#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/CodeGen/PseudoSourceValue.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Analysis/DebugInfo.h"
-#include "llvm/Target/TargetRegisterInfo.h"
#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetFrameInfo.h"
+#include "llvm/Target/TargetFrameLowering.h"
#include "llvm/Target/TargetInstrInfo.h"
#include "llvm/Target/TargetIntrinsicInfo.h"
#include "llvm/Target/TargetLowering.h"
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 cl::opt<unsigned>
+MaxParallelChains("dag-chain-limit", cl::desc("Max parallel isel dag chains"),
+ cl::init(64), cl::Hidden);
+
+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,
+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.isVector() && ValueVT.isInteger()) {
+ if (ValueVT.isInteger()) {
unsigned PartBits = PartVT.getSizeInBits();
unsigned ValueBits = ValueVT.getSizeInBits();
EVT HalfVT = EVT::getIntegerVT(*DAG.getContext(), RoundBits/2);
if (RoundParts > 2) {
- Lo = getCopyFromParts(DAG, dl, Parts, RoundParts / 2,
+ Lo = getCopyFromParts(DAG, DL, Parts, RoundParts / 2,
PartVT, HalfVT);
- Hi = getCopyFromParts(DAG, dl, Parts + RoundParts / 2,
+ Hi = getCopyFromParts(DAG, DL, Parts + RoundParts / 2,
RoundParts / 2, PartVT, HalfVT);
} else {
- Lo = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[0]);
- Hi = DAG.getNode(ISD::BIT_CONVERT, dl, HalfVT, Parts[1]);
+ 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);
+ 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,
+ Hi = getCopyFromParts(DAG, DL,
Parts + RoundParts, OddParts, PartVT, OddVT);
// Combine the round and odd parts.
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,
+ 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 (ValueVT.isVector()) {
- // Handle a multi-element vector.
- 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);
+ Lo = DAG.getNode(ISD::ZERO_EXTEND, DL, TotalVT, Lo);
+ Val = DAG.getNode(ISD::OR, DL, TotalVT, Lo, Hi);
}
-
- // 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);
} 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::BIT_CONVERT, dl, EVT(MVT::f64), Parts[0]);
- Hi = DAG.getNode(ISD::BIT_CONVERT, dl, EVT(MVT::f64), Parts[1]);
+ 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);
+ 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);
+ Val = getCopyFromParts(DAG, DL, Parts, NumParts, PartVT, IntVT);
}
}
if (PartVT == ValueVT)
return Val;
- if (PartVT.isVector()) {
- assert(ValueVT.isVector() && "Unknown vector conversion!");
- return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
- }
-
- if (ValueVT.isVector()) {
- assert(ValueVT.getVectorElementType() == PartVT &&
- ValueVT.getVectorNumElements() == 1 &&
- "Only trivial scalar-to-vector conversions should get here!");
- return DAG.getNode(ISD::BUILD_VECTOR, dl, ValueVT, Val);
- }
-
- if (PartVT.isInteger() &&
- ValueVT.isInteger()) {
+ 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,
+ Val = DAG.getNode(AssertOp, DL, PartVT, Val,
DAG.getValueType(ValueVT));
- return DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
- } else {
- return DAG.getNode(ISD::ANY_EXTEND, dl, ValueVT, Val);
+ return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
}
+ return DAG.getNode(ISD::ANY_EXTEND, DL, ValueVT, Val);
}
if (PartVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
- if (ValueVT.bitsLT(Val.getValueType())) {
- // FP_ROUND's are always exact here.
- return DAG.getNode(ISD::FP_ROUND, dl, ValueVT, Val,
+ // FP_ROUND's are always exact here.
+ if (ValueVT.bitsLT(Val.getValueType()))
+ return DAG.getNode(ISD::FP_ROUND, DL, ValueVT, Val,
DAG.getIntPtrConstant(1));
- }
- return DAG.getNode(ISD::FP_EXTEND, dl, ValueVT, Val);
+ return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
}
if (PartVT.getSizeInBits() == ValueVT.getSizeInBits())
- return DAG.getNode(ISD::BIT_CONVERT, dl, ValueVT, Val);
+ return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
llvm_unreachable("Unknown mismatch!");
return SDValue();
}
+/// 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.
+ return DAG.getNode(ISD::BITCAST, DL, ValueVT, Val);
+ }
+
+ assert(ValueVT.getVectorElementType() == PartVT &&
+ ValueVT.getVectorNumElements() == 1 &&
+ "Only trivial scalar-to-vector conversions should get here!");
+ 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,
+static void getCopyToParts(SelectionDAG &DAG, DebugLoc DL,
SDValue Val, SDValue *Parts, unsigned NumParts,
EVT PartVT,
ISD::NodeType ExtendKind = ISD::ANY_EXTEND) {
- const TargetLowering &TLI = DAG.getTargetLoweringInfo();
- EVT PtrVT = TLI.getPointerTy();
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)
+ if (NumParts == 0)
return;
- if (!ValueVT.isVector()) {
- 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 if (PartVT.isInteger() && ValueVT.isInteger()) {
- ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
- Val = DAG.getNode(ExtendKind, dl, ValueVT, Val);
- } else {
- llvm_unreachable("Unknown mismatch!");
- }
- } else if (PartBits == ValueVT.getSizeInBits()) {
- // Different types of the same size.
- assert(NumParts == 1 && PartVT != ValueVT);
- Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
- } else if (NumParts * PartBits < ValueVT.getSizeInBits()) {
- // If the parts cover less bits than value has, truncate the value.
- if (PartVT.isInteger() && ValueVT.isInteger()) {
- ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
- Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
- } else {
- llvm_unreachable("Unknown mismatch!");
- }
- }
-
- // 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;
- }
+ assert(!ValueVT.isVector() && "Vector case handled elsewhere");
+ if (PartVT == ValueVT) {
+ assert(NumParts == 1 && "No-op copy with multiple parts!");
+ 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.
+ 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() && 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.getConstant(RoundBits,
- TLI.getPointerTy()));
- 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;
+ "Unknown mismatch!");
ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
- Val = DAG.getNode(ISD::TRUNCATE, dl, ValueVT, Val);
+ Val = DAG.getNode(ExtendKind, DL, ValueVT, 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() && ValueVT.isInteger() &&
+ "Unknown mismatch!");
+ ValueVT = EVT::getIntegerVT(*DAG.getContext(), NumParts * PartBits);
+ Val = DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
+ }
+
+ // The value may have changed - recompute ValueVT.
+ ValueVT = Val.getValueType();
+ assert(NumParts * PartBits == ValueVT.getSizeInBits() &&
+ "Failed to tile the value with PartVT!");
- // The number of parts is a power of 2. Repeatedly bisect the value using
- // EXTRACT_ELEMENT.
- Parts[0] = DAG.getNode(ISD::BIT_CONVERT, 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.getConstant(1, PtrVT));
- Part0 = DAG.getNode(ISD::EXTRACT_ELEMENT, dl,
- ThisVT, Part0,
- DAG.getConstant(0, PtrVT));
-
- if (ThisBits == PartBits && ThisVT != PartVT) {
- Part0 = DAG.getNode(ISD::BIT_CONVERT, dl,
- PartVT, Part0);
- Part1 = DAG.getNode(ISD::BIT_CONVERT, dl,
- PartVT, Part1);
- }
+ 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);
+ if (TLI.isBigEndian())
+ std::reverse(Parts, Parts + OrigNumParts);
+}
- return;
- }
- // Vector ValueVT.
+/// 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) {
- if (PartVT.getSizeInBits() == ValueVT.getSizeInBits()) {
- Val = DAG.getNode(ISD::BIT_CONVERT, dl, PartVT, Val);
- } else {
- assert(ValueVT.getVectorElementType() == PartVT &&
- ValueVT.getVectorNumElements() == 1 &&
- "Only trivial vector-to-scalar conversions should get here!");
- Val = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
- PartVT, Val,
- DAG.getConstant(0, PtrVT));
- }
+ 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 {
+ // Vector -> scalar conversion.
