//===----------------------------------------------------------------------===//
#include "PPCISelLowering.h"
+#include "PPCMachineFunctionInfo.h"
+#include "PPCPredicates.h"
#include "PPCTargetMachine.h"
#include "PPCPerfectShuffle.h"
#include "llvm/ADT/VectorExtras.h"
#include "llvm/Intrinsics.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Target/TargetOptions.h"
+#include "llvm/Support/CommandLine.h"
using namespace llvm;
-PPCTargetLowering::PPCTargetLowering(TargetMachine &TM)
- : TargetLowering(TM) {
+static cl::opt<bool> EnablePPCPreinc("enable-ppc-preinc");
+
+PPCTargetLowering::PPCTargetLowering(PPCTargetMachine &TM)
+ : TargetLowering(TM), PPCSubTarget(*TM.getSubtargetImpl()) {
- // Fold away setcc operations if possible.
- setSetCCIsExpensive();
setPow2DivIsCheap();
// Use _setjmp/_longjmp instead of setjmp/longjmp.
- setUseUnderscoreSetJmpLongJmp(true);
+ setUseUnderscoreSetJmp(true);
+ setUseUnderscoreLongJmp(true);
// Set up the register classes.
addRegisterClass(MVT::i32, PPC::GPRCRegisterClass);
addRegisterClass(MVT::f32, PPC::F4RCRegisterClass);
addRegisterClass(MVT::f64, PPC::F8RCRegisterClass);
+ // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
+ setLoadXAction(ISD::SEXTLOAD, MVT::i1, Expand);
+ setLoadXAction(ISD::SEXTLOAD, MVT::i8, Expand);
+
+ // PowerPC does not have truncstore for i1.
+ setStoreXAction(MVT::i1, Promote);
+
+ // PowerPC has pre-inc load and store's.
+ setIndexedLoadAction(ISD::PRE_INC, MVT::i1, Legal);
+ setIndexedLoadAction(ISD::PRE_INC, MVT::i8, Legal);
+ setIndexedLoadAction(ISD::PRE_INC, MVT::i16, Legal);
+ setIndexedLoadAction(ISD::PRE_INC, MVT::i32, Legal);
+ setIndexedLoadAction(ISD::PRE_INC, MVT::i64, Legal);
+ setIndexedStoreAction(ISD::PRE_INC, MVT::i1, Legal);
+ setIndexedStoreAction(ISD::PRE_INC, MVT::i8, Legal);
+ setIndexedStoreAction(ISD::PRE_INC, MVT::i16, Legal);
+ setIndexedStoreAction(ISD::PRE_INC, MVT::i32, Legal);
+ setIndexedStoreAction(ISD::PRE_INC, MVT::i64, Legal);
+
setOperationAction(ISD::ConstantFP, MVT::f64, Expand);
setOperationAction(ISD::ConstantFP, MVT::f32, Expand);
setOperationAction(ISD::MEMSET, MVT::Other, Expand);
setOperationAction(ISD::MEMCPY, MVT::Other, Expand);
- // PowerPC has an i16 but no i8 (or i1) SEXTLOAD
- setOperationAction(ISD::SEXTLOAD, MVT::i1, Expand);
- setOperationAction(ISD::SEXTLOAD, MVT::i8, Expand);
-
// PowerPC has no SREM/UREM instructions
setOperationAction(ISD::SREM, MVT::i32, Expand);
setOperationAction(ISD::UREM, MVT::i32, Expand);
setOperationAction(ISD::BSWAP, MVT::i32 , Expand);
setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
+ setOperationAction(ISD::BSWAP, MVT::i64 , Expand);
+ setOperationAction(ISD::CTPOP, MVT::i64 , Expand);
+ setOperationAction(ISD::CTTZ , MVT::i64 , Expand);
// PowerPC does not have ROTR
setOperationAction(ISD::ROTR, MVT::i32 , Expand);
// PowerPC does not have Select
setOperationAction(ISD::SELECT, MVT::i32, Expand);
+ setOperationAction(ISD::SELECT, MVT::i64, Expand);
setOperationAction(ISD::SELECT, MVT::f32, Expand);
setOperationAction(ISD::SELECT, MVT::f64, Expand);
// PowerPC does not have BRCOND which requires SetCC
setOperationAction(ISD::BRCOND, MVT::Other, Expand);
+
+ setOperationAction(ISD::BR_JT, MVT::Other, Expand);
// PowerPC turns FP_TO_SINT into FCTIWZ and some load/stores.
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
setOperationAction(ISD::BIT_CONVERT, MVT::f32, Expand);
setOperationAction(ISD::BIT_CONVERT, MVT::i32, Expand);
-
- // PowerPC does not have truncstore for i1.
- setOperationAction(ISD::TRUNCSTORE, MVT::i1, Promote);
+ setOperationAction(ISD::BIT_CONVERT, MVT::i64, Expand);
+ setOperationAction(ISD::BIT_CONVERT, MVT::f64, Expand);
// We cannot sextinreg(i1). Expand to shifts.
setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);
setOperationAction(ISD::DEBUG_LOC, MVT::Other, Expand);
// FIXME - use subtarget debug flags
if (!TM.getSubtarget<PPCSubtarget>().isDarwin())
- setOperationAction(ISD::DEBUG_LABEL, MVT::Other, Expand);
+ setOperationAction(ISD::LABEL, MVT::Other, Expand);
// We want to legalize GlobalAddress and ConstantPool nodes into the
// appropriate instructions to materialize the address.
setOperationAction(ISD::VACOPY , MVT::Other, Expand);
setOperationAction(ISD::VAEND , MVT::Other, Expand);
setOperationAction(ISD::STACKSAVE , MVT::Other, Expand);
- setOperationAction(ISD::STACKRESTORE , MVT::Other, Expand);
- setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Expand);
-
+ setOperationAction(ISD::STACKRESTORE , MVT::Other, Custom);
+ setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32 , Custom);
+ setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64 , Custom);
+
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
// They also have instructions for converting between i64 and fp.
setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
+ setOperationAction(ISD::FP_TO_UINT, MVT::i64, Expand);
setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
-
+ setOperationAction(ISD::UINT_TO_FP, MVT::i64, Expand);
+ setOperationAction(ISD::FP_TO_UINT, MVT::i32, Expand);
+
// FIXME: disable this lowered code. This generates 64-bit register values,
// and we don't model the fact that the top part is clobbered by calls. We
// need to flag these together so that the value isn't live across a call.
setOperationAction(ISD::BUILD_PAIR, MVT::i64, Expand);
} else {
// 32 bit PowerPC wants to expand i64 shifts itself.
- setOperationAction(ISD::SHL, MVT::i64, Custom);
- setOperationAction(ISD::SRL, MVT::i64, Custom);
- setOperationAction(ISD::SRA, MVT::i64, Custom);
+ setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
+ setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
}
if (TM.getSubtarget<PPCSubtarget>().hasAltivec()) {
setSetCCResultType(MVT::i32);
setShiftAmountType(MVT::i32);
setSetCCResultContents(ZeroOrOneSetCCResult);
- setStackPointerRegisterToSaveRestore(PPC::R1);
+
+ if (TM.getSubtarget<PPCSubtarget>().isPPC64())
+ setStackPointerRegisterToSaveRestore(PPC::X1);
+ else
+ setStackPointerRegisterToSaveRestore(PPC::R1);
// We have target-specific dag combine patterns for the following nodes:
setTargetDAGCombine(ISD::SINT_TO_FP);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::BR_CC);
+ setTargetDAGCombine(ISD::BSWAP);
computeRegisterProperties();
}
case PPCISD::VPERM: return "PPCISD::VPERM";
case PPCISD::Hi: return "PPCISD::Hi";
case PPCISD::Lo: return "PPCISD::Lo";
+ case PPCISD::DYNALLOC: return "PPCISD::DYNALLOC";
case PPCISD::GlobalBaseReg: return "PPCISD::GlobalBaseReg";
case PPCISD::SRL: return "PPCISD::SRL";
case PPCISD::SRA: return "PPCISD::SRA";
case PPCISD::MFCR: return "PPCISD::MFCR";
case PPCISD::VCMP: return "PPCISD::VCMP";
case PPCISD::VCMPo: return "PPCISD::VCMPo";
+ case PPCISD::LBRX: return "PPCISD::LBRX";
+ case PPCISD::STBRX: return "PPCISD::STBRX";
case PPCISD::COND_BRANCH: return "PPCISD::COND_BRANCH";
}
}
static bool isFloatingPointZero(SDOperand Op) {
if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
- else if (Op.getOpcode() == ISD::EXTLOAD || Op.getOpcode() == ISD::LOAD) {
+ else if (ISD::isEXTLoad(Op.Val) || ISD::isNON_EXTLoad(Op.Val)) {
// Maybe this has already been legalized into the constant pool?
if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(Op.getOperand(1)))
- if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->get()))
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
return CFP->isExactlyValue(-0.0) || CFP->isExactlyValue(0.0);
}
return false;
return SDOperand();
}
+//===----------------------------------------------------------------------===//
+// Addressing Mode Selection
+//===----------------------------------------------------------------------===//
+
+/// isIntS16Immediate - This method tests to see if the node is either a 32-bit
+/// or 64-bit immediate, and if the value can be accurately represented as a
+/// sign extension from a 16-bit value. If so, this returns true and the
+/// immediate.
