1 //===- HexagonInstrInfo.cpp - Hexagon Instruction Information -------------===//
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 Hexagon implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "HexagonInstrInfo.h"
14 #include "Hexagon.h"
15 #include "HexagonFrameLowering.h"
16 #include "HexagonHazardRecognizer.h"
17 #include "HexagonRegisterInfo.h"
18 #include "HexagonSubtarget.h"
19 #include "llvm/ADT/ArrayRef.h"
20 #include "llvm/ADT/SmallPtrSet.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/CodeGen/DFAPacketizer.h"
24 #include "llvm/CodeGen/LivePhysRegs.h"
25 #include "llvm/CodeGen/MachineBasicBlock.h"
26 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
27 #include "llvm/CodeGen/MachineFrameInfo.h"
28 #include "llvm/CodeGen/MachineFunction.h"
29 #include "llvm/CodeGen/MachineInstr.h"
30 #include "llvm/CodeGen/MachineInstrBuilder.h"
31 #include "llvm/CodeGen/MachineInstrBundle.h"
32 #include "llvm/CodeGen/MachineLoopInfo.h"
33 #include "llvm/CodeGen/MachineMemOperand.h"
34 #include "llvm/CodeGen/MachineOperand.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/CodeGen/ScheduleDAG.h"
37 #include "llvm/CodeGen/TargetInstrInfo.h"
38 #include "llvm/CodeGen/TargetOpcodes.h"
39 #include "llvm/CodeGen/TargetRegisterInfo.h"
40 #include "llvm/CodeGen/TargetSubtargetInfo.h"
41 #include "llvm/IR/DebugLoc.h"
42 #include "llvm/MC/MCAsmInfo.h"
43 #include "llvm/MC/MCInstrDesc.h"
44 #include "llvm/MC/MCInstrItineraries.h"
45 #include "llvm/MC/MCRegisterInfo.h"
46 #include "llvm/Support/BranchProbability.h"
47 #include "llvm/Support/CommandLine.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/ErrorHandling.h"
50 #include "llvm/Support/MachineValueType.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/raw_ostream.h"
53 #include "llvm/Target/TargetMachine.h"
54 #include <cassert>
55 #include <cctype>
56 #include <cstdint>
57 #include <cstring>
58 #include <iterator>
59 #include <string>
60 #include <utility>
61 
62 using namespace llvm;
63 
64 #define DEBUG_TYPE "hexagon-instrinfo"
65 
66 #define GET_INSTRINFO_CTOR_DTOR
67 #define GET_INSTRMAP_INFO
68 #include "HexagonDepTimingClasses.h"
69 #include "HexagonGenDFAPacketizer.inc"
70 #include "HexagonGenInstrInfo.inc"
71 
72 cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden,
73   cl::init(false), cl::desc("Do not consider inline-asm a scheduling/"
74                             "packetization boundary."));
75 
76 static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction",
77   cl::Hidden, cl::init(true), cl::desc("Enable branch prediction"));
78 
79 static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule",
80   cl::Hidden, cl::ZeroOrMore, cl::init(false),
81   cl::desc("Disable schedule adjustment for new value stores."));
82 
83 static cl::opt<bool> EnableTimingClassLatency(
84   "enable-timing-class-latency", cl::Hidden, cl::init(false),
85   cl::desc("Enable timing class latency"));
86 
87 static cl::opt<bool> EnableALUForwarding(
88   "enable-alu-forwarding", cl::Hidden, cl::init(true),
89   cl::desc("Enable vec alu forwarding"));
90 
91 static cl::opt<bool> EnableACCForwarding(
92   "enable-acc-forwarding", cl::Hidden, cl::init(true),
93   cl::desc("Enable vec acc forwarding"));
94 
95 static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large",
96   cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm"));
97 
98 static cl::opt<bool> UseDFAHazardRec("dfa-hazard-rec",
99   cl::init(true), cl::Hidden, cl::ZeroOrMore,
100   cl::desc("Use the DFA based hazard recognizer."));
101 
102 /// Constants for Hexagon instructions.
103 const int Hexagon_MEMW_OFFSET_MAX = 4095;
104 const int Hexagon_MEMW_OFFSET_MIN = -4096;
105 const int Hexagon_MEMD_OFFSET_MAX = 8191;
106 const int Hexagon_MEMD_OFFSET_MIN = -8192;
107 const int Hexagon_MEMH_OFFSET_MAX = 2047;
108 const int Hexagon_MEMH_OFFSET_MIN = -2048;
109 const int Hexagon_MEMB_OFFSET_MAX = 1023;
110 const int Hexagon_MEMB_OFFSET_MIN = -1024;
111 const int Hexagon_ADDI_OFFSET_MAX = 32767;
112 const int Hexagon_ADDI_OFFSET_MIN = -32768;
113 
114 // Pin the vtable to this file.
anchor()115 void HexagonInstrInfo::anchor() {}
116 
HexagonInstrInfo(HexagonSubtarget & ST)117 HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST)
118   : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP),
119     Subtarget(ST) {}
120 
121 namespace llvm {
122 namespace HexagonFUnits {
123   bool isSlot0Only(unsigned units);
124 }
125 }
126 
isIntRegForSubInst(unsigned Reg)127 static bool isIntRegForSubInst(unsigned Reg) {
128   return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) ||
129          (Reg >= Hexagon::R16 && Reg <= Hexagon::R23);
130 }
131 
isDblRegForSubInst(unsigned Reg,const HexagonRegisterInfo & HRI)132 static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) {
133   return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_lo)) &&
134          isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_hi));
135 }
136 
137 /// Calculate number of instructions excluding the debug instructions.
nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB,MachineBasicBlock::const_instr_iterator MIE)138 static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB,
139                               MachineBasicBlock::const_instr_iterator MIE) {
140   unsigned Count = 0;
141   for (; MIB != MIE; ++MIB) {
142     if (!MIB->isDebugInstr())
143       ++Count;
144   }
145   return Count;
146 }
147 
148 /// Find the hardware loop instruction used to set-up the specified loop.
149 /// On Hexagon, we have two instructions used to set-up the hardware loop
150 /// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions
151 /// to indicate the end of a loop.
findLoopInstr(MachineBasicBlock * BB,unsigned EndLoopOp,MachineBasicBlock * TargetBB,SmallPtrSet<MachineBasicBlock *,8> & Visited) const152 MachineInstr *HexagonInstrInfo::findLoopInstr(MachineBasicBlock *BB,
153       unsigned EndLoopOp, MachineBasicBlock *TargetBB,
154       SmallPtrSet<MachineBasicBlock *, 8> &Visited) const {
155   unsigned LOOPi;
156   unsigned LOOPr;
157   if (EndLoopOp == Hexagon::ENDLOOP0) {
158     LOOPi = Hexagon::J2_loop0i;
159     LOOPr = Hexagon::J2_loop0r;
160   } else { // EndLoopOp == Hexagon::EndLOOP1
161     LOOPi = Hexagon::J2_loop1i;
162     LOOPr = Hexagon::J2_loop1r;
163   }
164 
165   // The loop set-up instruction will be in a predecessor block
166   for (MachineBasicBlock *PB : BB->predecessors()) {
167     // If this has been visited, already skip it.
168     if (!Visited.insert(PB).second)
169       continue;
170     if (PB == BB)
171       continue;
172     for (auto I = PB->instr_rbegin(), E = PB->instr_rend(); I != E; ++I) {
173       unsigned Opc = I->getOpcode();
174       if (Opc == LOOPi || Opc == LOOPr)
175         return &*I;
176       // We've reached a different loop, which means the loop01 has been
177       // removed.
178       if (Opc == EndLoopOp && I->getOperand(0).getMBB() != TargetBB)
179         return nullptr;
180     }
181     // Check the predecessors for the LOOP instruction.
182     if (MachineInstr *Loop = findLoopInstr(PB, EndLoopOp, TargetBB, Visited))
183       return Loop;
184   }
185   return nullptr;
186 }
187 
188 /// Gather register def/uses from MI.
189 /// This treats possible (predicated) defs as actually happening ones
190 /// (conservatively).
parseOperands(const MachineInstr & MI,SmallVector<unsigned,4> & Defs,SmallVector<unsigned,8> & Uses)191 static inline void parseOperands(const MachineInstr &MI,
192       SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) {
193   Defs.clear();
194   Uses.clear();
195 
196   for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
197     const MachineOperand &MO = MI.getOperand(i);
198 
199     if (!MO.isReg())
200       continue;
201 
202     Register Reg = MO.getReg();
203     if (!Reg)
204       continue;
205 
206     if (MO.isUse())
207       Uses.push_back(MO.getReg());
208 
209     if (MO.isDef())
210       Defs.push_back(MO.getReg());
211   }
212 }
213 
214 // Position dependent, so check twice for swap.
isDuplexPairMatch(unsigned Ga,unsigned Gb)215 static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) {
216   switch (Ga) {
217   case HexagonII::HSIG_None:
218   default:
219     return false;
220   case HexagonII::HSIG_L1:
221     return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A);
222   case HexagonII::HSIG_L2:
223     return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
224             Gb == HexagonII::HSIG_A);
225   case HexagonII::HSIG_S1:
226     return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
227             Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A);
228   case HexagonII::HSIG_S2:
229     return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
230             Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 ||
231             Gb == HexagonII::HSIG_A);
232   case HexagonII::HSIG_A:
233     return (Gb == HexagonII::HSIG_A);
234   case HexagonII::HSIG_Compound:
235     return (Gb == HexagonII::HSIG_Compound);
236   }
237   return false;
238 }
239 
240 /// isLoadFromStackSlot - If the specified machine instruction is a direct
241 /// load from a stack slot, return the virtual or physical register number of
242 /// the destination along with the FrameIndex of the loaded stack slot.  If
243 /// not, return 0.  This predicate must return 0 if the instruction has
244 /// any side effects other than loading from the stack slot.
isLoadFromStackSlot(const MachineInstr & MI,int & FrameIndex) const245 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
246                                                int &FrameIndex) const {
247   switch (MI.getOpcode()) {
248     default:
249       break;
250     case Hexagon::L2_loadri_io:
251     case Hexagon::L2_loadrd_io:
252     case Hexagon::V6_vL32b_ai:
253     case Hexagon::V6_vL32b_nt_ai:
254     case Hexagon::V6_vL32Ub_ai:
255     case Hexagon::LDriw_pred:
256     case Hexagon::LDriw_ctr:
257     case Hexagon::PS_vloadrq_ai:
258     case Hexagon::PS_vloadrw_ai:
259     case Hexagon::PS_vloadrw_nt_ai: {
260       const MachineOperand OpFI = MI.getOperand(1);
261       if (!OpFI.isFI())
262         return 0;
263       const MachineOperand OpOff = MI.getOperand(2);
264       if (!OpOff.isImm() || OpOff.getImm() != 0)
265         return 0;
266       FrameIndex = OpFI.getIndex();
267       return MI.getOperand(0).getReg();
268     }
269 
270     case Hexagon::L2_ploadrit_io:
271     case Hexagon::L2_ploadrif_io:
272     case Hexagon::L2_ploadrdt_io:
273     case Hexagon::L2_ploadrdf_io: {
274       const MachineOperand OpFI = MI.getOperand(2);
275       if (!OpFI.isFI())
276         return 0;
277       const MachineOperand OpOff = MI.getOperand(3);
278       if (!OpOff.isImm() || OpOff.getImm() != 0)
279         return 0;
280       FrameIndex = OpFI.getIndex();
281       return MI.getOperand(0).getReg();
282     }
283   }
284 
285   return 0;
286 }
287 
288 /// isStoreToStackSlot - If the specified machine instruction is a direct
289 /// store to a stack slot, return the virtual or physical register number of
290 /// the source reg along with the FrameIndex of the loaded stack slot.  If
291 /// not, return 0.  This predicate must return 0 if the instruction has
292 /// any side effects other than storing to the stack slot.
isStoreToStackSlot(const MachineInstr & MI,int & FrameIndex) const293 unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
294                                               int &FrameIndex) const {
295   switch (MI.getOpcode()) {
296     default:
297       break;
298     case Hexagon::S2_storerb_io:
299     case Hexagon::S2_storerh_io:
300     case Hexagon::S2_storeri_io:
301     case Hexagon::S2_storerd_io:
302     case Hexagon::V6_vS32b_ai:
303     case Hexagon::V6_vS32Ub_ai:
304     case Hexagon::STriw_pred:
305     case Hexagon::STriw_ctr:
306     case Hexagon::PS_vstorerq_ai:
307     case Hexagon::PS_vstorerw_ai: {
308       const MachineOperand &OpFI = MI.getOperand(0);
309       if (!OpFI.isFI())
310         return 0;
311       const MachineOperand &OpOff = MI.getOperand(1);
312       if (!OpOff.isImm() || OpOff.getImm() != 0)
313         return 0;
314       FrameIndex = OpFI.getIndex();
315       return MI.getOperand(2).getReg();
316     }
317 
318     case Hexagon::S2_pstorerbt_io:
319     case Hexagon::S2_pstorerbf_io:
320     case Hexagon::S2_pstorerht_io:
321     case Hexagon::S2_pstorerhf_io:
322     case Hexagon::S2_pstorerit_io:
323     case Hexagon::S2_pstorerif_io:
324     case Hexagon::S2_pstorerdt_io:
325     case Hexagon::S2_pstorerdf_io: {
326       const MachineOperand &OpFI = MI.getOperand(1);
327       if (!OpFI.isFI())
328         return 0;
329       const MachineOperand &OpOff = MI.getOperand(2);
330       if (!OpOff.isImm() || OpOff.getImm() != 0)
331         return 0;
332       FrameIndex = OpFI.getIndex();
333       return MI.getOperand(3).getReg();
334     }
335   }
336 
337   return 0;
338 }
339 
340 /// This function checks if the instruction or bundle of instructions
341 /// has load from stack slot and returns frameindex and machine memory
342 /// operand of that instruction if true.
hasLoadFromStackSlot(const MachineInstr & MI,SmallVectorImpl<const MachineMemOperand * > & Accesses) const343 bool HexagonInstrInfo::hasLoadFromStackSlot(
344     const MachineInstr &MI,
345     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
346   if (MI.isBundle()) {
347     const MachineBasicBlock *MBB = MI.getParent();
348     MachineBasicBlock::const_instr_iterator MII = MI.getIterator();
349     for (++MII; MII != MBB->instr_end() && MII->isInsideBundle(); ++MII)
350       if (TargetInstrInfo::hasLoadFromStackSlot(*MII, Accesses))
351         return true;
352     return false;
353   }
354 
355   return TargetInstrInfo::hasLoadFromStackSlot(MI, Accesses);
356 }
357 
358 /// This function checks if the instruction or bundle of instructions
359 /// has store to stack slot and returns frameindex and machine memory
360 /// operand of that instruction if true.
hasStoreToStackSlot(const MachineInstr & MI,SmallVectorImpl<const MachineMemOperand * > & Accesses) const361 bool HexagonInstrInfo::hasStoreToStackSlot(
362     const MachineInstr &MI,
363     SmallVectorImpl<const MachineMemOperand *> &Accesses) const {
364   if (MI.isBundle()) {
365     const MachineBasicBlock *MBB = MI.getParent();
366     MachineBasicBlock::const_instr_iterator MII = MI.getIterator();
367     for (++MII; MII != MBB->instr_end() && MII->isInsideBundle(); ++MII)
368       if (TargetInstrInfo::hasStoreToStackSlot(*MII, Accesses))
369         return true;
370     return false;
371   }
372 
373   return TargetInstrInfo::hasStoreToStackSlot(MI, Accesses);
374 }
375 
376 /// This function can analyze one/two way branching only and should (mostly) be
377 /// called by target independent side.
378 /// First entry is always the opcode of the branching instruction, except when
379 /// the Cond vector is supposed to be empty, e.g., when analyzeBranch fails, a
380 /// BB with only unconditional jump. Subsequent entries depend upon the opcode,
381 /// e.g. Jump_c p will have
382 /// Cond[0] = Jump_c
383 /// Cond[1] = p
384 /// HW-loop ENDLOOP:
385 /// Cond[0] = ENDLOOP
386 /// Cond[1] = MBB
387 /// New value jump:
388 /// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode
389 /// Cond[1] = R
390 /// Cond[2] = Imm
analyzeBranch(MachineBasicBlock & MBB,MachineBasicBlock * & TBB,MachineBasicBlock * & FBB,SmallVectorImpl<MachineOperand> & Cond,bool AllowModify) const391 bool HexagonInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
392                                      MachineBasicBlock *&TBB,
393                                      MachineBasicBlock *&FBB,
394                                      SmallVectorImpl<MachineOperand> &Cond,
395                                      bool AllowModify) const {
396   TBB = nullptr;
397   FBB = nullptr;
398   Cond.clear();
399 
400   // If the block has no terminators, it just falls into the block after it.
401   MachineBasicBlock::instr_iterator I = MBB.instr_end();
402   if (I == MBB.instr_begin())
403     return false;
404 
405   // A basic block may looks like this:
406   //
407   //  [   insn
408   //     EH_LABEL
409   //      insn
410   //      insn
411   //      insn
412   //     EH_LABEL
413   //      insn     ]
414   //
415   // It has two succs but does not have a terminator
416   // Don't know how to handle it.
417   do {
418     --I;
419     if (I->isEHLabel())
420       // Don't analyze EH branches.
421       return true;
422   } while (I != MBB.instr_begin());
423 
424   I = MBB.instr_end();
425   --I;
426 
427   while (I->isDebugInstr()) {
428     if (I == MBB.instr_begin())
429       return false;
430     --I;
431   }
432 
433   bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump &&
434                      I->getOperand(0).isMBB();
435   // Delete the J2_jump if it's equivalent to a fall-through.
436   if (AllowModify && JumpToBlock &&
437       MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
438     LLVM_DEBUG(dbgs() << "\nErasing the jump to successor block\n";);
439     I->eraseFromParent();
440     I = MBB.instr_end();
441     if (I == MBB.instr_begin())
442       return false;
443     --I;
444   }
445   if (!isUnpredicatedTerminator(*I))
446     return false;
447 
448   // Get the last instruction in the block.
449   MachineInstr *LastInst = &*I;
450   MachineInstr *SecondLastInst = nullptr;
451   // Find one more terminator if present.
452   while (true) {
453     if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) {
454       if (!SecondLastInst)
455         SecondLastInst = &*I;
456       else
457         // This is a third branch.
458         return true;
459     }
460     if (I == MBB.instr_begin())
461       break;
462     --I;
463   }
464 
465   int LastOpcode = LastInst->getOpcode();
466   int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0;
467   // If the branch target is not a basic block, it could be a tail call.
468   // (It is, if the target is a function.)
469   if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB())
470     return true;
471   if (SecLastOpcode == Hexagon::J2_jump &&
472       !SecondLastInst->getOperand(0).isMBB())
473     return true;
474 
475   bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode);
476   bool LastOpcodeHasNVJump = isNewValueJump(*LastInst);
477 
478   if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB())
479     return true;
480 
481   // If there is only one terminator instruction, process it.
482   if (LastInst && !SecondLastInst) {
483     if (LastOpcode == Hexagon::J2_jump) {
484       TBB = LastInst->getOperand(0).getMBB();
485       return false;
486     }
487     if (isEndLoopN(LastOpcode)) {
488       TBB = LastInst->getOperand(0).getMBB();
489       Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
490       Cond.push_back(LastInst->getOperand(0));
491       return false;
492     }
493     if (LastOpcodeHasJMP_c) {
494       TBB = LastInst->getOperand(1).getMBB();
495       Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
496       Cond.push_back(LastInst->getOperand(0));
497       return false;
498     }
499     // Only supporting rr/ri versions of new-value jumps.
500     if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) {
501       TBB = LastInst->getOperand(2).getMBB();
502       Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
503       Cond.push_back(LastInst->getOperand(0));
504       Cond.push_back(LastInst->getOperand(1));
505       return false;
506     }
507     LLVM_DEBUG(dbgs() << "\nCant analyze " << printMBBReference(MBB)
508                       << " with one jump\n";);
509     // Otherwise, don't know what this is.
510     return true;
511   }
512 
513   bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode);
514   bool SecLastOpcodeHasNVJump = isNewValueJump(*SecondLastInst);
515   if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) {
516     if (!SecondLastInst->getOperand(1).isMBB())
517       return true;
518     TBB =  SecondLastInst->getOperand(1).getMBB();
519     Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
520     Cond.push_back(SecondLastInst->getOperand(0));
521     FBB = LastInst->getOperand(0).getMBB();
522     return false;
523   }
524 
525   // Only supporting rr/ri versions of new-value jumps.
526   if (SecLastOpcodeHasNVJump &&
527       (SecondLastInst->getNumExplicitOperands() == 3) &&
528       (LastOpcode == Hexagon::J2_jump)) {
529     TBB = SecondLastInst->getOperand(2).getMBB();
530     Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
531     Cond.push_back(SecondLastInst->getOperand(0));
532     Cond.push_back(SecondLastInst->getOperand(1));
533     FBB = LastInst->getOperand(0).getMBB();
534     return false;
535   }
536 
537   // If the block ends with two Hexagon:JMPs, handle it.  The second one is not
538   // executed, so remove it.
539   if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) {
540     TBB = SecondLastInst->getOperand(0).getMBB();
541     I = LastInst->getIterator();
542     if (AllowModify)
543       I->eraseFromParent();
544     return false;
545   }
546 
547   // If the block ends with an ENDLOOP, and J2_jump, handle it.
548   if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) {
549     TBB = SecondLastInst->getOperand(0).getMBB();
550     Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
551     Cond.push_back(SecondLastInst->getOperand(0));
552     FBB = LastInst->getOperand(0).getMBB();
553     return false;
554   }
555   LLVM_DEBUG(dbgs() << "\nCant analyze " << printMBBReference(MBB)
556                     << " with two jumps";);
557   // Otherwise, can't handle this.
558   return true;
559 }
560 
removeBranch(MachineBasicBlock & MBB,int * BytesRemoved) const561 unsigned HexagonInstrInfo::removeBranch(MachineBasicBlock &MBB,
562                                         int *BytesRemoved) const {
563   assert(!BytesRemoved && "code size not handled");
564 
565   LLVM_DEBUG(dbgs() << "\nRemoving branches out of " << printMBBReference(MBB));
566   MachineBasicBlock::iterator I = MBB.end();
567   unsigned Count = 0;
568   while (I != MBB.begin()) {
569     --I;
570     if (I->isDebugInstr())
571       continue;
572     // Only removing branches from end of MBB.
573     if (!I->isBranch())
574       return Count;
575     if (Count && (I->getOpcode() == Hexagon::J2_jump))
576       llvm_unreachable("Malformed basic block: unconditional branch not last");
577     MBB.erase(&MBB.back());
578     I = MBB.end();
579     ++Count;
580   }
581   return Count;
582 }
583 
insertBranch(MachineBasicBlock & MBB,MachineBasicBlock * TBB,MachineBasicBlock * FBB,ArrayRef<MachineOperand> Cond,const DebugLoc & DL,int * BytesAdded) const584 unsigned HexagonInstrInfo::insertBranch(MachineBasicBlock &MBB,
585                                         MachineBasicBlock *TBB,
586                                         MachineBasicBlock *FBB,
587                                         ArrayRef<MachineOperand> Cond,
588                                         const DebugLoc &DL,
589                                         int *BytesAdded) const {
590   unsigned BOpc   = Hexagon::J2_jump;
591   unsigned BccOpc = Hexagon::J2_jumpt;
592   assert(validateBranchCond(Cond) && "Invalid branching condition");
593   assert(TBB && "insertBranch must not be told to insert a fallthrough");
594   assert(!BytesAdded && "code size not handled");
595 
596   // Check if reverseBranchCondition has asked to reverse this branch
597   // If we want to reverse the branch an odd number of times, we want
598   // J2_jumpf.
599   if (!Cond.empty() && Cond[0].isImm())
600     BccOpc = Cond[0].getImm();
601 
602   if (!FBB) {
603     if (Cond.empty()) {
604       // Due to a bug in TailMerging/CFG Optimization, we need to add a
605       // special case handling of a predicated jump followed by an
606       // unconditional jump. If not, Tail Merging and CFG Optimization go
607       // into an infinite loop.
608       MachineBasicBlock *NewTBB, *NewFBB;
609       SmallVector<MachineOperand, 4> Cond;
610       auto Term = MBB.getFirstTerminator();
611       if (Term != MBB.end() && isPredicated(*Term) &&
612           !analyzeBranch(MBB, NewTBB, NewFBB, Cond, false) &&
613           MachineFunction::iterator(NewTBB) == ++MBB.getIterator()) {
614         reverseBranchCondition(Cond);
615         removeBranch(MBB);
616         return insertBranch(MBB, TBB, nullptr, Cond, DL);
617       }
618       BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
619     } else if (isEndLoopN(Cond[0].getImm())) {
620       int EndLoopOp = Cond[0].getImm();
621       assert(Cond[1].isMBB());
622       // Since we're adding an ENDLOOP, there better be a LOOP instruction.
623       // Check for it, and change the BB target if needed.
624       SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
625       MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(),
626                                          VisitedBBs);
627       assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP");
628       Loop->getOperand(0).setMBB(TBB);
629       // Add the ENDLOOP after the finding the LOOP0.
630       BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
631     } else if (isNewValueJump(Cond[0].getImm())) {
632       assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump");
633       // New value jump
634       // (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset)
635       // (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset)
636       unsigned Flags1 = getUndefRegState(Cond[1].isUndef());
637       LLVM_DEBUG(dbgs() << "\nInserting NVJump for "
638                         << printMBBReference(MBB););
639       if (Cond[2].isReg()) {
640         unsigned Flags2 = getUndefRegState(Cond[2].isUndef());
641         BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
642           addReg(Cond[2].getReg(), Flags2).addMBB(TBB);
643       } else if(Cond[2].isImm()) {
644         BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
645           addImm(Cond[2].getImm()).addMBB(TBB);
646       } else
647         llvm_unreachable("Invalid condition for branching");
648     } else {
649       assert((Cond.size() == 2) && "Malformed cond vector");
650       const MachineOperand &RO = Cond[1];
651       unsigned Flags = getUndefRegState(RO.isUndef());
652       BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
653     }
654     return 1;
655   }
656   assert((!Cond.empty()) &&
657          "Cond. cannot be empty when multiple branchings are required");
658   assert((!isNewValueJump(Cond[0].getImm())) &&
659          "NV-jump cannot be inserted with another branch");
660   // Special case for hardware loops.  The condition is a basic block.
