This is an archive of the discontinued LLVM Phabricator instance.

Differential D137600

[mlir] Add hoisting of transfer ops in affine loops
ClosedPublic

Authored by jsetoain on Nov 7 2022, 4:29 PM.

Details

Reviewers
bondhugula
nicolasvasilache
dcaballe
hanchung
mravishankar
Commits
rGda291bab8120: [mlir] Add hoisting of transfer ops in affine loops
rG825da072a8ed: [mlir] Add hoisting of transfer ops in affine loops
Summary

The only way to do this with the current hoisting strategy is by
lowering Affine to Scf first, but that prevents further passes on
Affine.

Diff Detail

Event Timeline

jsetoain created this revision.Nov 7 2022, 4:29 PM
Herald added a reviewer: bondhugula. · View Herald TranscriptNov 7 2022, 4:29 PM
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: Moerafaat, zero9178, bzcheeseman and 22 others. · View Herald Transcript
Harbormaster completed remote builds in B196598: Diff 473827.Nov 7 2022, 4:59 PM
jsetoain updated this revision to Diff 476189.Nov 17 2022, 11:18 AM

Make utility function names more specific

Herald added a subscriber: hanchung. · View Herald TranscriptNov 17 2022, 11:18 AM
jsetoain updated this revision to Diff 476190.Nov 17 2022, 11:21 AM

Fix typo

Harbormaster completed remote builds in B198257: Diff 476190.Nov 17 2022, 11:44 AM
jsetoain updated this revision to Diff 476213.Nov 17 2022, 12:54 PM

clang-format

Harbormaster completed remote builds in B198277: Diff 476213.Nov 17 2022, 1:13 PM
jsetoain published this revision for review.Nov 17 2022, 1:14 PM
Herald added a reviewer: nicolasvasilache. · View Herald TranscriptNov 17 2022, 1:14 PM
Herald added a reviewer: dcaballe. · View Herald Transcript
Herald added a project: Restricted Project. · View Herald Transcript
Herald added subscribers: stephenneuendorffer, nicolasvasilache. · View Herald Transcript
bondhugula accepted this revision.Nov 23 2022, 5:25 AM

This looks great to me - thanks! Mostly minor comments.

mlir/include/mlir/Dialect/Affine/Utils.h
336

Nit: doesn't
Similarly, apostrophes missing below.

mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

dyn_cast here to avoid an extra check below.

526

dyn_cast here again.

532

You should actually check earlier and advance the walk if it's something other than affine.for or scf.for at line 428; either that or assert here with an extra else or use cast above for AffineForOp in the conditional.

mlir/lib/Dialect/SCF/Utils/Utils.cpp
41 ↗(On Diff #476213)

This should've been in the mlir::scf namespace. You can update it in this PR itself if you wish.

151 ↗(On Diff #476213)

SCFForLoop -> SCFFor or SCFLoop for conciseness.

This revision is now accepted and ready to land.Nov 23 2022, 5:25 AM
jsetoain updated this revision to Diff 477519.Nov 23 2022, 8:52 AM
jsetoain marked 6 inline comments as done.

Address comments

jsetoain marked an inline comment as not done.Nov 23 2022, 8:59 AM

Thanks for the review!

mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

I haven't found a way to dyn_cast the result of the dyn_cast. I'm not very familiar with llvm's static polymorphism constructs, so I'm not sure what the problem is, but I get errors like:

error: invalid cast of an rvalue expression of type ‘std::nullptr_t’ to type ‘llvm::CastInfo<mlir::scf::ForOp, mlir::LoopLikeOpInterface, void>::CastReturnType’ {aka ‘mlir::scf::ForOp&’}

Alternatively, I can write: dyn_cast<scf::ForOp>(transferRead->getParentOp()) but I'm not sure it's worth the loss of clarity.

mlir/lib/Dialect/SCF/Utils/Utils.cpp
41 ↗(On Diff #476213)

Indeed, this has been bugging me since I submitted the patch. I agree, it would be better if instead of using namespaces in the function names, we placed these utility functions in their respective namespaces. It's a straightforward NFC, but it would splash everywhere; I will push a separated patch as soon as I'm done with this review.

bondhugula added inline comments.Nov 23 2022, 9:07 AM
mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

Please see https://llvm.org/docs/ProgrammersManual.html#the-isa-cast-and-dyn-cast-templates.

Use dyn_cast instead of isa + cast, i.e.,

if (auto scfFor = dyn_cast<scf::ForOp>(loop)) {
  ...
mlir/lib/Dialect/SCF/Utils/Utils.cpp
148 ↗(On Diff #477519)

innermost

bondhugula requested changes to this revision.Nov 23 2022, 9:12 AM
bondhugula added inline comments.
mlir/lib/Dialect/Affine/Utils/Utils.cpp
1877–1941

Can we templatize this function and make it work for both AffineForOp and scf::ForOp? You'll need YieldOpTy (can be scf::YieldOp and AffineYieldOp) and ForOpTy (can be scf::ForOp or AffineForOp). Other than the op creation calls, isn't everything else the same?

This revision now requires changes to proceed.Nov 23 2022, 9:12 AM
Harbormaster completed remote builds in B199221: Diff 477519.Nov 23 2022, 9:26 AM
jsetoain updated this revision to Diff 477603.Nov 23 2022, 2:13 PM

Try dyn_cast

mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

I believe I didn't explain myself very well there. I understand how LLVM's RTTI is supposed to work, I don't understand why it's not working. I'm getting that error in my previous comment, but I don't understand enough about the implementation to figure out why.

I'm going to push the change anyway. Best case scenario, it's something wrong with my compiler. Worst case scenario, it will leave a nice log of the error I'm seeing.

Harbormaster completed remote builds in B199286: Diff 477603.Nov 23 2022, 2:32 PM
mravishankar requested changes to this revision.Nov 23 2022, 2:56 PM
mravishankar added inline comments.
mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

I am missing overall context, but you cant dyn_cast result of the dyn_cast....

IIUC what is being suggested here is

if (auto scfForOp = dyn_cast<scf::ForOp>(loop)) {
  ....
} else if (auto affineForOp = dyn_cast<AffineForOp>(loop)) {

}

or better yet, use TypeSwitch

return TypeSwitch<Operation *, WalkResult>(loop)
  .Case<scf::ForOp>([](scf::ForOp scfForOp) {...})
  .Case<AffineForOp>([](AffineForOp affineForOp) {...})
  .Default([](Operation *) { return WalkResult::interrupt(); }

(and move the loop.erase into the if in first case or callback for the Case in the second example).

This revision now requires changes to proceed.Nov 23 2022, 2:56 PM
jsetoain updated this revision to Diff 477649.Nov 23 2022, 4:28 PM

Address comments (1/n)

jsetoain marked 3 inline comments as done.Nov 23 2022, 4:34 PM
jsetoain added inline comments.
mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

That's what I suspected, and the problem is that loop is the result of a dyn_cast, so it can't be dyn_cast-ed again. Thankfully, TypeSwitch does the trick.

mravishankar added inline comments.Nov 23 2022, 4:49 PM
mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp
519–532

dyn_cast, etc. works only on Operation *. When you define TypeSwitch<Operation *, ....>, the LoopLikeInterface op gets implicitly cast to Operation *, and then TypeSwitch works the same way as dyn_cast. So with the if version you would need if (auto scfForOp = dyn_cast<scf::ForOp>(loop.getOperation()))

mravishankar resigned from this revision.Nov 23 2022, 4:52 PM

I dont have further comments on this patch. This only landed on my radar cause of this statement `I'm going to push the change anyway. Best case scenario, it's something wrong with my compiler. Worst case scenario, it will leave a nice log of the error I'm seeing.`
I might have mis-judged the sentiment, but pushes that could cause failures would cause pain downstream (and would probably get reverted anyway). Apologies if I misread this.