+static bool isIntS16Immediate(SDNode *N, short &Imm) {
+ if (N->getOpcode() != ISD::Constant)
+ return false;
+
+ Imm = (short)cast<ConstantSDNode>(N)->getValue();
+ if (N->getValueType(0) == MVT::i32)
+ return Imm == (int32_t)cast<ConstantSDNode>(N)->getValue();
+ else
+ return Imm == (int64_t)cast<ConstantSDNode>(N)->getValue();
+}
+static bool isIntS16Immediate(SDOperand Op, short &Imm) {
+ return isIntS16Immediate(Op.Val, Imm);
+}
+
+
+/// SelectAddressRegReg - Given the specified addressed, check to see if it
+/// can be represented as an indexed [r+r] operation. Returns false if it
+/// can be more efficiently represented with [r+imm].
+bool PPCTargetLowering::SelectAddressRegReg(SDOperand N, SDOperand &Base,
+ SDOperand &Index,
+ SelectionDAG &DAG) {
+ short imm = 0;
+ if (N.getOpcode() == ISD::ADD) {
+ if (isIntS16Immediate(N.getOperand(1), imm))
+ return false; // r+i
+ if (N.getOperand(1).getOpcode() == PPCISD::Lo)
+ return false; // r+i
+
+ Base = N.getOperand(0);
+ Index = N.getOperand(1);
+ return true;
+ } else if (N.getOpcode() == ISD::OR) {
+ if (isIntS16Immediate(N.getOperand(1), imm))
+ return false; // r+i can fold it if we can.
+
+ // If this is an or of disjoint bitfields, we can codegen this as an add
+ // (for better address arithmetic) if the LHS and RHS of the OR are provably
+ // disjoint.
+ uint64_t LHSKnownZero, LHSKnownOne;
+ uint64_t RHSKnownZero, RHSKnownOne;
+ ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
+
+ if (LHSKnownZero) {
+ ComputeMaskedBits(N.getOperand(1), ~0U, RHSKnownZero, RHSKnownOne);
+ // If all of the bits are known zero on the LHS or RHS, the add won't
+ // carry.
+ if ((LHSKnownZero | RHSKnownZero) == ~0U) {
+ Base = N.getOperand(0);
+ Index = N.getOperand(1);
+ return true;
+ }
+ }
+ }
+
+ return false;
+}
+
+/// Returns true if the address N can be represented by a base register plus
+/// a signed 16-bit displacement [r+imm], and if it is not better
+/// represented as reg+reg.
+bool PPCTargetLowering::SelectAddressRegImm(SDOperand N, SDOperand &Disp,
+ SDOperand &Base, SelectionDAG &DAG){
+ // If this can be more profitably realized as r+r, fail.
+ if (SelectAddressRegReg(N, Disp, Base, DAG))
+ return false;
+
+ if (N.getOpcode() == ISD::ADD) {
+ short imm = 0;
+ if (isIntS16Immediate(N.getOperand(1), imm)) {
+ Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
+ if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
+ Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
+ } else {
+ Base = N.getOperand(0);
+ }
+ return true; // [r+i]
+ } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
+ // Match LOAD (ADD (X, Lo(G))).
+ assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
+ && "Cannot handle constant offsets yet!");
+ Disp = N.getOperand(1).getOperand(0); // The global address.
+ assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
+ Disp.getOpcode() == ISD::TargetConstantPool ||
+ Disp.getOpcode() == ISD::TargetJumpTable);
+ Base = N.getOperand(0);
+ return true; // [&g+r]
+ }
+ } else if (N.getOpcode() == ISD::OR) {
+ short imm = 0;
+ if (isIntS16Immediate(N.getOperand(1), imm)) {
+ // If this is an or of disjoint bitfields, we can codegen this as an add
+ // (for better address arithmetic) if the LHS and RHS of the OR are
+ // provably disjoint.
+ uint64_t LHSKnownZero, LHSKnownOne;
+ ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
+ if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
+ // If all of the bits are known zero on the LHS or RHS, the add won't
+ // carry.
+ Base = N.getOperand(0);
+ Disp = DAG.getTargetConstant((int)imm & 0xFFFF, MVT::i32);
+ return true;
+ }
+ }
+ } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
+ // Loading from a constant address.
+
+ // If this address fits entirely in a 16-bit sext immediate field, codegen
+ // this as "d, 0"
+ short Imm;
+ if (isIntS16Immediate(CN, Imm)) {
+ Disp = DAG.getTargetConstant(Imm, CN->getValueType(0));
+ Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
+ return true;
+ }
+
+ // FIXME: Handle small sext constant offsets in PPC64 mode also!
+ if (CN->getValueType(0) == MVT::i32) {
+ int Addr = (int)CN->getValue();
+
+ // Otherwise, break this down into an LIS + disp.
+ Disp = DAG.getTargetConstant((short)Addr, MVT::i32);
+ Base = DAG.getConstant(Addr - (signed short)Addr, MVT::i32);
+ return true;
+ }
+ }
+
+ Disp = DAG.getTargetConstant(0, getPointerTy());
+ if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
+ Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
+ else
+ Base = N;
+ return true; // [r+0]
+}
+
+/// SelectAddressRegRegOnly - Given the specified addressed, force it to be
+/// represented as an indexed [r+r] operation.
+bool PPCTargetLowering::SelectAddressRegRegOnly(SDOperand N, SDOperand &Base,
+ SDOperand &Index,
+ SelectionDAG &DAG) {
+ // Check to see if we can easily represent this as an [r+r] address. This
+ // will fail if it thinks that the address is more profitably represented as
+ // reg+imm, e.g. where imm = 0.
+ if (SelectAddressRegReg(N, Base, Index, DAG))
+ return true;
+
+ // If the operand is an addition, always emit this as [r+r], since this is
+ // better (for code size, and execution, as the memop does the add for free)
+ // than emitting an explicit add.
+ if (N.getOpcode() == ISD::ADD) {
+ Base = N.getOperand(0);
+ Index = N.getOperand(1);
+ return true;
+ }
+
+ // Otherwise, do it the hard way, using R0 as the base register.
+ Base = DAG.getRegister(PPC::R0, N.getValueType());
+ Index = N;
+ return true;
+}
+
+/// SelectAddressRegImmShift - Returns true if the address N can be
+/// represented by a base register plus a signed 14-bit displacement
+/// [r+imm*4]. Suitable for use by STD and friends.
+bool PPCTargetLowering::SelectAddressRegImmShift(SDOperand N, SDOperand &Disp,
+ SDOperand &Base,
+ SelectionDAG &DAG) {
+ // If this can be more profitably realized as r+r, fail.
+ if (SelectAddressRegReg(N, Disp, Base, DAG))
+ return false;
+
+ if (N.getOpcode() == ISD::ADD) {
+ short imm = 0;
+ if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
+ Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
+ if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N.getOperand(0))) {
+ Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
+ } else {
+ Base = N.getOperand(0);
+ }
+ return true; // [r+i]
+ } else if (N.getOperand(1).getOpcode() == PPCISD::Lo) {
+ // Match LOAD (ADD (X, Lo(G))).
+ assert(!cast<ConstantSDNode>(N.getOperand(1).getOperand(1))->getValue()
+ && "Cannot handle constant offsets yet!");
+ Disp = N.getOperand(1).getOperand(0); // The global address.
+ assert(Disp.getOpcode() == ISD::TargetGlobalAddress ||
+ Disp.getOpcode() == ISD::TargetConstantPool ||
+ Disp.getOpcode() == ISD::TargetJumpTable);
+ Base = N.getOperand(0);
+ return true; // [&g+r]
+ }
+ } else if (N.getOpcode() == ISD::OR) {
+ short imm = 0;
+ if (isIntS16Immediate(N.getOperand(1), imm) && (imm & 3) == 0) {
+ // If this is an or of disjoint bitfields, we can codegen this as an add
+ // (for better address arithmetic) if the LHS and RHS of the OR are
+ // provably disjoint.
+ uint64_t LHSKnownZero, LHSKnownOne;
+ ComputeMaskedBits(N.getOperand(0), ~0U, LHSKnownZero, LHSKnownOne);
+ if ((LHSKnownZero|~(unsigned)imm) == ~0U) {
+ // If all of the bits are known zero on the LHS or RHS, the add won't
+ // carry.
+ Base = N.getOperand(0);
+ Disp = DAG.getTargetConstant(((int)imm & 0xFFFF) >> 2, MVT::i32);
+ return true;
+ }
+ }
+ } else if (ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N)) {
+ // Loading from a constant address.
+
+ // If this address fits entirely in a 14-bit sext immediate field, codegen
+ // this as "d, 0"
+ short Imm;
+ if (isIntS16Immediate(CN, Imm)) {
+ Disp = DAG.getTargetConstant((unsigned short)Imm >> 2, getPointerTy());
+ Base = DAG.getRegister(PPC::R0, CN->getValueType(0));
+ return true;
+ }
+
+ // FIXME: Handle small sext constant offsets in PPC64 mode also!
+ if (CN->getValueType(0) == MVT::i32) {
+ int Addr = (int)CN->getValue();
+
+ // Otherwise, break this down into an LIS + disp.
