Google Git
Sign in
llvm / llvm / refs/heads/release_31 / . / lib / CodeGen / SelectionDAG / SelectionDAGBuilder.cpp
blob: f1e879be9567ee44aca9ad119f1a8ddfdcec1751 [file] [log] [blame]
//===-- SelectionDAGBuilder.cpp - Selection-DAG building ------------------===//
//
// The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This implements routines for translating from LLVM IR into SelectionDAG IR.
//
//===----------------------------------------------------------------------===//
#define DEBUG_TYPE "isel"
#include "SDNodeDbgValue.h"
#include "SelectionDAGBuilder.h"
#include "llvm/ADT/BitVector.h"
#include "llvm/ADT/PostOrderIterator.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/Analysis/AliasAnalysis.h"
#include "llvm/Analysis/ConstantFolding.h"
#include "llvm/Constants.h"
#include "llvm/CallingConv.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/LLVMContext.h"
#include "llvm/Module.h"
#include "llvm/CodeGen/Analysis.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/FunctionLoweringInfo.h"
#include "llvm/CodeGen/GCStrategy.h"
#include "llvm/CodeGen/GCMetadata.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Analysis/DebugInfo.h"
#include "llvm/Target/TargetData.h"
#include "llvm/Target/TargetFrameLowering.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetIntrinsicInfo.h"
#include "llvm/Target/TargetLibraryInfo.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include <algorithm>
using namespace llvm;
/// LimitFloatPrecision - Generate low-precision inline sequences for
/// some float libcalls (6, 8 or 12 bits).
static unsigned LimitFloatPrecision;
static cl::opt<unsigned, true>
LimitFPPrecision("limit-float-precision",
cl::desc("Generate low-precision inline sequences "
"for some float libcalls"),
cl::location(LimitFloatPrecision),
cl::init(0));
// Limit the width of DAG chains. This is important in general to prevent
// prevent DAG-based analysis from blowing up. For example, alias analysis and
// load clustering may not complete in reasonable time. It is difficult to
// recognize and avoid this situation within each individual analysis, and
// future analyses are likely to have the same behavior. Limiting DAG width is
// the safe approach, and will be especially important with global DAGs.
//
// MaxParallelChains default is arbitrarily high to avoid affecting
// optimization, but could be lowered to improve compile time. Any ld-ld-st-st
// sequence over this should have been converted to llvm.memcpy by the
// frontend. It easy to induce this behavior with .ll code such as:
// %buffer = alloca [4096 x i8]
// %data = load [4096 x i8]* %argPtr
// store [4096 x i8] %data, [4096 x i8]* %buffer
static const unsigned MaxParallelChains = 64;
static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
const SDValue *Parts, unsigned NumParts,
EVT PartVT, EVT ValueVT);
/// getCopyFromParts - Create a value that contains the specified legal parts
/// combined into the value they represent. If the parts combine to a type
/// larger then ValueVT then AssertOp can be used to specify whether the extra
/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
/// (ISD::AssertSext).
static SDValue getCopyFromParts(SelectionDAG &DAG, DebugLoc DL,
const SDValue *Parts,
unsigned NumParts, EVT PartVT, EVT ValueVT,
ISD::NodeType AssertOp = ISD::DELETED_NODE) {
if (ValueVT.isVector())
return getCopyFromPartsVector(DAG, DL, Parts, NumParts, PartVT, ValueVT);
assert(NumParts > 0 && "No parts to assemble!");
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
SDValue Val = Parts[0];
if (NumParts > 1) {
// Assemble the value from multiple parts.
if (ValueVT.isInteger()) {
unsigned PartBits = PartVT.getSizeInBits();
unsigned ValueBits = ValueVT.getSizeInBits();
// Assemble the power of 2 part.
unsigned RoundParts = NumParts & (NumParts - 1) ?
1 << Log2_32(NumParts) : NumParts;
unsigned RoundBits = PartBits * RoundParts;
EVT RoundVT = RoundBits == ValueBits ?
ValueVT : EVT::getIntegerVT(*DAG.getContext(), RoundBits);
SDValue Lo, Hi;
EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
if (RoundParts > 2) {
Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
PartVT, HalfVT);
Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
RoundParts / 2, PartVT, HalfVT);
} else {
Lo = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[0]);
Hi = DAG.getNode(ISD::BITCAST, DL, HalfVT, Parts[1]);
}
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Val = DAG.getNode(ISD::BUILD_PAIR, DL, RoundVT, Lo, Hi);
if (RoundParts < NumParts) {
// Assemble the trailing non-power-of-2 part.
unsigned OddParts = NumParts - RoundParts;
EVT OddVT = EVT::getIntegerVT(*DAG.getContext(), OddParts * PartBits);
Hi = getCopyFromParts(DAG, DL,
Parts + RoundParts, OddParts, PartVT, OddVT);
// Combine the round and odd parts.
Lo = Val;
if (TLI.isBigEndian())
std::swap(Lo, Hi);
EVT TotalVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
Hi = DAG.getNode(ISD::ANY_EXTEND, DL, TotalVT, Hi);
Hi = DAG.getNode(ISD::SHL, DL, TotalVT, Hi,
DAG.getConstant(Lo.getValueType().getSizeInBits(),
TLI.getPointerTy()));
Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
}
} else if (PartVT.isFloatingPoint()) {
// FP split into multiple FP parts (for ppcf128)
assert(ValueVT == EVT(MVT::ppcf128) && PartVT == EVT(MVT::f64) &&
"Unexpected split");
SDValue Lo, Hi;
Lo = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[0]);
Hi = DAG.getNode(ISD::BITCAST, DL, EVT(MVT::f64), Parts[1]);
if (TLI.isBigEndian())
std::swap(Lo, Hi);
Val = DAG.getNode(ISD::BUILD_PAIR, DL, ValueVT, Lo, Hi);
} else {
// FP split into integer parts (soft fp)
assert(ValueVT.isFloatingPoint() && PartVT.isInteger() &&
!PartVT.isVector() && "Unexpected split");
EVT IntVT = EVT::getIntegerVT(*DAG.getContext(), ValueVT.getSizeInBits());
Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
}
}
// There is now one part, held in Val. Correct it to match ValueVT.
PartVT = Val.getValueType();
if (PartVT == ValueVT)
return Val;
if (PartVT.isInteger() && ValueVT.isInteger()) {
if (ValueVT.bitsLT(PartVT)) {
// For a truncate, see if we have any information to
// indicate whether the truncated bits will always be
// zero or sign-extension.
if (AssertOp != ISD::DELETED_NODE)
Val = DAG.getNode(AssertOp, DL, PartVT, Val,
DAG.getValueType(ValueVT));
return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
}
return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
}
if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
// FP_ROUND's are always exact here.
if (ValueVT.bitsLT(Val.getValueType()))
return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
DAG.getTargetConstant(1, TLI.getPointerTy()));
return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
}
if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
llvm_unreachable("Unknown mismatch!");
}
/// getCopyFromParts - Create a value that contains the specified legal parts
/// combined into the value they represent. If the parts combine to a type
/// larger then ValueVT then AssertOp can be used to specify whether the extra
/// bits are known to be zero (ISD::AssertZext) or sign extended from ValueVT
/// (ISD::AssertSext).
static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
const SDValue *Parts, unsigned NumParts,
EVT PartVT, EVT ValueVT) {
assert(ValueVT.isVector() && "Not a vector value");
assert(NumParts > 0 && "No parts to assemble!");
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
SDValue Val = Parts[0];
// Handle a multi-element vector.
if (NumParts > 1) {
EVT IntermediateVT, RegisterVT;
unsigned NumIntermediates;
unsigned NumRegs =
TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT, IntermediateVT,
NumIntermediates, RegisterVT);
assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
NumParts = NumRegs; // Silence a compiler warning.
assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
assert(RegisterVT == Parts[0].getValueType() &&
"Part type doesn't match part!");
// Assemble the parts into intermediate operands.
SmallVector<SDValue, 8> Ops(NumIntermediates);
if (NumIntermediates == NumParts) {
// If the register was not expanded, truncate or copy the value,
// as appropriate.
for (unsigned i = 0; i != NumParts; ++i)
Ops[i] = getCopyFromParts(DAG, DL, &Parts[i], 1,
PartVT, IntermediateVT);
} else if (NumParts > 0) {
// If the intermediate type was expanded, build the intermediate
// operands from the parts.
assert(NumParts % NumIntermediates == 0 &&
"Must expand into a divisible number of parts!");
unsigned Factor = NumParts / NumIntermediates;
for (unsigned i = 0; i != NumIntermediates; ++i)
Ops[i] = getCopyFromParts(DAG, DL, &Parts[i * Factor], Factor,
PartVT, IntermediateVT);
}
// Build a vector with BUILD_VECTOR or CONCAT_VECTORS from the
// intermediate operands.
Val = DAG.getNode(IntermediateVT.isVector() ?
ISD::CONCAT_VECTORS : ISD::BUILD_VECTOR, DL,
ValueVT, &Ops[0], NumIntermediates);
}
// There is now one part, held in Val. Correct it to match ValueVT.
