summaryrefslogtreecommitdiff
path: root/lib
diff options
context:
space:
mode:
authorJuergen Ributzka <juergen@apple.com>2014-01-24 18:23:08 +0000
committerJuergen Ributzka <juergen@apple.com>2014-01-24 18:23:08 +0000
commitfb282c68b7a31e8ee2810c81f5a6ed06e77cf7d1 (patch)
tree6b63d9b1a0642aff1c8bc52a62876348e27588bb /lib
parent8346f147ab6f06be4dac4af5c0e451a22bccf475 (diff)
downloadllvm-fb282c68b7a31e8ee2810c81f5a6ed06e77cf7d1.tar.gz
llvm-fb282c68b7a31e8ee2810c81f5a6ed06e77cf7d1.tar.bz2
llvm-fb282c68b7a31e8ee2810c81f5a6ed06e77cf7d1.tar.xz
Add Constant Hoisting Pass
This pass identifies expensive constants to hoist and coalesces them to better prepare it for SelectionDAG-based code generation. This works around the limitations of the basic-block-at-a-time approach. First it scans all instructions for integer constants and calculates its cost. If the constant can be folded into the instruction (the cost is TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't consider it expensive and leave it alone. This is the default behavior and the default implementation of getIntImmCost will always return TCC_Free. If the cost is more than TCC_BASIC, then the integer constant can't be folded into the instruction and it might be beneficial to hoist the constant. Similar constants are coalesced to reduce register pressure and materialization code. When a constant is hoisted, it is also hidden behind a bitcast to force it to be live-out of the basic block. Otherwise the constant would be just duplicated and each basic block would have its own copy in the SelectionDAG. The SelectionDAG recognizes such constants as opaque and doesn't perform certain transformations on them, which would create a new expensive constant. This optimization is only applied to integer constants in instructions and simple (this means not nested) constant cast experessions. For example: %0 = load i64* inttoptr (i64 big_constant to i64*) Reviewed by Eric git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200022 91177308-0d34-0410-b5e6-96231b3b80d8
Diffstat (limited to 'lib')
-rw-r--r--lib/Analysis/TargetTransformInfo.cpp22
-rw-r--r--lib/CodeGen/Passes.cpp6
-rw-r--r--lib/CodeGen/SelectionDAG/DAGCombiner.cpp9
-rw-r--r--lib/CodeGen/SelectionDAG/SelectionDAG.cpp41
-rw-r--r--lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp3
-rw-r--r--lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp10
-rw-r--r--lib/CodeGen/SelectionDAG/TargetLowering.cpp18
-rw-r--r--lib/Target/X86/X86TargetTransformInfo.cpp95
-rw-r--r--lib/Transforms/Scalar/CMakeLists.txt1
-rw-r--r--lib/Transforms/Scalar/CodeGenPrepare.cpp4
-rw-r--r--lib/Transforms/Scalar/ConstantHoisting.cpp429
-rw-r--r--lib/Transforms/Scalar/Scalar.cpp1
12 files changed, 611 insertions, 28 deletions
diff --git a/lib/Analysis/TargetTransformInfo.cpp b/lib/Analysis/TargetTransformInfo.cpp
index 360c4f5f1b..decb9aec4d 100644
--- a/lib/Analysis/TargetTransformInfo.cpp
+++ b/lib/Analysis/TargetTransformInfo.cpp
@@ -158,6 +158,16 @@ unsigned TargetTransformInfo::getIntImmCost(const APInt &Imm, Type *Ty) const {
return PrevTTI->getIntImmCost(Imm, Ty);
}
+unsigned TargetTransformInfo::getIntImmCost(unsigned Opcode, const APInt &Imm,
+ Type *Ty) const {
+ return PrevTTI->getIntImmCost(Opcode, Imm, Ty);
+}
+
+unsigned TargetTransformInfo::getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
+ Type *Ty) const {
+ return PrevTTI->getIntImmCost(IID, Imm, Ty);
+}
+
unsigned TargetTransformInfo::getNumberOfRegisters(bool Vector) const {
return PrevTTI->getNumberOfRegisters(Vector);
}
@@ -541,7 +551,17 @@ struct NoTTI LLVM_FINAL : ImmutablePass, TargetTransformInfo {
}
unsigned getIntImmCost(const APInt &Imm, Type *Ty) const LLVM_OVERRIDE {
- return 1;
+ return TCC_Basic;
+ }
+
+ unsigned getIntImmCost(unsigned Opcode, const APInt &Imm,
+ Type *Ty) const LLVM_OVERRIDE {
+ return TCC_Free;
+ }
+
+ unsigned getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
+ Type *Ty) const LLVM_OVERRIDE {
+ return TCC_Free;
}
unsigned getNumberOfRegisters(bool Vector) const LLVM_OVERRIDE {
diff --git a/lib/CodeGen/Passes.cpp b/lib/CodeGen/Passes.cpp
index 0e8e50eca7..7a7c42bfcc 100644
--- a/lib/CodeGen/Passes.cpp
+++ b/lib/CodeGen/Passes.cpp
@@ -70,6 +70,8 @@ static cl::opt<bool> DisableMachineSink("disable-machine-sink", cl::Hidden,
cl::desc("Disable Machine Sinking"));
static cl::opt<bool> DisableLSR("disable-lsr", cl::Hidden,
cl::desc("Disable Loop Strength Reduction Pass"));
+static cl::opt<bool> DisableConstantHoisting("disable-constant-hoisting",
+ cl::Hidden, cl::desc("Disable ConstantHoisting"));
static cl::opt<bool> DisableCGP("disable-cgp", cl::Hidden,
cl::desc("Disable Codegen Prepare"));
static cl::opt<bool> DisableCopyProp("disable-copyprop", cl::Hidden,
@@ -396,6 +398,10 @@ void TargetPassConfig::addIRPasses() {
// Make sure that no unreachable blocks are instruction selected.
