#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86ISelLowering.h"
-#include "X86ShuffleDecode.h"
#include "X86TargetMachine.h"
#include "X86TargetObjectFile.h"
+#include "Utils/X86ShuffleDecode.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
#include "llvm/Intrinsics.h"
#include "llvm/LLVMContext.h"
+#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
STATISTIC(NumTailCalls, "Number of tail calls");
static cl::opt<bool>
-DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
+Disable256Bit("disable-256bit", cl::Hidden,
+ cl::desc("Disable use of 256-bit vectors"));
// Forward declarations.
static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
SDValue V2);
-static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
+static SDValue Insert128BitVector(SDValue Result,
+ SDValue Vec,
+ SDValue Idx,
+ SelectionDAG &DAG,
+ DebugLoc dl);
+
+static SDValue Extract128BitVector(SDValue Vec,
+ SDValue Idx,
+ SelectionDAG &DAG,
+ DebugLoc dl);
+
+static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG);
+
+
+/// Generate a DAG to grab 128-bits from a vector > 128 bits. This
+/// sets things up to match to an AVX VEXTRACTF128 instruction or a
+/// simple subregister reference. Idx is an index in the 128 bits we
+/// want. It need not be aligned to a 128-bit bounday. That makes
+/// lowering EXTRACT_VECTOR_ELT operations easier.
+static SDValue Extract128BitVector(SDValue Vec,
+ SDValue Idx,
+ SelectionDAG &DAG,
+ DebugLoc dl) {
+ EVT VT = Vec.getValueType();
+ assert(VT.getSizeInBits() == 256 && "Unexpected vector size!");
+
+ EVT ElVT = VT.getVectorElementType();
+
+ int Factor = VT.getSizeInBits() / 128;
+
+ EVT ResultVT = EVT::getVectorVT(*DAG.getContext(),
+ ElVT,
+ VT.getVectorNumElements() / Factor);
+
+ // Extract from UNDEF is UNDEF.
+ if (Vec.getOpcode() == ISD::UNDEF)
+ return DAG.getNode(ISD::UNDEF, dl, ResultVT);
+
+ if (isa<ConstantSDNode>(Idx)) {
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+ // Extract the relevant 128 bits. Generate an EXTRACT_SUBVECTOR
+ // we can match to VEXTRACTF128.
+ unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
+
+ // This is the index of the first element of the 128-bit chunk
+ // we want.
+ unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
+ * ElemsPerChunk);
+
+ SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
+
+ SDValue Result = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec,
+ VecIdx);
+
+ return Result;
+ }
+
+ return SDValue();
+}
- bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
+/// Generate a DAG to put 128-bits into a vector > 128 bits. This
+/// sets things up to match to an AVX VINSERTF128 instruction or a
+/// simple superregister reference. Idx is an index in the 128 bits
+/// we want. It need not be aligned to a 128-bit bounday. That makes
+/// lowering INSERT_VECTOR_ELT operations easier.
+static SDValue Insert128BitVector(SDValue Result,
+ SDValue Vec,
+ SDValue Idx,
+ SelectionDAG &DAG,
+ DebugLoc dl) {
+ if (isa<ConstantSDNode>(Idx)) {
+ EVT VT = Vec.getValueType();
+ assert(VT.getSizeInBits() == 128 && "Unexpected vector size!");
- if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
- if (is64Bit) return new X8664_MachoTargetObjectFile();
+ EVT ElVT = VT.getVectorElementType();
+
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+ EVT ResultVT = Result.getValueType();
+
+ // Insert the relevant 128 bits.
+ unsigned ElemsPerChunk = 128 / ElVT.getSizeInBits();
+
+ // This is the index of the first element of the 128-bit chunk
+ // we want.
+ unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / 128)
+ * ElemsPerChunk);
+
+ SDValue VecIdx = DAG.getConstant(NormalizedIdxVal, MVT::i32);
+
+ Result = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec,
+ VecIdx);
+ return Result;
+ }
+
+ return SDValue();
+}
+
+/// Given two vectors, concat them.
+static SDValue ConcatVectors(SDValue Lower, SDValue Upper, SelectionDAG &DAG) {
+ DebugLoc dl = Lower.getDebugLoc();
+
+ assert(Lower.getValueType() == Upper.getValueType() && "Mismatched vectors!");
+
+ EVT VT = EVT::getVectorVT(*DAG.getContext(),
+ Lower.getValueType().getVectorElementType(),
+ Lower.getValueType().getVectorNumElements() * 2);
+
+ // TODO: Generalize to arbitrary vector length (this assumes 256-bit vectors).
+ assert(VT.getSizeInBits() == 256 && "Unsupported vector concat!");
+
+ // Insert the upper subvector.
+ SDValue Vec = Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, VT), Upper,
+ DAG.getConstant(
+ // This is half the length of the result
+ // vector. Start inserting the upper 128
+ // bits here.
+ Lower.getValueType().getVectorNumElements(),
+ MVT::i32),
+ DAG, dl);
+
+ // Insert the lower subvector.
+ Vec = Insert128BitVector(Vec, Lower, DAG.getConstant(0, MVT::i32), DAG, dl);
+ return Vec;
+}
+
+static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
+ const X86Subtarget *Subtarget = &TM.getSubtarget<X86Subtarget>();
+ bool is64Bit = Subtarget->is64Bit();
+
+ if (Subtarget->isTargetEnvMacho()) {
+ if (is64Bit)
+ return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
- } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
- if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
+ }
+
+ if (Subtarget->isTargetELF()) {
+ if (is64Bit)
+ return new X8664_ELFTargetObjectFile(TM);
return new X8632_ELFTargetObjectFile(TM);
- } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
- return new TargetLoweringObjectFileCOFF();
}
+ if (Subtarget->isTargetCOFF() && !Subtarget->isTargetEnvMacho())
+ return new TargetLoweringObjectFileCOFF();
llvm_unreachable("unknown subtarget type");
}
// (low) operations are left as Legal, as there are single-result
// instructions for this in x86. Using the two-result multiply instructions
// when both high and low results are needed must be arranged by dagcombine.
