2 //***************************************************************************
7 // This file defines stuff that is to be private to the Sparc
8 // backend, but is shared among different portions of the backend.
9 //**************************************************************************/
12 #ifndef SPARC_INTERNALS_H
13 #define SPARC_INTERNALS_H
16 #include "SparcRegClassInfo.h"
17 #include "llvm/Target/TargetMachine.h"
18 #include "llvm/Target/MachineInstrInfo.h"
19 #include "llvm/Target/MachineSchedInfo.h"
20 #include "llvm/Target/MachineFrameInfo.h"
21 #include "llvm/Target/MachineCacheInfo.h"
22 #include "llvm/CodeGen/RegClass.h"
23 #include "llvm/Type.h"
25 #include <sys/types.h>
29 // OpCodeMask definitions for the Sparc V9
31 const OpCodeMask Immed = 0x00002000; // immed or reg operand?
32 const OpCodeMask Annul = 0x20000000; // annul delay instr?
33 const OpCodeMask PredictTaken = 0x00080000; // predict branch taken?
36 enum SparcInstrSchedClass {
37 SPARC_NONE, /* Instructions with no scheduling restrictions */
38 SPARC_IEUN, /* Integer class that can use IEU0 or IEU1 */
39 SPARC_IEU0, /* Integer class IEU0 */
40 SPARC_IEU1, /* Integer class IEU1 */
41 SPARC_FPM, /* FP Multiply or Divide instructions */
42 SPARC_FPA, /* All other FP instructions */
43 SPARC_CTI, /* Control-transfer instructions */
44 SPARC_LD, /* Load instructions */
45 SPARC_ST, /* Store instructions */
46 SPARC_SINGLE, /* Instructions that must issue by themselves */
48 SPARC_INV, /* This should stay at the end for the next value */
49 SPARC_NUM_SCHED_CLASSES = SPARC_INV
53 //---------------------------------------------------------------------------
54 // enum SparcMachineOpCode.
55 // const MachineInstrDescriptor SparcMachineInstrDesc[]
58 // Description of UltraSparc machine instructions.
60 //---------------------------------------------------------------------------
62 enum SparcMachineOpCode {
63 #define I(ENUM, OPCODESTRING, NUMOPERANDS, RESULTPOS, MAXIMM, IMMSE, \
64 NUMDELAYSLOTS, LATENCY, SCHEDCLASS, INSTFLAGS) \
66 #include "SparcInstr.def"
68 // End-of-array marker
70 NUM_REAL_OPCODES = PHI, // number of valid opcodes
71 NUM_TOTAL_OPCODES = INVALID_OPCODE
75 // Array of machine instruction descriptions...
76 extern const MachineInstrDescriptor SparcMachineInstrDesc[];
79 //---------------------------------------------------------------------------
80 // class UltraSparcInstrInfo
83 // Information about individual instructions.
84 // Most information is stored in the SparcMachineInstrDesc array above.
85 // Other information is computed on demand, and most such functions
86 // default to member functions in base class MachineInstrInfo.
87 //---------------------------------------------------------------------------
89 class UltraSparcInstrInfo : public MachineInstrInfo {
91 /*ctor*/ UltraSparcInstrInfo(const TargetMachine& tgt);
94 // All immediate constants are in position 0 except the
95 // store instructions.
97 virtual int getImmmedConstantPos(MachineOpCode opCode) const {
99 if (this->maxImmedConstant(opCode, ignore) != 0)
101 assert(! this->isStore((MachineOpCode) STB - 1)); // first store is STB
102 assert(! this->isStore((MachineOpCode) STD + 1)); // last store is STD
103 return (opCode >= STB || opCode <= STD)? 2 : 1;
109 virtual bool hasResultInterlock (MachineOpCode opCode) const
111 // All UltraSPARC instructions have interlocks (note that delay slots
112 // are not considered here).
113 // However, instructions that use the result of an FCMP produce a
114 // 9-cycle stall if they are issued less than 3 cycles after the FCMP.
115 // Force the compiler to insert a software interlock (i.e., gap of
116 // 2 other groups, including NOPs if necessary).
117 return (opCode == FCMPS || opCode == FCMPD || opCode == FCMPQ);
120 //-------------------------------------------------------------------------
121 // Code generation support for creating individual machine instructions
122 //-------------------------------------------------------------------------
124 // Create an instruction sequence to put the constant `val' into
125 // the virtual register `dest'. The generated instructions are
126 // returned in `minstrVec'. Any temporary registers (TmpInstruction)
127 // created are returned in `tempVec'.
129 virtual void CreateCodeToLoadConst(Value* val,
131 vector<MachineInstr*>& minstrVec,
132 vector<TmpInstruction*>& tempVec) const;
135 // Create an instruction sequence to copy an integer value `val'
136 // to a floating point value `dest' by copying to memory and back.
137 // val must be an integral type. dest must be a Float or Double.
138 // The generated instructions are returned in `minstrVec'.
139 // Any temp. registers (TmpInstruction) created are returned in `tempVec'.
141 virtual void CreateCodeToCopyIntToFloat(Method* method,
144 vector<MachineInstr*>& minstrVec,
145 vector<TmpInstruction*>& tempVec,
146 TargetMachine& target) const;
148 // Similarly, create an instruction sequence to copy an FP value
149 // `val' to an integer value `dest' by copying to memory and back.
150 // See the previous function for information about return values.
152 virtual void CreateCodeToCopyFloatToInt(Method* method,
155 vector<MachineInstr*>& minstrVec,
156 vector<TmpInstruction*>& tempVec,
157 TargetMachine& target) const;
159 // create copy instruction(s)
161 CreateCopyInstructionsByType(const TargetMachine& target,
164 vector<MachineInstr*>& minstrVec) const;
170 //----------------------------------------------------------------------------
171 // class UltraSparcRegInfo
173 // This class implements the virtual class MachineRegInfo for Sparc.
