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, LiveRangeInfo& LRI,
175 vector<RegClass *> RCList) const;
178 Value *getValue4ReturnAddr( const MachineInstr * MInst ) const ;
180 int getRegType(const LiveRange *const LR) const {
184 switch( (LR->getRegClass())->getID() ) {
186 case IntRegClassID: return IntRegType;
188 case FloatRegClassID:
189 Typ = LR->getTypeID();
190 if( Typ == Type::FloatTyID )
191 return FPSingleRegType;
192 else if( Typ == Type::DoubleTyID )
193 return FPDoubleRegType;
194 else assert(0 && "Unknown type in FloatRegClass");
196 case IntCCRegClassID: return IntCCRegType;
198 case FloatCCRegClassID: return FloatCCRegType ;
200 default: assert( 0 && "Unknown reg class ID");
206 int getRegType(const Value *const Val) const {
210 switch( getRegClassIDOfValue(Val) ) {
212 case IntRegClassID: return IntRegType;
214 case FloatRegClassID:
215 Typ = (Val->getType())->getPrimitiveID();
216 if( Typ == Type::FloatTyID )
217 return FPSingleRegType;
218 else if( Typ == Type::DoubleTyID )
219 return FPDoubleRegType;
220 else assert(0 && "Unknown type in FloatRegClass");
222 case IntCCRegClassID: return IntCCRegType;
224 case FloatCCRegClassID: return FloatCCRegType ;
226 default: assert( 0 && "Unknown reg class ID");
234 // ***TODO: See this method is necessary
236 MachineInstr * cpValue2RegMI(Value * Val, const unsigned DestReg,
237 const int RegType) const;
239 const Value *getCallInstRetAddr(const MachineInstr *CallMI) const;
240 const unsigned getCallInstNumArgs(const MachineInstr *CallMI) const;
246 UltraSparcRegInfo(const UltraSparc *const USI ) : UltraSparcInfo(USI),
248 NumOfFloatArgRegs(32),
251 MachineRegClassArr.push_back( new SparcIntRegClass(IntRegClassID) );
252 MachineRegClassArr.push_back( new SparcFloatRegClass(FloatRegClassID) );
253 MachineRegClassArr.push_back( new SparcIntCCRegClass(IntCCRegClassID) );
254 MachineRegClassArr.push_back( new SparcFloatCCRegClass(FloatCCRegClassID));
256 assert( SparcFloatRegOrder::StartOfNonVolatileRegs == 32 &&
257 "32 Float regs are used for float arg passing");
261 ~UltraSparcRegInfo(void) { } // empty destructor
264 inline const UltraSparc & getUltraSparcInfo() const {
265 return *UltraSparcInfo;
270 inline unsigned getRegClassIDOfValue (const Value *const Val,
271 bool isCCReg = false) const {
273 Type::PrimitiveID ty = (Val->getType())->getPrimitiveID();
277 if( (ty && ty <= Type::LongTyID) || (ty == Type::LabelTyID) ||
278 (ty == Type::MethodTyID) || (ty == Type::PointerTyID) )
279 res = IntRegClassID; // sparc int reg (ty=0: void)
280 else if( ty <= Type::DoubleTyID)
281 res = FloatRegClassID; // sparc float reg class
283 cerr << "TypeID: " << ty << endl;
284 assert(0 && "Cannot resolve register class for type");
288 return res + 2; // corresponidng condition code regiser
293 // returns the register tha contains always zero
294 // this is the unified register number
295 inline int getZeroRegNum() const { return SparcIntRegOrder::g0; }
297 // returns the reg used for pushing the address when a method is called.
298 // This can be used for other purposes between calls
299 unsigned getCallAddressReg() const { return SparcIntRegOrder::o7; }
302 // and when we return from a method. It should be made sure that this
303 // register contains the return value when a return instruction is reached.
