Add CallSiteSplitting pass

[junbuml]

Summary

This change add a pass which tries to split a call-site to pass
more constrained arguments if its argument is predicated in the control flow
so that we can expose better context to the later passes (e.g, inliner, jump
threading, or IPA-CP based function cloning, etc.).
As of now we support two cases :

1. If a call site is dominated by an OR condition and if any of its arguments

are predicated on this OR condition, try to split the condition with more
constrained arguments. For example, in the code below, we try to split the
call site since we can predicate the argument (ptr) based on the OR condition.

Split from :

if (!ptr || c)
  callee(ptr);

to :

if (!ptr)
  callee(null ptr)  // set the known constant value
else if (c)
  callee(nonnull ptr)  // set non-null attribute in the argument

2. We can also split a call-site based on constant incoming values of a PHI

For example,
from :

BB0:
 %c = icmp eq i32 %i1, %i2
 br i1 %c, label %BB2, label %BB1
BB1:
 br label %BB2
BB2:
 %p = phi i32 [ 0, %BB0 ], [ 1, %BB1 ]
 call void @bar(i32 %p)

to

BB0:
 %c = icmp eq i32 %i1, %i2
 br i1 %c, label %BB2-split0, label %BB1
BB1:
 br label %BB2-split1
BB2-split0:
 call void @bar(i32 0)
 br label %BB2
BB2-split1:
 call void @bar(i32 1)
 br label %BB2
BB2:
 %p = phi i32 [ 0, %BB2-split0 ], [ 1, %BB2-split1 ]

Enabled this for O3 and LTO. I didn't see any significant code size increase in my spec2000/2006/2017 tests on aarch64. Observed 20% performance gain in spec2017/gcc without regressions in other benchmarks.

I've added only two simple tests to demonstrate this pass, but I will add more
tests covering what this patch doing if the overall approach is reasonable.

[mcrosier]

Graham,
This is the inlining work we discussed at the dev meeting. Would you mind adding the Manchester developer that work on the related patch internally?

[davidxl]

lib/Transforms/IPO/PassManagerBuilder.cpp

485

should this be moved after CFG simplification pass ? The latter may undo the splitting.

lib/Transforms/Scalar/CallSiteSplitting.cpp

170

Document this method -- describe parameters etc.

322

move the check to the caller.

327

move the check to caller.

366

if splitCallSite returns false, how can CallInst1 be non-null?

376

nit: shorten the name to just 'doCallSiteSplitting'

378

use range for loop: for (BasicBlock &BB: F)

380

use range loop.

383

Skip intrinsicInst. Or if there are folding opportunities, add test cases to cover it.

test/Transforms/CallSiteSplitting/callsite-split.ll

2

add test for new PM.

[junbuml]

lib/Transforms/IPO/PassManagerBuilder.cpp

485

This patch will split a call-site into two call-sites with different arguments. So, I doubt if SimplifyCFG can merge it back. Moreover, if SimpliyCFG cannot merge this back, then don't you think we need to somehow merge the call-site split in this pass back to original in later pass in case when not inlined?

Regarding the pass order, I think we need to place this pass even before constant propagation to allow more propagation.

[junbuml]

Move this pass before IPSCCP and addressed David's comments.

[junbuml]

lib/Transforms/Scalar/CallSiteSplitting.cpp

366

CallInst1 or CallInst2 could be created when detecting constrained arguments from an OR condition in createCallSitesOnOrPredicatedArgument(), but SplitBlockPredecessors() not always guarantee splitting predecessors. In case where we cannot split, we should clean up CallInst1 or CallInst2 if created already.

380

I don't think we can use range because we modify the IR (remove the call instruction and add blocks) over the iterations.

383

Added check for IntrinsicInst and InstructionTriviallyDead. Can you give me little more detail about folding opportunities ?

[fhahn]

I'll have a closer look later.

For now, CC @sdesmalen , he worked on a similar pass.

[mcrosier]

lib/Transforms/Scalar/CallSiteSplitting.cpp

26

I believe this should be:

//  if (!ptr)
//    callee(null)         // set the known constant value
//  else if (c)
//    callee(nonnull ptr)  // set non-null attribute in the argument

114

assert(PBI->isConditional() && "Expected predecessor block to end in a conditional branch.");

276

You might consider skipping Constant arguments before checking for the NonNull attribute.

303

Improve indent:

if (match(Cond, m_ICmp(Pred, m_Value(), m_Constant())))
  return;
ICmpInst *Cmp = cast<ICmpInst>(Cond);
if (isCondRelevantToAnyCallArgument(Cmp, CS))
  if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)
    BranchInsts.push_back(PBI);

315

BasicBlock *Parent = Instr->getParent();
if (Instr != (Parent->getFirstNonPHI()))
  return false;

for (BasicBlock::iterator BI = Parent->begin(), BE = Parent->end();

426

I think you can just use a reference to BB:

BasicBlock &BB = *BI++;
for (BasicBlock::iterator II = BB.begin(), IE = BB.end(); II != IE;) {

480

You can eliminate the Changed boolean by:

if (!doCallSiteSplitting(F, TLI))
  return PreservedAnalyses::all();

[mcrosier]

In D39137#907663, @fhahn wrote:

I'll have a closer look later.

For now, CC @sdesmalen , he worked on a similar pass.

Thanks, Florian.

[junbuml]

Addressed Chad's comments.
Thanks for the review.

[fhahn]

lib/Passes/PassBuilder.cpp

553

Is there any reason you use Level == 3 here and Level > 2 in lib/Transforms/IPO/PassManagerBuilder.cpp?

lib/Transforms/Scalar/CallSiteSplitting.cpp

78

ConstValue seems unused?

83

A range-based-for loop might make things easier here.

116

Would it make sense to add an assertion making sure PBI->getCondition() is an ICmpInst? I assume it could be a FCmpInst as well, if the caller passes a wrong BranchInsts. Same forthe cast to Constant later on.

117

I think Op0 and Op1 could do with more expressive names. Op0 is the argument that gets replaced by the constant Op1, right?

145

You could use continue here to get rid of the else { and reduce the nest in the loop.

245

range based for loop?

252

We have the same loop at 3 points now. Would it make sense to add a helper function for that?

297

just !HeaderBB?

308

maybe have this earlier as a cheap check before we do the more expensive checks?

317

Unneeded braces

320

range based for loop?

353

!Header?

398

Unneeded braces?

[fhahn]

test/Transforms/CallSiteSplitting/callsite-split.ll

2

Do we really need to run -inline -instcombine and -jump-threading? Wouldn't it be enough to just declare your callee functions and then just check if it has been split properly and the new call sites have the correct arguments?

Also, it might be a good idea to add a few negative tests, e.g. conditions with non-constants and so on

[junbuml]

test/Transforms/CallSiteSplitting/callsite-split.ll

2

I totally agree with this. I will add tests simply shows how we split and many negative cases as well.
The intention of this test was to demonstrate the motivation case we observed.

