[Mlir-commits] [mlir] [mlir][tensor] Make tensor::PadOp a ReifyRankedShapedTypeOpInterface (PR #145867)

[Nicolas Vasilache]

https://github.com/nicolasvasilache updated https://github.com/llvm/llvm-project/pull/145867

>From 07b70b68b174ed7fb80c9d030ff81c4335cb1304 Mon Sep 17 00:00:00 2001
From: Nicolas Vasilache <nico.vasilache at amd.com>
Date: Wed, 25 Jun 2025 16:46:44 +0200
Subject: [PATCH] [mlir][tensor] Make tensor::PadOp a
 ReifyRankedShapedTypeOpInterface

---
 .../mlir/Dialect/Tensor/IR/TensorOps.td       |  1 +
 mlir/lib/Dialect/Tensor/IR/TensorOps.cpp      | 25 +++++++++++++++++++
 .../value-bounds-op-interface-impl.mlir       | 17 +++++++++++++
 3 files changed, 43 insertions(+)

diff --git a/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td b/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td
index 35d0b16628417..47962f75558ea 100644
--- a/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td
+++ b/mlir/include/mlir/Dialect/Tensor/IR/TensorOps.td
@@ -1256,6 +1256,7 @@ def Tensor_CollapseShapeOp : Tensor_ReassociativeReshapeOp<"collapse_shape"> {
 
 def Tensor_PadOp : Tensor_Op<"pad", [
     DeclareOpInterfaceMethods<OpAsmOpInterface, ["getAsmResultNames"]>,
+    DeclareOpInterfaceMethods<ReifyRankedShapedTypeOpInterface>,
     AttrSizedOperandSegments,
     Pure,
     SingleBlockImplicitTerminator<"mlir::tensor::YieldOp">]> {
diff --git a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
index 72144ec71c5d2..20f48436b1901 100644
--- a/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
+++ b/mlir/lib/Dialect/Tensor/IR/TensorOps.cpp
@@ -7,6 +7,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "mlir/Dialect/Affine/IR/AffineOps.h"
+#include "mlir/Dialect/Affine/Utils.h"
 #include "mlir/Dialect/Arith/IR/Arith.h"
 #include "mlir/Dialect/Arith/Utils/Utils.h"
 #include "mlir/Dialect/Complex/IR/Complex.h"
@@ -3793,6 +3794,30 @@ struct FoldConsecutiveConstantPadding : public OpRewritePattern<tensor::PadOp> {
 
 } // namespace
 
+LogicalResult
+PadOp::reifyResultShapes(OpBuilder &b,
+                         ReifiedRankedShapedTypeDims &reifiedReturnShapes) {
+  reifiedReturnShapes.resize(1, SmallVector<OpFoldResult>(getType().getRank()));
+  SmallVector<OpFoldResult> lp = getMixedLowPad();
+  SmallVector<OpFoldResult> hp = getMixedHighPad();
+  for (int64_t i = 0; i < getResultType().getRank(); ++i) {
+    if (!getType().isDynamicDim(i)) {
+      reifiedReturnShapes[0][i] = b.getIndexAttr(getType().getDimSize(i));
+      continue;
+    }
+    Location loc = getLoc();
+    Value dim = b.createOrFold<tensor::DimOp>(
+        loc, getSource(), b.create<arith::ConstantIndexOp>(loc, i));
+
+    affine::AffineBuilder ab(b, loc);
+    AffineExpr d0, d1, d2;
+    bindDims(b.getContext(), d0, d1, d2);
+    reifiedReturnShapes[0][i] = affine::makeComposedFoldedAffineApply(
+        b, loc, {d0 + d1 + d2}, {dim, lp[i], hp[i]});
+  }
+  return success();
+}
+
 void PadOp::getCanonicalizationPatterns(RewritePatternSet &results,
                                         MLIRContext *context) {
   results.add<FoldStaticZeroPadding, FoldSourceTensorCast, FoldTargetTensorCast,
diff --git a/mlir/test/Dialect/Tensor/value-bounds-op-interface-impl.mlir b/mlir/test/Dialect/Tensor/value-bounds-op-interface-impl.mlir
index 6610d3180cf02..e526adb18cf1e 100644
--- a/mlir/test/Dialect/Tensor/value-bounds-op-interface-impl.mlir
+++ b/mlir/test/Dialect/Tensor/value-bounds-op-interface-impl.mlir
@@ -213,3 +213,20 @@ func.func @dynamic_dims_are_maybe_equal_2(%t: tensor<?x?xf32>) {
   "test.compare"(%dim0, %dim1) : (index, index) -> ()
   return
 }
+
+// -----
+
+// CHECK-LABEL:  func.func @pad_reification
+func.func @pad_reification(%cst : f32, %idx : index, %t: tensor<64x?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>
+
+  %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>
+
+  // CHECK: arith.constant 256 : index
+  %1 = "test.reify_bound"(%padded) {dim = 1, constant} : (tensor<1x?x64xf32>) -> (index)
+  return
+}