304 unsigned getReturnAddressReg() const { return SparcIntRegOrder::i7; }
306 void suggestRegs4MethodArgs(const Method *const Meth,
307 LiveRangeInfo& LRI) const;
309 void suggestRegs4CallArgs(const MachineInstr *const CallMI,
310 LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
312 void suggestReg4RetValue(const MachineInstr *const RetMI,
313 LiveRangeInfo& LRI ) const;
316 void colorMethodArgs(const Method *const Meth, LiveRangeInfo& LRI,
317 AddedInstrns *const FirstAI) const;
319 void colorCallArgs(const MachineInstr *const CallMI, LiveRangeInfo& LRI,
320 AddedInstrns *const CallAI) const;
322 void colorRetValue(const MachineInstr *const RetI, LiveRangeInfo& LRI,
323 AddedInstrns *const RetAI) const;
326 // bool handleSpecialMInstr(const MachineInstr * MInst,
327 // LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
330 static void printReg(const LiveRange *const LR) ;
332 // this method provides a unique number for each register
333 inline int getUnifiedRegNum(int RegClassID, int reg) const {
335 if( RegClassID == IntRegClassID && reg < 32 )
337 else if ( RegClassID == FloatRegClassID && reg < 64)
338 return reg + 32; // we have 32 int regs
339 else if( RegClassID == FloatCCRegClassID && reg < 4)
340 return reg + 32 + 64; // 32 int, 64 float
341 else if( RegClassID == IntCCRegClassID )
342 return 4+ 32 + 64; // only int cc reg
343 else if (reg==InvalidRegNum)
344 return InvalidRegNum;
346 assert(0 && "Invalid register class or reg number");
350 // given the unified register number, this gives the name
351 inline const string getUnifiedRegName(int reg) const {
353 return SparcIntRegOrder::getRegName(reg);
354 else if ( reg < (64 + 32) )
355 return SparcFloatRegOrder::getRegName( reg - 32);
356 else if( reg < (64+32+4) )
357 return SparcFloatCCRegOrder::getRegName( reg -32 - 64);
358 else if ( reg == 64+32+4)
359 return "xcc"; // only integer cc reg
361 else if (reg== InvalidRegNum) //****** TODO: Remove
364 assert(0 && "Invalid register number");
368 const Value * getCallInstRetVal(const MachineInstr *CallMI) const;
370 MachineInstr * cpReg2RegMI(const unsigned SrcReg, const unsigned DestReg,
371 const int RegType) const;
373 MachineInstr * cpReg2MemMI(const unsigned SrcReg, const unsigned DestPtrReg,
374 const int Offset, const int RegType) const;
376 MachineInstr * cpMem2RegMI(const unsigned SrcPtrReg, const int Offset,
377 const unsigned DestReg, const int RegType) const;
379 inline bool isRegVolatile(const int RegClassID, const int Reg) const {
380 return (MachineRegClassArr[RegClassID])->isRegVolatile(Reg);
384 inline unsigned getFramePointer() const {
385 return SparcIntRegOrder::i6;
388 inline unsigned getStackPointer() const {
389 return SparcIntRegOrder::o6;
392 inline int getInvalidRegNum() const {
393 return InvalidRegNum;
400 /*---------------------------------------------------------------------------
401 Scheduling guidelines for SPARC IIi:
403 I-Cache alignment rules (pg 326)
404 -- Align a branch target instruction so that it's entire group is within
405 the same cache line (may be 1-4 instructions).
406 ** Don't let a branch that is predicted taken be the last instruction
407 on an I-cache line: delay slot will need an entire line to be fetched
408 -- Make a FP instruction or a branch be the 4th instruction in a group.
409 For branches, there are tradeoffs in reordering to make this happen
411 ** Don't put a branch in a group that crosses a 32-byte boundary!
412 An artificial branch is inserted after every 32 bytes, and having
413 another branch will force the group to be broken into 2 groups.
416 -- Don't let a loop span two memory pages, if possible
418 Branch prediction performance:
419 -- Don't make the branch in a delay slot the target of a branch
420 -- Try not to have 2 predicted branches within a group of 4 instructions
421 (because each such group has a single branch target field).
422 -- Try to align branches in slots 0, 2, 4 or 6 of a cache line (to avoid
423 the wrong prediction bits being used in some cases).
425 D-Cache timing constraints:
426 -- Signed int loads of less than 64 bits have 3 cycle latency, not 2
427 -- All other loads that hit in D-Cache have 2 cycle latency
428 -- All loads are returned IN ORDER, so a D-Cache miss will delay a later hit
429 -- Mis-aligned loads or stores cause a trap. In particular, replace
430 mis-aligned FP double precision l/s with 2 single-precision l/s.
