[LoopInstSimplify] Re-implement the core logic of loop-instsimplify to be both simpler and substantially more efficient.

[chandlerc]

Summary

Rather than use a hand-rolled iteration technique that isn't quite the
same as RPO, use the pre-built RPO loop body traversal utility.

Once visiting the loop body in RPO, we can assert that we visit defs
before uses reliably. When this is the case, the only need to iterate is
when simplifying a def that is used by a PHI node along a back-edge.
With this patch, the first pass over the loop body is just a complete
simplification of every instruction across the loop body. When we
encounter a use of a simplified instruction that stems from a PHI node
in the loop body that has already been visited (due to some cyclic CFG,
potentially the loop itself, or a nested loop, or unstructured control
flow), we recall that specific PHI node for the second iteration.
Nothing else needs to be preserved from iteration to iteration.

On the second and later iterations, only instructions known to have
simplified inputs are considered, each time starting from a set of PHIs
that had simplified inputs along the backedges.

Dead instructions are collected along the way, but deleted in a batch at
the end of each iteration making the iterations themselves substantially
simpler. This uses a new batch API for recursively deleting dead
instructions.

This alsa changes the routine to visit subloops. Because simplification
is fundamentally transitive, we may need to visit the entire loop body,
including subloops, to handle knock-on simplification.

I've added a basic test file that helps demonstrate that all of these
changes work. It includes both straight-forward loops with
simplifications as well as interesting PHI-structures, CFG-structures,
and a nested loop case.

[chandlerc]

Fix some formatting goofs that snuck in w/o a run of clang-format to clean them up.

[asbirlea]

LGTM. Looks like a clean simplification to me.

llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp

131 ↗

(On Diff #148711)

Nit: "an iteration over all instructions in all blocks." perhaps.
For code readability it would be easier to understand IMO. Alternatively " end loop over instructions", " end loop over blocks" for above closing braces, but that seems too much.

[sanjoy]

llvm/lib/Transforms/Scalar/LoopInstSimplify.cpp

92 ↗

(On Diff #148711)

How about pulling out ToSimplify->empty() into a bool IsFirstIteration? I think that will make the intent clearer.

118 ↗

(On Diff #148711)

Might be worth asserting that if !L.contains(UserI) then UserI is a PHI node. Up to you.

[asbirlea]

llvm/lib/Transforms/Utils/Local.cpp

448 ↗

(On Diff #148711)

assert for isInstructionTriviallyDead too?

[chandlerc]

All suggestions implemented, submitting now, thanks!

[mzolotukhin]

llvm/trunk/lib/Transforms/Scalar/LoopInstSimplify.cpp

82

I wonder if it would be more efficient to iterate through ToSimplify instead of all instructions in all blocks. What do you think?

[chandlerc]

llvm/trunk/lib/Transforms/Scalar/LoopInstSimplify.cpp

82

I think it's trickier than that, but let me know if you see a way that seems more promising.

We need to visit defs before uses to reliably converge on the simplified result without revisiting instructions even more times. And we don't know all of the instructions we will need to visit, we only know the first ones. Each time we simplify an instruction we (potentially) grow the ToSimplify set.

So to only look at the instructions in the ToSimplify set, I think we'd need a system like the following:

• First build a mapping from every instruction in the loop body to an integer so that they can be cheaply sorted in RPO. We could do this in the first iteration.
• Build a sorted worklist in addition to a ToSimplify set, and every time we add an instruction to the ToSimplify set, insert it into the worklist in the correct (based on sort) position.
• Walk that worklist rather than all the instructions in the loop body.

To build a worklist like this which has good algorithmic properties, what you'd probably want is to make ToSimplify actually a std::set or some other ordered search tree with cheap in-order insertion because we'll constantly be inserting into all kinds of different positions.

A std::set (or similar) tree data structure combined with the std::less needing to do two hash table lookups to get the sort key seemed like it would end up having a really frustratingly high overhead, and that's why I didn't immediately jump to this solution.

