#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
+#include "llvm/GlobalValue.h"
#include "llvm/Instruction.h"
#include "llvm/Intrinsics.h"
-#include "llvm/GlobalValue.h"
+#include "llvm/Type.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/ADT/VectorExtras.h"
+#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
#include <sstream>
using namespace llvm;
setOperationAction(ISD::VSETCC, VT.getSimpleVT(), Custom);
if (ElemTy == MVT::i8 || ElemTy == MVT::i16)
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
+ if (ElemTy != MVT::i32) {
+ setOperationAction(ISD::SINT_TO_FP, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::UINT_TO_FP, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Expand);
+ }
setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
AddPromotedToType (ISD::XOR, VT.getSimpleVT(),
PromotedBitwiseVT.getSimpleVT());
}
+
+ // Neon does not support vector divide/remainder operations.
+ setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::FDIV, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
+ setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);
}
void ARMTargetLowering::addDRTypeForNEON(EVT VT) {
}
ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
- : TargetLowering(TM, createTLOF(TM)), ARMPCLabelIndex(0) {
+ : TargetLowering(TM, createTLOF(TM)) {
Subtarget = &TM.getSubtarget<ARMSubtarget>();
if (Subtarget->isTargetDarwin()) {
addQRTypeForNEON(MVT::v4i32);
addQRTypeForNEON(MVT::v2i64);
+ // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
+ // neither Neon nor VFP support any arithmetic operations on it.
+ setOperationAction(ISD::FADD, MVT::v2f64, Expand);
+ setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
+ setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
+ setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
+ setOperationAction(ISD::FREM, MVT::v2f64, Expand);
+ setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
+ setOperationAction(ISD::VSETCC, MVT::v2f64, Expand);
+ setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
+ setOperationAction(ISD::FABS, MVT::v2f64, Expand);
+ setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
+ setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
+ setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
+ setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
+ setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
+ setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
+ setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
+ setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
+ setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
+ setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
+ setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
+ setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
+ setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
+ setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
+ setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
+
+ // Neon does not support some operations on v1i64 and v2i64 types.
+ setOperationAction(ISD::MUL, MVT::v1i64, Expand);
+ setOperationAction(ISD::MUL, MVT::v2i64, Expand);
+ setOperationAction(ISD::VSETCC, MVT::v1i64, Expand);
+ setOperationAction(ISD::VSETCC, MVT::v2i64, Expand);
+
setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
setTargetDAGCombine(ISD::SHL);
setTargetDAGCombine(ISD::SRL);
if (!Subtarget->hasV6Ops())
setOperationAction(ISD::MULHS, MVT::i32, Expand);
}
- setOperationAction(ISD::SHL_PARTS, MVT::i32, Expand);
- setOperationAction(ISD::SRA_PARTS, MVT::i32, Expand);
- setOperationAction(ISD::SRL_PARTS, MVT::i32, Expand);
+ setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
setOperationAction(ISD::SRL, MVT::i64, Custom);
setOperationAction(ISD::SRA, MVT::i64, Custom);
setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
- // Support label based line numbers.
- setOperationAction(ISD::DBG_STOPPOINT, MVT::Other, Expand);
- setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
-
setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
setOperationAction(ISD::ConstantPool, MVT::i32, Custom);
setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
+ setOperationAction(ISD::BlockAddress, MVT::i32, Custom);
// Use the default implementation.
setOperationAction(ISD::VASTART, MVT::Other, Custom);
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
else
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
- setOperationAction(ISD::MEMBARRIER, MVT::Other, Expand);
+ setOperationAction(ISD::MEMBARRIER, MVT::Other, Custom);
if (!Subtarget->hasV6Ops() && !Subtarget->isThumb2()) {
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
if (!UseSoftFloat && Subtarget->hasVFP2() && !Subtarget->isThumb1Only())
- // Turn f64->i64 into FMRRD, i64 -> f64 to FMDRR iff target supports vfp2.
+ // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR iff target supports vfp2.
setOperationAction(ISD::BIT_CONVERT, MVT::i64, Custom);
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
- setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
- setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
setOperationAction(ISD::SETCC, MVT::i32, Expand);
setOperationAction(ISD::SETCC, MVT::f32, Expand);
}
// We have target-specific dag combine patterns for the following nodes:
- // ARMISD::FMRRD - No need to call setTargetDAGCombine
+ // ARMISD::VMOVRRD - No need to call setTargetDAGCombine
setTargetDAGCombine(ISD::ADD);
setTargetDAGCombine(ISD::SUB);
case ARMISD::SRA_FLAG: return "ARMISD::SRA_FLAG";
case ARMISD::RRX: return "ARMISD::RRX";
- case ARMISD::FMRRD: return "ARMISD::FMRRD";
- case ARMISD::FMDRR: return "ARMISD::FMDRR";
+ case ARMISD::VMOVRRD: return "ARMISD::VMOVRRD";
+ case ARMISD::VMOVDRR: return "ARMISD::VMOVDRR";
+
+ case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
+ case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";
case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";
case ARMISD::DYN_ALLOC: return "ARMISD::DYN_ALLOC";
+ case ARMISD::MEMBARRIER: return "ARMISD::MEMBARRIER";
+ case ARMISD::SYNCBARRIER: return "ARMISD::SYNCBARRIER";
+
case ARMISD::VCEQ: return "ARMISD::VCEQ";
case ARMISD::VCGE: return "ARMISD::VCGE";
case ARMISD::VCGEU: return "ARMISD::VCGEU";
/// getFunctionAlignment - Return the Log2 alignment of this function.
unsigned ARMTargetLowering::getFunctionAlignment(const Function *F) const {
- return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 1 : 2;
+ return getTargetMachine().getSubtarget<ARMSubtarget>().isThumb() ? 0 : 1;
}
//===----------------------------------------------------------------------===//
}
}
-/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC. It
-/// returns true if the operands should be inverted to form the proper
-/// comparison.
-static bool FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
+/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
+static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
ARMCC::CondCodes &CondCode2) {
- bool Invert = false;
CondCode2 = ARMCC::AL;
switch (CC) {
default: llvm_unreachable("Unknown FP condition!");
case ISD::SETGE:
case ISD::SETOGE: CondCode = ARMCC::GE; break;
case ISD::SETOLT: CondCode = ARMCC::MI; break;
- case ISD::SETOLE: CondCode = ARMCC::GT; Invert = true; break;
+ case ISD::SETOLE: CondCode = ARMCC::LS; break;
case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
case ISD::SETO: CondCode = ARMCC::VC; break;
case ISD::SETUO: CondCode = ARMCC::VS; break;
case ISD::SETNE:
case ISD::SETUNE: CondCode = ARMCC::NE; break;
}
- return Invert;
}
//===----------------------------------------------------------------------===//
/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
/// given CallingConvention value.
