Modify the ppc backend to use two register classes for FP: F8RC and F4RC.
These are used to represent float and double values, and the two regclasses
contain the same physical registers.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@23577 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/PowerPC/PPC32RegisterInfo.td b/lib/Target/PowerPC/PPC32RegisterInfo.td
index 2b26ea1..ef2e623 100644
--- a/lib/Target/PowerPC/PPC32RegisterInfo.td
+++ b/lib/Target/PowerPC/PPC32RegisterInfo.td
@@ -39,8 +39,12 @@
}];
}
-def FPRC : RegisterClass<"PPC32", f64, 64, [F0, F1, F2, F3, F4, F5, F6, F7,
+def F8RC : RegisterClass<"PPC32", f64, 64, [F0, F1, F2, F3, F4, F5, F6, F7,
F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21,
F22, F23, F24, F25, F26, F27, F28, F29, F30, F31]>;
+def F4RC : RegisterClass<"PPC32", f32, 32, [F0, F1, F2, F3, F4, F5, F6, F7,
+ F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21,
+ F22, F23, F24, F25, F26, F27, F28, F29, F30, F31]>;
+
def CRRC : RegisterClass<"PPC32", i32, 32, [CR0, CR1, CR5, CR6, CR7, CR2, CR3, CR4]>;
diff --git a/lib/Target/PowerPC/PPCBranchSelector.cpp b/lib/Target/PowerPC/PPCBranchSelector.cpp
index 24187e3..6e0e36f 100644
--- a/lib/Target/PowerPC/PPCBranchSelector.cpp
+++ b/lib/Target/PowerPC/PPCBranchSelector.cpp
@@ -41,7 +41,8 @@
// keeping the offsets up to date later when we emit long branch glue.
return 12;
case PPC::IMPLICIT_DEF_GPR: // no asm emitted
- case PPC::IMPLICIT_DEF_FP: // no asm emitted
+ case PPC::IMPLICIT_DEF_F4: // no asm emitted
+ case PPC::IMPLICIT_DEF_F8: // no asm emitted
return 0;
default:
break;
diff --git a/lib/Target/PowerPC/PPCCodeEmitter.cpp b/lib/Target/PowerPC/PPCCodeEmitter.cpp
index 25cbda5..af9c4cf 100644
--- a/lib/Target/PowerPC/PPCCodeEmitter.cpp
+++ b/lib/Target/PowerPC/PPCCodeEmitter.cpp
@@ -125,7 +125,8 @@
emitWord(getBinaryCodeForInstr(*I));
break;
case PPC::IMPLICIT_DEF_GPR:
- case PPC::IMPLICIT_DEF_FP:
+ case PPC::IMPLICIT_DEF_F8:
+ case PPC::IMPLICIT_DEF_F4:
break; // pseudo opcode, no side effects
case PPC::MovePCtoLR:
assert(0 && "CodeEmitter does not support MovePCtoLR instruction");
diff --git a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
index c8850cd..ab6c14f 100644
--- a/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
+++ b/lib/Target/PowerPC/PPCISelDAGToDAG.cpp
@@ -443,8 +443,10 @@
LHS, getI32Imm(Lo16(Imm)));
return CurDAG->getTargetNode(U ? PPC::CMPLW : PPC::CMPW, MVT::i32,
LHS, Select(RHS));
+ } else if (LHS.getValueType() == MVT::f32) {
+ return CurDAG->getTargetNode(PPC::FCMPUS, MVT::i32, LHS, Select(RHS));
} else {
- return CurDAG->getTargetNode(PPC::FCMPU, MVT::i32, LHS, Select(RHS));
+ return CurDAG->getTargetNode(PPC::FCMPUD, MVT::i32, LHS, Select(RHS));
}
}
@@ -679,8 +681,10 @@
case ISD::UNDEF:
if (N->getValueType(0) == MVT::i32)
CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_GPR, MVT::i32);
- else
- CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_FP, N->getValueType(0));
+ else if (N->getValueType(0) == MVT::f32)
+ CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_F4, MVT::f32);
+ else
+ CurDAG->SelectNodeTo(N, PPC::IMPLICIT_DEF_F8, MVT::f64);
return SDOperand(N, 0);
case ISD::FrameIndex: {
int FI = cast<FrameIndexSDNode>(N)->getIndex();
@@ -749,10 +753,16 @@
return SDOperand(Result.Val, Op.ResNo);
}
case PPCISD::FSEL:
- CurDAG->SelectNodeTo(N, PPC::FSEL, N->getValueType(0),
- Select(N->getOperand(0)),
- Select(N->getOperand(1)),
- Select(N->getOperand(2)));
+ if (N->getValueType(0) == MVT::f32)
+ CurDAG->SelectNodeTo(N, PPC::FSELS, MVT::f32,
+ Select(N->getOperand(0)),
+ Select(N->getOperand(1)),
+ Select(N->getOperand(2)));
+ else
+ CurDAG->SelectNodeTo(N, PPC::FSELD, MVT::f64,
+ Select(N->getOperand(0)),
+ Select(N->getOperand(1)),
+ Select(N->getOperand(2)));
return SDOperand(N, 0);
case PPCISD::FCFID:
CurDAG->SelectNodeTo(N, PPC::FCFID, N->getValueType(0),
@@ -956,15 +966,17 @@
return SDOperand(N, 0);
}
case ISD::FABS:
- CurDAG->SelectNodeTo(N, PPC::FABS, N->getValueType(0),
- Select(N->getOperand(0)));
+ if (N->getValueType(0) == MVT::f32)
+ CurDAG->SelectNodeTo(N, PPC::FABSS, MVT::f32, Select(N->getOperand(0)));
+ else
+ CurDAG->SelectNodeTo(N, PPC::FABSD, MVT::f64, Select(N->getOperand(0)));
return SDOperand(N, 0);
case ISD::FP_EXTEND:
assert(MVT::f64 == N->getValueType(0) &&
MVT::f32 == N->getOperand(0).getValueType() && "Illegal FP_EXTEND");
// We need to emit an FMR to make sure that the result has the right value
// type.
