I've extracted the immediately useful changes into D73460, but it doesn't include the RDA initialisation changes. With those changes, I think being able to re-run the analysis is also likely to make this wrapper redundant.

Nice. LGTM.

• Rebased.
• Broke up some of the initialisation code in RDA so we can re-run after finalization.

Reduced test case.

Using isDebugInstr instead.

In D73117#1833090, @samparker wrote:

Why isn't this handled properly by the scheduling infrastructure instead? This sounds a bit like a hack...

Why isn't this handled properly by the scheduling infrastructure instead? This sounds a bit like a hack...

Changing the threshold within the pass.

I will proceed with getting some tests together and upload a patch for controlling the threshold. The numbers look even nicer for thumb2 targets:

Thanks, you both are indeed right that it can reduce code size... I've now tried reducing BBDuplicateThreshold and it looks like halving this to 3 is the best option (for aarch64 at least). There's a couple of outliers, but most changes are -0.1%:
test-suite...otout/Shootout-nestedloop.test 484 468 -3.3%
test-suite...+/Shootout-C++-nestedloop.test 532 516 -3.0%
test-suite...Olden/perimeter/perimeter.test 2108 2148 1.9%
test-suite...urce/Applications/aha/aha.test 2964 2932 -1.1%
test-suite...ce/Benchmarks/PAQ8p/paq8p.test 53748 53964 0.4%
test-suite...ch/g721/g721encode/encode.test 5468 5452 -0.3%
test-suite...marks/Ptrdist/yacr2/yacr2.test 15204 15236 0.2%
test-suite...ications/JM/ldecod/ldecod.test 161116 160788 -0.2%
test-suite...marks/7zip/7zip-benchmark.test 469784 470728 0.2%
test-suite...chmarks/MallocBench/gs/gs.test 104228 104428 0.2%
test-suite...oxyApps-C/miniAMR/miniAMR.test 34260 34196 -0.2%
test-suite...ications/JM/lencod/lencod.test 380324 380996 0.2%
test-suite...plications/d/make_dparser.test 62816 62912 0.2%
test-suite...s-C/Pathfinder/PathFinder.test 15996 15972 -0.2%
test-suite...ks/Prolangs-C++/city/city.test 5428 5436 0.1%
test-suite.../Benchmarks/Ptrdist/bc/bc.test 27300 27260 -0.1%
test-suite...pplications/oggenc/oggenc.test 105776 105632 -0.1%
test-suite...-typeset/consumer-typeset.test 327944 327520 -0.1%
test-suite...lications/sqlite3/sqlite3.test 244704 244400 -0.1%
test-suite...tions/lambda-0.1.3/lambda.test 20092 20068 -0.1%
test-suite.../Prolangs-C/bison/mybison.test 35828 35788 -0.1%
test-suite...peg2/mpeg2dec/mpeg2decode.test 29916 29884 -0.1%
test-suite.../Applications/SPASS/SPASS.test 287832 288136 0.1%
test-suite...arching-flt/Searching-flt.test 8292 8284 -0.1%
test-suite...arching-dbl/Searching-dbl.test 8500 8492 -0.1%
test-suite...C/Packing-flt/Packing-flt.test 8548 8540 -0.1%
test-suite...ences-flt/Recurrences-flt.test 8644 8636 -0.1%
test-suite...t/StatementReordering-flt.test 8660 8652 -0.1%
test-suite...encode/alacconvert-encode.test 26044 26020 -0.1%
test-suite...decode/alacconvert-decode.test 26044 26020 -0.1%
test-suite...C/Packing-dbl/Packing-dbl.test 8756 8748 -0.1%
test-suite...lications/SIBsim4/SIBsim4.test 26476 26500 0.1%
test-suite...ences-dbl/Recurrences-dbl.test 8852 8844 -0.1%
test-suite...l/StatementReordering-dbl.test 8868 8860 -0.1%
test-suite...ing-flt/LoopRerolling-flt.test 9116 9108 -0.1%
test-suite...ing-flt/Equivalencing-flt.test 9268 9260 -0.1%
test-suite...ing-flt/NodeSplitting-flt.test 9276 9268 -0.1%
test-suite...ing-dbl/LoopRerolling-dbl.test 9340 9332 -0.1%
test-suite...mbolics-flt/Symbolics-flt.test 9364 9356 -0.1%
test-suite...ing-dbl/Equivalencing-dbl.test 9420 9412 -0.1%
test-suite...ing-dbl/NodeSplitting-dbl.test 9476 9468 -0.1%
test-suite...mbolics-dbl/Symbolics-dbl.test 9588 9580 -0.1%
test-suite...lt/IndirectAddressing-flt.test 9708 9700 -0.1%
test-suite...bl/IndirectAddressing-dbl.test 9924 9916 -0.1%
test-suite...lt/CrossingThresholds-flt.test 9940 9932 -0.1%
test-suite...cations/hexxagon/hexxagon.test 10124 10132 0.1%
test-suite...bl/CrossingThresholds-dbl.test 10148 10140 -0.1%
test-suite...flt/InductionVariable-flt.test 10204 10196 -0.1%
test-suite...flt/LoopRestructuring-flt.test 10436 10428 -0.1%
test-suite...ow-flt/GlobalDataFlow-flt.test 10452 10444 -0.1%
test-suite...dbl/InductionVariable-dbl.test 10468 10460 -0.1%
test-suite...oops-flt/ControlLoops-flt.test 10588 10580 -0.1%
test-suite...dbl/LoopRestructuring-dbl.test 10644 10636 -0.1%
test-suite...ow-dbl/GlobalDataFlow-dbl.test 10660 10652 -0.1%
test-suite...oops-dbl/ControlLoops-dbl.test 10780 10772 -0.1%
test-suite...pansion-flt/Expansion-flt.test 10996 10988 -0.1%
test-suite...-flt/LinearDependence-flt.test 11148 11140 -0.1%
test-suite...pansion-dbl/Expansion-dbl.test 11196 11188 -0.1%
test-suite...ctions-flt/Reductions-flt.test 11316 11308 -0.1%
test-suite...-dbl/LinearDependence-dbl.test 11364 11356 -0.1%
test-suite...ctions-dbl/Reductions-dbl.test 11524 11516 -0.1%
test-suite...sc-C++/stepanov_container.test 13132 13140 0.1%
test-suite...lFlow-flt/ControlFlow-flt.test 13764 13756 -0.1%
test-suite...lFlow-dbl/ControlFlow-dbl.test 13972 13964 -0.1%
Geomean difference -0.0%