Harbormaster completed remote builds in B199321: Diff 477649.Nov 23 2022, 6:19 PM
jsetoain updated this revision to Diff 477769.Nov 24 2022, 6:22 AM
jsetoain marked 2 inline comments as done.

Templatize function

jsetoain added a comment.EditedNov 24 2022, 6:23 AM
In D137600#3948191, @mravishankar wrote:

I dont have further comments on this patch. This only landed on my radar cause of this statement `I'm going to push the change anyway. Best case scenario, it's something wrong with my compiler. Worst case scenario, it will leave a nice log of the error I'm seeing.`
I might have mis-judged the sentiment, but pushes that could cause failures would cause pain downstream (and would probably get reverted anyway). Apologies if I misread this.

My apologies. English is not my native language, I didn't mean it like that, but I see where the confusion may have come from. What I was trying to say is "I will update this patch in revision, let the automatic checks fail where they fail for me locally, and then I can point to those logs to explain my problem". You can rest assure, I never had the intention to land this patch until all issues are clarified and everyone is happy with my code, and I would never land a patch that doesn't pass all tests in my machine :-) I appreciate the clarification on dyn_cast issue, this was a new one for me, thanks!

mlir/lib/Dialect/Affine/Utils/Utils.cpp
1877–1941

That and the access to the YieldOp operands (which I find strange). I considered the template early on, but was not convinced with the result. I'm uploading my templated version, let me know what you think. My issues are:

  1. It only really works for scf::ForOp and AffineForOp
  2. I'm getting away with not making SCF dependent on Affine only because that's the only alternative to scf::ForOp. It's creating a non-explicit dependency that I don't like.
  3. The whole thing should probably live somewhere else, but at the same time, this is code depends on a lot of for loop-related functions and methods that don't work for all possible loops.
  4. It only "coincidentally" works for both Affine and SCF fors because they have methods with the same identifiers. It feels like abusing templates to avoid copying a couple dozen lines of code.

That's why I originally scrapped it and decided in favour of total isolation. Whether this is preferable or not I will leave it up to you, you have the experience to judge :-)

jsetoain updated this revision to Diff 477770.Nov 24 2022, 6:26 AM

Remove unnecessary empty line

Harbormaster completed remote builds in B199412: Diff 477770.Nov 24 2022, 7:13 AM
bondhugula accepted this revision.Nov 24 2022, 12:31 PM

LGTM - thanks.

This revision is now accepted and ready to land.Nov 24 2022, 12:31 PM
bondhugula added inline comments.Nov 24 2022, 12:33 PM
mlir/include/mlir/Dialect/SCF/Utils/Utils.h
57 ↗(On Diff #477770)

This much C++ code in the header file isn't okay. I understand this was added here because it's a template function. Instead, you can just define it out of line and explicitly instantiate for both scf::ForOp and AffineForOp in the C++ file (two extra lines right below the definition).

jsetoain updated this revision to Diff 478976.Nov 30 2022, 9:05 AM

Replace the template with an existing function

jsetoain marked an inline comment as done.Nov 30 2022, 9:11 AM
jsetoain added inline comments.
mlir/include/mlir/Dialect/SCF/Utils/Utils.h
57 ↗(On Diff #477770)

I don't think that would have worked. Since I needed to instantiate the template in two different C++ files, the template had to be visible from both. In any case, I found a function that almost does what I need, and I can get by using that one plus some additional code right after that, which I believe is preferable.

Harbormaster completed remote builds in B200296: Diff 478976.Nov 30 2022, 9:56 AM
bondhugula added inline comments.Nov 30 2022, 6:03 PM
mlir/include/mlir/Dialect/SCF/Utils/Utils.h
57 ↗(On Diff #477770)

It doesn't matter how many different places you need to use the instantiation. What I meant was to add an explicit instantiation in Utils.cpp. Anything linking with the library would have the necessary things to link to. I don't see why it wouldn't work. For eg. see SideEffectInterfaces.cpp:

template bool mlir::hasEffect<MemoryEffects::Allocate>(Operation *, Value);
template bool mlir::hasEffect<MemoryEffects::Free>(Operation *, Value);
template bool mlir::hasEffect<MemoryEffects::Read>(Operation *, Value);
template bool mlir::hasEffect<MemoryEffects::Write>(Operation *, Value);
Herald added a reviewer: hanchung. · View Herald TranscriptNov 30 2022, 6:03 PM
jsetoain marked an inline comment as done.Dec 1 2022, 2:55 AM
jsetoain added inline comments.
mlir/include/mlir/Dialect/SCF/Utils/Utils.h
57 ↗(On Diff #477770)

Ah! I completely misunderstood what you were asking for, my apologies. Nevertheless, am I right to assume you're okay with the alternative? If so, I will leave this here for a couple more days and land it. Thanks for all your help (and patience)!

Closed by commit rG825da072a8ed: [mlir] Add hoisting of transfer ops in affine loops (authored by jsetoain). · Explain WhyDec 6 2022, 2:09 AM
This revision was automatically updated to reflect the committed changes.
jsetoain added a commit: rG825da072a8ed: [mlir] Add hoisting of transfer ops in affine loops.
hctim added a subscriber: hctim.Dec 6 2022, 9:10 AM

It looks like the update to Dialect/Linalg/hoisting.mlir breaks ASan, MSan and UBSan: https://lab.llvm.org/buildbot/#/builders/5/builds/29773

You can find instructions on reproducing the sanitizer bots here: https://github.com/google/sanitizers/wiki/SanitizerBotReproduceBuild. The buildbot_fast is the one is probably easiest to use and reflects all the breakages.

