Add straight-line strength reduction to LLVM

[jingyue]

Summary

Straight-line strength reduction (SLSR) is implemented in GCC but not yet in
LLVM. It has proven to effectively simplify statements derived from an unrolled
loop, and can potentially benefit many other cases too. For example,

LLVM unrolls

#pragma unroll
foo (int i = 0; i < 3; ++i) {
  sum += foo((b + i) * s);
}

into

sum += foo(b * s);
sum += foo((b + 1) * s);
sum += foo((b + 2) * s);

However, no optimizations yet reduce the internal redundancy of the three
expressions:

b * s
(b + 1) * s
(b + 2) * s

With SLSR, LLVM can optimize these three expressions into:

t1 = b * s
t2 = t1 + s
t3 = t2 + s

This commit is only an initial step towards implementing a series of such
optimizations. I will implement more (see TODO in the file commentary) in the
near future. This optimization is enabled for the NVPTX backend for now.
However, I am more than happy to push it to the standard optimization pipeline
after more thorough performance tests.

[jingyue]

FYI, the link to the GCC's implementation of SLSR is https://gcc.gnu.org/viewcvs/gcc/trunk/gcc/gimple-ssa-strength-reduction.c?view=markup.

[hfinkel]

Thanks for working on this! We need to be careful here, however, because:

1. We don't want to turn "free" computations into non-free computations (because we remove the ability to do address-mode folding). LoopStrengthReduce uses part of the TTI interface for dealing with target addressing modes, and we should do that here too.
2. We don't want to lengthen the critical path by unnecessarily decreasing the amount of available ILP. We don't currently have a good interface for asking a target how much ILP is available for simple integer operations (we have one for vectorization interleaving factors, which is similar in spirit, but on several targets with which I work, use a different number than would be appropriate here).

Regarding (2), we could also decide the long chain is the canonical form, and targets should split these late for ILP. I've not given this a lot of thought, and so I'm not sure how easy or hard that might be. Using the MachineCombiner might be a good place to do such splitting.

What do you think?

lib/Transforms/Scalar/StraightLineStrengthReduce.cpp

11

We don't need to mention GCC here.

182

You can remove the "TODO" part of this. There are a lot of things in LLVM that only handle the canonical operand ordering.

210

If (!C.Ins->getParent()) is fine. Also no { } needed.

[jingyue]

Hal, thanks for the comments!

(1) Given the cases I consider so far, I am not aware of any target that can fold (B + i) * S where B and S are variables into an addressing mode. I don't think such form can even fit into LLVM's AddrMode struct. However, once we start considering more cases (e.g., B + i * S), we should definitely check for addressing modes. Does this make sense?

(2) Thanks for pointing this out. I hadn't considered this issue when preparing this diff. I now understand your concern, but don't have a good solution yet. I'll investigate how LSR handles this issue and how MachineCombiner works to get a better picture. Any suggestions from other folks are appreciated!

Jingyue

[atrick]

Mainly, I would like to reiterate both of Hal's very good points above.

Out of curiosity, can you explain what sort of real-world source code results in this IR pattern?

You should keep in mind, that this level of optimization is currently the job of LSR. If you're doing this within a loop the IVChain part of LSR will try to do something similar.

I only reviewed it quickly, but your implementation looks good. Thank you for putting a bound on the number of candidate checks. I hate finding those quadratic behaviors later when compile time blows up.

LGTM given that it's not enabled in the standard pipeline yet.

[jingyue]

Hi Andrew,

The code example in my file commentary

#pragma unroll
foo (int i = 0; i < 3; ++i) {
  sum += foo((b + i) * s);
}

pretty much reflects real-world source code. More often, I see (b + i) * s used as an array index which creates the same issue.

LSR cannot yet solve the problem I'm facing here due to phase ordering. Loop unrolling, which happens way before LSR, destroys the loop structure. I agree with you that, if the loop is not unrolled, LSR should take care of this pattern.

[meheff]

LGTM

lib/Transforms/Scalar/StraightLineStrengthReduce.cpp

166

Since there can be more than one candidate would there ever be any advantage to looking deeper in the list? Some bases could be cheaper (eg, where i' - i == 1 and you don't need a multiply). However, I'd guess in the common unrolled case the latest candidate on the list is best. On the other hand though, from an ILP perspective the earliest candidate might be best.

