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 unsigned 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");
225 MachineInstr * cpReg2RegMI(const unsigned SrcReg, const unsigned DestReg,
226 const int RegType) const;
228 MachineInstr * cpValue2RegMI(Value * Val, const unsigned DestReg,
229 const int RegType) const;
235 UltraSparcRegInfo(const UltraSparc *const USI ) : UltraSparcInfo(USI),
237 NumOfFloatArgRegs(32),
240 MachineRegClassArr.push_back( new SparcIntRegClass(IntRegClassID) );
241 MachineRegClassArr.push_back( new SparcFloatRegClass(FloatRegClassID) );
242 MachineRegClassArr.push_back( new SparcIntCCRegClass(IntCCRegClassID) );
243 MachineRegClassArr.push_back( new SparcFloatCCRegClass(FloatCCRegClassID));
245 assert( SparcFloatRegOrder::StartOfNonVolatileRegs == 32 &&
246 "32 Float regs are used for float arg passing");
250 ~UltraSparcRegInfo(void) { } // empty destructor
253 inline const UltraSparc & getUltraSparcInfo() const {
254 return *UltraSparcInfo;
259 inline unsigned getRegClassIDOfValue (const Value *const Val,
260 bool isCCReg = false) const {
262 Type::PrimitiveID ty = (Val->getType())->getPrimitiveID();
266 if( (ty && ty <= Type::LongTyID) || (ty == Type::LabelTyID) ||
267 (ty == Type::MethodTyID) || (ty == Type::PointerTyID) )
268 res = IntRegClassID; // sparc int reg (ty=0: void)
269 else if( ty <= Type::DoubleTyID)
270 res = FloatRegClassID; // sparc float reg class
272 cout << "TypeID: " << ty << endl;
273 assert(0 && "Cannot resolve register class for type");
278 return res + 2; // corresponidng condition code regiser
285 // returns the register tha contains always zero
286 // this is the unified register number
287 inline int getZeroRegNum() const { return SparcIntRegOrder::g0; }
289 // returns the reg used for pushing the address when a method is called.
290 // This can be used for other purposes between calls
291 unsigned getCallAddressReg() const { return SparcIntRegOrder::o7; }
294 // and when we return from a method. It should be made sure that this
295 // register contains the return value when a return instruction is reached.
296 unsigned getReturnAddressReg() const { return SparcIntRegOrder::i7; }
298 void suggestRegs4MethodArgs(const Method *const Meth,
299 LiveRangeInfo& LRI) const;
301 void suggestRegs4CallArgs(const MachineInstr *const CallMI,
302 LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
304 void suggestReg4RetValue(const MachineInstr *const RetMI,
305 LiveRangeInfo& LRI ) const;
308 void colorMethodArgs(const Method *const Meth, LiveRangeInfo& LRI,
309 AddedInstrns *const FirstAI) const;
311 void colorCallArgs(const MachineInstr *const CallMI, LiveRangeInfo& LRI,
312 AddedInstrns *const CallAI) const;
314 void colorRetValue(const MachineInstr *const RetI, LiveRangeInfo& LRI,
315 AddedInstrns *const RetAI) const;
318 // bool handleSpecialMInstr(const MachineInstr * MInst,
319 // LiveRangeInfo& LRI, vector<RegClass *> RCL) const;
322 static void printReg(const LiveRange *const LR) ;
324 // this method provides a unique number for each register
325 inline int getUnifiedRegNum(int RegClassID, int reg) const {
327 if( RegClassID == IntRegClassID && reg < 32 )
329 else if ( RegClassID == FloatRegClassID && reg < 64)
330 return reg + 32; // we have 32 int regs
331 else if( RegClassID == FloatCCRegClassID && reg < 4)
332 return reg + 32 + 64; // 32 int, 64 float
333 else if( RegClassID == IntCCRegClassID )
334 return 4+ 32 + 64; // only int cc reg
335 else if (reg==1000) //****** TODO: Remove
338 assert(0 && "Invalid register class or reg number");
342 // given the unified register number, this gives the name
343 inline const string getUnifiedRegName(int reg) const {
345 return SparcIntRegOrder::getRegName(reg);
346 else if ( reg < (64 + 32) )
347 return SparcFloatRegOrder::getRegName( reg - 32);
348 else if( reg < (64+32+4) )
349 return SparcFloatCCRegOrder::getRegName( reg -32 - 64);
350 else if ( reg == 64+32+4)
351 return "xcc"; // only integer cc reg
353 else if (reg==1000) //****** TODO: Remove
356 assert(0 && "Invalid register number");
364 /*---------------------------------------------------------------------------
365 Scheduling guidelines for SPARC IIi:
367 I-Cache alignment rules (pg 326)
368 -- Align a branch target instruction so that it's entire group is within
369 the same cache line (may be 1-4 instructions).
