llvm-6502/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp

3433 lines
133 KiB
C++
Raw Normal View History

//===-- SelectionDAGISel.cpp - Implement the SelectionDAGISel class -------===//
//
// The LLVM Compiler Infrastructure
//
// This file was developed by the LLVM research group and is distributed under
// the University of Illinois Open Source License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This implements the SelectionDAGISel class.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "isel"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/CodeGen/ScheduleDAG.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/GlobalVariable.h"
#include "llvm/InlineAsm.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineDebugInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/Target/MRegisterInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetFrameInfo.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/Debug.h"
#include <map>
#include <set>
#include <iostream>
#include <algorithm>
using namespace llvm;
#ifndef NDEBUG
static cl::opt<bool>
ViewISelDAGs("view-isel-dags", cl::Hidden,
cl::desc("Pop up a window to show isel dags as they are selected"));
static cl::opt<bool>
ViewSchedDAGs("view-sched-dags", cl::Hidden,
cl::desc("Pop up a window to show sched dags as they are processed"));
#else
static const bool ViewISelDAGs = 0, ViewSchedDAGs = 0;
#endif
// Scheduling heuristics
enum SchedHeuristics {
defaultScheduling, // Let the target specify its preference.
noScheduling, // No scheduling, emit breadth first sequence.
simpleScheduling, // Two pass, min. critical path, max. utilization.
simpleNoItinScheduling, // Same as above exact using generic latency.
listSchedulingBURR, // Bottom up reg reduction list scheduling.
listSchedulingTD // Top-down list scheduler.
};
namespace {
cl::opt<SchedHeuristics>
ISHeuristic(
"sched",
cl::desc("Choose scheduling style"),
cl::init(defaultScheduling),
cl::values(
clEnumValN(defaultScheduling, "default",
"Target preferred scheduling style"),
clEnumValN(noScheduling, "none",
"No scheduling: breadth first sequencing"),
clEnumValN(simpleScheduling, "simple",
"Simple two pass scheduling: minimize critical path "
"and maximize processor utilization"),
clEnumValN(simpleNoItinScheduling, "simple-noitin",
"Simple two pass scheduling: Same as simple "
"except using generic latency"),
clEnumValN(listSchedulingBURR, "list-burr",
"Bottom up register reduction list scheduling"),
clEnumValN(listSchedulingTD, "list-td",
"Top-down list scheduler"),
clEnumValEnd));
} // namespace
namespace {
/// RegsForValue - This struct represents the physical registers that a
/// particular value is assigned and the type information about the value.
/// This is needed because values can be promoted into larger registers and
/// expanded into multiple smaller registers than the value.
struct RegsForValue {
/// Regs - This list hold the register (for legal and promoted values)
/// or register set (for expanded values) that the value should be assigned
/// to.
std::vector<unsigned> Regs;
/// RegVT - The value type of each register.
///
MVT::ValueType RegVT;
/// ValueVT - The value type of the LLVM value, which may be promoted from
/// RegVT or made from merging the two expanded parts.
MVT::ValueType ValueVT;
RegsForValue() : RegVT(MVT::Other), ValueVT(MVT::Other) {}
RegsForValue(unsigned Reg, MVT::ValueType regvt, MVT::ValueType valuevt)
: RegVT(regvt), ValueVT(valuevt) {
Regs.push_back(Reg);
}
RegsForValue(const std::vector<unsigned> &regs,
MVT::ValueType regvt, MVT::ValueType valuevt)
: Regs(regs), RegVT(regvt), ValueVT(valuevt) {
}
/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
/// this value and returns the result as a ValueVT value. This uses
/// Chain/Flag as the input and updates them for the output Chain/Flag.
SDOperand getCopyFromRegs(SelectionDAG &DAG,
SDOperand &Chain, SDOperand &Flag) const;
/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
/// specified value into the registers specified by this object. This uses
/// Chain/Flag as the input and updates them for the output Chain/Flag.
void getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
SDOperand &Chain, SDOperand &Flag) const;
/// AddInlineAsmOperands - Add this value to the specified inlineasm node
/// operand list. This adds the code marker and includes the number of
/// values added into it.
void AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
std::vector<SDOperand> &Ops) const;
};
}
namespace llvm {
//===--------------------------------------------------------------------===//
/// FunctionLoweringInfo - This contains information that is global to a
/// function that is used when lowering a region of the function.
class FunctionLoweringInfo {
public:
TargetLowering &TLI;
Function &Fn;
MachineFunction &MF;
SSARegMap *RegMap;
FunctionLoweringInfo(TargetLowering &TLI, Function &Fn,MachineFunction &MF);
/// MBBMap - A mapping from LLVM basic blocks to their machine code entry.
std::map<const BasicBlock*, MachineBasicBlock *> MBBMap;
/// ValueMap - Since we emit code for the function a basic block at a time,
/// we must remember which virtual registers hold the values for
/// cross-basic-block values.
std::map<const Value*, unsigned> ValueMap;
/// StaticAllocaMap - Keep track of frame indices for fixed sized allocas in
/// the entry block. This allows the allocas to be efficiently referenced
/// anywhere in the function.
std::map<const AllocaInst*, int> StaticAllocaMap;
unsigned MakeReg(MVT::ValueType VT) {
return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
}
unsigned CreateRegForValue(const Value *V);
unsigned InitializeRegForValue(const Value *V) {
unsigned &R = ValueMap[V];
assert(R == 0 && "Already initialized this value register!");
return R = CreateRegForValue(V);
}
};
}
/// isUsedOutsideOfDefiningBlock - Return true if this instruction is used by
/// PHI nodes or outside of the basic block that defines it, or used by a
/// switch instruction, which may expand to multiple basic blocks.
static bool isUsedOutsideOfDefiningBlock(Instruction *I) {
if (isa<PHINode>(I)) return true;
BasicBlock *BB = I->getParent();
for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E; ++UI)
if (cast<Instruction>(*UI)->getParent() != BB || isa<PHINode>(*UI) ||
isa<SwitchInst>(*UI))
return true;
return false;
}
/// isOnlyUsedInEntryBlock - If the specified argument is only used in the
/// entry block, return true. This includes arguments used by switches, since
/// the switch may expand into multiple basic blocks.
static bool isOnlyUsedInEntryBlock(Argument *A) {
BasicBlock *Entry = A->getParent()->begin();
for (Value::use_iterator UI = A->use_begin(), E = A->use_end(); UI != E; ++UI)
if (cast<Instruction>(*UI)->getParent() != Entry || isa<SwitchInst>(*UI))
return false; // Use not in entry block.
return true;
}
FunctionLoweringInfo::FunctionLoweringInfo(TargetLowering &tli,
Function &fn, MachineFunction &mf)
: TLI(tli), Fn(fn), MF(mf), RegMap(MF.getSSARegMap()) {
// Create a vreg for each argument register that is not dead and is used
// outside of the entry block for the function.
for (Function::arg_iterator AI = Fn.arg_begin(), E = Fn.arg_end();
AI != E; ++AI)
if (!isOnlyUsedInEntryBlock(AI))
InitializeRegForValue(AI);
// Initialize the mapping of values to registers. This is only set up for
// instruction values that are used outside of the block that defines
// them.
Function::iterator BB = Fn.begin(), EB = Fn.end();
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (AllocaInst *AI = dyn_cast<AllocaInst>(I))
if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(AI->getArraySize())) {
const Type *Ty = AI->getAllocatedType();
uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
unsigned Align =
std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
AI->getAlignment());
// If the alignment of the value is smaller than the size of the value,
// and if the size of the value is particularly small (<= 8 bytes),
// round up to the size of the value for potentially better performance.
//
// FIXME: This could be made better with a preferred alignment hook in
// TargetData. It serves primarily to 8-byte align doubles for X86.
if (Align < TySize && TySize <= 8) Align = TySize;
TySize *= CUI->getValue(); // Get total allocated size.
if (TySize == 0) TySize = 1; // Don't create zero-sized stack objects.
StaticAllocaMap[AI] =
MF.getFrameInfo()->CreateStackObject((unsigned)TySize, Align);
}
for (; BB != EB; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
if (!I->use_empty() && isUsedOutsideOfDefiningBlock(I))
if (!isa<AllocaInst>(I) ||
!StaticAllocaMap.count(cast<AllocaInst>(I)))
InitializeRegForValue(I);
// Create an initial MachineBasicBlock for each LLVM BasicBlock in F. This
// also creates the initial PHI MachineInstrs, though none of the input
// operands are populated.
for (BB = Fn.begin(), EB = Fn.end(); BB != EB; ++BB) {
MachineBasicBlock *MBB = new MachineBasicBlock(BB);
MBBMap[BB] = MBB;
MF.getBasicBlockList().push_back(MBB);
// Create Machine PHI nodes for LLVM PHI nodes, lowering them as
// appropriate.
PHINode *PN;
for (BasicBlock::iterator I = BB->begin();
(PN = dyn_cast<PHINode>(I)); ++I)
if (!PN->use_empty()) {
MVT::ValueType VT = TLI.getValueType(PN->getType());
unsigned NumElements;
if (VT != MVT::Vector)
NumElements = TLI.getNumElements(VT);
else {
MVT::ValueType VT1,VT2;
NumElements =
TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
VT1, VT2);
}
unsigned PHIReg = ValueMap[PN];
assert(PHIReg &&"PHI node does not have an assigned virtual register!");
for (unsigned i = 0; i != NumElements; ++i)
BuildMI(MBB, TargetInstrInfo::PHI, PN->getNumOperands(), PHIReg+i);
}
}
}
/// CreateRegForValue - Allocate the appropriate number of virtual registers of
/// the correctly promoted or expanded types. Assign these registers
/// consecutive vreg numbers and return the first assigned number.
unsigned FunctionLoweringInfo::CreateRegForValue(const Value *V) {
MVT::ValueType VT = TLI.getValueType(V->getType());
// The number of multiples of registers that we need, to, e.g., split up
// a <2 x int64> -> 4 x i32 registers.
unsigned NumVectorRegs = 1;
// If this is a packed type, figure out what type it will decompose into
// and how many of the elements it will use.
if (VT == MVT::Vector) {
const PackedType *PTy = cast<PackedType>(V->getType());
unsigned NumElts = PTy->getNumElements();
MVT::ValueType EltTy = TLI.getValueType(PTy->getElementType());
// Divide the input until we get to a supported size. This will always
// end with a scalar if the target doesn't support vectors.
while (NumElts > 1 && !TLI.isTypeLegal(getVectorType(EltTy, NumElts))) {
NumElts >>= 1;
NumVectorRegs <<= 1;
}
if (NumElts == 1)
VT = EltTy;
else
VT = getVectorType(EltTy, NumElts);
}
// The common case is that we will only create one register for this
// value. If we have that case, create and return the virtual register.
unsigned NV = TLI.getNumElements(VT);
if (NV == 1) {
// If we are promoting this value, pick the next largest supported type.
MVT::ValueType PromotedType = TLI.getTypeToTransformTo(VT);
unsigned Reg = MakeReg(PromotedType);
// If this is a vector of supported or promoted types (e.g. 4 x i16),
// create all of the registers.
for (unsigned i = 1; i != NumVectorRegs; ++i)
MakeReg(PromotedType);
return Reg;
}
// If this value is represented with multiple target registers, make sure
// to create enough consecutive registers of the right (smaller) type.
unsigned NT = VT-1; // Find the type to use.
while (TLI.getNumElements((MVT::ValueType)NT) != 1)
--NT;
unsigned R = MakeReg((MVT::ValueType)NT);
for (unsigned i = 1; i != NV*NumVectorRegs; ++i)
MakeReg((MVT::ValueType)NT);
return R;
}
//===----------------------------------------------------------------------===//
/// SelectionDAGLowering - This is the common target-independent lowering
/// implementation that is parameterized by a TargetLowering object.
/// Also, targets can overload any lowering method.
///
namespace llvm {
class SelectionDAGLowering {
MachineBasicBlock *CurMBB;
std::map<const Value*, SDOperand> NodeMap;
/// PendingLoads - Loads are not emitted to the program immediately. We bunch
/// them up and then emit token factor nodes when possible. This allows us to
/// get simple disambiguation between loads without worrying about alias
/// analysis.
std::vector<SDOperand> PendingLoads;
/// Case - A pair of values to record the Value for a switch case, and the
/// case's target basic block.
typedef std::pair<Constant*, MachineBasicBlock*> Case;
typedef std::vector<Case>::iterator CaseItr;
typedef std::pair<CaseItr, CaseItr> CaseRange;
/// CaseRec - A struct with ctor used in lowering switches to a binary tree
/// of conditional branches.
struct CaseRec {
CaseRec(MachineBasicBlock *bb, Constant *lt, Constant *ge, CaseRange r) :
CaseBB(bb), LT(lt), GE(ge), Range(r) {}
/// CaseBB - The MBB in which to emit the compare and branch
MachineBasicBlock *CaseBB;
/// LT, GE - If nonzero, we know the current case value must be less-than or
/// greater-than-or-equal-to these Constants.
Constant *LT;
Constant *GE;
/// Range - A pair of iterators representing the range of case values to be
/// processed at this point in the binary search tree.
CaseRange Range;
};
/// The comparison function for sorting Case values.
struct CaseCmp {
bool operator () (const Case& C1, const Case& C2) {
if (const ConstantUInt* U1 = dyn_cast<const ConstantUInt>(C1.first))
return U1->getValue() < cast<const ConstantUInt>(C2.first)->getValue();
const ConstantSInt* S1 = dyn_cast<const ConstantSInt>(C1.first);
return S1->getValue() < cast<const ConstantSInt>(C2.first)->getValue();
}
};
public:
// TLI - This is information that describes the available target features we
// need for lowering. This indicates when operations are unavailable,
// implemented with a libcall, etc.
TargetLowering &TLI;
SelectionDAG &DAG;
const TargetData *TD;
/// SwitchCases - Vector of CaseBlock structures used to communicate
/// SwitchInst code generation information.
std::vector<SelectionDAGISel::CaseBlock> SwitchCases;
SelectionDAGISel::JumpTable JT;
/// FuncInfo - Information about the function as a whole.
///
FunctionLoweringInfo &FuncInfo;
SelectionDAGLowering(SelectionDAG &dag, TargetLowering &tli,
FunctionLoweringInfo &funcinfo)
: TLI(tli), DAG(dag), TD(DAG.getTarget().getTargetData()),
JT(0,0,0,0), FuncInfo(funcinfo) {
}
/// getRoot - Return the current virtual root of the Selection DAG.
///
SDOperand getRoot() {
if (PendingLoads.empty())
return DAG.getRoot();
if (PendingLoads.size() == 1) {
SDOperand Root = PendingLoads[0];
DAG.setRoot(Root);
PendingLoads.clear();
return Root;
}
// Otherwise, we have to make a token factor node.
SDOperand Root = DAG.getNode(ISD::TokenFactor, MVT::Other, PendingLoads);
PendingLoads.clear();
DAG.setRoot(Root);
return Root;
}
void visit(Instruction &I) { visit(I.getOpcode(), I); }
void visit(unsigned Opcode, User &I) {
switch (Opcode) {
default: assert(0 && "Unknown instruction type encountered!");
abort();
// Build the switch statement using the Instruction.def file.
#define HANDLE_INST(NUM, OPCODE, CLASS) \
case Instruction::OPCODE:return visit##OPCODE((CLASS&)I);
#include "llvm/Instruction.def"
}
}
void setCurrentBasicBlock(MachineBasicBlock *MBB) { CurMBB = MBB; }
SDOperand getLoadFrom(const Type *Ty, SDOperand Ptr,
SDOperand SrcValue, SDOperand Root,
bool isVolatile);
SDOperand getIntPtrConstant(uint64_t Val) {
return DAG.getConstant(Val, TLI.getPointerTy());
}
SDOperand getValue(const Value *V);
const SDOperand &setValue(const Value *V, SDOperand NewN) {
SDOperand &N = NodeMap[V];
assert(N.Val == 0 && "Already set a value for this node!");
return N = NewN;
}
RegsForValue GetRegistersForValue(const std::string &ConstrCode,
MVT::ValueType VT,
bool OutReg, bool InReg,
std::set<unsigned> &OutputRegs,
std::set<unsigned> &InputRegs);
// Terminator instructions.
void visitRet(ReturnInst &I);
void visitBr(BranchInst &I);
void visitSwitch(SwitchInst &I);
void visitUnreachable(UnreachableInst &I) { /* noop */ }
// Helper for visitSwitch
void visitSwitchCase(SelectionDAGISel::CaseBlock &CB);
void visitJumpTable(SelectionDAGISel::JumpTable &JT);
// These all get lowered before this pass.
void visitInvoke(InvokeInst &I) { assert(0 && "TODO"); }
void visitUnwind(UnwindInst &I) { assert(0 && "TODO"); }
void visitBinary(User &I, unsigned IntOp, unsigned FPOp, unsigned VecOp);
void visitShift(User &I, unsigned Opcode);
void visitAdd(User &I) {
visitBinary(I, ISD::ADD, ISD::FADD, ISD::VADD);
}
void visitSub(User &I);
void visitMul(User &I) {
visitBinary(I, ISD::MUL, ISD::FMUL, ISD::VMUL);
}
void visitDiv(User &I) {
const Type *Ty = I.getType();
visitBinary(I,
Ty->isSigned() ? ISD::SDIV : ISD::UDIV, ISD::FDIV,
Ty->isSigned() ? ISD::VSDIV : ISD::VUDIV);
}
void visitRem(User &I) {
const Type *Ty = I.getType();
visitBinary(I, Ty->isSigned() ? ISD::SREM : ISD::UREM, ISD::FREM, 0);
}
void visitAnd(User &I) { visitBinary(I, ISD::AND, 0, ISD::VAND); }
void visitOr (User &I) { visitBinary(I, ISD::OR, 0, ISD::VOR); }
void visitXor(User &I) { visitBinary(I, ISD::XOR, 0, ISD::VXOR); }
void visitShl(User &I) { visitShift(I, ISD::SHL); }
void visitShr(User &I) {
visitShift(I, I.getType()->isUnsigned() ? ISD::SRL : ISD::SRA);
}
void visitSetCC(User &I, ISD::CondCode SignedOpc, ISD::CondCode UnsignedOpc);
void visitSetEQ(User &I) { visitSetCC(I, ISD::SETEQ, ISD::SETEQ); }
void visitSetNE(User &I) { visitSetCC(I, ISD::SETNE, ISD::SETNE); }
void visitSetLE(User &I) { visitSetCC(I, ISD::SETLE, ISD::SETULE); }
void visitSetGE(User &I) { visitSetCC(I, ISD::SETGE, ISD::SETUGE); }
void visitSetLT(User &I) { visitSetCC(I, ISD::SETLT, ISD::SETULT); }
void visitSetGT(User &I) { visitSetCC(I, ISD::SETGT, ISD::SETUGT); }
void visitExtractElement(User &I);
void visitInsertElement(User &I);
void visitShuffleVector(User &I);
void visitGetElementPtr(User &I);
void visitCast(User &I);
void visitSelect(User &I);
void visitMalloc(MallocInst &I);
void visitFree(FreeInst &I);
void visitAlloca(AllocaInst &I);
void visitLoad(LoadInst &I);
void visitStore(StoreInst &I);
void visitPHI(PHINode &I) { } // PHI nodes are handled specially.
