[Alignment][NFC] Use Align with TargetLowering::setMinFunctionAlignment
[llvm-core.git] / lib / Target / PowerPC / PPCInstrInfo.h
blob5150650439bc62cc747a5e88efd5a7b4fe8a480e
1 //===-- PPCInstrInfo.h - PowerPC Instruction Information --------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file contains the PowerPC implementation of the TargetInstrInfo class.
11 //===----------------------------------------------------------------------===//
13 #ifndef LLVM_LIB_TARGET_POWERPC_PPCINSTRINFO_H
14 #define LLVM_LIB_TARGET_POWERPC_PPCINSTRINFO_H
16 #include "PPCRegisterInfo.h"
17 #include "llvm/CodeGen/TargetInstrInfo.h"
19 #define GET_INSTRINFO_HEADER
20 #include "PPCGenInstrInfo.inc"
22 namespace llvm {
24 /// PPCII - This namespace holds all of the PowerPC target-specific
25 /// per-instruction flags. These must match the corresponding definitions in
26 /// PPC.td and PPCInstrFormats.td.
27 namespace PPCII {
28 enum {
29 // PPC970 Instruction Flags. These flags describe the characteristics of the
30 // PowerPC 970 (aka G5) dispatch groups and how they are formed out of
31 // raw machine instructions.
33 /// PPC970_First - This instruction starts a new dispatch group, so it will
34 /// always be the first one in the group.
35 PPC970_First = 0x1,
37 /// PPC970_Single - This instruction starts a new dispatch group and
38 /// terminates it, so it will be the sole instruction in the group.
39 PPC970_Single = 0x2,
41 /// PPC970_Cracked - This instruction is cracked into two pieces, requiring
42 /// two dispatch pipes to be available to issue.
43 PPC970_Cracked = 0x4,
45 /// PPC970_Mask/Shift - This is a bitmask that selects the pipeline type that
46 /// an instruction is issued to.
47 PPC970_Shift = 3,
48 PPC970_Mask = 0x07 << PPC970_Shift
50 enum PPC970_Unit {
51 /// These are the various PPC970 execution unit pipelines. Each instruction
52 /// is one of these.
53 PPC970_Pseudo = 0 << PPC970_Shift, // Pseudo instruction
54 PPC970_FXU = 1 << PPC970_Shift, // Fixed Point (aka Integer/ALU) Unit
55 PPC970_LSU = 2 << PPC970_Shift, // Load Store Unit
56 PPC970_FPU = 3 << PPC970_Shift, // Floating Point Unit
57 PPC970_CRU = 4 << PPC970_Shift, // Control Register Unit
58 PPC970_VALU = 5 << PPC970_Shift, // Vector ALU
59 PPC970_VPERM = 6 << PPC970_Shift, // Vector Permute Unit
60 PPC970_BRU = 7 << PPC970_Shift // Branch Unit
63 enum {
64 /// Shift count to bypass PPC970 flags
65 NewDef_Shift = 6,
67 /// This instruction is an X-Form memory operation.
68 XFormMemOp = 0x1 << (NewDef_Shift+1)
70 } // end namespace PPCII
72 // Instructions that have an immediate form might be convertible to that
73 // form if the correct input is a result of a load immediate. In order to
74 // know whether the transformation is special, we might need to know some
75 // of the details of the two forms.
76 struct ImmInstrInfo {
77 // Is the immediate field in the immediate form signed or unsigned?
78 uint64_t SignedImm : 1;
79 // Does the immediate need to be a multiple of some value?
80 uint64_t ImmMustBeMultipleOf : 5;
81 // Is R0/X0 treated specially by the original r+r instruction?
82 // If so, in which operand?
83 uint64_t ZeroIsSpecialOrig : 3;
84 // Is R0/X0 treated specially by the new r+i instruction?
85 // If so, in which operand?
86 uint64_t ZeroIsSpecialNew : 3;
87 // Is the operation commutative?