431 -- Simulations of integer codes show increase in avg. group size of
432 33% when code (including esp. non-faulting loads) is moved across
433 one branch, and 50% across 2 branches.
435 E-Cache timing constraints:
436 -- Scheduling for E-cache (D-Cache misses) is effective (due to load buffering)
438 Store buffer timing constraints:
439 -- Stores can be executed in same cycle as instruction producing the value
440 -- Stores are buffered and have lower priority for E-cache until
441 highwater mark is reached in the store buffer (5 stores)
443 Pipeline constraints:
444 -- Shifts can only use IEU0.
445 -- CC setting instructions can only use IEU1.
446 -- Several other instructions must only use IEU1:
447 EDGE(?), ARRAY(?), CALL, JMPL, BPr, PST, and FCMP.
448 -- Two instructions cannot store to the same register file in a single cycle
449 (single write port per file).
451 Issue and grouping constraints:
452 -- FP and branch instructions must use slot 4.
453 -- Shift instructions cannot be grouped with other IEU0-specific instructions.
454 -- CC setting instructions cannot be grouped with other IEU1-specific instrs.
455 -- Several instructions must be issued in a single-instruction group:
456 MOVcc or MOVr, MULs/x and DIVs/x, SAVE/RESTORE, many others
457 -- A CALL or JMPL breaks a group, ie, is not combined with subsequent instrs.
461 Branch delay slot scheduling rules:
462 -- A CTI couple (two back-to-back CTI instructions in the dynamic stream)
463 has a 9-instruction penalty: the entire pipeline is flushed when the
464 second instruction reaches stage 9 (W-Writeback).
465 -- Avoid putting multicycle instructions, and instructions that may cause
466 load misses, in the delay slot of an annulling branch.
467 -- Avoid putting WR, SAVE..., RESTORE and RETURN instructions in the
468 delay slot of an annulling branch.
470 *--------------------------------------------------------------------------- */
472 //---------------------------------------------------------------------------
473 // List of CPUResources for UltraSPARC IIi.
474 //---------------------------------------------------------------------------
476 const CPUResource AllIssueSlots( "All Instr Slots", 4);
477 const CPUResource IntIssueSlots( "Int Instr Slots", 3);
478 const CPUResource First3IssueSlots("Instr Slots 0-3", 3);
479 const CPUResource LSIssueSlots( "Load-Store Instr Slot", 1);
480 const CPUResource CTIIssueSlots( "Ctrl Transfer Instr Slot", 1);
481 const CPUResource FPAIssueSlots( "Int Instr Slot 1", 1);
482 const CPUResource FPMIssueSlots( "Int Instr Slot 1", 1);
484 // IEUN instructions can use either Alu and should use IAluN.
485 // IEU0 instructions must use Alu 1 and should use both IAluN and IAlu0.
486 // IEU1 instructions must use Alu 2 and should use both IAluN and IAlu1.
487 const CPUResource IAluN("Int ALU 1or2", 2);
488 const CPUResource IAlu0("Int ALU 1", 1);
489 const CPUResource IAlu1("Int ALU 2", 1);
491 const CPUResource LSAluC1("Load/Store Unit Addr Cycle", 1);
492 const CPUResource LSAluC2("Load/Store Unit Issue Cycle", 1);
493 const CPUResource LdReturn("Load Return Unit", 1);
495 const CPUResource FPMAluC1("FP Mul/Div Alu Cycle 1", 1);
496 const CPUResource FPMAluC2("FP Mul/Div Alu Cycle 2", 1);
497 const CPUResource FPMAluC3("FP Mul/Div Alu Cycle 3", 1);
499 const CPUResource FPAAluC1("FP Other Alu Cycle 1", 1);
500 const CPUResource FPAAluC2("FP Other Alu Cycle 2", 1);
501 const CPUResource FPAAluC3("FP Other Alu Cycle 3", 1);
503 const CPUResource IRegReadPorts("Int Reg ReadPorts", INT_MAX); // CHECK
504 const CPUResource IRegWritePorts("Int Reg WritePorts", 2); // CHECK
505 const CPUResource FPRegReadPorts("FP Reg Read Ports", INT_MAX); // CHECK
506 const CPUResource FPRegWritePorts("FP Reg Write Ports", 1); // CHECK
508 const CPUResource CTIDelayCycle( "CTI delay cycle", 1);
509 const CPUResource FCMPDelayCycle("FCMP delay cycle", 1);
512 //---------------------------------------------------------------------------
513 // const InstrClassRUsage SparcRUsageDesc[]
516 // Resource usage information for instruction in each scheduling class.