370 ** Don't let a branch that is predicted taken be the last instruction
371 on an I-cache line: delay slot will need an entire line to be fetched
372 -- Make a FP instruction or a branch be the 4th instruction in a group.
373 For branches, there are tradeoffs in reordering to make this happen
375 ** Don't put a branch in a group that crosses a 32-byte boundary!
376 An artificial branch is inserted after every 32 bytes, and having
377 another branch will force the group to be broken into 2 groups.
380 -- Don't let a loop span two memory pages, if possible
382 Branch prediction performance:
383 -- Don't make the branch in a delay slot the target of a branch
384 -- Try not to have 2 predicted branches within a group of 4 instructions
385 (because each such group has a single branch target field).
386 -- Try to align branches in slots 0, 2, 4 or 6 of a cache line (to avoid
387 the wrong prediction bits being used in some cases).
389 D-Cache timing constraints:
390 -- Signed int loads of less than 64 bits have 3 cycle latency, not 2
391 -- All other loads that hit in D-Cache have 2 cycle latency
392 -- All loads are returned IN ORDER, so a D-Cache miss will delay a later hit
393 -- Mis-aligned loads or stores cause a trap. In particular, replace
394 mis-aligned FP double precision l/s with 2 single-precision l/s.
395 -- Simulations of integer codes show increase in avg. group size of
396 33% when code (including esp. non-faulting loads) is moved across
397 one branch, and 50% across 2 branches.
399 E-Cache timing constraints:
400 -- Scheduling for E-cache (D-Cache misses) is effective (due to load buffering)
402 Store buffer timing constraints:
403 -- Stores can be executed in same cycle as instruction producing the value
404 -- Stores are buffered and have lower priority for E-cache until
405 highwater mark is reached in the store buffer (5 stores)
407 Pipeline constraints:
408 -- Shifts can only use IEU0.
409 -- CC setting instructions can only use IEU1.
410 -- Several other instructions must only use IEU1:
411 EDGE(?), ARRAY(?), CALL, JMPL, BPr, PST, and FCMP.
412 -- Two instructions cannot store to the same register file in a single cycle
413 (single write port per file).
415 Issue and grouping constraints:
416 -- FP and branch instructions must use slot 4.
417 -- Shift instructions cannot be grouped with other IEU0-specific instructions.
418 -- CC setting instructions cannot be grouped with other IEU1-specific instrs.
419 -- Several instructions must be issued in a single-instruction group:
420 MOVcc or MOVr, MULs/x and DIVs/x, SAVE/RESTORE, many others
421 -- A CALL or JMPL breaks a group, ie, is not combined with subsequent instrs.
425 Branch delay slot scheduling rules:
426 -- A CTI couple (two back-to-back CTI instructions in the dynamic stream)
427 has a 9-instruction penalty: the entire pipeline is flushed when the
428 second instruction reaches stage 9 (W-Writeback).
429 -- Avoid putting multicycle instructions, and instructions that may cause
430 load misses, in the delay slot of an annulling branch.
431 -- Avoid putting WR, SAVE..., RESTORE and RETURN instructions in the
432 delay slot of an annulling branch.
434 *--------------------------------------------------------------------------- */
436 //---------------------------------------------------------------------------
437 // List of CPUResources for UltraSPARC IIi.
