RPCS3/llvm-mirror
1
0
Fork 0
mirror of https://github.com/RPCS3/llvm-mirror.git synced 2024-10-19 02:52:53 +02:00
Code Issues Projects Releases Wiki Activity

Add Constant Hoisting Pass

Browse Source 
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

llvm-svn: 200022
This commit is contained in:
Juergen Ributzka 2014-01-24 18:23:08 +00:00
parent e4d29eb495
commit fa4fb4d6a4
20 changed files with 701 additions and 40 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
include/llvm/Analysis/TargetTransformInfo.h
View File

@ -92,6 +92,7 @@ public:
enum TargetCostConstants {
TCC_Free = 0, ///< Expected to fold away in lowering.
TCC_Basic = 1, ///< The cost of a typical 'add' instruction.
TCC_Load = 3,
TCC_Expensive = 4 ///< The cost of a 'div' instruction on x86.
};
@ -299,6 +300,13 @@ public:
/// immediate of the specified type.
virtual unsigned getIntImmCost(const APInt &Imm, Type *Ty) const;
/// \brief Return the expected cost of materialization for the given integer
/// immediate of the specified type for a given instruction. The cost can be
/// zero if the immediate can be folded into the specified instruction.
virtual unsigned getIntImmCost(unsigned Opcode, const APInt &Imm,
Type *Ty) const;
virtual unsigned getIntImmCost(Intrinsic::ID IID, const APInt &Imm,
Type *Ty) const;
/// @}
/// \name Vector Target Information

22
include/llvm/CodeGen/SelectionDAG.h
View File

@ -401,18 +401,22 @@ public:
//===--------------------------------------------------------------------===//
// Node creation methods.
//
SDValue getConstant(uint64_t Val, EVT VT, bool isTarget = false);
SDValue getConstant(const APInt &Val, EVT VT, bool isTarget = false);
SDValue getConstant(const ConstantInt &Val, EVT VT, bool isTarget = false);
SDValue getConstant(uint64_t Val, EVT VT, bool isTarget = false,
bool isOpaque = false);
SDValue getConstant(const APInt &Val, EVT VT, bool isTarget = false,
bool isOpaque = false);
SDValue getConstant(const ConstantInt &Val, EVT VT, bool isTarget = false,
bool isOpaque = false);
SDValue getIntPtrConstant(uint64_t Val, bool isTarget = false);
SDValue getTargetConstant(uint64_t Val, EVT VT) {
return getConstant(Val, VT, true);
SDValue getTargetConstant(uint64_t Val, EVT VT, bool isOpaque = false) {
return getConstant(Val, VT, true, isOpaque);
}
SDValue getTargetConstant(const APInt &Val, EVT VT) {
return getConstant(Val, VT, true);
SDValue getTargetConstant(const APInt &Val, EVT VT, bool isOpaque = false) {
return getConstant(Val, VT, true, isOpaque);
}
SDValue getTargetConstant(const ConstantInt &Val, EVT VT) {
return getConstant(Val, VT, true);
SDValue getTargetConstant(const ConstantInt &Val, EVT VT,
bool isOpaque = false) {
return getConstant(Val, VT, true, isOpaque);
}
// The forms below that take a double should only be used for simple
// constants that can be exactly represented in VT. No checks are made.

5
include/llvm/CodeGen/SelectionDAGNodes.h
View File

@ -1250,9 +1250,10 @@ public:
class ConstantSDNode : public SDNode {
const ConstantInt *Value;
friend class SelectionDAG;
ConstantSDNode(bool isTarget, const ConstantInt *val, EVT VT)
ConstantSDNode(bool isTarget, bool isOpaque, const ConstantInt *val, EVT VT)
: SDNode(isTarget ? ISD::TargetConstant : ISD::Constant,
0, DebugLoc(), getSDVTList(VT)), Value(val) {
SubclassData |= isOpaque;
}
public:
@ -1265,6 +1266,8 @@ public:
bool isNullValue() const { return Value->isNullValue(); }
bool isAllOnesValue() const { return Value->isAllOnesValue(); }
bool isOpaque() const { return SubclassData & 1; }
static bool classof(const SDNode *N) {
return N->getOpcode() == ISD::Constant ||
N->getOpcode() == ISD::TargetConstant;

1
include/llvm/InitializePasses.h
View File

@ -90,6 +90,7 @@ void initializeCFGSimplifyPassPass(PassRegistry&);
void initializeFlattenCFGPassPass(PassRegistry&);
void initializeStructurizeCFGPass(PassRegistry&);
void initializeCFGViewerPass(PassRegistry&);
void initializeConstantHoistingPass(PassRegistry&);
void initializeCodeGenPreparePass(PassRegistry&);
void initializeConstantMergePass(PassRegistry&);
void initializeConstantPropagationPass(PassRegistry&);

1
include/llvm/LinkAllPasses.h
View File

@ -129,6 +129,7 @@ namespace {
(void) llvm::createJumpThreadingPass();
(void) llvm::createUnifyFunctionExitNodesPass();
(void) llvm::createInstCountPass();
(void) llvm::createConstantHoistingPass();
(void) llvm::createCodeGenPreparePass();
(void) llvm::createEarlyCSEPass();
(void) llvm::createGVNPass();

6
include/llvm/Transforms/Scalar.h
View File

@ -310,6 +310,12 @@ FunctionPass *createMemCpyOptPass();
//
Pass *createLoopDeletionPass();
//===----------------------------------------------------------------------===//
//
// ConstantHoisting - This pass prepares a function for expensive constants.
//
FunctionPass *createConstantHoistingPass();
//===----------------------------------------------------------------------===//
//
// CodeGenPrepare - This pass prepares a function for instruction selection.