[junbuml]

Fixed Florian's comments. I will add more tests soon.
Thanks for the review.

lib/Passes/PassBuilder.cpp

553

Level in lib/Passes/PassBuilder.cpp is an enum defined :
enum OptimizationLevel {

  O0,
  O1,
  O2,
  O3,
  Os,
  Oz
};

while, OptLevel in lib/Transforms/IPO/PassManagerBuilder.cpp is defined :

/// The Optimization Level - Specify the basic optimization level.
///    0 = -O0, 1 = -O1, 2 = -O2, 3 = -O3
unsigned OptLevel;

lib/Transforms/Scalar/CallSiteSplitting.cpp

252

Changed to a range-based loop, but didn't add helper function because the purpose of each loop slightly different: first one adds an attribute, and the second one set argument value for one call-site, and the third one set arguments for two call-sites. IMHO, iterating for each case doesn't make a big confusion.

308

Agree.

[junbuml]

Added a test file to show simple cases covered by this pass.

[junbuml]

Hi David,
Currently I placed this pass before IPSCCP, so ran performance tests (spec2000/2006/2017/llvm-test-suite) on aarch64. No performance / size regression was observed. Still keep 20% performance gain in spec2017/gcc. Please let me know if you are okay with this pass order.

Thanks,
Jun

[davidxl]

lib/Transforms/Scalar/CallSiteSplitting.cpp

104

Takend --> Taken

409

Assert CallIInst1 and CallInst2 are null here before return false.

411

It seems better to not to speculatively clone the call instructions and delete them later when not needed. The analysis function should instead return 'clone' records indicating how to do the cloning during the transformation phase.

[junbuml]

Addressed David's comments. Please let me know any comment.
Thanks for the review.

lib/Transforms/Scalar/CallSiteSplitting.cpp

411

Added canSplitCallSite() and checked if the call-site can be split first. So, we do not speculatively clone call instructions. The new call-sites cloned should be split in splitCallSite() without any trouble when splitting preds (added assert for this). Now, I changed splitCallSite() to return void.

[junbuml]

Rebased and added debug messages.

[davidxl]

lgtm

[fhahn]

Great, thanks! lgtm too.

[junbuml]

Thanks for the review.

[thakis]

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

370 ↗

(On Diff #121523)

Was this break supposed to be in the if above? As-is, it unconditionally breaks after the first iteration of the loop.

But even if it was a line further up, we'd only check the first Phi in the parent and check if it's an arg. I don't know if that's enough.

[fhahn]

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

370 ↗

(On Diff #121523)

It looks like the code indeed bails out too early. I think we should check all PHIs. I've put up D77089. It should not matter if there are unrelated PHIs in between, but maybe the goal was to be extremely conservative @junbuml ?

include/llvm/InitializePasses.h

 void initializeBasicAAWrapperPassPass(PassRegistry&);
 void initializeBlockExtractorPassPass(PassRegistry&);
 void initializeBlockFrequencyInfoWrapperPassPass(PassRegistry&);
 void initializeBoundsCheckingPass(PassRegistry&);
 void initializeBranchFolderPassPass(PassRegistry&);
 void initializeBranchProbabilityInfoWrapperPassPass(PassRegistry&);
 void initializeBranchRelaxationPass(PassRegistry&);
 void initializeBreakCriticalEdgesPass(PassRegistry&);
+void initializeCallSiteSplittingLegacyPassPass(PassRegistry&);
 void initializeCFGOnlyPrinterLegacyPassPass(PassRegistry&);
 void initializeCFGOnlyViewerLegacyPassPass(PassRegistry&);
 void initializeCFGPrinterLegacyPassPass(PassRegistry&);
 void initializeCFGSimplifyPassPass(PassRegistry&);
 void initializeCFGViewerLegacyPassPass(PassRegistry&);
 void initializeCFLAndersAAWrapperPassPass(PassRegistry&);
 void initializeCFLSteensAAWrapperPassPass(PassRegistry&);
 void initializeCallGraphDOTPrinterPass(PassRegistry&);

include/llvm/Transforms/Scalar.h

 
 //===----------------------------------------------------------------------===//
 //
 // DeadStoreElimination - This pass deletes stores that are post-dominated by
 // must-aliased stores and are not loaded used between the stores.
 //
 FunctionPass *createDeadStoreEliminationPass();
 
+
+//===----------------------------------------------------------------------===//
+//
+// CallSiteSplitting - This pass split call-site based on its known argument
+// values.
+FunctionPass *createCallSiteSplittingPass();
+
+
 //===----------------------------------------------------------------------===//
 //
 // AggressiveDCE - This pass uses the SSA based Aggressive DCE algorithm.  This
 // algorithm assumes instructions are dead until proven otherwise, which makes
 // it more successful are removing non-obviously dead instructions.
 //
 FunctionPass *createAggressiveDCEPass();
 

include/llvm/Transforms/Scalar/CallSiteSplitting.h

+//===- CallSiteSplitting..h - Callsite Splitting ------------*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TRANSFORMS_SCALAR_CALLSITESPLITTING__H
+#define LLVM_TRANSFORMS_SCALAR_CALLSITESPLITTING__H
+
+#include "llvm/ADT/SetVector.h"
+#include "llvm/Analysis/AssumptionCache.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/PassManager.h"
+#include "llvm/Support/Compiler.h"
+#include <vector>
+
+namespace llvm {
+
+struct CallSiteSplittingPass : PassInfoMixin<CallSiteSplittingPass> {
+  /// \brief Run the pass over the function.
+  PreservedAnalyses run(Function &F, FunctionAnalysisManager &AM);
+};
+} // end namespace llvm
+
+#endif // LLVM_TRANSFORMS_SCALAR_CALLSITESPLITTING__H

lib/Passes/PassBuilder.cpp

 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
 #include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/InstrProfiling.h"
 #include "llvm/Transforms/PGOInstrumentation.h"
 #include "llvm/Transforms/SampleProfile.h"
 #include "llvm/Transforms/Scalar/ADCE.h"
 #include "llvm/Transforms/Scalar/AlignmentFromAssumptions.h"
 #include "llvm/Transforms/Scalar/BDCE.h"
+#include "llvm/Transforms/Scalar/CallSiteSplitting.h"
 #include "llvm/Transforms/Scalar/ConstantHoisting.h"
 #include "llvm/Transforms/Scalar/CorrelatedValuePropagation.h"
 #include "llvm/Transforms/Scalar/DCE.h"
 #include "llvm/Transforms/Scalar/DeadStoreElimination.h"
 #include "llvm/Transforms/Scalar/DivRemPairs.h"
 #include "llvm/Transforms/Scalar/EarlyCSE.h"
 #include "llvm/Transforms/Scalar/Float2Int.h"
 #include "llvm/Transforms/Scalar/GVN.h"
PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
 
   // Create an early function pass manager to cleanup the output of the
   // frontend.
   FunctionPassManager EarlyFPM(DebugLogging);
   EarlyFPM.addPass(SimplifyCFGPass());
   EarlyFPM.addPass(SROA());
   EarlyFPM.addPass(EarlyCSEPass());
   EarlyFPM.addPass(LowerExpectIntrinsicPass());
+  if (Level == O3)
+    EarlyFPM.addPass(CallSiteSplittingPass());
+
   // In SamplePGO ThinLTO backend, we need instcombine before profile annotation
   // to convert bitcast to direct calls so that they can be inlined during the
   // profile annotation prepration step.
   // More details about SamplePGO design can be found in:
   // https://research.google.com/pubs/pub45290.html
   // FIXME: revisit how SampleProfileLoad/Inliner/ICP is structured.
   if (PGOOpt && !PGOOpt->SampleProfileFile.empty() &&
       Phase == ThinLTOPhase::PostLink)
ModulePassManager PassBuilder::buildLTODefaultPipeline(OptimizationLevel Level,
   // Force any function attributes we want the rest of the pipeline to observe.
   MPM.addPass(ForceFunctionAttrsPass());
 