88 uint64_t IsCommutative : 1;
89 // The operand number to check for add-immediate def.
90 uint64_t OpNoForForwarding : 3;
91 // The operand number for the immediate.
92 uint64_t ImmOpNo : 3;
93 // The opcode of the new instruction.
94 uint64_t ImmOpcode : 16;
95 // The size of the immediate.
96 uint64_t ImmWidth : 5;
97 // The immediate should be truncated to N bits.
98 uint64_t TruncateImmTo : 5;
99 // Is the instruction summing the operand
100 uint64_t IsSummingOperands : 1;
103 // Information required to convert an instruction to just a materialized
104 // immediate.
105 struct LoadImmediateInfo {
106 unsigned Imm : 16;
107 unsigned Is64Bit : 1;
108 unsigned SetCR : 1;
111 class PPCSubtarget;
112 class PPCInstrInfo : public PPCGenInstrInfo {
113 PPCSubtarget &Subtarget;
114 const PPCRegisterInfo RI;
116 void StoreRegToStackSlot(MachineFunction &MF, unsigned SrcReg, bool isKill,
117 int FrameIdx, const TargetRegisterClass *RC,
118 SmallVectorImpl<MachineInstr *> &NewMIs) const;
119 void LoadRegFromStackSlot(MachineFunction &MF, const DebugLoc &DL,
120 unsigned DestReg, int FrameIdx,
121 const TargetRegisterClass *RC,
122 SmallVectorImpl<MachineInstr *> &NewMIs) const;
124 // If the inst has imm-form and one of its operand is produced by a LI,
125 // put the imm into the inst directly and remove the LI if possible.
126 bool transformToImmFormFedByLI(MachineInstr &MI, const ImmInstrInfo &III,
127 unsigned ConstantOpNo, MachineInstr &DefMI,
128 int64_t Imm) const;
129 // If the inst has imm-form and one of its operand is produced by an
130 // add-immediate, try to transform it when possible.
131 bool transformToImmFormFedByAdd(MachineInstr &MI, const ImmInstrInfo &III,
132 unsigned ConstantOpNo, MachineInstr &DefMI,
133 bool KillDefMI) const;
134 // Try to find that, if the instruction 'MI' contains any operand that
135 // could be forwarded from some inst that feeds it. If yes, return the
136 // Def of that operand. And OpNoForForwarding is the operand index in
137 // the 'MI' for that 'Def'. If we see another use of this Def between
138 // the Def and the MI, SeenIntermediateUse becomes 'true'.
139 MachineInstr *getForwardingDefMI(MachineInstr &MI,
140 unsigned &OpNoForForwarding,
141 bool &SeenIntermediateUse) const;
143 // Can the user MI have it's source at index \p OpNoForForwarding
144 // forwarded from an add-immediate that feeds it?
145 bool isUseMIElgibleForForwarding(MachineInstr &MI, const ImmInstrInfo &III,
146 unsigned OpNoForForwarding) const;
147 bool isDefMIElgibleForForwarding(MachineInstr &DefMI,
148 const ImmInstrInfo &III,
149 MachineOperand *&ImmMO,
150 MachineOperand *&RegMO) const;
151 bool isImmElgibleForForwarding(const MachineOperand &ImmMO,
152 const MachineInstr &DefMI,
153 const ImmInstrInfo &III,
154 int64_t &Imm) const;
155 bool isRegElgibleForForwarding(const MachineOperand &RegMO,
156 const MachineInstr &DefMI,
157 const MachineInstr &MI, bool KillDefMI,
158 bool &IsFwdFeederRegKilled) const;
159 const unsigned *getStoreOpcodesForSpillArray() const;
160 const unsigned *getLoadOpcodesForSpillArray() const;
161 virtual void anchor();
163 protected:
164 /// Commutes the operands in the given instruction.
165 /// The commutable operands are specified by their indices OpIdx1 and OpIdx2.
167 /// Do not call this method for a non-commutable instruction or for
168 /// non-commutable pair of operand indices OpIdx1 and OpIdx2.