- CurDAG->SelectNodeTo(N, PPC::FMR, MVT::f64, Select(N->getOperand(0)));
+ CurDAG->SelectNodeTo(N, PPC::FMRSD, MVT::f64, Select(N->getOperand(0)));
return SDOperand(N, 0);
case ISD::FP_ROUND:
assert(MVT::f32 == N->getValueType(0) &&
@@ -978,7 +990,8 @@
unsigned Opc;
switch (Val.isTargetOpcode() ? Val.getTargetOpcode() : 0) {
default: Opc = 0; break;
- case PPC::FABS: Opc = PPC::FNABS; break;
+ case PPC::FABSS: Opc = PPC::FNABSS; break;
+ case PPC::FABSD: Opc = PPC::FNABSD; break;
case PPC::FMADD: Opc = PPC::FNMADD; break;
case PPC::FMADDS: Opc = PPC::FNMADDS; break;
case PPC::FMSUB: Opc = PPC::FNMSUB; break;
@@ -988,7 +1001,7 @@
// inverted opcode and the original instruction's operands. Otherwise,
// fall through and generate a fneg instruction.
if (Opc) {
- if (PPC::FNABS == Opc)
+ if (Opc == PPC::FNABSS || Opc == PPC::FNABSD)
CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0));
else
CurDAG->SelectNodeTo(N, Opc, Ty, Val.getOperand(0),
@@ -996,7 +1009,10 @@
return SDOperand(N, 0);
}
}
- CurDAG->SelectNodeTo(N, PPC::FNEG, Ty, Val);
+ if (Ty == MVT::f32)
+ CurDAG->SelectNodeTo(N, PPC::FNEGS, MVT::f32, Val);
+ else
+ CurDAG->SelectNodeTo(N, PPC::FNEGS, MVT::f64, Val);
return SDOperand(N, 0);
}
case ISD::FSQRT: {
@@ -1090,9 +1106,26 @@
case MVT::f64: Opc = isIdx ? PPC::LFDX : PPC::LFD; break;
}
- CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
- Op1, Op2, Select(N->getOperand(0)));
- return SDOperand(N, Op.ResNo);
+ // If this is an f32 -> f64 load, emit the f32 load, then use an 'extending
+ // copy'.
+ if (TypeBeingLoaded != MVT::f32 || N->getOpcode() == ISD::LOAD) {
+ CurDAG->SelectNodeTo(N, Opc, N->getValueType(0), MVT::Other,
+ Op1, Op2, Select(N->getOperand(0)));
+ return SDOperand(N, Op.ResNo);
+ } else {
+ std::vector<SDOperand> Ops;
+ Ops.push_back(Op1);
+ Ops.push_back(Op2);
+ Ops.push_back(Select(N->getOperand(0)));
+ SDOperand Res = CurDAG->getTargetNode(Opc, MVT::f32, MVT::Other, Ops);
+ SDOperand Ext = CurDAG->getTargetNode(PPC::FMRSD, MVT::f64, Res);
+ CodeGenMap[Op.getValue(0)] = Ext;
+ CodeGenMap[Op.getValue(1)] = Res.getValue(1);
+ if (Op.ResNo)
+ return Res.getValue(1);
+ else
+ return Ext;
+ }
}
case ISD::TRUNCSTORE:
@@ -1250,7 +1283,13 @@
unsigned BROpc = getBCCForSetCC(CC);
bool isFP = MVT::isFloatingPoint(N->getValueType(0));
- unsigned SelectCCOp = isFP ? PPC::SELECT_CC_FP : PPC::SELECT_CC_Int;
+ unsigned SelectCCOp;
+ if (MVT::isInteger(N->getValueType(0)))
+ SelectCCOp = PPC::SELECT_CC_Int;
+ else if (N->getValueType(0) == MVT::f32)
+ SelectCCOp = PPC::SELECT_CC_F4;
+ else
+ SelectCCOp = PPC::SELECT_CC_F8;
CurDAG->SelectNodeTo(N, SelectCCOp, N->getValueType(0), CCReg,
Select(N->getOperand(2)), Select(N->getOperand(3)),
getI32Imm(BROpc));
diff --git a/lib/Target/PowerPC/PPCISelLowering.cpp b/lib/Target/PowerPC/PPCISelLowering.cpp
index 23d7966..fdefef7 100644
--- a/lib/Target/PowerPC/PPCISelLowering.cpp
+++ b/lib/Target/PowerPC/PPCISelLowering.cpp
@@ -33,8 +33,8 @@
// Set up the register classes.
