1 //===-- SparcInternals.h - Header file for Sparc backend ---------*- C++ -*--=//
3 // This file defines stuff that is to be private to the Sparc backend, but is
4 // shared among different portions of the backend.
6 //===----------------------------------------------------------------------===//
8 #ifndef SPARC_INTERNALS_H
9 #define SPARC_INTERNALS_H
12 #include "SparcRegClassInfo.h"
13 #include "llvm/Target/TargetMachine.h"
14 #include "llvm/Target/MachineInstrInfo.h"
16 #include "llvm/Target/MachineSchedInfo.h"
17 #include "llvm/CodeGen/RegClass.h"
18 #include "llvm/Type.h"
20 #include <sys/types.h>
24 // OpCodeMask definitions for the Sparc V9
26 const OpCodeMask Immed = 0x00002000; // immed or reg operand?
27 const OpCodeMask Annul = 0x20000000; // annul delay instr?
28 const OpCodeMask PredictTaken = 0x00080000; // predict branch taken?
31 enum SparcInstrSchedClass {
32 SPARC_NONE, /* Instructions with no scheduling restrictions */
33 SPARC_IEUN, /* Integer class that can use IEU0 or IEU1 */
34 SPARC_IEU0, /* Integer class IEU0 */
35 SPARC_IEU1, /* Integer class IEU1 */
36 SPARC_FPM, /* FP Multiply or Divide instructions */
37 SPARC_FPA, /* All other FP instructions */
38 SPARC_CTI, /* Control-transfer instructions */
39 SPARC_LD, /* Load instructions */
40 SPARC_ST, /* Store instructions */
41 SPARC_SINGLE, /* Instructions that must issue by themselves */
43 SPARC_INV, /* This should stay at the end for the next value */
44 SPARC_NUM_SCHED_CLASSES = SPARC_INV
48 //---------------------------------------------------------------------------
49 // enum SparcMachineOpCode.
50 // const MachineInstrDescriptor SparcMachineInstrDesc[]
53 // Description of UltraSparc machine instructions.
55 //---------------------------------------------------------------------------
57 enum SparcMachineOpCode {
58 #define I(ENUM, OPCODESTRING, NUMOPERANDS, RESULTPOS, MAXIMM, IMMSE, \
59 NUMDELAYSLOTS, LATENCY, SCHEDCLASS, INSTFLAGS) \
61 #include "SparcInstr.def"
63 // End-of-array marker
65 NUM_REAL_OPCODES = RETURN+1, // number of valid opcodes
66 NUM_TOTAL_OPCODES = INVALID_OPCODE
70 // Array of machine instruction descriptions...
71 extern const MachineInstrDescriptor SparcMachineInstrDesc[];
74 //---------------------------------------------------------------------------
75 // class UltraSparcInstrInfo
78 // Information about individual instructions.
79 // Most information is stored in the SparcMachineInstrDesc array above.
80 // Other information is computed on demand, and most such functions
81 // default to member functions in base class MachineInstrInfo.
82 //---------------------------------------------------------------------------
84 class UltraSparcInstrInfo : public MachineInstrInfo {
86 /*ctor*/ UltraSparcInstrInfo();
88 virtual bool hasResultInterlock (MachineOpCode opCode) const
90 // All UltraSPARC instructions have interlocks (note that delay slots
91 // are not considered here).
92 // However, instructions that use the result of an FCMP produce a
93 // 9-cycle stall if they are issued less than 3 cycles after the FCMP.
94 // Force the compiler to insert a software interlock (i.e., gap of
95 // 2 other groups, including NOPs if necessary).
96 return (opCode == FCMPS || opCode == FCMPD || opCode == FCMPQ);
99 //-------------------------------------------------------------------------
100 // Code generation support for creating individual machine instructions
101 //-------------------------------------------------------------------------
103 // Create an instruction sequence to put the constant `val' into
104 // the virtual register `dest'. The generated instructions are
105 // returned in `minstrVec'. Any temporary registers (TmpInstruction)
106 // created are returned in `tempVec'.
108 virtual void CreateCodeToLoadConst(Value* val,
110 vector<MachineInstr*>& minstrVec,
111 vector<TmpInstruction*>& tempVec) const;
121 class UltraSparcRegInfo : public MachineRegInfo
126 // The actual register classes in the Sparc
136 // Type of registers available in Sparc. There can be several reg types
137 // in the same class. For instace, the float reg class has Single/Double
148 // WARNING: If the above enum order must be changed, also modify
149 // getRegisterClassOfValue method below since it assumes this particular
150 // order for efficiency.