661   if (isEndLoopN(Cond[0].getImm())) {
662     int EndLoopOp = Cond[0].getImm();
663     assert(Cond[1].isMBB());
664     // Since we're adding an ENDLOOP, there better be a LOOP instruction.
665     // Check for it, and change the BB target if needed.
666     SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
667     MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, Cond[1].getMBB(),
668                                        VisitedBBs);
669     assert(Loop != nullptr && "Inserting an ENDLOOP without a LOOP");
670     Loop->getOperand(0).setMBB(TBB);
671     // Add the ENDLOOP after the finding the LOOP0.
672     BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
673   } else {
674     const MachineOperand &RO = Cond[1];
675     unsigned Flags = getUndefRegState(RO.isUndef());
676     BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
677   }
678   BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
679 
680   return 2;
681 }
682 
683 namespace {
684 class HexagonPipelinerLoopInfo : public TargetInstrInfo::PipelinerLoopInfo {
685   MachineInstr *Loop, *EndLoop;
686   MachineFunction *MF;
687   const HexagonInstrInfo *TII;
688   int64_t TripCount;
689   Register LoopCount;
690   DebugLoc DL;
691 
692 public:
HexagonPipelinerLoopInfo(MachineInstr * Loop,MachineInstr * EndLoop)693   HexagonPipelinerLoopInfo(MachineInstr *Loop, MachineInstr *EndLoop)
694       : Loop(Loop), EndLoop(EndLoop), MF(Loop->getParent()->getParent()),
695         TII(MF->getSubtarget<HexagonSubtarget>().getInstrInfo()),
696         DL(Loop->getDebugLoc()) {
697     // Inspect the Loop instruction up-front, as it may be deleted when we call
698     // createTripCountGreaterCondition.
699     TripCount = Loop->getOpcode() == Hexagon::J2_loop0r
700                     ? -1
701                     : Loop->getOperand(1).getImm();
702     if (TripCount == -1)
703       LoopCount = Loop->getOperand(1).getReg();
704   }
705 
shouldIgnoreForPipelining(const MachineInstr * MI) const706   bool shouldIgnoreForPipelining(const MachineInstr *MI) const override {
707     // Only ignore the terminator.
708     return MI == EndLoop;
709   }
710 
711   Optional<bool>
createTripCountGreaterCondition(int TC,MachineBasicBlock & MBB,SmallVectorImpl<MachineOperand> & Cond)712   createTripCountGreaterCondition(int TC, MachineBasicBlock &MBB,
713                                   SmallVectorImpl<MachineOperand> &Cond) override {
714     if (TripCount == -1) {
715       // Check if we're done with the loop.
716       unsigned Done = TII->createVR(MF, MVT::i1);
717       MachineInstr *NewCmp = BuildMI(&MBB, DL,
718                                      TII->get(Hexagon::C2_cmpgtui), Done)
719                                  .addReg(LoopCount)
720                                  .addImm(TC);
721       Cond.push_back(MachineOperand::CreateImm(Hexagon::J2_jumpf));
722       Cond.push_back(NewCmp->getOperand(0));
723       return {};
724     }
725 
726     return TripCount > TC;
727   }
728 
setPreheader(MachineBasicBlock * NewPreheader)729   void setPreheader(MachineBasicBlock *NewPreheader) override {
730     NewPreheader->splice(NewPreheader->getFirstTerminator(), Loop->getParent(),
731                          Loop);
732   }
733 
adjustTripCount(int TripCountAdjust)734   void adjustTripCount(int TripCountAdjust) override {
735     // If the loop trip count is a compile-time value, then just change the
736     // value.
737     if (Loop->getOpcode() == Hexagon::J2_loop0i ||
738         Loop->getOpcode() == Hexagon::J2_loop1i) {
739       int64_t TripCount = Loop->getOperand(1).getImm() + TripCountAdjust;
740       assert(TripCount > 0 && "Can't create an empty or negative loop!");
741       Loop->getOperand(1).setImm(TripCount);
742       return;
743     }
744 
745     // The loop trip count is a run-time value. We generate code to subtract
746     // one from the trip count, and update the loop instruction.
747     Register LoopCount = Loop->getOperand(1).getReg();
748     Register NewLoopCount = TII->createVR(MF, MVT::i32);
749     BuildMI(*Loop->getParent(), Loop, Loop->getDebugLoc(),
750             TII->get(Hexagon::A2_addi), NewLoopCount)
751         .addReg(LoopCount)
752         .addImm(TripCountAdjust);
753     Loop->getOperand(1).setReg(NewLoopCount);
754   }
755 
disposed()756   void disposed() override { Loop->eraseFromParent(); }
757 };
758 } // namespace
759 
760 std::unique_ptr<TargetInstrInfo::PipelinerLoopInfo>
analyzeLoopForPipelining(MachineBasicBlock * LoopBB) const761 HexagonInstrInfo::analyzeLoopForPipelining(MachineBasicBlock *LoopBB) const {
762   // We really "analyze" only hardware loops right now.
763   MachineBasicBlock::iterator I = LoopBB->getFirstTerminator();
764 
765   if (I != LoopBB->end() && isEndLoopN(I->getOpcode())) {
766     SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
767     MachineInstr *LoopInst = findLoopInstr(
768         LoopBB, I->getOpcode(), I->getOperand(0).getMBB(), VisitedBBs);
769     if (LoopInst)
770       return std::make_unique<HexagonPipelinerLoopInfo>(LoopInst, &*I);
771   }
772   return nullptr;
773 }
774 
isProfitableToIfCvt(MachineBasicBlock & MBB,unsigned NumCycles,unsigned ExtraPredCycles,BranchProbability Probability) const775 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
776       unsigned NumCycles, unsigned ExtraPredCycles,
777       BranchProbability Probability) const {
778   return nonDbgBBSize(&MBB) <= 3;
779 }
780 
isProfitableToIfCvt(MachineBasicBlock & TMBB,unsigned NumTCycles,unsigned ExtraTCycles,MachineBasicBlock & FMBB,unsigned NumFCycles,unsigned ExtraFCycles,BranchProbability Probability) const781 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
782       unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB,
783       unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability)
784       const {
785   return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3;
786 }
787 
isProfitableToDupForIfCvt(MachineBasicBlock & MBB,unsigned NumInstrs,BranchProbability Probability) const788 bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
789       unsigned NumInstrs, BranchProbability Probability) const {
790   return NumInstrs <= 4;
791 }
792 
getLiveInRegsAt(LivePhysRegs & Regs,const MachineInstr & MI)793 static void getLiveInRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) {
794   SmallVector<std::pair<MCPhysReg, const MachineOperand*>,2> Clobbers;
795   const MachineBasicBlock &B = *MI.getParent();
796   Regs.addLiveIns(B);
797   auto E = MachineBasicBlock::const_iterator(MI.getIterator());
798   for (auto I = B.begin(); I != E; ++I) {
799     Clobbers.clear();
800     Regs.stepForward(*I, Clobbers);
801   }
802 }
803 
getLiveOutRegsAt(LivePhysRegs & Regs,const MachineInstr & MI)804 static void getLiveOutRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) {
805   const MachineBasicBlock &B = *MI.getParent();
806   Regs.addLiveOuts(B);
807   auto E = ++MachineBasicBlock::const_iterator(MI.getIterator()).getReverse();
808   for (auto I = B.rbegin(); I != E; ++I)
809     Regs.stepBackward(*I);
810 }
811 
copyPhysReg(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,const DebugLoc & DL,MCRegister DestReg,MCRegister SrcReg,bool KillSrc) const812 void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
813                                    MachineBasicBlock::iterator I,
814                                    const DebugLoc &DL, MCRegister DestReg,
815                                    MCRegister SrcReg, bool KillSrc) const {
816   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
817   unsigned KillFlag = getKillRegState(KillSrc);
818 
819   if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) {
820     BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg)
821       .addReg(SrcReg, KillFlag);
822     return;
823   }
824   if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) {
825     BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg)
826       .addReg(SrcReg, KillFlag);
827     return;
828   }
829   if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) {
830     // Map Pd = Ps to Pd = or(Ps, Ps).
831     BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg)
832       .addReg(SrcReg).addReg(SrcReg, KillFlag);
833     return;
834   }
835   if (Hexagon::CtrRegsRegClass.contains(DestReg) &&
836       Hexagon::IntRegsRegClass.contains(SrcReg)) {
837     BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
838       .addReg(SrcReg, KillFlag);
839     return;
840   }
841   if (Hexagon::IntRegsRegClass.contains(DestReg) &&
842       Hexagon::CtrRegsRegClass.contains(SrcReg)) {
843     BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg)
844       .addReg(SrcReg, KillFlag);
845     return;
846   }
847   if (Hexagon::ModRegsRegClass.contains(DestReg) &&
848       Hexagon::IntRegsRegClass.contains(SrcReg)) {
849     BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
850       .addReg(SrcReg, KillFlag);
851     return;
852   }
853   if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
854       Hexagon::IntRegsRegClass.contains(DestReg)) {
855     BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
856       .addReg(SrcReg, KillFlag);
857     return;
858   }
859   if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
860       Hexagon::PredRegsRegClass.contains(DestReg)) {
861     BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg)
862       .addReg(SrcReg, KillFlag);
863     return;
864   }
865   if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
866       Hexagon::IntRegsRegClass.contains(DestReg)) {
867     BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
868       .addReg(SrcReg, KillFlag);
869     return;
870   }
871   if (Hexagon::HvxVRRegClass.contains(SrcReg, DestReg)) {
872     BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg).
873       addReg(SrcReg, KillFlag);
874     return;
875   }
876   if (Hexagon::HvxWRRegClass.contains(SrcReg, DestReg)) {
877     LivePhysRegs LiveAtMI(HRI);
878     getLiveInRegsAt(LiveAtMI, *I);
879     Register SrcLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
880     Register SrcHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
881     unsigned UndefLo = getUndefRegState(!LiveAtMI.contains(SrcLo));
882     unsigned UndefHi = getUndefRegState(!LiveAtMI.contains(SrcHi));
883     BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg)
884       .addReg(SrcHi, KillFlag | UndefHi)
885       .addReg(SrcLo, KillFlag | UndefLo);
886     return;
887   }
888   if (Hexagon::HvxQRRegClass.contains(SrcReg, DestReg)) {
889     BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg)
890       .addReg(SrcReg)
891       .addReg(SrcReg, KillFlag);
892     return;
893   }
894   if (Hexagon::HvxQRRegClass.contains(SrcReg) &&
895       Hexagon::HvxVRRegClass.contains(DestReg)) {
896     llvm_unreachable("Unimplemented pred to vec");
897     return;
898   }
899   if (Hexagon::HvxQRRegClass.contains(DestReg) &&
900       Hexagon::HvxVRRegClass.contains(SrcReg)) {
901     llvm_unreachable("Unimplemented vec to pred");
902     return;
903   }
904 
905 #ifndef NDEBUG
906   // Show the invalid registers to ease debugging.
907   dbgs() << "Invalid registers for copy in " << printMBBReference(MBB) << ": "
908          << printReg(DestReg, &HRI) << " = " << printReg(SrcReg, &HRI) << '\n';
909 #endif
910   llvm_unreachable("Unimplemented");
911 }
912 
storeRegToStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,Register SrcReg,bool isKill,int FI,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const913 void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
914       MachineBasicBlock::iterator I, Register SrcReg, bool isKill, int FI,
915       const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const {
916   DebugLoc DL = MBB.findDebugLoc(I);
917   MachineFunction &MF = *MBB.getParent();
918   MachineFrameInfo &MFI = MF.getFrameInfo();
919   unsigned KillFlag = getKillRegState(isKill);
920 
921   MachineMemOperand *MMO = MF.getMachineMemOperand(
922       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
923       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
924 
925   if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
926     BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io))
927       .addFrameIndex(FI).addImm(0)
928       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
929   } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
930     BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io))
931       .addFrameIndex(FI).addImm(0)
932       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
933   } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
934     BuildMI(MBB, I, DL, get(Hexagon::STriw_pred))
935       .addFrameIndex(FI).addImm(0)
936       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
937   } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
938     BuildMI(MBB, I, DL, get(Hexagon::STriw_ctr))
939       .addFrameIndex(FI).addImm(0)
940       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
941   } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) {
942     BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai))
943       .addFrameIndex(FI).addImm(0)
944       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
945   } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) {
946     BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerv_ai))
947       .addFrameIndex(FI).addImm(0)
948       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
949   } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) {
950     BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerw_ai))
951       .addFrameIndex(FI).addImm(0)
952       .addReg(SrcReg, KillFlag).addMemOperand(MMO);
953   } else {
954     llvm_unreachable("Unimplemented");
955   }
956 }
957 
loadRegFromStackSlot(MachineBasicBlock & MBB,MachineBasicBlock::iterator I,Register DestReg,int FI,const TargetRegisterClass * RC,const TargetRegisterInfo * TRI) const958 void HexagonInstrInfo::loadRegFromStackSlot(
959     MachineBasicBlock &MBB, MachineBasicBlock::iterator I, Register DestReg,
960     int FI, const TargetRegisterClass *RC,
961     const TargetRegisterInfo *TRI) const {
962   DebugLoc DL = MBB.findDebugLoc(I);
963   MachineFunction &MF = *MBB.getParent();
964   MachineFrameInfo &MFI = MF.getFrameInfo();
965 
966   MachineMemOperand *MMO = MF.getMachineMemOperand(
967       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
968       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
969 
970   if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
971     BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg)
972       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
973   } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
974     BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg)
975       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
976   } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
977     BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg)
978       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
979   } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
980     BuildMI(MBB, I, DL, get(Hexagon::LDriw_ctr), DestReg)
981       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
982   } else if (Hexagon::HvxQRRegClass.hasSubClassEq(RC)) {
983     BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai), DestReg)
984       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
985   } else if (Hexagon::HvxVRRegClass.hasSubClassEq(RC)) {
986     BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrv_ai), DestReg)
987       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
988   } else if (Hexagon::HvxWRRegClass.hasSubClassEq(RC)) {
989     BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrw_ai), DestReg)
990       .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
991   } else {
992     llvm_unreachable("Can't store this register to stack slot");
993   }
994 }
995 
996 /// expandPostRAPseudo - This function is called for all pseudo instructions
997 /// that remain after register allocation. Many pseudo instructions are
998 /// created to help register allocation. This is the place to convert them
999 /// into real instructions. The target can edit MI in place, or it can insert
1000 /// new instructions and erase MI. The function should return true if
1001 /// anything was changed.
expandPostRAPseudo(MachineInstr & MI) const1002 bool HexagonInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
1003   MachineBasicBlock &MBB = *MI.getParent();
1004   MachineFunction &MF = *MBB.getParent();
1005   MachineRegisterInfo &MRI = MF.getRegInfo();
1006   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1007   LivePhysRegs LiveIn(HRI), LiveOut(HRI);
1008   DebugLoc DL = MI.getDebugLoc();
1009   unsigned Opc = MI.getOpcode();
1010 
1011   auto RealCirc = [&](unsigned Opc, bool HasImm, unsigned MxOp) {
1012     Register Mx = MI.getOperand(MxOp).getReg();
1013     unsigned CSx = (Mx == Hexagon::M0 ? Hexagon::CS0 : Hexagon::CS1);
1014     BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrrcr), CSx)
1015         .add(MI.getOperand((HasImm ? 5 : 4)));
1016     auto MIB = BuildMI(MBB, MI, DL, get(Opc)).add(MI.getOperand(0))
1017         .add(MI.getOperand(1)).add(MI.getOperand(2)).add(MI.getOperand(3));
1018     if (HasImm)
1019       MIB.add(MI.getOperand(4));
1020     MIB.addReg(CSx, RegState::Implicit);
1021     MBB.erase(MI);
1022     return true;
1023   };
1024 
1025   auto UseAligned = [&](const MachineInstr &MI, Align NeedAlign) {
1026     if (MI.memoperands().empty())
1027       return false;
1028     return all_of(MI.memoperands(), [NeedAlign](const MachineMemOperand *MMO) {
1029       return MMO->getAlign() >= NeedAlign;
1030     });
1031   };
1032 
1033   switch (Opc) {
1034     case TargetOpcode::COPY: {
1035       MachineOperand &MD = MI.getOperand(0);
1036       MachineOperand &MS = MI.getOperand(1);
1037       MachineBasicBlock::iterator MBBI = MI.getIterator();
1038       if (MD.getReg() != MS.getReg() && !MS.isUndef()) {
1039         copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill());
1040         std::prev(MBBI)->copyImplicitOps(*MBB.getParent(), MI);
1041       }
1042       MBB.erase(MBBI);
1043       return true;
1044     }
1045     case Hexagon::PS_aligna:
1046       BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI.getOperand(0).getReg())
1047           .addReg(HRI.getFrameRegister())
1048           .addImm(-MI.getOperand(1).getImm());
1049       MBB.erase(MI);
1050       return true;
1051     case Hexagon::V6_vassignp: {
1052       Register SrcReg = MI.getOperand(1).getReg();
1053       Register DstReg = MI.getOperand(0).getReg();
1054       Register SrcLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
1055       Register SrcHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
1056       getLiveInRegsAt(LiveIn, MI);
1057       unsigned UndefLo = getUndefRegState(!LiveIn.contains(SrcLo));
1058       unsigned UndefHi = getUndefRegState(!LiveIn.contains(SrcHi));
1059       unsigned Kill = getKillRegState(MI.getOperand(1).isKill());
1060       BuildMI(MBB, MI, DL, get(Hexagon::V6_vcombine), DstReg)
1061           .addReg(SrcHi, UndefHi)
1062           .addReg(SrcLo, Kill | UndefLo);
1063       MBB.erase(MI);
1064       return true;
1065     }
1066     case Hexagon::V6_lo: {
1067       Register SrcReg = MI.getOperand(1).getReg();
1068       Register DstReg = MI.getOperand(0).getReg();
1069       Register SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
1070       copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI.getOperand(1).isKill());
1071       MBB.erase(MI);
1072       MRI.clearKillFlags(SrcSubLo);
1073       return true;
1074     }
1075     case Hexagon::V6_hi: {
1076       Register SrcReg = MI.getOperand(1).getReg();
1077       Register DstReg = MI.getOperand(0).getReg();
1078       Register SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
1079       copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI.getOperand(1).isKill());
1080       MBB.erase(MI);
1081       MRI.clearKillFlags(SrcSubHi);
1082       return true;
1083     }
1084     case Hexagon::PS_vloadrv_ai: {
1085       Register DstReg = MI.getOperand(0).getReg();
1086       const MachineOperand &BaseOp = MI.getOperand(1);
1087       assert(BaseOp.getSubReg() == 0);
1088       int Offset = MI.getOperand(2).getImm();
1089       Align NeedAlign = HRI.getSpillAlign(Hexagon::HvxVRRegClass);
1090       unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vL32b_ai
1091                                                   : Hexagon::V6_vL32Ub_ai;
1092       BuildMI(MBB, MI, DL, get(NewOpc), DstReg)
1093           .addReg(BaseOp.getReg(), getRegState(BaseOp))
1094           .addImm(Offset)
1095           .cloneMemRefs(MI);
1096       MBB.erase(MI);
1097       return true;
1098     }
1099     case Hexagon::PS_vloadrw_ai: {
1100       Register DstReg = MI.getOperand(0).getReg();
1101       const MachineOperand &BaseOp = MI.getOperand(1);
1102       assert(BaseOp.getSubReg() == 0);
1103       int Offset = MI.getOperand(2).getImm();
1104       unsigned VecOffset = HRI.getSpillSize(Hexagon::HvxVRRegClass);
1105       Align NeedAlign = HRI.getSpillAlign(Hexagon::HvxVRRegClass);
1106       unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vL32b_ai
1107                                                   : Hexagon::V6_vL32Ub_ai;
1108       BuildMI(MBB, MI, DL, get(NewOpc),
1109               HRI.getSubReg(DstReg, Hexagon::vsub_lo))
1110           .addReg(BaseOp.getReg(), getRegState(BaseOp) & ~RegState::Kill)
1111           .addImm(Offset)
1112           .cloneMemRefs(MI);
1113       BuildMI(MBB, MI, DL, get(NewOpc),
1114               HRI.getSubReg(DstReg, Hexagon::vsub_hi))
1115           .addReg(BaseOp.getReg(), getRegState(BaseOp))
1116           .addImm(Offset + VecOffset)
1117           .cloneMemRefs(MI);
1118       MBB.erase(MI);
1119       return true;
1120     }
1121     case Hexagon::PS_vstorerv_ai: {
1122       const MachineOperand &SrcOp = MI.getOperand(2);
1123       assert(SrcOp.getSubReg() == 0);
1124       const MachineOperand &BaseOp = MI.getOperand(0);
1125       assert(BaseOp.getSubReg() == 0);
1126       int Offset = MI.getOperand(1).getImm();
1127       Align NeedAlign = HRI.getSpillAlign(Hexagon::HvxVRRegClass);
1128       unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vS32b_ai
1129                                                   : Hexagon::V6_vS32Ub_ai;
1130       BuildMI(MBB, MI, DL, get(NewOpc))
1131           .addReg(BaseOp.getReg(), getRegState(BaseOp))
1132           .addImm(Offset)
1133           .addReg(SrcOp.getReg(), getRegState(SrcOp))
1134           .cloneMemRefs(MI);
1135       MBB.erase(MI);
1136       return true;
1137     }
1138     case Hexagon::PS_vstorerw_ai: {
1139       Register SrcReg = MI.getOperand(2).getReg();
1140       const MachineOperand &BaseOp = MI.getOperand(0);
1141       assert(BaseOp.getSubReg() == 0);
1142       int Offset = MI.getOperand(1).getImm();
1143       unsigned VecOffset = HRI.getSpillSize(Hexagon::HvxVRRegClass);
1144       Align NeedAlign = HRI.getSpillAlign(Hexagon::HvxVRRegClass);
1145       unsigned NewOpc = UseAligned(MI, NeedAlign) ? Hexagon::V6_vS32b_ai
1146                                                   : Hexagon::V6_vS32Ub_ai;
1147       BuildMI(MBB, MI, DL, get(NewOpc))
1148           .addReg(BaseOp.getReg(), getRegState(BaseOp) & ~RegState::Kill)
1149           .addImm(Offset)
1150           .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_lo))
1151           .cloneMemRefs(MI);
1152       BuildMI(MBB, MI, DL, get(NewOpc))
1153           .addReg(BaseOp.getReg(), getRegState(BaseOp))
1154           .addImm(Offset + VecOffset)
1155           .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_hi))
1156           .cloneMemRefs(MI);
1157       MBB.erase(MI);
1158       return true;
1159     }
1160     case Hexagon::PS_true: {
1161       Register Reg = MI.getOperand(0).getReg();
1162       BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg)
1163         .addReg(Reg, RegState::Undef)
1164         .addReg(Reg, RegState::Undef);
1165       MBB.erase(MI);
1166       return true;
1167     }
1168     case Hexagon::PS_false: {
1169       Register Reg = MI.getOperand(0).getReg();
1170       BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg)
1171         .addReg(Reg, RegState::Undef)
1172         .addReg(Reg, RegState::Undef);
1173       MBB.erase(MI);
1174       return true;
1175     }
1176     case Hexagon::PS_qtrue: {
1177       BuildMI(MBB, MI, DL, get(Hexagon::V6_veqw), MI.getOperand(0).getReg())
1178         .addReg(Hexagon::V0, RegState::Undef)
1179         .addReg(Hexagon::V0, RegState::Undef);
1180       MBB.erase(MI);
1181       return true;
1182     }
1183     case Hexagon::PS_qfalse: {
1184       BuildMI(MBB, MI, DL, get(Hexagon::V6_vgtw), MI.getOperand(0).getReg())
1185         .addReg(Hexagon::V0, RegState::Undef)
1186         .addReg(Hexagon::V0, RegState::Undef);
1187       MBB.erase(MI);
1188       return true;
1189     }
1190     case Hexagon::PS_vdd0: {
1191       Register Vd = MI.getOperand(0).getReg();
1192       BuildMI(MBB, MI, DL, get(Hexagon::V6_vsubw_dv), Vd)
1193         .addReg(Vd, RegState::Undef)
1194         .addReg(Vd, RegState::Undef);
1195       MBB.erase(MI);
1196       return true;
1197     }
1198     case Hexagon::PS_vmulw: {
1199       // Expand a 64-bit vector multiply into 2 32-bit scalar multiplies.
1200       Register DstReg = MI.getOperand(0).getReg();
1201       Register Src1Reg = MI.getOperand(1).getReg();
1202       Register Src2Reg = MI.getOperand(2).getReg();
1203       Register Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
1204       Register Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
1205       Register Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
1206       Register Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
1207       BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
1208               HRI.getSubReg(DstReg, Hexagon::isub_hi))
1209           .addReg(Src1SubHi)
1210           .addReg(Src2SubHi);
1211       BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
1212               HRI.getSubReg(DstReg, Hexagon::isub_lo))
1213           .addReg(Src1SubLo)
1214           .addReg(Src2SubLo);
1215       MBB.erase(MI);
1216       MRI.clearKillFlags(Src1SubHi);
1217       MRI.clearKillFlags(Src1SubLo);
1218       MRI.clearKillFlags(Src2SubHi);
1219       MRI.clearKillFlags(Src2SubLo);
1220       return true;
1221     }
1222     case Hexagon::PS_vmulw_acc: {
1223       // Expand 64-bit vector multiply with addition into 2 scalar multiplies.