517 // The InstrRUsage Objects for individual classes are specified first.
518 // Note that fetch and decode are decoupled from the execution pipelines
519 // via an instr buffer, so they are not included in the cycles below.
520 //---------------------------------------------------------------------------
522 const InstrClassRUsage NoneClassRUsage = {
527 /* isSingleIssue */ false,
528 /* breaksGroup */ false,
532 /* feasibleSlots[] */ { 0, 1, 2, 3 },
546 const InstrClassRUsage IEUNClassRUsage = {
551 /* isSingleIssue */ false,
552 /* breaksGroup */ false,
556 /* feasibleSlots[] */ { 0, 1, 2 },
560 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
561 { IntIssueSlots.rid, 0, 1 },
562 /*Cycle E */ { IAluN.rid, 1, 1 },
567 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
571 const InstrClassRUsage IEU0ClassRUsage = {
576 /* isSingleIssue */ false,
577 /* breaksGroup */ false,
581 /* feasibleSlots[] */ { 0, 1, 2 },
585 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
586 { IntIssueSlots.rid, 0, 1 },
587 /*Cycle E */ { IAluN.rid, 1, 1 },
593 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
597 const InstrClassRUsage IEU1ClassRUsage = {
602 /* isSingleIssue */ false,
603 /* breaksGroup */ false,
607 /* feasibleSlots[] */ { 0, 1, 2 },
611 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
612 { IntIssueSlots.rid, 0, 1 },
613 /*Cycle E */ { IAluN.rid, 1, 1 },
619 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
623 const InstrClassRUsage FPMClassRUsage = {
628 /* isSingleIssue */ false,
629 /* breaksGroup */ false,
633 /* feasibleSlots[] */ { 0, 1, 2, 3 },
637 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
638 { FPMIssueSlots.rid, 0, 1 },
639 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
640 /*Cycle C */ { FPMAluC1.rid, 2, 1 },
641 /*Cycle N1*/ { FPMAluC2.rid, 3, 1 },
642 /*Cycle N1*/ { FPMAluC3.rid, 4, 1 },
644 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
648 const InstrClassRUsage FPAClassRUsage = {
653 /* isSingleIssue */ false,
654 /* breaksGroup */ false,
658 /* feasibleSlots[] */ { 0, 1, 2, 3 },
662 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
663 { FPAIssueSlots.rid, 0, 1 },
664 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
665 /*Cycle C */ { FPAAluC1.rid, 2, 1 },
666 /*Cycle N1*/ { FPAAluC2.rid, 3, 1 },
667 /*Cycle N1*/ { FPAAluC3.rid, 4, 1 },
669 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
673 const InstrClassRUsage LDClassRUsage = {
678 /* isSingleIssue */ false,
679 /* breaksGroup */ false,
683 /* feasibleSlots[] */ { 0, 1, 2, },
687 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
688 { First3IssueSlots.rid, 0, 1 },
689 { LSIssueSlots.rid, 0, 1 },
690 /*Cycle E */ { LSAluC1.rid, 1, 1 },
691 /*Cycle C */ { LSAluC2.rid, 2, 1 },
692 { LdReturn.rid, 2, 1 },
696 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
700 const InstrClassRUsage STClassRUsage = {
705 /* isSingleIssue */ false,
706 /* breaksGroup */ false,
710 /* feasibleSlots[] */ { 0, 1, 2 },
714 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
715 { First3IssueSlots.rid, 0, 1 },
716 { LSIssueSlots.rid, 0, 1 },
717 /*Cycle E */ { LSAluC1.rid, 1, 1 },
718 /*Cycle C */ { LSAluC2.rid, 2, 1 }
726 const InstrClassRUsage CTIClassRUsage = {
731 /* isSingleIssue */ false,
732 /* breaksGroup */ false,
736 /* feasibleSlots[] */ { 0, 1, 2, 3 },
740 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
741 { CTIIssueSlots.rid, 0, 1 },
742 /*Cycle E */ { IAlu0.rid, 1, 1 },
743 /*Cycles E-C */ { CTIDelayCycle.rid, 1, 2 }
752 const InstrClassRUsage SingleClassRUsage = {
757 /* isSingleIssue */ true,
758 /* breaksGroup */ false,
762 /* feasibleSlots[] */ { 0 },
766 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
767 { AllIssueSlots.rid, 0, 1 },
768 { AllIssueSlots.rid, 0, 1 },
769 { AllIssueSlots.rid, 0, 1 },
770 /*Cycle E */ { IAlu0.rid, 1, 1 }
780 const InstrClassRUsage SparcRUsageDesc[] = {
794 //---------------------------------------------------------------------------
795 // const InstrIssueDelta SparcInstrIssueDeltas[]
798 // Changes to issue restrictions information in InstrClassRUsage for
799 // instructions that differ from other instructions in their class.