-CCAssignFn *ARMTargetLowering::CCAssignFnForNode(unsigned CC,
+CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
bool Return,
bool isVarArg) const {
switch (CC) {
/// appropriate copies out of appropriate physical registers.
SDValue
ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg> &Ins,
DebugLoc dl, SelectionDAG &DAG,
SmallVectorImpl<SDValue> &InVals) {
InFlag);
Chain = Hi.getValue(1);
InFlag = Hi.getValue(2);
- Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
+ Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
if (VA.getLocVT() == MVT::v2f64) {
SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
Chain = Hi.getValue(1);
InFlag = Hi.getValue(2);
- Val = DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
+ Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
DAG.getConstant(1, MVT::i32));
}
SmallVector<SDValue, 8> &MemOpChains,
ISD::ArgFlagsTy Flags) {
- SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
+ SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
DAG.getVTList(MVT::i32, MVT::i32), Arg);
RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));
/// nodes.
SDValue
ARMTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
bool isTailCall,
const SmallVectorImpl<ISD::OutputArg> &Outs,
const SmallVectorImpl<ISD::InputArg> &Ins,
bool isDirect = false;
bool isARMFunc = false;
bool isLocalARMFunc = false;
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
GlobalValue *GV = G->getGlobal();
isDirect = true;
isLocalARMFunc = !Subtarget->isThumb() && !isExt;
// tBX takes a register source operand.
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV,
ARMPCLabelIndex,
ARMCP::CPValue, 4);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
Callee = DAG.getLoad(getPointerTy(), dl,
- DAG.getEntryNode(), CPAddr, NULL, 0);
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
getPointerTy(), Callee, PICLabel);
} else
// tBX takes a register source operand.
const char *Sym = S->getSymbol();
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
Sym, ARMPCLabelIndex, 4);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
Callee = DAG.getLoad(getPointerTy(), dl,
- DAG.getEntryNode(), CPAddr, NULL, 0);
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
getPointerTy(), Callee, PICLabel);
} else
SDValue
ARMTargetLowering::LowerReturn(SDValue Chain,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
DebugLoc dl, SelectionDAG &DAG) {
// Extract the first half and return it in two registers.
SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
DAG.getConstant(0, MVT::i32));
- SDValue HalfGPRs = DAG.getNode(ARMISD::FMRRD, dl,
+ SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
DAG.getVTList(MVT::i32, MVT::i32), Half);
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
}
// Legalize ret f64 -> ret 2 x i32. We always have fmrrd if f64 is
// available.
- SDValue fmrrd = DAG.getNode(ARMISD::FMRRD, dl,
+ SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
Flag = Chain.getValue(1);
return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
}
+SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = 0;
+ DebugLoc DL = Op.getDebugLoc();
+ EVT PtrVT = getPointerTy();
+ BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
+ Reloc::Model RelocM = getTargetMachine().getRelocationModel();
+ SDValue CPAddr;
+ if (RelocM == Reloc::Static) {
+ CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
+ } else {
+ unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
+ ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ ARMConstantPoolValue *CPV = new ARMConstantPoolValue(BA, ARMPCLabelIndex,
+ ARMCP::CPBlockAddress,
+ PCAdj);
+ CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
+ }
+ CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
+ SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
+ if (RelocM == Reloc::Static)
+ return Result;
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
+ return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
+}
+
// Lower ISD::GlobalTLSAddress using the "general dynamic" model
SDValue
ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
DebugLoc dl = GA->getDebugLoc();
EVT PtrVT = getPointerTy();
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
ARMConstantPoolValue *CPV =
new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
ARMCP::CPValue, PCAdj, "tlsgd", true);
SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
- Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument, NULL, 0);
+ Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
+ PseudoSourceValue::getConstantPool(), 0);
SDValue Chain = Argument.getValue(1);
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);
// call __tls_get_addr.
LowerCallTo(Chain, (const Type *) Type::getInt32Ty(*DAG.getContext()),
false, false, false, false,
0, CallingConv::C, false, /*isReturnValueUsed=*/true,
- DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
+ DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl,
+ DAG.GetOrdering(Chain.getNode()));
return CallResult.first;
}
SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
if (GV->isDeclaration()) {
- // initial exec model
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
+ // Initial exec model.
unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
ARMConstantPoolValue *CPV =
new ARMConstantPoolValue(GA->getGlobal(), ARMPCLabelIndex,
ARMCP::CPValue, PCAdj, "gottpoff", true);
Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
- Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
+ Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
+ PseudoSourceValue::getConstantPool(), 0);
Chain = Offset.getValue(1);
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);
- Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
+ Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
+ PseudoSourceValue::getConstantPool(), 0);
} else {
// local exec model
ARMConstantPoolValue *CPV = new ARMConstantPoolValue(GV, "tpoff");
Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
- Offset = DAG.getLoad(PtrVT, dl, Chain, Offset, NULL, 0);
+ Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
+ PseudoSourceValue::getConstantPool(), 0);
}
// The address of the thread local variable is the add of the thread
SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
- CPAddr, NULL, 0);
+ CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
SDValue Chain = Result.getValue(1);
SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
if (!UseGOTOFF)
- Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
+ Result = DAG.getLoad(PtrVT, dl, Chain, Result,
+ PseudoSourceValue::getGOT(), 0);
return Result;
} else {
- SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
- CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
- return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
+ // If we have T2 ops, we can materialize the address directly via movt/movw
+ // pair. This is always cheaper.
+ if (Subtarget->useMovt()) {
+ return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
+ DAG.getTargetGlobalAddress(GV, PtrVT));
+ } else {
+ SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
+ CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
+ return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
+ }
}
}
SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
SelectionDAG &DAG) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = 0;
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
if (RelocM == Reloc::Static)
CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
else {
+ ARMPCLabelIndex = AFI->createConstPoolEntryUId();
unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
ARMConstantPoolValue *CPV =
new ARMConstantPoolValue(GV, ARMPCLabelIndex, ARMCP::CPValue, PCAdj);
}
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
- SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
+ SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
SDValue Chain = Result.getValue(1);
if (RelocM == Reloc::PIC_) {
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
}
- if (Subtarget->GVIsIndirectSymbol(GV, RelocM == Reloc::Static))
- Result = DAG.getLoad(PtrVT, dl, Chain, Result, NULL, 0);
+ if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
+ Result = DAG.getLoad(PtrVT, dl, Chain, Result,
+ PseudoSourceValue::getGOT(), 0);
return Result;
}
SelectionDAG &DAG){
assert(Subtarget->isTargetELF() &&
"GLOBAL OFFSET TABLE not implemented for non-ELF targets");
+ MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
ARMPCLabelIndex, PCAdj);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
- SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
}
-static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
- unsigned NumVecs) {
- SDNode *Node = Op.getNode();
- EVT VT = Node->getValueType(0);
-
- // No expansion needed for 64-bit vectors.
