.3%)  X86 DAG->DAG Instruction Selection
  27.4382 (  6.3%)  16.2437 ( 53.9%)  43.6819 (  9.4%)  43.6823 (
9.4%)  X86 Assembly / Object Emitter
  34.9581 (  8.1%)   0.5274 (  1.8%)  35.4855 (  7.6%)  35.4679 (
7.6%)  Function Integration/Inlining
  27.8556 (  6.4%)   0.3419 (  1.1%)  28.1975 (  6.1%)  28.1824 (
6.1%)  Global Value Numbering
  22.1479 (  5.1%)   0.2258 (  0.7%)  22.3737 (  4.8%)  22.3593 (
4.8%)  Combine redundant instructions
  19.2346 (  4.4%)   0.3639 (  1.2%)  19.5985 (  4.2%)  19.5870 (
4.2%)  Post RA top-down list latency scheduler
  15.8085 (  3.6%)   0.2675 (  0.9%)  16.0760 (  3.5%)  16.0614 (
3.5%)  Combine redundant instructions

Dec 16th:

Total Execution Time: 861.0898 seconds (860.5808 wall clock)

  135.7207 ( 15.7%)   0.2484 (  0.8%)  135.9692 ( 15.2%)  135.9531 (
15.2%)  Combine redundant instructions
  103.6609 ( 12.0%)   0.4566 (  1.4%)  104.1175 ( 11.7%)  104.1014 (
11.7%)  Combine redundant instructions
  97.1083 ( 11.3%)   6.9183 ( 21.8%)  104.0266 ( 11.6%)  104.0181 (
11.6%)  X86 DAG->DAG Instruction Selection
  72.6125 (  8.4%)   0.1701 (  0.5%)  72.7826 (  8.1%)  72.7678 (
8.1%)  Combine redundant instructions
  69.2144 (  8.0%)   0.6060 (  1.9%)  69.8204 (  7.8%)  69.8007 (
7.8%)  Function Integration/Inlining
  60.7837 (  7.1%)   0.3783 (  1.2%)  61.1620 (  6.8%)  61.1455 (
6.8%)  Global Value Numbering
  56.5650 (  6.6%)   0.1980 (  0.6%)  56.7630 (  6.4%)  56.7476 (
6.4%)  Combine redundant instructions

so, using LTO, lld takes 2x to build what it used to take (and all the
extra time seems spent in the optimizer).

As an (extra) experiment, I decided to take the unoptimized output of
game7 (via lld -save-temps) and pass to -opt -O2. That shows another
significant regression (with different characteristics).

June 2nd:
time opt -O2
real    6m23.016s
user   6m20.900s
sys     0m2.113s

35.9071 ( 10.0%)   0.7996 ( 10.9%)  36.7066 ( 10.0%)  36.6900 ( 10.1%)
 Function Integration/Inlining
33.4045 (  9.3%)   0.4053 (  5.5%)  33.8098 (  9.3%)  33.7919 (  9.3%)
 Global Value Numbering
27.1053 (  7.6%)   0.5940 (  8.1%)  27.6993 (  7.6%)  27.6995 (  7.6%)
 Bitcode Writer
25.6492 (  7.2%)   0.2491 (  3.4%)  25.8984 (  7.1%)  25.8805 (  7.1%)
 Combine redundant instructions
19.2686 (  5.4%)   0.2956 (  4.0%)  19.5642 (  5.4%)  19.5471 (  5.4%)
 Combine redundant instructions
18.6697 (  5.2%)   0.2625 (  3.6%)  18.9323 (  5.2%)  18.9148 (  5.2%)
 Combine redundant instructions
16.1294 (  4.5%)   0.2320 (  3.2%)  16.3614 (  4.5%)  16.3434 (  4.5%)
 Combine redundant instructions
13.5476 (  3.8%)   0.3945 (  5.4%)  13.9421 (  3.8%)  13.9295 (  3.8%)
 Combine redundant instructions
13.1746 (  3.7%)   0.1767 (  2.4%)  13.3512 (  3.7%)  13.3405 (  3.7%)
 Combine redundant instructions

