[mlir][vector] Fix bug in extractOp folding

[ThomasRaoux]

Summary

We were missing to check for transpose when folding.
Also add a new file to test folding independently of canonicalization as canonicalization was hiding the bug.

mlir/lib/Dialect/Vector/IR/VectorOps.cpp

ExtractFromInsertTransposeChainState::handleInsertOpWithMatchingPos(
     Value &res) {
   auto insertedPos = extractVector<int64_t>(nextInsertOp.getPosition());
   if (makeArrayRef(insertedPos) !=
       llvm::makeArrayRef(extractPosition).take_front(extractedRank))
     return failure();
   // Case 2.a. early-exit fold.
   res = nextInsertOp.getSource();
   // Case 2.b. if internal transposition is present, canFold will be false.
-  return success();
+  return success(canFold());
 }
 
 /// Case 3: if inserted position is a prefix of extractPosition,
 /// extract a portion of the source of the insertion.
 /// This method updates the internal state.
 LogicalResult
 ExtractFromInsertTransposeChainState::handleInsertOpWithPrefixPos(Value &res) {
   auto insertedPos = extractVector<int64_t>(nextInsertOp.getPosition());

mlir/test/Dialect/Vector/constant-fold.mlir

+// RUN: mlir-opt %s -split-input-file -test-constant-fold | FileCheck %s
+
+// CHECK-LABEL: fold_extract_transpose_negative
+func.func @fold_extract_transpose_negative(%arg0: vector<4x4xf16>) -> vector<4x4xf16> {
+  %cst = arith.constant dense<0.000000e+00> : vector<1x4x4xf16>
+  %0 = vector.insert %arg0, %cst [0] : vector<4x4xf16> into vector<1x4x4xf16>
+  // Verify that the transpose didn't get dropped.
+  // CHECK: %[[T:.+]] = vector.transpose
+  %1 = vector.transpose %0, [0, 2, 1] : vector<1x4x4xf16> to vector<1x4x4xf16>
+  // CHECK: vector.extract %[[T]][0]
+  %2 = vector.extract %1[0] : vector<1x4x4xf16>
+  return %2 : vector<4x4xf16>
+}