[Mlir-commits] [llvm] [mlir] [MLIR] Add initial convert-memref-to-emitc pass (PR #85389)

[Matthias Gehre]

================
@@ -0,0 +1,163 @@
+//===- MemRefToEmitC.cpp - MemRef to EmitC conversion ---------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a pass to convert memref ops into emitc ops.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Conversion/MemRefToEmitC/MemRefToEmitC.h"
+
+#include "mlir/Dialect/EmitC/IR/EmitC.h"
+#include "mlir/Dialect/Func/IR/FuncOps.h"
+#include "mlir/Dialect/Func/Transforms/FuncConversions.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/IR/Builders.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/DialectConversion.h"
+
+namespace mlir {
+#define GEN_PASS_DEF_CONVERTMEMREFTOEMITC
+#include "mlir/Conversion/Passes.h.inc"
+} // namespace mlir
+
+using namespace mlir;
+
+namespace {
+
+/// Disallow all memrefs even though we only have conversions
+/// for memrefs with static shape right now to have good diagnostics.
+bool isLegal(Type t) { return !isa<BaseMemRefType>(t); }
+
+template <typename RangeT>
+bool areLegal(RangeT &&range) {
+  return llvm::all_of(range, [](Type type) { return isLegal(type); });
+}
+
+bool isLegal(Operation *op) {
+  return areLegal(op->getOperandTypes()) && areLegal(op->getResultTypes());
+}
+
+bool isSignatureLegal(FunctionType ty) {
+  return areLegal(ty.getInputs()) && areLegal(ty.getResults());
+}
+
+struct ConvertAlloca final : public OpConversionPattern<memref::AllocaOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(memref::AllocaOp op, OpAdaptor operands,
+                  ConversionPatternRewriter &rewriter) const override {
+
+    if (!op.getType().hasStaticShape()) {
+      return rewriter.notifyMatchFailure(
+          op.getLoc(), "cannot transform alloca with dynamic shape");
+    }
+
+    if (op.getAlignment().value_or(1) > 1) {
+      // TODO: Allow alignment if it is not more than the natural alignment
+      // of the C array.
+      return rewriter.notifyMatchFailure(
+          op.getLoc(), "cannot transform alloca with alignment requirement");
+    }
+
+    auto resultTy = getTypeConverter()->convertType(op.getType());
+    auto noInit = emitc::OpaqueAttr::get(getContext(), "");
+    rewriter.replaceOpWithNewOp<emitc::VariableOp>(op, resultTy, noInit);
+    return success();
+  }
+};
+
+struct ConvertLoad final : public OpConversionPattern<memref::LoadOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(memref::LoadOp op, OpAdaptor operands,
+                  ConversionPatternRewriter &rewriter) const override {
+
+    rewriter.replaceOpWithNewOp<emitc::SubscriptOp>(op, operands.getMemref(),
+                                                    operands.getIndices());
+    return success();
+  }
+};
+
+struct ConvertStore final : public OpConversionPattern<memref::StoreOp> {
+  using OpConversionPattern::OpConversionPattern;
+
+  LogicalResult
+  matchAndRewrite(memref::StoreOp op, OpAdaptor operands,
+                  ConversionPatternRewriter &rewriter) const override {
+
+    auto subscript = rewriter.create<emitc::SubscriptOp>(
+        op.getLoc(), operands.getMemref(), operands.getIndices());
+    rewriter.replaceOpWithNewOp<emitc::AssignOp>(op, subscript,
+                                                 operands.getValue());
+    return success();
+  }
+};
+
+struct ConvertMemRefToEmitCPass
+    : public impl::ConvertMemRefToEmitCBase<ConvertMemRefToEmitCPass> {
+  void runOnOperation() override {
+    TypeConverter converter;
+    // Fallback for other types.
+    converter.addConversion([](Type type) { return type; });
+    populateMemRefToEmitCTypeConversion(converter);
+    converter.addConversion(
+        [&converter](FunctionType ty) -> std::optional<Type> {
+          SmallVector<Type> inputs;
+          if (failed(converter.convertTypes(ty.getInputs(), inputs)))
+            return std::nullopt;
+
+          SmallVector<Type> results;
+          if (failed(converter.convertTypes(ty.getResults(), results)))
+            return std::nullopt;
+
+          return FunctionType::get(ty.getContext(), inputs, results);
+        });
+
+    RewritePatternSet patterns(&getContext());
+    populateMemRefToEmitCConversionPatterns(patterns, converter);
+    populateFunctionOpInterfaceTypeConversionPattern<func::FuncOp>(patterns,
+                                                                   converter);
+    populateCallOpTypeConversionPattern(patterns, converter);
+    populateReturnOpTypeConversionPattern(patterns, converter);
+
+    ConversionTarget target(getContext());
+    target.addDynamicallyLegalOp<func::FuncOp>(
+        [](func::FuncOp op) { return isSignatureLegal(op.getFunctionType()); });
+    target.addDynamicallyLegalDialect<func::FuncDialect>(
+        [](Operation *op) { return isLegal(op); });
+    target.addIllegalDialect<memref::MemRefDialect>();
+    target.addLegalDialect<emitc::EmitCDialect>();
+
+    if (failed(applyPartialConversion(getOperation(), target,
+                                      std::move(patterns))))
+      return signalPassFailure();
+  }
+};
+} // namespace
+
+void mlir::populateMemRefToEmitCTypeConversion(TypeConverter &typeConverter) {
+  typeConverter.addConversion([](MemRefType memRefType) -> std::optional<Type> {
+    if (memRefType.hasStaticShape()) {
----------------
mgehre-amd wrote:

I was thinking about this, too, but I didn't see an example where this mattered. If all accesses are through memref.load and memref.store, is the layout of the memref observable?
Do you have an example where this is a problem?

https://github.com/llvm/llvm-project/pull/85389