[LSR] Fold terminating condition to other IV when possible

[eopXD]

Summary

When the IV is only used by the terminating condition (say IV-A) and the loop
has a predictable back-edge count and we have another IV (say IV-B) that is an
affine add recursion, we will be able to calculate the terminating value of
IV-B in the loop pre-header. This patch adds attempts to replace IV-B as the
new terminating condition and remove IV-A. It is safe to do so since IV-A is
only used as the terminating condition.

This transformation is suitable to be appended after LSR as it may optimize the
loop into the situation mentioned above. The transformation can reduce number of
IV-s in the loop by one.

A cli option lsr-term-fold is added and default disabled.

[fhahn]

llvm/test/Transforms/LoopStrengthReduce/lsr-fold-iv-complicate-add-rec.ll

46 ↗

(On Diff #454725)

is this needed?

50 ↗

(On Diff #454725)

is this needed? Would be good to try to simplify the test as much as possible. Also, it would be good to add a patch just for the tests and then only include the diff here.

[eopXD]

llvm/test/Transforms/LoopStrengthReduce/lsr-fold-iv-complicate-add-rec.ll

46 ↗

(On Diff #454725)

Thanks for dropping by. Yes, my code checks if the phi node has a value that comes from the loop pre-header. Without this the condition check won't pass.

50 ↗

(On Diff #454725)

Thanks for the reminder. Let me add a pre-commit test for this.

[fhahn]

llvm/test/Transforms/LoopStrengthReduce/lsr-fold-iv-complicate-add-rec.ll

46 ↗

(On Diff #454725)

right, but it doesn't need to be defined in the preheader, this could just be %mark I assume?

50 ↗

(On Diff #454725)

It would be good to clean them up before committing the test. Also, it looks like it is missing negative tests.

59 ↗

(On Diff #454725)

It looks like this function doesn't need to return a pointer?

[eopXD]

Update code based on pre-commit test.

[eopXD]

llvm/test/Transforms/LoopStrengthReduce/lsr-fold-iv-complicate-add-rec.ll

46 ↗

(On Diff #454725)

Done.

50 ↗

(On Diff #454725)

Will add negative test in pre-commit test.

59 ↗

(On Diff #454725)

Yeah, this is real code extracted, will prune in pre-commit test.

[eopXD]

Update testcase effected by this patch.

[mcberg2021]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6593

With isLoopSimplifyForm() you do not need to check for LoopLatch and LoopPreheader here, it already does.

6612

What happens if there is more than one use here? Why not use hasOneUse() and return if false?

6621

If you use hasOneUse, you will not need a loop here, you can still get the beginning and check

[eopXD]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6612

The main intention here is that the primary IV of the loop, may be transformed by multiple LLVM-IR instructions, but is only used to determine the terminal condition. In this case the use-def chain will be a long chain until the final value that is for

1. used by the terminal condition
2. goes into the phi-node in the loop header

So loop here is to travel through the long chain until meeting value that is used exactly twice.

When the scenario 1 and 2 is satisfied, we can say that this IV (and its corresponding PHINode) is the one we are looking to fold. We will need to traverse through the chain to verify the final usage is for 1 and 2.

Here is a illustration.

[eopXD]

Add negative testcase.

[eopXD]

Since LSR is used on many targets, adding this transformation will affect many
testcases and it will make this patch hard to review. For now we can open up
an option for the added transformation and only observe on the middle-end
testcases.

After this patch lands, we can then go on and create a TTI to allow reviewers
from each target to see if they want this transformation in their backend.

[eopXD]

Recover testcase LoopStrengthReduce/opaque-ptr.ll since an option is now added
to guard the transformation.

[eopXD]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6593

Done. Thank you for the tip.

[eopXD]

Only perform transformation and set Changed to true when the SCEV is safe to expand.

[eopXD]

Add comments and illustration on how the transformation is qualified.
Add TODO for follow-ups.

[mcberg2021]

LGTM, you might allow time for others though...

[eopXD]

Update code based on slight adjustment in pre-commit test.

[eopXD]

Fix minor type-o.

[hiraditya]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

191–193

cl::desc("Attempt to replace primary IV with other IV."));

[craig.topper]

Should "terminal condition" in the description by "terminating condition"?

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

189

Terminal -> Terminating? The word terminal does not appear in LSR prior to this patch.

6579

std::tuple?

6785

Is this loop just ToHelpFold->getIncomingValueForBlock(LoopPreheader)?