Another approach which doesn't have any better *algorithmic* properties in the worst case than the current one but might have hilariously better practical properties would be as follows:

• Build a vector of instruction pointers in RPO order during the initial traversal, and a map from instruction to vector index.
• Instead of using a ToSimplify set, use a sparse bitvector where the bit represents that the instruction at this index is in the set.
• Walk the sparse bitvector in order in all subsequent iterations, and then use the index of the set bits to find the instruction at that position in the RPO.

Technically, in the worst case, this is the same as the current approach -- for a loop with N instructions that we iterate on M times we do O(N*M) work. But with this approach, that work involves looking a bit in a fairly cache-optimized data structure rather than walking to a linked list node and testing it in a hash table. And as the iteration becomes sparse, we actually get algorithmic improvements that would likely help in the average case.

The down side is that it is much more complex and requires a decent amount of memory, which is why I didn't implement it at first.

I'm happy to pursue either of these if you and others thing it is worthwhile. I'm not sure we'll ever have a test case where this shows up as a practical problem, but that doesn't mean one won't show up. =D I mostly had trouble making the trade-offs here, and this patch seemed a strict improvement.

I'll admit I'm more reluctant about the first approach as I suspect its algorithmic scaling will be completely undermined by the practical cost of maintaining the sorted data structure.

But the second approach is super appealing. I didn't think super carefully about it when I wrote this the first time to convince myself it would be effective, but I'm happy to go do this work if you think its worthwhile based on my description.

[mzolotukhin]

llvm/trunk/lib/Transforms/Scalar/LoopInstSimplify.cpp

82

A std::set (or similar) tree data structure combined with the std::less needing to do two hash table lookups to get the sort key seemed like it would end up having a really frustratingly high overhead, and that's why I didn't immediately jump to this solution.

I was thinking about priority_queue here, which should have all the desired properties and supposedly less overhead, than set. We still will have to spend a lot of memory for enumerating all instructions, so maybe it's not worth it after all.

...this patch seemed a strict improvement.

I don't argue with that :)

llvm/trunk/include/llvm/Analysis/Utils/Local.h

bool wouldInstructionBeTriviallyDead(Instruction *I,
                                      const TargetLibraryInfo *TLI = nullptr);
 
 /// If the specified value is a trivially dead instruction, delete it.
 /// If that makes any of its operands trivially dead, delete them too,
 /// recursively. Return true if any instructions were deleted.
 bool RecursivelyDeleteTriviallyDeadInstructions(Value *V,
                                         const TargetLibraryInfo *TLI = nullptr);
 
+/// Delete all of the instructions in `DeadInsts`, and all other instructions
+/// that deleting these in turn causes to be trivially dead.
+///
+/// The initial instructions in the provided vector must all have empty use
+/// lists and satisfy `isInstructionTriviallyDead`.
+///
+/// `DeadInsts` will be used as scratch storage for this routine and will be
+/// empty afterward.
+void RecursivelyDeleteTriviallyDeadInstructions(
+    SmallVectorImpl<Instruction *> &DeadInsts,
+    const TargetLibraryInfo *TLI = nullptr);
+
 /// If the specified value is an effectively dead PHI node, due to being a
 /// def-use chain of single-use nodes that either forms a cycle or is terminated
 /// by a trivially dead instruction, delete it. If that makes any of its
 /// operands trivially dead, delete them too, recursively. Return true if a
 /// change was made.
 bool RecursivelyDeleteDeadPHINode(PHINode *PN,
                                   const TargetLibraryInfo *TLI = nullptr);
 

llvm/trunk/lib/Transforms/Scalar/LoopInstSimplify.cpp

 #include "llvm/ADT/PointerIntPair.h"
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/InstructionSimplify.h"
 #include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopIterator.h"
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/Utils/Local.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/CFG.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Dominators.h"
 #include "llvm/IR/Instruction.h"

 #include <utility>
 
 using namespace llvm;
 
 #define DEBUG_TYPE "loop-instsimplify"
 
 STATISTIC(NumSimplified, "Number of redundant instructions simplified");
 
