Lowering Through LLVM #26
Conversation
j2kun
force-pushed
the
compile-to-llvm
branch
from
40d7c65 to
f2c74c6
Compare
j2kun
force-pushed
the
compile-to-llvm
branch
from
f2c74c6 to
a54f23f
Compare
j2kun
force-pushed
the
compile-to-llvm
branch
from
0df8dd9 to
8422826
Compare
j2kun
force-pushed
the
compile-to-llvm
branch
from
577dbd9 to
7660ef0
Compare
| @@ -46,6 +46,12 @@ void polyToLLVMPipelineBuilder(mlir::OpPassManager &manager) { | |||
| manager.addPass(mlir::createConvertFuncToLLVMPass()); | |||
| manager.addPass(mlir::createFinalizeMemRefToLLVMConversionPass()); | |||
| manager.addPass(mlir::createReconcileUnrealizedCastsPass()); | |||
|
|
|||
| // Cleanup | |||
| manager.addPass(mlir::createCanonicalizerPass()); | |||
There was a problem hiding this comment.
Is there a particular reason for adding mlir::createCanonicalizerPass() two times? There's one on the top which I think was for the passes we had created and there's this one which to my understanding should be dealing with the canonicalization of the other upstream passes we added to this pipeline.
I tried to remove the first one and just keep this last one here and the binary still works correctly (just that there is some difference in the instructions being generated in the end).
There was a problem hiding this comment.
In general canonicalization must not change the behavior of the program (or else it would be buggy). It will only make things more efficient. Canonicalization patterns apply to particular ops in particular dialects, and higher-level canonicalization patterns tend to remove more unnecessary work than applying lower-level canonicalization patterns, since they capture a larger breadth of the overall program. Running canonicalization after every pass would be even better. But the reason we don't is because we expect it to do nothing, and it adds to the total runtime.
|
I wanted to add for anyone having the same problem I had while following the tutorial on your website (brilliant job by the way) that in the current version of MLIR, you may have to add the Pass to the lowering pipeline in order to eliminate all of the bufferization operations. I'm not 100% clear as to why, but it is an issue that I had when lowering my tensors. Thank you for the amazing tutorial! |
No description provided.