1224       Register DstReg = MI.getOperand(0).getReg();
1225       Register Src1Reg = MI.getOperand(1).getReg();
1226       Register Src2Reg = MI.getOperand(2).getReg();
1227       Register Src3Reg = MI.getOperand(3).getReg();
1228       Register Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
1229       Register Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
1230       Register Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
1231       Register Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
1232       Register Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::isub_hi);
1233       Register Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::isub_lo);
1234       BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
1235               HRI.getSubReg(DstReg, Hexagon::isub_hi))
1236           .addReg(Src1SubHi)
1237           .addReg(Src2SubHi)
1238           .addReg(Src3SubHi);
1239       BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
1240               HRI.getSubReg(DstReg, Hexagon::isub_lo))
1241           .addReg(Src1SubLo)
1242           .addReg(Src2SubLo)
1243           .addReg(Src3SubLo);
1244       MBB.erase(MI);
1245       MRI.clearKillFlags(Src1SubHi);
1246       MRI.clearKillFlags(Src1SubLo);
1247       MRI.clearKillFlags(Src2SubHi);
1248       MRI.clearKillFlags(Src2SubLo);
1249       MRI.clearKillFlags(Src3SubHi);
1250       MRI.clearKillFlags(Src3SubLo);
1251       return true;
1252     }
1253     case Hexagon::PS_pselect: {
1254       const MachineOperand &Op0 = MI.getOperand(0);
1255       const MachineOperand &Op1 = MI.getOperand(1);
1256       const MachineOperand &Op2 = MI.getOperand(2);
1257       const MachineOperand &Op3 = MI.getOperand(3);
1258       Register Rd = Op0.getReg();
1259       Register Pu = Op1.getReg();
1260       Register Rs = Op2.getReg();
1261       Register Rt = Op3.getReg();
1262       DebugLoc DL = MI.getDebugLoc();
1263       unsigned K1 = getKillRegState(Op1.isKill());
1264       unsigned K2 = getKillRegState(Op2.isKill());
1265       unsigned K3 = getKillRegState(Op3.isKill());
1266       if (Rd != Rs)
1267         BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd)
1268           .addReg(Pu, (Rd == Rt) ? K1 : 0)
1269           .addReg(Rs, K2);
1270       if (Rd != Rt)
1271         BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd)
1272           .addReg(Pu, K1)
1273           .addReg(Rt, K3);
1274       MBB.erase(MI);
1275       return true;
1276     }
1277     case Hexagon::PS_vselect: {
1278       const MachineOperand &Op0 = MI.getOperand(0);
1279       const MachineOperand &Op1 = MI.getOperand(1);
1280       const MachineOperand &Op2 = MI.getOperand(2);
1281       const MachineOperand &Op3 = MI.getOperand(3);
1282       getLiveOutRegsAt(LiveOut, MI);
1283       bool IsDestLive = !LiveOut.available(MRI, Op0.getReg());
1284       Register PReg = Op1.getReg();
1285       assert(Op1.getSubReg() == 0);
1286       unsigned PState = getRegState(Op1);
1287 
1288       if (Op0.getReg() != Op2.getReg()) {
1289         unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill
1290                                                   : PState;
1291         auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov))
1292                      .add(Op0)
1293                      .addReg(PReg, S)
1294                      .add(Op2);
1295         if (IsDestLive)
1296           T.addReg(Op0.getReg(), RegState::Implicit);
1297         IsDestLive = true;
1298       }
1299       if (Op0.getReg() != Op3.getReg()) {
1300         auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov))
1301                      .add(Op0)
1302                      .addReg(PReg, PState)
1303                      .add(Op3);
1304         if (IsDestLive)
1305           T.addReg(Op0.getReg(), RegState::Implicit);
1306       }
1307       MBB.erase(MI);
1308       return true;
1309     }
1310     case Hexagon::PS_wselect: {
1311       MachineOperand &Op0 = MI.getOperand(0);
1312       MachineOperand &Op1 = MI.getOperand(1);
1313       MachineOperand &Op2 = MI.getOperand(2);
1314       MachineOperand &Op3 = MI.getOperand(3);
1315       getLiveOutRegsAt(LiveOut, MI);
1316       bool IsDestLive = !LiveOut.available(MRI, Op0.getReg());
1317       Register PReg = Op1.getReg();
1318       assert(Op1.getSubReg() == 0);
1319       unsigned PState = getRegState(Op1);
1320 
1321       if (Op0.getReg() != Op2.getReg()) {
1322         unsigned S = Op0.getReg() != Op3.getReg() ? PState & ~RegState::Kill
1323                                                   : PState;
1324         Register SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_lo);
1325         Register SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_hi);
1326         auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine))
1327                      .add(Op0)
1328                      .addReg(PReg, S)
1329                      .addReg(SrcHi)
1330                      .addReg(SrcLo);
1331         if (IsDestLive)
1332           T.addReg(Op0.getReg(), RegState::Implicit);
1333         IsDestLive = true;
1334       }
1335       if (Op0.getReg() != Op3.getReg()) {
1336         Register SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_lo);
1337         Register SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_hi);
1338         auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine))
1339                      .add(Op0)
1340                      .addReg(PReg, PState)
1341                      .addReg(SrcHi)
1342                      .addReg(SrcLo);
1343         if (IsDestLive)
1344           T.addReg(Op0.getReg(), RegState::Implicit);
1345       }
1346       MBB.erase(MI);
1347       return true;
1348     }
1349 
1350     case Hexagon::PS_crash: {
1351       // Generate a misaligned load that is guaranteed to cause a crash.
1352       class CrashPseudoSourceValue : public PseudoSourceValue {
1353       public:
1354         CrashPseudoSourceValue(const TargetInstrInfo &TII)
1355           : PseudoSourceValue(TargetCustom, TII) {}
1356 
1357         bool isConstant(const MachineFrameInfo *) const override {
1358           return false;
1359         }
1360         bool isAliased(const MachineFrameInfo *) const override {
1361           return false;
1362         }
1363         bool mayAlias(const MachineFrameInfo *) const override {
1364           return false;
1365         }
1366         void printCustom(raw_ostream &OS) const override {
1367           OS << "MisalignedCrash";
1368         }
1369       };
1370 
1371       static const CrashPseudoSourceValue CrashPSV(*this);
1372       MachineMemOperand *MMO = MF.getMachineMemOperand(
1373           MachinePointerInfo(&CrashPSV),
1374           MachineMemOperand::MOLoad | MachineMemOperand::MOVolatile, 8,
1375           Align(1));
1376       BuildMI(MBB, MI, DL, get(Hexagon::PS_loadrdabs), Hexagon::D13)
1377         .addImm(0xBADC0FEE)  // Misaligned load.
1378         .addMemOperand(MMO);
1379       MBB.erase(MI);
1380       return true;
1381     }
1382 
1383     case Hexagon::PS_tailcall_i:
1384       MI.setDesc(get(Hexagon::J2_jump));
1385       return true;
1386     case Hexagon::PS_tailcall_r:
1387     case Hexagon::PS_jmpret:
1388       MI.setDesc(get(Hexagon::J2_jumpr));
1389       return true;
1390     case Hexagon::PS_jmprett:
1391       MI.setDesc(get(Hexagon::J2_jumprt));
1392       return true;
1393     case Hexagon::PS_jmpretf:
1394       MI.setDesc(get(Hexagon::J2_jumprf));
1395       return true;
1396     case Hexagon::PS_jmprettnewpt:
1397       MI.setDesc(get(Hexagon::J2_jumprtnewpt));
1398       return true;
1399     case Hexagon::PS_jmpretfnewpt:
1400       MI.setDesc(get(Hexagon::J2_jumprfnewpt));
1401       return true;
1402     case Hexagon::PS_jmprettnew:
1403       MI.setDesc(get(Hexagon::J2_jumprtnew));
1404       return true;
1405     case Hexagon::PS_jmpretfnew:
1406       MI.setDesc(get(Hexagon::J2_jumprfnew));
1407       return true;
1408 
1409     case Hexagon::PS_loadrub_pci:
1410       return RealCirc(Hexagon::L2_loadrub_pci, /*HasImm*/true,  /*MxOp*/4);
1411     case Hexagon::PS_loadrb_pci:
1412       return RealCirc(Hexagon::L2_loadrb_pci,  /*HasImm*/true,  /*MxOp*/4);
1413     case Hexagon::PS_loadruh_pci:
1414       return RealCirc(Hexagon::L2_loadruh_pci, /*HasImm*/true,  /*MxOp*/4);
1415     case Hexagon::PS_loadrh_pci:
1416       return RealCirc(Hexagon::L2_loadrh_pci,  /*HasImm*/true,  /*MxOp*/4);
1417     case Hexagon::PS_loadri_pci:
1418       return RealCirc(Hexagon::L2_loadri_pci,  /*HasImm*/true,  /*MxOp*/4);
1419     case Hexagon::PS_loadrd_pci:
1420       return RealCirc(Hexagon::L2_loadrd_pci,  /*HasImm*/true,  /*MxOp*/4);
1421     case Hexagon::PS_loadrub_pcr:
1422       return RealCirc(Hexagon::L2_loadrub_pcr, /*HasImm*/false, /*MxOp*/3);
1423     case Hexagon::PS_loadrb_pcr:
1424       return RealCirc(Hexagon::L2_loadrb_pcr,  /*HasImm*/false, /*MxOp*/3);
1425     case Hexagon::PS_loadruh_pcr:
1426       return RealCirc(Hexagon::L2_loadruh_pcr, /*HasImm*/false, /*MxOp*/3);
1427     case Hexagon::PS_loadrh_pcr:
1428       return RealCirc(Hexagon::L2_loadrh_pcr,  /*HasImm*/false, /*MxOp*/3);
1429     case Hexagon::PS_loadri_pcr:
1430       return RealCirc(Hexagon::L2_loadri_pcr,  /*HasImm*/false, /*MxOp*/3);
1431     case Hexagon::PS_loadrd_pcr:
1432       return RealCirc(Hexagon::L2_loadrd_pcr,  /*HasImm*/false, /*MxOp*/3);
1433     case Hexagon::PS_storerb_pci:
1434       return RealCirc(Hexagon::S2_storerb_pci, /*HasImm*/true,  /*MxOp*/3);
1435     case Hexagon::PS_storerh_pci:
1436       return RealCirc(Hexagon::S2_storerh_pci, /*HasImm*/true,  /*MxOp*/3);
1437     case Hexagon::PS_storerf_pci:
1438       return RealCirc(Hexagon::S2_storerf_pci, /*HasImm*/true,  /*MxOp*/3);
1439     case Hexagon::PS_storeri_pci:
1440       return RealCirc(Hexagon::S2_storeri_pci, /*HasImm*/true,  /*MxOp*/3);
1441     case Hexagon::PS_storerd_pci:
1442       return RealCirc(Hexagon::S2_storerd_pci, /*HasImm*/true,  /*MxOp*/3);
1443     case Hexagon::PS_storerb_pcr:
1444       return RealCirc(Hexagon::S2_storerb_pcr, /*HasImm*/false, /*MxOp*/2);
1445     case Hexagon::PS_storerh_pcr:
1446       return RealCirc(Hexagon::S2_storerh_pcr, /*HasImm*/false, /*MxOp*/2);
1447     case Hexagon::PS_storerf_pcr:
1448       return RealCirc(Hexagon::S2_storerf_pcr, /*HasImm*/false, /*MxOp*/2);
1449     case Hexagon::PS_storeri_pcr:
1450       return RealCirc(Hexagon::S2_storeri_pcr, /*HasImm*/false, /*MxOp*/2);
1451     case Hexagon::PS_storerd_pcr:
1452       return RealCirc(Hexagon::S2_storerd_pcr, /*HasImm*/false, /*MxOp*/2);
1453   }
1454 
1455   return false;
1456 }
1457 
1458 MachineBasicBlock::instr_iterator
expandVGatherPseudo(MachineInstr & MI) const1459 HexagonInstrInfo::expandVGatherPseudo(MachineInstr &MI) const {
1460   MachineBasicBlock &MBB = *MI.getParent();
1461   const DebugLoc &DL = MI.getDebugLoc();
1462   unsigned Opc = MI.getOpcode();
1463   MachineBasicBlock::iterator First;
1464 
1465   switch (Opc) {
1466     case Hexagon::V6_vgathermh_pseudo:
1467       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermh))
1468                   .add(MI.getOperand(1))
1469                   .add(MI.getOperand(2))
1470                   .add(MI.getOperand(3));
1471       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1472           .add(MI.getOperand(0))
1473           .addImm(0)
1474           .addReg(Hexagon::VTMP);
1475       MBB.erase(MI);
1476       return First.getInstrIterator();
1477 
1478     case Hexagon::V6_vgathermw_pseudo:
1479       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermw))
1480                   .add(MI.getOperand(1))
1481                   .add(MI.getOperand(2))
1482                   .add(MI.getOperand(3));
1483       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1484           .add(MI.getOperand(0))
1485           .addImm(0)
1486           .addReg(Hexagon::VTMP);
1487       MBB.erase(MI);
1488       return First.getInstrIterator();
1489 
1490     case Hexagon::V6_vgathermhw_pseudo:
1491       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhw))
1492                   .add(MI.getOperand(1))
1493                   .add(MI.getOperand(2))
1494                   .add(MI.getOperand(3));
1495       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1496           .add(MI.getOperand(0))
1497           .addImm(0)
1498           .addReg(Hexagon::VTMP);
1499       MBB.erase(MI);
1500       return First.getInstrIterator();
1501 
1502     case Hexagon::V6_vgathermhq_pseudo:
1503       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhq))
1504                   .add(MI.getOperand(1))
1505                   .add(MI.getOperand(2))
1506                   .add(MI.getOperand(3))
1507                   .add(MI.getOperand(4));
1508       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1509           .add(MI.getOperand(0))
1510           .addImm(0)
1511           .addReg(Hexagon::VTMP);
1512       MBB.erase(MI);
1513       return First.getInstrIterator();
1514 
1515     case Hexagon::V6_vgathermwq_pseudo:
1516       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermwq))
1517                   .add(MI.getOperand(1))
1518                   .add(MI.getOperand(2))
1519                   .add(MI.getOperand(3))
1520                   .add(MI.getOperand(4));
1521       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1522           .add(MI.getOperand(0))
1523           .addImm(0)
1524           .addReg(Hexagon::VTMP);
1525       MBB.erase(MI);
1526       return First.getInstrIterator();
1527 
1528     case Hexagon::V6_vgathermhwq_pseudo:
1529       First = BuildMI(MBB, MI, DL, get(Hexagon::V6_vgathermhwq))
1530                   .add(MI.getOperand(1))
1531                   .add(MI.getOperand(2))
1532                   .add(MI.getOperand(3))
1533                   .add(MI.getOperand(4));
1534       BuildMI(MBB, MI, DL, get(Hexagon::V6_vS32b_new_ai))
1535           .add(MI.getOperand(0))
1536           .addImm(0)
1537           .addReg(Hexagon::VTMP);
1538       MBB.erase(MI);
1539       return First.getInstrIterator();
1540   }
1541 
1542   return MI.getIterator();
1543 }
1544 
1545 // We indicate that we want to reverse the branch by
1546 // inserting the reversed branching opcode.
reverseBranchCondition(SmallVectorImpl<MachineOperand> & Cond) const1547 bool HexagonInstrInfo::reverseBranchCondition(
1548       SmallVectorImpl<MachineOperand> &Cond) const {
1549   if (Cond.empty())
1550     return true;
1551   assert(Cond[0].isImm() && "First entry in the cond vector not imm-val");
1552   unsigned opcode = Cond[0].getImm();
1553   //unsigned temp;
1554   assert(get(opcode).isBranch() && "Should be a branching condition.");
1555   if (isEndLoopN(opcode))
1556     return true;
1557   unsigned NewOpcode = getInvertedPredicatedOpcode(opcode);
1558   Cond[0].setImm(NewOpcode);
1559   return false;
1560 }
1561 
insertNoop(MachineBasicBlock & MBB,MachineBasicBlock::iterator MI) const1562 void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB,
1563       MachineBasicBlock::iterator MI) const {
1564   DebugLoc DL;
1565   BuildMI(MBB, MI, DL, get(Hexagon::A2_nop));
1566 }
1567 
isPostIncrement(const MachineInstr & MI) const1568 bool HexagonInstrInfo::isPostIncrement(const MachineInstr &MI) const {
1569   return getAddrMode(MI) == HexagonII::PostInc;
1570 }
1571 
1572 // Returns true if an instruction is predicated irrespective of the predicate
1573 // sense. For example, all of the following will return true.
1574 // if (p0) R1 = add(R2, R3)
1575 // if (!p0) R1 = add(R2, R3)
1576 // if (p0.new) R1 = add(R2, R3)
1577 // if (!p0.new) R1 = add(R2, R3)
1578 // Note: New-value stores are not included here as in the current
1579 // implementation, we don't need to check their predicate sense.
isPredicated(const MachineInstr & MI) const1580 bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const {
1581   const uint64_t F = MI.getDesc().TSFlags;
1582   return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
1583 }
1584 
PredicateInstruction(MachineInstr & MI,ArrayRef<MachineOperand> Cond) const1585 bool HexagonInstrInfo::PredicateInstruction(
1586     MachineInstr &MI, ArrayRef<MachineOperand> Cond) const {
1587   if (Cond.empty() || isNewValueJump(Cond[0].getImm()) ||
1588       isEndLoopN(Cond[0].getImm())) {
1589     LLVM_DEBUG(dbgs() << "\nCannot predicate:"; MI.dump(););
1590     return false;
1591   }
1592   int Opc = MI.getOpcode();
1593   assert (isPredicable(MI) && "Expected predicable instruction");
1594   bool invertJump = predOpcodeHasNot(Cond);
1595 
1596   // We have to predicate MI "in place", i.e. after this function returns,
1597   // MI will need to be transformed into a predicated form. To avoid com-
1598   // plicated manipulations with the operands (handling tied operands,
1599   // etc.), build a new temporary instruction, then overwrite MI with it.
1600 
1601   MachineBasicBlock &B = *MI.getParent();
1602   DebugLoc DL = MI.getDebugLoc();
1603   unsigned PredOpc = getCondOpcode(Opc, invertJump);
1604   MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc));
1605   unsigned NOp = 0, NumOps = MI.getNumOperands();
1606   while (NOp < NumOps) {
1607     MachineOperand &Op = MI.getOperand(NOp);
1608     if (!Op.isReg() || !Op.isDef() || Op.isImplicit())
1609       break;
1610     T.add(Op);
1611     NOp++;
1612   }
1613 
1614   unsigned PredReg, PredRegPos, PredRegFlags;
1615   bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags);
1616   (void)GotPredReg;
1617   assert(GotPredReg);
1618   T.addReg(PredReg, PredRegFlags);
1619   while (NOp < NumOps)
1620     T.add(MI.getOperand(NOp++));
1621 
1622   MI.setDesc(get(PredOpc));
1623   while (unsigned n = MI.getNumOperands())
1624     MI.RemoveOperand(n-1);
1625   for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i)
1626     MI.addOperand(T->getOperand(i));
1627 
1628   MachineBasicBlock::instr_iterator TI = T->getIterator();
1629   B.erase(TI);
1630 
1631   MachineRegisterInfo &MRI = B.getParent()->getRegInfo();
1632   MRI.clearKillFlags(PredReg);
1633   return true;
1634 }
1635 
SubsumesPredicate(ArrayRef<MachineOperand> Pred1,ArrayRef<MachineOperand> Pred2) const1636 bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1637       ArrayRef<MachineOperand> Pred2) const {
1638   // TODO: Fix this
1639   return false;
1640 }
1641 
ClobbersPredicate(MachineInstr & MI,std::vector<MachineOperand> & Pred,bool SkipDead) const1642 bool HexagonInstrInfo::ClobbersPredicate(MachineInstr &MI,
1643                                          std::vector<MachineOperand> &Pred,
1644                                          bool SkipDead) const {
1645   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
1646 
1647   for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) {
1648     MachineOperand MO = MI.getOperand(oper);
1649     if (MO.isReg()) {
1650       if (!MO.isDef())
1651         continue;
1652       const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg());
1653       if (RC == &Hexagon::PredRegsRegClass) {
1654         Pred.push_back(MO);
1655         return true;
1656       }
1657       continue;
1658     } else if (MO.isRegMask()) {
1659       for (unsigned PR : Hexagon::PredRegsRegClass) {
1660         if (!MI.modifiesRegister(PR, &HRI))
1661           continue;
1662         Pred.push_back(MO);
1663         return true;
1664       }
1665     }
1666   }
1667   return false;
1668 }
1669 
isPredicable(const MachineInstr & MI) const1670 bool HexagonInstrInfo::isPredicable(const MachineInstr &MI) const {
1671   if (!MI.getDesc().isPredicable())
1672     return false;
1673 
1674   if (MI.isCall() || isTailCall(MI)) {
1675     if (!Subtarget.usePredicatedCalls())
1676       return false;
1677   }
1678 
1679   // HVX loads are not predicable on v60, but are on v62.
1680   if (!Subtarget.hasV62Ops()) {
1681     switch (MI.getOpcode()) {
1682       case Hexagon::V6_vL32b_ai:
1683       case Hexagon::V6_vL32b_pi:
1684       case Hexagon::V6_vL32b_ppu:
1685       case Hexagon::V6_vL32b_cur_ai:
1686       case Hexagon::V6_vL32b_cur_pi:
1687       case Hexagon::V6_vL32b_cur_ppu:
1688       case Hexagon::V6_vL32b_nt_ai:
1689       case Hexagon::V6_vL32b_nt_pi:
1690       case Hexagon::V6_vL32b_nt_ppu:
1691       case Hexagon::V6_vL32b_tmp_ai:
1692       case Hexagon::V6_vL32b_tmp_pi:
1693       case Hexagon::V6_vL32b_tmp_ppu:
1694       case Hexagon::V6_vL32b_nt_cur_ai:
1695       case Hexagon::V6_vL32b_nt_cur_pi:
1696       case Hexagon::V6_vL32b_nt_cur_ppu:
1697       case Hexagon::V6_vL32b_nt_tmp_ai:
1698       case Hexagon::V6_vL32b_nt_tmp_pi:
1699       case Hexagon::V6_vL32b_nt_tmp_ppu:
1700         return false;
1701     }
1702   }
1703   return true;
1704 }
1705 
isSchedulingBoundary(const MachineInstr & MI,const MachineBasicBlock * MBB,const MachineFunction & MF) const1706 bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
1707                                             const MachineBasicBlock *MBB,
1708                                             const MachineFunction &MF) const {
1709   // Debug info is never a scheduling boundary. It's necessary to be explicit
1710   // due to the special treatment of IT instructions below, otherwise a
1711   // dbg_value followed by an IT will result in the IT instruction being
1712   // considered a scheduling hazard, which is wrong. It should be the actual
1713   // instruction preceding the dbg_value instruction(s), just like it is
1714   // when debug info is not present.
1715   if (MI.isDebugInstr())
1716     return false;
1717 
1718   // Throwing call is a boundary.
1719   if (MI.isCall()) {
1720     // Don't mess around with no return calls.
1721     if (doesNotReturn(MI))
1722       return true;
1723     // If any of the block's successors is a landing pad, this could be a
1724     // throwing call.
1725     for (auto I : MBB->successors())
1726       if (I->isEHPad())
1727         return true;
1728   }
1729 
1730   // Terminators and labels can't be scheduled around.
1731   if (MI.getDesc().isTerminator() || MI.isPosition())
1732     return true;
1733 
1734   // INLINEASM_BR can jump to another block
1735   if (MI.getOpcode() == TargetOpcode::INLINEASM_BR)
1736     return true;
1737 
1738   if (MI.isInlineAsm() && !ScheduleInlineAsm)
1739     return true;
1740 
1741   return false;
1742 }
1743 
1744 /// Measure the specified inline asm to determine an approximation of its
1745 /// length.
1746 /// Comments (which run till the next SeparatorString or newline) do not
1747 /// count as an instruction.
1748 /// Any other non-whitespace text is considered an instruction, with
1749 /// multiple instructions separated by SeparatorString or newlines.
1750 /// Variable-length instructions are not handled here; this function
1751 /// may be overloaded in the target code to do that.
1752 /// Hexagon counts the number of ##'s and adjust for that many
1753 /// constant exenders.
getInlineAsmLength(const char * Str,const MCAsmInfo & MAI,const TargetSubtargetInfo * STI) const1754 unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str,
1755                                               const MCAsmInfo &MAI,
1756                                               const TargetSubtargetInfo *STI) const {
1757   StringRef AStr(Str);
1758   // Count the number of instructions in the asm.
1759   bool atInsnStart = true;
1760   unsigned Length = 0;
1761   const unsigned MaxInstLength = MAI.getMaxInstLength(STI);
1762   for (; *Str; ++Str) {
1763     if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
1764                                 strlen(MAI.getSeparatorString())) == 0)
1765       atInsnStart = true;
1766     if (atInsnStart && !isSpace(static_cast<unsigned char>(*Str))) {
1767       Length += MaxInstLength;
1768       atInsnStart = false;
1769     }
1770     if (atInsnStart && strncmp(Str, MAI.getCommentString().data(),
1771                                MAI.getCommentString().size()) == 0)
1772       atInsnStart = false;
1773   }
1774 
1775   // Add to size number of constant extenders seen * 4.
1776   StringRef Occ("##");
1777   Length += AStr.count(Occ)*4;
1778   return Length;
1779 }
1780 
1781 ScheduleHazardRecognizer*
CreateTargetPostRAHazardRecognizer(const InstrItineraryData * II,const ScheduleDAG * DAG) const1782 HexagonInstrInfo::CreateTargetPostRAHazardRecognizer(
1783       const InstrItineraryData *II, const ScheduleDAG *DAG) const {
1784   if (UseDFAHazardRec)
1785     return new HexagonHazardRecognizer(II, this, Subtarget);
1786   return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
1787 }
1788 
1789 /// For a comparison instruction, return the source registers in
1790 /// \p SrcReg and \p SrcReg2 if having two register operands, and the value it
1791 /// compares against in CmpValue. Return true if the comparison instruction
1792 /// can be analyzed.
analyzeCompare(const MachineInstr & MI,Register & SrcReg,Register & SrcReg2,int & Mask,int & Value) const1793 bool HexagonInstrInfo::analyzeCompare(const MachineInstr &MI, Register &SrcReg,
1794                                       Register &SrcReg2, int &Mask,
1795                                       int &Value) const {
1796   unsigned Opc = MI.getOpcode();
1797 
1798   // Set mask and the first source register.