******************** TEST 'MLIR :: Dialect/Linalg/hoisting.mlir' FAILED ********************
Script:
--
: 'RUN: at line 1';   /b/sanitizer-x86_64-linux-fast/build/llvm_build_asan/bin/mlir-opt /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/Dialect/Linalg/hoisting.mlir -test-linalg-hoisting=test-hoist-redundant-transfers -allow-unregistered-dialect -split-input-file | /b/sanitizer-x86_64-linux-fast/build/llvm_build_asan/bin/FileCheck /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/Dialect/Linalg/hoisting.mlir
--
Exit Code: 1
Command Output (stderr):
--
=================================================================
==455613==ERROR: AddressSanitizer: stack-use-after-scope on address 0x7f43c2dbea50 at pc 0x5607ff79ca82 bp 0x7ffd712d54a0 sp 0x7ffd712d5498
READ of size 8 at 0x7f43c2dbea50 thread T0
    #0 0x5607ff79ca81 in mlir::ValueRange::dereference_iterator(llvm::PointerUnion<mlir::Value const*, mlir::OpOperand*, mlir::detail::OpResultImpl*> const&, long) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/OperationSupport.cpp:613:12
    #1 0x5607feb11ea4 in operator* /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:1281:14
    #2 0x5607feb11ea4 in deref<0UL, 1UL> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:768:23
    #3 0x5607feb11ea4 in operator* /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:793:12
    #4 0x5607feb11ea4 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:543:26
    #5 0x5607feb11ea4 in Case<mlir::AffineForOp, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:534:30)> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/TypeSwitch.h:122:16
    #6 0x5607feb11ea4 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:534:12
    #7 0x5607feb11ea4 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:230:14
    #8 0x5607feb11ea4 in mlir::WalkResult llvm::function_ref<mlir::WalkResult (mlir::Operation*)>::callback_fn<std::__1::enable_if<!llvm::is_one_of<mlir::vector::TransferReadOp, mlir::Operation*, mlir::Region*, mlir::Block*>::value && std::is_same<mlir::WalkResult, mlir::WalkResult>::value, mlir::WalkResult>::type mlir::detail::walk<(mlir::WalkOrder)1, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2, mlir::vector::TransferReadOp, mlir::WalkResult>(mlir::Operation*, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2&&)::'lambda'(mlir::Operation*)>(long, mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #9 0x5607ff7e8a04 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #10 0x5607ff7e8a04 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #11 0x5607ff7e8a04 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #12 0x5607ff7e8a04 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #13 0x5607feb060c4 in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::vector::TransferReadOp, mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:233:10
    #14 0x5607feb060c4 in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:579:12
    #15 0x5607feb060c4 in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/OpDefinition.h:147:19
    #16 0x5607feb060c4 in mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:10
    #17 0x5607feb050f7 in (anonymous namespace)::TestLinalgHoisting::runOnOperation() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/lib/Dialect/Linalg/TestLinalgHoisting.cpp:48:5
    #18 0x5607ff3b525a in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:465:11
    #19 0x5607ff3b6164 in mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #20 0x5607ff3b9655 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:749:36
    #21 0x5607ff3b9655 in failableParallelForEach<std::__1::__wrap_iter<OpPMInfo *>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:46:18
    #22 0x5607ff3b9655 in failableParallelForEach<std::__1::vector<OpPMInfo, std::__1::allocator<OpPMInfo> > &, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:92:10
    #23 0x5607ff3b9655 in mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:759:14
    #24 0x5607ff3b52e3 in runOnOperation /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:650:5
    #25 0x5607ff3b52e3 in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:463:14
    #26 0x5607ff3bcfb4 in runPipeline /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #27 0x5607ff3bcfb4 in runPasses /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:834:10
    #28 0x5607ff3bcfb4 in mlir::PassManager::run(mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:814:60
    #29 0x5607ff3abb7a in performActions(llvm::raw_ostream&, bool, bool, llvm::SourceMgr&, mlir::MLIRContext*, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:91:17
    #30 0x5607ff3ab260 in processBuffer /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:139:12
    #31 0x5607ff3ab260 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:181:12
    #32 0x5607ff3ab260 in mlir::LogicalResult llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>::callback_fn<mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool)::$_0>(long, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #33 0x5607ff5341a5 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #34 0x5607ff5341a5 in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool)::$_0::operator()(llvm::StringRef) const /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:88:16
    #35 0x5607ff532f46 in interleave<const llvm::StringRef *, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:43), void> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2067:5
    #36 0x5607ff532f46 in interleave<llvm::SmallVector<llvm::StringRef, 8U>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), llvm::raw_ostream, llvm::StringRef> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:3
    #37 0x5607ff532f46 in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:91:3
    #38 0x5607ff3a51f5 in mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:186:10
    #39 0x5607ff3a65ef in MlirOptMain /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:209:10
    #40 0x5607ff3a65ef in mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:306:14
    #41 0x5607fac120ff in main /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/tools/mlir-opt/mlir-opt.cpp:244:7
    #42 0x7f43c481dd8f  (/lib/x86_64-linux-gnu/libc.so.6+0x29d8f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #43 0x7f43c481de3f in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e3f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #44 0x5607fab5b5e4 in _start (/b/sanitizer-x86_64-linux-fast/build/llvm_build_asan/bin/mlir-opt+0x60ce5e4)
Address 0x7f43c2dbea50 is located in stack of thread T0 at offset 592 in frame
    #0 0x5607feb0fc9f in mlir::WalkResult llvm::function_ref<mlir::WalkResult (mlir::Operation*)>::callback_fn<std::__1::enable_if<!llvm::is_one_of<mlir::vector::TransferReadOp, mlir::Operation*, mlir::Region*, mlir::Block*>::value && std::is_same<mlir::WalkResult, mlir::WalkResult>::value, mlir::WalkResult>::type mlir::detail::walk<(mlir::WalkOrder)1, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2, mlir::vector::TransferReadOp, mlir::WalkResult>(mlir::Operation*, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2&&)::'lambda'(mlir::Operation*)>(long, mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:44
  This frame has 31 object(s):
    [32, 56) 'ref.tmp.i.i'
    [96, 104) 'newForOp.i.i24.i'
    [128, 136) 'ref.tmp.i.i25.i'
    [160, 216) 'ref.tmp14.i.i.i'
    [256, 258) 'ref.tmp3.i.i.i.i.i.i'
    [272, 280) 'newForOp.i.i.i'
    [304, 312) 'ref.tmp.i.i16.i'
    [336, 352) 'agg.tmp7.i.i.i'
    [368, 432) 'ref.tmp8.i.i.i'
    [464, 472) 'ref.tmp10.i.i.i'
    [496, 512) 'agg.tmp15.i.i.i'
    [528, 536) 'ref.tmp16.i.i.i'
    [560, 576) 'ref.tmp35.i.i.i'
    [592, 600) 'ref.tmp37.i.i.i' <== Memory access at offset 592 is inside this variable
    [624, 648) '__begin3.i.i.i'
    [688, 704) 'it.i.i.i'
    [720, 728) 'ref.tmp53.i.i.i'
    [752, 808) 'ref.tmp56.i.i.i'
    [848, 872) 'ref.tmp.i517.i.i'
    [912, 936) 'ref.tmp.i444.i.i'
    [976, 1000) 'ref.tmp.i365.i.i'
    [1040, 1048) 'ref.tmp.i.i.i.i'
    [1072, 1088) 'ref.tmp.i.i.i'
    [1104, 1112) 'transferRead.i.i'
    [1136, 1152) 'loop.i.i'
    [1168, 1224) 'forwardSlice.i.i'
    [1264, 1272) 'transferWrite.i.i'
    [1296, 1304) 'candidateWrite.i.i'
    [1328, 1360) 'dom.i.i'
    [1392, 1424) 'b.i.i'
    [1456, 1504) 'yieldFn.i.i'
HINT: this may be a false positive if your program uses some custom stack unwind mechanism, swapcontext or vfork