153 // reverse pointer to get info about the ultra sparc machine
154 const UltraSparc *const UltraSparcInfo;
156 // Both int and float rguments can be passed in 6 int regs -
157 // %o0 to %o5 (cannot be changed)
158 unsigned const NumOfIntArgRegs;
159 unsigned const NumOfFloatArgRegs;
160 int const InvalidRegNum;
162 //void setCallArgColor(LiveRange *const LR, const unsigned RegNo) const;
164 void setCallOrRetArgCol(LiveRange *const LR, const unsigned RegNo,
165 const MachineInstr *MI,AddedInstrMapType &AIMap)const;
167 MachineInstr * getCopy2RegMI(const Value *SrcVal, const unsigned Reg,
168 unsigned RegClassID) const ;
171 void suggestReg4RetAddr(const MachineInstr * RetMI,
172 LiveRangeInfo& LRI) const;
174 void suggestReg4CallAddr(const MachineInstr * CallMI) const;
177 Value *getValue4ReturnAddr( const MachineInstr * MInst ) const ;
179 int getRegType(const LiveRange *const LR) const {
183 switch( (LR->getRegClass())->getID() ) {
185 case IntRegClassID: return IntRegType;
187 case FloatRegClassID:
188 Typ = LR->getTypeID();
189 if( Typ == Type::FloatTyID )
190 return FPSingleRegType;
191 else if( Typ == Type::DoubleTyID )
192 return FPDoubleRegType;
193 else assert(0 && "Unknown type in FloatRegClass");
195 case IntCCRegClassID: return IntCCRegType;
197 case FloatCCRegClassID: return FloatCCRegType ;
199 default: assert( 0 && "Unknown reg class ID");
205 int getRegType(const Value *const Val) const {
209 switch( getRegClassIDOfValue(Val) ) {
211 case IntRegClassID: return IntRegType;
213 case FloatRegClassID:
214 Typ = (Val->getType())->getPrimitiveID();
215 if( Typ == Type::FloatTyID )
216 return FPSingleRegType;
217 else if( Typ == Type::DoubleTyID )
218 return FPDoubleRegType;
219 else assert(0 && "Unknown type in FloatRegClass");
221 case IntCCRegClassID: return IntCCRegType;
223 case FloatCCRegClassID: return FloatCCRegType ;
225 default: assert( 0 && "Unknown reg class ID");
233 // ***TODO: See this method is necessary
235 MachineInstr * cpValue2RegMI(Value * Val, const unsigned DestReg,
236 const int RegType) const;
242 UltraSparcRegInfo(const UltraSparc *const USI ) : UltraSparcInfo(USI),
244 NumOfFloatArgRegs(32),
247 MachineRegClassArr.push_back( new SparcIntRegClass(IntRegClassID) );
248 MachineRegClassArr.push_back( new SparcFloatRegClass(FloatRegClassID) );
249 MachineRegClassArr.push_back( new SparcIntCCRegClass(IntCCRegClassID) );
250 MachineRegClassArr.push_back( new SparcFloatCCRegClass(FloatCCRegClassID));
252 assert( SparcFloatRegOrder::StartOfNonVolatileRegs == 32 &&
253 "32 Float regs are used for float arg passing");
257 ~UltraSparcRegInfo(void) { } // empty destructor
260 inline const UltraSparc & getUltraSparcInfo() const {
261 return *UltraSparcInfo;
266 inline unsigned getRegClassIDOfValue (const Value *const Val,
267 bool isCCReg = false) const {
269 Type::PrimitiveID ty = (Val->getType())->getPrimitiveID();
273 if( (ty && ty <= Type::LongTyID) || (ty == Type::LabelTyID) ||
274 (ty == Type::MethodTyID) || (ty == Type::PointerTyID) )
275 res = IntRegClassID; // sparc int reg (ty=0: void)
276 else if( ty <= Type::DoubleTyID)
277 res = FloatRegClassID; // sparc float reg class
279 cerr << "TypeID: " << ty << endl;
280 assert(0 && "Cannot resolve register class for type");
284 return res + 2; // corresponidng condition code regiser
289 // returns the register tha contains always zero
290 // this is the unified register number
291 inline int getZeroRegNum() const { return SparcIntRegOrder::g0; }
293 // returns the reg used for pushing the address when a method is called.
