[llvm] [AMDGPU] Improve StructurizeCFG pass performance: avoid redundant DebugLoc map initialization. NFC. (PR #130568)

[Valery Pykhtin via llvm-commits]

================
@@ -990,10 +996,10 @@ BasicBlock *StructurizeCFG::getNextFlow(BasicBlock *Dominator) {
                                         Func, Insert);
   FlowSet.insert(Flow);
 
-  // use a temporary variable to avoid a use-after-free if the map's storage is
-  // reallocated
-  DebugLoc DL = TermDL[Dominator];
-  TermDL[Flow] = std::move(DL);
+  auto *Term = Dominator->getTerminator();
+  if (const DebugLoc &DL =
----------------
vpykhtin wrote:

Thinking about this more, expression:

Term ? Term->getDebugLoc() : TermDL.lookup(Dominator)

returns temporary DebugLoc object because this is the common type for LHS and RHS of the operator? because LHS returns reference and RHS - temporary object. This can work as is without problems, but I don't like that temporary DL would require redundant "retrack" of the internal MDNode pointer and I better avoid it by rewriting the statement.

https://github.com/llvm/llvm-project/pull/130568