void visitCall(CallInst &I);
void visitInlineAsm(CallInst &I);
const char *visitIntrinsicCall(CallInst &I, unsigned Intrinsic);
void visitTargetIntrinsic(CallInst &I, unsigned Intrinsic);
void visitVAStart(CallInst &I);
void visitVAArg(VAArgInst &I);
void visitVAEnd(CallInst &I);
void visitVACopy(CallInst &I);
void visitFrameReturnAddress(CallInst &I, bool isFrameAddress);
void visitMemIntrinsic(CallInst &I, unsigned Op);
void visitUserOp1(Instruction &I) {
assert(0 && "UserOp1 should not exist at instruction selection time!");
abort();
}
void visitUserOp2(Instruction &I) {
assert(0 && "UserOp2 should not exist at instruction selection time!");
abort();
}
};
} // end namespace llvm
SDOperand SelectionDAGLowering::getValue(const Value *V) {
SDOperand &N = NodeMap[V];
if (N.Val) return N;
const Type *VTy = V->getType();
MVT::ValueType VT = TLI.getValueType(VTy);
if (Constant *C = const_cast<Constant*>(dyn_cast<Constant>(V))) {
if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
visit(CE->getOpcode(), *CE);
assert(N.Val && "visit didn't populate the ValueMap!");
return N;
} else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
return N = DAG.getGlobalAddress(GV, VT);
} else if (isa<ConstantPointerNull>(C)) {
return N = DAG.getConstant(0, TLI.getPointerTy());
} else if (isa<UndefValue>(C)) {
if (!isa<PackedType>(VTy))
return N = DAG.getNode(ISD::UNDEF, VT);
// Create a VBUILD_VECTOR of undef nodes.
const PackedType *PTy = cast<PackedType>(VTy);
unsigned NumElements = PTy->getNumElements();
MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
std::vector<SDOperand> Ops;
Ops.assign(NumElements, DAG.getNode(ISD::UNDEF, PVT));
// Create a VConstant node with generic Vector type.
Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
Ops.push_back(DAG.getValueType(PVT));
return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops);
} else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
return N = DAG.getConstantFP(CFP->getValue(), VT);
} else if (const PackedType *PTy = dyn_cast<PackedType>(VTy)) {
unsigned NumElements = PTy->getNumElements();
MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
// Now that we know the number and type of the elements, push a
// Constant or ConstantFP node onto the ops list for each element of
// the packed constant.
std::vector<SDOperand> Ops;
if (ConstantPacked *CP = dyn_cast<ConstantPacked>(C)) {
for (unsigned i = 0; i != NumElements; ++i)
Ops.push_back(getValue(CP->getOperand(i)));
} else {
assert(isa<ConstantAggregateZero>(C) && "Unknown packed constant!");
SDOperand Op;
if (MVT::isFloatingPoint(PVT))
Op = DAG.getConstantFP(0, PVT);
else
Op = DAG.getConstant(0, PVT);
Ops.assign(NumElements, Op);
}
// Create a VBUILD_VECTOR node with generic Vector type.
Ops.push_back(DAG.getConstant(NumElements, MVT::i32));
Ops.push_back(DAG.getValueType(PVT));
return N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops);
} else {
// Canonicalize all constant ints to be unsigned.
return N = DAG.getConstant(cast<ConstantIntegral>(C)->getRawValue(),VT);
}
}
if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
std::map<const AllocaInst*, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end())
return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
}
std::map<const Value*, unsigned>::const_iterator VMI =
FuncInfo.ValueMap.find(V);
assert(VMI != FuncInfo.ValueMap.end() && "Value not in map!");
unsigned InReg = VMI->second;
// If this type is not legal, make it so now.
if (VT != MVT::Vector) {
MVT::ValueType DestVT = TLI.getTypeToTransformTo(VT);
N = DAG.getCopyFromReg(DAG.getEntryNode(), InReg, DestVT);
if (DestVT < VT) {
// Source must be expanded. This input value is actually coming from the
// register pair VMI->second and VMI->second+1.
N = DAG.getNode(ISD::BUILD_PAIR, VT, N,
DAG.getCopyFromReg(DAG.getEntryNode(), InReg+1, DestVT));
} else if (DestVT > VT) { // Promotion case
if (MVT::isFloatingPoint(VT))
N = DAG.getNode(ISD::FP_ROUND, VT, N);
else
N = DAG.getNode(ISD::TRUNCATE, VT, N);
}
} else {
// Otherwise, if this is a vector, make it available as a generic vector
// here.
MVT::ValueType PTyElementVT, PTyLegalElementVT;
const PackedType *PTy = cast<PackedType>(VTy);
unsigned NE = TLI.getPackedTypeBreakdown(PTy, PTyElementVT,
PTyLegalElementVT);
// Build a VBUILD_VECTOR with the input registers.
std::vector<SDOperand> Ops;
if (PTyElementVT == PTyLegalElementVT) {
// If the value types are legal, just VBUILD the CopyFromReg nodes.
for (unsigned i = 0; i != NE; ++i)
Ops.push_back(DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
PTyElementVT));
} else if (PTyElementVT < PTyLegalElementVT) {
// If the register was promoted, use TRUNCATE of FP_ROUND as appropriate.
for (unsigned i = 0; i != NE; ++i) {
SDOperand Op = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
PTyElementVT);
if (MVT::isFloatingPoint(PTyElementVT))
Op = DAG.getNode(ISD::FP_ROUND, PTyElementVT, Op);
else
Op = DAG.getNode(ISD::TRUNCATE, PTyElementVT, Op);
Ops.push_back(Op);
}
} else {
// If the register was expanded, use BUILD_PAIR.
assert((NE & 1) == 0 && "Must expand into a multiple of 2 elements!");
for (unsigned i = 0; i != NE/2; ++i) {
SDOperand Op0 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
PTyElementVT);
SDOperand Op1 = DAG.getCopyFromReg(DAG.getEntryNode(), InReg++,
PTyElementVT);
Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Op0, Op1));
}
}
Ops.push_back(DAG.getConstant(NE, MVT::i32));
Ops.push_back(DAG.getValueType(PTyLegalElementVT));
N = DAG.getNode(ISD::VBUILD_VECTOR, MVT::Vector, Ops);
// Finally, use a VBIT_CONVERT to make this available as the appropriate
// vector type.
N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
DAG.getConstant(PTy->getNumElements(),
MVT::i32),
DAG.getValueType(TLI.getValueType(PTy->getElementType())));
}
return N;
}
void SelectionDAGLowering::visitRet(ReturnInst &I) {
if (I.getNumOperands() == 0) {
DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, getRoot()));
return;
}
std::vector<SDOperand> NewValues;
NewValues.push_back(getRoot());
for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i) {
SDOperand RetOp = getValue(I.getOperand(i));
// If this is an integer return value, we need to promote it ourselves to
// the full width of a register, since LegalizeOp will use ANY_EXTEND rather
// than sign/zero.
if (MVT::isInteger(RetOp.getValueType()) &&
RetOp.getValueType() < MVT::i64) {
MVT::ValueType TmpVT;
if (TLI.getTypeAction(MVT::i32) == TargetLowering::Promote)
TmpVT = TLI.getTypeToTransformTo(MVT::i32);
else
TmpVT = MVT::i32;
if (I.getOperand(i)->getType()->isSigned())
RetOp = DAG.getNode(ISD::SIGN_EXTEND, TmpVT, RetOp);
else
RetOp = DAG.getNode(ISD::ZERO_EXTEND, TmpVT, RetOp);
}
NewValues.push_back(RetOp);
}
DAG.setRoot(DAG.getNode(ISD::RET, MVT::Other, NewValues));
}
void SelectionDAGLowering::visitBr(BranchInst &I) {
// Update machine-CFG edges.
MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
CurMBB->addSuccessor(Succ0MBB);
// Figure out which block is immediately after the current one.
MachineBasicBlock *NextBlock = 0;
MachineFunction::iterator BBI = CurMBB;
if (++BBI != CurMBB->getParent()->end())
NextBlock = BBI;
if (I.isUnconditional()) {
// If this is not a fall-through branch, emit the branch.
if (Succ0MBB != NextBlock)
DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
DAG.getBasicBlock(Succ0MBB)));
} else {
MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
CurMBB->addSuccessor(Succ1MBB);
SDOperand Cond = getValue(I.getCondition());
if (Succ1MBB == NextBlock) {
// If the condition is false, fall through. This means we should branch
// if the condition is true to Succ #0.
DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(),
Cond, DAG.getBasicBlock(Succ0MBB)));
} else if (Succ0MBB == NextBlock) {
// If the condition is true, fall through. This means we should branch if
// the condition is false to Succ #1. Invert the condition first.
SDOperand True = DAG.getConstant(1, Cond.getValueType());
Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(),
Cond, DAG.getBasicBlock(Succ1MBB)));
} else {
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
// If the false case is the current basic block, then this is a self
// loop. We do not want to emit "Loop: ... brcond Out; br Loop", as it
// adds an extra instruction in the loop. Instead, invert the
// condition and emit "Loop: ... br!cond Loop; br Out.
if (CurMBB == Succ1MBB) {
std::swap(Succ0MBB, Succ1MBB);
SDOperand True = DAG.getConstant(1, Cond.getValueType());
Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
}
SDOperand True = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
DAG.getBasicBlock(Succ0MBB));
DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, True,
DAG.getBasicBlock(Succ1MBB)));
}
}
}
/// visitSwitchCase - Emits the necessary code to represent a single node in
/// the binary search tree resulting from lowering a switch instruction.
void SelectionDAGLowering::visitSwitchCase(SelectionDAGISel::CaseBlock &CB) {
SDOperand SwitchOp = getValue(CB.SwitchV);
SDOperand CaseOp = getValue(CB.CaseC);
SDOperand Cond = DAG.getSetCC(MVT::i1, SwitchOp, CaseOp, CB.CC);
// Set NextBlock to be the MBB immediately after the current one, if any.
// This is used to avoid emitting unnecessary branches to the next block.
MachineBasicBlock *NextBlock = 0;
MachineFunction::iterator BBI = CurMBB;
if (++BBI != CurMBB->getParent()->end())
NextBlock = BBI;
// If the lhs block is the next block, invert the condition so that we can
// fall through to the lhs instead of the rhs block.
if (CB.LHSBB == NextBlock) {
std::swap(CB.LHSBB, CB.RHSBB);
SDOperand True = DAG.getConstant(1, Cond.getValueType());
Cond = DAG.getNode(ISD::XOR, Cond.getValueType(), Cond, True);
}
SDOperand BrCond = DAG.getNode(ISD::BRCOND, MVT::Other, getRoot(), Cond,
DAG.getBasicBlock(CB.LHSBB));
if (CB.RHSBB == NextBlock)
DAG.setRoot(BrCond);
else
DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, BrCond,
DAG.getBasicBlock(CB.RHSBB)));
// Update successor info
CurMBB->addSuccessor(CB.LHSBB);
CurMBB->addSuccessor(CB.RHSBB);
}
/// visitSwitchCase - Emits the necessary code to represent a single node in
/// the binary search tree resulting from lowering a switch instruction.
void SelectionDAGLowering::visitJumpTable(SelectionDAGISel::JumpTable &JT) {
// FIXME: Need to emit different code for PIC vs. Non-PIC, specifically,
// we need to add the address of the jump table to the value loaded, since
// the entries in the jump table will be differences rather than absolute
// addresses.
// Emit the code for the jump table
MVT::ValueType PTy = TLI.getPointerTy();
unsigned PTyBytes = MVT::getSizeInBits(PTy)/8;
SDOperand Copy = DAG.getCopyFromReg(getRoot(), JT.Reg, PTy);
SDOperand IDX = DAG.getNode(ISD::MUL, PTy, Copy,
DAG.getConstant(PTyBytes, PTy));
SDOperand ADD = DAG.getNode(ISD::ADD, PTy, IDX, DAG.getJumpTable(JT.JTI,PTy));
SDOperand LD = DAG.getLoad(PTy, Copy.getValue(1), ADD, DAG.getSrcValue(0));
DAG.setRoot(DAG.getNode(ISD::BRIND, MVT::Other, LD.getValue(1), LD));
}
void SelectionDAGLowering::visitSwitch(SwitchInst &I) {
// Figure out which block is immediately after the current one.
MachineBasicBlock *NextBlock = 0;
MachineFunction::iterator BBI = CurMBB;
if (++BBI != CurMBB->getParent()->end())
NextBlock = BBI;
// If there is only the default destination, branch to it if it is not the
// next basic block. Otherwise, just fall through.
if (I.getNumOperands() == 2) {
// Update machine-CFG edges.
MachineBasicBlock *DefaultMBB = FuncInfo.MBBMap[I.getDefaultDest()];
// If this is not a fall-through branch, emit the branch.
if (DefaultMBB != NextBlock)
DAG.setRoot(DAG.getNode(ISD::BR, MVT::Other, getRoot(),
DAG.getBasicBlock(DefaultMBB)));
return;
}
// If there are any non-default case statements, create a vector of Cases
// representing each one, and sort the vector so that we can efficiently
// create a binary search tree from them.
std::vector<Case> Cases;
for (unsigned i = 1; i < I.getNumSuccessors(); ++i) {
MachineBasicBlock *SMBB = FuncInfo.MBBMap[I.getSuccessor(i)];
Cases.push_back(Case(I.getSuccessorValue(i), SMBB));
}
std::sort(Cases.begin(), Cases.end(), CaseCmp());
// Get the Value to be switched on and default basic blocks, which will be
// inserted into CaseBlock records, representing basic blocks in the binary
// search tree.
Value *SV = I.getOperand(0);
MachineBasicBlock *Default = FuncInfo.MBBMap[I.getDefaultDest()];
// Get the MachineFunction which holds the current MBB. This is used during
// emission of jump tables, and when inserting any additional MBBs necessary
// to represent the switch.
MachineFunction *CurMF = CurMBB->getParent();
const BasicBlock *LLVMBB = CurMBB->getBasicBlock();
Reloc::Model Relocs = TLI.getTargetMachine().getRelocationModel();
// If the switch has more than 5 blocks, and at least 75% dense, then emit a
// jump table rather than lowering the switch to a binary tree of conditional
// branches.
// FIXME: Make this work with PIC code
if (TLI.isOperationLegal(ISD::BRIND, TLI.getPointerTy()) &&
(Relocs == Reloc::Static || Relocs == Reloc::DynamicNoPIC) &&
Cases.size() > 5) {
uint64_t First = cast<ConstantIntegral>(Cases.front().first)->getRawValue();
uint64_t Last = cast<ConstantIntegral>(Cases.back().first)->getRawValue();
double Density = (double)Cases.size() / (double)((Last - First) + 1ULL);
if (Density >= 0.75) {
// Create a new basic block to hold the code for loading the address
// of the jump table, and jumping to it. Update successor information;
// we will either branch to the default case for the switch, or the jump
// table.
MachineBasicBlock *JumpTableBB = new MachineBasicBlock(LLVMBB);
CurMF->getBasicBlockList().insert(BBI, JumpTableBB);
CurMBB->addSuccessor(Default);
CurMBB->addSuccessor(JumpTableBB);
// Subtract the lowest switch case value from the value being switched on
// and conditional branch to default mbb if the result is greater than the
// difference between smallest and largest cases.
SDOperand SwitchOp = getValue(SV);
MVT::ValueType VT = SwitchOp.getValueType();
SDOperand SUB = DAG.getNode(ISD::SUB, VT, SwitchOp,
DAG.getConstant(First, VT));
// The SDNode we just created, which holds the value being switched on
// minus the the smallest case value, needs to be copied to a virtual
// register so it can be used as an index into the jump table in a
// subsequent basic block. This value may be smaller or larger than the
// target's pointer type, and therefore require extension or truncating.
if (VT > TLI.getPointerTy())
SwitchOp = DAG.getNode(ISD::TRUNCATE, TLI.getPointerTy(), SUB);
else
SwitchOp = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(), SUB);
unsigned JumpTableReg = FuncInfo.MakeReg(TLI.getPointerTy());
SDOperand CopyTo = DAG.getCopyToReg(getRoot(), JumpTableReg, SwitchOp);
// Emit the range check for the jump table, and branch to the default
// block for the switch statement if the value being switched on exceeds
// the largest case in the switch.