294 // This can be used for other purposes between calls
295 unsigned getCallAddressReg() const { return SparcIntRegOrder::o7; }
298 // and when we return from a method. It should be made sure that this
299 // register contains the return value when a return instruction is reached.
300 unsigned getReturnAddressReg() const { return SparcIntRegOrder::i7; }
302 void suggestRegs4MethodArgs(const Method *const Meth,
303 LiveRangeInfo& LRI) const;
305 void suggestRegs4CallArgs(const MachineInstr *const CallMI,
306 LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
308 void suggestReg4RetValue(const MachineInstr *const RetMI,
309 LiveRangeInfo& LRI ) const;
312 void colorMethodArgs(const Method *const Meth, LiveRangeInfo& LRI,
313 AddedInstrns *const FirstAI) const;
315 void colorCallArgs(const MachineInstr *const CallMI, LiveRangeInfo& LRI,
316 AddedInstrns *const CallAI) const;
318 void colorRetValue(const MachineInstr *const RetI, LiveRangeInfo& LRI,
319 AddedInstrns *const RetAI) const;
322 // bool handleSpecialMInstr(const MachineInstr * MInst,
323 // LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
326 static void printReg(const LiveRange *const LR) ;
328 // this method provides a unique number for each register
329 inline int getUnifiedRegNum(int RegClassID, int reg) const {
331 if( RegClassID == IntRegClassID && reg < 32 )
333 else if ( RegClassID == FloatRegClassID && reg < 64)
334 return reg + 32; // we have 32 int regs
335 else if( RegClassID == FloatCCRegClassID && reg < 4)
336 return reg + 32 + 64; // 32 int, 64 float
337 else if( RegClassID == IntCCRegClassID )
338 return 4+ 32 + 64; // only int cc reg
339 else if (reg==InvalidRegNum)
340 return InvalidRegNum;
342 assert(0 && "Invalid register class or reg number");
346 // given the unified register number, this gives the name
347 inline const string getUnifiedRegName(int reg) const {
349 return SparcIntRegOrder::getRegName(reg);
350 else if ( reg < (64 + 32) )
351 return SparcFloatRegOrder::getRegName( reg - 32);
352 else if( reg < (64+32+4) )
353 return SparcFloatCCRegOrder::getRegName( reg -32 - 64);
354 else if ( reg == 64+32+4)
355 return "xcc"; // only integer cc reg
357 else if (reg== InvalidRegNum) //****** TODO: Remove
360 assert(0 && "Invalid register number");
364 MachineInstr * cpReg2RegMI(const unsigned SrcReg, const unsigned DestReg,
365 const int RegType) const;
367 MachineInstr * cpReg2MemMI(const unsigned SrcReg, const unsigned DestPtrReg,
368 const int Offset, const int RegType) const;
370 MachineInstr * cpMem2RegMI(const unsigned SrcPtrReg, const int Offset,
371 const unsigned DestReg, const int RegType) const;
373 inline bool isRegVolatile(const int RegClassID, const int Reg) const {
374 return (MachineRegClassArr[RegClassID])->isRegVolatile(Reg);
378 inline unsigned getFramePointer() const {
379 return SparcIntRegOrder::i6;
382 inline unsigned getStackPointer() const {
383 return SparcIntRegOrder::o6;
386 inline int getInvalidRegNum() const {
387 return InvalidRegNum;
394 /*---------------------------------------------------------------------------
395 Scheduling guidelines for SPARC IIi:
397 I-Cache alignment rules (pg 326)
398 -- Align a branch target instruction so that it's entire group is within
399 the same cache line (may be 1-4 instructions).
400 ** Don't let a branch that is predicted taken be the last instruction
401 on an I-cache line: delay slot will need an entire line to be fetched
402 -- Make a FP instruction or a branch be the 4th instruction in a group.
403 For branches, there are tradeoffs in reordering to make this happen
405 ** Don't put a branch in a group that crosses a 32-byte boundary!
406 An artificial branch is inserted after every 32 bytes, and having
407 another branch will force the group to be broken into 2 groups.
410 -- Don't let a loop span two memory pages, if possible
412 Branch prediction performance:
413 -- Don't make the branch in a delay slot the target of a branch
414 -- Try not to have 2 predicted branches within a group of 4 instructions
415 (because each such group has a single branch target field).
