#include "PowerPCInstrBuilder.h"
#include "PowerPCInstrInfo.h"
#include "PPC32TargetMachine.h"
+#include "PPC32ISelLowering.h"
#include "llvm/Constants.h"
#include "llvm/Function.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/CodeGen/SSARegMap.h"
#include "llvm/Target/TargetData.h"
-#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/MathExtras.h"
#include <algorithm>
using namespace llvm;
-
-//===----------------------------------------------------------------------===//
-// PPC32TargetLowering - PPC32 Implementation of the TargetLowering interface
-namespace {
- class PPC32TargetLowering : public TargetLowering {
- int VarArgsFrameIndex; // FrameIndex for start of varargs area.
- int ReturnAddrIndex; // FrameIndex for return slot.
- public:
- PPC32TargetLowering(TargetMachine &TM) : TargetLowering(TM) {
- // Fold away setcc operations if possible.
- setSetCCIsExpensive();
-
- // Set up the register classes.
- addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
- addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
- addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
-
- // PowerPC has no intrinsics for these particular operations
- setOperationAction(ISD::MEMMOVE, MVT::Other, 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);
-
- // We don't support sin/cos/sqrt/fmod
- setOperationAction(ISD::FSIN , MVT::f64, Expand);
- setOperationAction(ISD::FCOS , MVT::f64, Expand);
- setOperationAction(ISD::SREM , MVT::f64, Expand);
- setOperationAction(ISD::FSIN , MVT::f32, Expand);
- setOperationAction(ISD::FCOS , MVT::f32, Expand);
- setOperationAction(ISD::SREM , MVT::f32, Expand);
-
- // If we're enabling GP optimizations, use hardware square root
- if (!TM.getSubtarget<PPCSubtarget>().isGigaProcessor()) {
- setOperationAction(ISD::FSQRT, MVT::f64, Expand);
- setOperationAction(ISD::FSQRT, MVT::f32, Expand);
- }
-
- // PowerPC does not have CTPOP or CTTZ
- setOperationAction(ISD::CTPOP, MVT::i32 , Expand);
- setOperationAction(ISD::CTTZ , MVT::i32 , Expand);
-
- // PowerPC does not have Select
- setOperationAction(ISD::SELECT, MVT::i32, Expand);
- setOperationAction(ISD::SELECT, MVT::f32, Expand);
- setOperationAction(ISD::SELECT, MVT::f64, Expand);
-
- setSetCCResultContents(ZeroOrOneSetCCResult);
- addLegalFPImmediate(+0.0); // Necessary for FSEL
- addLegalFPImmediate(-0.0); //
-
- computeRegisterProperties();
- }
-
- /// LowerArguments - This hook must be implemented to indicate how we should
- /// lower the arguments for the specified function, into the specified DAG.
- virtual std::vector<SDOperand>
- LowerArguments(Function &F, SelectionDAG &DAG);
-
- /// LowerCallTo - This hook lowers an abstract call to a function into an
- /// actual call.
- virtual std::pair<SDOperand, SDOperand>
- LowerCallTo(SDOperand Chain, const Type *RetTy, bool isVarArg, unsigned CC,
- bool isTailCall, SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG);
-
- virtual SDOperand LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG);
-
- virtual std::pair<SDOperand,SDOperand>
- LowerVAArg(SDOperand Chain, SDOperand VAListP, Value *VAListV,
- const Type *ArgTy, SelectionDAG &DAG);
-
- virtual std::pair<SDOperand, SDOperand>
- LowerFrameReturnAddress(bool isFrameAddr, SDOperand Chain, unsigned Depth,
- SelectionDAG &DAG);
- };
-}
-
-
-std::vector<SDOperand>
-PPC32TargetLowering::LowerArguments(Function &F, SelectionDAG &DAG) {
- //
- // add beautiful description of PPC stack frame format, or at least some docs
- //
- MachineFunction &MF = DAG.getMachineFunction();
- MachineFrameInfo *MFI = MF.getFrameInfo();
- MachineBasicBlock& BB = MF.front();
- std::vector<SDOperand> ArgValues;
-
- // Due to the rather complicated nature of the PowerPC ABI, rather than a
- // fixed size array of physical args, for the sake of simplicity let the STL
- // handle tracking them for us.
