Normalize reinterpret_cast op

Normalize reinterpret_cast op for statically shaped input
and output memrefs.

Author: arnab-polymage

mlir/include/mlir/Dialect/Affine/Utils.h

@@ -32,6 +32,7 @@ class FuncOp;
 namespace memref {
 class AllocOp;
 class AllocaOp;
+class ReinterpretCastOp;
 } // namespace memref
 
 namespace affine {
@@ -243,16 +244,18 @@ LogicalResult replaceAllMemRefUsesWith(Value oldMemRef, Value newMemRef,
                                        ArrayRef<Value> symbolOperands = {},
                                        bool allowNonDereferencingOps = false);
 
-/// Rewrites the memref defined by this alloc op to have an identity layout map
-/// and updates all its indexing uses. Returns failure if any of its uses
-/// escape (while leaving the IR in a valid state).
+/// Rewrites the memref defined by alloc or reinterpret_cast op to have an
+/// identity layout map and updates all its indexing uses. Returns failure if
+/// any of its uses escape (while leaving the IR in a valid state).
 template <typename AllocLikeOp>
 LogicalResult normalizeMemRef(AllocLikeOp *op);
 extern template LogicalResult
 normalizeMemRef<memref::AllocaOp>(memref::AllocaOp *op);
 extern template LogicalResult
 normalizeMemRef<memref::AllocOp>(memref::AllocOp *op);
 
+LogicalResult normalizeMemRef(memref::ReinterpretCastOp *op);
+
 /// Normalizes `memrefType` so that the affine layout map of the memref is
 /// transformed to an identity map with a new shape being computed for the
 /// normalized memref type and returns it. The old memref type is simplify

mlir/lib/Dialect/Affine/Utils/Utils.cpp

@@ -1101,7 +1101,7 @@ void mlir::affine::affineScalarReplace(func::FuncOp f, DominanceInfo &domInfo,
 // Private helper function to transform memref.load with reduced rank.
 // This function will modify the indices of the memref.load to match the
 // newMemRef.
-LogicalResult transformMemRefLoadWithReducedRank(
+LogicalResult transformMemRefLoadOrStoreWithReducedRank(
     Operation *op, Value oldMemRef, Value newMemRef, unsigned memRefOperandPos,
     ArrayRef<Value> extraIndices, ArrayRef<Value> extraOperands,
     ArrayRef<Value> symbolOperands, AffineMap indexRemap) {
@@ -1182,6 +1182,14 @@ LogicalResult transformMemRefLoadWithReducedRank(
 
   return success();
 }
+
+// Checks if `op` is non dereferencing.
+// TODO: This hardcoded check will be removed once the right interface is added.
+static bool isDereferencingOp(Operation *op) {
+  return isa<AffineMapAccessInterface, memref::LoadOp, memref::StoreOp>(
+      op);
+}
+
 // Perform the replacement in `op`.
 LogicalResult mlir::affine::replaceAllMemRefUsesWith(
     Value oldMemRef, Value newMemRef, Operation *op,
@@ -1228,41 +1236,44 @@ LogicalResult mlir::affine::replaceAllMemRefUsesWith(
   // The following checks if op is dereferencing memref and performs the access
   // index rewrites.
   auto affMapAccInterface = dyn_cast<AffineMapAccessInterface>(op);
-  if (!affMapAccInterface) {
+  if (!isDereferencingOp(op)) {
     if (!allowNonDereferencingOps) {
       // Failure: memref used in a non-dereferencing context (potentially
       // escapes); no replacement in these cases unless allowNonDereferencingOps
       // is set.
       return failure();
     }
+    op->setOperand(memRefOperandPos, newMemRef);
+    return success();
+  }
 
-    // Check if it is a memref.load
-    auto memrefLoad = dyn_cast<memref::LoadOp>(op);
-    bool isReductionLike =
-        indexRemap.getNumResults() < indexRemap.getNumInputs();
-    if (!memrefLoad || !isReductionLike) {
-      op->setOperand(memRefOperandPos, newMemRef);
-      return success();
-    }
-
-    return transformMemRefLoadWithReducedRank(
-        op, oldMemRef, newMemRef, memRefOperandPos, extraIndices, extraOperands,
-        symbolOperands, indexRemap);
+  // Perform index rewrites for the dereferencing op and then replace the op.
+  SmallVector<Value, 4> oldMapOperands;
+  AffineMap oldMap;
+  unsigned oldMemRefNumIndices = oldMemRefRank;
+  if (affMapAccInterface) {
+    // If `op` implements AffineMapAccessInterface, we can get the indices by
+    // quering the number of map operands from the operand list from a certain
+    // offset (`memRefOperandPos` in this case).
+    NamedAttribute oldMapAttrPair =
+        affMapAccInterface.getAffineMapAttrForMemRef(oldMemRef);
+    oldMap = cast<AffineMapAttr>(oldMapAttrPair.getValue()).getValue();
+    oldMemRefNumIndices = oldMap.getNumInputs();
+    oldMapOperands.assign(op->operand_begin() + memRefOperandPos + 1,
+                          op->operand_begin() + memRefOperandPos + 1 +
+                              oldMemRefNumIndices);
+  } else {
+    oldMapOperands.assign(op->operand_begin() + memRefOperandPos + 1,
+                          op->operand_begin() + memRefOperandPos + 1 +
+                              oldMemRefRank);
   }
-  // Perform index rewrites for the dereferencing op and then replace the op
-  NamedAttribute oldMapAttrPair =
-      affMapAccInterface.getAffineMapAttrForMemRef(oldMemRef);
-  AffineMap oldMap = cast<AffineMapAttr>(oldMapAttrPair.getValue()).getValue();
-  unsigned oldMapNumInputs = oldMap.getNumInputs();
-  SmallVector<Value, 4> oldMapOperands(
-      op->operand_begin() + memRefOperandPos + 1,
-      op->operand_begin() + memRefOperandPos + 1 + oldMapNumInputs);
 