addRegisterClass(MVT::i32, PPC32::GPRCRegisterClass);
- addRegisterClass(MVT::f32, PPC32::FPRCRegisterClass);
- addRegisterClass(MVT::f64, PPC32::FPRCRegisterClass);
+ addRegisterClass(MVT::f32, PPC32::F4RCRegisterClass);
+ addRegisterClass(MVT::f64, PPC32::F8RCRegisterClass);
// PowerPC has no intrinsics for these particular operations
setOperationAction(ISD::MEMMOVE, MVT::Other, Expand);
@@ -396,7 +396,11 @@
ObjSize = (ObjectVT == MVT::f64) ? 8 : 4;
if (!ArgLive) break;
if (FPR_remaining > 0) {
- unsigned VReg = RegMap->createVirtualRegister(&PPC32::FPRCRegClass);
+ unsigned VReg;
+ if (ObjectVT == MVT::f32)
+ VReg = RegMap->createVirtualRegister(&PPC32::F4RCRegClass);
+ else
+ VReg = RegMap->createVirtualRegister(&PPC32::F8RCRegClass);
MF.addLiveIn(FPR[FPR_idx], VReg);
argt = newroot = DAG.getCopyFromReg(DAG.getRoot(), VReg, ObjectVT);
--FPR_remaining;
@@ -724,7 +728,8 @@
PPC32TargetLowering::InsertAtEndOfBasicBlock(MachineInstr *MI,
MachineBasicBlock *BB) {
assert((MI->getOpcode() == PPC::SELECT_CC_Int ||
- MI->getOpcode() == PPC::SELECT_CC_FP) &&
+ MI->getOpcode() == PPC::SELECT_CC_F4 ||
+ MI->getOpcode() == PPC::SELECT_CC_F8) &&
"Unexpected instr type to insert");
// To "insert" a SELECT_CC instruction, we actually have to insert the diamond
diff --git a/lib/Target/PowerPC/PPCISelPattern.cpp b/lib/Target/PowerPC/PPCISelPattern.cpp
index 9be7984..0571049 100644
--- a/lib/Target/PowerPC/PPCISelPattern.cpp
+++ b/lib/Target/PowerPC/PPCISelPattern.cpp
@@ -87,9 +87,6 @@
inline unsigned MakeIntReg() {
return RegMap->createVirtualRegister(PPC32::GPRCRegisterClass);
}
- inline unsigned MakeFPReg() {
- return RegMap->createVirtualRegister(PPC32::FPRCRegisterClass);
- }
// dag -> dag expanders for integer divide by constant
SDOperand BuildSDIVSequence(SDOperand N);
@@ -593,8 +590,6 @@
unsigned ISel::SelectCC(SDOperand LHS, SDOperand RHS, ISD::CondCode CC) {
unsigned Result, Tmp1, Tmp2;
bool AlreadySelected = false;
- static const unsigned CompareOpcodes[] =
- { PPC::FCMPU, PPC::FCMPU, PPC::CMPW, PPC::CMPLW };
// Allocate a condition register for this expression
Result = RegMap->createVirtualRegister(PPC32::CRRCRegisterClass);
@@ -626,8 +621,13 @@
else
BuildMI(BB, PPC::CMPWI, 2, Result).addReg(Tmp1).addSImm(Tmp2);
} else {
- bool IsInteger = MVT::isInteger(LHS.getValueType());
- unsigned CompareOpc = CompareOpcodes[2 * IsInteger + U];
+ unsigned CompareOpc;
+ if (MVT::isInteger(LHS.getValueType()))
+ CompareOpc = U ? PPC::CMPLW : PPC::CMPW;
+ else if (LHS.getValueType() == MVT::f32)
+ CompareOpc = PPC::FCMPUS;
+ else
+ CompareOpc = PPC::FCMPUD;
Tmp1 = SelectExpr(LHS);
Tmp2 = SelectExpr(RHS);
BuildMI(BB, CompareOpc, 2, Result).addReg(Tmp1).addReg(Tmp2);
@@ -815,7 +815,10 @@
Tmp1 = SelectExpr(N.getOperand(0));
Tmp2 = SelectExpr(N.getOperand(1));
Tmp3 = SelectExpr(N.getOperand(2));
- BuildMI(BB, PPC::FSEL, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ if (N.getOperand(0).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::FSELS, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
+ else
+ BuildMI(BB, PPC::FSELD, 3, Result).addReg(Tmp1).addReg(Tmp2).addReg(Tmp3);
return Result;
case PPCISD::FCFID:
Tmp1 = SelectExpr(N.getOperand(0));
@@ -832,8 +835,10 @@
case ISD::UNDEF:
if (Node->getValueType(0) == MVT::i32)
BuildMI(BB, PPC::IMPLICIT_DEF_GPR, 0, Result);
+ else if (Node->getValueType(0) == MVT::f32)
+ BuildMI(BB, PPC::IMPLICIT_DEF_F4, 0, Result);
else
- BuildMI(BB, PPC::IMPLICIT_DEF_FP, 0, Result);
+ BuildMI(BB, PPC::IMPLICIT_DEF_F8, 0, Result);
return Result;
case ISD::DYNAMIC_STACKALLOC:
// Generate both result values. FIXME: Need a better commment here?