+ Disp = DAG.getTargetConstant((short)Addr >> 2, MVT::i32);
+ Base = DAG.getConstant(Addr - (signed short)Addr, MVT::i32);
+ return true;
+ }
+ }
+
+ Disp = DAG.getTargetConstant(0, getPointerTy());
+ if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(N))
+ Base = DAG.getTargetFrameIndex(FI->getIndex(), N.getValueType());
+ else
+ Base = N;
+ return true; // [r+0]
+}
+
+
+/// getPreIndexedAddressParts - returns true by value, base pointer and
+/// offset pointer and addressing mode by reference if the node's address
+/// can be legally represented as pre-indexed load / store address.
+bool PPCTargetLowering::getPreIndexedAddressParts(SDNode *N, SDOperand &Base,
+ SDOperand &Offset,
+ ISD::MemIndexedMode &AM,
+ SelectionDAG &DAG) {
+ // Disabled by default for now.
+ if (!EnablePPCPreinc) return false;
+
+ SDOperand Ptr;
+ MVT::ValueType VT;
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
+ Ptr = LD->getBasePtr();
+ VT = LD->getLoadedVT();
+
+ } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) {
+ ST = ST;
+ Ptr = ST->getBasePtr();
+ VT = ST->getStoredVT();
+ } else
+ return false;
+
+ // PowerPC doesn't have preinc load/store instructions for vectors.
+ if (MVT::isVector(VT))
+ return false;
+
+ // TODO: Check reg+reg first.
+
+ // LDU/STU use reg+imm*4, others use reg+imm.
+ if (VT != MVT::i64) {
+ // reg + imm
+ if (!SelectAddressRegImm(Ptr, Offset, Base, DAG))
+ return false;
+ } else {
+ // reg + imm * 4.
+ if (!SelectAddressRegImmShift(Ptr, Offset, Base, DAG))
+ return false;
+ }
+
+ if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N)) {
+ // PPC64 doesn't have lwau, but it does have lwaux. Reject preinc load of
+ // sext i32 to i64 when addr mode is r+i.
+ if (LD->getValueType(0) == MVT::i64 && LD->getLoadedVT() == MVT::i32 &&
+ LD->getExtensionType() == ISD::SEXTLOAD &&
+ isa<ConstantSDNode>(Offset))
+ return false;
+ }
+
+ AM = ISD::PRE_INC;
+ return true;
+}
+
//===----------------------------------------------------------------------===//
// LowerOperation implementation
//===----------------------------------------------------------------------===//
static SDOperand LowerConstantPool(SDOperand Op, SelectionDAG &DAG) {
MVT::ValueType PtrVT = Op.getValueType();
ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
- Constant *C = CP->get();
+ Constant *C = CP->getConstVal();
SDOperand CPI = DAG.getTargetConstantPool(C, PtrVT, CP->getAlignment());
SDOperand Zero = DAG.getConstant(0, PtrVT);
return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
}
- if (TM.getRelocationModel() == Reloc::PIC) {
+ if (TM.getRelocationModel() == Reloc::PIC_) {
// With PIC, the first instruction is actually "GR+hi(&G)".
Hi = DAG.getNode(ISD::ADD, PtrVT,
DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
}
- if (TM.getRelocationModel() == Reloc::PIC) {
+ if (TM.getRelocationModel() == Reloc::PIC_) {
// With PIC, the first instruction is actually "GR+hi(&G)".
Hi = DAG.getNode(ISD::ADD, PtrVT,
- DAG.getNode(PPCISD::GlobalBaseReg, MVT::i32), Hi);
+ DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
}
Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
return DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
}
- if (TM.getRelocationModel() == Reloc::PIC) {
+ if (TM.getRelocationModel() == Reloc::PIC_) {
// With PIC, the first instruction is actually "GR+hi(&G)".
Hi = DAG.getNode(ISD::ADD, PtrVT,
DAG.getNode(PPCISD::GlobalBaseReg, PtrVT), Hi);
Lo = DAG.getNode(ISD::ADD, PtrVT, Hi, Lo);
- if (!GV->hasWeakLinkage() && !GV->hasLinkOnceLinkage() &&
- (!GV->isExternal() || GV->hasNotBeenReadFromBytecode()))
+ if (!TM.getSubtarget<PPCSubtarget>().hasLazyResolverStub(GV))
return Lo;
// If the global is weak or external, we have to go through the lazy
// resolution stub.
- return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, DAG.getSrcValue(0));
+ return DAG.getLoad(PtrVT, DAG.getEntryNode(), Lo, NULL, 0);
}
static SDOperand LowerSETCC(SDOperand Op, SelectionDAG &DAG) {
}
// If we have an integer seteq/setne, turn it into a compare against zero
- // by subtracting the rhs from the lhs, which is faster than setting a
- // condition register, reading it back out, and masking the correct bit.
+ // by xor'ing the rhs with the lhs, which is faster than setting a
+ // condition register, reading it back out, and masking the correct bit. The
+ // normal approach here uses sub to do this instead of xor. Using xor exposes
+ // the result to other bit-twiddling opportunities.
MVT::ValueType LHSVT = Op.getOperand(0).getValueType();
if (MVT::isInteger(LHSVT) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
MVT::ValueType VT = Op.getValueType();
- SDOperand Sub = DAG.getNode(ISD::SUB, LHSVT, Op.getOperand(0),
+ SDOperand Sub = DAG.getNode(ISD::XOR, LHSVT, Op.getOperand(0),
Op.getOperand(1));
return DAG.getSetCC(VT, Sub, DAG.getConstant(0, LHSVT), CC);
}
unsigned VarArgsFrameIndex) {
// vastart just stores the address of the VarArgsFrameIndex slot into the
// memory location argument.
- SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
- return DAG.getNode(ISD::STORE, MVT::Other, Op.getOperand(0), FR,
- Op.getOperand(1), Op.getOperand(2));
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+ SDOperand FR = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
+ SrcValueSDNode *SV = cast<SrcValueSDNode>(Op.getOperand(2));
+ return DAG.getStore(Op.getOperand(0), FR, Op.getOperand(1), SV->getValue(),
+ SV->getOffset());
}
static SDOperand LowerFORMAL_ARGUMENTS(SDOperand Op, SelectionDAG &DAG,
MachineFunction &MF = DAG.getMachineFunction();
MachineFrameInfo *MFI = MF.getFrameInfo();
SSARegMap *RegMap = MF.getSSARegMap();
- std::vector<SDOperand> ArgValues;
+ SmallVector<SDOperand, 8> ArgValues;
SDOperand Root = Op.getOperand(0);
- unsigned ArgOffset = 24;
- const unsigned Num_GPR_Regs = 8;
- const unsigned Num_FPR_Regs = 13;
- const unsigned Num_VR_Regs = 12;
- unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+ bool isPPC64 = PtrVT == MVT::i64;
+ unsigned PtrByteSize = isPPC64 ? 8 : 4;
+
+ unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64);
static const unsigned GPR_32[] = { // 32-bit registers.
PPC::R3, PPC::R4, PPC::R5, PPC::R6,
PPC::V9, PPC::V10, PPC::V11, PPC::V12, PPC::V13
};
- MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
- bool isPPC64 = PtrVT == MVT::i64;
+ const unsigned Num_GPR_Regs = sizeof(GPR_32)/sizeof(GPR_32[0]);
+ const unsigned Num_FPR_Regs = sizeof(FPR)/sizeof(FPR[0]);
+ const unsigned Num_VR_Regs = sizeof( VR)/sizeof( VR[0]);
+
+ unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
+
const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
// Add DAG nodes to load the arguments or copy them out of registers. On
- // entry to a function on PPC, the arguments start at offset 24, although the
- // first ones are often in registers.
+ // entry to a function on PPC, the arguments start after the linkage area,
+ // although the first ones are often in registers.
for (unsigned ArgNo = 0, e = Op.Val->getNumValues()-1; ArgNo != e; ++ArgNo) {
SDOperand ArgVal;
bool needsLoad = false;
MVT::ValueType ObjectVT = Op.getValue(ArgNo).getValueType();
unsigned ObjSize = MVT::getSizeInBits(ObjectVT)/8;
+ unsigned ArgSize = ObjSize;
unsigned CurArgOffset = ArgOffset;
switch (ObjectVT) {
default: assert(0 && "Unhandled argument type!");
case MVT::i32:
// All int arguments reserve stack space.
- ArgOffset += isPPC64 ? 8 : 4;
+ ArgOffset += PtrByteSize;
if (GPR_idx != Num_GPR_Regs) {
unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
++GPR_idx;
} else {
needsLoad = true;
+ ArgSize = PtrByteSize;
}
break;
case MVT::i64: // PPC64
case MVT::f32:
case MVT::f64:
// All FP arguments reserve stack space.
