[llvm] [NVPTX] Do not mark move of global address as cheap enabling more CSE (PR #153730)

[Alex MacLean via llvm-commits]

================
@@ -1510,18 +1510,22 @@ let hasSideEffects = false in {
                                     "mov.b64 \t$d, __local_depot$num;">;
 }
 
-
-// copyPhysreg is hard-coded in NVPTXInstrInfo.cpp
-let hasSideEffects = false, isAsCheapAsAMove = true in {
-  let isMoveReg = true in
+let hasSideEffects = false in {
+  let isMoveReg = true, isAsCheapAsAMove = true in
     class MOVr<RegisterClass RC, string OpStr> :
       BasicNVPTXInst<(outs RC:$dst), (ins RC:$src), "mov." # OpStr>;
 
-  let isMoveImm = true in
+  let isMoveImm = true, isAsCheapAsAMove = true in
     class MOVi<RegTyInfo t, string suffix> :
       BasicNVPTXInst<(outs t.RC:$dst), (ins t.Imm:$src),
               "mov." # suffix,
               [(set t.Ty:$dst, t.ImmNode:$src)]>;
+
+  // We don't want to set isAsCheapAsAMove to true for these instructions as
+  // this would prevent CSE and resulted in regressions.
----------------
AlexMaclean wrote:

Added!

https://github.com/llvm/llvm-project/pull/153730