WIP: [MLIR] Lower vector slices while lowering linalg to LLVM.

[aartbik]

mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp

580

nit: I would move this down with white space around it. Reason, all others populate for the apply method below, this one already applies

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

1233

nit: slightly more general name, we are packing more and more vector-vector lowerings into this set? (I have another one pending :-)

[aartbik]

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

1241

can you keep these comments too please in the moved code?

[poechsel]

Rebase, address comments

[poechsel]

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

1233

Actually I think putting these in a separate pass might be a good idea:

• lowerVectorSlicesToLLVM already looks like an independent pass (we are first collecting a bunch of patterns and then applying them greedily)
• Testing is also made harder as lowerVectorSlicesToLLVM also does dead code elimination.

mlir/include/mlir/Conversion/VectorToLLVM/ConvertVectorToLLVM.h

 
 /// Collect a set of patterns to convert from the Vector dialect to LLVM.
 void populateVectorToLLVMConversionPatterns(LLVMTypeConverter &converter,
                                             OwningRewritePatternList &patterns);
 
 /// Create a pass to convert vector operations to the LLVMIR dialect.
 std::unique_ptr<OperationPass<ModuleOp>> createConvertVectorToLLVMPass();
 
+// Perform progressive lowering of operations on "slices" and
+// all contraction operations. Also applies folding and DCE.
+void lowerVectorToVector(Operation *op, MLIRContext *context);
+
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_VECTORTOLLVM_CONVERTVECTORTOLLVM_H_

mlir/lib/Conversion/LinalgToLLVM/LinalgToLLVM.cpp

struct ConvertLinalgToLLVMPass
     : public ConvertLinalgToLLVMBase<ConvertLinalgToLLVMPass> {
   void runOnOperation() override;
 };
 } // namespace
 
 void ConvertLinalgToLLVMPass::runOnOperation() {
   auto module = getOperation();
 
+  lowerVectorToVector(module, &getContext());
   // Convert to the LLVM IR dialect using the converter defined above.
   OwningRewritePatternList patterns;
   LLVMTypeConverter converter(&getContext());
   populateAffineToStdConversionPatterns(patterns, &getContext());
   populateLoopToStdConversionPatterns(patterns, &getContext());
   populateStdToLLVMConversionPatterns(converter, patterns);
   populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
   populateVectorToLLVMConversionPatterns(converter, patterns);
   populateLinalgToStandardConversionPatterns(patterns, &getContext());
   populateLinalgToLLVMConversionPatterns(converter, patterns, &getContext());
 
   LLVMConversionTarget target(getContext());
   target.addDynamicallyLegalOp<FuncOp>(
       [&](FuncOp op) { return converter.isSignatureLegal(op.getType()); });
   target.addLegalOp<ModuleOp, ModuleTerminatorOp>();
   if (failed(applyFullConversion(module, target, patterns, &converter)))
     signalPassFailure();
 }
 
 std::unique_ptr<OperationPass<ModuleOp>> mlir::createConvertLinalgToLLVMPass() {
   return std::make_unique<ConvertLinalgToLLVMPass>();
 }

mlir/lib/Conversion/VectorToLLVM/ConvertVectorToLLVM.cpp

 }
 
 void mlir::populateVectorToLLVMMatrixConversionPatterns(
     LLVMTypeConverter &converter, OwningRewritePatternList &patterns) {
   MLIRContext *ctx = converter.getDialect()->getContext();
   patterns.insert<VectorMatmulOpConversion>(ctx, converter);
 }
 
+void mlir::lowerVectorToVector(Operation *op, MLIRContext *context) {
+  // Perform progressive lowering of operations on slices and
+  // all contraction operations. Also applies folding and DCE.
+  OwningRewritePatternList patterns;
+  populateVectorSlicesLoweringPatterns(patterns, context);
+  populateVectorContractLoweringPatterns(patterns, context);
+  applyPatternsAndFoldGreedily(op, patterns);
+}
+
 namespace {
 struct LowerVectorToLLVMPass
     : public ConvertVectorToLLVMBase<LowerVectorToLLVMPass> {
   void runOnOperation() override;
 };
 } // namespace
 
 void LowerVectorToLLVMPass::runOnOperation() {
-  // Perform progressive lowering of operations on slices and
-  // all contraction operations. Also applies folding and DCE.
-  {
-    OwningRewritePatternList patterns;
-    populateVectorSlicesLoweringPatterns(patterns, &getContext());
-    populateVectorContractLoweringPatterns(patterns, &getContext());
-    applyPatternsAndFoldGreedily(getOperation(), patterns);
-  }
+  lowerVectorToVector(getOperation(), &getContext());
 
   // Convert to the LLVM IR dialect.
   LLVMTypeConverter converter(&getContext());
   OwningRewritePatternList patterns;
   populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
   populateVectorToLLVMConversionPatterns(converter, patterns);
   populateVectorToLLVMMatrixConversionPatterns(converter, patterns);
   populateStdToLLVMConversionPatterns(converter, patterns);