[Linalg] Fix linalg.pack canonicalization priority issue (llvm#160340)

The current canonicalization prioritizes unpack->pack folder over
dropping padding if not needed but that folder fails if there is padding
and hence blocks all canonicalizations. We now put the failures in the
if statement so that we can proceed if the unpack->pack folder
conditions are not met.

Signed-off-by: Nirvedh Meshram <[email protected]>

Author: nirvedhmeshram

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

@@ -5583,14 +5583,13 @@ static bool inferStaticShape(PackOp packOp, SmallVectorImpl<int64_t> &srcShape,
 LogicalResult PackOp::canonicalize(PackOp packOp, PatternRewriter &rewriter) {
   // Fold an pack(unpack(x)) to x.
   if (auto unPackOp = packOp.getSource().getDefiningOp<UnPackOp>()) {
-    if (unPackOp.getSourceType() != packOp.getDestType())
-      return failure();
-    if (packOp.getPaddingValue() ||
-        !hasSameInnerOuterAttribute(packOp, unPackOp) ||
-        !haveSameTiles(packOp, unPackOp))
-      return failure();
-    rewriter.replaceOp(packOp, unPackOp.getSource());
-    return success();
+    if (unPackOp.getSourceType() == packOp.getDestType() &&
+        !packOp.getPaddingValue() &&
+        hasSameInnerOuterAttribute(packOp, unPackOp) &&
+        haveSameTiles(packOp, unPackOp)) {
+      rewriter.replaceOp(packOp, unPackOp.getSource());
+      return success();
+    }
   }
 
   // Fold optional PaddingValue operand away if padding is not needed.

mlir/test/Dialect/Linalg/canonicalize.mlir

@@ -1756,10 +1756,11 @@ func.func @pack_unpack(%t: tensor<16x16x?x?xf32>, %tile1: index, %tile2: index)
 // CHECK-SAME: %[[T:.+]]: tensor<16x16x8x8xf32>
 // CHECK: return %[[T]] : tensor<16x16x8x8xf32>
 func.func @pack_unpack(%t: tensor<16x16x8x8xf32>) -> tensor<16x16x8x8xf32> {
+  %cst = arith.constant 0.000000e+00 : f32
   %tensor_empty = tensor.empty() : tensor<128x128xf32>
   %unpacked = linalg.unpack %t inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %tensor_empty : tensor<16x16x8x8xf32> -> tensor<128x128xf32>
   %tensor_empty1 = tensor.empty() : tensor<16x16x8x8xf32>
-  %packed = linalg.pack %unpacked inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %tensor_empty1 : tensor<128x128xf32> -> tensor<16x16x8x8xf32>
+  %packed = linalg.pack %unpacked padding_value(%cst : f32) inner_dims_pos = [0, 1] inner_tiles = [8, 8] into %tensor_empty1 : tensor<128x128xf32> -> tensor<16x16x8x8xf32>
   return %packed : tensor<16x16x8x8xf32>
 }