- setOperationAction(ISD::MULHS , MVT::i8 , Expand);
- setOperationAction(ISD::MULHU , MVT::i8 , Expand);
- setOperationAction(ISD::SDIV , MVT::i8 , Expand);
- setOperationAction(ISD::UDIV , MVT::i8 , Expand);
- setOperationAction(ISD::SREM , MVT::i8 , Expand);
- setOperationAction(ISD::UREM , MVT::i8 , Expand);
- setOperationAction(ISD::MULHS , MVT::i16 , Expand);
- setOperationAction(ISD::MULHU , MVT::i16 , Expand);
- setOperationAction(ISD::SDIV , MVT::i16 , Expand);
- setOperationAction(ISD::UDIV , MVT::i16 , Expand);
- setOperationAction(ISD::SREM , MVT::i16 , Expand);
- setOperationAction(ISD::UREM , MVT::i16 , Expand);
- setOperationAction(ISD::MULHS , MVT::i32 , Expand);
- setOperationAction(ISD::MULHU , MVT::i32 , Expand);
- setOperationAction(ISD::SDIV , MVT::i32 , Expand);
- setOperationAction(ISD::UDIV , MVT::i32 , Expand);
- setOperationAction(ISD::SREM , MVT::i32 , Expand);
- setOperationAction(ISD::UREM , MVT::i32 , Expand);
- setOperationAction(ISD::MULHS , MVT::i64 , Expand);
- setOperationAction(ISD::MULHU , MVT::i64 , Expand);
- setOperationAction(ISD::SDIV , MVT::i64 , Expand);
- setOperationAction(ISD::UDIV , MVT::i64 , Expand);
- setOperationAction(ISD::SREM , MVT::i64 , Expand);
- setOperationAction(ISD::UREM , MVT::i64 , Expand);
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::MULHS, VT, Expand);
+ setOperationAction(ISD::MULHU, VT, Expand);
+ setOperationAction(ISD::SDIV, VT, Expand);
+ setOperationAction(ISD::UDIV, VT, Expand);
+ setOperationAction(ISD::SREM, VT, Expand);
+ setOperationAction(ISD::UREM, VT, Expand);
+
+ // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
+ setOperationAction(ISD::ADDC, VT, Custom);
+ setOperationAction(ISD::ADDE, VT, Custom);
+ setOperationAction(ISD::SUBC, VT, Custom);
+ setOperationAction(ISD::SUBE, VT, Custom);
+ }
setOperationAction(ISD::BR_JT , MVT::Other, Expand);
setOperationAction(ISD::BRCOND , MVT::Other, Custom);
setOperationAction(ISD::SELECT , MVT::i1 , Promote);
// X86 wants to expand cmov itself.
setOperationAction(ISD::SELECT , MVT::i8 , Custom);
- setOperationAction(ISD::SELECT , MVT::i16 , Custom);
+ setOperationAction(ISD::SELECT , MVT::i16 , Custom);
setOperationAction(ISD::SELECT , MVT::i32 , Custom);
setOperationAction(ISD::SELECT , MVT::f32 , Custom);
setOperationAction(ISD::SELECT , MVT::f64 , Custom);
setShouldFoldAtomicFences(true);
// Expand certain atomics
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
-
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
+ }
if (!Subtarget->is64Bit()) {
setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::VECTOR_SHUFFLE, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::EXTRACT_VECTOR_ELT,(MVT::SimpleValueType)VT,Expand);
- setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
setOperationAction(ISD::INSERT_VECTOR_ELT,(MVT::SimpleValueType)VT, Expand);
+ setOperationAction(ISD::EXTRACT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
+ setOperationAction(ISD::INSERT_SUBVECTOR,(MVT::SimpleValueType)VT,Expand);
setOperationAction(ISD::FABS, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FSIN, (MVT::SimpleValueType)VT, Expand);
setOperationAction(ISD::FCOS, (MVT::SimpleValueType)VT, Expand);
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
// with -msoft-float, disable use of MMX as well.
- if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
+ if (!UseSoftFloat && Subtarget->hasMMX()) {
addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
// No operations on x86mmx supported, everything uses intrinsics.
}
setOperationAction(ISD::LOAD, MVT::v8i32, Legal);
setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
+
setOperationAction(ISD::FADD, MVT::v8f32, Legal);
setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v8f32, Custom);
- //setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8f32, Custom);
- //setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8f32, Custom);
- //setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
- //setOperationAction(ISD::VSETCC, MVT::v8f32, Custom);
-
- // Operations to consider commented out -v16i16 v32i8
- //setOperationAction(ISD::ADD, MVT::v16i16, Legal);
- setOperationAction(ISD::ADD, MVT::v8i32, Custom);
- setOperationAction(ISD::ADD, MVT::v4i64, Custom);
- //setOperationAction(ISD::SUB, MVT::v32i8, Legal);
- //setOperationAction(ISD::SUB, MVT::v16i16, Legal);
- setOperationAction(ISD::SUB, MVT::v8i32, Custom);
- setOperationAction(ISD::SUB, MVT::v4i64, Custom);
- //setOperationAction(ISD::MUL, MVT::v16i16, Legal);
+
setOperationAction(ISD::FADD, MVT::v4f64, Legal);
setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
- setOperationAction(ISD::VSETCC, MVT::v4f64, Custom);
- // setOperationAction(ISD::VSETCC, MVT::v32i8, Custom);
- // setOperationAction(ISD::VSETCC, MVT::v16i16, Custom);
- setOperationAction(ISD::VSETCC, MVT::v8i32, Custom);
-
- // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v32i8, Custom);
- // setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i16, Custom);
- // setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i16, Custom);
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i32, Custom);
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8f32, Custom);
-
- setOperationAction(ISD::BUILD_VECTOR, MVT::v4f64, Custom);
- setOperationAction(ISD::BUILD_VECTOR, MVT::v4i64, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f64, Custom);
- setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i64, Custom);
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f64, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f64, Custom);
-
-#if 0
- // Not sure we want to do this since there are no 256-bit integer
- // operations in AVX
-
- // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
- // This includes 256-bit vectors
- for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; ++i) {
- EVT VT = (MVT::SimpleValueType)i;
-
- // Do not attempt to custom lower non-power-of-2 vectors
- if (!isPowerOf2_32(VT.getVectorNumElements()))
+ // Custom lower build_vector, vector_shuffle, scalar_to_vector,
+ // insert_vector_elt extract_subvector and extract_vector_elt for
+ // 256-bit types.
+ for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
+ ++i) {
+ MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
+ // Do not attempt to custom lower non-256-bit vectors
+ if (!isPowerOf2_32(MVT(VT).getVectorNumElements())
+ || (MVT(VT).getSizeInBits() < 256))
continue;
-
setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
- }
+ setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
+ }
+ // Custom-lower insert_subvector and extract_subvector based on
+ // the result type.
+ for (unsigned i = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ i <= (unsigned)MVT::LAST_VECTOR_VALUETYPE;
+ ++i) {
+ MVT::SimpleValueType VT = (MVT::SimpleValueType)i;
+ // Do not attempt to custom lower non-256-bit vectors
+ if (!isPowerOf2_32(MVT(VT).getVectorNumElements()))
+ continue;
- if (Subtarget->is64Bit()) {
- setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i64, Custom);
- setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i64, Custom);
+ if (MVT(VT).getSizeInBits() == 128) {
+ setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
+ }
+ else if (MVT(VT).getSizeInBits() == 256) {
+ setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
+ }
}
-#endif
-#if 0
- // Not sure we want to do this since there are no 256-bit integer
- // operations in AVX
-
- // Promote v32i8, v16i16, v8i32 load, select, and, or, xor to v4i64.