6786

Do you really need a cast here? If you do, I'd use U.get() instead.

6788

Same here

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6611

Combine this if with the previous if

if (!isa<ICmpInst>(TermCond) || !cast<ICmpInst>(TermCond)->isEquality())

6634

std::function<bool(PHINode &)> -> auto

6637

This cast is unecessary.

6661

std::function<bool(PHINode &)> -> auto

6667

Can this use PN.getIncomingValueForBlock or getBasicBlockIndex?

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6697

Combine this with the previous if

[eopXD]

Address comments from @craig.topper.

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

189

Replaced all use of "terminal condition" to "terminating condition".

191–193

Thank you for the simplification. Was trying to explain in detail but this is much better.

6579

Thank you for the tip, done.

6611

Done.

6667

This simplifies the code very much. Thank you for the tip.

6697

Done. Thank you.

6785

Yes, thank you.

6786

The problem is elegantly solved by using PHINode:: getIncomingValueForBlock you've mentioned.

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

189

The option name still says Terminal

Review needs to re-titled and the description needs to be updated

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6604

neq -> ne

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6634

Is PN.getNumIncomingValues() == 2 implied by LoopSimplifyForm?

6661

Is PN.getNumIncomingValues() == 2 implied by LoopSimplifyForm?

6668

Are the LoopPreheader and LoopLatch being predecessors of the phi implied by LoopSimplify form?

[eopXD]

Address review comments.

[eopXD]

By definition of LoopSimplifyForm (contains exactly 1 preheader, exactly 1 latch, and has dedicated exit), the code can be simplified.
Thank you @craig.topper for the review.

[eopXD]

Rebased upon change in D132443.

[eopXD]

Rename testcase.

[fhahn]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6609

Looks like a negative test with different branch conditions is missing.

6670

it looks like there's a negative test missing with types that aren't scev-able (e.g. vector types I think).

6675

It looks like there's a test case missing where one of the phis isn't an AddRec/non-affine AddRec.

[nikic]

Could you please update the patch description to describe in more detailed what this patch does?

[eopXD]

Added missing negative testcases.

[eopXD]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6670

Done, thank you for the reminder.

6675

Done, thank you for the reminder.

[eopXD]

Fix type-o.

[eopXD]

Some statistics to help push the patch, here are the number of loops affected when compiling spec2k6 with downstream compiler and this patch applied:

400.perlbench 56
401.bzip2 6
403.gcc 155
429.mcf 1
445.gobmk 45
456.hmmer 44
458.sjeng 9
462.libquantum 4
464.h264ref 125
471.omnetpp 7
473.astar 14
483.xalancbmk 81

[eopXD]

Add counter to record number of folding done.

[eopXD]

Gentle ping, thank you. @fhahn @craig.topper

Is the description acceptable now? @nikic

[eopXD]

ping, thank you.

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6579

Looking at this again. Could this be an Optional<std::pair<PHINode *, PHINode *>> Then you could return None for CantFold.

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6840

What are the other users of OldTermCond and why is it ok to replace them with poison?

6845

Is there a DCE after this? Earlier optimizations in this function used DeleteDeadPHIs to recursively clean up.

[eopXD]

Address comment.

[eopXD]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6579

Modified, thank you for the tip.

[eopXD]

Address comment.

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6840

Nice catch. Added check and negative test for it.

6845

Yes. Removing the terminal condition and leveraging DeleteDeadPHIs is a much better approach. Thank you.

[craig.topper]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6840

What if OldTermCond isn't the direct user of the ToFold phi? Do we clean everything?

[eopXD]

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

6840

Yes, DeleteDeadPHIs is able to delete the chain of uses of the ToFold phi.

[eopXD]

Fix testcase.

[craig.topper]

LGTM

[eopXD]

Rebase to latest main.

[eopXD]

Rebase to latest main.

[reames]

I recently stumbled into a test case which would benefit from this, and as a result, took a close look at the code in tree.

I posted one generalization https://reviews.llvm.org/D142240 so that this covers my case of interest.

However, I also think we've got a couple of correctness issues here which need addressed.

First, I don't see anything in the code which checks that the alternate IV is at least as wide as the starting IV. As a result, we can try to rewrite a loop with a BE count of say 1024 to use an IV which is an i8. This can change the BE count of the loop.

Second, we're adding a use to an existing IV. When doing that, we need to worry about about poison correctness. The case to consider is the alternate IV overflowing on the last iteration, and that poison value previously being unused (since we don't take backedge), and now used in the latch test (resulting in UB).

