[mlir][tensor] Fix tensor.pad to remove newly static values (llvm#79938)

The canonicalization incrementally converts foldable dynamic hi/lo
padding to static hi/lo values. During this canonicalization the
static-fied valued should be removed from the dynamic values.

Author: rsuderman

mlir/lib/Dialect/Tensor/IR/TensorOps.cpp

@@ -3158,19 +3158,23 @@ struct FoldStaticPadding : public OpRewritePattern<PadOp> {
 
     // Extract the static info from the high and low operands.
     SmallVector<int64_t> constOperandsLow;
+    SmallVector<Value> newLows;
     for (auto operand : padTensorOp.getLow()) {
       APSInt intOp;
       if (!matchPattern(operand, m_ConstantInt(&intOp))) {
         constOperandsLow.push_back(ShapedType::kDynamic);
+        newLows.push_back(operand);
         continue;
       }
       constOperandsLow.push_back(intOp.getExtValue());
     }
     SmallVector<int64_t> constOperandsHigh;
+    SmallVector<Value> newHighs;
     for (auto operand : padTensorOp.getHigh()) {
       APSInt intOp;
       if (!matchPattern(operand, m_ConstantInt(&intOp))) {
         constOperandsHigh.push_back(ShapedType::kDynamic);
+        newHighs.push_back(operand);
         continue;
       }
       constOperandsHigh.push_back(intOp.getExtValue());
@@ -3222,7 +3226,7 @@ struct FoldStaticPadding : public OpRewritePattern<PadOp> {
         newOutDims, padTensorOp.getType().getElementType());
     auto newOp = rewriter.create<PadOp>(
         padTensorOp->getLoc(), newResultType, input, staticLow, staticHigh,
-        padTensorOp.getLow(), padTensorOp.getHigh(), padTensorOp.getNofold(),
+        newLows, newHighs, padTensorOp.getNofold(),
         getPrunedAttributeList(padTensorOp, PadOp::getAttributeNames()));
 
     IRMapping mapper;

mlir/test/Dialect/Tensor/canonicalize.mlir

@@ -1361,7 +1361,7 @@ func.func @pad_same_static_shape(%arg0: tensor<5x6xf32>, %a: index)
 // CHECK-LABEL:   func @pad_fold_static(
 // CHECK-SAME:      %[[INPUT:.*]]: tensor<?x64x?x?xf32>) -> tensor<?x?x?x?xf32> {
 // CHECK:           %[[CST:.*]] = arith.constant 0.000000e+00 : f32
-// CHECK:           %[[PADDING:.*]] = arith.constant 4 : index
+// CHECK-NOT:       arith.constant 4 : index
 // CHECK:           %[[PADDED:.*]] = tensor.pad %[[INPUT]]
 // CHECK-SAME:        low[0, 4, 1, 1] high[0, 4, 1, 1]  {
 // CHECK:           ^bb0(%[[ARG1:.*]]: index, %[[ARG2:.*]]: index, %[[ARG3:.*]]: index, %[[ARG4:.*]]: index):