175 //----------------------------------------------------------------------------
183 class UltraSparcRegInfo : public MachineRegInfo
187 // The actual register classes in the Sparc
190 IntRegClassID, // Integer
191 FloatRegClassID, // Float (both single/double)
192 IntCCRegClassID, // Int Condition Code
193 FloatCCRegClassID // Float Condition code
197 // Type of registers available in Sparc. There can be several reg types
198 // in the same class. For instace, the float reg class has Single/Double
209 // **** WARNING: If the above enum order is changed, also modify
210 // getRegisterClassOfValue method below since it assumes this particular
211 // order for efficiency.
214 // reverse pointer to get info about the ultra sparc machine
216 const UltraSparc *const UltraSparcInfo;
218 // Number of registers used for passing int args (usually 6: %o0 - %o5)
220 unsigned const NumOfIntArgRegs;
222 // Number of registers used for passing float args (usually 32: %f0 - %f31)
224 unsigned const NumOfFloatArgRegs;
226 // An out of bound register number that can be used to initialize register
227 // numbers. Useful for error detection.
229 int const InvalidRegNum;
232 // ======================== Private Methods =============================
234 // The following methods are used to color special live ranges (e.g.
235 // method args and return values etc.) with specific hardware registers
236 // as required. See SparcRegInfo.cpp for the implementation.
238 void setCallOrRetArgCol(LiveRange *const LR, const unsigned RegNo,
239 const MachineInstr *MI,AddedInstrMapType &AIMap)const;
241 MachineInstr * getCopy2RegMI(const Value *SrcVal, const unsigned Reg,
242 unsigned RegClassID) const ;
244 void suggestReg4RetAddr(const MachineInstr * RetMI,
245 LiveRangeInfo& LRI) const;
247 void suggestReg4CallAddr(const MachineInstr * CallMI, LiveRangeInfo& LRI,
248 vector<RegClass *> RCList) const;
252 // The following methods are used to find the addresses etc. contained
253 // in specail machine instructions like CALL/RET
255 Value *getValue4ReturnAddr( const MachineInstr * MInst ) const ;
256 const Value *getCallInstRetAddr(const MachineInstr *CallMI) const;
257 const unsigned getCallInstNumArgs(const MachineInstr *CallMI) const;
260 // The following 3 methods are used to find the RegType (see enum above)
261 // of a LiveRange, Value and using the unified RegClassID
263 int getRegType(const LiveRange *const LR) const {
267 switch( (LR->getRegClass())->getID() ) {
269 case IntRegClassID: return IntRegType;
271 case FloatRegClassID:
272 Typ = LR->getTypeID();
273 if( Typ == Type::FloatTyID )
274 return FPSingleRegType;
275 else if( Typ == Type::DoubleTyID )
276 return FPDoubleRegType;
277 else assert(0 && "Unknown type in FloatRegClass");
279 case IntCCRegClassID: return IntCCRegType;
281 case FloatCCRegClassID: return FloatCCRegType ;
283 default: assert( 0 && "Unknown reg class ID");
289 int getRegType(const Value *const Val) const {
293 switch( getRegClassIDOfValue(Val) ) {
295 case IntRegClassID: return IntRegType;
297 case FloatRegClassID:
298 Typ = (Val->getType())->getPrimitiveID();
299 if( Typ == Type::FloatTyID )
300 return FPSingleRegType;
301 else if( Typ == Type::DoubleTyID )
302 return FPDoubleRegType;
303 else assert(0 && "Unknown type in FloatRegClass");
305 case IntCCRegClassID: return IntCCRegType;
307 case FloatCCRegClassID: return FloatCCRegType ;
309 default: assert( 0 && "Unknown reg class ID");
316 int getRegType(int reg) const {
319 else if ( reg < (32 + 32) )
320 return FPSingleRegType;
321 else if ( reg < (64 + 32) )
322 return FPDoubleRegType;
323 else if( reg < (64+32+4) )
324 return FloatCCRegType;
325 else if( reg < (64+32+4+2) )
328 assert(0 && "Invalid register number in getRegType");
334 // The following methods are used to generate copy instructions to move
335 // data between condition code registers
337 MachineInstr * cpCCR2IntMI(const unsigned IntReg) const;
338 MachineInstr * cpInt2CCRMI(const unsigned IntReg) const;
340 // Used to generate a copy instruction based on the register class of
343 MachineInstr * cpValue2RegMI(Value * Val, const unsigned DestReg,
344 const int RegType) const;
347 // The following 2 methods are used to order the instructions addeed by
348 // the register allocator in association with method calling. See
349 // SparcRegInfo.cpp for more details
351 void moveInst2OrdVec(vector<MachineInstr *> &OrdVec, MachineInstr *UnordInst,
352 PhyRegAlloc &PRA ) const;
354 void OrderAddedInstrns( vector<MachineInstr *> &UnordVec,
355 vector<MachineInstr *> &OrdVec,
356 PhyRegAlloc &PRA) const;
359 // To find whether a particular call is to a var arg method
361 bool isVarArgCall(const MachineInstr *CallMI) const;
369 UltraSparcRegInfo(const TargetMachine& tgt ) :
371 UltraSparcInfo(& (const UltraSparc&) tgt),
373 NumOfFloatArgRegs(32),
374 InvalidRegNum(1000) {
376 MachineRegClassArr.push_back( new SparcIntRegClass(IntRegClassID) );
377 MachineRegClassArr.push_back( new SparcFloatRegClass(FloatRegClassID) );
378 MachineRegClassArr.push_back( new SparcIntCCRegClass(IntCCRegClassID) );
379 MachineRegClassArr.push_back( new SparcFloatCCRegClass(FloatCCRegClassID));
381 assert( SparcFloatRegOrder::StartOfNonVolatileRegs == 32 &&
382 "32 Float regs are used for float arg passing");
387 ~UltraSparcRegInfo(void) { } // empty destructor
390 // To get complete machine information structure using the machine register
393 inline const UltraSparc & getUltraSparcInfo() const {
394 return *UltraSparcInfo;
398 // To find the register class of a Value
400 inline unsigned getRegClassIDOfValue (const Value *const Val,
401 bool isCCReg = false) const {
403 Type::PrimitiveID ty = (Val->getType())->getPrimitiveID();
407 if( (ty && ty <= Type::LongTyID) || (ty == Type::LabelTyID) ||
408 (ty == Type::MethodTyID) || (ty == Type::PointerTyID) )
409 res = IntRegClassID; // sparc int reg (ty=0: void)
410 else if( ty <= Type::DoubleTyID)
411 res = FloatRegClassID; // sparc float reg class
413 cerr << "TypeID: " << ty << endl;
414 assert(0 && "Cannot resolve register class for type");
419 return res + 2; // corresponidng condition code regiser
426 // returns the register that contains always zero
427 // this is the unified register number
429 inline int getZeroRegNum() const { return SparcIntRegOrder::g0; }
431 // returns the reg used for pushing the address when a method is called.
