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)
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);
112 class UltraSparcRegInfo : public MachineRegInfo
117 // The actual register classes in the Sparc
127 // Type of registers available in Sparc. There can be several reg types
128 // in the same class. For instace, the float reg class has Single/Double
139 // WARNING: If the above enum order must be changed, also modify
140 // getRegisterClassOfValue method below since it assumes this particular
141 // order for efficiency.
144 // reverse pointer to get info about the ultra sparc machine
145 const UltraSparc *const UltraSparcInfo;
147 // Both int and float rguments can be passed in 6 int regs -
148 // %o0 to %o5 (cannot be changed)
149 unsigned const NumOfIntArgRegs;
150 unsigned const NumOfFloatArgRegs;
151 int const InvalidRegNum;
153 //void setCallArgColor(LiveRange *const LR, const unsigned RegNo) const;
155 void setCallOrRetArgCol(LiveRange *const LR, const unsigned RegNo,
156 const MachineInstr *MI,AddedInstrMapType &AIMap)const;
158 MachineInstr * getCopy2RegMI(const Value *SrcVal, const unsigned Reg,
159 unsigned RegClassID) const ;
162 void suggestReg4RetAddr(const MachineInstr * RetMI,
163 LiveRangeInfo& LRI) const;
165 void suggestReg4CallAddr(const MachineInstr * CallMI) const;
168 Value *getValue4ReturnAddr( const MachineInstr * MInst ) const ;
170 int getRegType(const LiveRange *const LR) const {
174 switch( (LR->getRegClass())->getID() ) {
176 case IntRegClassID: return IntRegType;
178 case FloatRegClassID:
179 Typ = LR->getTypeID();
180 if( Typ == Type::FloatTyID )
181 return FPSingleRegType;
182 else if( Typ == Type::DoubleTyID )
183 return FPDoubleRegType;
184 else assert(0 && "Unknown type in FloatRegClass");
186 case IntCCRegClassID: return IntCCRegType;
188 case FloatCCRegClassID: return FloatCCRegType ;
190 default: assert( 0 && "Unknown reg class ID");
196 int getRegType(const Value *const Val) const {
200 switch( getRegClassIDOfValue(Val) ) {
202 case IntRegClassID: return IntRegType;
204 case FloatRegClassID:
205 Typ = (Val->getType())->getPrimitiveID();
206 if( Typ == Type::FloatTyID )
207 return FPSingleRegType;
208 else if( Typ == Type::DoubleTyID )
209 return FPDoubleRegType;
210 else assert(0 && "Unknown type in FloatRegClass");
212 case IntCCRegClassID: return IntCCRegType;
214 case FloatCCRegClassID: return FloatCCRegType ;
216 default: assert( 0 && "Unknown reg class ID");
224 // ***TODO: See this method is necessary
226 MachineInstr * cpValue2RegMI(Value * Val, const unsigned DestReg,
227 const int RegType) const;
233 UltraSparcRegInfo(const UltraSparc *const USI ) : UltraSparcInfo(USI),
235 NumOfFloatArgRegs(32),
238 MachineRegClassArr.push_back( new SparcIntRegClass(IntRegClassID) );
239 MachineRegClassArr.push_back( new SparcFloatRegClass(FloatRegClassID) );
240 MachineRegClassArr.push_back( new SparcIntCCRegClass(IntCCRegClassID) );
241 MachineRegClassArr.push_back( new SparcFloatCCRegClass(FloatCCRegClassID));
243 assert( SparcFloatRegOrder::StartOfNonVolatileRegs == 32 &&
244 "32 Float regs are used for float arg passing");
248 ~UltraSparcRegInfo(void) { } // empty destructor
251 inline const UltraSparc & getUltraSparcInfo() const {
252 return *UltraSparcInfo;
257 inline unsigned getRegClassIDOfValue (const Value *const Val,
258 bool isCCReg = false) const {
260 Type::PrimitiveID ty = (Val->getType())->getPrimitiveID();
264 if( (ty && ty <= Type::LongTyID) || (ty == Type::LabelTyID) ||
265 (ty == Type::MethodTyID) || (ty == Type::PointerTyID) )
266 res = IntRegClassID; // sparc int reg (ty=0: void)
267 else if( ty <= Type::DoubleTyID)
268 res = FloatRegClassID; // sparc float reg class
270 cerr << "TypeID: " << ty << endl;
271 assert(0 && "Cannot resolve register class for type");
275 return res + 2; // corresponidng condition code regiser
280 // returns the register tha contains always zero
281 // this is the unified register number
282 inline int getZeroRegNum() const { return SparcIntRegOrder::g0; }
284 // returns the reg used for pushing the address when a method is called.