1799   switch (Opc) {
1800     case Hexagon::C2_cmpeq:
1801     case Hexagon::C2_cmpeqp:
1802     case Hexagon::C2_cmpgt:
1803     case Hexagon::C2_cmpgtp:
1804     case Hexagon::C2_cmpgtu:
1805     case Hexagon::C2_cmpgtup:
1806     case Hexagon::C4_cmpneq:
1807     case Hexagon::C4_cmplte:
1808     case Hexagon::C4_cmplteu:
1809     case Hexagon::C2_cmpeqi:
1810     case Hexagon::C2_cmpgti:
1811     case Hexagon::C2_cmpgtui:
1812     case Hexagon::C4_cmpneqi:
1813     case Hexagon::C4_cmplteui:
1814     case Hexagon::C4_cmpltei:
1815       SrcReg = MI.getOperand(1).getReg();
1816       Mask = ~0;
1817       break;
1818     case Hexagon::A4_cmpbeq:
1819     case Hexagon::A4_cmpbgt:
1820     case Hexagon::A4_cmpbgtu:
1821     case Hexagon::A4_cmpbeqi:
1822     case Hexagon::A4_cmpbgti:
1823     case Hexagon::A4_cmpbgtui:
1824       SrcReg = MI.getOperand(1).getReg();
1825       Mask = 0xFF;
1826       break;
1827     case Hexagon::A4_cmpheq:
1828     case Hexagon::A4_cmphgt:
1829     case Hexagon::A4_cmphgtu:
1830     case Hexagon::A4_cmpheqi:
1831     case Hexagon::A4_cmphgti:
1832     case Hexagon::A4_cmphgtui:
1833       SrcReg = MI.getOperand(1).getReg();
1834       Mask = 0xFFFF;
1835       break;
1836   }
1837 
1838   // Set the value/second source register.
1839   switch (Opc) {
1840     case Hexagon::C2_cmpeq:
1841     case Hexagon::C2_cmpeqp:
1842     case Hexagon::C2_cmpgt:
1843     case Hexagon::C2_cmpgtp:
1844     case Hexagon::C2_cmpgtu:
1845     case Hexagon::C2_cmpgtup:
1846     case Hexagon::A4_cmpbeq:
1847     case Hexagon::A4_cmpbgt:
1848     case Hexagon::A4_cmpbgtu:
1849     case Hexagon::A4_cmpheq:
1850     case Hexagon::A4_cmphgt:
1851     case Hexagon::A4_cmphgtu:
1852     case Hexagon::C4_cmpneq:
1853     case Hexagon::C4_cmplte:
1854     case Hexagon::C4_cmplteu:
1855       SrcReg2 = MI.getOperand(2).getReg();
1856       return true;
1857 
1858     case Hexagon::C2_cmpeqi:
1859     case Hexagon::C2_cmpgtui:
1860     case Hexagon::C2_cmpgti:
1861     case Hexagon::C4_cmpneqi:
1862     case Hexagon::C4_cmplteui:
1863     case Hexagon::C4_cmpltei:
1864     case Hexagon::A4_cmpbeqi:
1865     case Hexagon::A4_cmpbgti:
1866     case Hexagon::A4_cmpbgtui:
1867     case Hexagon::A4_cmpheqi:
1868     case Hexagon::A4_cmphgti:
1869     case Hexagon::A4_cmphgtui: {
1870       SrcReg2 = 0;
1871       const MachineOperand &Op2 = MI.getOperand(2);
1872       if (!Op2.isImm())
1873         return false;
1874       Value = MI.getOperand(2).getImm();
1875       return true;
1876     }
1877   }
1878 
1879   return false;
1880 }
1881 
getInstrLatency(const InstrItineraryData * ItinData,const MachineInstr & MI,unsigned * PredCost) const1882 unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1883                                            const MachineInstr &MI,
1884                                            unsigned *PredCost) const {
1885   return getInstrTimingClassLatency(ItinData, MI);
1886 }
1887 
CreateTargetScheduleState(const TargetSubtargetInfo & STI) const1888 DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState(
1889     const TargetSubtargetInfo &STI) const {
1890   const InstrItineraryData *II = STI.getInstrItineraryData();
1891   return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II);
1892 }
1893 
1894 // Inspired by this pair:
1895 //  %r13 = L2_loadri_io %r29, 136; mem:LD4[FixedStack0]
1896 //  S2_storeri_io %r29, 132, killed %r1; flags:  mem:ST4[FixedStack1]
1897 // Currently AA considers the addresses in these instructions to be aliasing.
areMemAccessesTriviallyDisjoint(const MachineInstr & MIa,const MachineInstr & MIb) const1898 bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint(
1899     const MachineInstr &MIa, const MachineInstr &MIb) const {
1900   if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() ||
1901       MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
1902     return false;
1903 
1904   // Instructions that are pure loads, not loads and stores like memops are not
1905   // dependent.
1906   if (MIa.mayLoad() && !isMemOp(MIa) && MIb.mayLoad() && !isMemOp(MIb))
1907     return true;
1908 
1909   // Get the base register in MIa.
1910   unsigned BasePosA, OffsetPosA;
1911   if (!getBaseAndOffsetPosition(MIa, BasePosA, OffsetPosA))
1912     return false;
1913   const MachineOperand &BaseA = MIa.getOperand(BasePosA);
1914   Register BaseRegA = BaseA.getReg();
1915   unsigned BaseSubA = BaseA.getSubReg();
1916 
1917   // Get the base register in MIb.
1918   unsigned BasePosB, OffsetPosB;
1919   if (!getBaseAndOffsetPosition(MIb, BasePosB, OffsetPosB))
1920     return false;
1921   const MachineOperand &BaseB = MIb.getOperand(BasePosB);
1922   Register BaseRegB = BaseB.getReg();
1923   unsigned BaseSubB = BaseB.getSubReg();
1924 
1925   if (BaseRegA != BaseRegB || BaseSubA != BaseSubB)
1926     return false;
1927 
1928   // Get the access sizes.
1929   unsigned SizeA = getMemAccessSize(MIa);
1930   unsigned SizeB = getMemAccessSize(MIb);
1931 
1932   // Get the offsets. Handle immediates only for now.
1933   const MachineOperand &OffA = MIa.getOperand(OffsetPosA);
1934   const MachineOperand &OffB = MIb.getOperand(OffsetPosB);
1935   if (!MIa.getOperand(OffsetPosA).isImm() ||
1936       !MIb.getOperand(OffsetPosB).isImm())
1937     return false;
1938   int OffsetA = isPostIncrement(MIa) ? 0 : OffA.getImm();
1939   int OffsetB = isPostIncrement(MIb) ? 0 : OffB.getImm();
1940 
1941   // This is a mem access with the same base register and known offsets from it.
1942   // Reason about it.
1943   if (OffsetA > OffsetB) {
1944     uint64_t OffDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB);
1945     return SizeB <= OffDiff;
1946   }
1947   if (OffsetA < OffsetB) {
1948     uint64_t OffDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA);
1949     return SizeA <= OffDiff;
1950   }
1951 
1952   return false;
1953 }
1954 
1955 /// If the instruction is an increment of a constant value, return the amount.
getIncrementValue(const MachineInstr & MI,int & Value) const1956 bool HexagonInstrInfo::getIncrementValue(const MachineInstr &MI,
1957       int &Value) const {
1958   if (isPostIncrement(MI)) {
1959     unsigned BasePos = 0, OffsetPos = 0;
1960     if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos))
1961       return false;
1962     const MachineOperand &OffsetOp = MI.getOperand(OffsetPos);
1963     if (OffsetOp.isImm()) {
1964       Value = OffsetOp.getImm();
1965       return true;
1966     }
1967   } else if (MI.getOpcode() == Hexagon::A2_addi) {
1968     const MachineOperand &AddOp = MI.getOperand(2);
1969     if (AddOp.isImm()) {
1970       Value = AddOp.getImm();
1971       return true;
1972     }
1973   }
1974 
1975   return false;
1976 }
1977 
1978 std::pair<unsigned, unsigned>
decomposeMachineOperandsTargetFlags(unsigned TF) const1979 HexagonInstrInfo::decomposeMachineOperandsTargetFlags(unsigned TF) const {
1980   return std::make_pair(TF & ~HexagonII::MO_Bitmasks,
1981                         TF & HexagonII::MO_Bitmasks);
1982 }
1983 
1984 ArrayRef<std::pair<unsigned, const char*>>
getSerializableDirectMachineOperandTargetFlags() const1985 HexagonInstrInfo::getSerializableDirectMachineOperandTargetFlags() const {
1986   using namespace HexagonII;
1987 
1988   static const std::pair<unsigned, const char*> Flags[] = {
1989     {MO_PCREL,  "hexagon-pcrel"},
1990     {MO_GOT,    "hexagon-got"},
1991     {MO_LO16,   "hexagon-lo16"},
1992     {MO_HI16,   "hexagon-hi16"},
1993     {MO_GPREL,  "hexagon-gprel"},
1994     {MO_GDGOT,  "hexagon-gdgot"},
1995     {MO_GDPLT,  "hexagon-gdplt"},
1996     {MO_IE,     "hexagon-ie"},
1997     {MO_IEGOT,  "hexagon-iegot"},
1998     {MO_TPREL,  "hexagon-tprel"}
1999   };
2000   return makeArrayRef(Flags);
2001 }
2002 
2003 ArrayRef<std::pair<unsigned, const char*>>
getSerializableBitmaskMachineOperandTargetFlags() const2004 HexagonInstrInfo::getSerializableBitmaskMachineOperandTargetFlags() const {
2005   using namespace HexagonII;
2006 
2007   static const std::pair<unsigned, const char*> Flags[] = {
2008     {HMOTF_ConstExtended, "hexagon-ext"}
2009   };
2010   return makeArrayRef(Flags);
2011 }
2012 
createVR(MachineFunction * MF,MVT VT) const2013 unsigned HexagonInstrInfo::createVR(MachineFunction *MF, MVT VT) const {
2014   MachineRegisterInfo &MRI = MF->getRegInfo();
2015   const TargetRegisterClass *TRC;
2016   if (VT == MVT::i1) {
2017     TRC = &Hexagon::PredRegsRegClass;
2018   } else if (VT == MVT::i32 || VT == MVT::f32) {
2019     TRC = &Hexagon::IntRegsRegClass;
2020   } else if (VT == MVT::i64 || VT == MVT::f64) {
2021     TRC = &Hexagon::DoubleRegsRegClass;
2022   } else {
2023     llvm_unreachable("Cannot handle this register class");
2024   }
2025 
2026   Register NewReg = MRI.createVirtualRegister(TRC);
2027   return NewReg;
2028 }
2029 
isAbsoluteSet(const MachineInstr & MI) const2030 bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr &MI) const {
2031   return (getAddrMode(MI) == HexagonII::AbsoluteSet);
2032 }
2033 
isAccumulator(const MachineInstr & MI) const2034 bool HexagonInstrInfo::isAccumulator(const MachineInstr &MI) const {
2035   const uint64_t F = MI.getDesc().TSFlags;
2036   return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask);
2037 }
2038 
isBaseImmOffset(const MachineInstr & MI) const2039 bool HexagonInstrInfo::isBaseImmOffset(const MachineInstr &MI) const {
2040   return getAddrMode(MI) == HexagonII::BaseImmOffset;
2041 }
2042 
isComplex(const MachineInstr & MI) const2043 bool HexagonInstrInfo::isComplex(const MachineInstr &MI) const {
2044   return !isTC1(MI) && !isTC2Early(MI) && !MI.getDesc().mayLoad() &&
2045          !MI.getDesc().mayStore() &&
2046          MI.getDesc().getOpcode() != Hexagon::S2_allocframe &&
2047          MI.getDesc().getOpcode() != Hexagon::L2_deallocframe &&
2048          !isMemOp(MI) && !MI.isBranch() && !MI.isReturn() && !MI.isCall();
2049 }
2050 
2051 // Return true if the instruction is a compund branch instruction.
isCompoundBranchInstr(const MachineInstr & MI) const2052 bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr &MI) const {
2053   return getType(MI) == HexagonII::TypeCJ && MI.isBranch();
2054 }
2055 
2056 // TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle
2057 // isFPImm and later getFPImm as well.
isConstExtended(const MachineInstr & MI) const2058 bool HexagonInstrInfo::isConstExtended(const MachineInstr &MI) const {
2059   const uint64_t F = MI.getDesc().TSFlags;
2060   unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask;
2061   if (isExtended) // Instruction must be extended.
2062     return true;
2063 
2064   unsigned isExtendable =
2065     (F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask;
2066   if (!isExtendable)
2067     return false;
2068 
2069   if (MI.isCall())
2070     return false;
2071 
2072   short ExtOpNum = getCExtOpNum(MI);
2073   const MachineOperand &MO = MI.getOperand(ExtOpNum);
2074   // Use MO operand flags to determine if MO
2075   // has the HMOTF_ConstExtended flag set.
2076   if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended)
2077     return true;
2078   // If this is a Machine BB address we are talking about, and it is
2079   // not marked as extended, say so.
2080   if (MO.isMBB())
2081     return false;
2082 
2083   // We could be using an instruction with an extendable immediate and shoehorn
2084   // a global address into it. If it is a global address it will be constant
2085   // extended. We do this for COMBINE.
2086   if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() ||
2087       MO.isJTI() || MO.isCPI() || MO.isFPImm())
2088     return true;
2089 
2090   // If the extendable operand is not 'Immediate' type, the instruction should
2091   // have 'isExtended' flag set.
2092   assert(MO.isImm() && "Extendable operand must be Immediate type");
2093 
2094   int MinValue = getMinValue(MI);
2095   int MaxValue = getMaxValue(MI);
2096   int ImmValue = MO.getImm();
2097 
2098   return (ImmValue < MinValue || ImmValue > MaxValue);
2099 }
2100 
isDeallocRet(const MachineInstr & MI) const2101 bool HexagonInstrInfo::isDeallocRet(const MachineInstr &MI) const {
2102   switch (MI.getOpcode()) {
2103   case Hexagon::L4_return:
2104   case Hexagon::L4_return_t:
2105   case Hexagon::L4_return_f:
2106   case Hexagon::L4_return_tnew_pnt:
2107   case Hexagon::L4_return_fnew_pnt:
2108   case Hexagon::L4_return_tnew_pt:
2109   case Hexagon::L4_return_fnew_pt:
2110     return true;
2111   }
2112   return false;
2113 }
2114 
2115 // Return true when ConsMI uses a register defined by ProdMI.
isDependent(const MachineInstr & ProdMI,const MachineInstr & ConsMI) const2116 bool HexagonInstrInfo::isDependent(const MachineInstr &ProdMI,
2117       const MachineInstr &ConsMI) const {
2118   if (!ProdMI.getDesc().getNumDefs())
2119     return false;
2120   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
2121 
2122   SmallVector<unsigned, 4> DefsA;
2123   SmallVector<unsigned, 4> DefsB;
2124   SmallVector<unsigned, 8> UsesA;
2125   SmallVector<unsigned, 8> UsesB;
2126 
2127   parseOperands(ProdMI, DefsA, UsesA);
2128   parseOperands(ConsMI, DefsB, UsesB);
2129 
2130   for (auto &RegA : DefsA)
2131     for (auto &RegB : UsesB) {
2132       // True data dependency.
2133       if (RegA == RegB)
2134         return true;
2135 
2136       if (Register::isPhysicalRegister(RegA))
2137         for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs)
2138           if (RegB == *SubRegs)
2139             return true;
2140 
2141       if (Register::isPhysicalRegister(RegB))
2142         for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs)
2143           if (RegA == *SubRegs)
2144             return true;
2145     }
2146 
2147   return false;
2148 }
2149 
2150 // Returns true if the instruction is alread a .cur.
isDotCurInst(const MachineInstr & MI) const2151 bool HexagonInstrInfo::isDotCurInst(const MachineInstr &MI) const {
2152   switch (MI.getOpcode()) {
2153   case Hexagon::V6_vL32b_cur_pi:
2154   case Hexagon::V6_vL32b_cur_ai:
2155     return true;
2156   }
2157   return false;
2158 }
2159 
2160 // Returns true, if any one of the operands is a dot new
2161 // insn, whether it is predicated dot new or register dot new.
isDotNewInst(const MachineInstr & MI) const2162 bool HexagonInstrInfo::isDotNewInst(const MachineInstr &MI) const {
2163   if (isNewValueInst(MI) || (isPredicated(MI) && isPredicatedNew(MI)))
2164     return true;
2165 
2166   return false;
2167 }
2168 
2169 /// Symmetrical. See if these two instructions are fit for duplex pair.
isDuplexPair(const MachineInstr & MIa,const MachineInstr & MIb) const2170 bool HexagonInstrInfo::isDuplexPair(const MachineInstr &MIa,
2171       const MachineInstr &MIb) const {
2172   HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa);
2173   HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb);
2174   return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG));
2175 }
2176 
isEarlySourceInstr(const MachineInstr & MI) const2177 bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr &MI) const {
2178   if (MI.mayLoadOrStore() || MI.isCompare())
2179     return true;
2180 
2181   // Multiply
2182   unsigned SchedClass = MI.getDesc().getSchedClass();
2183   return is_TC4x(SchedClass) || is_TC3x(SchedClass);
2184 }
2185 
isEndLoopN(unsigned Opcode) const2186 bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const {
2187   return (Opcode == Hexagon::ENDLOOP0 ||
2188           Opcode == Hexagon::ENDLOOP1);
2189 }
2190 
isExpr(unsigned OpType) const2191 bool HexagonInstrInfo::isExpr(unsigned OpType) const {
2192   switch(OpType) {
2193   case MachineOperand::MO_MachineBasicBlock:
2194   case MachineOperand::MO_GlobalAddress:
2195   case MachineOperand::MO_ExternalSymbol:
2196   case MachineOperand::MO_JumpTableIndex:
2197   case MachineOperand::MO_ConstantPoolIndex:
2198   case MachineOperand::MO_BlockAddress:
2199     return true;
2200   default:
2201     return false;
2202   }
2203 }
2204 
isExtendable(const MachineInstr & MI) const2205 bool HexagonInstrInfo::isExtendable(const MachineInstr &MI) const {
2206   const MCInstrDesc &MID = MI.getDesc();
2207   const uint64_t F = MID.TSFlags;
2208   if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask)
2209     return true;
2210 
2211   // TODO: This is largely obsolete now. Will need to be removed
2212   // in consecutive patches.
2213   switch (MI.getOpcode()) {
2214     // PS_fi and PS_fia remain special cases.
2215     case Hexagon::PS_fi:
2216     case Hexagon::PS_fia:
2217       return true;
2218     default:
2219       return false;
2220   }
2221   return  false;
2222 }
2223 
2224 // This returns true in two cases:
2225 // - The OP code itself indicates that this is an extended instruction.
2226 // - One of MOs has been marked with HMOTF_ConstExtended flag.
isExtended(const MachineInstr & MI) const2227 bool HexagonInstrInfo::isExtended(const MachineInstr &MI) const {
2228   // First check if this is permanently extended op code.
2229   const uint64_t F = MI.getDesc().TSFlags;
2230   if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask)
2231     return true;
2232   // Use MO operand flags to determine if one of MI's operands
2233   // has HMOTF_ConstExtended flag set.
2234   for (const MachineOperand &MO : MI.operands())
2235     if (MO.getTargetFlags() & HexagonII::HMOTF_ConstExtended)
2236       return true;
2237   return  false;
2238 }
2239 
isFloat(const MachineInstr & MI) const2240 bool HexagonInstrInfo::isFloat(const MachineInstr &MI) const {
2241   unsigned Opcode = MI.getOpcode();
2242   const uint64_t F = get(Opcode).TSFlags;
2243   return (F >> HexagonII::FPPos) & HexagonII::FPMask;
2244 }
2245 
2246 // No V60 HVX VMEM with A_INDIRECT.
isHVXMemWithAIndirect(const MachineInstr & I,const MachineInstr & J) const2247 bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr &I,
2248       const MachineInstr &J) const {
2249   if (!isHVXVec(I))
2250     return false;
2251   if (!I.mayLoad() && !I.mayStore())
2252     return false;
2253   return J.isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J);
2254 }
2255 
isIndirectCall(const MachineInstr & MI) const2256 bool HexagonInstrInfo::isIndirectCall(const MachineInstr &MI) const {
2257   switch (MI.getOpcode()) {
2258   case Hexagon::J2_callr:
2259   case Hexagon::J2_callrf:
2260   case Hexagon::J2_callrt:
2261   case Hexagon::PS_call_nr:
2262     return true;
2263   }
2264   return false;
2265 }
2266 
isIndirectL4Return(const MachineInstr & MI) const2267 bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr &MI) const {
2268   switch (MI.getOpcode()) {
2269   case Hexagon::L4_return:
2270   case Hexagon::L4_return_t:
2271   case Hexagon::L4_return_f:
2272   case Hexagon::L4_return_fnew_pnt:
2273   case Hexagon::L4_return_fnew_pt:
2274   case Hexagon::L4_return_tnew_pnt:
2275   case Hexagon::L4_return_tnew_pt:
2276     return true;
2277   }
2278   return false;
2279 }
2280 
isJumpR(const MachineInstr & MI) const2281 bool HexagonInstrInfo::isJumpR(const MachineInstr &MI) const {
2282   switch (MI.getOpcode()) {
2283   case Hexagon::J2_jumpr:
2284   case Hexagon::J2_jumprt:
2285   case Hexagon::J2_jumprf:
2286   case Hexagon::J2_jumprtnewpt:
2287   case Hexagon::J2_jumprfnewpt:
2288   case Hexagon::J2_jumprtnew:
2289   case Hexagon::J2_jumprfnew:
2290     return true;
2291   }
2292   return false;
2293 }
2294 
2295 // Return true if a given MI can accommodate given offset.
2296 // Use abs estimate as oppose to the exact number.
2297 // TODO: This will need to be changed to use MC level
2298 // definition of instruction extendable field size.
isJumpWithinBranchRange(const MachineInstr & MI,unsigned offset) const2299 bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr &MI,
2300       unsigned offset) const {
2301   // This selection of jump instructions matches to that what
2302   // analyzeBranch can parse, plus NVJ.
2303   if (isNewValueJump(MI)) // r9:2
2304     return isInt<11>(offset);
2305 
2306   switch (MI.getOpcode()) {
2307   // Still missing Jump to address condition on register value.
2308   default:
2309     return false;
2310   case Hexagon::J2_jump: // bits<24> dst; // r22:2
2311   case Hexagon::J2_call:
2312   case Hexagon::PS_call_nr:
2313     return isInt<24>(offset);
2314   case Hexagon::J2_jumpt: //bits<17> dst; // r15:2
2315   case Hexagon::J2_jumpf:
2316   case Hexagon::J2_jumptnew:
2317   case Hexagon::J2_jumptnewpt:
2318   case Hexagon::J2_jumpfnew:
2319   case Hexagon::J2_jumpfnewpt:
2320   case Hexagon::J2_callt:
2321   case Hexagon::J2_callf:
2322     return isInt<17>(offset);
2323   case Hexagon::J2_loop0i:
2324   case Hexagon::J2_loop0iext:
2325   case Hexagon::J2_loop0r:
2326   case Hexagon::J2_loop0rext:
2327   case Hexagon::J2_loop1i:
2328   case Hexagon::J2_loop1iext:
2329   case Hexagon::J2_loop1r:
2330   case Hexagon::J2_loop1rext:
2331     return isInt<9>(offset);
2332   // TODO: Add all the compound branches here. Can we do this in Relation model?
2333   case Hexagon::J4_cmpeqi_tp0_jump_nt:
2334   case Hexagon::J4_cmpeqi_tp1_jump_nt:
2335   case Hexagon::J4_cmpeqn1_tp0_jump_nt:
2336   case Hexagon::J4_cmpeqn1_tp1_jump_nt:
2337     return isInt<11>(offset);
2338   }
2339 }
2340 
isLateInstrFeedsEarlyInstr(const MachineInstr & LRMI,const MachineInstr & ESMI) const2341 bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr &LRMI,
2342       const MachineInstr &ESMI) const {
2343   bool isLate = isLateResultInstr(LRMI);
2344   bool isEarly = isEarlySourceInstr(ESMI);
2345 
2346   LLVM_DEBUG(dbgs() << "V60" << (isLate ? "-LR  " : " --  "));
2347   LLVM_DEBUG(LRMI.dump());
2348   LLVM_DEBUG(dbgs() << "V60" << (isEarly ? "-ES  " : " --  "));
2349   LLVM_DEBUG(ESMI.dump());
2350 
2351   if (isLate && isEarly) {
2352     LLVM_DEBUG(dbgs() << "++Is Late Result feeding Early Source\n");
2353     return true;
2354   }
2355 
2356   return false;
2357 }
2358 
isLateResultInstr(const MachineInstr & MI) const2359 bool HexagonInstrInfo::isLateResultInstr(const MachineInstr &MI) const {
2360   switch (MI.getOpcode()) {
2361   case TargetOpcode::EXTRACT_SUBREG:
2362   case TargetOpcode::INSERT_SUBREG:
2363   case TargetOpcode::SUBREG_TO_REG:
2364   case TargetOpcode::REG_SEQUENCE:
2365   case TargetOpcode::IMPLICIT_DEF:
2366   case TargetOpcode::COPY:
2367   case TargetOpcode::INLINEASM:
2368   case TargetOpcode::PHI:
2369     return false;
2370   default:
2371     break;
2372   }
2373 
2374   unsigned SchedClass = MI.getDesc().getSchedClass();
2375   return !is_TC1(SchedClass);
2376 }
2377 
isLateSourceInstr(const MachineInstr & MI) const2378 bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr &MI) const {
2379   // Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply
2380   // resource, but all operands can be received late like an ALU instruction.