SDOperand CMP = DAG.getSetCC(TLI.getSetCCResultTy(), SUB,
DAG.getConstant(Last-First,VT), ISD::SETUGT);
DAG.setRoot(DAG.getNode(ISD::BRCOND, MVT::Other, CopyTo, CMP,
DAG.getBasicBlock(Default)));
// Build a vector of destination BBs, corresponding to each target
// of the jump table. If the value of the jump table slot corresponds to
// a case statement, push the case's BB onto the vector, otherwise, push
// the default BB.
std::set<MachineBasicBlock*> UniqueBBs;
std::vector<MachineBasicBlock*> DestBBs;
uint64_t TEI = First;
for (CaseItr ii = Cases.begin(), ee = Cases.end(); ii != ee; ++TEI) {
if (cast<ConstantIntegral>(ii->first)->getRawValue() == TEI) {
DestBBs.push_back(ii->second);
UniqueBBs.insert(ii->second);
++ii;
} else {
DestBBs.push_back(Default);
UniqueBBs.insert(Default);
}
}
// Update successor info
for (std::set<MachineBasicBlock*>::iterator ii = UniqueBBs.begin(),
ee = UniqueBBs.end(); ii != ee; ++ii)
JumpTableBB->addSuccessor(*ii);
// Create a jump table index for this jump table, or return an existing
// one.
unsigned JTI = CurMF->getJumpTableInfo()->getJumpTableIndex(DestBBs);
// Set the jump table information so that we can codegen it as a second
// MachineBasicBlock
JT.Reg = JumpTableReg;
JT.JTI = JTI;
JT.MBB = JumpTableBB;
JT.Default = Default;
return;
}
}
// Push the initial CaseRec onto the worklist
std::vector<CaseRec> CaseVec;
CaseVec.push_back(CaseRec(CurMBB,0,0,CaseRange(Cases.begin(),Cases.end())));
while (!CaseVec.empty()) {
// Grab a record representing a case range to process off the worklist
CaseRec CR = CaseVec.back();
CaseVec.pop_back();
// Size is the number of Cases represented by this range. If Size is 1,
// then we are processing a leaf of the binary search tree. Otherwise,
// we need to pick a pivot, and push left and right ranges onto the
// worklist.
unsigned Size = CR.Range.second - CR.Range.first;
if (Size == 1) {
// Create a CaseBlock record representing a conditional branch to
// the Case's target mbb if the value being switched on SV is equal
// to C. Otherwise, branch to default.
Constant *C = CR.Range.first->first;
MachineBasicBlock *Target = CR.Range.first->second;
SelectionDAGISel::CaseBlock CB(ISD::SETEQ, SV, C, Target, Default,
CR.CaseBB);
// If the MBB representing the leaf node is the current MBB, then just
// call visitSwitchCase to emit the code into the current block.
// Otherwise, push the CaseBlock onto the vector to be later processed
// by SDISel, and insert the node's MBB before the next MBB.
if (CR.CaseBB == CurMBB)
visitSwitchCase(CB);
else {
SwitchCases.push_back(CB);
CurMF->getBasicBlockList().insert(BBI, CR.CaseBB);
}
} else {
// split case range at pivot
CaseItr Pivot = CR.Range.first + (Size / 2);
CaseRange LHSR(CR.Range.first, Pivot);
CaseRange RHSR(Pivot, CR.Range.second);
Constant *C = Pivot->first;
MachineBasicBlock *RHSBB = 0, *LHSBB = 0;
// We know that we branch to the LHS if the Value being switched on is
// less than the Pivot value, C. We use this to optimize our binary
// tree a bit, by recognizing that if SV is greater than or equal to the
// LHS's Case Value, and that Case Value is exactly one less than the
// Pivot's Value, then we can branch directly to the LHS's Target,
// rather than creating a leaf node for it.
if ((LHSR.second - LHSR.first) == 1 &&
LHSR.first->first == CR.GE &&
cast<ConstantIntegral>(C)->getRawValue() ==
(cast<ConstantIntegral>(CR.GE)->getRawValue() + 1ULL)) {
LHSBB = LHSR.first->second;
} else {
LHSBB = new MachineBasicBlock(LLVMBB);
CaseVec.push_back(CaseRec(LHSBB,C,CR.GE,LHSR));
}
// Similar to the optimization above, if the Value being switched on is
// known to be less than the Constant CR.LT, and the current Case Value
// is CR.LT - 1, then we can branch directly to the target block for
// the current Case Value, rather than emitting a RHS leaf node for it.
if ((RHSR.second - RHSR.first) == 1 && CR.LT &&
cast<ConstantIntegral>(RHSR.first->first)->getRawValue() ==
(cast<ConstantIntegral>(CR.LT)->getRawValue() - 1ULL)) {
RHSBB = RHSR.first->second;
} else {
RHSBB = new MachineBasicBlock(LLVMBB);
CaseVec.push_back(CaseRec(RHSBB,CR.LT,C,RHSR));
}
// Create a CaseBlock record representing a conditional branch to
// the LHS node if the value being switched on SV is less than C.
// Otherwise, branch to LHS.
ISD::CondCode CC = C->getType()->isSigned() ? ISD::SETLT : ISD::SETULT;
SelectionDAGISel::CaseBlock CB(CC, SV, C, LHSBB, RHSBB, CR.CaseBB);
if (CR.CaseBB == CurMBB)
visitSwitchCase(CB);
else {
SwitchCases.push_back(CB);
CurMF->getBasicBlockList().insert(BBI, CR.CaseBB);
}
}
}
}
void SelectionDAGLowering::visitSub(User &I) {
// -0.0 - X --> fneg
if (I.getType()->isFloatingPoint()) {
if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
if (CFP->isExactlyValue(-0.0)) {
SDOperand Op2 = getValue(I.getOperand(1));
setValue(&I, DAG.getNode(ISD::FNEG, Op2.getValueType(), Op2));
return;
}
}
visitBinary(I, ISD::SUB, ISD::FSUB, ISD::VSUB);
}
void SelectionDAGLowering::visitBinary(User &I, unsigned IntOp, unsigned FPOp,
unsigned VecOp) {
const Type *Ty = I.getType();
SDOperand Op1 = getValue(I.getOperand(0));
SDOperand Op2 = getValue(I.getOperand(1));
if (Ty->isIntegral()) {
setValue(&I, DAG.getNode(IntOp, Op1.getValueType(), Op1, Op2));
} else if (Ty->isFloatingPoint()) {
setValue(&I, DAG.getNode(FPOp, Op1.getValueType(), Op1, Op2));
} else {
const PackedType *PTy = cast<PackedType>(Ty);
SDOperand Num = DAG.getConstant(PTy->getNumElements(), MVT::i32);
SDOperand Typ = DAG.getValueType(TLI.getValueType(PTy->getElementType()));
setValue(&I, DAG.getNode(VecOp, MVT::Vector, Op1, Op2, Num, Typ));
}
}
void SelectionDAGLowering::visitShift(User &I, unsigned Opcode) {
SDOperand Op1 = getValue(I.getOperand(0));
SDOperand Op2 = getValue(I.getOperand(1));
Op2 = DAG.getNode(ISD::ANY_EXTEND, TLI.getShiftAmountTy(), Op2);
setValue(&I, DAG.getNode(Opcode, Op1.getValueType(), Op1, Op2));
}
void SelectionDAGLowering::visitSetCC(User &I,ISD::CondCode SignedOpcode,
ISD::CondCode UnsignedOpcode) {
SDOperand Op1 = getValue(I.getOperand(0));
SDOperand Op2 = getValue(I.getOperand(1));
ISD::CondCode Opcode = SignedOpcode;
if (I.getOperand(0)->getType()->isUnsigned())
Opcode = UnsignedOpcode;
setValue(&I, DAG.getSetCC(MVT::i1, Op1, Op2, Opcode));
}
void SelectionDAGLowering::visitSelect(User &I) {
SDOperand Cond = getValue(I.getOperand(0));
SDOperand TrueVal = getValue(I.getOperand(1));
SDOperand FalseVal = getValue(I.getOperand(2));
if (!isa<PackedType>(I.getType())) {
setValue(&I, DAG.getNode(ISD::SELECT, TrueVal.getValueType(), Cond,
TrueVal, FalseVal));
} else {
setValue(&I, DAG.getNode(ISD::VSELECT, MVT::Vector, Cond, TrueVal, FalseVal,
*(TrueVal.Val->op_end()-2),
*(TrueVal.Val->op_end()-1)));
}
}
void SelectionDAGLowering::visitCast(User &I) {
SDOperand N = getValue(I.getOperand(0));
MVT::ValueType SrcVT = N.getValueType();
MVT::ValueType DestVT = TLI.getValueType(I.getType());
if (DestVT == MVT::Vector) {
// This is a cast to a vector from something else. This is always a bit
// convert. Get information about the input vector.
const PackedType *DestTy = cast<PackedType>(I.getType());
MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N,
DAG.getConstant(DestTy->getNumElements(),MVT::i32),
DAG.getValueType(EltVT)));
} else if (SrcVT == DestVT) {
setValue(&I, N); // noop cast.
} else if (DestVT == MVT::i1) {
// Cast to bool is a comparison against zero, not truncation to zero.
SDOperand Zero = isInteger(SrcVT) ? DAG.getConstant(0, N.getValueType()) :
DAG.getConstantFP(0.0, N.getValueType());
setValue(&I, DAG.getSetCC(MVT::i1, N, Zero, ISD::SETNE));
} else if (isInteger(SrcVT)) {
if (isInteger(DestVT)) { // Int -> Int cast
if (DestVT < SrcVT) // Truncating cast?
setValue(&I, DAG.getNode(ISD::TRUNCATE, DestVT, N));
else if (I.getOperand(0)->getType()->isSigned())
setValue(&I, DAG.getNode(ISD::SIGN_EXTEND, DestVT, N));
else
setValue(&I, DAG.getNode(ISD::ZERO_EXTEND, DestVT, N));
} else if (isFloatingPoint(DestVT)) { // Int -> FP cast
if (I.getOperand(0)->getType()->isSigned())
setValue(&I, DAG.getNode(ISD::SINT_TO_FP, DestVT, N));
else
setValue(&I, DAG.getNode(ISD::UINT_TO_FP, DestVT, N));
} else {
assert(0 && "Unknown cast!");
}
} else if (isFloatingPoint(SrcVT)) {
if (isFloatingPoint(DestVT)) { // FP -> FP cast
if (DestVT < SrcVT) // Rounding cast?
setValue(&I, DAG.getNode(ISD::FP_ROUND, DestVT, N));
else
setValue(&I, DAG.getNode(ISD::FP_EXTEND, DestVT, N));
} else if (isInteger(DestVT)) { // FP -> Int cast.
if (I.getType()->isSigned())
setValue(&I, DAG.getNode(ISD::FP_TO_SINT, DestVT, N));
else
setValue(&I, DAG.getNode(ISD::FP_TO_UINT, DestVT, N));
} else {
assert(0 && "Unknown cast!");
}
} else {
assert(SrcVT == MVT::Vector && "Unknown cast!");
assert(DestVT != MVT::Vector && "Casts to vector already handled!");
// This is a cast from a vector to something else. This is always a bit
// convert. Get information about the input vector.
setValue(&I, DAG.getNode(ISD::VBIT_CONVERT, DestVT, N));
}
}
void SelectionDAGLowering::visitInsertElement(User &I) {
SDOperand InVec = getValue(I.getOperand(0));
SDOperand InVal = getValue(I.getOperand(1));
SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
getValue(I.getOperand(2)));
SDOperand Num = *(InVec.Val->op_end()-2);
SDOperand Typ = *(InVec.Val->op_end()-1);
setValue(&I, DAG.getNode(ISD::VINSERT_VECTOR_ELT, MVT::Vector,
InVec, InVal, InIdx, Num, Typ));
}
void SelectionDAGLowering::visitExtractElement(User &I) {
SDOperand InVec = getValue(I.getOperand(0));
SDOperand InIdx = DAG.getNode(ISD::ZERO_EXTEND, TLI.getPointerTy(),
getValue(I.getOperand(1)));
SDOperand Typ = *(InVec.Val->op_end()-1);
setValue(&I, DAG.getNode(ISD::VEXTRACT_VECTOR_ELT,
TLI.getValueType(I.getType()), InVec, InIdx));
}
void SelectionDAGLowering::visitShuffleVector(User &I) {
SDOperand V1 = getValue(I.getOperand(0));
SDOperand V2 = getValue(I.getOperand(1));
SDOperand Mask = getValue(I.getOperand(2));
SDOperand Num = *(V1.Val->op_end()-2);
SDOperand Typ = *(V2.Val->op_end()-1);
setValue(&I, DAG.getNode(ISD::VVECTOR_SHUFFLE, MVT::Vector,
V1, V2, Mask, Num, Typ));
}
void SelectionDAGLowering::visitGetElementPtr(User &I) {
SDOperand N = getValue(I.getOperand(0));
const Type *Ty = I.getOperand(0)->getType();
const Type *UIntPtrTy = TD->getIntPtrType();
for (GetElementPtrInst::op_iterator OI = I.op_begin()+1, E = I.op_end();
OI != E; ++OI) {
Value *Idx = *OI;
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
unsigned Field = cast<ConstantUInt>(Idx)->getValue();
if (Field) {
// N = N + Offset
uint64_t Offset = TD->getStructLayout(StTy)->MemberOffsets[Field];
N = DAG.getNode(ISD::ADD, N.getValueType(), N,
getIntPtrConstant(Offset));
}
Ty = StTy->getElementType(Field);
} else {
Ty = cast<SequentialType>(Ty)->getElementType();
// If this is a constant subscript, handle it quickly.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (CI->getRawValue() == 0) continue;
uint64_t Offs;
if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(CI))
Offs = (int64_t)TD->getTypeSize(Ty)*CSI->getValue();
else
Offs = TD->getTypeSize(Ty)*cast<ConstantUInt>(CI)->getValue();
N = DAG.getNode(ISD::ADD, N.getValueType(), N, getIntPtrConstant(Offs));
continue;
}
// N = N + Idx * ElementSize;
uint64_t ElementSize = TD->getTypeSize(Ty);
SDOperand IdxN = getValue(Idx);
// If the index is smaller or larger than intptr_t, truncate or extend
// it.
if (IdxN.getValueType() < N.getValueType()) {
if (Idx->getType()->isSigned())
IdxN = DAG.getNode(ISD::SIGN_EXTEND, N.getValueType(), IdxN);
else
IdxN = DAG.getNode(ISD::ZERO_EXTEND, N.getValueType(), IdxN);
} else if (IdxN.getValueType() > N.getValueType())
IdxN = DAG.getNode(ISD::TRUNCATE, N.getValueType(), IdxN);
// If this is a multiply by a power of two, turn it into a shl
// immediately. This is a very common case.
if (isPowerOf2_64(ElementSize)) {
unsigned Amt = Log2_64(ElementSize);
IdxN = DAG.getNode(ISD::SHL, N.getValueType(), IdxN,
DAG.getConstant(Amt, TLI.getShiftAmountTy()));
N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
continue;
}
SDOperand Scale = getIntPtrConstant(ElementSize);
IdxN = DAG.getNode(ISD::MUL, N.getValueType(), IdxN, Scale);
N = DAG.getNode(ISD::ADD, N.getValueType(), N, IdxN);
}
}
setValue(&I, N);
}
void SelectionDAGLowering::visitAlloca(AllocaInst &I) {
// If this is a fixed sized alloca in the entry block of the function,
// allocate it statically on the stack.
if (FuncInfo.StaticAllocaMap.count(&I))
return; // getValue will auto-populate this.
const Type *Ty = I.getAllocatedType();
uint64_t TySize = TLI.getTargetData()->getTypeSize(Ty);
unsigned Align = std::max((unsigned)TLI.getTargetData()->getTypeAlignment(Ty),
I.getAlignment());
SDOperand AllocSize = getValue(I.getArraySize());
MVT::ValueType IntPtr = TLI.getPointerTy();
if (IntPtr < AllocSize.getValueType())
AllocSize = DAG.getNode(ISD::TRUNCATE, IntPtr, AllocSize);
else if (IntPtr > AllocSize.getValueType())
AllocSize = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, AllocSize);
AllocSize = DAG.getNode(ISD::MUL, IntPtr, AllocSize,
getIntPtrConstant(TySize));
// Handle alignment. If the requested alignment is less than or equal to the
// stack alignment, ignore it and round the size of the allocation up to the
// stack alignment size. If the size is greater than the stack alignment, we
// note this in the DYNAMIC_STACKALLOC node.
unsigned StackAlign =
TLI.getTargetMachine().getFrameInfo()->getStackAlignment();
if (Align <= StackAlign) {
Align = 0;
// Add SA-1 to the size.
AllocSize = DAG.getNode(ISD::ADD, AllocSize.getValueType(), AllocSize,
getIntPtrConstant(StackAlign-1));
// Mask out the low bits for alignment purposes.
AllocSize = DAG.getNode(ISD::AND, AllocSize.getValueType(), AllocSize,
getIntPtrConstant(~(uint64_t)(StackAlign-1)));
}
std::vector<MVT::ValueType> VTs;
VTs.push_back(AllocSize.getValueType());
VTs.push_back(MVT::Other);
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
Ops.push_back(AllocSize);
Ops.push_back(getIntPtrConstant(Align));
SDOperand DSA = DAG.getNode(ISD::DYNAMIC_STACKALLOC, VTs, Ops);
DAG.setRoot(setValue(&I, DSA).getValue(1));
// Inform the Frame Information that we have just allocated a variable-sized
// object.
CurMBB->getParent()->getFrameInfo()->CreateVariableSizedObject();
}
void SelectionDAGLowering::visitLoad(LoadInst &I) {
SDOperand Ptr = getValue(I.getOperand(0));
SDOperand Root;
if (I.isVolatile())
Root = getRoot();
else {
// Do not serialize non-volatile loads against each other.
