[mlir][memref] Add a new ReifyResultShapes pass (#145927)

nicolasvasilache · fabianmcg · web-flow · commit 08cf6ae53785 · 2025-07-01T15:39:21.000+02:00
This pass reifies the shapes of a subset of `ReifyRankedShapedTypeOpInterface` ops with `tensor` results. The pass currently only supports result shape type reification for: - tensor::PadOp - tensor::ConcatOp It addresses a representation gap where implicit op semantics are needed to infer static result types from dynamic operands. But it does so by using `ReifyRankedShapedTypeOpInterface` as the source of truth rather than the op itself. As a consequence, this cannot generalize today. TODO: in the future, we should consider coupling this information with op "transfer functions" (e.g. `IndexingMapOpInterface`) to provide a source of truth that can work across result shape inference, canonicalization and op verifiers. The pass replaces the operations with their reified versions, when more static information can be derived, and inserts casts when results shapes are updated. Example: ```mlir #map = affine_map<(d0) -> (-d0 + 256)> func.func @func(%arg0: f32, %arg1: index, %arg2: tensor<64x?x64xf32>) -> tensor<1x?x64xf32> { %0 = affine.apply #map(%arg1) %extracted_slice = tensor.extract_slice %arg2[0, 0, 0] [1, %arg1, 64] [1, 1, 1] : tensor<64x?x64xf32> to tensor<1x?x64xf32> %padded = tensor.pad %extracted_slice low[0, 0, 0] high[0, %0, 0] { ^bb0(%arg3: index, %arg4: index, %arg5: index): tensor.yield %arg0 : f32 } : tensor<1x?x64xf32> to tensor<1x?x64xf32> return %padded : tensor<1x?x64xf32> } // mlir-opt --reify-result-shapes #map = affine_map<()[s0] -> (-s0 + 256)> func.func @func(%arg0: f32, %arg1: index, %arg2: tensor<64x?x64xf32>) -> tensor<1x?x64xf32> { %0 = affine.apply #map()[%arg1] %extracted_slice = tensor.extract_slice %arg2[0, 0, 0] [1, %arg1, 64] [1, 1, 1] : tensor<64x?x64xf32> to tensor<1x?x64xf32> %padded = tensor.pad %extracted_slice low[0, 0, 0] high[0, %0, 0] { ^bb0(%arg3: index, %arg4: index, %arg5: index): tensor.yield %arg0 : f32 } : tensor<1x?x64xf32> to tensor<1x256x64xf32> %cast = tensor.cast %padded : tensor<1x256x64xf32> to tensor<1x?x64xf32> return %cast : tensor<1x?x64xf32> } ``` --------- Co-authored-by: Fabian Mora <fabian.mora-cordero@amd.com>
diff --git a/mlir/include/mlir/Dialect/MemRef/Transforms/Passes.td b/mlir/include/mlir/Dialect/MemRef/Transforms/Passes.td
@@ -182,6 +182,68 @@ def ResolveShapedTypeResultDimsPass : Pass<"resolve-shaped-type-result-dims"> {
   ];
 }
 
+def ReifyResultShapesPass : Pass<"reify-result-shapes"> {
+  let summary ="Reifies the results of `tensor::PadOp` and `tensor::ConcatOp`.";
+  let description = [{
+    This pass reifies the shapes of a subset of `ReifyRankedShapedTypeOpInterface`
+    ops with `tensor` results.
+    
+    The pass currently only supports result shape type reification for:
+      - tensor::PadOp
+      - tensor::ConcatOp
+    It addresses a representation gap where implicit op semantics are needed to
+    infer static result types from dynamic operands.
+    But it does so by using `ReifyRankedShapedTypeOpInterface` as the source of
+    truth rather than the op itself. As a consequence, this cannot generalize
+    today.
+
+    TODO: in the future, we should consider coupling this information with op
+    "transfer functions" (e.g. `IndexingMapOpInterface`) to provide a source of
+    truth that can work across result shape inference, canonicalization and op
+    verifiers.
+
+    The pass replaces the operations with their reified versions, when more
+    static information can be derived, and inserts casts when results shapes
+    are updated.