PartVT = Val.getValueType();
if (PartVT == ValueVT)
return Val;
if (PartVT.isVector()) {
// If the element type of the source/dest vectors are the same, but the
// parts vector has more elements than the value vector, then we have a
// vector widening case (e.g. <2 x float> -> <4 x float>). Extract the
// elements we want.
if (PartVT.getVectorElementType() == ValueVT.getVectorElementType()) {
assert(PartVT.getVectorNumElements() > ValueVT.getVectorNumElements() &&
"Cannot narrow, it would be a lossy transformation");
return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, ValueVT, Val,
DAG.getIntPtrConstant(0));
}
// Vector/Vector bitcast.
if (ValueVT.getSizeInBits() == PartVT.getSizeInBits())
return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
assert(PartVT.getVectorNumElements() == ValueVT.getVectorNumElements() &&
"Cannot handle this kind of promotion");
// Promoted vector extract
bool Smaller = ValueVT.bitsLE(PartVT);
return DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
DL, ValueVT, Val);
}
// Trivial bitcast if the types are the same size and the destination
// vector type is legal.
if (PartVT.getSizeInBits() == ValueVT.getSizeInBits() &&
TLI.isTypeLegal(ValueVT))
return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
// Handle cases such as i8 -> <1 x i1>
assert(ValueVT.getVectorNumElements() == 1 &&
"Only trivial scalar-to-vector conversions should get here!");
if (ValueVT.getVectorNumElements() == 1 &&
ValueVT.getVectorElementType() != PartVT) {
bool Smaller = ValueVT.bitsLE(PartVT);
Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
DL, ValueVT.getScalarType(), Val);
}
return DAG.getNode(ISD::BUILD_VECTOR, DL, ValueVT, Val);
}
static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc dl,
SDValue Val, SDValue *Parts, unsigned NumParts,
EVT PartVT);
/// getCopyToParts - Create a series of nodes that contain the specified value
/// split into legal parts. If the parts contain more bits than Val, then, for
/// integers, ExtendKind can be used to specify how to generate the extra bits.
static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
SDValue Val, SDValue *Parts, unsigned NumParts,
EVT PartVT,
ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
EVT ValueVT = Val.getValueType();
// Handle the vector case separately.
if (ValueVT.isVector())
return getCopyToPartsVector(DAG, DL, Val, Parts, NumParts, PartVT);
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
unsigned PartBits = PartVT.getSizeInBits();
unsigned OrigNumParts = NumParts;
assert(TLI.isTypeLegal(PartVT) && "Copying to an illegal type!");
if (NumParts == 0)
return;
assert(!ValueVT.isVector() && "Vector case handled elsewhere");
if (PartVT == ValueVT) {
assert(NumParts == 1 && "No-op copy with multiple parts!");
Parts[0] = Val;
return;
}
if (NumParts * PartBits > ValueVT.getSizeInBits()) {
// If the parts cover more bits than the value has, promote the value.
if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
assert(NumParts == 1 && "Do not know what to promote to!");
Val = DAG.getNode(ISD::FP_EXTEND, DL, PartVT, Val);
} else {
assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
ValueVT.isInteger() &&
"Unknown mismatch!");
ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
Val = DAG.getNode(ExtendKind, DL, ValueVT, Val);
if (PartVT == MVT::x86mmx)
Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
}
} else if (PartBits == ValueVT.getSizeInBits()) {
// Different types of the same size.
assert(NumParts == 1 && PartVT != ValueVT);
Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
} else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
// If the parts cover less bits than value has, truncate the value.
assert((PartVT.isInteger() || PartVT == MVT::x86mmx) &&
ValueVT.isInteger() &&
"Unknown mismatch!");
ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
if (PartVT == MVT::x86mmx)
Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
}
// The value may have changed - recompute ValueVT.
ValueVT = Val.getValueType();
assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
"Failed to tile the value with PartVT!");
if (NumParts == 1) {
assert(PartVT == ValueVT && "Type conversion failed!");
Parts[0] = Val;
return;
}
// Expand the value into multiple parts.
if (NumParts & (NumParts - 1)) {
// The number of parts is not a power of 2. Split off and copy the tail.
assert(PartVT.isInteger() && ValueVT.isInteger() &&
"Do not know what to expand to!");
unsigned RoundParts = 1 << Log2_32(NumParts);
unsigned RoundBits = RoundParts * PartBits;
unsigned OddParts = NumParts - RoundParts;
SDValue OddVal = DAG.getNode(ISD::SRL, DL, ValueVT, Val,
DAG.getIntPtrConstant(RoundBits));
getCopyToParts(DAG, DL, OddVal, Parts + RoundParts, OddParts, PartVT);
if (TLI.isBigEndian())
// The odd parts were reversed by getCopyToParts - unreverse them.
std::reverse(Parts + RoundParts, Parts + NumParts);
NumParts = RoundParts;
ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
}
// The number of parts is a power of 2. Repeatedly bisect the value using
// EXTRACT_ELEMENT.
Parts[0] = DAG.getNode(ISD::BITCAST, DL,
EVT::getIntegerVT(*DAG.getContext(),
ValueVT.getSizeInBits()),
Val);
for (unsigned StepSize = NumParts; StepSize > 1; StepSize /= 2) {
for (unsigned i = 0; i < NumParts; i += StepSize) {
unsigned ThisBits = StepSize * PartBits / 2;
EVT ThisVT = EVT::getIntegerVT(*DAG.getContext(), ThisBits);
SDValue &Part0 = Parts[i];
SDValue &Part1 = Parts[i+StepSize/2];
Part1 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
ThisVT, Part0, DAG.getIntPtrConstant(1));
Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, DL,
ThisVT, Part0, DAG.getIntPtrConstant(0));
if (ThisBits == PartBits && ThisVT != PartVT) {
Part0 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part0);
Part1 = DAG.getNode(ISD::BITCAST, DL, PartVT, Part1);
}
}
}
if (TLI.isBigEndian())
std::reverse(Parts, Parts + OrigNumParts);
}
/// getCopyToPartsVector - Create a series of nodes that contain the specified
/// value split into legal parts.
static void getCopyToPartsVector(SelectionDAG &DAG, DebugLoc DL,
SDValue Val, SDValue *Parts, unsigned NumParts,
EVT PartVT) {
EVT ValueVT = Val.getValueType();
assert(ValueVT.isVector() && "Not a vector");
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
if (NumParts == 1) {
if (PartVT == ValueVT) {
// Nothing to do.
} else if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
// Bitconvert vector->vector case.
Val = DAG.getNode(ISD::BITCAST, DL, PartVT, Val);
} else if (PartVT.isVector() &&
PartVT.getVectorElementType() == ValueVT.getVectorElementType() &&
PartVT.getVectorNumElements() > ValueVT.getVectorNumElements()) {
EVT ElementVT = PartVT.getVectorElementType();
// Vector widening case, e.g. <2 x float> -> <4 x float>. Shuffle in
// undef elements.
SmallVector<SDValue, 16> Ops;
for (unsigned i = 0, e = ValueVT.getVectorNumElements(); i != e; ++i)
Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
ElementVT, Val, DAG.getIntPtrConstant(i)));
for (unsigned i = ValueVT.getVectorNumElements(),
e = PartVT.getVectorNumElements(); i != e; ++i)
Ops.push_back(DAG.getUNDEF(ElementVT));
Val = DAG.getNode(ISD::BUILD_VECTOR, DL, PartVT, &Ops[0], Ops.size());
// FIXME: Use CONCAT for 2x -> 4x.
//SDValue UndefElts = DAG.getUNDEF(VectorTy);
//Val = DAG.getNode(ISD::CONCAT_VECTORS, DL, PartVT, Val, UndefElts);
} else if (PartVT.isVector() &&
PartVT.getVectorElementType().bitsGE(
ValueVT.getVectorElementType()) &&
PartVT.getVectorNumElements() == ValueVT.getVectorNumElements()) {
// Promoted vector extract
bool Smaller = PartVT.bitsLE(ValueVT);
Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
DL, PartVT, Val);
} else{
// Vector -> scalar conversion.
assert(ValueVT.getVectorNumElements() == 1 &&
"Only trivial vector-to-scalar conversions should get here!");
Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
PartVT, Val, DAG.getIntPtrConstant(0));
bool Smaller = ValueVT.bitsLE(PartVT);
Val = DAG.getNode((Smaller ? ISD::TRUNCATE : ISD::ANY_EXTEND),
DL, PartVT, Val);
}
Parts[0] = Val;
return;
}
// Handle a multi-element vector.
EVT IntermediateVT, RegisterVT;
unsigned NumIntermediates;
unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
IntermediateVT,
NumIntermediates, RegisterVT);
unsigned NumElements = ValueVT.getVectorNumElements();
assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
NumParts = NumRegs; // Silence a compiler warning.
assert(RegisterVT == PartVT && "Part type doesn't match vector breakdown!");
// Split the vector into intermediate operands.
SmallVector<SDValue, 8> Ops(NumIntermediates);
for (unsigned i = 0; i != NumIntermediates; ++i) {
if (IntermediateVT.isVector())
Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
IntermediateVT, Val,
DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
else
Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
IntermediateVT, Val, DAG.getIntPtrConstant(i));
}
// Split the intermediate operands into legal parts.
if (NumParts == NumIntermediates) {
// If the register was not expanded, promote or copy the value,
// as appropriate.
for (unsigned i = 0; i != NumParts; ++i)
getCopyToParts(DAG, DL, Ops[i], &Parts[i], 1, PartVT);
} else if (NumParts > 0) {
// If the intermediate type was expanded, split each the value into
// legal parts.
assert(NumParts % NumIntermediates == 0 &&
"Must expand into a divisible number of parts!");
unsigned Factor = NumParts / NumIntermediates;
for (unsigned i = 0; i != NumIntermediates; ++i)
getCopyToParts(DAG, DL, Ops[i], &Parts[i*Factor], Factor, PartVT);
}
}
namespace {
/// RegsForValue - This struct represents the registers (physical or virtual)
/// that a particular set of values is assigned, and the type information
/// about the value. The most common situation is to represent one value at a
/// time, but struct or array values are handled element-wise as multiple
/// values. The splitting of aggregates is performed recursively, so that we
/// never have aggregate-typed registers. The values at this point do not
/// necessarily have legal types, so each value may require one or more
/// registers of some legal type.
///
struct RegsForValue {
/// ValueVTs - The value types of the values, which may not be legal, and
/// may need be promoted or synthesized from one or more registers.
///
SmallVector<EVT, 4> ValueVTs;
/// RegVTs - The value types of the registers. This is the same size as
/// ValueVTs and it records, for each value, what the type of the assigned
/// register or registers are. (Individual values are never synthesized
/// from more than one type of register.)
///
/// With virtual registers, the contents of RegVTs is redundant with TLI's
/// getRegisterType member function, however when with physical registers
/// it is necessary to have a separate record of the types.
///
SmallVector<EVT, 4> RegVTs;
/// Regs - This list holds the registers assigned to the values.
/// Each legal or promoted value requires one register, and each
/// expanded value requires multiple registers.
///
SmallVector<unsigned, 4> Regs;
RegsForValue() {}
RegsForValue(const SmallVector<unsigned, 4> &regs,
EVT regvt, EVT valuevt)
: ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
RegsForValue(LLVMContext &Context, const TargetLowering &tli,
unsigned Reg, Type *Ty) {
ComputeValueVTs(tli, Ty, ValueVTs);
for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
EVT ValueVT = ValueVTs[Value];
unsigned NumRegs = tli.getNumRegisters(Context, ValueVT);
EVT RegisterVT = tli.getRegisterType(Context, ValueVT);
for (unsigned i = 0; i != NumRegs; ++i)
Regs.push_back(Reg + i);
RegVTs.push_back(RegisterVT);
Reg += NumRegs;
}
}
/// areValueTypesLegal - Return true if types of all the values are legal.
bool areValueTypesLegal(const TargetLowering &TLI) {
for (unsigned Value = 0, e = ValueVTs.size(); Value != e; ++Value) {
EVT RegisterVT = RegVTs[Value];
if (!TLI.isTypeLegal(RegisterVT))
return false;
}
return true;
}
/// append - Add the specified values to this one.
void append(const RegsForValue &RHS) {
ValueVTs.append(RHS.ValueVTs.begin(), RHS.ValueVTs.end());
RegVTs.append(RHS.RegVTs.begin(), RHS.RegVTs.end());
Regs.append(RHS.Regs.begin(), RHS.Regs.end());
}
/// getCopyFromRegs - Emit a series of CopyFromReg nodes that copies from
/// this value and returns the result as a ValueVTs value. This uses
/// Chain/Flag as the input and updates them for the output Chain/Flag.
