#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 "llvm/ADT/Statistic.h"
// 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);
}
// 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())
+ 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) {
+ BuildMI(&BB, PPC::IMPLICIT_DEF, 0, 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);
+ 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));
+ }
return ArgValues;
}
namespace {
Statistic<>NotLogic("ppc-codegen", "Number of inverted logical ops");
+Statistic<>FusedFP("ppc-codegen", "Number of fused fp operations");
//===--------------------------------------------------------------------===//
/// ISel - PPC32 specific code to select PPC32 machine instructions for
/// SelectionDAG operations.
void SelectBranchCC(SDOperand N);
};
+/// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
+/// returns zero when the input is not exactly a power of two.
+static unsigned ExactLog2(unsigned Val) {
+ if (Val == 0 || (Val & (Val-1))) return 0;
+ unsigned Count = 0;
+ while (Val != 1) {
+ Val >>= 1;
+ ++Count;
+ }
+ return Count;
+}
+
/// canUseAsImmediateForOpcode - This method returns a value indicating whether
/// the ConstantSDNode N can be used as an immediate to Opcode. The return
/// values are either 0, 1 or 2. 0 indicates that either N is not a
if (U && (v >= 0 && v <= 65535)) { Imm = v & 0xFFFF; return 1; }
if (!U && (v <= 32767 && v >= -32768)) { Imm = v & 0xFFFF; return 1; }
break;
+ case ISD::SDIV:
+ if (0 != ExactLog2(v)) { Imm = ExactLog2(v); return 1; }
+ break;
}
return 0;
}
}
case ISD::FNEG:
- if (ISD::FABS == N.getOperand(0).getOpcode()) {
+ if (!NoExcessFPPrecision &&
+ ISD::ADD == N.getOperand(0).getOpcode() &&
+ N.getOperand(0).Val->hasOneUse() &&
+ ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
+ N.getOperand(0).getOperand(0).Val->hasOneUse()) {
+ ++FusedFP; // Statistic
+ Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
+ Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
+ Opc = DestType == MVT::f64 ? PPC::FNMADD : PPC::FNMADDS;
+ BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ } else if (!NoExcessFPPrecision &&
+ ISD::SUB == N.getOperand(0).getOpcode() &&
+ N.getOperand(0).Val->hasOneUse() &&
+ ISD::MUL == N.getOperand(0).getOperand(0).getOpcode() &&
+ N.getOperand(0).getOperand(0).Val->hasOneUse()) {
+ ++FusedFP; // Statistic
+ Tmp1 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(0).getOperand(1));
+ Tmp3 = SelectExpr(N.getOperand(0).getOperand(1));
+ Opc = DestType == MVT::f64 ? PPC::FNMSUB : PPC::FNMSUBS;
+ BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ } else if (ISD::FABS == N.getOperand(0).getOpcode()) {
Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
BuildMI(BB, PPC::FNABS, 1, Result).addReg(Tmp1);
} else {
return Result;
}
- case ISD::MUL:
case ISD::ADD:
+ if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
+ N.getOperand(0).Val->hasOneUse()) {
+ ++FusedFP; // Statistic
+ Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
+ Tmp3 = SelectExpr(N.getOperand(1));
+ Opc = DestType == MVT::f64 ? PPC::FMADD : PPC::FMADDS;
+ BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ return Result;
+ }
+ Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS;
+ Tmp1 = SelectExpr(N.getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(1));
+ BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
+ return Result;
+
case ISD::SUB:
+ if (!NoExcessFPPrecision && N.getOperand(0).getOpcode() == ISD::MUL &&
+ N.getOperand(0).Val->hasOneUse()) {
+ ++FusedFP; // Statistic
+ Tmp1 = SelectExpr(N.getOperand(0).getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(0).getOperand(1));
+ Tmp3 = SelectExpr(N.getOperand(1));
+ Opc = DestType == MVT::f64 ? PPC::FMSUB : PPC::FMSUBS;
+ BuildMI(BB, Opc, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ return Result;
+ }
+ Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS;
+ Tmp1 = SelectExpr(N.getOperand(0));
+ Tmp2 = SelectExpr(N.getOperand(1));
+ BuildMI(BB, Opc, 2, Result).addReg(Tmp1).addReg(Tmp2);
+ return Result;
+
+ case ISD::MUL:
case ISD::SDIV:
switch( opcode ) {
case ISD::MUL: Opc = DestType == MVT::f64 ? PPC::FMUL : PPC::FMULS; break;
- case ISD::ADD: Opc = DestType == MVT::f64 ? PPC::FADD : PPC::FADDS; break;
- case ISD::SUB: Opc = DestType == MVT::f64 ? PPC::FSUB : PPC::FSUBS; break;
case ISD::SDIV: Opc = DestType == MVT::f64 ? PPC::FDIV : PPC::FDIVS; break;
};
Tmp1 = SelectExpr(N.getOperand(0));
break;
}
+ if (ISD::CopyFromReg == opcode)
+ DestType = N.getValue(0).getValueType();
+
if (DestType == MVT::f64 || DestType == MVT::f32)
if (ISD::LOAD != opcode && ISD::EXTLOAD != opcode && ISD::UNDEF != opcode)
return SelectExprFP(N, Result);
Select(N.getOperand(0));
ExprMap[N.getValue(Node->getNumValues()-1)] = 1;
+ MachineInstr *CallMI;
+ // Emit the correct call instruction based on the type of symbol called.