+ assert(ValueVT.getVectorElementType() == PartVT &&
+ 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));
}
Parts[0] = Val;
EVT IntermediateVT, RegisterVT;
unsigned NumIntermediates;
unsigned NumRegs = TLI.getVectorTypeBreakdown(*DAG.getContext(), ValueVT,
- IntermediateVT, NumIntermediates, RegisterVT);
+ IntermediateVT,
+ NumIntermediates, RegisterVT);
unsigned NumElements = ValueVT.getVectorNumElements();
assert(NumRegs == NumParts && "Part count doesn't match vector breakdown!");
SmallVector<SDValue, 8> Ops(NumIntermediates);
for (unsigned i = 0; i != NumIntermediates; ++i) {
if (IntermediateVT.isVector())
- Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl,
+ Ops[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL,
IntermediateVT, Val,
- DAG.getConstant(i * (NumElements / NumIntermediates),
- PtrVT));
+ DAG.getIntPtrConstant(i * (NumElements / NumIntermediates)));
else
- Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
- IntermediateVT, Val,
- DAG.getConstant(i, PtrVT));
+ Ops[i] = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
+ IntermediateVT, Val, DAG.getIntPtrConstant(i));
}
// Split the intermediate operands into legal parts.
// 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);
+ 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.
"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);
+ 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
EVT regvt, EVT valuevt)
: ValueVTs(1, valuevt), RegVTs(1, regvt), Regs(regs) {}
- RegsForValue(const SmallVector<unsigned, 4> ®s,
- const SmallVector<EVT, 4> ®vts,
- const SmallVector<EVT, 4> &valuevts)
- : ValueVTs(valuevts), RegVTs(regvts), Regs(regs) {}
-
RegsForValue(LLVMContext &Context, const TargetLowering &tli,
unsigned Reg, const Type *Ty) {
ComputeValueVTs(tli, Ty, ValueVTs);
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.
}
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()) {
- unsigned SlotNo = Regs[Part+i]-TargetRegisterInfo::FirstVirtualRegister;
- if (FuncInfo.LiveOutRegInfo.size() > SlotNo) {
- const FunctionLoweringInfo::LiveOutInfo &LOI =
- FuncInfo.LiveOutRegInfo[SlotNo];
-
- 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
-
- if (FromVT != MVT::Other)
- P = DAG.getNode(isSExt ? ISD::AssertSext : ISD::AssertZext, dl,
- RegisterVT, P, DAG.getValueType(FromVT));
- }
- }
+ if (!TargetRegisterInfo::isVirtualRegister(Regs[Part+i]) ||
+ !RegisterVT.isInteger() || RegisterVT.isVector() ||
+ !FuncInfo.LiveOutRegInfo.inBounds(Regs[Part+i]))
+ continue;
+
+ const FunctionLoweringInfo::LiveOutInfo &LOI =
+ FuncInfo.LiveOutRegInfo[Regs[Part+i]];
+
+ 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;
- Parts[i] = P;
+ // 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(),
unsigned NumParts = TLI.getNumRegisters(*DAG.getContext(), ValueVT);
EVT RegisterVT = RegVTs[Value];
- getCopyToParts(DAG, dl,
- Val.getValue(Val.getResNo() + Value),
+ getCopyToParts(DAG, dl, Val.getValue(Val.getResNo() + Value),
&Parts[Part], NumParts, RegisterVT);
Part += NumParts;
}
UnusedArgNodeMap.clear();
PendingLoads.clear();
PendingExports.clear();
+ DanglingDebugInfoMap.clear();
CurDebugLoc = DebugLoc();
HasTailCall = false;
}
}
}
+// 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);
+ 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
unsigned InReg = It->second;
RegsForValue RFV(*DAG.getContext(), TLI, InReg, V->getType());
SDValue Chain = DAG.getEntryNode();
- return N = RFV.getCopyFromRegs(DAG, FuncInfo, getCurDebugLoc(), Chain, NULL);
+ 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;
}
// 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 getMaterializedValue.
+/// 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)) {
unsigned NumValues = ValueVTs.size();
SmallVector<SDValue, 4> Chains(NumValues);
- EVT PtrVT = PtrValueVTs[0];
for (unsigned i = 0; i != NumValues; ++i) {
- SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, RetPtr,
- DAG.getConstant(Offsets[i], PtrVT));
+ 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),
- Add, NULL, Offsets[i], false, false, 0);
+ // FIXME: better loc info would be nice.
+ Add, MachinePointerInfo(), false, false, 0);
}
Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
if (Cases[0].CC == ISD::SETNE && Cases[0].FalseBB == Cases[1].ThisBB)
return false;
}
-
+
return true;
}
// 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
// jle foo
//
if (const BinaryOperator *BOp = dyn_cast<BinaryOperator>(CondVal)) {
- if (BOp->hasOneUse() &&
+ if (!TLI.isJumpExpensive() &&
+ BOp->hasOneUse() &&
(BOp->getOpcode() == Instruction::And ||
BOp->getOpcode() == Instruction::Or)) {
FindMergedConditions(BOp, Succ0MBB, Succ1MBB, BrMBB, BrMBB,
MVT::Other, getControlRoot(), Cond,
DAG.getBasicBlock(CB.TrueBB));
- // Insert the false branch.
- if (CB.FalseBB != NextBlock)
- BrCond = DAG.getNode(ISD::BR, dl, MVT::Other, BrCond,
- DAG.getBasicBlock(CB.FalseBB));
+ // 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);
}
Sub, DAG.getConstant(B.Range, VT),
ISD::SETUGT);
- SDValue ShiftOp = DAG.getZExtOrTrunc(Sub, getCurDebugLoc(),
- TLI.getPointerTy());
+ // 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 ((uint64_t)((int64_t)B.Cases[i].Mask >> VT.getSizeInBits()) + 1 >= 2) {
+ // 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.Reg = FuncInfo.CreateReg(TLI.getPointerTy());
+ B.RegVT = VT;
+ B.Reg = FuncInfo.CreateReg(VT);
SDValue CopyTo = DAG.getCopyToReg(getControlRoot(), getCurDebugLoc(),
- B.Reg, ShiftOp);
+ 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.
}
/// visitBitTestCase - this function produces one "bit test"
-void SelectionDAGBuilder::visitBitTestCase(MachineBasicBlock* NextMBB,
+void SelectionDAGBuilder::visitBitTestCase(BitTestBlock &BB,
+ MachineBasicBlock* NextMBB,
unsigned Reg,
BitTestCase &B,
MachineBasicBlock *SwitchBB) {
- SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(), Reg,
- TLI.getPointerTy());
+ EVT VT = BB.RegVT;
+ SDValue ShiftOp = DAG.getCopyFromReg(getControlRoot(), getCurDebugLoc(),
+ Reg, VT);
SDValue Cmp;
if (CountPopulation_64(B.Mask) == 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(ShiftOp.getValueType()),
+ TLI.getSetCCResultType(VT),
ShiftOp,
- DAG.getConstant(CountTrailingZeros_64(B.Mask),
- TLI.getPointerTy()),
+ DAG.getConstant(CountTrailingZeros_64(B.Mask), VT),
ISD::SETEQ);
} else {
// Make desired shift
- SDValue SwitchVal = DAG.getNode(ISD::SHL, getCurDebugLoc(),
- TLI.getPointerTy(),
- DAG.getConstant(1, TLI.getPointerTy()),
- ShiftOp);
+ 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(),
- TLI.getPointerTy(), SwitchVal,
- DAG.getConstant(B.Mask, TLI.getPointerTy()));
+ VT, SwitchVal, DAG.getConstant(B.Mask, VT));
Cmp = DAG.getSetCC(getCurDebugLoc(),
- TLI.getSetCCResultType(AndOp.getValueType()),
- AndOp, DAG.getConstant(0, TLI.getPointerTy()),
+ TLI.getSetCCResultType(VT),
+ AndOp, DAG.getConstant(0, VT),
ISD::SETNE);
}
if (++BBI != FuncInfo.MF->end())
NextBlock = BBI;
- // TODO: If any two of the cases has the same destination, and if one value
+ // 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.