- ArgOffset += ObjSize;
+ ArgOffset += isPPC64 ? 8 : ObjSize;
// Every 4 bytes of argument space consumes one of the GPRs available for
// argument passing.
if (GPR_idx != Num_GPR_Regs) {
++GPR_idx;
- if (ObjSize == 8 && GPR_idx != Num_GPR_Regs)
+ if (ObjSize == 8 && GPR_idx != Num_GPR_Regs && !isPPC64)
++GPR_idx;
}
if (FPR_idx != Num_FPR_Regs) {
// If the argument is actually used, emit a load from the right stack
// slot.
if (!Op.Val->hasNUsesOfValue(0, ArgNo)) {
- int FI = MFI->CreateFixedObject(ObjSize, CurArgOffset);
+ int FI = MFI->CreateFixedObject(ObjSize,
+ CurArgOffset + (ArgSize - ObjSize));
SDOperand FIN = DAG.getFrameIndex(FI, PtrVT);
- ArgVal = DAG.getLoad(ObjectVT, Root, FIN,
- DAG.getSrcValue(NULL));
+ ArgVal = DAG.getLoad(ObjectVT, Root, FIN, NULL, 0);
} else {
// Don't emit a dead load.
ArgVal = DAG.getNode(ISD::UNDEF, ObjectVT);
// 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.
- std::vector<SDOperand> MemOps;
+ SmallVector<SDOperand, 8> MemOps;
for (; GPR_idx != Num_GPR_Regs; ++GPR_idx) {
- unsigned VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
+ unsigned VReg;
+ if (isPPC64)
+ VReg = RegMap->createVirtualRegister(&PPC::G8RCRegClass);
+ else
+ VReg = RegMap->createVirtualRegister(&PPC::GPRCRegClass);
+
MF.addLiveIn(GPR[GPR_idx], VReg);
SDOperand Val = DAG.getCopyFromReg(Root, VReg, PtrVT);
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
- Val, FIN, DAG.getSrcValue(NULL));
+ SDOperand Store = DAG.getStore(Val.getValue(1), Val, FIN, NULL, 0);
MemOps.push_back(Store);
// Increment the address by four for the next argument to store
SDOperand PtrOff = DAG.getConstant(MVT::getSizeInBits(PtrVT)/8, PtrVT);
FIN = DAG.getNode(ISD::ADD, PtrOff.getValueType(), FIN, PtrOff);
}
if (!MemOps.empty())
- Root = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
+ Root = DAG.getNode(ISD::TokenFactor, MVT::Other,&MemOps[0],MemOps.size());
}
ArgValues.push_back(Root);
// Return the new list of results.
std::vector<MVT::ValueType> RetVT(Op.Val->value_begin(),
Op.Val->value_end());
- return DAG.getNode(ISD::MERGE_VALUES, RetVT, ArgValues);
+ return DAG.getNode(ISD::MERGE_VALUES, RetVT, &ArgValues[0], ArgValues.size());
}
/// isCallCompatibleAddress - Return the immediate to use if the specified
return DAG.getConstant((int)C->getValue() >> 2, MVT::i32).Val;
}
-
static SDOperand LowerCALL(SDOperand Op, SelectionDAG &DAG) {
SDOperand Chain = Op.getOperand(0);
- unsigned CallingConv= cast<ConstantSDNode>(Op.getOperand(1))->getValue();
bool isVarArg = cast<ConstantSDNode>(Op.getOperand(2))->getValue() != 0;
- bool isTailCall = cast<ConstantSDNode>(Op.getOperand(3))->getValue() != 0;
SDOperand Callee = Op.getOperand(4);
unsigned NumOps = (Op.getNumOperands() - 5) / 2;
MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
bool isPPC64 = PtrVT == MVT::i64;
unsigned PtrByteSize = isPPC64 ? 8 : 4;
-
// args_to_use will accumulate outgoing args for the PPCISD::CALL case in
// SelectExpr to use to put the arguments in the appropriate registers.
// Count how many bytes are to be pushed on the stack, including the linkage
// area, and parameter passing area. We start with 24/48 bytes, which is
// prereserved space for [SP][CR][LR][3 x unused].
- unsigned NumBytes = 6*PtrByteSize;
+ unsigned NumBytes = PPCFrameInfo::getLinkageSize(isPPC64);
// Add up all the space actually used.
- for (unsigned i = 0; i != NumOps; ++i)
- NumBytes += MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
+ for (unsigned i = 0; i != NumOps; ++i) {
+ unsigned ArgSize =MVT::getSizeInBits(Op.getOperand(5+2*i).getValueType())/8;
+ ArgSize = std::max(ArgSize, PtrByteSize);
+ NumBytes += ArgSize;
+ }
// The prolog code of the callee may store up to 8 GPR argument registers to
// the stack, allowing va_start to index over them in memory if its varargs.
// Because we cannot tell if this is needed on the caller side, we have to
// conservatively assume that it is needed. As such, make sure we have at
// least enough stack space for the caller to store the 8 GPRs.
- if (NumBytes < 6*PtrByteSize+8*PtrByteSize)
- NumBytes = 6*PtrByteSize+8*PtrByteSize;
+ NumBytes = std::max(NumBytes, PPCFrameInfo::getMinCallFrameSize(isPPC64));
// Adjust the stack pointer for the new arguments...
// These operations are automatically eliminated by the prolog/epilog pass
// memory. Also, if this is a vararg function, floating point operations
// must be stored to our stack, and loaded into integer regs as well, if
// any integer regs are available for argument passing.
- unsigned ArgOffset = 6*PtrByteSize;
+ unsigned ArgOffset = PPCFrameInfo::getLinkageSize(isPPC64);
unsigned GPR_idx = 0, FPR_idx = 0, VR_idx = 0;
+
static const unsigned GPR_32[] = { // 32-bit registers.
PPC::R3, PPC::R4, PPC::R5, PPC::R6,
PPC::R7, PPC::R8, PPC::R9, PPC::R10,
const unsigned *GPR = isPPC64 ? GPR_64 : GPR_32;
std::vector<std::pair<unsigned, SDOperand> > RegsToPass;
- std::vector<SDOperand> MemOpChains;
+ SmallVector<SDOperand, 8> MemOpChains;
for (unsigned i = 0; i != NumOps; ++i) {
SDOperand Arg = Op.getOperand(5+2*i);
if (GPR_idx != NumGPRs) {
RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Arg));
} else {
- MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Arg, PtrOff, DAG.getSrcValue(NULL)));
+ MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
}
ArgOffset += PtrByteSize;
break;
case MVT::f32:
case MVT::f64:
+ if (isVarArg && isPPC64) {
+ // Float varargs need to be promoted to double.
+ if (Arg.getValueType() == MVT::f32)
+ Arg = DAG.getNode(ISD::FP_EXTEND, MVT::f64, Arg);
+ }
+
if (FPR_idx != NumFPRs) {
RegsToPass.push_back(std::make_pair(FPR[FPR_idx++], Arg));
if (isVarArg) {
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Arg, PtrOff,
- DAG.getSrcValue(NULL));
+ SDOperand Store = DAG.getStore(Chain, Arg, PtrOff, NULL, 0);
MemOpChains.push_back(Store);
// Float varargs are always shadowed in available integer registers
if (GPR_idx != NumGPRs) {
- SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff,
- DAG.getSrcValue(NULL));
+ SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
MemOpChains.push_back(Load.getValue(1));
RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
}
- if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64) {
+ if (GPR_idx != NumGPRs && Arg.getValueType() == MVT::f64 && !isPPC64){
SDOperand ConstFour = DAG.getConstant(4, PtrOff.getValueType());
PtrOff = DAG.getNode(ISD::ADD, PtrVT, PtrOff, ConstFour);
- SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff,
- DAG.getSrcValue(NULL));
+ SDOperand Load = DAG.getLoad(PtrVT, Store, PtrOff, NULL, 0);
MemOpChains.push_back(Load.getValue(1));
RegsToPass.push_back(std::make_pair(GPR[GPR_idx++], Load));
}
++GPR_idx;
}
} else {
- MemOpChains.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Arg, PtrOff, DAG.getSrcValue(NULL)));
+ MemOpChains.push_back(DAG.getStore(Chain, Arg, PtrOff, NULL, 0));
}
if (isPPC64)
ArgOffset += 8;
}
}
if (!MemOpChains.empty())
- Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOpChains);
+ Chain = DAG.getNode(ISD::TokenFactor, MVT::Other,
+ &MemOpChains[0], MemOpChains.size());
// Build a sequence of copy-to-reg nodes chained together with token chain
// and flag operands which copy the outgoing args into the appropriate regs.
NodeTys.push_back(MVT::Other); // Returns a chain
NodeTys.push_back(MVT::Flag); // Returns a flag for retval copy to use.
- std::vector<SDOperand> Ops;
+ SmallVector<SDOperand, 8> Ops;
unsigned CallOpc = PPCISD::CALL;
// If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
else {
// Otherwise, this is an indirect call. We have to use a MTCTR/BCTRL pair
// to do the call, we can't use PPCISD::CALL.
- Ops.push_back(Chain);
- Ops.push_back(Callee);
-
- if (InFlag.Val)
- Ops.push_back(InFlag);
- Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, Ops);
+ SDOperand MTCTROps[] = {Chain, Callee, InFlag};
+ Chain = DAG.getNode(PPCISD::MTCTR, NodeTys, MTCTROps, 2+(InFlag.Val!=0));
InFlag = Chain.getValue(1);
// Copy the callee address into R12 on darwin.