addPass(createUnreachableBlockEliminationPass());
+
+ // Prepare expensive constants for SelectionDAG.
+ if (getOptLevel() != CodeGenOpt::None && !DisableConstantHoisting)
+ addPass(createConstantHoistingPass());
}
/// Turn exception handling constructs into something the code generators can
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 2dafcb9fa9..04fe316ca6 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -3212,11 +3212,14 @@ SDValue DAGCombiner::visitOR(SDNode *N) {
if (N1C && N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
isa<ConstantSDNode>(N0.getOperand(1))) {
ConstantSDNode *C1 = cast<ConstantSDNode>(N0.getOperand(1));
- if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0)
+ if ((C1->getAPIntValue() & N1C->getAPIntValue()) != 0) {
+ SDValue COR = DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1);
+ if (!COR.getNode())
+ return SDValue();
return DAG.getNode(ISD::AND, SDLoc(N), VT,
DAG.getNode(ISD::OR, SDLoc(N0), VT,
- N0.getOperand(0), N1),
- DAG.FoldConstantArithmetic(ISD::OR, VT, N1C, C1));
+ N0.getOperand(0), N1), COR);
+ }
}
// fold (or (setcc x), (setcc y)) -> (setcc (or x, y))
if (isSetCCEquivalent(N0, LL, LR, CC0) && isSetCCEquivalent(N1, RL, RR, CC1)){
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index e003caeddb..430736aa46 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -384,9 +384,12 @@ static void AddNodeIDCustom(FoldingSetNodeID &ID, const SDNode *N) {
llvm_unreachable("Should only be used on nodes with operands");
default: break; // Normal nodes don't need extra info.
case ISD::TargetConstant:
- case ISD::Constant:
- ID.AddPointer(cast<ConstantSDNode>(N)->getConstantIntValue());
+ case ISD::Constant: {
+ const ConstantSDNode *C = cast<ConstantSDNode>(N);
+ ID.AddPointer(C->getConstantIntValue());
+ ID.AddBoolean(C->isOpaque());
break;
+ }
case ISD::TargetConstantFP:
case ISD::ConstantFP: {
ID.AddPointer(cast<ConstantFPSDNode>(N)->getConstantFPValue());
@@ -971,19 +974,21 @@ SDValue SelectionDAG::getNOT(SDLoc DL, SDValue Val, EVT VT) {
return getNode(ISD::XOR, DL, VT, Val, NegOne);
}
-SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT) {
+SDValue SelectionDAG::getConstant(uint64_t Val, EVT VT, bool isT, bool isO) {
EVT EltVT = VT.getScalarType();
assert((EltVT.getSizeInBits() >= 64 ||
(uint64_t)((int64_t)Val >> EltVT.getSizeInBits()) + 1 < 2) &&
"getConstant with a uint64_t value that doesn't fit in the type!");
- return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT);
+ return getConstant(APInt(EltVT.getSizeInBits(), Val), VT, isT, isO);
}
-SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT) {
- return getConstant(*ConstantInt::get(*Context, Val), VT, isT);
+SDValue SelectionDAG::getConstant(const APInt &Val, EVT VT, bool isT, bool isO)
+{
+ return getConstant(*ConstantInt::get(*Context, Val), VT, isT, isO);
}
-SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
+SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT,
+ bool isO) {
assert(VT.isInteger() && "Cannot create FP integer constant!");
EVT EltVT = VT.getScalarType();
@@ -1025,7 +1030,7 @@ SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
for (unsigned i = 0; i < ViaVecNumElts / VT.getVectorNumElements(); ++i) {
EltParts.push_back(getConstant(NewVal.lshr(i * ViaEltSizeInBits)
.trunc(ViaEltSizeInBits),
- ViaEltVT, isT));
+ ViaEltVT, isT, isO));
}
// EltParts is currently in little endian order. If we actually want
@@ -1056,6 +1061,7 @@ SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
FoldingSetNodeID ID;
AddNodeIDNode(ID, Opc, getVTList(EltVT), 0, 0);
ID.AddPointer(Elt);
+ ID.AddBoolean(isO);
void *IP = 0;
SDNode *N = NULL;
if ((N = CSEMap.FindNodeOrInsertPos(ID, IP)))
@@ -1063,7 +1069,7 @@ SDValue SelectionDAG::getConstant(const ConstantInt &Val, EVT VT, bool isT) {
return SDValue(N, 0);
if (!N) {
- N = new (NodeAllocator) ConstantSDNode(isT, Elt, EltVT);
+ N = new (NodeAllocator) ConstantSDNode(isT, isO, Elt, EltVT);
CSEMap.InsertNode(N, IP);
AllNodes.push_back(N);
}
@@ -2789,10 +2795,13 @@ SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, EVT VT,
ConstantSDNode *Scalar1 = dyn_cast<ConstantSDNode>(Cst1);
ConstantSDNode *Scalar2 = dyn_cast<ConstantSDNode>(Cst2);
- if (Scalar1 && Scalar2) {
+ if (Scalar1 && Scalar2 && (Scalar1->isOpaque() || Scalar2->isOpaque()))
+ return SDValue();
+
+ if (Scalar1 && Scalar2)
// Scalar instruction.