416 -- Try to align branches in slots 0, 2, 4 or 6 of a cache line (to avoid
417 the wrong prediction bits being used in some cases).
419 D-Cache timing constraints:
420 -- Signed int loads of less than 64 bits have 3 cycle latency, not 2
421 -- All other loads that hit in D-Cache have 2 cycle latency
422 -- All loads are returned IN ORDER, so a D-Cache miss will delay a later hit
423 -- Mis-aligned loads or stores cause a trap. In particular, replace
424 mis-aligned FP double precision l/s with 2 single-precision l/s.
425 -- Simulations of integer codes show increase in avg. group size of
426 33% when code (including esp. non-faulting loads) is moved across
427 one branch, and 50% across 2 branches.
429 E-Cache timing constraints:
430 -- Scheduling for E-cache (D-Cache misses) is effective (due to load buffering)
432 Store buffer timing constraints:
433 -- Stores can be executed in same cycle as instruction producing the value
434 -- Stores are buffered and have lower priority for E-cache until
435 highwater mark is reached in the store buffer (5 stores)
437 Pipeline constraints:
438 -- Shifts can only use IEU0.
439 -- CC setting instructions can only use IEU1.
440 -- Several other instructions must only use IEU1:
441 EDGE(?), ARRAY(?), CALL, JMPL, BPr, PST, and FCMP.
442 -- Two instructions cannot store to the same register file in a single cycle
443 (single write port per file).
445 Issue and grouping constraints:
446 -- FP and branch instructions must use slot 4.
447 -- Shift instructions cannot be grouped with other IEU0-specific instructions.
448 -- CC setting instructions cannot be grouped with other IEU1-specific instrs.
449 -- Several instructions must be issued in a single-instruction group:
450 MOVcc or MOVr, MULs/x and DIVs/x, SAVE/RESTORE, many others
451 -- A CALL or JMPL breaks a group, ie, is not combined with subsequent instrs.
455 Branch delay slot scheduling rules:
456 -- A CTI couple (two back-to-back CTI instructions in the dynamic stream)
457 has a 9-instruction penalty: the entire pipeline is flushed when the
458 second instruction reaches stage 9 (W-Writeback).
459 -- Avoid putting multicycle instructions, and instructions that may cause
460 load misses, in the delay slot of an annulling branch.
461 -- Avoid putting WR, SAVE..., RESTORE and RETURN instructions in the
462 delay slot of an annulling branch.
464 *--------------------------------------------------------------------------- */
466 //---------------------------------------------------------------------------
467 // List of CPUResources for UltraSPARC IIi.
468 //---------------------------------------------------------------------------
470 const CPUResource AllIssueSlots( "All Instr Slots", 4);
471 const CPUResource IntIssueSlots( "Int Instr Slots", 3);
472 const CPUResource First3IssueSlots("Instr Slots 0-3", 3);
473 const CPUResource LSIssueSlots( "Load-Store Instr Slot", 1);
474 const CPUResource CTIIssueSlots( "Ctrl Transfer Instr Slot", 1);
475 const CPUResource FPAIssueSlots( "Int Instr Slot 1", 1);
476 const CPUResource FPMIssueSlots( "Int Instr Slot 1", 1);
478 // IEUN instructions can use either Alu and should use IAluN.
479 // IEU0 instructions must use Alu 1 and should use both IAluN and IAlu0.
480 // IEU1 instructions must use Alu 2 and should use both IAluN and IAlu1.
481 const CPUResource IAluN("Int ALU 1or2", 2);
482 const CPUResource IAlu0("Int ALU 1", 1);
483 const CPUResource IAlu1("Int ALU 2", 1);
485 const CPUResource LSAluC1("Load/Store Unit Addr Cycle", 1);
486 const CPUResource LSAluC2("Load/Store Unit Issue Cycle", 1);
487 const CPUResource LdReturn("Load Return Unit", 1);
489 const CPUResource FPMAluC1("FP Mul/Div Alu Cycle 1", 1);
490 const CPUResource FPMAluC2("FP Mul/Div Alu Cycle 2", 1);
491 const CPUResource FPMAluC3("FP Mul/Div Alu Cycle 3", 1);
493 const CPUResource FPAAluC1("FP Other Alu Cycle 1", 1);
494 const CPUResource FPAAluC2("FP Other Alu Cycle 2", 1);
495 const CPUResource FPAAluC3("FP Other Alu Cycle 3", 1);
497 const CPUResource IRegReadPorts("Int Reg ReadPorts", INT_MAX); // CHECK
498 const CPUResource IRegWritePorts("Int Reg WritePorts", 2); // CHECK
499 const CPUResource FPRegReadPorts("FP Reg Read Ports", INT_MAX); // CHECK
500 const CPUResource FPRegWritePorts("FP Reg Write Ports", 1); // CHECK
502 const CPUResource CTIDelayCycle( "CTI delay cycle", 1);
503 const CPUResource FCMPDelayCycle("FCMP delay cycle", 1);
506 //---------------------------------------------------------------------------
507 // const InstrClassRUsage SparcRUsageDesc[]
510 // Resource usage information for instruction in each scheduling class.