438 //---------------------------------------------------------------------------
440 const CPUResource AllIssueSlots( "All Instr Slots", 4);
441 const CPUResource IntIssueSlots( "Int Instr Slots", 3);
442 const CPUResource First3IssueSlots("Instr Slots 0-3", 3);
443 const CPUResource LSIssueSlots( "Load-Store Instr Slot", 1);
444 const CPUResource CTIIssueSlots( "Ctrl Transfer Instr Slot", 1);
445 const CPUResource FPAIssueSlots( "Int Instr Slot 1", 1);
446 const CPUResource FPMIssueSlots( "Int Instr Slot 1", 1);
448 // IEUN instructions can use either Alu and should use IAluN.
449 // IEU0 instructions must use Alu 1 and should use both IAluN and IAlu0.
450 // IEU1 instructions must use Alu 2 and should use both IAluN and IAlu1.
451 const CPUResource IAluN("Int ALU 1or2", 2);
452 const CPUResource IAlu0("Int ALU 1", 1);
453 const CPUResource IAlu1("Int ALU 2", 1);
455 const CPUResource LSAluC1("Load/Store Unit Addr Cycle", 1);
456 const CPUResource LSAluC2("Load/Store Unit Issue Cycle", 1);
457 const CPUResource LdReturn("Load Return Unit", 1);
459 const CPUResource FPMAluC1("FP Mul/Div Alu Cycle 1", 1);
460 const CPUResource FPMAluC2("FP Mul/Div Alu Cycle 2", 1);
461 const CPUResource FPMAluC3("FP Mul/Div Alu Cycle 3", 1);
463 const CPUResource FPAAluC1("FP Other Alu Cycle 1", 1);
464 const CPUResource FPAAluC2("FP Other Alu Cycle 2", 1);
465 const CPUResource FPAAluC3("FP Other Alu Cycle 3", 1);
467 const CPUResource IRegReadPorts("Int Reg ReadPorts", INT_MAX); // CHECK
468 const CPUResource IRegWritePorts("Int Reg WritePorts", 2); // CHECK
469 const CPUResource FPRegReadPorts("FP Reg Read Ports", INT_MAX); // CHECK
470 const CPUResource FPRegWritePorts("FP Reg Write Ports", 1); // CHECK
472 const CPUResource CTIDelayCycle( "CTI delay cycle", 1);
473 const CPUResource FCMPDelayCycle("FCMP delay cycle", 1);
476 //---------------------------------------------------------------------------
477 // const InstrClassRUsage SparcRUsageDesc[]
480 // Resource usage information for instruction in each scheduling class.
481 // The InstrRUsage Objects for individual classes are specified first.
482 // Note that fetch and decode are decoupled from the execution pipelines
483 // via an instr buffer, so they are not included in the cycles below.
484 //---------------------------------------------------------------------------
486 const InstrClassRUsage NoneClassRUsage = {
491 /* isSingleIssue */ false,
492 /* breaksGroup */ false,
496 /* feasibleSlots[] */ { 0, 1, 2, 3 },
510 const InstrClassRUsage IEUNClassRUsage = {
515 /* isSingleIssue */ false,
516 /* breaksGroup */ false,
520 /* feasibleSlots[] */ { 0, 1, 2 },
524 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
525 { IntIssueSlots.rid, 0, 1 },
526 /*Cycle E */ { IAluN.rid, 1, 1 },
531 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
535 const InstrClassRUsage IEU0ClassRUsage = {
540 /* isSingleIssue */ false,
541 /* breaksGroup */ false,
545 /* feasibleSlots[] */ { 0, 1, 2 },
549 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
550 { IntIssueSlots.rid, 0, 1 },
551 /*Cycle E */ { IAluN.rid, 1, 1 },
557 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
561 const InstrClassRUsage IEU1ClassRUsage = {