285 // This can be used for other purposes between calls
286 unsigned getCallAddressReg() const { return SparcIntRegOrder::o7; }
289 // and when we return from a method. It should be made sure that this
290 // register contains the return value when a return instruction is reached.
291 unsigned getReturnAddressReg() const { return SparcIntRegOrder::i7; }
293 void suggestRegs4MethodArgs(const Method *const Meth,
294 LiveRangeInfo& LRI) const;
296 void suggestRegs4CallArgs(const MachineInstr *const CallMI,
297 LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
299 void suggestReg4RetValue(const MachineInstr *const RetMI,
300 LiveRangeInfo& LRI ) const;
303 void colorMethodArgs(const Method *const Meth, LiveRangeInfo& LRI,
304 AddedInstrns *const FirstAI) const;
306 void colorCallArgs(const MachineInstr *const CallMI, LiveRangeInfo& LRI,
307 AddedInstrns *const CallAI) const;
309 void colorRetValue(const MachineInstr *const RetI, LiveRangeInfo& LRI,
310 AddedInstrns *const RetAI) const;
313 // bool handleSpecialMInstr(const MachineInstr * MInst,
314 // LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
317 static void printReg(const LiveRange *const LR) ;
319 // this method provides a unique number for each register
320 inline int getUnifiedRegNum(int RegClassID, int reg) const {
322 if( RegClassID == IntRegClassID && reg < 32 )
324 else if ( RegClassID == FloatRegClassID && reg < 64)
325 return reg + 32; // we have 32 int regs
326 else if( RegClassID == FloatCCRegClassID && reg < 4)
327 return reg + 32 + 64; // 32 int, 64 float
328 else if( RegClassID == IntCCRegClassID )
329 return 4+ 32 + 64; // only int cc reg
330 else if (reg==InvalidRegNum)
331 return InvalidRegNum;
333 assert(0 && "Invalid register class or reg number");
337 // given the unified register number, this gives the name
338 inline const string getUnifiedRegName(int reg) const {
340 return SparcIntRegOrder::getRegName(reg);
341 else if ( reg < (64 + 32) )
342 return SparcFloatRegOrder::getRegName( reg - 32);
343 else if( reg < (64+32+4) )
344 return SparcFloatCCRegOrder::getRegName( reg -32 - 64);
345 else if ( reg == 64+32+4)
346 return "xcc"; // only integer cc reg
348 else if (reg== InvalidRegNum) //****** TODO: Remove
351 assert(0 && "Invalid register number");
355 MachineInstr * cpReg2RegMI(const unsigned SrcReg, const unsigned DestReg,
356 const int RegType) const;
358 MachineInstr * cpReg2MemMI(const unsigned SrcReg, const unsigned DestPtrReg,
359 const int Offset, const int RegType) const;
361 MachineInstr * cpMem2RegMI(const unsigned SrcPtrReg, const int Offset,
362 const unsigned DestReg, const int RegType) const;
364 inline bool isRegVolatile(const int RegClassID, const int Reg) const {
365 return (MachineRegClassArr[RegClassID])->isRegVolatile(Reg);
369 inline unsigned getFramePointer() const {
370 return SparcIntRegOrder::i6;
373 inline unsigned getStackPointer() const {
374 return SparcIntRegOrder::o6;
377 inline int getInvalidRegNum() const {
378 return InvalidRegNum;
385 /*---------------------------------------------------------------------------
386 Scheduling guidelines for SPARC IIi:
388 I-Cache alignment rules (pg 326)
389 -- Align a branch target instruction so that it's entire group is within
390 the same cache line (may be 1-4 instructions).