+
+    Example:
+    ```mlir
+    #map = affine_map<(d0) -> (-d0 + 256)>
+    func.func @func(%arg0: f32, %arg1: index, %arg2: tensor<64x?x64xf32>)
+        -> tensor<1x?x64xf32>
+    {
+      %0 = affine.apply #map(%arg1)
+      %extracted_slice = tensor.extract_slice %arg2[0, 0, 0] [1, %arg1, 64] [1, 1, 1]
+        : tensor<64x?x64xf32> to tensor<1x?x64xf32>
+      %padded = tensor.pad %extracted_slice low[0, 0, 0] high[0, %0, 0] {
+      ^bb0(%arg3: index, %arg4: index, %arg5: index):
+        tensor.yield %arg0 : f32
+      } : tensor<1x?x64xf32> to tensor<1x?x64xf32>
+      return %padded : tensor<1x?x64xf32>
+    }
+
+    // mlir-opt --reify-result-shapes
+    #map = affine_map<()[s0] -> (-s0 + 256)>
+    func.func @func(%arg0: f32, %arg1: index, %arg2: tensor<64x?x64xf32>)
+        -> tensor<1x?x64xf32>
+    {
+      %0 = affine.apply #map()[%arg1]
+      %extracted_slice = tensor.extract_slice %arg2[0, 0, 0] [1, %arg1, 64] [1, 1, 1]
+        : tensor<64x?x64xf32> to tensor<1x?x64xf32>
+      %padded = tensor.pad %extracted_slice low[0, 0, 0] high[0, %0, 0] {
+      ^bb0(%arg3: index, %arg4: index, %arg5: index):
+        tensor.yield %arg0 : f32
+      } : tensor<1x?x64xf32> to tensor<1x256x64xf32>
+      %cast = tensor.cast %padded : tensor<1x256x64xf32> to tensor<1x?x64xf32>
+      return %cast : tensor<1x?x64xf32>
+    }
+    ```
+  }];
+  let dependentDialects = [
+    "affine::AffineDialect", "memref::MemRefDialect", "tensor::TensorDialect"
+  ];
+}
+
 def ExpandStridedMetadataPass : Pass<"expand-strided-metadata"> {
   let summary = "Expand memref operations into easier to analyze constructs";
   let description = [{
diff --git a/mlir/include/mlir/Dialect/MemRef/Transforms/Transforms.h b/mlir/include/mlir/Dialect/MemRef/Transforms/Transforms.h
@@ -23,6 +23,7 @@ class RewritePatternSet;
 class RewriterBase;
 class Value;
 class ValueRange;
+class ReifyRankedShapedTypeOpInterface;
 
 namespace arith {
 class WideIntEmulationConverter;
@@ -208,7 +209,6 @@ FailureOr<Value> replaceWithIndependentOp(RewriterBase &rewriter,
 memref::AllocaOp allocToAlloca(
     RewriterBase &rewriter, memref::AllocOp alloc,
     function_ref<bool(memref::AllocOp, memref::DeallocOp)> filter = nullptr);
-
 } // namespace memref
 } // namespace mlir
 
diff --git a/mlir/lib/Dialect/MemRef/Transforms/CMakeLists.txt b/mlir/lib/Dialect/MemRef/Transforms/CMakeLists.txt
@@ -13,6 +13,7 @@ add_mlir_dialect_library(MLIRMemRefTransforms
   IndependenceTransforms.cpp
   MultiBuffer.cpp
   NormalizeMemRefs.cpp
+  ReifyResultShapes.cpp
   ResolveShapedTypeResultDims.cpp
   RuntimeOpVerification.cpp
 
diff --git a/mlir/lib/Dialect/MemRef/Transforms/ReifyResultShapes.cpp b/mlir/lib/Dialect/MemRef/Transforms/ReifyResultShapes.cpp
@@ -0,0 +1,159 @@
+//===- ReifyResultShapes.cpp - Reify result shapes ------------------------===//
+//
+// 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 transform reifies result shapes of `ReifyRankedShapedTypeOpInterface`
+// operations with ranked `memref` and `tensor` results.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/MemRef/Transforms/Passes.h"
+
+#include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/MemRef/IR/MemRef.h"
+#include "mlir/Dialect/MemRef/Transforms/Transforms.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Interfaces/DestinationStyleOpInterface.h"
+#include "mlir/Interfaces/InferTypeOpInterface.h"
+#include "llvm/Support/InterleavedRange.h"
+
+#define DEBUG_TYPE "reify-result-shapes"
+#define DBGS() (llvm::dbgs() << "[" DEBUG_TYPE << "]: ")
+
+namespace mlir {
+namespace memref {
+#define GEN_PASS_DEF_REIFYRESULTSHAPESPASS
+#include "mlir/Dialect/MemRef/Transforms/Passes.h.inc"
+} // namespace memref
+} // namespace mlir
+
+using namespace mlir;
+
+/// Reifies the results of `op`, potentially replacing `op` with a reified
+/// version. Returns `failure` if `mlir::reifyResultShapes` returned failure,
+/// otherwise it always succeeds. Users of this transform should always expect
+/// it to modify the IR, even when it fails. If any of the result types changes,
+/// the transform will insert cast operations to the old type to keep the IR
+/// consistent.