[jingyue]

Disable SLSR if the target schedules for ILP.

Also, added one more test where the bump is a multiple of S.

[atrick]

I don't like piggybacking on the scheduling mode. When a target selects the scheduler it doesn't expect to get different IR out of the optimizer. Don't surprise users.

It would be fine to add a separate hook for this pass. But I don't think it's an issue until the pass goes into the standard pipeline.

[jingyue]

lib/Transforms/Scalar/StraightLineStrengthReduce.cpp

11

done

166

We may want to improve the strategy in the future. The patterns I have seen so far are derived from loop unrolling, and in those cases choosing the latest candidate very likely produces the best code. Also note that in the cases where the value grows by a multiple of the stride, e.g.,

b * s
(b + 2) * s
(b + 4) * s

choosing the latest candidate leads to a uniform bump (2 * s in this example) for every iteration, and this bump can be GVN'ed later.

182

done

210

done

[jingyue]

stop piggybacking scheduling info

[jingyue]

Andrew, thanks for pointing this out! I reverted my change on using scheduling info, and left a TODO in the file comment.

[atrick]

lgtm

[jholewinski]

LGTM! Thanks for working on this!

[jingyue]

Minor changes, NFC:
Merge from upstream
Do not put usernames in TODO

[jingyue]

Hal, do you think this patch is good to go, given we don't push it to the standard pipeline yet? I left a TODO on the ILP issue you raised. I like your point a lot, but I feel I should do it in a separate diff.

[hfinkel]

In D7310#117774, @jingyue wrote:

Hal, do you think this patch is good to go, given we don't push it to the standard pipeline yet? I left a TODO on the ILP issue you raised. I like your point a lot, but I feel I should do it in a separate diff.

Yes, that's fine with me. Let's iterate in-tree. Please proceed.

include/llvm/InitializePasses.h

 void initializeSingleLoopExtractorPass(PassRegistry&);
 void initializeSinkingPass(PassRegistry&);
 void initializeSeparateConstOffsetFromGEPPass(PassRegistry &);
 void initializeSlotIndexesPass(PassRegistry&);
 void initializeSpillPlacementPass(PassRegistry&);
 void initializeStackProtectorPass(PassRegistry&);
 void initializeStackColoringPass(PassRegistry&);
 void initializeStackSlotColoringPass(PassRegistry&);
+void initializeStraightLineStrengthReducePass(PassRegistry &);
 void initializeStripDeadDebugInfoPass(PassRegistry&);
 void initializeStripDeadPrototypesPassPass(PassRegistry&);
 void initializeStripDebugDeclarePass(PassRegistry&);
 void initializeStripNonDebugSymbolsPass(PassRegistry&);
 void initializeStripSymbolsPass(PassRegistry&);
 void initializeTailCallElimPass(PassRegistry&);
 void initializeTailDuplicatePassPass(PassRegistry&);
 void initializeTargetPassConfigPass(PassRegistry&);

include/llvm/LinkAllPasses.h

ForcePassLinking() {
       (void) llvm::createInstructionSimplifierPass();
       (void) llvm::createLoopVectorizePass();
       (void) llvm::createSLPVectorizerPass();
       (void) llvm::createBBVectorizePass();
       (void) llvm::createPartiallyInlineLibCallsPass();
       (void) llvm::createScalarizerPass();
       (void) llvm::createSeparateConstOffsetFromGEPPass();
       (void) llvm::createRewriteSymbolsPass();
+      (void) llvm::createStraightLineStrengthReducePass();
 
       (void)new llvm::IntervalPartition();
       (void)new llvm::ScalarEvolution();
       ((llvm::Function*)nullptr)->viewCFGOnly();
       llvm::RGPassManager RGM;
       ((llvm::RegionPass*)nullptr)->runOnRegion((llvm::Region*)nullptr, RGM);
       llvm::AliasSetTracker X(*(llvm::AliasAnalysis*)nullptr);
       X.add(nullptr, 0, llvm::AAMDNodes()); // for -print-alias-sets
     }
   } ForcePassLinking; // Force link by creating a global definition.
 }
 
 #endif

include/llvm/Transforms/Scalar.h

createSeparateConstOffsetFromGEPPass(const TargetMachine *TM = nullptr,
                                      bool LowerGEP = false);
 
