[GSoC 2021] Utilize LoopNest Pass

Project Infrastructure LLVM Dev List Archives
#1

Hello,

I’m thinking of participating in GSoC at LLVM.
I’ve already talked about ideas with some project mentors but a newly added project idea “Utilize LoopNest Pass” attracted me.

I want to get in touch with its mentors (Whitney Tsang, Ettore Tiotto), whose email I couldn’t find out.

Sincerely,
Toshiki Maekawa

#2
  • Whitney, Ettore
#3

I’m glad to get in touch with you two.
I’m planning to utilize LoopNest pass for some passes like loop unrolling.
Could you tell me more about this project if any?

2021年3月9日(火) 20:47 Whitney Tsang <whitney.uwaterloo@gmail.com>:

#4

Sure, in this project, we plan to utilize the LoopNest pass in some existing passes.
As an example, here is one of the patches to transform the loop interchange pass: https://reviews.llvm.org/D97847.
On top of doing the code changes, we want to understand what are the pros and cons of the change.
We can do that by collecting some data, e.g. compile time.

There are two loop unroller, LoopFullUnrollPass and LoopUnrollPass.
The LoopFullUnrollPass is not a good candidate for the change, as it is a loop pass and it is being used as a loop pass.
It is written to be traverse from inner to outer loop, with other loop transformations in between in the loop pipeline.
The LoopUnrollPass could be a candidate, but the LoopUnrollAndJamPass can be an easier to understand candidate,
as unroll and jam transformation naturally operate on more than one loop (LoopNest).

Regards,
Whitney Tsang

graycol.gifuint256_t —2021/03/10 07:52:17 AM—I’m glad to get in touch with you two. I’m planning to utilize LoopNest pass for some passes like lo

#5

Thank you for your advice. As you mentioned, LoopUnrollAndJam seems to be a good candidate.

I’m going to find another analyse/transform pass suitable for loop nest pass.

Best,
Toshiki Maekawa

2021年3月10日(水) 23:32 Whitney T Tsang <whitneyt@ca.ibm.com>:

graycol.gif

#6

Hello,

I have a question about finding good passes for utilizing LoopNest pass.

The LoopFullUnrollPass is not a good candidate for the change, as it is a loop pass and it is being used as a loop pass.
Then you mean that passes which are implemented as not a loop pass (= function pass) and not already used as a loop pass are good for this change?
I didn’t get that. Could you tell me why it is?

Thanks in advance,Toshiki Maekawa

graycol.gif

#7

Passes implemented as a loop pass can still be good for this change.
I am referring to passes implemented as a loop pass and used as a loop pass in a loop pipeline may not be a good candidate.
Take the LoopFullUnrollPass as an example, LoopFullUnrollPass is added in LPM2 with other loop passes, e.g. LoopIdiomRecognizePass, IndVarSimplifyPass, …
Loops are processed from inner to outer loop, where each loop is processed by all the passes in LPM2, before process its outer loops.
If we change LoopFullUnrollPass to a LoopNest pass, then there is no way we can preserve the same order of processing loops.
It may still be good to change such passes, but there can be huge performance impact, which we need to analyze and handle, and probably not the best choice to look at first.

[LoopPassManager](https://llvm.org/doxygen/classllvm_1_1PassManager_3_01Loop_00_01LoopAnalysisManager_00_01LoopStandardAnalysisResults_01_6_00_01LPMUpdater_01_6_01_4.html) LPM2(DebugLogging);

LPM2.addPass([LoopIdiomRecognizePass](https://llvm.org/doxygen/classllvm_1_1LoopIdiomRecognizePass.html)());
LPM2.addPass([IndVarSimplifyPass](https://llvm.org/doxygen/classllvm_1_1IndVarSimplifyPass.html)());

for (auto &[C](https://llvm.org/doxygen/README__ALTIVEC_8txt.html#a9aacd9146afe44bf656cd664e2a88c8c) : LateLoopOptimizationsEPCallbacks)
[C](https://llvm.org/doxygen/README__ALTIVEC_8txt.html#a9aacd9146afe44bf656cd664e2a88c8c)(LPM2, [Level](https://llvm.org/doxygen/namespacellvm_1_1PICLevel.html#a66ddbf1bb21f90ddc44260d1ca677b6b));