      (longjmp and C++ exceptions *are* supported)
SUMMARY: AddressSanitizer: stack-use-after-scope /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/OperationSupport.cpp:613:12 in mlir::ValueRange::dereference_iterator(llvm::PointerUnion<mlir::Value const*, mlir::OpOperand*, mlir::detail::OpResultImpl*> const&, long)
Shadow bytes around the buggy address:
  0x0fe8f85afcf0: f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5 f5
  0x0fe8f85afd00: f1 f1 f1 f1 f8 f8 f8 f2 f2 f2 f2 f2 f8 f2 f2 f2
  0x0fe8f85afd10: f8 f2 f2 f2 f8 f8 f8 f8 f8 f8 f8 f2 f2 f2 f2 f2
  0x0fe8f85afd20: f8 f2 00 f2 f2 f2 f8 f2 f2 f2 00 00 f2 f2 f8 f8
  0x0fe8f85afd30: f8 f8 f8 f8 f8 f8 f2 f2 f2 f2 f8 f2 f2 f2 00 00
=>0x0fe8f85afd40: f2 f2 f8 f2 f2 f2 f8 f8 f2 f2[f8]f2 f2 f2 00 00
  0x0fe8f85afd50: 00 f2 f2 f2 f2 f2 00 00 f2 f2 f8 f2 f2 f2 f8 f8
  0x0fe8f85afd60: f8 f8 f8 f8 f8 f2 f2 f2 f2 f2 f8 f8 f8 f2 f2 f2
  0x0fe8f85afd70: f2 f2 f8 f8 f8 f2 f2 f2 f2 f2 f8 f8 f8 f2 f2 f2
  0x0fe8f85afd80: f2 f2 f8 f2 f2 f2 f8 f8 f2 f2 00 f2 f2 f2 00 00
  0x0fe8f85afd90: f2 f2 00 00 00 00 00 00 00 f2 f2 f2 f2 f2 00 f2
Shadow byte legend (one shadow byte represents 8 application bytes):
  Addressable:           00
  Partially addressable: 01 02 03 04 05 06 07 
  Heap left redzone:       fa
  Freed heap region:       fd
  Stack left redzone:      f1
  Stack mid redzone:       f2
  Stack right redzone:     f3
  Stack after return:      f5
  Stack use after scope:   f8
  Global redzone:          f9
  Global init order:       f6
  Poisoned by user:        f7
  Container overflow:      fc
  Array cookie:            ac
  Intra object redzone:    bb
  ASan internal:           fe
  Left alloca redzone:     ca
  Right alloca redzone:    cb
==455613==ABORTING
******************** TEST 'MLIR :: Dialect/Linalg/hoisting.mlir' FAILED ********************
Script:
--
: 'RUN: at line 1';   /b/sanitizer-x86_64-linux-fast/build/llvm_build_ubsan/bin/mlir-opt /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/Dialect/Linalg/hoisting.mlir -test-linalg-hoisting=test-hoist-redundant-transfers -allow-unregistered-dialect -split-input-file | /b/sanitizer-x86_64-linux-fast/build/llvm_build_ubsan/bin/FileCheck /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/Dialect/Linalg/hoisting.mlir
--
Exit Code: 1
Command Output (stderr):
--
/b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:165:52: runtime error: member call on misaligned address 0x000000000051 for type 'mlir::Block', which requires 8 byte alignment
0x000000000051: note: pointer points here
<memory cannot be printed>
    #0 0x5583a6b717ba in getParentOp /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:165:52
    #1 0x5583a6b717ba in mlir::Operation::isProperAncestor(mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Operation.cpp:177:26
    #2 0x5583a6b99283 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #3 0x5583a6b99283 in mlir::Value::replaceUsesWithIf(mlir::Value, llvm::function_ref<bool (mlir::OpOperand&)>) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Value.cpp:85:9
    #4 0x5583a6718589 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:545:31
    #5 0x5583a6718589 in llvm::TypeSwitch<mlir::Operation*, mlir::WalkResult>& llvm::TypeSwitch<mlir::Operation*, mlir::WalkResult>::Case<mlir::AffineForOp, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2::operator()(mlir::vector::TransferReadOp) const::'lambda'(mlir::AffineForOp)>(mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2::operator()(mlir::vector::TransferReadOp) const::'lambda'(mlir::AffineForOp)&&) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/TypeSwitch.h:122:16
    #6 0x5583a6717ddf in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:534:12
    #7 0x5583a6717ddf in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:230:14
    #8 0x5583a6717ddf in mlir::WalkResult llvm::function_ref<mlir::WalkResult (mlir::Operation*)>::callback_fn<std::__1::enable_if<!llvm::is_one_of<mlir::vector::TransferReadOp, mlir::Operation*, mlir::Region*, mlir::Block*>::value && std::is_same<mlir::WalkResult, mlir::WalkResult>::value, mlir::WalkResult>::type mlir::detail::walk<(mlir::WalkOrder)1, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2, mlir::vector::TransferReadOp, mlir::WalkResult>(mlir::Operation*, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2&&)::'lambda'(mlir::Operation*)>(long, mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #9 0x5583a6b9e8d6 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #10 0x5583a6b9e8d6 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:181:12
    #11 0x5583a6b9e852 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #12 0x5583a6b9e852 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #13 0x5583a6b9e852 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #14 0x5583a6b9e852 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #15 0x5583a67143be in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::vector::TransferReadOp, mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:233:10
    #16 0x5583a67143be in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:579:12
    #17 0x5583a67143be in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/OpDefinition.h:147:19
    #18 0x5583a67143be in mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:10
    #19 0x5583a671402b in (anonymous namespace)::TestLinalgHoisting::runOnOperation() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/lib/Dialect/Linalg/TestLinalgHoisting.cpp:48:5
    #20 0x5583a69eda15 in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:465:11
    #21 0x5583a69ee212 in mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #22 0x5583a69f61bb in mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::$_14::operator()(mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool)::OpPMInfo&) const /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:749:36
    #23 0x5583a69ef9fe in failableParallelForEach<std::__1::__wrap_iter<OpPMInfo *>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:46:18
    #24 0x5583a69ef9fe in failableParallelForEach<std::__1::vector<OpPMInfo, std::__1::allocator<OpPMInfo> > &, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:92:10
    #25 0x5583a69ef9fe in mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:759:14
    #26 0x5583a69ed94a in runOnOperation /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:650:5
    #27 0x5583a69ed94a in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:463:14
    #28 0x5583a69ee212 in mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #29 0x5583a69f0fa5 in runPasses /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:834:10
    #30 0x5583a69f0fa5 in mlir::PassManager::run(mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:814:60
    #31 0x5583a69e9ff7 in performActions(llvm::raw_ostream&, bool, bool, llvm::SourceMgr&, mlir::MLIRContext*, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:91:17
    #32 0x5583a69e9c47 in processBuffer /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:139:12
    #33 0x5583a69e9c47 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:181:12
    #34 0x5583a69e9c47 in mlir::LogicalResult llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>::callback_fn<mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool)::$_0>(long, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #35 0x5583a6a9253c in llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>::operator()(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) const /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #36 0x5583a6a92703 in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool)::$_0::operator()(llvm::StringRef) const /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:88:16
    #37 0x5583a6a923af in interleave<const llvm::StringRef *, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:43), void> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2067:5
    #38 0x5583a6a923af in interleave<llvm::SmallVector<llvm::StringRef, 8U>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), llvm::raw_ostream, llvm::StringRef> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:3
    #39 0x5583a6a923af in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:91:3
    #40 0x5583a69e8556 in mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:186:10
    #41 0x5583a69e8648 in mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, mlir::PassPipelineCLParser const&, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:209:10