Inputs.push_back(std::make_pair(Scalar1, Scalar2));
- } else {
+ else {
// For vectors extract each constant element into Inputs so we can constant
// fold them individually.
BuildVectorSDNode *BV1 = dyn_cast<BuildVectorSDNode>(Cst1);
@@ -2808,6 +2817,9 @@ SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, EVT VT,
if (!V1 || !V2) // Not a constant, bail.
return SDValue();
+ if (V1->isOpaque() || V2->isOpaque())
+ return SDValue();
+
// Avoid BUILD_VECTOR nodes that perform implicit truncation.
// FIXME: This is valid and could be handled by truncating the APInts.
if (V1->getValueType(0) != SVT || V2->getValueType(0) != SVT)
@@ -3561,10 +3573,11 @@ static SDValue getMemsetStringVal(EVT VT, SDLoc dl, SelectionDAG &DAG,
Val |= (uint64_t)(unsigned char)Str[i] << (NumVTBytes-i-1)*8;
}
- // If the "cost" of materializing the integer immediate is 1 or free, then
- // it is cost effective to turn the load into the immediate.
+ // If the "cost" of materializing the integer immediate is less than the cost
+ // of a load, then it is cost effective to turn the load into the immediate.
const TargetTransformInfo *TTI = DAG.getTargetTransformInfo();
- if (TTI->getIntImmCost(Val, VT.getTypeForEVT(*DAG.getContext())) < 2)
+ if (TTI->getIntImmCost(Val, VT.getTypeForEVT(*DAG.getContext())) <
+ TargetTransformInfo::TCC_Load)
return DAG.getConstant(Val, VT);
return SDValue(0, 0);
}
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index 6c8e2f83a4..a561c5190a 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -2945,6 +2945,9 @@ void SelectionDAGBuilder::visitBitCast(const User &I) {
if (DestVT != N.getValueType())
setValue(&I, DAG.getNode(ISD::BITCAST, getCurSDLoc(),
DestVT, N)); // convert types.
+ else if(ConstantSDNode *C = dyn_cast<ConstantSDNode>(N))
+ setValue(&I, DAG.getConstant(C->getAPIntValue(), C->getValueType(0),
+ /*isTarget=*/false, /*isOpaque*/true));
else
setValue(&I, N); // noop cast.
}
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
index 79377f7424..6f980884be 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
@@ -81,7 +81,10 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
case ISD::VALUETYPE: return "ValueType";
case ISD::Register: return "Register";
case ISD::RegisterMask: return "RegisterMask";
- case ISD::Constant: return "Constant";
+ case ISD::Constant:
+ if (cast<ConstantSDNode>(this)->isOpaque())
+ return "OpaqueConstant";
+ return "Constant";
case ISD::ConstantFP: return "ConstantFP";
case ISD::GlobalAddress: return "GlobalAddress";
case ISD::GlobalTLSAddress: return "GlobalTLSAddress";
@@ -111,7 +114,10 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
}
case ISD::BUILD_VECTOR: return "BUILD_VECTOR";
- case ISD::TargetConstant: return "TargetConstant";
+ case ISD::TargetConstant:
+ if (cast<ConstantSDNode>(this)->isOpaque())
+ return "OpaqueTargetConstant";
+ return "TargetConstant";
case ISD::TargetConstantFP: return "TargetConstantFP";
case ISD::TargetGlobalAddress: return "TargetGlobalAddress";
case ISD::TargetGlobalTLSAddress: return "TargetGlobalTLSAddress";
diff --git a/lib/CodeGen/SelectionDAG/TargetLowering.cpp b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
index fa5f9b4c94..6052a48640 100644
--- a/lib/CodeGen/SelectionDAG/TargetLowering.cpp
+++ b/lib/CodeGen/SelectionDAG/TargetLowering.cpp
@@ -1470,17 +1470,23 @@ TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1,
if (Cond == ISD::SETGE || Cond == ISD::SETUGE) {
if (C1 == MinVal) return DAG.getConstant(1, VT); // X >= MIN --> true
// X >= C0 --> X > (C0-1)
- return DAG.getSetCC(dl, VT, N0,
- DAG.getConstant(C1-1, N1.getValueType()),
- (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
+ APInt C = C1-1;
+ if (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
+ isLegalICmpImmediate(C.getSExtValue())))
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(C, N1.getValueType()),
+ (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT);
}
if (Cond == ISD::SETLE || Cond == ISD::SETULE) {
if (C1 == MaxVal) return DAG.getConstant(1, VT); // X <= MAX --> true
// X <= C0 --> X < (C0+1)
- return DAG.getSetCC(dl, VT, N0,
- DAG.getConstant(C1+1, N1.getValueType()),
- (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
+ APInt C = C1+1;
+ if (!N1C->isOpaque() || (N1C->isOpaque() && C.getBitWidth() <= 64 &&
+ isLegalICmpImmediate(C.getSExtValue())))
+ return DAG.getSetCC(dl, VT, N0,
+ DAG.getConstant(C, N1.getValueType()),