- std::vector<unsigned> argVR, argPR, argOp;
- unsigned ArgOffset = 24;
- unsigned GPR_remaining = 8;
- unsigned FPR_remaining = 13;
- unsigned GPR_idx = 0, FPR_idx = 0;
- static const unsigned GPR[] = {
- PPC::R3, PPC::R4, PPC::R5, PPC::R6,
- PPC::R7, PPC::R8, PPC::R9, PPC::R10,
- };
- static const unsigned FPR[] = {
- 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
- };
-
- // Add DAG nodes to load the arguments... On entry to a function on PPC,
- // the arguments start at offset 24, although they are likely to be passed
- // in registers.
- for (Function::arg_iterator I = F.arg_begin(), E = F.arg_end(); I != E; ++I) {
- SDOperand newroot, argt;
- unsigned ObjSize;
- bool needsLoad = false;
- bool ArgLive = !I->use_empty();
- MVT::ValueType ObjectVT = getValueType(I->getType());
-
- switch (ObjectVT) {
- default: assert(0 && "Unhandled argument type!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- ObjSize = 4;
- if (!ArgLive) break;
- if (GPR_remaining > 0) {
- MF.addLiveIn(GPR[GPR_idx]);
- argt = newroot = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32,
- DAG.getRoot());
- if (ObjectVT != MVT::i32)
- argt = DAG.getNode(ISD::TRUNCATE, ObjectVT, newroot);
- } else {
- needsLoad = true;
- }
- break;
- case MVT::i64: ObjSize = 8;
- if (!ArgLive) break;
- if (GPR_remaining > 0) {
- SDOperand argHi, argLo;
- MF.addLiveIn(GPR[GPR_idx]);
- argHi = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
- // If we have two or more remaining argument registers, then both halves
- // of the i64 can be sourced from there. Otherwise, the lower half will
- // have to come off the stack. This can happen when an i64 is preceded
- // by 28 bytes of arguments.
- if (GPR_remaining > 1) {
- MF.addLiveIn(GPR[GPR_idx+1]);
- argLo = DAG.getCopyFromReg(GPR[GPR_idx+1], MVT::i32, argHi);
- } else {
- int FI = MFI->CreateFixedObject(4, ArgOffset+4);
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
- argLo = DAG.getLoad(MVT::i32, DAG.getEntryNode(), FIN,
- DAG.getSrcValue(NULL));
- }
- // Build the outgoing arg thingy
- argt = DAG.getNode(ISD::BUILD_PAIR, MVT::i64, argLo, argHi);
- newroot = argLo;
- } else {
- needsLoad = true;
- }
- break;
- case MVT::f32:
- case MVT::f64:
- ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
- if (!ArgLive) break;
- if (FPR_remaining > 0) {
- MF.addLiveIn(FPR[FPR_idx]);
- argt = newroot = DAG.getCopyFromReg(FPR[FPR_idx], ObjectVT,
- DAG.getRoot());
- --FPR_remaining;
- ++FPR_idx;
- } else {
- needsLoad = true;
- }
- break;
- }
-
- // We need to load the argument to a virtual register if we determined above
- // that we ran out of physical registers of the appropriate type
- if (needsLoad) {
- unsigned SubregOffset = 0;
- if (ObjectVT == MVT::i8 || ObjectVT == MVT::i1) SubregOffset = 3;
- if (ObjectVT == MVT::i16) SubregOffset = 2;
- int FI = MFI->CreateFixedObject(ObjSize, ArgOffset);
- SDOperand FIN = DAG.getFrameIndex(FI, MVT::i32);
- FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN,
- DAG.getConstant(SubregOffset, MVT::i32));
- argt = newroot = DAG.getLoad(ObjectVT, DAG.getEntryNode(), FIN,
- DAG.getSrcValue(NULL));
- }
-
- // Every 4 bytes of argument space consumes one of the GPRs available for
- // argument passing.
- if (GPR_remaining > 0) {
- unsigned delta = (GPR_remaining > 1 && ObjSize == 8) ? 2 : 1;
- GPR_remaining -= delta;
- GPR_idx += delta;
- }
- ArgOffset += ObjSize;
- if (newroot.Val)
- DAG.setRoot(newroot.getValue(1));
-
- ArgValues.push_back(argt);
- }
-
- // If the function takes variable number of arguments, make a frame index for
- // the start of the first vararg value... for expansion of llvm.va_start.