NodeTys.clear();
NodeTys.push_back(MVT::Other);
NodeTys.push_back(MVT::Flag);
- Ops.clear();
Ops.push_back(Chain);
CallOpc = PPCISD::BCTRL;
Callee.Val = 0;
if (InFlag.Val)
Ops.push_back(InFlag);
- Chain = DAG.getNode(CallOpc, NodeTys, Ops);
+ Chain = DAG.getNode(CallOpc, NodeTys, &Ops[0], Ops.size());
InFlag = Chain.getValue(1);
- std::vector<SDOperand> ResultVals;
+ SDOperand ResultVals[3];
+ unsigned NumResults = 0;
NodeTys.clear();
// If the call has results, copy the values out of the ret val registers.
case MVT::i32:
if (Op.Val->getValueType(1) == MVT::i32) {
Chain = DAG.getCopyFromReg(Chain, PPC::R4, MVT::i32, InFlag).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[0] = Chain.getValue(0);
Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32,
Chain.getValue(2)).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[1] = Chain.getValue(0);
+ NumResults = 2;
NodeTys.push_back(MVT::i32);
} else {
Chain = DAG.getCopyFromReg(Chain, PPC::R3, MVT::i32, InFlag).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[0] = Chain.getValue(0);
+ NumResults = 1;
}
NodeTys.push_back(MVT::i32);
break;
case MVT::i64:
Chain = DAG.getCopyFromReg(Chain, PPC::X3, MVT::i64, InFlag).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[0] = Chain.getValue(0);
+ NumResults = 1;
NodeTys.push_back(MVT::i64);
break;
case MVT::f32:
case MVT::f64:
Chain = DAG.getCopyFromReg(Chain, PPC::F1, Op.Val->getValueType(0),
InFlag).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[0] = Chain.getValue(0);
+ NumResults = 1;
NodeTys.push_back(Op.Val->getValueType(0));
break;
case MVT::v4f32:
case MVT::v16i8:
Chain = DAG.getCopyFromReg(Chain, PPC::V2, Op.Val->getValueType(0),
InFlag).getValue(1);
- ResultVals.push_back(Chain.getValue(0));
+ ResultVals[0] = Chain.getValue(0);
+ NumResults = 1;
NodeTys.push_back(Op.Val->getValueType(0));
break;
}
NodeTys.push_back(MVT::Other);
// If the function returns void, just return the chain.
- if (ResultVals.empty())
+ if (NumResults == 0)
return Chain;
// Otherwise, merge everything together with a MERGE_VALUES node.
- ResultVals.push_back(Chain);
- SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys, ResultVals);
+ ResultVals[NumResults++] = Chain;
+ SDOperand Res = DAG.getNode(ISD::MERGE_VALUES, NodeTys,
+ ResultVals, NumResults);
return Res.getValue(Op.ResNo);
}
ArgReg = PPC::R3;
} else if (ArgVT == MVT::i64) {
ArgReg = PPC::X3;
- } else if (MVT::isFloatingPoint(ArgVT)) {
- ArgReg = PPC::F1;
- } else {
- assert(MVT::isVector(ArgVT));
+ } else if (MVT::isVector(ArgVT)) {
ArgReg = PPC::V2;
+ } else {
+ assert(MVT::isFloatingPoint(ArgVT));
+ ArgReg = PPC::F1;
}
Copy = DAG.getCopyToReg(Op.getOperand(0), ArgReg, Op.getOperand(1),
return DAG.getNode(PPCISD::RET_FLAG, MVT::Other, Copy, Copy.getValue(1));
}
+static SDOperand LowerSTACKRESTORE(SDOperand Op, SelectionDAG &DAG,
+ const PPCSubtarget &Subtarget) {
+ // When we pop the dynamic allocation we need to restore the SP link.
+
+ // Get the corect type for pointers.
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+
+ // Construct the stack pointer operand.
+ bool IsPPC64 = Subtarget.isPPC64();
+ unsigned SP = IsPPC64 ? PPC::X1 : PPC::R1;
+ SDOperand StackPtr = DAG.getRegister(SP, PtrVT);
+
+ // Get the operands for the STACKRESTORE.
+ SDOperand Chain = Op.getOperand(0);
+ SDOperand SaveSP = Op.getOperand(1);
+
+ // Load the old link SP.
+ SDOperand LoadLinkSP = DAG.getLoad(PtrVT, Chain, StackPtr, NULL, 0);
+
+ // Restore the stack pointer.
+ Chain = DAG.getCopyToReg(LoadLinkSP.getValue(1), SP, SaveSP);
+
+ // Store the old link SP.
+ return DAG.getStore(Chain, LoadLinkSP, StackPtr, NULL, 0);
+}
+
+static SDOperand LowerDYNAMIC_STACKALLOC(SDOperand Op, SelectionDAG &DAG,
+ const PPCSubtarget &Subtarget) {
+ MachineFunction &MF = DAG.getMachineFunction();
+ bool IsPPC64 = Subtarget.isPPC64();
+
+ // Get current frame pointer save index. The users of this index will be
+ // primarily DYNALLOC instructions.
+ PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>();
+ int FPSI = FI->getFramePointerSaveIndex();
+
+ // If the frame pointer save index hasn't been defined yet.
+ if (!FPSI) {
+ // Find out what the fix offset of the frame pointer save area.
+ int Offset = PPCFrameInfo::getFramePointerSaveOffset(IsPPC64);
+ // Allocate the frame index for frame pointer save area.
+ FPSI = MF.getFrameInfo()->CreateFixedObject(IsPPC64? 8 : 4, Offset);
+ // Save the result.
+ FI->setFramePointerSaveIndex(FPSI);
+ }
+
+ // Get the inputs.
+ SDOperand Chain = Op.getOperand(0);
+ SDOperand Size = Op.getOperand(1);
+
+ // Get the corect type for pointers.
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+ // Negate the size.
+ SDOperand NegSize = DAG.getNode(ISD::SUB, PtrVT,
+ DAG.getConstant(0, PtrVT), Size);
+ // Construct a node for the frame pointer save index.
+ SDOperand FPSIdx = DAG.getFrameIndex(FPSI, PtrVT);
+ // Build a DYNALLOC node.
+ SDOperand Ops[3] = { Chain, NegSize, FPSIdx };
+ SDVTList VTs = DAG.getVTList(PtrVT, MVT::Other);
+ return DAG.getNode(PPCISD::DYNALLOC, VTs, Ops, 3);
+}
+
+
/// LowerSELECT_CC - Lower floating point select_cc's into fsel instruction when
/// possible.
static SDOperand LowerSELECT_CC(SDOperand Op, SelectionDAG &DAG) {
// then lfd it and fcfid it.
MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
int FrameIdx = FrameInfo->CreateStackObject(8, 8);
- SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+ SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
SDOperand Ext64 = DAG.getNode(PPCISD::EXTSW_32, MVT::i32,
Op.getOperand(0));
DAG.getEntryNode(), Ext64, FIdx,
DAG.getSrcValue(NULL));
// Load the value as a double.
- SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, DAG.getSrcValue(NULL));
+ SDOperand Ld = DAG.getLoad(MVT::f64, Store, FIdx, NULL, 0);
// FCFID it and return it.
SDOperand FP = DAG.getNode(PPCISD::FCFID, MVT::f64, Ld);
return FP;
}
-static SDOperand LowerSHL(SDOperand Op, SelectionDAG &DAG,
- MVT::ValueType PtrVT) {
- assert(Op.getValueType() == MVT::i64 &&
+static SDOperand LowerSHL_PARTS(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
- // The generic code does a fine job expanding shift by a constant.
- if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
- // Otherwise, expand into a bunch of logical ops. Note that these ops
+ // Expand into a bunch of logical ops. Note that these ops
// depend on the PPC behavior for oversized shift amounts.
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(0, PtrVT));
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(1, PtrVT));
- SDOperand Amt = Op.getOperand(1);
+ SDOperand Lo = Op.getOperand(0);
+ SDOperand Hi = Op.getOperand(1);
+ SDOperand Amt = Op.getOperand(2);
SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
DAG.getConstant(32, MVT::i32), Amt);
SDOperand Tmp6 = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Tmp5);
SDOperand OutHi = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
SDOperand OutLo = DAG.getNode(PPCISD::SHL, MVT::i32, Lo, Amt);
- return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
+ SDOperand OutOps[] = { OutLo, OutHi };
+ return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
+ OutOps, 2);
}
-static SDOperand LowerSRL(SDOperand Op, SelectionDAG &DAG,
- MVT::ValueType PtrVT) {
- assert(Op.getValueType() == MVT::i64 &&
- Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SHL!");
- // The generic code does a fine job expanding shift by a constant.
- if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
+static SDOperand LowerSRL_PARTS(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
+ Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRL!");
// Otherwise, expand into a bunch of logical ops. Note that these ops
// depend on the PPC behavior for oversized shift amounts.