2381   return getType(MI) == HexagonII::TypeCVI_VX_LATE;
2382 }
2383 
isLoopN(const MachineInstr & MI) const2384 bool HexagonInstrInfo::isLoopN(const MachineInstr &MI) const {
2385   unsigned Opcode = MI.getOpcode();
2386   return Opcode == Hexagon::J2_loop0i    ||
2387          Opcode == Hexagon::J2_loop0r    ||
2388          Opcode == Hexagon::J2_loop0iext ||
2389          Opcode == Hexagon::J2_loop0rext ||
2390          Opcode == Hexagon::J2_loop1i    ||
2391          Opcode == Hexagon::J2_loop1r    ||
2392          Opcode == Hexagon::J2_loop1iext ||
2393          Opcode == Hexagon::J2_loop1rext;
2394 }
2395 
isMemOp(const MachineInstr & MI) const2396 bool HexagonInstrInfo::isMemOp(const MachineInstr &MI) const {
2397   switch (MI.getOpcode()) {
2398     default: return false;
2399     case Hexagon::L4_iadd_memopw_io:
2400     case Hexagon::L4_isub_memopw_io:
2401     case Hexagon::L4_add_memopw_io:
2402     case Hexagon::L4_sub_memopw_io:
2403     case Hexagon::L4_and_memopw_io:
2404     case Hexagon::L4_or_memopw_io:
2405     case Hexagon::L4_iadd_memoph_io:
2406     case Hexagon::L4_isub_memoph_io:
2407     case Hexagon::L4_add_memoph_io:
2408     case Hexagon::L4_sub_memoph_io:
2409     case Hexagon::L4_and_memoph_io:
2410     case Hexagon::L4_or_memoph_io:
2411     case Hexagon::L4_iadd_memopb_io:
2412     case Hexagon::L4_isub_memopb_io:
2413     case Hexagon::L4_add_memopb_io:
2414     case Hexagon::L4_sub_memopb_io:
2415     case Hexagon::L4_and_memopb_io:
2416     case Hexagon::L4_or_memopb_io:
2417     case Hexagon::L4_ior_memopb_io:
2418     case Hexagon::L4_ior_memoph_io:
2419     case Hexagon::L4_ior_memopw_io:
2420     case Hexagon::L4_iand_memopb_io:
2421     case Hexagon::L4_iand_memoph_io:
2422     case Hexagon::L4_iand_memopw_io:
2423     return true;
2424   }
2425   return false;
2426 }
2427 
isNewValue(const MachineInstr & MI) const2428 bool HexagonInstrInfo::isNewValue(const MachineInstr &MI) const {
2429   const uint64_t F = MI.getDesc().TSFlags;
2430   return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
2431 }
2432 
isNewValue(unsigned Opcode) const2433 bool HexagonInstrInfo::isNewValue(unsigned Opcode) const {
2434   const uint64_t F = get(Opcode).TSFlags;
2435   return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
2436 }
2437 
isNewValueInst(const MachineInstr & MI) const2438 bool HexagonInstrInfo::isNewValueInst(const MachineInstr &MI) const {
2439   return isNewValueJump(MI) || isNewValueStore(MI);
2440 }
2441 
isNewValueJump(const MachineInstr & MI) const2442 bool HexagonInstrInfo::isNewValueJump(const MachineInstr &MI) const {
2443   return isNewValue(MI) && MI.isBranch();
2444 }
2445 
isNewValueJump(unsigned Opcode) const2446 bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const {
2447   return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode);
2448 }
2449 
isNewValueStore(const MachineInstr & MI) const2450 bool HexagonInstrInfo::isNewValueStore(const MachineInstr &MI) const {
2451   const uint64_t F = MI.getDesc().TSFlags;
2452   return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
2453 }
2454 
isNewValueStore(unsigned Opcode) const2455 bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const {
2456   const uint64_t F = get(Opcode).TSFlags;
2457   return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
2458 }
2459 
2460 // Returns true if a particular operand is extendable for an instruction.
isOperandExtended(const MachineInstr & MI,unsigned OperandNum) const2461 bool HexagonInstrInfo::isOperandExtended(const MachineInstr &MI,
2462     unsigned OperandNum) const {
2463   const uint64_t F = MI.getDesc().TSFlags;
2464   return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask)
2465           == OperandNum;
2466 }
2467 
isPredicatedNew(const MachineInstr & MI) const2468 bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const {
2469   const uint64_t F = MI.getDesc().TSFlags;
2470   assert(isPredicated(MI));
2471   return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
2472 }
2473 
isPredicatedNew(unsigned Opcode) const2474 bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const {
2475   const uint64_t F = get(Opcode).TSFlags;
2476   assert(isPredicated(Opcode));
2477   return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
2478 }
2479 
isPredicatedTrue(const MachineInstr & MI) const2480 bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const {
2481   const uint64_t F = MI.getDesc().TSFlags;
2482   return !((F >> HexagonII::PredicatedFalsePos) &
2483            HexagonII::PredicatedFalseMask);
2484 }
2485 
isPredicatedTrue(unsigned Opcode) const2486 bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const {
2487   const uint64_t F = get(Opcode).TSFlags;
2488   // Make sure that the instruction is predicated.
2489   assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask);
2490   return !((F >> HexagonII::PredicatedFalsePos) &
2491            HexagonII::PredicatedFalseMask);
2492 }
2493 
isPredicated(unsigned Opcode) const2494 bool HexagonInstrInfo::isPredicated(unsigned Opcode) const {
2495   const uint64_t F = get(Opcode).TSFlags;
2496   return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
2497 }
2498 
isPredicateLate(unsigned Opcode) const2499 bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const {
2500   const uint64_t F = get(Opcode).TSFlags;
2501   return (F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask;
2502 }
2503 
isPredictedTaken(unsigned Opcode) const2504 bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const {
2505   const uint64_t F = get(Opcode).TSFlags;
2506   assert(get(Opcode).isBranch() &&
2507          (isPredicatedNew(Opcode) || isNewValue(Opcode)));
2508   return (F >> HexagonII::TakenPos) & HexagonII::TakenMask;
2509 }
2510 
isSaveCalleeSavedRegsCall(const MachineInstr & MI) const2511 bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr &MI) const {
2512   return MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 ||
2513          MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT ||
2514          MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC ||
2515          MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC;
2516 }
2517 
isSignExtendingLoad(const MachineInstr & MI) const2518 bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr &MI) const {
2519   switch (MI.getOpcode()) {
2520   // Byte
2521   case Hexagon::L2_loadrb_io:
2522   case Hexagon::L4_loadrb_ur:
2523   case Hexagon::L4_loadrb_ap:
2524   case Hexagon::L2_loadrb_pr:
2525   case Hexagon::L2_loadrb_pbr:
2526   case Hexagon::L2_loadrb_pi:
2527   case Hexagon::L2_loadrb_pci:
2528   case Hexagon::L2_loadrb_pcr:
2529   case Hexagon::L2_loadbsw2_io:
2530   case Hexagon::L4_loadbsw2_ur:
2531   case Hexagon::L4_loadbsw2_ap:
2532   case Hexagon::L2_loadbsw2_pr:
2533   case Hexagon::L2_loadbsw2_pbr:
2534   case Hexagon::L2_loadbsw2_pi:
2535   case Hexagon::L2_loadbsw2_pci:
2536   case Hexagon::L2_loadbsw2_pcr:
2537   case Hexagon::L2_loadbsw4_io:
2538   case Hexagon::L4_loadbsw4_ur:
2539   case Hexagon::L4_loadbsw4_ap:
2540   case Hexagon::L2_loadbsw4_pr:
2541   case Hexagon::L2_loadbsw4_pbr:
2542   case Hexagon::L2_loadbsw4_pi:
2543   case Hexagon::L2_loadbsw4_pci:
2544   case Hexagon::L2_loadbsw4_pcr:
2545   case Hexagon::L4_loadrb_rr:
2546   case Hexagon::L2_ploadrbt_io:
2547   case Hexagon::L2_ploadrbt_pi:
2548   case Hexagon::L2_ploadrbf_io:
2549   case Hexagon::L2_ploadrbf_pi:
2550   case Hexagon::L2_ploadrbtnew_io:
2551   case Hexagon::L2_ploadrbfnew_io:
2552   case Hexagon::L4_ploadrbt_rr:
2553   case Hexagon::L4_ploadrbf_rr:
2554   case Hexagon::L4_ploadrbtnew_rr:
2555   case Hexagon::L4_ploadrbfnew_rr:
2556   case Hexagon::L2_ploadrbtnew_pi:
2557   case Hexagon::L2_ploadrbfnew_pi:
2558   case Hexagon::L4_ploadrbt_abs:
2559   case Hexagon::L4_ploadrbf_abs:
2560   case Hexagon::L4_ploadrbtnew_abs:
2561   case Hexagon::L4_ploadrbfnew_abs:
2562   case Hexagon::L2_loadrbgp:
2563   // Half
2564   case Hexagon::L2_loadrh_io:
2565   case Hexagon::L4_loadrh_ur:
2566   case Hexagon::L4_loadrh_ap:
2567   case Hexagon::L2_loadrh_pr:
2568   case Hexagon::L2_loadrh_pbr:
2569   case Hexagon::L2_loadrh_pi:
2570   case Hexagon::L2_loadrh_pci:
2571   case Hexagon::L2_loadrh_pcr:
2572   case Hexagon::L4_loadrh_rr:
2573   case Hexagon::L2_ploadrht_io:
2574   case Hexagon::L2_ploadrht_pi:
2575   case Hexagon::L2_ploadrhf_io:
2576   case Hexagon::L2_ploadrhf_pi:
2577   case Hexagon::L2_ploadrhtnew_io:
2578   case Hexagon::L2_ploadrhfnew_io:
2579   case Hexagon::L4_ploadrht_rr:
2580   case Hexagon::L4_ploadrhf_rr:
2581   case Hexagon::L4_ploadrhtnew_rr:
2582   case Hexagon::L4_ploadrhfnew_rr:
2583   case Hexagon::L2_ploadrhtnew_pi:
2584   case Hexagon::L2_ploadrhfnew_pi:
2585   case Hexagon::L4_ploadrht_abs:
2586   case Hexagon::L4_ploadrhf_abs:
2587   case Hexagon::L4_ploadrhtnew_abs:
2588   case Hexagon::L4_ploadrhfnew_abs:
2589   case Hexagon::L2_loadrhgp:
2590     return true;
2591   default:
2592     return false;
2593   }
2594 }
2595 
isSolo(const MachineInstr & MI) const2596 bool HexagonInstrInfo::isSolo(const MachineInstr &MI) const {
2597   const uint64_t F = MI.getDesc().TSFlags;
2598   return (F >> HexagonII::SoloPos) & HexagonII::SoloMask;
2599 }
2600 
isSpillPredRegOp(const MachineInstr & MI) const2601 bool HexagonInstrInfo::isSpillPredRegOp(const MachineInstr &MI) const {
2602   switch (MI.getOpcode()) {
2603   case Hexagon::STriw_pred:
2604   case Hexagon::LDriw_pred:
2605     return true;
2606   default:
2607     return false;
2608   }
2609 }
2610 
isTailCall(const MachineInstr & MI) const2611 bool HexagonInstrInfo::isTailCall(const MachineInstr &MI) const {
2612   if (!MI.isBranch())
2613     return false;
2614 
2615   for (auto &Op : MI.operands())
2616     if (Op.isGlobal() || Op.isSymbol())
2617       return true;
2618   return false;
2619 }
2620 
2621 // Returns true when SU has a timing class TC1.
isTC1(const MachineInstr & MI) const2622 bool HexagonInstrInfo::isTC1(const MachineInstr &MI) const {
2623   unsigned SchedClass = MI.getDesc().getSchedClass();
2624   return is_TC1(SchedClass);
2625 }
2626 
isTC2(const MachineInstr & MI) const2627 bool HexagonInstrInfo::isTC2(const MachineInstr &MI) const {
2628   unsigned SchedClass = MI.getDesc().getSchedClass();
2629   return is_TC2(SchedClass);
2630 }
2631 
isTC2Early(const MachineInstr & MI) const2632 bool HexagonInstrInfo::isTC2Early(const MachineInstr &MI) const {
2633   unsigned SchedClass = MI.getDesc().getSchedClass();
2634   return is_TC2early(SchedClass);
2635 }
2636 
isTC4x(const MachineInstr & MI) const2637 bool HexagonInstrInfo::isTC4x(const MachineInstr &MI) const {
2638   unsigned SchedClass = MI.getDesc().getSchedClass();
2639   return is_TC4x(SchedClass);
2640 }
2641 
2642 // Schedule this ASAP.
isToBeScheduledASAP(const MachineInstr & MI1,const MachineInstr & MI2) const2643 bool HexagonInstrInfo::isToBeScheduledASAP(const MachineInstr &MI1,
2644       const MachineInstr &MI2) const {
2645   if (mayBeCurLoad(MI1)) {
2646     // if (result of SU is used in Next) return true;
2647     Register DstReg = MI1.getOperand(0).getReg();
2648     int N = MI2.getNumOperands();
2649     for (int I = 0; I < N; I++)
2650       if (MI2.getOperand(I).isReg() && DstReg == MI2.getOperand(I).getReg())
2651         return true;
2652   }
2653   if (mayBeNewStore(MI2))
2654     if (MI2.getOpcode() == Hexagon::V6_vS32b_pi)
2655       if (MI1.getOperand(0).isReg() && MI2.getOperand(3).isReg() &&
2656           MI1.getOperand(0).getReg() == MI2.getOperand(3).getReg())
2657         return true;
2658   return false;
2659 }
2660 
isHVXVec(const MachineInstr & MI) const2661 bool HexagonInstrInfo::isHVXVec(const MachineInstr &MI) const {
2662   const uint64_t V = getType(MI);
2663   return HexagonII::TypeCVI_FIRST <= V && V <= HexagonII::TypeCVI_LAST;
2664 }
2665 
2666 // Check if the Offset is a valid auto-inc imm by Load/Store Type.
isValidAutoIncImm(const EVT VT,int Offset) const2667 bool HexagonInstrInfo::isValidAutoIncImm(const EVT VT, int Offset) const {
2668   int Size = VT.getSizeInBits() / 8;
2669   if (Offset % Size != 0)
2670     return false;
2671   int Count = Offset / Size;
2672 
2673   switch (VT.getSimpleVT().SimpleTy) {
2674     // For scalars the auto-inc is s4
2675     case MVT::i8:
2676     case MVT::i16:
2677     case MVT::i32:
2678     case MVT::i64:
2679     case MVT::f32:
2680     case MVT::f64:
2681     case MVT::v2i16:
2682     case MVT::v2i32:
2683     case MVT::v4i8:
2684     case MVT::v4i16:
2685     case MVT::v8i8:
2686       return isInt<4>(Count);
2687     // For HVX vectors the auto-inc is s3
2688     case MVT::v64i8:
2689     case MVT::v32i16:
2690     case MVT::v16i32:
2691     case MVT::v8i64:
2692     case MVT::v128i8:
2693     case MVT::v64i16:
2694     case MVT::v32i32:
2695     case MVT::v16i64:
2696       return isInt<3>(Count);
2697     default:
2698       break;
2699   }
2700 
2701   llvm_unreachable("Not an valid type!");
2702 }
2703 
isValidOffset(unsigned Opcode,int Offset,const TargetRegisterInfo * TRI,bool Extend) const2704 bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset,
2705       const TargetRegisterInfo *TRI, bool Extend) const {
2706   // This function is to check whether the "Offset" is in the correct range of
2707   // the given "Opcode". If "Offset" is not in the correct range, "A2_addi" is
2708   // inserted to calculate the final address. Due to this reason, the function
2709   // assumes that the "Offset" has correct alignment.
2710   // We used to assert if the offset was not properly aligned, however,
2711   // there are cases where a misaligned pointer recast can cause this
2712   // problem, and we need to allow for it. The front end warns of such
2713   // misaligns with respect to load size.
2714   switch (Opcode) {
2715   case Hexagon::PS_vstorerq_ai:
2716   case Hexagon::PS_vstorerv_ai:
2717   case Hexagon::PS_vstorerw_ai:
2718   case Hexagon::PS_vstorerw_nt_ai:
2719   case Hexagon::PS_vloadrq_ai:
2720   case Hexagon::PS_vloadrv_ai:
2721   case Hexagon::PS_vloadrw_ai:
2722   case Hexagon::PS_vloadrw_nt_ai:
2723   case Hexagon::V6_vL32b_ai:
2724   case Hexagon::V6_vS32b_ai:
2725   case Hexagon::V6_vS32b_qpred_ai:
2726   case Hexagon::V6_vS32b_nqpred_ai:
2727   case Hexagon::V6_vL32b_nt_ai:
2728   case Hexagon::V6_vS32b_nt_ai:
2729   case Hexagon::V6_vL32Ub_ai:
2730   case Hexagon::V6_vS32Ub_ai: {
2731     unsigned VectorSize = TRI->getSpillSize(Hexagon::HvxVRRegClass);
2732     assert(isPowerOf2_32(VectorSize));
2733     if (Offset & (VectorSize-1))
2734       return false;
2735     return isInt<4>(Offset >> Log2_32(VectorSize));
2736   }
2737 
2738   case Hexagon::J2_loop0i:
2739   case Hexagon::J2_loop1i:
2740     return isUInt<10>(Offset);
2741 
2742   case Hexagon::S4_storeirb_io:
2743   case Hexagon::S4_storeirbt_io:
2744   case Hexagon::S4_storeirbf_io:
2745     return isUInt<6>(Offset);
2746 
2747   case Hexagon::S4_storeirh_io:
2748   case Hexagon::S4_storeirht_io:
2749   case Hexagon::S4_storeirhf_io:
2750     return isShiftedUInt<6,1>(Offset);
2751 
2752   case Hexagon::S4_storeiri_io:
2753   case Hexagon::S4_storeirit_io:
2754   case Hexagon::S4_storeirif_io:
2755     return isShiftedUInt<6,2>(Offset);
2756   }
2757 
2758   if (Extend)
2759     return true;
2760 
2761   switch (Opcode) {
2762   case Hexagon::L2_loadri_io:
2763   case Hexagon::S2_storeri_io:
2764     return (Offset >= Hexagon_MEMW_OFFSET_MIN) &&
2765       (Offset <= Hexagon_MEMW_OFFSET_MAX);
2766 
2767   case Hexagon::L2_loadrd_io:
2768   case Hexagon::S2_storerd_io:
2769     return (Offset >= Hexagon_MEMD_OFFSET_MIN) &&
2770       (Offset <= Hexagon_MEMD_OFFSET_MAX);
2771 
2772   case Hexagon::L2_loadrh_io:
2773   case Hexagon::L2_loadruh_io:
2774   case Hexagon::S2_storerh_io:
2775   case Hexagon::S2_storerf_io:
2776     return (Offset >= Hexagon_MEMH_OFFSET_MIN) &&
2777       (Offset <= Hexagon_MEMH_OFFSET_MAX);
2778 
2779   case Hexagon::L2_loadrb_io:
2780   case Hexagon::L2_loadrub_io:
2781   case Hexagon::S2_storerb_io:
2782     return (Offset >= Hexagon_MEMB_OFFSET_MIN) &&
2783       (Offset <= Hexagon_MEMB_OFFSET_MAX);
2784 
2785   case Hexagon::A2_addi:
2786     return (Offset >= Hexagon_ADDI_OFFSET_MIN) &&
2787       (Offset <= Hexagon_ADDI_OFFSET_MAX);
2788 
2789   case Hexagon::L4_iadd_memopw_io:
2790   case Hexagon::L4_isub_memopw_io:
2791   case Hexagon::L4_add_memopw_io:
2792   case Hexagon::L4_sub_memopw_io:
2793   case Hexagon::L4_and_memopw_io:
2794   case Hexagon::L4_or_memopw_io:
2795     return (0 <= Offset && Offset <= 255);
2796 
2797   case Hexagon::L4_iadd_memoph_io:
2798   case Hexagon::L4_isub_memoph_io:
2799   case Hexagon::L4_add_memoph_io:
2800   case Hexagon::L4_sub_memoph_io:
2801   case Hexagon::L4_and_memoph_io:
2802   case Hexagon::L4_or_memoph_io:
2803     return (0 <= Offset && Offset <= 127);
2804 
2805   case Hexagon::L4_iadd_memopb_io:
2806   case Hexagon::L4_isub_memopb_io:
2807   case Hexagon::L4_add_memopb_io:
2808   case Hexagon::L4_sub_memopb_io:
2809   case Hexagon::L4_and_memopb_io:
2810   case Hexagon::L4_or_memopb_io:
2811     return (0 <= Offset && Offset <= 63);
2812 
2813   // LDriw_xxx and STriw_xxx are pseudo operations, so it has to take offset of
2814   // any size. Later pass knows how to handle it.
2815   case Hexagon::STriw_pred:
2816   case Hexagon::LDriw_pred:
2817   case Hexagon::STriw_ctr:
2818   case Hexagon::LDriw_ctr:
2819     return true;
2820 
2821   case Hexagon::PS_fi:
2822   case Hexagon::PS_fia:
2823   case Hexagon::INLINEASM:
2824     return true;
2825 
2826   case Hexagon::L2_ploadrbt_io:
2827   case Hexagon::L2_ploadrbf_io:
2828   case Hexagon::L2_ploadrubt_io:
2829   case Hexagon::L2_ploadrubf_io:
2830   case Hexagon::S2_pstorerbt_io:
2831   case Hexagon::S2_pstorerbf_io:
2832     return isUInt<6>(Offset);
2833 
2834   case Hexagon::L2_ploadrht_io:
2835   case Hexagon::L2_ploadrhf_io:
2836   case Hexagon::L2_ploadruht_io:
2837   case Hexagon::L2_ploadruhf_io:
2838   case Hexagon::S2_pstorerht_io:
2839   case Hexagon::S2_pstorerhf_io:
2840     return isShiftedUInt<6,1>(Offset);
2841 
2842   case Hexagon::L2_ploadrit_io:
2843   case Hexagon::L2_ploadrif_io:
2844   case Hexagon::S2_pstorerit_io:
2845   case Hexagon::S2_pstorerif_io:
2846     return isShiftedUInt<6,2>(Offset);
2847 
2848   case Hexagon::L2_ploadrdt_io:
2849   case Hexagon::L2_ploadrdf_io:
2850   case Hexagon::S2_pstorerdt_io:
2851   case Hexagon::S2_pstorerdf_io:
2852     return isShiftedUInt<6,3>(Offset);
2853   } // switch
2854 
2855   llvm_unreachable("No offset range is defined for this opcode. "
2856                    "Please define it in the above switch statement!");
2857 }
2858 
isVecAcc(const MachineInstr & MI) const2859 bool HexagonInstrInfo::isVecAcc(const MachineInstr &MI) const {
2860   return isHVXVec(MI) && isAccumulator(MI);
2861 }
2862 
isVecALU(const MachineInstr & MI) const2863 bool HexagonInstrInfo::isVecALU(const MachineInstr &MI) const {
2864   const uint64_t F = get(MI.getOpcode()).TSFlags;
2865   const uint64_t V = ((F >> HexagonII::TypePos) & HexagonII::TypeMask);
2866   return
2867     V == HexagonII::TypeCVI_VA         ||
2868     V == HexagonII::TypeCVI_VA_DV;
2869 }
2870 
isVecUsableNextPacket(const MachineInstr & ProdMI,const MachineInstr & ConsMI) const2871 bool HexagonInstrInfo::isVecUsableNextPacket(const MachineInstr &ProdMI,
2872       const MachineInstr &ConsMI) const {
2873   if (EnableACCForwarding && isVecAcc(ProdMI) && isVecAcc(ConsMI))
2874     return true;
2875 
2876   if (EnableALUForwarding && (isVecALU(ConsMI) || isLateSourceInstr(ConsMI)))
2877     return true;
2878 
2879   if (mayBeNewStore(ConsMI))
2880     return true;
2881 
2882   return false;
2883 }
2884 
isZeroExtendingLoad(const MachineInstr & MI) const2885 bool HexagonInstrInfo::isZeroExtendingLoad(const MachineInstr &MI) const {
2886   switch (MI.getOpcode()) {
2887   // Byte
2888   case Hexagon::L2_loadrub_io:
2889   case Hexagon::L4_loadrub_ur:
2890   case Hexagon::L4_loadrub_ap:
2891   case Hexagon::L2_loadrub_pr:
2892   case Hexagon::L2_loadrub_pbr:
2893   case Hexagon::L2_loadrub_pi:
2894   case Hexagon::L2_loadrub_pci:
2895   case Hexagon::L2_loadrub_pcr:
2896   case Hexagon::L2_loadbzw2_io:
2897   case Hexagon::L4_loadbzw2_ur:
2898   case Hexagon::L4_loadbzw2_ap:
2899   case Hexagon::L2_loadbzw2_pr:
2900   case Hexagon::L2_loadbzw2_pbr:
2901   case Hexagon::L2_loadbzw2_pi:
2902   case Hexagon::L2_loadbzw2_pci:
2903   case Hexagon::L2_loadbzw2_pcr:
2904   case Hexagon::L2_loadbzw4_io:
2905   case Hexagon::L4_loadbzw4_ur:
2906   case Hexagon::L4_loadbzw4_ap:
2907   case Hexagon::L2_loadbzw4_pr:
2908   case Hexagon::L2_loadbzw4_pbr:
2909   case Hexagon::L2_loadbzw4_pi:
2910   case Hexagon::L2_loadbzw4_pci:
2911   case Hexagon::L2_loadbzw4_pcr:
2912   case Hexagon::L4_loadrub_rr:
2913   case Hexagon::L2_ploadrubt_io:
2914   case Hexagon::L2_ploadrubt_pi:
2915   case Hexagon::L2_ploadrubf_io:
2916   case Hexagon::L2_ploadrubf_pi:
2917   case Hexagon::L2_ploadrubtnew_io:
2918   case Hexagon::L2_ploadrubfnew_io:
2919   case Hexagon::L4_ploadrubt_rr:
2920   case Hexagon::L4_ploadrubf_rr:
2921   case Hexagon::L4_ploadrubtnew_rr:
2922   case Hexagon::L4_ploadrubfnew_rr:
2923   case Hexagon::L2_ploadrubtnew_pi:
2924   case Hexagon::L2_ploadrubfnew_pi:
2925   case Hexagon::L4_ploadrubt_abs:
2926   case Hexagon::L4_ploadrubf_abs:
2927   case Hexagon::L4_ploadrubtnew_abs:
2928   case Hexagon::L4_ploadrubfnew_abs:
2929   case Hexagon::L2_loadrubgp:
2930   // Half
2931   case Hexagon::L2_loadruh_io:
2932   case Hexagon::L4_loadruh_ur:
2933   case Hexagon::L4_loadruh_ap:
2934   case Hexagon::L2_loadruh_pr:
2935   case Hexagon::L2_loadruh_pbr:
2936   case Hexagon::L2_loadruh_pi:
2937   case Hexagon::L2_loadruh_pci:
2938   case Hexagon::L2_loadruh_pcr:
2939   case Hexagon::L4_loadruh_rr:
2940   case Hexagon::L2_ploadruht_io:
2941   case Hexagon::L2_ploadruht_pi:
2942   case Hexagon::L2_ploadruhf_io:
2943   case Hexagon::L2_ploadruhf_pi:
2944   case Hexagon::L2_ploadruhtnew_io:
2945   case Hexagon::L2_ploadruhfnew_io:
2946   case Hexagon::L4_ploadruht_rr:
2947   case Hexagon::L4_ploadruhf_rr:
2948   case Hexagon::L4_ploadruhtnew_rr:
2949   case Hexagon::L4_ploadruhfnew_rr:
2950   case Hexagon::L2_ploadruhtnew_pi:
2951   case Hexagon::L2_ploadruhfnew_pi:
2952   case Hexagon::L4_ploadruht_abs:
2953   case Hexagon::L4_ploadruhf_abs:
2954   case Hexagon::L4_ploadruhtnew_abs:
2955   case Hexagon::L4_ploadruhfnew_abs:
2956   case Hexagon::L2_loadruhgp:
2957     return true;
2958   default:
2959     return false;
2960   }
2961 }
2962 
2963 // Add latency to instruction.