169 /// Even though the instruction is commutable, the method may still
170 /// fail to commute the operands, null pointer is returned in such cases.
172 /// For example, we can commute rlwimi instructions, but only if the
173 /// rotate amt is zero. We also have to munge the immediates a bit.
174 MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
175 unsigned OpIdx1,
176 unsigned OpIdx2) const override;
178 public:
179 explicit PPCInstrInfo(PPCSubtarget &STI);
181 /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info. As
182 /// such, whenever a client has an instance of instruction info, it should
183 /// always be able to get register info as well (through this method).
185 const PPCRegisterInfo &getRegisterInfo() const { return RI; }
187 bool isXFormMemOp(unsigned Opcode) const {
188 return get(Opcode).TSFlags & PPCII::XFormMemOp;
190 static bool isSameClassPhysRegCopy(unsigned Opcode) {
191 unsigned CopyOpcodes[] =
192 { PPC::OR, PPC::OR8, PPC::FMR, PPC::VOR, PPC::XXLOR, PPC::XXLORf,
193 PPC::XSCPSGNDP, PPC::MCRF, PPC::QVFMR, PPC::QVFMRs, PPC::QVFMRb,
194 PPC::CROR, PPC::EVOR, -1U };
195 for (int i = 0; CopyOpcodes[i] != -1U; i++)
196 if (Opcode == CopyOpcodes[i])
197 return true;
198 return false;
201 ScheduleHazardRecognizer *
202 CreateTargetHazardRecognizer(const TargetSubtargetInfo *STI,
203 const ScheduleDAG *DAG) const override;
204 ScheduleHazardRecognizer *
205 CreateTargetPostRAHazardRecognizer(const InstrItineraryData *II,
206 const ScheduleDAG *DAG) const override;
208 unsigned getInstrLatency(const InstrItineraryData *ItinData,
209 const MachineInstr &MI,
210 unsigned *PredCost = nullptr) const override;
212 int getOperandLatency(const InstrItineraryData *ItinData,
213 const MachineInstr &DefMI, unsigned DefIdx,
214 const MachineInstr &UseMI,
215 unsigned UseIdx) const override;
216 int getOperandLatency(const InstrItineraryData *ItinData,
217 SDNode *DefNode, unsigned DefIdx,
218 SDNode *UseNode, unsigned UseIdx) const override {
219 return PPCGenInstrInfo::getOperandLatency(ItinData, DefNode, DefIdx,
220 UseNode, UseIdx);
223 bool hasLowDefLatency(const TargetSchedModel &SchedModel,
224 const MachineInstr &DefMI,
225 unsigned DefIdx) const override {
226 // Machine LICM should hoist all instructions in low-register-pressure
227 // situations; none are sufficiently free to justify leaving in a loop
228 // body.
229 return false;
232 bool useMachineCombiner() const override {
233 return true;
236 /// Return true when there is potentially a faster code sequence
237 /// for an instruction chain ending in <Root>. All potential patterns are
238 /// output in the <Pattern> array.
239 bool getMachineCombinerPatterns(
240 MachineInstr &Root,
241 SmallVectorImpl<MachineCombinerPattern> &P) const override;
243 bool isAssociativeAndCommutative(const MachineInstr &Inst) const override;
245 bool isCoalescableExtInstr(const MachineInstr &MI,
246 unsigned &SrcReg, unsigned &DstReg,
247 unsigned &SubIdx) const override;
248 unsigned isLoadFromStackSlot(const MachineInstr &MI,
249 int &FrameIndex) const override;
250 bool isReallyTriviallyReMaterializable(const MachineInstr &MI,
251 AliasAnalysis *AA) const override;
252 unsigned isStoreToStackSlot(const MachineInstr &MI,
253 int &FrameIndex) const override;
255 bool findCommutedOpIndices(MachineInstr &MI, unsigned &SrcOpIdx1,
256 unsigned &SrcOpIdx2) const override;
258 void insertNoop(MachineBasicBlock &MBB,
259 MachineBasicBlock::iterator MI) const override;
262 // Branch analysis.
