diff options
Diffstat (limited to 'llvm')
| -rw-r--r-- | llvm/lib/CodeGen/RegAlloc/PhyRegAlloc.cpp | 407 |
1 files changed, 326 insertions, 81 deletions
diff --git a/llvm/lib/CodeGen/RegAlloc/PhyRegAlloc.cpp b/llvm/lib/CodeGen/RegAlloc/PhyRegAlloc.cpp index 4d68b15dea8..41b27a66184 100644 --- a/llvm/lib/CodeGen/RegAlloc/PhyRegAlloc.cpp +++ b/llvm/lib/CodeGen/RegAlloc/PhyRegAlloc.cpp @@ -14,6 +14,7 @@ #include "llvm/CodeGen/MachineFunction.h" #include "llvm/CodeGen/MachineFunctionInfo.h" #include "llvm/CodeGen/FunctionLiveVarInfo.h" +#include "llvm/CodeGen/InstrSelection.h" #include "llvm/Analysis/LoopInfo.h" #include "llvm/Target/TargetMachine.h" #include "llvm/Target/TargetFrameInfo.h" @@ -489,13 +490,12 @@ AppendInstructions(std::vector<MachineInstr *> &IAft, } } -void PhyRegAlloc::updateInstruction(MachineInstr* MInst, BasicBlock* BB) +static bool MarkAllocatedRegs(MachineInstr* MInst, + LiveRangeInfo& LRI, + const TargetRegInfo& MRI) { - unsigned Opcode = MInst->getOpCode(); - - // Reset tmp stack positions so they can be reused for each machine instr. - MF.getInfo()->popAllTempValues(); - + bool instrNeedsSpills = false; + // First, set the registers for operands in the machine instruction // if a register was successfully allocated. Do this first because we // will need to know which registers are already used by this instr'n. @@ -507,19 +507,41 @@ void PhyRegAlloc::updateInstruction(MachineInstr* MInst, BasicBlock* BB) Op.getType() == MachineOperand::MO_CCRegister) { const Value *const Val = Op.getVRegValue(); - if (const LiveRange* LR = LRI.getLiveRangeForValue(Val)) + if (const LiveRange* LR = LRI.getLiveRangeForValue(Val)) { + // Remember if any operand needs spilling + instrNeedsSpills |= LR->isMarkedForSpill(); + + // An operand may have a color whether or not it needs spilling if (LR->hasColor()) MInst->SetRegForOperand(OpNum, MRI.getUnifiedRegNum(LR->getRegClass()->getID(), LR->getColor())); + } } } // for each operand - - // Mark that the operands have been updated. setRelRegsUsedByThisInst() - // is called to find registers used by each MachineInst, and it should not - // be used for an instruction until this is done. This flag just serves - // as a sanity check. + + return instrNeedsSpills; +} + +void PhyRegAlloc::updateInstruction(MachineBasicBlock::iterator& MII, + MachineBasicBlock &MBB) +{ + MachineInstr* MInst = *MII; + unsigned Opcode = MInst->getOpCode(); + + // Reset tmp stack positions so they can be reused for each machine instr. + MF.getInfo()->popAllTempValues(); + + // Mark the operands for which regs have been allocated. + bool instrNeedsSpills = MarkAllocatedRegs(*MII, LRI, MRI); + +#ifndef NDEBUG + // Mark that the operands have been updated. Later, + // setRelRegsUsedByThisInst() is called to find registers used by each + // MachineInst, and it should not be used for an instruction until + // this is done. This flag just serves as a sanity check. OperandsColoredMap[MInst] = true; +#endif // Now insert caller-saving code before/after the call. // Do this before inserting spill code since some registers must be @@ -527,25 +549,26 @@ void PhyRegAlloc::updateInstruction(MachineInstr* MInst, BasicBlock* BB) // if (TM.getInstrInfo().isCall(Opcode)) { AddedInstrns &AI = AddedInstrMap[MInst]; - MRI.insertCallerSavingCode(AI.InstrnsBefore, AI.InstrnsAfter, - MInst, BB, *this); + insertCallerSavingCode(AI.InstrnsBefore, AI.InstrnsAfter, MInst, + MBB.getBasicBlock()); } // Now insert spill code for remaining operands not allocated to // registers. This must be done even for call return instructions // since those are not handled by the special code above. - for (unsigned OpNum=0; OpNum < MInst->getNumOperands(); ++OpNum) - { - MachineOperand& Op = MInst->getOperand(OpNum); - if (Op.getType() == MachineOperand::MO_VirtualRegister || - Op.getType() == MachineOperand::MO_CCRegister) - { - const Value* Val = Op.getVRegValue(); - if (const LiveRange *LR = LRI.getLiveRangeForValue(Val)) - if (LR->isMarkedForSpill()) - insertCode4SpilledLR(LR, MInst, BB, OpNum); - } - } // for each operand + if (instrNeedsSpills) + for (unsigned OpNum=0; OpNum < MInst->getNumOperands(); ++OpNum) + { + MachineOperand& Op = MInst->getOperand(OpNum); + if (Op.getType() == MachineOperand::MO_VirtualRegister || + Op.getType() == MachineOperand::MO_CCRegister) + { + const Value* Val = Op.getVRegValue(); + if (const LiveRange *LR = LRI.getLiveRangeForValue(Val)) + if (LR->isMarkedForSpill()) + insertCode4SpilledLR(LR, MII, MBB, OpNum); + } + } // for each operand } void PhyRegAlloc::updateMachineCode() @@ -566,10 +589,9 @@ void PhyRegAlloc::updateMachineCode() // Iterate over all machine instructions in BB and mark operands with // their assigned registers or insert spill code, as appropriate. // Also, fix operands of call/return instructions. - // for (MachineBasicBlock::iterator MII = MBB.begin(); MII != MBB.end(); ++MII) - if (!TM.getInstrInfo().isDummyPhiInstr((*MII)->getOpCode()))// ignore Phis - updateInstruction(*MII, const_cast<BasicBlock*>(MBB.getBasicBlock())); + if (! TM.getInstrInfo().isDummyPhiInstr((*MII)->getOpCode())) + updateInstruction(MII, MBB); // Now, move code out of delay slots of branches and returns if needed. // (Also, move "after" code from calls to the last delay slot instruction.) @@ -591,50 +613,48 @@ void PhyRegAlloc::updateMachineCode() if (unsigned delaySlots = TM.getInstrInfo().getNumDelaySlots((*MII)->getOpCode())) { - assert(delaySlots==1 && "Not handling multiple delay slots!"); - - MachineInstr *MInst = *MII; - MachineInstr *MDelayInst = *(MII+1); - + MachineInstr *MInst = *MII, *DelaySlotMI = *(MII+1); + // Check the 2 conditions above: // (1) Does a branch need instructions added after it? // (2) O/w does delay slot instr. need instrns before or after? - bool isBranch = (TM.getInstrInfo().isBranch((*MII)->getOpCode()) || - TM.getInstrInfo().isReturn((*MII)->getOpCode())); - bool cond1 = isBranch && AddedInstrMap[MInst].InstrnsAfter.size() > 0; - bool cond2 = (AddedInstrMap.count(MDelayInst) || - AddedInstrMap[MDelayInst].InstrnsAfter.size() > 0); + bool isBranch = (TM.getInstrInfo().isBranch(MInst->getOpCode()) || + TM.getInstrInfo().isReturn(MInst->getOpCode())); + bool cond1 = (isBranch && + AddedInstrMap.count(MInst) && + AddedInstrMap[MInst].InstrnsAfter.size() > 0); + bool cond2 = (AddedInstrMap.count(DelaySlotMI) && + (AddedInstrMap[DelaySlotMI].InstrnsBefore.size() > 0 || + AddedInstrMap[DelaySlotMI].InstrnsAfter.size() > 0)); if (cond1 || cond2) { - // Move delay slot instrn before the preceding branch. - // InsertBefore() modifies MII to point to the branch again. - assert(((*MII)->getOpCodeFlags() & AnnulFlag) == 0 && - "FIXME: Annul bit must be turned off here!"); - InsertBefore(MDelayInst, MBB, MII); + assert((MInst->getOpCodeFlags() & AnnulFlag) == 0 && + "FIXME: Moving an annulled delay slot instruction!"); + assert(delaySlots==1 && + "InsertBefore does not yet handle >1 delay slots!"); + InsertBefore(DelaySlotMI, MBB, MII); // MII pts back to branch // In case (1), delete it and don't replace with anything! // Otherwise (i.e., case (2) only) replace it with a NOP. if (cond1) { - assert(AddedInstrMap[MInst].InstrnsAfter.size() <= delaySlots && - "Cannot put more than #delaySlots spill instrns after " - "branch or