511 // The InstrRUsage Objects for individual classes are specified first.
512 // Note that fetch and decode are decoupled from the execution pipelines
513 // via an instr buffer, so they are not included in the cycles below.
514 //---------------------------------------------------------------------------
516 const InstrClassRUsage NoneClassRUsage = {
521 /* isSingleIssue */ false,
522 /* breaksGroup */ false,
526 /* feasibleSlots[] */ { 0, 1, 2, 3 },
540 const InstrClassRUsage IEUNClassRUsage = {
545 /* isSingleIssue */ false,
546 /* breaksGroup */ false,
550 /* feasibleSlots[] */ { 0, 1, 2 },
554 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
555 { IntIssueSlots.rid, 0, 1 },
556 /*Cycle E */ { IAluN.rid, 1, 1 },
561 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
565 const InstrClassRUsage IEU0ClassRUsage = {
570 /* isSingleIssue */ false,
571 /* breaksGroup */ false,
575 /* feasibleSlots[] */ { 0, 1, 2 },
579 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
580 { IntIssueSlots.rid, 0, 1 },
581 /*Cycle E */ { IAluN.rid, 1, 1 },
587 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
591 const InstrClassRUsage IEU1ClassRUsage = {
596 /* isSingleIssue */ false,
597 /* breaksGroup */ false,
601 /* feasibleSlots[] */ { 0, 1, 2 },
605 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
606 { IntIssueSlots.rid, 0, 1 },
607 /*Cycle E */ { IAluN.rid, 1, 1 },
613 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
617 const InstrClassRUsage FPMClassRUsage = {
622 /* isSingleIssue */ false,
623 /* breaksGroup */ false,
627 /* feasibleSlots[] */ { 0, 1, 2, 3 },
631 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
632 { FPMIssueSlots.rid, 0, 1 },
633 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
634 /*Cycle C */ { FPMAluC1.rid, 2, 1 },
635 /*Cycle N1*/ { FPMAluC2.rid, 3, 1 },
636 /*Cycle N1*/ { FPMAluC3.rid, 4, 1 },
638 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
642 const InstrClassRUsage FPAClassRUsage = {
647 /* isSingleIssue */ false,
648 /* breaksGroup */ false,
652 /* feasibleSlots[] */ { 0, 1, 2, 3 },
656 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
657 { FPAIssueSlots.rid, 0, 1 },
658 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
659 /*Cycle C */ { FPAAluC1.rid, 2, 1 },
660 /*Cycle N1*/ { FPAAluC2.rid, 3, 1 },
661 /*Cycle N1*/ { FPAAluC3.rid, 4, 1 },
663 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
667 const InstrClassRUsage LDClassRUsage = {
672 /* isSingleIssue */ false,
673 /* breaksGroup */ false,
677 /* feasibleSlots[] */ { 0, 1, 2, },
681 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
682 { First3IssueSlots.rid, 0, 1 },
683 { LSIssueSlots.rid, 0, 1 },
684 /*Cycle E */ { LSAluC1.rid, 1, 1 },
685 /*Cycle C */ { LSAluC2.rid, 2, 1 },
686 { LdReturn.rid, 2, 1 },
690 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
694 const InstrClassRUsage STClassRUsage = {
699 /* isSingleIssue */ false,
700 /* breaksGroup */ false,
704 /* feasibleSlots[] */ { 0, 1, 2 },
708 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
709 { First3IssueSlots.rid, 0, 1 },
710 { LSIssueSlots.rid, 0, 1 },
711 /*Cycle E */ { LSAluC1.rid, 1, 1 },
712 /*Cycle C */ { LSAluC2.rid, 2, 1 }
720 const InstrClassRUsage CTIClassRUsage = {
725 /* isSingleIssue */ false,
726 /* breaksGroup */ false,
730 /* feasibleSlots[] */ { 0, 1, 2, 3 },
734 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
735 { CTIIssueSlots.rid, 0, 1 },
736 /*Cycle E */ { IAlu0.rid, 1, 1 },
737 /*Cycles E-C */ { CTIDelayCycle.rid, 1, 2 }
746 const InstrClassRUsage SingleClassRUsage = {
751 /* isSingleIssue */ true,
752 /* breaksGroup */ false,
756 /* feasibleSlots[] */ { 0 },
760 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
761 { AllIssueSlots.rid, 0, 1 },
762 { AllIssueSlots.rid, 0, 1 },
763 { AllIssueSlots.rid, 0, 1 },
764 /*Cycle E */ { IAlu0.rid, 1, 1 }
774 const InstrClassRUsage SparcRUsageDesc[] = {
788 //---------------------------------------------------------------------------
789 // const InstrIssueDelta SparcInstrIssueDeltas[]
792 // Changes to issue restrictions information in InstrClassRUsage for
793 // instructions that differ from other instructions in their class.