Root = DAG.getRoot();
}
setValue(&I, getLoadFrom(I.getType(), Ptr, DAG.getSrcValue(I.getOperand(0)),
Root, I.isVolatile()));
}
SDOperand SelectionDAGLowering::getLoadFrom(const Type *Ty, SDOperand Ptr,
SDOperand SrcValue, SDOperand Root,
bool isVolatile) {
SDOperand L;
if (const PackedType *PTy = dyn_cast<PackedType>(Ty)) {
MVT::ValueType PVT = TLI.getValueType(PTy->getElementType());
L = DAG.getVecLoad(PTy->getNumElements(), PVT, Root, Ptr, SrcValue);
} else {
L = DAG.getLoad(TLI.getValueType(Ty), Root, Ptr, SrcValue);
}
if (isVolatile)
DAG.setRoot(L.getValue(1));
else
PendingLoads.push_back(L.getValue(1));
return L;
}
void SelectionDAGLowering::visitStore(StoreInst &I) {
Value *SrcV = I.getOperand(0);
SDOperand Src = getValue(SrcV);
SDOperand Ptr = getValue(I.getOperand(1));
DAG.setRoot(DAG.getNode(ISD::STORE, MVT::Other, getRoot(), Src, Ptr,
DAG.getSrcValue(I.getOperand(1))));
}
/// IntrinsicCannotAccessMemory - Return true if the specified intrinsic cannot
/// access memory and has no other side effects at all.
static bool IntrinsicCannotAccessMemory(unsigned IntrinsicID) {
#define GET_NO_MEMORY_INTRINSICS
#include "llvm/Intrinsics.gen"
#undef GET_NO_MEMORY_INTRINSICS
return false;
}
// IntrinsicOnlyReadsMemory - Return true if the specified intrinsic doesn't
// have any side-effects or if it only reads memory.
static bool IntrinsicOnlyReadsMemory(unsigned IntrinsicID) {
#define GET_SIDE_EFFECT_INFO
#include "llvm/Intrinsics.gen"
#undef GET_SIDE_EFFECT_INFO
return false;
}
/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
/// node.
void SelectionDAGLowering::visitTargetIntrinsic(CallInst &I,
unsigned Intrinsic) {
bool HasChain = !IntrinsicCannotAccessMemory(Intrinsic);
bool OnlyLoad = HasChain && IntrinsicOnlyReadsMemory(Intrinsic);
// Build the operand list.
std::vector<SDOperand> Ops;
if (HasChain) { // If this intrinsic has side-effects, chainify it.
if (OnlyLoad) {
// We don't need to serialize loads against other loads.
Ops.push_back(DAG.getRoot());
} else {
Ops.push_back(getRoot());
}
}
// Add the intrinsic ID as an integer operand.
Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
// Add all operands of the call to the operand list.
for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
SDOperand Op = getValue(I.getOperand(i));
// If this is a vector type, force it to the right packed type.
if (Op.getValueType() == MVT::Vector) {
const PackedType *OpTy = cast<PackedType>(I.getOperand(i)->getType());
MVT::ValueType EltVT = TLI.getValueType(OpTy->getElementType());
MVT::ValueType VVT = MVT::getVectorType(EltVT, OpTy->getNumElements());
assert(VVT != MVT::Other && "Intrinsic uses a non-legal type?");
Op = DAG.getNode(ISD::VBIT_CONVERT, VVT, Op);
}
assert(TLI.isTypeLegal(Op.getValueType()) &&
"Intrinsic uses a non-legal type?");
Ops.push_back(Op);
}
std::vector<MVT::ValueType> VTs;
if (I.getType() != Type::VoidTy) {
MVT::ValueType VT = TLI.getValueType(I.getType());
if (VT == MVT::Vector) {
const PackedType *DestTy = cast<PackedType>(I.getType());
MVT::ValueType EltVT = TLI.getValueType(DestTy->getElementType());
VT = MVT::getVectorType(EltVT, DestTy->getNumElements());
assert(VT != MVT::Other && "Intrinsic uses a non-legal type?");
}
assert(TLI.isTypeLegal(VT) && "Intrinsic uses a non-legal type?");
VTs.push_back(VT);
}
if (HasChain)
VTs.push_back(MVT::Other);
// Create the node.
SDOperand Result;
if (!HasChain)
Result = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, VTs, Ops);
else if (I.getType() != Type::VoidTy)
Result = DAG.getNode(ISD::INTRINSIC_W_CHAIN, VTs, Ops);
else
Result = DAG.getNode(ISD::INTRINSIC_VOID, VTs, Ops);
if (HasChain) {
SDOperand Chain = Result.getValue(Result.Val->getNumValues()-1);
if (OnlyLoad)
PendingLoads.push_back(Chain);
else
DAG.setRoot(Chain);
}
if (I.getType() != Type::VoidTy) {
if (const PackedType *PTy = dyn_cast<PackedType>(I.getType())) {
MVT::ValueType EVT = TLI.getValueType(PTy->getElementType());
Result = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Result,
DAG.getConstant(PTy->getNumElements(), MVT::i32),
DAG.getValueType(EVT));
}
setValue(&I, Result);
}
}
/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
/// we want to emit this as a call to a named external function, return the name
/// otherwise lower it and return null.
const char *
SelectionDAGLowering::visitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
switch (Intrinsic) {
default:
// By default, turn this into a target intrinsic node.
visitTargetIntrinsic(I, Intrinsic);
return 0;
case Intrinsic::vastart: visitVAStart(I); return 0;
case Intrinsic::vaend: visitVAEnd(I); return 0;
case Intrinsic::vacopy: visitVACopy(I); return 0;
case Intrinsic::returnaddress: visitFrameReturnAddress(I, false); return 0;
case Intrinsic::frameaddress: visitFrameReturnAddress(I, true); return 0;
case Intrinsic::setjmp:
return "_setjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
break;
case Intrinsic::longjmp:
return "_longjmp"+!TLI.usesUnderscoreSetJmpLongJmp();
break;
case Intrinsic::memcpy_i32:
case Intrinsic::memcpy_i64:
visitMemIntrinsic(I, ISD::MEMCPY);
return 0;
case Intrinsic::memset_i32:
case Intrinsic::memset_i64:
visitMemIntrinsic(I, ISD::MEMSET);
return 0;
case Intrinsic::memmove_i32:
case Intrinsic::memmove_i64:
visitMemIntrinsic(I, ISD::MEMMOVE);
return 0;
case Intrinsic::dbg_stoppoint: {
MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
DbgStopPointInst &SPI = cast<DbgStopPointInst>(I);
if (DebugInfo && SPI.getContext() && DebugInfo->Verify(SPI.getContext())) {
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
Ops.push_back(getValue(SPI.getLineValue()));
Ops.push_back(getValue(SPI.getColumnValue()));
DebugInfoDesc *DD = DebugInfo->getDescFor(SPI.getContext());
assert(DD && "Not a debug information descriptor");
CompileUnitDesc *CompileUnit = cast<CompileUnitDesc>(DD);
Ops.push_back(DAG.getString(CompileUnit->getFileName()));
Ops.push_back(DAG.getString(CompileUnit->getDirectory()));
DAG.setRoot(DAG.getNode(ISD::LOCATION, MVT::Other, Ops));
}
return 0;
}
case Intrinsic::dbg_region_start: {
MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
DbgRegionStartInst &RSI = cast<DbgRegionStartInst>(I);
if (DebugInfo && RSI.getContext() && DebugInfo->Verify(RSI.getContext())) {
std::vector<SDOperand> Ops;
unsigned LabelID = DebugInfo->RecordRegionStart(RSI.getContext());
Ops.push_back(getRoot());
Ops.push_back(DAG.getConstant(LabelID, MVT::i32));
DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, Ops));
}
return 0;
}
case Intrinsic::dbg_region_end: {
MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
DbgRegionEndInst &REI = cast<DbgRegionEndInst>(I);
if (DebugInfo && REI.getContext() && DebugInfo->Verify(REI.getContext())) {
std::vector<SDOperand> Ops;
unsigned LabelID = DebugInfo->RecordRegionEnd(REI.getContext());
Ops.push_back(getRoot());
Ops.push_back(DAG.getConstant(LabelID, MVT::i32));
DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, Ops));
}
return 0;
}
case Intrinsic::dbg_func_start: {
MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
DbgFuncStartInst &FSI = cast<DbgFuncStartInst>(I);
if (DebugInfo && FSI.getSubprogram() &&
DebugInfo->Verify(FSI.getSubprogram())) {
std::vector<SDOperand> Ops;
unsigned LabelID = DebugInfo->RecordRegionStart(FSI.getSubprogram());
Ops.push_back(getRoot());
Ops.push_back(DAG.getConstant(LabelID, MVT::i32));
DAG.setRoot(DAG.getNode(ISD::DEBUG_LABEL, MVT::Other, Ops));
}
return 0;
}
case Intrinsic::dbg_declare: {
MachineDebugInfo *DebugInfo = DAG.getMachineDebugInfo();
DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
if (DebugInfo && DI.getVariable() && DebugInfo->Verify(DI.getVariable())) {
std::vector<SDOperand> Ops;
SDOperand AddressOp = getValue(DI.getAddress());
if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(AddressOp)) {
DebugInfo->RecordVariable(DI.getVariable(), FI->getIndex());
}
}
return 0;
}
case Intrinsic::isunordered_f32:
case Intrinsic::isunordered_f64:
setValue(&I, DAG.getSetCC(MVT::i1,getValue(I.getOperand(1)),
getValue(I.getOperand(2)), ISD::SETUO));
return 0;
case Intrinsic::sqrt_f32:
case Intrinsic::sqrt_f64:
setValue(&I, DAG.getNode(ISD::FSQRT,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1))));
return 0;
case Intrinsic::pcmarker: {
SDOperand Tmp = getValue(I.getOperand(1));
DAG.setRoot(DAG.getNode(ISD::PCMARKER, MVT::Other, getRoot(), Tmp));
return 0;
}
case Intrinsic::readcyclecounter: {
std::vector<MVT::ValueType> VTs;
VTs.push_back(MVT::i64);
VTs.push_back(MVT::Other);
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
SDOperand Tmp = DAG.getNode(ISD::READCYCLECOUNTER, VTs, Ops);
setValue(&I, Tmp);
DAG.setRoot(Tmp.getValue(1));
return 0;
}
case Intrinsic::bswap_i16:
case Intrinsic::bswap_i32:
case Intrinsic::bswap_i64:
setValue(&I, DAG.getNode(ISD::BSWAP,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1))));
return 0;
case Intrinsic::cttz_i8:
case Intrinsic::cttz_i16:
case Intrinsic::cttz_i32:
case Intrinsic::cttz_i64:
setValue(&I, DAG.getNode(ISD::CTTZ,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1))));
return 0;
case Intrinsic::ctlz_i8:
case Intrinsic::ctlz_i16:
case Intrinsic::ctlz_i32:
case Intrinsic::ctlz_i64:
setValue(&I, DAG.getNode(ISD::CTLZ,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1))));
return 0;
case Intrinsic::ctpop_i8:
case Intrinsic::ctpop_i16:
case Intrinsic::ctpop_i32:
case Intrinsic::ctpop_i64:
setValue(&I, DAG.getNode(ISD::CTPOP,
getValue(I.getOperand(1)).getValueType(),
getValue(I.getOperand(1))));
return 0;
case Intrinsic::stacksave: {
std::vector<MVT::ValueType> VTs;
VTs.push_back(TLI.getPointerTy());
VTs.push_back(MVT::Other);
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
SDOperand Tmp = DAG.getNode(ISD::STACKSAVE, VTs, Ops);
setValue(&I, Tmp);
DAG.setRoot(Tmp.getValue(1));
return 0;
}
case Intrinsic::stackrestore: {
SDOperand Tmp = getValue(I.getOperand(1));
DAG.setRoot(DAG.getNode(ISD::STACKRESTORE, MVT::Other, getRoot(), Tmp));
return 0;
}
case Intrinsic::prefetch:
// FIXME: Currently discarding prefetches.
return 0;
}
}
void SelectionDAGLowering::visitCall(CallInst &I) {
const char *RenameFn = 0;
if (Function *F = I.getCalledFunction()) {
if (F->isExternal())
if (unsigned IID = F->getIntrinsicID()) {
RenameFn = visitIntrinsicCall(I, IID);
if (!RenameFn)
return;
} else { // Not an LLVM intrinsic.
const std::string &Name = F->getName();
if (Name[0] == 'c' && (Name == "copysign" || Name == "copysignf")) {
if (I.getNumOperands() == 3 && // Basic sanity checks.
I.getOperand(1)->getType()->isFloatingPoint() &&
I.getType() == I.getOperand(1)->getType() &&
I.getType() == I.getOperand(2)->getType()) {
SDOperand LHS = getValue(I.getOperand(1));
SDOperand RHS = getValue(I.getOperand(2));
setValue(&I, DAG.getNode(ISD::FCOPYSIGN, LHS.getValueType(),
LHS, RHS));
return;
}
} else if (Name[0] == 'f' && (Name == "fabs" || Name == "fabsf")) {
if (I.getNumOperands() == 2 && // Basic sanity checks.
I.getOperand(1)->getType()->isFloatingPoint() &&
I.getType() == I.getOperand(1)->getType()) {
SDOperand Tmp = getValue(I.getOperand(1));
setValue(&I, DAG.getNode(ISD::FABS, Tmp.getValueType(), Tmp));
return;
}
} else if (Name[0] == 's' && (Name == "sin" || Name == "sinf")) {
if (I.getNumOperands() == 2 && // Basic sanity checks.
I.getOperand(1)->getType()->isFloatingPoint() &&
I.getType() == I.getOperand(1)->getType()) {
SDOperand Tmp = getValue(I.getOperand(1));
setValue(&I, DAG.getNode(ISD::FSIN, Tmp.getValueType(), Tmp));
return;
}
} else if (Name[0] == 'c' && (Name == "cos" || Name == "cosf")) {
if (I.getNumOperands() == 2 && // Basic sanity checks.
I.getOperand(1)->getType()->isFloatingPoint() &&
I.getType() == I.getOperand(1)->getType()) {
SDOperand Tmp = getValue(I.getOperand(1));
setValue(&I, DAG.getNode(ISD::FCOS, Tmp.getValueType(), Tmp));
return;
}
}
}
} else if (isa<InlineAsm>(I.getOperand(0))) {
visitInlineAsm(I);
return;
}
SDOperand Callee;
if (!RenameFn)
Callee = getValue(I.getOperand(0));
else
Callee = DAG.getExternalSymbol(RenameFn, TLI.getPointerTy());
std::vector<std::pair<SDOperand, const Type*> > Args;
Args.reserve(I.getNumOperands());
for (unsigned i = 1, e = I.getNumOperands(); i != e; ++i) {
Value *Arg = I.getOperand(i);
SDOperand ArgNode = getValue(Arg);
Args.push_back(std::make_pair(ArgNode, Arg->getType()));
}
const PointerType *PT = cast<PointerType>(I.getCalledValue()->getType());
const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
std::pair<SDOperand,SDOperand> Result =
TLI.LowerCallTo(getRoot(), I.getType(), FTy->isVarArg(), I.getCallingConv(),
I.isTailCall(), Callee, Args, DAG);
if (I.getType() != Type::VoidTy)
setValue(&I, Result.first);
DAG.setRoot(Result.second);
}
SDOperand RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
SDOperand &Chain, SDOperand &Flag)const{
SDOperand Val = DAG.getCopyFromReg(Chain, Regs[0], RegVT, Flag);
Chain = Val.getValue(1);
Flag = Val.getValue(2);
// If the result was expanded, copy from the top part.
if (Regs.size() > 1) {
assert(Regs.size() == 2 &&
"Cannot expand to more than 2 elts yet!");
SDOperand Hi = DAG.getCopyFromReg(Chain, Regs[1], RegVT, Flag);
Chain = Val.getValue(1);
Flag = Val.getValue(2);
if (DAG.getTargetLoweringInfo().isLittleEndian())
return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Val, Hi);
else
return DAG.getNode(ISD::BUILD_PAIR, ValueVT, Hi, Val);
}
// Otherwise, if the return value was promoted, truncate it to the
// appropriate type.
if (RegVT == ValueVT)
return Val;
if (MVT::isInteger(RegVT))
return DAG.getNode(ISD::TRUNCATE, ValueVT, Val);
else
return DAG.getNode(ISD::FP_ROUND, ValueVT, Val);
}
/// getCopyToRegs - Emit a series of CopyToReg nodes that copies the
/// specified value into the registers specified by this object. This uses
/// Chain/Flag as the input and updates them for the output Chain/Flag.
void RegsForValue::getCopyToRegs(SDOperand Val, SelectionDAG &DAG,
SDOperand &Chain, SDOperand &Flag) const {
if (Regs.size() == 1) {
// If there is a single register and the types differ, this must be
// a promotion.
if (RegVT != ValueVT) {
if (MVT::isInteger(RegVT))
Val = DAG.getNode(ISD::ANY_EXTEND, RegVT, Val);
else
Val = DAG.getNode(ISD::FP_EXTEND, RegVT, Val);
}
Chain = DAG.getCopyToReg(Chain, Regs[0], Val, Flag);
Flag = Chain.getValue(1);
} else {
std::vector<unsigned> R(Regs);
if (!DAG.getTargetLoweringInfo().isLittleEndian())
std::reverse(R.begin(), R.end());
for (unsigned i = 0, e = R.size(); i != e; ++i) {
SDOperand Part = DAG.getNode(ISD::EXTRACT_ELEMENT, RegVT, Val,
DAG.getConstant(i, MVT::i32));
Chain = DAG.getCopyToReg(Chain, R[i], Part, Flag);
Flag = Chain.getValue(1);
}
}
}
/// AddInlineAsmOperands - Add this value to the specified inlineasm node
/// operand list. This adds the code marker and includes the number of
/// values added into it.
void RegsForValue::AddInlineAsmOperands(unsigned Code, SelectionDAG &DAG,
std::vector<SDOperand> &Ops) const {
Ops.push_back(DAG.getConstant(Code | (Regs.size() << 3), MVT::i32));
for (unsigned i = 0, e = Regs.size(); i != e; ++i)
Ops.push_back(DAG.getRegister(Regs[i], RegVT));
}
/// isAllocatableRegister - If the specified register is safe to allocate,
/// i.e. it isn't a stack pointer or some other special register, return the
/// register class for the register. Otherwise, return null.
static const TargetRegisterClass *
isAllocatableRegister(unsigned Reg, MachineFunction &MF,
const TargetLowering &TLI, const MRegisterInfo *MRI) {
MVT::ValueType FoundVT = MVT::Other;
const TargetRegisterClass *FoundRC = 0;
for (MRegisterInfo::regclass_iterator RCI = MRI->regclass_begin(),
E = MRI->regclass_end(); RCI != E; ++RCI) {
MVT::ValueType ThisVT = MVT::Other;
const TargetRegisterClass *RC = *RCI;
// If none of the the value types for this register class are valid, we
// can't use it. For example, 64-bit reg classes on 32-bit targets.
for (TargetRegisterClass::vt_iterator I = RC->vt_begin(), E = RC->vt_end();
I != E; ++I) {
if (TLI.isTypeLegal(*I)) {
// If we have already found this register in a different register class,
// choose the one with the largest VT specified. For example, on
// PowerPC, we favor f64 register classes over f32.
if (FoundVT == MVT::Other ||
MVT::getSizeInBits(FoundVT) < MVT::getSizeInBits(*I)) {
ThisVT = *I;
break;
}
}
}
if (ThisVT == MVT::Other) continue;
// NOTE: This isn't ideal. In particular, this might allocate the
// frame pointer in functions that need it (due to them not being taken
// out of allocation, because a variable sized allocation hasn't been seen
// yet). This is a slight code pessimization, but should still work.
for (TargetRegisterClass::iterator I = RC->allocation_order_begin(MF),
E = RC->allocation_order_end(MF); I != E; ++I)
if (*I == Reg) {
// We found a matching register class. Keep looking at others in case
// we find one with larger registers that this physreg is also in.