    #42 0x5583a69e8ca9 in mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:306:14
    #43 0x5583a5124295 in main /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/tools/mlir-opt/mlir-opt.cpp:244:7
    #44 0x7fe2f378ad8f  (/lib/x86_64-linux-gnu/libc.so.6+0x29d8f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #45 0x7fe2f378ae3f in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e3f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #46 0x5583a510b364 in _start (/b/sanitizer-x86_64-linux-fast/build/llvm_build_ubsan/bin/mlir-opt+0x404e364)
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:165:52 in
==1418730==WARNING: MemorySanitizer: use-of-uninitialized-value
    #0 0x55da63fd5ae3 in use_begin /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/UseDefLists.h:202:56
    #1 0x55da63fd5ae3 in use_begin /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Value.h:189:49
    #2 0x55da63fd5ae3 in getUses /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Value.h:193:39
    #3 0x55da63fd5ae3 in mlir::Value::replaceUsesWithIf(mlir::Value, llvm::function_ref<bool (mlir::OpOperand&)>) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Value.cpp:84:52
    #4 0x55da63431716 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:545:31
    #5 0x55da63431716 in Case<mlir::AffineForOp, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:534:30)> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/TypeSwitch.h:122:16
    #6 0x55da63431716 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:534:12
    #7 0x55da63431716 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:230:14
    #8 0x55da63431716 in mlir::WalkResult llvm::function_ref<mlir::WalkResult (mlir::Operation*)>::callback_fn<std::__1::enable_if<!llvm::is_one_of<mlir::vector::TransferReadOp, mlir::Operation*, mlir::Region*, mlir::Block*>::value && std::is_same<mlir::WalkResult, mlir::WalkResult>::value, mlir::WalkResult>::type mlir::detail::walk<(mlir::WalkOrder)1, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2, mlir::vector::TransferReadOp, mlir::WalkResult>(mlir::Operation*, mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp)::$_2&&)::'lambda'(mlir::Operation*)>(long, mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #9 0x55da63fdfb44 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #10 0x55da63fdfb44 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:181:12
    #11 0x55da63fdfa78 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #12 0x55da63fdfa78 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #13 0x55da63fdfa78 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #14 0x55da63fdfa78 in mlir::detail::walk(mlir::Operation*, llvm::function_ref<mlir::WalkResult (mlir::Operation*)>, mlir::WalkOrder) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/IR/Visitors.cpp:174:13
    #15 0x55da6342641a in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::vector::TransferReadOp, mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Visitors.h:233:10
    #16 0x55da6342641a in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Operation.h:579:12
    #17 0x55da6342641a in walk<(mlir::WalkOrder)1, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:15), mlir::WalkResult> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/OpDefinition.h:147:19
    #18 0x55da6342641a in mlir::linalg::hoistRedundantVectorTransfers(mlir::func::FuncOp) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp:425:10
    #19 0x55da63425946 in (anonymous namespace)::TestLinalgHoisting::runOnOperation() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/test/lib/Dialect/Linalg/TestLinalgHoisting.cpp:48:5
    #20 0x55da63bf1614 in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:465:11
    #21 0x55da63bf258b in mlir::detail::OpToOpPassAdaptor::runPipeline(mlir::OpPassManager&, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int, mlir::PassInstrumentor*, mlir::PassInstrumentation::PipelineParentInfo const*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #22 0x55da63bf62b9 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:749:36
    #23 0x55da63bf62b9 in failableParallelForEach<std::__1::__wrap_iter<OpPMInfo *>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:46:18
    #24 0x55da63bf62b9 in failableParallelForEach<std::__1::vector<OpPMInfo, std::__1::allocator<OpPMInfo> > &, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:739:20) &> /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/Threading.h:92:10
    #25 0x55da63bf62b9 in mlir::detail::OpToOpPassAdaptor::runOnOperationAsyncImpl(bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:759:14
    #26 0x55da63bf16a2 in runOnOperation /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:650:5
    #27 0x55da63bf16a2 in mlir::detail::OpToOpPassAdaptor::run(mlir::Pass*, mlir::Operation*, mlir::AnalysisManager, bool, unsigned int) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:463:14
    #28 0x55da63bf99d4 in runPipeline /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:529:16
    #29 0x55da63bf99d4 in runPasses /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:834:10
    #30 0x55da63bf99d4 in mlir::PassManager::run(mlir::Operation*) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Pass/Pass.cpp:814:60
    #31 0x55da63be8316 in performActions(llvm::raw_ostream&, bool, bool, llvm::SourceMgr&, mlir::MLIRContext*, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:91:17
    #32 0x55da63be79a8 in processBuffer /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:139:12
    #33 0x55da63be79a8 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:181:12
    #34 0x55da63be79a8 in mlir::LogicalResult llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>::callback_fn<mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool)::$_0>(long, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&) /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:45:12
    #35 0x55da63d5aec0 in operator() /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLFunctionalExtras.h:68:12
    #36 0x55da63d5aec0 in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool)::$_0::operator()(llvm::StringRef) const /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:88:16
    #37 0x55da63d5a712 in interleave<const llvm::StringRef *, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:43), void> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2067:5
    #38 0x55da63d5a712 in interleave<llvm::SmallVector<llvm::StringRef, 8U>, (lambda at /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:81:23), llvm::raw_ostream, llvm::StringRef> /b/sanitizer-x86_64-linux-fast/build/llvm-project/llvm/include/llvm/ADT/STLExtras.h:2085:3
    #39 0x55da63d5a712 in mlir::splitAndProcessBuffer(std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::raw_ostream&)>, llvm::raw_ostream&, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Support/ToolUtilities.cpp:91:3
    #40 0x55da63be2c46 in mlir::MlirOptMain(llvm::raw_ostream&, std::__1::unique_ptr<llvm::MemoryBuffer, std::__1::default_delete<llvm::MemoryBuffer>>, llvm::function_ref<mlir::LogicalResult (mlir::PassManager&)>, mlir::DialectRegistry&, bool, bool, bool, bool, bool, bool, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:186:10
    #41 0x55da63be3de7 in MlirOptMain /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:209:10
    #42 0x55da63be3de7 in mlir::MlirOptMain(int, char**, llvm::StringRef, mlir::DialectRegistry&, bool) /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/lib/Tools/mlir-opt/MlirOptMain.cpp:306:14
    #43 0x55da5fe0ae09 in main /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/tools/mlir-opt/mlir-opt.cpp:244:7
    #44 0x7ff04c884d8f  (/lib/x86_64-linux-gnu/libc.so.6+0x29d8f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #45 0x7ff04c884e3f in __libc_start_main (/lib/x86_64-linux-gnu/libc.so.6+0x29e3f) (BuildId: 69389d485a9793dbe873f0ea2c93e02efaa9aa3d)
    #46 0x55da5fd8d324 in _start (/b/sanitizer-x86_64-linux-fast/build/llvm_build_msan/bin/mlir-opt+0x22bb324)
SUMMARY: MemorySanitizer: use-of-uninitialized-value /b/sanitizer-x86_64-linux-fast/build/llvm-project/mlir/include/mlir/IR/UseDefLists.h:202:56 in use_begin
hctim added a reverting change: rG969f0cba7edf: Revert "[mlir] Add hoisting of transfer ops in affine loops".Dec 6 2022, 9:45 AM
jsetoain added a comment.Dec 6 2022, 9:50 AM
In D137600#3974885, @hctim wrote:

It looks like the update to Dialect/Linalg/hoisting.mlir breaks ASan, MSan and UBSan: https://lab.llvm.org/buildbot/#/builders/5/builds/29773

Yes, I saw that, I'm working on it right now. Thanks for the suggestions!

jsetoain reopened this revision.Dec 6 2022, 4:04 PM
This revision is now accepted and ready to land.Dec 6 2022, 4:04 PM
jsetoain updated this revision to Diff 480670.Dec 6 2022, 4:04 PM

Fix sanitizer issues

Harbormaster completed remote builds in B201528: Diff 480670.Dec 7 2022, 12:27 AM
jsetoain updated this revision to Diff 480816.Dec 7 2022, 1:59 AM

clang-format

Harbormaster completed remote builds in B201641: Diff 480816.Dec 7 2022, 11:11 AM
Closed by commit rGda291bab8120: [mlir] Add hoisting of transfer ops in affine loops (authored by jsetoain). · Explain WhyDec 7 2022, 12:10 PM
This revision was automatically updated to reflect the committed changes.
jsetoain added a commit: rGda291bab8120: [mlir] Add hoisting of transfer ops in affine loops.

Revision Contents

PathSize
mlir/
include/
mlir/
Dialect/
Affine/
8 lines
lib/
Dialect/
Affine/
Utils/
69 lines
Linalg/
Transforms/
23 lines
test/
Dialect/
Linalg/
36 lines

Diff 473827

mlir/include/mlir/Dialect/Affine/Utils.h

Show First 20 Lines • Show All 321 Lines • ▼ Show 20 Lines
/// (noninclusive) and prior to `memOp` (e.g. on a control flow/op path /// (noninclusive) and prior to `memOp` (e.g. on a control flow/op path
/// between the operations) do not have the potential memory effect /// between the operations) do not have the potential memory effect
/// `EffectType` on `memOp`. `memOp` is an operation that reads or writes to /// `EffectType` on `memOp`. `memOp` is an operation that reads or writes to
/// a memref. For example, if `EffectType` is MemoryEffects::Write, this method /// a memref. For example, if `EffectType` is MemoryEffects::Write, this method
/// will check if there is no write to the memory between `start` and `memOp` /// will check if there is no write to the memory between `start` and `memOp`
/// that would change the read within `memOp`. /// that would change the read within `memOp`.
template <typename EffectType, typename T> template <typename EffectType, typename T>
bool hasNoInterveningEffect(Operation *start, T memOp); bool hasNoInterveningEffect(Operation *start, T memOp);
using NewYieldValueFn = std::function<SmallVector<Value>(
OpBuilder &b, Location loc, ArrayRef<BlockArgument> newBBArgs)>;
AffineForOp replaceLoopWithNewYields(OpBuilder &builder, AffineForOp loop,
ValueRange newIterOperands,
const NewYieldValueFn &newYieldValuesFn,
bool replaceIterOperandsUsesInLoop = true);
bondhugulaUnsubmitted
Done
ReplyInline Actions

Nit: doesn't
Similarly, apostrophes missing below.

bondhugula: Nit: doesn't Similarly, apostrophes missing below.
} // namespace mlir } // namespace mlir
#endif // MLIR_DIALECT_AFFINE_UTILS_H #endif // MLIR_DIALECT_AFFINE_UTILS_H

mlir/lib/Dialect/Affine/Utils/Utils.cpp

Show First 20 Lines • Show All 1,866 Lines • ▼ Show 20 LinesFailureOr<SmallVector<Value>> mlir::delinearizeIndex(OpBuilder &b, Location loc,
for (Value divisor : divisors) { for (Value divisor : divisors) {
DivModValue divMod = getDivMod(b, loc, residual, divisor); DivModValue divMod = getDivMod(b, loc, residual, divisor);
results.push_back(divMod.quotient); results.push_back(divMod.quotient);
residual = divMod.remainder; residual = divMod.remainder;
} }
results.push_back(residual); results.push_back(residual);
return results; return results;
} }
AffineForOp
mlir::replaceLoopWithNewYields(OpBuilder &builder, AffineForOp loop,
ValueRange newIterOperands,
const NewYieldValueFn &newYieldValuesFn,
bool replaceIterOperandsUsesInLoop) {
// Create a new loop before the existing one, with the extra operands.
OpBuilder::InsertionGuard g(builder);
builder.setInsertionPoint(loop);
auto operands = llvm::to_vector(loop.getIterOperands());
operands.append(newIterOperands.begin(), newIterOperands.end());
AffineForOp newLoop = builder.create<AffineForOp>(
loop.getLoc(), loop.getLowerBound().getOperands(), loop.getLowerBound().getMap(),
loop.getUpperBound().getOperands(), loop.getUpperBound().getMap(), loop.getStep(),
operands, [](OpBuilder &, Location, Value, ValueRange) {});
Block *loopBody = loop.getBody();
Block *newLoopBody = newLoop.getBody();
// Move the body of the original loop to the new loop.
newLoopBody->getOperations().splice(newLoopBody->end(),
loopBody->getOperations());
// Generate the new yield values to use by using the callback and append the
// yield values to the scf.yield operation.
auto yield = cast<AffineYieldOp>(newLoopBody->getTerminator());
ArrayRef<BlockArgument> newBBArgs =
newLoopBody->getArguments().take_back(newIterOperands.size());
{
OpBuilder::InsertionGuard g(builder);
builder.setInsertionPoint(yield);
SmallVector<Value> newYieldedValues =
newYieldValuesFn(builder, loop.getLoc(), newBBArgs);
assert(newIterOperands.size() == newYieldedValues.size() &&
"expected as many new yield values as new iter operands");
yield.operandsMutable().append(newYieldedValues);
}
// Remap the BlockArguments from the original loop to the new loop
// BlockArguments.
ArrayRef<BlockArgument> bbArgs = loopBody->getArguments();
for (auto it :
llvm::zip(bbArgs, newLoopBody->getArguments().take_front(bbArgs.size())))
std::get<0>(it).replaceAllUsesWith(std::get<1>(it));
if (replaceIterOperandsUsesInLoop) {
// Replace all uses of `newIterOperands` with the corresponding basic block
// arguments.
for (auto it : llvm::zip(newIterOperands, newBBArgs)) {
std::get<0>(it).replaceUsesWithIf(std::get<1>(it), [&](OpOperand &use) {
Operation *user = use.getOwner();
return newLoop->isProperAncestor(user);
});
}
}
// Replace all uses of the original loop with corresponding values from the
// new loop.
loop.replaceAllUsesWith(
newLoop.getResults().take_front(loop.getNumResults()));
// Add a fake yield to the original loop body that just returns the
// BlockArguments corresponding to the iter_args. This makes it a no-op loop.
// The loop is dead. The caller is expected to erase it.
builder.setInsertionPointToEnd(loopBody);
builder.create<AffineYieldOp>(loop->getLoc(), loop.getRegionIterArgs());
bondhugulaUnsubmitted
Not Done
ReplyInline Actions

Can we templatize this function and make it work for both AffineForOp and scf::ForOp? You'll need YieldOpTy (can be scf::YieldOp and AffineYieldOp) and ForOpTy (can be scf::ForOp or AffineForOp). Other than the op creation calls, isn't everything else the same?

bondhugula: Can we templatize this function and make it work for both `AffineForOp` and scf::ForOp? You'll…
jsetoainAuthorUnsubmitted
Done
ReplyInline Actions

That and the access to the YieldOp operands (which I find strange). I considered the template early on, but was not convinced with the result. I'm uploading my templated version, let me know what you think. My issues are:

  1. It only really works for scf::ForOp and AffineForOp
  2. I'm getting away with not making SCF dependent on Affine only because that's the only alternative to scf::ForOp. It's creating a non-explicit dependency that I don't like.
  3. The whole thing should probably live somewhere else, but at the same time, this is code depends on a lot of for loop-related functions and methods that don't work for all possible loops.
  4. It only "coincidentally" works for both Affine and SCF fors because they have methods with the same identifiers. It feels like abusing templates to avoid copying a couple dozen lines of code.