@@ -1011,7 +1016,8 @@
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]);
+ BuildMI(BB, N.getOperand(i+2).getValueType() == MVT::f32 ? PPC::FMRS :
+ PPC::FMRD, 1, FPR[FPR_idx]).addReg(ArgVR[i]);
CallMI->addRegOperand(FPR[FPR_idx], MachineOperand::Use);
++FPR_idx;
break;
@@ -1033,14 +1039,15 @@
}
break;
case MVT::f32:
+ BuildMI(BB, PPC::FMRS, 1, Result).addReg(PPC::F1);
+ break;
case MVT::f64:
- BuildMI(BB, PPC::FMR, 1, Result).addReg(PPC::F1);
+ BuildMI(BB, PPC::FMRD, 1, Result).addReg(PPC::F1);
break;
}
return Result+N.ResNo;
}
- case ISD::SIGN_EXTEND:
case ISD::SIGN_EXTEND_INREG:
Tmp1 = SelectExpr(N.getOperand(0));
switch(cast<VTSDNode>(Node->getOperand(1))->getVT()) {
@@ -1063,8 +1070,10 @@
Tmp1 = dyn_cast<RegisterSDNode>(Node->getOperand(1))->getReg();
if (MVT::isInteger(DestType))
BuildMI(BB, PPC::OR, 2, Result).addReg(Tmp1).addReg(Tmp1);
+ else if (DestType == MVT::f32)
+ BuildMI(BB, PPC::FMRS, 1, Result).addReg(Tmp1);
else
- BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
+ BuildMI(BB, PPC::FMRD, 1, Result).addReg(Tmp1);
return Result;
case ISD::SHL:
@@ -1655,16 +1664,26 @@
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);
+ if (N.getOperand(0).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::FNABSS, 1, Result).addReg(Tmp1);
+ else
+ BuildMI(BB, PPC::FNABSD, 1, Result).addReg(Tmp1);
+
} else {
Tmp1 = SelectExpr(N.getOperand(0));
- BuildMI(BB, PPC::FNEG, 1, Result).addReg(Tmp1);
+ if (N.getOperand(0).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::FNEGS, 1, Result).addReg(Tmp1);
+ else
+ BuildMI(BB, PPC::FNEGD, 1, Result).addReg(Tmp1);
}
return Result;
case ISD::FABS:
Tmp1 = SelectExpr(N.getOperand(0));
- BuildMI(BB, PPC::FABS, 1, Result).addReg(Tmp1);
+ if (N.getOperand(0).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::FABSS, 1, Result).addReg(Tmp1);
+ else
+ BuildMI(BB, PPC::FABSD, 1, Result).addReg(Tmp1);
return Result;
case ISD::FSQRT:
@@ -1686,7 +1705,7 @@
N.getOperand(0).getValueType() == MVT::f32 &&
"only f32 to f64 conversion supported here");
Tmp1 = SelectExpr(N.getOperand(0));
- BuildMI(BB, PPC::FMR, 1, Result).addReg(Tmp1);
+ BuildMI(BB, PPC::FMRSD, 1, Result).addReg(Tmp1);
return Result;
}
return 0;
@@ -1735,9 +1754,10 @@
Tmp2 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
if (Tmp1 != Tmp2) {
- if (N.getOperand(2).getValueType() == MVT::f64 ||
- N.getOperand(2).getValueType() == MVT::f32)
- BuildMI(BB, PPC::FMR, 1, Tmp2).addReg(Tmp1);
+ if (N.getOperand(2).getValueType() == MVT::f64)
+ BuildMI(BB, PPC::FMRD, 1, Tmp2).addReg(Tmp1);
+ else if (N.getOperand(2).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::FMRS, 1, Tmp2).addReg(Tmp1);
else
BuildMI(BB, PPC::OR, 2, Tmp2).addReg(Tmp1).addReg(Tmp1);
}
@@ -1747,8 +1767,10 @@
Tmp1 = cast<RegisterSDNode>(N.getOperand(1))->getReg();
if (N.getOperand(1).getValueType() == MVT::i32)
BuildMI(BB, PPC::IMPLICIT_DEF_GPR, 0, Tmp1);
+ else if (N.getOperand(1).getValueType() == MVT::f32)
+ BuildMI(BB, PPC::IMPLICIT_DEF_F4, 0, Tmp1);
else
- BuildMI(BB, PPC::IMPLICIT_DEF_FP, 0, Tmp1);
+ BuildMI(BB, PPC::IMPLICIT_DEF_F8, 0, Tmp1);
return;
case ISD::RET:
switch (N.getNumOperands()) {
@@ -1771,8 +1793,10 @@
default:
assert(0 && "Unknown return type!");
case MVT::f64:
+ BuildMI(BB, PPC::FMRD, 1, PPC::F1).addReg(Tmp1);
+ break;
case MVT::f32:
- BuildMI(BB, PPC::FMR, 1, PPC::F1).addReg(Tmp1);
+ BuildMI(BB, PPC::FMRS, 1, PPC::F1).addReg(Tmp1);
break;
case MVT::i32:
BuildMI(BB, PPC::OR, 2, PPC::R3).addReg(Tmp1).addReg(Tmp1);
diff --git a/lib/Target/PowerPC/PPCInstrInfo.cpp b/lib/Target/PowerPC/PPCInstrInfo.cpp
index 3ec78e3..dc84075 100644
--- a/lib/Target/PowerPC/PPCInstrInfo.cpp
+++ b/lib/Target/PowerPC/PPCInstrInfo.cpp
@@ -57,7 +57,7 @@
destReg = MI.getOperand(0).getReg();
return true;
}
- } else if (oc == PPC::FMR) { // fmr r1, r2
+ } else if (oc == PPC::FMRS || oc == PPC::FMRD) { // fmr r1, r2
assert(MI.getNumOperands() == 2 &&
MI.getOperand(0).isRegister() &&
MI.getOperand(1).isRegister() &&
diff --git a/lib/Target/PowerPC/PPCInstrInfo.td b/lib/Target/PowerPC/PPCInstrInfo.td
index 98f6359..bc4278c 100644
--- a/lib/Target/PowerPC/PPCInstrInfo.td
+++ b/lib/Target/PowerPC/PPCInstrInfo.td
@@ -317,14 +317,17 @@
def ADJCALLSTACKUP : Pseudo<(ops u16imm:$amt), "; ADJCALLSTACKUP">;
}
def IMPLICIT_DEF_GPR : Pseudo<(ops GPRC:$rD), "; $rD = IMPLICIT_DEF_GPRC">;
-def IMPLICIT_DEF_FP : Pseudo<(ops FPRC:$rD), "; %rD = IMPLICIT_DEF_FP">;
+def IMPLICIT_DEF_F8 : Pseudo<(ops F8RC:$rD), "; %rD = IMPLICIT_DEF_F8">;
+def IMPLICIT_DEF_F4 : Pseudo<(ops F4RC:$rD), "; %rD = IMPLICIT_DEF_F4">;
// SELECT_CC_* - Used to implement the SELECT_CC DAG operation. Expanded by the
// scheduler into a branch sequence.
