#include "ARMConstantPoolValue.h"
#include "ARMISelLowering.h"
#include "ARMMachineFunctionInfo.h"
+#include "ARMPerfectShuffle.h"
#include "ARMRegisterInfo.h"
#include "ARMSubtarget.h"
#include "ARMTargetMachine.h"
setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
setOperationAction(ISD::SCALAR_TO_VECTOR, VT.getSimpleVT(), Custom);
setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Custom);
+ setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Expand);
if (VT.isInteger()) {
setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
setLibcallName(RTLIB::SRL_I128, 0);
setLibcallName(RTLIB::SRA_I128, 0);
+ // Libcalls should use the AAPCS base standard ABI, even if hard float
+ // is in effect, as per the ARM RTABI specification, section 4.1.2.
+ if (Subtarget->isAAPCS_ABI()) {
+ for (int i = 0; i < RTLIB::UNKNOWN_LIBCALL; ++i) {
+ setLibcallCallingConv(static_cast<RTLIB::Libcall>(i),
+ CallingConv::ARM_AAPCS);
+ }
+ }
+
if (Subtarget->isThumb1Only())
addRegisterClass(MVT::i32, ARM::tGPRRegisterClass);
else
setOperationAction(ISD::VAEND, MVT::Other, Expand);
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
+ setOperationAction(ISD::EHSELECTION, MVT::i32, Expand);
+ // FIXME: Shouldn't need this, since no register is used, but the legalizer
+ // doesn't yet know how to not do that for SjLj.
+ setExceptionSelectorRegister(ARM::R0);
if (Subtarget->isThumb())
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
else
setStackPointerRegisterToSaveRestore(ARM::SP);
setSchedulingPreference(SchedulingForRegPressure);
- setIfCvtBlockSizeLimit(Subtarget->isThumb() ? 0 : 10);
- setIfCvtDupBlockSizeLimit(Subtarget->isThumb() ? 0 : 2);
-
- if (!Subtarget->isThumb()) {
- // Use branch latency information to determine if-conversion limits.
- // FIXME: If-converter should use instruction latency of the branch being
- // eliminated to compute the threshold. For ARMv6, the branch "latency"
- // varies depending on whether it's dynamically or statically predicted
- // and on whether the destination is in the prefetch buffer.
- const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
- const InstrItineraryData &InstrItins = Subtarget->getInstrItineraryData();
- unsigned Latency= InstrItins.getLatency(TII->get(ARM::Bcc).getSchedClass());
- if (Latency > 1) {
- setIfCvtBlockSizeLimit(Latency-1);
- if (Latency > 2)
- setIfCvtDupBlockSizeLimit(Latency-2);
- } else {
- setIfCvtBlockSizeLimit(10);
- setIfCvtDupBlockSizeLimit(2);
- }
+
+ // FIXME: If-converter should use instruction latency to determine
+ // profitability rather than relying on fixed limits.
+ if (Subtarget->getCPUString() == "generic") {
+ // Generic (and overly aggressive) if-conversion limits.
+ setIfCvtBlockSizeLimit(10);
+ setIfCvtDupBlockSizeLimit(2);
+ } else if (Subtarget->hasV6Ops()) {
+ setIfCvtBlockSizeLimit(2);
+ setIfCvtDupBlockSizeLimit(1);
+ } else {
+ setIfCvtBlockSizeLimit(3);
+ setIfCvtDupBlockSizeLimit(2);
}
maxStoresPerMemcpy = 1; //// temporary - rewrite interface to use type
case ARMISD::VQRSHRNsu: return "ARMISD::VQRSHRNsu";
case ARMISD::VGETLANEu: return "ARMISD::VGETLANEu";
case ARMISD::VGETLANEs: return "ARMISD::VGETLANEs";
- case ARMISD::VDUPLANEQ: return "ARMISD::VDUPLANEQ";
+ case ARMISD::VDUP: return "ARMISD::VDUP";
+ case ARMISD::VDUPLANE: return "ARMISD::VDUPLANE";
case ARMISD::VLD2D: return "ARMISD::VLD2D";
case ARMISD::VLD3D: return "ARMISD::VLD3D";
case ARMISD::VLD4D: return "ARMISD::VLD4D";
case ARMISD::VST2D: return "ARMISD::VST2D";
case ARMISD::VST3D: return "ARMISD::VST3D";
case ARMISD::VST4D: return "ARMISD::VST4D";
+ case ARMISD::VEXT: return "ARMISD::VEXT";
+ case ARMISD::VREV64: return "ARMISD::VREV64";
+ case ARMISD::VREV32: return "ARMISD::VREV32";
+ case ARMISD::VREV16: return "ARMISD::VREV16";
+ case ARMISD::VZIP: return "ARMISD::VZIP";
+ case ARMISD::VUZP: return "ARMISD::VUZP";
+ case ARMISD::VTRN: return "ARMISD::VTRN";
}
}
// tBX takes a register source operand.