[reames]

(More observations)

For example of code which gets both of the above issues correct, look at the LFTR and rewriteLoopExitValues transformations.

Third, we need some form of cost model. The simplest would be to add a isHighCostExpansion check on the new bound. SCEV expressions can (in edge cases) be quite involved, and we probably don't want to do this rewrite universally.

llvm/lib/Transforms/Scalar/LoopStrengthReduce.cpp

static cl::opt<unsigned> ComplexityLimit(
   "lsr-complexity-limit", cl::Hidden,
   cl::init(std::numeric_limits<uint16_t>::max()),
   cl::desc("LSR search space complexity limit"));
 
 static cl::opt<unsigned> SetupCostDepthLimit(
     "lsr-setupcost-depth-limit", cl::Hidden, cl::init(7),
     cl::desc("The limit on recursion depth for LSRs setup cost"));
 
+static cl::opt<bool> AllowTerminatingConditionFoldingAfterLSR(
+    "lsr-term-fold", cl::Hidden, cl::init(false),
+    cl::desc("Attempt to replace primary IV with other IV."));
+
 #ifndef NDEBUG
 // Stress test IV chain generation.
 static cl::opt<bool> StressIVChain(
   "stress-ivchain", cl::Hidden, cl::init(false),
   cl::desc("Stress test LSR IV chains"));
 #else
 static bool StressIVChain = false;
 #endif
 
static llvm::PHINode *GetInductionVariable(const Loop &L, ScalarEvolution &SE,
 
   for (PHINode &P : L.getHeader()->phis()) {
     if (IsSuitableIV(&P))
       return &P;
   }
   return nullptr;
 }
 