800 //---------------------------------------------------------------------------
802 const InstrIssueDelta SparcInstrIssueDeltas[] = {
804 // opCode, isSingleIssue, breaksGroup, numBubbles
806 // Special cases for single-issue only
807 // Other single issue cases are below.
808 //{ LDDA, true, true, 0 },
809 //{ STDA, true, true, 0 },
810 //{ LDDF, true, true, 0 },
811 //{ LDDFA, true, true, 0 },
812 { ADDC, true, true, 0 },
813 { ADDCcc, true, true, 0 },
814 { SUBC, true, true, 0 },
815 { SUBCcc, true, true, 0 },
816 //{ SAVE, true, true, 0 },
817 //{ RESTORE, true, true, 0 },
818 //{ LDSTUB, true, true, 0 },
819 //{ SWAP, true, true, 0 },
820 //{ SWAPA, true, true, 0 },
821 //{ CAS, true, true, 0 },
822 //{ CASA, true, true, 0 },
823 //{ CASX, true, true, 0 },
824 //{ CASXA, true, true, 0 },
825 //{ LDFSR, true, true, 0 },
826 //{ LDFSRA, true, true, 0 },
827 //{ LDXFSR, true, true, 0 },
828 //{ LDXFSRA, true, true, 0 },
829 //{ STFSR, true, true, 0 },
830 //{ STFSRA, true, true, 0 },
831 //{ STXFSR, true, true, 0 },
832 //{ STXFSRA, true, true, 0 },
833 //{ SAVED, true, true, 0 },
834 //{ RESTORED, true, true, 0 },
835 //{ FLUSH, true, true, 9 },
836 //{ FLUSHW, true, true, 9 },
837 //{ ALIGNADDR, true, true, 0 },
838 { RETURN, true, true, 0 },
839 //{ DONE, true, true, 0 },
840 //{ RETRY, true, true, 0 },
841 //{ WR, true, true, 0 },
842 //{ WRPR, true, true, 4 },
843 //{ RD, true, true, 0 },
844 //{ RDPR, true, true, 0 },
845 //{ TCC, true, true, 0 },
846 //{ SHUTDOWN, true, true, 0 },
848 // Special cases for breaking group *before*
849 // CURRENTLY NOT SUPPORTED!
850 { CALL, false, false, 0 },
851 { JMPL, false, false, 0 },
853 // Special cases for breaking the group *after*
854 { MULX, true, true, (4+34)/2 },
855 { FDIVS, false, true, 0 },
856 { FDIVD, false, true, 0 },
857 { FDIVQ, false, true, 0 },
858 { FSQRTS, false, true, 0 },
859 { FSQRTD, false, true, 0 },
860 { FSQRTQ, false, true, 0 },
861 //{ FCMP{LE,GT,NE,EQ}, false, true, 0 },
863 // Instructions that introduce bubbles
864 //{ MULScc, true, true, 2 },
865 //{ SMULcc, true, true, (4+18)/2 },
866 //{ UMULcc, true, true, (4+19)/2 },
867 { SDIVX, true, true, 68 },
868 { UDIVX, true, true, 68 },
869 //{ SDIVcc, true, true, 36 },
870 //{ UDIVcc, true, true, 37 },
871 //{ WR, false, false, 4 },
872 //{ WRPR, false, false, 4 },
876 //---------------------------------------------------------------------------
877 // const InstrRUsageDelta SparcInstrUsageDeltas[]
880 // Changes to resource usage information in InstrClassRUsage for
881 // instructions that differ from other instructions in their class.
