[lsr][term-fold] Restrict expansion budget, particularly for profiled loops

This is a follow up to an item I noted in my submission comment for
e947f95.  I don't have a real world example where this is triggering
unprofitably, but avoiding the transform when we estimate the loop
to be short running from profiling seems quite reasonable.

A couple of observations:
* Because we already checked for affine addrecs, the worst expansion we're
  going to see is the unknown trip count * unknown step case (both are
  loop invariant).  I picked a fallback cost (2 * SCEV's default expansion
  budget of 4) which was high enough to allow this.  These cases will
  end up being somewhat rare anyways as finding one where we can prove
  legality takes some work.
* Short running loops with constant trip counts produce either constant
  expressions (if the step is constant) or multiplications by small
  constant (if the step is unknown).  Both of these are cheap by any
  reasonable budget and thus not wroth checking for explicitly.
* Estimated trip counts come from profile data, loops without profile
  data will not have one.

Overall, this doesn't end up making much of a difference in practical
behavior of the transform.  Probably still worth landing, but I expect
this to be pretty low impact.

Possible future work (which I have no plans of doing at this time):
* Ppick the lowest cost alternate IV if we have multiple ones under the
  budget.
* Use SCEV to prove a constant upper bound on the trip count, and restrict
  the transform for unknown-but-small trip counts.
* I think we can now remove the affine restriction if we wanted.  Would
  need to think carefully about legality on this one.

I don't have examples where any of these would be profitable in practice,
but we'll see what comes up in the future.

Author: preames

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

@@ -6762,7 +6762,7 @@ static llvm::PHINode *GetInductionVariable(const Loop &L, ScalarEvolution &SE,
 
 static std::optional<std::tuple<PHINode *, PHINode *, const SCEV *, bool>>
 canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
-                      const LoopInfo &LI) {
+                      const LoopInfo &LI, const TargetTransformInfo &TTI) {
   if (!L->isInnermost()) {
     LLVM_DEBUG(dbgs() << "Cannot fold on non-innermost loop\n");
     return std::nullopt;
@@ -6813,13 +6813,23 @@ canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
   if (!isAlmostDeadIV(ToFold, LoopLatch, TermCond))
     return std::nullopt;
 
+  // Inserting instructions in the preheader has a runtime cost, scale
+  // the allowed cost with the loops trip count as best we can.
+  const unsigned ExpansionBudget = [&]() {
+    if (std::optional<unsigned> SmallTC = getLoopEstimatedTripCount(L))
+      return std::min(2*SCEVCheapExpansionBudget, *SmallTC);
+    // Unknown trip count, assume long running by default.
+    return 2*SCEVCheapExpansionBudget;
+  }();
+
   const SCEV *BECount = SE.getBackedgeTakenCount(L);
   const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
   SCEVExpander Expander(SE, DL, "lsr_fold_term_cond");
 
   PHINode *ToHelpFold = nullptr;
   const SCEV *TermValueS = nullptr;
   bool MustDropPoison = false;
+  auto InsertPt = L->getLoopPreheader()->getTerminator();
   for (PHINode &PN : L->getHeader()->phis()) {
     if (ToFold == &PN)
       continue;
@@ -6861,6 +6871,14 @@ canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
       continue;
     }
 
+    if (Expander.isHighCostExpansion(TermValueSLocal, L, ExpansionBudget,
+                                     &TTI, InsertPt)) {
+      LLVM_DEBUG(
+          dbgs() << "Is too expensive to expand terminating value for phi node"
+                 << PN << "\n");
+      continue;
+    }
+
     // The candidate IV may have been otherwise dead and poison from the
     // very first iteration.  If we can't disprove that, we can't use the IV.
     if (!mustExecuteUBIfPoisonOnPathTo(&PN, LoopLatch->getTerminator(), &DT)) {
@@ -6986,7 +7004,7 @@ static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
   }();
 
   if (EnableFormTerm) {
-    if (auto Opt = canFoldTermCondOfLoop(L, SE, DT, LI)) {
+    if (auto Opt = canFoldTermCondOfLoop(L, SE, DT, LI, TTI)) {
       auto [ToFold, ToHelpFold, TermValueS, MustDrop] = *Opt;
 
       Changed = true;