[llvm] [indvars] Missing variables at Og (PR #88270)

[Nikita Popov via llvm-commits]

================
@@ -1611,11 +1657,30 @@ int llvm::rewriteLoopExitValues(Loop *L, LoopInfo *LI, TargetLibraryInfo *TLI,
     // Replace PN with ExitVal if that is legal and does not break LCSSA.
     if (PN->getNumIncomingValues() == 1 &&
         LI->replacementPreservesLCSSAForm(PN, ExitVal)) {
+      addDebugValuesToLoopVariable(PN->getParent(), ExitVal, PN);
       PN->replaceAllUsesWith(ExitVal);
       PN->eraseFromParent();
     }
   }
 
+  // If the loop can be deleted and there are no PHIs to be rewritten (there
+  // are no loop live-out values), record debug variables corresponding to the
+  // induction variable with their constant exit-values. Those values will be
+  // inserted by the 'deletion loop' logic.
+  if (LoopCanBeDel && RewritePhiSet.empty()) {
+    if (auto *IndVar = L->getInductionVariable(*SE)) {
+      const SCEV *PNSCEV = SE->getSCEVAtScope(IndVar, L->getParentLoop());
+      if (auto *Const = dyn_cast<SCEVConstant>(PNSCEV)) {
+        Value *FinalIVValue = Const->getValue();
+        if (L->getUniqueExitBlock()) {
+          SmallVector<DbgVariableIntrinsic *> DbgUsers;
+          findDbgUsers(DbgUsers, IndVar);
+          L->preserveDebugInductionVariableInfo(FinalIVValue, DbgUsers);
----------------
nikic wrote:

TBH I still think the whole design here makes little sense. I think this case should either just be Won't Fix, or LoopDeletion needs to compute the exit value, rather than relying on some complex interaction with IndVars (which is not the only user of LoopDeletion!)

https://github.com/llvm/llvm-project/pull/88270