[SCFToAffine] Add a pass to raise scf to affine ops.

mlir/include/mlir/Conversion/Passes.h

@@ -58,6 +58,7 @@
 #include "mlir/Conversion/OpenMPToLLVM/ConvertOpenMPToLLVM.h"
 #include "mlir/Conversion/PDLToPDLInterp/PDLToPDLInterp.h"
 #include "mlir/Conversion/ReconcileUnrealizedCasts/ReconcileUnrealizedCasts.h"
+#include "mlir/Conversion/SCFToAffine/SCFToAffine.h"
 #include "mlir/Conversion/SCFToControlFlow/SCFToControlFlow.h"
 #include "mlir/Conversion/SCFToEmitC/SCFToEmitC.h"
 #include "mlir/Conversion/SCFToGPU/SCFToGPUPass.h"

mlir/include/mlir/Conversion/Passes.td

@@ -1027,6 +1027,18 @@ def ReconcileUnrealizedCastsPass : Pass<"reconcile-unrealized-casts"> {
   }];
 }
 
+//===----------------------------------------------------------------------===//
+// SCFToAffine
+//===----------------------------------------------------------------------===//
+
+def RaiseSCFToAffinePass : Pass<"raise-scf-to-affine"> {
+  let summary = "Raise SCF to affine ops";
+  let dependentDialects = [
+    "affine::AffineDialect",
+    "scf::SCFDialect",
+  ];
+}
+
 //===----------------------------------------------------------------------===//
 // SCFToControlFlow
 //===----------------------------------------------------------------------===//

mlir/include/mlir/Conversion/SCFToAffine/SCFToAffine.h

@@ -0,0 +1,26 @@
+//===- SCFToAffine.h - SCF to Affine Pass entrypoint ------------*- C++ -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef MLIR_CONVERSION_SCFTOAFFINE_SCFTOAFFINE_H_
+#define MLIR_CONVERSION_SCFTOAFFINE_SCFTOAFFINE_H_
+
+#include <memory>
+
+namespace mlir {
+class Pass;
+class RewritePatternSet;
+
+#define GEN_PASS_DECL_RAISESCFTOAFFINEPASS
+#include "mlir/Conversion/Passes.h.inc"
+
+/// Collect a set of patterns to convert SCF operations to Affine operations.
+void populateSCFToAffineConversionPatterns(RewritePatternSet &patterns);
+
+} // namespace mlir
+
+#endif // MLIR_CONVERSION_SCFTOAFFINE_SCFTOAFFINE_H_

mlir/lib/Conversion/CMakeLists.txt

@@ -51,6 +51,7 @@ add_subdirectory(OpenACCToSCF)
 add_subdirectory(OpenMPToLLVM)
 add_subdirectory(PDLToPDLInterp)
 add_subdirectory(ReconcileUnrealizedCasts)
+add_subdirectory(SCFToAffine)
 add_subdirectory(SCFToControlFlow)
 add_subdirectory(SCFToEmitC)
 add_subdirectory(SCFToGPU)

mlir/lib/Conversion/SCFToAffine/CMakeLists.txt

@@ -0,0 +1,17 @@
+add_mlir_conversion_library(MLIRSCFToAffine
+  SCFToAffine.cpp
+
+  ADDITIONAL_HEADER_DIRS
+  ${MLIR_MAIN_INCLUDE_DIR}/mlir/Conversion/SCFToAffine
+
+  DEPENDS
+  MLIRConversionPassIncGen
+
+  LINK_LIBS PUBLIC
+  MLIRArithDialect
+  MLIRAffineDialect
+  MLIRLLVMDialect
+  MLIRSCFDialect
+  MLIRSCFTransforms
+  MLIRTransforms
+  )

