This is a follow-up from a question I asked last month. I'm evaluating the performance of two pass sequences that resemble (but are not) -O3.
With -O3, -debug-pass=Structure prints several independent blocks that seem to represent several calls to opt. I focused on two of these blocks, say S1 and S2, and compared the following optimization methods:
1. Executing them separately, ie. opt -S1 | opt -S2
2. Executing them in a single call, ie. opt -S1 -S2
I built the test suite with each of these configurations, then measured the performance of the compiled programs with perf, over 10 runs.
I'm attaching a plot of the speedup of method 1 over method 2. The intervals represent the standard deviation of the performance measures.
As you can see, programs compiled with method 1 are significantly slower than their counterparts compiled with method 2. However, if passes were applied in order using function composition, their performance should be the same.
I'd like to know if there is a way to recover this property in the pass manager, or at least explain the difference. If needed, I can provide scripts to reproduce the measurements.
Do you have more information? What were the exact command lines you
used? Do you have an example program that demonstrates the difference
than you can share?
Sébastien Michelland via llvm-dev <email@example.com> writes:
I reproduced the test on many individual files and got very variable results... it seems the computer's workload when running the test suite influenced the execution speed a lot more than standard deviation shows. I'll withdraw the performance claim until I can get consistent results (changed subject line), apologies for the confusion.
What I can still show easily is that the code generated by these two methods is different (which is already weird). For a simple example, grab a copy of bilateral_grid.bc:
Then you can generate my sequences with [opt -O3 -debug-pass=Arguments] and diff the outputs. Please see the attached script.
The differences seem to be mainly on variable indices (are they randomized?); on some test (namely jacobi-2d-imper) I have seen calling convention differences.
I'd like to optimize programs by greedily selecting optimizations, making a call to opt at each step. If I don't have equality between the two methods, I can't be sure that the sequence I'm building will make much sense.
build-bileteral-grid.sh (409 Bytes)
Passing -print-after-all to opt should print the IR after each pass. That may help figure out what’s going on.
This answer is a bit slow; I tried to look into the sequence details but 250 passes plus the complex bitcode of test suite examples makes this pretty hard.
In the meantime I stumbled upon llvm-diff which abstracts away the most significant difference, namely instruction renaming. It also ignores function attributes so calling conventions are silently unified; but at least it gives empty diffs when comparing the two methods. This means that my performance differences are mostly measurement errors...
Some of the differences might be "normal", eg. caused by randomized data structures. I don't have that much experience with LLVM code so I'm not sure how probable this is.
I'll stick to llvm-diff for now and maybe come back to this when I have a clearer understanding of the pass management process. ^^
Thanks for your time and help!
We want LLVM to be deterministic and there have been efforts to fix
problems related to data structures causing different generated code
sequences. It's certainly possible something like that is going on, but
it shouldn't just be dismissed. It would be best if we could get to the
bottom of it and see what needs fixing.
Sébastien Michelland <firstname.lastname@example.org> writes:
Alright, since it's deemed important I'll do my best to help. I've built an up-to-date LLVM from Git, I'll make more tests once I setup the test suite instrumentation I need to replace the optimization sequences.
So far I can't say it looks like a determinism issue to me. The two methods are deterministic on their own as far as I can test; they just don't have the same output (which might have an explanation in terms of how the pass manager works?).
Are you able to reproduce the test case from earlier (the one with the attached shell script)?