+ (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT);
}
if ((Cond == ISD::SETLT || Cond == ISD::SETULT) && C1 == MinVal)
diff --git a/lib/Target/X86/X86TargetTransformInfo.cpp b/lib/Target/X86/X86TargetTransformInfo.cpp
index da2b021da9..781be2fddd 100644
--- a/lib/Target/X86/X86TargetTransformInfo.cpp
+++ b/lib/Target/X86/X86TargetTransformInfo.cpp
@@ -18,6 +18,7 @@
#include "X86.h"
#include "X86TargetMachine.h"
#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/IntrinsicInst.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/CostTable.h"
#include "llvm/Target/TargetLowering.h"
@@ -107,6 +108,14 @@ public:
virtual unsigned getReductionCost(unsigned Opcode, Type *Ty,
bool IsPairwiseForm) const LLVM_OVERRIDE;
+ virtual unsigned getIntImmCost(const APInt &Imm,
+ Type *Ty) const LLVM_OVERRIDE;
+
+ virtual unsigned getIntImmCost(unsigned Opcode, const APInt &Imm,
+ Type *Ty) const LLVM_OVERRIDE;
+ virtual unsigned getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
+ Type *Ty) const LLVM_OVERRIDE;
+
/// @}
};
@@ -694,3 +703,89 @@ unsigned X86TTI::getReductionCost(unsigned Opcode, Type *ValTy,
return TargetTransformInfo::getReductionCost(Opcode, ValTy, IsPairwise);
}
+unsigned X86TTI::getIntImmCost(const APInt &Imm, Type *Ty) const {
+ assert(Ty->isIntegerTy());
+
+ unsigned BitSize = Ty->getPrimitiveSizeInBits();
+ if (BitSize == 0)
+ return ~0U;
+
+ if (Imm.getBitWidth() <= 64 &&
+ (isInt<32>(Imm.getSExtValue()) || isUInt<32>(Imm.getZExtValue())))
+ return TCC_Basic;
+ else
+ return 2 * TCC_Basic;
+}
+
+unsigned X86TTI::getIntImmCost(unsigned Opcode, const APInt &Imm,
+ Type *Ty) const {
+ assert(Ty->isIntegerTy());
+
+ unsigned BitSize = Ty->getPrimitiveSizeInBits();
+ if (BitSize == 0)
+ return ~0U;
+
+ switch (Opcode) {
+ case Instruction::Add:
+ case Instruction::Sub:
+ case Instruction::Mul:
+ case Instruction::UDiv:
+ case Instruction::SDiv:
+ case Instruction::URem:
+ case Instruction::SRem:
+ case Instruction::Shl:
+ case Instruction::LShr:
+ case Instruction::AShr:
+ case Instruction::And:
+ case Instruction::Or:
+ case Instruction::Xor:
+ case Instruction::ICmp:
+ if (Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
+ return TCC_Free;
+ else
+ return X86TTI::getIntImmCost(Imm, Ty);
+ case Instruction::Trunc:
+ case Instruction::ZExt:
+ case Instruction::SExt:
+ case Instruction::IntToPtr:
+ case Instruction::PtrToInt:
+ case Instruction::BitCast:
+ case Instruction::Call:
+ case Instruction::Select:
+ case Instruction::Ret:
+ case Instruction::Load:
+ case Instruction::Store:
+ return X86TTI::getIntImmCost(Imm, Ty);
+ }
+ return TargetTransformInfo::getIntImmCost(Opcode, Imm, Ty);
+}
+
+unsigned X86TTI::getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
+ Type *Ty) const {
+ assert(Ty->isIntegerTy());
+
+ unsigned BitSize = Ty->getPrimitiveSizeInBits();
+ if (BitSize == 0)
+ return ~0U;
+
+ switch (IID) {
+ default: return TargetTransformInfo::getIntImmCost(IID, Imm, Ty);
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::uadd_with_overflow:
+ case Intrinsic::ssub_with_overflow:
+ case Intrinsic::usub_with_overflow:
+ case Intrinsic::smul_with_overflow:
+ case Intrinsic::umul_with_overflow:
+ if (Imm.getBitWidth() <= 64 && isInt<32>(Imm.getSExtValue()))
+ return TCC_Free;
+ else
+ return X86TTI::getIntImmCost(Imm, Ty);
+ case Intrinsic::experimental_stackmap:
+ case Intrinsic::experimental_patchpoint_void:
+ case Intrinsic::experimental_patchpoint_i64:
+ if (Imm.getBitWidth() <= 64 && isInt<64>(Imm.getSExtValue()))
+ return TCC_Free;
+ else
+ return X86TTI::getIntImmCost(Imm, Ty);
+ }
+}
diff --git a/lib/Transforms/Scalar/CMakeLists.txt b/lib/Transforms/Scalar/CMakeLists.txt
index 0b2928677c..8a29b0c48c 100644
--- a/lib/Transforms/Scalar/CMakeLists.txt
+++ b/lib/Transforms/Scalar/CMakeLists.txt
@@ -1,6 +1,7 @@
add_llvm_library(LLVMScalarOpts
ADCE.cpp
CodeGenPrepare.cpp
+ ConstantHoisting.cpp
ConstantProp.cpp
CorrelatedValuePropagation.cpp
DCE.cpp
diff --git a/lib/Transforms/Scalar/CodeGenPrepare.cpp b/lib/Transforms/Scalar/CodeGenPrepare.cpp
index 38f587b2cc..6acbd5eaa1 100644
--- a/lib/Transforms/Scalar/CodeGenPrepare.cpp
+++ b/lib/Transforms/Scalar/CodeGenPrepare.cpp
@@ -240,7 +240,7 @@ bool CodeGenPrepare::runOnFunction(Function &F) {
bool CodeGenPrepare::EliminateFallThrough(Function &F) {
bool Changed = false;
// Scan all of the blocks in the function, except for the entry block.