mlir/lib/Conversion/SCFToAffine/SCFToAffine.cpp

@@ -0,0 +1,134 @@
+//===- SCFToAffine.cpp - SCF to Affine 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 raise scf.for, scf.if and loop.terminator
+// ops into affine ops.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Conversion/SCFToAffine/SCFToAffine.h"
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/SCF/IR/SCF.h"
+#include "mlir/IR/Verifier.h"
+#include "mlir/Transforms/Passes.h"
+#include "mlir/Transforms/WalkPatternRewriteDriver.h"
+#include "llvm/Support/DebugLog.h"
+
+namespace mlir {
+#define GEN_PASS_DEF_RAISESCFTOAFFINEPASS
+#include "mlir/Conversion/Passes.h.inc"
+} // namespace mlir
+
+#define DEBUG_TYPE "raise-scf-to-affine"
+
+using namespace mlir;
+
+namespace {
+
+struct SCFToAffinePass
+    : public impl::RaiseSCFToAffinePassBase<SCFToAffinePass> {
+  void runOnOperation() override;
+};
+
+bool canRaiseToAffine(scf::ForOp op) {
+  return affine::isValidDim(op.getLowerBound()) &&
+         affine::isValidDim(op.getUpperBound()) &&
+         affine::isValidSymbol(op.getStep());
+}
+
+struct ForOpRewrite : public OpRewritePattern<scf::ForOp> {
+  using OpRewritePattern<scf::ForOp>::OpRewritePattern;
+
+  std::pair<affine::AffineForOp, Value>
+  createAffineFor(scf::ForOp op, PatternRewriter &rewriter) const {
+    IntegerAttr constAttr;
+    if (matchPattern(op.getStep(), m_Constant(&constAttr))) {
+      int64_t step = constAttr.getInt();
+      if (step > 0)
+        return positiveConstantStep(op, step, rewriter);
+    }
+    return genericBounds(op, rewriter);
+  }
+
+  std::pair<affine::AffineForOp, Value>
+  positiveConstantStep(scf::ForOp op, int64_t step,
+                       PatternRewriter &rewriter) const {
+    auto affineFor = affine::AffineForOp::create(
+        rewriter, op.getLoc(), ValueRange(op.getLowerBound()),
+        AffineMap::get(1, 0, rewriter.getAffineDimExpr(0)),
+        ValueRange(op.getUpperBound()),
+        AffineMap::get(1, 0, rewriter.getAffineDimExpr(0)), step,
+        op.getInits());
+    return std::make_pair(affineFor, affineFor.getInductionVar());
+  }
+
+  std::pair<affine::AffineForOp, Value>
+  genericBounds(scf::ForOp op, PatternRewriter &rewriter) const {
+    Value lower = op.getLowerBound();
+    Value upper = op.getUpperBound();
+    Value step = op.getStep();
+    AffineExpr lowerExpr = rewriter.getAffineDimExpr(0);
+    AffineExpr upperExpr = rewriter.getAffineDimExpr(1);
+    AffineExpr stepExpr = rewriter.getAffineSymbolExpr(0);
+    auto affineFor = affine::AffineForOp::create(
+        rewriter, op.getLoc(), ValueRange(), rewriter.getConstantAffineMap(0),
+        ValueRange({lower, upper, step}),
+        AffineMap::get(
+            2, 1, (upperExpr - lowerExpr + stepExpr - 1).floorDiv(stepExpr)),
+        1, op.getInits());
+
+    rewriter.setInsertionPointToStart(affineFor.getBody());
+    auto actualIndexMap = AffineMap::get(
+        2, 1, lowerExpr + rewriter.getAffineDimExpr(1) * stepExpr);
+    auto actualIndex = affine::AffineApplyOp::create(
+        rewriter, op.getLoc(), actualIndexMap,
+        ValueRange({lower, affineFor.getInductionVar(), step}));
+    return std::make_pair(affineFor, actualIndex.getResult());
+  }
+
+  LogicalResult matchAndRewrite(scf::ForOp op,
+                                PatternRewriter &rewriter) const override {
+    if (!canRaiseToAffine(op)) {
+      LDBG() << "[affine] Cannot raise scf op: " << op << "\n";
+      return failure();
+    }
+
+    auto [affineFor, actualIndex] = createAffineFor(op, rewriter);
+    Block *affineBody = affineFor.getBody();
+
+    if (affineBody->mightHaveTerminator())
+      rewriter.eraseOp(affineBody->getTerminator());
+
+    SmallVector<Value> argValues;
+    argValues.push_back(actualIndex);
+    llvm::append_range(argValues, affineFor.getRegionIterArgs());
+    rewriter.inlineBlockBefore(op.getBody(), affineBody, affineBody->end(),
+                               argValues);
+
+    auto scfYieldOp = cast<scf::YieldOp>(affineBody->getTerminator());
+    rewriter.setInsertionPointToEnd(affineBody);
+    rewriter.replaceOpWithNewOp<affine::AffineYieldOp>(
+        scfYieldOp, scfYieldOp->getOperands());
+
+    rewriter.replaceOp(op, affineFor);
+    return success();
+  }
+};
+
+} // namespace
+
+void mlir::populateSCFToAffineConversionPatterns(RewritePatternSet &patterns) {
+  patterns.add<ForOpRewrite>(patterns.getContext());
+}
+
+void SCFToAffinePass::runOnOperation() {
+  MLIRContext &ctx = getContext();
+  RewritePatternSet patterns(&ctx);
+  populateSCFToAffineConversionPatterns(patterns);
+  walkAndApplyPatterns(getOperation(), std::move(patterns));
+}

