[Mlir-commits] [mlir] [mlir][ArmSME] Use liveness information in the tile allocator (PR #90448)
Benjamin Maxwell
llvmlistbot at llvm.org
Tue Apr 30 08:21:12 PDT 2024
================
@@ -116,4 +116,50 @@ VectorType getSMETileTypeForElement(Type elementType) {
return VectorType::get({minNumElts, minNumElts}, elementType, {true, true});
}
+void eraseTriviallyDeadTileOps(IRRewriter &rewriter,
+ FunctionOpInterface function) {
+ SmallVector<Operation *> worklist;
+ function->walk([&](Operation *op) {
+ auto armSMEOp = dyn_cast<arm_sme::ArmSMETileOpInterface>(op);
+ if (armSMEOp && isOpTriviallyDead(armSMEOp))
+ worklist.push_back(armSMEOp);
+ });
+ while (!worklist.empty()) {
+ Operation *op = worklist.pop_back_val();
+ if (!isOpTriviallyDead(op))
+ continue;
+ for (Value value : op->getOperands()) {
+ if (auto armSMEOp = value.getDefiningOp<arm_sme::ArmSMETileOpInterface>())
+ worklist.push_back(armSMEOp);
+ }
+ rewriter.eraseOp(op);
+ }
+}
+
+bool isTriviallyCloneableTileOp(arm_sme::ArmSMETileOpInterface tileOp) {
+ return tileOp && tileOp->getNumResults() == 1 &&
+ tileOp->getNumOperands() == 0 && isPure(tileOp);
+}
----------------
MacDue wrote:
It could have side-effects or inputs.
- If it has inputs you may need to consider those too (and clone them)
- If it has side-effects you don't know how moving the operation could change the semantics of the program
Both cases seem non-trivial to me :)
https://github.com/llvm/llvm-project/pull/90448
More information about the Mlir-commits
mailing list