return! Need to handle spill differently."); - DeleteInstruction(MBB, MII); // MII now points to next inst. + DeleteInstruction(MBB, ++MII); // MII now points to next inst. + --MII; // reset MII for ++MII of loop } - else { - MachineInstr* nopI =BuildMI(TM.getInstrInfo().getNOPOpCode(),1); - SubstituteInPlace(nopI, MBB, MII+1); // replace with NOP + else + SubstituteInPlace(BuildMI(TM.getInstrInfo().getNOPOpCode(),1), + MBB, MII+1); // replace with NOP + + if (DEBUG_RA) { + cerr << "\nRegAlloc: Moved instr. with added code: " + << *DelaySlotMI + << " out of delay slots of instr: " << *MInst; } } - - // If this is not a branch or return (probably a call), - // the Instrnsafter, if any, must really go after the last - // delay slot. Move the InstrAfter to the instr. in that slot. - // We must do this after the previous code because the instructions - // in delay slots may get moved out by that code. - // - if (!isBranch) - move2DelayedInstr(MInst, *(MII+delaySlots)); + else + // For non-branch instr with delay slots (probably a call), move + // InstrAfter to the instr. in the last delay slot. + move2DelayedInstr(*MII, *(MII+delaySlots)); } // Finally iterate over all instructions in BB and insert before/after @@ -651,6 +671,16 @@ void PhyRegAlloc::updateMachineCode() AddedInstrns &CallAI = AddedInstrMap[MInst]; #ifndef NDEBUG + bool isBranch = (TM.getInstrInfo().isBranch(MInst->getOpCode()) || + TM.getInstrInfo().isReturn(MInst->getOpCode())); + assert((!isBranch || + AddedInstrMap[MInst].InstrnsAfter.size() <= + TM.getInstrInfo().getNumDelaySlots(MInst->getOpCode())) && + "Cannot put more than #delaySlots instrns after " + "branch or return! Need to handle temps differently."); +#endif + +#ifndef NDEBUG // Temporary sanity checking code to detect whether the same machine // instruction is ever inserted twice before/after a call. // I suspect this is happening but am not sure. --Vikram, 7/1/03. @@ -681,6 +711,7 @@ void PhyRegAlloc::updateMachineCode() } // if there are any added instructions } // for each machine instruction + } } @@ -696,10 +727,13 @@ void PhyRegAlloc::updateMachineCode() //---------------------------------------------------------------------------- void PhyRegAlloc::insertCode4SpilledLR(const LiveRange *LR, - MachineInstr *MInst, - const BasicBlock *BB, + MachineBasicBlock::iterator& MII, + MachineBasicBlock &MBB, const unsigned OpNum) { + MachineInstr *MInst = *MII; + const BasicBlock *BB = MBB.getBasicBlock(); + assert((! TM.getInstrInfo().isCall(MInst->getOpCode()) || OpNum == 0) && "Outgoing arg of a call must be handled elsewhere (func arg ok)"); assert(! TM.getInstrInfo().isReturn(MInst->getOpCode()) && @@ -711,7 +745,20 @@ void PhyRegAlloc::insertCode4SpilledLR(const LiveRange *LR, unsigned RegType = MRI.getRegTypeForLR(LR); int SpillOff = LR->getSpillOffFromFP(); RegClass *RC = LR->getRegClass(); - const ValueSet &LVSetBef = LVI->getLiveVarSetBeforeMInst(MInst, BB); + + // Get the live-variable set to find registers free before this instr. + // If this instr. is in the delay slot of a branch or return, use the live + // var set before that branch or return -- we don't want to trample those! + // + MachineInstr *LiveBeforeThisMI = MInst; + if (MII != MBB.begin()) { + MachineInstr *PredMI = *(MII-1); + if (unsigned DS = TM.getInstrInfo().getNumDelaySlots(PredMI->getOpCode())) { + assert(DS == 1 && "Only checking immediate pred. for delay slots!"); + LiveBeforeThisMI = PredMI; + } + } + const ValueSet &LVSetBef = LVI->getLiveVarSetBeforeMInst(LiveBeforeThisMI,BB); MF.getInfo()->pushTempValue(MRI.getSpilledRegSize(RegType) ); @@ -751,8 +798,8 @@ void