[llvm] [BOLT] Gadget scanner: reformulate the state for data-flow analysis (PR #131898)

[Anatoly Trosinenko via llvm-commits]

================
@@ -169,16 +186,26 @@ struct State {
   std::vector<SmallPtrSet<const MCInst *, 4>> LastInstWritingReg;
   State() {}
   State(unsigned NumRegs, unsigned NumRegsToTrack)
-      : NonAutClobRegs(NumRegs), LastInstWritingReg(NumRegsToTrack) {}
-  State &operator|=(const State &StateIn) {
-    NonAutClobRegs |= StateIn.NonAutClobRegs;
+      : SafeToDerefRegs(NumRegs), LastInstWritingReg(NumRegsToTrack) {}
+
+  /// Returns S, so that S.merge(S1) == S1.merge(S) == S1.
+  static State getMergeNeutralElement(unsigned NumRegs,
----------------
atrosinenko wrote:

After thinking a bit more, finally dropped this function and switched to handling "empty" state explicitly. It is indeed rather hard to reason about when/whether the state of entry block (which is the only state known at the beginning on the analysis) is finally propagated into every other program point's state.

https://github.com/llvm/llvm-project/pull/131898