882 //---------------------------------------------------------------------------
884 const InstrRUsageDelta SparcInstrUsageDeltas[] = {
886 // MachineOpCode, Resource, Start cycle, Num cycles
889 // JMPL counts as a load/store instruction for issue!
891 { JMPL, LSIssueSlots.rid, 0, 1 },
894 // Many instructions cannot issue for the next 2 cycles after an FCMP
895 // We model that with a fake resource FCMPDelayCycle.
897 { FCMPS, FCMPDelayCycle.rid, 1, 3 },
898 { FCMPD, FCMPDelayCycle.rid, 1, 3 },
899 { FCMPQ, FCMPDelayCycle.rid, 1, 3 },
901 { MULX, FCMPDelayCycle.rid, 1, 1 },
902 { SDIVX, FCMPDelayCycle.rid, 1, 1 },
903 { UDIVX, FCMPDelayCycle.rid, 1, 1 },
904 //{ SMULcc, FCMPDelayCycle.rid, 1, 1 },
905 //{ UMULcc, FCMPDelayCycle.rid, 1, 1 },
906 //{ SDIVcc, FCMPDelayCycle.rid, 1, 1 },
907 //{ UDIVcc, FCMPDelayCycle.rid, 1, 1 },
908 { STD, FCMPDelayCycle.rid, 1, 1 },
909 { FMOVRSZ, FCMPDelayCycle.rid, 1, 1 },
910 { FMOVRSLEZ,FCMPDelayCycle.rid, 1, 1 },
911 { FMOVRSLZ, FCMPDelayCycle.rid, 1, 1 },
912 { FMOVRSNZ, FCMPDelayCycle.rid, 1, 1 },
913 { FMOVRSGZ, FCMPDelayCycle.rid, 1, 1 },
914 { FMOVRSGEZ,FCMPDelayCycle.rid, 1, 1 },
917 // Some instructions are stalled in the GROUP stage if a CTI is in
920 { LDD, CTIDelayCycle.rid, 1, 1 },
921 //{ LDDA, CTIDelayCycle.rid, 1, 1 },
922 //{ LDDSTUB, CTIDelayCycle.rid, 1, 1 },
923 //{ LDDSTUBA, CTIDelayCycle.rid, 1, 1 },
924 //{ SWAP, CTIDelayCycle.rid, 1, 1 },
925 //{ SWAPA, CTIDelayCycle.rid, 1, 1 },
926 //{ CAS, CTIDelayCycle.rid, 1, 1 },
927 //{ CASA, CTIDelayCycle.rid, 1, 1 },
928 //{ CASX, CTIDelayCycle.rid, 1, 1 },
929 //{ CASXA, CTIDelayCycle.rid, 1, 1 },
932 // Signed int loads of less than dword size return data in cycle N1 (not C)
933 // and put all loads in consecutive cycles into delayed load return mode.
935 { LDSB, LdReturn.rid, 2, -1 },
936 { LDSB, LdReturn.rid, 3, 1 },
938 { LDSH, LdReturn.rid, 2, -1 },
939 { LDSH, LdReturn.rid, 3, 1 },
941 { LDSW, LdReturn.rid, 2, -1 },
942 { LDSW, LdReturn.rid, 3, 1 },
945 #undef EXPLICIT_BUBBLES_NEEDED
946 #ifdef EXPLICIT_BUBBLES_NEEDED
948 // MULScc inserts one bubble.
949 // This means it breaks the current group (captured in UltraSparcSchedInfo)
950 // *and occupies all issue slots for the next cycle
952 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
953 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
954 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
955 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
958 // SMULcc inserts between 4 and 18 bubbles, depending on #leading 0s in rs1.
959 // We just model this with a simple average.
961 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
962 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
963 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
964 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
966 // SMULcc inserts between 4 and 19 bubbles, depending on #leading 0s in rs1.
967 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
968 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
969 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
970 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
973 // MULX inserts between 4 and 34 bubbles, depending on #leading 0s in rs1.
975 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
976 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
977 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
978 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
981 // SDIVcc inserts 36 bubbles.
983 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
984 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
985 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
986 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
988 // UDIVcc inserts 37 bubbles.
989 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
990 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
991 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
992 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
995 // SDIVX inserts 68 bubbles.