   // Do basic inference of function attributes from known properties of system
   // libraries and other oracles.
   MPM.addPass(InferFunctionAttrsPass());
 
   if (Level > 1) {
+    FunctionPassManager EarlyFPM(DebugLogging);
+    EarlyFPM.addPass(CallSiteSplittingPass());
+    MPM.addPass(createModuleToFunctionPassAdaptor(std::move(EarlyFPM)));
+
     // Indirect call promotion. This should promote all the targets that are
     // left by the earlier promotion pass that promotes intra-module targets.
     // This two-step promotion is to save the compile time. For LTO, it should
     // produce the same result as if we only do promotion here.
     MPM.addPass(PGOIndirectCallPromotion(
         true /* InLTO */, PGOOpt && !PGOOpt->SampleProfileFile.empty()));
-
     // Propagate constants at call sites into the functions they call.  This
     // opens opportunities for globalopt (and inlining) by substituting function
     // pointers passed as arguments to direct uses of functions.
    MPM.addPass(IPSCCPPass());
 
    // Attach metadata to indirect call sites indicating the set of functions
    // they may target at run-time. This should follow IPSCCP.
    MPM.addPass(CalledValuePropagationPass());

lib/Passes/PassRegistry.def

 #define FUNCTION_PASS(NAME, CREATE_PASS)
 #endif
 FUNCTION_PASS("aa-eval", AAEvaluator())
 FUNCTION_PASS("adce", ADCEPass())
 FUNCTION_PASS("add-discriminators", AddDiscriminatorsPass())
 FUNCTION_PASS("alignment-from-assumptions", AlignmentFromAssumptionsPass())
 FUNCTION_PASS("bdce", BDCEPass())
 FUNCTION_PASS("break-crit-edges", BreakCriticalEdgesPass())
+FUNCTION_PASS("callsite-splitting", CallSiteSplittingPass())
 FUNCTION_PASS("consthoist", ConstantHoistingPass())
 FUNCTION_PASS("correlated-propagation", CorrelatedValuePropagationPass())
 FUNCTION_PASS("dce", DCEPass())
 FUNCTION_PASS("div-rem-pairs", DivRemPairsPass())
 FUNCTION_PASS("dse", DSEPass())
 FUNCTION_PASS("dot-cfg", CFGPrinterPass())
 FUNCTION_PASS("dot-cfg-only", CFGOnlyPrinterPass())
 FUNCTION_PASS("early-cse", EarlyCSEPass(/*UseMemorySSA=*/false))

lib/Transforms/IPO/PassManagerBuilder.cpp

void PassManagerBuilder::populateModulePassManager(
   if (PrepareForThinLTOUsingPGOSampleProfile)
     DisableUnrollLoops = true;
 
   // Infer attributes about declarations if possible.
   MPM.add(createInferFunctionAttrsLegacyPass());
 
   addExtensionsToPM(EP_ModuleOptimizerEarly, MPM);
 
+  if (OptLevel > 2)
+    MPM.add(createCallSiteSplittingPass());
+
   MPM.add(createIPSCCPPass());          // IP SCCP
   MPM.add(createCalledValuePropagationPass());
   MPM.add(createGlobalOptimizerPass()); // Optimize out global vars
   // Promote any localized global vars.
   MPM.add(createPromoteMemoryToRegisterPass());
 
   MPM.add(createDeadArgEliminationPass()); // Dead argument elimination
 
   addInstructionCombiningPass(MPM); // Clean up after IPCP & DAE
   addExtensionsToPM(EP_Peephole, MPM);
   MPM.add(createCFGSimplificationPass()); // Clean up after IPCP & DAE
 
   // For SamplePGO in ThinLTO compile phase, we do not want to do indirect
   // call promotion as it will change the CFG too much to make the 2nd
   // profile annotation in backend more difficult.
   // PGO instrumentation is added during the compile phase for ThinLTO, do
   // not run it a second time
   if (!PerformThinLTO && !PrepareForThinLTOUsingPGOSampleProfile)
     addPGOInstrPasses(MPM);
 
   // We add a module alias analysis pass here. In part due to bugs in the
void PassManagerBuilder::addLTOOptimizationPasses(legacy::PassManagerBase &PM) {
 
   // Allow forcing function attributes as a debugging and tuning aid.
   PM.add(createForceFunctionAttrsLegacyPass());
 
   // Infer attributes about declarations if possible.
   PM.add(createInferFunctionAttrsLegacyPass());
 
   if (OptLevel > 1) {
+    // Split call-site with more constrained arguments.
+    PM.add(createCallSiteSplittingPass());
+
     // Indirect call promotion. This should promote all the targets that are
     // left by the earlier promotion pass that promotes intra-module targets.
     // This two-step promotion is to save the compile time. For LTO, it should
     // produce the same result as if we only do promotion here.
     PM.add(
         createPGOIndirectCallPromotionLegacyPass(true, !PGOSampleUse.empty()));
 
     // Propagate constants at call sites into the functions they call.  This

lib/Transforms/Scalar/CMakeLists.txt

 add_llvm_library(LLVMScalarOpts
   ADCE.cpp
   AlignmentFromAssumptions.cpp
   BDCE.cpp
+  CallSiteSplitting.cpp
   ConstantHoisting.cpp
   ConstantProp.cpp
   CorrelatedValuePropagation.cpp
   DCE.cpp
   DeadStoreElimination.cpp
   DivRemPairs.cpp
   EarlyCSE.cpp
   FlattenCFGPass.cpp