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(0, PtrVT));
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(1, PtrVT));
- SDOperand Amt = Op.getOperand(1);
+ SDOperand Lo = Op.getOperand(0);
+ SDOperand Hi = Op.getOperand(1);
+ SDOperand Amt = Op.getOperand(2);
SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
DAG.getConstant(32, MVT::i32), Amt);
SDOperand Tmp6 = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Tmp5);
SDOperand OutLo = DAG.getNode(ISD::OR, MVT::i32, Tmp4, Tmp6);
SDOperand OutHi = DAG.getNode(PPCISD::SRL, MVT::i32, Hi, Amt);
- return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
+ SDOperand OutOps[] = { OutLo, OutHi };
+ return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
+ OutOps, 2);
}
-static SDOperand LowerSRA(SDOperand Op, SelectionDAG &DAG,
- MVT::ValueType PtrVT) {
- assert(Op.getValueType() == MVT::i64 &&
+static SDOperand LowerSRA_PARTS(SDOperand Op, SelectionDAG &DAG) {
+ assert(Op.getNumOperands() == 3 && Op.getValueType() == MVT::i32 &&
Op.getOperand(1).getValueType() == MVT::i32 && "Unexpected SRA!");
- // The generic code does a fine job expanding shift by a constant.
- if (isa<ConstantSDNode>(Op.getOperand(1))) return SDOperand();
// Otherwise, expand into a bunch of logical ops, followed by a select_cc.
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(0, PtrVT));
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32, Op.getOperand(0),
- DAG.getConstant(1, PtrVT));
- SDOperand Amt = Op.getOperand(1);
+ SDOperand Lo = Op.getOperand(0);
+ SDOperand Hi = Op.getOperand(1);
+ SDOperand Amt = Op.getOperand(2);
SDOperand Tmp1 = DAG.getNode(ISD::SUB, MVT::i32,
DAG.getConstant(32, MVT::i32), Amt);
SDOperand OutHi = DAG.getNode(PPCISD::SRA, MVT::i32, Hi, Amt);
SDOperand OutLo = DAG.getSelectCC(Tmp5, DAG.getConstant(0, MVT::i32),
Tmp4, Tmp6, ISD::SETLE);
- return DAG.getNode(ISD::BUILD_PAIR, MVT::i64, OutLo, OutHi);
+ SDOperand OutOps[] = { OutLo, OutHi };
+ return DAG.getNode(ISD::MERGE_VALUES, DAG.getVTList(MVT::i32, MVT::i32),
+ OutOps, 2);
}
//===----------------------------------------------------------------------===//
static SDOperand BuildSplatI(int Val, unsigned SplatSize, MVT::ValueType VT,
SelectionDAG &DAG) {
assert(Val >= -16 && Val <= 15 && "vsplti is out of range!");
-
- // Force vspltis[hw] -1 to vspltisb -1.
- if (Val == -1) SplatSize = 1;
-
+
static const MVT::ValueType VTys[] = { // canonical VT to use for each size.
MVT::v16i8, MVT::v8i16, MVT::Other, MVT::v4i32
};
+
+ MVT::ValueType ReqVT = VT != MVT::Other ? VT : VTys[SplatSize-1];
+
+ // Force vspltis[hw] -1 to vspltisb -1 to canonicalize.
+ if (Val == -1)
+ SplatSize = 1;
+
MVT::ValueType CanonicalVT = VTys[SplatSize-1];
// Build a canonical splat for this value.
SDOperand Elt = DAG.getConstant(Val, MVT::getVectorBaseType(CanonicalVT));
- std::vector<SDOperand> Ops(MVT::getVectorNumElements(CanonicalVT), Elt);
- SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT, Ops);
- return DAG.getNode(ISD::BIT_CONVERT, VT, Res);
+ SmallVector<SDOperand, 8> Ops;
+ Ops.assign(MVT::getVectorNumElements(CanonicalVT), Elt);
+ SDOperand Res = DAG.getNode(ISD::BUILD_VECTOR, CanonicalVT,
+ &Ops[0], Ops.size());
+ return DAG.getNode(ISD::BIT_CONVERT, ReqVT, Res);
}
/// BuildIntrinsicOp - Return a binary operator intrinsic node with the
LHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, LHS);
RHS = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, RHS);
- std::vector<SDOperand> Ops;
+ SDOperand Ops[16];
for (unsigned i = 0; i != 16; ++i)
- Ops.push_back(DAG.getConstant(i+Amt, MVT::i32));
+ Ops[i] = DAG.getConstant(i+Amt, MVT::i32);
SDOperand T = DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, LHS, RHS,
- DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops));
+ DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops,16));
return DAG.getNode(ISD::BIT_CONVERT, VT, T);
}
-1, 1, -2, 2, -3, 3, -4, 4, -5, 5, -6, 6, -7, 7,
-8, 8, -9, 9, -10, 10, -11, 11, -12, 12, -13, 13, 14, -14, 15, -15, -16
};
+
for (unsigned idx = 0; idx < sizeof(SplatCsts)/sizeof(SplatCsts[0]); ++idx){
// Indirect through the SplatCsts array so that we favor 'vsplti -1' for
// cases which are ambiguous (e.g. formation of 0x8000_0000). 'vsplti -1'
// vsplti + shl self.
if (SextVal == (i << (int)TypeShiftAmt)) {
- Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
+ SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
static const unsigned IIDs[] = { // Intrinsic to use for each size.
Intrinsic::ppc_altivec_vslb, Intrinsic::ppc_altivec_vslh, 0,
Intrinsic::ppc_altivec_vslw
};
- return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
+ Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
}
// vsplti + srl self.
if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
- Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
+ SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
static const unsigned IIDs[] = { // Intrinsic to use for each size.
Intrinsic::ppc_altivec_vsrb, Intrinsic::ppc_altivec_vsrh, 0,
Intrinsic::ppc_altivec_vsrw
};
- return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
+ Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
}
// vsplti + sra self.
if (SextVal == (int)((unsigned)i >> TypeShiftAmt)) {
- Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
+ SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
static const unsigned IIDs[] = { // Intrinsic to use for each size.
Intrinsic::ppc_altivec_vsrab, Intrinsic::ppc_altivec_vsrah, 0,
Intrinsic::ppc_altivec_vsraw
};
- return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
+ Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
}
// vsplti + rol self.
if (SextVal == (int)(((unsigned)i << TypeShiftAmt) |
((unsigned)i >> (SplatBitSize-TypeShiftAmt)))) {
- Op = BuildSplatI(i, SplatSize, Op.getValueType(), DAG);
+ SDOperand Res = BuildSplatI(i, SplatSize, MVT::Other, DAG);
static const unsigned IIDs[] = { // Intrinsic to use for each size.
Intrinsic::ppc_altivec_vrlb, Intrinsic::ppc_altivec_vrlh, 0,
Intrinsic::ppc_altivec_vrlw
};
- return BuildIntrinsicOp(IIDs[SplatSize-1], Op, Op, DAG);
+ Res = BuildIntrinsicOp(IIDs[SplatSize-1], Res, Res, DAG);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), Res);
}
// t = vsplti c, result = vsldoi t, t, 1
// Odd, in range [17,31]: (vsplti C)-(vsplti -16).
if (SextVal >= 0 && SextVal <= 31) {
- SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, Op.getValueType(),DAG);
- SDOperand RHS = BuildSplatI(-16, SplatSize, Op.getValueType(), DAG);
- return DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
+ SDOperand LHS = BuildSplatI(SextVal-16, SplatSize, MVT::Other, DAG);
+ SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
+ LHS = DAG.getNode(ISD::SUB, Op.getValueType(), LHS, RHS);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
}
// Odd, in range [-31,-17]: (vsplti C)+(vsplti -16).
if (SextVal >= -31 && SextVal <= 0) {
- SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, Op.getValueType(),DAG);
- SDOperand RHS = BuildSplatI(-16, SplatSize, Op.getValueType(), DAG);
- return DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
+ SDOperand LHS = BuildSplatI(SextVal+16, SplatSize, MVT::Other, DAG);
+ SDOperand RHS = BuildSplatI(-16, SplatSize, MVT::Other, DAG);
+ LHS = DAG.getNode(ISD::ADD, Op.getValueType(), LHS, RHS);
+ return DAG.getNode(ISD::BIT_CONVERT, Op.getValueType(), LHS);
}
}
case OP_VSLDOI12:
return BuildVSLDOI(OpLHS, OpRHS, 12, OpLHS.getValueType(), DAG);
}
- std::vector<SDOperand> Ops;
+ SDOperand Ops[16];
for (unsigned i = 0; i != 16; ++i)
- Ops.push_back(DAG.getConstant(ShufIdxs[i], MVT::i32));
+ Ops[i] = DAG.getConstant(ShufIdxs[i], MVT::i32);
return DAG.getNode(ISD::VECTOR_SHUFFLE, OpLHS.getValueType(), OpLHS, OpRHS,
- DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops));
+ DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
}
/// LowerVECTOR_SHUFFLE - Return the code we lower for VECTOR_SHUFFLE. If this
MVT::ValueType EltVT = MVT::getVectorBaseType(V1.getValueType());
unsigned BytesPerElement = MVT::getSizeInBits(EltVT)/8;
- std::vector<SDOperand> ResultMask;
+ SmallVector<SDOperand, 16> ResultMask;
for (unsigned i = 0, e = PermMask.getNumOperands(); i != e; ++i) {
unsigned SrcElt;
if (PermMask.getOperand(i).getOpcode() == ISD::UNDEF)
MVT::i8));
}
- SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, ResultMask);
+ SDOperand VPermMask = DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8,
+ &ResultMask[0], ResultMask.size());
return DAG.getNode(PPCISD::VPERM, V1.getValueType(), V1, V2, VPermMask);
}
}
// Create the PPCISD altivec 'dot' comparison node.