 //===----------------------------------------------------------------------===//
 //
 // LoadCombine - Combine loads into bigger loads.
 //
 BasicBlockPass *createLoadCombinePass();
 
+FunctionPass *createStraightLineStrengthReducePass();
+
 } // End llvm namespace
 
 #endif

lib/Target/NVPTX/NVPTXTargetMachine.cpp

void NVPTXPassConfig::addIRPasses() {
   disablePass(&BranchFolderPassID);
   disablePass(&TailDuplicateID);
 
   addPass(createNVPTXImageOptimizerPass());
   TargetPassConfig::addIRPasses();
   addPass(createNVPTXAssignValidGlobalNamesPass());
   addPass(createGenericToNVVMPass());
   addPass(createNVPTXFavorNonGenericAddrSpacesPass());
+  addPass(createStraightLineStrengthReducePass());
   addPass(createSeparateConstOffsetFromGEPPass());
   // The SeparateConstOffsetFromGEP pass creates variadic bases that can be used
   // by multiple GEPs. Run GVN or EarlyCSE to really reuse them. GVN generates
   // significantly better code than EarlyCSE for some of our benchmarks.
   if (getOptLevel() == CodeGenOpt::Aggressive)
     addPass(createGVNPass());
   else
     addPass(createEarlyCSEPass());

lib/Transforms/Scalar/CMakeLists.txt

add_llvm_library(LLVMScalarOpts
   SROA.cpp
   SampleProfile.cpp
   Scalar.cpp
   ScalarReplAggregates.cpp
   Scalarizer.cpp
   SeparateConstOffsetFromGEP.cpp
   SimplifyCFGPass.cpp
   Sink.cpp
+  StraightLineStrengthReduce.cpp
   StructurizeCFG.cpp
   TailRecursionElimination.cpp
   )
 
 add_dependencies(LLVMScalarOpts intrinsics_gen)

lib/Transforms/Scalar/Scalar.cpp

void llvm::initializeScalarOpts(PassRegistry &Registry) {
   initializeSROAPass(Registry);
   initializeSROA_DTPass(Registry);
   initializeSROA_SSAUpPass(Registry);
   initializeCFGSimplifyPassPass(Registry);
   initializeStructurizeCFGPass(Registry);
   initializeSinkingPass(Registry);
   initializeTailCallElimPass(Registry);
   initializeSeparateConstOffsetFromGEPPass(Registry);
+  initializeStraightLineStrengthReducePass(Registry);
   initializeLoadCombinePass(Registry);
 }
 
 void LLVMInitializeScalarOpts(LLVMPassRegistryRef R) {
   initializeScalarOpts(*unwrap(R));
 }
 