FoundRC = RC;
FoundVT = ThisVT;
break;
}
}
return FoundRC;
}
RegsForValue SelectionDAGLowering::
GetRegistersForValue(const std::string &ConstrCode,
MVT::ValueType VT, bool isOutReg, bool isInReg,
std::set<unsigned> &OutputRegs,
std::set<unsigned> &InputRegs) {
std::pair<unsigned, const TargetRegisterClass*> PhysReg =
TLI.getRegForInlineAsmConstraint(ConstrCode, VT);
std::vector<unsigned> Regs;
unsigned NumRegs = VT != MVT::Other ? TLI.getNumElements(VT) : 1;
MVT::ValueType RegVT;
MVT::ValueType ValueVT = VT;
if (PhysReg.first) {
if (VT == MVT::Other)
ValueVT = *PhysReg.second->vt_begin();
RegVT = VT;
// This is a explicit reference to a physical register.
Regs.push_back(PhysReg.first);
// If this is an expanded reference, add the rest of the regs to Regs.
if (NumRegs != 1) {
RegVT = *PhysReg.second->vt_begin();
TargetRegisterClass::iterator I = PhysReg.second->begin();
TargetRegisterClass::iterator E = PhysReg.second->end();
for (; *I != PhysReg.first; ++I)
assert(I != E && "Didn't find reg!");
// Already added the first reg.
--NumRegs; ++I;
for (; NumRegs; --NumRegs, ++I) {
assert(I != E && "Ran out of registers to allocate!");
Regs.push_back(*I);
}
}
return RegsForValue(Regs, RegVT, ValueVT);
}
// This is a reference to a register class. Allocate NumRegs consecutive,
// available, registers from the class.
std::vector<unsigned> RegClassRegs =
TLI.getRegClassForInlineAsmConstraint(ConstrCode, VT);
const MRegisterInfo *MRI = DAG.getTarget().getRegisterInfo();
MachineFunction &MF = *CurMBB->getParent();
unsigned NumAllocated = 0;
for (unsigned i = 0, e = RegClassRegs.size(); i != e; ++i) {
unsigned Reg = RegClassRegs[i];
// See if this register is available.
if ((isOutReg && OutputRegs.count(Reg)) || // Already used.
(isInReg && InputRegs.count(Reg))) { // Already used.
// Make sure we find consecutive registers.
NumAllocated = 0;
continue;
}
// Check to see if this register is allocatable (i.e. don't give out the
// stack pointer).
const TargetRegisterClass *RC = isAllocatableRegister(Reg, MF, TLI, MRI);
if (!RC) {
// Make sure we find consecutive registers.
NumAllocated = 0;
continue;
}
// Okay, this register is good, we can use it.
++NumAllocated;
// If we allocated enough consecutive
if (NumAllocated == NumRegs) {
unsigned RegStart = (i-NumAllocated)+1;
unsigned RegEnd = i+1;
// Mark all of the allocated registers used.
for (unsigned i = RegStart; i != RegEnd; ++i) {
unsigned Reg = RegClassRegs[i];
Regs.push_back(Reg);
if (isOutReg) OutputRegs.insert(Reg); // Mark reg used.
if (isInReg) InputRegs.insert(Reg); // Mark reg used.
}
return RegsForValue(Regs, *RC->vt_begin(), VT);
}
}
// Otherwise, we couldn't allocate enough registers for this.
return RegsForValue();
}
/// visitInlineAsm - Handle a call to an InlineAsm object.
///
void SelectionDAGLowering::visitInlineAsm(CallInst &I) {
InlineAsm *IA = cast<InlineAsm>(I.getOperand(0));
SDOperand AsmStr = DAG.getTargetExternalSymbol(IA->getAsmString().c_str(),
MVT::Other);
// Note, we treat inline asms both with and without side-effects as the same.
// If an inline asm doesn't have side effects and doesn't access memory, we
// could not choose to not chain it.
bool hasSideEffects = IA->hasSideEffects();
std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
std::vector<MVT::ValueType> ConstraintVTs;
/// AsmNodeOperands - A list of pairs. The first element is a register, the
/// second is a bitfield where bit #0 is set if it is a use and bit #1 is set
/// if it is a def of that register.
std::vector<SDOperand> AsmNodeOperands;
AsmNodeOperands.push_back(SDOperand()); // reserve space for input chain
AsmNodeOperands.push_back(AsmStr);
SDOperand Chain = getRoot();
SDOperand Flag;
// We fully assign registers here at isel time. This is not optimal, but
// should work. For register classes that correspond to LLVM classes, we
// could let the LLVM RA do its thing, but we currently don't. Do a prepass
// over the constraints, collecting fixed registers that we know we can't use.
std::set<unsigned> OutputRegs, InputRegs;
unsigned OpNum = 1;
for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
std::string &ConstraintCode = Constraints[i].Codes[0];
MVT::ValueType OpVT;
// Compute the value type for each operand and add it to ConstraintVTs.
switch (Constraints[i].Type) {
case InlineAsm::isOutput:
if (!Constraints[i].isIndirectOutput) {
assert(I.getType() != Type::VoidTy && "Bad inline asm!");
OpVT = TLI.getValueType(I.getType());
} else {
const Type *OpTy = I.getOperand(OpNum)->getType();
OpVT = TLI.getValueType(cast<PointerType>(OpTy)->getElementType());
OpNum++; // Consumes a call operand.
}
break;
case InlineAsm::isInput:
OpVT = TLI.getValueType(I.getOperand(OpNum)->getType());
OpNum++; // Consumes a call operand.
break;
case InlineAsm::isClobber:
OpVT = MVT::Other;
break;
}
ConstraintVTs.push_back(OpVT);
if (TLI.getRegForInlineAsmConstraint(ConstraintCode, OpVT).first == 0)
continue; // Not assigned a fixed reg.
// Build a list of regs that this operand uses. This always has a single
// element for promoted/expanded operands.
RegsForValue Regs = GetRegistersForValue(ConstraintCode, OpVT,
false, false,
OutputRegs, InputRegs);
switch (Constraints[i].Type) {
case InlineAsm::isOutput:
// We can't assign any other output to this register.
OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
// If this is an early-clobber output, it cannot be assigned to the same
// value as the input reg.
if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
break;
case InlineAsm::isInput:
// We can't assign any other input to this register.
InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
break;
case InlineAsm::isClobber:
// Clobbered regs cannot be used as inputs or outputs.
InputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
OutputRegs.insert(Regs.Regs.begin(), Regs.Regs.end());
break;
}
}
// Loop over all of the inputs, copying the operand values into the
// appropriate registers and processing the output regs.
RegsForValue RetValRegs;
std::vector<std::pair<RegsForValue, Value*> > IndirectStoresToEmit;
OpNum = 1;
for (unsigned i = 0, e = Constraints.size(); i != e; ++i) {
assert(Constraints[i].Codes.size() == 1 && "Only handles one code so far!");
std::string &ConstraintCode = Constraints[i].Codes[0];
switch (Constraints[i].Type) {
case InlineAsm::isOutput: {
TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
if (ConstraintCode.size() == 1) // not a physreg name.
CTy = TLI.getConstraintType(ConstraintCode[0]);
if (CTy == TargetLowering::C_Memory) {
// Memory output.
SDOperand InOperandVal = getValue(I.getOperand(OpNum));
// Check that the operand (the address to store to) isn't a float.
if (!MVT::isInteger(InOperandVal.getValueType()))
assert(0 && "MATCH FAIL!");
if (!Constraints[i].isIndirectOutput)
assert(0 && "MATCH FAIL!");
OpNum++; // Consumes a call operand.
// Extend/truncate to the right pointer type if needed.
MVT::ValueType PtrType = TLI.getPointerTy();
if (InOperandVal.getValueType() < PtrType)
InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
else if (InOperandVal.getValueType() > PtrType)
InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
// Add information to the INLINEASM node to know about this output.
unsigned ResOpType = 4/*MEM*/ | (1 << 3);
AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
AsmNodeOperands.push_back(InOperandVal);
break;
}
// Otherwise, this is a register output.
assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
// If this is an early-clobber output, or if there is an input
// constraint that matches this, we need to reserve the input register
// so no other inputs allocate to it.
bool UsesInputRegister = false;
if (Constraints[i].isEarlyClobber || Constraints[i].hasMatchingInput)
UsesInputRegister = true;
// Copy the output from the appropriate register. Find a register that
// we can use.
RegsForValue Regs =
GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
true, UsesInputRegister,
OutputRegs, InputRegs);
assert(!Regs.Regs.empty() && "Couldn't allocate output reg!");
if (!Constraints[i].isIndirectOutput) {
assert(RetValRegs.Regs.empty() &&
"Cannot have multiple output constraints yet!");
assert(I.getType() != Type::VoidTy && "Bad inline asm!");
RetValRegs = Regs;
} else {
IndirectStoresToEmit.push_back(std::make_pair(Regs,
I.getOperand(OpNum)));
OpNum++; // Consumes a call operand.
}
// Add information to the INLINEASM node to know that this register is
// set.
Regs.AddInlineAsmOperands(2 /*REGDEF*/, DAG, AsmNodeOperands);
break;
}
case InlineAsm::isInput: {
SDOperand InOperandVal = getValue(I.getOperand(OpNum));
OpNum++; // Consumes a call operand.
if (isdigit(ConstraintCode[0])) { // Matching constraint?
// If this is required to match an output register we have already set,
// just use its register.
unsigned OperandNo = atoi(ConstraintCode.c_str());
// Scan until we find the definition we already emitted of this operand.
// When we find it, create a RegsForValue operand.
unsigned CurOp = 2; // The first operand.
for (; OperandNo; --OperandNo) {
// Advance to the next operand.
unsigned NumOps =
cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
assert((NumOps & 7) == 2 /*REGDEF*/ &&
"Skipped past definitions?");
CurOp += (NumOps>>3)+1;
}
unsigned NumOps =
cast<ConstantSDNode>(AsmNodeOperands[CurOp])->getValue();
assert((NumOps & 7) == 2 /*REGDEF*/ &&
"Skipped past definitions?");
// Add NumOps>>3 registers to MatchedRegs.
RegsForValue MatchedRegs;
MatchedRegs.ValueVT = InOperandVal.getValueType();
MatchedRegs.RegVT = AsmNodeOperands[CurOp+1].getValueType();
for (unsigned i = 0, e = NumOps>>3; i != e; ++i) {
unsigned Reg=cast<RegisterSDNode>(AsmNodeOperands[++CurOp])->getReg();
MatchedRegs.Regs.push_back(Reg);
}
// Use the produced MatchedRegs object to
MatchedRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag);
MatchedRegs.AddInlineAsmOperands(1 /*REGUSE*/, DAG, AsmNodeOperands);
break;
}
TargetLowering::ConstraintType CTy = TargetLowering::C_RegisterClass;
if (ConstraintCode.size() == 1) // not a physreg name.
CTy = TLI.getConstraintType(ConstraintCode[0]);
if (CTy == TargetLowering::C_Other) {
if (!TLI.isOperandValidForConstraint(InOperandVal, ConstraintCode[0]))
assert(0 && "MATCH FAIL!");
// Add information to the INLINEASM node to know about this input.
unsigned ResOpType = 3 /*IMM*/ | (1 << 3);
AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
AsmNodeOperands.push_back(InOperandVal);
break;
} else if (CTy == TargetLowering::C_Memory) {
// Memory input.
// Check that the operand isn't a float.
if (!MVT::isInteger(InOperandVal.getValueType()))
assert(0 && "MATCH FAIL!");
// Extend/truncate to the right pointer type if needed.
MVT::ValueType PtrType = TLI.getPointerTy();
if (InOperandVal.getValueType() < PtrType)
InOperandVal = DAG.getNode(ISD::ZERO_EXTEND, PtrType, InOperandVal);
else if (InOperandVal.getValueType() > PtrType)
InOperandVal = DAG.getNode(ISD::TRUNCATE, PtrType, InOperandVal);
// Add information to the INLINEASM node to know about this input.
unsigned ResOpType = 4/*MEM*/ | (1 << 3);
AsmNodeOperands.push_back(DAG.getConstant(ResOpType, MVT::i32));
AsmNodeOperands.push_back(InOperandVal);
break;
}
assert(CTy == TargetLowering::C_RegisterClass && "Unknown op type!");
// Copy the input into the appropriate registers.
RegsForValue InRegs =
GetRegistersForValue(ConstraintCode, ConstraintVTs[i],
false, true, OutputRegs, InputRegs);
// FIXME: should be match fail.
assert(!InRegs.Regs.empty() && "Couldn't allocate input reg!");
InRegs.getCopyToRegs(InOperandVal, DAG, Chain, Flag);
InRegs.AddInlineAsmOperands(1/*REGUSE*/, DAG, AsmNodeOperands);
break;
}
case InlineAsm::isClobber: {
RegsForValue ClobberedRegs =
GetRegistersForValue(ConstraintCode, MVT::Other, false, false,
OutputRegs, InputRegs);
// Add the clobbered value to the operand list, so that the register
// allocator is aware that the physreg got clobbered.
if (!ClobberedRegs.Regs.empty())
ClobberedRegs.AddInlineAsmOperands(2/*REGDEF*/, DAG, AsmNodeOperands);
break;
}
}
}
// Finish up input operands.
AsmNodeOperands[0] = Chain;
if (Flag.Val) AsmNodeOperands.push_back(Flag);
std::vector<MVT::ValueType> VTs;
VTs.push_back(MVT::Other);
VTs.push_back(MVT::Flag);
Chain = DAG.getNode(ISD::INLINEASM, VTs, AsmNodeOperands);
Flag = Chain.getValue(1);
// If this asm returns a register value, copy the result from that register
// and set it as the value of the call.
if (!RetValRegs.Regs.empty())
setValue(&I, RetValRegs.getCopyFromRegs(DAG, Chain, Flag));
std::vector<std::pair<SDOperand, Value*> > StoresToEmit;
// Process indirect outputs, first output all of the flagged copies out of
// physregs.
for (unsigned i = 0, e = IndirectStoresToEmit.size(); i != e; ++i) {
RegsForValue &OutRegs = IndirectStoresToEmit[i].first;
Value *Ptr = IndirectStoresToEmit[i].second;
SDOperand OutVal = OutRegs.getCopyFromRegs(DAG, Chain, Flag);
StoresToEmit.push_back(std::make_pair(OutVal, Ptr));
}
// Emit the non-flagged stores from the physregs.
std::vector<SDOperand> OutChains;
for (unsigned i = 0, e = StoresToEmit.size(); i != e; ++i)
OutChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
StoresToEmit[i].first,
getValue(StoresToEmit[i].second),
DAG.getSrcValue(StoresToEmit[i].second)));
if (!OutChains.empty())
Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains);
DAG.setRoot(Chain);
}
void SelectionDAGLowering::visitMalloc(MallocInst &I) {
SDOperand Src = getValue(I.getOperand(0));
MVT::ValueType IntPtr = TLI.getPointerTy();
if (IntPtr < Src.getValueType())
Src = DAG.getNode(ISD::TRUNCATE, IntPtr, Src);
else if (IntPtr > Src.getValueType())
Src = DAG.getNode(ISD::ZERO_EXTEND, IntPtr, Src);
// Scale the source by the type size.
uint64_t ElementSize = TD->getTypeSize(I.getType()->getElementType());
Src = DAG.getNode(ISD::MUL, Src.getValueType(),
Src, getIntPtrConstant(ElementSize));
std::vector<std::pair<SDOperand, const Type*> > Args;
Args.push_back(std::make_pair(Src, TLI.getTargetData()->getIntPtrType()));
std::pair<SDOperand,SDOperand> Result =
TLI.LowerCallTo(getRoot(), I.getType(), false, CallingConv::C, true,
DAG.getExternalSymbol("malloc", IntPtr),
Args, DAG);
setValue(&I, Result.first); // Pointers always fit in registers
DAG.setRoot(Result.second);
}
void SelectionDAGLowering::visitFree(FreeInst &I) {
std::vector<std::pair<SDOperand, const Type*> > Args;
Args.push_back(std::make_pair(getValue(I.getOperand(0)),
TLI.getTargetData()->getIntPtrType()));
MVT::ValueType IntPtr = TLI.getPointerTy();
std::pair<SDOperand,SDOperand> Result =
TLI.LowerCallTo(getRoot(), Type::VoidTy, false, CallingConv::C, true,
DAG.getExternalSymbol("free", IntPtr), Args, DAG);
DAG.setRoot(Result.second);
}
// InsertAtEndOfBasicBlock - This method should be implemented by targets that
// mark instructions with the 'usesCustomDAGSchedInserter' flag. These
// instructions are special in various ways, which require special support to
// insert. The specified MachineInstr is created but not inserted into any
// basic blocks, and the scheduler passes ownership of it to this method.