lib/Transforms/Scalar/CallSiteSplitting.cpp

+//===- CallSiteSplitting.cpp ----------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements a transformation that tries to split a call-site to pass
+// more constrained arguments if its argument is predicated in the control flow
+// so that we can expose better context to the later passes (e.g, inliner, jump
+// threading, or IPA-CP based function cloning, etc.).
+// As of now we support two cases :
+//
+// 1) If a call site is dominated by an OR condition and if any of its arguments
+// are predicated on this OR condition, try to split the condition with more
+// constrained arguments. For example, in the code below, we try to split the
+// call site since we can predicate the argument(ptr) based on the OR condition.
+//
+// Split from :
+//   if (!ptr || c)
+//     callee(ptr);
+// to :
+//   if (!ptr)
+//     callee(null)         // set the known constant value
+//   else if (c)
+//     callee(nonnull ptr)  // set non-null attribute in the argument
+//
+// 2) We can also split a call-site based on constant incoming values of a PHI
+// For example,
+// from :
+//   Header:
+//    %c = icmp eq i32 %i1, %i2
+//    br i1 %c, label %Tail, label %TBB
+//   TBB:
+//    br label Tail%
+//   Tail:
+//    %p = phi i32 [ 0, %Header], [ 1, %TBB]
+//    call void @bar(i32 %p)
+// to
+//   Header:
+//    %c = icmp eq i32 %i1, %i2
+//    br i1 %c, label %Tail-split0, label %TBB
+//   TBB:
+//    br label %Tail-split1
+//   Tail-split0:
+//    call void @bar(i32 0)
+//    br label %Tail
+//   Tail-split1:
+//    call void @bar(i32 1)
+//    br label %Tail
+//   Tail:
+//    %p = phi i32 [ 0, %Tail-split0 ], [ 1, %Tail-split1 ]
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/Scalar/CallSiteSplitting.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
+#include "llvm/Transforms/Utils/Local.h"
+
+using namespace llvm;
+using namespace PatternMatch;
+
+#define DEBUG_TYPE "callsite-splitting"
+
+STATISTIC(NumCallSiteSplit, "Number of call-site split");
+
+static void addNonNullAttribute(Instruction *CallI, Instruction *&NewCallI,
+                                Value *Op) {
+  if (!NewCallI)
+    NewCallI = CallI->clone();
+  CallSite CS(NewCallI);
+  unsigned ArgNo = 0;
+  for (auto &I : CS.args()) {
+    if (&*I == Op)
+      CS.addParamAttr(ArgNo, Attribute::NonNull);
+    ++ArgNo;
+  }
+}
+
+static void setConstantInArgument(Instruction *CallI, Instruction *&NewCallI,
+                                  Value *Op, Constant *ConstValue) {
+  if (!NewCallI)
+    NewCallI = CallI->clone();
+  CallSite CS(NewCallI);
+  unsigned ArgNo = 0;
+  for (auto &I : CS.args()) {
+    if (&*I == Op)
+      CS.setArgument(ArgNo, ConstValue);
+    ++ArgNo;
+  }
+}
+
+static bool createCallSitesOnOrPredicatedArgument(
+    CallSite CS, Instruction *&NewCSTakenFromHeader,
+    Instruction *&NewCSTakenFromNextCond,
+    SmallVectorImpl<BranchInst *> &BranchInsts, BasicBlock *HeaderBB) {
+  assert(BranchInsts.size() <= 2 &&
+         "Unexpected number of blocks in the OR predicated condition");
+  Instruction *Instr = CS.getInstruction();
+  BasicBlock *CallSiteBB = Instr->getParent();
+  TerminatorInst *HeaderTI = HeaderBB->getTerminator();
+  bool IsCSInTakenPath = CallSiteBB == HeaderTI->getSuccessor(0);
+
+  for (unsigned I = 0, E = BranchInsts.size(); I != E; ++I) {
+    BranchInst *PBI = BranchInsts[I];
+    assert(isa<ICmpInst>(PBI->getCondition()) &&
+           "Unexpected condition in a conditional branch.");
+    ICmpInst *Cmp = cast<ICmpInst>(PBI->getCondition());
+    Value *Arg = Cmp->getOperand(0);
+    assert(isa<Constant>(Cmp->getOperand(1)) &&
+           "Expected op1 to be a constant.");
+    Constant *ConstVal = cast<Constant>(Cmp->getOperand(1));
+    CmpInst::Predicate Pred = Cmp->getPredicate();
+
+    if (PBI->getParent() == HeaderBB) {
+      Instruction *&CallTakenFromHeader =
+          IsCSInTakenPath ? NewCSTakenFromHeader : NewCSTakenFromNextCond;
+      Instruction *&CallUntakenFromHeader =
+          IsCSInTakenPath ? NewCSTakenFromNextCond : NewCSTakenFromHeader;
+
+      assert(Pred == ICmpInst::ICMP_EQ ||
+             Pred == ICmpInst::ICMP_NE &&
+                 "Unexpected predicate in an OR condition");
+
+      // Set the constant value for agruments in the call predicated based on
+      // the OR condition.
+      Instruction *&CallToSetConst = Pred == ICmpInst::ICMP_EQ
+                                         ? CallTakenFromHeader
+                                         : CallUntakenFromHeader;
+      setConstantInArgument(Instr, CallToSetConst, Arg, ConstVal);
+
+      // Add the NonNull attribute if compared with the null pointer.
+      if (ConstVal->getType()->isPointerTy() && ConstVal->isNullValue()) {
+        Instruction *&CallToSetAttr = Pred == ICmpInst::ICMP_EQ
+                                          ? CallUntakenFromHeader
+                                          : CallTakenFromHeader;
+        addNonNullAttribute(Instr, CallToSetAttr, Arg);
+      }
+      continue;
+    }
+
+    if (Pred == ICmpInst::ICMP_EQ) {
+      if (PBI->getSuccessor(0) == Instr->getParent()) {
+        // Set the constant value for the call taken from the second block in
+        // the OR condition.
+        setConstantInArgument(Instr, NewCSTakenFromNextCond, Arg, ConstVal);
+      } else {
+        // Add the NonNull attribute if compared with the null pointer for the
+        // call taken from the second block in the OR condition.
+        if (ConstVal->getType()->isPointerTy() && ConstVal->isNullValue())
+          addNonNullAttribute(Instr, NewCSTakenFromNextCond, Arg);
+      }
+    } else {
+      if (PBI->getSuccessor(0) == Instr->getParent()) {
+        // Add the NonNull attribute if compared with the null pointer for the
+        // call taken from the second block in the OR condition.
+        if (ConstVal->getType()->isPointerTy() && ConstVal->isNullValue())
+          addNonNullAttribute(Instr, NewCSTakenFromNextCond, Arg);
+      } else if (Pred == ICmpInst::ICMP_NE) {
+        // Set the constant value for the call in the untaken path from the
+        // header block.
+        setConstantInArgument(Instr, NewCSTakenFromNextCond, Arg, ConstVal);
+      } else
+        llvm_unreachable("Unexpected condition");
+    }
+  }
+  return NewCSTakenFromHeader || NewCSTakenFromNextCond;
+}
+
+static bool canSplitCallSite(CallSite CS) {
+  // FIXME: As of now we handle only CallInst. InvokeInst could be handled
+  // without too much effort.
+  Instruction *Instr = CS.getInstruction();
+  if (!isa<CallInst>(Instr))
+    return false;
+
+  // Allow splitting a call-site only when there is no instruction before the
+  // call-site in the basic block. Based on this constraint, we only clone the
+  // call instruction, and we do not move a call-site across any other
+  // instruction.
+  BasicBlock *CallSiteBB = Instr->getParent();
+  if (Instr != CallSiteBB->getFirstNonPHI())
+    return false;
+
+  pred_iterator PII = pred_begin(CallSiteBB);
+  pred_iterator PIE = pred_end(CallSiteBB);
+  unsigned NumPreds = std::distance(PII, PIE);
+
+  // Allow only one extra call-site. No more than two from one call-site.
+  if (NumPreds != 2)
+    return false;
+
+  // Cannot split an edge from an IndirectBrInst.
+  BasicBlock *Preds[2] = {*PII++, *PII};
+  if (isa<IndirectBrInst>(Preds[0]->getTerminator()) ||
+      isa<IndirectBrInst>(Preds[1]->getTerminator()))
+    return false;
+
+  return CallSiteBB->canSplitPredecessors();
+}
+
+/// Return true if the CS is split into its new predecessors which are directly
+/// hooked to each of its orignial predecessors pointed by PredBB1 and PredBB2.
+/// Note that PredBB1 and PredBB2 are decided in findPredicatedArgument(),
+/// especially for the OR predicated case where PredBB1 will point the header,
+/// and PredBB2 will point the the second compare block. CallInst1 and CallInst2
+/// will be the new call-sites placed in the new predecessors split for PredBB1
+/// and PredBB2, repectively. Therefore, CallInst1 will be the call-site placed
+/// between Header and Tail, and CallInst2 will be the call-site between TBB and
+/// Tail. For example, in the IR below with an OR condition, the call-site can
+/// be split
+///
+/// from :
+///
+///   Header:
+///     %c = icmp eq i32* %a, null
+///     br i1 %c %Tail, %TBB
+///   TBB:
+///     %c2 = icmp eq i32* %b, null
+///     br i1 %c %Tail, %End
+///   Tail:
+///     %ca = call i1  @callee (i32* %a, i32* %b)
+///
+///  to :
+///
+///   Header:                          // PredBB1 is Header
+///     %c = icmp eq i32* %a, null
+///     br i1 %c %Tail-split1, %TBB
+///   TBB:                             // PredBB2 is TBB
+///     %c2 = icmp eq i32* %b, null
+///     br i1 %c %Tail-split2, %End
+///   Tail-split1:
+///     %ca1 = call @callee (i32* null, i32* %b)         // CallInst1
+///    br %Tail
+///   Tail-split2:
+///     %ca2 = call @callee (i32* nonnull %a, i32* null) // CallInst2
+///    br %Tail
+///   Tail:
+///    %p = phi i1 [%ca1, %Tail-split1],[%ca2, %Tail-split2]
+///
+/// Note that for an OR predicated case, CallInst1 and CallInst2 should be
+/// created with more constrained arguments in
+/// createCallSitesOnOrPredicatedArgument().
+static void splitCallSite(CallSite CS, BasicBlock *PredBB1, BasicBlock *PredBB2,
+                          Instruction *CallInst1, Instruction *CallInst2) {
+  Instruction *Instr = CS.getInstruction();
+  BasicBlock *TailBB = Instr->getParent();
+  assert(Instr == (TailBB->getFirstNonPHI()) && "Unexpected call-site");
+
+  BasicBlock *SplitBlock1 =
+      SplitBlockPredecessors(TailBB, PredBB1, ".predBB1.split");
+  BasicBlock *SplitBlock2 =
+      SplitBlockPredecessors(TailBB, PredBB2, ".predBB2.split");
+
+  assert((SplitBlock1 && SplitBlock2) && "Unexpected new basic block split.");
+
+  if (!CallInst1)
+    CallInst1 = Instr->clone();
+  if (!CallInst2)
+    CallInst2 = Instr->clone();
+
+  CallInst1->insertBefore(&*SplitBlock1->getFirstInsertionPt());
+  CallInst2->insertBefore(&*SplitBlock2->getFirstInsertionPt());
+
+  CallSite CS1(CallInst1);
+  CallSite CS2(CallInst2);
+
+  // Handle PHIs used as arguments in the call-site.
+  for (auto &PI : *TailBB) {
+    PHINode *PN = dyn_cast<PHINode>(&PI);
+    if (!PN)
+      break;
+    unsigned ArgNo = 0;
+    for (auto &CI : CS.args()) {
+      if (&*CI == PN) {
+        CS1.setArgument(ArgNo, PN->getIncomingValueForBlock(SplitBlock1));
+        CS2.setArgument(ArgNo, PN->getIncomingValueForBlock(SplitBlock2));
+      }
+      ++ArgNo;
+    }
+  }
+
+  // Replace users of the original call with a PHI mering call-sites split.
+  if (Instr->getNumUses()) {
+    PHINode *PN = PHINode::Create(Instr->getType(), 2, "phi.call", Instr);
+    PN->addIncoming(CallInst1, SplitBlock1);
+    PN->addIncoming(CallInst2, SplitBlock2);
+    Instr->replaceAllUsesWith(PN);
+  }
+  DEBUG(dbgs() << "split call-site : " << *Instr << " into \n");
+  DEBUG(dbgs() << "    " << *CallInst1 << " in " << SplitBlock1->getName()
+               << "\n");
+  DEBUG(dbgs() << "    " << *CallInst2 << " in " << SplitBlock2->getName()
+               << "\n");
+  Instr->eraseFromParent();
+  NumCallSiteSplit++;
+}
+
+static bool isCondRelevantToAnyCallArgument(ICmpInst *Cmp, CallSite CS) {
+  assert(isa<Constant>(Cmp->getOperand(1)) && "Expected a constant operand.");
+  Value *Op0 = Cmp->getOperand(0);
+  unsigned ArgNo = 0;
+  for (CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end(); I != E;
+       ++I, ++ArgNo) {
+    // Don't consider constant or arguments that are already known non-null.
+    if (isa<Constant>(*I) || CS.paramHasAttr(ArgNo, Attribute::NonNull))
+      continue;
+
+    if (*I == Op0)
+      return true;
+  }
+  return false;
+}
+
+static void findOrCondRelevantToCallArgument(
+    CallSite CS, BasicBlock *PredBB, BasicBlock *OtherPredBB,
+    SmallVectorImpl<BranchInst *> &BranchInsts, BasicBlock *&HeaderBB) {
+  auto *PBI = dyn_cast<BranchInst>(PredBB->getTerminator());
+  if (!PBI || !PBI->isConditional())
+    return;
+
+  if (PBI->getSuccessor(0) == OtherPredBB ||
+      PBI->getSuccessor(1) == OtherPredBB)
+    if (PredBB == OtherPredBB->getSinglePredecessor()) {
+      assert(!HeaderBB && "Expect to find only a single header block");
+      HeaderBB = PredBB;
+    }
+
+  CmpInst::Predicate Pred;
+  Value *Cond = PBI->getCondition();
+  if (!match(Cond, m_ICmp(Pred, m_Value(), m_Constant())))
+    return;
+  ICmpInst *Cmp = cast<ICmpInst>(Cond);
+  if (Pred == ICmpInst::ICMP_EQ || Pred == ICmpInst::ICMP_NE)
+    if (isCondRelevantToAnyCallArgument(Cmp, CS))
+      BranchInsts.push_back(PBI);
+}
+
+// Return true if the call-site has an argument which is a PHI with only
+// constant incoming values.
+static bool isPredicatedOnPHI(CallSite CS) {
+  Instruction *Instr = CS.getInstruction();
+  BasicBlock *Parent = Instr->getParent();
+  if (Instr != Parent->getFirstNonPHI())
+    return false;
+
+  for (auto &BI : *Parent) {
+    if (PHINode *PN = dyn_cast<PHINode>(&BI)) {
+      for (auto &I : CS.args())
+        if (&*I == PN) {
+          assert(PN->getNumIncomingValues() == 2 &&
+                 "Unexpected number of incoming values");
+          if (PN->getIncomingBlock(0) == PN->getIncomingBlock(1))
+            return false;
+          if (PN->getIncomingValue(0) == PN->getIncomingValue(1))
+            continue;
+          if (isa<Constant>(PN->getIncomingValue(0)) &&
+              isa<Constant>(PN->getIncomingValue(1)))
+            return true;
+        }
+    }
+    break;
+  }
+  return false;
+}
+
+// Return true if an agument in CS is predicated on an 'or' condition.
+// Create new call-site with arguments constrained based on the OR condition.
+static bool findPredicatedOnOrCondition(CallSite CS, BasicBlock *PredBB1,
+                                        BasicBlock *PredBB2,
+                                        Instruction *&NewCallTakenFromHeader,
+                                        Instruction *&NewCallTakenFromNextCond,
+                                        BasicBlock *&HeaderBB) {
+  SmallVector<BranchInst *, 4> BranchInsts;
+  findOrCondRelevantToCallArgument(CS, PredBB1, PredBB2, BranchInsts, HeaderBB);
+  findOrCondRelevantToCallArgument(CS, PredBB2, PredBB1, BranchInsts, HeaderBB);
+  if (BranchInsts.empty() || !HeaderBB)
+    return false;
+
+  // If an OR condition is detected, try to create call sites with constrained
+  // arguments (e.g., NonNull attribute or constant value).
+  return createCallSitesOnOrPredicatedArgument(CS, NewCallTakenFromHeader,
+                                               NewCallTakenFromNextCond,
+                                               BranchInsts, HeaderBB);
+}
+
+static bool findPredicatedArgument(CallSite CS, Instruction *&CallInst1,
+                                   Instruction *&CallInst2,
+                                   BasicBlock *&PredBB1, BasicBlock *&PredBB2) {
+  BasicBlock *CallSiteBB = CS.getInstruction()->getParent();
+  pred_iterator PII = pred_begin(CallSiteBB);
+  pred_iterator PIE = pred_end(CallSiteBB);
+  assert(std::distance(PII, PIE) == 2 && "Expect only two predecessors.");
+  BasicBlock *Preds[2] = {*PII++, *PII};
+  BasicBlock *&HeaderBB = PredBB1;
+  if (!findPredicatedOnOrCondition(CS, Preds[0], Preds[1], CallInst1, CallInst2,
+                                   HeaderBB) &&
+      !isPredicatedOnPHI(CS))
+    return false;
+
+  if (!PredBB1)
+    PredBB1 = Preds[0];
+
+  PredBB2 = PredBB1 == Preds[0] ? Preds[1] : Preds[0];
+  return true;
+}
+
+static bool tryToSplitCallSite(CallSite CS) {
+  if (!CS.arg_size())
+    return false;
+
+  BasicBlock *PredBB1 = nullptr;
+  BasicBlock *PredBB2 = nullptr;
+  Instruction *CallInst1 = nullptr;
+  Instruction *CallInst2 = nullptr;
+  if (!canSplitCallSite(CS) ||
+      !findPredicatedArgument(CS, CallInst1, CallInst2, PredBB1, PredBB2)) {
+    assert(!CallInst1 && !CallInst2 && "Unexpected new call-sites cloned.");
+    return false;
+  }
+  splitCallSite(CS, PredBB1, PredBB2, CallInst1, CallInst2);
+  return true;
+}
+
+static bool doCallSiteSplitting(Function &F, TargetLibraryInfo &TLI) {
+  bool Changed = false;
+  for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE;) {
+    BasicBlock &BB = *BI++;
+    for (BasicBlock::iterator II = BB.begin(), IE = BB.end(); II != IE;) {
+      Instruction *I = &*II++;
+      CallSite CS(cast<Value>(I));
+      if (!CS || isa<IntrinsicInst>(I) || isInstructionTriviallyDead(I, &TLI))
+        continue;
+
+      Function *Callee = CS.getCalledFunction();
+      if (!Callee || Callee->isDeclaration())
+        continue;
+      Changed |= tryToSplitCallSite(CS);
+    }
+  }
+  return Changed;
+}
+
+namespace {
+struct CallSiteSplittingLegacyPass : public FunctionPass {
+  static char ID;
+  CallSiteSplittingLegacyPass() : FunctionPass(ID) {
+    initializeCallSiteSplittingLegacyPassPass(*PassRegistry::getPassRegistry());
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<TargetLibraryInfoWrapperPass>();
+    FunctionPass::getAnalysisUsage(AU);
+  }
+
+  bool runOnFunction(Function &F) override {
+    if (skipFunction(F))
+      return false;
+
+    auto &TLI = getAnalysis<TargetLibraryInfoWrapperPass>().getTLI();
+    return doCallSiteSplitting(F, TLI);
+  }
+};
+} // namespace
+
+char CallSiteSplittingLegacyPass::ID = 0;
+INITIALIZE_PASS_BEGIN(CallSiteSplittingLegacyPass, "callsite-splitting",
+                      "Call-site splitting", false, false)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
+INITIALIZE_PASS_END(CallSiteSplittingLegacyPass, "callsite-splitting",
+                    "Call-site splitting", false, false)
+FunctionPass *llvm::createCallSiteSplittingPass() {
+  return new CallSiteSplittingLegacyPass();
+}
+
+PreservedAnalyses CallSiteSplittingPass::run(Function &F,
+                                             FunctionAnalysisManager &AM) {
+  auto &TLI = AM.getResult<TargetLibraryAnalysis>(F);
+
+  if (!doCallSiteSplitting(F, TLI))
+    return PreservedAnalyses::all();
+  PreservedAnalyses PA;
+  return PA;
+}

