Address post review comment #4344 (#4348)

In the scope of `satisfies2DBlockReadAlignment`, it expects make tensor
pointer op to be found.

`satisfies2DBlockReadAlignment` is called at
https://github.com/intel/intel-xpu-backend-for-triton/blob/main/third_party/intel/lib/TritonIntelGPUTransforms/MaterializeBlockPointer.cpp#L74,
and check of make tensor pointer is done earlier at
https://github.com/intel/intel-xpu-backend-for-triton/blob/main/third_party/intel/lib/TritonIntelGPUTransforms/MaterializeBlockPointer.cpp#L60.

---------

Signed-off-by: Whitney Tsang <[email protected]>

Author: whitneywhtsang

third_party/intel/lib/Dialect/Triton/Transforms/TensorDescToBlockPointer.cpp

@@ -127,39 +127,41 @@ struct TritonIntelTensorDescToBlockPointer
   }
 
   void propagateToLoops(Operation *op) {
-    if (auto loopOp = dyn_cast<LoopLikeOpInterface>(op)) {
-      bool updated = false;
-      for (auto [initArg, rgnInitArg, yieldVal, loopRes] :
-           llvm::zip(loopOp.getInits(), loopOp.getRegionIterArgs(),
-                     loopOp.getYieldedValues(), loopOp->getResults())) {
-        Type initArgType = initArg.getType();
-        Type rgnInitArgType = rgnInitArg.getType();
-        assert(rgnInitArgType == loopRes.getType() &&
-               rgnInitArgType == yieldVal.getType() && "Type mismatch");
-        if (rgnInitArgType != initArgType) {
-          rgnInitArg.setType(initArgType);
-          yieldVal.setType(initArgType);
-          loopRes.setType(initArgType);
-          updated = true;
-        }
-      }
-      if (!updated)
-        return;
-
-      // For while loops we also need to update the "after" region arguments.
-      if (auto loopOp = dyn_cast<scf::WhileOp>(op)) {
-        for (auto [initArg, rgnAfterArg] :
-             llvm::zip(loopOp.getInits(), loopOp.getAfterArguments())) {
-          Type initArgType = initArg.getType();
-          if (rgnAfterArg.getType() != initArgType)
-            rgnAfterArg.setType(initArgType);
-        }
+    auto loopOp = dyn_cast<LoopLikeOpInterface>(op);
+    if (!loopOp)
+      return;
+
+    bool updated = false;
+    for (auto [initArg, rgnInitArg, yieldVal, loopRes] :
+         llvm::zip(loopOp.getInits(), loopOp.getRegionIterArgs(),
+                   loopOp.getYieldedValues(), loopOp->getResults())) {
+      Type initArgType = initArg.getType();
+      Type rgnInitArgType = rgnInitArg.getType();
+      assert(rgnInitArgType == loopRes.getType() &&
+             rgnInitArgType == yieldVal.getType() && "Type mismatch");
+      if (rgnInitArgType != initArgType) {
+        rgnInitArg.setType(initArgType);
+        yieldVal.setType(initArgType);
+        loopRes.setType(initArgType);
+        updated = true;
       }
+    }
+    if (!updated)
+      return;
 
-      // Propagate the loop results to their users.
-      for (Operation *user : loopOp->getUsers())
-        propagateToLoops(user);
+    // For while loops we also need to update the "after" region arguments.
+    if (auto loopOp = dyn_cast<scf::WhileOp>(op)) {
+      for (auto [initArg, rgnAfterArg] :
+           llvm::zip(loopOp.getInits(), loopOp.getAfterArguments())) {
+        Type initArgType = initArg.getType();
+        if (rgnAfterArg.getType() != initArgType)
+          rgnAfterArg.setType(initArgType);
+      }
     }
+
+    // Propagate the loop results to their users.
+    for (Operation *user : loopOp->getUsers())
+      propagateToLoops(user);
   }
 
   LogicalResult rewriteMakeTensorDescriptorOp(tt::MakeTensorDescOp op) {

third_party/intel/lib/TritonIntelGPUTransforms/MaterializeBlockPointer.cpp

@@ -295,9 +295,7 @@ struct TritonIntelGPUMaterializeBlockPointerPass
     // operation.
     std::optional<tt::MakeTensorPtrOp> defOp =
         tt::intel::findDefiningMakeTensorPtrOp(ptr);
-    if (!defOp)
-      return false;
-
+    assert(defOp && "Expected a make tensor ptr op.");
     tt::MakeTensorPtrOp makeTensorPtrOp = *defOp;
     Operation::operand_range shape = makeTensorPtrOp.getShape();
     if (shape.size() == 1)