566 /* isSingleIssue */ false,
567 /* breaksGroup */ false,
571 /* feasibleSlots[] */ { 0, 1, 2 },
575 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
576 { IntIssueSlots.rid, 0, 1 },
577 /*Cycle E */ { IAluN.rid, 1, 1 },
583 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
587 const InstrClassRUsage FPMClassRUsage = {
592 /* isSingleIssue */ false,
593 /* breaksGroup */ false,
597 /* feasibleSlots[] */ { 0, 1, 2, 3 },
601 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
602 { FPMIssueSlots.rid, 0, 1 },
603 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
604 /*Cycle C */ { FPMAluC1.rid, 2, 1 },
605 /*Cycle N1*/ { FPMAluC2.rid, 3, 1 },
606 /*Cycle N1*/ { FPMAluC3.rid, 4, 1 },
608 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
612 const InstrClassRUsage FPAClassRUsage = {
617 /* isSingleIssue */ false,
618 /* breaksGroup */ false,
622 /* feasibleSlots[] */ { 0, 1, 2, 3 },
626 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
627 { FPAIssueSlots.rid, 0, 1 },
628 /*Cycle E */ { FPRegReadPorts.rid, 1, 1 },
629 /*Cycle C */ { FPAAluC1.rid, 2, 1 },
630 /*Cycle N1*/ { FPAAluC2.rid, 3, 1 },
631 /*Cycle N1*/ { FPAAluC3.rid, 4, 1 },
633 /*Cycle W */ { FPRegWritePorts.rid, 6, 1 }
637 const InstrClassRUsage LDClassRUsage = {
642 /* isSingleIssue */ false,
643 /* breaksGroup */ false,
647 /* feasibleSlots[] */ { 0, 1, 2, },
651 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
652 { First3IssueSlots.rid, 0, 1 },
653 { LSIssueSlots.rid, 0, 1 },
654 /*Cycle E */ { LSAluC1.rid, 1, 1 },
655 /*Cycle C */ { LSAluC2.rid, 2, 1 },
656 { LdReturn.rid, 2, 1 },
660 /*Cycle W */ { IRegWritePorts.rid, 6, 1 }
664 const InstrClassRUsage STClassRUsage = {
669 /* isSingleIssue */ false,
670 /* breaksGroup */ false,
674 /* feasibleSlots[] */ { 0, 1, 2 },
678 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
679 { First3IssueSlots.rid, 0, 1 },
680 { LSIssueSlots.rid, 0, 1 },
681 /*Cycle E */ { LSAluC1.rid, 1, 1 },
682 /*Cycle C */ { LSAluC2.rid, 2, 1 }
690 const InstrClassRUsage CTIClassRUsage = {
695 /* isSingleIssue */ false,
696 /* breaksGroup */ false,
700 /* feasibleSlots[] */ { 0, 1, 2, 3 },
704 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
705 { CTIIssueSlots.rid, 0, 1 },
706 /*Cycle E */ { IAlu0.rid, 1, 1 },
707 /*Cycles E-C */ { CTIDelayCycle.rid, 1, 2 }
716 const InstrClassRUsage SingleClassRUsage = {
721 /* isSingleIssue */ true,
722 /* breaksGroup */ false,
726 /* feasibleSlots[] */ { 0 },
730 /*Cycle G */ { AllIssueSlots.rid, 0, 1 },
731 { AllIssueSlots.rid, 0, 1 },
732 { AllIssueSlots.rid, 0, 1 },
733 { AllIssueSlots.rid, 0, 1 },
734 /*Cycle E */ { IAlu0.rid, 1, 1 }
744 const InstrClassRUsage SparcRUsageDesc[] = {
758 //---------------------------------------------------------------------------
759 // const InstrIssueDelta SparcInstrIssueDeltas[]
762 // Changes to issue restrictions information in InstrClassRUsage for
763 // instructions that differ from other instructions in their class.
764 //---------------------------------------------------------------------------
766 const InstrIssueDelta SparcInstrIssueDeltas[] = {
768 // opCode, isSingleIssue, breaksGroup, numBubbles
770 // Special cases for single-issue only
771 // Other single issue cases are below.
