diff --git a/lib/Target/PowerPC/PPCInstrInfo.cpp b/lib/Target/PowerPC/PPCInstrInfo.cpp index f28d07c8a61..fe103c65771 100644 --- a/lib/Target/PowerPC/PPCInstrInfo.cpp +++ b/lib/Target/PowerPC/PPCInstrInfo.cpp @@ -511,7 +511,7 @@ PPCInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB, NewMIs.back()->addMemOperand(MF, MMO); } -void +bool PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, unsigned DestReg, int FrameIdx, const TargetRegisterClass *RC, @@ -541,28 +541,36 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LFS), DestReg), FrameIdx)); } else if (PPC::CRRCRegisterClass->hasSubClassEq(RC)) { - // FIXME: We need a scatch reg here. The trouble with using R0 is that - // it's possible for the stack frame to be so big the save location is - // out of range of immediate offsets, necessitating another register. - // We hack this on Darwin by reserving R2. It's probably broken on Linux - // at the moment. - unsigned ScratchReg = TM.getSubtargetImpl()->isDarwinABI() ? - PPC::R2 : PPC::R0; - NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), - ScratchReg), FrameIdx)); - - // If the reloaded register isn't CR0, shift the bits right so that they are - // in the right CR's slot. - if (DestReg != PPC::CR0) { - unsigned ShiftBits = getPPCRegisterNumbering(DestReg)*4; - // rlwinm r11, r11, 32-ShiftBits, 0, 31. - NewMIs.push_back(BuildMI(MF, DL, get(PPC::RLWINM), ScratchReg) - .addReg(ScratchReg).addImm(32-ShiftBits).addImm(0) - .addImm(31)); + if ((!DisablePPC32RS && !TM.getSubtargetImpl()->isPPC64()) || + (!DisablePPC64RS && TM.getSubtargetImpl()->isPPC64())) { + NewMIs.push_back(addFrameReference(BuildMI(MF, DL, + get(PPC::RESTORE_CR), DestReg) + , FrameIdx)); + return true; + } else { + // FIXME: We need a scatch reg here. The trouble with using R0 is that + // it's possible for the stack frame to be so big the save location is + // out of range of immediate offsets, necessitating another register. + // We hack this on Darwin by reserving R2. It's probably broken on Linux + // at the moment. + unsigned ScratchReg = TM.getSubtargetImpl()->isDarwinABI() ? + PPC::R2 : PPC::R0; + NewMIs.push_back(addFrameReference(BuildMI(MF, DL, get(PPC::LWZ), + ScratchReg), FrameIdx)); + + // If the reloaded register isn't CR0, shift the bits right so that they are + // in the right CR's slot. + if (DestReg != PPC::CR0) { + unsigned ShiftBits = getPPCRegisterNumbering(DestReg)*4; + // rlwinm r11, r11, 32-ShiftBits, 0, 31. + NewMIs.push_back(BuildMI(MF, DL, get(PPC::RLWINM), ScratchReg) + .addReg(ScratchReg).addImm(32-ShiftBits).addImm(0) + .addImm(31)); + } + + NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTCRF), DestReg) + .addReg(ScratchReg)); } - - NewMIs.push_back(BuildMI(MF, DL, get(PPC::MTCRF), DestReg) - .addReg(ScratchReg)); } else if (PPC::CRBITRCRegisterClass->hasSubClassEq(RC)) { unsigned Reg = 0; @@ -607,6 +615,8 @@ PPCInstrInfo::LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, } else { llvm_unreachable("Unknown regclass!"); } + + return false; } void @@ -619,7 +629,10 @@ PPCInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB, SmallVector NewMIs; DebugLoc DL; if (MI != MBB.end()) DL = MI->getDebugLoc(); - LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs); + if (LoadRegFromStackSlot(MF, DL, DestReg, FrameIdx, RC, NewMIs)) { + PPCFunctionInfo *FuncInfo = MF.getInfo(); + FuncInfo->setSpillsCR(); + } for (unsigned i = 0, e = NewMIs.size(); i != e; ++i) MBB.insert(MI, NewMIs[i]); diff --git a/lib/Target/PowerPC/PPCInstrInfo.h b/lib/Target/PowerPC/PPCInstrInfo.h index 8b35fb23a1b..e90f8cb242c 100644 --- a/lib/Target/PowerPC/PPCInstrInfo.h +++ b/lib/Target/PowerPC/PPCInstrInfo.h @@ -72,7 +72,7 @@ class PPCInstrInfo : public PPCGenInstrInfo { unsigned SrcReg, bool isKill, int FrameIdx, const TargetRegisterClass *RC, SmallVectorImpl &NewMIs) const; - void LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, + bool LoadRegFromStackSlot(MachineFunction &MF, DebugLoc DL, unsigned DestReg, int FrameIdx, const TargetRegisterClass *RC, SmallVectorImpl &NewMIs) const; diff --git a/lib/Target/PowerPC/PPCInstrInfo.td b/lib/Target/PowerPC/PPCInstrInfo.td index 