391 ** Don't let a branch that is predicted taken be the last instruction
392 on an I-cache line: delay slot will need an entire line to be fetched
393 -- Make a FP instruction or a branch be the 4th instruction in a group.
394 For branches, there are tradeoffs in reordering to make this happen
396 ** Don't put a branch in a group that crosses a 32-byte boundary!
397 An artificial branch is inserted after every 32 bytes, and having
398 another branch will force the group to be broken into 2 groups.
401 -- Don't let a loop span two memory pages, if possible
403 Branch prediction performance:
404 -- Don't make the branch in a delay slot the target of a branch
405 -- Try not to have 2 predicted branches within a group of 4 instructions
406 (because each such group has a single branch target field).
407 -- Try to align branches in slots 0, 2, 4 or 6 of a cache line (to avoid
408 the wrong prediction bits being used in some cases).
410 D-Cache timing constraints:
411 -- Signed int loads of less than 64 bits have 3 cycle latency, not 2
412 -- All other loads that hit in D-Cache have 2 cycle latency
413 -- All loads are returned IN ORDER, so a D-Cache miss will delay a later hit
414 -- Mis-aligned loads or stores cause a trap. In particular, replace
415 mis-aligned FP double precision l/s with 2 single-precision l/s.
416 -- Simulations of integer codes show increase in avg. group size of
417 33% when code (including esp. non-faulting loads) is moved across
418 one branch, and 50% across 2 branches.
420 E-Cache timing constraints:
421 -- Scheduling for E-cache (D-Cache misses) is effective (due to load buffering)
423 Store buffer timing constraints:
424 -- Stores can be executed in same cycle as instruction producing the value
425 -- Stores are buffered and have lower priority for E-cache until
426 highwater mark is reached in the store buffer (5 stores)
428 Pipeline constraints:
429 -- Shifts can only use IEU0.
430 -- CC setting instructions can only use IEU1.
431 -- Several other instructions must only use IEU1:
432 EDGE(?), ARRAY(?), CALL, JMPL, BPr, PST, and FCMP.
433 -- Two instructions cannot store to the same register file in a single cycle
434 (single write port per file).
436 Issue and grouping constraints:
437 -- FP and branch instructions must use slot 4.
438 -- Shift instructions cannot be grouped with other IEU0-specific instructions.
439 -- CC setting instructions cannot be grouped with other IEU1-specific instrs.
440 -- Several instructions must be issued in a single-instruction group:
441 MOVcc or MOVr, MULs/x and DIVs/x, SAVE/RESTORE, many others
442 -- A CALL or JMPL breaks a group, ie, is not combined with subsequent instrs.
446 Branch delay slot scheduling rules:
447 -- A CTI couple (two back-to-back CTI instructions in the dynamic stream)
448 has a 9-instruction penalty: the entire pipeline is flushed when the
449 second instruction reaches stage 9 (W-Writeback).
450 -- Avoid putting multicycle instructions, and instructions that may cause
451 load misses, in the delay slot of an annulling branch.
452 -- Avoid putting WR, SAVE..., RESTORE and RETURN instructions in the
453 delay slot of an annulling branch.
455 *--------------------------------------------------------------------------- */
457 //---------------------------------------------------------------------------
458 // List of CPUResources for UltraSPARC IIi.