- if (VT.is64BitVector())
- return SDValue();
-
- // FIXME: We need to expand VLD3 and VLD4 of 128-bit vectors into separate
- // operations to load the even and odd registers.
- return SDValue();
-}
-
-static SDValue LowerNeonVSTIntrinsic(SDValue Op, SelectionDAG &DAG,
- unsigned NumVecs) {
- SDNode *Node = Op.getNode();
- EVT VT = Node->getOperand(3).getValueType();
-
- // No expansion needed for 64-bit vectors.
- if (VT.is64BitVector())
- return SDValue();
-
- // FIXME: We need to expand VST3 and VST4 of 128-bit vectors into separate
- // operations to store the even and odd registers.
- return SDValue();
-}
-
-static SDValue LowerNeonVLDLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
- unsigned NumVecs) {
- SDNode *Node = Op.getNode();
- EVT VT = Node->getValueType(0);
-
- if (!VT.is64BitVector())
- return SDValue(); // unimplemented
-
- // Change the lane number operand to be a TargetConstant; otherwise it
- // will be legalized into a register.
- ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
- if (!Lane) {
- assert(false && "vld lane number must be a constant");
- return SDValue();
- }
- SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
- Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
- return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
-}
-
-static SDValue LowerNeonVSTLaneIntrinsic(SDValue Op, SelectionDAG &DAG,
- unsigned NumVecs) {
- SDNode *Node = Op.getNode();
- EVT VT = Node->getOperand(3).getValueType();
-
- if (!VT.is64BitVector())
- return SDValue(); // unimplemented
-
- // Change the lane number operand to be a TargetConstant; otherwise it
- // will be legalized into a register.
- ConstantSDNode *Lane = dyn_cast<ConstantSDNode>(Node->getOperand(NumVecs+3));
- if (!Lane) {
- assert(false && "vst lane number must be a constant");
- return SDValue();
- }
- SmallVector<SDValue, 8> Ops(Node->op_begin(), Node->op_end());
- Ops[NumVecs+3] = DAG.getTargetConstant(Lane->getZExtValue(), MVT::i32);
- return DAG.UpdateNodeOperands(Op, &Ops[0], Ops.size());
-}
-
-SDValue
-ARMTargetLowering::LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) {
- unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
- switch (IntNo) {
- case Intrinsic::arm_neon_vld3:
- return LowerNeonVLDIntrinsic(Op, DAG, 3);
- case Intrinsic::arm_neon_vld4:
- return LowerNeonVLDIntrinsic(Op, DAG, 4);
- case Intrinsic::arm_neon_vld2lane:
- return LowerNeonVLDLaneIntrinsic(Op, DAG, 2);
- case Intrinsic::arm_neon_vld3lane:
- return LowerNeonVLDLaneIntrinsic(Op, DAG, 3);
- case Intrinsic::arm_neon_vld4lane:
- return LowerNeonVLDLaneIntrinsic(Op, DAG, 4);
- case Intrinsic::arm_neon_vst3:
- return LowerNeonVSTIntrinsic(Op, DAG, 3);
- case Intrinsic::arm_neon_vst4:
- return LowerNeonVSTIntrinsic(Op, DAG, 4);
- case Intrinsic::arm_neon_vst2lane:
- return LowerNeonVSTLaneIntrinsic(Op, DAG, 2);
- case Intrinsic::arm_neon_vst3lane:
- return LowerNeonVSTLaneIntrinsic(Op, DAG, 3);
- case Intrinsic::arm_neon_vst4lane:
- return LowerNeonVSTLaneIntrinsic(Op, DAG, 4);
- default: return SDValue(); // Don't custom lower most intrinsics.
- }
-}
-
SDValue
ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) {
unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
}
case Intrinsic::eh_sjlj_lsda: {
MachineFunction &MF = DAG.getMachineFunction();
+ ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
+ unsigned ARMPCLabelIndex = AFI->createConstPoolEntryUId();
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
Reloc::Model RelocM = getTargetMachine().getRelocationModel();
CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
SDValue Result =
- DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr, NULL, 0);
+ DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
+ PseudoSourceValue::getConstantPool(), 0);
SDValue Chain = Result.getValue(1);
if (RelocM == Reloc::PIC_) {
- SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex++, MVT::i32);
+ SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
}
return Result;
}
}
+static SDValue LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG,
+ const ARMSubtarget *Subtarget) {
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue Op5 = Op.getOperand(5);
+ SDValue Res;
+ unsigned isDeviceBarrier = cast<ConstantSDNode>(Op5)->getZExtValue();
+ if (isDeviceBarrier) {
+ if (Subtarget->hasV7Ops())
+ Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0));
+ else
+ Res = DAG.getNode(ARMISD::SYNCBARRIER, dl, MVT::Other, Op.getOperand(0),
+ DAG.getConstant(0, MVT::i32));
+ } else {
+ if (Subtarget->hasV7Ops())
+ Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
+ else
+ Res = DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
+ DAG.getConstant(0, MVT::i32));
+ }
+ return Res;
+}
+
static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG,
unsigned VarArgsFrameIndex) {
// vastart just stores the address of the VarArgsFrameIndex slot into the
if (NextVA.isMemLoc()) {
unsigned ArgSize = NextVA.getLocVT().getSizeInBits()/8;
MachineFrameInfo *MFI = MF.getFrameInfo();
- int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset());
+ int FI = MFI->CreateFixedObject(ArgSize, NextVA.getLocMemOffset(),
+ true, false);
// Create load node to retrieve arguments from the stack.
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
- ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN, NULL, 0);
+ ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
+ PseudoSourceValue::getFixedStack(FI), 0);
} else {
Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
}
- return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, ArgValue, ArgValue2);
+ return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
}
SDValue
ARMTargetLowering::LowerFormalArguments(SDValue Chain,
- unsigned CallConv, bool isVarArg,
+ CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::InputArg>
&Ins,
DebugLoc dl, SelectionDAG &DAG,
assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");
unsigned ArgSize = VA.getLocVT().getSizeInBits()/8;
- int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset());
+ int FI = MFI->CreateFixedObject(ArgSize, VA.getLocMemOffset(),
+ true, false);
// Create load nodes to retrieve arguments from the stack.
SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
- InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN, NULL, 0));
+ InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
+ PseudoSourceValue::getFixedStack(FI), 0));
}
}
unsigned Align = MF.getTarget().getFrameInfo()->getStackAlignment();
unsigned VARegSize = (4 - NumGPRs) * 4;
unsigned VARegSaveSize = (VARegSize + Align - 1) & ~(Align - 1);
- unsigned ArgOffset = 0;
+ unsigned ArgOffset = CCInfo.getNextStackOffset();
if (VARegSaveSize) {
// If this function is vararg, store any remaining integer argument regs
// to their spots on the stack so that they may be loaded by deferencing
// the result of va_next.
AFI->setVarArgsRegSaveSize(VARegSaveSize);
- ArgOffset = CCInfo.getNextStackOffset();
VarArgsFrameIndex = MFI->CreateFixedObject(VARegSaveSize, ArgOffset +
- VARegSaveSize - VARegSize);
+ VARegSaveSize - VARegSize,
+ true, false);
SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, getPointerTy());
SmallVector<SDValue, 4> MemOps;
unsigned VReg = MF.addLiveIn(GPRArgRegs[NumGPRs], RC);
SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
- SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN, NULL, 0);
+ SDValue Store = DAG.getStore(Val.getValue(1), dl, Val, FIN,
+ PseudoSourceValue::getFixedStack(VarArgsFrameIndex), 0);
MemOps.push_back(Store);
FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
DAG.getConstant(4, getPointerTy()));
&MemOps[0], MemOps.size());
} else
// This will point to the next argument passed via stack.
- VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
+ VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset, true, false);
}
return Chain;
return false;
}
-static bool isLegalCmpImmediate(unsigned C, bool isThumb1Only) {
- return ( isThumb1Only && (C & ~255U) == 0) ||
- (!isThumb1Only && ARM_AM::getSOImmVal(C) != -1);
-}
-
/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
/// the given operands.
-static SDValue getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
- SDValue &ARMCC, SelectionDAG &DAG, bool isThumb1Only,
- DebugLoc dl) {
+SDValue
+ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
+ SDValue &ARMCC, SelectionDAG &DAG, DebugLoc dl) {
if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
unsigned C = RHSC->getZExtValue();
- if (!isLegalCmpImmediate(C, isThumb1Only)) {
+ if (!isLegalICmpImmediate(C)) {
// Constant does not fit, try adjusting it by one?
switch (CC) {
default: break;
case ISD::SETLT:
case ISD::SETGE:
- if (isLegalCmpImmediate(C-1, isThumb1Only)) {
+ if (isLegalICmpImmediate(C-1)) {
CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
RHS = DAG.getConstant(C-1, MVT::i32);
}
break;
case ISD::SETULT:
case ISD::SETUGE:
- if (C > 0 && isLegalCmpImmediate(C-1, isThumb1Only)) {
+ if (C > 0 && isLegalICmpImmediate(C-1)) {
CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
RHS = DAG.getConstant(C-1, MVT::i32);
}
break;
case ISD::SETLE:
case ISD::SETGT:
- if (isLegalCmpImmediate(C+1, isThumb1Only)) {
+ if (isLegalICmpImmediate(C+1)) {
CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
RHS = DAG.getConstant(C+1, MVT::i32);
}
break;
case ISD::SETULE:
case ISD::SETUGT:
- if (C < 0xffffffff && isLegalCmpImmediate(C+1, isThumb1Only)) {
+ if (C < 0xffffffff && isLegalICmpImmediate(C+1)) {
CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
RHS = DAG.getConstant(C+1, MVT::i32);
}
return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Flag, Cmp);
}
-static SDValue LowerSELECT_CC(SDValue Op, SelectionDAG &DAG,
- const ARMSubtarget *ST) {
+SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) {
EVT VT = Op.getValueType();
SDValue LHS = Op.getOperand(0);
SDValue RHS = Op.getOperand(1);
if (LHS.getValueType() == MVT::i32) {
SDValue ARMCC;
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
- SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
+ SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC, CCR,Cmp);
}
ARMCC::CondCodes CondCode, CondCode2;
- if (FPCCToARMCC(CC, CondCode, CondCode2))
- std::swap(TrueVal, FalseVal);
+ FPCCToARMCC(CC, CondCode, CondCode2);
SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
return Result;
}
-static SDValue LowerBR_CC(SDValue Op, SelectionDAG &DAG,
- const ARMSubtarget *ST) {
+SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) {
SDValue Chain = Op.getOperand(0);
ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
SDValue LHS = Op.getOperand(2);
if (LHS.getValueType() == MVT::i32) {
SDValue ARMCC;
SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
- SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, ST->isThumb1Only(), dl);
+ SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMCC, DAG, dl);
return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
Chain, Dest, ARMCC, CCR,Cmp);
}
assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
ARMCC::CondCodes CondCode, CondCode2;
- if (FPCCToARMCC(CC, CondCode, CondCode2))
- // Swap the LHS/RHS of the comparison if needed.
- std::swap(LHS, RHS);
+ FPCCToARMCC(CC, CondCode, CondCode2);
SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
SDValue ARMCC = DAG.getConstant(CondCode, MVT::i32);
Addr, Op.getOperand(2), JTI, UId);
}
if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
- Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr, NULL, 0);
+ Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
+ PseudoSourceValue::getJumpTable(), 0);
Chain = Addr.getValue(1);
Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
} else {
- Addr = DAG.getLoad(PTy, dl, Chain, Addr, NULL, 0);
+ Addr = DAG.getLoad(PTy, dl, Chain, Addr,
+ PseudoSourceValue::getJumpTable(), 0);
Chain = Addr.getValue(1);
return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
}
SDValue Op = N->getOperand(0);
DebugLoc dl = N->getDebugLoc();
if (N->getValueType(0) == MVT::f64) {
- // Turn i64->f64 into FMDRR.
+ // Turn i64->f64 into VMOVDRR.
SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
DAG.getConstant(0, MVT::i32));
SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
DAG.getConstant(1, MVT::i32));
- return DAG.getNode(ARMISD::FMDRR, dl, MVT::f64, Lo, Hi);
+ return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
}
- // Turn f64->i64 into FMRRD.
- SDValue Cvt = DAG.getNode(ARMISD::FMRRD, dl,
+ // Turn f64->i64 into VMOVRRD.
+ SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
// Merge the pieces into a single i64 value.
// will be implemented with the NEON VNEG instruction. However, VNEG does
// not support i64 elements, so sometimes the zero vectors will need to be
// explicitly constructed. For those cases, and potentially other uses in
- // the future, always build zero vectors as <4 x i32> or <2 x i32> bitcasted
+ // the future, always build zero vectors as <16 x i8> or <8 x i8> bitcasted
// to their dest type. This ensures they get CSE'd.