That's why I originally scrapped it and decided in favour of total isolation. Whether this is preferable or not I will leave it up to you, you have the experience to judge :-)

jsetoain: That and the access to the `YieldOp` operands (which I find strange). I considered the template…
return newLoop;
}

mlir/lib/Dialect/Linalg/Transforms/Hoisting.cpp

Show All 9 Lines
// in the context of Linalg transformations. // in the context of Linalg transformations.
// //
//===----------------------------------------------------------------------===// //===----------------------------------------------------------------------===//
#include "mlir/Dialect/Linalg/Transforms/Hoisting.h" #include "mlir/Dialect/Linalg/Transforms/Hoisting.h"
#include "mlir/Analysis/SliceAnalysis.h" #include "mlir/Analysis/SliceAnalysis.h"
#include "mlir/Dialect/Affine/Analysis/AffineStructures.h" #include "mlir/Dialect/Affine/Analysis/AffineStructures.h"
#include "mlir/Dialect/Affine/IR/AffineValueMap.h" #include "mlir/Dialect/Affine/IR/AffineValueMap.h"
#include "mlir/Dialect/Affine/IR/AffineOps.h"
#include "mlir/Dialect/Affine/Utils.h"
#include "mlir/Dialect/Arith/IR/Arith.h" #include "mlir/Dialect/Arith/IR/Arith.h"
#include "mlir/Dialect/Func/IR/FuncOps.h" #include "mlir/Dialect/Func/IR/FuncOps.h"
#include "mlir/Dialect/Linalg/IR/Linalg.h" #include "mlir/Dialect/Linalg/IR/Linalg.h"
#include "mlir/Dialect/Linalg/Transforms/Transforms.h" #include "mlir/Dialect/Linalg/Transforms/Transforms.h"
#include "mlir/Dialect/SCF/IR/SCF.h" #include "mlir/Dialect/SCF/IR/SCF.h"
#include "mlir/Dialect/SCF/Utils/Utils.h" #include "mlir/Dialect/SCF/Utils/Utils.h"
#include "mlir/Dialect/Tensor/IR/Tensor.h" #include "mlir/Dialect/Tensor/IR/Tensor.h"
#include "mlir/Dialect/Vector/IR/VectorOps.h" #include "mlir/Dialect/Vector/IR/VectorOps.h"
▲ Show 20 Lines • Show All 394 Lines • ▼ Show 20 Lines func.walk(
[&](LoopLikeOpInterface loopLike) { moveLoopInvariantCode(loopLike); }); [&](LoopLikeOpInterface loopLike) { moveLoopInvariantCode(loopLike); });
func.walk([&](vector::TransferReadOp transferRead) { func.walk([&](vector::TransferReadOp transferRead) {
if (!transferRead.getShapedType().isa<MemRefType>()) if (!transferRead.getShapedType().isa<MemRefType>())
return WalkResult::advance(); return WalkResult::advance();
LLVM_DEBUG(DBGS() << "Candidate for hoisting: " LLVM_DEBUG(DBGS() << "Candidate for hoisting: "
<< *transferRead.getOperation() << "\n"); << *transferRead.getOperation() << "\n");
auto loop = dyn_cast<scf::ForOp>(transferRead->getParentOp()); auto loop = dyn_cast<LoopLikeOpInterface>(transferRead->getParentOp());
LLVM_DEBUG(DBGS() << "Parent op: " << *transferRead->getParentOp() LLVM_DEBUG(DBGS() << "Parent op: " << *transferRead->getParentOp()
<< "\n"); << "\n");
if (!loop) if (!loop)
return WalkResult::advance(); return WalkResult::advance();
LLVM_DEBUG(DBGS() << "Candidate read: " << *transferRead.getOperation() LLVM_DEBUG(DBGS() << "Candidate read: " << *transferRead.getOperation()
<< "\n"); << "\n");
▲ Show 20 Lines • Show All 71 Lines • ▼ Show 20 Lines func.walk([&](vector::TransferReadOp transferRead) {
transferWrite->moveAfter(loop); transferWrite->moveAfter(loop);
// Rewrite `loop` with new yields by cloning and erase the original loop. // Rewrite `loop` with new yields by cloning and erase the original loop.
OpBuilder b(transferRead); OpBuilder b(transferRead);
NewYieldValueFn yieldFn = [&](OpBuilder &b, Location loc, NewYieldValueFn yieldFn = [&](OpBuilder &b, Location loc,
ArrayRef<BlockArgument> newBBArgs) { ArrayRef<BlockArgument> newBBArgs) {
return SmallVector<Value>{transferWrite.getVector()}; return SmallVector<Value>{transferWrite.getVector()};
}; };
if (isa<scf::ForOp>(loop)) {
auto newForOp = auto newForOp =
replaceLoopWithNewYields(b, loop, transferRead.getVector(), yieldFn); replaceLoopWithNewYields(b, cast<scf::ForOp>(loop), transferRead.getVector(), yieldFn);
// Transfer write has been hoisted, need to update the written vector by // Transfer write has been hoisted, need to update the written vector by
// the value yielded by the newForOp. // the value yielded by the newForOp.
transferWrite.getVectorMutable().assign(newForOp.getResults().back()); transferWrite.getVectorMutable().assign(newForOp.getResults().back());
} else if (isa<AffineForOp>(loop)) {
bondhugulaUnsubmitted
Done
ReplyInline Actions

dyn_cast here again.

bondhugula: `dyn_cast` here again.
auto newForOp =
replaceLoopWithNewYields(b, cast<AffineForOp>(loop), transferRead.getVector(), yieldFn);
// Transfer write has been hoisted, need to update the written vector by
// the value yielded by the newForOp.
transferWrite.getVectorMutable().assign(newForOp.getResults().back());
}
bondhugulaUnsubmitted
Done
ReplyInline Actions

dyn_cast here to avoid an extra check below.

bondhugula: `dyn_cast` here to avoid an extra check below.
jsetoainAuthorUnsubmitted
Done
ReplyInline Actions

I haven't found a way to dyn_cast the result of the dyn_cast. I'm not very familiar with llvm's static polymorphism constructs, so I'm not sure what the problem is, but I get errors like:

error: invalid cast of an rvalue expression of type ‘std::nullptr_t’ to type ‘llvm::CastInfo<mlir::scf::ForOp, mlir::LoopLikeOpInterface, void>::CastReturnType’ {aka ‘mlir::scf::ForOp&’}

Alternatively, I can write: dyn_cast<scf::ForOp>(transferRead->getParentOp()) but I'm not sure it's worth the loss of clarity.

jsetoain: I haven't found a way to `dyn_cast` the result of the `dyn_cast`. I'm not very familiar with…
bondhugulaUnsubmitted
Done
ReplyInline Actions

Please see https://llvm.org/docs/ProgrammersManual.html#the-isa-cast-and-dyn-cast-templates.

