Introduce "expect" intrinsic instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@134516 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 183 insertions, 0 deletions

diff --git a/lib/CodeGen/IntrinsicLowering.cpp b/lib/CodeGen/IntrinsicLowering.cpp
index 3861ddadf6..b0a823042d 100644
--- a/lib/CodeGen/IntrinsicLowering.cpp
+++ b/lib/CodeGen/IntrinsicLowering.cpp
@@ -353,6 +353,13 @@ void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
     report_fatal_error("Code generator does not support intrinsic function '"+
                       Callee->getName()+"'!");
 
+  case Intrinsic::expect: {
+    // Just replace __builtin_expect(exp, c) with EXP.
+    Value *V = CI->getArgOperand(0);
+    CI->replaceAllUsesWith(V);
+    break;
+  }
+
     // The setjmp/longjmp intrinsics should only exist in the code if it was
     // never optimized (ie, right out of the CFE), or if it has been hacked on
     // by the lowerinvoke pass.  In both cases, the right thing to do is to
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index ea59ca14d8..fc8f0d2996 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -4771,6 +4771,13 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
   case Intrinsic::flt_rounds:
     setValue(&I, DAG.getNode(ISD::FLT_ROUNDS_, dl, MVT::i32));
     return 0;
+
+  case Intrinsic::expect: {
+    // Just replace __builtin_expect(exp, c) with EXP.
+    setValue(&I, getValue(I.getArgOperand(0)));
+    return 0;
+  }
+
   case Intrinsic::trap: {
     StringRef TrapFuncName = getTrapFunctionName();
     if (TrapFuncName.empty()) {
diff --git a/lib/Transforms/Scalar/Scalar.cpp b/lib/Transforms/Scalar/Scalar.cpp
index 158d65352f..302c287d3c 100644
--- a/lib/Transforms/Scalar/Scalar.cpp
+++ b/lib/Transforms/Scalar/Scalar.cpp
@@ -48,6 +48,7 @@ void llvm::initializeScalarOpts(PassRegistry &Registry) {
   initializeLoopUnswitchPass(Registry);
   initializeLoopIdiomRecognizePass(Registry);
   initializeLowerAtomicPass(Registry);
+  initializeLowerExpectIntrinsicPass(Registry);
   initializeMemCpyOptPass(Registry);
   initializeObjCARCAliasAnalysisPass(Registry);
   initializeObjCARCExpandPass(Registry);
diff --git a/lib/Transforms/Utils/CMakeLists.txt b/lib/Transforms/Utils/CMakeLists.txt
index 2df534f232..204c2c63e1 100644
--- a/lib/Transforms/Utils/CMakeLists.txt
+++ b/lib/Transforms/Utils/CMakeLists.txt
@@ -14,6 +14,7 @@ add_llvm_library(LLVMTransformUtils
   Local.cpp
   LoopSimplify.cpp
   LoopUnroll.cpp
+  LowerExpectIntrinsic.cpp
   LowerInvoke.cpp
   LowerSwitch.cpp
   Mem2Reg.cpp
diff --git a/lib/Transforms/Utils/LowerExpectIntrinsic.cpp b/lib/Transforms/Utils/LowerExpectIntrinsic.cpp
new file mode 100644
index 0000000000..fd94c84413
--- /dev/null
+++ b/lib/Transforms/Utils/LowerExpectIntrinsic.cpp
@@ -0,0 +1,163 @@
+#define DEBUG_TYPE "lower-expect-intrinsic"
+#include "llvm/Constants.h"
+#include "llvm/Function.h"
+#include "llvm/BasicBlock.h"
+#include "llvm/LLVMContext.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/Metadata.h"
+#include "llvm/Pass.h"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/ADT/Statistic.h"
+#include <vector>
+
+using namespace llvm;
+
+STATISTIC(IfHandled, "Number of 'expect' intrinsic intructions handled");
+
+static cl::opt<uint32_t>
+LikelyBranchWeight("likely-branch-weight", cl::Hidden, cl::init(64),
+                   cl::desc("Weight of the branch likely to be taken (default = 64)"));
+static cl::opt<uint32_t>
+UnlikelyBranchWeight("unlikely-branch-weight", cl::Hidden, cl::init(4),
+                   cl::desc("Weight of the branch unlikely to be taken (default = 4)"));
+
+namespace {
+
+  class LowerExpectIntrinsic : public FunctionPass {
+
+    bool HandleSwitchExpect(SwitchInst *SI);
+
+    bool HandleIfExpect(BranchInst *BI);
+
+  public:
+    static char ID;
+    LowerExpectIntrinsic() : FunctionPass(ID) {
+      initializeLowerExpectIntrinsicPass(*PassRegistry::getPassRegistry());
+    }
+
+    bool runOnFunction(Function &F);
+  };
+}
+
+
+bool LowerExpectIntrinsic::HandleSwitchExpect(SwitchInst *SI) {
+  CallInst *CI = dyn_cast<CallInst>(SI->getCondition());
+  if (!CI)
+    return false;
+