   // Apply 'oldMemRefOperands = oldMap(oldMapOperands)'.
   SmallVector<Value, 4> oldMemRefOperands;
   SmallVector<Value, 4> affineApplyOps;
   oldMemRefOperands.reserve(oldMemRefRank);
-  if (oldMap != builder.getMultiDimIdentityMap(oldMap.getNumDims())) {
+  if (affMapAccInterface &&
+      oldMap != builder.getMultiDimIdentityMap(oldMap.getNumDims())) {
     for (auto resultExpr : oldMap.getResults()) {
       auto singleResMap = AffineMap::get(oldMap.getNumDims(),
                                          oldMap.getNumSymbols(), resultExpr);
@@ -1287,7 +1298,6 @@ LogicalResult mlir::affine::replaceAllMemRefUsesWith(
 
   SmallVector<Value, 4> remapOutputs;
   remapOutputs.reserve(oldMemRefRank);
-
   if (indexRemap &&
       indexRemap != builder.getMultiDimIdentityMap(indexRemap.getNumDims())) {
     // Remapped indices.
@@ -1303,7 +1313,6 @@ LogicalResult mlir::affine::replaceAllMemRefUsesWith(
     // No remapping specified.
     remapOutputs.assign(remapOperands.begin(), remapOperands.end());
   }
-
   SmallVector<Value, 4> newMapOperands;
   newMapOperands.reserve(newMemRefRank);
 
@@ -1338,13 +1347,26 @@ LogicalResult mlir::affine::replaceAllMemRefUsesWith(
   state.operands.push_back(newMemRef);
 
   // Insert the new memref map operands.
-  state.operands.append(newMapOperands.begin(), newMapOperands.end());
+  if (affMapAccInterface) {
+    state.operands.append(newMapOperands.begin(), newMapOperands.end());
+  } else {
+    // In the case of dereferencing ops not implementing
+    // AffineMapAccessInterface, we need to apply the values of `newMapOperands`
+    // to the `newMap` to get the correct indices.
+    for (unsigned i = 0; i < newMemRefRank; i++)
+      state.operands.push_back(builder.create<AffineApplyOp>(
+          op->getLoc(),
+          AffineMap::get(newMap.getNumDims(), newMap.getNumSymbols(),
+                         newMap.getResult(i)),
+          newMapOperands));
+  }
 
   // Insert the remaining operands unmodified.
+  unsigned oldMapNumInputs = oldMapOperands.size();
+
   state.operands.append(op->operand_begin() + memRefOperandPos + 1 +
                             oldMapNumInputs,
                         op->operand_end());
-
   // Result types don't change. Both memref's are of the same elemental type.
   state.types.reserve(op->getNumResults());
   for (auto result : op->getResults())
@@ -1353,7 +1375,9 @@ LogicalResult mlir::affine::replaceAllMemRefUsesWith(
   // Add attribute for 'newMap', other Attributes do not change.
   auto newMapAttr = AffineMapAttr::get(newMap);
   for (auto namedAttr : op->getAttrs()) {
-    if (namedAttr.getName() == oldMapAttrPair.getName())
+    if (affMapAccInterface &&
+        namedAttr.getName() ==
+            affMapAccInterface.getAffineMapAttrForMemRef(oldMemRef).getName())
       state.attributes.push_back({namedAttr.getName(), newMapAttr});
     else
       state.attributes.push_back(namedAttr);
@@ -1846,6 +1870,93 @@ LogicalResult mlir::affine::normalizeMemRef(AllocLikeOp *allocOp) {
   return success();
 }
 
