[Mlir-commits] [mlir] [mlir][tensor] Fix PadOp::getConstantPaddingValue (PR #121205)

[Longsheng Mou]

https://github.com/CoTinker updated https://github.com/llvm/llvm-project/pull/121205

>From 2cccd2bda51be947d13c78e24bb41b72df9d3497 Mon Sep 17 00:00:00 2001
From: Longsheng Mou <longshengmou at gmail.com>
Date: Tue, 8 Apr 2025 19:27:50 +0800
Subject: [PATCH] [mlir][tensor] Fix `createFillOrGenerateOp`

This PR clones the padding value defined inside the PadOp block to outside.
---
 .../Dialect/Linalg/Transforms/Transforms.cpp  |  6 +++++-
 .../TensorToLinalg/tensor-ops-to-linalg.mlir  | 19 +++++++++++++++++++
 2 files changed, 24 insertions(+), 1 deletion(-)

diff --git a/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp b/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp
index dcd50cc44f81b..95605297afc69 100644
--- a/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp
+++ b/mlir/lib/Dialect/Linalg/Transforms/Transforms.cpp
@@ -927,8 +927,12 @@ Value DecomposePadOpPattern::createFillOrGenerateOp(
     RewriterBase &rewriter, tensor::PadOp padOp, Value dest,
     const SmallVector<Value> &dynSizes) const {
   auto padValue = padOp.getConstantPaddingValue();
-  if (padValue)
+  if (padValue) {
+    // Clone the padding value defined inside the PadOp block to outside.
+    if (padValue.getParentBlock() == &padOp.getRegion().front())
+      padValue = rewriter.clone(*padValue.getDefiningOp())->getResult(0);
     return rewriter.create<FillOp>(padOp.getLoc(), padValue, dest).result();
+  }
 
   // Fill could not be optimized: Lower to tensor::GenerateOp with region.
   auto generateOp = rewriter.create<tensor::GenerateOp>(
diff --git a/mlir/test/Conversion/TensorToLinalg/tensor-ops-to-linalg.mlir b/mlir/test/Conversion/TensorToLinalg/tensor-ops-to-linalg.mlir
index a0a676edceb74..b58d407733f23 100644
--- a/mlir/test/Conversion/TensorToLinalg/tensor-ops-to-linalg.mlir
+++ b/mlir/test/Conversion/TensorToLinalg/tensor-ops-to-linalg.mlir
@@ -44,3 +44,22 @@ func.func @generalize_pad_tensor_dynamic_shape(%arg0: tensor<4x?x2x?xf32>, %arg1
   } : tensor<4x?x2x?xf32> to tensor<4x?x?x?xf32>
   return %out : tensor<4x?x?x?xf32>
 }
+
+// -----
+
+// CHECK-LABEL:   func.func @generalize_pad_tensor_constant_inside(
+// CHECK-SAME:                                                     %[[SRC:.*]]: tensor<1x28x28x1xf32>) -> tensor<1x32x32x1xf32> {
+// CHECK:           %[[INIT:.*]] = tensor.empty() : tensor<1x32x32x1xf32>
+// CHECK:           %[[CST:.*]] = arith.constant 0.000000e+00 : f32
+// CHECK:           %[[FILL:.*]] = linalg.fill ins(%[[CST]] : f32) outs(%[[INIT]] : tensor<1x32x32x1xf32>) -> tensor<1x32x32x1xf32>
+// CHECK:           %[[PADDED:.*]] = tensor.insert_slice %[[SRC]] into %[[FILL]][0, 2, 2, 0] [1, 28, 28, 1] [1, 1, 1, 1] : tensor<1x28x28x1xf32> into tensor<1x32x32x1xf32>
+// CHECK:           return %[[PADDED]] : tensor<1x32x32x1xf32>
+// CHECK:         }
+func.func @generalize_pad_tensor_constant_inside(%arg0: tensor<1x28x28x1xf32>) -> tensor<1x32x32x1xf32> {
+  %0 = tensor.pad %arg0 low[0, 2, 2, 0] high[0, 2, 2, 0]  {
+  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
+    %cst = arith.constant 0.000000e+00 : f32
+    tensor.yield %cst : f32
+  } : tensor<1x28x28x1xf32> to tensor<1x32x32x1xf32>
+  return %0 : tensor<1x32x32x1xf32>
+}