let usesCustomDAGSchedInserter = 1 in { // Expanded by the scheduler.
def SELECT_CC_Int : Pseudo<(ops GPRC:$dst, CRRC:$cond, GPRC:$T, GPRC:$F,
i32imm:$BROPC), "; SELECT_CC PSEUDO!">;
- def SELECT_CC_FP : Pseudo<(ops FPRC:$dst, CRRC:$cond, FPRC:$T, FPRC:$F,
+ def SELECT_CC_F4 : Pseudo<(ops F4RC:$dst, CRRC:$cond, F4RC:$T, F4RC:$F,
+ i32imm:$BROPC), "; SELECT_CC PSEUDO!">;
+ def SELECT_CC_F8 : Pseudo<(ops F8RC:$dst, CRRC:$cond, F8RC:$T, F8RC:$F,
i32imm:$BROPC), "; SELECT_CC PSEUDO!">;
}
@@ -463,15 +466,15 @@
def CMPLDI : DForm_6_ext<10, (ops CRRC:$dst, GPRC:$src1, u16imm:$src2),
"cmpldi $dst, $src1, $src2">, isPPC64;
let isLoad = 1 in {
-def LFS : DForm_8<48, (ops FPRC:$rD, symbolLo:$disp, GPRC:$rA),
+def LFS : DForm_8<48, (ops F4RC:$rD, symbolLo:$disp, GPRC:$rA),
"lfs $rD, $disp($rA)">;
-def LFD : DForm_8<50, (ops FPRC:$rD, symbolLo:$disp, GPRC:$rA),
+def LFD : DForm_8<50, (ops F8RC:$rD, symbolLo:$disp, GPRC:$rA),
"lfd $rD, $disp($rA)">;
}
let isStore = 1 in {
-def STFS : DForm_9<52, (ops FPRC:$rS, symbolLo:$disp, GPRC:$rA),
+def STFS : DForm_9<52, (ops F4RC:$rS, symbolLo:$disp, GPRC:$rA),
"stfs $rS, $disp($rA)">;
-def STFD : DForm_9<54, (ops FPRC:$rS, symbolLo:$disp, GPRC:$rA),
+def STFD : DForm_9<54, (ops F8RC:$rS, symbolLo:$disp, GPRC:$rA),
"stfd $rS, $disp($rA)">;
}
@@ -596,51 +599,74 @@
"cmplw $crD, $rA, $rB">;
def CMPLD : XForm_16_ext<31, 32, (ops CRRC:$crD, GPRC:$rA, GPRC:$rB),
"cmpld $crD, $rA, $rB">, isPPC64;
-def FCMPO : XForm_17<63, 32, (ops CRRC:$crD, FPRC:$fA, FPRC:$fB),
- "fcmpo $crD, $fA, $fB">;
-def FCMPU : XForm_17<63, 0, (ops CRRC:$crD, FPRC:$fA, FPRC:$fB),
+//def FCMPO : XForm_17<63, 32, (ops CRRC:$crD, FPRC:$fA, FPRC:$fB),
+// "fcmpo $crD, $fA, $fB">;
+def FCMPUS : XForm_17<63, 0, (ops CRRC:$crD, F4RC:$fA, F4RC:$fB),
"fcmpu $crD, $fA, $fB">;
+def FCMPUD : XForm_17<63, 0, (ops CRRC:$crD, F8RC:$fA, F8RC:$fB),
+ "fcmpu $crD, $fA, $fB">;
+
let isLoad = 1 in {
-def LFSX : XForm_25<31, 535, (ops FPRC:$dst, GPRC:$base, GPRC:$index),
+def LFSX : XForm_25<31, 535, (ops F4RC:$dst, GPRC:$base, GPRC:$index),
"lfsx $dst, $base, $index">;
-def LFDX : XForm_25<31, 599, (ops FPRC:$dst, GPRC:$base, GPRC:$index),
+def LFDX : XForm_25<31, 599, (ops F8RC:$dst, GPRC:$base, GPRC:$index),
"lfdx $dst, $base, $index">;
}
-def FCFID : XForm_26<63, 846, (ops FPRC:$frD, FPRC:$frB),
+def FCFID : XForm_26<63, 846, (ops F8RC:$frD, F8RC:$frB),
"fcfid $frD, $frB",
[]>, isPPC64;
-def FCTIDZ : XForm_26<63, 815, (ops FPRC:$frD, FPRC:$frB),
+def FCTIDZ : XForm_26<63, 815, (ops F8RC:$frD, F8RC:$frB),
"fctidz $frD, $frB",
[]>, isPPC64;
-def FCTIWZ : XForm_26<63, 15, (ops FPRC:$frD, FPRC:$frB),
+def FCTIWZ : XForm_26<63, 15, (ops F8RC:$frD, F8RC:$frB),