- if (F.isVarArg()) {
- VarArgsFrameIndex = MFI->CreateFixedObject(4, ArgOffset);
- SDOperand FIN = DAG.getFrameIndex(VarArgsFrameIndex, MVT::i32);
- // 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;
- for (; GPR_remaining > 0; --GPR_remaining, ++GPR_idx) {
- MF.addLiveIn(GPR[GPR_idx]);
- SDOperand Val = DAG.getCopyFromReg(GPR[GPR_idx], MVT::i32, DAG.getRoot());
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Val.getValue(1),
- Val, FIN, DAG.getSrcValue(NULL));
- MemOps.push_back(Store);
- // Increment the address by four for the next argument to store
- SDOperand PtrOff = DAG.getConstant(4, getPointerTy());
- FIN = DAG.getNode(ISD::ADD, MVT::i32, FIN, PtrOff);
- }
- DAG.setRoot(DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps));
- }
-
- // Finally, inform the code generator which regs we return values in.
- switch (getValueType(F.getReturnType())) {
- default: assert(0 && "Unknown type!");
- case MVT::isVoid: break;
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- MF.addLiveOut(PPC::R3);
- break;
- case MVT::i64:
- MF.addLiveOut(PPC::R3);
- MF.addLiveOut(PPC::R4);
- break;
- case MVT::f32:
- case MVT::f64:
- MF.addLiveOut(PPC::F1);
- break;
- }
-
- return ArgValues;
-}
-
-std::pair<SDOperand, SDOperand>
-PPC32TargetLowering::LowerCallTo(SDOperand Chain,
- const Type *RetTy, bool isVarArg,
- unsigned CallingConv, bool isTailCall,
- SDOperand Callee, ArgListTy &Args,
- SelectionDAG &DAG) {
- // args_to_use will accumulate outgoing args for the ISD::CALL case in
- // SelectExpr to use to put the arguments in the appropriate registers.
- std::vector<SDOperand> args_to_use;
-
- // Count how many bytes are to be pushed on the stack, including the linkage
- // area, and parameter passing area.
- unsigned NumBytes = 24;
-
- if (Args.empty()) {
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
- } else {
- for (unsigned i = 0, e = Args.size(); i != e; ++i)
- switch (getValueType(Args[i].second)) {
- default: assert(0 && "Unknown value type!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- case MVT::i32:
- case MVT::f32:
- NumBytes += 4;
- break;
- case MVT::i64:
- case MVT::f64:
- NumBytes += 8;
- break;
- }
-
- // Just to be safe, we'll always reserve the full 24 bytes of linkage area
- // plus 32 bytes of argument space in case any called code gets funky on us.
- // (Required by ABI to support var arg)
- if (NumBytes < 56) NumBytes = 56;
-
- // Adjust the stack pointer for the new arguments...
- // These operations are automatically eliminated by the prolog/epilog pass
- Chain = DAG.getNode(ISD::CALLSEQ_START, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
-
- // Set up a copy of the stack pointer for use loading and storing any
- // arguments that may not fit in the registers available for argument
- // passing.
- SDOperand StackPtr = DAG.getCopyFromReg(PPC::R1, MVT::i32,
- DAG.getEntryNode());
-
- // Figure out which arguments are going to go in registers, and which in
- // 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 = 24;
- unsigned GPR_remaining = 8;
- unsigned FPR_remaining = 13;
-
- std::vector<SDOperand> MemOps;
- for (unsigned i = 0, e = Args.size(); i != e; ++i) {
- // PtrOff will be used to store the current argument to the stack if a
- // register cannot be found for it.
- SDOperand PtrOff = DAG.getConstant(ArgOffset, getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, MVT::i32, StackPtr, PtrOff);
- MVT::ValueType ArgVT = getValueType(Args[i].second);
-
- switch (ArgVT) {
- default: assert(0 && "Unexpected ValueType for argument!");
- case MVT::i1:
- case MVT::i8:
- case MVT::i16:
- // Promote the integer to 32 bits. If the input type is signed use a
- // sign extend, otherwise use a zero extend.