794 //---------------------------------------------------------------------------
796 const InstrIssueDelta SparcInstrIssueDeltas[] = {
798 // opCode, isSingleIssue, breaksGroup, numBubbles
800 // Special cases for single-issue only
801 // Other single issue cases are below.
802 //{ LDDA, true, true, 0 },
803 //{ STDA, true, true, 0 },
804 //{ LDDF, true, true, 0 },
805 //{ LDDFA, true, true, 0 },
806 { ADDC, true, true, 0 },
807 { ADDCcc, true, true, 0 },
808 { SUBC, true, true, 0 },
809 { SUBCcc, true, true, 0 },
810 //{ SAVE, true, true, 0 },
811 //{ RESTORE, true, true, 0 },
812 //{ LDSTUB, true, true, 0 },
813 //{ SWAP, true, true, 0 },
814 //{ SWAPA, true, true, 0 },
815 //{ CAS, true, true, 0 },
816 //{ CASA, true, true, 0 },
817 //{ CASX, true, true, 0 },
818 //{ CASXA, true, true, 0 },
819 //{ LDFSR, true, true, 0 },
820 //{ LDFSRA, true, true, 0 },
821 //{ LDXFSR, true, true, 0 },
822 //{ LDXFSRA, true, true, 0 },
823 //{ STFSR, true, true, 0 },
824 //{ STFSRA, true, true, 0 },
825 //{ STXFSR, true, true, 0 },
826 //{ STXFSRA, true, true, 0 },
827 //{ SAVED, true, true, 0 },
828 //{ RESTORED, true, true, 0 },
829 //{ FLUSH, true, true, 9 },
830 //{ FLUSHW, true, true, 9 },
831 //{ ALIGNADDR, true, true, 0 },
832 { RETURN, true, true, 0 },
833 //{ DONE, true, true, 0 },
834 //{ RETRY, true, true, 0 },
835 //{ WR, true, true, 0 },
836 //{ WRPR, true, true, 4 },
837 //{ RD, true, true, 0 },
838 //{ RDPR, true, true, 0 },
839 //{ TCC, true, true, 0 },
840 //{ SHUTDOWN, true, true, 0 },
842 // Special cases for breaking group *before*
843 // CURRENTLY NOT SUPPORTED!
844 { CALL, false, false, 0 },
845 { JMPL, false, false, 0 },
847 // Special cases for breaking the group *after*
848 { MULX, true, true, (4+34)/2 },
849 { FDIVS, false, true, 0 },
850 { FDIVD, false, true, 0 },
851 { FDIVQ, false, true, 0 },
852 { FSQRTS, false, true, 0 },
853 { FSQRTD, false, true, 0 },
854 { FSQRTQ, false, true, 0 },
855 //{ FCMP{LE,GT,NE,EQ}, false, true, 0 },
857 // Instructions that introduce bubbles
858 //{ MULScc, true, true, 2 },
859 //{ SMULcc, true, true, (4+18)/2 },
860 //{ UMULcc, true, true, (4+19)/2 },
861 { SDIVX, true, true, 68 },
862 { UDIVX, true, true, 68 },
863 //{ SDIVcc, true, true, 36 },
864 //{ UDIVcc, true, true, 37 },
865 //{ WR, false, false, 4 },
866 //{ WRPR, false, false, 4 },
870 //---------------------------------------------------------------------------
871 // const InstrRUsageDelta SparcInstrUsageDeltas[]
874 // Changes to resource usage information in InstrClassRUsage for
875 // instructions that differ from other instructions in their class.