+  Function *Fn = CI->getCalledFunction();
+  if (!Fn || Fn->getIntrinsicID() != Intrinsic::expect)
+    return false;
+
+  Value *ArgValue = CI->getArgOperand(0);
+  ConstantInt *ExpectedValue = dyn_cast<ConstantInt>(CI->getArgOperand(1));
+  if (!ExpectedValue)
+    return false;
+
+  LLVMContext &Context = CI->getContext();
+  const Type *Int32Ty = Type::getInt32Ty(Context);
+
+  unsigned caseNo = SI->findCaseValue(ExpectedValue);
+  std::vector<Value *> Vec;
+  unsigned n = SI->getNumCases();
+  Vec.resize(n + 1); // +1 for MDString
+
+  Vec[0] = MDString::get(Context, "branch_weights");
+  for (unsigned i = 0; i < n; ++i) {
+    Vec[i + 1] = ConstantInt::get(Int32Ty, i == caseNo ? LikelyBranchWeight : UnlikelyBranchWeight);
+  }
+
+  MDNode *WeightsNode = llvm::MDNode::get(Context, Vec);
+  SI->setMetadata(LLVMContext::MD_prof, WeightsNode);
+
+  SI->setCondition(ArgValue);
+  return true;
+}
+
+
+bool LowerExpectIntrinsic::HandleIfExpect(BranchInst *BI) {
+  if (BI->isUnconditional())
+    return false;
+
+  // Handle non-optimized IR code like:
+  //   %expval = call i64 @llvm.expect.i64.i64(i64 %conv1, i64 1)
+  //   %tobool = icmp ne i64 %expval, 0
+  //   br i1 %tobool, label %if.then, label %if.end
+
+  ICmpInst *CmpI = dyn_cast<ICmpInst>(BI->getCondition());
+  if (!CmpI || CmpI->getPredicate() != CmpInst::ICMP_NE)
+    return false;
+
+  CallInst *CI = dyn_cast<CallInst>(CmpI->getOperand(0));
+  if (!CI)
+    return false;
+
+  Function *Fn = CI->getCalledFunction();
+  if (!Fn || Fn->getIntrinsicID() != Intrinsic::expect)
+    return false;
+
+  Value *ArgValue = CI->getArgOperand(0);
+  ConstantInt *ExpectedValue = dyn_cast<ConstantInt>(CI->getArgOperand(1));
+  if (!ExpectedValue)
+    return false;
+
+  LLVMContext &Context = CI->getContext();
+  const Type *Int32Ty = Type::getInt32Ty(Context);
+  bool Likely = ExpectedValue->isOne();
+
+  // If expect value is equal to 1 it means that we are more likely to take
+  // branch 0, in other case more likely is branch 1.
+  Value *Ops[] = {
+    MDString::get(Context, "branch_weights"),
+    ConstantInt::get(Int32Ty, Likely ? LikelyBranchWeight : UnlikelyBranchWeight),
+    ConstantInt::get(Int32Ty, Likely ? UnlikelyBranchWeight : LikelyBranchWeight)
+  };
+
+  MDNode *WeightsNode = MDNode::get(Context, ArrayRef<Value *>(Ops, 3));
+  BI->setMetadata(LLVMContext::MD_prof, WeightsNode);
+
+  CmpI->setOperand(0, ArgValue);
+  return true;
+}
+
+
+bool LowerExpectIntrinsic::runOnFunction(Function &F) {
+  for (Function::iterator I = F.begin(), E = F.end(); I != E;) {
+    BasicBlock *BB = I++;
+
+    // Create "block_weights" metadata.
+    if (BranchInst *BI = dyn_cast<BranchInst>(BB->getTerminator())) {
+      if (HandleIfExpect(BI))
+        IfHandled++;
+    } else if (SwitchInst *SI = dyn_cast<SwitchInst>(BB->getTerminator())) {
+      if (HandleSwitchExpect(SI))
+        IfHandled++;
+    }
+
+    // remove llvm.expect intrinsics.
+    for (BasicBlock::iterator BI = BB->begin(), BE = BB->end();
+         BI != BE; ) {
+      CallInst *CI = dyn_cast<CallInst>(BI++);
+      if (!CI)
+        continue;
+
+      Function *Fn = CI->getCalledFunction();
+      if (Fn && Fn->getIntrinsicID() == Intrinsic::expect)
+        CI->eraseFromParent();
+    }
+  }
+
+  return false;
+}
+
+
+char LowerExpectIntrinsic::ID = 0;
+INITIALIZE_PASS(LowerExpectIntrinsic, "lower-expect", "Lower 'expect' "
+                "Intrinsics", false, false)
+
+FunctionPass *llvm::createLowerExpectIntrinsicPass() {
+  return new LowerExpectIntrinsic();
+}
diff --git a/lib/VMCore/LLVMContext.cpp b/lib/VMCore/LLVMContext.cpp
index 1bd497d05d..ebd1e0aa1b 100644
--- a/lib/VMCore/LLVMContext.cpp
+++ b/lib/VMCore/LLVMContext.cpp
@@ -39,6 +39,10 @@ LLVMContext::LLVMContext() : pImpl(new LLVMContextImpl(*this)) {
   // Create the 'tbaa' metadata kind.
   unsigned TBAAID = getMDKindID("tbaa");
   assert(TBAAID == MD_tbaa && "tbaa kind id drifted"); (void)TBAAID;
+
+  // Create the 'prof' metadata kind.
+  unsigned ProfID = getMDKindID("prof");
+  assert(ProfID == MD_prof && "prof kind id drifted"); (void)ProfID;
 }
 LLVMContext::~LLVMContext() { delete pImpl; }