- if (Args[i].second->isSigned())
- Args[i].first =DAG.getNode(ISD::SIGN_EXTEND, MVT::i32, Args[i].first);
- else
- Args[i].first =DAG.getNode(ISD::ZERO_EXTEND, MVT::i32, Args[i].first);
- // FALL THROUGH
- case MVT::i32:
- if (GPR_remaining > 0) {
- args_to_use.push_back(Args[i].first);
- --GPR_remaining;
- } else {
- MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Args[i].first, PtrOff,
- DAG.getSrcValue(NULL)));
- }
- ArgOffset += 4;
- break;
- case MVT::i64:
- // If we have one free GPR left, we can place the upper half of the i64
- // in it, and store the other half to the stack. If we have two or more
- // free GPRs, then we can pass both halves of the i64 in registers.
- if (GPR_remaining > 0) {
- SDOperand Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
- Args[i].first, DAG.getConstant(1, MVT::i32));
- SDOperand Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, MVT::i32,
- Args[i].first, DAG.getConstant(0, MVT::i32));
- args_to_use.push_back(Hi);
- --GPR_remaining;
- if (GPR_remaining > 0) {
- args_to_use.push_back(Lo);
- --GPR_remaining;
- } else {
- SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
- MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Lo, PtrOff, DAG.getSrcValue(NULL)));
- }
- } else {
- MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Args[i].first, PtrOff,
- DAG.getSrcValue(NULL)));
- }
- ArgOffset += 8;
- break;
- case MVT::f32:
- case MVT::f64:
- if (FPR_remaining > 0) {
- args_to_use.push_back(Args[i].first);
- --FPR_remaining;
- if (isVarArg) {
- SDOperand Store = DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Args[i].first, PtrOff,
- DAG.getSrcValue(NULL));
- MemOps.push_back(Store);
- // Float varargs are always shadowed in available integer registers
- if (GPR_remaining > 0) {
- SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
- DAG.getSrcValue(NULL));
- MemOps.push_back(Load);
- args_to_use.push_back(Load);
- --GPR_remaining;
- }
- if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
- SDOperand ConstFour = DAG.getConstant(4, getPointerTy());
- PtrOff = DAG.getNode(ISD::ADD, MVT::i32, PtrOff, ConstFour);
- SDOperand Load = DAG.getLoad(MVT::i32, Store, PtrOff,
- DAG.getSrcValue(NULL));
- MemOps.push_back(Load);
- args_to_use.push_back(Load);
- --GPR_remaining;
- }
- } else {
- // If we have any FPRs remaining, we may also have GPRs remaining.
- // Args passed in FPRs consume either 1 (f32) or 2 (f64) available
- // GPRs.
- if (GPR_remaining > 0) {
- args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
- --GPR_remaining;
- }
- if (GPR_remaining > 0 && MVT::f64 == ArgVT) {
- args_to_use.push_back(DAG.getNode(ISD::UNDEF, MVT::i32));
- --GPR_remaining;
- }
- }
- } else {
- MemOps.push_back(DAG.getNode(ISD::STORE, MVT::Other, Chain,
- Args[i].first, PtrOff,
- DAG.getSrcValue(NULL)));
- }
- ArgOffset += (ArgVT == MVT::f32) ? 4 : 8;
- break;
- }
- }
- if (!MemOps.empty())
- Chain = DAG.getNode(ISD::TokenFactor, MVT::Other, MemOps);
- }
-
- std::vector<MVT::ValueType> RetVals;
- MVT::ValueType RetTyVT = getValueType(RetTy);
- if (RetTyVT != MVT::isVoid)
- RetVals.push_back(RetTyVT);
- RetVals.push_back(MVT::Other);
-
- SDOperand TheCall = SDOperand(DAG.getCall(RetVals,
- Chain, Callee, args_to_use), 0);
- Chain = TheCall.getValue(RetTyVT != MVT::isVoid);
- Chain = DAG.getNode(ISD::CALLSEQ_END, MVT::Other, Chain,
- DAG.getConstant(NumBytes, getPointerTy()));
- return std::make_pair(TheCall, Chain);
-}
-
-SDOperand PPC32TargetLowering::LowerVAStart(SDOperand Chain, SDOperand VAListP,