- for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
+ for (Function::iterator I = llvm::next(F.begin()), E = F.end(); I != E; ) {
BasicBlock *BB = I++;
// If the destination block has a single pred, then this is a trivial
// edge, just collapse it.
@@ -276,7 +276,7 @@ bool CodeGenPrepare::EliminateFallThrough(Function &F) {
bool CodeGenPrepare::EliminateMostlyEmptyBlocks(Function &F) {
bool MadeChange = false;
// Note that this intentionally skips the entry block.
- for (Function::iterator I = ++F.begin(), E = F.end(); I != E; ) {
+ for (Function::iterator I = llvm::next(F.begin()), E = F.end(); I != E; ) {
BasicBlock *BB = I++;
// If this block doesn't end with an uncond branch, ignore it.
diff --git a/lib/Transforms/Scalar/ConstantHoisting.cpp b/lib/Transforms/Scalar/ConstantHoisting.cpp
new file mode 100644
index 0000000000..b6191bbfbf
--- /dev/null
+++ b/lib/Transforms/Scalar/ConstantHoisting.cpp
@@ -0,0 +1,429 @@
+//===- ConstantHoisting.cpp - Prepare code for expensive constants --------===//
+//
+// The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass identifies expensive constants to hoist and coalesces them to
+// better prepare it for SelectionDAG-based code generation. This works around
+// the limitations of the basic-block-at-a-time approach.
+//
+// First it scans all instructions for integer constants and calculates its
+// cost. If the constant can be folded into the instruction (the cost is
+// TCC_Free) or the cost is just a simple operation (TCC_BASIC), then we don't
+// consider it expensive and leave it alone. This is the default behavior and
+// the default implementation of getIntImmCost will always return TCC_Free.
+//
+// If the cost is more than TCC_BASIC, then the integer constant can't be folded
+// into the instruction and it might be beneficial to hoist the constant.
+// Similar constants are coalesced to reduce register pressure and
+// materialization code.
+//
+// When a constant is hoisted, it is also hidden behind a bitcast to force it to
+// be live-out of the basic block. Otherwise the constant would be just
+// duplicated and each basic block would have its own copy in the SelectionDAG.
+// The SelectionDAG recognizes such constants as opaque and doesn't perform
+// certain transformations on them, which would create a new expensive constant.
+//
+// This optimization is only applied to integer constants in instructions and
+// simple (this means not nested) constant cast experessions. For example:
+// %0 = load i64* inttoptr (i64 big_constant to i64*)
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "consthoist"
+#include "llvm/Transforms/Scalar.h"
+#include "llvm/ADT/MapVector.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/TargetTransformInfo.h"
+#include "llvm/IR/Constants.h"
+#include "llvm/IR/Dominators.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/Pass.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+
+using namespace llvm;
+
+STATISTIC(NumConstantsHoisted, "Number of constants hoisted");
+STATISTIC(NumConstantsRebased, "Number of constants rebased");
+
+
+namespace {
+typedef SmallVector<User *, 4> ConstantUseListType;
+struct ConstantCandidate {
+ unsigned CumulativeCost;
+ ConstantUseListType Uses;
+};
+
+struct ConstantInfo {
+ ConstantInt *BaseConstant;
+ struct RebasedConstantInfo {
+ ConstantInt *OriginalConstant;
+ Constant *Offset;
+ ConstantUseListType Uses;
+ };
+ typedef SmallVector<RebasedConstantInfo, 4> RebasedConstantListType;
+ RebasedConstantListType RebasedConstants;
+};
+
+class ConstantHoisting : public FunctionPass {
+ const TargetTransformInfo *TTI;
+ DominatorTree *DT;
+
+ /// Keeps track of expensive constants found in the function.
+ typedef MapVector<ConstantInt *, ConstantCandidate> ConstantMapType;
+ ConstantMapType ConstantMap;
+
+ /// These are the final constants we decided to hoist.