+ SwitchBB->addSuccessor(Small.BB);
+ SwitchBB->addSuccessor(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) {
}
static APInt ComputeRange(const APInt &First, const APInt &Last) {
- APInt LastExt(Last), FirstExt(First);
uint32_t BitWidth = std::max(Last.getBitWidth(), First.getBitWidth()) + 1;
- LastExt.sext(BitWidth); FirstExt.sext(BitWidth);
+ APInt LastExt = Last.sext(BitWidth), FirstExt = First.sext(BitWidth);
return (LastExt - FirstExt + 1ULL);
}
}
BitTestBlock BTB(lowBound, cmpRange, SV,
- -1U, (CR.CaseBB == SwitchBB),
+ -1U, MVT::Other, (CR.CaseBB == SwitchBB),
CR.CaseBB, Default, BTC);
if (CR.CaseBB == SwitchBB)
if (Cases.size() >= 2)
// Must recompute end() each iteration because it may be
// invalidated by erase if we hold on to it
- for (CaseItr I = Cases.begin(), J = ++(Cases.begin()); J != Cases.end(); ) {
+ for (CaseItr I = Cases.begin(), J = llvm::next(Cases.begin());
+ J != Cases.end(); ) {
const APInt& nextValue = cast<ConstantInt>(J->Low)->getValue();
const APInt& currentValue = cast<ConstantInt>(I->High)->getValue();
MachineBasicBlock* nextBB = J->BB;
return numCmps;
}
+void SelectionDAGBuilder::UpdateSplitBlock(MachineBasicBlock *First,
+ MachineBasicBlock *Last) {
+ // Update JTCases.
+ for (unsigned i = 0, e = JTCases.size(); i != e; ++i)
+ if (JTCases[i].first.HeaderBB == First)
+ JTCases[i].first.HeaderBB = Last;
+
+ // Update BitTestCases.
+ for (unsigned i = 0, e = BitTestCases.size(); i != e; ++i)
+ if (BitTestCases[i].Parent == First)
+ BitTestCases[i].Parent = Last;
+}
+
void SelectionDAGBuilder::visitSwitch(const SwitchInst &SI) {
MachineBasicBlock *SwitchMBB = FuncInfo.MBB;
void SelectionDAGBuilder::visitFSub(const User &I) {
// -0.0 - X --> fneg
const Type *Ty = I.getType();
- if (Ty->isVectorTy()) {
- if (ConstantVector *CV = dyn_cast<ConstantVector>(I.getOperand(0))) {
- const VectorType *DestTy = cast<VectorType>(I.getType());
- const Type *ElTy = DestTy->getElementType();
- unsigned VL = DestTy->getNumElements();
- std::vector<Constant*> NZ(VL, ConstantFP::getNegativeZero(ElTy));
- Constant *CNZ = ConstantVector::get(&NZ[0], NZ.size());
- if (CV == CNZ) {
- SDValue Op2 = getValue(I.getOperand(1));
- setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
- Op2.getValueType(), Op2));
- return;
- }
- }
+ if (isa<Constant>(I.getOperand(0)) &&
+ I.getOperand(0) == ConstantFP::getZeroValueForNegation(Ty)) {
+ SDValue Op2 = getValue(I.getOperand(1));
+ setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
+ Op2.getValueType(), Op2));
+ return;
}
- if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0)))
- if (CFP->isExactlyValue(ConstantFP::getNegativeZero(Ty)->getValueAPF())) {
- SDValue Op2 = getValue(I.getOperand(1));
- setValue(&I, DAG.getNode(ISD::FNEG, getCurDebugLoc(),
- Op2.getValueType(), Op2));
- return;
- }
-
visitBinary(I, ISD::FSUB);
}
void SelectionDAGBuilder::visitShift(const User &I, unsigned Opcode) {
SDValue Op1 = getValue(I.getOperand(0));
SDValue Op2 = getValue(I.getOperand(1));
- if (!I.getType()->isVectorTy() &&
- Op2.getValueType() != TLI.getShiftAmountTy()) {
+
+ MVT ShiftTy = TLI.getShiftAmountTy();
+
+ // Coerce the shift amount to the right type if we can.
+ if (!I.getType()->isVectorTy() && Op2.getValueType() != ShiftTy) {
+ unsigned ShiftSize = ShiftTy.getSizeInBits();
+ unsigned Op2Size = Op2.getValueType().getSizeInBits();
+ DebugLoc DL = getCurDebugLoc();
+
// If the operand is smaller than the shift count type, promote it.
- EVT PTy = TLI.getPointerTy();
- EVT STy = TLI.getShiftAmountTy();
- if (STy.bitsGT(Op2.getValueType()))
- Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
- TLI.getShiftAmountTy(), Op2);
+ if (ShiftSize > Op2Size)
+ Op2 = DAG.getNode(ISD::ZERO_EXTEND, DL, ShiftTy, Op2);
+
// If the operand is larger than the shift count type but the shift
// count type has enough bits to represent any shift value, truncate
// it now. This is a common case and it exposes the truncate to
// optimization early.
- else if (STy.getSizeInBits() >=
- Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
- Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
- TLI.getShiftAmountTy(), Op2);
- // Otherwise we'll need to temporarily settle for some other
- // convenient type; type legalization will make adjustments as
- // needed.
- else if (PTy.bitsLT(Op2.getValueType()))
- Op2 = DAG.getNode(ISD::TRUNCATE, getCurDebugLoc(),
- TLI.getPointerTy(), Op2);
- else if (PTy.bitsGT(Op2.getValueType()))
- Op2 = DAG.getNode(ISD::ANY_EXTEND, getCurDebugLoc(),
- TLI.getPointerTy(), Op2);
+ else if (ShiftSize >= Log2_32_Ceil(Op2.getValueType().getSizeInBits()))
+ Op2 = DAG.getNode(ISD::TRUNCATE, DL, ShiftTy, Op2);
+ // Otherwise we'll need to temporarily settle for some other convenient
+ // type. Type legalization will make adjustments once the shiftee is split.
+ else
+ Op2 = DAG.getZExtOrTrunc(Op2, DL, MVT::i32);
}
setValue(&I, DAG.getNode(Opcode, getCurDebugLoc(),
EVT DestVT = TLI.getValueType(I.getType());
// BitCast assures us that source and destination are the same size so this is
- // either a BIT_CONVERT or a no-op.
+ // either a BITCAST or a no-op.
if (DestVT != N.getValueType())
- setValue(&I, DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
+ setValue(&I, DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
DestVT, N)); // convert types.
else
setValue(&I, N); // noop cast.
} else {
StartIdx[Input] = (MinRange[Input]/MaskNumElts)*MaskNumElts;
if (MaxRange[Input] - StartIdx[Input] < (int)MaskNumElts &&
- StartIdx[Input] + MaskNumElts < SrcNumElts)
+ StartIdx[Input] + MaskNumElts <= SrcNumElts)
RangeUse[Input] = 1; // Extract from a multiple of the mask length.
}
}
bool IntoUndef = isa<UndefValue>(Op0);
bool FromUndef = isa<UndefValue>(Op1);
- unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
- I.idx_begin(), I.idx_end());
+ unsigned LinearIndex = ComputeLinearIndex(AggTy, I.idx_begin(), I.idx_end());
SmallVector<EVT, 4> AggValueVTs;
ComputeValueVTs(TLI, AggTy, AggValueVTs);
const Type *ValTy = I.getType();
bool OutOfUndef = isa<UndefValue>(Op0);
- unsigned LinearIndex = ComputeLinearIndex(TLI, AggTy,
- I.idx_begin(), I.idx_end());
+ unsigned LinearIndex = ComputeLinearIndex(AggTy, I.idx_begin(), I.idx_end());
SmallVector<EVT, 4> ValValueVTs;
ComputeValueVTs(TLI, ValTy, ValValueVTs);
}
Ty = StTy->getElementType(Field);
- } else if (const UnionType *UnTy = dyn_cast<UnionType>(Ty)) {
- unsigned Field = cast<ConstantInt>(Idx)->getZExtValue();
-
- // Offset canonically 0 for unions, but type changes
- Ty = UnTy->getElementType(Field);
} else {
Ty = cast<SequentialType>(Ty)->getElementType();
// Handle alignment. If the requested alignment is less than or equal to
// the stack alignment, ignore it. If the size is greater than or equal to
// the stack alignment, we note this in the DYNAMIC_STACKALLOC node.
- unsigned StackAlign = TM.getFrameInfo()->getStackAlignment();
+ unsigned StackAlign = TM.getFrameLowering()->getStackAlignment();
if (Align <= StackAlign)
Align = 0;
// Inform the Frame Information that we have just allocated a variable-sized
// object.
- FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject();
+ FuncInfo.MF->getFrameInfo()->CreateVariableSizedObject(Align ? Align : 1);
}
void SelectionDAGBuilder::visitLoad(const LoadInst &I) {
bool isVolatile = I.isVolatile();
bool isNonTemporal = I.getMetadata("nontemporal") != 0;
unsigned Alignment = I.getAlignment();
+ const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
SmallVector<EVT, 4> ValueVTs;
SmallVector<uint64_t, 4> Offsets;
SDValue Root;
bool ConstantMemory = false;
- if (I.isVolatile())
+ if (I.isVolatile() || NumValues > MaxParallelChains)
// Serialize volatile loads with other side effects.