- Value *VAListV, SelectionDAG &DAG) {
- // 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, Chain, FR, VAListP,
- DAG.getSrcValue(VAListV));
-}
-
-std::pair<SDOperand,SDOperand>
-PPC32TargetLowering::LowerVAArg(SDOperand Chain,
- SDOperand VAListP, Value *VAListV,
- const Type *ArgTy, SelectionDAG &DAG) {
- MVT::ValueType ArgVT = getValueType(ArgTy);
-
- SDOperand VAList =
- DAG.getLoad(MVT::i32, Chain, VAListP, DAG.getSrcValue(VAListV));
- SDOperand Result = DAG.getLoad(ArgVT, Chain, VAList, DAG.getSrcValue(NULL));
- unsigned Amt;
- if (ArgVT == MVT::i32 || ArgVT == MVT::f32)
- Amt = 4;
- else {
- assert((ArgVT == MVT::i64 || ArgVT == MVT::f64) &&
- "Other types should have been promoted for varargs!");
- Amt = 8;
- }
- VAList = DAG.getNode(ISD::ADD, VAList.getValueType(), VAList,
- DAG.getConstant(Amt, VAList.getValueType()));
- Chain = DAG.getNode(ISD::STORE, MVT::Other, Chain,
- VAList, VAListP, DAG.getSrcValue(VAListV));
- return std::make_pair(Result, Chain);
-}
-
-
-std::pair<SDOperand, SDOperand> PPC32TargetLowering::
-LowerFrameReturnAddress(bool isFrameAddress, SDOperand Chain, unsigned Depth,
- SelectionDAG &DAG) {
- assert(0 && "LowerFrameReturnAddress unimplemented");
- abort();
-}
-
namespace {
Statistic<>Recorded("ppc-codegen", "Number of recording ops emitted");
Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
}
// if the mask doesn't intersect any Indeterminant bits
- if (!(Mask & Indeterminant)) {
+ if (Mask && !(Mask & Indeterminant)) {
SH = Shift;
// make sure the mask is still a mask (wrap arounds may not be)
return isRunOfOnes(Mask, MB, ME);
void ISel::SelectBranchCC(SDOperand N)
{
MachineBasicBlock *Dest =
- cast<BasicBlockSDNode>(N.getOperand(2))->getBasicBlock();
+ cast<BasicBlockSDNode>(N.getOperand(4))->getBasicBlock();
Select(N.getOperand(0)); //chain
-
- // FIXME: Until we have Branch_CC and Branch_Twoway_CC, we're going to have to
- // Fake it up by hand by checking to see if op 1 is a SetCC, or a boolean.
- unsigned CCReg;
- ISD::CondCode CC;
- SDOperand Cond = N.getOperand(1);
- if (Cond.getOpcode() == ISD::SETCC) {
- CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
- CCReg = SelectCC(Cond.getOperand(0), Cond.getOperand(1), CC);
- } else {
- CC = ISD::SETNE;
- CCReg = SelectCC(Cond, ISelDAG->getConstant(0, Cond.getValueType()), CC);
- }
+ ISD::CondCode CC = cast<CondCodeSDNode>(N.getOperand(1))->get();
+ unsigned CCReg = SelectCC(N.getOperand(2), N.getOperand(3), CC);
unsigned Opc = getBCCForSetCC(CC);
// Iterate to the next basic block
// and build a PowerPC branch pseudo-op, suitable for long branch conversion
// if necessary by the branch selection pass. Otherwise, emit a standard
// conditional branch.
- if (N.getOpcode() == ISD::BRCONDTWOWAY) {
+ if (N.getOpcode() == ISD::BRTWOWAY_CC) {
MachineBasicBlock *Fallthrough =
- cast<BasicBlockSDNode>(N.getOperand(3))->getBasicBlock();
+ cast<BasicBlockSDNode>(N.getOperand(5))->getBasicBlock();
if (Dest != It) {
BuildMI(BB, PPC::COND_BRANCH, 4).addReg(CCReg).addImm(Opc)
.addMBB(Dest).addMBB(Fallthrough);
SDNode *Node = N.Val;
MVT::ValueType DestType = N.getValueType();
- if (Node->getOpcode() == ISD::CopyFromReg &&
- (MRegisterInfo::isVirtualRegister(cast<RegSDNode>(Node)->getReg()) ||
- cast<RegSDNode>(Node)->getReg() == PPC::R1))
+ if (Node->getOpcode() == ISD::CopyFromReg) {
+ unsigned Reg = cast<RegisterSDNode>(Node->getOperand(1))->getReg();
// Just use the specified register as our input.