+ SmallVector<ConstantInfo, 4> Constants;
+public:
+ static char ID; // Pass identification, replacement for typeid
+ ConstantHoisting() : FunctionPass(ID), TTI(0) {
+ initializeConstantHoistingPass(*PassRegistry::getPassRegistry());
+ }
+
+ bool runOnFunction(Function &F);
+
+ const char *getPassName() const { return "Constant Hoisting"; }
+
+ virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+ AU.setPreservesCFG();
+ AU.addRequired<DominatorTreeWrapperPass>();
+ AU.addRequired<TargetTransformInfo>();
+ }
+
+private:
+ void CollectConstant(User *U, unsigned Opcode, Intrinsic::ID IID,
+ ConstantInt *C);
+ void CollectConstants(Instruction *I);
+ void CollectConstants(Function &F);
+ void FindAndMakeBaseConstant(ConstantMapType::iterator S,
+ ConstantMapType::iterator E);
+ void FindBaseConstants();
+ Instruction *FindConstantInsertionPoint(Function &F,
+ const ConstantInfo &CI) const;
+ void EmitBaseConstants(Function &F, User *U, Instruction *Base,
+ Constant *Offset, ConstantInt *OriginalConstant);
+ bool EmitBaseConstants(Function &F);
+ bool OptimizeConstants(Function &F);
+};
+}
+
+char ConstantHoisting::ID = 0;
+INITIALIZE_PASS_BEGIN(ConstantHoisting, "consthoist", "Constant Hoisting",
+ false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
+INITIALIZE_AG_DEPENDENCY(TargetTransformInfo)
+INITIALIZE_PASS_END(ConstantHoisting, "consthoist", "Constant Hoisting",
+ false, false)
+
+FunctionPass *llvm::createConstantHoistingPass() {
+ return new ConstantHoisting();
+}
+
+/// \brief Perform the constant hoisting optimization for the given function.
+bool ConstantHoisting::runOnFunction(Function &F) {
+ DEBUG(dbgs() << "********** Constant Hoisting **********\n");
+ DEBUG(dbgs() << "********** Function: " << F.getName() << '\n');
+
+ DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
+ TTI = &getAnalysis<TargetTransformInfo>();
+
+ return OptimizeConstants(F);
+}
+
+void ConstantHoisting::CollectConstant(User * U, unsigned Opcode,
+ Intrinsic::ID IID, ConstantInt *C) {
+ unsigned Cost;
+ if (Opcode)
+ Cost = TTI->getIntImmCost(Opcode, C->getValue(), C->getType());
+ else
+ Cost = TTI->getIntImmCost(IID, C->getValue(), C->getType());
+
+ if (Cost > TargetTransformInfo::TCC_Basic) {
+ ConstantCandidate &CC = ConstantMap[C];
+ CC.CumulativeCost += Cost;
+ CC.Uses.push_back(U);
+ }
+}
+
+/// \brief Scan the instruction or constant expression for expensive integer
+/// constants and record them in the constant map.
+void ConstantHoisting::CollectConstants(Instruction *I) {
+ unsigned Opcode = 0;
+ Intrinsic::ID IID = Intrinsic::not_intrinsic;
+ if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
+ IID = II->getIntrinsicID();
+ else
+ Opcode = I->getOpcode();
+
+ // Scan all operands.
+ for (User::op_iterator O = I->op_begin(), E = I->op_end(); O != E; ++O) {
+ if (ConstantInt *C = dyn_cast<ConstantInt>(O)) {
+ CollectConstant(I, Opcode, IID, C);
+ continue;
+ }
+ if (ConstantExpr *CE = dyn_cast<ConstantExpr>(O)) {
+ // We only handle constant cast expressions.
+ if (!CE->isCast())
+ continue;
+
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CE->getOperand(0))) {
+ // Ignore the cast expression and use the opcode of the instruction.
+ CollectConstant(CE, Opcode, IID, C);
+ continue;
+ }
+ }
+ }
+}
+
+/// \brief Collect all integer constants in the function that cannot be folded
+/// into an instruction itself.
+void ConstantHoisting::CollectConstants(Function &F) {
+ for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
+ for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
+ CollectConstants(I);
+}
+
+/// \brief Compare function for sorting integer constants by type and by value
+/// within a type in ConstantMaps.
+static bool
+ConstantMapLessThan(const std::pair<ConstantInt *, ConstantCandidate> &LHS,
+ const std::pair<ConstantInt *, ConstantCandidate> &RHS) {
+ if (LHS.first->getType() == RHS.first->getType())
+ return LHS.first->getValue().ult(RHS.first->getValue());
+ else
+ return LHS.first->getType()->getBitWidth() <
+ RHS.first->getType()->getBitWidth();
+}
+
+/// \brief Find the base constant within the given range and rebase all other
+/// constants with respect to the base constant.
+void ConstantHoisting::FindAndMakeBaseConstant(ConstantMapType::iterator S,
+ ConstantMapType::iterator E) {
+ ConstantMapType::iterator MaxCostItr = S;
+ unsigned NumUses = 0;
+ // Use the constant that has the maximum cost as base constant.