lib/Transforms/Scalar/Scalar.cpp

 using namespace llvm;
 
 /// initializeScalarOptsPasses - Initialize all passes linked into the
 /// ScalarOpts library.
 void llvm::initializeScalarOpts(PassRegistry &Registry) {
   initializeADCELegacyPassPass(Registry);
   initializeBDCELegacyPassPass(Registry);
   initializeAlignmentFromAssumptionsPass(Registry);
+  initializeCallSiteSplittingLegacyPassPass(Registry);
   initializeConstantHoistingLegacyPassPass(Registry);
   initializeConstantPropagationPass(Registry);
   initializeCorrelatedValuePropagationPass(Registry);
   initializeDCELegacyPassPass(Registry);
   initializeDeadInstEliminationPass(Registry);
   initializeDivRemPairsLegacyPassPass(Registry);
   initializeScalarizerPass(Registry);
   initializeDSELegacyPassPass(Registry);

test/Other/new-pm-defaults.ll

 ; CHECK-O-NEXT: Running pass: SimplifyCFGPass
 ; CHECK-O-NEXT: Running analysis: TargetIRAnalysis
 ; CHECK-O-NEXT: Running analysis: AssumptionAnalysis
 ; CHECK-O-NEXT: Running pass: SROA
 ; CHECK-O-NEXT: Running analysis: DominatorTreeAnalysis
 ; CHECK-O-NEXT: Running pass: EarlyCSEPass
 ; CHECK-O-NEXT: Running analysis: TargetLibraryAnalysis
 ; CHECK-O-NEXT: Running pass: LowerExpectIntrinsicPass
+; CHECK-O3-NEXT: Running pass: CallSiteSplittingPass
 ; CHECK-O-NEXT: Finished llvm::Function pass manager run.
 ; CHECK-O-NEXT: Running pass: IPSCCPPass
 ; CHECK-O-NEXT: Running pass: CalledValuePropagationPass
 ; CHECK-O-NEXT: Running pass: GlobalOptPass
 ; CHECK-O-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PromotePass>
 ; CHECK-O-NEXT: Running pass: DeadArgumentEliminationPass
 ; CHECK-O-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PassManager{{.*}}>
 ; CHECK-O-NEXT: Starting llvm::Function pass manager run.