- return cast<RegSDNode>(Node)->getReg();
+ if (MRegisterInfo::isVirtualRegister(Reg) || Reg == PPC::R1)
+ return Reg;
+ }
unsigned &Reg = ExprMap[N];
if (Reg) return Reg;
switch (opcode) {
default:
- Node->dump();
- assert(0 && "\nNode not handled!\n");
+ Node->dump(); std::cerr << '\n';
+ assert(0 && "Node not handled!\n");
case ISD::UNDEF:
BuildMI(BB, PPC::IMPLICIT_DEF, 0, Result);
return Result;
DestType = N.getValue(0).getValueType();
if (Result == 1)
Result = ExprMap[N.getValue(0)] = MakeReg(DestType);
- Tmp1 = dyn_cast<RegSDNode>(Node)->getReg();
+ Tmp1 = dyn_cast<RegisterSDNode>(Node->getOperand(1))->getReg();
if (MVT::isInteger(DestType))
BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
else
case ISD::SUB_PARTS: {
assert(N.getNumOperands() == 4 && N.getValueType() == MVT::i32 &&
"Not an i64 add/sub!");
- // Emit all of the operands.
- std::vector<unsigned> InVals;
- for (unsigned i = 0, e = N.getNumOperands(); i != e; ++i)
- InVals.push_back(SelectExpr(N.getOperand(i)));
+ unsigned Tmp4 = 0;
+ bool ME = isIntImmediate(N.getOperand(3),Tmp3) && ((signed)Tmp3 == -1);
+ bool ZE = isIntImmediate(N.getOperand(3),Tmp3) && (Tmp3 == 0);
+ bool IM = isIntImmediate(N.getOperand(2),Tmp3) && ((signed)Tmp3 >= -32768 ||
+ (signed)Tmp3 < 32768);
+ Tmp1 = SelectExpr(N.getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(1));
+ if (!IM || N.getOpcode() == ISD::SUB_PARTS)
+ Tmp3 = SelectExpr(N.getOperand(2));
+ if ((!ME && !ZE) || N.getOpcode() == ISD::SUB_PARTS)
+ Tmp4 = SelectExpr(N.getOperand(3));
+
if (N.getOpcode() == ISD::ADD_PARTS) {
- BuildMI(BB, PPC::ADDC, 2, Result).addReg(InVals[0]).addReg(InVals[2]);
- BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(InVals[1]).addReg(InVals[3]);
+ // Codegen the low 32 bits of the add. Interestingly, there is no shifted
+ // form of add immediate carrying.
+ if (IM)
+ BuildMI(BB, PPC::ADDIC, 2, Result).addReg(Tmp1).addSImm(Tmp3);
+ else
+ BuildMI(BB, PPC::ADDC, 2, Result).addReg(Tmp1).addReg(Tmp3);
+ // Codegen the high 32 bits, adding zero, minus one, or the full value
+ // along with the carry flag produced by addc/addic to tmp2.