/// If the Flag pointer is NULL, no flag is used.
SDValue getCopyFromRegs(SelectionDAG &DAG, FunctionLoweringInfo &FuncInfo,
DebugLoc dl,
SDValue &Chain, SDValue *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.
/// If the Flag pointer is NULL, no flag is used.
void getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
SDValue &Chain, SDValue *Flag) const;
/// AddInlineAsmOperands - Add this value to the specified inlineasm node
/// operand list. This adds the code marker, matching input operand index
/// (if applicable), and includes the number of values added into it.
void AddInlineAsmOperands(unsigned Kind,
bool HasMatching, unsigned MatchingIdx,
SelectionDAG &DAG,
std::vector<SDValue> &Ops) const;
};
}
/// 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.
/// If the Flag pointer is NULL, no flag is used.
SDValue RegsForValue::getCopyFromRegs(SelectionDAG &DAG,
FunctionLoweringInfo &FuncInfo,
DebugLoc dl,
SDValue &Chain, SDValue *Flag) const {
// A Value with type {} or [0 x %t] needs no registers.
if (ValueVTs.empty())
return SDValue();
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// Assemble the legal parts into the final values.
SmallVector<SDValue, 4> Values(ValueVTs.size());
SmallVector<SDValue, 8> Parts;
for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
// Copy the legal parts from the registers.
EVT ValueVT = ValueVTs[Value];
unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
EVT RegisterVT = RegVTs[Value];
Parts.resize(NumRegs);
for (unsigned i = 0; i != NumRegs; ++i) {
SDValue P;
if (Flag == 0) {
P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT);
} else {
P = DAG.getCopyFromReg(Chain, dl, Regs[Part+i], RegisterVT, *Flag);
*Flag = P.getValue(2);
}
Chain = P.getValue(1);
Parts[i] = P;
// If the source register was virtual and if we know something about it,
// add an assert node.
if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
!RegisterVT.isInteger() || RegisterVT.isVector())
continue;
const FunctionLoweringInfo::LiveOutInfo *LOI =
FuncInfo.GetLiveOutRegInfo(Regs[Part+i]);
if (!LOI)
continue;
unsigned RegSize = RegisterVT.getSizeInBits();
unsigned NumSignBits = LOI->NumSignBits;
unsigned NumZeroBits = LOI->KnownZero.countLeadingOnes();
// FIXME: We capture more information than the dag can represent. For
// now, just use the tightest assertzext/assertsext possible.
bool isSExt = true;
EVT FromVT(MVT::Other);
if (NumSignBits == RegSize)
isSExt = true, FromVT = MVT::i1; // ASSERT SEXT 1
else if (NumZeroBits >= RegSize-1)
isSExt = false, FromVT = MVT::i1; // ASSERT ZEXT 1
else if (NumSignBits > RegSize-8)
isSExt = true, FromVT = MVT::i8; // ASSERT SEXT 8
else if (NumZeroBits >= RegSize-8)
isSExt = false, FromVT = MVT::i8; // ASSERT ZEXT 8
else if (NumSignBits > RegSize-16)
isSExt = true, FromVT = MVT::i16; // ASSERT SEXT 16
else if (NumZeroBits >= RegSize-16)
isSExt = false, FromVT = MVT::i16; // ASSERT ZEXT 16
else if (NumSignBits > RegSize-32)
isSExt = true, FromVT = MVT::i32; // ASSERT SEXT 32
else if (NumZeroBits >= RegSize-32)
isSExt = false, FromVT = MVT::i32; // ASSERT ZEXT 32
else
continue;
// Add an assertion node.
assert(FromVT != MVT::Other);
Parts[i] = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
RegisterVT, P, DAG.getValueType(FromVT));
}
Values[Value] = getCopyFromParts(DAG, dl, Parts.begin(),
NumRegs, RegisterVT, ValueVT);
Part += NumRegs;
Parts.clear();
}
return DAG.getNode(ISD::MERGE_VALUES, dl,
DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
&Values[0], ValueVTs.size());
}
/// 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.
/// If the Flag pointer is NULL, no flag is used.
void RegsForValue::getCopyToRegs(SDValue Val, SelectionDAG &DAG, DebugLoc dl,
SDValue &Chain, SDValue *Flag) const {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
// Get the list of the values's legal parts.
unsigned NumRegs = Regs.size();
SmallVector<SDValue, 8> Parts(NumRegs);
for (unsigned Value = 0, Part = 0, e = ValueVTs.size(); Value != e; ++Value) {
EVT ValueVT = ValueVTs[Value];
unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
EVT RegisterVT = RegVTs[Value];
getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
&Parts[Part], NumParts, RegisterVT);
Part += NumParts;
}
// Copy the parts into the registers.
SmallVector<SDValue, 8> Chains(NumRegs);
for (unsigned i = 0; i != NumRegs; ++i) {
SDValue Part;
if (Flag == 0) {
Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i]);
} else {
Part = DAG.getCopyToReg(Chain, dl, Regs[i], Parts[i], *Flag);
*Flag = Part.getValue(1);
}
Chains[i] = Part.getValue(0);
}
if (NumRegs == 1 || Flag)
// If NumRegs > 1 && Flag is used then the use of the last CopyToReg is
// flagged to it. That is the CopyToReg nodes and the user are considered
// a single scheduling unit. If we create a TokenFactor and return it as
// chain, then the TokenFactor is both a predecessor (operand) of the
// user as well as a successor (the TF operands are flagged to the user).
// c1, f1 = CopyToReg
// c2, f2 = CopyToReg
// c3 = TokenFactor c1, c2
// ...
// = op c3, ..., f2
Chain = Chains[NumRegs-1];
else
Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, &Chains[0], NumRegs);
}
/// 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, bool HasMatching,
unsigned MatchingIdx,
SelectionDAG &DAG,
std::vector<SDValue> &Ops) const {
const TargetLowering &TLI = DAG.getTargetLoweringInfo();
unsigned Flag = InlineAsm::getFlagWord(Code, Regs.size());
if (HasMatching)
Flag = InlineAsm::getFlagWordForMatchingOp(Flag, MatchingIdx);
else if (!Regs.empty() &&
TargetRegisterInfo::isVirtualRegister(Regs.front())) {
// Put the register class of the virtual registers in the flag word. That
// way, later passes can recompute register class constraints for inline
// assembly as well as normal instructions.
// Don't do this for tied operands that can use the regclass information
// from the def.
const MachineRegisterInfo &MRI = DAG.getMachineFunction().getRegInfo();
const TargetRegisterClass *RC = MRI.getRegClass(Regs.front());
Flag = InlineAsm::getFlagWordForRegClass(Flag, RC->getID());
}
SDValue Res = DAG.getTargetConstant(Flag, MVT::i32);
Ops.push_back(Res);
for (unsigned Value = 0, Reg = 0, e = ValueVTs.size(); Value != e; ++Value) {
unsigned NumRegs = TLI.getNumRegisters(*DAG.getContext(), ValueVTs[Value]);
EVT RegisterVT = RegVTs[Value];
for (unsigned i = 0; i != NumRegs; ++i) {
assert(Reg < Regs.size() && "Mismatch in # registers expected");
Ops.push_back(DAG.getRegister(Regs[Reg++], RegisterVT));
}
}
}
void SelectionDAGBuilder::init(GCFunctionInfo *gfi, AliasAnalysis &aa,
const TargetLibraryInfo *li) {
AA = &aa;
GFI = gfi;
LibInfo = li;
TD = DAG.getTarget().getTargetData();
LPadToCallSiteMap.clear();
}
/// clear - Clear out the current SelectionDAG and the associated
/// state and prepare this SelectionDAGBuilder object to be used
/// for a new block. This doesn't clear out information about
/// additional blocks that are needed to complete switch lowering
/// or PHI node updating; that information is cleared out as it is
/// consumed.
void SelectionDAGBuilder::clear() {
NodeMap.clear();
UnusedArgNodeMap.clear();
PendingLoads.clear();
PendingExports.clear();
CurDebugLoc = DebugLoc();
HasTailCall = false;
}
/// clearDanglingDebugInfo - Clear the dangling debug information
/// map. This function is seperated from the clear so that debug
/// information that is dangling in a basic block can be properly
/// resolved in a different basic block. This allows the
/// SelectionDAG to resolve dangling debug information attached
/// to PHI nodes.
void SelectionDAGBuilder::clearDanglingDebugInfo() {
DanglingDebugInfoMap.clear();
}
/// getRoot - Return the current virtual root of the Selection DAG,
/// flushing any PendingLoad items. This must be done before emitting
/// a store or any other node that may need to be ordered after any
/// prior load instructions.
///
SDValue SelectionDAGBuilder::getRoot() {
if (PendingLoads.empty())
return DAG.getRoot();
if (PendingLoads.size() == 1) {
SDValue Root = PendingLoads[0];
DAG.setRoot(Root);
PendingLoads.clear();
return Root;
}
// Otherwise, we have to make a token factor node.
SDValue Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
&PendingLoads[0], PendingLoads.size());
PendingLoads.clear();
DAG.setRoot(Root);
return Root;
}
/// getControlRoot - Similar to getRoot, but instead of flushing all the
/// PendingLoad items, flush all the PendingExports items. It is necessary
/// to do this before emitting a terminator instruction.
///
SDValue SelectionDAGBuilder::getControlRoot() {
SDValue Root = DAG.getRoot();
if (PendingExports.empty())
return Root;
// Turn all of the CopyToReg chains into one factored node.
if (Root.getOpcode() != ISD::EntryToken) {
unsigned i = 0, e = PendingExports.size();
for (; i != e; ++i) {
assert(PendingExports[i].getNode()->getNumOperands() > 1);
if (PendingExports[i].getNode()->getOperand(0) == Root)
break; // Don't add the root if we already indirectly depend on it.