876 //---------------------------------------------------------------------------
878 const InstrRUsageDelta SparcInstrUsageDeltas[] = {
880 // MachineOpCode, Resource, Start cycle, Num cycles
883 // JMPL counts as a load/store instruction for issue!
885 { JMPL, LSIssueSlots.rid, 0, 1 },
888 // Many instructions cannot issue for the next 2 cycles after an FCMP
889 // We model that with a fake resource FCMPDelayCycle.
891 { FCMPS, FCMPDelayCycle.rid, 1, 3 },
892 { FCMPD, FCMPDelayCycle.rid, 1, 3 },
893 { FCMPQ, FCMPDelayCycle.rid, 1, 3 },
895 { MULX, FCMPDelayCycle.rid, 1, 1 },
896 { SDIVX, FCMPDelayCycle.rid, 1, 1 },
897 { UDIVX, FCMPDelayCycle.rid, 1, 1 },
898 //{ SMULcc, FCMPDelayCycle.rid, 1, 1 },
899 //{ UMULcc, FCMPDelayCycle.rid, 1, 1 },
900 //{ SDIVcc, FCMPDelayCycle.rid, 1, 1 },
901 //{ UDIVcc, FCMPDelayCycle.rid, 1, 1 },
902 { STD, FCMPDelayCycle.rid, 1, 1 },
903 { FMOVRSZ, FCMPDelayCycle.rid, 1, 1 },
904 { FMOVRSLEZ,FCMPDelayCycle.rid, 1, 1 },
905 { FMOVRSLZ, FCMPDelayCycle.rid, 1, 1 },
906 { FMOVRSNZ, FCMPDelayCycle.rid, 1, 1 },
907 { FMOVRSGZ, FCMPDelayCycle.rid, 1, 1 },
908 { FMOVRSGEZ,FCMPDelayCycle.rid, 1, 1 },
911 // Some instructions are stalled in the GROUP stage if a CTI is in
914 { LDD, CTIDelayCycle.rid, 1, 1 },
915 //{ LDDA, CTIDelayCycle.rid, 1, 1 },
916 //{ LDDSTUB, CTIDelayCycle.rid, 1, 1 },
917 //{ LDDSTUBA, CTIDelayCycle.rid, 1, 1 },
918 //{ SWAP, CTIDelayCycle.rid, 1, 1 },
919 //{ SWAPA, CTIDelayCycle.rid, 1, 1 },
920 //{ CAS, CTIDelayCycle.rid, 1, 1 },
921 //{ CASA, CTIDelayCycle.rid, 1, 1 },
922 //{ CASX, CTIDelayCycle.rid, 1, 1 },
923 //{ CASXA, CTIDelayCycle.rid, 1, 1 },
926 // Signed int loads of less than dword size return data in cycle N1 (not C)
927 // and put all loads in consecutive cycles into delayed load return mode.
929 { LDSB, LdReturn.rid, 2, -1 },
930 { LDSB, LdReturn.rid, 3, 1 },
932 { LDSH, LdReturn.rid, 2, -1 },
933 { LDSH, LdReturn.rid, 3, 1 },
935 { LDSW, LdReturn.rid, 2, -1 },
936 { LDSW, LdReturn.rid, 3, 1 },
939 #undef EXPLICIT_BUBBLES_NEEDED
940 #ifdef EXPLICIT_BUBBLES_NEEDED
942 // MULScc inserts one bubble.
943 // This means it breaks the current group (captured in UltraSparcSchedInfo)
944 // *and occupies all issue slots for the next cycle
946 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
947 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
948 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
949 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
952 // SMULcc inserts between 4 and 18 bubbles, depending on #leading 0s in rs1.
953 // We just model this with a simple average.
955 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
956 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
957 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
958 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
960 // SMULcc inserts between 4 and 19 bubbles, depending on #leading 0s in rs1.
961 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
962 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
963 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
964 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
967 // MULX inserts between 4 and 34 bubbles, depending on #leading 0s in rs1.
969 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
970 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
971 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
972 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
975 // SDIVcc inserts 36 bubbles.