+static std::tuple<bool, PHINode *, PHINode *>
+canFoldTermCondOfLoop(Loop *L, ScalarEvolution &SE, DominatorTree &DT,
+                      const LoopInfo &LI) {
+  const std::tuple<bool, PHINode *, PHINode *> CantFold = {false, nullptr,
+                                                           nullptr};
+
+  if (!L->isInnermost()) {
+    LLVM_DEBUG(dbgs() << "Cannot fold on non-innermost loop\n");
+    return CantFold;
+  }
+  // Only inspect on simple loop structure
+  if (!L->isLoopSimplifyForm()) {
+    LLVM_DEBUG(dbgs() << "Cannot fold on non-simple loop\n");
+    return CantFold;
+  }
+
+  if (!SE.hasLoopInvariantBackedgeTakenCount(L)) {
+    LLVM_DEBUG(dbgs() << "Cannot fold on backedge that is loop variant\n");
+    return CantFold;
+  }
+
+  BasicBlock *LoopPreheader = L->getLoopPreheader();
+  BasicBlock *LoopLatch = L->getLoopLatch();
+
+  // TODO: Can we do something for greater than and less than?
+  // Terminating condition is foldable when it is an eq/ne icmp
+  BranchInst *BI = cast<BranchInst>(LoopLatch->getTerminator());
+  if (BI->isUnconditional())
+    return CantFold;
+  Value *TermCond = BI->getCondition();
+  if (!isa<ICmpInst>(TermCond) || !cast<ICmpInst>(TermCond)->isEquality()) {
+    LLVM_DEBUG(dbgs() << "Cannot fold on branching condition that is not an "
+                         "ICmpInst::eq / ICmpInst::ne\n");
+    return CantFold;
+  }
+
+  // For `IsToFold`, a primary IV can be replaced by other affine AddRec when it
+  // is only used by the terminating condition. To check for this, we may need
+  // to traverse through a chain of use-def until we can examine the final
+  // usage.
+  //         *----------------------*
+  //   *---->|  LoopHeader:         |
+  //   |     |  PrimaryIV = phi ... |
+  //   |     *----------------------*
+  //   |              |
+  //   |              |
+  //   |           chain of
+  //   |          single use
+  // used by          |
+  //  phi             |
+  //   |            Value
+  //   |          /       \
+  //   |     chain of     chain of
+  //   |    single use     single use
+  //   |      /               \
+  //   |     /                 \
+  //   *- Value                Value --> used by terminating condition
+  auto IsToFold = [&](PHINode &PN) -> bool {
+    Value *V = &PN;
+
+    while (V->getNumUses() == 1)
+      V = *V->user_begin();
+
+    if (V->getNumUses() != 2)
+      return false;
+
+    Value *VToPN = nullptr;
+    Value *VToTermCond = nullptr;
+    for (User *U : V->users()) {
+      while (U->getNumUses() == 1) {
+        if (isa<PHINode>(U))
+          VToPN = U;
+        if (U == TermCond)
+          VToTermCond = U;
+        U = *U->user_begin();
+      }
+    }
+    return VToPN && VToTermCond;
+  };
+
+  // For `IsToHelpFold`, other IV that is an affine AddRec will be sufficient to
+  // replace the terminating condition
+  auto IsToHelpFold = [&](PHINode &PN) -> bool {
+    // TODO: Right now we limit the phi node to help the folding be of a start
+    // value of getelementptr. We can extend to any kinds of IV as long as it is
+    // an affine AddRec. Add a switch to cover more types of instructions here
+    // and down in the actual transformation.
+    return isa<GetElementPtrInst>(PN.getIncomingValueForBlock(LoopPreheader));
+  };
+
+  PHINode *ToFold = nullptr;
+  PHINode *ToHelpFold = nullptr;
+
+  for (PHINode &PN : L->getHeader()->phis()) {
+    if (!SE.isSCEVable(PN.getType())) {
+      LLVM_DEBUG(dbgs() << "IV of phi '" << PN
+                        << "' is not SCEV-able, not qualified for the "
+                           "terminating condition folding.\n");
+      continue;
+    }
+    const SCEV *S = SE.getSCEV(&PN);
+    const SCEVAddRecExpr *AddRec = dyn_cast<SCEVAddRecExpr>(S);
+    // Only speculate on affine AddRec
+    if (!AddRec || !AddRec->isAffine()) {
+      LLVM_DEBUG(dbgs() << "SCEV of phi '" << PN
+                        << "' is not an affine add recursion, not qualified "
+                           "for the terminating condition folding.\n");
+      continue;
+    }
+
+    if (IsToFold(PN))
+      ToFold = &PN;
+    else if (IsToHelpFold(PN))
+      ToHelpFold = &PN;
+  }
+
+  LLVM_DEBUG(if (ToFold && !ToHelpFold) dbgs()
+                 << "Cannot find other AddRec IV to help folding\n";);
+
+  LLVM_DEBUG(if (ToFold && ToHelpFold) dbgs()
+             << "\nFound loop that can fold terminating condition\n"
+             << "  BECount (SCEV): " << *SE.getBackedgeTakenCount(L) << "\n"
+             << "  TermCond: " << *TermCond << "\n"
+             << "  BrandInst: " << *BI << "\n"
+             << "  ToFold: " << *ToFold << "\n"
+             << "  ToHelpFold: " << *ToHelpFold << "\n");
+
+  return {ToFold != nullptr && ToHelpFold != nullptr, ToFold, ToHelpFold};
+}
+
 static bool ReduceLoopStrength(Loop *L, IVUsers &IU, ScalarEvolution &SE,
                                DominatorTree &DT, LoopInfo &LI,
                                const TargetTransformInfo &TTI,
                                AssumptionCache &AC, TargetLibraryInfo &TLI,
                                MemorySSA *MSSA) {
 
   // Debug preservation - before we start removing anything identify which DVI
   // meet the salvageable criteria and store their DIExpression and SCEVs.
if (L->isRecursivelyLCSSAForm(DT, LI) && L->getExitBlock()) {
     if (Rewrites) {
       Changed = true;
       RecursivelyDeleteTriviallyDeadInstructionsPermissive(DeadInsts, &TLI,
                                                            MSSAU.get());
       DeleteDeadPHIs(L->getHeader(), &TLI, MSSAU.get());
     }
   }
 