-static bool SimplifyLoopInst(Loop *L, DominatorTree *DT, LoopInfo *LI,
-                             AssumptionCache *AC,
-                             const TargetLibraryInfo *TLI) {
-  SmallVector<BasicBlock *, 8> ExitBlocks;
-  L->getUniqueExitBlocks(ExitBlocks);
-  array_pod_sort(ExitBlocks.begin(), ExitBlocks.end());
-
+static bool simplifyLoopInst(Loop &L, DominatorTree &DT, LoopInfo &LI,
+                             AssumptionCache &AC,
+                             const TargetLibraryInfo &TLI) {
+  const DataLayout &DL = L.getHeader()->getModule()->getDataLayout();
+  SimplifyQuery SQ(DL, &TLI, &DT, &AC);
+
+  // On the first pass over the loop body we try to simplify every instruction.
+  // On subsequent passes, we can restrict this to only simplifying instructions
+  // where the inputs have been updated. We end up needing two sets: one
+  // containing the instructions we are simplifying in *this* pass, and one for
+  // the instructions we will want to simplify in the *next* pass. We use
+  // pointers so we can swap between two stably allocated sets.
   SmallPtrSet<const Instruction *, 8> S1, S2, *ToSimplify = &S1, *Next = &S2;
 
-  // The bit we are stealing from the pointer represents whether this basic
-  // block is the header of a subloop, in which case we only process its phis.
-  using WorklistItem = PointerIntPair<BasicBlock *, 1>;
-  SmallVector<WorklistItem, 16> VisitStack;
-  SmallPtrSet<BasicBlock *, 32> Visited;
+  // Track the PHI nodes that have already been visited during each iteration so
+  // that we can identify when it is necessary to iterate.
+  SmallPtrSet<PHINode *, 4> VisitedPHIs;
+
+  // While simplifying we may discover dead code or cause code to become dead.
+  // Keep track of all such instructions and we will delete them at the end.
+  SmallVector<Instruction *, 8> DeadInsts;
+
+  // First we want to create an RPO traversal of the loop body. By processing in
+  // RPO we can ensure that definitions are processed prior to uses (for non PHI
+  // uses) in all cases. This ensures we maximize the simplifications in each
+  // iteration over the loop and minimizes the possible causes for continuing to
+  // iterate.
+  LoopBlocksRPO RPOT(&L);
+  RPOT.perform(&LI);
 
   bool Changed = false;
-  bool LocalChanged;
-  do {
-    LocalChanged = false;
-
-    VisitStack.clear();
-    Visited.clear();
-
-    VisitStack.push_back(WorklistItem(L->getHeader(), false));
-
-    while (!VisitStack.empty()) {
-      WorklistItem Item = VisitStack.pop_back_val();
-      BasicBlock *BB = Item.getPointer();
-      bool IsSubloopHeader = Item.getInt();
-      const DataLayout &DL = L->getHeader()->getModule()->getDataLayout();
-
-      // Simplify instructions in the current basic block.
-      for (BasicBlock::iterator BI = BB->begin(), BE = BB->end(); BI != BE;) {
-        Instruction *I = &*BI++;
-
-        // The first time through the loop ToSimplify is empty and we try to
-        // simplify all instructions. On later iterations ToSimplify is not
-        // empty and we only bother simplifying instructions that are in it.
-        if (!ToSimplify->empty() && !ToSimplify->count(I))
+  for (;;) {
+    for (BasicBlock *BB : RPOT) {
+      for (Instruction &I : *BB) {
+        if (auto *PI = dyn_cast<PHINode>(&I))
+          VisitedPHIs.insert(PI);
+
+        if (I.use_empty()) {
+          if (isInstructionTriviallyDead(&I, &TLI))
+            DeadInsts.push_back(&I);
           continue;
-
-        // Don't bother simplifying unused instructions.
-        if (!I->use_empty()) {
-          Value *V = SimplifyInstruction(I, {DL, TLI, DT, AC});
-          if (V && LI->replacementPreservesLCSSAForm(I, V)) {
-            // Mark all uses for resimplification next time round the loop.
-            for (User *U : I->users())
-              Next->insert(cast<Instruction>(U));
-
-            I->replaceAllUsesWith(V);
-            LocalChanged = true;
-            ++NumSimplified;
-          }
-        }
-        if (RecursivelyDeleteTriviallyDeadInstructions(I, TLI)) {
-          // RecursivelyDeleteTriviallyDeadInstruction can remove more than one
-          // instruction, so simply incrementing the iterator does not work.
-          // When instructions get deleted re-iterate instead.
-          BI = BB->begin();
-          BE = BB->end();
-          LocalChanged = true;
         }
 