Use dyn_cast instead of isa + cast, i.e.,

if (auto scfFor = dyn_cast<scf::ForOp>(loop)) {
  ...
bondhugula: Please see https://llvm.org/docs/ProgrammersManual.html#the-isa-cast-and-dyn-cast-templates.
jsetoainAuthorUnsubmitted
Done
ReplyInline Actions

I believe I didn't explain myself very well there. I understand how LLVM's RTTI is supposed to work, I don't understand why it's not working. I'm getting that error in my previous comment, but I don't understand enough about the implementation to figure out why.

I'm going to push the change anyway. Best case scenario, it's something wrong with my compiler. Worst case scenario, it will leave a nice log of the error I'm seeing.

jsetoain: I believe I didn't explain myself very well there. I understand how LLVM's RTTI is supposed to…
mravishankarUnsubmitted
Done
ReplyInline Actions

I am missing overall context, but you cant dyn_cast result of the dyn_cast....

IIUC what is being suggested here is

if (auto scfForOp = dyn_cast<scf::ForOp>(loop)) {
  ....
} else if (auto affineForOp = dyn_cast<AffineForOp>(loop)) {

}

or better yet, use TypeSwitch

return TypeSwitch<Operation *, WalkResult>(loop)
  .Case<scf::ForOp>([](scf::ForOp scfForOp) {...})
  .Case<AffineForOp>([](AffineForOp affineForOp) {...})
  .Default([](Operation *) { return WalkResult::interrupt(); }

(and move the loop.erase into the if in first case or callback for the Case in the second example).

mravishankar: I am missing overall context, but you cant `dyn_cast` result of the `dyn_cast`.... IIUC what…
jsetoainAuthorUnsubmitted
Done
ReplyInline Actions

That's what I suspected, and the problem is that loop is the result of a dyn_cast, so it can't be dyn_cast-ed again. Thankfully, TypeSwitch does the trick.

jsetoain: That's what I suspected, and the problem is that `loop` is the result of a `dyn_cast`, so it…
mravishankarUnsubmitted
Done
ReplyInline Actions

dyn_cast, etc. works only on Operation *. When you define TypeSwitch<Operation *, ....>, the LoopLikeInterface op gets implicitly cast to Operation *, and then TypeSwitch works the same way as dyn_cast. So with the if version you would need if (auto scfForOp = dyn_cast<scf::ForOp>(loop.getOperation()))

mravishankar: `dyn_cast`, etc. works only on `Operation *`. When you define `TypeSwitch<Operation *, ....>`…
bondhugulaUnsubmitted
Done
ReplyInline Actions

You should actually check earlier and advance the walk if it's something other than affine.for or scf.for at line 428; either that or assert here with an extra else or use cast above for AffineForOp in the conditional.

bondhugula: You should actually check earlier and advance the walk if it's something other than affine.for…
changed = true; changed = true;
loop.erase(); loop.erase();
// Need to interrupt and restart because erasing the loop messes up the // Need to interrupt and restart because erasing the loop messes up the
// walk. // walk.
return WalkResult::interrupt(); return WalkResult::interrupt();
}); });
} }
} }

mlir/test/Dialect/Linalg/hoisting.mlir

Show First 20 Lines • Show All 463 Lines • ▼ Show 20 Lines %1 = scf.for %j = %lb to %ub step %step iter_args(%arg5 = %tensor)
%r01 = vector.transfer_read %w10[%c0, %c3], %cst: tensor<?x?xf32>, vector<2xf32> %r01 = vector.transfer_read %w10[%c0, %c3], %cst: tensor<?x?xf32>, vector<2xf32>
%u01 = "some_use"(%r01) : (vector<2xf32>) -> vector<2xf32> %u01 = "some_use"(%r01) : (vector<2xf32>) -> vector<2xf32>
%w11 = vector.transfer_write %u01, %w10[%c0, %c3] : vector<2xf32>, tensor<?x?xf32> %w11 = vector.transfer_write %u01, %w10[%c0, %c3] : vector<2xf32>, tensor<?x?xf32>
scf.yield %w11 : tensor<?x?xf32> scf.yield %w11 : tensor<?x?xf32>
} }
return %1 : tensor<?x?xf32> return %1 : tensor<?x?xf32>
} }
// -----
// CHECK-LABEL: func @hoist_vector_transfer_pairs_in_affine_loops(
// CHECK-SAME: %[[MEMREF0:[a-zA-Z0-9]+]]: memref<64x64xi32>,
// CHECK-SAME: %[[MEMREF1:[a-zA-Z0-9]+]]: memref<64x64xi32>,
// CHECK-SAME: %[[MEMREF2:[a-zA-Z0-9]+]]: memref<64x64xi32>) {
// CHECK: %[[C0:.*]] = arith.constant 0 : i32
// CHECK: affine.for %[[I:.*]] = 0 to 64 {
// CHECK: affine.for %[[J:.*]] = 0 to 64 step 16 {
// CHECK: %[[R0:.*]] = vector.transfer_read %[[MEMREF2]][%[[I]], %[[J]]], %[[C0]] : memref<64x64xi32>, vector<16xi32>
// CHECK: %[[R:.*]] = affine.for %[[K:.*]] = 0 to 64 iter_args(%[[ACC:.*]] = %[[R0]]) -> (vector<16xi32>) {
// CHECK: %[[AV:.*]] = vector.transfer_read %[[MEMREF0]][%[[I]], %[[K]]], %[[C0]] {{.*}}: memref<64x64xi32>, vector<16xi32>
// CHECK: %[[BV:.*]] = vector.transfer_read %[[MEMREF1]][%[[K]], %[[J]]], %[[C0]] {{.*}}: memref<64x64xi32>, vector<16xi32>
// CHECK: %[[T0:.*]] = arith.muli %[[AV]], %[[BV]] : vector<16xi32>
// CHECK: %[[T1:.*]] = arith.addi %[[ACC]], %[[T0]] : vector<16xi32>
// CHECK: affine.yield %[[T1]] : vector<16xi32>
// CHECK: }
// CHECK: vector.transfer_write %[[R]], %[[MEMREF2]][%[[I]], %[[J]]] : vector<16xi32>, memref<64x64xi32>
// CHECK: }
// CHECK: }
func.func @hoist_vector_transfer_pairs_in_affine_loops(%memref0: memref<64x64xi32>, %memref1: memref<64x64xi32>, %memref2: memref<64x64xi32>) {
%c0_i32 = arith.constant 0 : i32
affine.for %arg3 = 0 to 64 {
affine.for %arg4 = 0 to 64 step 16 {
affine.for %arg5 = 0 to 64 {
%0 = vector.transfer_read %memref0[%arg3, %arg5], %c0_i32 {permutation_map = affine_map<(d0, d1) -> (0)>} : memref<64x64xi32>, vector<16xi32>
%1 = vector.transfer_read %memref1[%arg5, %arg4], %c0_i32 : memref<64x64xi32>, vector<16xi32>
%2 = vector.transfer_read %memref2[%arg3, %arg4], %c0_i32 : memref<64x64xi32>, vector<16xi32>
%3 = arith.muli %0, %1 : vector<16xi32>
%4 = arith.addi %2, %3 : vector<16xi32>
vector.transfer_write %4, %memref2[%arg3, %arg4] : vector<16xi32>, memref<64x64xi32>
}
}
}
return
}