addLatencyToSchedule(const MachineInstr & MI1,const MachineInstr & MI2) const2964 bool HexagonInstrInfo::addLatencyToSchedule(const MachineInstr &MI1,
2965       const MachineInstr &MI2) const {
2966   if (isHVXVec(MI1) && isHVXVec(MI2))
2967     if (!isVecUsableNextPacket(MI1, MI2))
2968       return true;
2969   return false;
2970 }
2971 
2972 /// Get the base register and byte offset of a load/store instr.
getMemOperandsWithOffsetWidth(const MachineInstr & LdSt,SmallVectorImpl<const MachineOperand * > & BaseOps,int64_t & Offset,bool & OffsetIsScalable,unsigned & Width,const TargetRegisterInfo * TRI) const2973 bool HexagonInstrInfo::getMemOperandsWithOffsetWidth(
2974     const MachineInstr &LdSt, SmallVectorImpl<const MachineOperand *> &BaseOps,
2975     int64_t &Offset, bool &OffsetIsScalable, unsigned &Width,
2976     const TargetRegisterInfo *TRI) const {
2977   OffsetIsScalable = false;
2978   const MachineOperand *BaseOp = getBaseAndOffset(LdSt, Offset, Width);
2979   if (!BaseOp || !BaseOp->isReg())
2980     return false;
2981   BaseOps.push_back(BaseOp);
2982   return true;
2983 }
2984 
2985 /// Can these instructions execute at the same time in a bundle.
canExecuteInBundle(const MachineInstr & First,const MachineInstr & Second) const2986 bool HexagonInstrInfo::canExecuteInBundle(const MachineInstr &First,
2987       const MachineInstr &Second) const {
2988   if (Second.mayStore() && First.getOpcode() == Hexagon::S2_allocframe) {
2989     const MachineOperand &Op = Second.getOperand(0);
2990     if (Op.isReg() && Op.isUse() && Op.getReg() == Hexagon::R29)
2991       return true;
2992   }
2993   if (DisableNVSchedule)
2994     return false;
2995   if (mayBeNewStore(Second)) {
2996     // Make sure the definition of the first instruction is the value being
2997     // stored.
2998     const MachineOperand &Stored =
2999       Second.getOperand(Second.getNumOperands() - 1);
3000     if (!Stored.isReg())
3001       return false;
3002     for (unsigned i = 0, e = First.getNumOperands(); i < e; ++i) {
3003       const MachineOperand &Op = First.getOperand(i);
3004       if (Op.isReg() && Op.isDef() && Op.getReg() == Stored.getReg())
3005         return true;
3006     }
3007   }
3008   return false;
3009 }
3010 
doesNotReturn(const MachineInstr & CallMI) const3011 bool HexagonInstrInfo::doesNotReturn(const MachineInstr &CallMI) const {
3012   unsigned Opc = CallMI.getOpcode();
3013   return Opc == Hexagon::PS_call_nr || Opc == Hexagon::PS_callr_nr;
3014 }
3015 
hasEHLabel(const MachineBasicBlock * B) const3016 bool HexagonInstrInfo::hasEHLabel(const MachineBasicBlock *B) const {
3017   for (auto &I : *B)
3018     if (I.isEHLabel())
3019       return true;
3020   return false;
3021 }
3022 
3023 // Returns true if an instruction can be converted into a non-extended
3024 // equivalent instruction.
hasNonExtEquivalent(const MachineInstr & MI) const3025 bool HexagonInstrInfo::hasNonExtEquivalent(const MachineInstr &MI) const {
3026   short NonExtOpcode;
3027   // Check if the instruction has a register form that uses register in place
3028   // of the extended operand, if so return that as the non-extended form.
3029   if (Hexagon::getRegForm(MI.getOpcode()) >= 0)
3030     return true;
3031 
3032   if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) {
3033     // Check addressing mode and retrieve non-ext equivalent instruction.
3034 
3035     switch (getAddrMode(MI)) {
3036     case HexagonII::Absolute:
3037       // Load/store with absolute addressing mode can be converted into
3038       // base+offset mode.
3039       NonExtOpcode = Hexagon::changeAddrMode_abs_io(MI.getOpcode());
3040       break;
3041     case HexagonII::BaseImmOffset:
3042       // Load/store with base+offset addressing mode can be converted into
3043       // base+register offset addressing mode. However left shift operand should
3044       // be set to 0.
3045       NonExtOpcode = Hexagon::changeAddrMode_io_rr(MI.getOpcode());
3046       break;
3047     case HexagonII::BaseLongOffset:
3048       NonExtOpcode = Hexagon::changeAddrMode_ur_rr(MI.getOpcode());
3049       break;
3050     default:
3051       return false;
3052     }
3053     if (NonExtOpcode < 0)
3054       return false;
3055     return true;
3056   }
3057   return false;
3058 }
3059 
hasPseudoInstrPair(const MachineInstr & MI) const3060 bool HexagonInstrInfo::hasPseudoInstrPair(const MachineInstr &MI) const {
3061   return Hexagon::getRealHWInstr(MI.getOpcode(),
3062                                  Hexagon::InstrType_Pseudo) >= 0;
3063 }
3064 
hasUncondBranch(const MachineBasicBlock * B) const3065 bool HexagonInstrInfo::hasUncondBranch(const MachineBasicBlock *B)
3066       const {
3067   MachineBasicBlock::const_iterator I = B->getFirstTerminator(), E = B->end();
3068   while (I != E) {
3069     if (I->isBarrier())
3070       return true;
3071     ++I;
3072   }
3073   return false;
3074 }
3075 
3076 // Returns true, if a LD insn can be promoted to a cur load.
mayBeCurLoad(const MachineInstr & MI) const3077 bool HexagonInstrInfo::mayBeCurLoad(const MachineInstr &MI) const {
3078   const uint64_t F = MI.getDesc().TSFlags;
3079   return ((F >> HexagonII::mayCVLoadPos) & HexagonII::mayCVLoadMask) &&
3080          Subtarget.hasV60Ops();
3081 }
3082 
3083 // Returns true, if a ST insn can be promoted to a new-value store.
mayBeNewStore(const MachineInstr & MI) const3084 bool HexagonInstrInfo::mayBeNewStore(const MachineInstr &MI) const {
3085   if (MI.mayStore() && !Subtarget.useNewValueStores())
3086     return false;
3087 
3088   const uint64_t F = MI.getDesc().TSFlags;
3089   return (F >> HexagonII::mayNVStorePos) & HexagonII::mayNVStoreMask;
3090 }
3091 
producesStall(const MachineInstr & ProdMI,const MachineInstr & ConsMI) const3092 bool HexagonInstrInfo::producesStall(const MachineInstr &ProdMI,
3093       const MachineInstr &ConsMI) const {
3094   // There is no stall when ProdMI is not a V60 vector.
3095   if (!isHVXVec(ProdMI))
3096     return false;
3097 
3098   // There is no stall when ProdMI and ConsMI are not dependent.
3099   if (!isDependent(ProdMI, ConsMI))
3100     return false;
3101 
3102   // When Forward Scheduling is enabled, there is no stall if ProdMI and ConsMI
3103   // are scheduled in consecutive packets.
3104   if (isVecUsableNextPacket(ProdMI, ConsMI))
3105     return false;
3106 
3107   return true;
3108 }
3109 
producesStall(const MachineInstr & MI,MachineBasicBlock::const_instr_iterator BII) const3110 bool HexagonInstrInfo::producesStall(const MachineInstr &MI,
3111       MachineBasicBlock::const_instr_iterator BII) const {
3112   // There is no stall when I is not a V60 vector.
3113   if (!isHVXVec(MI))
3114     return false;
3115 
3116   MachineBasicBlock::const_instr_iterator MII = BII;
3117   MachineBasicBlock::const_instr_iterator MIE = MII->getParent()->instr_end();
3118 
3119   if (!MII->isBundle())
3120     return producesStall(*MII, MI);
3121 
3122   for (++MII; MII != MIE && MII->isInsideBundle(); ++MII) {
3123     const MachineInstr &J = *MII;
3124     if (producesStall(J, MI))
3125       return true;
3126   }
3127   return false;
3128 }
3129 
predCanBeUsedAsDotNew(const MachineInstr & MI,unsigned PredReg) const3130 bool HexagonInstrInfo::predCanBeUsedAsDotNew(const MachineInstr &MI,
3131       unsigned PredReg) const {
3132   for (const MachineOperand &MO : MI.operands()) {
3133     // Predicate register must be explicitly defined.
3134     if (MO.isRegMask() && MO.clobbersPhysReg(PredReg))
3135       return false;
3136     if (MO.isReg() && MO.isDef() && MO.isImplicit() && (MO.getReg() == PredReg))
3137       return false;
3138   }
3139 
3140   // Instruction that produce late predicate cannot be used as sources of
3141   // dot-new.
3142   switch (MI.getOpcode()) {
3143     case Hexagon::A4_addp_c:
3144     case Hexagon::A4_subp_c:
3145     case Hexagon::A4_tlbmatch:
3146     case Hexagon::A5_ACS:
3147     case Hexagon::F2_sfinvsqrta:
3148     case Hexagon::F2_sfrecipa:
3149     case Hexagon::J2_endloop0:
3150     case Hexagon::J2_endloop01:
3151     case Hexagon::J2_ploop1si:
3152     case Hexagon::J2_ploop1sr:
3153     case Hexagon::J2_ploop2si:
3154     case Hexagon::J2_ploop2sr:
3155     case Hexagon::J2_ploop3si:
3156     case Hexagon::J2_ploop3sr:
3157     case Hexagon::S2_cabacdecbin:
3158     case Hexagon::S2_storew_locked:
3159     case Hexagon::S4_stored_locked:
3160       return false;
3161   }
3162   return true;
3163 }
3164 
PredOpcodeHasJMP_c(unsigned Opcode) const3165 bool HexagonInstrInfo::PredOpcodeHasJMP_c(unsigned Opcode) const {
3166   return Opcode == Hexagon::J2_jumpt      ||
3167          Opcode == Hexagon::J2_jumptpt    ||
3168          Opcode == Hexagon::J2_jumpf      ||
3169          Opcode == Hexagon::J2_jumpfpt    ||
3170          Opcode == Hexagon::J2_jumptnew   ||
3171          Opcode == Hexagon::J2_jumpfnew   ||
3172          Opcode == Hexagon::J2_jumptnewpt ||
3173          Opcode == Hexagon::J2_jumpfnewpt;
3174 }
3175 
predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const3176 bool HexagonInstrInfo::predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const {
3177   if (Cond.empty() || !isPredicated(Cond[0].getImm()))
3178     return false;
3179   return !isPredicatedTrue(Cond[0].getImm());
3180 }
3181 
getAddrMode(const MachineInstr & MI) const3182 unsigned HexagonInstrInfo::getAddrMode(const MachineInstr &MI) const {
3183   const uint64_t F = MI.getDesc().TSFlags;
3184   return (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask;
3185 }
3186 
3187 // Returns the base register in a memory access (load/store). The offset is
3188 // returned in Offset and the access size is returned in AccessSize.
3189 // If the base operand has a subregister or the offset field does not contain
3190 // an immediate value, return nullptr.
getBaseAndOffset(const MachineInstr & MI,int64_t & Offset,unsigned & AccessSize) const3191 MachineOperand *HexagonInstrInfo::getBaseAndOffset(const MachineInstr &MI,
3192                                                    int64_t &Offset,
3193                                                    unsigned &AccessSize) const {
3194   // Return if it is not a base+offset type instruction or a MemOp.
3195   if (getAddrMode(MI) != HexagonII::BaseImmOffset &&
3196       getAddrMode(MI) != HexagonII::BaseLongOffset &&
3197       !isMemOp(MI) && !isPostIncrement(MI))
3198     return nullptr;
3199 
3200   AccessSize = getMemAccessSize(MI);
3201 
3202   unsigned BasePos = 0, OffsetPos = 0;
3203   if (!getBaseAndOffsetPosition(MI, BasePos, OffsetPos))
3204     return nullptr;
3205 
3206   // Post increment updates its EA after the mem access,
3207   // so we need to treat its offset as zero.
3208   if (isPostIncrement(MI)) {
3209     Offset = 0;
3210   } else {
3211     const MachineOperand &OffsetOp = MI.getOperand(OffsetPos);
3212     if (!OffsetOp.isImm())
3213       return nullptr;
3214     Offset = OffsetOp.getImm();
3215   }
3216 
3217   const MachineOperand &BaseOp = MI.getOperand(BasePos);
3218   if (BaseOp.getSubReg() != 0)
3219     return nullptr;
3220   return &const_cast<MachineOperand&>(BaseOp);
3221 }
3222 
3223 /// Return the position of the base and offset operands for this instruction.
getBaseAndOffsetPosition(const MachineInstr & MI,unsigned & BasePos,unsigned & OffsetPos) const3224 bool HexagonInstrInfo::getBaseAndOffsetPosition(const MachineInstr &MI,
3225       unsigned &BasePos, unsigned &OffsetPos) const {
3226   if (!isAddrModeWithOffset(MI) && !isPostIncrement(MI))
3227     return false;
3228 
3229   // Deal with memops first.
3230   if (isMemOp(MI)) {
3231     BasePos = 0;
3232     OffsetPos = 1;
3233   } else if (MI.mayStore()) {
3234     BasePos = 0;
3235     OffsetPos = 1;
3236   } else if (MI.mayLoad()) {
3237     BasePos = 1;
3238     OffsetPos = 2;
3239   } else
3240     return false;
3241 
3242   if (isPredicated(MI)) {
3243     BasePos++;
3244     OffsetPos++;
3245   }
3246   if (isPostIncrement(MI)) {
3247     BasePos++;
3248     OffsetPos++;
3249   }
3250 
3251   if (!MI.getOperand(BasePos).isReg() || !MI.getOperand(OffsetPos).isImm())
3252     return false;
3253 
3254   return true;
3255 }
3256 
3257 // Inserts branching instructions in reverse order of their occurrence.
3258 // e.g. jump_t t1 (i1)
3259 // jump t2        (i2)
3260 // Jumpers = {i2, i1}
getBranchingInstrs(MachineBasicBlock & MBB) const3261 SmallVector<MachineInstr*, 2> HexagonInstrInfo::getBranchingInstrs(
3262       MachineBasicBlock& MBB) const {
3263   SmallVector<MachineInstr*, 2> Jumpers;
3264   // If the block has no terminators, it just falls into the block after it.
3265   MachineBasicBlock::instr_iterator I = MBB.instr_end();
3266   if (I == MBB.instr_begin())
3267     return Jumpers;
3268 
3269   // A basic block may looks like this:
3270   //
3271   //  [   insn
3272   //     EH_LABEL
3273   //      insn
3274   //      insn
3275   //      insn
3276   //     EH_LABEL
3277   //      insn     ]
3278   //
3279   // It has two succs but does not have a terminator
3280   // Don't know how to handle it.
3281   do {
3282     --I;
3283     if (I->isEHLabel())
3284       return Jumpers;
3285   } while (I != MBB.instr_begin());
3286 
3287   I = MBB.instr_end();
3288   --I;
3289 
3290   while (I->isDebugInstr()) {
3291     if (I == MBB.instr_begin())
3292       return Jumpers;
3293     --I;
3294   }
3295   if (!isUnpredicatedTerminator(*I))
3296     return Jumpers;
3297 
3298   // Get the last instruction in the block.
3299   MachineInstr *LastInst = &*I;
3300   Jumpers.push_back(LastInst);
3301   MachineInstr *SecondLastInst = nullptr;
3302   // Find one more terminator if present.
3303   do {
3304     if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) {
3305       if (!SecondLastInst) {
3306         SecondLastInst = &*I;
3307         Jumpers.push_back(SecondLastInst);
3308       } else // This is a third branch.
3309         return Jumpers;
3310     }
3311     if (I == MBB.instr_begin())
3312       break;
3313     --I;
3314   } while (true);
3315   return Jumpers;
3316 }
3317 
3318 // Returns Operand Index for the constant extended instruction.
getCExtOpNum(const MachineInstr & MI) const3319 unsigned HexagonInstrInfo::getCExtOpNum(const MachineInstr &MI) const {
3320   const uint64_t F = MI.getDesc().TSFlags;
3321   return (F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask;
3322 }
3323 
3324 // See if instruction could potentially be a duplex candidate.
3325 // If so, return its group. Zero otherwise.
getCompoundCandidateGroup(const MachineInstr & MI) const3326 HexagonII::CompoundGroup HexagonInstrInfo::getCompoundCandidateGroup(
3327       const MachineInstr &MI) const {
3328   unsigned DstReg, SrcReg, Src1Reg, Src2Reg;
3329 
3330   switch (MI.getOpcode()) {
3331   default:
3332     return HexagonII::HCG_None;
3333   //
3334   // Compound pairs.
3335   // "p0=cmp.eq(Rs16,Rt16); if (p0.new) jump:nt #r9:2"
3336   // "Rd16=#U6 ; jump #r9:2"
3337   // "Rd16=Rs16 ; jump #r9:2"
3338   //
3339   case Hexagon::C2_cmpeq:
3340   case Hexagon::C2_cmpgt:
3341   case Hexagon::C2_cmpgtu:
3342     DstReg = MI.getOperand(0).getReg();
3343     Src1Reg = MI.getOperand(1).getReg();
3344     Src2Reg = MI.getOperand(2).getReg();
3345     if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3346         (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3347         isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg))
3348       return HexagonII::HCG_A;
3349     break;
3350   case Hexagon::C2_cmpeqi:
3351   case Hexagon::C2_cmpgti:
3352   case Hexagon::C2_cmpgtui:
3353     // P0 = cmp.eq(Rs,#u2)
3354     DstReg = MI.getOperand(0).getReg();
3355     SrcReg = MI.getOperand(1).getReg();
3356     if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3357         (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3358         isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() &&
3359         ((isUInt<5>(MI.getOperand(2).getImm())) ||
3360          (MI.getOperand(2).getImm() == -1)))
3361       return HexagonII::HCG_A;
3362     break;
3363   case Hexagon::A2_tfr:
3364     // Rd = Rs
3365     DstReg = MI.getOperand(0).getReg();
3366     SrcReg = MI.getOperand(1).getReg();
3367     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
3368       return HexagonII::HCG_A;
3369     break;
3370   case Hexagon::A2_tfrsi:
3371     // Rd = #u6
3372     // Do not test for #u6 size since the const is getting extended
3373     // regardless and compound could be formed.
3374     DstReg = MI.getOperand(0).getReg();
3375     if (isIntRegForSubInst(DstReg))
3376       return HexagonII::HCG_A;
3377     break;
3378   case Hexagon::S2_tstbit_i:
3379     DstReg = MI.getOperand(0).getReg();
3380     Src1Reg = MI.getOperand(1).getReg();
3381     if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3382         (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3383         MI.getOperand(2).isImm() &&
3384         isIntRegForSubInst(Src1Reg) && (MI.getOperand(2).getImm() == 0))
3385       return HexagonII::HCG_A;
3386     break;
3387   // The fact that .new form is used pretty much guarantees
3388   // that predicate register will match. Nevertheless,
3389   // there could be some false positives without additional
3390   // checking.
3391   case Hexagon::J2_jumptnew:
3392   case Hexagon::J2_jumpfnew:
3393   case Hexagon::J2_jumptnewpt:
3394   case Hexagon::J2_jumpfnewpt:
3395     Src1Reg = MI.getOperand(0).getReg();
3396     if (Hexagon::PredRegsRegClass.contains(Src1Reg) &&
3397         (Hexagon::P0 == Src1Reg || Hexagon::P1 == Src1Reg))
3398       return HexagonII::HCG_B;
3399     break;
3400   // Transfer and jump:
3401   // Rd=#U6 ; jump #r9:2
3402   // Rd=Rs ; jump #r9:2
3403   // Do not test for jump range here.
3404   case Hexagon::J2_jump:
3405   case Hexagon::RESTORE_DEALLOC_RET_JMP_V4:
3406   case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
3407     return HexagonII::HCG_C;
3408   }
3409 
3410   return HexagonII::HCG_None;
3411 }
3412 
3413 // Returns -1 when there is no opcode found.
getCompoundOpcode(const MachineInstr & GA,const MachineInstr & GB) const3414 unsigned HexagonInstrInfo::getCompoundOpcode(const MachineInstr &GA,
3415       const MachineInstr &GB) const {
3416   assert(getCompoundCandidateGroup(GA) == HexagonII::HCG_A);
3417   assert(getCompoundCandidateGroup(GB) == HexagonII::HCG_B);
3418   if ((GA.getOpcode() != Hexagon::C2_cmpeqi) ||
3419       (GB.getOpcode() != Hexagon::J2_jumptnew))
3420     return -1u;
3421   Register DestReg = GA.getOperand(0).getReg();
3422   if (!GB.readsRegister(DestReg))
3423     return -1u;
3424   if (DestReg != Hexagon::P0 && DestReg != Hexagon::P1)
3425     return -1u;
3426   // The value compared against must be either u5 or -1.
3427   const MachineOperand &CmpOp = GA.getOperand(2);
3428   if (!CmpOp.isImm())
3429     return -1u;
3430   int V = CmpOp.getImm();
3431   if (V == -1)
3432     return DestReg == Hexagon::P0 ? Hexagon::J4_cmpeqn1_tp0_jump_nt
3433                                   : Hexagon::J4_cmpeqn1_tp1_jump_nt;
3434   if (!isUInt<5>(V))
3435     return -1u;
3436   return DestReg == Hexagon::P0 ? Hexagon::J4_cmpeqi_tp0_jump_nt
3437                                 : Hexagon::J4_cmpeqi_tp1_jump_nt;
3438 }
3439 
3440 // Returns -1 if there is no opcode found.
getDuplexOpcode(const MachineInstr & MI,bool ForBigCore) const3441 int HexagonInstrInfo::getDuplexOpcode(const MachineInstr &MI,
3442                                       bool ForBigCore) const {
3443   // Static table to switch the opcodes across Tiny Core and Big Core.
3444   // dup_ opcodes are Big core opcodes.
3445   // NOTE: There are special instructions that need to handled later.
3446   // L4_return* instructions, they will only occupy SLOT0 (on big core too).
3447   // PS_jmpret - This pseudo translates to J2_jumpr which occupies only SLOT2.
3448   // The compiler need to base the root instruction to L6_return_map_to_raw
3449   // which can go any slot.