+ for (ConstantMapType::iterator I = S; I != E; ++I) {
+ NumUses += I->second.Uses.size();
+ if (I->second.CumulativeCost > MaxCostItr->second.CumulativeCost)
+ MaxCostItr = I;
+ }
+
+ // Don't hoist constants that have only one use.
+ if (NumUses <= 1)
+ return;
+
+ ConstantInfo CI;
+ CI.BaseConstant = MaxCostItr->first;
+ Type *Ty = CI.BaseConstant->getType();
+ // Rebase the constants with respect to the base constant.
+ for (ConstantMapType::iterator I = S; I != E; ++I) {
+ APInt Diff = I->first->getValue() - CI.BaseConstant->getValue();
+ ConstantInfo::RebasedConstantInfo RCI;
+ RCI.OriginalConstant = I->first;
+ RCI.Offset = ConstantInt::get(Ty, Diff);
+ RCI.Uses = llvm_move(I->second.Uses);
+ CI.RebasedConstants.push_back(RCI);
+ }
+ Constants.push_back(CI);
+}
+
+/// \brief Finds and combines constants that can be easily rematerialized with
+/// an add from a common base constant.
+void ConstantHoisting::FindBaseConstants() {
+ // Sort the constants by value and type. This invalidates the mapping.
+ std::sort(ConstantMap.begin(), ConstantMap.end(), ConstantMapLessThan);
+
+ // Simple linear scan through the sorted constant map for viable merge
+ // candidates.
+ ConstantMapType::iterator MinValItr = ConstantMap.begin();
+ for (ConstantMapType::iterator I = llvm::next(ConstantMap.begin()),
+ E = ConstantMap.end(); I != E; ++I) {
+ if (MinValItr->first->getType() == I->first->getType()) {
+ // Check if the constant is in range of an add with immediate.
+ APInt Diff = I->first->getValue() - MinValItr->first->getValue();
+ if ((Diff.getBitWidth() <= 64) &&
+ TTI->isLegalAddImmediate(Diff.getSExtValue()))
+ continue;
+ }
+ // We either have now a different constant type or the constant is not in
+ // range of an add with immediate anymore.
+ FindAndMakeBaseConstant(MinValItr, I);
+ // Start a new base constant search.
+ MinValItr = I;
+ }
+ // Finalize the last base constant search.
+ FindAndMakeBaseConstant(MinValItr, ConstantMap.end());
+}
+
+/// \brief Records the basic block of the instruction or all basic blocks of the
+/// users of the constant expression.
+static void CollectBasicBlocks(SmallPtrSet<BasicBlock *, 4> &BBs, User *U) {
+ if (Instruction *I = dyn_cast<Instruction>(U))
+ BBs.insert(I->getParent());
+ else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U))
+ // Find all users of this constant expression.
+ for (Value::use_iterator UU = CE->use_begin(), E = CE->use_end();
+ UU != E; ++UU)
+ // Only record users that are instructions. We don't want to go down a
+ // nested constant expression chain.
+ if (Instruction *I = dyn_cast<Instruction>(*UU))
+ BBs.insert(I->getParent());
+}
+
+/// \brief Find an insertion point that dominates all uses.
+Instruction *ConstantHoisting::
+FindConstantInsertionPoint(Function &F, const ConstantInfo &CI) const {
+ BasicBlock *Entry = &F.getEntryBlock();
+
+ // Collect all basic blocks.
+ SmallPtrSet<BasicBlock *, 4> BBs;
+ ConstantInfo::RebasedConstantListType::const_iterator RCI, RCE;
+ for (RCI = CI.RebasedConstants.begin(), RCE = CI.RebasedConstants.end();
+ RCI != RCE; ++RCI)
+ for (SmallVectorImpl<User *>::const_iterator U = RCI->Uses.begin(),
+ E = RCI->Uses.end(); U != E; ++U)
+ CollectBasicBlocks(BBs, *U);
+
+ if (BBs.count(Entry))
+ return Entry->getFirstInsertionPt();
+
+ while (BBs.size() >= 2) {
+ BasicBlock *BB, *BB1, *BB2;
+ BB1 = *BBs.begin();
+ BB2 = *llvm::next(BBs.begin());
+ BB = DT->findNearestCommonDominator(BB1, BB2);
+ if (BB == Entry)
+ return Entry->getFirstInsertionPt();
+ BBs.erase(BB1);
+ BBs.erase(BB2);
+ BBs.insert(BB);
+ }
+ assert((BBs.size() == 1) && "Expected only one element.");
+ return (*BBs.begin())->getFirstInsertionPt();
+}
+
+/// \brief Emit materialization code for all rebased constants and update their
+/// users.
+void ConstantHoisting::EmitBaseConstants(Function &F, User *U,
+ Instruction *Base, Constant *Offset,
+ ConstantInt *OriginalConstant) {
+ if (Instruction *I = dyn_cast<Instruction>(U)) {
+ Instruction *Mat = Base;
+ if (!Offset->isNullValue()) {
+ Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
+ "const_mat", I);
+
+ // Use the same debug location as the instruction we are about to update.