-        if (IsSubloopHeader && !isa<PHINode>(I))
-          break;
-      }
+        // We special case the first iteration which we can detect due to the
+        // empty `ToSimplify` set.
+        bool IsFirstIteration = ToSimplify->empty();
 
-      // Add all successors to the worklist, except for loop exit blocks and the
-      // bodies of subloops. We visit the headers of loops so that we can
-      // process
-      // their phis, but we contract the rest of the subloop body and only
-      // follow
-      // edges leading back to the original loop.
-      for (succ_iterator SI = succ_begin(BB), SE = succ_end(BB); SI != SE;
-           ++SI) {
-        BasicBlock *SuccBB = *SI;
-        if (!Visited.insert(SuccBB).second)
+        if (!IsFirstIteration && !ToSimplify->count(&I))
           continue;
 
-        const Loop *SuccLoop = LI->getLoopFor(SuccBB);
-        if (SuccLoop && SuccLoop->getHeader() == SuccBB &&
-            L->contains(SuccLoop)) {
-          VisitStack.push_back(WorklistItem(SuccBB, true));
-
-          SmallVector<BasicBlock *, 8> SubLoopExitBlocks;
-          SuccLoop->getExitBlocks(SubLoopExitBlocks);
-
-          for (unsigned i = 0; i < SubLoopExitBlocks.size(); ++i) {
-            BasicBlock *ExitBB = SubLoopExitBlocks[i];
-            if (LI->getLoopFor(ExitBB) == L && Visited.insert(ExitBB).second)
-              VisitStack.push_back(WorklistItem(ExitBB, false));
-          }
+        Value *V = SimplifyInstruction(&I, SQ.getWithInstruction(&I));
+        if (!V || !LI.replacementPreservesLCSSAForm(&I, V))
+          continue;
 
+        for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
+             UI != UE;) {
+          Use &U = *UI++;
+          auto *UserI = cast<Instruction>(U.getUser());
+          U.set(V);
+
+          // If the instruction is used by a PHI node we have already processed
+          // we'll need to iterate on the loop body to converge, so add it to
+          // the next set.
+          if (auto *UserPI = dyn_cast<PHINode>(UserI))
+            if (VisitedPHIs.count(UserPI)) {
+              Next->insert(UserPI);
               continue;
             }
 
-        bool IsExitBlock =
-            std::binary_search(ExitBlocks.begin(), ExitBlocks.end(), SuccBB);
-        if (IsExitBlock)
-          continue;
+          // If we are only simplifying targeted instructions and the user is an
+          // instruction in the loop body, add it to our set of targeted
+          // instructions. Because we process defs before uses (outside of PHIs)
+          // we won't have visited it yet.
+          //
+          // We also skip any uses outside of the loop being simplified. Those
+          // should always be PHI nodes due to LCSSA form, and we don't want to
+          // try to simplify those away.
+          assert((L.contains(UserI) || isa<PHINode>(UserI)) &&
+                 "Uses outside the loop should be PHI nodes due to LCSSA!");
+          if (!IsFirstIteration && L.contains(UserI))
+            ToSimplify->insert(UserI);
+        }
 
-        VisitStack.push_back(WorklistItem(SuccBB, false));
+        assert(I.use_empty() && "Should always have replaced all uses!");
+        if (isInstructionTriviallyDead(&I, &TLI))
+          DeadInsts.push_back(&I);
+        ++NumSimplified;
+        Changed = true;
       }
     }
 
-    // Place the list of instructions to simplify on the next loop iteration
-    // into ToSimplify.
-    std::swap(ToSimplify, Next);
-    Next->clear();
+    // Delete any dead instructions found thus far now that we've finished an
+    // iteration over all instructions in all the loop blocks.
+    if (!DeadInsts.empty()) {
+      Changed = true;
+      RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, &TLI);
+    }
+
+    // If we never found a PHI that needs to be simplified in the next
+    // iteration, we're done.
+    if (Next->empty())
+      break;
 