test/Other/new-pm-lto-defaults.ll

 
 ; CHECK-O: Starting llvm::Module pass manager run.
 ; CHECK-O-NEXT: Running pass: PassManager<{{.*}}Module
 ; CHECK-O-NEXT: Starting llvm::Module pass manager run.
 ; CHECK-O-NEXT: Running pass: GlobalDCEPass
 ; CHECK-O-NEXT: Running pass: ForceFunctionAttrsPass
 ; CHECK-O-NEXT: Running pass: InferFunctionAttrsPass
 ; CHECK-O-NEXT: Running analysis: TargetLibraryAnalysis
+; CHECK-O2-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PassManager{{.*}}>
+; CHECK-O2-NEXT: Running analysis: InnerAnalysisManagerProxy<{{.*}}Module
+; CHECK-O2-NEXT: Starting llvm::Function pass manager run.
+; CHECK-O2-NEXT: Running pass: CallSiteSplittingPass on foo
+; CHECK-O2-NEXT: Running analysis: TargetLibraryAnalysis on foo
+; CHECK-O2-NEXT: Finished llvm::Function pass manager run.
 ; CHECK-O2-NEXT: PGOIndirectCallPromotion
 ; CHECK-O2-NEXT: Running analysis: ProfileSummaryAnalysis
-; CHECK-O2-NEXT: Running analysis: InnerAnalysisManagerProxy<{{.*}}Function
 ; CHECK-O2-NEXT: Running analysis: OptimizationRemarkEmitterAnalysis
 ; CHECK-O2-NEXT: Running pass: IPSCCPPass
 ; CHECK-O2-NEXT: Running pass: CalledValuePropagationPass
 ; CHECK-O-NEXT: Running pass: ModuleToPostOrderCGSCCPassAdaptor<{{.*}}PostOrderFunctionAttrsPass>
 ; CHECK-O-NEXT: Running analysis: InnerAnalysisManagerProxy<{{.*}}SCC
 ; CHECK-O1-NEXT: Running analysis: InnerAnalysisManagerProxy<{{.*}}Function
 ; CHECK-O-NEXT: Running analysis: LazyCallGraphAnalysis
 ; CHECK-O-NEXT: Running analysis: FunctionAnalysisManagerCGSCCProxy
 ; CHECK-O-NEXT: Running analysis: OuterAnalysisManagerProxy<{{.*}}LazyCallGraph{{.*}}>
 ; CHECK-O-NEXT: Running analysis: AAManager
-; CHECK-O-NEXT: Running analysis: TargetLibraryAnalysis
+; CHECK-O1-NEXT: Running analysis: TargetLibraryAnalysis
 ; CHECK-O-NEXT: Running pass: ReversePostOrderFunctionAttrsPass
 ; CHECK-O-NEXT: Running analysis: CallGraphAnalysis
 ; CHECK-O-NEXT: Running pass: GlobalSplitPass
 ; CHECK-O-NEXT: Running pass: WholeProgramDevirtPass
 ; CHECK-O2-NEXT: Running pass: GlobalOptPass
 ; CHECK-O2-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PromotePass>
 ; CHECK-O2-NEXT: Running analysis: DominatorTreeAnalysis
 ; CHECK-O2-NEXT: Running analysis: AssumptionAnalysis

test/Other/new-pm-thinlto-defaults.ll

 ; CHECK-O-NEXT: Running pass: SimplifyCFGPass
 ; CHECK-O-NEXT: Running analysis: TargetIRAnalysis
 ; CHECK-O-NEXT: Running analysis: AssumptionAnalysis
 ; CHECK-O-NEXT: Running pass: SROA
 ; CHECK-O-NEXT: Running analysis: DominatorTreeAnalysis
 ; CHECK-O-NEXT: Running pass: EarlyCSEPass
 ; CHECK-O-NEXT: Running analysis: TargetLibraryAnalysis
 ; CHECK-O-NEXT: Running pass: LowerExpectIntrinsicPass
+; CHECK-O3-NEXT: Running pass: CallSiteSplittingPass
 ; CHECK-O-NEXT: Finished llvm::Function pass manager run.
 ; CHECK-O-NEXT: Running pass: IPSCCPPass
 ; CHECK-O-NEXT: Running pass: CalledValuePropagationPass
 ; CHECK-O-NEXT: Running pass: GlobalOptPass
 ; CHECK-O-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PromotePass>
 ; CHECK-O-NEXT: Running pass: DeadArgumentEliminationPass
 ; CHECK-O-NEXT: Running pass: ModuleToFunctionPassAdaptor<{{.*}}PassManager{{.*}}>
 ; CHECK-O-NEXT: Starting llvm::Function pass manager run.