+ Mat->setDebugLoc(I->getDebugLoc());
+
+ DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
+ << " + " << *Offset << ") in BB "
+ << I->getParent()->getName() << '\n' << *Mat << '\n');
+ }
+ DEBUG(dbgs() << "Update: " << *I << '\n');
+ I->replaceUsesOfWith(OriginalConstant, Mat);
+ DEBUG(dbgs() << "To: " << *I << '\n');
+ return;
+ }
+ assert(isa<ConstantExpr>(U) && "Expected a ConstantExpr.");
+ ConstantExpr *CE = cast<ConstantExpr>(U);
+ for (Value::use_iterator UU = CE->use_begin(), E = CE->use_end();
+ UU != E; ++UU) {
+ // We only handel instructions here and won't walk down a ConstantExpr chain
+ // to replace all ConstExpr with instructions.
+ if (Instruction *I = dyn_cast<Instruction>(*UU)) {
+ Instruction *Mat = Base;
+ if (!Offset->isNullValue()) {
+ Mat = BinaryOperator::Create(Instruction::Add, Base, Offset,
+ "const_mat", I);
+
+ // Use the same debug location as the instruction we are about to
+ // update.
+ Mat->setDebugLoc(I->getDebugLoc());
+
+ DEBUG(dbgs() << "Materialize constant (" << *Base->getOperand(0)
+ << " + " << *Offset << ") in BB "
+ << I->getParent()->getName() << '\n' << *Mat << '\n');
+ }
+ Instruction *ICE = CE->getAsInstruction();
+ ICE->replaceUsesOfWith(OriginalConstant, Mat);
+ ICE->insertBefore(I);
+
+ // Use the same debug location as the instruction we are about to update.
+ ICE->setDebugLoc(I->getDebugLoc());
+
+ DEBUG(dbgs() << "Create instruction: " << *ICE << '\n');
+ DEBUG(dbgs() << "Update: " << *I << '\n');
+ I->replaceUsesOfWith(CE, ICE);
+ DEBUG(dbgs() << "To: " << *I << '\n');
+ }
+ }
+}
+
+/// \brief Hoist and hide the base constant behind a bitcast and emit
+/// materialization code for derived constants.
+bool ConstantHoisting::EmitBaseConstants(Function &F) {
+ bool MadeChange = false;
+ SmallVectorImpl<ConstantInfo>::iterator CI, CE;
+ for (CI = Constants.begin(), CE = Constants.end(); CI != CE; ++CI) {
+ // Hoist and hide the base constant behind a bitcast.
+ Instruction *IP = FindConstantInsertionPoint(F, *CI);
+ IntegerType *Ty = CI->BaseConstant->getType();
+ Instruction *Base = new BitCastInst(CI->BaseConstant, Ty, "const", IP);
+ DEBUG(dbgs() << "Hoist constant (" << *CI->BaseConstant << ") to BB "
+ << IP->getParent()->getName() << '\n');
+ NumConstantsHoisted++;
+
+ // Emit materialization code for all rebased constants.
+ ConstantInfo::RebasedConstantListType::iterator RCI, RCE;
+ for (RCI = CI->RebasedConstants.begin(), RCE = CI->RebasedConstants.end();
+ RCI != RCE; ++RCI) {
+ NumConstantsRebased++;
+ for (SmallVectorImpl<User *>::iterator U = RCI->Uses.begin(),
+ E = RCI->Uses.end(); U != E; ++U)
+ EmitBaseConstants(F, *U, Base, RCI->Offset, RCI->OriginalConstant);
+ }
+
+ // Use the same debug location as the last user of the constant.
+ assert(!Base->use_empty() && "The use list is empty!?");
+ assert(isa<Instruction>(Base->use_back()) &&
+ "All uses should be instructions.");
+ Base->setDebugLoc(cast<Instruction>(Base->use_back())->getDebugLoc());
+
+ // Correct for base constant, which we counted above too.
+ NumConstantsRebased--;
+ MadeChange = true;
+ }
+ return MadeChange;
+}
+
+/// \brief Optimize expensive integer constants in the given function.
+bool ConstantHoisting::OptimizeConstants(Function &F) {
+ bool MadeChange = false;
+
+ // Collect all constant candidates.
+ CollectConstants(F);
+
+ // There are no constants to worry about.
+ if (ConstantMap.empty())
+ return MadeChange;
+
+ // Combine constants that can be easily materialized with an add from a common
+ // base constant.
+ FindBaseConstants();
+
+ // Finaly hoist the base constant and emit materializating code for dependent
+ // constants.
+ MadeChange |= EmitBaseConstants(F);
+
+ ConstantMap.clear();
+ Constants.clear();
+
+ return MadeChange;
+}
diff --git a/lib/Transforms/Scalar/Scalar.cpp b/lib/Transforms/Scalar/Scalar.cpp
index 6983d06b68..f774874cb4 100644
--- a/lib/Transforms/Scalar/Scalar.cpp
+++ b/lib/Transforms/Scalar/Scalar.cpp
@@ -30,6 +30,7 @@ void llvm::initializeScalarOpts(PassRegistry &Registry) {
initializeADCEPass(Registry);
initializeSampleProfileLoaderPass(Registry);
initializeCodeGenPreparePass(Registry);
+ initializeConstantHoistingPass(Registry);
initializeConstantPropagationPass(Registry);
initializeCorrelatedValuePropagationPass(Registry);
initializeDCEPass(Registry);