432 // This can be used for other purposes between calls
434 unsigned getCallAddressReg() const { return SparcIntRegOrder::o7; }
436 // Returns the register containing the return address.
437 // It should be made sure that this register contains the return
438 // value when a return instruction is reached.
440 unsigned getReturnAddressReg() const { return SparcIntRegOrder::i7; }
444 // The following methods are used to color special live ranges (e.g.
445 // method args and return values etc.) with specific hardware registers
446 // as required. See SparcRegInfo.cpp for the implementation for Sparc.
448 void suggestRegs4MethodArgs(const Method *const Meth,
449 LiveRangeInfo& LRI) const;
451 void suggestRegs4CallArgs(const MachineInstr *const CallMI,
452 LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
454 void suggestReg4RetValue(const MachineInstr *const RetMI,
455 LiveRangeInfo& LRI ) const;
458 void colorMethodArgs(const Method *const Meth, LiveRangeInfo& LRI,
459 AddedInstrns *const FirstAI) const;
461 void colorCallArgs(const MachineInstr *const CallMI, LiveRangeInfo& LRI,
462 AddedInstrns *const CallAI, PhyRegAlloc &PRA,
463 const BasicBlock *BB) const;
465 void colorRetValue(const MachineInstr *const RetI, LiveRangeInfo& LRI,
466 AddedInstrns *const RetAI) const;
470 // method used for printing a register for debugging purposes
472 static void printReg(const LiveRange *const LR) ;
474 // this method provides a unique number for each register
476 inline int getUnifiedRegNum(int RegClassID, int reg) const {
478 if( RegClassID == IntRegClassID && reg < 32 )
480 else if ( RegClassID == FloatRegClassID && reg < 64)
481 return reg + 32; // we have 32 int regs
482 else if( RegClassID == FloatCCRegClassID && reg < 4)
483 return reg + 32 + 64; // 32 int, 64 float
484 else if( RegClassID == IntCCRegClassID )
485 return 4+ 32 + 64; // only int cc reg
486 else if (reg==InvalidRegNum)
487 return InvalidRegNum;
489 assert(0 && "Invalid register class or reg number");
493 // given the unified register number, this gives the name
494 // for generating assembly code or debugging.
496 inline const string getUnifiedRegName(int reg) const {
498 return SparcIntRegOrder::getRegName(reg);
499 else if ( reg < (64 + 32) )
500 return SparcFloatRegOrder::getRegName( reg - 32);
501 else if( reg < (64+32+4) )
502 return SparcFloatCCRegOrder::getRegName( reg -32 - 64);
503 else if( reg < (64+32+4+2) ) // two names: %xcc and %ccr
504 return SparcIntCCRegOrder::getRegName( reg -32 - 64 - 4);
505 else if (reg== InvalidRegNum) //****** TODO: Remove */
508 assert(0 && "Invalid register number");
514 // The fllowing methods are used by instruction selection
516 inline unsigned int getRegNumInCallersWindow(int reg) {
517 if (reg == InvalidRegNum || reg >= 32)
519 return SparcIntRegOrder::getRegNumInCallersWindow(reg);
522 inline bool mustBeRemappedInCallersWindow(int reg) {
523 return (reg != InvalidRegNum && reg < 32);
528 // returns the # of bytes of stack space allocated for each register
529 // type. For Sparc, currently we allocate 8 bytes on stack for all
530 // register types. We can optimize this later if necessary to save stack
531 // space (However, should make sure that stack alignment is correct)
533 inline int getSpilledRegSize(const int RegType) const {
538 // To obtain the return value contained in a CALL machine instruction
540 const Value * getCallInstRetVal(const MachineInstr *CallMI) const;
543 // The following methods are used to generate "copy" machine instructions
544 // for an architecture.
