[llvm] [AArch64] Prevent the AArch64LoadStoreOptimizer from reordering CFI instructions (PR #101317)

[Momchil Velikov via llvm-commits]

================
@@ -2243,6 +2266,20 @@ MachineBasicBlock::iterator AArch64LoadStoreOpt::findMatchingUpdateInsnBackward(
     if (!ModifiedRegUnits.available(BaseReg) ||
         !UsedRegUnits.available(BaseReg))
       return E;
+
+    // If we have a destination register (i.e. a load instruction) and a
+    // destination register is used or modified, then we can only merge forward,
+    // i.e. the combined instruction is put in the place of the memory
+    // instruction. Same applies if we see a memory access or side effects.
+    if (MI.mayLoadOrStore() || MI.hasUnmodeledSideEffects() ||
+        (DestReg[0] != AArch64::NoRegister &&
+         !(ModifiedRegUnits.available(DestReg[0]) &&
+           UsedRegUnits.available(DestReg[0]))) ||
+        (DestReg[1] != AArch64::NoRegister &&
+         !(ModifiedRegUnits.available(DestReg[1]) &&
+           UsedRegUnits.available(DestReg[1]))))
+      MergeEither = false;
+
----------------
momchil-velikov wrote:

Added tests, cover most of it.

https://github.com/llvm/llvm-project/pull/101317