SDValue Vec;
- SDValue Cst = DAG.getTargetConstant(0, MVT::i32);
- if (VT.getSizeInBits() == 64)
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
- else
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+ SDValue Cst = DAG.getTargetConstant(0, MVT::i8);
+ SmallVector<SDValue, 8> Ops;
+ MVT TVT;
+
+ if (VT.getSizeInBits() == 64) {
+ Ops.assign(8, Cst); TVT = MVT::v8i8;
+ } else {
+ Ops.assign(16, Cst); TVT = MVT::v16i8;
+ }
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
}
static SDValue getOnesVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
assert(VT.isVector() && "Expected a vector type");
- // Always build ones vectors as <4 x i32> or <2 x i32> bitcasted to their dest
- // type. This ensures they get CSE'd.
+ // Always build ones vectors as <16 x i8> or <8 x i8> bitcasted to their
+ // dest type. This ensures they get CSE'd.
SDValue Vec;
- SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
- if (VT.getSizeInBits() == 64)
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2i32, Cst, Cst);
- else
- Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
+ SDValue Cst = DAG.getTargetConstant(0xFF, MVT::i8);
+ SmallVector<SDValue, 8> Ops;
+ MVT TVT;
+
+ if (VT.getSizeInBits() == 64) {
+ Ops.assign(8, Cst); TVT = MVT::v8i8;
+ } else {
+ Ops.assign(16, Cst); TVT = MVT::v16i8;
+ }
+ Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, TVT, &Ops[0], Ops.size());
return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
}
+/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
+/// i32 values and take a 2 x i32 value to shift plus a shift amount.
+SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && "Not a double-shift!");
+ EVT VT = Op.getValueType();
+ unsigned VTBits = VT.getSizeInBits();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue ShOpLo = Op.getOperand(0);
+ SDValue ShOpHi = Op.getOperand(1);
+ SDValue ShAmt = Op.getOperand(2);
+ SDValue ARMCC;
+ unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;
+
+ assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);
+
+ SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
+ DAG.getConstant(VTBits, MVT::i32), ShAmt);
+ SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
+ SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
+ DAG.getConstant(VTBits, MVT::i32));
+ SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
+ SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
+ SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
+
+ SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
+ SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
+ ARMCC, DAG, dl);
+ SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
+ SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMCC,
+ CCR, Cmp);
+
+ SDValue Ops[2] = { Lo, Hi };
+ return DAG.getMergeValues(Ops, 2, dl);
+}
+
+/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
+/// i32 values and take a 2 x i32 value to shift plus a shift amount.
+SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && "Not a double-shift!");
+ EVT VT = Op.getValueType();
+ unsigned VTBits = VT.getSizeInBits();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue ShOpLo = Op.getOperand(0);
+ SDValue ShOpHi = Op.getOperand(1);
+ SDValue ShAmt = Op.getOperand(2);
+ SDValue ARMCC;
+
+ assert(Op.getOpcode() == ISD::SHL_PARTS);
+ SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
+ DAG.getConstant(VTBits, MVT::i32), ShAmt);
+ SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
+ SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
+ DAG.getConstant(VTBits, MVT::i32));
+ SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
+ SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);
+
+ SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
+ SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
+ SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
+ ARMCC, DAG, dl);
+ SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
+ SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMCC,
+ CCR, Cmp);
+
+ SDValue Ops[2] = { Lo, Hi };
+ return DAG.getMergeValues(Ops, 2, dl);
+}
+
static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
const ARMSubtarget *ST) {
EVT VT = N->getValueType(0);
assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
"Only possible block sizes for VREV are: 16, 32, 64");
- unsigned NumElts = VT.getVectorNumElements();
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
unsigned BlockElts = M[0] + 1;
if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
unsigned NumElts = VT.getVectorNumElements();
WhichResult = (M[0] == 0 ? 0 : 1);
for (unsigned i = 0; i < NumElts; i += 2) {
return true;
}
+/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
+static bool isVTRN_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ for (unsigned i = 0; i < NumElts; i += 2) {
+ if ((unsigned) M[i] != i + WhichResult ||
+ (unsigned) M[i+1] != i + WhichResult)
+ return false;
+ }
+ return true;
+}
+
static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
unsigned NumElts = VT.getVectorNumElements();
WhichResult = (M[0] == 0 ? 0 : 1);
for (unsigned i = 0; i != NumElts; ++i) {
}
// VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
- if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
+ if (VT.is64BitVector() && EltSz == 32)
+ return false;
+
+ return true;
+}
+
+/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
+static bool isVUZP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned Half = VT.getVectorNumElements() / 2;
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ for (unsigned j = 0; j != 2; ++j) {
+ unsigned Idx = WhichResult;
+ for (unsigned i = 0; i != Half; ++i) {
+ if ((unsigned) M[i + j * Half] != Idx)
+ return false;
+ Idx += 2;
+ }
+ }
+
+ // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && EltSz == 32)
return false;
return true;
static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
unsigned NumElts = VT.getVectorNumElements();
WhichResult = (M[0] == 0 ? 0 : 1);
unsigned Idx = WhichResult * NumElts / 2;
}
// VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
- if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
+ if (VT.is64BitVector() && EltSz == 32)
+ return false;
+
+ return true;
+}
+
+/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
+/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
+/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
+static bool isVZIP_v_undef_Mask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned EltSz = VT.getVectorElementType().getSizeInBits();
+ if (EltSz == 64)
+ return false;
+
+ unsigned NumElts = VT.getVectorNumElements();
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ unsigned Idx = WhichResult * NumElts / 2;
+ for (unsigned i = 0; i != NumElts; i += 2) {
+ if ((unsigned) M[i] != Idx ||
+ (unsigned) M[i+1] != Idx)
+ return false;
+ Idx += 1;
+ }
+
+ // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && EltSz == 32)
return false;
return true;
}
+
static SDValue BuildSplat(SDValue Val, EVT VT, SelectionDAG &DAG, DebugLoc dl) {
// Canonicalize all-zeros and all-ones vectors.
ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
isVEXTMask(M, VT, ReverseVEXT, Imm) ||
isVTRNMask(M, VT, WhichResult) ||
isVUZPMask(M, VT, WhichResult) ||
- isVZIPMask(M, VT, WhichResult));
+ isVZIPMask(M, VT, WhichResult) ||
+ isVTRN_v_undef_Mask(M, VT, WhichResult) ||
+ isVUZP_v_undef_Mask(M, VT, WhichResult) ||
+ isVZIP_v_undef_Mask(M, VT, WhichResult));
}
/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
int Lane = SVN->getSplatIndex();
+ // If this is undef splat, generate it via "just" vdup, if possible.
