[BACKEND] Cleanup IR after backward remat in RemoveLayoutConversions (triton-lang#5659)

Backward remat creates a lot of dummy convert ops that should be removed
before running the hosting phase otherwise we spend a lot of time
looking at dummy convert.

This doesn't change the result of the pass but it helps reduce compile
time and prevent combinatory explosion.

Author: ThomasRaoux

lib/Dialect/TritonGPU/Transforms/RemoveLayoutConversions.cpp

@@ -1173,6 +1173,22 @@ class TritonGPURemoveLayoutConversionsPass
     : public impl::TritonGPURemoveLayoutConversionsBase<
           TritonGPURemoveLayoutConversionsPass> {
 public:
+  // Cleanup convert ops.
+  void cleanupConvertOps() {
+    MLIRContext *context = &getContext();
+    ModuleOp m = getOperation();
+    RewritePatternSet cleanUpPatterns(context);
+    ConvertLayoutOp::getCanonicalizationPatterns(cleanUpPatterns, context);
+    if (applyPatternsAndFoldGreedily(m, std::move(cleanUpPatterns)).failed()) {
+      signalPassFailure();
+    }
+
+    LLVM_DEBUG({
+      DBGS() << "Module after canonicalizing:\n";
+      m.dump();
+    });
+  }
+
   void runOnOperation() override {
     MLIRContext *context = &getContext();
     ModuleOp m = getOperation();
@@ -1191,16 +1207,7 @@ class TritonGPURemoveLayoutConversionsPass
       m.dump();
     });
 
-    RewritePatternSet cleanUpPatterns(context);
-    ConvertLayoutOp::getCanonicalizationPatterns(cleanUpPatterns, context);
-    if (applyPatternsAndFoldGreedily(m, std::move(cleanUpPatterns)).failed()) {
-      signalPassFailure();
-    }
-
-    LLVM_DEBUG({
-      DBGS() << "Module after canonicalizing:\n";
-      m.dump();
-    });
+    cleanupConvertOps();
 
     // 2. For remaining convert ops, try to rematerialize the slice of producer
     // operation to avoid having to convert.
@@ -1210,6 +1217,9 @@ class TritonGPURemoveLayoutConversionsPass
       m.dump();
     });
 
+    // Cleanup dummy converts created during backward remat.
+    cleanupConvertOps();
+
     // 3. For remaining converts, try to hoist them above cast generating larger
     // size types in order to reduce the cost of the convert op.
     hoistConvert(m);