MachineBasicBlock *TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
MachineBasicBlock *MBB) {
std::cerr << "If a target marks an instruction with "
"'usesCustomDAGSchedInserter', it must implement "
"TargetLowering::InsertAtEndOfBasicBlock!\n";
abort();
return 0;
}
void SelectionDAGLowering::visitVAStart(CallInst &I) {
DAG.setRoot(DAG.getNode(ISD::VASTART, MVT::Other, getRoot(),
getValue(I.getOperand(1)),
DAG.getSrcValue(I.getOperand(1))));
}
void SelectionDAGLowering::visitVAArg(VAArgInst &I) {
SDOperand V = DAG.getVAArg(TLI.getValueType(I.getType()), getRoot(),
getValue(I.getOperand(0)),
DAG.getSrcValue(I.getOperand(0)));
setValue(&I, V);
DAG.setRoot(V.getValue(1));
}
void SelectionDAGLowering::visitVAEnd(CallInst &I) {
DAG.setRoot(DAG.getNode(ISD::VAEND, MVT::Other, getRoot(),
getValue(I.getOperand(1)),
DAG.getSrcValue(I.getOperand(1))));
}
void SelectionDAGLowering::visitVACopy(CallInst &I) {
DAG.setRoot(DAG.getNode(ISD::VACOPY, MVT::Other, getRoot(),
getValue(I.getOperand(1)),
getValue(I.getOperand(2)),
DAG.getSrcValue(I.getOperand(1)),
DAG.getSrcValue(I.getOperand(2))));
}
/// TargetLowering::LowerArguments - This is the default LowerArguments
/// implementation, which just inserts a FORMAL_ARGUMENTS node. FIXME: When all
/// targets are migrated to using FORMAL_ARGUMENTS, this hook should be removed.
std::vector<SDOperand>
TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
// Add CC# and isVararg as operands to the FORMAL_ARGUMENTS node.
std::vector<SDOperand> Ops;
Ops.push_back(DAG.getConstant(F.getCallingConv(), getPointerTy()));
Ops.push_back(DAG.getConstant(F.isVarArg(), getPointerTy()));
// Add one result value for each formal argument.
std::vector<MVT::ValueType> RetVals;
for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
MVT::ValueType VT = getValueType(I->getType());
switch (getTypeAction(VT)) {
default: assert(0 && "Unknown type action!");
case Legal:
RetVals.push_back(VT);
break;
case Promote:
RetVals.push_back(getTypeToTransformTo(VT));
break;
case Expand:
if (VT != MVT::Vector) {
// If this is a large integer, it needs to be broken up into small
// integers. Figure out what the destination type is and how many small
// integers it turns into.
MVT::ValueType NVT = getTypeToTransformTo(VT);
unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
for (unsigned i = 0; i != NumVals; ++i)
RetVals.push_back(NVT);
} else {
// Otherwise, this is a vector type. We only support legal vectors
// right now.
unsigned NumElems = cast<PackedType>(I->getType())->getNumElements();
const Type *EltTy = cast<PackedType>(I->getType())->getElementType();
// Figure out if there is a Packed type corresponding to this Vector
// type. If so, convert to the packed type.
MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
if (TVT != MVT::Other && isTypeLegal(TVT)) {
RetVals.push_back(TVT);
} else {
assert(0 && "Don't support illegal by-val vector arguments yet!");
}
}
break;
}
}
if (RetVals.size() == 0)
RetVals.push_back(MVT::isVoid);
// Create the node.
SDNode *Result = DAG.getNode(ISD::FORMAL_ARGUMENTS, RetVals, Ops).Val;
// Set up the return result vector.
Ops.clear();
unsigned i = 0;
for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
MVT::ValueType VT = getValueType(I->getType());
switch (getTypeAction(VT)) {
default: assert(0 && "Unknown type action!");
case Legal:
Ops.push_back(SDOperand(Result, i++));
break;
case Promote: {
SDOperand Op(Result, i++);
if (MVT::isInteger(VT)) {
unsigned AssertOp = I->getType()->isSigned() ? ISD::AssertSext
: ISD::AssertZext;
Op = DAG.getNode(AssertOp, Op.getValueType(), Op, DAG.getValueType(VT));
Op = DAG.getNode(ISD::TRUNCATE, VT, Op);
} else {
assert(MVT::isFloatingPoint(VT) && "Not int or FP?");
Op = DAG.getNode(ISD::FP_ROUND, VT, Op);
}
Ops.push_back(Op);
break;
}
case Expand:
if (VT != MVT::Vector) {
// If this is a large integer, it needs to be reassembled from small
// integers. Figure out what the source elt type is and how many small
// integers it is.
MVT::ValueType NVT = getTypeToTransformTo(VT);
unsigned NumVals = MVT::getSizeInBits(VT)/MVT::getSizeInBits(NVT);
if (NumVals == 2) {
SDOperand Lo = SDOperand(Result, i++);
SDOperand Hi = SDOperand(Result, i++);
if (!isLittleEndian())
std::swap(Lo, Hi);
Ops.push_back(DAG.getNode(ISD::BUILD_PAIR, VT, Lo, Hi));
} else {
// Value scalarized into many values. Unimp for now.
assert(0 && "Cannot expand i64 -> i16 yet!");
}
} else {
// Otherwise, this is a vector type. We only support legal vectors
// right now.
const PackedType *PTy = cast<PackedType>(I->getType());
unsigned NumElems = PTy->getNumElements();
const Type *EltTy = PTy->getElementType();
// Figure out if there is a Packed type corresponding to this Vector
// type. If so, convert to the packed type.
MVT::ValueType TVT = MVT::getVectorType(getValueType(EltTy), NumElems);
if (TVT != MVT::Other && isTypeLegal(TVT)) {
SDOperand N = SDOperand(Result, i++);
// Handle copies from generic vectors to registers.
MVT::ValueType PTyElementVT, PTyLegalElementVT;
unsigned NE = getPackedTypeBreakdown(PTy, PTyElementVT,
PTyLegalElementVT);
// Insert a VBIT_CONVERT of the FORMAL_ARGUMENTS to a
// "N x PTyElementVT" MVT::Vector type.
N = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, N,
DAG.getConstant(NE, MVT::i32),
DAG.getValueType(PTyElementVT));
Ops.push_back(N);
} else {
assert(0 && "Don't support illegal by-val vector arguments yet!");
}
}
break;
}
}
return Ops;
}
// It is always conservatively correct for llvm.returnaddress and
// llvm.frameaddress to return 0.
std::pair<SDOperand, SDOperand>
TargetLowering::LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain,
unsigned Depth, SelectionDAG &DAG) {
return std::make_pair(DAG.getConstant(0, getPointerTy()), Chain);
}
SDOperand TargetLowering::LowerOperation(SDOperand Op, SelectionDAG &DAG) {
assert(0 && "LowerOperation not implemented for this target!");
abort();
return SDOperand();
}
SDOperand TargetLowering::CustomPromoteOperation(SDOperand Op,
SelectionDAG &DAG) {
assert(0 && "CustomPromoteOperation not implemented for this target!");
abort();
return SDOperand();
}
void SelectionDAGLowering::visitFrameReturnAddress(CallInst &I, bool isFrame) {
unsigned Depth = (unsigned)cast<ConstantUInt>(I.getOperand(1))->getValue();
std::pair<SDOperand,SDOperand> Result =
TLI.LowerFrameReturnAddress(isFrame, getRoot(), Depth, DAG);
setValue(&I, Result.first);
DAG.setRoot(Result.second);
}
/// getMemsetValue - Vectorized representation of the memset value
/// operand.
static SDOperand getMemsetValue(SDOperand Value, MVT::ValueType VT,
SelectionDAG &DAG) {
MVT::ValueType CurVT = VT;
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Value)) {
uint64_t Val = C->getValue() & 255;
unsigned Shift = 8;
while (CurVT != MVT::i8) {
Val = (Val << Shift) | Val;
Shift <<= 1;
CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
}
return DAG.getConstant(Val, VT);
} else {
Value = DAG.getNode(ISD::ZERO_EXTEND, VT, Value);
unsigned Shift = 8;
while (CurVT != MVT::i8) {
Value =
DAG.getNode(ISD::OR, VT,
DAG.getNode(ISD::SHL, VT, Value,
DAG.getConstant(Shift, MVT::i8)), Value);
Shift <<= 1;
CurVT = (MVT::ValueType)((unsigned)CurVT - 1);
}
return Value;
}
}
/// getMemsetStringVal - Similar to getMemsetValue. Except this is only
/// used when a memcpy is turned into a memset when the source is a constant
/// string ptr.
static SDOperand getMemsetStringVal(MVT::ValueType VT,
SelectionDAG &DAG, TargetLowering &TLI,
std::string &Str, unsigned Offset) {
MVT::ValueType CurVT = VT;
uint64_t Val = 0;
unsigned MSB = getSizeInBits(VT) / 8;
if (TLI.isLittleEndian())
Offset = Offset + MSB - 1;
for (unsigned i = 0; i != MSB; ++i) {
Val = (Val << 8) | Str[Offset];
Offset += TLI.isLittleEndian() ? -1 : 1;
}
return DAG.getConstant(Val, VT);
}
/// getMemBasePlusOffset - Returns base and offset node for the
static SDOperand getMemBasePlusOffset(SDOperand Base, unsigned Offset,
SelectionDAG &DAG, TargetLowering &TLI) {
MVT::ValueType VT = Base.getValueType();
return DAG.getNode(ISD::ADD, VT, Base, DAG.getConstant(Offset, VT));
}
/// MeetsMaxMemopRequirement - Determines if the number of memory ops required
/// to replace the memset / memcpy is below the threshold. It also returns the
/// types of the sequence of memory ops to perform memset / memcpy.
static bool MeetsMaxMemopRequirement(std::vector<MVT::ValueType> &MemOps,
unsigned Limit, uint64_t Size,
unsigned Align, TargetLowering &TLI) {
MVT::ValueType VT;
if (TLI.allowsUnalignedMemoryAccesses()) {
VT = MVT::i64;
} else {
switch (Align & 7) {
case 0:
VT = MVT::i64;
break;
case 4:
VT = MVT::i32;
break;
case 2:
VT = MVT::i16;
break;
default:
VT = MVT::i8;
break;
}
}
MVT::ValueType LVT = MVT::i64;
while (!TLI.isTypeLegal(LVT))
LVT = (MVT::ValueType)((unsigned)LVT - 1);
assert(MVT::isInteger(LVT));
if (VT > LVT)
VT = LVT;
unsigned NumMemOps = 0;
while (Size != 0) {
unsigned VTSize = getSizeInBits(VT) / 8;
while (VTSize > Size) {
VT = (MVT::ValueType)((unsigned)VT - 1);
VTSize >>= 1;
}
assert(MVT::isInteger(VT));
if (++NumMemOps > Limit)
return false;
MemOps.push_back(VT);
Size -= VTSize;
}
return true;
}
void SelectionDAGLowering::visitMemIntrinsic(CallInst &I, unsigned Op) {
SDOperand Op1 = getValue(I.getOperand(1));
SDOperand Op2 = getValue(I.getOperand(2));
SDOperand Op3 = getValue(I.getOperand(3));
SDOperand Op4 = getValue(I.getOperand(4));
unsigned Align = (unsigned)cast<ConstantSDNode>(Op4)->getValue();
if (Align == 0) Align = 1;
if (ConstantSDNode *Size = dyn_cast<ConstantSDNode>(Op3)) {
std::vector<MVT::ValueType> MemOps;
// Expand memset / memcpy to a series of load / store ops
// if the size operand falls below a certain threshold.
std::vector<SDOperand> OutChains;
switch (Op) {
default: break; // Do nothing for now.
case ISD::MEMSET: {
if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemset(),
Size->getValue(), Align, TLI)) {
unsigned NumMemOps = MemOps.size();
unsigned Offset = 0;
for (unsigned i = 0; i < NumMemOps; i++) {
MVT::ValueType VT = MemOps[i];
unsigned VTSize = getSizeInBits(VT) / 8;
SDOperand Value = getMemsetValue(Op2, VT, DAG);
SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, getRoot(),
Value,
getMemBasePlusOffset(Op1, Offset, DAG, TLI),
DAG.getSrcValue(I.getOperand(1), Offset));
OutChains.push_back(Store);
Offset += VTSize;
}
}
break;
}
case ISD::MEMCPY: {
if (MeetsMaxMemopRequirement(MemOps, TLI.getMaxStoresPerMemcpy(),
Size->getValue(), Align, TLI)) {
unsigned NumMemOps = MemOps.size();
unsigned SrcOff = 0, DstOff = 0, SrcDelta = 0;
GlobalAddressSDNode *G = NULL;
std::string Str;
bool CopyFromStr = false;
if (Op2.getOpcode() == ISD::GlobalAddress)
G = cast<GlobalAddressSDNode>(Op2);
else if (Op2.getOpcode() == ISD::ADD &&
Op2.getOperand(0).getOpcode() == ISD::GlobalAddress &&
Op2.getOperand(1).getOpcode() == ISD::Constant) {
G = cast<GlobalAddressSDNode>(Op2.getOperand(0));
SrcDelta = cast<ConstantSDNode>(Op2.getOperand(1))->getValue();
}
if (G) {
GlobalVariable *GV = dyn_cast<GlobalVariable>(G->getGlobal());
if (GV) {
Str = GV->getStringValue(false);
if (!Str.empty()) {
CopyFromStr = true;
SrcOff += SrcDelta;
}
}
}
for (unsigned i = 0; i < NumMemOps; i++) {
MVT::ValueType VT = MemOps[i];
unsigned VTSize = getSizeInBits(VT) / 8;
SDOperand Value, Chain, Store;
if (CopyFromStr) {
Value = getMemsetStringVal(VT, DAG, TLI, Str, SrcOff);
Chain = getRoot();
Store =
DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
DAG.getSrcValue(I.getOperand(1), DstOff));
} else {
Value = DAG.getLoad(VT, getRoot(),
getMemBasePlusOffset(Op2, SrcOff, DAG, TLI),
DAG.getSrcValue(I.getOperand(2), SrcOff));
Chain = Value.getValue(1);
Store =
DAG.getNode(ISD::STORE, MVT::Other, Chain, Value,
getMemBasePlusOffset(Op1, DstOff, DAG, TLI),
DAG.getSrcValue(I.getOperand(1), DstOff));
}
OutChains.push_back(Store);
SrcOff += VTSize;
DstOff += VTSize;
}
}
break;
}
}
if (!OutChains.empty()) {
DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains));
return;
}
}
std::vector<SDOperand> Ops;
Ops.push_back(getRoot());
Ops.push_back(Op1);
Ops.push_back(Op2);
Ops.push_back(Op3);
Ops.push_back(Op4);
DAG.setRoot(DAG.getNode(Op, MVT::Other, Ops));
}
//===----------------------------------------------------------------------===//
// SelectionDAGISel code
//===----------------------------------------------------------------------===//
unsigned SelectionDAGISel::MakeReg(MVT::ValueType VT) {
return RegMap->createVirtualRegister(TLI.getRegClassFor(VT));
}
void SelectionDAGISel::getAnalysisUsage(AnalysisUsage &AU) const {
// FIXME: we only modify the CFG to split critical edges. This
// updates dom and loop info.
}
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
/// InsertGEPComputeCode - Insert code into BB to compute Ptr+PtrOffset,
/// casting to the type of GEPI.
static Value *InsertGEPComputeCode(Value *&V, BasicBlock *BB, Instruction *GEPI,
Value *Ptr, Value *PtrOffset) {
if (V) return V; // Already computed.
BasicBlock::iterator InsertPt;
if (BB == GEPI->getParent()) {
// If insert into the GEP's block, insert right after the GEP.
InsertPt = GEPI;
++InsertPt;
} else {
// Otherwise, insert at the top of BB, after any PHI nodes
InsertPt = BB->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
}
// If Ptr is itself a cast, but in some other BB, emit a copy of the cast into
// BB so that there is only one value live across basic blocks (the cast
// operand).
if (CastInst *CI = dyn_cast<CastInst>(Ptr))
if (CI->getParent() != BB && isa<PointerType>(CI->getOperand(0)->getType()))
Ptr = new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// Add the offset, cast it to the right type.
Ptr = BinaryOperator::createAdd(Ptr, PtrOffset, "", InsertPt);
Ptr = new CastInst(Ptr, GEPI->getType(), "", InsertPt);
return V = Ptr;
}
Sink noop copies into the basic block that uses them. This reduces the number of cross-block live ranges, and allows the bb-at-a-time selector to always coallesce these away, at isel time. This reduces the load on the coallescer and register allocator. For example on a codec on X86, we went from: 1643 asm-printer - Number of machine instrs printed 419 liveintervals - Number of loads/stores folded into instructions 1144 liveintervals - Number of identity moves eliminated after coalescing 1022 liveintervals - Number of interval joins performed 282 liveintervals - Number of intervals after coalescing 1304 liveintervals - Number of original intervals 86 regalloc - Number of times we had to backtrack 1.90232 regalloc - Ratio of intervals processed over total intervals 40 spiller - Number of values reused 182 spiller - Number of loads added 121 spiller - Number of stores added 132 spiller - Number of register spills 6 twoaddressinstruction - Number of instructions commuted to coalesce 360 twoaddressinstruction - Number of two-address instructions to: 1636 asm-printer - Number of machine instrs printed 403 liveintervals - Number of loads/stores folded into instructions 1155 liveintervals - Number of identity moves eliminated after coalescing 1033 liveintervals - Number of interval joins performed 279 liveintervals - Number of intervals after coalescing 1312 liveintervals - Number of original intervals 76 regalloc - Number of times we had to backtrack 1.88998 regalloc - Ratio of intervals processed over total intervals 1 spiller - Number of copies elided 41 spiller - Number of values reused 191 spiller - Number of loads added 114 spiller - Number of stores added 128 spiller - Number of register spills 4 twoaddressinstruction - Number of instructions commuted to coalesce 356 twoaddressinstruction - Number of two-address instructions On this testcase, this change provides a modest reduction in spill code, regalloc iterations, and total instructions emitted. It increases the number of register coallesces. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28115 91177308-0d34-0410-b5e6-96231b3b80d8
2006-05-05 01:04:50 +00:00
/// OptimizeNoopCopyExpression - We have determined that the specified cast
/// instruction is a noop copy (e.g. it's casting from one pointer type to
/// another, int->uint, or int->sbyte on PPC.
static void OptimizeNoopCopyExpression(CastInst *CI) {
BasicBlock *DefBB = CI->getParent();
/// InsertedCasts - Only insert a cast in each block once.
std::map<BasicBlock*, CastInst*> InsertedCasts;
for (Value::use_iterator UI = CI->use_begin(), E = CI->use_end();
UI != E; ) {
Use &TheUse = UI.getUse();
Instruction *User = cast<Instruction>(*UI);
// Figure out which BB this cast is used in. For PHI's this is the
// appropriate predecessor block.