+ if (Lane == -1) Lane = 0;
+
if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
}
return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
V1, V2).getValue(WhichResult);
+ if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+ if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+ if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
+ V1, V1).getValue(WhichResult);
+
// If the shuffle is not directly supported and it has 4 elements, use
// the PerfectShuffle-generated table to synthesize it from other shuffles.
if (VT.getVectorNumElements() == 4 &&
DebugLoc dl = Op.getDebugLoc();
SDValue Vec = Op.getOperand(0);
SDValue Lane = Op.getOperand(1);
-
- // FIXME: This is invalid for 8 and 16-bit elements - the information about
- // sign / zero extension is lost!
- Op = DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
- Op = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Op, DAG.getValueType(VT));
-
- if (VT.bitsLT(MVT::i32))
- Op = DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
- else if (VT.bitsGT(MVT::i32))
- Op = DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op);
-
- return Op;
+ assert(VT == MVT::i32 &&
+ Vec.getValueType().getVectorElementType().getSizeInBits() < 32 &&
+ "unexpected type for custom-lowering vector extract");
+ return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
}
static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
switch (Op.getOpcode()) {
default: llvm_unreachable("Don't know how to custom lower this!");
case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
+ case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
case ISD::GlobalAddress:
return Subtarget->isTargetDarwin() ? LowerGlobalAddressDarwin(Op, DAG) :
LowerGlobalAddressELF(Op, DAG);
case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
- case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG, Subtarget);
- case ISD::BR_CC: return LowerBR_CC(Op, DAG, Subtarget);
+ case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
+ case ISD::BR_CC: return LowerBR_CC(Op, DAG);
case ISD::BR_JT: return LowerBR_JT(Op, DAG);
case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
case ISD::VASTART: return LowerVASTART(Op, DAG, VarArgsFrameIndex);
+ case ISD::MEMBARRIER: return LowerMEMBARRIER(Op, DAG, Subtarget);
case ISD::SINT_TO_FP:
case ISD::UINT_TO_FP: return LowerINT_TO_FP(Op, DAG);
case ISD::FP_TO_SINT:
case ISD::RETURNADDR: break;
case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
case ISD::GLOBAL_OFFSET_TABLE: return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
- case ISD::INTRINSIC_VOID:
- case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, DAG);
case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, DAG);
case ISD::BIT_CONVERT: return ExpandBIT_CONVERT(Op.getNode(), DAG);
case ISD::SHL:
case ISD::SRL:
case ISD::SRA: return LowerShift(Op.getNode(), DAG, Subtarget);
+ case ISD::SHL_PARTS: return LowerShiftLeftParts(Op, DAG);
+ case ISD::SRL_PARTS:
+ case ISD::SRA_PARTS: return LowerShiftRightParts(Op, DAG);
case ISD::VSETCC: return LowerVSETCC(Op, DAG);
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
// ARM Scheduler Hooks
//===----------------------------------------------------------------------===//
+MachineBasicBlock *
+ARMTargetLowering::EmitAtomicCmpSwap(MachineInstr *MI,
+ MachineBasicBlock *BB,
+ unsigned Size) const {
+ unsigned dest = MI->getOperand(0).getReg();
+ unsigned ptr = MI->getOperand(1).getReg();
+ unsigned oldval = MI->getOperand(2).getReg();
+ unsigned newval = MI->getOperand(3).getReg();
+ unsigned scratch = BB->getParent()->getRegInfo()
+ .createVirtualRegister(ARM::GPRRegisterClass);
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ DebugLoc dl = MI->getDebugLoc();
+ bool isThumb2 = Subtarget->isThumb2();
+
+ unsigned ldrOpc, strOpc;
+ switch (Size) {
+ default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
+ case 1:
+ ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
+ strOpc = isThumb2 ? ARM::t2LDREXB : ARM::STREXB;
+ break;
+ case 2:
+ ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
+ strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
+ break;
+ case 4:
+ ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
+ strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
+ break;
+ }
+
+ MachineFunction *MF = BB->getParent();
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ MachineFunction::iterator It = BB;
+ ++It; // insert the new blocks after the current block
+
+ MachineBasicBlock *loop1MBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *loop2MBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *exitMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ MF->insert(It, loop1MBB);
+ MF->insert(It, loop2MBB);
+ MF->insert(It, exitMBB);
+ exitMBB->transferSuccessors(BB);
+
+ // thisMBB:
+ // ...
+ // fallthrough --> loop1MBB
+ BB->addSuccessor(loop1MBB);
+
+ // loop1MBB:
+ // ldrex dest, [ptr]
+ // cmp dest, oldval
+ // bne exitMBB
+ BB = loop1MBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPrr : ARM::CMPrr))
+ .addReg(dest).addReg(oldval));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(exitMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ BB->addSuccessor(loop2MBB);
+ BB->addSuccessor(exitMBB);
+
+ // loop2MBB:
+ // strex scratch, newval, [ptr]
+ // cmp scratch, #0
+ // bne loop1MBB
+ BB = loop2MBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(newval)
+ .addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
+ .addReg(scratch).addImm(0));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(loop1MBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+ BB->addSuccessor(loop1MBB);
+ BB->addSuccessor(exitMBB);
+
+ // exitMBB:
+ // ...
+ BB = exitMBB;
+
+ MF->DeleteMachineInstr(MI); // The instruction is gone now.
+
+ return BB;
+}
+
+MachineBasicBlock *
+ARMTargetLowering::EmitAtomicBinary(MachineInstr *MI, MachineBasicBlock *BB,
+ unsigned Size, unsigned BinOpcode) const {
+ // This also handles ATOMIC_SWAP, indicated by BinOpcode==0.