459 //---------------------------------------------------------------------------
461 const CPUResource AllIssueSlots( "All Instr Slots", 4);
462 const CPUResource IntIssueSlots( "Int Instr Slots", 3);
463 const CPUResource First3IssueSlots("Instr Slots 0-3", 3);
464 const CPUResource LSIssueSlots( "Load-Store Instr Slot", 1);
465 const CPUResource CTIIssueSlots( "Ctrl Transfer Instr Slot", 1);
466 const CPUResource FPAIssueSlots( "Int Instr Slot 1", 1);
467 const CPUResource FPMIssueSlots( "Int Instr Slot 1", 1);
469 // IEUN instructions can use either Alu and should use IAluN.
470 // IEU0 instructions must use Alu 1 and should use both IAluN and IAlu0.
471 // IEU1 instructions must use Alu 2 and should use both IAluN and IAlu1.
472 const CPUResource IAluN("Int ALU 1or2", 2);
473 const CPUResource IAlu0("Int ALU 1", 1);
474 const CPUResource IAlu1("Int ALU 2", 1);
476 const CPUResource LSAluC1("Load/Store Unit Addr Cycle", 1);
477 const CPUResource LSAluC2("Load/Store Unit Issue Cycle", 1);
478 const CPUResource LdReturn("Load Return Unit", 1);
480 const CPUResource FPMAluC1("FP Mul/Div Alu Cycle 1", 1);
481 const CPUResource FPMAluC2("FP Mul/Div Alu Cycle 2", 1);
482 const CPUResource FPMAluC3("FP Mul/Div Alu Cycle 3", 1);
484 const CPUResource FPAAluC1("FP Other Alu Cycle 1", 1);
485 const CPUResource FPAAluC2("FP Other Alu Cycle 2", 1);
486 const CPUResource FPAAluC3("FP Other Alu Cycle 3", 1);
488 const CPUResource IRegReadPorts("Int Reg ReadPorts", INT_MAX); // CHECK
489 const CPUResource IRegWritePorts("Int Reg WritePorts", 2); // CHECK
490 const CPUResource FPRegReadPorts("FP Reg Read Ports", INT_MAX); // CHECK
491 const CPUResource FPRegWritePorts("FP Reg Write Ports", 1); // CHECK
493 const CPUResource CTIDelayCycle( "CTI delay cycle", 1);
494 const CPUResource FCMPDelayCycle("FCMP delay cycle", 1);
497 //---------------------------------------------------------------------------
498 // const InstrClassRUsage SparcRUsageDesc[]
501 // Resource usage information for instruction in each scheduling class.
502 // The InstrRUsage Objects for individual classes are specified first.
503 // Note that fetch and decode are decoupled from the execution pipelines
504 // via an instr buffer, so they are not included in the cycles below.
505 //---------------------------------------------------------------------------
507 const InstrClassRUsage NoneClassRUsage = {
512 /* isSingleIssue */ false,
513 /* breaksGroup */ false,
517 /* feasibleSlots[] */ { 0, 1, 2, 3 },
531 const InstrClassRUsage IEUNClassRUsage = {
536 /* isSingleIssue */ false,
537 /* breaksGroup */ false,
541 /* feasibleSlots[] */ { 0, 1, 2 },
545 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
546 { IntIssueSlots.rid, 0, 1 },
547 /*Cycle E */ { IAluN.rid, 1, 1 },
552 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
556 const InstrClassRUsage IEU0ClassRUsage = {
561 /* isSingleIssue */ false,
562 /* breaksGroup */ false,
566 /* feasibleSlots[] */ { 0, 1, 2 },
570 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
571 { IntIssueSlots.rid, 0, 1 },
572 /*Cycle E */ { IAluN.rid, 1, 1 },
578 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
582 const InstrClassRUsage IEU1ClassRUsage = {
587 /* isSingleIssue */ false,