}
if (i == e)
PendingExports.push_back(Root);
}
Root = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(), MVT::Other,
&PendingExports[0],
PendingExports.size());
PendingExports.clear();
DAG.setRoot(Root);
return Root;
}
void SelectionDAGBuilder::AssignOrderingToNode(const SDNode *Node) {
if (DAG.GetOrdering(Node) != 0) return; // Already has ordering.
DAG.AssignOrdering(Node, SDNodeOrder);
for (unsigned I = 0, E = Node->getNumOperands(); I != E; ++I)
AssignOrderingToNode(Node->getOperand(I).getNode());
}
void SelectionDAGBuilder::visit(const Instruction &I) {
// Set up outgoing PHI node register values before emitting the terminator.
if (isa<TerminatorInst>(&I))
HandlePHINodesInSuccessorBlocks(I.getParent());
CurDebugLoc = I.getDebugLoc();
visit(I.getOpcode(), I);
if (!isa<TerminatorInst>(&I) && !HasTailCall)
CopyToExportRegsIfNeeded(&I);
CurDebugLoc = DebugLoc();
}
void SelectionDAGBuilder::visitPHI(const PHINode &) {
llvm_unreachable("SelectionDAGBuilder shouldn't visit PHI nodes!");
}
void SelectionDAGBuilder::visit(unsigned Opcode, const User &I) {
// Note: this doesn't use InstVisitor, because it has to work with
// ConstantExpr's in addition to instructions.
switch (Opcode) {
default: llvm_unreachable("Unknown instruction type encountered!");
// Build the switch statement using the Instruction.def file.
#define HANDLE_INST(NUM, OPCODE, CLASS) \
case Instruction::OPCODE: visit##OPCODE((CLASS&)I); break;
#include "llvm/Instruction.def"
}
// Assign the ordering to the freshly created DAG nodes.
if (NodeMap.count(&I)) {
++SDNodeOrder;
AssignOrderingToNode(getValue(&I).getNode());
}
}
// resolveDanglingDebugInfo - if we saw an earlier dbg_value referring to V,
// generate the debug data structures now that we've seen its definition.
void SelectionDAGBuilder::resolveDanglingDebugInfo(const Value *V,
SDValue Val) {
DanglingDebugInfo &DDI = DanglingDebugInfoMap[V];
if (DDI.getDI()) {
const DbgValueInst *DI = DDI.getDI();
DebugLoc dl = DDI.getdl();
unsigned DbgSDNodeOrder = DDI.getSDNodeOrder();
MDNode *Variable = DI->getVariable();
uint64_t Offset = DI->getOffset();
SDDbgValue *SDV;
if (Val.getNode()) {
if (!EmitFuncArgumentDbgValue(V, Variable, Offset, Val)) {
SDV = DAG.getDbgValue(Variable, Val.getNode(),
Val.getResNo(), Offset, dl, DbgSDNodeOrder);
DAG.AddDbgValue(SDV, Val.getNode(), false);
}
} else
DEBUG(dbgs() << "Dropping debug info for " << DI << "\n");
DanglingDebugInfoMap[V] = DanglingDebugInfo();
}
}
/// getValue - Return an SDValue for the given Value.
SDValue SelectionDAGBuilder::getValue(const Value *V) {
// If we already have an SDValue for this value, use it. It's important
// to do this first, so that we don't create a CopyFromReg if we already
// have a regular SDValue.
SDValue &N = NodeMap[V];
if (N.getNode()) return N;
// If there's a virtual register allocated and initialized for this
// value, use it.
DenseMap<const Value *, unsigned>::iterator It = FuncInfo.ValueMap.find(V);
if (It != FuncInfo.ValueMap.end()) {
unsigned InReg = It->second;
RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
SDValue Chain = DAG.getEntryNode();
N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
resolveDanglingDebugInfo(V, N);
return N;
}
// Otherwise create a new SDValue and remember it.
SDValue Val = getValueImpl(V);
NodeMap[V] = Val;
resolveDanglingDebugInfo(V, Val);
return Val;
}
/// getNonRegisterValue - Return an SDValue for the given Value, but
/// don't look in FuncInfo.ValueMap for a virtual register.
SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
// If we already have an SDValue for this value, use it.
SDValue &N = NodeMap[V];
if (N.getNode()) return N;
// Otherwise create a new SDValue and remember it.
SDValue Val = getValueImpl(V);
NodeMap[V] = Val;
resolveDanglingDebugInfo(V, Val);
return Val;
}
/// getValueImpl - Helper function for getValue and getNonRegisterValue.
/// Create an SDValue for the given value.
SDValue SelectionDAGBuilder::getValueImpl(const Value *V) {
if (const Constant *C = dyn_cast<Constant>(V)) {
EVT VT = TLI.getValueType(V->getType(), true);
if (const ConstantInt *CI = dyn_cast<ConstantInt>(C))
return DAG.getConstant(*CI, VT);
if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
return DAG.getGlobalAddress(GV, getCurDebugLoc(), VT);
if (isa<ConstantPointerNull>(C))
return DAG.getConstant(0, TLI.getPointerTy());
if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
return DAG.getConstantFP(*CFP, VT);
if (isa<UndefValue>(C) && !V->getType()->isAggregateType())
return DAG.getUNDEF(VT);
if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
visit(CE->getOpcode(), *CE);
SDValue N1 = NodeMap[V];
assert(N1.getNode() && "visit didn't populate the NodeMap!");
return N1;
}
if (isa<ConstantStruct>(C) || isa<ConstantArray>(C)) {
SmallVector<SDValue, 4> Constants;
for (User::const_op_iterator OI = C->op_begin(), OE = C->op_end();
OI != OE; ++OI) {
SDNode *Val = getValue(*OI).getNode();
// If the operand is an empty aggregate, there are no values.
if (!Val) continue;
// Add each leaf value from the operand to the Constants list
// to form a flattened list of all the values.
for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
Constants.push_back(SDValue(Val, i));
}
return DAG.getMergeValues(&Constants[0], Constants.size(),
getCurDebugLoc());
}
if (const ConstantDataSequential *CDS =
dyn_cast<ConstantDataSequential>(C)) {
SmallVector<SDValue, 4> Ops;
for (unsigned i = 0, e = CDS->getNumElements(); i != e; ++i) {
SDNode *Val = getValue(CDS->getElementAsConstant(i)).getNode();
// Add each leaf value from the operand to the Constants list
// to form a flattened list of all the values.
for (unsigned i = 0, e = Val->getNumValues(); i != e; ++i)
Ops.push_back(SDValue(Val, i));
}
if (isa<ArrayType>(CDS->getType()))
return DAG.getMergeValues(&Ops[0], Ops.size(), getCurDebugLoc());
return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
VT, &Ops[0], Ops.size());
}
if (C->getType()->isStructTy() || C->getType()->isArrayTy()) {
assert((isa<ConstantAggregateZero>(C) || isa<UndefValue>(C)) &&
"Unknown struct or array constant!");
SmallVector<EVT, 4> ValueVTs;
ComputeValueVTs(TLI, C->getType(), ValueVTs);
unsigned NumElts = ValueVTs.size();
if (NumElts == 0)
return SDValue(); // empty struct
SmallVector<SDValue, 4> Constants(NumElts);
for (unsigned i = 0; i != NumElts; ++i) {
EVT EltVT = ValueVTs[i];
if (isa<UndefValue>(C))
Constants[i] = DAG.getUNDEF(EltVT);
else if (EltVT.isFloatingPoint())
Constants[i] = DAG.getConstantFP(0, EltVT);
else
Constants[i] = DAG.getConstant(0, EltVT);
}
return DAG.getMergeValues(&Constants[0], NumElts,
getCurDebugLoc());
}
if (const BlockAddress *BA = dyn_cast<BlockAddress>(C))
return DAG.getBlockAddress(BA, VT);
VectorType *VecTy = cast<VectorType>(V->getType());
unsigned NumElements = VecTy->getNumElements();
// Now that we know the number and type of the elements, get that number of
// elements into the Ops array based on what kind of constant it is.
SmallVector<SDValue, 16> Ops;
if (const ConstantVector *CV = dyn_cast<ConstantVector>(C)) {
for (unsigned i = 0; i != NumElements; ++i)
Ops.push_back(getValue(CV->getOperand(i)));
} else {
assert(isa<ConstantAggregateZero>(C) && "Unknown vector constant!");
EVT EltVT = TLI.getValueType(VecTy->getElementType());
SDValue Op;
if (EltVT.isFloatingPoint())
Op = DAG.getConstantFP(0, EltVT);
else
Op = DAG.getConstant(0, EltVT);
Ops.assign(NumElements, Op);
}
// Create a BUILD_VECTOR node.
return NodeMap[V] = DAG.getNode(ISD::BUILD_VECTOR, getCurDebugLoc(),
VT, &Ops[0], Ops.size());
}
// If this is a static alloca, generate it as the frameindex instead of
// computation.
if (const AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
DenseMap<const AllocaInst*, int>::iterator SI =
FuncInfo.StaticAllocaMap.find(AI);
if (SI != FuncInfo.StaticAllocaMap.end())
return DAG.getFrameIndex(SI->second, TLI.getPointerTy());
}
// If this is an instruction which fast-isel has deferred, select it now.
if (const Instruction *Inst = dyn_cast<Instruction>(V)) {
unsigned InReg = FuncInfo.InitializeRegForValue(Inst);
RegsForValue RFV(*DAG.getContext(), TLI, InReg, Inst->getType());
SDValue Chain = DAG.getEntryNode();
return RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
}
llvm_unreachable("Can't get register for value!");
}
void SelectionDAGBuilder::visitRet(const ReturnInst &I) {
SDValue Chain = getControlRoot();
SmallVector<ISD::OutputArg, 8> Outs;
SmallVector<SDValue, 8> OutVals;
if (!FuncInfo.CanLowerReturn) {
unsigned DemoteReg = FuncInfo.DemoteRegister;
const Function *F = I.getParent()->getParent();
// Emit a store of the return value through the virtual register.
// Leave Outs empty so that LowerReturn won't try to load return
// registers the usual way.