test/Transforms/CallSiteSplitting/callsite-split-or-phi.ll

+; RUN: opt < %s -callsite-splitting -S | FileCheck %s
+; RUN: opt < %s  -passes='function(callsite-splitting)' -S | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-linaro-linux-gnueabi"
+
+;CHECK-LABEL: @test_eq_eq
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* null, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* nonnull %a, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_eq_eq(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp eq i32* %a, null
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_ne_eq
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* nonnull %a, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* null, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_ne_eq(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp ne i32* %a, null
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_ne_ne
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* nonnull %a, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* null, i32 %v, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_ne_ne(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp ne i32* %a, null
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp ne i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_eq_eq_untaken
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* nonnull %a, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* null, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_eq_eq_untaken(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp eq i32* %a, null
+  br i1 %tobool1, label %TBB, label %Tail
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_ne_eq_untaken
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* null, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* nonnull %a, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_ne_eq_untaken(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp ne i32* %a, null
+  br i1 %tobool1, label %TBB, label %Tail
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_ne_ne_untaken
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* null, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* nonnull %a, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_ne_ne_untaken(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp ne i32* %a, null
+  br i1 %tobool1, label %TBB, label %Tail
+
+TBB:
+  %cmp = icmp ne i32 %v, 1
+  br i1 %cmp, label %End, label %Tail
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_nonconst_const_phi
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* %a, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* %a, i32 1, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_nonconst_const_phi(i32* %a, i32* %b, i32 %v) {
+Header:
+  %tobool1 = icmp eq i32* %a, %b
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_nonconst_nonconst_phi
+;CHECK-LABEL: Tail.predBB1.split:
+;CHECK: %[[CALL1:.*]] = call i32 @callee(i32* %a, i32 %v, i32 1)
+;CHECK-LABEL: Tail.predBB2.split:
+;CHECK: %[[CALL2:.*]] = call i32 @callee(i32* %a, i32 %v, i32 2)
+;CHECK-LABEL: Tail
+;CHECK: %p = phi i32 [ 1, %Tail.predBB1.split ], [ 2, %Tail.predBB2.split ]
+;CHECK: %[[MERGED:.*]] = phi i32 [ %[[CALL1]], %Tail.predBB1.split ], [ %[[CALL2]], %Tail.predBB2.split ]
+;CHECK: ret i32 %[[MERGED]]
+define i32 @test_nonconst_nonconst_phi(i32* %a, i32* %b, i32 %v, i32 %v2) {
+Header:
+  %tobool1 = icmp eq i32* %a, %b
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, %v2 
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_nonconst_nonconst_phi_noncost
+;CHECK-NOT: Tail.predBB1.split:
+;CHECK-NOT: Tail.predBB2.split:
+;CHECK-LABEL: Tail:
+;CHECK: %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+;CHECK: ret i32 %r
+define i32 @test_nonconst_nonconst_phi_noncost(i32* %a, i32* %b, i32 %v, i32 %v2) {
+Header:
+  %tobool1 = icmp eq i32* %a, %b
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, %v2 
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[%v,%Header], [%v2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_fisrtnonphi
+;CHECK-NOT: Tail.predBB1.split:
+;CHECK-NOT: Tail.predBB2.split:
+;CHECK-LABEL: Tail:
+;CHECK: %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+;CHECK: ret i32 %r
+define i32 @test_fisrtnonphi(i32* %a, i32 %v) {
+Header:
+  %tobool1 = icmp eq i32* %a, null
+  br i1 %tobool1, label %Tail, label %TBB
+
+TBB:
+  %cmp = icmp eq i32 %v, 1
+  br i1 %cmp, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  store i32 %v, i32* %a
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_3preds_constphi
+;CHECK-NOT: Tail.predBB1.split:
+;CHECK-NOT: Tail.predBB2.split:
+;CHECK-LABEL: Tail:
+;CHECK: %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+;CHECK: ret i32 %r
+define i32 @test_3preds_constphi(i32* %a, i32 %v, i1 %c1, i1 %c2, i1 %c3) {
+Header:
+  br i1 %c1, label %Tail, label %TBB1
+
+TBB1:
+  br i1 %c2, label %Tail, label %TBB2
+
+TBB2:
+  br i1 %c3, label %Tail, label %End
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB1], [3, %TBB2]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+;CHECK-LABEL: @test_indirectbr_phi
+;CHECK-NOT: Tail.predBB1.split:
+;CHECK-NOT: Tail.predBB2.split:
+;CHECK-LABEL: Tail:
+;CHECK: %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+;CHECK: ret i32 %r
+define i32 @test_indirectbr_phi(i8* %address, i32* %a, i32* %b, i32 %v) {
+Header:
+   %indirect.goto.dest = select i1 undef, i8* blockaddress(@test_indirectbr_phi, %End), i8* %address
+   indirectbr i8* %indirect.goto.dest, [label %TBB, label %Tail]
+
+TBB:
+  %indirect.goto.dest2 = select i1 undef, i8* blockaddress(@test_indirectbr_phi, %End), i8* %address
+  indirectbr i8* %indirect.goto.dest2, [label %Tail, label %End]
+
+Tail:
+  %p = phi i32[1,%Header], [2, %TBB]
+  %r = call i32 @callee(i32* %a, i32 %v, i32 %p)
+  ret i32 %r
+
+End:
+  ret i32 %v
+}
+
+define i32 @callee(i32* %a, i32 %v, i32 %p) {
+entry:
+  %c = icmp ne i32* %a, null
+  br i1 %c, label %BB1, label %BB2
+
+BB1:
+  call void @dummy(i32* %a, i32 %p)
+  br label %End
+
+BB2:
+  call void @dummy2(i32 %v, i32 %p)
+  br label %End
+
+End:
+  ret i32 %p
+}
+
+declare void @dummy(i32*, i32)
+declare void @dummy2(i32, i32)

test/Transforms/CallSiteSplitting/callsite-split.ll

+; RUN: opt < %s -callsite-splitting -inline -instcombine -jump-threading -S | FileCheck %s
+; RUN: opt < %s  -passes='function(callsite-splitting),cgscc(inline),function(instcombine,jump-threading)' -S | FileCheck %s
+
+target datalayout = "e-m:e-i8:8:32-i16:16:32-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64-linaro-linux-gnueabi"
+
+%struct.bitmap = type { i32, %struct.bitmap* }
+
+;CHECK-LABEL: @caller
+;CHECK-LABEL: NextCond:
+;CHECK: br {{.*}} label %callee.exit
+;CHECK-LABEL: CallSiteBB.predBB1.split:
+;CHECK: call void @callee(%struct.bitmap* null, %struct.bitmap* null, %struct.bitmap* %b_elt, i1 false)
+;CHECK-LABEL: callee.exit:
+;CHECK: call void @dummy2(%struct.bitmap* %a_elt)
+
+define void @caller(i1 %c, %struct.bitmap* %a_elt, %struct.bitmap* %b_elt) {
+entry:
+  br label %Top
+
+Top:
+  %tobool1 = icmp eq %struct.bitmap* %a_elt, null
+  br i1 %tobool1, label %CallSiteBB, label %NextCond
+
+NextCond:
+  %cmp = icmp ne %struct.bitmap* %b_elt, null
+  br i1 %cmp, label %CallSiteBB, label %End
+
+CallSiteBB:
+  %p = phi i1 [0, %Top], [%c, %NextCond]
+  call void @callee(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %b_elt, i1 %p)
+  br label %End
+
+End:
+  ret void
+}
+
+define void @callee(%struct.bitmap* %dst_elt, %struct.bitmap* %a_elt, %struct.bitmap* %b_elt, i1 %c) {
+entry:
+  %tobool = icmp ne %struct.bitmap* %a_elt, null
+  %tobool1 = icmp ne %struct.bitmap* %b_elt, null
+  %or.cond = and i1 %tobool, %tobool1
+  br i1 %or.cond, label %Cond, label %Big
+
+Cond:
+  %cmp = icmp eq %struct.bitmap*  %dst_elt, %a_elt
+  br i1 %cmp, label %Small, label %Big
+
+Small:
+  call void @dummy2(%struct.bitmap* %a_elt)
+  br label %End
+
+Big:
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  call void @dummy1(%struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt, %struct.bitmap* %a_elt)
+  br label %End
+
+End:
+  ret void
+}
+
+declare void @dummy2(%struct.bitmap*)
+declare void @dummy1(%struct.bitmap*, %struct.bitmap*, %struct.bitmap*, %struct.bitmap*, %struct.bitmap*, %struct.bitmap*)
+
+
+;CHECK-LABEL: @caller2
+;CHECK-LABEL: CallSiteBB.predBB1.split:
+;CHECK: call void @dummy4()
+;CHECK-LABEL: CallSiteBB.predBB2.split:
+;CHECK: call void @dummy3()
+;CheCK-LABEL: CallSiteBB:
+;CHECK: %phi.call = phi i1 [ false, %CallSiteBB.predBB1.split ], [ true, %CallSiteBB.predBB2.split ]
+;CHECK: call void @foo(i1 %phi.call)
+define void @caller2(i1 %c, %struct.bitmap* %a_elt, %struct.bitmap* %b_elt, %struct.bitmap* %c_elt) {
+entry:
+  br label %Top
+
+Top:
+  %tobool1 = icmp eq %struct.bitmap* %a_elt, %b_elt
+  br i1 %tobool1, label %CallSiteBB, label %NextCond
+
+NextCond:
+  %cmp = icmp ne %struct.bitmap* %b_elt, %c_elt
+  br i1 %cmp, label %CallSiteBB, label %End
+
+CallSiteBB:
+  %phi = phi i1 [0, %Top],[1, %NextCond]
+  %u = call i1 @callee2(i1 %phi)
+  call void @foo(i1 %u)
+  br label %End
+
+End:
+  ret void
+}
+
+define i1 @callee2(i1 %b) {
+entry:
+  br i1 %b, label %BB1, label %BB2
+
+BB1:
+  call void @dummy3()
+  br label %End
+
+BB2:
+  call void @dummy4()
+  br label %End
+
+End:
+  ret i1 %b
+}
+
+declare void @dummy3()
+declare void @dummy4()
+declare void @foo(i1)