+ if (ZE)
+ BuildMI(BB, PPC::ADDZE, 1, Result+1).addReg(Tmp2);
+ else if (ME)
+ BuildMI(BB, PPC::ADDME, 1, Result+1).addReg(Tmp2);
+ else
+ BuildMI(BB, PPC::ADDE, 2, Result+1).addReg(Tmp2).addReg(Tmp4);
} else {
- BuildMI(BB, PPC::SUBFC, 2, Result).addReg(InVals[2]).addReg(InVals[0]);
- BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(InVals[3]).addReg(InVals[1]);
+ BuildMI(BB, PPC::SUBFC, 2, Result).addReg(Tmp3).addReg(Tmp1);
+ BuildMI(BB, PPC::SUBFE, 2, Result+1).addReg(Tmp4).addReg(Tmp2);
}
return Result+N.ResNo;
}
return Result+N.ResNo;
}
- case ISD::FP_TO_UINT:
case ISD::FP_TO_SINT: {
- bool U = (ISD::FP_TO_UINT == opcode);
Tmp1 = SelectExpr(N.getOperand(0));
- if (!U) {
- Tmp2 = MakeFPReg();
- BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
- int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
- addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
- addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
- return Result;
- } else {
- unsigned Zero = getConstDouble(0.0);
- unsigned MaxInt = getConstDouble((1LL << 32) - 1);
- unsigned Border = getConstDouble(1LL << 31);
- unsigned UseZero = MakeFPReg();
- unsigned UseMaxInt = MakeFPReg();
- unsigned UseChoice = MakeFPReg();
- unsigned TmpReg = MakeFPReg();
- unsigned TmpReg2 = MakeFPReg();
- unsigned ConvReg = MakeFPReg();
- unsigned IntTmp = MakeIntReg();
- unsigned XorReg = MakeIntReg();
- MachineFunction *F = BB->getParent();
- int FrameIdx = F->getFrameInfo()->CreateStackObject(8, 8);
- // Update machine-CFG edges
- MachineBasicBlock *XorMBB = new MachineBasicBlock(BB->getBasicBlock());
- MachineBasicBlock *PhiMBB = new MachineBasicBlock(BB->getBasicBlock());
- MachineBasicBlock *OldMBB = BB;
- ilist<MachineBasicBlock>::iterator It = BB; ++It;
- F->getBasicBlockList().insert(It, XorMBB);
- F->getBasicBlockList().insert(It, PhiMBB);
- BB->addSuccessor(XorMBB);
- BB->addSuccessor(PhiMBB);
- // Convert from floating point to unsigned 32-bit value
- // Use 0 if incoming value is < 0.0
- BuildMI(BB, PPC::FSEL, 3, UseZero).addReg(Tmp1).addReg(Tmp1).addReg(Zero);
- // Use 2**32 - 1 if incoming value is >= 2**32
- BuildMI(BB, PPC::FSUB, 2, UseMaxInt).addReg(MaxInt).addReg(Tmp1);
- BuildMI(BB, PPC::FSEL, 3, UseChoice).addReg(UseMaxInt).addReg(UseZero)
- .addReg(MaxInt);
- // Subtract 2**31
- BuildMI(BB, PPC::FSUB, 2, TmpReg).addReg(UseChoice).addReg(Border);
- // Use difference if >= 2**31
- BuildMI(BB, PPC::FCMPU, 2, PPC::CR0).addReg(UseChoice).addReg(Border);
- BuildMI(BB, PPC::FSEL, 3, TmpReg2).addReg(TmpReg).addReg(TmpReg)
- .addReg(UseChoice);
- // Convert to integer
- BuildMI(BB, PPC::FCTIWZ, 1, ConvReg).addReg(TmpReg2);
- addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(ConvReg), FrameIdx);
- addFrameReference(BuildMI(BB, PPC::LWZ, 2, IntTmp), FrameIdx, 4);
- BuildMI(BB, PPC::BLT, 2).addReg(PPC::CR0).addMBB(PhiMBB);
- BuildMI(BB, PPC::B, 1).addMBB(XorMBB);
-
- // XorMBB:
- // add 2**31 if input was >= 2**31
- BB = XorMBB;
- BuildMI(BB, PPC::XORIS, 2, XorReg).addReg(IntTmp).addImm(0x8000);
- XorMBB->addSuccessor(PhiMBB);
-
- // PhiMBB:
- // DestReg = phi [ IntTmp, OldMBB ], [ XorReg, XorMBB ]
- BB = PhiMBB;
- BuildMI(BB, PPC::PHI, 4, Result).addReg(IntTmp).addMBB(OldMBB)
- .addReg(XorReg).addMBB(XorMBB);
- return Result;
- }
- assert(0 && "Should never get here");
- return 0;
+ Tmp2 = MakeFPReg();
+ BuildMI(BB, PPC::FCTIWZ, 1, Tmp2).addReg(Tmp1);
+ int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
+ addFrameReference(BuildMI(BB, PPC::STFD, 3).addReg(Tmp2), FrameIdx);
+ addFrameReference(BuildMI(BB, PPC::LWZ, 2, Result), FrameIdx, 4);
+ return Result;
}
case ISD::SETCC: {
assert(0 && "Should never get here");
}
- unsigned TrueValue = SelectExpr(N.getOperand(2)); //Use if TRUE
- unsigned FalseValue = SelectExpr(N.getOperand(3)); //Use if FALSE
+ // If the False value only has one use, we can generate better code by
+ // selecting it in the fallthrough basic block rather than here, which
+ // increases register pressure.