263 bool analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
264 MachineBasicBlock *&FBB,
265 SmallVectorImpl<MachineOperand> &Cond,
266 bool AllowModify) const override;
267 unsigned removeBranch(MachineBasicBlock &MBB,
268 int *BytesRemoved = nullptr) const override;
269 unsigned insertBranch(MachineBasicBlock &MBB, MachineBasicBlock *TBB,
270 MachineBasicBlock *FBB, ArrayRef<MachineOperand> Cond,
271 const DebugLoc &DL,
272 int *BytesAdded = nullptr) const override;
274 // Select analysis.
275 bool canInsertSelect(const MachineBasicBlock &, ArrayRef<MachineOperand> Cond,
276 unsigned, unsigned, int &, int &, int &) const override;
277 void insertSelect(MachineBasicBlock &MBB, MachineBasicBlock::iterator MI,
278 const DebugLoc &DL, unsigned DstReg,
279 ArrayRef<MachineOperand> Cond, unsigned TrueReg,
280 unsigned FalseReg) const override;
282 void copyPhysReg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
283 const DebugLoc &DL, unsigned DestReg, unsigned SrcReg,
284 bool KillSrc) const override;
286 void storeRegToStackSlot(MachineBasicBlock &MBB,
287 MachineBasicBlock::iterator MBBI,
288 unsigned SrcReg, bool isKill, int FrameIndex,
289 const TargetRegisterClass *RC,
290 const TargetRegisterInfo *TRI) const override;
292 void loadRegFromStackSlot(MachineBasicBlock &MBB,
293 MachineBasicBlock::iterator MBBI,
294 unsigned DestReg, int FrameIndex,
295 const TargetRegisterClass *RC,
296 const TargetRegisterInfo *TRI) const override;
298 unsigned getStoreOpcodeForSpill(unsigned Reg,
299 const TargetRegisterClass *RC = nullptr) const;
301 unsigned getLoadOpcodeForSpill(unsigned Reg,
302 const TargetRegisterClass *RC = nullptr) const;
304 bool
305 reverseBranchCondition(SmallVectorImpl<MachineOperand> &Cond) const override;
307 bool FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI, unsigned Reg,
308 MachineRegisterInfo *MRI) const override;
310 // If conversion by predication (only supported by some branch instructions).
311 // All of the profitability checks always return true; it is always
312 // profitable to use the predicated branches.
313 bool isProfitableToIfCvt(MachineBasicBlock &MBB,
314 unsigned NumCycles, unsigned ExtraPredCycles,
315 BranchProbability Probability) const override {
316 return true;
319 bool isProfitableToIfCvt(MachineBasicBlock &TMBB,
320 unsigned NumT, unsigned ExtraT,
321 MachineBasicBlock &FMBB,
322 unsigned NumF, unsigned ExtraF,
323 BranchProbability Probability) const override;
325 bool isProfitableToDupForIfCvt(MachineBasicBlock &MBB, unsigned NumCycles,
326 BranchProbability Probability) const override {
327 return true;
330 bool isProfitableToUnpredicate(MachineBasicBlock &TMBB,
331 MachineBasicBlock &FMBB) const override {
332 return false;
335 // Predication support.
336 bool isPredicated(const MachineInstr &MI) const override;
338 bool isUnpredicatedTerminator(const MachineInstr &MI) const override;
340 bool PredicateInstruction(MachineInstr &MI,
341 ArrayRef<MachineOperand> Pred) const override;
343 bool SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
344 ArrayRef<MachineOperand> Pred2) const override;
346 bool DefinesPredicate(MachineInstr &MI,
347 std::vector<MachineOperand> &Pred) const override;
349 bool isPredicable(const MachineInstr &MI) const override;
351 // Comparison optimization.