+  if (AllowTerminatingConditionFoldingAfterLSR) {
+    auto CanFoldTerminatingCondition = canFoldTermCondOfLoop(L, SE, DT, LI);
+    if (std::get<0>(CanFoldTerminatingCondition) == true) {
+      BasicBlock *LoopPreheader = L->getLoopPreheader();
+      BasicBlock *LoopLatch = L->getLoopLatch();
+
+      PHINode *ToFold = std::get<1>(CanFoldTerminatingCondition);
+      PHINode *ToHelpFold = std::get<2>(CanFoldTerminatingCondition);
+
+      LLVM_DEBUG(dbgs() << "To fold phi-node:\n"
+                        << *ToFold << "\n"
+                        << "New term-cond phi-node:\n"
+                        << *ToHelpFold << "\n");
+
+      Value *StartValue = ToHelpFold->getIncomingValueForBlock(LoopPreheader);
+      Value *LoopValue = ToHelpFold->getIncomingValueForBlock(LoopLatch);
+
+      // SCEVExpander for both use in preheader and latch
+      const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
+      SCEVExpander Expander(SE, DL, "lsr_fold_term_cond");
+      SCEVExpanderCleaner ExpCleaner(Expander);
+
+      // Create new terminating value at loop header
+      GetElementPtrInst *StartValueGEP = cast<GetElementPtrInst>(StartValue);
+      Type *PtrTy = StartValueGEP->getPointerOperand()->getType();
+
+      const SCEV *BECount = SE.getBackedgeTakenCount(L);
+      const SCEVAddRecExpr *AddRec =
+          cast<SCEVAddRecExpr>(SE.getSCEV(ToHelpFold));
+
+      // TermValue = Start + Stride * (BackedgeCount + 1)
+      const SCEV *TermValueS = SE.getAddExpr(
+          AddRec->getOperand(0),
+          SE.getTruncateOrZeroExtend(
+              SE.getMulExpr(
+                  AddRec->getOperand(1),
+                  SE.getTruncateOrZeroExtend(
+                      SE.getAddExpr(BECount, SE.getOne(BECount->getType())),
+                      AddRec->getOperand(1)->getType())),
+              AddRec->getOperand(0)->getType()));
+
+      // NOTE: If this is triggered, we should add this into predicate
+      if (!Expander.isSafeToExpand(TermValueS)) {
+        LLVMContext &Ctx = L->getHeader()->getContext();
+        Ctx.emitError(
+            "Terminating value is not safe to expand, need to add it to "
+            "predicate");
+      } else { // Now we replace the condition with ToHelpFold and remove ToFold
+        Changed = true;
+
+        Value *TermValue = Expander.expandCodeFor(
+            TermValueS, PtrTy, LoopPreheader->getTerminator());
+
+        LLVM_DEBUG(dbgs() << "Start value of new term-cond phi-node:\n"
+                          << *StartValue << "\n"
+                          << "Terminating value of new term-cond phi-node:\n"
+                          << *TermValue << "\n");
+
+        // Create new terminating condition at loop latch
+        BranchInst *BI = cast<BranchInst>(LoopLatch->getTerminator());
+        ICmpInst *OldTermCond = cast<ICmpInst>(BI->getCondition());
+        IRBuilder<> LatchBuilder(LoopLatch->getTerminator());
+        Value *NewTermCond = LatchBuilder.CreateICmp(
+            OldTermCond->getPredicate(), LoopValue, TermValue,
+            "lsr_fold_term_cond.replaced_term_cond");
+
+        LLVM_DEBUG(dbgs() << "Old term-cond:\n"
+                          << *OldTermCond << "\n"
+                          << "New term-cond:\b" << *NewTermCond << "\n");
+
+        BI->setCondition(NewTermCond);
+
+        OldTermCond->replaceAllUsesWith(
+            PoisonValue::get(OldTermCond->getType()));
+        OldTermCond->eraseFromParent();
+
+        // Cleanup the old terminating condition that is no longer used
+        // Clear the PHINode and DCE will do the rest...
+        while (ToFold->getNumIncomingValues())
+          ToFold->removeIncomingValue(0u);
+      }
+
+      ExpCleaner.markResultUsed();
+    }
+  }
+
   if (SalvageableDVIRecords.empty())
     return Changed;
 
   // Obtain relevant IVs and attempt to rewrite the salvageable DVIs with
   // expressions composed using the derived iteration count.
   // TODO: Allow for multiple IV references for nested AddRecSCEVs
   for (const auto &L : LI) {
     if (llvm::PHINode *IV = GetInductionVariable(*L, SE, Reducer))