- std::vector<SDOperand> Ops;
+ SDOperand Ops[] = {
+ Op.getOperand(2), // LHS
+ Op.getOperand(3), // RHS
+ DAG.getConstant(CompareOpc, MVT::i32)
+ };
std::vector<MVT::ValueType> VTs;
- Ops.push_back(Op.getOperand(2)); // LHS
- Ops.push_back(Op.getOperand(3)); // RHS
- Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
VTs.push_back(Op.getOperand(2).getValueType());
VTs.push_back(MVT::Flag);
- SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
+ SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
// Now that we have the comparison, emit a copy from the CR to a GPR.
// This is flagged to the above dot comparison.
// Create a stack slot that is 16-byte aligned.
MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
int FrameIdx = FrameInfo->CreateStackObject(16, 16);
- SDOperand FIdx = DAG.getFrameIndex(FrameIdx, MVT::i32);
+ MVT::ValueType PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
+ SDOperand FIdx = DAG.getFrameIndex(FrameIdx, PtrVT);
// Store the input value into Value#0 of the stack slot.
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, DAG.getEntryNode(),
- Op.getOperand(0), FIdx,DAG.getSrcValue(NULL));
+ SDOperand Store = DAG.getStore(DAG.getEntryNode(),
+ Op.getOperand(0), FIdx, NULL, 0);
// Load it out.
- return DAG.getLoad(Op.getValueType(), Store, FIdx, DAG.getSrcValue(NULL));
+ return DAG.getLoad(Op.getValueType(), Store, FIdx, NULL, 0);
}
static SDOperand LowerMUL(SDOperand Op, SelectionDAG &DAG) {
OddParts = DAG.getNode(ISD::BIT_CONVERT, MVT::v16i8, OddParts);
// Merge the results together.
- std::vector<SDOperand> Ops;
+ SDOperand Ops[16];
for (unsigned i = 0; i != 8; ++i) {
- Ops.push_back(DAG.getConstant(2*i+1, MVT::i8));
- Ops.push_back(DAG.getConstant(2*i+1+16, MVT::i8));
+ Ops[i*2 ] = DAG.getConstant(2*i+1, MVT::i8);
+ Ops[i*2+1] = DAG.getConstant(2*i+1+16, MVT::i8);
}
-
return DAG.getNode(ISD::VECTOR_SHUFFLE, MVT::v16i8, EvenParts, OddParts,
- DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops));
+ DAG.getNode(ISD::BUILD_VECTOR, MVT::v16i8, Ops, 16));
} else {
assert(0 && "Unknown mul to lower!");
abort();
return LowerFORMAL_ARGUMENTS(Op, DAG, VarArgsFrameIndex);
case ISD::CALL: return LowerCALL(Op, DAG);
case ISD::RET: return LowerRET(Op, DAG);
+ case ISD::STACKRESTORE: return LowerSTACKRESTORE(Op, DAG, PPCSubTarget);
+ case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG,
+ PPCSubTarget);
case ISD::SELECT_CC: return LowerSELECT_CC(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
// Lower 64-bit shifts.
- case ISD::SHL: return LowerSHL(Op, DAG, getPointerTy());
- case ISD::SRL: return LowerSRL(Op, DAG, getPointerTy());
- case ISD::SRA: return LowerSRA(Op, DAG, getPointerTy());
+ case ISD::SHL_PARTS: return LowerSHL_PARTS(Op, DAG);
+ case ISD::SRL_PARTS: return LowerSRL_PARTS(Op, DAG);
+ case ISD::SRA_PARTS: return LowerSRA_PARTS(Op, DAG);
// Vector-related lowering.
case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
MachineBasicBlock *
PPCTargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
MachineBasicBlock *BB) {
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
assert((MI->getOpcode() == PPC::SELECT_CC_I4 ||
MI->getOpcode() == PPC::SELECT_CC_I8 ||
MI->getOpcode() == PPC::SELECT_CC_F4 ||
MachineBasicBlock *thisMBB = BB;
MachineBasicBlock *copy0MBB = new MachineBasicBlock(LLVM_BB);
MachineBasicBlock *sinkMBB = new MachineBasicBlock(LLVM_BB);
- BuildMI(BB, MI->getOperand(4).getImmedValue(), 2)
- .addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
+ unsigned SelectPred = MI->getOperand(4).getImm();
+ BuildMI(BB, TII->get(PPC::BCC))
+ .addImm(SelectPred).addReg(MI->getOperand(1).getReg()).addMBB(sinkMBB);
MachineFunction *F = BB->getParent();
F->getBasicBlockList().insert(It, copy0MBB);
F->getBasicBlockList().insert(It, sinkMBB);
// %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
// ...
BB = sinkMBB;
- BuildMI(BB, PPC::PHI, 4, MI->getOperand(0).getReg())
+ BuildMI(BB, TII->get(PPC::PHI), MI->getOperand(0).getReg())
.addReg(MI->getOperand(3).getReg()).addMBB(copy0MBB)
.addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
SelectionDAG &DAG = DCI.DAG;
switch (N->getOpcode()) {
default: break;
+ case PPCISD::SHL:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
+ if (C->getValue() == 0) // 0 << V -> 0.
+ return N->getOperand(0);
+ }
+ break;
+ case PPCISD::SRL:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
+ if (C->getValue() == 0) // 0 >>u V -> 0.
+ return N->getOperand(0);
+ }
+ break;
+ case PPCISD::SRA:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(0))) {
+ if (C->getValue() == 0 || // 0 >>s V -> 0.
+ C->isAllOnesValue()) // -1 >>s V -> -1.
+ return N->getOperand(0);
+ }
+ break;
+
case ISD::SINT_TO_FP:
if (TM.getSubtarget<PPCSubtarget>().has64BitSupport()) {
if (N->getOperand(0).getOpcode() == ISD::FP_TO_SINT) {
DCI.AddToWorklist(Val.Val);
return Val;
}
+
+ // Turn STORE (BSWAP) -> sthbrx/stwbrx.
+ if (N->getOperand(1).getOpcode() == ISD::BSWAP &&
+ N->getOperand(1).Val->hasOneUse() &&
+ (N->getOperand(1).getValueType() == MVT::i32 ||
+ N->getOperand(1).getValueType() == MVT::i16)) {
+ SDOperand BSwapOp = N->getOperand(1).getOperand(0);
+ // Do an any-extend to 32-bits if this is a half-word input.
+ if (BSwapOp.getValueType() == MVT::i16)
+ BSwapOp = DAG.getNode(ISD::ANY_EXTEND, MVT::i32, BSwapOp);
+
+ return DAG.getNode(PPCISD::STBRX, MVT::Other, N->getOperand(0), BSwapOp,
+ N->getOperand(2), N->getOperand(3),
+ DAG.getValueType(N->getOperand(1).getValueType()));
+ }
+ break;
+ case ISD::BSWAP:
+ // Turn BSWAP (LOAD) -> lhbrx/lwbrx.
+ if (ISD::isNON_EXTLoad(N->getOperand(0).Val) &&
+ N->getOperand(0).hasOneUse() &&
+ (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i16)) {
+ SDOperand Load = N->getOperand(0);
+ LoadSDNode *LD = cast<LoadSDNode>(Load);
+ // Create the byte-swapping load.
+ std::vector<MVT::ValueType> VTs;
+ VTs.push_back(MVT::i32);
+ VTs.push_back(MVT::Other);
+ SDOperand SV = DAG.getSrcValue(LD->getSrcValue(), LD->getSrcValueOffset());
+ SDOperand Ops[] = {
+ LD->getChain(), // Chain
+ LD->getBasePtr(), // Ptr
+ SV, // SrcValue
+ DAG.getValueType(N->getValueType(0)) // VT
+ };
+ SDOperand BSLoad = DAG.getNode(PPCISD::LBRX, VTs, Ops, 4);
+
+ // If this is an i16 load, insert the truncate.
+ SDOperand ResVal = BSLoad;
+ if (N->getValueType(0) == MVT::i16)
+ ResVal = DAG.getNode(ISD::TRUNCATE, MVT::i16, BSLoad);
+
+ // First, combine the bswap away. This makes the value produced by the
+ // load dead.
+ DCI.CombineTo(N, ResVal);
+
+ // Next, combine the load away, we give it a bogus result value but a real
+ // chain result. The result value is dead because the bswap is dead.
+ DCI.CombineTo(Load.Val, ResVal, BSLoad.getValue(1));
+
+ // Return N so it doesn't get rechecked!
+ return SDOperand(N, 0);
+ }
+
break;
case PPCISD::VCMP: {
// If a VCMPo node already exists with exactly the same operands as this
bool BranchOnWhenPredTrue = (CC == ISD::SETEQ) ^ (Val == 0);
// Create the PPCISD altivec 'dot' comparison node.
- std::vector<SDOperand> Ops;
std::vector<MVT::ValueType> VTs;
- Ops.push_back(LHS.getOperand(2)); // LHS of compare
- Ops.push_back(LHS.getOperand(3)); // RHS of compare
- Ops.push_back(DAG.getConstant(CompareOpc, MVT::i32));
+ SDOperand Ops[] = {
+ LHS.getOperand(2), // LHS of compare
+ LHS.getOperand(3), // RHS of compare
+ DAG.getConstant(CompareOpc, MVT::i32)
+ };
VTs.push_back(LHS.getOperand(2).getValueType());
VTs.push_back(MVT::Flag);
- SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops);
+ SDOperand CompNode = DAG.getNode(PPCISD::VCMPo, VTs, Ops, 3);
// Unpack the result based on how the target uses it.