-    Changed |= LocalChanged;
-  } while (LocalChanged);
+    // Otherwise, put the next set in place for the next iteration and reset it
+    // and the visited PHIs for that iteration.
+    std::swap(Next, ToSimplify);
+    Next->clear();
+    VisitedPHIs.clear();
+    DeadInsts.clear();
+  }
 
   return Changed;
 }
 
 namespace {
 
 class LoopInstSimplifyLegacyPass : public LoopPass {
 public:
   static char ID; // Pass ID, replacement for typeid
 
   LoopInstSimplifyLegacyPass() : LoopPass(ID) {
     initializeLoopInstSimplifyLegacyPassPass(*PassRegistry::getPassRegistry());
   }
 
   bool runOnLoop(Loop *L, LPPassManager &LPM) override {
     if (skipLoop(L))
       return false;
-    DominatorTreeWrapperPass *DTWP =
-        getAnalysisIfAvailable<DominatorTreeWrapperPass>();
-    DominatorTree *DT = DTWP ? &DTWP->getDomTree() : nullptr;
-    LoopInfo *LI = &getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
-    AssumptionCache *AC =
-        &getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
+    DominatorTree &DT = getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+    LoopInfo &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
+    AssumptionCache &AC =
+        getAnalysis<AssumptionCacheTracker>().getAssumptionCache(
             *L->getHeader()->getParent());
-    const TargetLibraryInfo *TLI =
-        &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
+    const TargetLibraryInfo &TLI =
+        getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
 
-    return SimplifyLoopInst(L, DT, LI, AC, TLI);
+    return simplifyLoopInst(*L, DT, LI, AC, TLI);
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.addRequired<AssumptionCacheTracker>();
+    AU.addRequired<DominatorTreeWrapperPass>();
     AU.addRequired<TargetLibraryInfoWrapperPass>();
     AU.setPreservesCFG();
     getLoopAnalysisUsage(AU);
   }
 };
 
 } // end anonymous namespace
 
 PreservedAnalyses LoopInstSimplifyPass::run(Loop &L, LoopAnalysisManager &AM,
                                             LoopStandardAnalysisResults &AR,
                                             LPMUpdater &) {
-  if (!SimplifyLoopInst(&L, &AR.DT, &AR.LI, &AR.AC, &AR.TLI))
+  if (!simplifyLoopInst(L, AR.DT, AR.LI, AR.AC, AR.TLI))
     return PreservedAnalyses::all();
 
   auto PA = getLoopPassPreservedAnalyses();
   PA.preserveSet<CFGAnalyses>();
   return PA;
 }
 
 char LoopInstSimplifyLegacyPass::ID = 0;

llvm/trunk/lib/Transforms/Utils/Local.cpp

 llvm::RecursivelyDeleteTriviallyDeadInstructions(Value *V,
                                                  const TargetLibraryInfo *TLI) {
   Instruction *I = dyn_cast<Instruction>(V);
   if (!I || !I->use_empty() || !isInstructionTriviallyDead(I, TLI))
     return false;
 
   SmallVector<Instruction*, 16> DeadInsts;
   DeadInsts.push_back(I);
+  RecursivelyDeleteTriviallyDeadInstructions(DeadInsts, TLI);
 
-  do {
-    I = DeadInsts.pop_back_val();
-    salvageDebugInfo(*I);
+  return true;
+}
+
+void llvm::RecursivelyDeleteTriviallyDeadInstructions(
+    SmallVectorImpl<Instruction *> &DeadInsts, const TargetLibraryInfo *TLI) {
+  // Process the dead instruction list until empty.
+  while (!DeadInsts.empty()) {
+    Instruction &I = *DeadInsts.pop_back_val();
+    assert(I.use_empty() && "Instructions with uses are not dead.");
+    assert(isInstructionTriviallyDead(&I, TLI) &&
+           "Live instruction found in dead worklist!");
+
+    // Don't lose the debug info while deleting the instructions.
+    salvageDebugInfo(I);
 
     // Null out all of the instruction's operands to see if any operand becomes
     // dead as we go.
-    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
-      Value *OpV = I->getOperand(i);
-      I->setOperand(i, nullptr);
+    for (Use &OpU : I.operands()) {
+      Value *OpV = OpU.get();
+      OpU.set(nullptr);
 