default-results    threshold-3        diff

count 310.000000 310.000000 310.000000
mean 26375.587097 26377.445161 -0.000273
std 65478.305060 65501.068235 0.002870
min 460.000000 460.000000 -0.033058
25% 1286.000000 1286.000000 0.000000
50% 3128.000000 3128.000000 0.000000
75% 11262.000000 11260.000000 0.000000
max 469784.000000 470728.000000 0.018975

LGTM, but please address nits before committing.

Removed some accidentally added files.

Rebased, cleaned up some tests and addressed comments.

Ah, thanks, then maybe I need to look elsewhere. I appear to be running into const issues when trying to sort in print and dump.

Ok, I'll get some test suite numbers across some cores. I would hope long latency instructions, such as div, would be predicated in the same way as armv7 so the performance would be comparable... I definitely share your concern there though, it will be the main source of the performance changes.

Good to see those looks finally loosing those subs and movs!

Are there any cores where we don't want FeatureDontRestrictIT?

How would you feel about unrestricting by default..? One of the GCC guys I was talking too also said it's hampering performance compared to armv7. I've got little bandwidth and I want to get some more numbers, and I'm not quite trusting my chromebook at the minute... but its' A72/A53 and I'm seeing good improvements on my data set of 1 (coremark) of 14% improvement.

I vaguely remember something that will go through and remove the one operand phis too, but I'm not sure what that is called (other than instcombine).

I'm hoping that this close to codegen, the phis won't be a problem..?

Please ignore the RDA and BaseInstrInfo changes, I've just copied them from the parent.

Cheers.

• Added explicit check for VPNOT.
• Added TODO comment.
• Removed duplicate test.
• Added comments to tests and had a small shuffle around of one of them.

Ping.

Does VPNOT not need to be explicitly excluded because they are def's anyway?

It's actually the multiple vpr uses which prevent them from being included. I think I'll explicitly disallow them for now, as well as adding the comment.

Even if you only decrease the size you can have a distance increase due to aligment constraints

Ah... thanks for raising this. I'll look into it.

I reckon adding something to that unroll test that would consistently drive the unroller, like a triple, should be added. Otherwise, LGTM.

I had to revert this due to buildbot failures. It seems the print order of liveins is non-deterministic... D72441 hopefully addresses this.

I can add liveness info into the outlined function, but we will need to do the same on AArch64 even if is not necessary.

Hmm, would it be possible to pass a bool to the pass which will control whether liveness is updated? The other, suboptimal, option would be to not do outlining when we have the LOB extension.

LGTM

It might be possible to rearrange Low Overhead Loops to run before ConstantIslands, but you'd probably need to do more to make it work properly. I don't think ConstantIslands knows how to handle the branches generated by LowOverheadLoop.

It's currently the backends responsibility to lower and handle any of the intrinsics that are inserted for hardware loops, so I don't think we should be concerned with other transforms triggering. I think the benefit of SCEV still working for these loops far outweighs the codegen effort - especially since this effort already has to be done because of isel and machine optimisations. That said, it probably wouldn't hurt to add some tests for common transforms especially ones where the loop body would be duplicated.

A couple of changes to RDA:

• It how has getLiveInUses which collects all the users of a live-in register.
• It now has getAllUses, which collects locals users and users in the successor blocks (using getLiveInUses).

LGTM

Regenerated the disjoint vcmp test.