588 /* breaksGroup */ false,
592 /* feasibleSlots[] */ { 0, 1, 2 },
596 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
597 { IntIssueSlots.rid, 0, 1 },
598 /*Cycle E */ { IAluN.rid, 1, 1 },
604 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
608 const InstrClassRUsage FPMClassRUsage = {
613 /* isSingleIssue */ false,
614 /* breaksGroup */ false,
618 /* feasibleSlots[] */ { 0, 1, 2, 3 },
622 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
623 { FPMIssueSlots.rid, 0, 1 },
624 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
625 /*Cycle C */ { FPMAluC1.rid, 2, 1 },
626 /*Cycle N1*/ { FPMAluC2.rid, 3, 1 },
627 /*Cycle N1*/ { FPMAluC3.rid, 4, 1 },
629 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
633 const InstrClassRUsage FPAClassRUsage = {
638 /* isSingleIssue */ false,
639 /* breaksGroup */ false,
643 /* feasibleSlots[] */ { 0, 1, 2, 3 },
647 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
648 { FPAIssueSlots.rid, 0, 1 },
649 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
650 /*Cycle C */ { FPAAluC1.rid, 2, 1 },
651 /*Cycle N1*/ { FPAAluC2.rid, 3, 1 },
652 /*Cycle N1*/ { FPAAluC3.rid, 4, 1 },
654 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
658 const InstrClassRUsage LDClassRUsage = {
663 /* isSingleIssue */ false,
664 /* breaksGroup */ false,
668 /* feasibleSlots[] */ { 0, 1, 2, },
672 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
673 { First3IssueSlots.rid, 0, 1 },
674 { LSIssueSlots.rid, 0, 1 },
675 /*Cycle E */ { LSAluC1.rid, 1, 1 },
676 /*Cycle C */ { LSAluC2.rid, 2, 1 },
677 { LdReturn.rid, 2, 1 },
681 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
685 const InstrClassRUsage STClassRUsage = {
690 /* isSingleIssue */ false,
691 /* breaksGroup */ false,
695 /* feasibleSlots[] */ { 0, 1, 2 },
699 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
700 { First3IssueSlots.rid, 0, 1 },
701 { LSIssueSlots.rid, 0, 1 },
702 /*Cycle E */ { LSAluC1.rid, 1, 1 },
703 /*Cycle C */ { LSAluC2.rid, 2, 1 }
711 const InstrClassRUsage CTIClassRUsage = {
716 /* isSingleIssue */ false,
717 /* breaksGroup */ false,
721 /* feasibleSlots[] */ { 0, 1, 2, 3 },
725 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
726 { CTIIssueSlots.rid, 0, 1 },
727 /*Cycle E */ { IAlu0.rid, 1, 1 },
728 /*Cycles E-C */ { CTIDelayCycle.rid, 1, 2 }
737 const InstrClassRUsage SingleClassRUsage = {
742 /* isSingleIssue */ true,
743 /* breaksGroup */ false,
747 /* feasibleSlots[] */ { 0 },
751 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
752 { AllIssueSlots.rid, 0, 1 },
753 { AllIssueSlots.rid, 0, 1 },
754 { AllIssueSlots.rid, 0, 1 },
755 /*Cycle E */ { IAlu0.rid, 1, 1 }
765 const InstrClassRUsage SparcRUsageDesc[] = {
779 //---------------------------------------------------------------------------
780 // const InstrIssueDelta SparcInstrIssueDeltas[]
783 // Changes to issue restrictions information in InstrClassRUsage for
784 // instructions that differ from other instructions in their class.
785 //---------------------------------------------------------------------------
787 const InstrIssueDelta SparcInstrIssueDeltas[] = {
789 // opCode, isSingleIssue, breaksGroup, numBubbles
791 // Special cases for single-issue only
792 // Other single issue cases are below.