Root = getRoot();
- else if (AA->pointsToConstantMemory(SV)) {
+ else if (AA->pointsToConstantMemory(
+ AliasAnalysis::Location(SV, AA->getTypeStoreSize(Ty), TBAAInfo))) {
// Do not serialize (non-volatile) loads of constant memory with anything.
Root = DAG.getEntryNode();
ConstantMemory = true;
}
SmallVector<SDValue, 4> Values(NumValues);
- SmallVector<SDValue, 4> Chains(NumValues);
+ SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
+ NumValues));
EVT PtrVT = Ptr.getValueType();
- for (unsigned i = 0; i != NumValues; ++i) {
+ unsigned ChainI = 0;
+ for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
+ // Serializing loads here may result in excessive register pressure, and
+ // TokenFactor places arbitrary choke points on the scheduler. SD scheduling
+ // could recover a bit by hoisting nodes upward in the chain by recognizing
+ // they are side-effect free or do not alias. The optimizer should really
+ // avoid this case by converting large object/array copies to llvm.memcpy
+ // (MaxParallelChains should always remain as failsafe).
+ if (ChainI == MaxParallelChains) {
+ assert(PendingLoads.empty() && "PendingLoads must be serialized first");
+ SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
+ MVT::Other, &Chains[0], ChainI);
+ Root = Chain;
+ ChainI = 0;
+ }
SDValue A = DAG.getNode(ISD::ADD, getCurDebugLoc(),
PtrVT, Ptr,
DAG.getConstant(Offsets[i], PtrVT));
SDValue L = DAG.getLoad(ValueVTs[i], getCurDebugLoc(), Root,
- A, SV, Offsets[i], isVolatile,
- isNonTemporal, Alignment);
+ A, MachinePointerInfo(SV, Offsets[i]), isVolatile,
+ isNonTemporal, Alignment, TBAAInfo);
Values[i] = L;
- Chains[i] = L.getValue(1);
+ Chains[ChainI] = L.getValue(1);
}
if (!ConstantMemory) {
SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
- MVT::Other, &Chains[0], NumValues);
+ MVT::Other, &Chains[0], ChainI);
if (isVolatile)
DAG.setRoot(Chain);
else
SDValue Ptr = getValue(PtrV);
SDValue Root = getRoot();
- SmallVector<SDValue, 4> Chains(NumValues);
+ SmallVector<SDValue, 4> Chains(std::min(unsigned(MaxParallelChains),
+ NumValues));
EVT PtrVT = Ptr.getValueType();
bool isVolatile = I.isVolatile();
bool isNonTemporal = I.getMetadata("nontemporal") != 0;
unsigned Alignment = I.getAlignment();
-
- for (unsigned i = 0; i != NumValues; ++i) {
+ const MDNode *TBAAInfo = I.getMetadata(LLVMContext::MD_tbaa);
+
+ unsigned ChainI = 0;
+ for (unsigned i = 0; i != NumValues; ++i, ++ChainI) {
+ // See visitLoad comments.
+ if (ChainI == MaxParallelChains) {
+ SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
+ MVT::Other, &Chains[0], ChainI);
+ Root = Chain;
+ ChainI = 0;
+ }
SDValue Add = DAG.getNode(ISD::ADD, getCurDebugLoc(), PtrVT, Ptr,
DAG.getConstant(Offsets[i], PtrVT));
- Chains[i] = DAG.getStore(Root, getCurDebugLoc(),
- SDValue(Src.getNode(), Src.getResNo() + i),
- Add, PtrV, Offsets[i], isVolatile,
- isNonTemporal, Alignment);
- }
-
- DAG.setRoot(DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
- MVT::Other, &Chains[0], NumValues));
+ SDValue St = DAG.getStore(Root, getCurDebugLoc(),
+ SDValue(Src.getNode(), Src.getResNo() + i),
+ Add, MachinePointerInfo(PtrV, Offsets[i]),
+ isVolatile, isNonTemporal, Alignment, TBAAInfo);
+ Chains[ChainI] = St;
+ }
+
+ SDValue StoreNode = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
+ MVT::Other, &Chains[0], ChainI);
+ ++SDNodeOrder;
+ AssignOrderingToNode(StoreNode.getNode());
+ DAG.setRoot(StoreNode);
}
/// visitTargetIntrinsic - Lower a call of a target intrinsic to an INTRINSIC
bool IsTgtIntrinsic = TLI.getTgtMemIntrinsic(Info, I, Intrinsic);
// Add the intrinsic ID as an integer operand if it's not a target intrinsic.
- if (!IsTgtIntrinsic)
+ if (!IsTgtIntrinsic || Info.opc == ISD::INTRINSIC_VOID ||
+ Info.opc == ISD::INTRINSIC_W_CHAIN)
Ops.push_back(DAG.getConstant(Intrinsic, TLI.getPointerTy()));
// Add all operands of the call to the operand list.
// This is target intrinsic that touches memory
Result = DAG.getMemIntrinsicNode(Info.opc, getCurDebugLoc(),
VTs, &Ops[0], Ops.size(),
- Info.memVT, Info.ptrVal, Info.offset,
+ Info.memVT,
+ MachinePointerInfo(Info.ptrVal, Info.offset),
Info.align, Info.vol,
Info.readMem, Info.writeMem);
} else if (!HasChain) {
if (!I.getType()->isVoidTy()) {
if (const VectorType *PTy = dyn_cast<VectorType>(I.getType())) {
EVT VT = TLI.getValueType(PTy);
- Result = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(), VT, Result);
+ Result = DAG.getNode(ISD::BITCAST, getCurDebugLoc(), VT, Result);
}
setValue(&I, Result);
DAG.getConstant(0x007fffff, MVT::i32));
SDValue t2 = DAG.getNode(ISD::OR, dl, MVT::i32, t1,
DAG.getConstant(0x3f800000, MVT::i32));
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t2);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::f32, t2);
}
/// GetExponent - Get the exponent:
SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
getF32Constant(DAG, 0x3f7f5e7e));
- SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t5);
+ SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t5);
// Add the exponent into the result in integer domain.
SDValue t6 = DAG.getNode(ISD::ADD, dl, MVT::i32,
TwoToFracPartOfX, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t6);
+ result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t6);
} else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
// For floating-point precision of 12:
//
SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
getF32Constant(DAG, 0x3f7ff8fd));
- SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,MVT::i32, t7);
+ SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,MVT::i32, t7);
// Add the exponent into the result in integer domain.
SDValue t8 = DAG.getNode(ISD::ADD, dl, MVT::i32,
TwoToFracPartOfX, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t8);
+ result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t8);
} else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
// For floating-point precision of 18:
//
SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
getF32Constant(DAG, 0x3f800000));
- SDValue TwoToFracPartOfX = DAG.getNode(ISD::BIT_CONVERT, dl,
+ SDValue TwoToFracPartOfX = DAG.getNode(ISD::BITCAST, dl,
MVT::i32, t13);
// Add the exponent into the result in integer domain.
SDValue t14 = DAG.getNode(ISD::ADD, dl, MVT::i32,
TwoToFracPartOfX, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, t14);
+ result = DAG.getNode(ISD::BITCAST, dl, MVT::f32, t14);
}
} else {
// No special expansion.
if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
SDValue Op = getValue(I.getArgOperand(0));
- SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
+ SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
// Scale the exponent by log(2) [0.69314718f].
SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
SDValue Op = getValue(I.getArgOperand(0));
- SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
+ SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
// Get the exponent.
SDValue LogOfExponent = GetExponent(DAG, Op1, TLI, dl);
if (getValue(I.getArgOperand(0)).getValueType() == MVT::f32 &&
LimitFloatPrecision > 0 && LimitFloatPrecision <= 18) {
SDValue Op = getValue(I.getArgOperand(0));
- SDValue Op1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, Op);
+ SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
// Scale the exponent by log10(2) [0.30102999f].