const char *Sym = S->getSymbol();
if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
- ARMConstantPoolValue *CPV = new ARMConstantPoolValue(Sym, ARMPCLabelIndex,
+ ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
+ Sym, ARMPCLabelIndex,
ARMCP::CPStub, 4);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
ArgListTy Args;
ArgListEntry Entry;
Entry.Node = Argument;
- Entry.Ty = (const Type *) Type::Int32Ty;
+ Entry.Ty = (const Type *) Type::getInt32Ty(*DAG.getContext());
Args.push_back(Entry);
// FIXME: is there useful debug info available here?
std::pair<SDValue, SDValue> CallResult =
- LowerCallTo(Chain, (const Type *) Type::Int32Ty, false, false, false, false,
+ 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);
return CallResult.first;
EVT PtrVT = getPointerTy();
DebugLoc dl = Op.getDebugLoc();
unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
- ARMConstantPoolValue *CPV = new ARMConstantPoolValue("_GLOBAL_OFFSET_TABLE_",
+ ARMConstantPoolValue *CPV = new ARMConstantPoolValue(*DAG.getContext(),
+ "_GLOBAL_OFFSET_TABLE_",
ARMPCLabelIndex,
ARMCP::CPValue, PCAdj);
SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
std::string LSDAName = "L_lsda_";
LSDAName += MF.getFunction()->getName();
ARMConstantPoolValue *CPV =
- new ARMConstantPoolValue(LSDAName.c_str(), ARMPCLabelIndex, Kind, PCAdj);
+ new ARMConstantPoolValue(*DAG.getContext(), LSDAName.c_str(),
+ ARMPCLabelIndex, Kind, PCAdj);
CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
SDValue Result =
SplatBitSize, DAG);
}
+static bool isVEXTMask(const SmallVectorImpl<int> &M, EVT VT,
+ bool &ReverseVEXT, unsigned &Imm) {
+ unsigned NumElts = VT.getVectorNumElements();
+ ReverseVEXT = false;
+ Imm = M[0];
+
+ // If this is a VEXT shuffle, the immediate value is the index of the first
+ // element. The other shuffle indices must be the successive elements after
+ // the first one.
+ unsigned ExpectedElt = Imm;
+ for (unsigned i = 1; i < NumElts; ++i) {
+ // Increment the expected index. If it wraps around, it may still be
+ // a VEXT but the source vectors must be swapped.
+ ExpectedElt += 1;
+ if (ExpectedElt == NumElts * 2) {
+ ExpectedElt = 0;
+ ReverseVEXT = true;
+ }
+
+ if (ExpectedElt != static_cast<unsigned>(M[i]))
+ return false;
+ }
+
+ // Adjust the index value if the source operands will be swapped.
+ if (ReverseVEXT)
+ Imm -= NumElts;
+
+ return true;
+}
+
/// isVREVMask - Check if a vector shuffle corresponds to a VREV
/// instruction with the specified blocksize. (The order of the elements
/// within each block of the vector is reversed.)
-bool ARM::isVREVMask(ShuffleVectorSDNode *N, unsigned BlockSize) {
+static bool isVREVMask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned BlockSize) {
assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
"Only possible block sizes for VREV are: 16, 32, 64");
- EVT VT = N->getValueType(0);
unsigned NumElts = VT.getVectorNumElements();
unsigned EltSz = VT.getVectorElementType().getSizeInBits();
- unsigned BlockElts = N->getMaskElt(0) + 1;
+ unsigned BlockElts = M[0] + 1;
if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
return false;
for (unsigned i = 0; i < NumElts; ++i) {
- if ((unsigned) N->getMaskElt(i) !=
+ if ((unsigned) M[i] !=
(i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
return false;
}
return true;
}
+static bool isVTRNMask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ 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 + NumElts + WhichResult)
+ return false;
+ }
+ return true;
+}
+
+static bool isVUZPMask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ unsigned NumElts = VT.getVectorNumElements();
+ WhichResult = (M[0] == 0 ? 0 : 1);
+ for (unsigned i = 0; i != NumElts; ++i) {
+ if ((unsigned) M[i] != 2 * i + WhichResult)
+ return false;
+ }
+
+ // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 32)
+ return false;
+
+ return true;
+}
+
+static bool isVZIPMask(const SmallVectorImpl<int> &M, EVT VT,
+ unsigned &WhichResult) {
+ 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 + NumElts)
+ return false;
+ Idx += 1;
+ }
+
+ // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
+ if (VT.is64BitVector() && VT.getVectorElementType().getSizeInBits() == 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 = dyn_cast<ConstantSDNode>(Val.getNode());
+ ConstantSDNode *ConstVal = cast<ConstantSDNode>(Val.getNode());
if (ConstVal->isNullValue())
return getZeroVector(VT, DAG, dl);
if (ConstVal->isAllOnesValue())
// If this is a case we can't handle, return null and let the default
// expansion code take care of it.
static SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) {
- BuildVectorSDNode *BVN = dyn_cast<BuildVectorSDNode>(Op.getNode());
- assert(BVN != 0 && "Expected a BuildVectorSDNode in LowerBUILD_VECTOR");
+ BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
return SDValue();
}
+/// isShuffleMaskLegal - Targets can use this to indicate that they only
+/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
+/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
+/// are assumed to be legal.
+bool
+ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
+ EVT VT) const {
+ if (VT.getVectorNumElements() == 4 &&
+ (VT.is128BitVector() || VT.is64BitVector())) {
+ unsigned PFIndexes[4];
+ for (unsigned i = 0; i != 4; ++i) {
+ if (M[i] < 0)
+ PFIndexes[i] = 8;
+ else
+ PFIndexes[i] = M[i];
+ }
+
+ // Compute the index in the perfect shuffle table.
+ unsigned PFTableIndex =
+ PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
+ unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
+ unsigned Cost = (PFEntry >> 30);
+
+ if (Cost <= 4)
+ return true;
+ }
+
+ bool ReverseVEXT;
+ unsigned Imm, WhichResult;
+
+ return (ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
+ isVREVMask(M, VT, 64) ||
+ isVREVMask(M, VT, 32) ||
+ isVREVMask(M, VT, 16) ||
+ isVEXTMask(M, VT, ReverseVEXT, Imm) ||
+ isVTRNMask(M, VT, WhichResult) ||
+ isVUZPMask(M, VT, WhichResult) ||
+ isVZIPMask(M, VT, WhichResult));
+}
+
+/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
+/// the specified operations to build the shuffle.
+static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
+ SDValue RHS, SelectionDAG &DAG,
+ DebugLoc dl) {
+ unsigned OpNum = (PFEntry >> 26) & 0x0F;
+ unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
+ unsigned RHSID = (PFEntry >> 0) & ((1 << 13)-1);
+
+ enum {
+ OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
+ OP_VREV,
+ OP_VDUP0,
+ OP_VDUP1,
+ OP_VDUP2,
+ OP_VDUP3,
+ OP_VEXT1,
+ OP_VEXT2,
+ OP_VEXT3,
+ OP_VUZPL, // VUZP, left result
+ OP_VUZPR, // VUZP, right result
+ OP_VZIPL, // VZIP, left result
+ OP_VZIPR, // VZIP, right result
+ OP_VTRNL, // VTRN, left result
+ OP_VTRNR // VTRN, right result
+ };
+
+ if (OpNum == OP_COPY) {
+ if (LHSID == (1*9+2)*9+3) return LHS;
+ assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
+ return RHS;
+ }
+
+ SDValue OpLHS, OpRHS;
+ OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
+ OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
+ EVT VT = OpLHS.getValueType();
+
+ switch (OpNum) {
+ default: llvm_unreachable("Unknown shuffle opcode!");
+ case OP_VREV:
+ return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
+ case OP_VDUP0:
+ case OP_VDUP1:
+ case OP_VDUP2:
+ case OP_VDUP3:
+ return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
+ OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
+ case OP_VEXT1:
+ case OP_VEXT2:
+ case OP_VEXT3:
+ return DAG.getNode(ARMISD::VEXT, dl, VT,
+ OpLHS, OpRHS,
+ DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
+ case OP_VUZPL:
+ case OP_VUZPR:
+ return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
+ OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
+ case OP_VZIPL:
+ case OP_VZIPR:
+ return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
+ OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
+ case OP_VTRNL:
+ case OP_VTRNR:
+ return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
+ OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
+ }
+}
+
static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
- return Op;
+ SDValue V1 = Op.getOperand(0);
+ SDValue V2 = Op.getOperand(1);
+ DebugLoc dl = Op.getDebugLoc();
+ EVT VT = Op.getValueType();
+ ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
+ SmallVector<int, 8> ShuffleMask;
+
+ // Convert shuffles that are directly supported on NEON to target-specific
+ // DAG nodes, instead of keeping them as shuffles and matching them again
+ // during code selection. This is more efficient and avoids the possibility
+ // of inconsistencies between legalization and selection.