- // Including 256-bit vectors
- for (unsigned i = (unsigned)MVT::v16i8; i != (unsigned)MVT::v4i64; i++) {
- EVT VT = (MVT::SimpleValueType)i;
+ // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
+ // Don't promote loads because we need them for VPERM vector index versions.
- if (!VT.is256BitVector()) {
+ for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
+ VT != (unsigned)MVT::LAST_VECTOR_VALUETYPE;
+ VT++) {
+ if (!isPowerOf2_32(MVT((MVT::SimpleValueType)VT).getVectorNumElements())
+ || (MVT((MVT::SimpleValueType)VT).getSizeInBits() < 256))
continue;
- }
- setOperationAction(ISD::AND, VT, Promote);
- AddPromotedToType (ISD::AND, VT, MVT::v4i64);
- setOperationAction(ISD::OR, VT, Promote);
- AddPromotedToType (ISD::OR, VT, MVT::v4i64);
- setOperationAction(ISD::XOR, VT, Promote);
- AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
- setOperationAction(ISD::LOAD, VT, Promote);
- AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
- setOperationAction(ISD::SELECT, VT, Promote);
- AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
+ setOperationAction(ISD::AND, (MVT::SimpleValueType)VT, Promote);
+ AddPromotedToType (ISD::AND, (MVT::SimpleValueType)VT, MVT::v4i64);
+ setOperationAction(ISD::OR, (MVT::SimpleValueType)VT, Promote);
+ AddPromotedToType (ISD::OR, (MVT::SimpleValueType)VT, MVT::v4i64);
+ setOperationAction(ISD::XOR, (MVT::SimpleValueType)VT, Promote);
+ AddPromotedToType (ISD::XOR, (MVT::SimpleValueType)VT, MVT::v4i64);
+ //setOperationAction(ISD::LOAD, (MVT::SimpleValueType)VT, Promote);
+ //AddPromotedToType (ISD::LOAD, (MVT::SimpleValueType)VT, MVT::v4i64);
+ setOperationAction(ISD::SELECT, (MVT::SimpleValueType)VT, Promote);
+ AddPromotedToType (ISD::SELECT, (MVT::SimpleValueType)VT, MVT::v4i64);
}
-
- setTruncStoreAction(MVT::f64, MVT::f32, Expand);
-#endif
}
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
-
+
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
// handle type legalization for these operations here.
//
setOperationAction(ISD::SMULO, VT, Custom);
setOperationAction(ISD::UMULO, VT, Custom);
}
-
+
// There are no 8-bit 3-address imul/mul instructions
setOperationAction(ISD::SMULO, MVT::i8, Expand);
setOperationAction(ISD::UMULO, MVT::i8, Expand);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::OR);
setTargetDAGCombine(ISD::AND);
+ setTargetDAGCombine(ISD::ADD);
+ setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
computeRegisterProperties();
- // FIXME: These should be based on subtarget info. Plus, the values should
- // be smaller when we are in optimizing for size mode.
+ // On Darwin, -Os means optimize for size without hurting performance,
+ // do not reduce the limit.
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
+ maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
- maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
+ maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
}
return F->hasFnAttr(Attribute::OptimizeForSize) ? 0 : 4;
}
+// FIXME: Why this routine is here? Move to RegInfo!
std::pair<const TargetRegisterClass*, uint8_t>
X86TargetLowering::findRepresentativeClass(EVT VT) const{
const TargetRegisterClass *RRC = 0;
return std::make_pair(RRC, Cost);
}
+// FIXME: Why this routine is here? Move to RegInfo!
unsigned
X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
MachineFunction &MF) const {
- const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
+ const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
switch (RC->getID()) {
if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
SDValue Ops[] = { Chain, InFlag };
- Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
+ Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
Ops, 2), 1);
Val = Chain.getValue(0);
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
+
+ // Allocate shadow area for Win64
+ if (IsWin64) {
+ CCInfo.AllocateStack(32, 8);
+ }
+
CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
unsigned LastVal = ~0U;
TotalNumXMMRegs = 0;
if (IsWin64) {
- const TargetFrameInfo &TFI = *getTargetMachine().getFrameInfo();
+ const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
// Get to the caller-allocated home save location. Add 8 to account
// for the return address.
int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
DebugLoc dl, SelectionDAG &DAG,
const CCValAssign &VA,
ISD::ArgFlagsTy Flags) const {
- const unsigned FirstStackArgOffset = (Subtarget->isTargetWin64() ? 32 : 0);
- unsigned LocMemOffset = FirstStackArgOffset + VA.getLocMemOffset();
+ unsigned LocMemOffset = VA.getLocMemOffset();
SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
if (Flags.isByVal())
SmallVectorImpl<SDValue> &InVals) const {
MachineFunction &MF = DAG.getMachineFunction();
bool Is64Bit = Subtarget->is64Bit();
+ bool IsWin64 = Subtarget->isTargetWin64();
bool IsStructRet = CallIsStructReturn(Outs);
bool IsSibcall = false;
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
+
+ // Allocate shadow area for Win64
+ if (IsWin64) {
+ CCInfo.AllocateStack(32, 8);
+ }
+
CCInfo.AnalyzeCallOperands(Outs, CC_X86);
// Get a count of how many bytes are to be pushed on the stack.
if (VA.isRegLoc()) {
RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
- if (isVarArg && Subtarget->isTargetWin64()) {
+ if (isVarArg && IsWin64) {
// Win64 ABI requires argument XMM reg to be copied to the corresponding
// shadow reg if callee is a varargs function.
unsigned ShadowReg = 0;
}
}
- if (Is64Bit && isVarArg && !Subtarget->isTargetWin64()) {
+ if (Is64Bit && isVarArg && !IsWin64) {
// From AMD64 ABI document:
// For calls that may call functions that use varargs or stdargs
// (prototype-less calls or calls to functions containing ellipsis (...) in
}
// Returns a chain & a flag for retval copy to use.
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
SmallVector<SDValue, 8> Ops;
if (!IsSibcall && isTailCall) {
Ops.push_back(DAG.getRegister(X86::EBX, getPointerTy()));
// Add an implicit use of AL for non-Windows x86 64-bit vararg functions.