546 MachineInstr * cpReg2RegMI(const unsigned SrcReg, const unsigned DestReg,
547 const int RegType) const;
549 MachineInstr * cpReg2MemMI(const unsigned SrcReg, const unsigned DestPtrReg,
550 const int Offset, const int RegType) const;
552 MachineInstr * cpMem2RegMI(const unsigned SrcPtrReg, const int Offset,
553 const unsigned DestReg, const int RegType) const;
555 MachineInstr* cpValue2Value(Value *Src, Value *Dest) const;
558 // To see whether a register is a volatile (i.e., whehter it must be
559 // preserved acorss calls)
561 inline bool isRegVolatile(const int RegClassID, const int Reg) const {
562 return (MachineRegClassArr[RegClassID])->isRegVolatile(Reg);
566 inline unsigned getFramePointer() const {
567 return SparcIntRegOrder::i6;
570 inline unsigned getStackPointer() const {
571 return SparcIntRegOrder::o6;
574 inline int getInvalidRegNum() const {
575 return InvalidRegNum;
580 // This method inserts the caller saving code for call instructions
582 void insertCallerSavingCode(const MachineInstr *MInst,
583 const BasicBlock *BB, PhyRegAlloc &PRA ) const;
589 /*---------------------------------------------------------------------------
590 Scheduling guidelines for SPARC IIi:
592 I-Cache alignment rules (pg 326)
593 -- Align a branch target instruction so that it's entire group is within
594 the same cache line (may be 1-4 instructions).
595 ** Don't let a branch that is predicted taken be the last instruction
596 on an I-cache line: delay slot will need an entire line to be fetched
597 -- Make a FP instruction or a branch be the 4th instruction in a group.
598 For branches, there are tradeoffs in reordering to make this happen
600 ** Don't put a branch in a group that crosses a 32-byte boundary!
601 An artificial branch is inserted after every 32 bytes, and having
602 another branch will force the group to be broken into 2 groups.
605 -- Don't let a loop span two memory pages, if possible
607 Branch prediction performance:
608 -- Don't make the branch in a delay slot the target of a branch
609 -- Try not to have 2 predicted branches within a group of 4 instructions
610 (because each such group has a single branch target field).
611 -- Try to align branches in slots 0, 2, 4 or 6 of a cache line (to avoid
612 the wrong prediction bits being used in some cases).
614 D-Cache timing constraints:
615 -- Signed int loads of less than 64 bits have 3 cycle latency, not 2
616 -- All other loads that hit in D-Cache have 2 cycle latency
617 -- All loads are returned IN ORDER, so a D-Cache miss will delay a later hit
618 -- Mis-aligned loads or stores cause a trap. In particular, replace
619 mis-aligned FP double precision l/s with 2 single-precision l/s.
620 -- Simulations of integer codes show increase in avg. group size of
621 33% when code (including esp. non-faulting loads) is moved across
622 one branch, and 50% across 2 branches.
624 E-Cache timing constraints:
625 -- Scheduling for E-cache (D-Cache misses) is effective (due to load buffering)
627 Store buffer timing constraints:
628 -- Stores can be executed in same cycle as instruction producing the value
629 -- Stores are buffered and have lower priority for E-cache until
630 highwater mark is reached in the store buffer (5 stores)
632 Pipeline constraints:
633 -- Shifts can only use IEU0.
634 -- CC setting instructions can only use IEU1.
635 -- Several other instructions must only use IEU1:
636 EDGE(?), ARRAY(?), CALL, JMPL, BPr, PST, and FCMP.
637 -- Two instructions cannot store to the same register file in a single cycle
638 (single write port per file).
640 Issue and grouping constraints:
641 -- FP and branch instructions must use slot 4.
642 -- Shift instructions cannot be grouped with other IEU0-specific instructions.
643 -- CC setting instructions cannot be grouped with other IEU1-specific instrs.
644 -- Several instructions must be issued in a single-instruction group:
645 MOVcc or MOVr, MULs/x and DIVs/x, SAVE/RESTORE, many others
646 -- A CALL or JMPL breaks a group, ie, is not combined with subsequent instrs.
650 Branch delay slot scheduling rules:
651 -- A CTI couple (two back-to-back CTI instructions in the dynamic stream)
652 has a 9-instruction penalty: the entire pipeline is flushed when the
653 second instruction reaches stage 9 (W-Writeback).
654 -- Avoid putting multicycle instructions, and instructions that may cause
655 load misses, in the delay slot of an annulling branch.
656 -- Avoid putting WR, SAVE..., RESTORE and RETURN instructions in the
657 delay slot of an annulling branch.
659 *--------------------------------------------------------------------------- */
661 //---------------------------------------------------------------------------
662 // List of CPUResources for UltraSPARC IIi.
663 //---------------------------------------------------------------------------
665 const CPUResource AllIssueSlots( "All Instr Slots", 4);
666 const CPUResource IntIssueSlots( "Int Instr Slots", 3);
667 const CPUResource First3IssueSlots("Instr Slots 0-3", 3);
668 const CPUResource LSIssueSlots( "Load-Store Instr Slot", 1);
669 const CPUResource CTIIssueSlots( "Ctrl Transfer Instr Slot", 1);
670 const CPUResource FPAIssueSlots( "Int Instr Slot 1", 1);
671 const CPUResource FPMIssueSlots( "Int Instr Slot 1", 1);
673 // IEUN instructions can use either Alu and should use IAluN.
674 // IEU0 instructions must use Alu 1 and should use both IAluN and IAlu0.
675 // IEU1 instructions must use Alu 2 and should use both IAluN and IAlu1.
676 const CPUResource IAluN("Int ALU 1or2", 2);
677 const CPUResource IAlu0("Int ALU 1", 1);
678 const CPUResource IAlu1("Int ALU 2", 1);
680 const CPUResource LSAluC1("Load/Store Unit Addr Cycle", 1);
681 const CPUResource LSAluC2("Load/Store Unit Issue Cycle", 1);
682 const CPUResource LdReturn("Load Return Unit", 1);
684 const CPUResource FPMAluC1("FP Mul/Div Alu Cycle 1", 1);
685 const CPUResource FPMAluC2("FP Mul/Div Alu Cycle 2", 1);
686 const CPUResource FPMAluC3("FP Mul/Div Alu Cycle 3", 1);
688 const CPUResource FPAAluC1("FP Other Alu Cycle 1", 1);
689 const CPUResource FPAAluC2("FP Other Alu Cycle 2", 1);
690 const CPUResource FPAAluC3("FP Other Alu Cycle 3", 1);
692 const CPUResource IRegReadPorts("Int Reg ReadPorts", INT_MAX); // CHECK
693 const CPUResource IRegWritePorts("Int Reg WritePorts", 2); // CHECK
694 const CPUResource FPRegReadPorts("FP Reg Read Ports", INT_MAX); // CHECK
695 const CPUResource FPRegWritePorts("FP Reg Write Ports", 1); // CHECK
697 const CPUResource CTIDelayCycle( "CTI delay cycle", 1);
698 const CPUResource FCMPDelayCycle("FCMP delay cycle", 1);
701 //---------------------------------------------------------------------------
702 // const InstrClassRUsage SparcRUsageDesc[]
705 // Resource usage information for instruction in each scheduling class.