LPM2.addPass([LoopDeletionPass](https://llvm.org/doxygen/classllvm_1_1LoopDeletionPass.html)());

if ([Phase](https://llvm.org/doxygen/AArch64FalkorHWPFFix_8cpp.html#a780618ccf661aebc12f8d991d294c950) != [ThinOrFullLTOPhase::ThinLTOPreLink](https://llvm.org/doxygen/namespacellvm.html#a5d737fb4258bb27586a1bffd557fbb49ad94cc56b0a9155d607f2609b0f5c39d3) || !PGOOpt ||
PGOOpt->Action != [PGOOptions::SampleUse](https://llvm.org/doxygen/structllvm_1_1PGOOptions.html#a13bd589bcabdfc073bac5711f76dd2b6a9cbf1cb1b0731fc75926a6930592968a))
LPM2.addPass([LoopFullUnrollPass](https://llvm.org/doxygen/classllvm_1_1LoopFullUnrollPass.html)([Level](https://llvm.org/doxygen/namespacellvm_1_1PICLevel.html#a66ddbf1bb21f90ddc44260d1ca677b6b).getSpeedupLevel(),
/* OnlyWhenForced= */ !PTO.[LoopUnrolling](https://llvm.org/doxygen/classllvm_1_1PipelineTuningOptions.html#af5b2b612c2769e79040e34bc4be7a77f),
PTO.[ForgetAllSCEVInLoopUnroll](https://llvm.org/doxygen/classllvm_1_1PipelineTuningOptions.html#ac4261b710708d3ffd351693d268dbc2b)));

for (auto &[C](https://llvm.org/doxygen/README__ALTIVEC_8txt.html#a9aacd9146afe44bf656cd664e2a88c8c) : LoopOptimizerEndEPCallbacks)
(LPM2, [Level](https://llvm.org/doxygen/namespacellvm_1_1PICLevel.html#a66ddbf1bb21f90ddc44260d1ca677b6b));

FPM.[addPass](https://llvm.org/doxygen/classllvm_1_1PassManager.html#a314ff184ce4ace8801e1158ef909e22e)([createFunctionToLoopPassAdaptor](https://llvm.org/doxygen/namespacellvm.html#a5118c8cfac85abdec7366d89d023a1ef)(
[std::move](https://llvm.org/doxygen/lib_2Target_2ARM_2README_8txt.html#ad3a99906764c35b2694ae90fa57744a5)(LPM2), /*UseMemorySSA=*/false, /*UseBlockFrequencyInfo=*/false,
DebugLogging));

Regards,
Whitney Tsang

graycol.gifuint256_t —2021/03/14 08:48:39 PM—Hello, I have a question about finding good passes for utilizing LoopNest pass.

#8

Thank you for your advice. I understand which passes are good for this change.
I’ll avoid transformation passes already used in a loop pipeline as much as possible.

Best,
Toshiki Maekawa

2021年3月15日(月) 22:17 Whitney T Tsang <whitneyt@ca.ibm.com>:

graycol.gif

#9

Hi,

I tried to utilize LoopNest pass for LoopUnrollAndJam pass, but found it difficult.
LoopUnrollAndJam is currently implemented as FunctionPass, and using some functions that may add new loops like simplifyLoop().

We need to track such mutations by LPMUpdater, so need to make the functions that may add loops use the updater.

This could lead to a big code change. What do you think about this change?

Thanks,
Toshiki Maekawa

2021年3月17日(水) 0:47 uint256_t <konndennsa@gmail.com>:

graycol.gif

#10

LoopUnrollAndJam was changed from LoopPass to FunctionPass, which was a good choice as LoopNestPass didn’t exists at that time.
https://reviews.llvm.org/D72230

You can see that simplifyLoop and formLCSSARecursively was adding in that patch.
Those changes may not be needed anymore if we are using LoopNestPass, and if the PassManager itself can make sure the loops are simplified and in LCSSA form.

Regards,
Whitney Tsang

graycol.gif

#11

I didn’t realize LoopUnrollAndJam used to be implemented as LoopPass. I should have checked the commit log.

Thanks,
Toshiki Maekawa

2021年3月23日(火) 1:01 Whitney T Tsang <whitneyt@ca.ibm.com>:

graycol.gif

#12

Hi, it’s been a while.

I submitted a draft of project proposal to the website of GSoC.
Could you give me some advice about it?

(or should I share a draft on this ML?)

Best,
Toshiki Maekawa

graycol.gif

#13

Here are some comments:

  1. Not sure LoopFusionPass can be changed to be a LoopNest pass, as it operates on more than 1 LoopNest.
  2. We can also consider changing analysis on top of transformation. For example, you mentioned LoopCachePrinterPass, which is a transformation that print the result of the analysis/utility LoopCacheAnalysis.
  3. Some loop passes used in a loop pipeline can be good candidate in a later phase of the project, when you are more familiar with the idea, e.g. LoopIdiomRecognizePass.
  4. I prefer to study the impact and collect data needed for each pass, instead of at the end of the project.
    Best,
    Whitney Tsang

graycol.gif

#14

Thank you for your comments.
I’ve totally forgot that utilities for analysis that aren’t implemented as analysis passes can be candidate. It’ll be nice for them to utilize LoopNest as a primary object.

Best,
Toshiki Maekawa

graycol.gif

graycol.gif