- if (Is64Bit && isVarArg && !Subtarget->isTargetWin64())
+ if (Is64Bit && isVarArg && !IsWin64)
Ops.push_back(DAG.getRegister(X86::AL, MVT::i8));
if (InFlag.getNode())
SelectionDAG& DAG) const {
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const TargetFrameInfo &TFI = *TM.getFrameInfo();
+ const TargetFrameLowering &TFI = *TM.getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
uint64_t AlignMask = StackAlignment - 1;
int64_t Offset = StackSize;
int FI = INT_MAX;
if (Arg.getOpcode() == ISD::CopyFromReg) {
unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
- if (!VR || TargetRegisterInfo::isPhysicalRegister(VR))
+ if (!TargetRegisterInfo::isVirtualRegister(VR))
return false;
MachineInstr *Def = MRI->getVRegDef(VR);
if (!Def)
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
+
+ // Allocate shadow area for Win64
+ if (Subtarget->isTargetWin64()) {
+ CCInfo.AllocateStack(32, 8);
+ }
+
CCInfo.AnalyzeCallOperands(Outs, CC_X86);
if (CCInfo.getNextStackOffset()) {
MachineFunction &MF = DAG.getMachineFunction();
if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
return false;
- if (Subtarget->isTargetWin64())
- // Win64 ABI has additional complications.
- return false;
// Check if the arguments are already laid out in the right way as
// the caller's fixed stack objects.
// First determine if it is required or is profitable to flip the operands.
// If LHS is a foldable load, but RHS is not, flip the condition.
- if ((ISD::isNON_EXTLoad(LHS.getNode()) && LHS.hasOneUse()) &&
- !(ISD::isNON_EXTLoad(RHS.getNode()) && RHS.hasOneUse())) {
+ if (ISD::isNON_EXTLoad(LHS.getNode()) &&
+ !ISD::isNON_EXTLoad(RHS.getNode())) {
SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
std::swap(LHS, RHS);
}
return true;
}
+/// isVEXTRACTF128Index - Return true if the specified
+/// EXTRACT_SUBVECTOR operand specifies a vector extract that is
+/// suitable for input to VEXTRACTF128.
+bool X86::isVEXTRACTF128Index(SDNode *N) {
+ if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
+ return false;
+
+ // The index should be aligned on a 128-bit boundary.
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
+
+ unsigned VL = N->getValueType(0).getVectorNumElements();
+ unsigned VBits = N->getValueType(0).getSizeInBits();
+ unsigned ElSize = VBits / VL;
+ bool Result = (Index * ElSize) % 128 == 0;
+
+ return Result;
+}
+
+/// isVINSERTF128Index - Return true if the specified INSERT_SUBVECTOR
+/// operand specifies a subvector insert that is suitable for input to
+/// VINSERTF128.
+bool X86::isVINSERTF128Index(SDNode *N) {
+ if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
+ return false;
+
+ // The index should be aligned on a 128-bit boundary.
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
+
+ unsigned VL = N->getValueType(0).getVectorNumElements();
+ unsigned VBits = N->getValueType(0).getSizeInBits();
+ unsigned ElSize = VBits / VL;
+ bool Result = (Index * ElSize) % 128 == 0;
+
+ return Result;
+}
+
/// getShuffleSHUFImmediate - Return the appropriate immediate to shuffle
/// the specified VECTOR_SHUFFLE mask with PSHUF* and SHUFP* instructions.
unsigned X86::getShuffleSHUFImmediate(SDNode *N) {
return (Val - i) * EltSize;
}
+/// getExtractVEXTRACTF128Immediate - Return the appropriate immediate
+/// to extract the specified EXTRACT_SUBVECTOR index with VEXTRACTF128
+/// instructions.
+unsigned X86::getExtractVEXTRACTF128Immediate(SDNode *N) {
+ if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
+ llvm_unreachable("Illegal extract subvector for VEXTRACTF128");
+
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
+
+ EVT VecVT = N->getOperand(0).getValueType();
+ EVT ElVT = VecVT.getVectorElementType();
+
+ unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
+
+ return Index / NumElemsPerChunk;
+}
+
+/// getInsertVINSERTF128Immediate - Return the appropriate immediate
+/// to insert at the specified INSERT_SUBVECTOR index with VINSERTF128
+/// instructions.
+unsigned X86::getInsertVINSERTF128Immediate(SDNode *N) {
+ if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
+ llvm_unreachable("Illegal insert subvector for VINSERTF128");
+
+ uint64_t Index =
+ cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
+
+ EVT VecVT = N->getValueType(0);
+ EVT ElVT = VecVT.getVectorElementType();
+
+ unsigned NumElemsPerChunk = 128 / ElVT.getSizeInBits();
+
+ return Index / NumElemsPerChunk;
+}
+
/// isZeroNode - Returns true if Elt is a constant zero or a floating point
/// constant +0.0.
bool X86::isZeroNode(SDValue Elt) {
if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
FI = FINode->getIndex();
Offset = 0;
- } else if (Ptr.getOpcode() == ISD::ADD &&
- isa<ConstantSDNode>(Ptr.getOperand(1)) &&
+ } else if (DAG.isBaseWithConstantOffset(Ptr) &&
isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
Offset = Ptr.getConstantOperandVal(1);
SDValue
X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
DebugLoc dl = Op.getDebugLoc();
+
+ EVT VT = Op.getValueType();
+ EVT ExtVT = VT.getVectorElementType();
+
+ unsigned NumElems = Op.getNumOperands();
+
+ // For AVX-length vectors, build the individual 128-bit pieces and
+ // use shuffles to put them in place.
+ if (VT.getSizeInBits() > 256 &&
+ Subtarget->hasAVX() &&
+ !Disable256Bit &&
+ !ISD::isBuildVectorAllZeros(Op.getNode())) {
+ SmallVector<SDValue, 8> V;
+ V.resize(NumElems);
+ for (unsigned i = 0; i < NumElems; ++i) {
+ V[i] = Op.getOperand(i);
+ }
+
+ EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
+
+ // Build the lower subvector.
+ SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[0], NumElems/2);
+ // Build the upper subvector.
+ SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT, &V[NumElems / 2],
+ NumElems/2);
+
+ return ConcatVectors(Lower, Upper, DAG);
+ }
+
// All zero's are handled with pxor in SSE2 and above, xorps in SSE1.
// All one's are handled with pcmpeqd. In AVX, zero's are handled with
// vpxor in 128-bit and xor{pd,ps} in 256-bit, but no 256 version of pcmpeqd
return getZeroVector(Op.getValueType(), Subtarget->hasSSE2(), DAG, dl);
}
- EVT VT = Op.getValueType();
- EVT ExtVT = VT.getVectorElementType();
unsigned EVTBits = ExtVT.getSizeInBits();
- unsigned NumElems = Op.getNumOperands();
unsigned NumZero = 0;
unsigned NumNonZero = 0;
unsigned NonZeros = 0;
if (MayFoldVectorLoad(V1) && MayFoldIntoStore(Op))
CanFoldLoad = true;
+ // Both of them can't be memory operations though.