772 //{ LDDA, true, true, 0 },
773 //{ STDA, true, true, 0 },
774 //{ LDDF, true, true, 0 },
775 //{ LDDFA, true, true, 0 },
776 { ADDC, true, true, 0 },
777 { ADDCcc, true, true, 0 },
778 { SUBC, true, true, 0 },
779 { SUBCcc, true, true, 0 },
780 //{ SAVE, true, true, 0 },
781 //{ RESTORE, true, true, 0 },
782 //{ LDSTUB, true, true, 0 },
783 //{ SWAP, true, true, 0 },
784 //{ SWAPA, true, true, 0 },
785 //{ CAS, true, true, 0 },
786 //{ CASA, true, true, 0 },
787 //{ CASX, true, true, 0 },
788 //{ CASXA, true, true, 0 },
789 //{ LDFSR, true, true, 0 },
790 //{ LDFSRA, true, true, 0 },
791 //{ LDXFSR, true, true, 0 },
792 //{ LDXFSRA, true, true, 0 },
793 //{ STFSR, true, true, 0 },
794 //{ STFSRA, true, true, 0 },
795 //{ STXFSR, true, true, 0 },
796 //{ STXFSRA, true, true, 0 },
797 //{ SAVED, true, true, 0 },
798 //{ RESTORED, true, true, 0 },
799 //{ FLUSH, true, true, 9 },
800 //{ FLUSHW, true, true, 9 },
801 //{ ALIGNADDR, true, true, 0 },
802 { RETURN, true, true, 0 },
803 //{ DONE, true, true, 0 },
804 //{ RETRY, true, true, 0 },
805 //{ WR, true, true, 0 },
806 //{ WRPR, true, true, 4 },
807 //{ RD, true, true, 0 },
808 //{ RDPR, true, true, 0 },
809 //{ TCC, true, true, 0 },
810 //{ SHUTDOWN, true, true, 0 },
812 // Special cases for breaking group *before*
813 // CURRENTLY NOT SUPPORTED!
814 { CALL, false, false, 0 },
815 { JMPL, false, false, 0 },
817 // Special cases for breaking the group *after*
818 { MULX, true, true, (4+34)/2 },
819 { FDIVS, false, true, 0 },
820 { FDIVD, false, true, 0 },
821 { FDIVQ, false, true, 0 },
822 { FSQRTS, false, true, 0 },
823 { FSQRTD, false, true, 0 },
824 { FSQRTQ, false, true, 0 },
825 //{ FCMP{LE,GT,NE,EQ}, false, true, 0 },
827 // Instructions that introduce bubbles
828 //{ MULScc, true, true, 2 },
829 //{ SMULcc, true, true, (4+18)/2 },
830 //{ UMULcc, true, true, (4+19)/2 },
831 { SDIVX, true, true, 68 },
832 { UDIVX, true, true, 68 },
833 //{ SDIVcc, true, true, 36 },
834 //{ UDIVcc, true, true, 37 },
835 //{ WR, false, false, 4 },
836 //{ WRPR, false, false, 4 },
840 //---------------------------------------------------------------------------
841 // const InstrRUsageDelta SparcInstrUsageDeltas[]
844 // Changes to resource usage information in InstrClassRUsage for
845 // instructions that differ from other instructions in their class.
846 //---------------------------------------------------------------------------
848 const InstrRUsageDelta SparcInstrUsageDeltas[] = {
850 // MachineOpCode, Resource, Start cycle, Num cycles
853 // JMPL counts as a load/store instruction for issue!
855 { JMPL, LSIssueSlots.rid, 0, 1 },
858 // Many instructions cannot issue for the next 2 cycles after an FCMP
859 // We model that with a fake resource FCMPDelayCycle.