793 //{ LDDA, true, true, 0 },
794 //{ STDA, true, true, 0 },
795 //{ LDDF, true, true, 0 },
796 //{ LDDFA, true, true, 0 },
797 { ADDC, true, true, 0 },
798 { ADDCcc, true, true, 0 },
799 { SUBC, true, true, 0 },
800 { SUBCcc, true, true, 0 },
801 //{ SAVE, true, true, 0 },
802 //{ RESTORE, true, true, 0 },
803 //{ LDSTUB, true, true, 0 },
804 //{ SWAP, true, true, 0 },
805 //{ SWAPA, true, true, 0 },
806 //{ CAS, true, true, 0 },
807 //{ CASA, true, true, 0 },
808 //{ CASX, true, true, 0 },
809 //{ CASXA, true, true, 0 },
810 //{ LDFSR, true, true, 0 },
811 //{ LDFSRA, true, true, 0 },
812 //{ LDXFSR, true, true, 0 },
813 //{ LDXFSRA, true, true, 0 },
814 //{ STFSR, true, true, 0 },
815 //{ STFSRA, true, true, 0 },
816 //{ STXFSR, true, true, 0 },
817 //{ STXFSRA, true, true, 0 },
818 //{ SAVED, true, true, 0 },
819 //{ RESTORED, true, true, 0 },
820 //{ FLUSH, true, true, 9 },
821 //{ FLUSHW, true, true, 9 },
822 //{ ALIGNADDR, true, true, 0 },
823 { RETURN, true, true, 0 },
824 //{ DONE, true, true, 0 },
825 //{ RETRY, true, true, 0 },
826 //{ WR, true, true, 0 },
827 //{ WRPR, true, true, 4 },
828 //{ RD, true, true, 0 },
829 //{ RDPR, true, true, 0 },
830 //{ TCC, true, true, 0 },
831 //{ SHUTDOWN, true, true, 0 },
833 // Special cases for breaking group *before*
834 // CURRENTLY NOT SUPPORTED!
835 { CALL, false, false, 0 },
836 { JMPL, false, false, 0 },
838 // Special cases for breaking the group *after*
839 { MULX, true, true, (4+34)/2 },
840 { FDIVS, false, true, 0 },
841 { FDIVD, false, true, 0 },
842 { FDIVQ, false, true, 0 },
843 { FSQRTS, false, true, 0 },
844 { FSQRTD, false, true, 0 },
845 { FSQRTQ, false, true, 0 },
846 //{ FCMP{LE,GT,NE,EQ}, false, true, 0 },
848 // Instructions that introduce bubbles
849 //{ MULScc, true, true, 2 },
850 //{ SMULcc, true, true, (4+18)/2 },
851 //{ UMULcc, true, true, (4+19)/2 },
852 { SDIVX, true, true, 68 },
853 { UDIVX, true, true, 68 },
854 //{ SDIVcc, true, true, 36 },
855 //{ UDIVcc, true, true, 37 },
856 //{ WR, false, false, 4 },
857 //{ WRPR, false, false, 4 },
861 //---------------------------------------------------------------------------
862 // const InstrRUsageDelta SparcInstrUsageDeltas[]
865 // Changes to resource usage information in InstrClassRUsage for
866 // instructions that differ from other instructions in their class.
867 //---------------------------------------------------------------------------
869 const InstrRUsageDelta SparcInstrUsageDeltas[] = {
871 // MachineOpCode, Resource, Start cycle, Num cycles
874 // JMPL counts as a load/store instruction for issue!
876 { JMPL, LSIssueSlots.rid, 0, 1 },
879 // Many instructions cannot issue for the next 2 cycles after an FCMP
880 // We model that with a fake resource FCMPDelayCycle.