- unsigned CompOpc;
+ PPC::Predicate CompOpc;
switch (cast<ConstantSDNode>(LHS.getOperand(1))->getValue()) {
default: // Can't happen, don't crash on invalid number though.
case 0: // Branch on the value of the EQ bit of CR6.
- CompOpc = BranchOnWhenPredTrue ? PPC::BEQ : PPC::BNE;
+ CompOpc = BranchOnWhenPredTrue ? PPC::PRED_EQ : PPC::PRED_NE;
break;
case 1: // Branch on the inverted value of the EQ bit of CR6.
- CompOpc = BranchOnWhenPredTrue ? PPC::BNE : PPC::BEQ;
+ CompOpc = BranchOnWhenPredTrue ? PPC::PRED_NE : PPC::PRED_EQ;
break;
case 2: // Branch on the value of the LT bit of CR6.
- CompOpc = BranchOnWhenPredTrue ? PPC::BLT : PPC::BGE;
+ CompOpc = BranchOnWhenPredTrue ? PPC::PRED_LT : PPC::PRED_GE;
break;
case 3: // Branch on the inverted value of the LT bit of CR6.
- CompOpc = BranchOnWhenPredTrue ? PPC::BGE : PPC::BLT;
+ CompOpc = BranchOnWhenPredTrue ? PPC::PRED_GE : PPC::PRED_LT;
break;
}
return DAG.getNode(PPCISD::COND_BRANCH, MVT::Other, N->getOperand(0),
- DAG.getRegister(PPC::CR6, MVT::i32),
DAG.getConstant(CompOpc, MVT::i32),
+ DAG.getRegister(PPC::CR6, MVT::i32),
N->getOperand(4), CompNode.getValue(1));
}
break;
KnownOne = 0;
switch (Op.getOpcode()) {
default: break;
+ case PPCISD::LBRX: {
+ // lhbrx is known to have the top bits cleared out.
+ if (cast<VTSDNode>(Op.getOperand(3))->getVT() == MVT::i16)
+ KnownZero = 0xFFFF0000;
+ break;
+ }
case ISD::INTRINSIC_WO_CHAIN: {
switch (cast<ConstantSDNode>(Op.getOperand(0))->getValue()) {
default: break;
return TargetLowering::getConstraintType(ConstraintLetter);
}
-
-std::vector<unsigned> PPCTargetLowering::
-getRegClassForInlineAsmConstraint(const std::string &Constraint,
- MVT::ValueType VT) const {
+std::pair<unsigned, const TargetRegisterClass*>
+PPCTargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint,
+ MVT::ValueType VT) const {
if (Constraint.size() == 1) {
- switch (Constraint[0]) { // GCC RS6000 Constraint Letters
- default: break; // Unknown constriant letter
- case 'b':
- return make_vector<unsigned>(/*no R0*/ PPC::R1 , PPC::R2 , PPC::R3 ,
- PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
- PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
- PPC::R12, PPC::R13, PPC::R14, PPC::R15,
- PPC::R16, PPC::R17, PPC::R18, PPC::R19,
- PPC::R20, PPC::R21, PPC::R22, PPC::R23,
- PPC::R24, PPC::R25, PPC::R26, PPC::R27,
- PPC::R28, PPC::R29, PPC::R30, PPC::R31,
- 0);
- case 'r':
- return make_vector<unsigned>(PPC::R0 , PPC::R1 , PPC::R2 , PPC::R3 ,
- PPC::R4 , PPC::R5 , PPC::R6 , PPC::R7 ,
- PPC::R8 , PPC::R9 , PPC::R10, PPC::R11,
- PPC::R12, PPC::R13, PPC::R14, PPC::R15,
- PPC::R16, PPC::R17, PPC::R18, PPC::R19,
- PPC::R20, PPC::R21, PPC::R22, PPC::R23,
- PPC::R24, PPC::R25, PPC::R26, PPC::R27,
- PPC::R28, PPC::R29, PPC::R30, PPC::R31,
- 0);
- case 'f':
- return make_vector<unsigned>(PPC::F0 , PPC::F1 , PPC::F2 , PPC::F3 ,
- PPC::F4 , PPC::F5 , PPC::F6 , PPC::F7 ,
- PPC::F8 , PPC::F9 , PPC::F10, PPC::F11,
- PPC::F12, PPC::F13, PPC::F14, PPC::F15,
- PPC::F16, PPC::F17, PPC::F18, PPC::F19,
- PPC::F20, PPC::F21, PPC::F22, PPC::F23,
- PPC::F24, PPC::F25, PPC::F26, PPC::F27,
- PPC::F28, PPC::F29, PPC::F30, PPC::F31,
- 0);
+ // GCC RS6000 Constraint Letters
+ switch (Constraint[0]) {
+ case 'b': // R1-R31
+ case 'r': // R0-R31
+ if (VT == MVT::i64 && PPCSubTarget.isPPC64())
+ return std::make_pair(0U, PPC::G8RCRegisterClass);
+ return std::make_pair(0U, PPC::GPRCRegisterClass);
+ case 'f':
+ if (VT == MVT::f32)
+ return std::make_pair(0U, PPC::F4RCRegisterClass);
+ else if (VT == MVT::f64)
+ return std::make_pair(0U, PPC::F8RCRegisterClass);
+ break;
case 'v':
- return make_vector<unsigned>(PPC::V0 , PPC::V1 , PPC::V2 , PPC::V3 ,
- PPC::V4 , PPC::V5 , PPC::V6 , PPC::V7 ,
- PPC::V8 , PPC::V9 , PPC::V10, PPC::V11,
- PPC::V12, PPC::V13, PPC::V14, PPC::V15,
- PPC::V16, PPC::V17, PPC::V18, PPC::V19,
- PPC::V20, PPC::V21, PPC::V22, PPC::V23,
- PPC::V24, PPC::V25, PPC::V26, PPC::V27,
- PPC::V28, PPC::V29, PPC::V30, PPC::V31,
- 0);
- case 'y':
- return make_vector<unsigned>(PPC::CR0, PPC::CR1, PPC::CR2, PPC::CR3,
- PPC::CR4, PPC::CR5, PPC::CR6, PPC::CR7,
- 0);
+ return std::make_pair(0U, PPC::VRRCRegisterClass);
+ case 'y': // crrc
+ return std::make_pair(0U, PPC::CRRCRegisterClass);
}
}
- return std::vector<unsigned>();
+ return TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
}
+
// isOperandValidForConstraint
-bool PPCTargetLowering::
-isOperandValidForConstraint(SDOperand Op, char Letter) {
+SDOperand PPCTargetLowering::
+isOperandValidForConstraint(SDOperand Op, char Letter, SelectionDAG &DAG) {
switch (Letter) {
default: break;
case 'I':
case 'N':
case 'O':
case 'P': {
- if (!isa<ConstantSDNode>(Op)) return false; // Must be an immediate.
+ if (!isa<ConstantSDNode>(Op)) return SDOperand(0,0);// Must be an immediate.
unsigned Value = cast<ConstantSDNode>(Op)->getValue();
switch (Letter) {
default: assert(0 && "Unknown constraint letter!");
case 'I': // "I" is a signed 16-bit constant.
- return (short)Value == (int)Value;
+ if ((short)Value == (int)Value) return Op;
+ break;
case 'J': // "J" is a constant with only the high-order 16 bits nonzero.
case 'L': // "L" is a signed 16-bit constant shifted left 16 bits.
- return (short)Value == 0;
+ if ((short)Value == 0) return Op;
+ break;
case 'K': // "K" is a constant with only the low-order 16 bits nonzero.
- return (Value >> 16) == 0;
+ if ((Value >> 16) == 0) return Op;
+ break;
case 'M': // "M" is a constant that is greater than 31.
- return Value > 31;
+ if (Value > 31) return Op;
+ break;
case 'N': // "N" is a positive constant that is an exact power of two.
- return (int)Value > 0 && isPowerOf2_32(Value);
+ if ((int)Value > 0 && isPowerOf2_32(Value)) return Op;
+ break;
case 'O': // "O" is the constant zero.
- return Value == 0;
+ if (Value == 0) return Op;
+ break;
case 'P': // "P" is a constant whose negation is a signed 16-bit constant.
- return (short)-Value == (int)-Value;
+ if ((short)-Value == (int)-Value) return Op;
+ break;
}
break;
}
}
// Handle standard constraint letters.
- return TargetLowering::isOperandValidForConstraint(Op, Letter);
+ return TargetLowering::isOperandValidForConstraint(Op, Letter, DAG);
}
/// isLegalAddressImmediate - Return true if the integer value can be used
// PPC allows a sign-extended 16-bit immediate field.
return (V > -(1 << 16) && V < (1 << 16)-1);
}
+
+bool PPCTargetLowering::isLegalAddressImmediate(llvm::GlobalValue* GV) const {
+ return TargetLowering::isLegalAddressImmediate(GV);
+}
projects / oota-llvm.git / blobdiff