706 // The InstrRUsage Objects for individual classes are specified first.
707 // Note that fetch and decode are decoupled from the execution pipelines
708 // via an instr buffer, so they are not included in the cycles below.
709 //---------------------------------------------------------------------------
711 const InstrClassRUsage NoneClassRUsage = {
716 /* isSingleIssue */ false,
717 /* breaksGroup */ false,
721 /* feasibleSlots[] */ { 0, 1, 2, 3 },
735 const InstrClassRUsage IEUNClassRUsage = {
740 /* isSingleIssue */ false,
741 /* breaksGroup */ false,
745 /* feasibleSlots[] */ { 0, 1, 2 },
749 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
750 { IntIssueSlots.rid, 0, 1 },
751 /*Cycle E */ { IAluN.rid, 1, 1 },
756 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
760 const InstrClassRUsage IEU0ClassRUsage = {
765 /* isSingleIssue */ false,
766 /* breaksGroup */ false,
770 /* feasibleSlots[] */ { 0, 1, 2 },
774 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
775 { IntIssueSlots.rid, 0, 1 },
776 /*Cycle E */ { IAluN.rid, 1, 1 },
782 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
786 const InstrClassRUsage IEU1ClassRUsage = {
791 /* isSingleIssue */ false,
792 /* breaksGroup */ false,
796 /* feasibleSlots[] */ { 0, 1, 2 },
800 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
801 { IntIssueSlots.rid, 0, 1 },
802 /*Cycle E */ { IAluN.rid, 1, 1 },
808 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
812 const InstrClassRUsage FPMClassRUsage = {
817 /* isSingleIssue */ false,
818 /* breaksGroup */ false,
822 /* feasibleSlots[] */ { 0, 1, 2, 3 },
826 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
827 { FPMIssueSlots.rid, 0, 1 },
828 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
829 /*Cycle C */ { FPMAluC1.rid, 2, 1 },
830 /*Cycle N1*/ { FPMAluC2.rid, 3, 1 },
831 /*Cycle N1*/ { FPMAluC3.rid, 4, 1 },
833 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
837 const InstrClassRUsage FPAClassRUsage = {
842 /* isSingleIssue */ false,
843 /* breaksGroup */ false,
847 /* feasibleSlots[] */ { 0, 1, 2, 3 },
851 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
852 { FPAIssueSlots.rid, 0, 1 },
853 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
854 /*Cycle C */ { FPAAluC1.rid, 2, 1 },
855 /*Cycle N1*/ { FPAAluC2.rid, 3, 1 },
856 /*Cycle N1*/ { FPAAluC3.rid, 4, 1 },
858 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
862 const InstrClassRUsage LDClassRUsage = {
867 /* isSingleIssue */ false,
868 /* breaksGroup */ false,
872 /* feasibleSlots[] */ { 0, 1, 2, },
876 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
877 { First3IssueSlots.rid, 0, 1 },
878 { LSIssueSlots.rid, 0, 1 },
879 /*Cycle E */ { LSAluC1.rid, 1, 1 },
880 /*Cycle C */ { LSAluC2.rid, 2, 1 },
881 { LdReturn.rid, 2, 1 },
885 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
889 const InstrClassRUsage STClassRUsage = {
894 /* isSingleIssue */ false,
895 /* breaksGroup */ false,
899 /* feasibleSlots[] */ { 0, 1, 2 },
903 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
904 { First3IssueSlots.rid, 0, 1 },
905 { LSIssueSlots.rid, 0, 1 },
906 /*Cycle E */ { LSAluC1.rid, 1, 1 },
907 /*Cycle C */ { LSAluC2.rid, 2, 1 }
915 const InstrClassRUsage CTIClassRUsage = {
920 /* isSingleIssue */ false,
921 /* breaksGroup */ false,
925 /* feasibleSlots[] */ { 0, 1, 2, 3 },
929 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
930 { CTIIssueSlots.rid, 0, 1 },
931 /*Cycle E */ { IAlu0.rid, 1, 1 },
932 /*Cycles E-C */ { CTIDelayCycle.rid, 1, 2 }
941 const InstrClassRUsage SingleClassRUsage = {
946 /* isSingleIssue */ true,
947 /* breaksGroup */ false,
951 /* feasibleSlots[] */ { 0 },
955 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
956 { AllIssueSlots.rid, 0, 1 },
957 { AllIssueSlots.rid, 0, 1 },
958 { AllIssueSlots.rid, 0, 1 },
959 /*Cycle E */ { IAlu0.rid, 1, 1 }
969 const InstrClassRUsage SparcRUsageDesc[] = {
983 //---------------------------------------------------------------------------
984 // const InstrIssueDelta SparcInstrIssueDeltas[]
987 // Changes to issue restrictions information in InstrClassRUsage for
988 // instructions that differ from other instructions in their class.
989 //---------------------------------------------------------------------------
991 const InstrIssueDelta SparcInstrIssueDeltas[] = {
993 // opCode, isSingleIssue, breaksGroup, numBubbles
995 // Special cases for single-issue only
996 // Other single issue cases are below.