+ if (MayFoldVectorLoad(V1) && MayFoldVectorLoad(V2))
+ CanFoldLoad = false;
+
if (CanFoldLoad) {
if (HasSSE2 && NumElems == 2)
return getTargetShuffleNode(X86ISD::MOVLPD, dl, VT, V1, V2, DAG);
case MVT::v16i8: return X86ISD::PUNPCKLBW;
case MVT::v8i16: return X86ISD::PUNPCKLWD;
default:
- llvm_unreachable("Unknow type for unpckl");
+ llvm_unreachable("Unknown type for unpckl");
}
return 0;
}
case MVT::v16i8: return X86ISD::PUNPCKHBW;
case MVT::v8i16: return X86ISD::PUNPCKHWD;
default:
- llvm_unreachable("Unknow type for unpckh");
+ llvm_unreachable("Unknown type for unpckh");
}
return 0;
}
if (!isa<ConstantSDNode>(Op.getOperand(1)))
return SDValue();
+ SDValue Vec = Op.getOperand(0);
+ EVT VecVT = Vec.getValueType();
+
+ // If this is a 256-bit vector result, first extract the 128-bit
+ // vector and then extract from the 128-bit vector.
+ if (VecVT.getSizeInBits() > 128) {
+ DebugLoc dl = Op.getNode()->getDebugLoc();
+ unsigned NumElems = VecVT.getVectorNumElements();
+ SDValue Idx = Op.getOperand(1);
+
+ if (!isa<ConstantSDNode>(Idx))
+ return SDValue();
+
+ unsigned ExtractNumElems = NumElems / (VecVT.getSizeInBits() / 128);
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+ // Get the 128-bit vector.
+ bool Upper = IdxVal >= ExtractNumElems;
+ Vec = Extract128BitVector(Vec, Idx, DAG, dl);
+
+ // Extract from it.
+ SDValue ScaledIdx = Idx;
+ if (Upper)
+ ScaledIdx = DAG.getNode(ISD::SUB, dl, Idx.getValueType(), Idx,
+ DAG.getConstant(ExtractNumElems,
+ Idx.getValueType()));
+ return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
+ ScaledIdx);
+ }
+
+ assert(Vec.getValueSizeInBits() <= 128 && "Unexpected vector length");
+
if (Subtarget->hasSSE41()) {
SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG);
if (Res.getNode())
EVT VT = Op.getValueType();
EVT EltVT = VT.getVectorElementType();
+ DebugLoc dl = Op.getDebugLoc();
+ SDValue N0 = Op.getOperand(0);
+ SDValue N1 = Op.getOperand(1);
+ SDValue N2 = Op.getOperand(2);
+
+ // If this is a 256-bit vector result, first insert into a 128-bit
+ // vector and then insert into the 256-bit vector.
+ if (VT.getSizeInBits() > 128) {
+ if (!isa<ConstantSDNode>(N2))
+ return SDValue();
+
+ // Get the 128-bit vector.
+ unsigned NumElems = VT.getVectorNumElements();
+ unsigned IdxVal = cast<ConstantSDNode>(N2)->getZExtValue();
+ bool Upper = IdxVal >= NumElems / 2;
+
+ SDValue SubN0 = Extract128BitVector(N0, N2, DAG, dl);
+
+ // Insert into it.
+ SDValue ScaledN2 = N2;
+ if (Upper)
+ ScaledN2 = DAG.getNode(ISD::SUB, dl, N2.getValueType(), N2,
+ DAG.getConstant(NumElems /
+ (VT.getSizeInBits() / 128),
+ N2.getValueType()));
+ Op = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, SubN0.getValueType(), SubN0,
+ N1, ScaledN2);
+
+ // Insert the 128-bit vector
+ // FIXME: Why UNDEF?
+ return Insert128BitVector(N0, Op, N2, DAG, dl);
+ }
+
if (Subtarget->hasSSE41())
return LowerINSERT_VECTOR_ELT_SSE4(Op, DAG);
if (EltVT == MVT::i8)
return SDValue();
- DebugLoc dl = Op.getDebugLoc();
- SDValue N0 = Op.getOperand(0);
- SDValue N1 = Op.getOperand(1);
- SDValue N2 = Op.getOperand(2);
-
if (EltVT.getSizeInBits() == 16 && isa<ConstantSDNode>(N2)) {
// Transform it so it match pinsrw which expects a 16-bit value in a GR32
// as its second argument.
SDValue
X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
+ LLVMContext *Context = DAG.getContext();
DebugLoc dl = Op.getDebugLoc();
+ EVT OpVT = Op.getValueType();
+
+ // If this is a 256-bit vector result, first insert into a 128-bit
+ // vector and then insert into the 256-bit vector.
+ if (OpVT.getSizeInBits() > 128) {
+ // Insert into a 128-bit vector.
+ EVT VT128 = EVT::getVectorVT(*Context,
+ OpVT.getVectorElementType(),
+ OpVT.getVectorNumElements() / 2);
+
+ Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
+
+ // Insert the 128-bit vector.
+ return Insert128BitVector(DAG.getNode(ISD::UNDEF, dl, OpVT), Op,
+ DAG.getConstant(0, MVT::i32),
+ DAG, dl);
+ }
if (Op.getValueType() == MVT::v1i64 &&
Op.getOperand(0).getValueType() == MVT::i64)
DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
}
+// Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
+// a simple subregister reference or explicit instructions to grab
+// upper bits of a vector.
+SDValue
+X86TargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
+ if (Subtarget->hasAVX()) {
+ DebugLoc dl = Op.getNode()->getDebugLoc();
+ SDValue Vec = Op.getNode()->getOperand(0);
+ SDValue Idx = Op.getNode()->getOperand(1);
+
+ if (Op.getNode()->getValueType(0).getSizeInBits() == 128
+ && Vec.getNode()->getValueType(0).getSizeInBits() == 256) {
+ return Extract128BitVector(Vec, Idx, DAG, dl);
+ }
+ }
+ return SDValue();
+}
+
+// Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
+// simple superregister reference or explicit instructions to insert
+// the upper bits of a vector.
+SDValue
+X86TargetLowering::LowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const {
+ if (Subtarget->hasAVX()) {
+ DebugLoc dl = Op.getNode()->getDebugLoc();
+ SDValue Vec = Op.getNode()->getOperand(0);
+ SDValue SubVec = Op.getNode()->getOperand(1);
+ SDValue Idx = Op.getNode()->getOperand(2);
+
+ if (Op.getNode()->getValueType(0).getSizeInBits() == 256
+ && SubVec.getNode()->getValueType(0).getSizeInBits() == 128) {
+ return Insert128BitVector(Vec, SubVec, Idx, DAG, dl);
+ }
+ }
+ return SDValue();
+}
+
// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
// their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
// one of the above mentioned nodes. It has to be wrapped because otherwise
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
unsigned char OperandFlags) {
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
DebugLoc dl = GA->getDebugLoc();
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
GA->getValueType(0),
// Lowering the machine isd will make sure everything is in the right
// location.