353 bool analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
354 unsigned &SrcReg2, int &Mask, int &Value) const override;
356 bool optimizeCompareInstr(MachineInstr &CmpInstr, unsigned SrcReg,
357 unsigned SrcReg2, int Mask, int Value,
358 const MachineRegisterInfo *MRI) const override;
361 /// Return true if get the base operand, byte offset of an instruction and
362 /// the memory width. Width is the size of memory that is being
363 /// loaded/stored (e.g. 1, 2, 4, 8).
364 bool getMemOperandWithOffsetWidth(const MachineInstr &LdSt,
365 const MachineOperand *&BaseOp,
366 int64_t &Offset, unsigned &Width,
367 const TargetRegisterInfo *TRI) const;
369 /// Return true if two MIs access different memory addresses and false
370 /// otherwise
371 bool
372 areMemAccessesTriviallyDisjoint(const MachineInstr &MIa,
373 const MachineInstr &MIb,
374 AliasAnalysis *AA = nullptr) const override;
376 /// GetInstSize - Return the number of bytes of code the specified
377 /// instruction may be. This returns the maximum number of bytes.
379 unsigned getInstSizeInBytes(const MachineInstr &MI) const override;
381 void getNoop(MCInst &NopInst) const override;
383 std::pair<unsigned, unsigned>
384 decomposeMachineOperandsTargetFlags(unsigned TF) const override;
386 ArrayRef<std::pair<unsigned, const char *>>
387 getSerializableDirectMachineOperandTargetFlags() const override;
389 ArrayRef<std::pair<unsigned, const char *>>
390 getSerializableBitmaskMachineOperandTargetFlags() const override;
392 // Expand VSX Memory Pseudo instruction to either a VSX or a FP instruction.
393 bool expandVSXMemPseudo(MachineInstr &MI) const;
395 // Lower pseudo instructions after register allocation.
396 bool expandPostRAPseudo(MachineInstr &MI) const override;
398 static bool isVFRegister(unsigned Reg) {
399 return Reg >= PPC::VF0 && Reg <= PPC::VF31;
401 static bool isVRRegister(unsigned Reg) {
402 return Reg >= PPC::V0 && Reg <= PPC::V31;
404 const TargetRegisterClass *updatedRC(const TargetRegisterClass *RC) const;
405 static int getRecordFormOpcode(unsigned Opcode);
407 bool isTOCSaveMI(const MachineInstr &MI) const;
409 bool isSignOrZeroExtended(const MachineInstr &MI, bool SignExt,
410 const unsigned PhiDepth) const;
412 /// Return true if the output of the instruction is always a sign-extended,
413 /// i.e. 0 to 31-th bits are same as 32-th bit.
414 bool isSignExtended(const MachineInstr &MI, const unsigned depth = 0) const {
415 return isSignOrZeroExtended(MI, true, depth);
418 /// Return true if the output of the instruction is always zero-extended,
419 /// i.e. 0 to 31-th bits are all zeros
420 bool isZeroExtended(const MachineInstr &MI, const unsigned depth = 0) const {
421 return isSignOrZeroExtended(MI, false, depth);
424 bool convertToImmediateForm(MachineInstr &MI,
425 MachineInstr **KilledDef = nullptr) const;
427 /// Fixup killed/dead flag for register \p RegNo between instructions [\p
428 /// StartMI, \p EndMI]. Some PostRA transformations may violate register
429 /// killed/dead flags semantics, this function can be called to fix up. Before
430 /// calling this function,
431 /// 1. Ensure that \p RegNo liveness is killed after instruction \p EndMI.
432 /// 2. Ensure that there is no new definition between (\p StartMI, \p EndMI)
433 /// and possible definition for \p RegNo is \p StartMI or \p EndMI.
434 /// 3. Ensure that all instructions between [\p StartMI, \p EndMI] are in same
435 /// basic block.