PhyRegAlloc::insertCode4SpilledLR(const LiveRange *LR, // and use the TmpReg as one operand of instruction // actual loading instruction(s) - MRI.cpMem2RegMI(AdIMid, MRI.getFramePointer(), SpillOff, TmpRegU, RegType, - scratchReg); + MRI.cpMem2RegMI(AdIMid, MRI.getFramePointer(), SpillOff, TmpRegU, + RegType, scratchReg); // the actual load should be after the instructions to free up TmpRegU MIBef.insert(MIBef.end(), AdIMid.begin(), AdIMid.end()); @@ -764,8 +811,8 @@ void PhyRegAlloc::insertCode4SpilledLR(const LiveRange *LR, // on the stack position allocated for this LR // actual storing instruction(s) - MRI.cpReg2MemMI(AdIMid, TmpRegU, MRI.getFramePointer(), SpillOff, RegType, - scratchReg); + MRI.cpReg2MemMI(AdIMid, TmpRegU, MRI.getFramePointer(), SpillOff, + RegType, scratchReg); MIAft.insert(MIAft.begin(), AdIMid.begin(), AdIMid.end()); } // if !DEF @@ -784,6 +831,195 @@ void PhyRegAlloc::insertCode4SpilledLR(const LiveRange *LR, } + +//---------------------------------------------------------------------------- +// This method inserts caller saving/restoring instructons before/after +// a call machine instruction. The caller saving/restoring instructions are +// inserted like: +// ** caller saving instructions +// other instructions inserted for the call by ColorCallArg +// CALL instruction +// other instructions inserted for the call ColorCallArg +// ** caller restoring instructions +//---------------------------------------------------------------------------- + +void +PhyRegAlloc::insertCallerSavingCode(std::vector<MachineInstr*> &instrnsBefore, + std::vector<MachineInstr*> &instrnsAfter, + MachineInstr *CallMI, + const BasicBlock *BB) +{ + assert(TM.getInstrInfo().isCall(CallMI->getOpCode())); + + // has set to record which registers were saved/restored + // + hash_set<unsigned> PushedRegSet; + + CallArgsDescriptor* argDesc = CallArgsDescriptor::get(CallMI); + + // if the call is to a instrumentation function, do not insert save and + // restore instructions the instrumentation function takes care of save + // restore for volatile regs. + // + // FIXME: this should be made general, not specific to the reoptimizer! + // + const Function *Callee = argDesc->getCallInst()->getCalledFunction(); + bool isLLVMFirstTrigger = Callee && Callee->getName() == "llvm_first_trigger"; + + // Now check if the call has a return value (using argDesc) and if so, + // find the LR of the TmpInstruction representing the return value register. + // (using the last or second-last *implicit operand* of the call MI). + // Insert it to to the PushedRegSet since we must not save that register + // and restore it after the call. + // We do this because, we look at the LV set *after* the instruction + // to determine, which LRs must be saved across calls. The return value + // of the call is live in this set - but we must not save/restore it. + // + if (const Value *origRetVal = argDesc->getReturnValue()) { + unsigned retValRefNum = (CallMI->getNumImplicitRefs() - + (argDesc->getIndirectFuncPtr()? 1 : 2)); + const TmpInstruction* tmpRetVal = + cast<TmpInstruction>(CallMI->getImplicitRef(retValRefNum)); + assert(tmpRetVal->getOperand(0) == origRetVal && + tmpRetVal->getType() == origRetVal->getType() && + "Wrong implicit ref?"); + LiveRange *RetValLR = LRI.getLiveRangeForValue(tmpRetVal); + assert(RetValLR && "No LR for RetValue of call"); + + if (! RetValLR->isMarkedForSpill()) + PushedRegSet.insert(MRI.getUnifiedRegNum(RetValLR->getRegClassID(), + RetValLR->getColor())); + } + + const ValueSet &LVSetAft = LVI->getLiveVarSetAfterMInst(CallMI, BB); + ValueSet::const_iterator LIt = LVSetAft.begin(); + + // for