Skip values with existing conversions in getConvertBackwardSlice

`getConvertBackwardSlice` currently includes any existing
rematerialisation it finds in the returned slice. However, this
shouldn't be necessary because that value does not need to be
rematerialised or included in the cost calculation.

Instead, once we find a valid rematerialisation, we can stop the
traversal right away.

Author: neildhar

include/triton/Dialect/TritonGPU/Transforms/Utility.h

@@ -171,8 +171,16 @@ void appendToForOpYield(scf::ForOp forOp, ArrayRef<Value> newOperands);
 Operation *cloneWithInferType(mlir::OpBuilder &rewriter, Operation *op,
                               IRMapping &mapping);
 
-// Get backward slice of tensor values starting from the root node along with
-// encoding propagation.
+/// For a given \p root value with desired layout \p rootEncoding, get the
+/// backward slice of values that would have to be recreated to produce the
+/// value of \p root with that layout (without an intervening layout
+/// conversion). The traversal stops once we reach an operand that meets one of
+/// the following:
+///   1. has the desired layout
+///   2. \p getExistingConversion returns an existing converted value
+///   3. \p stopPropagation returns true for an op.
+/// The slice is returned in \p slice, and the desired layout of each value in
+/// the slice is stored in \p layouts.
 LogicalResult getConvertBackwardSlice(
     OpOperand &root, SetVector<Value> &slice, Attribute rootEncoding,
     DenseMap<Value, Attribute> &layout,

lib/Dialect/TritonGPU/Transforms/Utility.cpp

@@ -894,12 +894,13 @@ LogicalResult getConvertBackwardSlice(
     if (failed(updateLayout(currentValue, encoding)))
       return failure();
 
-    Value existing;
+    // If there is already an existing conversion to the target layout, we don't
+    // need to propagate to the operands.
+    // Note that this is per-use rather than per-value, so if another use fails
+    // the getExistingConversion check, we may still traverse the operands.
     if (getExistingConversion &&
-        (existing = getExistingConversion(*currentValueUse, encoding))) {
-      if (failed(updateLayout(existing, encoding)))
-        return failure();
-      currentValue = existing;
+        getExistingConversion(*currentValueUse, encoding)) {
+      continue;
     }
 
     if (auto ifOp = currentValue.getDefiningOp<scf::IfOp>()) {