SDValue Chain = DAG.getEntryNode();
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue Args[] = { Chain, Offset };
Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
// TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
MFI->setAdjustsStack(true);
-
+
// And our return value (tls address) is in the standard call return value
// location.
unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
SDVTList Tys;
bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
if (useSSE)
- Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
+ Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
else
Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
// Load the value out, extending it from f32 to f80.
// FIXME: Avoid the extend by constructing the right constant pool?
- SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, MVT::f80, dl, DAG.getEntryNode(),
+ SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
FudgePtr, MachinePointerInfo::getConstantPool(),
MVT::f32, false, false, 4);
// Extend everything to 80 bits to force it to be done on x87.
// Lower (X & (1 << N)) == 0 to BT(X, N).
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
- if (Op0.getOpcode() == ISD::AND &&
- Op0.hasOneUse() &&
+ if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
Op1.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(Op1)->isNullValue() &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
return NewSetCC;
}
- // Look for "(setcc) == / != 1" to avoid unnecessary setcc.
- if (Op0.getOpcode() == X86ISD::SETCC &&
- Op1.getOpcode() == ISD::Constant &&
+ // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
+ // these.
+ if (Op1.getOpcode() == ISD::Constant &&
(cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
cast<ConstantSDNode>(Op1)->isNullValue()) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
- X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
- bool Invert = (CC == ISD::SETNE) ^
- cast<ConstantSDNode>(Op1)->isNullValue();
- if (Invert)
+
+ // If the input is a setcc, then reuse the input setcc or use a new one with
+ // the inverted condition.
+ if (Op0.getOpcode() == X86ISD::SETCC) {
+ X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
+ bool Invert = (CC == ISD::SETNE) ^
+ cast<ConstantSDNode>(Op1)->isNullValue();
+ if (!Invert) return Op0;
+
CCode = X86::GetOppositeBranchCondition(CCode);
- return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ }
}
bool isFP = Op1.getValueType().isFloatingPoint();
if (X86CC == X86::COND_INVALID)
return SDValue();
- SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
-
- // Use sbb x, x to materialize carry bit into a GPR.
- if (X86CC == X86::COND_B)
- return DAG.getNode(ISD::AND, dl, MVT::i8,
- DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond),
- DAG.getConstant(1, MVT::i8));
-
+ SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond);
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
}
SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
if (Op.getResNo() == 1 &&
(Opc == X86ISD::ADD ||
Opc == X86ISD::SUB ||
+ Opc == X86ISD::ADC ||
+ Opc == X86ISD::SBB ||
Opc == X86ISD::SMUL ||
Opc == X86ISD::UMUL ||
Opc == X86ISD::INC ||
if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
return true;
-
+
return false;
}
Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
isZero(Cond.getOperand(1).getOperand(1))) {
SDValue Cmp = Cond.getOperand(1);
-
+
unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
-
- if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
+
+ if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
(CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
SDValue CmpOp0 = Cmp.getOperand(0);
Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
-
+
SDValue Res = // Res = 0 or -1.
DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
-
+
if (isAllOnes(Op1) != (CondCode == X86::COND_E))
Res = DAG.getNOT(DL, Res, Res.getValueType());
-
+
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
if (N2C == 0 || !N2C->isNullValue())
Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
+ // a < b ? -1 : 0 -> RES = ~setcc_carry
+ // a < b ? 0 : -1 -> RES = setcc_carry
+ // a >= b ? -1 : 0 -> RES = setcc_carry
+ // a >= b ? 0 : -1 -> RES = ~setcc_carry
+ if (Cond.getOpcode() == X86ISD::CMP) {
+ unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
+
+ if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
+ (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
+ SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
+ DAG.getConstant(X86::COND_B, MVT::i8), Cond);
+ if (isAllOnes(Op1) != (CondCode == X86::COND_B))
+ return DAG.getNOT(DL, Res, Res.getValueType());
+ return Res;
+ }
+ }
+
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
- SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
SDValue Ops[] = { Op2, Op1, CC, Cond };
return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
}
Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
Flag = Chain.getValue(1);
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
Flag = Chain.getValue(1);
MachineFunction &MF = DAG.getMachineFunction();
const TargetMachine &TM = MF.getTarget();
- const TargetFrameInfo &TFI = *TM.getFrameInfo();
+ const TargetFrameLowering &TFI = *TM.getFrameLowering();
unsigned StackAlignment = TFI.getStackAlignment();
EVT VT = Op.getValueType();
DebugLoc DL = Op.getDebugLoc();
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(4, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(2, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
// return pblendv(r, r+r, a);
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
return R;
}
SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
MVT::i32);
SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
-
+
SDValue SetCC =
DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
DAG.getConstant(X86::COND_O, MVT::i32),
SDValue(Sum.getNode(), 2));
-
+
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
return Sum;
}
Op.getOperand(3),
DAG.getTargetConstant(size, MVT::i8),
cpIn.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
Ops, 5, T, MMO);
SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
SelectionDAG &DAG) const {
assert(Subtarget->is64Bit() && "Result not type legalized?");
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
SelectionDAG &DAG) const {
EVT SrcVT = Op.getOperand(0).getValueType();
EVT DstVT = Op.getValueType();
- assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
- Subtarget->hasMMX() && !DisableMMX) &&
- "Unexpected custom BITCAST");
+ assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
+ Subtarget->hasMMX() && "Unexpected custom BITCAST");
assert((DstVT == MVT::i64 ||
(DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
"Unexpected custom BITCAST");
// All other conversions need to be expanded.
return SDValue();
}
+
SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
SDNode *Node = Op.getNode();
DebugLoc dl = Node->getDebugLoc();
cast<AtomicSDNode>(Node)->getAlignment());
}
+static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getNode()->getValueType(0);
+
+ // Let legalize expand this if it isn't a legal type yet.
+ if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
+ return SDValue();
+
+ SDVTList VTs = DAG.getVTList(VT, MVT::i32);
+
+ unsigned Opc;
+ bool ExtraOp = false;
+ switch (Op.getOpcode()) {
+ default: assert(0 && "Invalid code");
+ case ISD::ADDC: Opc = X86ISD::ADD; break;
+ case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
+ case ISD::SUBC: Opc = X86ISD::SUB; break;
+ case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
+ }
+
+ if (!ExtraOp)
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1));
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1), Op.getOperand(2));
+}
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
+ case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op, DAG);
+ case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, DAG);
case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
case ISD::UMULO: return LowerXALUO(Op, DAG);
case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
case ISD::BITCAST: return LowerBITCAST(Op, DAG);
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
}
}
default:
assert(false && "Do not know how to custom type legalize this operation!");
return;
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE:
+ // We don't want to expand or promote these.