 void LLVMAddAggressiveDCEPass(LLVMPassManagerRef PM) {

lib/Transforms/Scalar/StraightLineStrengthReduce.cpp

+//===-- StraightLineStrengthReduce.cpp - ------------------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements straight-line strength reduction (SLSR), which already
+// exists in GCC. Unlike loop strength reduction, this algorithm is designed to
+// reduce arithmetic redundancy in straight-line code instead of loops. It has
+// proven to be effective in simplifying arithmetic statements derived from an
+// unrolled loop. It can also simplify the logic of SeparateConstOffsetFromGEP.
+//
+// There are many optimizations we can perform in the domain of SLSR. This file
+// for now contains only an initial step. Specifically, we look for strength
+// reduction candidate in the form of
+//
+// (B + i) * S
+//
+// where B and S are integer constants or variables, and i is a constant
+// integer. If we found two such candidates
+//
+// S1: X = (B + i) * S S2: Y = (B + i') * S
+//
+// and S1 dominates S2, we call S1 a basis of S2, and can replace S2 with
+//
+// Y = X + (i' - i) * S
+//
+// where (i' - i) * S is folded to the extent possible. When S2 has multiple
+// bases, we pick the one that is closest to S2, or S2's "immediate" basis.
+//
+// TODO(jingyue):
+//
+// - Handle candidates in the form of B + i * S
+//
+// - Handle candidates in the form of pointer arithmetics. e.g., B[i * S]
+//
+// - Floating point arithmetics when fast math is enabled.
+#include <vector>
+
+#include "llvm/ADT/DenseSet.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Scalar.h"
+
+using namespace llvm;
+using namespace PatternMatch;
+
+namespace {
+
+class StraightLineStrengthReduce : public FunctionPass {
+ public:
+  // SLSR candidate. Such a candidate must be in the form of
+  //   (Base + Index) * Stride
+  struct Candidate : public ilist_node<Candidate> {
+    Candidate(Value *B = nullptr, ConstantInt *Idx = nullptr,
+              Value *S = nullptr, Instruction *I = nullptr)
+        : Base(B), Index(Idx), Stride(S), Ins(I), Basis(nullptr) {}
+    Value *Base;
+    ConstantInt *Index;
+    Value *Stride;
+    // The instruction this candidate corresponds to. It helps us to rewrite a
+    // candidate with respect to its immediate basis. Note that one instruction
+    // can corresponds to multiple candidates depending on how you associate the
+    // expression. For instance,
+    //
+    // (a + 1) * (b + 2)
+    //
+    // can be treated as
+    //
+    // <Base: a, Index: 1, Stride: b + 2>
+    //
+    // or
+    //
+    // <Base: b, Index: 2, Stride: a + 1>
+    Instruction *Ins;
+    // Points to the immediate basis of this candidate, or nullptr if we cannot
+    // find any basis for this candidate.
+    Candidate *Basis;
+  };
+
+  static char ID;
+
+  StraightLineStrengthReduce(): FunctionPass(ID), DT(nullptr) {
+    initializeStraightLineStrengthReducePass(*PassRegistry::getPassRegistry());
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<DominatorTreeWrapperPass>();
+    // We do not modify the shape of the CFG.
+    AU.setPreservesCFG();
+  }
+
+  bool runOnFunction(Function &F) override;
+
+ private:
+  // Returns true if Basis is a basis for C, i.e., Basis dominates C and they
+  // share the same base and stride.
+  bool isBasisFor(const Candidate &Basis, const Candidate &C);
+  // Checks whether I is in a candidate form. If so, adds all the matching forms
+  // to Candidates, and tries to find the immediate basis for each of them.
+  void allocateCandidateAndFindBasis(Instruction *I);
+  // Given that I is in the form of "(B + Idx) * S", adds this form to
+  // Candidates, and finds its immediate basis.
+  void allocateCandidateAndFindBasis(Value *B, ConstantInt *Idx, Value *S,
+                                     Instruction *I);
+  // Rewrites candidate C with respect to Basis.
+  void rewriteCandidateWithBasis(const Candidate &C, const Candidate &Basis);
+
+  DominatorTree *DT;
+  ilist<Candidate> Candidates;
+  // Temporarily holds all instructions that are unlinked (but not deleted) by
+  // rewriteCandidateWithBasis. These instructions will be actually removed
+  // after all rewriting finishes.
+  DenseSet<Instruction *> UnlinkedInstructions;
+};
+}  // anonymous namespace
+
+char StraightLineStrengthReduce::ID = 0;
+INITIALIZE_PASS_BEGIN(StraightLineStrengthReduce, "slsr",
+                      "Straight line strength reduction", false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
+INITIALIZE_PASS_END(StraightLineStrengthReduce, "slsr",
+                    "Straight line strength reduction", false, false)
+
+FunctionPass *llvm::createStraightLineStrengthReducePass() {
+  return new StraightLineStrengthReduce();
+}
+
+bool StraightLineStrengthReduce::isBasisFor(const Candidate &Basis,
+                                            const Candidate &C) {
+  return (Basis.Ins != C.Ins && // skip the same instruction
+          // Basis must dominate C in order to rewrite C with respect to Basis.
+          DT->dominates(Basis.Ins->getParent(), C.Ins->getParent()) &&
+          // They share the same base and stride.
+          Basis.Base == C.Base &&
+          Basis.Stride == C.Stride);
+}
+
+// TODO(jingyue): We currently implement an algorithm whose time complexity is