BasicBlock *UserBB = User->getParent();
if (PHINode *PN = dyn_cast<PHINode>(User)) {
unsigned OpVal = UI.getOperandNo()/2;
UserBB = PN->getIncomingBlock(OpVal);
}
// Preincrement use iterator so we don't invalidate it.
++UI;
// If this user is in the same block as the cast, don't change the cast.
if (UserBB == DefBB) continue;
// If we have already inserted a cast into this block, use it.
CastInst *&InsertedCast = InsertedCasts[UserBB];
if (!InsertedCast) {
BasicBlock::iterator InsertPt = UserBB->begin();
while (isa<PHINode>(InsertPt)) ++InsertPt;
InsertedCast =
new CastInst(CI->getOperand(0), CI->getType(), "", InsertPt);
}
// Replace a use of the cast with a use of the new casat.
TheUse = InsertedCast;
}
}
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
/// OptimizeGEPExpression - Since we are doing basic-block-at-a-time instruction
/// selection, we want to be a bit careful about some things. In particular, if
/// we have a GEP instruction that is used in a different block than it is
/// defined, the addressing expression of the GEP cannot be folded into loads or
/// stores that use it. In this case, decompose the GEP and move constant
/// indices into blocks that use it.
static void OptimizeGEPExpression(GetElementPtrInst *GEPI,
const TargetData *TD) {
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// If this GEP is only used inside the block it is defined in, there is no
// need to rewrite it.
bool isUsedOutsideDefBB = false;
BasicBlock *DefBB = GEPI->getParent();
for (Value::use_iterator UI = GEPI->use_begin(), E = GEPI->use_end();
UI != E; ++UI) {
if (cast<Instruction>(*UI)->getParent() != DefBB) {
isUsedOutsideDefBB = true;
break;
}
}
if (!isUsedOutsideDefBB) return;
// If this GEP has no non-zero constant indices, there is nothing we can do,
// ignore it.
bool hasConstantIndex = false;
for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
E = GEPI->op_end(); OI != E; ++OI) {
if (ConstantInt *CI = dyn_cast<ConstantInt>(*OI))
if (CI->getRawValue()) {
hasConstantIndex = true;
break;
}
}
// If this is a GEP &Alloca, 0, 0, forward subst the frame index into uses.
if (!hasConstantIndex && !isa<AllocaInst>(GEPI->getOperand(0))) return;
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// Otherwise, decompose the GEP instruction into multiplies and adds. Sum the
// constant offset (which we now know is non-zero) and deal with it later.
uint64_t ConstantOffset = 0;
const Type *UIntPtrTy = TD->getIntPtrType();
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
Value *Ptr = new CastInst(GEPI->getOperand(0), UIntPtrTy, "", GEPI);
const Type *Ty = GEPI->getOperand(0)->getType();
for (GetElementPtrInst::op_iterator OI = GEPI->op_begin()+1,
E = GEPI->op_end(); OI != E; ++OI) {
Value *Idx = *OI;
if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
unsigned Field = cast<ConstantUInt>(Idx)->getValue();
if (Field)
ConstantOffset += TD->getStructLayout(StTy)->MemberOffsets[Field];
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
Ty = StTy->getElementType(Field);
} else {
Ty = cast<SequentialType>(Ty)->getElementType();
// Handle constant subscripts.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Idx)) {
if (CI->getRawValue() == 0) continue;
if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(CI))
ConstantOffset += (int64_t)TD->getTypeSize(Ty)*CSI->getValue();
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
else
ConstantOffset+=TD->getTypeSize(Ty)*cast<ConstantUInt>(CI)->getValue();
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
continue;
}
// Ptr = Ptr + Idx * ElementSize;
// Cast Idx to UIntPtrTy if needed.
Idx = new CastInst(Idx, UIntPtrTy, "", GEPI);
uint64_t ElementSize = TD->getTypeSize(Ty);
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// Mask off bits that should not be set.
ElementSize &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
Constant *SizeCst = ConstantUInt::get(UIntPtrTy, ElementSize);
// Multiply by the element size and add to the base.
Idx = BinaryOperator::createMul(Idx, SizeCst, "", GEPI);
Ptr = BinaryOperator::createAdd(Ptr, Idx, "", GEPI);
}
}
// Make sure that the offset fits in uintptr_t.
ConstantOffset &= ~0ULL >> (64-UIntPtrTy->getPrimitiveSizeInBits());
Constant *PtrOffset = ConstantUInt::get(UIntPtrTy, ConstantOffset);
// Okay, we have now emitted all of the variable index parts to the BB that
// the GEP is defined in. Loop over all of the using instructions, inserting
// an "add Ptr, ConstantOffset" into each block that uses it and update the
// instruction to use the newly computed value, making GEPI dead. When the
// user is a load or store instruction address, we emit the add into the user
// block, otherwise we use a canonical version right next to the gep (these
// won't be foldable as addresses, so we might as well share the computation).
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
std::map<BasicBlock*,Value*> InsertedExprs;
while (!GEPI->use_empty()) {
Instruction *User = cast<Instruction>(GEPI->use_back());
// If this use is not foldable into the addressing mode, use a version
// emitted in the GEP block.
Value *NewVal;
if (!isa<LoadInst>(User) &&
(!isa<StoreInst>(User) || User->getOperand(0) == GEPI)) {
NewVal = InsertGEPComputeCode(InsertedExprs[DefBB], DefBB, GEPI,
Ptr, PtrOffset);
} else {
// Otherwise, insert the code in the User's block so it can be folded into
// any users in that block.
NewVal = InsertGEPComputeCode(InsertedExprs[User->getParent()],
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
User->getParent(), GEPI,
Ptr, PtrOffset);
}
User->replaceUsesOfWith(GEPI, NewVal);
}
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// Finally, the GEP is dead, remove it.
GEPI->eraseFromParent();
}
bool SelectionDAGISel::runOnFunction(Function &Fn) {
MachineFunction &MF = MachineFunction::construct(&Fn, TLI.getTargetMachine());
RegMap = MF.getSSARegMap();
DEBUG(std::cerr << "\n\n\n=== " << Fn.getName() << "\n");
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
// First, split all critical edges for PHI nodes with incoming values that are
// constants, this way the load of the constant into a vreg will not be placed
// into MBBs that are used some other way.
//
Sink noop copies into the basic block that uses them. This reduces the number of cross-block live ranges, and allows the bb-at-a-time selector to always coallesce these away, at isel time. This reduces the load on the coallescer and register allocator. For example on a codec on X86, we went from: 1643 asm-printer - Number of machine instrs printed 419 liveintervals - Number of loads/stores folded into instructions 1144 liveintervals - Number of identity moves eliminated after coalescing 1022 liveintervals - Number of interval joins performed 282 liveintervals - Number of intervals after coalescing 1304 liveintervals - Number of original intervals 86 regalloc - Number of times we had to backtrack 1.90232 regalloc - Ratio of intervals processed over total intervals 40 spiller - Number of values reused 182 spiller - Number of loads added 121 spiller - Number of stores added 132 spiller - Number of register spills 6 twoaddressinstruction - Number of instructions commuted to coalesce 360 twoaddressinstruction - Number of two-address instructions to: 1636 asm-printer - Number of machine instrs printed 403 liveintervals - Number of loads/stores folded into instructions 1155 liveintervals - Number of identity moves eliminated after coalescing 1033 liveintervals - Number of interval joins performed 279 liveintervals - Number of intervals after coalescing 1312 liveintervals - Number of original intervals 76 regalloc - Number of times we had to backtrack 1.88998 regalloc - Ratio of intervals processed over total intervals 1 spiller - Number of copies elided 41 spiller - Number of values reused 191 spiller - Number of loads added 114 spiller - Number of stores added 128 spiller - Number of register spills 4 twoaddressinstruction - Number of instructions commuted to coalesce 356 twoaddressinstruction - Number of two-address instructions On this testcase, this change provides a modest reduction in spill code, regalloc iterations, and total instructions emitted. It increases the number of register coallesces. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28115 91177308-0d34-0410-b5e6-96231b3b80d8
2006-05-05 01:04:50 +00:00
// In this pass we also look for GEP and cast instructions that are used
// across basic blocks and rewrite them to improve basic-block-at-a-time
// selection.
//
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
//
for (Function::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
PHINode *PN;
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
BasicBlock::iterator BBI;
for (BBI = BB->begin(); (PN = dyn_cast<PHINode>(BBI)); ++BBI)
for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
if (isa<Constant>(PN->getIncomingValue(i)))
SplitCriticalEdge(PN->getIncomingBlock(i), BB);
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
Sink noop copies into the basic block that uses them. This reduces the number of cross-block live ranges, and allows the bb-at-a-time selector to always coallesce these away, at isel time. This reduces the load on the coallescer and register allocator. For example on a codec on X86, we went from: 1643 asm-printer - Number of machine instrs printed 419 liveintervals - Number of loads/stores folded into instructions 1144 liveintervals - Number of identity moves eliminated after coalescing 1022 liveintervals - Number of interval joins performed 282 liveintervals - Number of intervals after coalescing 1304 liveintervals - Number of original intervals 86 regalloc - Number of times we had to backtrack 1.90232 regalloc - Ratio of intervals processed over total intervals 40 spiller - Number of values reused 182 spiller - Number of loads added 121 spiller - Number of stores added 132 spiller - Number of register spills 6 twoaddressinstruction - Number of instructions commuted to coalesce 360 twoaddressinstruction - Number of two-address instructions to: 1636 asm-printer - Number of machine instrs printed 403 liveintervals - Number of loads/stores folded into instructions 1155 liveintervals - Number of identity moves eliminated after coalescing 1033 liveintervals - Number of interval joins performed 279 liveintervals - Number of intervals after coalescing 1312 liveintervals - Number of original intervals 76 regalloc - Number of times we had to backtrack 1.88998 regalloc - Ratio of intervals processed over total intervals 1 spiller - Number of copies elided 41 spiller - Number of values reused 191 spiller - Number of loads added 114 spiller - Number of stores added 128 spiller - Number of register spills 4 twoaddressinstruction - Number of instructions commuted to coalesce 356 twoaddressinstruction - Number of two-address instructions On this testcase, this change provides a modest reduction in spill code, regalloc iterations, and total instructions emitted. It increases the number of register coallesces. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28115 91177308-0d34-0410-b5e6-96231b3b80d8
2006-05-05 01:04:50 +00:00
for (BasicBlock::iterator E = BB->end(); BBI != E; ) {
Instruction *I = BBI++;
if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(I)) {
Fix the #1 code quality problem that I have seen on X86 (and it also affects PPC and other targets). In a particular, consider code like this: struct Vector3 { double x, y, z; }; struct Matrix3 { Vector3 a, b, c; }; double dot(Vector3 &a, Vector3 &b) { return a.x * b.x + a.y * b.y + a.z * b.z; } Vector3 mul(Vector3 &a, Matrix3 &b) { Vector3 r; r.x = dot( a, b.a ); r.y = dot( a, b.b ); r.z = dot( a, b.c ); return r; } void transform(Matrix3 &m, Vector3 *x, int n) { for (int i = 0; i < n; i++) x[i] = mul( x[i], m ); } we compile transform to a loop with all of the GEP instructions for indexing into 'm' pulled out of the loop (9 of them). Because isel occurs a bb at a time we are unable to fold the constant index into the loads in the loop, leading to PPC code that looks like this: LBB3_1: ; no_exit.preheader li r2, 0 addi r6, r3, 64 ;; 9 values live across the loop body! addi r7, r3, 56 addi r8, r3, 48 addi r9, r3, 40 addi r10, r3, 32 addi r11, r3, 24 addi r12, r3, 16 addi r30, r3, 8 LBB3_2: ; no_exit lfd f0, 0(r30) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) ;; no constant indices folded into the loads! lfd f3, 0(r4) lfd f4, 0(r10) lfd f5, 0(r6) lfd f6, 0(r7) lfd f7, 0(r8) lfd f8, 0(r9) lfd f9, 0(r11) lfd f10, 0(r12) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r29, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r29, r29 bne cr0, LBB3_2 ; no_exit uh, yuck. With this patch, we now sink the constant offsets into the loop, producing this code: LBB3_1: ; no_exit.preheader li r2, 0 LBB3_2: ; no_exit lfd f0, 8(r3) lfd f1, 8(r4) fmul f0, f1, f0 lfd f2, 0(r3) lfd f3, 0(r4) lfd f4, 32(r3) ;; much nicer. lfd f5, 64(r3) lfd f6, 56(r3) lfd f7, 48(r3) lfd f8, 40(r3) lfd f9, 24(r3) lfd f10, 16(r3) lfd f11, 16(r4) fmadd f0, f3, f2, f0 fmul f2, f1, f4 fmadd f0, f11, f10, f0 fmadd f2, f3, f9, f2 fmul f1, f1, f6 stfd f0, 0(r4) fmadd f0, f11, f8, f2 fmadd f1, f3, f7, f1 stfd f0, 8(r4) fmadd f0, f11, f5, f1 addi r6, r4, 24 stfd f0, 16(r4) addi r2, r2, 1 cmpw cr0, r2, r5 or r4, r6, r6 bne cr0, LBB3_2 ; no_exit This is much nicer as it reduces register pressure in the loop a lot. On X86, this takes the function from having 9 spilled registers to 2. This should help some spec programs on X86 (gzip?) This is currently only enabled with -enable-gep-isel-opt to allow perf testing tonight. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@24606 91177308-0d34-0410-b5e6-96231b3b80d8
2005-12-05 07:10:48 +00:00
OptimizeGEPExpression(GEPI, TLI.getTargetData());
Sink noop copies into the basic block that uses them. This reduces the number of cross-block live ranges, and allows the bb-at-a-time selector to always coallesce these away, at isel time. This reduces the load on the coallescer and register allocator. For example on a codec on X86, we went from: 1643 asm-printer - Number of machine instrs printed 419 liveintervals - Number of loads/stores folded into instructions 1144 liveintervals - Number of identity moves eliminated after coalescing 1022 liveintervals - Number of interval joins performed 282 liveintervals - Number of intervals after coalescing 1304 liveintervals - Number of original intervals 86 regalloc - Number of times we had to backtrack 1.90232 regalloc - Ratio of intervals processed over total intervals 40 spiller - Number of values reused 182 spiller - Number of loads added 121 spiller - Number of stores added 132 spiller - Number of register spills 6 twoaddressinstruction - Number of instructions commuted to coalesce 360 twoaddressinstruction - Number of two-address instructions to: 1636 asm-printer - Number of machine instrs printed 403 liveintervals - Number of loads/stores folded into instructions 1155 liveintervals - Number of identity moves eliminated after coalescing 1033 liveintervals - Number of interval joins performed 279 liveintervals - Number of intervals after coalescing 1312 liveintervals - Number of original intervals 76 regalloc - Number of times we had to backtrack 1.88998 regalloc - Ratio of intervals processed over total intervals 1 spiller - Number of copies elided 41 spiller - Number of values reused 191 spiller - Number of loads added 114 spiller - Number of stores added 128 spiller - Number of register spills 4 twoaddressinstruction - Number of instructions commuted to coalesce 356 twoaddressinstruction - Number of two-address instructions On this testcase, this change provides a modest reduction in spill code, regalloc iterations, and total instructions emitted. It increases the number of register coallesces. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@28115 91177308-0d34-0410-b5e6-96231b3b80d8
2006-05-05 01:04:50 +00:00
} else if (CastInst *CI = dyn_cast<CastInst>(I)) {
// If this is a noop copy, sink it into user blocks to reduce the number
// of virtual registers that must be created and coallesced.
MVT::ValueType SrcVT = TLI.getValueType(CI->getOperand(0)->getType());
MVT::ValueType DstVT = TLI.getValueType(CI->getType());
// This is an fp<->int conversion?
if (MVT::isInteger(SrcVT) != MVT::isInteger(DstVT))
continue;
// If this is an extension, it will be a zero or sign extension, which
// isn't a noop.
if (SrcVT < DstVT) continue;
// If these values will be promoted, find out what they will be promoted
// to. This helps us consider truncates on PPC as noop copies when they
// are.
if (TLI.getTypeAction(SrcVT) == TargetLowering::Promote)
SrcVT = TLI.getTypeToTransformTo(SrcVT);
if (TLI.getTypeAction(DstVT) == TargetLowering::Promote)
DstVT = TLI.getTypeToTransformTo(DstVT);
// If, after promotion, these are the same types, this is a noop copy.
if (SrcVT == DstVT)
OptimizeNoopCopyExpression(CI);
}
}
}
FunctionLoweringInfo FuncInfo(TLI, Fn, MF);
for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
SelectBasicBlock(I, MF, FuncInfo);
return true;
}
SDOperand SelectionDAGISel::
CopyValueToVirtualRegister(SelectionDAGLowering &SDL, Value *V, unsigned Reg) {
SDOperand Op = SDL.getValue(V);
assert((Op.getOpcode() != ISD::CopyFromReg ||
cast<RegisterSDNode>(Op.getOperand(1))->getReg() != Reg) &&
"Copy from a reg to the same reg!");
// If this type is not legal, we must make sure to not create an invalid
// register use.