SDValue Exp = GetExponent(DAG, Op1, TLI, dl);
SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
getF32Constant(DAG, 0x3f7f5e7e));
- SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
+ SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
} else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
// For floating-point precision of 12:
SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
getF32Constant(DAG, 0x3f7ff8fd));
- SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
+ SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
} else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
// For floating-point precision of 18:
SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
getF32Constant(DAG, 0x3f800000));
- SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
+ SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
}
} else {
SDValue t4 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t3, X);
SDValue t5 = DAG.getNode(ISD::FADD, dl, MVT::f32, t4,
getF32Constant(DAG, 0x3f7f5e7e));
- SDValue t6 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t5);
+ SDValue t6 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t5);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t6, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
} else if (LimitFloatPrecision > 6 && LimitFloatPrecision <= 12) {
// For floating-point precision of 12:
SDValue t6 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t5, X);
SDValue t7 = DAG.getNode(ISD::FADD, dl, MVT::f32, t6,
getF32Constant(DAG, 0x3f7ff8fd));
- SDValue t8 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t7);
+ SDValue t8 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t7);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t8, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
} else { // LimitFloatPrecision > 12 && LimitFloatPrecision <= 18
// For floating-point precision of 18:
SDValue t12 = DAG.getNode(ISD::FMUL, dl, MVT::f32, t11, X);
SDValue t13 = DAG.getNode(ISD::FADD, dl, MVT::f32, t12,
getF32Constant(DAG, 0x3f800000));
- SDValue t14 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i32, t13);
+ SDValue t14 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, t13);
SDValue TwoToFractionalPartOfX =
DAG.getNode(ISD::ADD, dl, MVT::i32, t14, IntegerPartOfX);
- result = DAG.getNode(ISD::BIT_CONVERT, dl,
+ result = DAG.getNode(ISD::BITCAST, dl,
MVT::f32, TwoToFractionalPartOfX);
}
} else {
/// argument, create the corresponding DBG_VALUE machine instruction for it now.
/// At the end of instruction selection, they will be inserted to the entry BB.
bool
-SelectionDAGBuilder::EmitFuncArgumentDbgValue(const DbgValueInst &DI,
- const Value *V, MDNode *Variable,
- uint64_t Offset,
+SelectionDAGBuilder::EmitFuncArgumentDbgValue(const Value *V, MDNode *Variable,
+ int64_t Offset,
const SDValue &N) {
- if (!isa<Argument>(V))
+ const Argument *Arg = dyn_cast<Argument>(V);
+ if (!Arg)
return false;
MachineFunction &MF = DAG.getMachineFunction();
+ const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
+ const TargetRegisterInfo *TRI = DAG.getTarget().getRegisterInfo();
+
// Ignore inlined function arguments here.
DIVariable DV(Variable);
if (DV.isInlinedFnArgument(MF.getFunction()))
return false;
unsigned Reg = 0;
- if (N.getOpcode() == ISD::CopyFromReg) {
+ if (Arg->hasByValAttr()) {
+ // Byval arguments' frame index is recorded during argument lowering.
+ // Use this info directly.
+ Reg = TRI->getFrameRegister(MF);
+ Offset = FuncInfo.getByValArgumentFrameIndex(Arg);
+ // If byval argument ofset is not recorded then ignore this.
+ if (!Offset)
+ Reg = 0;
+ }
+
+ if (N.getNode() && N.getOpcode() == ISD::CopyFromReg) {
Reg = cast<RegisterSDNode>(N.getOperand(1))->getReg();
- if (Reg && TargetRegisterInfo::isVirtualRegister(Reg)) {
+ if (TargetRegisterInfo::isVirtualRegister(Reg)) {
MachineRegisterInfo &RegInfo = MF.getRegInfo();
unsigned PR = RegInfo.getLiveInPhysReg(Reg);
if (PR)
}
if (!Reg) {
+ // Check if ValueMap has reg number.
DenseMap<const Value *, unsigned>::iterator VMI = FuncInfo.ValueMap.find(V);
- if (VMI == FuncInfo.ValueMap.end())
- return false;
- Reg = VMI->second;
+ if (VMI != FuncInfo.ValueMap.end())
+ Reg = VMI->second;
}
- const TargetInstrInfo *TII = DAG.getTarget().getInstrInfo();
+ if (!Reg && N.getNode()) {
+ // Check if frame index is available.
+ if (LoadSDNode *LNode = dyn_cast<LoadSDNode>(N.getNode()))
+ if (FrameIndexSDNode *FINode =
+ dyn_cast<FrameIndexSDNode>(LNode->getBasePtr().getNode())) {
+ Reg = TRI->getFrameRegister(MF);
+ Offset = FINode->getIndex();
+ }
+ }
+
+ if (!Reg)
+ return false;
+
MachineInstrBuilder MIB = BuildMI(MF, getCurDebugLoc(),
TII->get(TargetOpcode::DBG_VALUE))
.addReg(Reg, RegState::Debug).addImm(Offset).addMetadata(Variable);
}
// VisualStudio defines setjmp as _setjmp
-#if defined(_MSC_VER) && defined(setjmp)
-#define setjmp_undefined_for_visual_studio
-#undef setjmp
+#if defined(_MSC_VER) && defined(setjmp) && \
+ !defined(setjmp_undefined_for_msvc)
+# pragma push_macro("setjmp")
+# undef setjmp
+# define setjmp_undefined_for_msvc
#endif
/// visitIntrinsicCall - Lower the call to the specified intrinsic function. If
unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol, false,
- I.getArgOperand(0), 0, I.getArgOperand(1), 0));
+ MachinePointerInfo(I.getArgOperand(0)),
+ MachinePointerInfo(I.getArgOperand(1))));
return 0;
}
case Intrinsic::memset: {
unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
DAG.setRoot(DAG.getMemset(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
- I.getArgOperand(0), 0));
+ MachinePointerInfo(I.getArgOperand(0))));
return 0;
}
case Intrinsic::memmove: {
SDValue Op3 = getValue(I.getArgOperand(2));
unsigned Align = cast<ConstantInt>(I.getArgOperand(3))->getZExtValue();
bool isVol = cast<ConstantInt>(I.getArgOperand(4))->getZExtValue();
-
- // If the source and destination are known to not be aliases, we can
- // lower memmove as memcpy.
- uint64_t Size = -1ULL;
- if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op3))
- Size = C->getZExtValue();
- if (AA->alias(I.getArgOperand(0), Size, I.getArgOperand(1), Size) ==
- AliasAnalysis::NoAlias) {
- DAG.setRoot(DAG.getMemcpy(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
- false, I.getArgOperand(0), 0, I.getArgOperand(1), 0));
- return 0;
- }
-
DAG.setRoot(DAG.getMemmove(getRoot(), dl, Op1, Op2, Op3, Align, isVol,
- I.getArgOperand(0), 0, I.getArgOperand(1), 0));
+ MachinePointerInfo(I.getArgOperand(0)),
+ MachinePointerInfo(I.getArgOperand(1))));
return 0;
}
case Intrinsic::dbg_declare: {
const DbgDeclareInst &DI = cast<DbgDeclareInst>(I);
- if (!DIVariable(DI.getVariable()).Verify())
- return 0;
-
MDNode *Variable = DI.getVariable();
- // Parameters are handled specially.
- bool isParameter =
- DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
const Value *Address = DI.getAddress();
- if (!Address)
+ if (!Address || !DIVariable(DI.getVariable()).Verify())
return 0;
- if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
- Address = BCI->getOperand(0);
- const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
- if (AI) {
- // Don't handle byval arguments or VLAs, for example.
- // Non-byval arguments are handled here (they refer to the stack temporary
- // alloca at this point).
- DenseMap<const AllocaInst*, int>::iterator SI =
- FuncInfo.StaticAllocaMap.find(AI);
- if (SI == FuncInfo.StaticAllocaMap.end())
- return 0; // VLAs.
- int FI = SI->second;
-
- MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
- if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
- MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
- }
// Build an entry in DbgOrdering. Debug info input nodes get an SDNodeOrder
// but do not always have a corresponding SDNode built. The SDNodeOrder
// absolute, but not relative, values are different depending on whether
// debug info exists.
++SDNodeOrder;
+
+ // Check if address has undef value.
+ if (isa<UndefValue>(Address) ||
+ (Address->use_empty() && !isa<Argument>(Address))) {
+ DEBUG(dbgs() << "Dropping debug info for " << DI);
+ return 0;
+ }
+
SDValue &N = NodeMap[Address];
+ if (!N.getNode() && isa<Argument>(Address))
+ // Check unused arguments map.
+ N = UnusedArgNodeMap[Address];
SDDbgValue *SDV;
if (N.getNode()) {
+ // Parameters are handled specially.