"fctiwz $frD, $frB",
[]>;
-def FABS : XForm_26<63, 264, (ops FPRC:$frD, FPRC:$frB),
- "fabs $frD, $frB",
- [(set FPRC:$frD, (fabs FPRC:$frB))]>;
-def FMR : XForm_26<63, 72, (ops FPRC:$frD, FPRC:$frB),
- "fmr $frD, $frB",
- []>; // (set FPRC:$frD, FPRC:$frB)
-def FNABS : XForm_26<63, 136, (ops FPRC:$frD, FPRC:$frB),
- "fnabs $frD, $frB",
- [(set FPRC:$frD, (fneg (fabs FPRC:$frB)))]>;
-def FNEG : XForm_26<63, 40, (ops FPRC:$frD, FPRC:$frB),
- "fneg $frD, $frB",
- [(set FPRC:$frD, (fneg FPRC:$frB))]>;
-def FRSP : XForm_26<63, 12, (ops FPRC:$frD, FPRC:$frB),
+def FRSP : XForm_26<63, 12, (ops F4RC:$frD, F8RC:$frB),
"frsp $frD, $frB",
[]>;
-def FSQRT : XForm_26<63, 22, (ops FPRC:$frD, FPRC:$frB),
+def FSQRT : XForm_26<63, 22, (ops F8RC:$frD, F8RC:$frB),
"fsqrt $frD, $frB",
- [(set FPRC:$frD, (fsqrt FPRC:$frB))]>;
-def FSQRTS : XForm_26<59, 22, (ops FPRC:$frD, FPRC:$frB),
+ [(set F8RC:$frD, (fsqrt F8RC:$frB))]>;
+def FSQRTS : XForm_26<59, 22, (ops F4RC:$frD, F4RC:$frB),
"fsqrts $frD, $frB",
[]>;
+
+/// FMR is split into 3 versions, one for 4/8 byte FP, and one for extending.
+def FMRS : XForm_26<63, 72, (ops F4RC:$frD, F4RC:$frB),
+ "fmr $frD, $frB",
+ []>; // (set F4RC:$frD, F4RC:$frB)
+def FMRD : XForm_26<63, 72, (ops F8RC:$frD, F8RC:$frB),
+ "fmr $frD, $frB",
+ []>; // (set F8RC:$frD, F8RC:$frB)
+def FMRSD : XForm_26<63, 72, (ops F8RC:$frD, F4RC:$frB),
+ "fmr $frD, $frB",
+ []>; // (set F8RC:$frD, (fpextend F4RC:$frB))
+
+// These are artificially split into two different forms, for 4/8 byte FP.
+def FABSS : XForm_26<63, 264, (ops F4RC:$frD, F4RC:$frB),
+ "fabs $frD, $frB",
+ [(set F4RC:$frD, (fabs F4RC:$frB))]>;
+def FABSD : XForm_26<63, 264, (ops F8RC:$frD, F8RC:$frB),
+ "fabs $frD, $frB",
+ [(set F8RC:$frD, (fabs F8RC:$frB))]>;
+def FNABSS : XForm_26<63, 136, (ops F4RC:$frD, F4RC:$frB),
+ "fnabs $frD, $frB",
+ [(set F4RC:$frD, (fneg (fabs F4RC:$frB)))]>;
+def FNABSD : XForm_26<63, 136, (ops F8RC:$frD, F8RC:$frB),
+ "fnabs $frD, $frB",
+ [(set F8RC:$frD, (fneg (fabs F8RC:$frB)))]>;
+def FNEGS : XForm_26<63, 40, (ops F4RC:$frD, F4RC:$frB),
+ "fneg $frD, $frB",
+ [(set F4RC:$frD, (fneg F4RC:$frB))]>;
+def FNEGD : XForm_26<63, 40, (ops F8RC:$frD, F8RC:$frB),
+ "fneg $frD, $frB",
+ [(set F8RC:$frD, (fneg F8RC:$frB))]>;
+
let isStore = 1 in {
-def STFSX : XForm_28<31, 663, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB),
+def STFSX : XForm_28<31, 663, (ops F4RC:$frS, GPRC:$rA, GPRC:$rB),
"stfsx $frS, $rA, $rB">;
-def STFDX : XForm_28<31, 727, (ops FPRC:$frS, GPRC:$rA, GPRC:$rB),
+def STFDX : XForm_28<31, 727, (ops F8RC:$frS, GPRC:$rA, GPRC:$rB),
"stfdx $frS, $rA, $rB">;
}
@@ -730,75 +756,80 @@
// this type.
