[llvm] [NVPTX] Add some basic folds for ADDRSPACECAST (PR #129157)

[Artem Belevich via llvm-commits]

================
@@ -5209,6 +5209,26 @@ PerformBUILD_VECTORCombine(SDNode *N, TargetLowering::DAGCombinerInfo &DCI) {
   return DAG.getNode(ISD::BITCAST, DL, VT, PRMT);
 }
 
+static SDValue combineADDRSPACECAST(SDNode *N,
----------------
Artem-B wrote:

I'm not against transitive casts in principle. If we have a valid combinations, collapsing them is fine. It's just PTX does not give us that much wiggle room, as we can only cast to generic and back, and we can't use that for casts between different AS, so A->B->A will already handle them.

The only possibility seems to be casts between shared::cta and shared::cluster. PTX docs are not quite clear on that (or I didn't find the right place in the docs yet). If you can clarify whether casts between shared variants are legal, then we could restore transitive cast folding for them.


https://github.com/llvm/llvm-project/pull/129157