3450   static const std::map<unsigned, unsigned> DupMap = {
3451       {Hexagon::A2_add, Hexagon::dup_A2_add},
3452       {Hexagon::A2_addi, Hexagon::dup_A2_addi},
3453       {Hexagon::A2_andir, Hexagon::dup_A2_andir},
3454       {Hexagon::A2_combineii, Hexagon::dup_A2_combineii},
3455       {Hexagon::A2_sxtb, Hexagon::dup_A2_sxtb},
3456       {Hexagon::A2_sxth, Hexagon::dup_A2_sxth},
3457       {Hexagon::A2_tfr, Hexagon::dup_A2_tfr},
3458       {Hexagon::A2_tfrsi, Hexagon::dup_A2_tfrsi},
3459       {Hexagon::A2_zxtb, Hexagon::dup_A2_zxtb},
3460       {Hexagon::A2_zxth, Hexagon::dup_A2_zxth},
3461       {Hexagon::A4_combineii, Hexagon::dup_A4_combineii},
3462       {Hexagon::A4_combineir, Hexagon::dup_A4_combineir},
3463       {Hexagon::A4_combineri, Hexagon::dup_A4_combineri},
3464       {Hexagon::C2_cmoveif, Hexagon::dup_C2_cmoveif},
3465       {Hexagon::C2_cmoveit, Hexagon::dup_C2_cmoveit},
3466       {Hexagon::C2_cmovenewif, Hexagon::dup_C2_cmovenewif},
3467       {Hexagon::C2_cmovenewit, Hexagon::dup_C2_cmovenewit},
3468       {Hexagon::C2_cmpeqi, Hexagon::dup_C2_cmpeqi},
3469       {Hexagon::L2_deallocframe, Hexagon::dup_L2_deallocframe},
3470       {Hexagon::L2_loadrb_io, Hexagon::dup_L2_loadrb_io},
3471       {Hexagon::L2_loadrd_io, Hexagon::dup_L2_loadrd_io},
3472       {Hexagon::L2_loadrh_io, Hexagon::dup_L2_loadrh_io},
3473       {Hexagon::L2_loadri_io, Hexagon::dup_L2_loadri_io},
3474       {Hexagon::L2_loadrub_io, Hexagon::dup_L2_loadrub_io},
3475       {Hexagon::L2_loadruh_io, Hexagon::dup_L2_loadruh_io},
3476       {Hexagon::S2_allocframe, Hexagon::dup_S2_allocframe},
3477       {Hexagon::S2_storerb_io, Hexagon::dup_S2_storerb_io},
3478       {Hexagon::S2_storerd_io, Hexagon::dup_S2_storerd_io},
3479       {Hexagon::S2_storerh_io, Hexagon::dup_S2_storerh_io},
3480       {Hexagon::S2_storeri_io, Hexagon::dup_S2_storeri_io},
3481       {Hexagon::S4_storeirb_io, Hexagon::dup_S4_storeirb_io},
3482       {Hexagon::S4_storeiri_io, Hexagon::dup_S4_storeiri_io},
3483   };
3484   unsigned OpNum = MI.getOpcode();
3485   // Conversion to Big core.
3486   if (ForBigCore) {
3487     auto Iter = DupMap.find(OpNum);
3488     if (Iter != DupMap.end())
3489       return Iter->second;
3490   } else { // Conversion to Tiny core.
3491     for (auto Iter = DupMap.begin(), End = DupMap.end(); Iter != End; ++Iter)
3492       if (Iter->second == OpNum)
3493         return Iter->first;
3494   }
3495   return -1;
3496 }
3497 
getCondOpcode(int Opc,bool invertPredicate) const3498 int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const {
3499   enum Hexagon::PredSense inPredSense;
3500   inPredSense = invertPredicate ? Hexagon::PredSense_false :
3501                                   Hexagon::PredSense_true;
3502   int CondOpcode = Hexagon::getPredOpcode(Opc, inPredSense);
3503   if (CondOpcode >= 0) // Valid Conditional opcode/instruction
3504     return CondOpcode;
3505 
3506   llvm_unreachable("Unexpected predicable instruction");
3507 }
3508 
3509 // Return the cur value instruction for a given store.
getDotCurOp(const MachineInstr & MI) const3510 int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const {
3511   switch (MI.getOpcode()) {
3512   default: llvm_unreachable("Unknown .cur type");
3513   case Hexagon::V6_vL32b_pi:
3514     return Hexagon::V6_vL32b_cur_pi;
3515   case Hexagon::V6_vL32b_ai:
3516     return Hexagon::V6_vL32b_cur_ai;
3517   case Hexagon::V6_vL32b_nt_pi:
3518     return Hexagon::V6_vL32b_nt_cur_pi;
3519   case Hexagon::V6_vL32b_nt_ai:
3520     return Hexagon::V6_vL32b_nt_cur_ai;
3521   }
3522   return 0;
3523 }
3524 
3525 // Return the regular version of the .cur instruction.
getNonDotCurOp(const MachineInstr & MI) const3526 int HexagonInstrInfo::getNonDotCurOp(const MachineInstr &MI) const {
3527   switch (MI.getOpcode()) {
3528   default: llvm_unreachable("Unknown .cur type");
3529   case Hexagon::V6_vL32b_cur_pi:
3530     return Hexagon::V6_vL32b_pi;
3531   case Hexagon::V6_vL32b_cur_ai:
3532     return Hexagon::V6_vL32b_ai;
3533   case Hexagon::V6_vL32b_nt_cur_pi:
3534     return Hexagon::V6_vL32b_nt_pi;
3535   case Hexagon::V6_vL32b_nt_cur_ai:
3536     return Hexagon::V6_vL32b_nt_ai;
3537   }
3538   return 0;
3539 }
3540 
3541 // The diagram below shows the steps involved in the conversion of a predicated
3542 // store instruction to its .new predicated new-value form.
3543 //
3544 // Note: It doesn't include conditional new-value stores as they can't be
3545 // converted to .new predicate.
3546 //
3547 //               p.new NV store [ if(p0.new)memw(R0+#0)=R2.new ]
3548 //                ^           ^
3549 //               /             \ (not OK. it will cause new-value store to be
3550 //              /               X conditional on p0.new while R2 producer is
3551 //             /                 \ on p0)
3552 //            /                   \.
3553 //     p.new store                 p.old NV store
3554 // [if(p0.new)memw(R0+#0)=R2]    [if(p0)memw(R0+#0)=R2.new]
3555 //            ^                  ^
3556 //             \                /
3557 //              \              /
3558 //               \            /
3559 //                 p.old store
3560 //             [if (p0)memw(R0+#0)=R2]
3561 //
3562 // The following set of instructions further explains the scenario where
3563 // conditional new-value store becomes invalid when promoted to .new predicate
3564 // form.
3565 //
3566 // { 1) if (p0) r0 = add(r1, r2)
3567 //   2) p0 = cmp.eq(r3, #0) }
3568 //
3569 //   3) if (p0) memb(r1+#0) = r0  --> this instruction can't be grouped with
3570 // the first two instructions because in instr 1, r0 is conditional on old value
3571 // of p0 but its use in instr 3 is conditional on p0 modified by instr 2 which
3572 // is not valid for new-value stores.
3573 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded
3574 // from the "Conditional Store" list. Because a predicated new value store
3575 // would NOT be promoted to a double dot new store. See diagram below:
3576 // This function returns yes for those stores that are predicated but not
3577 // yet promoted to predicate dot new instructions.
3578 //
3579 //                          +---------------------+
3580 //                    /-----| if (p0) memw(..)=r0 |---------\~
3581 //                   ||     +---------------------+         ||
3582 //          promote  ||       /\       /\                   ||  promote
3583 //                   ||      /||\     /||\                  ||
3584 //                  \||/    demote     ||                  \||/
3585 //                   \/       ||       ||                   \/
3586 //       +-------------------------+   ||   +-------------------------+
3587 //       | if (p0.new) memw(..)=r0 |   ||   | if (p0) memw(..)=r0.new |
3588 //       +-------------------------+   ||   +-------------------------+
3589 //                        ||           ||         ||
3590 //                        ||         demote      \||/
3591 //                      promote        ||         \/ NOT possible
3592 //                        ||           ||         /\~
3593 //                       \||/          ||        /||\~
3594 //                        \/           ||         ||
3595 //                      +-----------------------------+
3596 //                      | if (p0.new) memw(..)=r0.new |
3597 //                      +-----------------------------+
3598 //                           Double Dot New Store
3599 //
3600 // Returns the most basic instruction for the .new predicated instructions and
3601 // new-value stores.
3602 // For example, all of the following instructions will be converted back to the
3603 // same instruction:
3604 // 1) if (p0.new) memw(R0+#0) = R1.new  --->
3605 // 2) if (p0) memw(R0+#0)= R1.new      -------> if (p0) memw(R0+#0) = R1
3606 // 3) if (p0.new) memw(R0+#0) = R1      --->
3607 //
3608 // To understand the translation of instruction 1 to its original form, consider
3609 // a packet with 3 instructions.
3610 // { p0 = cmp.eq(R0,R1)
3611 //   if (p0.new) R2 = add(R3, R4)
3612 //   R5 = add (R3, R1)
3613 // }
3614 // if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet
3615 //
3616 // This instruction can be part of the previous packet only if both p0 and R2
3617 // are promoted to .new values. This promotion happens in steps, first
3618 // predicate register is promoted to .new and in the next iteration R2 is
3619 // promoted. Therefore, in case of dependence check failure (due to R5) during
3620 // next iteration, it should be converted back to its most basic form.
3621 
3622 // Return the new value instruction for a given store.
getDotNewOp(const MachineInstr & MI) const3623 int HexagonInstrInfo::getDotNewOp(const MachineInstr &MI) const {
3624   int NVOpcode = Hexagon::getNewValueOpcode(MI.getOpcode());
3625   if (NVOpcode >= 0) // Valid new-value store instruction.
3626     return NVOpcode;
3627 
3628   switch (MI.getOpcode()) {
3629   default:
3630     report_fatal_error(std::string("Unknown .new type: ") +
3631       std::to_string(MI.getOpcode()));
3632   case Hexagon::S4_storerb_ur:
3633     return Hexagon::S4_storerbnew_ur;
3634 
3635   case Hexagon::S2_storerb_pci:
3636     return Hexagon::S2_storerb_pci;
3637 
3638   case Hexagon::S2_storeri_pci:
3639     return Hexagon::S2_storeri_pci;
3640 
3641   case Hexagon::S2_storerh_pci:
3642     return Hexagon::S2_storerh_pci;
3643 
3644   case Hexagon::S2_storerd_pci:
3645     return Hexagon::S2_storerd_pci;
3646 
3647   case Hexagon::S2_storerf_pci:
3648     return Hexagon::S2_storerf_pci;
3649 
3650   case Hexagon::V6_vS32b_ai:
3651     return Hexagon::V6_vS32b_new_ai;
3652 
3653   case Hexagon::V6_vS32b_pi:
3654     return Hexagon::V6_vS32b_new_pi;
3655   }
3656   return 0;
3657 }
3658 
3659 // Returns the opcode to use when converting MI, which is a conditional jump,
3660 // into a conditional instruction which uses the .new value of the predicate.
3661 // We also use branch probabilities to add a hint to the jump.
3662 // If MBPI is null, all edges will be treated as equally likely for the
3663 // purposes of establishing a predication hint.
getDotNewPredJumpOp(const MachineInstr & MI,const MachineBranchProbabilityInfo * MBPI) const3664 int HexagonInstrInfo::getDotNewPredJumpOp(const MachineInstr &MI,
3665       const MachineBranchProbabilityInfo *MBPI) const {
3666   // We assume that block can have at most two successors.
3667   const MachineBasicBlock *Src = MI.getParent();
3668   const MachineOperand &BrTarget = MI.getOperand(1);
3669   bool Taken = false;
3670   const BranchProbability OneHalf(1, 2);
3671 
3672   auto getEdgeProbability = [MBPI] (const MachineBasicBlock *Src,
3673                                     const MachineBasicBlock *Dst) {
3674     if (MBPI)
3675       return MBPI->getEdgeProbability(Src, Dst);
3676     return BranchProbability(1, Src->succ_size());
3677   };
3678 
3679   if (BrTarget.isMBB()) {
3680     const MachineBasicBlock *Dst = BrTarget.getMBB();
3681     Taken = getEdgeProbability(Src, Dst) >= OneHalf;
3682   } else {
3683     // The branch target is not a basic block (most likely a function).
3684     // Since BPI only gives probabilities for targets that are basic blocks,
3685     // try to identify another target of this branch (potentially a fall-
3686     // -through) and check the probability of that target.
3687     //
3688     // The only handled branch combinations are:
3689     // - one conditional branch,
3690     // - one conditional branch followed by one unconditional branch.
3691     // Otherwise, assume not-taken.
3692     assert(MI.isConditionalBranch());
3693     const MachineBasicBlock &B = *MI.getParent();
3694     bool SawCond = false, Bad = false;
3695     for (const MachineInstr &I : B) {
3696       if (!I.isBranch())
3697         continue;
3698       if (I.isConditionalBranch()) {
3699         SawCond = true;
3700         if (&I != &MI) {
3701           Bad = true;
3702           break;
3703         }
3704       }
3705       if (I.isUnconditionalBranch() && !SawCond) {
3706         Bad = true;
3707         break;
3708       }
3709     }
3710     if (!Bad) {
3711       MachineBasicBlock::const_instr_iterator It(MI);
3712       MachineBasicBlock::const_instr_iterator NextIt = std::next(It);
3713       if (NextIt == B.instr_end()) {
3714         // If this branch is the last, look for the fall-through block.
3715         for (const MachineBasicBlock *SB : B.successors()) {
3716           if (!B.isLayoutSuccessor(SB))
3717             continue;
3718           Taken = getEdgeProbability(Src, SB) < OneHalf;
3719           break;
3720         }
3721       } else {
3722         assert(NextIt->isUnconditionalBranch());
3723         // Find the first MBB operand and assume it's the target.
3724         const MachineBasicBlock *BT = nullptr;
3725         for (const MachineOperand &Op : NextIt->operands()) {
3726           if (!Op.isMBB())
3727             continue;
3728           BT = Op.getMBB();
3729           break;
3730         }
3731         Taken = BT && getEdgeProbability(Src, BT) < OneHalf;
3732       }
3733     } // if (!Bad)
3734   }
3735 
3736   // The Taken flag should be set to something reasonable by this point.
3737 
3738   switch (MI.getOpcode()) {
3739   case Hexagon::J2_jumpt:
3740     return Taken ? Hexagon::J2_jumptnewpt : Hexagon::J2_jumptnew;
3741   case Hexagon::J2_jumpf:
3742     return Taken ? Hexagon::J2_jumpfnewpt : Hexagon::J2_jumpfnew;
3743 
3744   default:
3745     llvm_unreachable("Unexpected jump instruction.");
3746   }
3747 }
3748 
3749 // Return .new predicate version for an instruction.
getDotNewPredOp(const MachineInstr & MI,const MachineBranchProbabilityInfo * MBPI) const3750 int HexagonInstrInfo::getDotNewPredOp(const MachineInstr &MI,
3751       const MachineBranchProbabilityInfo *MBPI) const {
3752   switch (MI.getOpcode()) {
3753   // Condtional Jumps
3754   case Hexagon::J2_jumpt:
3755   case Hexagon::J2_jumpf:
3756     return getDotNewPredJumpOp(MI, MBPI);
3757   }
3758 
3759   int NewOpcode = Hexagon::getPredNewOpcode(MI.getOpcode());
3760   if (NewOpcode >= 0)
3761     return NewOpcode;
3762   return 0;
3763 }
3764 
getDotOldOp(const MachineInstr & MI) const3765 int HexagonInstrInfo::getDotOldOp(const MachineInstr &MI) const {
3766   int NewOp = MI.getOpcode();
3767   if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form
3768     NewOp = Hexagon::getPredOldOpcode(NewOp);
3769     // All Hexagon architectures have prediction bits on dot-new branches,
3770     // but only Hexagon V60+ has prediction bits on dot-old ones. Make sure
3771     // to pick the right opcode when converting back to dot-old.
3772     if (!Subtarget.getFeatureBits()[Hexagon::ArchV60]) {
3773       switch (NewOp) {
3774       case Hexagon::J2_jumptpt:
3775         NewOp = Hexagon::J2_jumpt;
3776         break;
3777       case Hexagon::J2_jumpfpt:
3778         NewOp = Hexagon::J2_jumpf;
3779         break;
3780       case Hexagon::J2_jumprtpt:
3781         NewOp = Hexagon::J2_jumprt;
3782         break;
3783       case Hexagon::J2_jumprfpt:
3784         NewOp = Hexagon::J2_jumprf;
3785         break;
3786       }
3787     }
3788     assert(NewOp >= 0 &&
3789            "Couldn't change predicate new instruction to its old form.");
3790   }
3791 
3792   if (isNewValueStore(NewOp)) { // Convert into non-new-value format
3793     NewOp = Hexagon::getNonNVStore(NewOp);
3794     assert(NewOp >= 0 && "Couldn't change new-value store to its old form.");
3795   }
3796 
3797   if (Subtarget.hasV60Ops())
3798     return NewOp;
3799 
3800   // Subtargets prior to V60 didn't support 'taken' forms of predicated jumps.
3801   switch (NewOp) {
3802   case Hexagon::J2_jumpfpt:
3803     return Hexagon::J2_jumpf;
3804   case Hexagon::J2_jumptpt:
3805     return Hexagon::J2_jumpt;
3806   case Hexagon::J2_jumprfpt:
3807     return Hexagon::J2_jumprf;
3808   case Hexagon::J2_jumprtpt:
3809     return Hexagon::J2_jumprt;
3810   }
3811   return NewOp;
3812 }
3813 
3814 // See if instruction could potentially be a duplex candidate.
3815 // If so, return its group. Zero otherwise.
getDuplexCandidateGroup(const MachineInstr & MI) const3816 HexagonII::SubInstructionGroup HexagonInstrInfo::getDuplexCandidateGroup(
3817       const MachineInstr &MI) const {
3818   unsigned DstReg, SrcReg, Src1Reg, Src2Reg;
3819   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
3820 
3821   switch (MI.getOpcode()) {
3822   default:
3823     return HexagonII::HSIG_None;
3824   //
3825   // Group L1:
3826   //
3827   // Rd = memw(Rs+#u4:2)
3828   // Rd = memub(Rs+#u4:0)
3829   case Hexagon::L2_loadri_io:
3830   case Hexagon::dup_L2_loadri_io:
3831     DstReg = MI.getOperand(0).getReg();
3832     SrcReg = MI.getOperand(1).getReg();
3833     // Special case this one from Group L2.
3834     // Rd = memw(r29+#u5:2)
3835     if (isIntRegForSubInst(DstReg)) {
3836       if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
3837           HRI.getStackRegister() == SrcReg &&
3838           MI.getOperand(2).isImm() &&
3839           isShiftedUInt<5,2>(MI.getOperand(2).getImm()))
3840         return HexagonII::HSIG_L2;
3841       // Rd = memw(Rs+#u4:2)
3842       if (isIntRegForSubInst(SrcReg) &&
3843           (MI.getOperand(2).isImm() &&
3844           isShiftedUInt<4,2>(MI.getOperand(2).getImm())))
3845         return HexagonII::HSIG_L1;
3846     }
3847     break;
3848   case Hexagon::L2_loadrub_io:
3849   case Hexagon::dup_L2_loadrub_io:
3850     // Rd = memub(Rs+#u4:0)
3851     DstReg = MI.getOperand(0).getReg();
3852     SrcReg = MI.getOperand(1).getReg();
3853     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3854         MI.getOperand(2).isImm() && isUInt<4>(MI.getOperand(2).getImm()))
3855       return HexagonII::HSIG_L1;
3856     break;
3857   //
3858   // Group L2:
3859   //
3860   // Rd = memh/memuh(Rs+#u3:1)
3861   // Rd = memb(Rs+#u3:0)
3862   // Rd = memw(r29+#u5:2) - Handled above.
3863   // Rdd = memd(r29+#u5:3)
3864   // deallocframe
3865   // [if ([!]p0[.new])] dealloc_return
3866   // [if ([!]p0[.new])] jumpr r31
3867   case Hexagon::L2_loadrh_io:
3868   case Hexagon::L2_loadruh_io:
3869   case Hexagon::dup_L2_loadrh_io:
3870   case Hexagon::dup_L2_loadruh_io:
3871     // Rd = memh/memuh(Rs+#u3:1)
3872     DstReg = MI.getOperand(0).getReg();
3873     SrcReg = MI.getOperand(1).getReg();
3874     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3875         MI.getOperand(2).isImm() &&
3876         isShiftedUInt<3,1>(MI.getOperand(2).getImm()))
3877       return HexagonII::HSIG_L2;
3878     break;
3879   case Hexagon::L2_loadrb_io:
3880   case Hexagon::dup_L2_loadrb_io:
3881     // Rd = memb(Rs+#u3:0)
3882     DstReg = MI.getOperand(0).getReg();
3883     SrcReg = MI.getOperand(1).getReg();
3884     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3885         MI.getOperand(2).isImm() &&
3886         isUInt<3>(MI.getOperand(2).getImm()))
3887       return HexagonII::HSIG_L2;
3888     break;
3889   case Hexagon::L2_loadrd_io:
3890   case Hexagon::dup_L2_loadrd_io:
3891     // Rdd = memd(r29+#u5:3)
3892     DstReg = MI.getOperand(0).getReg();
3893     SrcReg = MI.getOperand(1).getReg();
3894     if (isDblRegForSubInst(DstReg, HRI) &&
3895         Hexagon::IntRegsRegClass.contains(SrcReg) &&
3896         HRI.getStackRegister() == SrcReg &&
3897         MI.getOperand(2).isImm() &&
3898         isShiftedUInt<5,3>(MI.getOperand(2).getImm()))
3899       return HexagonII::HSIG_L2;
3900     break;
3901   // dealloc_return is not documented in Hexagon Manual, but marked
3902   // with A_SUBINSN attribute in iset_v4classic.py.
3903   case Hexagon::RESTORE_DEALLOC_RET_JMP_V4:
3904   case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
3905   case Hexagon::L4_return:
3906   case Hexagon::L2_deallocframe:
3907   case Hexagon::dup_L2_deallocframe:
3908     return HexagonII::HSIG_L2;
3909   case Hexagon::EH_RETURN_JMPR:
3910   case Hexagon::PS_jmpret:
3911   case Hexagon::SL2_jumpr31:
3912     // jumpr r31
3913     // Actual form JMPR implicit-def %pc, implicit %r31, implicit internal %r0
3914     DstReg = MI.getOperand(0).getReg();
3915     if (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg))
3916       return HexagonII::HSIG_L2;
3917     break;
3918   case Hexagon::PS_jmprett:
3919   case Hexagon::PS_jmpretf:
3920   case Hexagon::PS_jmprettnewpt:
3921   case Hexagon::PS_jmpretfnewpt:
3922   case Hexagon::PS_jmprettnew:
3923   case Hexagon::PS_jmpretfnew:
3924   case Hexagon::SL2_jumpr31_t:
3925   case Hexagon::SL2_jumpr31_f:
3926   case Hexagon::SL2_jumpr31_tnew:
3927   case Hexagon::SL2_jumpr31_fnew:
3928     DstReg = MI.getOperand(1).getReg();
3929     SrcReg = MI.getOperand(0).getReg();
3930     // [if ([!]p0[.new])] jumpr r31
3931     if ((Hexagon::PredRegsRegClass.contains(SrcReg) &&
3932         (Hexagon::P0 == SrcReg)) &&
3933         (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg)))
3934       return HexagonII::HSIG_L2;
3935     break;
3936   case Hexagon::L4_return_t:
3937   case Hexagon::L4_return_f:
3938   case Hexagon::L4_return_tnew_pnt:
3939   case Hexagon::L4_return_fnew_pnt:
3940   case Hexagon::L4_return_tnew_pt:
3941   case Hexagon::L4_return_fnew_pt:
3942     // [if ([!]p0[.new])] dealloc_return
3943     SrcReg = MI.getOperand(0).getReg();
3944     if (Hexagon::PredRegsRegClass.contains(SrcReg) && (Hexagon::P0 == SrcReg))
3945       return HexagonII::HSIG_L2;
3946     break;
3947   //
3948   // Group S1:
3949   //
3950   // memw(Rs+#u4:2) = Rt
3951   // memb(Rs+#u4:0) = Rt
3952   case Hexagon::S2_storeri_io:
3953   case Hexagon::dup_S2_storeri_io:
3954     // Special case this one from Group S2.