llvm/test/Transforms/LoopStrengthReduce/lsr-term-fold-negative-testcase.ll

+; REQUIRES: asserts
+; RUN: opt < %s -passes="loop-reduce" -S -debug -lsr-term-fold 2>&1 | FileCheck %s
+
+target datalayout = "e-p:32:32:32-n32"
+
+define i32 @loop_variant(ptr %ar, i32 %n, i32 %m) {
+; CHECK: Cannot fold on backedge that is loop variant
+entry:
+  br label %for.cond
+
+for.cond:                                         ; preds = %for.cond, %entry
+  %n.addr.0 = phi i32 [ %n, %entry ], [ %mul, %for.cond ]
+  %cmp = icmp slt i32 %n.addr.0, %m
+  %mul = shl nsw i32 %n.addr.0, 1
+  br i1 %cmp, label %for.cond, label %for.end
+
+for.end:                                          ; preds = %for.cond
+  ret i32 %n.addr.0
+}
+
+define i32 @nested_loop(ptr %ar, i32 %n, i32 %m, i32 %o) {
+; CHECK: Cannot fold on backedge that is loop variant
+; CHECK: Cannot fold on non-innermost loop
+entry:
+  %cmp15 = icmp sgt i32 %o, 0
+  br i1 %cmp15, label %for.body, label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.cleanup3, %entry
+  %cnt.0.lcssa = phi i32 [ 0, %entry ], [ %cnt.1.lcssa, %for.cond.cleanup3 ]
+  ret i32 %cnt.0.lcssa
+
+for.body:                                         ; preds = %entry, %for.cond.cleanup3
+  %i.017 = phi i32 [ %inc6, %for.cond.cleanup3 ], [ 0, %entry ]
+  %cnt.016 = phi i32 [ %cnt.1.lcssa, %for.cond.cleanup3 ], [ 0, %entry ]
+  %sub = sub nsw i32 %n, %i.017
+  %cmp212 = icmp slt i32 %sub, %m
+  br i1 %cmp212, label %for.body4, label %for.cond.cleanup3
+
+for.cond.cleanup3:                                ; preds = %for.body4, %for.body
+  %cnt.1.lcssa = phi i32 [ %cnt.016, %for.body ], [ %inc, %for.body4 ]
+  %inc6 = add nuw nsw i32 %i.017, 1
+  %cmp = icmp slt i32 %inc6, %o
+  br i1 %cmp, label %for.body, label %for.cond.cleanup
+
+for.body4:                                        ; preds = %for.body, %for.body4
+  %j.014 = phi i32 [ %mul, %for.body4 ], [ %sub, %for.body ]
+  %cnt.113 = phi i32 [ %inc, %for.body4 ], [ %cnt.016, %for.body ]
+  %inc = add nsw i32 %cnt.113, 1
+  %mul = shl nsw i32 %j.014, 1
+  %cmp2 = icmp slt i32 %mul, %m
+  br i1 %cmp2, label %for.body4, label %for.cond.cleanup3
+}
+
+; The terminating condition folding transformation cannot find the ptr IV
+; because it checks if the value comes from the LoopPreheader. %mark is from
+; the function argument, so it is not qualified for the transformation.
+define void @no_iv_to_help(ptr %mark, i32 signext %length) {
+; CHECK: Cannot find other AddRec IV to help folding
+entry:
+  %tobool.not3 = icmp eq i32 %length, 0
+  br i1 %tobool.not3, label %for.cond.cleanup, label %for.body
+
+for.cond.cleanup:                                 ; preds = %for.body, %entry
+  ret void
+
+for.body:                                         ; preds = %entry, %for.body
+  %i.05 = phi i32 [ %dec, %for.body ], [ %length, %entry ]
+  %dst.04 = phi ptr [ %incdec.ptr, %for.body ], [ %mark, %entry ]
+  %0 = load ptr, ptr %dst.04, align 8
+  call ptr @foo(ptr %0)
+  %incdec.ptr = getelementptr inbounds ptr, ptr %dst.04, i64 1
+  %dec = add nsw i32 %i.05, -1
+  %tobool.not = icmp eq i32 %dec, 0
+  br i1 %tobool.not, label %for.cond.cleanup, label %for.body
+}
+
+declare void @foo(ptr)
+
+define void @NonAddRecIV(ptr %a) {
+; CHECK: SCEV of phi '  %lsr.iv = phi i32 [ %lsr.iv.next, %for.body ], [ 1, %entry ]'
+; CHECK: is not an affine add recursion, not qualified for the terminating condition folding.
+entry:
+  %uglygep = getelementptr i8, ptr %a, i32 84
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %lsr.iv1 = phi ptr [ %uglygep2, %for.body ], [ %uglygep, %entry ]
+  %lsr.iv = phi i32 [ %lsr.iv.next, %for.body ], [ 1, %entry ]
+  store i32 1, ptr %lsr.iv1, align 4
+  %lsr.iv.next = mul nsw i32 %lsr.iv, 2
+  %uglygep2 = getelementptr i8, ptr %lsr.iv1, i64 4
+  %exitcond.not = icmp eq i32 %lsr.iv.next, 65536
+  br i1 %exitcond.not, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret void
+}
+
+@fp_inc = common global float 0.000000e+00, align 4
+
+define void @NonSCEVableIV(float %init, float* %A, i32 %N) {
+; CHECK: IV of phi '  %x.05 = phi float [ %init, %entry ], [ %add, %for.body ]'
+; CHECK: is not SCEV-able, not qualified for the terminating condition folding.
+entry:
+  %0 = load float, float* @fp_inc, align 4
+  br label %for.body
+
+for.body:                                         ; preds = %entry
+  %indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
+  %x.05 = phi float [ %init, %entry ], [ %add, %for.body ]
+  %arrayidx = getelementptr inbounds float, float* %A, i64 %indvars.iv
+  store float %x.05, float* %arrayidx, align 4
+  %add = fsub float %x.05, %0
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+  %lftr.wideiv = trunc i64 %indvars.iv.next to i32
+  %exitcond = icmp eq i32 %lftr.wideiv, %N
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.end
+  ret void
+}
+
+define void @NonIcmpEqNe(ptr %a) {
+; CHECK: Cannot fold on branching condition that is not an ICmpInst::eq / ICmpInst::ne
+entry:
+  %uglygep = getelementptr i8, ptr %a, i64 84
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %lsr.iv1 = phi ptr [ %uglygep2, %for.body ], [ %uglygep, %entry ]
+  %lsr.iv = phi i64 [ %lsr.iv.next, %for.body ], [ 379, %entry ]
+  store i32 1, ptr %lsr.iv1, align 4
+  %lsr.iv.next = add nsw i64 %lsr.iv, -1
+  %uglygep2 = getelementptr i8, ptr %lsr.iv1, i64 4
+  %exitcond.not = icmp sle i64 %lsr.iv.next, 0
+  br i1 %exitcond.not, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret void
+}