+ bool isParameter =
+ DIVariable(Variable).getTag() == dwarf::DW_TAG_arg_variable;
+ if (const BitCastInst *BCI = dyn_cast<BitCastInst>(Address))
+ Address = BCI->getOperand(0);
+ const AllocaInst *AI = dyn_cast<AllocaInst>(Address);
+
if (isParameter && !AI) {
FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(N.getNode());
if (FINode)
// Byval parameter. We have a frame index at this point.
SDV = DAG.getDbgValue(Variable, FINode->getIndex(),
0, dl, SDNodeOrder);
- else
+ else {
// Can't do anything with other non-AI cases yet. This might be a
// parameter of a callee function that got inlined, for example.
+ DEBUG(dbgs() << "Dropping debug info for " << DI);
return 0;
+ }
} else if (AI)
SDV = DAG.getDbgValue(Variable, N.getNode(), N.getResNo(),
0, dl, SDNodeOrder);
- else
+ else {
// Can't do anything with other non-AI cases yet.
+ DEBUG(dbgs() << "Dropping debug info for " << DI);
return 0;
+ }
DAG.AddDbgValue(SDV, N.getNode(), isParameter);
} else {
- // This isn't useful, but it shows what we're missing.
- SDV = DAG.getDbgValue(Variable, UndefValue::get(Address->getType()),
- 0, dl, SDNodeOrder);
- DAG.AddDbgValue(SDV, 0, isParameter);
+ // If Address is an argument then try to emit its dbg value using
+ // virtual register info from the FuncInfo.ValueMap.
+ if (!EmitFuncArgumentDbgValue(Address, Variable, 0, N)) {
+ // If variable is pinned by a alloca in dominating bb then
+ // use StaticAllocaMap.
+ if (const AllocaInst *AI = dyn_cast<AllocaInst>(Address)) {
+ if (AI->getParent() != DI.getParent()) {
+ DenseMap<const AllocaInst*, int>::iterator SI =
+ FuncInfo.StaticAllocaMap.find(AI);
+ if (SI != FuncInfo.StaticAllocaMap.end()) {
+ SDV = DAG.getDbgValue(Variable, SI->second,
+ 0, dl, SDNodeOrder);
+ DAG.AddDbgValue(SDV, 0, false);
+ return 0;
+ }
+ }
+ }
+ DEBUG(dbgs() << "Dropping debug info for " << DI);
+ }
}
return 0;
}
SDV = DAG.getDbgValue(Variable, V, Offset, dl, SDNodeOrder);
DAG.AddDbgValue(SDV, 0, false);
} else {
- bool createUndef = false;
- // FIXME : Why not use getValue() directly ?
+ // Do not use getValue() in here; we don't want to generate code at
+ // this point if it hasn't been done yet.
SDValue N = NodeMap[V];
if (!N.getNode() && isa<Argument>(V))
// Check unused arguments map.
N = UnusedArgNodeMap[V];
if (N.getNode()) {
- if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) {
+ if (!EmitFuncArgumentDbgValue(V, Variable, Offset, N)) {
SDV = DAG.getDbgValue(Variable, N.getNode(),
N.getResNo(), Offset, dl, SDNodeOrder);
DAG.AddDbgValue(SDV, N.getNode(), false);
}
- } else if (isa<PHINode>(V) && !V->use_empty()) {
- SDValue N = getValue(V);
- if (N.getNode()) {
- if (!EmitFuncArgumentDbgValue(DI, V, Variable, Offset, N)) {
- SDV = DAG.getDbgValue(Variable, N.getNode(),
- N.getResNo(), Offset, dl, SDNodeOrder);
- DAG.AddDbgValue(SDV, N.getNode(), false);
- }
- } else
- createUndef = true;
- } else
- createUndef = true;
- if (createUndef) {
+ } else if (!V->use_empty() ) {
+ // Do not call getValue(V) yet, as we don't want to generate code.
+ // Remember it for later.
+ DanglingDebugInfo DDI(&DI, dl, SDNodeOrder);
+ DanglingDebugInfoMap[V] = DDI;
+ } else {
// We may expand this to cover more cases. One case where we have no
- // data available is an unreferenced parameter; we need this fallback.
- SDV = DAG.getDbgValue(Variable, UndefValue::get(V->getType()),
- Offset, dl, SDNodeOrder);
- DAG.AddDbgValue(SDV, 0, false);
+ // data available is an unreferenced parameter.
+ DEBUG(dbgs() << "Dropping debug info for " << DI);
}
}
if (SI == FuncInfo.StaticAllocaMap.end())
return 0; // VLAs.
int FI = SI->second;
-
+
MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
if (!DI.getDebugLoc().isUnknown() && MMI.hasDebugInfo())
MMI.setVariableDbgInfo(Variable, FI, DI.getDebugLoc());
}
case Intrinsic::eh_sjlj_longjmp: {
DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_LONGJMP, dl, MVT::Other,
- getRoot(),
- getValue(I.getArgOperand(0))));
+ getRoot(), getValue(I.getArgOperand(0))));
return 0;
}
+ case Intrinsic::eh_sjlj_dispatch_setup: {
+ DAG.setRoot(DAG.getNode(ISD::EH_SJLJ_DISPATCHSETUP, dl, MVT::Other,
+ getRoot(), getValue(I.getArgOperand(0))));
+ return 0;
+ }
+
+ case Intrinsic::x86_mmx_pslli_w:
+ case Intrinsic::x86_mmx_pslli_d:
+ case Intrinsic::x86_mmx_pslli_q:
+ case Intrinsic::x86_mmx_psrli_w:
+ case Intrinsic::x86_mmx_psrli_d:
+ case Intrinsic::x86_mmx_psrli_q:
+ case Intrinsic::x86_mmx_psrai_w:
+ case Intrinsic::x86_mmx_psrai_d: {
+ SDValue ShAmt = getValue(I.getArgOperand(1));
+ if (isa<ConstantSDNode>(ShAmt)) {
+ visitTargetIntrinsic(I, Intrinsic);
+ return 0;
+ }
+ unsigned NewIntrinsic = 0;
+ EVT ShAmtVT = MVT::v2i32;
+ switch (Intrinsic) {
+ case Intrinsic::x86_mmx_pslli_w:
+ NewIntrinsic = Intrinsic::x86_mmx_psll_w;
+ break;
+ case Intrinsic::x86_mmx_pslli_d:
+ NewIntrinsic = Intrinsic::x86_mmx_psll_d;
+ break;
+ case Intrinsic::x86_mmx_pslli_q:
+ NewIntrinsic = Intrinsic::x86_mmx_psll_q;
+ break;
+ case Intrinsic::x86_mmx_psrli_w:
+ NewIntrinsic = Intrinsic::x86_mmx_psrl_w;
+ break;
+ case Intrinsic::x86_mmx_psrli_d:
+ NewIntrinsic = Intrinsic::x86_mmx_psrl_d;
+ break;
+ case Intrinsic::x86_mmx_psrli_q:
+ NewIntrinsic = Intrinsic::x86_mmx_psrl_q;
+ break;
+ case Intrinsic::x86_mmx_psrai_w:
+ NewIntrinsic = Intrinsic::x86_mmx_psra_w;
+ break;
+ case Intrinsic::x86_mmx_psrai_d:
+ NewIntrinsic = Intrinsic::x86_mmx_psra_d;
+ break;
+ default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
+ }
+ // The vector shift intrinsics with scalars uses 32b shift amounts but
+ // the sse2/mmx shift instructions reads 64 bits. Set the upper 32 bits
+ // to be zero.
+ // We must do this early because v2i32 is not a legal type.
+ DebugLoc dl = getCurDebugLoc();
+ SDValue ShOps[2];
+ ShOps[0] = ShAmt;
+ ShOps[1] = DAG.getConstant(0, MVT::i32);
+ ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, ShAmtVT, &ShOps[0], 2);
+ EVT DestVT = TLI.getValueType(I.getType());
+ ShAmt = DAG.getNode(ISD::BITCAST, dl, DestVT, ShAmt);
+ Res = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, DestVT,
+ DAG.getConstant(NewIntrinsic, MVT::i32),
+ getValue(I.getArgOperand(0)), ShAmt);
+ setValue(&I, Res);
+ return 0;
+ }
case Intrinsic::convertff:
case Intrinsic::convertfsi:
case Intrinsic::convertfui:
// Store the stack protector onto the stack.