997 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
998 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
999 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1000 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
1003 // UDIVX inserts 68 bubbles.
1005 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1006 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1007 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1008 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
1011 // WR inserts 4 bubbles.
1013 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1014 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1015 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1016 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1019 // WRPR inserts 4 bubbles.
1021 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1022 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1023 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1024 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1027 // DONE inserts 9 bubbles.
1029 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1030 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1031 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1032 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1035 // RETRY inserts 9 bubbles.
1037 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1038 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1039 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1040 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1042 #endif /*EXPLICIT_BUBBLES_NEEDED */
1047 // Additional delays to be captured in code:
1048 // 1. RDPR from several state registers (page 349)
1049 // 2. RD from *any* register (page 349)
1050 // 3. Writes to TICK, PSTATE, TL registers and FLUSH{W} instr (page 349)
1051 // 4. Integer store can be in same group as instr producing value to store.
1052 // 5. BICC and BPICC can be in the same group as instr producing CC (pg 350)
1053 // 6. FMOVr cannot be in the same or next group as an IEU instr (pg 351).
1054 // 7. The second instr. of a CTI group inserts 9 bubbles (pg 351)
1055 // 8. WR{PR}, SVAE, SAVED, RESTORE, RESTORED, RETURN, RETRY, and DONE that
1056 // follow an annulling branch cannot be issued in the same group or in
1057 // the 3 groups following the branch.
1058 // 9. A predicted annulled load does not stall dependent instructions.
1059 // Other annulled delay slot instructions *do* stall dependents, so
1060 // nothing special needs to be done for them during scheduling.
1061 //10. Do not put a load use that may be annulled in the same group as the
1062 // branch. The group will stall until the load returns.
1063 //11. Single-prec. FP loads lock 2 registers, for dependency checking.
1066 // Additional delays we cannot or will not capture:
1067 // 1. If DCTI is last word of cache line, it is delayed until next line can be
1068 // fetched. Also, other DCTI alignment-related delays (pg 352)
1069 // 2. Load-after-store is delayed by 7 extra cycles if load hits in D-Cache.
1070 // Also, several other store-load and load-store conflicts (pg 358)
1071 // 3. MEMBAR, LD{X}FSR, LDD{A} and a bunch of other load stalls (pg 358)
1072 // 4. There can be at most 8 outstanding buffered store instructions
1073 // (including some others like MEMBAR, LDSTUB, CAS{AX}, and FLUSH)
1077 //---------------------------------------------------------------------------
1078 // class UltraSparcSchedInfo
1081 // Interface to instruction scheduling information for UltraSPARC.
1082 // The parameter values above are based on UltraSPARC IIi.
1083 //---------------------------------------------------------------------------
1086 class UltraSparcSchedInfo: public MachineSchedInfo {
1088 /*ctor*/ UltraSparcSchedInfo (const MachineInstrInfo* mii);
1089 /*dtor*/ virtual ~UltraSparcSchedInfo () {}
1091 virtual void initializeResources ();
1095 //---------------------------------------------------------------------------
1096 // class UltraSparcMachine
1099 // Primary interface to machine description for the UltraSPARC.
1100 // Primarily just initializes machine-dependent parameters in
1101 // class TargetMachine, and creates machine-dependent subclasses
1102 // for classes such as InstrInfo, SchedInfo and RegInfo.
1103 //---------------------------------------------------------------------------
1105 class UltraSparc : public TargetMachine {
1107 UltraSparcInstrInfo instrInfo;
1108 UltraSparcSchedInfo schedInfo;
1109 UltraSparcRegInfo regInfo;
1112 virtual ~UltraSparc() {}
1114 virtual const MachineInstrInfo &getInstrInfo() const { return instrInfo; }
1115 virtual const MachineSchedInfo &getSchedInfo() const { return schedInfo; }
1116 virtual const MachineRegInfo &getRegInfo() const { return regInfo; }
1118 // compileMethod - For the sparc, we do instruction selection, followed by
1119 // delay slot scheduling, then register allocation.
1121 virtual bool compileMethod(Method *M);
1124 // emitAssembly - Output assembly language code (a .s file) for the specified
1125 // module. The specified module must have been compiled before this may be
1128 virtual void emitAssembly(const Module *M, ostream &OutStr) const;