Dec 16th:

real    20m10.734s
user    20m8.523s
sys     0m2.197s

  208.8113 ( 17.6%)   0.1703 (  1.9%)  208.9815 ( 17.5%)  208.9698 (
17.5%)  Value Propagation
  179.6863 ( 15.2%)   0.1215 (  1.3%)  179.8077 ( 15.1%)  179.7974 (
15.1%)  Value Propagation
  92.0158 (  7.8%)   0.2674 (  3.0%)  92.2832 (  7.7%)  92.2613 (
7.7%)  Combine redundant instructions
  72.3330 (  6.1%)   0.6026 (  6.7%)  72.9356 (  6.1%)  72.9210 (
6.1%)  Combine redundant instructions
  72.2505 (  6.1%)   0.2167 (  2.4%)  72.4672 (  6.1%)  72.4539 (
6.1%)  Combine redundant instructions
  66.6765 (  5.6%)   0.3482 (  3.9%)  67.0247 (  5.6%)  67.0040 (
5.6%)  Combine redundant instructions
  65.5029 (  5.5%)   0.4092 (  4.5%)  65.9121 (  5.5%)  65.8913 (
5.5%)  Combine redundant instructions
  61.8355 (  5.2%)   0.8150 (  9.0%)  62.6505 (  5.2%)  62.6315 (
5.2%)  Function Integration/Inlining
  54.9184 (  4.6%)   0.3359 (  3.7%)  55.2543 (  4.6%)  55.2332 (
4.6%)  Combine redundant instructions
  50.2597 (  4.2%)   0.2187 (  2.4%)  50.4784 (  4.2%)  50.4654 (
4.2%)  Combine redundant instructions
  47.2597 (  4.0%)   0.3719 (  4.1%)  47.6316 (  4.0%)  47.6105 (
4.0%)  Global Value Numbering

I don't have an infrastructure to measure the runtime performance
benefits/regression of clang, but I have for `game7`.
I wasn't able to notice any fundamental speedup (at least, not
something that justifies a 2x build-time).

tl;dr:
There are quite a few things to notice:
1) GVN used to be the top pass in the middle-end, in some cases, and
pretty much always in the top-3. This is not the case anymore, but
it's still a pass where we spend a lot of time. This is being worked
on by Daniel Berlin and me) https://reviews.llvm.org/D26224 so there's
some hope that will be sorted out (or at least there's a plan for it).
2) For clang, we spend 35% more time inside instcombine, and for game7
instcombine seems to largely dominate the amount of time we spend
optimizing IR. I tried to bisect (which is not easy considering the
test takes a long time to run), but I wasn't able to identify a single
point in time responsible for the regression. It seems to be an
additive effect. My wild (or not so wild) guess is that every day
somebody adds a matcher of two because that improves their testcase,
and at some point all this things add up. I'll try to do some
additional profiling but I guess large part of our time is spent
solving bitwise-domain dataflow problems (ComputeKnownBits et
similia). Once GVN will be in a more stable state, I plan to
experiment with caching results.
3) Something peculiar is that we spend 2x time in the inliner. I'm not
familiar with the inliner, IIRC there were some changes to threshold
recently, so any help here will be appreciated (both in reproducing
the results and with analysis).
4) For the last testcase (opt -O2 on large unoptimized chunk of
bitcode) llvm spends 33% of its time in CVP, and very likely in LVI. I
think it's not as lazy as it claims to be (or at least, the way we use
it). This doesn't show up in a full LTO run because we don't run CVP
as part of the default LTO pipeline, but the case shows how LVI can be
a bottleneck for large TUs (or at least how -O2 compile time degraded
on large TUs). I haven't thought about the problem very carefully, but
there seems to be some progress on this front (
https://llvm.org/bugs/show_bug.cgi?id=10584). I can't share the
original bitcode file but I can probably do some profiling on it as
well.

As next steps I'll try to get a more detailed analysis of the
problems. In particular, try to do what Rui did for lld but with more
coarse granularity (every week) to have a chart of the compile time
trend for these cases over the last 6 months, and post here.

I think (I know) some people are aware of the problems I outline in
this e-mail. But apparently not everybody. We're in a situation where
compile time is increasing without real control. I'm happy that Apple
is doing a serious effort to track build-time, so hopefully things
will improve. There are, although, some cases (like adding matchers in
instcombine or knobs) where the compile time regression is hard to
track until it's too late. LLVM as a project tries not to stop people
trying to get things done and that's great, but from time to time it's
good to take a step back and re-evaluate approaches.
The purpose of this e-mail was to outline where we regressed, for
those interested.

Thanks for your time, and of course, feedback welcome!

-- 
Davide

"There are no solved problems; there are only problems that are more
or less solved" -- Henri Poincare


More information about the llvm-dev mailing list