882 { FCMPS, FCMPDelayCycle.rid, 1, 3 },
883 { FCMPD, FCMPDelayCycle.rid, 1, 3 },
884 { FCMPQ, FCMPDelayCycle.rid, 1, 3 },
886 { MULX, FCMPDelayCycle.rid, 1, 1 },
887 { SDIVX, FCMPDelayCycle.rid, 1, 1 },
888 { UDIVX, FCMPDelayCycle.rid, 1, 1 },
889 //{ SMULcc, FCMPDelayCycle.rid, 1, 1 },
890 //{ UMULcc, FCMPDelayCycle.rid, 1, 1 },
891 //{ SDIVcc, FCMPDelayCycle.rid, 1, 1 },
892 //{ UDIVcc, FCMPDelayCycle.rid, 1, 1 },
893 { STD, FCMPDelayCycle.rid, 1, 1 },
894 { FMOVRSZ, FCMPDelayCycle.rid, 1, 1 },
895 { FMOVRSLEZ,FCMPDelayCycle.rid, 1, 1 },
896 { FMOVRSLZ, FCMPDelayCycle.rid, 1, 1 },
897 { FMOVRSNZ, FCMPDelayCycle.rid, 1, 1 },
898 { FMOVRSGZ, FCMPDelayCycle.rid, 1, 1 },
899 { FMOVRSGEZ,FCMPDelayCycle.rid, 1, 1 },
902 // Some instructions are stalled in the GROUP stage if a CTI is in
905 { LDD, CTIDelayCycle.rid, 1, 1 },
906 //{ LDDA, CTIDelayCycle.rid, 1, 1 },
907 //{ LDDSTUB, CTIDelayCycle.rid, 1, 1 },
908 //{ LDDSTUBA, CTIDelayCycle.rid, 1, 1 },
909 //{ SWAP, CTIDelayCycle.rid, 1, 1 },
910 //{ SWAPA, CTIDelayCycle.rid, 1, 1 },
911 //{ CAS, CTIDelayCycle.rid, 1, 1 },
912 //{ CASA, CTIDelayCycle.rid, 1, 1 },
913 //{ CASX, CTIDelayCycle.rid, 1, 1 },
914 //{ CASXA, CTIDelayCycle.rid, 1, 1 },
917 // Signed int loads of less than dword size return data in cycle N1 (not C)
918 // and put all loads in consecutive cycles into delayed load return mode.
920 { LDSB, LdReturn.rid, 2, -1 },
921 { LDSB, LdReturn.rid, 3, 1 },
923 { LDSH, LdReturn.rid, 2, -1 },
924 { LDSH, LdReturn.rid, 3, 1 },
926 { LDSW, LdReturn.rid, 2, -1 },
927 { LDSW, LdReturn.rid, 3, 1 },
930 #undef EXPLICIT_BUBBLES_NEEDED
931 #ifdef EXPLICIT_BUBBLES_NEEDED
933 // MULScc inserts one bubble.
934 // This means it breaks the current group (captured in UltraSparcSchedInfo)
935 // *and occupies all issue slots for the next cycle
937 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
938 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
939 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
940 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
943 // SMULcc inserts between 4 and 18 bubbles, depending on #leading 0s in rs1.
944 // We just model this with a simple average.
946 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
947 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
948 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
949 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
951 // SMULcc inserts between 4 and 19 bubbles, depending on #leading 0s in rs1.
952 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
953 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
954 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
955 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
958 // MULX inserts between 4 and 34 bubbles, depending on #leading 0s in rs1.
960 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
961 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
962 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
963 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
966 // SDIVcc inserts 36 bubbles.
968 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
969 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
970 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
971 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
973 // UDIVcc inserts 37 bubbles.
974 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
975 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
976 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
977 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
980 // SDIVX inserts 68 bubbles.
982 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
983 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
984 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
985 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
988 // UDIVX inserts 68 bubbles.
990 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
991 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
992 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
993 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
996 // WR inserts 4 bubbles.
998 //{ WR, AllIssueSlots.rid, 2, 68-1 },
999 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1000 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1001 //{ WR, AllIssueSlots.rid, 2, 68-1 },
1004 // WRPR inserts 4 bubbles.
1006 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1007 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1008 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1009 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
1012 // DONE inserts 9 bubbles.