llvm/test/Transforms/LoopStrengthReduce/lsr-term-fold.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt < %s -passes="loop-reduce" -S | FileCheck %s
+; RUN: opt < %s -passes="loop-reduce" -S -lsr-term-fold | FileCheck %s
 
 ; There are 3 test cases here regarding the replacing the primary iv with
 ; other affine AddRec IV.
 ; 1. Loop with constant tripcount
 ; 2. Loop with runtime tripcount
 ; 3. The replacing AddRec IV is a complicated AddRec. This tests whether
 ;    the fold terminating condition transformation is writing new terminating
 ;    condition in the correct type.
 
 target datalayout = "e-p:32:32:32-n32"
 
 ; This is equivalent to the following code, loop guard is removed.
 ;   void const_tripcount(int *a) {
 ;     for (int i = 21; i < 400; i++)
 ;       a[i] = 1;
 ;   }
 define void @const_tripcount(ptr %a) {
 ; CHECK-LABEL: @const_tripcount(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[UGLYGEP:%.*]] = getelementptr i8, ptr [[A:%.*]], i64 84
+; CHECK-NEXT:    [[UGLYGEP1:%.*]] = getelementptr i8, ptr [[A]], i32 1600
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[LSR_IV1:%.*]] = phi ptr [ [[UGLYGEP2:%.*]], [[FOR_BODY]] ], [ [[UGLYGEP]], [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[LSR_IV:%.*]] = phi i64 [ [[LSR_IV_NEXT:%.*]], [[FOR_BODY]] ], [ 379, [[ENTRY]] ]
 ; CHECK-NEXT:    store i32 1, ptr [[LSR_IV1]], align 4
-; CHECK-NEXT:    [[LSR_IV_NEXT]] = add nsw i64 [[LSR_IV]], -1
+; CHECK-NEXT:    [[LSR_IV_NEXT:%.*]] = add nsw i64 poison, -1
 ; CHECK-NEXT:    [[UGLYGEP2]] = getelementptr i8, ptr [[LSR_IV1]], i64 4
-; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i64 [[LSR_IV_NEXT]], 0
-; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; CHECK-NEXT:    [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND:%.*]] = icmp eq ptr [[UGLYGEP2]], [[UGLYGEP1]]
+; CHECK-NEXT:    br i1 [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %uglygep = getelementptr i8, ptr %a, i64 84
   br label %for.body
 