997 //{ LDDA, true, true, 0 },
998 //{ STDA, true, true, 0 },
999 //{ LDDF, true, true, 0 },
1000 //{ LDDFA, true, true, 0 },
1001 { ADDC, true, true, 0 },
1002 { ADDCcc, true, true, 0 },
1003 { SUBC, true, true, 0 },
1004 { SUBCcc, true, true, 0 },
1005 //{ LDSTUB, true, true, 0 },
1006 //{ SWAP, true, true, 0 },
1007 //{ SWAPA, true, true, 0 },
1008 //{ CAS, true, true, 0 },
1009 //{ CASA, true, true, 0 },
1010 //{ CASX, true, true, 0 },
1011 //{ CASXA, true, true, 0 },
1012 //{ LDFSR, true, true, 0 },
1013 //{ LDFSRA, true, true, 0 },
1014 //{ LDXFSR, true, true, 0 },
1015 //{ LDXFSRA, true, true, 0 },
1016 //{ STFSR, true, true, 0 },
1017 //{ STFSRA, true, true, 0 },
1018 //{ STXFSR, true, true, 0 },
1019 //{ STXFSRA, true, true, 0 },
1020 //{ SAVED, true, true, 0 },
1021 //{ RESTORED, true, true, 0 },
1022 //{ FLUSH, true, true, 9 },
1023 //{ FLUSHW, true, true, 9 },
1024 //{ ALIGNADDR, true, true, 0 },
1025 { RETURN, true, true, 0 },
1026 //{ DONE, true, true, 0 },
1027 //{ RETRY, true, true, 0 },
1028 //{ TCC, true, true, 0 },
1029 //{ SHUTDOWN, true, true, 0 },
1031 // Special cases for breaking group *before*
1032 // CURRENTLY NOT SUPPORTED!
1033 { CALL, false, false, 0 },
1034 { JMPLCALL, false, false, 0 },
1035 { JMPLRET, false, false, 0 },
1037 // Special cases for breaking the group *after*
1038 { MULX, true, true, (4+34)/2 },
1039 { FDIVS, false, true, 0 },
1040 { FDIVD, false, true, 0 },
1041 { FDIVQ, false, true, 0 },
1042 { FSQRTS, false, true, 0 },
1043 { FSQRTD, false, true, 0 },
1044 { FSQRTQ, false, true, 0 },
1045 //{ FCMP{LE,GT,NE,EQ}, false, true, 0 },
1047 // Instructions that introduce bubbles
1048 //{ MULScc, true, true, 2 },
1049 //{ SMULcc, true, true, (4+18)/2 },
1050 //{ UMULcc, true, true, (4+19)/2 },
1051 { SDIVX, true, true, 68 },
1052 { UDIVX, true, true, 68 },
1053 //{ SDIVcc, true, true, 36 },
1054 //{ UDIVcc, true, true, 37 },
1055 { WRCCR, true, true, 4 },
1056 //{ WRPR, true, true, 4 },
1057 //{ RDCCR, true, true, 0 }, // no bubbles after, but see below
1058 //{ RDPR, true, true, 0 },
1062 //---------------------------------------------------------------------------
1063 // const InstrRUsageDelta SparcInstrUsageDeltas[]
1066 // Changes to resource usage information in InstrClassRUsage for
1067 // instructions that differ from other instructions in their class.
1068 //---------------------------------------------------------------------------
1070 const InstrRUsageDelta SparcInstrUsageDeltas[] = {
1072 // MachineOpCode, Resource, Start cycle, Num cycles
1075 // JMPL counts as a load/store instruction for issue!
1077 { JMPLCALL, LSIssueSlots.rid, 0, 1 },
1078 { JMPLRET, LSIssueSlots.rid, 0, 1 },
1081 // Many instructions cannot issue for the next 2 cycles after an FCMP
1082 // We model that with a fake resource FCMPDelayCycle.
1084 { FCMPS, FCMPDelayCycle.rid, 1, 3 },
1085 { FCMPD, FCMPDelayCycle.rid, 1, 3 },
1086 { FCMPQ, FCMPDelayCycle.rid, 1, 3 },
1088 { MULX, FCMPDelayCycle.rid, 1, 1 },
1089 { SDIVX, FCMPDelayCycle.rid, 1, 1 },
1090 { UDIVX, FCMPDelayCycle.rid, 1, 1 },
1091 //{ SMULcc, FCMPDelayCycle.rid, 1, 1 },
1092 //{ UMULcc, FCMPDelayCycle.rid, 1, 1 },
1093 //{ SDIVcc, FCMPDelayCycle.rid, 1, 1 },
1094 //{ UDIVcc, FCMPDelayCycle.rid, 1, 1 },
1095 { STD, FCMPDelayCycle.rid, 1, 1 },
1096 { FMOVRSZ, FCMPDelayCycle.rid, 1, 1 },
1097 { FMOVRSLEZ,FCMPDelayCycle.rid, 1, 1 },
1098 { FMOVRSLZ, FCMPDelayCycle.rid, 1, 1 },
1099 { FMOVRSNZ, FCMPDelayCycle.rid, 1, 1 },
1100 { FMOVRSGZ, FCMPDelayCycle.rid, 1, 1 },
1101 { FMOVRSGEZ,FCMPDelayCycle.rid, 1, 1 },
1104 // Some instructions are stalled in the GROUP stage if a CTI is in
1105 // the E or C stage. We model that with a fake resource CTIDelayCycle.