mlir/test/Conversion/SCFToAffine/scf-to-affine.mlir

@@ -0,0 +1,51 @@
+// RUN: mlir-opt -raise-scf-to-affine -split-input-file %s | FileCheck %s
+
+// CHECK: #[[$UB_MAP:.+]] = affine_map<(d0, d1)[s0] -> ((d1 - d0 + s0 - 1) floordiv s0)>
+// CHECK: #[[$IV_MAP:.+]] = affine_map<(d0, d1)[s0] -> (d0 + d1 * s0)>
+// CHECK-LABEL: @generic_loop
+// CHECK-SAME: %[[ARR:.*]]: memref<?xi32>, %[[LOWER:.*]]: index, %[[UPPER:.*]]: index, %[[STEP:.*]]: index
+func.func @generic_loop(%arr: memref<?xi32>, %lower: index, %upper: index, %step: index) {
+// CHECK: affine.for %[[IV:.*]] = 0 to #[[$UB_MAP]](%[[LOWER]], %[[UPPER]])[%[[STEP]]] {
+// CHECK:   %[[IDX:.*]] = affine.apply #[[$IV_MAP]](%[[LOWER]], %[[IV]])[%[[STEP]]]
+// CHECK:   memref.store %{{.*}}, %[[ARR]][%[[IDX]]] : memref<?xi32>
+// CHECK: }
+  %c0_i32 = arith.constant 0 : i32
+  scf.for %idx = %lower to %upper step %step {
+    memref.store %c0_i32, %arr[%idx] : memref<?xi32>
+  }
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP:.+]] = affine_map<(d0) -> (d0)>
+// CHECK-LABEL: @loop_with_constant_step
+// CHECK-SAME: %[[ARR:.*]]: memref<?xi32>, %[[LOWER:.*]]: index, %[[UPPER:.*]]: index
+func.func @loop_with_constant_step(%arr: memref<?xi32>, %lower: index, %upper: index) {
+// CHECK: affine.for %[[IDX:.*]] = #[[$MAP]](%[[LOWER]]) to #[[$MAP]](%[[UPPER]]) step 3 {
+// CHECK:   memref.store %{{.*}}, %[[ARR]][%[[IDX]]] : memref<?xi32>
+// CHECK: }
+  %c0_i32 = arith.constant 0 : i32
+  %c3 = arith.constant 3 : index
+  scf.for %idx = %lower to %upper step %c3 {
+    memref.store %c0_i32, %arr[%idx] : memref<?xi32>
+  }
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @nested_loop
+func.func @nested_loop(%arg0: memref<?x?xi32>, %upper1: index, %upper2: index) {
+// CHECK: affine.for
+// CHECK:   affine.for
+  %c0_i32 = arith.constant 0 : i32
+  %c0 = arith.constant 0 : index
+  %c1 = arith.constant 1 : index
+  scf.for %i = %c0 to %upper1 step %c1 {
+    scf.for %j = %c0 to %upper2 step %c1 {
+      memref.store %c0_i32, %arg0[%i, %j] : memref<?x?xi32>
+    }
+  }
+  return
+}