+ // FIXME: floating-point vectors should be canonicalized to integer vectors
+ // of the same time so that they get CSEd properly.
+ SVN->getMask(ShuffleMask);
+
+ if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
+ int Lane = SVN->getSplatIndex();
+ if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
+ return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
+ }
+ return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
+ DAG.getConstant(Lane, MVT::i32));
+ }
+
+ bool ReverseVEXT;
+ unsigned Imm;
+ if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
+ if (ReverseVEXT)
+ std::swap(V1, V2);
+ return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
+ DAG.getConstant(Imm, MVT::i32));
+ }
+
+ if (isVREVMask(ShuffleMask, VT, 64))
+ return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
+ if (isVREVMask(ShuffleMask, VT, 32))
+ return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
+ if (isVREVMask(ShuffleMask, VT, 16))
+ return DAG.getNode(ARMISD::VREV16, dl, VT, V1);
+
+ // Check for Neon shuffles that modify both input vectors in place.
+ // If both results are used, i.e., if there are two shuffles with the same
+ // source operands and with masks corresponding to both results of one of
+ // these operations, DAG memoization will ensure that a single node is
+ // used for both shuffles.
+ unsigned WhichResult;
+ if (isVTRNMask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
+ V1, V2).getValue(WhichResult);
+ if (isVUZPMask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
+ V1, V2).getValue(WhichResult);
+ if (isVZIPMask(ShuffleMask, VT, WhichResult))
+ return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
+ V1, V2).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 &&
+ (VT.is128BitVector() || VT.is64BitVector())) {
+ unsigned PFIndexes[4];
+ for (unsigned i = 0; i != 4; ++i) {
+ if (ShuffleMask[i] < 0)
+ PFIndexes[i] = 8;
+ else
+ PFIndexes[i] = ShuffleMask[i];
+ }
+
+ // Compute the index in the perfect shuffle table.
+ unsigned PFTableIndex =
+ PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
+
+ unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
+ unsigned Cost = (PFEntry >> 30);
+
+ if (Cost <= 4)
+ return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
+ }
+
+ return SDValue();
}
static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
switch (MI->getOpcode()) {
default:
llvm_unreachable("Unexpected instr type to insert");
- case ARM::tMOVCCr: {
+ 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
// destination vreg to set, the condition code register to branch on, the
return SDValue();
}
+bool ARMTargetLowering::allowsUnalignedMemoryAccesses(EVT VT) const {
+ if (!Subtarget->hasV6Ops())
+ // Pre-v6 does not support unaligned mem access.
+ return false;
+ else if (!Subtarget->hasV6Ops()) {
+ // v6 may or may not support unaligned mem access.
+ if (!Subtarget->isTargetDarwin())
+ return false;
+ }
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default:
+ return false;
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ return true;
+ // FIXME: VLD1 etc with standard alignment is legal.
+ }
+}
+
+static bool isLegalT1AddressImmediate(int64_t V, EVT VT) {
+ if (V < 0)
+ return false;
+
+ unsigned Scale = 1;
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i1:
+ case MVT::i8:
+ // Scale == 1;
+ break;
+ case MVT::i16:
+ // Scale == 2;
+ Scale = 2;
+ break;
+ case MVT::i32:
+ // Scale == 4;
+ Scale = 4;
+ break;
+ }
+
+ if ((V & (Scale - 1)) != 0)
+ return false;
+ V /= Scale;
+ return V == (V & ((1LL << 5) - 1));
+}
+
+static bool isLegalT2AddressImmediate(int64_t V, EVT VT,
+ const ARMSubtarget *Subtarget) {
+ bool isNeg = false;
+ if (V < 0) {
+ isNeg = true;
+ V = - V;
+ }
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ // + imm12 or - imm8
+ if (isNeg)
+ return V == (V & ((1LL << 8) - 1));
+ return V == (V & ((1LL << 12) - 1));
+ case MVT::f32:
+ case MVT::f64:
+ // Same as ARM mode. FIXME: NEON?