//
def FMADD : AForm_1<63, 29,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRC, F8RC:$FRB),
"fmadd $FRT, $FRA, $FRC, $FRB",
- [(set FPRC:$FRT, (fadd (fmul FPRC:$FRA, FPRC:$FRC),
- FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fadd (fmul F8RC:$FRA, F8RC:$FRC),
+ F8RC:$FRB))]>;
def FMADDS : AForm_1<59, 29,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
"fmadds $FRT, $FRA, $FRC, $FRB",
[]>;
def FMSUB : AForm_1<63, 28,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRC, F8RC:$FRB),
"fmsub $FRT, $FRA, $FRC, $FRB",
- [(set FPRC:$FRT, (fsub (fmul FPRC:$FRA, FPRC:$FRC),
- FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fsub (fmul F8RC:$FRA, F8RC:$FRC),
+ F8RC:$FRB))]>;
def FMSUBS : AForm_1<59, 28,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
"fmsubs $FRT, $FRA, $FRC, $FRB",
[]>;
def FNMADD : AForm_1<63, 31,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRC, F8RC:$FRB),
"fnmadd $FRT, $FRA, $FRC, $FRB",
- [(set FPRC:$FRT, (fneg (fadd (fmul FPRC:$FRA, FPRC:$FRC),
- FPRC:$FRB)))]>;
+ [(set F8RC:$FRT, (fneg (fadd (fmul F8RC:$FRA, F8RC:$FRC),
+ F8RC:$FRB)))]>;
def FNMADDS : AForm_1<59, 31,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
"fnmadds $FRT, $FRA, $FRC, $FRB",
[]>;
def FNMSUB : AForm_1<63, 30,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRC, F8RC:$FRB),
"fnmsub $FRT, $FRA, $FRC, $FRB",
- [(set FPRC:$FRT, (fneg (fsub (fmul FPRC:$FRA, FPRC:$FRC),
- FPRC:$FRB)))]>;
+ [(set F8RC:$FRT, (fneg (fsub (fmul F8RC:$FRA, F8RC:$FRC),
+ F8RC:$FRB)))]>;
def FNMSUBS : AForm_1<59, 30,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
"fnmsubs $FRT, $FRA, $FRC, $FRB",
[]>;
-def FSEL : AForm_1<63, 23,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRC, FPRC:$FRB),
+// FSEL is artificially split into 4 and 8-byte forms.
+def FSELD : AForm_1<63, 23,
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRC, F8RC:$FRB),
+ "fsel $FRT, $FRA, $FRC, $FRB",
+ []>;
+def FSELS : AForm_1<63, 23,
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
"fsel $FRT, $FRA, $FRC, $FRB",
[]>;
def FADD : AForm_2<63, 21,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRB),
"fadd $FRT, $FRA, $FRB",
- [(set FPRC:$FRT, (fadd FPRC:$FRA, FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fadd F8RC:$FRA, F8RC:$FRB))]>;
def FADDS : AForm_2<59, 21,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
"fadds $FRT, $FRA, $FRB",
[]>;
def FDIV : AForm_2<63, 18,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRB),
"fdiv $FRT, $FRA, $FRB",
- [(set FPRC:$FRT, (fdiv FPRC:$FRA, FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fdiv F8RC:$FRA, F8RC:$FRB))]>;
def FDIVS : AForm_2<59, 18,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
"fdivs $FRT, $FRA, $FRB",
[]>;
def FMUL : AForm_3<63, 25,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRB),
"fmul $FRT, $FRA, $FRB",
- [(set FPRC:$FRT, (fmul FPRC:$FRA, FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fmul F8RC:$FRA, F8RC:$FRB))]>;
def FMULS : AForm_3<59, 25,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
"fmuls $FRT, $FRA, $FRB",
[]>;
def FSUB : AForm_2<63, 20,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F8RC:$FRT, F8RC:$FRA, F8RC:$FRB),
"fsub $FRT, $FRA, $FRB",
- [(set FPRC:$FRT, (fsub FPRC:$FRA, FPRC:$FRB))]>;
+ [(set F8RC:$FRT, (fsub F8RC:$FRA, F8RC:$FRB))]>;
def FSUBS : AForm_2<59, 20,
- (ops FPRC:$FRT, FPRC:$FRA, FPRC:$FRB),
+ (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
"fsubs $FRT, $FRA, $FRB",
[]>;
diff --git a/lib/Target/PowerPC/PPCRegisterInfo.cpp b/lib/Target/PowerPC/PPCRegisterInfo.cpp
index 76b0c59..955cb6b 100644
--- a/lib/Target/PowerPC/PPCRegisterInfo.cpp
+++ b/lib/Target/PowerPC/PPCRegisterInfo.cpp
@@ -43,47 +43,26 @@
ImmToIdxMap[PPC::ADDI] = PPC::ADD;
}
-static unsigned getIdx(const TargetRegisterClass *RC) {
- if (RC == PPC32::GPRCRegisterClass) {
- switch (RC->getSize()) {
- default: assert(0 && "Invalid data size!");
- case 1: return 0;
- case 2: return 1;
- case 4: return 2;
- }
- } else if (RC == PPC32::FPRCRegisterClass) {
- switch (RC->getSize()) {
- default: assert(0 && "Invalid data size!");
- case 4: return 3;
- case 8: return 4;
- }
- } else if (RC == PPC32::CRRCRegisterClass) {
- switch (RC->getSize()) {
- default: assert(0 && "Invalid data size!");
- case 4: return 2;
- }
- }
- std::cerr << "Invalid register class to getIdx()!\n";
- abort();