SmallVector<EVT, 1> PtrValueVTs;
ComputeValueVTs(TLI, PointerType::getUnqual(F->getReturnType()),
PtrValueVTs);
SDValue RetPtr = DAG.getRegister(DemoteReg, PtrValueVTs[0]);
SDValue RetOp = getValue(I.getOperand(0));
SmallVector<EVT, 4> ValueVTs;
SmallVector<uint64_t, 4> Offsets;
ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs, &Offsets);
unsigned NumValues = ValueVTs.size();
SmallVector<SDValue, 4> Chains(NumValues);
for (unsigned i = 0; i != NumValues; ++i) {
SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(),
RetPtr.getValueType(), RetPtr,
DAG.getIntPtrConstant(Offsets[i]));
Chains[i] =
DAG.getStore(Chain, getCurDebugLoc(),
SDValue(RetOp.getNode(), RetOp.getResNo() + i),
// FIXME: better loc info would be nice.
Add, MachinePointerInfo(), false, false, 0);
}
Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
MVT::Other, &Chains[0], NumValues);
} else if (I.getNumOperands() != 0) {
SmallVector<EVT, 4> ValueVTs;
ComputeValueVTs(TLI, I.getOperand(0)->getType(), ValueVTs);
unsigned NumValues = ValueVTs.size();
if (NumValues) {
SDValue RetOp = getValue(I.getOperand(0));
for (unsigned j = 0, f = NumValues; j != f; ++j) {
EVT VT = ValueVTs[j];
ISD::NodeType ExtendKind = ISD::ANY_EXTEND;
const Function *F = I.getParent()->getParent();
if (F->paramHasAttr(0, Attribute::SExt))
ExtendKind = ISD::SIGN_EXTEND;
else if (F->paramHasAttr(0, Attribute::ZExt))
ExtendKind = ISD::ZERO_EXTEND;
if (ExtendKind != ISD::ANY_EXTEND && VT.isInteger())
VT = TLI.getTypeForExtArgOrReturn(*DAG.getContext(), VT, ExtendKind);
unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), VT);
EVT PartVT = TLI.getRegisterType(*DAG.getContext(), VT);
SmallVector<SDValue, 4> Parts(NumParts);
getCopyToParts(DAG, getCurDebugLoc(),
SDValue(RetOp.getNode(), RetOp.getResNo() + j),
&Parts[0], NumParts, PartVT, ExtendKind);
// 'inreg' on function refers to return value
ISD::ArgFlagsTy Flags = ISD::ArgFlagsTy();
if (F->paramHasAttr(0, Attribute::InReg))
Flags.setInReg();
// Propagate extension type if any
if (ExtendKind == ISD::SIGN_EXTEND)
Flags.setSExt();
else if (ExtendKind == ISD::ZERO_EXTEND)
Flags.setZExt();
for (unsigned i = 0; i < NumParts; ++i) {
Outs.push_back(ISD::OutputArg(Flags, Parts[i].getValueType(),
/*isfixed=*/true));
OutVals.push_back(Parts[i]);
}
}
}
}
bool isVarArg = DAG.getMachineFunction().getFunction()->isVarArg();
CallingConv::ID CallConv =
DAG.getMachineFunction().getFunction()->getCallingConv();
Chain = TLI.LowerReturn(Chain, CallConv, isVarArg,
Outs, OutVals, getCurDebugLoc(), DAG);
// Verify that the target's LowerReturn behaved as expected.
assert(Chain.getNode() && Chain.getValueType() == MVT::Other &&
"LowerReturn didn't return a valid chain!");
// Update the DAG with the new chain value resulting from return lowering.
DAG.setRoot(Chain);
}
/// CopyToExportRegsIfNeeded - If the given value has virtual registers
/// created for it, emit nodes to copy the value into the virtual
/// registers.
void SelectionDAGBuilder::CopyToExportRegsIfNeeded(const Value *V) {
// Skip empty types
if (V->getType()->isEmptyTy())
return;
DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
if (VMI != FuncInfo.ValueMap.end()) {
assert(!V->use_empty() && "Unused value assigned virtual registers!");
CopyValueToVirtualRegister(V, VMI->second);
}
}
/// ExportFromCurrentBlock - If this condition isn't known to be exported from
/// the current basic block, add it to ValueMap now so that we'll get a
/// CopyTo/FromReg.
void SelectionDAGBuilder::ExportFromCurrentBlock(const Value *V) {
// No need to export constants.
if (!isa<Instruction>(V) && !isa<Argument>(V)) return;
// Already exported?
if (FuncInfo.isExportedInst(V)) return;
unsigned Reg = FuncInfo.InitializeRegForValue(V);
CopyValueToVirtualRegister(V, Reg);
}
bool SelectionDAGBuilder::isExportableFromCurrentBlock(const Value *V,
const BasicBlock *FromBB) {
// The operands of the setcc have to be in this block. We don't know
// how to export them from some other block.
if (const Instruction *VI = dyn_cast<Instruction>(V)) {
// Can export from current BB.
if (VI->getParent() == FromBB)
return true;
// Is already exported, noop.
return FuncInfo.isExportedInst(V);
}
// If this is an argument, we can export it if the BB is the entry block or
// if it is already exported.
if (isa<Argument>(V)) {
if (FromBB == &FromBB->getParent()->getEntryBlock())
return true;
// Otherwise, can only export this if it is already exported.
return FuncInfo.isExportedInst(V);
}
// Otherwise, constants can always be exported.
return true;
}
/// Return branch probability calculated by BranchProbabilityInfo for IR blocks.
uint32_t SelectionDAGBuilder::getEdgeWeight(const MachineBasicBlock *Src,
const MachineBasicBlock *Dst) const {
BranchProbabilityInfo *BPI = FuncInfo.BPI;
if (!BPI)
return 0;
const BasicBlock *SrcBB = Src->getBasicBlock();
const BasicBlock *DstBB = Dst->getBasicBlock();
return BPI->getEdgeWeight(SrcBB, DstBB);
}
void SelectionDAGBuilder::
addSuccessorWithWeight(MachineBasicBlock *Src, MachineBasicBlock *Dst,
uint32_t Weight /* = 0 */) {
if (!Weight)
Weight = getEdgeWeight(Src, Dst);
Src->addSuccessor(Dst, Weight);
}
static bool InBlock(const Value *V, const BasicBlock *BB) {
if (const Instruction *I = dyn_cast<Instruction>(V))
return I->getParent() == BB;
return true;
}
/// EmitBranchForMergedCondition - Helper method for FindMergedConditions.
/// This function emits a branch and is used at the leaves of an OR or an
/// AND operator tree.
///
void
SelectionDAGBuilder::EmitBranchForMergedCondition(const Value *Cond,
MachineBasicBlock *TBB,
MachineBasicBlock *FBB,
MachineBasicBlock *CurBB,
MachineBasicBlock *SwitchBB) {
const BasicBlock *BB = CurBB->getBasicBlock();
// If the leaf of the tree is a comparison, merge the condition into
// the caseblock.
if (const CmpInst *BOp = dyn_cast<CmpInst>(Cond)) {
// The operands of the cmp have to be in this block. We don't know
// how to export them from some other block. If this is the first block
// of the sequence, no exporting is needed.
if (CurBB == SwitchBB ||
(isExportableFromCurrentBlock(BOp->getOperand(0), BB) &&
isExportableFromCurrentBlock(BOp->getOperand(1), BB))) {
ISD::CondCode Condition;
if (const ICmpInst *IC = dyn_cast<ICmpInst>(Cond)) {
Condition = getICmpCondCode(IC->getPredicate());
} else if (const FCmpInst *FC = dyn_cast<FCmpInst>(Cond)) {
Condition = getFCmpCondCode(FC->getPredicate());
if (TM.Options.NoNaNsFPMath)
Condition = getFCmpCodeWithoutNaN(Condition);
} else {
Condition = ISD::SETEQ; // silence warning.
llvm_unreachable("Unknown compare instruction");
}
CaseBlock CB(Condition, BOp->getOperand(0),
BOp->getOperand(1), NULL, TBB, FBB, CurBB);
SwitchCases.push_back(CB);
return;
}
}
// Create a CaseBlock record representing this branch.
CaseBlock CB(ISD::SETEQ, Cond, ConstantInt::getTrue(*DAG.getContext()),
NULL, TBB, FBB, CurBB);
SwitchCases.push_back(CB);
}
/// FindMergedConditions - If Cond is an expression like
void SelectionDAGBuilder::FindMergedConditions(const Value *Cond,
MachineBasicBlock *TBB,
MachineBasicBlock *FBB,
MachineBasicBlock *CurBB,
MachineBasicBlock *SwitchBB,
unsigned Opc) {
// If this node is not part of the or/and tree, emit it as a branch.
const Instruction *BOp = dyn_cast<Instruction>(Cond);
if (!BOp || !(isa<BinaryOperator>(BOp) || isa<CmpInst>(BOp)) ||
(unsigned)BOp->getOpcode() != Opc || !BOp->hasOneUse() ||
BOp->getParent() != CurBB->getBasicBlock() ||
!InBlock(BOp->getOperand(0), CurBB->getBasicBlock()) ||
!InBlock(BOp->getOperand(1), CurBB->getBasicBlock())) {
EmitBranchForMergedCondition(Cond, TBB, FBB, CurBB, SwitchBB);
return;
}
// Create TmpBB after CurBB.
MachineFunction::iterator BBI = CurBB;
MachineFunction &MF = DAG.getMachineFunction();
MachineBasicBlock *TmpBB = MF.CreateMachineBasicBlock(CurBB->getBasicBlock());
CurBB->getParent()->insert(++BBI, TmpBB);
if (Opc == Instruction::Or) {
// Codegen X | Y as:
// jmp_if_X TBB
// jmp TmpBB
// TmpBB:
// jmp_if_Y TBB
// jmp FBB
//
// Emit the LHS condition.
FindMergedConditions(BOp->getOperand(0), TBB, TmpBB, CurBB, SwitchBB, Opc);
// Emit the RHS condition into TmpBB.
FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
} else {
assert(Opc == Instruction::And && "Unknown merge op!");
// Codegen X & Y as:
// jmp_if_X TmpBB
// jmp FBB
// TmpBB:
// jmp_if_Y TBB
// jmp FBB
//
// This requires creation of TmpBB after CurBB.
// Emit the LHS condition.
FindMergedConditions(BOp->getOperand(0), TmpBB, FBB, CurBB, SwitchBB, Opc);
// Emit the RHS condition into TmpBB.
FindMergedConditions(BOp->getOperand(1), TBB, FBB, TmpBB, SwitchBB, Opc);
}
}
/// If the set of cases should be emitted as a series of branches, return true.
