[Mlir-commits] [mlir] [mlir][acc] Add data clause deduplication pattern for kernel env (PR #231965)

[Razvan Lupusoru]

================
@@ -78,6 +78,87 @@ struct RemoveEmptyKernelEnvironment
   }
 };
 
+struct KernelEnvironmentDedupDataOperands
+    : public OpRewritePattern<acc::KernelEnvironmentOp> {
+  using OpRewritePattern<acc::KernelEnvironmentOp>::OpRewritePattern;
+
+  static bool isReductionDataOperand(Value operand) {
+    Operation *def = operand.getDefiningOp();
+    if (!def)
+      return false;
+    if (std::optional<DataClause> clause = getDataClause(def))
+      return *clause == DataClause::acc_reduction;
+    return false;
+  }
+
+  // Later clauses for one variable receive the same address, so the first
+  // remains. acc.map_info is a finalized mapping: its order and semantics
+  // stand. The variable alone does not identify it, since the flags and
+  // descriptors would have to be compared as well.
+  static bool isSameMappedVariable(Value lhs, Value rhs) {
+    if (lhs == rhs)
+      return true;
+    Operation *lhsDef = lhs.getDefiningOp();
+    Operation *rhsDef = rhs.getDefiningOp();
+    if (!lhsDef || !rhsDef || isa<MapInfoOp>(lhsDef) || isa<MapInfoOp>(rhsDef))
+      return false;
+    Value lhsVar = getVar(lhsDef);
+    return lhsVar && lhsVar == getVar(rhsDef) &&
+           getBounds(lhsDef) == getBounds(rhsDef);
+  }
----------------
razvanlupusoru wrote:

We do not care actually because subsequent clauses only increment present counter and provide existing mapping. However, it is important we do not change the order at this time because a create followed by a copyin is not the same as a copyin followed by create (that said, we may consider in future to make them equivalent and there is an open discussion about clause ordering from OpenACC spec perspective).

https://github.com/llvm/llvm-project/pull/231965