977 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
978 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
979 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
980 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
982 // UDIVcc inserts 37 bubbles.
983 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
984 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
985 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
986 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
989 // SDIVX inserts 68 bubbles.
991 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
992 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
993 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
994 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
997 // UDIVX inserts 68 bubbles.
999 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1000 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1001 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1002 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1005 // WR inserts 4 bubbles.
1007 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1008 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1009 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1010 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1013 // WRPR inserts 4 bubbles.
1015 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1016 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1017 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1018 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1021 // DONE inserts 9 bubbles.
1023 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1024 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1025 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1026 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1029 // RETRY inserts 9 bubbles.
1031 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1032 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1033 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1034 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1036 #endif /*EXPLICIT_BUBBLES_NEEDED */
1041 // Additional delays to be captured in code:
1042 // 1. RDPR from several state registers (page 349)
1043 // 2. RD from *any* register (page 349)
1044 // 3. Writes to TICK, PSTATE, TL registers and FLUSH{W} instr (page 349)
1045 // 4. Integer store can be in same group as instr producing value to store.
1046 // 5. BICC and BPICC can be in the same group as instr producing CC (pg 350)
1047 // 6. FMOVr cannot be in the same or next group as an IEU instr (pg 351).
1048 // 7. The second instr. of a CTI group inserts 9 bubbles (pg 351)
1049 // 8. WR{PR}, SVAE, SAVED, RESTORE, RESTORED, RETURN, RETRY, and DONE that
1050 // follow an annulling branch cannot be issued in the same group or in
1051 // the 3 groups following the branch.
1052 // 9. A predicted annulled load does not stall dependent instructions.
1053 // Other annulled delay slot instructions *do* stall dependents, so
1054 // nothing special needs to be done for them during scheduling.
1055 //10. Do not put a load use that may be annulled in the same group as the
1056 // branch. The group will stall until the load returns.
1057 //11. Single-prec. FP loads lock 2 registers, for dependency checking.
1060 // Additional delays we cannot or will not capture:
1061 // 1. If DCTI is last word of cache line, it is delayed until next line can be
1062 // fetched. Also, other DCTI alignment-related delays (pg 352)
1063 // 2. Load-after-store is delayed by 7 extra cycles if load hits in D-Cache.
1064 // Also, several other store-load and load-store conflicts (pg 358)
1065 // 3. MEMBAR, LD{X}FSR, LDD{A} and a bunch of other load stalls (pg 358)
1066 // 4. There can be at most 8 outstanding buffered store instructions
1067 // (including some others like MEMBAR, LDSTUB, CAS{AX}, and FLUSH)
1071 //---------------------------------------------------------------------------
1072 // class UltraSparcSchedInfo
1075 // Interface to instruction scheduling information for UltraSPARC.
1076 // The parameter values above are based on UltraSPARC IIi.
1077 //---------------------------------------------------------------------------
1080 class UltraSparcSchedInfo: public MachineSchedInfo {
1082 /*ctor*/ UltraSparcSchedInfo (const MachineInstrInfo* mii);
1083 /*dtor*/ virtual ~UltraSparcSchedInfo () {}
1085 virtual void initializeResources ();
1089 //---------------------------------------------------------------------------
1090 // class UltraSparcMachine
1093 // Primary interface to machine description for the UltraSPARC.
1094 // Primarily just initializes machine-dependent parameters in
1095 // class TargetMachine, and creates machine-dependent subclasses
1096 // for classes such as InstrInfo, SchedInfo and RegInfo.
1097 //---------------------------------------------------------------------------
1099 class UltraSparc : public TargetMachine {
1101 UltraSparcInstrInfo instrInfo;
1102 UltraSparcSchedInfo schedInfo;
1103 UltraSparcRegInfo regInfo;
1106 virtual ~UltraSparc() {}
1108 virtual const MachineInstrInfo &getInstrInfo() const { return instrInfo; }
1109 virtual const MachineSchedInfo &getSchedInfo() const { return schedInfo; }
1110 virtual const MachineRegInfo &getRegInfo() const { return regInfo; }
1112 // compileMethod - For the sparc, we do instruction selection, followed by
1113 // delay slot scheduling, then register allocation.
1115 virtual bool compileMethod(Method *M);
1118 // emitAssembly - Output assembly language code (a .s file) for the specified
1119 // module. The specified module must have been compiled before this may be
1122 virtual void emitAssembly(const Module *M, ostream &OutStr) const;