/// If we should emit this as a bunch of and/or'd together conditions, return
/// false.
bool
SelectionDAGBuilder::ShouldEmitAsBranches(const std::vector<CaseBlock> &Cases){
if (Cases.size() != 2) return true;
// If this is two comparisons of the same values or'd or and'd together, they
// will get folded into a single comparison, so don't emit two blocks.
if ((Cases[0].CmpLHS == Cases[1].CmpLHS &&
Cases[0].CmpRHS == Cases[1].CmpRHS) ||
(Cases[0].CmpRHS == Cases[1].CmpLHS &&
Cases[0].CmpLHS == Cases[1].CmpRHS)) {
return false;
}
// Handle: (X != null) | (Y != null) --> (X|Y) != 0
// Handle: (X == null) & (Y == null) --> (X|Y) == 0
if (Cases[0].CmpRHS == Cases[1].CmpRHS &&
Cases[0].CC == Cases[1].CC &&
isa<Constant>(Cases[0].CmpRHS) &&
cast<Constant>(Cases[0].CmpRHS)->isNullValue()) {
if (Cases[0].CC == ISD::SETEQ && Cases[0].TrueBB == Cases[1].ThisBB)
return false;
if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
return false;
}
return true;
}
void SelectionDAGBuilder::visitBr(const BranchInst &I) {
MachineBasicBlock *BrMBB = FuncInfo.MBB;
// Update machine-CFG edges.
MachineBasicBlock *Succ0MBB = FuncInfo.MBBMap[I.getSuccessor(0)];
// Figure out which block is immediately after the current one.
MachineBasicBlock *NextBlock = 0;
MachineFunction::iterator BBI = BrMBB;
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
if (I.isUnconditional()) {
// Update machine-CFG edges.
BrMBB->addSuccessor(Succ0MBB);
// If this is not a fall-through branch, emit the branch.
if (Succ0MBB != NextBlock)
DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
MVT::Other, getControlRoot(),
DAG.getBasicBlock(Succ0MBB)));
return;
}
// If this condition is one of the special cases we handle, do special stuff
// now.
const Value *CondVal = I.getCondition();
MachineBasicBlock *Succ1MBB = FuncInfo.MBBMap[I.getSuccessor(1)];
// If this is a series of conditions that are or'd or and'd together, emit
// this as a sequence of branches instead of setcc's with and/or operations.
// As long as jumps are not expensive, this should improve performance.
// For example, instead of something like:
// cmp A, B
// C = seteq
// cmp D, E
// F = setle
// or C, F
// jnz foo
// Emit:
// cmp A, B
// je foo
// cmp D, E
// jle foo
//
if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
if (!TLI.isJumpExpensive() &&
BOp->hasOneUse() &&
(BOp->getOpcode() == Instruction::And ||
BOp->getOpcode() == Instruction::Or)) {
FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
BOp->getOpcode());
// If the compares in later blocks need to use values not currently
// exported from this block, export them now. This block should always
// be the first entry.
assert(SwitchCases[0].ThisBB == BrMBB && "Unexpected lowering!");
// Allow some cases to be rejected.
if (ShouldEmitAsBranches(SwitchCases)) {
for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i) {
ExportFromCurrentBlock(SwitchCases[i].CmpLHS);
ExportFromCurrentBlock(SwitchCases[i].CmpRHS);
}
// Emit the branch for this block.
visitSwitchCase(SwitchCases[0], BrMBB);
SwitchCases.erase(SwitchCases.begin());
return;
}
// Okay, we decided not to do this, remove any inserted MBB's and clear
// SwitchCases.
for (unsigned i = 1, e = SwitchCases.size(); i != e; ++i)
FuncInfo.MF->erase(SwitchCases[i].ThisBB);
SwitchCases.clear();
}
}
// Create a CaseBlock record representing this branch.
CaseBlock CB(ISD::SETEQ, CondVal, ConstantInt::getTrue(*DAG.getContext()),
NULL, Succ0MBB, Succ1MBB, BrMBB);
// Use visitSwitchCase to actually insert the fast branch sequence for this
// cond branch.
visitSwitchCase(CB, BrMBB);
}
/// visitSwitchCase - Emits the necessary code to represent a single node in
/// the binary search tree resulting from lowering a switch instruction.
void SelectionDAGBuilder::visitSwitchCase(CaseBlock &CB,
MachineBasicBlock *SwitchBB) {
SDValue Cond;
SDValue CondLHS = getValue(CB.CmpLHS);
DebugLoc dl = getCurDebugLoc();
// Build the setcc now.
if (CB.CmpMHS == NULL) {
// Fold "(X == true)" to X and "(X == false)" to !X to
// handle common cases produced by branch lowering.
if (CB.CmpRHS == ConstantInt::getTrue(*DAG.getContext()) &&
CB.CC == ISD::SETEQ)
Cond = CondLHS;
else if (CB.CmpRHS == ConstantInt::getFalse(*DAG.getContext()) &&
CB.CC == ISD::SETEQ) {
SDValue True = DAG.getConstant(1, CondLHS.getValueType());
Cond = DAG.getNode(ISD::XOR, dl, CondLHS.getValueType(), CondLHS, True);
} else
Cond = DAG.getSetCC(dl, MVT::i1, CondLHS, getValue(CB.CmpRHS), CB.CC);
} else {
assert(CB.CC == ISD::SETLE && "Can handle only LE ranges now");
const APInt& Low = cast<ConstantInt>(CB.CmpLHS)->getValue();
const APInt& High = cast<ConstantInt>(CB.CmpRHS)->getValue();
SDValue CmpOp = getValue(CB.CmpMHS);
EVT VT = CmpOp.getValueType();
if (cast<ConstantInt>(CB.CmpLHS)->isMinValue(true)) {
Cond = DAG.getSetCC(dl, MVT::i1, CmpOp, DAG.getConstant(High, VT),
ISD::SETLE);
} else {
SDValue SUB = DAG.getNode(ISD::SUB, dl,
VT, CmpOp, DAG.getConstant(Low, VT));
Cond = DAG.getSetCC(dl, MVT::i1, SUB,
DAG.getConstant(High-Low, VT), ISD::SETULE);
}
}
// Update successor info
addSuccessorWithWeight(SwitchBB, CB.TrueBB, CB.TrueWeight);
addSuccessorWithWeight(SwitchBB, CB.FalseBB, CB.FalseWeight);
// 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 = SwitchBB;
if (++BBI != FuncInfo.MF->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.TrueBB == NextBlock) {
std::swap(CB.TrueBB, CB.FalseBB);
SDValue True = DAG.getConstant(1, Cond.getValueType());
Cond = DAG.getNode(ISD::XOR, dl, Cond.getValueType(), Cond, True);
}
SDValue BrCond = DAG.getNode(ISD::BRCOND, dl,
MVT::Other, getControlRoot(), Cond,
DAG.getBasicBlock(CB.TrueBB));
// Insert the false branch. Do this even if it's a fall through branch,
// this makes it easier to do DAG optimizations which require inverting
// the branch condition.
BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
DAG.getBasicBlock(CB.FalseBB));
DAG.setRoot(BrCond);
}
/// visitJumpTable - Emit JumpTable node in the current MBB
void SelectionDAGBuilder::visitJumpTable(JumpTable &JT) {
// Emit the code for the jump table
assert(JT.Reg != -1U && "Should lower JT Header first!");
EVT PTy = TLI.getPointerTy();
SDValue Index = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
JT.Reg, PTy);
SDValue Table = DAG.getJumpTable(JT.JTI, PTy);
SDValue BrJumpTable = DAG.getNode(ISD::BR_JT, getCurDebugLoc(),
MVT::Other, Index.getValue(1),
Table, Index);
DAG.setRoot(BrJumpTable);
}
/// visitJumpTableHeader - This function emits necessary code to produce index
/// in the JumpTable from switch case.
void SelectionDAGBuilder::visitJumpTableHeader(JumpTable &JT,
JumpTableHeader &JTH,
MachineBasicBlock *SwitchBB) {
// 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.
SDValue SwitchOp = getValue(JTH.SValue);
EVT VT = SwitchOp.getValueType();
SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
DAG.getConstant(JTH.First, VT));
// The SDNode we just created, which holds the value being switched on minus
// 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.
SwitchOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), TLI.getPointerTy());
unsigned JumpTableReg = FuncInfo.CreateReg(TLI.getPointerTy());
SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
JumpTableReg, SwitchOp);
JT.Reg = JumpTableReg;
// 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.
SDValue CMP = DAG.getSetCC(getCurDebugLoc(),
TLI.getSetCCResultType(Sub.getValueType()), Sub,
DAG.getConstant(JTH.Last-JTH.First,VT),
ISD::SETUGT);
// 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 = SwitchBB;
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
SDValue BrCond = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
MVT::Other, CopyTo, CMP,
DAG.getBasicBlock(JT.Default));
if (JT.MBB != NextBlock)
BrCond = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrCond,
DAG.getBasicBlock(JT.MBB));
DAG.setRoot(BrCond);
}
/// visitBitTestHeader - This function emits necessary code to produce value
/// suitable for "bit tests"
void SelectionDAGBuilder::visitBitTestHeader(BitTestBlock &B,
MachineBasicBlock *SwitchBB) {
// Subtract the minimum value
SDValue SwitchOp = getValue(B.SValue);
EVT VT = SwitchOp.getValueType();
SDValue Sub = DAG.getNode(ISD::SUB, getCurDebugLoc(), VT, SwitchOp,
DAG.getConstant(B.First, VT));
// Check range
SDValue RangeCmp = DAG.getSetCC(getCurDebugLoc(),
TLI.getSetCCResultType(Sub.getValueType()),
Sub, DAG.getConstant(B.Range, VT),
ISD::SETUGT);
// Determine the type of the test operands.
bool UsePtrType = false;
if (!TLI.isTypeLegal(VT))
UsePtrType = true;
else {
for (unsigned i = 0, e = B.Cases.size(); i != e; ++i)
if (!isUIntN(VT.getSizeInBits(), B.Cases[i].Mask)) {
// Switch table case range are encoded into series of masks.
// Just use pointer type, it's guaranteed to fit.
