Align the bufferization pipeline according to the upstream doc #295
Conversation
| pm.addPass(bufferization::createOneShotBufferizePass(options)); | ||
| pm.addPass(createCanonicalizerPass()); | ||
| pm.addPass(createCSEPass()); | ||
| pm.addPass(createCanonicalizerPass()); |
There was a problem hiding this comment.
L101 is already CanonicalizerPass, do we really need to do it again after CSE?
There was a problem hiding this comment.
Yes, the first CanonicalizerPass is mainly used to canonicalize the scf.for and scf.forall to the canonical form in memref, exposing more CSE optimization opportunities. The CSE pass is used to eliminate the common memref.subview so there will be some useless memref.copy memref.copy %a, %a. The last CanonicalizerPass is used to eliminate the useless memref.copy which copies from A to A like memref.copy %a, %a
lib/gc/Transforms/Pipeline.cpp
Outdated
| pm.addPass(bufferization::createDropEquivalentBufferResultsPass()); | ||
| pm.addNestedPass<func::FuncOp>( | ||
| bufferization::createPromoteBuffersToStackPass()); | ||
| mlir::bufferization::BufferDeallocationPipelineOptions deallocOption; |
There was a problem hiding this comment.
nit: no need to add mlir namespace here.
BTW, could you also put a note/link to the upstream bufferization pipeline?
There was a problem hiding this comment.
Thanks for the advice! the mlir namespace has been removed and the note has been added
Fix part of issue #288
empty_tensor_eliminationcould eliminate the copy introduced bybufferization.materialize_in_destination. And adjust the pipeline according to the doc(https://mlir.llvm.org/docs/Bufferization/#overview).