-      if (!OpV->use_empty()) continue;
+      if (!OpV->use_empty())
+        continue;
 
       // If the operand is an instruction that became dead as we nulled out the
       // operand, and if it is 'trivially' dead, delete it in a future loop
       // iteration.
       if (Instruction *OpI = dyn_cast<Instruction>(OpV))
         if (isInstructionTriviallyDead(OpI, TLI))
           DeadInsts.push_back(OpI);
     }
 
-    I->eraseFromParent();
-  } while (!DeadInsts.empty());
-
-  return true;
+    I.eraseFromParent();
+  }
 }
 
 /// areAllUsesEqual - Check whether the uses of a value are all the same.
 /// This is similar to Instruction::hasOneUse() except this will also return
 /// true when there are no uses or multiple uses that all refer to the same
 /// value.
 static bool areAllUsesEqual(Instruction *I) {
   Value::user_iterator UI = I->user_begin();

llvm/trunk/test/Transforms/LoopInstSimplify/basic.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S %s -passes=loop-instsimplify | FileCheck %s
+
+; Test very basic folding and propagation occurs within a loop body. This should
+; collapse to the loop iteration structure and the LCSSA PHI node.
+define i32 @test1(i32 %n, i32 %x) {
+; CHECK-LABEL: @test1(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[I_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[I_NEXT]] = add nsw i32 [[I]], 1
+; CHECK-NEXT:    [[I_CMP:%.*]] = icmp slt i32 [[I_NEXT]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[I_CMP]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[X_LCSSA:%.*]] = phi i32 [ [[X:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    ret i32 [[X_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %loop ]
+  %x.add = add nsw i32 %x, 0
+  %x.sub = sub i32 %x.add, 0
+  %x.and = and i32 %x.sub, -1
+  %i.next = add nsw i32 %i, 1
+  %i.cmp = icmp slt i32 %i.next, %n
+  br i1 %i.cmp, label %loop, label %exit
+
+exit:
+  %x.lcssa = phi i32 [ %x.and, %loop ]
+  ret i32 %x.lcssa
+}
+
+; Test basic loop structure that still has a simplification feed a prior PHI.
+define i32 @test2(i32 %n, i32 %x) {
+; CHECK-LABEL: @test2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[I_NEXT:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    [[I_NEXT]] = add nsw i32 [[I]], 1
+; CHECK-NEXT:    [[I_CMP:%.*]] = icmp slt i32 [[I_NEXT]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[I_CMP]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[X_LCSSA:%.*]] = phi i32 [ [[X:%.*]], [[LOOP]] ]
+; CHECK-NEXT:    ret i32 [[X_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %loop ]
+  %x.loop = phi i32 [ %x, %entry ], [ %x.next, %loop ]
+  %x.next = add nsw i32 %x.loop, 0
+  %i.next = add nsw i32 %i, 1
+  %i.cmp = icmp slt i32 %i.next, %n
+  br i1 %i.cmp, label %loop, label %exit
+
+exit:
+  %x.lcssa = phi i32 [ %x.loop, %loop ]
+  ret i32 %x.lcssa
+}
+
+; Test a diamond CFG with inner PHI nodes.
+define i32 @test3(i32 %n, i32 %x) {
+; CHECK-LABEL: @test3(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[I_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
+; CHECK-NEXT:    [[X_CMP:%.*]] = icmp slt i32 [[I]], 42