UsePtrType = true;
break;
}
}
if (UsePtrType) {
VT = TLI.getPointerTy();
Sub = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(), VT);
}
B.RegVT = VT;
B.Reg = FuncInfo.CreateReg(VT);
SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
B.Reg, Sub);
// 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 = SwitchBB;
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
MachineBasicBlock* MBB = B.Cases[0].ThisBB;
addSuccessorWithWeight(SwitchBB, B.Default);
addSuccessorWithWeight(SwitchBB, MBB);
SDValue BrRange = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
MVT::Other, CopyTo, RangeCmp,
DAG.getBasicBlock(B.Default));
if (MBB != NextBlock)
BrRange = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, CopyTo,
DAG.getBasicBlock(MBB));
DAG.setRoot(BrRange);
}
/// visitBitTestCase - this function produces one "bit test"
void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
MachineBasicBlock* NextMBB,
unsigned Reg,
BitTestCase &B,
MachineBasicBlock *SwitchBB) {
EVT VT = BB.RegVT;
SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
Reg, VT);
SDValue Cmp;
unsigned PopCount = CountPopulation_64(B.Mask);
if (PopCount == 1) {
// Testing for a single bit; just compare the shift count with what it
// would need to be to shift a 1 bit in that position.
Cmp = DAG.getSetCC(getCurDebugLoc(),
TLI.getSetCCResultType(VT),
ShiftOp,
DAG.getConstant(CountTrailingZeros_64(B.Mask), VT),
ISD::SETEQ);
} else if (PopCount == BB.Range) {
// There is only one zero bit in the range, test for it directly.
Cmp = DAG.getSetCC(getCurDebugLoc(),
TLI.getSetCCResultType(VT),
ShiftOp,
DAG.getConstant(CountTrailingOnes_64(B.Mask), VT),
ISD::SETNE);
} else {
// Make desired shift
SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(), VT,
DAG.getConstant(1, VT), ShiftOp);
// Emit bit tests and jumps
SDValue AndOp = DAG.getNode(ISD::AND, getCurDebugLoc(),
VT, SwitchVal, DAG.getConstant(B.Mask, VT));
Cmp = DAG.getSetCC(getCurDebugLoc(),
TLI.getSetCCResultType(VT),
AndOp, DAG.getConstant(0, VT),
ISD::SETNE);
}
addSuccessorWithWeight(SwitchBB, B.TargetBB);
addSuccessorWithWeight(SwitchBB, NextMBB);
SDValue BrAnd = DAG.getNode(ISD::BRCOND, getCurDebugLoc(),
MVT::Other, getControlRoot(),
Cmp, DAG.getBasicBlock(B.TargetBB));
// 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 = SwitchBB;
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
if (NextMBB != NextBlock)
BrAnd = DAG.getNode(ISD::BR, getCurDebugLoc(), MVT::Other, BrAnd,
DAG.getBasicBlock(NextMBB));
DAG.setRoot(BrAnd);
}
void SelectionDAGBuilder::visitInvoke(const InvokeInst &I) {
MachineBasicBlock *InvokeMBB = FuncInfo.MBB;
// Retrieve successors.
MachineBasicBlock *Return = FuncInfo.MBBMap[I.getSuccessor(0)];
MachineBasicBlock *LandingPad = FuncInfo.MBBMap[I.getSuccessor(1)];
const Value *Callee(I.getCalledValue());
if (isa<InlineAsm>(Callee))
visitInlineAsm(&I);
else
LowerCallTo(&I, getValue(Callee), false, LandingPad);
// If the value of the invoke is used outside of its defining block, make it
// available as a virtual register.
CopyToExportRegsIfNeeded(&I);
// Update successor info
addSuccessorWithWeight(InvokeMBB, Return);
addSuccessorWithWeight(InvokeMBB, LandingPad);
// Drop into normal successor.
DAG.setRoot(DAG.getNode(ISD::BR, getCurDebugLoc(),
MVT::Other, getControlRoot(),
DAG.getBasicBlock(Return)));
}
void SelectionDAGBuilder::visitResume(const ResumeInst &RI) {
llvm_unreachable("SelectionDAGBuilder shouldn't visit resume instructions!");
}
void SelectionDAGBuilder::visitLandingPad(const LandingPadInst &LP) {
assert(FuncInfo.MBB->isLandingPad() &&
"Call to landingpad not in landing pad!");
MachineBasicBlock *MBB = FuncInfo.MBB;
MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
AddLandingPadInfo(LP, MMI, MBB);
// If there aren't registers to copy the values into (e.g., during SjLj
// exceptions), then don't bother to create these DAG nodes.
if (TLI.getExceptionPointerRegister() == 0 &&
TLI.getExceptionSelectorRegister() == 0)
return;
SmallVector<EVT, 2> ValueVTs;
ComputeValueVTs(TLI, LP.getType(), ValueVTs);
// Insert the EXCEPTIONADDR instruction.
assert(FuncInfo.MBB->isLandingPad() &&
"Call to eh.exception not in landing pad!");
SDVTList VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
SDValue Ops[2];
Ops[0] = DAG.getRoot();
SDValue Op1 = DAG.getNode(ISD::EXCEPTIONADDR, getCurDebugLoc(), VTs, Ops, 1);
SDValue Chain = Op1.getValue(1);
// Insert the EHSELECTION instruction.
VTs = DAG.getVTList(TLI.getPointerTy(), MVT::Other);
Ops[0] = Op1;
Ops[1] = Chain;
SDValue Op2 = DAG.getNode(ISD::EHSELECTION, getCurDebugLoc(), VTs, Ops, 2);
Chain = Op2.getValue(1);
Op2 = DAG.getSExtOrTrunc(Op2, getCurDebugLoc(), MVT::i32);
Ops[0] = Op1;
Ops[1] = Op2;
SDValue Res = DAG.getNode(ISD::MERGE_VALUES, getCurDebugLoc(),
DAG.getVTList(&ValueVTs[0], ValueVTs.size()),
&Ops[0], 2);
std::pair<SDValue, SDValue> RetPair = std::make_pair(Res, Chain);
setValue(&LP, RetPair.first);
DAG.setRoot(RetPair.second);
}
/// handleSmallSwitchCaseRange - Emit a series of specific tests (suitable for
/// small case ranges).
bool SelectionDAGBuilder::handleSmallSwitchRange(CaseRec& CR,
CaseRecVector& WorkList,
const Value* SV,
MachineBasicBlock *Default,
MachineBasicBlock *SwitchBB) {
Case& BackCase = *(CR.Range.second-1);
// Size is the number of Cases represented by this range.
size_t Size = CR.Range.second - CR.Range.first;
if (Size > 3)
return false;
// Get the MachineFunction which holds the current MBB. This is used when
// inserting any additional MBBs necessary to represent the switch.
MachineFunction *CurMF = FuncInfo.MF;
// Figure out which block is immediately after the current one.
MachineBasicBlock *NextBlock = 0;
MachineFunction::iterator BBI = CR.CaseBB;
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
// If any two of the cases has the same destination, and if one value
// is the same as the other, but has one bit unset that the other has set,
// use bit manipulation to do two compares at once. For example:
// "if (X == 6 || X == 4)" -> "if ((X|2) == 6)"
// TODO: This could be extended to merge any 2 cases in switches with 3 cases.
// TODO: Handle cases where CR.CaseBB != SwitchBB.
if (Size == 2 && CR.CaseBB == SwitchBB) {
Case &Small = *CR.Range.first;
Case &Big = *(CR.Range.second-1);
if (Small.Low == Small.High && Big.Low == Big.High && Small.BB == Big.BB) {
const APInt& SmallValue = cast<ConstantInt>(Small.Low)->getValue();
const APInt& BigValue = cast<ConstantInt>(Big.Low)->getValue();
// Check that there is only one bit different.
if (BigValue.countPopulation() == SmallValue.countPopulation() + 1 &&
(SmallValue | BigValue) == BigValue) {
// Isolate the common bit.
APInt CommonBit = BigValue & ~SmallValue;
assert((SmallValue | CommonBit) == BigValue &&
CommonBit.countPopulation() == 1 && "Not a common bit?");
SDValue CondLHS = getValue(SV);
EVT VT = CondLHS.getValueType();
DebugLoc DL = getCurDebugLoc();
SDValue Or = DAG.getNode(ISD::OR, DL, VT, CondLHS,
DAG.getConstant(CommonBit, VT));
SDValue Cond = DAG.getSetCC(DL, MVT::i1,
Or, DAG.getConstant(BigValue, VT),
ISD::SETEQ);
// Update successor info.
addSuccessorWithWeight(SwitchBB, Small.BB);
addSuccessorWithWeight(SwitchBB, Default);
// Insert the true branch.
SDValue BrCond = DAG.getNode(ISD::BRCOND, DL, MVT::Other,
getControlRoot(), Cond,
DAG.getBasicBlock(Small.BB));
// Insert the false branch.
BrCond = DAG.getNode(ISD::BR, DL, MVT::Other, BrCond,
DAG.getBasicBlock(Default));
DAG.setRoot(BrCond);
return true;
}
}
}
// Rearrange the case blocks so that the last one falls through if possible.
if (NextBlock && Default != NextBlock && BackCase.BB != NextBlock) {
// The last case block won't fall through into 'NextBlock' if we emit the
// branches in this order. See if rearranging a case value would help.
for (CaseItr I = CR.Range.first, E = CR.Range.second-1; I != E; ++I) {
if (I->BB == NextBlock) {
std::swap(*I, BackCase);
break;
}
}
}
// 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.
MachineBasicBlock *CurBlock = CR.CaseBB;
for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I) {
MachineBasicBlock *FallThrough;
if (I != E-1) {
FallThrough = CurMF->CreateMachineBasicBlock(CurBlock->getBasicBlock());
CurMF->insert(BBI, FallThrough);
// Put SV in a virtual register to make it available from the new blocks.
ExportFromCurrentBlock(SV);
} else {
// If the last case doesn't match, go to the default block.