MVT::ValueType SrcVT = Op.getValueType();
MVT::ValueType DestVT = TLI.getTypeToTransformTo(SrcVT);
SelectionDAG &DAG = SDL.DAG;
if (SrcVT == DestVT) {
return DAG.getCopyToReg(SDL.getRoot(), Reg, Op);
} else if (SrcVT == MVT::Vector) {
// Handle copies from generic vectors to registers.
MVT::ValueType PTyElementVT, PTyLegalElementVT;
unsigned NE = TLI.getPackedTypeBreakdown(cast<PackedType>(V->getType()),
PTyElementVT, PTyLegalElementVT);
// Insert a VBIT_CONVERT of the input vector to a "N x PTyElementVT"
// MVT::Vector type.
Op = DAG.getNode(ISD::VBIT_CONVERT, MVT::Vector, Op,
DAG.getConstant(NE, MVT::i32),
DAG.getValueType(PTyElementVT));
// Loop over all of the elements of the resultant vector,
// VEXTRACT_VECTOR_ELT'ing them, converting them to PTyLegalElementVT, then
// copying them into output registers.
std::vector<SDOperand> OutChains;
SDOperand Root = SDL.getRoot();
for (unsigned i = 0; i != NE; ++i) {
SDOperand Elt = DAG.getNode(ISD::VEXTRACT_VECTOR_ELT, PTyElementVT,
Op, DAG.getConstant(i, MVT::i32));
if (PTyElementVT == PTyLegalElementVT) {
// Elements are legal.
OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
} else if (PTyLegalElementVT > PTyElementVT) {
// Elements are promoted.
if (MVT::isFloatingPoint(PTyLegalElementVT))
Elt = DAG.getNode(ISD::FP_EXTEND, PTyLegalElementVT, Elt);
else
Elt = DAG.getNode(ISD::ANY_EXTEND, PTyLegalElementVT, Elt);
OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Elt));
} else {
// Elements are expanded.
// The src value is expanded into multiple registers.
SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
Elt, DAG.getConstant(0, MVT::i32));
SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, PTyLegalElementVT,
Elt, DAG.getConstant(1, MVT::i32));
OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Lo));
OutChains.push_back(DAG.getCopyToReg(Root, Reg++, Hi));
}
}
return DAG.getNode(ISD::TokenFactor, MVT::Other, OutChains);
} else if (SrcVT < DestVT) {
// The src value is promoted to the register.
if (MVT::isFloatingPoint(SrcVT))
Op = DAG.getNode(ISD::FP_EXTEND, DestVT, Op);
else
Op = DAG.getNode(ISD::ANY_EXTEND, DestVT, Op);
return DAG.getCopyToReg(SDL.getRoot(), Reg, Op);
} else {
// The src value is expanded into multiple registers.
SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
Op, DAG.getConstant(0, MVT::i32));
SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, DestVT,
Op, DAG.getConstant(1, MVT::i32));
Op = DAG.getCopyToReg(SDL.getRoot(), Reg, Lo);
return DAG.getCopyToReg(Op, Reg+1, Hi);
}
}
void SelectionDAGISel::
LowerArguments(BasicBlock *BB, SelectionDAGLowering &SDL,
std::vector<SDOperand> &UnorderedChains) {
// If this is the entry block, emit arguments.
Function &F = *BB->getParent();
FunctionLoweringInfo &FuncInfo = SDL.FuncInfo;
SDOperand OldRoot = SDL.DAG.getRoot();
std::vector<SDOperand> Args = TLI.LowerArguments(F, SDL.DAG);
unsigned a = 0;
for (Function::arg_iterator AI = F.arg_begin(), E = F.arg_end();
AI != E; ++AI, ++a)
if (!AI->use_empty()) {
SDL.setValue(AI, Args[a]);
// If this argument is live outside of the entry block, insert a copy from
// whereever we got it to the vreg that other BB's will reference it as.
if (FuncInfo.ValueMap.count(AI)) {
SDOperand Copy =
CopyValueToVirtualRegister(SDL, AI, FuncInfo.ValueMap[AI]);
UnorderedChains.push_back(Copy);
}
}
// Next, if the function has live ins that need to be copied into vregs,
// emit the copies now, into the top of the block.
MachineFunction &MF = SDL.DAG.getMachineFunction();
if (MF.livein_begin() != MF.livein_end()) {
SSARegMap *RegMap = MF.getSSARegMap();
const MRegisterInfo &MRI = *MF.getTarget().getRegisterInfo();
for (MachineFunction::livein_iterator LI = MF.livein_begin(),
E = MF.livein_end(); LI != E; ++LI)
if (LI->second)
MRI.copyRegToReg(*MF.begin(), MF.begin()->end(), LI->second,
LI->first, RegMap->getRegClass(LI->second));
}
// Finally, if the target has anything special to do, allow it to do so.
EmitFunctionEntryCode(F, SDL.DAG.getMachineFunction());
}
void SelectionDAGISel::BuildSelectionDAG(SelectionDAG &DAG, BasicBlock *LLVMBB,
std::vector<std::pair<MachineInstr*, unsigned> > &PHINodesToUpdate,
FunctionLoweringInfo &FuncInfo) {
SelectionDAGLowering SDL(DAG, TLI, FuncInfo);
std::vector<SDOperand> UnorderedChains;
// Lower any arguments needed in this block if this is the entry block.
if (LLVMBB == &LLVMBB->getParent()->front())
LowerArguments(LLVMBB, SDL, UnorderedChains);
BB = FuncInfo.MBBMap[LLVMBB];
SDL.setCurrentBasicBlock(BB);
// Lower all of the non-terminator instructions.
for (BasicBlock::iterator I = LLVMBB->begin(), E = --LLVMBB->end();
I != E; ++I)
SDL.visit(*I);
// Ensure that all instructions which are used outside of their defining
// blocks are available as virtual registers.
for (BasicBlock::iterator I = LLVMBB->begin(), E = LLVMBB->end(); I != E;++I)
if (!I->use_empty() && !isa<PHINode>(I)) {
std::map<const Value*, unsigned>::iterator VMI =FuncInfo.ValueMap.find(I);
if (VMI != FuncInfo.ValueMap.end())
UnorderedChains.push_back(
CopyValueToVirtualRegister(SDL, I, VMI->second));
}
// Handle PHI nodes in successor blocks. Emit code into the SelectionDAG to
// ensure constants are generated when needed. Remember the virtual registers
// that need to be added to the Machine PHI nodes as input. We cannot just
// directly add them, because expansion might result in multiple MBB's for one
// BB. As such, the start of the BB might correspond to a different MBB than
// the end.
//
// Emit constants only once even if used by multiple PHI nodes.
std::map<Constant*, unsigned> ConstantsOut;
// Check successor nodes PHI nodes that expect a constant to be available from
// this block.
TerminatorInst *TI = LLVMBB->getTerminator();
for (unsigned succ = 0, e = TI->getNumSuccessors(); succ != e; ++succ) {
BasicBlock *SuccBB = TI->getSuccessor(succ);
MachineBasicBlock::iterator MBBI = FuncInfo.MBBMap[SuccBB]->begin();
PHINode *PN;
// At this point we know that there is a 1-1 correspondence between LLVM PHI
// nodes and Machine PHI nodes, but the incoming operands have not been
// emitted yet.
for (BasicBlock::iterator I = SuccBB->begin();
(PN = dyn_cast<PHINode>(I)); ++I)
if (!PN->use_empty()) {
unsigned Reg;
Value *PHIOp = PN->getIncomingValueForBlock(LLVMBB);
if (Constant *C = dyn_cast<Constant>(PHIOp)) {
unsigned &RegOut = ConstantsOut[C];
if (RegOut == 0) {
RegOut = FuncInfo.CreateRegForValue(C);
UnorderedChains.push_back(
CopyValueToVirtualRegister(SDL, C, RegOut));
}
Reg = RegOut;
} else {
Reg = FuncInfo.ValueMap[PHIOp];
if (Reg == 0) {
assert(isa<AllocaInst>(PHIOp) &&
FuncInfo.StaticAllocaMap.count(cast<AllocaInst>(PHIOp)) &&
"Didn't codegen value into a register!??");
Reg = FuncInfo.CreateRegForValue(PHIOp);
UnorderedChains.push_back(
CopyValueToVirtualRegister(SDL, PHIOp, Reg));
}
}
// Remember that this register needs to added to the machine PHI node as
// the input for this MBB.
MVT::ValueType VT = TLI.getValueType(PN->getType());
unsigned NumElements;
if (VT != MVT::Vector)
NumElements = TLI.getNumElements(VT);
else {
MVT::ValueType VT1,VT2;
NumElements =
TLI.getPackedTypeBreakdown(cast<PackedType>(PN->getType()),
VT1, VT2);
}
for (unsigned i = 0, e = NumElements; i != e; ++i)
PHINodesToUpdate.push_back(std::make_pair(MBBI++, Reg+i));
}
}
ConstantsOut.clear();
// Turn all of the unordered chains into one factored node.
if (!UnorderedChains.empty()) {
SDOperand Root = SDL.getRoot();
if (Root.getOpcode() != ISD::EntryToken) {
unsigned i = 0, e = UnorderedChains.size();
for (; i != e; ++i) {
assert(UnorderedChains[i].Val->getNumOperands() > 1);
if (UnorderedChains[i].Val->getOperand(0) == Root)
break; // Don't add the root if we already indirectly depend on it.
}
if (i == e)
UnorderedChains.push_back(Root);
}
DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, UnorderedChains));
}
// Lower the terminator after the copies are emitted.
SDL.visit(*LLVMBB->getTerminator());
// Copy over any CaseBlock records that may now exist due to SwitchInst
// lowering, as well as any jump table information.
SwitchCases.clear();
SwitchCases = SDL.SwitchCases;
JT = SDL.JT;
// Make sure the root of the DAG is up-to-date.
DAG.setRoot(SDL.getRoot());
}
void SelectionDAGISel::CodeGenAndEmitDAG(SelectionDAG &DAG) {
// Run the DAG combiner in pre-legalize mode.
DAG.Combine(false);
DEBUG(std::cerr << "Lowered selection DAG:\n");
DEBUG(DAG.dump());
// Second step, hack on the DAG until it only uses operations and types that
// the target supports.
DAG.Legalize();
DEBUG(std::cerr << "Legalized selection DAG:\n");
DEBUG(DAG.dump());
// Run the DAG combiner in post-legalize mode.
DAG.Combine(true);
if (ViewISelDAGs) DAG.viewGraph();
// Third, instruction select all of the operations to machine code, adding the
// code to the MachineBasicBlock.
InstructionSelectBasicBlock(DAG);
DEBUG(std::cerr << "Selected machine code:\n");
DEBUG(BB->dump());
}
void SelectionDAGISel::SelectBasicBlock(BasicBlock *LLVMBB, MachineFunction &MF,
FunctionLoweringInfo &FuncInfo) {
std::vector<std::pair<MachineInstr*, unsigned> > PHINodesToUpdate;
{
SelectionDAG DAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
CurDAG = &DAG;
// First step, lower LLVM code to some DAG. This DAG may use operations and
// types that are not supported by the target.
BuildSelectionDAG(DAG, LLVMBB, PHINodesToUpdate, FuncInfo);
// Second step, emit the lowered DAG as machine code.
CodeGenAndEmitDAG(DAG);
}
// Next, now that we know what the last MBB the LLVM BB expanded is, update
// PHI nodes in successors.
if (SwitchCases.empty() && JT.Reg == 0) {
for (unsigned i = 0, e = PHINodesToUpdate.size(); i != e; ++i) {
MachineInstr *PHI = PHINodesToUpdate[i].first;
assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
"This is not a machine PHI node that we are updating!");
PHI->addRegOperand(PHINodesToUpdate[i].second);
PHI->addMachineBasicBlockOperand(BB);
}
return;
}
// If the JumpTable record is filled in, then we need to emit a jump table.
// Updating the PHI nodes is tricky in this case, since we need to determine
// whether the PHI is a successor of the range check MBB or the jump table MBB
if (JT.Reg) {
assert(SwitchCases.empty() && "Cannot have jump table and lowered switch");
SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
CurDAG = &SDAG;
SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
MachineBasicBlock *RangeBB = BB;
// Set the current basic block to the mbb we wish to insert the code into
BB = JT.MBB;
SDL.setCurrentBasicBlock(BB);
// Emit the code
SDL.visitJumpTable(JT);
SDAG.setRoot(SDL.getRoot());
CodeGenAndEmitDAG(SDAG);
// Update PHI Nodes
for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
MachineInstr *PHI = PHINodesToUpdate[pi].first;
MachineBasicBlock *PHIBB = PHI->getParent();
assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
"This is not a machine PHI node that we are updating!");
if (PHIBB == JT.Default) {
PHI->addRegOperand(PHINodesToUpdate[pi].second);
PHI->addMachineBasicBlockOperand(RangeBB);
}
if (BB->succ_end() != std::find(BB->succ_begin(),BB->succ_end(), PHIBB)) {
PHI->addRegOperand(PHINodesToUpdate[pi].second);
PHI->addMachineBasicBlockOperand(BB);
}
}
return;
}
// If we generated any switch lowering information, build and codegen any
// additional DAGs necessary.
for(unsigned i = 0, e = SwitchCases.size(); i != e; ++i) {
SelectionDAG SDAG(TLI, MF, getAnalysisToUpdate<MachineDebugInfo>());
CurDAG = &SDAG;
SelectionDAGLowering SDL(SDAG, TLI, FuncInfo);
// Set the current basic block to the mbb we wish to insert the code into
BB = SwitchCases[i].ThisBB;
SDL.setCurrentBasicBlock(BB);
// Emit the code
SDL.visitSwitchCase(SwitchCases[i]);
SDAG.setRoot(SDL.getRoot());
CodeGenAndEmitDAG(SDAG);
// Iterate over the phi nodes, if there is a phi node in a successor of this
// block (for instance, the default block), then add a pair of operands to
// the phi node for this block, as if we were coming from the original
// BB before switch expansion.
for (unsigned pi = 0, pe = PHINodesToUpdate.size(); pi != pe; ++pi) {
MachineInstr *PHI = PHINodesToUpdate[pi].first;
MachineBasicBlock *PHIBB = PHI->getParent();
assert(PHI->getOpcode() == TargetInstrInfo::PHI &&
"This is not a machine PHI node that we are updating!");
if (PHIBB == SwitchCases[i].LHSBB || PHIBB == SwitchCases[i].RHSBB) {
PHI->addRegOperand(PHINodesToUpdate[pi].second);
PHI->addMachineBasicBlockOperand(BB);
}
}
}
}
//===----------------------------------------------------------------------===//
/// ScheduleAndEmitDAG - Pick a safe ordering and emit instructions for each
/// target node in the graph.
void SelectionDAGISel::ScheduleAndEmitDAG(SelectionDAG &DAG) {
if (ViewSchedDAGs) DAG.viewGraph();
ScheduleDAG *SL = NULL;
switch (ISHeuristic) {
default: assert(0 && "Unrecognized scheduling heuristic");
case defaultScheduling:
if (TLI.getSchedulingPreference() == TargetLowering::SchedulingForLatency)
SL = createTDListDAGScheduler(DAG, BB, CreateTargetHazardRecognizer());
else {
assert(TLI.getSchedulingPreference() ==
TargetLowering::SchedulingForRegPressure && "Unknown sched type!");
SL = createBURRListDAGScheduler(DAG, BB);
}
break;
case noScheduling:
SL = createBFS_DAGScheduler(DAG, BB);
break;
case simpleScheduling:
SL = createSimpleDAGScheduler(false, DAG, BB);
break;
case simpleNoItinScheduling:
SL = createSimpleDAGScheduler(true, DAG, BB);
break;
case listSchedulingBURR:
SL = createBURRListDAGScheduler(DAG, BB);
break;
case listSchedulingTD:
SL = createTDListDAGScheduler(DAG, BB, CreateTargetHazardRecognizer());
break;
}
BB = SL->Run();
delete SL;
}
HazardRecognizer *SelectionDAGISel::CreateTargetHazardRecognizer() {
return new HazardRecognizer();
}
/// SelectInlineAsmMemoryOperands - Calls to this are automatically generated
/// by tblgen. Others should not call it.
void SelectionDAGISel::
SelectInlineAsmMemoryOperands(std::vector<SDOperand> &Ops, SelectionDAG &DAG) {
std::vector<SDOperand> InOps;
std::swap(InOps, Ops);
Ops.push_back(InOps[0]); // input chain.
Ops.push_back(InOps[1]); // input asm string.
const char *AsmStr = cast<ExternalSymbolSDNode>(InOps[1])->getSymbol();
unsigned i = 2, e = InOps.size();
if (InOps[e-1].getValueType() == MVT::Flag)
--e; // Don't process a flag operand if it is here.
while (i != e) {
unsigned Flags = cast<ConstantSDNode>(InOps[i])->getValue();
if ((Flags & 7) != 4 /*MEM*/) {
// Just skip over this operand, copying the operands verbatim.
Ops.insert(Ops.end(), InOps.begin()+i, InOps.begin()+i+(Flags >> 3) + 1);
i += (Flags >> 3) + 1;
} else {
assert((Flags >> 3) == 1 && "Memory operand with multiple values?");
// Otherwise, this is a memory operand. Ask the target to select it.
std::vector<SDOperand> SelOps;
if (SelectInlineAsmMemoryOperand(InOps[i+1], 'm', SelOps, DAG)) {
std::cerr << "Could not match memory address. Inline asm failure!\n";
exit(1);
}
// Add this to the output node.
Ops.push_back(DAG.getConstant(4/*MEM*/ | (SelOps.size() << 3), MVT::i32));
Ops.insert(Ops.end(), SelOps.begin(), SelOps.end());
i += 2;
}
}
// Add the flag input back if present.
if (e != InOps.size())
Ops.push_back(InOps.back());
}