[llvm] [LoopInterchange] Enable it by default (WIP) (PR #124911)

[Sjoerd Meijer via llvm-commits]

sjoerdmeijer wrote:

> Could you provide numbers for performance changes (improvments and regressions), e.g. highlights from llvm-test-suite? Would help illustrating that we also gain something from slighlty increased compile-time.

Hi @Meinersbur, thanks for the comments, that's a very fair question!

I would like to use two metrics for success here: the number of times loop-interchange triggers in the test-suite, and performance changes, with a higher weight for the former, initially. Let me explain this. We see this as a first enablement step: i.e. we have prioritised correctness and made interchange more conservative, in order to get a first version running by default. Enablement of interchange, loopcache analysis, and data dependence analysis would a big step forward. Even before we made interchange more conservative, interchange was not triggering on our motivating examples. So after this first step, we will follow up with a second optimisation step and lift restrictions to let it trigger on our examples. 

The number of times interchange triggers is a good metric for the potential of the pass and how general it is. I have counted 37 occurrences of interchange if external benchmark ffmpeg is included and 22 times without. So that is really good, I think. I will paste the build log with optimisation remarks at the end of this message for more details. 

About perf changes: for me a good result is if this shows no performance regressions, and maybe a few minor uplifts here and there. And the reason is what I wrote above, we very much see this as a first enablement step, and know we have to work on making it more optimal. And when I measured the llvm test-suite, this is exactly what I saw: no regression, minor uplifts (in Polybench and its solvers). But that was before the latest interchange patches went in, so I will verify and confirm this with the top of trunk. I would like to add that I have issues with the llvm test-suite for evaluating this kind of codegen changes as they may not be very visible, and I briefly talked about that in my llvm dev conference talk [here](https://youtu.be/u7W3_xL_3NQ?t=1032) (sorry for the plug, but it's a link to the exact timestamp).  So llvm test-suite performance numbers need to be taken with a pinch of salt, but again, will do confirm. 


    test-suite-externals/ffmpeg/libavcodec/aacps.c:121:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/aacps_tablegen.h:77:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/dcaenc.c:971:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/dpx.c:700:13: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/dpx.c:665:13: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/ituh263enc.c:786:13: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/mpeg12enc.c:113:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/mpeg4videoenc.c:1176:13: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/mpegaudiodec_template.c:268:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/proresenc_anatoliy.c:525:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavcodec/xface.c:291:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavfilter/af_arnndn.c:1512:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavfilter/vf_datascope.c:1077:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavfilter/vf_datascope.c:678:9: remark: Loop interchanged
    test-suite-externals/ffmpeg/libavfilter/vf_photosensitivity.c:182:9: remark: Loop interchanged
    MultiSource/Applications/JM/ldecod/block.c:935:5: remark: Loop interchanged
    MultiSource/Applications/JM/ldecod/block.c:1265:5: remark: Loop interchanged
    MultiSource/Applications/JM/ldecod/macroblock.c:2594:5: remark: Loop interchanged
    MultiSource/Applications/JM/lencod/block.c:2139:5: remark: Loop interchanged
    MultiSource/Applications/JM/lencod/block.c:2860:5: remark: Loop interchanged
    MultiSource/Applications/JM/lencod/block.c:3016:5: remark: Loop interchanged
    MultiSource/Applications/obsequi/tables.c:224:7: remark: Loop interchanged
    MultiSource/Benchmarks/mediabench/mpeg2/mpeg2dec/spatscal.c:298:5: remark: Loop interchanged
    MultiSource/Benchmarks/nbench/nbench1.c:283:2: remark: Loop interchanged
    MultiSource/Benchmarks/nbench/nbench1.c:1906:1: remark: Loop interchanged
    MultiSource/Benchmarks/nbench/nbench1.c:1907:2: remark: Loop interchanged
    SingleSource/Benchmarks/Linpack/linpack-pc.c:595:17: remark: Loop interchanged
    SingleSource/Benchmarks/Linpack/linpack-pc.c:1204:17: remark: Loop interchanged
    SingleSource/Benchmarks/Misc/dt.c:15:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/datamining/correlation/correlation.c:103:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/datamining/covariance/covariance.c:81:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/cholesky/cholesky.c:63:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/cholesky/cholesky.c:64:7: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/lu/lu.c:51:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/lu/lu.c:52:7: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/ludcmp/ludcmp.c:62:5: remark: Loop interchanged
    SingleSource/Benchmarks/Polybench/linear-algebra/solvers/ludcmp/ludcmp.c:63:7: remark: Loop interchanged

https://github.com/llvm/llvm-project/pull/124911