+; CHECK-NEXT:    br i1 [[X_CMP]], label [[LOOP_LHS:%.*]], label [[LOOP_RHS:%.*]]
+; CHECK:       loop.lhs:
+; CHECK-NEXT:    br label [[LOOP_LATCH]]
+; CHECK:       loop.rhs:
+; CHECK-NEXT:    br label [[LOOP_LATCH]]
+; CHECK:       loop.latch:
+; CHECK-NEXT:    [[I_NEXT]] = add nsw i32 [[I]], 1
+; CHECK-NEXT:    [[I_CMP:%.*]] = icmp slt i32 [[I_NEXT]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[I_CMP]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[X_LCSSA:%.*]] = phi i32 [ [[X:%.*]], [[LOOP_LATCH]] ]
+; CHECK-NEXT:    ret i32 [[X_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %loop.latch ]
+  %x.loop = phi i32 [ %x, %entry ], [ %x.phi, %loop.latch ]
+  %x.add = add nsw i32 %x.loop, 0
+  %x.cmp = icmp slt i32 %i, 42
+  br i1 %x.cmp, label %loop.lhs, label %loop.rhs
+
+loop.lhs:
+  %x.l.add = add nsw i32 %x.add, 0
+  br label %loop.latch
+
+loop.rhs:
+  %x.r.sub = sub nsw i32 %x.add, 0
+  br label %loop.latch
+
+loop.latch:
+  %x.phi = phi i32 [ %x.l.add, %loop.lhs ], [ %x.r.sub, %loop.rhs ]
+  %i.next = add nsw i32 %i, 1
+  %i.cmp = icmp slt i32 %i.next, %n
+  br i1 %i.cmp, label %loop, label %exit
+
+exit:
+  %x.lcssa = phi i32 [ %x.loop, %loop.latch ]
+  ret i32 %x.lcssa
+}
+
+; Test an inner loop that is only simplified when processing the outer loop, and
+; an outer loop only simplified when processing the inner loop.
+define i32 @test4(i32 %n, i32 %m, i32 %x) {
+; CHECK-LABEL: @test4(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    br label [[LOOP:%.*]]
+; CHECK:       loop:
+; CHECK-NEXT:    [[I:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[I_NEXT:%.*]], [[LOOP_LATCH:%.*]] ]
+; CHECK-NEXT:    br label [[LOOP_INNER:%.*]]
+; CHECK:       loop.inner:
+; CHECK-NEXT:    [[J:%.*]] = phi i32 [ 0, [[LOOP]] ], [ [[J_NEXT:%.*]], [[LOOP_INNER]] ]
+; CHECK-NEXT:    [[J_NEXT]] = add nsw i32 [[J]], 1
+; CHECK-NEXT:    [[J_CMP:%.*]] = icmp slt i32 [[J_NEXT]], [[M:%.*]]
+; CHECK-NEXT:    br i1 [[J_CMP]], label [[LOOP_INNER]], label [[LOOP_LATCH]]
+; CHECK:       loop.latch:
+; CHECK-NEXT:    [[I_NEXT]] = add nsw i32 [[I]], 1
+; CHECK-NEXT:    [[I_CMP:%.*]] = icmp slt i32 [[I_NEXT]], [[N:%.*]]
+; CHECK-NEXT:    br i1 [[I_CMP]], label [[LOOP]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[X_LCSSA:%.*]] = phi i32 [ [[X:%.*]], [[LOOP_LATCH]] ]
+; CHECK-NEXT:    ret i32 [[X_LCSSA]]
+;
+entry:
+  br label %loop
+
+loop:
+  %i = phi i32 [ 0, %entry ], [ %i.next, %loop.latch ]
+  %x.loop = phi i32 [ %x, %entry ], [ %x.inner.lcssa, %loop.latch ]
+  %x.add = add nsw i32 %x.loop, 0
+  br label %loop.inner
+
+loop.inner:
+  %j = phi i32 [ 0, %loop ], [ %j.next, %loop.inner ]
+  %x.inner.loop = phi i32 [ %x.add, %loop ], [ %x.inner.add, %loop.inner ]
+  %x.inner.add = add nsw i32 %x.inner.loop, 0
+  %j.next = add nsw i32 %j, 1
+  %j.cmp = icmp slt i32 %j.next, %m
+  br i1 %j.cmp, label %loop.inner, label %loop.latch
+
+loop.latch:
+  %x.inner.lcssa = phi i32 [ %x.inner.loop, %loop.inner ]
+  %i.next = add nsw i32 %i, 1
+  %i.cmp = icmp slt i32 %i.next, %n
+  br i1 %i.cmp, label %loop, label %exit
+
+exit:
+  %x.lcssa = phi i32 [ %x.loop, %loop.latch ]
+  ret i32 %x.lcssa
+}