+// linear to the number of existing candidates. However, a better algorithm
+// exists. We could depth-first search the dominator tree, and maintain a hash
+// table that contains all candidates that dominate the node being traversed.
+// This hash table is indexed by the base and the stride of a candidate.
+// Therefore, finding the immediate basis of a candidate boils down to one
+// hash-table look up.
+void StraightLineStrengthReduce::allocateCandidateAndFindBasis(Value *B,
+                                                               ConstantInt *Idx,
+                                                               Value *S,
+                                                               Instruction *I) {
+  Candidate C(B, Idx, S, I);
+  // Try to compute the immediate basis of C.
+  unsigned NumIterations = 0;
+  // Limit the scan radius to avoid running forever.
+  static const int MaxNumIterations = 50;
+  for (auto Basis = Candidates.rbegin();
+       Basis != Candidates.rend() && NumIterations < MaxNumIterations;
+       ++Basis, ++NumIterations) {
+    if (isBasisFor(*Basis, C)) {
+      C.Basis = &(*Basis);
+      break;
+    }
+  }
+  // Regardless of whether we find a basis for C, we need to push C to the
+  // candidate list.
+  Candidates.push_back(C);
+}
+
+void StraightLineStrengthReduce::allocateCandidateAndFindBasis(Instruction *I) {
+  Value *B = nullptr;
+  ConstantInt *Idx = nullptr;
+  // "(Base + Index) * Stride" must be a Mul instruction at the first hand.
+  if (I->getOpcode() == Instruction::Mul) {
+    if (IntegerType *ITy = dyn_cast<IntegerType>(I->getType())) {
+      Value *LHS = I->getOperand(0), *RHS = I->getOperand(1);
+      for (unsigned Swapped = 0; Swapped < 2; ++Swapped) {
+        // TODO(jingyue): The reassociation pass places the constant in an
+        // addition at the right hand side. Therefore, we most likely see (a +
+        // 5) rather than (5 + a). If the other case turns out important, we
+        // should handle that too.
+        if (match(LHS, m_Add(m_Value(B), m_ConstantInt(Idx)))) {
+          // If LHS is in the form of "Base + Index", then I is in the form of
+          // "(Base + Index) * RHS".
+          allocateCandidateAndFindBasis(B, Idx, RHS, I);
+        } else {
+          // Otherwise, at least try the form (LHS + 0) * RHS.
+          allocateCandidateAndFindBasis(LHS, ConstantInt::get(ITy, 0), RHS, I);
+        }
+        // Swap LHS and RHS so that we also cover the cases where LHS is the
+        // stride.
+        if (LHS == RHS)
+          break;
+        std::swap(LHS, RHS);
+      }
+    }
+  }
+}
+
+void StraightLineStrengthReduce::rewriteCandidateWithBasis(
+    const Candidate &C, const Candidate &Basis) {
+  // An instruction can correspond to multiple candidates. Therefore, instead of
+  // simply deleting an instruction when we rewrite it, we mark its parent as
+  // nullptr (i.e. unlink it) so that we can skip the candidates whose
+  // instruction is already rewritten.
+  if (C.Ins->getParent() == nullptr) {
+    return;
+  }
+  assert(C.Base == Basis.Base && C.Stride == Basis.Stride);
+  // Basis = (B + i) * S
+  // C     = (B + i') * S
+  //   ==>
+  // C     = Basis + (i' - i) * S
+  IRBuilder<> Builder(C.Ins);
+  ConstantInt *IndexOffset = ConstantInt::get(
+      C.Ins->getContext(), C.Index->getValue() - Basis.Index->getValue());
+  Value *Reduced;
+  // TODO: preserve nsw/nuw in some cases.
+  if (IndexOffset->isOne()) {
+    // If (i' - i) is 1, fold C into Basis + S.
+    Reduced = Builder.CreateAdd(Basis.Ins, C.Stride);
+  } else if (IndexOffset->isMinusOne()) {
+    // If (i' - i) is -1, fold C into Basis - S.
+    Reduced = Builder.CreateSub(Basis.Ins, C.Stride);
+  } else {
+    Value *Bump = Builder.CreateMul(C.Stride, IndexOffset);
+    Reduced = Builder.CreateAdd(Basis.Ins, Bump);
+  }
+  Reduced->takeName(C.Ins);
+  C.Ins->replaceAllUsesWith(Reduced);
+  C.Ins->dropAllReferences();
+  // Unlink C.Ins so that we can skip other candidates also corresponding to
+  // C.Ins. The actual deletion is postponed to the end of runOnFunction.
+  C.Ins->removeFromParent();
+  UnlinkedInstructions.insert(C.Ins);
+}
+
+bool StraightLineStrengthReduce::runOnFunction(Function &F) {
+  if (skipOptnoneFunction(F))
+    return false;
+
+  DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+  // Traverse the dominator tree in the depth-first order. This order makes sure
+  // all bases of a candidate are in Candidates when we process it.
+  for (auto node = GraphTraits<DominatorTree *>::nodes_begin(DT);
+       node != GraphTraits<DominatorTree *>::nodes_end(DT); ++node) {
+    BasicBlock *B = node->getBlock();
+    for (auto I = B->begin(); I != B->end(); ++I) {
+      allocateCandidateAndFindBasis(I);
+    }
+  }
+
+  // Rewrite candidates in the reverse depth-first order. This order makes sure
+  // a candidate being rewritten is not a basis for any other candidate.
+  while (!Candidates.empty()) {
+    const Candidate &C = Candidates.back();
+    if (C.Basis != nullptr) {
+      rewriteCandidateWithBasis(C, *C.Basis);
+    }
+    Candidates.pop_back();
+  }
+
+  // Delete all unlink instructions.
+  for (auto I : UnlinkedInstructions) {
+    delete I;
+  }
+  bool Ret = !UnlinkedInstructions.empty();
+  UnlinkedInstructions.clear();
+  return Ret;
+}