1107 { LDD, CTIDelayCycle.rid, 1, 1 },
1108 //{ LDDA, CTIDelayCycle.rid, 1, 1 },
1109 //{ LDDSTUB, CTIDelayCycle.rid, 1, 1 },
1110 //{ LDDSTUBA, CTIDelayCycle.rid, 1, 1 },
1111 //{ SWAP, CTIDelayCycle.rid, 1, 1 },
1112 //{ SWAPA, CTIDelayCycle.rid, 1, 1 },
1113 //{ CAS, CTIDelayCycle.rid, 1, 1 },
1114 //{ CASA, CTIDelayCycle.rid, 1, 1 },
1115 //{ CASX, CTIDelayCycle.rid, 1, 1 },
1116 //{ CASXA, CTIDelayCycle.rid, 1, 1 },
1119 // Signed int loads of less than dword size return data in cycle N1 (not C)
1120 // and put all loads in consecutive cycles into delayed load return mode.
1122 { LDSB, LdReturn.rid, 2, -1 },
1123 { LDSB, LdReturn.rid, 3, 1 },
1125 { LDSH, LdReturn.rid, 2, -1 },
1126 { LDSH, LdReturn.rid, 3, 1 },
1128 { LDSW, LdReturn.rid, 2, -1 },
1129 { LDSW, LdReturn.rid, 3, 1 },
1132 // RDPR from certain registers and RD from any register are not dispatchable
1133 // until four clocks after they reach the head of the instr. buffer.
1134 // Together with their single-issue requirement, this means all four issue
1135 // slots are effectively blocked for those cycles, plus the issue cycle.
1136 // This does not increase the latency of the instruction itself.
1138 { RDCCR, AllIssueSlots.rid, 0, 5 },
1139 { RDCCR, AllIssueSlots.rid, 0, 5 },
1140 { RDCCR, AllIssueSlots.rid, 0, 5 },
1141 { RDCCR, AllIssueSlots.rid, 0, 5 },
1143 #undef EXPLICIT_BUBBLES_NEEDED
1144 #ifdef EXPLICIT_BUBBLES_NEEDED
1146 // MULScc inserts one bubble.
1147 // This means it breaks the current group (captured in UltraSparcSchedInfo)
1148 // *and occupies all issue slots for the next cycle
1150 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
1151 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
1152 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
1153 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
1156 // SMULcc inserts between 4 and 18 bubbles, depending on #leading 0s in rs1.
1157 // We just model this with a simple average.
1159 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
1160 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
1161 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
1162 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
1164 // SMULcc inserts between 4 and 19 bubbles, depending on #leading 0s in rs1.
1165 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
1166 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
1167 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
1168 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
1171 // MULX inserts between 4 and 34 bubbles, depending on #leading 0s in rs1.
1173 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
1174 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
1175 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
1176 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
1179 // SDIVcc inserts 36 bubbles.
1181 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
1182 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
1183 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
1184 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
1186 // UDIVcc inserts 37 bubbles.
1187 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
1188 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
1189 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
1190 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
1193 // SDIVX inserts 68 bubbles.
1195 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1196 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1197 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1198 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1201 // UDIVX inserts 68 bubbles.
1203 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1204 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1205 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1206 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1209 // WR inserts 4 bubbles.
1211 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1212 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1213 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1214 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1217 // WRPR inserts 4 bubbles.
1219 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1220 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1221 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1222 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1225 // DONE inserts 9 bubbles.
1227 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1228 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1229 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1230 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1233 // RETRY inserts 9 bubbles.
1235 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1236 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1237 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1238 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1240 #endif /*EXPLICIT_BUBBLES_NEEDED */
1245 // Additional delays to be captured in code:
1246 // 1. RDPR from several state registers (page 349)
1247 // 2. RD from *any* register (page 349)
1248 // 3. Writes to TICK, PSTATE, TL registers and FLUSH{W} instr (page 349)
1249 // 4. Integer store can be in same group as instr producing value to store.
1250 // 5. BICC and BPICC can be in the same group as instr producing CC (pg 350)
1251 // 6. FMOVr cannot be in the same or next group as an IEU instr (pg 351).
1252 // 7. The second instr. of a CTI group inserts 9 bubbles (pg 351)
1253 // 8. WR{PR}, SVAE, SAVED, RESTORE, RESTORED, RETURN, RETRY, and DONE that
1254 // follow an annulling branch cannot be issued in the same group or in
1255 // the 3 groups following the branch.
1256 // 9. A predicted annulled load does not stall dependent instructions.
1257 // Other annulled delay slot instructions *do* stall dependents, so
1258 // nothing special needs to be done for them during scheduling.
1259 //10. Do not put a load use that may be annulled in the same group as the
1260 // branch. The group will stall until the load returns.
1261 //11. Single-prec. FP loads lock 2 registers, for dependency checking.
1264 // Additional delays we cannot or will not capture:
1265 // 1. If DCTI is last word of cache line, it is delayed until next line can be
1266 // fetched. Also, other DCTI alignment-related delays (pg 352)
1267 // 2. Load-after-store is delayed by 7 extra cycles if load hits in D-Cache.
1268 // Also, several other store-load and load-store conflicts (pg 358)
1269 // 3. MEMBAR, LD{X}FSR, LDD{A} and a bunch of other load stalls (pg 358)
1270 // 4. There can be at most 8 outstanding buffered store instructions
1271 // (including some others like MEMBAR, LDSTUB, CAS{AX}, and FLUSH)
1275 //---------------------------------------------------------------------------
1276 // class UltraSparcSchedInfo
1279 // Interface to instruction scheduling information for UltraSPARC.
1280 // The parameter values above are based on UltraSPARC IIi.
1281 //---------------------------------------------------------------------------
1284 class UltraSparcSchedInfo: public MachineSchedInfo {
1286 /*ctor*/ UltraSparcSchedInfo (const TargetMachine& tgt);
1287 /*dtor*/ virtual ~UltraSparcSchedInfo () {}
1289 virtual void initializeResources ();
1293 //---------------------------------------------------------------------------
1294 // class UltraSparcFrameInfo
1297 // Interface to stack frame layout info for the UltraSPARC.
