Address PR comments

Author: dhernandez0

mlir/lib/Dialect/Rock/Transforms/BlockwiseLoadTileToThreadwise.cpp

@@ -270,8 +270,9 @@ class LoweringBlockwiseLoadTileOp final
       indicesNext[0] =
           arith::AddIOp::create(b, loc, indicesNext[0], one).getResult();
 
-      // it's ok if the indices are out of bounds because we use
-      // GLOBAL_PREFETCH_B8 with Speculative Prefetch
+      // it's acceptable if the indices are out of bounds because we use
+      // GLOBAL_PREFETCH_B8 with Speculative Prefetch. See llvm.prefetch
+      // documentation in AMDGPUUsage.rst
       rock::ThreadwisePrefetchOp::create(b, loc, wrappedSource,
                                          /*extraViews=*/b.getArrayAttr({}),
                                          /*extraIndices=*/indicesNext,

mlir/lib/Dialect/Rock/Transforms/SugarToLoops.cpp

@@ -1172,6 +1172,11 @@ struct GlobalPrefetchRewritePattern
     Value source = op.getSource();
 
     source = asGlobal(b, source);
+
+    // it's acceptable if the indices are out of bounds because we use
+    // GLOBAL_PREFETCH_B8 with Speculative Prefetch. See llvm.prefetch
+    // documentation in AMDGPUUsage.rst localityHint=3 is translated to memory
+    // scope SCOPE_SE.
     b.replaceOpWithNewOp<memref::PrefetchOp>(
         op, source, op.getSourceCoord(), /*isWrite=*/false, /*localityHint=*/3,
         /*isDataCache=*/true);