[llvm] [BOLT] Improve InsertNegateRAStatePass::inferUnknownStates (PR #163381)

[Gergely Bálint via llvm-commits]

================
@@ -200,16 +200,29 @@ This pass runs after optimizations. It performns the _inverse_ of MarkRAState pa
 Some BOLT passes can add new Instructions. In InsertNegateRAStatePass, we have
 to know what RA state these have.
 
-The current solution has the `inferUnknownStates` function to cover these, using
-a fairly simple strategy: unknown states inherit the last known state.
-
-This will be updated to a more robust solution.
-
 > [!important]
 > As issue #160989 describes, unwind info is incorrect in stubs with multiple callers.
 > For this same reason, we cannot generate correct pac-specific unwind info: the signess
 > of the _incorrect_ return address is meaningless.
 
+Assignment of RAStates to newly generated instructions is done in `inferUnknownStates`.
+We have three different cases to cover:
+
+1. If a BasicBlock has some instructions with known RA state, and some without, we
+   can copy the RAState of known instructions to the unknown ones. As the control
+   flow only changes between BasicBlocks, instructions in the same BasicBlock have the
+   same return address.
+
+2. If all instructions in a BasicBlock are unknown, we can look at all CFG neighbors
+   (that is predecessors/successors). The RAState should be the same as of the
+   neighboring blocks. Conflicting RAStates in neighbors indicate an error. Such
+   functions should be ignored.
+
+3. If a BasicBlock has no CFG neighbors, we have to copy the RAState of the previous
+    BasicBlock in layout order.
+
----------------
bgergely0 wrote:

exactly.

https://github.com/llvm/llvm-project/pull/163381