+ if (!Subtarget->hasVFP2())
+ return false;
+ if ((V & 3) != 0)
+ return false;
+ V >>= 2;
+ return V == (V & ((1LL << 8) - 1));
+ }
+}
+
/// isLegalAddressImmediate - Return true if the integer value can be used
/// as the offset of the target addressing mode for load / store of the
/// given type.
if (!VT.isSimple())
return false;
- if (Subtarget->isThumb()) { // FIXME for thumb2
- if (V < 0)
- return false;
-
- unsigned Scale = 1;
- switch (VT.getSimpleVT().SimpleTy) {
- default: return false;
- case MVT::i1:
- case MVT::i8:
- // Scale == 1;
- break;
- case MVT::i16:
- // Scale == 2;
- Scale = 2;
- break;
- case MVT::i32:
- // Scale == 4;
- Scale = 4;
- break;
- }
-
- if ((V & (Scale - 1)) != 0)
- return false;
- V /= Scale;
- return V == (V & ((1LL << 5) - 1));
- }
+ if (Subtarget->isThumb1Only())
+ return isLegalT1AddressImmediate(V, VT);
+ else if (Subtarget->isThumb2())
+ return isLegalT2AddressImmediate(V, VT, Subtarget);
+ // ARM mode.
if (V < 0)
V = - V;
switch (VT.getSimpleVT().SimpleTy) {
return V == (V & ((1LL << 8) - 1));
case MVT::f32:
case MVT::f64:
- if (!Subtarget->hasVFP2())
+ if (!Subtarget->hasVFP2()) // FIXME: NEON?
return false;
if ((V & 3) != 0)
return false;
}
}
+bool ARMTargetLowering::isLegalT2ScaledAddressingMode(const AddrMode &AM,
+ EVT VT) const {
+ int Scale = AM.Scale;
+ if (Scale < 0)
+ return false;
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return false;
+ case MVT::i1:
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ if (Scale == 1)
+ return true;
+ // r + r << imm
+ Scale = Scale & ~1;
+ return Scale == 2 || Scale == 4 || Scale == 8;
+ case MVT::i64:
+ // r + r
+ if (((unsigned)AM.HasBaseReg + Scale) <= 2)
+ return true;
+ return false;
+ case MVT::isVoid:
+ // Note, we allow "void" uses (basically, uses that aren't loads or
+ // stores), because arm allows folding a scale into many arithmetic
+ // operations. This should be made more precise and revisited later.
+
+ // Allow r << imm, but the imm has to be a multiple of two.
+ if (Scale & 1) return false;
+ return isPowerOf2_32(Scale);
+ }
+}
+
/// isLegalAddressingMode - Return true if the addressing mode represented
/// by AM is legal for this target, for a load/store of the specified type.
bool ARMTargetLowering::isLegalAddressingMode(const AddrMode &AM,
case 0: // no scale reg, must be "r+i" or "r", or "i".
break;
case 1:
- if (Subtarget->isThumb()) // FIXME for thumb2
+ if (Subtarget->isThumb1Only())
return false;
// FALL THROUGH.
default:
if (!VT.isSimple())
return false;
+ if (Subtarget->isThumb2())
+ return isLegalT2ScaledAddressingMode(AM, VT);
+
int Scale = AM.Scale;
switch (VT.getSimpleVT().SimpleTy) {
default: return false;
case MVT::i1:
case MVT::i8:
case MVT::i32:
- case MVT::i64:
- // This assumes i64 is legalized to a pair of i32. If not (i.e.
- // ldrd / strd are used, then its address mode is same as i16.
- // r + r
if (Scale < 0) Scale = -Scale;
if (Scale == 1)
return true;
// r + r << imm
return isPowerOf2_32(Scale & ~1);
case MVT::i16:
+ case MVT::i64:
// r + r
if (((unsigned)AM.HasBaseReg + Scale) <= 2)
return true;
// operations. This should be made more precise and revisited later.
// Allow r << imm, but the imm has to be a multiple of two.
- if (AM.Scale & 1) return false;
- return isPowerOf2_32(AM.Scale);
+ if (Scale & 1) return false;
+ return isPowerOf2_32(Scale);
}
break;
}
bool isInc;
bool isLegal = false;
- if (Subtarget->isThumb() && Subtarget->hasThumb2())
+ if (Subtarget->isThumb2())
isLegal = getT2IndexedAddressParts(Ptr.getNode(), VT, isSEXTLoad, Base,
Offset, isInc, DAG);
else
bool isInc;
bool isLegal = false;
- if (Subtarget->isThumb() && Subtarget->hasThumb2())
+ if (Subtarget->isThumb2())
isLegal = getT2IndexedAddressParts(Op, VT, isSEXTLoad, Base, Offset,
isInc, DAG);
else