Res = DAG.getStore(getRoot(), getCurDebugLoc(), Src, FIN,
- PseudoSourceValue::getFixedStack(FI),
- 0, true, false, 0);
+ MachinePointerInfo::getFixedStack(FI),
+ true, false, 0);
setValue(&I, Res);
DAG.setRoot(Res);
return 0;
case Intrinsic::prefetch: {
SDValue Ops[4];
+ unsigned rw = cast<ConstantInt>(I.getArgOperand(1))->getZExtValue();
Ops[0] = getRoot();
Ops[1] = getValue(I.getArgOperand(0));
Ops[2] = getValue(I.getArgOperand(1));
Ops[3] = getValue(I.getArgOperand(2));
- DAG.setRoot(DAG.getNode(ISD::PREFETCH, dl, MVT::Other, &Ops[0], 4));
+ DAG.setRoot(DAG.getMemIntrinsicNode(ISD::PREFETCH, dl,
+ DAG.getVTList(MVT::Other),
+ &Ops[0], 4,
+ EVT::getIntegerVT(*Context, 8),
+ MachinePointerInfo(I.getArgOperand(0)),
+ 0, /* align */
+ false, /* volatile */
+ rw==0, /* read */
+ rw==1)); /* write */
return 0;
}
-
case Intrinsic::memory_barrier: {
SDValue Ops[6];
Ops[0] = getRoot();
getValue(I.getArgOperand(0)),
getValue(I.getArgOperand(1)),
getValue(I.getArgOperand(2)),
- I.getArgOperand(0));
+ MachinePointerInfo(I.getArgOperand(0)));
setValue(&I, L);
DAG.setRoot(L.getValue(1));
return 0;
FTy->isVarArg(), Outs, FTy->getContext());
SDValue DemoteStackSlot;
+ int DemoteStackIdx = -100;
if (!CanLowerReturn) {
uint64_t TySize = TLI.getTargetData()->getTypeAllocSize(
unsigned Align = TLI.getTargetData()->getPrefTypeAlignment(
FTy->getReturnType());
MachineFunction &MF = DAG.getMachineFunction();
- int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
+ DemoteStackIdx = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
const Type *StackSlotPtrType = PointerType::getUnqual(FTy->getReturnType());
- DemoteStackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
+ DemoteStackSlot = DAG.getFrameIndex(DemoteStackIdx, TLI.getPointerTy());
Entry.Node = DemoteStackSlot;
Entry.Ty = StackSlotPtrType;
Entry.isSExt = false;
!isInTailCallPosition(CS, CS.getAttributes().getRetAttributes(), TLI))
isTailCall = false;
+ // If there's a possibility that fast-isel has already selected some amount
+ // of the current basic block, don't emit a tail call.
+ if (isTailCall && EnableFastISel)
+ isTailCall = false;
+
std::pair<SDValue,SDValue> Result =
TLI.LowerCallTo(getRoot(), RetTy,
CS.paramHasAttr(0, Attribute::SExt),
DemoteStackSlot,
DAG.getConstant(Offsets[i], PtrVT));
SDValue L = DAG.getLoad(Outs[i].VT, getCurDebugLoc(), Result.second,
- Add, NULL, Offsets[i], false, false, 1);
+ Add,
+ MachinePointerInfo::getFixedStack(DemoteStackIdx, Offsets[i]),
+ false, false, 1);
Values[i] = L;
Chains[i] = L.getValue(1);
}
SDValue Chain = DAG.getNode(ISD::TokenFactor, getCurDebugLoc(),
MVT::Other, &Chains[0], NumValues);
PendingLoads.push_back(Chain);
-
+
// Collect the legal value parts into potentially illegal values
// that correspond to the original function's return values.
SmallVector<EVT, 4> RetTys;
EVT VT = RetTys[I];
EVT RegisterVT = TLI.getRegisterType(RetTy->getContext(), VT);
unsigned NumRegs = TLI.getNumRegisters(RetTy->getContext(), VT);
-
+
SDValue ReturnValue =
getCopyFromParts(DAG, getCurDebugLoc(), &Values[CurReg], NumRegs,
RegisterVT, VT, AssertOp);
SDValue Ptr = Builder.getValue(PtrVal);
SDValue LoadVal = Builder.DAG.getLoad(LoadVT, Builder.getCurDebugLoc(), Root,
- Ptr, PtrVal /*SrcValue*/, 0/*SVOffset*/,
+ Ptr, MachinePointerInfo(PtrVal),
false /*volatile*/,
false /*nontemporal*/, 1 /* align=1 */);
visitInlineAsm(&I);
return;
}
-
+
+ // See if any floating point values are being passed to this function. This is
+ // used to emit an undefined reference to fltused on Windows.
+ const FunctionType *FT =
+ cast<FunctionType>(I.getCalledValue()->getType()->getContainedType(0));
+ MachineModuleInfo &MMI = DAG.getMachineFunction().getMMI();
+ if (FT->isVarArg() &&
+ !MMI.callsExternalVAFunctionWithFloatingPointArguments()) {
+ for (unsigned i = 0, e = I.getNumArgOperands(); i != e; ++i) {
+ const Type* T = I.getArgOperand(i)->getType();
+ for (po_iterator<const Type*> i = po_begin(T), e = po_end(T);
+ i != e; ++i) {
+ if (!i->isFloatingPointTy()) continue;
+ MMI.setCallsExternalVAFunctionWithFloatingPointArguments(true);
+ break;
+ }
+ }
+ }
+
const char *RenameFn = 0;
if (Function *F = I.getCalledFunction()) {
if (F->isDeclaration()) {
}
}
}
-
+
SDValue Callee;
if (!RenameFn)
Callee = getValue(I.getCalledValue());
/// contains the set of register corresponding to the operand.
RegsForValue AssignedRegs;
- explicit SDISelAsmOperandInfo(const InlineAsm::ConstraintInfo &info)
+ explicit SDISelAsmOperandInfo(const TargetLowering::AsmOperandInfo &info)
: TargetLowering::AsmOperandInfo(info), CallOperand(0,0) {
}
}
};
+typedef SmallVector<SDISelAsmOperandInfo,16> SDISelAsmOperandInfoVector;
+
} // end llvm namespace.
/// isAllocatableRegister - If the specified register is safe to allocate,
// vector types).
EVT RegVT = *PhysReg.second->vt_begin();
if (RegVT.getSizeInBits() == OpInfo.ConstraintVT.getSizeInBits()) {
- OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
+ OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
RegVT, OpInfo.CallOperand);
OpInfo.ConstraintVT = RegVT;
} else if (RegVT.isInteger() && OpInfo.ConstraintVT.isFloatingPoint()) {
// machine.
RegVT = EVT::getIntegerVT(Context,
OpInfo.ConstraintVT.getSizeInBits());
- OpInfo.CallOperand = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
+ OpInfo.CallOperand = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
RegVT, OpInfo.CallOperand);
OpInfo.ConstraintVT = RegVT;
}
const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue());
/// ConstraintOperands - Information about all of the constraints.
- std::vector<SDISelAsmOperandInfo> ConstraintOperands;
+ SDISelAsmOperandInfoVector ConstraintOperands;
std::set<unsigned> OutputRegs, InputRegs;
- // Do a prepass over the constraints, canonicalizing them, and building up the
- // ConstraintOperands list.
- std::vector<InlineAsm::ConstraintInfo>
- ConstraintInfos = IA->ParseConstraints();
-
- bool hasMemory = hasInlineAsmMemConstraint(ConstraintInfos, TLI);
-
- SDValue Chain, Flag;
-
- // We won't need to flush pending loads if this asm doesn't touch
- // memory and is nonvolatile.
- if (hasMemory || IA->hasSideEffects())
- Chain = getRoot();
- else
- Chain = DAG.getRoot();
+ TargetLowering::AsmOperandInfoVector TargetConstraints = TLI.ParseConstraints(CS);
+ bool hasMemory = false;
unsigned ArgNo = 0; // ArgNo - The argument of the CallInst.
unsigned ResNo = 0; // ResNo - The result number of the next output.
- for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
- ConstraintOperands.push_back(SDISelAsmOperandInfo(ConstraintInfos[i]));
+ for (unsigned i = 0, e = TargetConstraints.size(); i != e; ++i) {
+ ConstraintOperands.push_back(SDISelAsmOperandInfo(TargetConstraints[i]));
SDISelAsmOperandInfo &OpInfo = ConstraintOperands.back();
EVT OpVT = MVT::Other;
// If this is an input or an indirect output, process the call argument.
// BasicBlocks are labels, currently appearing only in asm's.
if (OpInfo.CallOperandVal) {
- // Strip bitcasts, if any. This mostly comes up for functions.