FallThrough = Default;
}
const Value *RHS, *LHS, *MHS;
ISD::CondCode CC;
if (I->High == I->Low) {
// This is just small small case range :) containing exactly 1 case
CC = ISD::SETEQ;
LHS = SV; RHS = I->High; MHS = NULL;
} else {
CC = ISD::SETLE;
LHS = I->Low; MHS = SV; RHS = I->High;
}
uint32_t ExtraWeight = I->ExtraWeight;
CaseBlock CB(CC, LHS, RHS, MHS, /* truebb */ I->BB, /* falsebb */ FallThrough,
/* me */ CurBlock,
/* trueweight */ ExtraWeight / 2, /* falseweight */ ExtraWeight / 2);
// If emitting the first comparison, 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 (CurBlock == SwitchBB)
visitSwitchCase(CB, SwitchBB);
else
SwitchCases.push_back(CB);
CurBlock = FallThrough;
}
return true;
}
static inline bool areJTsAllowed(const TargetLowering &TLI) {
return !TLI.getTargetMachine().Options.DisableJumpTables &&
(TLI.isOperationLegalOrCustom(ISD::BR_JT, MVT::Other) ||
TLI.isOperationLegalOrCustom(ISD::BRIND, MVT::Other));
}
static APInt ComputeRange(const APInt &First, const APInt &Last) {
uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
return (LastExt - FirstExt + 1ULL);
}
/// handleJTSwitchCase - Emit jumptable for current switch case range
bool SelectionDAGBuilder::handleJTSwitchCase(CaseRec &CR,
CaseRecVector &WorkList,
const Value *SV,
MachineBasicBlock *Default,
MachineBasicBlock *SwitchBB) {
Case& FrontCase = *CR.Range.first;
Case& BackCase = *(CR.Range.second-1);
const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
APInt TSize(First.getBitWidth(), 0);
for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++I)
TSize += I->size();
if (!areJTsAllowed(TLI) || TSize.ult(4))
return false;
APInt Range = ComputeRange(First, Last);
// The density is TSize / Range. Require at least 40%.
// It should not be possible for IntTSize to saturate for sane code, but make
// sure we handle Range saturation correctly.
uint64_t IntRange = Range.getLimitedValue(UINT64_MAX/10);
uint64_t IntTSize = TSize.getLimitedValue(UINT64_MAX/10);
if (IntTSize * 10 < IntRange * 4)
return false;
DEBUG(dbgs() << "Lowering jump table\n"
<< "First entry: " << First << ". Last entry: " << Last << '\n'
<< "Range: " << Range << ". Size: " << TSize << ".\n\n");
// Get the MachineFunction which holds the current MBB. This is used when
// inserting any additional MBBs necessary to represent the switch.
MachineFunction *CurMF = FuncInfo.MF;
// Figure out which block is immediately after the current one.
MachineFunction::iterator BBI = CR.CaseBB;
++BBI;
const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
// 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 = CurMF->CreateMachineBasicBlock(LLVMBB);
CurMF->insert(BBI, JumpTableBB);
addSuccessorWithWeight(CR.CaseBB, Default);
addSuccessorWithWeight(CR.CaseBB, JumpTableBB);
// 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::vector<MachineBasicBlock*> DestBBs;
APInt TEI = First;
for (CaseItr I = CR.Range.first, E = CR.Range.second; I != E; ++TEI) {
const APInt &Low = cast<ConstantInt>(I->Low)->getValue();
const APInt &High = cast<ConstantInt>(I->High)->getValue();
if (Low.sle(TEI) && TEI.sle(High)) {
DestBBs.push_back(I->BB);
if (TEI==High)
++I;
} else {
DestBBs.push_back(Default);
}
}
// Update successor info. Add one edge to each unique successor.
BitVector SuccsHandled(CR.CaseBB->getParent()->getNumBlockIDs());
for (std::vector<MachineBasicBlock*>::iterator I = DestBBs.begin(),
E = DestBBs.end(); I != E; ++I) {
if (!SuccsHandled[(*I)->getNumber()]) {
SuccsHandled[(*I)->getNumber()] = true;
addSuccessorWithWeight(JumpTableBB, *I);
}
}
// Create a jump table index for this jump table.
unsigned JTEncoding = TLI.getJumpTableEncoding();
unsigned JTI = CurMF->getOrCreateJumpTableInfo(JTEncoding)
->createJumpTableIndex(DestBBs);
// Set the jump table information so that we can codegen it as a second
// MachineBasicBlock
JumpTable JT(-1U, JTI, JumpTableBB, Default);
JumpTableHeader JTH(First, Last, SV, CR.CaseBB, (CR.CaseBB == SwitchBB));
if (CR.CaseBB == SwitchBB)
visitJumpTableHeader(JT, JTH, SwitchBB);
JTCases.push_back(JumpTableBlock(JTH, JT));
return true;
}
/// handleBTSplitSwitchCase - emit comparison and split binary search tree into
/// 2 subtrees.
bool SelectionDAGBuilder::handleBTSplitSwitchCase(CaseRec& CR,
CaseRecVector& WorkList,
const Value* SV,
MachineBasicBlock *Default,
MachineBasicBlock *SwitchBB) {
// Get the MachineFunction which holds the current MBB. This is used when
// inserting any additional MBBs necessary to represent the switch.
MachineFunction *CurMF = FuncInfo.MF;
// Figure out which block is immediately after the current one.
MachineFunction::iterator BBI = CR.CaseBB;
++BBI;
Case& FrontCase = *CR.Range.first;
Case& BackCase = *(CR.Range.second-1);
const BasicBlock *LLVMBB = CR.CaseBB->getBasicBlock();
// Size is the number of Cases represented by this range.
unsigned Size = CR.Range.second - CR.Range.first;
const APInt &First = cast<ConstantInt>(FrontCase.Low)->getValue();
const APInt &Last = cast<ConstantInt>(BackCase.High)->getValue();
double FMetric = 0;
CaseItr Pivot = CR.Range.first + Size/2;
// Select optimal pivot, maximizing sum density of LHS and RHS. This will
// (heuristically) allow us to emit JumpTable's later.
APInt TSize(First.getBitWidth(), 0);
for (CaseItr I = CR.Range.first, E = CR.Range.second;
I!=E; ++I)
TSize += I->size();
APInt LSize = FrontCase.size();
APInt RSize = TSize-LSize;
DEBUG(dbgs() << "Selecting best pivot: \n"
<< "First: " << First << ", Last: " << Last <<'\n'
<< "LSize: " << LSize << ", RSize: " << RSize << '\n');
for (CaseItr I = CR.Range.first, J=I+1, E = CR.Range.second;
J!=E; ++I, ++J) {
const APInt &LEnd = cast<ConstantInt>(I->High)->getValue();
const APInt &RBegin = cast<ConstantInt>(J->Low)->getValue();
APInt Range = ComputeRange(LEnd, RBegin);
assert((Range - 2ULL).isNonNegative() &&
"Invalid case distance");
// Use volatile double here to avoid excess precision issues on some hosts,
// e.g. that use 80-bit X87 registers.
volatile double LDensity =
(double)LSize.roundToDouble() /
(LEnd - First + 1ULL).roundToDouble();
volatile double RDensity =
(double)RSize.roundToDouble() /
(Last - RBegin + 1ULL).roundToDouble();
double Metric = Range.logBase2()*(LDensity+RDensity);
// Should always split in some non-trivial place
DEBUG(dbgs() <<"=>Step\n"
<< "LEnd: " << LEnd << ", RBegin: " << RBegin << '\n'
<< "LDensity: " << LDensity
<< ", RDensity: " << RDensity << '\n'
<< "Metric: " << Metric << '\n');
if (FMetric < Metric) {
Pivot = J;
FMetric = Metric;
DEBUG(dbgs() << "Current metric set to: " << FMetric << '\n');
}
LSize += J->size();
RSize -= J->size();
}
if (areJTsAllowed(TLI)) {
// If our case is dense we *really* should handle it earlier!
assert((FMetric > 0) && "Should handle dense range earlier!");
} else {
Pivot = CR.Range.first + Size/2;
}
CaseRange LHSR(CR.Range.first, Pivot);
CaseRange RHSR(Pivot, CR.Range.second);
const Constant *C = Pivot->Low;
MachineBasicBlock *FalseBB = 0, *TrueBB = 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->High == CR.GE &&
cast<ConstantInt>(C)->getValue() ==
(cast<ConstantInt>(CR.GE)->getValue() + 1LL)) {
TrueBB = LHSR.first->BB;
} else {
TrueBB = CurMF->CreateMachineBasicBlock(LLVMBB);
CurMF->insert(BBI, TrueBB);
WorkList.push_back(CaseRec(TrueBB, C, CR.GE, LHSR));
// Put SV in a virtual register to make it available from the new blocks.
ExportFromCurrentBlock(SV);
}
// 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<ConstantInt>(RHSR.first->Low)->getValue() ==
(cast<ConstantInt>(CR.LT)->getValue() - 1LL)) {
FalseBB = RHSR.first->BB;
} else {
FalseBB = CurMF->CreateMachineBasicBlock(LLVMBB);
CurMF->insert(BBI, FalseBB);
WorkList.push_back(CaseRec(FalseBB,CR.LT,C,RHSR));
// Put SV in a virtual register to make it available from the new blocks.
ExportFromCurrentBlock(SV);
}
// 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.
CaseBlock CB(ISD::SETLT, SV, C, NULL, TrueBB, FalseBB, CR.CaseBB);
if (CR.CaseBB == SwitchBB)
visitSwitchCase(CB, SwitchBB);
else
SwitchCases.push_back(CB);
return true;
}
/// handleBitTestsSwitchCase - if current case range has few destination and
/// range span less, than machine word bitwidth, encode case range into series
/// of masks and emit bit tests with these masks.
bool SelectionDAGBuilder::handleBitTestsSwitchCase(CaseRec& CR,
CaseRecVector& WorkList,
const Value* SV,
MachineBasicBlock* Default,
MachineBasicBlock *SwitchBB){
EVT PTy = TLI.getPointerTy();
unsigned IntPtrBits = PTy.getSizeInBits();
Case& FrontCase = *CR.Range.first;
Case& BackCase = *(CR.Range.second-1);
// Get the MachineFunction which holds the current MBB. This is used when
// inserting any additional MBBs necessary to represent the switch.
MachineFunction *CurMF = FuncInfo.MF;
// If target does not have legal shift left, do not emit bit tests at all.
if (!TLI.isOperationLegal(ISD::SHL, TLI.getPointerTy()))
return false;
size_t numCmps = 0;
for (CaseItr I = CR.Range.first, E = CR.Range.second;
I!=E; ++I) {
// Single case counts one, case range - two.
numCmps += (I->Low == I->High ? 1 : 2);
}
// Count unique destinations
SmallSet<MachineBasicBlock*, 4> Dests;
for (CaseItr I = CR.Range.first, E = CR.Range.second; I!=E; ++I) {
Dests.insert(I->BB);
if (Dests.size() > 3)
// Don't bother the code below, if there are too much unique destinations
return false;
}
DEBUG(dbgs() << "Total number of unique destinations: "
<< Dests.size() << '\n'
<< "Total number of comparisons: " << numCmps << '\n');
// Compute span of values.
const APInt& minValue = cast<ConstantInt>(FrontCase.Low)->getValue();
const APInt& maxValue = cast<ConstantInt>(BackCase.High)->getValue();