+ return;
case ISD::FP_TO_SINT: {
std::pair<SDValue,SDValue> Vals =
FP_TO_INTHelper(SDValue(N, 0), DAG, true);
return;
}
case ISD::READCYCLECOUNTER: {
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = N->getOperand(0);
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
SDValue Ops[] = { swapInH.getValue(0),
N->getOperand(1),
swapInH.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
Ops, 3, T, MMO);
case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
case X86ISD::PSIGND: return "X86ISD::PSIGND";
+ case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
case X86ISD::ADD: return "X86ISD::ADD";
case X86ISD::SUB: return "X86ISD::SUB";
+ case X86ISD::ADC: return "X86ISD::ADC";
+ case X86ISD::SBB: return "X86ISD::SBB";
case X86ISD::SMUL: return "X86ISD::SMUL";
case X86ISD::UMUL: return "X86ISD::UMUL";
case X86ISD::INC: return "X86ISD::INC";
X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
-
+
// Address into RAX/EAX, other two args into ECX, EDX.
unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
for (int i = 0; i < X86::AddrNumOperands; ++i)
MIB.addOperand(MI->getOperand(i));
-
+
unsigned ValOps = X86::AddrNumOperands;
BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
.addReg(MI->getOperand(ValOps).getReg());
// The instruction doesn't actually take any operands though.
BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
-
+
MI->eraseFromParent(); // The pseudo is gone now.
return BB;
}
X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
-
+
// First arg in ECX, the second in EAX.
BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
.addReg(MI->getOperand(0).getReg());
BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
.addReg(MI->getOperand(1).getReg());
-
+
// The instruction doesn't actually take any operands though.
BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
-
+
MI->eraseFromParent(); // The pseudo is gone now.
return BB;
}
MachineBasicBlock *BB) const {
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
+ case X86::TAILJMPd64:
+ case X86::TAILJMPr64:
+ case X86::TAILJMPm64:
+ assert(!"TAILJMP64 would not be touched here.");
+ case X86::TCRETURNdi64:
+ case X86::TCRETURNri64:
+ case X86::TCRETURNmi64:
+ // Defs of TCRETURNxx64 has Win64's callee-saved registers, as subset.
+ // On AMD64, additional defs should be added before register allocation.
+ if (!Subtarget->isTargetWin64()) {
+ MI->addRegisterDefined(X86::RSI);
+ MI->addRegisterDefined(X86::RDI);
+ MI->addRegisterDefined(X86::XMM6);
+ MI->addRegisterDefined(X86::XMM7);
+ MI->addRegisterDefined(X86::XMM8);
+ MI->addRegisterDefined(X86::XMM9);
+ MI->addRegisterDefined(X86::XMM10);
+ MI->addRegisterDefined(X86::XMM11);
+ MI->addRegisterDefined(X86::XMM12);
+ MI->addRegisterDefined(X86::XMM13);
+ MI->addRegisterDefined(X86::XMM14);
+ MI->addRegisterDefined(X86::XMM15);
+ }
+ return BB;
case X86::WIN_ALLOCA:
return EmitLoweredWinAlloca(MI, BB);
case X86::TLSCall_32:
// Thread synchronization.
case X86::MONITOR:
- return EmitMonitor(MI, BB);
+ return EmitMonitor(MI, BB);
case X86::MWAIT:
return EmitMwait(MI, BB);
default: break;
case X86ISD::ADD:
case X86ISD::SUB:
+ case X86ISD::ADC:
+ case X86ISD::SBB:
case X86ISD::SMUL:
case X86ISD::UMUL:
case X86ISD::INC:
/// if the load addresses are consecutive, non-overlapping, and in the right
/// order.
static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
- const TargetLowering &TLI) {
+ TargetLowering::DAGCombinerInfo &DCI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
if (VT.getSizeInBits() != 128)
return SDValue();
+ // Don't create instructions with illegal types after legalize types has run.
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
+ return SDValue();
+
SmallVector<SDValue, 16> Elts;
for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
const X86Subtarget *Subtarget) {
if (DCI.isBeforeLegalizeOps())
return SDValue();
-
+
// Want to form PANDN nodes, in the hopes of then easily combining them with
// OR and AND nodes to form PBLEND/PSIGN.
EVT VT = N->getValueType(0);
if (VT != MVT::v2i64)
return SDValue();
-
+
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
DebugLoc DL = N->getDebugLoc();
-
+
// Check LHS for vnot
- if (N0.getOpcode() == ISD::XOR &&
+ if (N0.getOpcode() == ISD::XOR &&
ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
if (N1.getOpcode() == ISD::XOR &&
ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
-
+
return SDValue();
}
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
-
+
// look for psign/blend
if (Subtarget->hasSSSE3()) {
if (VT == MVT::v2i64) {
Y = N0.getOperand(1);
if (N0.getOperand(1) == Mask)
Y = N0.getOperand(0);
-
+
// Check to see if the mask appeared in both the AND and PANDN and
if (!Y.getNode())
return SDValue();
-
+
// Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
if (Mask.getOpcode() != ISD::BITCAST ||
X.getOpcode() != ISD::BITCAST ||
Y.getOpcode() != ISD::BITCAST)
return SDValue();
-
+
// Look through mask bitcast.
Mask = Mask.getOperand(0);
EVT MaskVT = Mask.getValueType();
// will be an intrinsic.
if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
return SDValue();
-
+
// FIXME: what to do for bytes, since there is a psignb/pblendvb, but
// there is no psrai.b
switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
break;
default: return SDValue();
}
-
+
// Check that the SRA is all signbits.
SDValue SraC = Mask.getOperand(2);
unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
if ((SraAmt + 1) != EltBits)
return SDValue();
-
+
DebugLoc DL = N->getDebugLoc();
// Now we know we at least have a plendvb with the mask val. See if
// PBLENDVB only available on SSE 4.1
if (!Subtarget->hasSSE41())
return SDValue();
-
- unsigned IID = Intrinsic::x86_sse41_pblendvb;
+
X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
- Mask = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, DL, MVT::v16i8,
- DAG.getConstant(IID, MVT::i32), X, Y, Mask);
+ Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
}
}
}
-
+
// fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
std::swap(N0, N1);
DAG.getNode(ISD::TRUNCATE, DL,
MVT::i8, ShAmt0));
}
-
+
return SDValue();
}
return SDValue();
}
+// Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
+static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+ unsigned X86CC = N->getConstantOperandVal(0);
+ SDValue EFLAG = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
+ // a zext and produces an all-ones bit which is more useful than 0/1 in some
+ // cases.
+ if (X86CC == X86::COND_B)
+ return DAG.getNode(ISD::AND, DL, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), EFLAG),
+ DAG.getConstant(1, MVT::i8));
+
+ return SDValue();
+}
+
+// Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
+static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
+ X86TargetLowering::DAGCombinerInfo &DCI) {
+ // If the LHS and RHS of the ADC node are zero, then it can't overflow and
+ // the result is either zero or one (depending on the input carry bit).
+ // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
+ if (X86::isZeroNode(N->getOperand(0)) &&
+ X86::isZeroNode(N->getOperand(1)) &&
+ // We don't have a good way to replace an EFLAGS use, so only do this when
+ // dead right now.
+ SDValue(N, 1).use_empty()) {
+ DebugLoc DL = N->getDebugLoc();
+ EVT VT = N->getValueType(0);
+ SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
+ SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
+ DAG.getConstant(X86::COND_B,MVT::i8),
+ N->getOperand(2)),
+ DAG.getConstant(1, VT));
+ return DCI.CombineTo(N, Res1, CarryOut);
+ }
+
+ return SDValue();
+}
+
+// fold (add Y, (sete X, 0)) -> adc 0, Y
+// (add Y, (setne X, 0)) -> sbb -1, Y
+// (sub (sete X, 0), Y) -> sbb 0, Y
+// (sub (setne X, 0), Y) -> adc -1, Y
+static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
+ DebugLoc DL = N->getDebugLoc();
+
+ // Look through ZExts.
+ SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
+ if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
+ return SDValue();
+
+ SDValue SetCC = Ext.getOperand(0);
+ if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
+ return SDValue();
+
+ X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
+ if (CC != X86::COND_E && CC != X86::COND_NE)
+ return SDValue();
+
+ SDValue Cmp = SetCC.getOperand(1);
+ if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
+ !X86::isZeroNode(Cmp.getOperand(1)) ||
+ !Cmp.getOperand(0).getValueType().isInteger())
+ return SDValue();
+
+ SDValue CmpOp0 = Cmp.getOperand(0);
+ SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
+ DAG.getConstant(1, CmpOp0.getValueType()));
+
+ SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
+ if (CC == X86::COND_NE)
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
switch (N->getOpcode()) {
default: break;
case ISD::EXTRACT_VECTOR_ELT:
- return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
+ return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
+ case ISD::ADD:
+ case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
+ case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
+ case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
case X86ISD::SHUFPS: // Handle all target specific shuffles
case X86ISD::SHUFPD:
case X86ISD::PALIGN:
case X86ISD::PSHUFLW:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
+ case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
}
return SDValue();
// X86 Inline Assembly Support
//===----------------------------------------------------------------------===//
-static bool LowerToBSwap(CallInst *CI) {
- // FIXME: this should verify that we are targetting a 486 or better. If not,
- // we will turn this bswap into something that will be lowered to logical ops
- // instead of emitting the bswap asm. For now, we don't support 486 or lower
- // so don't worry about this.
-
- // Verify this is a simple bswap.
- if (CI->getNumArgOperands() != 1 ||
- CI->getType() != CI->getArgOperand(0)->getType() ||
- !CI->getType()->isIntegerTy())
- return false;
-
- const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
- if (!Ty || Ty->getBitWidth() % 16 != 0)
- return false;
-
- // Okay, we can do this xform, do so now.
- const Type *Tys[] = { Ty };
- Module *M = CI->getParent()->getParent()->getParent();
- Constant *Int = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
-
- Value *Op = CI->getArgOperand(0);
- Op = CallInst::Create(Int, Op, CI->getName(), CI);
-
- CI->replaceAllUsesWith(Op);
- CI->eraseFromParent();
- return true;
-}
-
bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
- InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
std::string AsmStr = IA->getAsmString();
AsmPieces.clear();
SplitString(AsmStr, AsmPieces, " \t"); // Split with whitespace.
+ // FIXME: this should verify that we are targetting a 486 or better. If not,
+ // we will turn this bswap into something that will be lowered to logical ops
+ // instead of emitting the bswap asm. For now, we don't support 486 or lower
+ // so don't worry about this.
// bswap $0
if (AsmPieces.size() == 2 &&
(AsmPieces[0] == "bswap" ||
AsmPieces[1] == "${0:q}")) {
// No need to check constraints, nothing other than the equivalent of
// "=r,0" would be valid here.
- return LowerToBSwap(CI);
+ const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
// rorw $$8, ${0:w} --> llvm.bswap.i16
if (CI->getType()->isIntegerTy(16) &&
AsmPieces[2] == "${0:w}" &&
IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
AsmPieces.clear();
- const std::string &Constraints = IA->getConstraintString();
- SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
+ const std::string &ConstraintsStr = IA->getConstraintString();
+ SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
std::sort(AsmPieces.begin(), AsmPieces.end());
if (AsmPieces.size() == 4 &&
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
AsmPieces[2] == "~{flags}" &&
AsmPieces[3] == "~{fpsr}") {
- return LowerToBSwap(CI);
+ const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
}
break;
if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
Words[2] == "${0:w}") {
AsmPieces.clear();
- const std::string &Constraints = IA->getConstraintString();
- SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
+ const std::string &ConstraintsStr = IA->getConstraintString();
+ SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
std::sort(AsmPieces.begin(), AsmPieces.end());
if (AsmPieces.size() == 4 &&
AsmPieces[0] == "~{cc}" &&
AsmPieces[1] == "~{dirflag}" &&
AsmPieces[2] == "~{flags}" &&
AsmPieces[3] == "~{fpsr}") {
- return LowerToBSwap(CI);
+ const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+ return IntrinsicLowering::LowerToByteSwap(CI);
}
}
}
}
}
- if (CI->getType()->isIntegerTy(64) &&
- Constraints.size() >= 2 &&
- Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
- Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
- // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
- SmallVector<StringRef, 4> Words;
- SplitString(AsmPieces[0], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
- Words.clear();
- SplitString(AsmPieces[1], Words, " \t");
- if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
+
+ if (CI->getType()->isIntegerTy(64)) {
+ InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
+ if (Constraints.size() >= 2 &&
+ Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
+ Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
+ // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
+ SmallVector<StringRef, 4> Words;
+ SplitString(AsmPieces[0], Words, " \t");
+ if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%eax") {
Words.clear();
- SplitString(AsmPieces[2], Words, " \t,");
- if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
- Words[2] == "%edx") {
- return LowerToBSwap(CI);
+ SplitString(AsmPieces[1], Words, " \t");
+ if (Words.size() == 2 && Words[0] == "bswap" && Words[1] == "%edx") {
+ Words.clear();
+ SplitString(AsmPieces[2], Words, " \t,");
+ if (Words.size() == 3 && Words[0] == "xchgl" && Words[1] == "%eax" &&
+ Words[2] == "%edx") {
+ const IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
+ if (!Ty || Ty->getBitWidth() % 16 != 0)
+ return false;
+ return IntrinsicLowering::LowerToByteSwap(CI);
+ }
}
}
}
weight = CW_SpecificReg;
break;
case 'y':
- if (type->isX86_MMXTy() && !DisableMMX && Subtarget->hasMMX())
+ if (type->isX86_MMXTy() && Subtarget->hasMMX())
weight = CW_SpecificReg;
break;
case 'x':