test/Transforms/StraightLineStrengthReduce/slsr.ll

+; RUN: opt < %s -slsr -gvn -S | FileCheck %s
+
+declare i32 @foo(i32 %a)
+
+define i32 @slsr1(i32 %b, i32 %s) {
+; CHECK-LABEL: @slsr1(
+  ; v0 = foo(b * s);
+  %mul0 = mul i32 %b, %s
+; CHECK: mul i32
+; CHECK-NOT: mul i32
+  %v0 = call i32 @foo(i32 %mul0)
+
+  ; v1 = foo((b + 1) * s);
+  %b1 = add i32 %b, 1
+  %mul1 = mul i32 %b1, %s
+  %v1 = call i32 @foo(i32 %mul1)
+
+  ; v2 = foo((b + 2) * s);
+  %b2 = add i32 %b, 2
+  %mul2 = mul i32 %b2, %s
+  %v2 = call i32 @foo(i32 %mul2)
+
+  ; return v0 + v1 + v2;
+  %1 = add i32 %v0, %v1
+  %2 = add i32 %1, %v2
+  ret i32 %2
+}
+
+; v0 = foo(a * b)
+; v1 = foo((a + 1) * b)
+; v2 = foo(a * (b + 1))
+; v3 = foo((a + 1) * (b + 1))
+define i32 @slsr2(i32 %a, i32 %b) {
+; CHECK-LABEL: @slsr2(
+  %a1 = add i32 %a, 1
+  %b1 = add i32 %b, 1
+  %mul0 = mul i32 %a, %b
+; CHECK: mul i32
+; CHECK-NOT: mul i32
+  %mul1 = mul i32 %a1, %b
+  %mul2 = mul i32 %a, %b1
+  %mul3 = mul i32 %a1, %b1
+
+  %v0 = call i32 @foo(i32 %mul0)
+  %v1 = call i32 @foo(i32 %mul1)
+  %v2 = call i32 @foo(i32 %mul2)
+  %v3 = call i32 @foo(i32 %mul3)
+
+  %1 = add i32 %v0, %v1
+  %2 = add i32 %1, %v2
+  %3 = add i32 %2, %v3
+  ret i32 %3
+}
+
+; Do not rewrite a candidate if its potential basis does not dominate it.
+; v0 = 0;
+; if (cond)
+;   v0 = foo(a * b);
+; v1 = foo((a + 1) * b);
+; return v0 + v1;
+define i32 @not_dominate(i1 %cond, i32 %a, i32 %b) {
+; CHECK-LABEL: @not_dominate(
+entry:
+  %a1 = add i32 %a, 1
+  br i1 %cond, label %then, label %merge
+
+then:
+  %mul0 = mul i32 %a, %b
+; CHECK: %mul0 = mul i32 %a, %b
+  %v0 = call i32 @foo(i32 %mul0)
+  br label %merge
+
+merge:
+  %v0.phi = phi i32 [ 0, %entry ], [ %mul0, %then ]
+  %mul1 = mul i32 %a1, %b
+; CHECK: %mul1 = mul i32 %a1, %b
+  %v1 = call i32 @foo(i32 %mul1)
+  %sum = add i32 %v0.phi, %v1
+  ret i32 %sum
+}