22
lib/Analysis/TargetTransformInfo.cpp
View File

@ -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 {

6
lib/CodeGen/Passes.cpp
View File

@ -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

9
lib/CodeGen/SelectionDAG/DAGCombiner.cpp
View File

@ -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)){

41
lib/CodeGen/SelectionDAG/SelectionDAG.cpp
View File

@ -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);
}

3
lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
View File

@ -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.
}

10
lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
View File

@ -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";

18
lib/CodeGen/SelectionDAG/TargetLowering.cpp
View File

@ -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)

95
lib/Target/X86/X86TargetTransformInfo.cpp
View File

@ -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);
}
}

1
lib/Transforms/Scalar/CMakeLists.txt
View File

@ -1,6 +1,7 @@
add_llvm_library(LLVMScalarOpts
ADCE.cpp
CodeGenPrepare.cpp
ConstantHoisting.cpp
ConstantProp.cpp
CorrelatedValuePropagation.cpp
DCE.cpp

4
lib/Transforms/Scalar/CodeGenPrepare.cpp
View File

@ -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.

429
lib/Transforms/Scalar/ConstantHoisting.cpp Normal file
View File

@ -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;
}

1
lib/Transforms/Scalar/Scalar.cpp
View File

@ -30,6 +30,7 @@ void llvm::initializeScalarOpts(PassRegistry &Registry) {
initializeADCEPass(Registry);
initializeSampleProfileLoaderPass(Registry);
initializeCodeGenPreparePass(Registry);
initializeConstantHoistingPass(Registry);
initializeConstantPropagationPass(Registry);
initializeCorrelatedValuePropagationPass(Registry);
initializeDCEPass(Registry);

6
test/CodeGen/ARM/memcpy-inline.ll
View File

@ -38,7 +38,8 @@ entry:
define void @t2(i8* nocapture %C) nounwind {
entry:
; CHECK-LABEL: t2:
; CHECK: ldr [[REG2:r[0-9]+]], [r1, #32]
; CHECK: movw [[REG2:r[0-9]+]], #16716
; CHECK: movt [[REG2:r[0-9]+]], #72
; CHECK: str [[REG2]], [r0, #32]
; CHECK: vld1.8 {d{{[0-9]+}}, d{{[0-9]+}}}, [r1]
; CHECK: vst1.8 {d{{[0-9]+}}, d{{[0-9]+}}}, [r0]
@ -79,7 +80,8 @@ entry:
; CHECK: strb [[REG5]], [r0, #6]
; CHECK: movw [[REG6:r[0-9]+]], #21587
; CHECK: strh [[REG6]], [r0, #4]
; CHECK: ldr [[REG7:r[0-9]+]],
; CHECK: movw [[REG7:r[0-9]+]], #18500
; CHECK: movt [[REG7:r[0-9]+]], #22866
; CHECK: str [[REG7]]
tail call void @llvm.memcpy.p0i8.p0i8.i64(i8* %C, i8* getelementptr inbounds ([7 x i8]* @.str5, i64 0, i64 0), i64 7, i32 1, i1 false)
ret void

53
test/CodeGen/X86/large-constants.ll Normal file
View File

@ -0,0 +1,53 @@
; RUN: llc < %s -mtriple=x86_64-darwin -mcpu=corei7 | grep movabsq | count 2
define i64 @constant_hoisting(i64 %o0, i64 %o1, i64 %o2, i64 %o3, i64 %o4, i64 %o5) {
entry:
%l0 = and i64 %o0, -281474976710654
%c0 = icmp ne i64 %l0, 0
br i1 %c0, label %fail, label %bb1
bb1:
%l1 = and i64 %o1, -281474976710654
%c1 = icmp ne i64 %l1, 0
br i1 %c1, label %fail, label %bb2
bb2:
%l2 = and i64 %o2, -281474976710654
%c2 = icmp ne i64 %l2, 0
br i1 %c2, label %fail, label %bb3
bb3:
%l3 = and i64 %o3, -281474976710654
%c3 = icmp ne i64 %l3, 0
br i1 %c3, label %fail, label %bb4
bb4:
%l4 = and i64 %o4, -281474976710653
%c4 = icmp ne i64 %l4, 0
br i1 %c4, label %fail, label %bb5
bb5:
%l5 = and i64 %o5, -281474976710652
%c5 = icmp ne i64 %l5, 0
br i1 %c5, label %fail, label %bb6
bb6:
ret i64 %l5
fail:
ret i64 -1
}
define void @constant_expressions() {
entry:
%0 = load i64* inttoptr (i64 add (i64 51250129900, i64 0) to i64*)
%1 = load i64* inttoptr (i64 add (i64 51250129900, i64 8) to i64*)
%2 = load i64* inttoptr (i64 add (i64 51250129900, i64 16) to i64*)
%3 = load i64* inttoptr (i64 add (i64 51250129900, i64 24) to i64*)
%4 = add i64 %0, %1
%5 = add i64 %2, %3
%6 = add i64 %4, %5
store i64 %6, i64* inttoptr (i64 add (i64 51250129900, i64 0) to i64*)
ret void
}