74a42ba8c42..ec54867d494 100644 --- a/lib/Target/PowerPC/PPCInstrInfo.td +++ b/lib/Target/PowerPC/PPCInstrInfo.td @@ -402,6 +402,11 @@ let usesCustomInserter = 1, // Expanded after instruction selection. def SPILL_CR : Pseudo<(outs), (ins GPRC:$cond, memri:$F), "", []>; +// RESTORE_CR - Indicate that we're restoring the CR register (previously +// spilled), so we'll need to scavenge a register for it. +def RESTORE_CR : Pseudo<(outs GPRC:$cond), (ins memri:$F), + "", []>; + let isTerminator = 1, isBarrier = 1, PPC970_Unit = 7 in { let isReturn = 1, Uses = [LR, RM] in def BLR : XLForm_2_br<19, 16, 0, (outs), (ins pred:$p), diff --git a/lib/Target/PowerPC/PPCRegisterInfo.cpp b/lib/Target/PowerPC/PPCRegisterInfo.cpp index f6c4a3d06b0..e47f65e2433 100644 --- a/lib/Target/PowerPC/PPCRegisterInfo.cpp +++ b/lib/Target/PowerPC/PPCRegisterInfo.cpp @@ -459,7 +459,7 @@ void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, unsigned FrameIndex, int SPAdj, RegScavenger *RS) const { // Get the instruction. - MachineInstr &MI = *II; // ; SPILL_CR , , + MachineInstr &MI = *II; // ; SPILL_CR , // Get the instruction's basic block. MachineBasicBlock &MBB = *MI.getParent(); DebugLoc dl = MI.getDebugLoc(); @@ -494,6 +494,44 @@ void PPCRegisterInfo::lowerCRSpilling(MachineBasicBlock::iterator II, MBB.erase(II); } +void PPCRegisterInfo::lowerCRRestore(MachineBasicBlock::iterator II, + unsigned FrameIndex, int SPAdj, + RegScavenger *RS) const { + // Get the instruction. + MachineInstr &MI = *II; // ; = RESTORE_CR + // Get the instruction's basic block. + MachineBasicBlock &MBB = *MI.getParent(); + DebugLoc dl = MI.getDebugLoc(); + + const TargetRegisterClass *G8RC = &PPC::G8RCRegClass; + const TargetRegisterClass *GPRC = &PPC::GPRCRegClass; + const TargetRegisterClass *RC = Subtarget.isPPC64() ? G8RC : GPRC; + unsigned Reg = findScratchRegister(II, RS, RC, SPAdj); + unsigned DestReg = MI.getOperand(0).getReg(); + assert(MI.definesRegister(DestReg) && + "RESTORE_CR does not define its destination"); + bool LP64 = Subtarget.isPPC64(); + + addFrameReference(BuildMI(MBB, II, dl, TII.get(LP64 ? PPC::LWZ8 : PPC::LWZ), + Reg), FrameIndex); + + // If the reloaded register isn't CR0, shift the bits right so that they are + // in the right CR's slot. + if (DestReg != PPC::CR0) { + unsigned ShiftBits = getPPCRegisterNumbering(DestReg)*4; + // rlwinm r11, r11, 32-ShiftBits, 0, 31. + BuildMI(MBB, II, dl, TII.get(PPC::RLWINM), Reg) + .addReg(Reg).addImm(32-ShiftBits).addImm(0) + .addImm(31); + } + + BuildMI(MBB, II, dl, TII.get(PPC::MTCRF), DestReg) + .addReg(Reg); + + // Discard the pseudo instruction. + MBB.erase(II); +} + void PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj, RegScavenger *RS) const { @@ -539,12 +577,16 @@ PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, return; } - // Special case for pseudo-op SPILL_CR. - if (requiresRegisterScavenging(MF)) + // Special case for pseudo-ops SPILL_CR and RESTORE_CR. + if (requiresRegisterScavenging(MF)) { if (OpC == PPC::SPILL_CR) { lowerCRSpilling(II, FrameIndex, SPAdj, RS); return; + } else if (OpC == PPC::RESTORE_CR) { + lowerCRRestore(II, FrameIndex, SPAdj, RS); + return; } + } // Replace the FrameIndex with base register with GPR1 (SP) or GPR31 (FP). MI.getOperand(FIOperandNo).ChangeToRegister(TFI->hasFP(MF) ? @@ -594,7 +636,6 @@ PPCRegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, // The offset doesn't fit into a single register, scavenge one to build the // offset in. - // FIXME: figure out what SPAdj is doing here. unsigned SReg; if (requiresRegisterScavenging(MF)) diff --git a/lib/Target/PowerPC/PPCRegisterInfo.h b/lib/Target/PowerPC/PPCRegisterInfo.h index f70a594e99b..faf690f48de 100644 --- a/lib/Target/PowerPC/PPCRegisterInfo.h +++ b/lib/Target/PowerPC/PPCRegisterInfo.h @@ -57,6 +57,8 @@ public: int SPAdj, RegScavenger *RS) const; void lowerCRSpilling(MachineBasicBlock::iterator II, unsigned FrameIndex, int SPAdj, RegScavenger *RS) const; + void lowerCRRestore(MachineBasicBlock::iterator II, unsigned FrameIndex, + int SPAdj, RegScavenger *RS) const; void eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj, RegScavenger *RS = NULL) const;