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ const BasicBlock *LLVM_BB = BB->getBasicBlock();
+ MachineFunction *F = BB->getParent();
+ MachineFunction::iterator It = BB;
+ ++It;
+
+ unsigned dest = MI->getOperand(0).getReg();
+ unsigned ptr = MI->getOperand(1).getReg();
+ unsigned incr = MI->getOperand(2).getReg();
+ DebugLoc dl = MI->getDebugLoc();
+
+ bool isThumb2 = Subtarget->isThumb2();
+ unsigned ldrOpc, strOpc;
+ switch (Size) {
+ default: llvm_unreachable("unsupported size for AtomicCmpSwap!");
+ case 1:
+ ldrOpc = isThumb2 ? ARM::t2LDREXB : ARM::LDREXB;
+ strOpc = isThumb2 ? ARM::t2STREXB : ARM::STREXB;
+ break;
+ case 2:
+ ldrOpc = isThumb2 ? ARM::t2LDREXH : ARM::LDREXH;
+ strOpc = isThumb2 ? ARM::t2STREXH : ARM::STREXH;
+ break;
+ case 4:
+ ldrOpc = isThumb2 ? ARM::t2LDREX : ARM::LDREX;
+ strOpc = isThumb2 ? ARM::t2STREX : ARM::STREX;
+ break;
+ }
+
+ MachineBasicBlock *loopMBB = F->CreateMachineBasicBlock(LLVM_BB);
+ MachineBasicBlock *exitMBB = F->CreateMachineBasicBlock(LLVM_BB);
+ F->insert(It, loopMBB);
+ F->insert(It, exitMBB);
+ exitMBB->transferSuccessors(BB);
+
+ MachineRegisterInfo &RegInfo = F->getRegInfo();
+ unsigned scratch = RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
+ unsigned scratch2 = (!BinOpcode) ? incr :
+ RegInfo.createVirtualRegister(ARM::GPRRegisterClass);
+
+ // thisMBB:
+ // ...
+ // fallthrough --> loopMBB
+ BB->addSuccessor(loopMBB);
+
+ // loopMBB:
+ // ldrex dest, ptr
+ // <binop> scratch2, dest, incr
+ // strex scratch, scratch2, ptr
+ // cmp scratch, #0
+ // bne- loopMBB
+ // fallthrough --> exitMBB
+ BB = loopMBB;
+ AddDefaultPred(BuildMI(BB, dl, TII->get(ldrOpc), dest).addReg(ptr));
+ if (BinOpcode) {
+ // operand order needs to go the other way for NAND
+ if (BinOpcode == ARM::BICrr || BinOpcode == ARM::t2BICrr)
+ AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
+ addReg(incr).addReg(dest)).addReg(0);
+ else
+ AddDefaultPred(BuildMI(BB, dl, TII->get(BinOpcode), scratch2).
+ addReg(dest).addReg(incr)).addReg(0);
+ }
+
+ AddDefaultPred(BuildMI(BB, dl, TII->get(strOpc), scratch).addReg(scratch2)
+ .addReg(ptr));
+ AddDefaultPred(BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2CMPri : ARM::CMPri))
+ .addReg(scratch).addImm(0));
+ BuildMI(BB, dl, TII->get(isThumb2 ? ARM::t2Bcc : ARM::Bcc))
+ .addMBB(loopMBB).addImm(ARMCC::NE).addReg(ARM::CPSR);
+
+ BB->addSuccessor(loopMBB);
+ BB->addSuccessor(exitMBB);
+
+ // exitMBB:
+ // ...
+ BB = exitMBB;
+
+ F->DeleteMachineInstr(MI); // The instruction is gone now.
+
+ return BB;
+}
+
MachineBasicBlock *
ARMTargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
- MachineBasicBlock *BB) const {
+ MachineBasicBlock *BB,
+ DenseMap<MachineBasicBlock*, MachineBasicBlock*> *EM) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc dl = MI->getDebugLoc();
+ bool isThumb2 = Subtarget->isThumb2();
switch (MI->getOpcode()) {
default:
+ MI->dump();
llvm_unreachable("Unexpected instr type to insert");
+
+ case ARM::ATOMIC_LOAD_ADD_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+ case ARM::ATOMIC_LOAD_ADD_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+ case ARM::ATOMIC_LOAD_ADD_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ADDrr : ARM::ADDrr);
+
+ case ARM::ATOMIC_LOAD_AND_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+ case ARM::ATOMIC_LOAD_AND_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+ case ARM::ATOMIC_LOAD_AND_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ANDrr : ARM::ANDrr);
+
+ case ARM::ATOMIC_LOAD_OR_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+ case ARM::ATOMIC_LOAD_OR_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+ case ARM::ATOMIC_LOAD_OR_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2ORRrr : ARM::ORRrr);
+
+ case ARM::ATOMIC_LOAD_XOR_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+ case ARM::ATOMIC_LOAD_XOR_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+ case ARM::ATOMIC_LOAD_XOR_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2EORrr : ARM::EORrr);
+
+ case ARM::ATOMIC_LOAD_NAND_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+ case ARM::ATOMIC_LOAD_NAND_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+ case ARM::ATOMIC_LOAD_NAND_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2BICrr : ARM::BICrr);
+
+ case ARM::ATOMIC_LOAD_SUB_I8:
+ return EmitAtomicBinary(MI, BB, 1, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+ case ARM::ATOMIC_LOAD_SUB_I16:
+ return EmitAtomicBinary(MI, BB, 2, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+ case ARM::ATOMIC_LOAD_SUB_I32:
+ return EmitAtomicBinary(MI, BB, 4, isThumb2 ? ARM::t2SUBrr : ARM::SUBrr);
+
+ case ARM::ATOMIC_SWAP_I8: return EmitAtomicBinary(MI, BB, 1, 0);
+ case ARM::ATOMIC_SWAP_I16: return EmitAtomicBinary(MI, BB, 2, 0);
+ case ARM::ATOMIC_SWAP_I32: return EmitAtomicBinary(MI, BB, 4, 0);
+
+ case ARM::ATOMIC_CMP_SWAP_I8: return EmitAtomicCmpSwap(MI, BB, 1);
+ case ARM::ATOMIC_CMP_SWAP_I16: return EmitAtomicCmpSwap(MI, BB, 2);
+ case ARM::ATOMIC_CMP_SWAP_I32: return EmitAtomicCmpSwap(MI, BB, 4);
+
case ARM::tMOVCCr_pseudo: {
// To "insert" a SELECT_CC instruction, we actually have to insert the
// diamond control-flow pattern. The incoming instruction knows the
F->insert(It, sinkMBB);
// Update machine-CFG edges by first adding all successors of the current
// block to the new block which will contain the Phi node for the select.
- for(MachineBasicBlock::succ_iterator i = BB->succ_begin(),
- e = BB->succ_end(); i != e; ++i)
- sinkMBB->addSuccessor(*i);
+ // Also inform sdisel of the edge changes.
+ for (MachineBasicBlock::succ_iterator I = BB->succ_begin(),
+ E = BB->succ_end(); I != E; ++I) {
+ EM->insert(std::make_pair(*I, sinkMBB));
+ sinkMBB->addSuccessor(*I);
+ }
// Next, remove all successors of the current block, and add the true
// and fallthrough blocks as its successors.
- while(!BB->succ_empty())
+ while (!BB->succ_empty())
BB->removeSuccessor(BB->succ_begin());
BB->addSuccessor(copy0MBB);
BB->addSuccessor(sinkMBB);
return SDValue();
}
-
-/// PerformFMRRDCombine - Target-specific dag combine xforms for ARMISD::FMRRD.