+ unsigned TrueValue = SelectExpr(N.getOperand(2));
+ unsigned FalseValue = SelectExpr(N.getOperand(3));
unsigned CCReg = SelectCC(N.getOperand(0), N.getOperand(1), CC);
Opc = getBCCForSetCC(CC);
case ISD::Constant:
switch (N.getValueType()) {
default: assert(0 && "Cannot use constants of this type!");
- case MVT::i1:
- BuildMI(BB, PPC::LI, 1, Result)
- .addSImm(!cast<ConstantSDNode>(N)->isNullValue());
- break;
case MVT::i32:
{
int v = (int)cast<ConstantSDNode>(N)->getSignExtended();
Tmp1 = SelectExpr(N.getOperand(0));
BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
return Result;
-
- case ISD::UINT_TO_FP:
- case ISD::SINT_TO_FP: {
- assert (N.getOperand(0).getValueType() == MVT::i32
- && "int to float must operate on i32");
- bool IsUnsigned = (ISD::UINT_TO_FP == opcode);
- Tmp1 = SelectExpr(N.getOperand(0)); // Get the operand register
- Tmp2 = MakeFPReg(); // temp reg to load the integer value into
- Tmp3 = MakeIntReg(); // temp reg to hold the conversion constant
-
- int FrameIdx = BB->getParent()->getFrameInfo()->CreateStackObject(8, 8);
- MachineConstantPool *CP = BB->getParent()->getConstantPool();
-
- if (IsUnsigned) {
- unsigned ConstF = getConstDouble(0x1.000000p52);
- // Store the hi & low halves of the fp value, currently in int regs
- BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
- addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
- addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp1), FrameIdx, 4);
- addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
- // Generate the return value with a subtract
- BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
- } else {
- unsigned ConstF = getConstDouble(0x1.000008p52);
- unsigned TmpL = MakeIntReg();
- // Store the hi & low halves of the fp value, currently in int regs
- BuildMI(BB, PPC::LIS, 1, Tmp3).addSImm(0x4330);
- addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(Tmp3), FrameIdx);
- BuildMI(BB, PPC::XORIS, 2, TmpL).addReg(Tmp1).addImm(0x8000);
- addFrameReference(BuildMI(BB, PPC::STW, 3).addReg(TmpL), FrameIdx, 4);
- addFrameReference(BuildMI(BB, PPC::LFD, 2, Tmp2), FrameIdx);
- // Generate the return value with a subtract
- BuildMI(BB, PPC::FSUB, 2, Result).addReg(Tmp2).addReg(ConstF);
- }
- return Result;
- }
}
return 0;
}
BuildMI(BB, PPC::B, 1).addMBB(Dest);
return;
}
- case ISD::BRCOND:
- case ISD::BRCONDTWOWAY:
+ case ISD::BR_CC:
+ case ISD::BRTWOWAY_CC:
SelectBranchCC(N);
return;
case ISD::CopyToReg:
Select(N.getOperand(0));
- Tmp1 = SelectExpr(N.getOperand(1));
- Tmp2 = cast<RegSDNode>(N)->getReg();
+ Tmp1 = SelectExpr(N.getOperand(2));
+ Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
if (Tmp1 != Tmp2) {
- if (N.getOperand(1).getValueType() == MVT::f64 ||
- N.getOperand(1).getValueType() == MVT::f32)
+ if (N.getOperand(2).getValueType() == MVT::f64 ||
+ N.getOperand(2).getValueType() == MVT::f32)
BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
else
BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
return;
case ISD::ImplicitDef:
Select(N.getOperand(0));
- BuildMI(BB, PPC::IMPLICIT_DEF, 0, cast<RegSDNode>(N)->getReg());
+ BuildMI(BB, PPC::IMPLICIT_DEF, 0,
+ cast<RegisterSDNode>(N.getOperand(1))->getReg());
return;
case ISD::RET:
switch (N.getNumOperands()) {