1014 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1015 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1016 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1017 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
1020 // RETRY inserts 9 bubbles.
1022 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1023 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1024 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1025 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1027 #endif /*EXPLICIT_BUBBLES_NEEDED */
1032 // Additional delays to be captured in code:
1033 // 1. RDPR from several state registers (page 349)
1034 // 2. RD from *any* register (page 349)
1035 // 3. Writes to TICK, PSTATE, TL registers and FLUSH{W} instr (page 349)
1036 // 4. Integer store can be in same group as instr producing value to store.
1037 // 5. BICC and BPICC can be in the same group as instr producing CC (pg 350)
1038 // 6. FMOVr cannot be in the same or next group as an IEU instr (pg 351).
1039 // 7. The second instr. of a CTI group inserts 9 bubbles (pg 351)
1040 // 8. WR{PR}, SVAE, SAVED, RESTORE, RESTORED, RETURN, RETRY, and DONE that
1041 // follow an annulling branch cannot be issued in the same group or in
1042 // the 3 groups following the branch.
1043 // 9. A predicted annulled load does not stall dependent instructions.
1044 // Other annulled delay slot instructions *do* stall dependents, so
1045 // nothing special needs to be done for them during scheduling.
1046 //10. Do not put a load use that may be annulled in the same group as the
1047 // branch. The group will stall until the load returns.
1048 //11. Single-prec. FP loads lock 2 registers, for dependency checking.
1051 // Additional delays we cannot or will not capture:
1052 // 1. If DCTI is last word of cache line, it is delayed until next line can be
1053 // fetched. Also, other DCTI alignment-related delays (pg 352)
1054 // 2. Load-after-store is delayed by 7 extra cycles if load hits in D-Cache.
1055 // Also, several other store-load and load-store conflicts (pg 358)
1056 // 3. MEMBAR, LD{X}FSR, LDD{A} and a bunch of other load stalls (pg 358)
1057 // 4. There can be at most 8 outstanding buffered store instructions
1058 // (including some others like MEMBAR, LDSTUB, CAS{AX}, and FLUSH)
1062 //---------------------------------------------------------------------------
1063 // class UltraSparcSchedInfo
1066 // Interface to instruction scheduling information for UltraSPARC.
1067 // The parameter values above are based on UltraSPARC IIi.
1068 //---------------------------------------------------------------------------
1071 class UltraSparcSchedInfo: public MachineSchedInfo {
1073 /*ctor*/ UltraSparcSchedInfo (const MachineInstrInfo* mii);
1074 /*dtor*/ virtual ~UltraSparcSchedInfo () {}
1076 virtual void initializeResources ();
1080 //---------------------------------------------------------------------------
1081 // class UltraSparcMachine
1084 // Primary interface to machine description for the UltraSPARC.
1085 // Primarily just initializes machine-dependent parameters in
1086 // class TargetMachine, and creates machine-dependent subclasses
1087 // for classes such as InstrInfo, SchedInfo and RegInfo.
1088 //---------------------------------------------------------------------------
1090 class UltraSparc : public TargetMachine {
1092 UltraSparcInstrInfo instrInfo;
1093 UltraSparcSchedInfo schedInfo;
1094 UltraSparcRegInfo regInfo;
1097 virtual ~UltraSparc() {}
1099 virtual const MachineInstrInfo &getInstrInfo() const { return instrInfo; }
1100 virtual const MachineSchedInfo &getSchedInfo() const { return schedInfo; }
1101 virtual const MachineRegInfo &getRegInfo() const { return regInfo; }
1103 // compileMethod - For the sparc, we do instruction selection, followed by
1104 // delay slot scheduling, then register allocation.
1106 virtual bool compileMethod(Method *M);
1109 // emitAssembly - Output assembly language code (a .s file) for the specified
1110 // module. The specified module must have been compiled before this may be
1113 virtual void emitAssembly(const Module *M, ostream &OutStr) const;