861 { FCMPS, FCMPDelayCycle.rid, 1, 3 },
862 { FCMPD, FCMPDelayCycle.rid, 1, 3 },
863 { FCMPQ, FCMPDelayCycle.rid, 1, 3 },
865 { MULX, FCMPDelayCycle.rid, 1, 1 },
866 { SDIVX, FCMPDelayCycle.rid, 1, 1 },
867 { UDIVX, FCMPDelayCycle.rid, 1, 1 },
868 //{ SMULcc, FCMPDelayCycle.rid, 1, 1 },
869 //{ UMULcc, FCMPDelayCycle.rid, 1, 1 },
870 //{ SDIVcc, FCMPDelayCycle.rid, 1, 1 },
871 //{ UDIVcc, FCMPDelayCycle.rid, 1, 1 },
872 { STD, FCMPDelayCycle.rid, 1, 1 },
873 { FMOVRSZ, FCMPDelayCycle.rid, 1, 1 },
874 { FMOVRSLEZ,FCMPDelayCycle.rid, 1, 1 },
875 { FMOVRSLZ, FCMPDelayCycle.rid, 1, 1 },
876 { FMOVRSNZ, FCMPDelayCycle.rid, 1, 1 },
877 { FMOVRSGZ, FCMPDelayCycle.rid, 1, 1 },
878 { FMOVRSGEZ,FCMPDelayCycle.rid, 1, 1 },
881 // Some instructions are stalled in the GROUP stage if a CTI is in
884 { LDD, CTIDelayCycle.rid, 1, 1 },
885 //{ LDDA, CTIDelayCycle.rid, 1, 1 },
886 //{ LDDSTUB, CTIDelayCycle.rid, 1, 1 },
887 //{ LDDSTUBA, CTIDelayCycle.rid, 1, 1 },
888 //{ SWAP, CTIDelayCycle.rid, 1, 1 },
889 //{ SWAPA, CTIDelayCycle.rid, 1, 1 },
890 //{ CAS, CTIDelayCycle.rid, 1, 1 },
891 //{ CASA, CTIDelayCycle.rid, 1, 1 },
892 //{ CASX, CTIDelayCycle.rid, 1, 1 },
893 //{ CASXA, CTIDelayCycle.rid, 1, 1 },
896 // Signed int loads of less than dword size return data in cycle N1 (not C)
897 // and put all loads in consecutive cycles into delayed load return mode.
899 { LDSB, LdReturn.rid, 2, -1 },
900 { LDSB, LdReturn.rid, 3, 1 },
902 { LDSH, LdReturn.rid, 2, -1 },
903 { LDSH, LdReturn.rid, 3, 1 },
905 { LDSW, LdReturn.rid, 2, -1 },
906 { LDSW, LdReturn.rid, 3, 1 },
909 #undef EXPLICIT_BUBBLES_NEEDED
910 #ifdef EXPLICIT_BUBBLES_NEEDED
912 // MULScc inserts one bubble.
913 // This means it breaks the current group (captured in UltraSparcSchedInfo)
914 // *and occupies all issue slots for the next cycle
916 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
917 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
918 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
919 //{ MULScc, AllIssueSlots.rid, 2, 2-1 },
922 // SMULcc inserts between 4 and 18 bubbles, depending on #leading 0s in rs1.
923 // We just model this with a simple average.
925 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
926 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
927 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
928 //{ SMULcc, AllIssueSlots.rid, 2, ((4+18)/2)-1 },
930 // SMULcc inserts between 4 and 19 bubbles, depending on #leading 0s in rs1.
931 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
932 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
933 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
934 //{ UMULcc, AllIssueSlots.rid, 2, ((4+19)/2)-1 },
937 // MULX inserts between 4 and 34 bubbles, depending on #leading 0s in rs1.
939 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
940 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
941 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
942 { MULX, AllIssueSlots.rid, 2, ((4+34)/2)-1 },
945 // SDIVcc inserts 36 bubbles.
947 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
948 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
949 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
950 //{ SDIVcc, AllIssueSlots.rid, 2, 36-1 },
952 // UDIVcc inserts 37 bubbles.
953 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
954 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
955 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
956 //{ UDIVcc, AllIssueSlots.rid, 2, 37-1 },
959 // SDIVX inserts 68 bubbles.
961 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
962 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
963 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
964 { SDIVX, AllIssueSlots.rid, 2, 68-1 },
967 // UDIVX inserts 68 bubbles.
969 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
970 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
971 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
972 { UDIVX, AllIssueSlots.rid, 2, 68-1 },
975 // WR inserts 4 bubbles.
977 //{ WR, AllIssueSlots.rid, 2, 68-1 },
978 //{ WR, AllIssueSlots.rid, 2, 68-1 },
979 //{ WR, AllIssueSlots.rid, 2, 68-1 },
980 //{ WR, AllIssueSlots.rid, 2, 68-1 },
983 // WRPR inserts 4 bubbles.