1298 // Starting offsets for each area of the stack frame are aligned at
1299 // a multiple of getStackFrameSizeAlignment().
1300 //---------------------------------------------------------------------------
1302 class UltraSparcFrameInfo: public MachineFrameInfo {
1304 /*ctor*/ UltraSparcFrameInfo(const TargetMachine& tgt) : MachineFrameInfo(tgt) {}
1307 int getStackFrameSizeAlignment () const { return StackFrameSizeAlignment;}
1308 int getMinStackFrameSize () const { return MinStackFrameSize; }
1309 int getNumFixedOutgoingArgs () const { return NumFixedOutgoingArgs; }
1310 int getSizeOfEachArgOnStack () const { return SizeOfEachArgOnStack; }
1311 bool argsOnStackHaveFixedSize () const { return true; }
1314 // These methods compute offsets using the frame contents for a
1315 // particular method. The frame contents are obtained from the
1316 // MachineCodeInfoForMethod object for the given method.
1318 int getFirstIncomingArgOffset (MachineCodeForMethod& mcInfo,
1321 pos = true; // arguments area grows upwards
1322 return FirstIncomingArgOffsetFromFP;
1324 int getFirstOutgoingArgOffset (MachineCodeForMethod& mcInfo,
1327 pos = true; // arguments area grows upwards
1328 return FirstOutgoingArgOffsetFromSP;
1330 int getFirstOptionalOutgoingArgOffset(MachineCodeForMethod& mcInfo,
1333 pos = true; // arguments area grows upwards
1334 return FirstOptionalOutgoingArgOffsetFromSP;
1337 int getFirstAutomaticVarOffset (MachineCodeForMethod& mcInfo,
1339 int getRegSpillAreaOffset (MachineCodeForMethod& mcInfo,
1341 int getTmpAreaOffset (MachineCodeForMethod& mcInfo,
1343 int getDynamicAreaOffset (MachineCodeForMethod& mcInfo,
1347 // These methods specify the base register used for each stack area
1348 // (generally FP or SP)
1350 virtual int getIncomingArgBaseRegNum() const {
1351 return (int) target.getRegInfo().getFramePointer();
1353 virtual int getOutgoingArgBaseRegNum() const {
1354 return (int) target.getRegInfo().getStackPointer();
1356 virtual int getOptionalOutgoingArgBaseRegNum() const {
1357 return (int) target.getRegInfo().getStackPointer();
1359 virtual int getAutomaticVarBaseRegNum() const {
1360 return (int) target.getRegInfo().getFramePointer();
1362 virtual int getRegSpillAreaBaseRegNum() const {
1363 return (int) target.getRegInfo().getFramePointer();
1365 virtual int getDynamicAreaBaseRegNum() const {
1366 return (int) target.getRegInfo().getStackPointer();
1370 // All stack addresses must be offset by 0x7ff (2047) on Sparc V9.
1371 static const int OFFSET = (int) 0x7ff;
1372 static const int StackFrameSizeAlignment = 16;
1373 static const int MinStackFrameSize = 176;
1374 static const int NumFixedOutgoingArgs = 6;
1375 static const int SizeOfEachArgOnStack = 8;
1376 static const int StaticAreaOffsetFromFP = 0 + OFFSET;
1377 static const int FirstIncomingArgOffsetFromFP = 128 + OFFSET;
1378 static const int FirstOptionalIncomingArgOffsetFromFP = 176 + OFFSET;
1379 static const int FirstOutgoingArgOffsetFromSP = 128 + OFFSET;
1380 static const int FirstOptionalOutgoingArgOffsetFromSP = 176 + OFFSET;
1384 //---------------------------------------------------------------------------
1385 // class UltraSparcCacheInfo
1388 // Interface to cache parameters for the UltraSPARC.
1389 // Just use defaults for now.
1390 //---------------------------------------------------------------------------
1392 class UltraSparcCacheInfo: public MachineCacheInfo {
1394 /*ctor*/ UltraSparcCacheInfo (const TargetMachine& target) :
1395 MachineCacheInfo(target) {}
1399 //---------------------------------------------------------------------------
1400 // class UltraSparcMachine
1403 // Primary interface to machine description for the UltraSPARC.
1404 // Primarily just initializes machine-dependent parameters in
1405 // class TargetMachine, and creates machine-dependent subclasses
1406 // for classes such as InstrInfo, SchedInfo and RegInfo.
1407 //---------------------------------------------------------------------------
1409 class UltraSparc : public TargetMachine {
1411 UltraSparcInstrInfo instrInfo;
1412 UltraSparcSchedInfo schedInfo;
1413 UltraSparcRegInfo regInfo;
1414 UltraSparcFrameInfo frameInfo;
1415 UltraSparcCacheInfo cacheInfo;
1418 virtual ~UltraSparc() {}
1420 virtual const MachineInstrInfo &getInstrInfo() const { return instrInfo; }
1421 virtual const MachineSchedInfo &getSchedInfo() const { return schedInfo; }
1422 virtual const MachineRegInfo &getRegInfo() const { return regInfo; }
1423 virtual const MachineFrameInfo &getFrameInfo() const { return frameInfo; }
1424 virtual const MachineCacheInfo &getCacheInfo() const { return cacheInfo; }
1426 // compileMethod - For the sparc, we do instruction selection, followed by
1427 // delay slot scheduling, then register allocation.
1429 virtual bool compileMethod(Method *M);
1432 // emitAssembly - Output assembly language code (a .s file) for the specified
1433 // module. The specified module must have been compiled before this may be
1436 virtual void emitAssembly(const Module *M, ostream &OutStr) const;