 for.body:                                         ; preds = %for.body, %entry

 ;     for (int i = 21; i < N; i++)
 ;       a[i] = 1;
 ;   }
 
 define void @runtime_tripcount(ptr %a, i32 %N) {
 ; CHECK-LABEL: @runtime_tripcount(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[UGLYGEP:%.*]] = getelementptr i8, ptr [[A:%.*]], i32 84
+; CHECK-NEXT:    [[TMP0:%.*]] = shl i32 [[N:%.*]], 2
+; CHECK-NEXT:    [[TMP1:%.*]] = add i32 [[TMP0]], 84
+; CHECK-NEXT:    [[UGLYGEP1:%.*]] = getelementptr i8, ptr [[A]], i32 [[TMP1]]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[LSR_IV1:%.*]] = phi ptr [ [[UGLYGEP2:%.*]], [[FOR_BODY]] ], [ [[UGLYGEP]], [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[LSR_IV:%.*]] = phi i32 [ [[LSR_IV_NEXT:%.*]], [[FOR_BODY]] ], [ [[N:%.*]], [[ENTRY]] ]
 ; CHECK-NEXT:    store i32 1, ptr [[LSR_IV1]], align 4
-; CHECK-NEXT:    [[LSR_IV_NEXT]] = add nsw i32 [[LSR_IV]], -1
+; CHECK-NEXT:    [[LSR_IV_NEXT:%.*]] = add nsw i32 poison, -1
 ; CHECK-NEXT:    [[UGLYGEP2]] = getelementptr i8, ptr [[LSR_IV1]], i64 4
-; CHECK-NEXT:    [[EXITCOND_NOT:%.*]] = icmp eq i32 [[LSR_IV_NEXT]], 0
-; CHECK-NEXT:    br i1 [[EXITCOND_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; CHECK-NEXT:    [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND:%.*]] = icmp eq ptr [[UGLYGEP2]], [[UGLYGEP1]]
+; CHECK-NEXT:    br i1 [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %uglygep = getelementptr i8, ptr %a, i32 84
   br label %for.body
 
 for.body:                                         ; preds = %for.body, %entry

 ;       for (int i = length; i; i--) {
 ;           foo(*it);
 ;           it++;
 ;       }
 ;   }
 define void @ptr_of_ptr_addrec(ptr %ptrptr, i32 %length) {
 ; CHECK-LABEL: @ptr_of_ptr_addrec(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    [[START_PTRPTR:%.*]] = getelementptr inbounds ptr, ptr [[PTRPTR:%.*]]
+; CHECK-NEXT:    [[START_PTRPTR:%.*]] = getelementptr ptr, ptr [[PTRPTR:%.*]]
+; CHECK-NEXT:    [[TMP0:%.*]] = shl i32 [[LENGTH:%.*]], 2
+; CHECK-NEXT:    [[UGLYGEP:%.*]] = getelementptr i8, ptr [[START_PTRPTR]], i32 [[TMP0]]
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
-; CHECK-NEXT:    [[I_05:%.*]] = phi i32 [ [[DEC:%.*]], [[FOR_BODY]] ], [ [[LENGTH:%.*]], [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[IT_04:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[FOR_BODY]] ], [ [[START_PTRPTR]], [[ENTRY]] ]
-; CHECK-NEXT:    [[TMP0:%.*]] = load ptr, ptr [[IT_04]], align 8
-; CHECK-NEXT:    tail call void @foo(ptr [[TMP0]])
+; CHECK-NEXT:    [[IT_04:%.*]] = phi ptr [ [[INCDEC_PTR:%.*]], [[FOR_BODY]] ], [ [[START_PTRPTR]], [[ENTRY:%.*]] ]
+; CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[IT_04]], align 8
+; CHECK-NEXT:    tail call void @foo(ptr [[TMP1]])
 ; CHECK-NEXT:    [[INCDEC_PTR]] = getelementptr inbounds ptr, ptr [[IT_04]], i64 1
-; CHECK-NEXT:    [[DEC]] = add i32 [[I_05]], -1
-; CHECK-NEXT:    [[TOBOOL_NOT:%.*]] = icmp eq i32 [[DEC]], 0
-; CHECK-NEXT:    br i1 [[TOBOOL_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; CHECK-NEXT:    [[DEC:%.*]] = add i32 poison, -1
+; CHECK-NEXT:    [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND:%.*]] = icmp eq ptr [[INCDEC_PTR]], [[UGLYGEP]]
+; CHECK-NEXT:    br i1 [[LSR_FOLD_TERM_COND_REPLACED_TERM_COND]], label [[FOR_END:%.*]], label [[FOR_BODY]]
 ; CHECK:       for.end:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %start.ptrptr = getelementptr inbounds ptr, ptr %ptrptr
   br label %for.body
 
 for.body:                                         ; preds = %entry, %for.body