985 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
986 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
987 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
988 //{ WRPR, AllIssueSlots.rid, 2, 68-1 },
991 // DONE inserts 9 bubbles.
993 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
994 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
995 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
996 //{ DONE, AllIssueSlots.rid, 2, 9-1 },
999 // RETRY inserts 9 bubbles.
1001 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1002 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1003 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1004 //{ RETRY, AllIssueSlots.rid, 2, 9-1 },
1006 #endif /*EXPLICIT_BUBBLES_NEEDED */
1011 // Additional delays to be captured in code:
1012 // 1. RDPR from several state registers (page 349)
1013 // 2. RD from *any* register (page 349)
1014 // 3. Writes to TICK, PSTATE, TL registers and FLUSH{W} instr (page 349)
1015 // 4. Integer store can be in same group as instr producing value to store.
1016 // 5. BICC and BPICC can be in the same group as instr producing CC (pg 350)
1017 // 6. FMOVr cannot be in the same or next group as an IEU instr (pg 351).
1018 // 7. The second instr. of a CTI group inserts 9 bubbles (pg 351)
1019 // 8. WR{PR}, SVAE, SAVED, RESTORE, RESTORED, RETURN, RETRY, and DONE that
1020 // follow an annulling branch cannot be issued in the same group or in
1021 // the 3 groups following the branch.
1022 // 9. A predicted annulled load does not stall dependent instructions.
1023 // Other annulled delay slot instructions *do* stall dependents, so
1024 // nothing special needs to be done for them during scheduling.
1025 //10. Do not put a load use that may be annulled in the same group as the
1026 // branch. The group will stall until the load returns.
1027 //11. Single-prec. FP loads lock 2 registers, for dependency checking.
1030 // Additional delays we cannot or will not capture:
1031 // 1. If DCTI is last word of cache line, it is delayed until next line can be
1032 // fetched. Also, other DCTI alignment-related delays (pg 352)
1033 // 2. Load-after-store is delayed by 7 extra cycles if load hits in D-Cache.
1034 // Also, several other store-load and load-store conflicts (pg 358)
1035 // 3. MEMBAR, LD{X}FSR, LDD{A} and a bunch of other load stalls (pg 358)
1036 // 4. There can be at most 8 outstanding buffered store instructions
1037 // (including some others like MEMBAR, LDSTUB, CAS{AX}, and FLUSH)
1041 //---------------------------------------------------------------------------
1042 // class UltraSparcSchedInfo
1045 // Interface to instruction scheduling information for UltraSPARC.
1046 // The parameter values above are based on UltraSPARC IIi.
1047 //---------------------------------------------------------------------------
1050 class UltraSparcSchedInfo: public MachineSchedInfo {
1052 /*ctor*/ UltraSparcSchedInfo (const MachineInstrInfo* mii);
1053 /*dtor*/ virtual ~UltraSparcSchedInfo () {}
1055 virtual void initializeResources ();
1059 //---------------------------------------------------------------------------
1060 // class UltraSparcMachine
1063 // Primary interface to machine description for the UltraSPARC.
1064 // Primarily just initializes machine-dependent parameters in
1065 // class TargetMachine, and creates machine-dependent subclasses
1066 // for classes such as InstrInfo, SchedInfo and RegInfo.
1067 //---------------------------------------------------------------------------
1069 class UltraSparc : public TargetMachine {
1071 UltraSparcInstrInfo instrInfo;
1072 UltraSparcSchedInfo schedInfo;
1073 UltraSparcRegInfo regInfo;
1076 virtual ~UltraSparc() {}
1078 virtual const MachineInstrInfo &getInstrInfo() const { return instrInfo; }
1079 virtual const MachineSchedInfo &getSchedInfo() const { return schedInfo; }
1080 virtual const MachineRegInfo &getRegInfo() const { return regInfo; }
1082 // compileMethod - For the sparc, we do instruction selection, followed by
1083 // delay slot scheduling, then register allocation.
1085 virtual bool compileMethod(Method *M);
1088 // emitAssembly - Output assembly language code (a .s file) for the specified
1089 // module. The specified module must have been compiled before this may be
1092 virtual void emitAssembly(const Module *M, ostream &OutStr) const;