each live var in live variable set after machine inst + for( ; LIt != LVSetAft.end(); ++LIt) { + + // get the live range corresponding to live var + LiveRange *const LR = LRI.getLiveRangeForValue(*LIt); + + // LR can be null if it is a const since a const + // doesn't have a dominating def - see Assumptions above + if( LR ) { + + if(! LR->isMarkedForSpill()) { + + assert(LR->hasColor() && "LR is neither spilled nor colored?"); + unsigned RCID = LR->getRegClassID(); + unsigned Color = LR->getColor(); + + if (MRI.isRegVolatile(RCID, Color) ) { + + //if the function is special LLVM function, + //And the register is not modified by call, don't save and restore + if (isLLVMFirstTrigger && !MRI.modifiedByCall(RCID, Color)) + continue; + + // if the value is in both LV sets (i.e., live before and after + // the call machine instruction) + + unsigned Reg = MRI.getUnifiedRegNum(RCID, Color); + + if( PushedRegSet.find(Reg) == PushedRegSet.end() ) { + + // if we haven't already pushed that register + + unsigned RegType = MRI.getRegTypeForLR(LR); + + // Now get two instructions - to push on stack and pop from stack + // and add them to InstrnsBefore and InstrnsAfter of the + // call instruction + // + int StackOff = + MF.getInfo()->pushTempValue(MRI.getSpilledRegSize(RegType)); + + //---- Insert code for pushing the reg on stack ---------- + + std::vector<MachineInstr*> AdIBef, AdIAft; + + // We may need a scratch register to copy the saved value + // to/from memory. This may itself have to insert code to + // free up a scratch register. Any such code should go before + // the save code. The scratch register, if any, is by default + // temporary and not "used" by the instruction unless the + // copy code itself decides to keep the value in the scratch reg. + int scratchRegType = -1; + int scratchReg = -1; + if (MRI.regTypeNeedsScratchReg(RegType, scratchRegType)) + { // Find a register not live in the LVSet before CallMI + const ValueSet &LVSetBef = + LVI->getLiveVarSetBeforeMInst(CallMI, BB); + scratchReg = getUsableUniRegAtMI(scratchRegType, &LVSetBef, + CallMI, AdIBef, AdIAft); + assert(scratchReg != MRI.getInvalidRegNum()); + } + + if (AdIBef.size() > 0) + instrnsBefore.insert(instrnsBefore.end(), + AdIBef.begin(), AdIBef.end()); + + MRI.cpReg2MemMI(instrnsBefore, Reg, MRI.getFramePointer(), + StackOff, RegType, scratchReg); + + if (AdIAft.size() > 0) + instrnsBefore.insert(instrnsBefore.end(), + AdIAft.begin(), AdIAft.end()); + + //---- Insert code for popping the reg from the stack ---------- + + AdIBef.clear(); + AdIAft.clear(); + + // We may need a scratch register to copy the saved value + // from memory. This may itself have to insert code to + // free up a scratch register. Any such code should go + // after the save code. As above, scratch is not marked "used". + // + scratchRegType = -1; + scratchReg = -1; + if (MRI.regTypeNeedsScratchReg(RegType, scratchRegType)) + { // Find a register not live in the LVSet after CallMI + scratchReg = getUsableUniRegAtMI(scratchRegType, &LVSetAft, + CallMI, AdIBef, AdIAft); + assert(scratchReg != MRI.getInvalidRegNum()); + } + + if (AdIBef.size() > 0) + instrnsAfter.insert(instrnsAfter.end(), + AdIBef.begin(), AdIBef.end()); + + MRI.cpMem2RegMI(instrnsAfter, MRI.getFramePointer(), StackOff, + Reg, RegType, scratchReg); + + if (AdIAft.size() > 0) + instrnsAfter.insert(instrnsAfter.end(), + AdIAft.begin(), AdIAft.end()); + + PushedRegSet.insert(Reg); + + if(DEBUG_RA) { + std::cerr << "\nFor call inst:" << *CallMI; + std::cerr << " -inserted caller saving instrs: Before:\n\t "; + for_each(instrnsBefore.begin(), instrnsBefore.end(), + std::mem_fun(&MachineInstr::dump)); + std::cerr << " -and After:\n\t "; + for_each(instrnsAfter.begin(), instrnsAfter.end(), + std::mem_fun(&MachineInstr::dump)); + } + } // if not already pushed + + } // if LR has a volatile color + + } // if LR has color + + } // if there is a LR for Var + + } // for each value in the LV set after instruction +} + + //---------------------------------------------------------------------------- // We can use the following method to get a temporary register to be used // BEFORE any given machine instruction. If there is a register available, @@ -837,24 +1073,28 @@ int PhyRegAlloc::getUsableUniRegAtMI(const int RegType, return RegU; } + //---------------------------------------------------------------------------- -// This method is called to get a new unused register that can be used to -// accomodate a spilled value. -// This method may be called several times for a single machine instruction -// if it contains many spilled operands. Each time it is called, it finds -// a register which is not live at that instruction and also which is not -// used by other spilled operands of the same instruction. -// Return register number is relative to the register class. NOT -// unified number +// This method is called to get a new unused register that can be used +// to accomodate a temporary value. This method may be called several times +// for a single machine instruction. Each time it is called, it finds a +// register which is not live at that instruction and also which is not used +// by other spilled operands of the same instruction. Return register number +// is relative to the register class, NOT the unified number. //---------------------------------------------------------------------------- int PhyRegAlloc::getUnusedUniRegAtMI(RegClass *RC, const int RegType, - const MachineInstr *MInst, - const ValueSet *LVSetBef) { + const MachineInstr *MInst, + const ValueSet* LVSetBef) { RC->clearColorsUsed(); // Reset array - + + if (LVSetBef == NULL) { + LVSetBef = &LVI->getLiveVarSetBeforeMInst(MInst); + assert(LVSetBef != NULL && "Unable to get live-var set before MInst?"); + } + ValueSet::const_iterator LIt = LVSetBef->begin(); // for each live var in live variable set after machine inst @@ -953,11 +1193,16 @@ void PhyRegAlloc::setRelRegsUsedByThisInst(RegClass *RC, void PhyRegAlloc::move2DelayedInstr(const MachineInstr *OrigMI, const MachineInstr *DelayedMI) { + if (DEBUG_RA) { + cerr << "\nRegAlloc: Moved InstrnsAfter for: " << *OrigMI; + cerr << " to last delay slot instrn: " << *DelayedMI; + } + // "added after" instructions of the original instr std::vector<MachineInstr *> &OrigAft = AddedInstrMap[OrigMI].InstrnsAfter; // "added after" instructions of the delayed instr - std::vector<MachineInstr *> &DelayedAft =AddedInstrMap[DelayedMI].InstrnsAfter; + std::vector<MachineInstr *> &DelayedAft=AddedInstrMap[DelayedMI].InstrnsAfter; // go thru all the "added after instructions" of the original instruction // and append them to the "added after instructions" of the delayed @@ -1063,7 +1308,8 @@ void PhyRegAlloc::printMachineCode() //---------------------------------------------------------------------------- void PhyRegAlloc::colorIncomingArgs() { - MRI.colorMethodArgs(Fn, LRI, &AddedInstrAtEntry); + MRI.colorMethodArgs(Fn, LRI, AddedInstrAtEntry.InstrnsBefore, + AddedInstrAtEntry.InstrnsAfter); } @@ -1138,7 +1384,6 @@ void PhyRegAlloc::allocateStackSpace4SpilledLRs() { } - //---------------------------------------------------------------------------- // The entry pont to Register Allocation //---------------------------------------------------------------------------- |