+static LogicalResult reifyOpResultShapes(RewriterBase &rewriter,
+                                         ReifyRankedShapedTypeOpInterface op) {
+  LLVM_DEBUG({ DBGS() << " reifying op: " << op << "\n"; });
+  // Get the reified out shapes.
+  ReifiedRankedShapedTypeDims reifiedResultShapes;
+  if (failed(mlir::reifyResultShapes(rewriter, op, reifiedResultShapes)) ||
+      reifiedResultShapes.empty()) {
+    return op->emitWarning() << "failed to get the reified shapes";
+  }
+
+  bool modified = false;
+  // Compute the new output types.
+  SmallVector<Type> outTypes;
+  for (const auto &[oldTy, reifiedShape] :
+       llvm::zip(op->getResultTypes(), reifiedResultShapes)) {
+    // Skip if it's not a memref or tensor type.
+    if (!isa<RankedTensorType, MemRefType>(oldTy)) {
+      outTypes.push_back(oldTy);
+      continue;
+    }
+
+    ShapedType shapedTy = dyn_cast<ShapedType>(oldTy);
+
+    SmallVector<int64_t> shape = llvm::to_vector(shapedTy.getShape());
+    for (auto &&[dim, ofr] : llvm::zip_equal(shape, reifiedShape)) {
+      std::optional<int64_t> maybeCst = getConstantIntValue(ofr);
+      // If the reified dim is dynamic set it appropriately.
+      if (!maybeCst.has_value()) {
+        dim = ShapedType::kDynamic;
+        continue;
+      }
+      // Set the static dim.
+      dim = *maybeCst;
+    }
+
+    // If the shape didn't change continue.
+    if (shape == shapedTy.getShape()) {
+      outTypes.push_back(oldTy);
+      continue;
+    }
+    modified = true;
+    outTypes.push_back(shapedTy.cloneWith(shape, shapedTy.getElementType()));
+  }
+
+  // Return if we don't need to update.
+  if (!modified) {
+    LLVM_DEBUG({ DBGS() << "- op doesn't require update\n"; });
+    return success();
+  }
+
+  LLVM_DEBUG({
+    DBGS() << "- oldTypes: " << llvm::interleaved_array(op->getResultTypes())
+           << " \n";
+    DBGS() << "- outTypes: " << llvm::interleaved_array(outTypes) << " \n";
+  });
+
+  // We now have outTypes that need to be turned to cast ops.
+  Location loc = op->getLoc();
+  SmallVector<Value> newResults;
+  // TODO: `mlir::reifyResultShapes` and op verifiers may not agree atm.
+  // This is a confluence problem that will need to be addressed.
+  // For now, we know PadOp and ConcatOp are fine.
+  assert((isa<tensor::PadOp, tensor::ConcatOp>(op.getOperation())) &&
+         "incorrect op");
+  Operation *newOp = rewriter.clone(*op);
+  for (auto [reifiedTy, oldRes] : llvm::zip(outTypes, op->getResults())) {
+    OpResult newRes = newOp->getResult(oldRes.getResultNumber());
+    Type oldTy = oldRes.getType();
+    // Continue if the type remained invariant or is not shaped.
+    if (oldTy == reifiedTy || !isa<MemRefType, RankedTensorType>(oldTy)) {
+      newResults.push_back(newRes);
+      continue;
+    }
+
+    // Update the type.