-static SDValue PerformFMRRDCombine(SDNode *N,
+/// PerformVMOVRRDCombine - Target-specific dag combine xforms for ARMISD::VMOVRRD.
+static SDValue PerformVMOVRRDCombine(SDNode *N,
TargetLowering::DAGCombinerInfo &DCI) {
// fmrrd(fmdrr x, y) -> x,y
SDValue InDouble = N->getOperand(0);
- if (InDouble.getOpcode() == ARMISD::FMDRR)
+ if (InDouble.getOpcode() == ARMISD::VMOVDRR)
return DCI.CombineTo(N, InDouble.getOperand(0), InDouble.getOperand(1));
return SDValue();
}
default: break;
case ISD::ADD: return PerformADDCombine(N, DCI);
case ISD::SUB: return PerformSUBCombine(N, DCI);
- case ARMISD::FMRRD: return PerformFMRRDCombine(N, DCI);
+ case ARMISD::VMOVRRD: return PerformVMOVRRDCombine(N, DCI);
case ISD::INTRINSIC_WO_CHAIN:
return PerformIntrinsicCombine(N, DCI.DAG);
case ISD::SHL:
return true;
}
+/// isLegalICmpImmediate - Return true if the specified immediate is legal
+/// icmp immediate, that is the target has icmp instructions which can compare
+/// a register against the immediate without having to materialize the
+/// immediate into a register.
+bool ARMTargetLowering::isLegalICmpImmediate(int64_t Imm) const {
+ if (!Subtarget->isThumb())
+ return ARM_AM::getSOImmVal(Imm) != -1;
+ if (Subtarget->isThumb2())
+ return ARM_AM::getT2SOImmVal(Imm) != -1;
+ return Imm >= 0 && Imm <= 255;
+}
+
static bool getARMIndexedAddressParts(SDNode *Ptr, EVT VT,
bool isSEXTLoad, SDValue &Base,
SDValue &Offset, bool &isInc,
return true;
}
- // FIXME: Use FLDM / FSTM to emulate indexed FP load / store.
+ // FIXME: Use VLDM / VSTM to emulate indexed FP load / store.
return false;
}
ARMTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
EVT VT) const {
if (Constraint.size() == 1) {
- // GCC RS6000 Constraint Letters
+ // GCC ARM Constraint Letters
switch (Constraint[0]) {
case 'l':
- if (Subtarget->isThumb1Only())
+ if (Subtarget->isThumb())
return std::make_pair(0U, ARM::tGPRRegisterClass);
else
return std::make_pair(0U, ARM::GPRRegisterClass);
case 'w':
if (VT == MVT::f32)
return std::make_pair(0U, ARM::SPRRegisterClass);
- if (VT == MVT::f64)
+ if (VT.getSizeInBits() == 64)
return std::make_pair(0U, ARM::DPRRegisterClass);
+ if (VT.getSizeInBits() == 128)
+ return std::make_pair(0U, ARM::QPRRegisterClass);
break;
}
}
ARM::S20,ARM::S21,ARM::S22,ARM::S23,
ARM::S24,ARM::S25,ARM::S26,ARM::S27,
ARM::S28,ARM::S29,ARM::S30,ARM::S31, 0);
- if (VT == MVT::f64)
+ if (VT.getSizeInBits() == 64)
return make_vector<unsigned>(ARM::D0, ARM::D1, ARM::D2, ARM::D3,
ARM::D4, ARM::D5, ARM::D6, ARM::D7,
ARM::D8, ARM::D9, ARM::D10,ARM::D11,
ARM::D12,ARM::D13,ARM::D14,ARM::D15, 0);
+ if (VT.getSizeInBits() == 128)
+ return make_vector<unsigned>(ARM::Q0, ARM::Q1, ARM::Q2, ARM::Q3,
+ ARM::Q4, ARM::Q5, ARM::Q6, ARM::Q7, 0);
break;
}
return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, hasMemory,
Ops, DAG);
}
+
+bool
+ARMTargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const {
+ // The ARM target isn't yet aware of offsets.
+ return false;
+}
+
+int ARM::getVFPf32Imm(const APFloat &FPImm) {
+ APInt Imm = FPImm.bitcastToAPInt();
+ uint32_t Sign = Imm.lshr(31).getZExtValue() & 1;
+ int32_t Exp = (Imm.lshr(23).getSExtValue() & 0xff) - 127; // -126 to 127
+ int64_t Mantissa = Imm.getZExtValue() & 0x7fffff; // 23 bits
+
+ // We can handle 4 bits of mantissa.
+ // mantissa = (16+UInt(e:f:g:h))/16.
+ if (Mantissa & 0x7ffff)
+ return -1;
+ Mantissa >>= 19;
+ if ((Mantissa & 0xf) != Mantissa)
+ return -1;
+
+ // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
+ if (Exp < -3 || Exp > 4)
+ return -1;
+ Exp = ((Exp+3) & 0x7) ^ 4;
+
+ return ((int)Sign << 7) | (Exp << 4) | Mantissa;
+}
+
+int ARM::getVFPf64Imm(const APFloat &FPImm) {
+ APInt Imm = FPImm.bitcastToAPInt();
+ uint64_t Sign = Imm.lshr(63).getZExtValue() & 1;
+ int64_t Exp = (Imm.lshr(52).getSExtValue() & 0x7ff) - 1023; // -1022 to 1023
+ uint64_t Mantissa = Imm.getZExtValue() & 0xfffffffffffffLL;
+
+ // We can handle 4 bits of mantissa.
+ // mantissa = (16+UInt(e:f:g:h))/16.
+ if (Mantissa & 0xffffffffffffLL)
+ return -1;
+ Mantissa >>= 48;
+ if ((Mantissa & 0xf) != Mantissa)
+ return -1;
+
+ // We can handle 3 bits of exponent: exp == UInt(NOT(b):c:d)-3
+ if (Exp < -3 || Exp > 4)
+ return -1;
+ Exp = ((Exp+3) & 0x7) ^ 4;
+
+ return ((int)Sign << 7) | (Exp << 4) | Mantissa;
+}
+
+/// isFPImmLegal - Returns true if the target can instruction select the
+/// specified FP immediate natively. If false, the legalizer will
+/// materialize the FP immediate as a load from a constant pool.
+bool ARMTargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
+ if (!Subtarget->hasVFP3())
+ return false;
+ if (VT == MVT::f32)
+ return ARM::getVFPf32Imm(Imm) != -1;
+ if (VT == MVT::f64)
+ return ARM::getVFPf64Imm(Imm) != -1;
+ return false;
+}
projects / oota-llvm.git / blobdiff