-}
-
void
PPC32RegisterInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
MachineBasicBlock::iterator MI,
unsigned SrcReg, int FrameIdx,
const TargetRegisterClass *RC) const {
- static const unsigned Opcode[] = {
- PPC::STB, PPC::STH, PPC::STW, PPC::STFS, PPC::STFD
- };
- unsigned OC = Opcode[getIdx(RC)];
if (SrcReg == PPC::LR) {
BuildMI(MBB, MI, PPC::MFLR, 1, PPC::R11);
- addFrameReference(BuildMI(MBB, MI, OC, 3).addReg(PPC::R11),FrameIdx);
+ addFrameReference(BuildMI(MBB, MI, PPC::STW, 3).addReg(PPC::R11), FrameIdx);
} else if (RC == PPC32::CRRCRegisterClass) {
BuildMI(MBB, MI, PPC::MFCR, 0, PPC::R11);
- addFrameReference(BuildMI(MBB, MI, OC, 3).addReg(PPC::R11),FrameIdx);
+ addFrameReference(BuildMI(MBB, MI, PPC::STW, 3).addReg(PPC::R11), FrameIdx);
+ } else if (RC == PPC32::GPRCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::STW, 3).addReg(SrcReg),FrameIdx);
+ } else if (RC == PPC32::F8RCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::STFD, 3).addReg(SrcReg),FrameIdx);
+ } else if (RC == PPC32::F4RCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::STFS, 3).addReg(SrcReg),FrameIdx);
} else {
- addFrameReference(BuildMI(MBB, MI, OC, 3).addReg(SrcReg),FrameIdx);
+ assert(0 && "Unknown regclass!");
+ abort();
}
}
@@ -92,18 +71,21 @@
MachineBasicBlock::iterator MI,
unsigned DestReg, int FrameIdx,
const TargetRegisterClass *RC) const {
- static const unsigned Opcode[] = {
- PPC::LBZ, PPC::LHZ, PPC::LWZ, PPC::LFS, PPC::LFD
- };
- unsigned OC = Opcode[getIdx(RC)];
if (DestReg == PPC::LR) {
- addFrameReference(BuildMI(MBB, MI, OC, 2, PPC::R11), FrameIdx);
+ addFrameReference(BuildMI(MBB, MI, PPC::LWZ, 2, PPC::R11), FrameIdx);
BuildMI(MBB, MI, PPC::MTLR, 1).addReg(PPC::R11);
} else if (RC == PPC32::CRRCRegisterClass) {
- addFrameReference(BuildMI(MBB, MI, OC, 2, PPC::R11), FrameIdx);
+ addFrameReference(BuildMI(MBB, MI, PPC::LWZ, 2, PPC::R11), FrameIdx);
BuildMI(MBB, MI, PPC::MTCRF, 1, DestReg).addReg(PPC::R11);
+ } else if (RC == PPC32::GPRCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::LWZ, 2, DestReg), FrameIdx);
+ } else if (RC == PPC32::F8RCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::LFD, 2, DestReg), FrameIdx);
+ } else if (RC == PPC32::F4RCRegisterClass) {
+ addFrameReference(BuildMI(MBB, MI, PPC::LFS, 2, DestReg), FrameIdx);
} else {
- addFrameReference(BuildMI(MBB, MI, OC, 2, DestReg), FrameIdx);
+ assert(0 && "Unknown regclass!");
+ abort();
}
}
@@ -115,8 +97,10 @@
if (RC == PPC32::GPRCRegisterClass) {
BuildMI(MBB, MI, PPC::OR, 2, DestReg).addReg(SrcReg).addReg(SrcReg);
- } else if (RC == PPC32::FPRCRegisterClass) {
- BuildMI(MBB, MI, PPC::FMR, 1, DestReg).addReg(SrcReg);
+ } else if (RC == PPC32::F4RCRegisterClass) {
+ BuildMI(MBB, MI, PPC::FMRS, 1, DestReg).addReg(SrcReg);
+ } else if (RC == PPC32::F8RCRegisterClass) {
+ BuildMI(MBB, MI, PPC::FMRD, 1, DestReg).addReg(SrcReg);
} else if (RC == PPC32::CRRCRegisterClass) {
BuildMI(MBB, MI, PPC::MCRF, 1, DestReg).addReg(SrcReg);
} else {
@@ -130,6 +114,9 @@
switch (MI->getOpcode()) {
default: break;
case PPC::LWZ:
+ //
+ // case PPC::LFS: // ENABLE!!
+ //
case PPC::LFD:
if (MI->getOperand(1).isImmediate() && !MI->getOperand(1).getImmedValue() &&
MI->getOperand(2).isFrameIndex()) {
@@ -161,8 +148,7 @@
return addFrameReference(BuildMI(PPC::LWZ, 2, OutReg), FrameIndex);
}
- } else if (Opc == PPC::FMR) {
- // We currently always spill FP values as doubles. :(
+ } else if (Opc == PPC::FMRD) {
if (OpNum == 0) { // move -> store
unsigned InReg = MI->getOperand(1).getReg();
return addFrameReference(BuildMI(PPC::STFD,
@@ -171,6 +157,15 @@
unsigned OutReg = MI->getOperand(0).getReg();
return addFrameReference(BuildMI(PPC::LFD, 2, OutReg), FrameIndex);
}
+ } else if (Opc == PPC::FMRS) {
+ if (OpNum == 0) { // move -> store
+ unsigned InReg = MI->getOperand(1).getReg();
+ return addFrameReference(BuildMI(PPC::STFS,
+ 3).addReg(InReg), FrameIndex);
+ } else { // move -> load
+ unsigned OutReg = MI->getOperand(0).getReg();
+ return addFrameReference(BuildMI(PPC::LFS, 2, OutReg), FrameIndex);
+ }
}
return 0;
}
@@ -262,7 +257,8 @@
MI.SetMachineOperandReg(1, MI.getOperand(i).getReg());
MI.SetMachineOperandReg(2, PPC::R0);
} else {
- MI.SetMachineOperandConst(OffIdx,MachineOperand::MO_SignExtendedImmed,Offset);
+ MI.SetMachineOperandConst(OffIdx, MachineOperand::MO_SignExtendedImmed,
+ Offset);
}
}