+LogicalResult
+mlir::affine::normalizeMemRef(memref::ReinterpretCastOp *reinterpretCastOp) {
+  MemRefType memrefType = reinterpretCastOp->getType();
+  AffineMap oldLayoutMap = memrefType.getLayout().getAffineMap();
+  Value oldMemRef = reinterpretCastOp->getResult();
+
+  // If `oldLayoutMap` is identity, `memrefType` is already normalized.
+  if (oldLayoutMap.isIdentity())
+    return success();
+
+  // Fetch a new memref type after normalizing the old memref to have an
+  // identity map layout.
+  MemRefType newMemRefType =
+      normalizeMemRefType(memrefType);
+  if (newMemRefType == memrefType)
+    // `oldLayoutMap` couldn't be transformed to an identity map.
+    return failure();
+
+  uint64_t newRank = newMemRefType.getRank();
+  SmallVector<Value> mapOperands(oldLayoutMap.getNumDims() +
+                                 oldLayoutMap.getNumSymbols());
+  SmallVector<Value> oldStrides = reinterpretCastOp->getStrides();
+  Location loc = reinterpretCastOp->getLoc();
+  // As `newMemRefType` is normalized, it is unit strided.
+  SmallVector<int64_t> newStaticStrides(newRank, 1);
+  ArrayRef<int64_t> oldShape = memrefType.getShape();
+  mlir::ValueRange oldSizes = reinterpretCastOp->getSizes();
+  unsigned idx = 0;
+  SmallVector<int64_t> newStaticSizes;
+  OpBuilder b(*reinterpretCastOp);
+  // Collectthe map operands which will be used to compute the new normalized
+  // memref shape.
+  for (unsigned i = 0, e = memrefType.getRank(); i < e; i++) {
+    if (oldShape[i] == ShapedType::kDynamic)
+      mapOperands[i] =
+          b.create<arith::SubIOp>(loc, oldSizes[0].getType(), oldSizes[idx++],
+                                  b.create<arith::ConstantIndexOp>(loc, 1));
+    else
+      mapOperands[i] = b.create<arith::ConstantIndexOp>(loc, oldShape[i] - 1);
+  }
+  for (unsigned i = 0, e = oldStrides.size(); i < e; i++)
+    mapOperands[memrefType.getRank() + i] = oldStrides[i];
+  SmallVector<Value> newSizes;
+  ArrayRef<int64_t> newShape = newMemRefType.getShape();
+  // Compute size along all the dimensions of the new normalized memref.
+  for (unsigned i = 0; i < newRank; i++) {
+    if (newShape[i] != ShapedType::kDynamic)
+      continue;
+    newSizes.push_back(b.create<AffineApplyOp>(
+        loc,
+        AffineMap::get(oldLayoutMap.getNumDims(), oldLayoutMap.getNumSymbols(),
+                       oldLayoutMap.getResult(i)),
+        mapOperands));
+  }
+  for (unsigned i = 0, e = newSizes.size(); i < e; i++)
+    newSizes[i] =
+        b.create<arith::AddIOp>(loc, newSizes[i].getType(), newSizes[i],
+                                b.create<arith::ConstantIndexOp>(loc, 1));
+  // Create the new reinterpret_cast op.
+  memref::ReinterpretCastOp newReinterpretCast =
+      b.create<memref::ReinterpretCastOp>(
+          loc, newMemRefType, reinterpretCastOp->getSource(),
+          reinterpretCastOp->getOffsets(), newSizes, mlir::ValueRange(),
+          /*static_offsets=*/reinterpretCastOp->getStaticOffsets(),
+          /*static_sizes=*/newShape,
+          /*static_strides=*/newStaticStrides);
+
+  // Replace all uses of the old memref.
+  if (failed(replaceAllMemRefUsesWith(oldMemRef,
+                                      /*newMemRef=*/newReinterpretCast,
+                                      /*extraIndices=*/{},
+                                      /*indexRemap=*/oldLayoutMap,
+                                      /*extraOperands=*/{},
+                                      /*symbolOperands=*/oldStrides,
+                                      /*domOpFilter=*/nullptr,
+                                      /*postDomOpFilter=*/nullptr,
+                                      /*allowNonDereferencingOps=*/true))) {
+    // If it failed (due to escapes for example), bail out.
+    newReinterpretCast->erase();
+    return failure();
+  }
+
+  oldMemRef.replaceAllUsesWith(newReinterpretCast);
+  reinterpretCastOp->erase();
+  return success();
+}
+
 template LogicalResult
 mlir::affine::normalizeMemRef<memref::AllocaOp>(memref::AllocaOp *op);
 template LogicalResult

mlir/lib/Dialect/MemRef/Transforms/NormalizeMemRefs.cpp

@@ -363,6 +363,15 @@ void NormalizeMemRefs::normalizeFuncOpMemRefs(func::FuncOp funcOp,
   for (memref::AllocaOp allocaOp : allocaOps)
     (void)normalizeMemRef(&allocaOp);
 
+  // Turn memrefs' non-identity layouts maps into ones with identity. Collect
+  // reinterpret_cast ops first and then process since normalizeMemRef
+  // replaces/erases ops during memref rewriting.
+  SmallVector<memref::ReinterpretCastOp> reinterpretCastOps;
+  funcOp.walk(
+      [&](memref::ReinterpretCastOp op) { reinterpretCastOps.push_back(op); });
+  for (memref::ReinterpretCastOp reinterpretCastOp : reinterpretCastOps)
+    (void)normalizeMemRef(&reinterpretCastOp);
+
   // We use this OpBuilder to create new memref layout later.
   OpBuilder b(funcOp);