+ if (GlobalAddressSDNode *GASD =
+ dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
+ CallMI = BuildMI(PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(),
+ true);
+ } else if (ExternalSymbolSDNode *ESSDN =
+ dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
+ CallMI = BuildMI(PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(),
+ true);
+ } else {
+ Tmp1 = SelectExpr(N.getOperand(1));
+ BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
+ BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
+ CallMI = BuildMI(PPC::CALLindirect, 3).addImm(20).addImm(0)
+ .addReg(PPC::R12);
+ }
+
// Load the register args to virtual regs
std::vector<unsigned> ArgVR;
for(int i = 2, e = Node->getNumOperands(); i < e; ++i)
case MVT::i16:
case MVT::i32:
assert(GPR_idx < 8 && "Too many int args");
- if (N.getOperand(i+2).getOpcode() != ISD::UNDEF)
+ if (N.getOperand(i+2).getOpcode() != ISD::UNDEF) {
BuildMI(BB, PPC::OR,2,GPR[GPR_idx]).addReg(ArgVR[i]).addReg(ArgVR[i]);
+ CallMI->addRegOperand(GPR[GPR_idx], MachineOperand::Use);
+ }
++GPR_idx;
break;
case MVT::f64:
case MVT::f32:
assert(FPR_idx < 13 && "Too many fp args");
BuildMI(BB, PPC::FMR, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
+ CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
++FPR_idx;
break;
}
}
-
- // Emit the correct call instruction based on the type of symbol called.
- if (GlobalAddressSDNode *GASD =
- dyn_cast<GlobalAddressSDNode>(N.getOperand(1))) {
- BuildMI(BB, PPC::CALLpcrel, 1).addGlobalAddress(GASD->getGlobal(), true);
- } else if (ExternalSymbolSDNode *ESSDN =
- dyn_cast<ExternalSymbolSDNode>(N.getOperand(1))) {
- BuildMI(BB, PPC::CALLpcrel, 1).addExternalSymbol(ESSDN->getSymbol(), true);
- } else {
- Tmp1 = SelectExpr(N.getOperand(1));
- BuildMI(BB, PPC::OR, 2, PPC::R12).addReg(Tmp1).addReg(Tmp1);
- BuildMI(BB, PPC::MTCTR, 1).addReg(PPC::R12);
- BuildMI(BB, PPC::CALLindirect, 3).addImm(20).addImm(0).addReg(PPC::R12);
- }
+
+ // Put the call instruction in the correct place in the MachineBasicBlock
+ BB->push_back(CallMI);
switch (Node->getValueType(0)) {
default: assert(0 && "Unknown value type for call result!");
case MVT::i8:
case MVT::i16:
case MVT::i32:
- BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
- if (Node->getValueType(1) == MVT::i32)
- BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R4).addReg(PPC::R4);
+ if (Node->getValueType(1) == MVT::i32) {
+ BuildMI(BB, PPC::OR, 2, Result+1).addReg(PPC::R3).addReg(PPC::R3);
+ BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R4).addReg(PPC::R4);
+ } else {
+ BuildMI(BB, PPC::OR, 2, Result).addReg(PPC::R3).addReg(PPC::R3);
+ }
break;
case MVT::f32:
case MVT::f64:
case ISD::SDIV:
case ISD::UDIV:
+ if (1 == canUseAsImmediateForOpcode(N.getOperand(1), opcode, Tmp3)) {
+ Tmp1 = MakeReg(MVT::i32);
+ Tmp2 = SelectExpr(N.getOperand(0));
+ BuildMI(BB, PPC::SRAWI, 2, Tmp1).addReg(Tmp2).addImm(Tmp3);
+ BuildMI(BB, PPC::ADDZE, 1, Result).addReg(Tmp1);
+ return Result;
+ }
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
Opc = (ISD::UDIV == opcode) ? PPC::DIVWU : PPC::DIVW;
BuildMI(BB, PPC::SRW, 2, Tmp3).addReg(ShiftOpLo).addReg(Tmp1);
BuildMI(BB, PPC::OR, 2, Tmp4).addReg(Tmp2).addReg(Tmp3);
BuildMI(BB, PPC::ADDI, 2, Tmp5).addReg(SHReg).addSImm(-32);
- BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpHi).addReg(Tmp5);
+ BuildMI(BB, PPC::SLW, 2, Tmp6).addReg(ShiftOpLo).addReg(Tmp5);
BuildMI(BB, PPC::OR, 2, Result+1).addReg(Tmp4).addReg(Tmp6);
BuildMI(BB, PPC::SLW, 2, Result).addReg(ShiftOpLo).addReg(SHReg);
} else if (ISD::SRL_PARTS == opcode) {