436 void fixupIsDeadOrKill(MachineInstr &StartMI, MachineInstr &EndMI,
437 unsigned RegNo) const;
438 void replaceInstrWithLI(MachineInstr &MI, const LoadImmediateInfo &LII) const;
439 void replaceInstrOperandWithImm(MachineInstr &MI, unsigned OpNo,
440 int64_t Imm) const;
442 bool instrHasImmForm(unsigned Opc, bool IsVFReg, ImmInstrInfo &III,
443 bool PostRA) const;
445 // In PostRA phase, try to find instruction defines \p Reg before \p MI.
446 // \p SeenIntermediate is set to true if uses between DefMI and \p MI exist.
447 MachineInstr *getDefMIPostRA(unsigned Reg, MachineInstr &MI,
448 bool &SeenIntermediateUse) const;
450 /// getRegNumForOperand - some operands use different numbering schemes
451 /// for the same registers. For example, a VSX instruction may have any of
452 /// vs0-vs63 allocated whereas an Altivec instruction could only have
453 /// vs32-vs63 allocated (numbered as v0-v31). This function returns the actual
454 /// register number needed for the opcode/operand number combination.
455 /// The operand number argument will be useful when we need to extend this
456 /// to instructions that use both Altivec and VSX numbering (for different
457 /// operands).
458 static unsigned getRegNumForOperand(const MCInstrDesc &Desc, unsigned Reg,
459 unsigned OpNo) {
460 int16_t regClass = Desc.OpInfo[OpNo].RegClass;
461 switch (regClass) {
462 // We store F0-F31, VF0-VF31 in MCOperand and it should be F0-F31,
463 // VSX32-VSX63 during encoding/disassembling
464 case PPC::VSSRCRegClassID:
465 case PPC::VSFRCRegClassID:
466 if (isVFRegister(Reg))
467 return PPC::VSX32 + (Reg - PPC::VF0);
468 break;
469 // We store VSL0-VSL31, V0-V31 in MCOperand and it should be VSL0-VSL31,
470 // VSX32-VSX63 during encoding/disassembling
471 case PPC::VSRCRegClassID:
472 if (isVRRegister(Reg))
473 return PPC::VSX32 + (Reg - PPC::V0);
474 break;
475 // Other RegClass doesn't need mapping
476 default:
477 break;
479 return Reg;
482 /// Check \p Opcode is BDNZ (Decrement CTR and branch if it is still nonzero).
483 bool isBDNZ(unsigned Opcode) const;
485 /// Find the hardware loop instruction used to set-up the specified loop.
486 /// On PPC, we have two instructions used to set-up the hardware loop
487 /// (MTCTRloop, MTCTR8loop) with corresponding endloop (BDNZ, BDNZ8)
488 /// instructions to indicate the end of a loop.
489 MachineInstr *findLoopInstr(MachineBasicBlock &PreHeader) const;
491 /// Analyze the loop code to find the loop induction variable and compare used
492 /// to compute the number of iterations. Currently, we analyze loop that are
493 /// controlled using hardware loops. In this case, the induction variable
494 /// instruction is null. For all other cases, this function returns true,
495 /// which means we're unable to analyze it. \p IndVarInst and \p CmpInst will
496 /// return new values when we can analyze the readonly loop \p L, otherwise,
497 /// nothing got changed
498 bool analyzeLoop(MachineLoop &L, MachineInstr *&IndVarInst,
499 MachineInstr *&CmpInst) const override;
500 /// Generate code to reduce the loop iteration by one and check if the loop
501 /// is finished. Return the value/register of the new loop count. We need
502 /// this function when peeling off one or more iterations of a loop. This
503 /// function assumes the last iteration is peeled first.
504 unsigned reduceLoopCount(MachineBasicBlock &MBB, MachineBasicBlock &PreHeader,
505 MachineInstr *IndVar, MachineInstr &Cmp,
506 SmallVectorImpl<MachineOperand> &Cond,
507 SmallVectorImpl<MachineInstr *> &PrevInsts,
508 unsigned Iter, unsigned MaxIter) const override;
513 #endif