+    newRes.setType(reifiedTy);
+    if (isa<RankedTensorType>(reifiedTy)) {
+      newResults.push_back(rewriter.create<tensor::CastOp>(loc, oldTy, newRes));
+    } else {
+      assert(isa<MemRefType>(reifiedTy) && "expected a memref type");
+      newResults.push_back(rewriter.create<memref::CastOp>(loc, oldTy, newRes));
+    }
+  }
+
+  LLVM_DEBUG({
+    DBGS() << "- reified results " << llvm::interleaved_array(newResults)
+           << "\n";
+  });
+  rewriter.replaceOp(op, newResults);
+  return success();
+}
+
+//===----------------------------------------------------------------------===//
+// Pass registration
+//===----------------------------------------------------------------------===//
+
+namespace {
+struct ReifyResultShapesPass final
+    : public memref::impl::ReifyResultShapesPassBase<ReifyResultShapesPass> {
+  void runOnOperation() override;
+};
+} // namespace
+
+void ReifyResultShapesPass::runOnOperation() {
+  SmallVector<ReifyRankedShapedTypeOpInterface> ops;
+  getOperation()->walk([&](ReifyRankedShapedTypeOpInterface op) {
+    // Handle ops that are not DPS and that do not carry an tied operand shapes.
+    // For now, limit to tensor::PadOp and tensor::ConcatOp.
+    if (!isa<tensor::PadOp, tensor::ConcatOp>(op.getOperation()))
+      return;
+    ops.push_back(op);
+  });
+  IRRewriter rewriter(&getContext());
+  for (ReifyRankedShapedTypeOpInterface op : ops) {
+    rewriter.setInsertionPoint(op);
+    (void)reifyOpResultShapes(rewriter, op);
+  }
+}
diff --git a/mlir/test/Dialect/Tensor/reify-shapes.mlir b/mlir/test/Dialect/Tensor/reify-shapes.mlir
@@ -0,0 +1,31 @@
+// RUN: mlir-opt -reify-result-shapes  %s | FileCheck %s
+
+// The test below checks concat op reification. In the first case, no cast is inserted while on the second a cast gets inserted.
+// CHECK-LABEL:  func.func @concat_reification
+func.func @concat_reification(%arg0: tensor<4x7x3xf32>, %arg1 : tensor<4x4x3xf32>, %arg2: tensor<?x?x?xf32>)
+  -> (tensor<4x11x3xf32>, tensor<?x?x?xf32>) {
+  // CHECK: %[[RES0:.*]] = tensor.concat dim(1) %{{.*}} : (tensor<4x7x3xf32>, tensor<4x4x3xf32>) -> tensor<4x11x3xf32>
+  %1 = tensor.concat dim(1) %arg0, %arg1 : (tensor<4x7x3xf32>, tensor<4x4x3xf32>) -> tensor<4x11x3xf32>
+  // CHECK: %[[V0:.*]] = tensor.concat dim(2) %{{.*}} : (tensor<4x7x3xf32>, tensor<?x?x?xf32>) -> tensor<4x7x?xf32>
+  // CHECK: %[[RES1:.*]] = tensor.cast %[[V0]] : tensor<4x7x?xf32> to tensor<?x?x?xf32>
+  %2 = tensor.concat dim(2) %arg0, %arg2 : (tensor<4x7x3xf32>, tensor<?x?x?xf32>) -> tensor<?x?x?xf32>
+  // CHECK: return %[[RES0]], %[[RES1]] : tensor<4x11x3xf32>, tensor<?x?x?xf32>
+  return %1, %2 : tensor<4x11x3xf32>, tensor<?x?x?xf32>
+}
+
+// CHECK-LABEL:  func.func @pad_reification
+func.func @pad_reification(%cst : f32, %idx : index, %t: tensor<64x?x64xf32>) -> tensor<1x?x64xf32> {
+  %pad_amt = affine.apply affine_map<(d0) -> (-d0 + 256)>(%idx)
+  %es = tensor.extract_slice %t[0, 0, 0] [1, %idx, 64] [1, 1, 1] 
+    : tensor<64x?x64xf32> to tensor<1x?x64xf32>
+
+  // CHECK: tensor.pad
+  // CHECK: : tensor<1x?x64xf32> to tensor<1x256x64xf32>
+  // CHECK: tensor.cast %{{.*}} : tensor<1x256x64xf32> to tensor<1x?x64xf32>
+  %padded = tensor.pad %es low[0, 0, 0] high[0, %pad_amt, 0] {
+    ^bb0(%a: index, %b: index, %c: index):
+    tensor.yield %cst : f32
+  } : tensor<1x?x64xf32> to tensor<1x?x64xf32>
+
+  return %padded : tensor<1x?x64xf32>
+}