- OpInfo.CallOperandVal = OpInfo.CallOperandVal->stripPointerCasts();
-
if (const BasicBlock *BB = dyn_cast<BasicBlock>(OpInfo.CallOperandVal)) {
OpInfo.CallOperand = DAG.getBasicBlock(FuncInfo.MBBMap[BB]);
} else {
}
OpInfo.ConstraintVT = OpVT;
+
+ // Indirect operand accesses access memory.
+ if (OpInfo.isIndirect)
+ hasMemory = true;
+ else {
+ for (unsigned j = 0, ee = OpInfo.Codes.size(); j != ee; ++j) {
+ TargetLowering::ConstraintType CType = TLI.getConstraintType(OpInfo.Codes[j]);
+ if (CType == TargetLowering::C_Memory) {
+ hasMemory = true;
+ break;
+ }
+ }
+ }
}
+ SDValue Chain, Flag;
+
+ // We won't need to flush pending loads if this asm doesn't touch
+ // memory and is nonvolatile.
+ if (hasMemory || IA->hasSideEffects())
+ Chain = getRoot();
+ else
+ Chain = DAG.getRoot();
+
// Second pass over the constraints: compute which constraint option to use
// and assign registers to constraints that want a specific physreg.
- for (unsigned i = 0, e = ConstraintInfos.size(); i != e; ++i) {
+ for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
SDISelAsmOperandInfo &OpInfo = ConstraintOperands[i];
// If this is an output operand with a matching input operand, look up the
// error.
if (OpInfo.hasMatchingInput()) {
SDISelAsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput];
-
+
if (OpInfo.ConstraintVT != Input.ConstraintVT) {
if ((OpInfo.ConstraintVT.isInteger() !=
Input.ConstraintVT.isInteger()) ||
// need to to provide an address for the memory input.
if (OpInfo.ConstraintType == TargetLowering::C_Memory &&
!OpInfo.isIndirect) {
- assert(OpInfo.Type == InlineAsm::isInput &&
+ assert((OpInfo.isMultipleAlternative || (OpInfo.Type == InlineAsm::isInput)) &&
"Can only indirectify direct input operands!");
// Memory operands really want the address of the value. If we don't have
int SSFI = MF.getFrameInfo()->CreateStackObject(TySize, Align, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, TLI.getPointerTy());
Chain = DAG.getStore(Chain, getCurDebugLoc(),
- OpInfo.CallOperand, StackSlot, NULL, 0,
+ OpInfo.CallOperand, StackSlot,
+ MachinePointerInfo::getFixedStack(SSFI),
false, false, 0);
OpInfo.CallOperand = StackSlot;
}
GetRegistersForValue(OpInfo, OutputRegs, InputRegs);
}
- ConstraintInfos.clear();
-
// Second pass - Loop over all of the operands, assigning virtual or physregs
// to register class operands.
for (unsigned i = 0, e = ConstraintOperands.size(); i != e; ++i) {
const MDNode *SrcLoc = CS.getInstruction()->getMetadata("srcloc");
AsmNodeOperands.push_back(DAG.getMDNode(SrcLoc));
- // Remember the AlignStack bit as operand 3.
- AsmNodeOperands.push_back(DAG.getTargetConstant(IA->isAlignStack() ? 1 : 0,
- MVT::i1));
+ // Remember the HasSideEffect and AlignStack bits as operand 3.
+ unsigned ExtraInfo = 0;
+ if (IA->hasSideEffects())
+ ExtraInfo |= InlineAsm::Extra_HasSideEffects;
+ if (IA->isAlignStack())
+ ExtraInfo |= InlineAsm::Extra_IsAlignStack;
+ AsmNodeOperands.push_back(DAG.getTargetConstant(ExtraInfo,
+ TLI.getPointerTy()));
// Loop over all of the inputs, copying the operand values into the
// appropriate registers and processing the output regs.
" don't know how to handle tied "
"indirect register inputs");
}
-
+
RegsForValue MatchedRegs;
MatchedRegs.ValueVTs.push_back(InOperandVal.getValueType());
EVT RegVT = AsmNodeOperands[CurOp+1].getValueType();
DAG, AsmNodeOperands);
break;
}
-
+
assert(InlineAsm::isMemKind(OpFlag) && "Unknown matching constraint!");
assert(InlineAsm::getNumOperandRegisters(OpFlag) == 1 &&
"Unexpected number of operands");
break;
}
- if (OpInfo.ConstraintType == TargetLowering::C_Other) {
- assert(!OpInfo.isIndirect &&
- "Don't know how to handle indirect other inputs yet!");
+ // Treat indirect 'X' constraint as memory.
+ if (OpInfo.ConstraintType == TargetLowering::C_Other &&
+ OpInfo.isIndirect)
+ OpInfo.ConstraintType = TargetLowering::C_Memory;
+ if (OpInfo.ConstraintType == TargetLowering::C_Other) {
std::vector<SDValue> Ops;
TLI.LowerAsmOperandForConstraint(InOperandVal, OpInfo.ConstraintCode[0],
Ops, DAG);
AsmNodeOperands.insert(AsmNodeOperands.end(), Ops.begin(), Ops.end());
break;
}
-
+
if (OpInfo.ConstraintType == TargetLowering::C_Memory) {
assert(OpInfo.isIndirect && "Operand must be indirect to be a mem!");
assert(InOperandVal.getValueType() == TLI.getPointerTy() &&
if (Flag.getNode()) AsmNodeOperands.push_back(Flag);
Chain = DAG.getNode(ISD::INLINEASM, getCurDebugLoc(),
- DAG.getVTList(MVT::Other, MVT::Flag),
+ DAG.getVTList(MVT::Other, MVT::Glue),
&AsmNodeOperands[0], AsmNodeOperands.size());
Flag = Chain.getValue(1);
// not have the same VT as was expected. Convert it to the right type
// with bit_convert.
if (ResultType != Val.getValueType() && Val.getValueType().isVector()) {
- Val = DAG.getNode(ISD::BIT_CONVERT, getCurDebugLoc(),
+ Val = DAG.getNode(ISD::BITCAST, getCurDebugLoc(),
ResultType, Val);
} else if (ResultType != Val.getValueType() &&
SDValue Val = DAG.getStore(Chain, getCurDebugLoc(),
StoresToEmit[i].first,
getValue(StoresToEmit[i].second),
- StoresToEmit[i].second, 0,
+ MachinePointerInfo(StoresToEmit[i].second),
false, false, 0);
OutChains.push_back(Val);
}
}
void SelectionDAGBuilder::visitVAArg(const VAArgInst &I) {
+ const TargetData &TD = *TLI.getTargetData();
SDValue V = DAG.getVAArg(TLI.getValueType(I.getType()), getCurDebugLoc(),
getRoot(), getValue(I.getOperand(0)),
- DAG.getSrcValue(I.getOperand(0)));
+ DAG.getSrcValue(I.getOperand(0)),
+ TD.getABITypeAlignment(I.getType()));
setValue(&I, V);
DAG.setRoot(V.getValue(1));
}
unsigned NumRegs = getNumRegisters(RetTy->getContext(), VT);
for (unsigned i = 0; i != NumRegs; ++i) {
ISD::InputArg MyFlags;
- MyFlags.VT = RegisterVT;
+ MyFlags.VT = RegisterVT.getSimpleVT();
MyFlags.Used = isReturnValueUsed;
if (RetSExt)
MyFlags.Flags.setSExt();
DEBUG(for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
assert(InVals[i].getNode() &&
"LowerCall emitted a null value!");
- assert(Ins[i].VT == InVals[i].getValueType() &&
+ assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
"LowerCall emitted a value with the wrong type!");
});
for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
assert(InVals[i].getNode() &&
"LowerFormalArguments emitted a null value!");
- assert(Ins[i].VT == InVals[i].getValueType() &&
+ assert(EVT(Ins[i].VT) == InVals[i].getValueType() &&
"LowerFormalArguments emitted a value with the wrong type!");
}
});
i += NumParts;
}
+ // Note down frame index for byval arguments.
+ if (I->hasByValAttr() && !ArgValues.empty())
+ if (FrameIndexSDNode *FI =
+ dyn_cast<FrameIndexSDNode>(ArgValues[0].getNode()))
+ FuncInfo->setByValArgumentFrameIndex(I, FI->getIndex());
+
if (!I->use_empty()) {
SDValue Res;
if (!ArgValues.empty())
projects / oota-llvm.git / blobdiff