3955     // memw(r29+#u5:2) = Rt
3956     Src1Reg = MI.getOperand(0).getReg();
3957     Src2Reg = MI.getOperand(2).getReg();
3958     if (Hexagon::IntRegsRegClass.contains(Src1Reg) &&
3959         isIntRegForSubInst(Src2Reg) &&
3960         HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() &&
3961         isShiftedUInt<5,2>(MI.getOperand(1).getImm()))
3962       return HexagonII::HSIG_S2;
3963     // memw(Rs+#u4:2) = Rt
3964     if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3965         MI.getOperand(1).isImm() &&
3966         isShiftedUInt<4,2>(MI.getOperand(1).getImm()))
3967       return HexagonII::HSIG_S1;
3968     break;
3969   case Hexagon::S2_storerb_io:
3970   case Hexagon::dup_S2_storerb_io:
3971     // memb(Rs+#u4:0) = Rt
3972     Src1Reg = MI.getOperand(0).getReg();
3973     Src2Reg = MI.getOperand(2).getReg();
3974     if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3975         MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()))
3976       return HexagonII::HSIG_S1;
3977     break;
3978   //
3979   // Group S2:
3980   //
3981   // memh(Rs+#u3:1) = Rt
3982   // memw(r29+#u5:2) = Rt
3983   // memd(r29+#s6:3) = Rtt
3984   // memw(Rs+#u4:2) = #U1
3985   // memb(Rs+#u4) = #U1
3986   // allocframe(#u5:3)
3987   case Hexagon::S2_storerh_io:
3988   case Hexagon::dup_S2_storerh_io:
3989     // memh(Rs+#u3:1) = Rt
3990     Src1Reg = MI.getOperand(0).getReg();
3991     Src2Reg = MI.getOperand(2).getReg();
3992     if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3993         MI.getOperand(1).isImm() &&
3994         isShiftedUInt<3,1>(MI.getOperand(1).getImm()))
3995       return HexagonII::HSIG_S1;
3996     break;
3997   case Hexagon::S2_storerd_io:
3998   case Hexagon::dup_S2_storerd_io:
3999     // memd(r29+#s6:3) = Rtt
4000     Src1Reg = MI.getOperand(0).getReg();
4001     Src2Reg = MI.getOperand(2).getReg();
4002     if (isDblRegForSubInst(Src2Reg, HRI) &&
4003         Hexagon::IntRegsRegClass.contains(Src1Reg) &&
4004         HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() &&
4005         isShiftedInt<6,3>(MI.getOperand(1).getImm()))
4006       return HexagonII::HSIG_S2;
4007     break;
4008   case Hexagon::S4_storeiri_io:
4009   case Hexagon::dup_S4_storeiri_io:
4010     // memw(Rs+#u4:2) = #U1
4011     Src1Reg = MI.getOperand(0).getReg();
4012     if (isIntRegForSubInst(Src1Reg) && MI.getOperand(1).isImm() &&
4013         isShiftedUInt<4,2>(MI.getOperand(1).getImm()) &&
4014         MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm()))
4015       return HexagonII::HSIG_S2;
4016     break;
4017   case Hexagon::S4_storeirb_io:
4018   case Hexagon::dup_S4_storeirb_io:
4019     // memb(Rs+#u4) = #U1
4020     Src1Reg = MI.getOperand(0).getReg();
4021     if (isIntRegForSubInst(Src1Reg) &&
4022         MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()) &&
4023         MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm()))
4024       return HexagonII::HSIG_S2;
4025     break;
4026   case Hexagon::S2_allocframe:
4027   case Hexagon::dup_S2_allocframe:
4028     if (MI.getOperand(2).isImm() &&
4029         isShiftedUInt<5,3>(MI.getOperand(2).getImm()))
4030       return HexagonII::HSIG_S1;
4031     break;
4032   //
4033   // Group A:
4034   //
4035   // Rx = add(Rx,#s7)
4036   // Rd = Rs
4037   // Rd = #u6
4038   // Rd = #-1
4039   // if ([!]P0[.new]) Rd = #0
4040   // Rd = add(r29,#u6:2)
4041   // Rx = add(Rx,Rs)
4042   // P0 = cmp.eq(Rs,#u2)
4043   // Rdd = combine(#0,Rs)
4044   // Rdd = combine(Rs,#0)
4045   // Rdd = combine(#u2,#U2)
4046   // Rd = add(Rs,#1)
4047   // Rd = add(Rs,#-1)
4048   // Rd = sxth/sxtb/zxtb/zxth(Rs)
4049   // Rd = and(Rs,#1)
4050   case Hexagon::A2_addi:
4051   case Hexagon::dup_A2_addi:
4052     DstReg = MI.getOperand(0).getReg();
4053     SrcReg = MI.getOperand(1).getReg();
4054     if (isIntRegForSubInst(DstReg)) {
4055       // Rd = add(r29,#u6:2)
4056       if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
4057         HRI.getStackRegister() == SrcReg && MI.getOperand(2).isImm() &&
4058         isShiftedUInt<6,2>(MI.getOperand(2).getImm()))
4059         return HexagonII::HSIG_A;
4060       // Rx = add(Rx,#s7)
4061       if ((DstReg == SrcReg) && MI.getOperand(2).isImm() &&
4062           isInt<7>(MI.getOperand(2).getImm()))
4063         return HexagonII::HSIG_A;
4064       // Rd = add(Rs,#1)
4065       // Rd = add(Rs,#-1)
4066       if (isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() &&
4067           ((MI.getOperand(2).getImm() == 1) ||
4068           (MI.getOperand(2).getImm() == -1)))
4069         return HexagonII::HSIG_A;
4070     }
4071     break;
4072   case Hexagon::A2_add:
4073   case Hexagon::dup_A2_add:
4074     // Rx = add(Rx,Rs)
4075     DstReg = MI.getOperand(0).getReg();
4076     Src1Reg = MI.getOperand(1).getReg();
4077     Src2Reg = MI.getOperand(2).getReg();
4078     if (isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) &&
4079         isIntRegForSubInst(Src2Reg))
4080       return HexagonII::HSIG_A;
4081     break;
4082   case Hexagon::A2_andir:
4083   case Hexagon::dup_A2_andir:
4084     // Same as zxtb.
4085     // Rd16=and(Rs16,#255)
4086     // Rd16=and(Rs16,#1)
4087     DstReg = MI.getOperand(0).getReg();
4088     SrcReg = MI.getOperand(1).getReg();
4089     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
4090         MI.getOperand(2).isImm() &&
4091         ((MI.getOperand(2).getImm() == 1) ||
4092         (MI.getOperand(2).getImm() == 255)))
4093       return HexagonII::HSIG_A;
4094     break;
4095   case Hexagon::A2_tfr:
4096   case Hexagon::dup_A2_tfr:
4097     // Rd = Rs
4098     DstReg = MI.getOperand(0).getReg();
4099     SrcReg = MI.getOperand(1).getReg();
4100     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
4101       return HexagonII::HSIG_A;
4102     break;
4103   case Hexagon::A2_tfrsi:
4104   case Hexagon::dup_A2_tfrsi:
4105     // Rd = #u6
4106     // Do not test for #u6 size since the const is getting extended
4107     // regardless and compound could be formed.
4108     // Rd = #-1
4109     DstReg = MI.getOperand(0).getReg();
4110     if (isIntRegForSubInst(DstReg))
4111       return HexagonII::HSIG_A;
4112     break;
4113   case Hexagon::C2_cmoveit:
4114   case Hexagon::C2_cmovenewit:
4115   case Hexagon::C2_cmoveif:
4116   case Hexagon::C2_cmovenewif:
4117   case Hexagon::dup_C2_cmoveit:
4118   case Hexagon::dup_C2_cmovenewit:
4119   case Hexagon::dup_C2_cmoveif:
4120   case Hexagon::dup_C2_cmovenewif:
4121     // if ([!]P0[.new]) Rd = #0
4122     // Actual form:
4123     // %r16 = C2_cmovenewit internal %p0, 0, implicit undef %r16;
4124     DstReg = MI.getOperand(0).getReg();
4125     SrcReg = MI.getOperand(1).getReg();
4126     if (isIntRegForSubInst(DstReg) &&
4127         Hexagon::PredRegsRegClass.contains(SrcReg) && Hexagon::P0 == SrcReg &&
4128         MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0)
4129       return HexagonII::HSIG_A;
4130     break;
4131   case Hexagon::C2_cmpeqi:
4132   case Hexagon::dup_C2_cmpeqi:
4133     // P0 = cmp.eq(Rs,#u2)
4134     DstReg = MI.getOperand(0).getReg();
4135     SrcReg = MI.getOperand(1).getReg();
4136     if (Hexagon::PredRegsRegClass.contains(DstReg) &&
4137         Hexagon::P0 == DstReg && isIntRegForSubInst(SrcReg) &&
4138         MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm()))
4139       return HexagonII::HSIG_A;
4140     break;
4141   case Hexagon::A2_combineii:
4142   case Hexagon::A4_combineii:
4143   case Hexagon::dup_A2_combineii:
4144   case Hexagon::dup_A4_combineii:
4145     // Rdd = combine(#u2,#U2)
4146     DstReg = MI.getOperand(0).getReg();
4147     if (isDblRegForSubInst(DstReg, HRI) &&
4148         ((MI.getOperand(1).isImm() && isUInt<2>(MI.getOperand(1).getImm())) ||
4149         (MI.getOperand(1).isGlobal() &&
4150         isUInt<2>(MI.getOperand(1).getOffset()))) &&
4151         ((MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) ||
4152         (MI.getOperand(2).isGlobal() &&
4153         isUInt<2>(MI.getOperand(2).getOffset()))))
4154       return HexagonII::HSIG_A;
4155     break;
4156   case Hexagon::A4_combineri:
4157   case Hexagon::dup_A4_combineri:
4158     // Rdd = combine(Rs,#0)
4159     // Rdd = combine(Rs,#0)
4160     DstReg = MI.getOperand(0).getReg();
4161     SrcReg = MI.getOperand(1).getReg();
4162     if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) &&
4163         ((MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) ||
4164         (MI.getOperand(2).isGlobal() && MI.getOperand(2).getOffset() == 0)))
4165       return HexagonII::HSIG_A;
4166     break;
4167   case Hexagon::A4_combineir:
4168   case Hexagon::dup_A4_combineir:
4169     // Rdd = combine(#0,Rs)
4170     DstReg = MI.getOperand(0).getReg();
4171     SrcReg = MI.getOperand(2).getReg();
4172     if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) &&
4173         ((MI.getOperand(1).isImm() && MI.getOperand(1).getImm() == 0) ||
4174         (MI.getOperand(1).isGlobal() && MI.getOperand(1).getOffset() == 0)))
4175       return HexagonII::HSIG_A;
4176     break;
4177   case Hexagon::A2_sxtb:
4178   case Hexagon::A2_sxth:
4179   case Hexagon::A2_zxtb:
4180   case Hexagon::A2_zxth:
4181   case Hexagon::dup_A2_sxtb:
4182   case Hexagon::dup_A2_sxth:
4183   case Hexagon::dup_A2_zxtb:
4184   case Hexagon::dup_A2_zxth:
4185     // Rd = sxth/sxtb/zxtb/zxth(Rs)
4186     DstReg = MI.getOperand(0).getReg();
4187     SrcReg = MI.getOperand(1).getReg();
4188     if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
4189       return HexagonII::HSIG_A;
4190     break;
4191   }
4192 
4193   return HexagonII::HSIG_None;
4194 }
4195 
getEquivalentHWInstr(const MachineInstr & MI) const4196 short HexagonInstrInfo::getEquivalentHWInstr(const MachineInstr &MI) const {
4197   return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Real);
4198 }
4199 
getInstrTimingClassLatency(const InstrItineraryData * ItinData,const MachineInstr & MI) const4200 unsigned HexagonInstrInfo::getInstrTimingClassLatency(
4201       const InstrItineraryData *ItinData, const MachineInstr &MI) const {
4202   // Default to one cycle for no itinerary. However, an "empty" itinerary may
4203   // still have a MinLatency property, which getStageLatency checks.
4204   if (!ItinData)
4205     return getInstrLatency(ItinData, MI);
4206 
4207   if (MI.isTransient())
4208     return 0;
4209   return ItinData->getStageLatency(MI.getDesc().getSchedClass());
4210 }
4211 
4212 /// getOperandLatency - Compute and return the use operand latency of a given
4213 /// pair of def and use.
4214 /// In most cases, the static scheduling itinerary was enough to determine the
4215 /// operand latency. But it may not be possible for instructions with variable
4216 /// number of defs / uses.
4217 ///
4218 /// This is a raw interface to the itinerary that may be directly overriden by
4219 /// a target. Use computeOperandLatency to get the best estimate of latency.
getOperandLatency(const InstrItineraryData * ItinData,const MachineInstr & DefMI,unsigned DefIdx,const MachineInstr & UseMI,unsigned UseIdx) const4220 int HexagonInstrInfo::getOperandLatency(const InstrItineraryData *ItinData,
4221                                         const MachineInstr &DefMI,
4222                                         unsigned DefIdx,
4223                                         const MachineInstr &UseMI,
4224                                         unsigned UseIdx) const {
4225   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
4226 
4227   // Get DefIdx and UseIdx for super registers.
4228   const MachineOperand &DefMO = DefMI.getOperand(DefIdx);
4229 
4230   if (DefMO.isReg() && Register::isPhysicalRegister(DefMO.getReg())) {
4231     if (DefMO.isImplicit()) {
4232       for (MCSuperRegIterator SR(DefMO.getReg(), &HRI); SR.isValid(); ++SR) {
4233         int Idx = DefMI.findRegisterDefOperandIdx(*SR, false, false, &HRI);
4234         if (Idx != -1) {
4235           DefIdx = Idx;
4236           break;
4237         }
4238       }
4239     }
4240 
4241     const MachineOperand &UseMO = UseMI.getOperand(UseIdx);
4242     if (UseMO.isImplicit()) {
4243       for (MCSuperRegIterator SR(UseMO.getReg(), &HRI); SR.isValid(); ++SR) {
4244         int Idx = UseMI.findRegisterUseOperandIdx(*SR, false, &HRI);
4245         if (Idx != -1) {
4246           UseIdx = Idx;
4247           break;
4248         }
4249       }
4250     }
4251   }
4252 
4253   int Latency = TargetInstrInfo::getOperandLatency(ItinData, DefMI, DefIdx,
4254                                                    UseMI, UseIdx);
4255   if (!Latency)
4256     // We should never have 0 cycle latency between two instructions unless
4257     // they can be packetized together. However, this decision can't be made
4258     // here.
4259     Latency = 1;
4260   return Latency;
4261 }
4262 
4263 // inverts the predication logic.
4264 // p -> NotP
4265 // NotP -> P
getInvertedPredSense(SmallVectorImpl<MachineOperand> & Cond) const4266 bool HexagonInstrInfo::getInvertedPredSense(
4267       SmallVectorImpl<MachineOperand> &Cond) const {
4268   if (Cond.empty())
4269     return false;
4270   unsigned Opc = getInvertedPredicatedOpcode(Cond[0].getImm());
4271   Cond[0].setImm(Opc);
4272   return true;
4273 }
4274 
getInvertedPredicatedOpcode(const int Opc) const4275 unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const {
4276   int InvPredOpcode;
4277   InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc)
4278                                         : Hexagon::getTruePredOpcode(Opc);
4279   if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate.
4280     return InvPredOpcode;
4281 
4282   llvm_unreachable("Unexpected predicated instruction");
4283 }
4284 
4285 // Returns the max value that doesn't need to be extended.
getMaxValue(const MachineInstr & MI) const4286 int HexagonInstrInfo::getMaxValue(const MachineInstr &MI) const {
4287   const uint64_t F = MI.getDesc().TSFlags;
4288   unsigned isSigned = (F >> HexagonII::ExtentSignedPos)
4289                     & HexagonII::ExtentSignedMask;
4290   unsigned bits =  (F >> HexagonII::ExtentBitsPos)
4291                     & HexagonII::ExtentBitsMask;
4292 
4293   if (isSigned) // if value is signed
4294     return ~(-1U << (bits - 1));
4295   else
4296     return ~(-1U << bits);
4297 }
4298 
4299 
isAddrModeWithOffset(const MachineInstr & MI) const4300 bool HexagonInstrInfo::isAddrModeWithOffset(const MachineInstr &MI) const {
4301   switch (MI.getOpcode()) {
4302   case Hexagon::L2_loadrbgp:
4303   case Hexagon::L2_loadrdgp:
4304   case Hexagon::L2_loadrhgp:
4305   case Hexagon::L2_loadrigp:
4306   case Hexagon::L2_loadrubgp:
4307   case Hexagon::L2_loadruhgp:
4308   case Hexagon::S2_storerbgp:
4309   case Hexagon::S2_storerbnewgp:
4310   case Hexagon::S2_storerhgp:
4311   case Hexagon::S2_storerhnewgp:
4312   case Hexagon::S2_storerigp:
4313   case Hexagon::S2_storerinewgp:
4314   case Hexagon::S2_storerdgp:
4315   case Hexagon::S2_storerfgp:
4316     return true;
4317   }
4318   const uint64_t F = MI.getDesc().TSFlags;
4319   unsigned addrMode =
4320     ((F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask);
4321   // Disallow any base+offset instruction. The assembler does not yet reorder
4322   // based up any zero offset instruction.
4323   return (addrMode == HexagonII::BaseRegOffset ||
4324           addrMode == HexagonII::BaseImmOffset ||
4325           addrMode == HexagonII::BaseLongOffset);
4326 }
4327 
isPureSlot0(const MachineInstr & MI) const4328 bool HexagonInstrInfo::isPureSlot0(const MachineInstr &MI) const {
4329   // Workaround for the Global Scheduler. Sometimes, it creates
4330   // A4_ext as a Pseudo instruction and calls this function to see if
4331   // it can be added to an existing bundle. Since the instruction doesn't
4332   // belong to any BB yet, we can't use getUnits API.
4333   if (MI.getOpcode() == Hexagon::A4_ext)
4334     return false;
4335 
4336   unsigned FuncUnits = getUnits(MI);
4337   return HexagonFUnits::isSlot0Only(FuncUnits);
4338 }
4339 
isRestrictNoSlot1Store(const MachineInstr & MI) const4340 bool HexagonInstrInfo::isRestrictNoSlot1Store(const MachineInstr &MI) const {
4341   const uint64_t F = MI.getDesc().TSFlags;
4342   return ((F >> HexagonII::RestrictNoSlot1StorePos) &
4343           HexagonII::RestrictNoSlot1StoreMask);
4344 }
4345 
changeDuplexOpcode(MachineBasicBlock::instr_iterator MII,bool ToBigInstrs) const4346 void HexagonInstrInfo::changeDuplexOpcode(MachineBasicBlock::instr_iterator MII,
4347                                           bool ToBigInstrs) const {
4348   int Opcode = -1;
4349   if (ToBigInstrs) { // To BigCore Instr.
4350     // Check if the instruction can form a Duplex.
4351     if (getDuplexCandidateGroup(*MII))
4352       // Get the opcode marked "dup_*" tag.
4353       Opcode = getDuplexOpcode(*MII, ToBigInstrs);
4354   } else // To TinyCore Instr.
4355     Opcode = getDuplexOpcode(*MII, ToBigInstrs);
4356 
4357   // Change the opcode of the instruction.
4358   if (Opcode >= 0)
4359     MII->setDesc(get(Opcode));
4360 }
4361 
4362 // This function is used to translate instructions to facilitate generating
4363 // Duplexes on TinyCore.
translateInstrsForDup(MachineFunction & MF,bool ToBigInstrs) const4364 void HexagonInstrInfo::translateInstrsForDup(MachineFunction &MF,
4365                                              bool ToBigInstrs) const {
4366   for (auto &MB : MF)
4367     for (MachineBasicBlock::instr_iterator Instr = MB.instr_begin(),
4368                                            End = MB.instr_end();
4369          Instr != End; ++Instr)
4370       changeDuplexOpcode(Instr, ToBigInstrs);
4371 }
4372 
4373 // This is a specialized form of above function.
translateInstrsForDup(MachineBasicBlock::instr_iterator MII,bool ToBigInstrs) const4374 void HexagonInstrInfo::translateInstrsForDup(
4375     MachineBasicBlock::instr_iterator MII, bool ToBigInstrs) const {
4376   MachineBasicBlock *MBB = MII->getParent();
4377   while ((MII != MBB->instr_end()) && MII->isInsideBundle()) {
4378     changeDuplexOpcode(MII, ToBigInstrs);
4379     ++MII;
4380   }
4381 }
4382 
getMemAccessSize(const MachineInstr & MI) const4383 unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr &MI) const {
4384   using namespace HexagonII;
4385 
4386   const uint64_t F = MI.getDesc().TSFlags;
4387   unsigned S = (F >> MemAccessSizePos) & MemAccesSizeMask;
4388   unsigned Size = getMemAccessSizeInBytes(MemAccessSize(S));
4389   if (Size != 0)
4390     return Size;
4391 
4392   // Handle vector access sizes.
4393   const HexagonRegisterInfo &HRI = *Subtarget.getRegisterInfo();
4394   switch (S) {
4395     case HexagonII::HVXVectorAccess:
4396       return HRI.getSpillSize(Hexagon::HvxVRRegClass);
4397     default:
4398       llvm_unreachable("Unexpected instruction");
4399   }
4400 }
4401 
4402 // Returns the min value that doesn't need to be extended.
getMinValue(const MachineInstr & MI) const4403 int HexagonInstrInfo::getMinValue(const MachineInstr &MI) const {
4404   const uint64_t F = MI.getDesc().TSFlags;
4405   unsigned isSigned = (F >> HexagonII::ExtentSignedPos)
4406                     & HexagonII::ExtentSignedMask;
4407   unsigned bits =  (F >> HexagonII::ExtentBitsPos)
4408                     & HexagonII::ExtentBitsMask;
4409 
4410   if (isSigned) // if value is signed
4411     return -1U << (bits - 1);
4412   else
4413     return 0;
4414 }
4415 
4416 // Returns opcode of the non-extended equivalent instruction.
getNonExtOpcode(const MachineInstr & MI) const4417 short HexagonInstrInfo::getNonExtOpcode(const MachineInstr &MI) const {
4418   // Check if the instruction has a register form that uses register in place
4419   // of the extended operand, if so return that as the non-extended form.
4420   short NonExtOpcode = Hexagon::getRegForm(MI.getOpcode());
4421     if (NonExtOpcode >= 0)
4422       return NonExtOpcode;
4423 
4424   if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) {
4425     // Check addressing mode and retrieve non-ext equivalent instruction.
4426     switch (getAddrMode(MI)) {
4427     case HexagonII::Absolute:
4428       return Hexagon::changeAddrMode_abs_io(MI.getOpcode());
4429     case HexagonII::BaseImmOffset:
4430       return Hexagon::changeAddrMode_io_rr(MI.getOpcode());
4431     case HexagonII::BaseLongOffset:
4432       return Hexagon::changeAddrMode_ur_rr(MI.getOpcode());
4433 
4434     default:
4435       return -1;
4436     }
4437   }
4438   return -1;
4439 }
4440 
getPredReg(ArrayRef<MachineOperand> Cond,unsigned & PredReg,unsigned & PredRegPos,unsigned & PredRegFlags) const4441 bool HexagonInstrInfo::getPredReg(ArrayRef<MachineOperand> Cond,
4442       unsigned &PredReg, unsigned &PredRegPos, unsigned &PredRegFlags) const {
4443   if (Cond.empty())
4444     return false;
4445   assert(Cond.size() == 2);
4446   if (isNewValueJump(Cond[0].getImm()) || Cond[1].isMBB()) {
4447     LLVM_DEBUG(dbgs() << "No predregs for new-value jumps/endloop");
4448     return false;
4449   }
4450   PredReg = Cond[1].getReg();
4451   PredRegPos = 1;
4452   // See IfConversion.cpp why we add RegState::Implicit | RegState::Undef
4453   PredRegFlags = 0;
4454   if (Cond[1].isImplicit())
4455     PredRegFlags = RegState::Implicit;
4456   if (Cond[1].isUndef())
4457     PredRegFlags |= RegState::Undef;
4458   return true;
4459 }
4460 
getPseudoInstrPair(const MachineInstr & MI) const4461 short HexagonInstrInfo::getPseudoInstrPair(const MachineInstr &MI) const {
4462   return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Pseudo);
4463 }
4464 
getRegForm(const MachineInstr & MI) const4465 short HexagonInstrInfo::getRegForm(const MachineInstr &MI) const {
4466   return Hexagon::getRegForm(MI.getOpcode());
4467 }
4468 
4469 // Return the number of bytes required to encode the instruction.
4470 // Hexagon instructions are fixed length, 4 bytes, unless they
4471 // use a constant extender, which requires another 4 bytes.
4472 // For debug instructions and prolog labels, return 0.
getSize(const MachineInstr & MI) const4473 unsigned HexagonInstrInfo::getSize(const MachineInstr &MI) const {
4474   if (MI.isDebugInstr() || MI.isPosition())
4475     return 0;
4476 
4477   unsigned Size = MI.getDesc().getSize();
4478   if (!Size)
4479     // Assume the default insn size in case it cannot be determined
4480     // for whatever reason.
4481     Size = HEXAGON_INSTR_SIZE;
4482 
4483   if (isConstExtended(MI) || isExtended(MI))
4484     Size += HEXAGON_INSTR_SIZE;
4485 
4486   // Try and compute number of instructions in asm.
4487   if (BranchRelaxAsmLarge && MI.getOpcode() == Hexagon::INLINEASM) {
4488     const MachineBasicBlock &MBB = *MI.getParent();
4489     const MachineFunction *MF = MBB.getParent();
4490     const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
4491 
4492     // Count the number of register definitions to find the asm string.
4493     unsigned NumDefs = 0;
4494     for (; MI.getOperand(NumDefs).isReg() && MI.getOperand(NumDefs).isDef();
4495          ++NumDefs)
4496       assert(NumDefs != MI.getNumOperands()-2 && "No asm string?");
4497 
4498     assert(MI.getOperand(NumDefs).isSymbol() && "No asm string?");
4499     // Disassemble the AsmStr and approximate number of instructions.
4500     const char *AsmStr = MI.getOperand(NumDefs).getSymbolName();
4501     Size = getInlineAsmLength(AsmStr, *MAI);
4502   }
4503 
4504   return Size;
4505 }
4506 
getType(const MachineInstr & MI) const4507 uint64_t HexagonInstrInfo::getType(const MachineInstr &MI) const {
4508   const uint64_t F = MI.getDesc().TSFlags;
4509   return (F >> HexagonII::TypePos) & HexagonII::TypeMask;
4510 }
4511 
getUnits(const MachineInstr & MI) const4512 InstrStage::FuncUnits HexagonInstrInfo::getUnits(const MachineInstr &MI) const {
4513   const InstrItineraryData &II = *Subtarget.getInstrItineraryData();
4514   const InstrStage &IS = *II.beginStage(MI.getDesc().getSchedClass());
4515 
4516   return IS.getUnits();
4517 }
4518 
4519 // Calculate size of the basic block without debug instructions.
nonDbgBBSize(const MachineBasicBlock * BB) const4520 unsigned HexagonInstrInfo::nonDbgBBSize(const MachineBasicBlock *BB) const {
4521   return nonDbgMICount(BB->instr_begin(), BB->instr_end());
4522 }
4523 
nonDbgBundleSize(MachineBasicBlock::const_iterator BundleHead) const4524 unsigned HexagonInstrInfo::nonDbgBundleSize(
4525       MachineBasicBlock::const_iterator BundleHead) const {
4526   assert(BundleHead->isBundle() && "Not a bundle header");
4527   auto MII = BundleHead.getInstrIterator();
4528   // Skip the bundle header.
4529   return nonDbgMICount(++MII, getBundleEnd(BundleHead.getInstrIterator()));
4530 }
4531 
4532 /// immediateExtend - Changes the instruction in place to one using an immediate
4533 /// extender.
immediateExtend(MachineInstr & MI) const4534 void HexagonInstrInfo::immediateExtend(MachineInstr &MI) const {
4535   assert((isExtendable(MI)||isConstExtended(MI)) &&
4536                                "Instruction must be extendable");
4537   // Find which operand is extendable.
4538   short ExtOpNum = getCExtOpNum(MI);
4539   MachineOperand &MO = MI.getOperand(ExtOpNum);
4540   // This needs to be something we understand.
4541   assert((MO.isMBB() || MO.isImm()) &&
4542          "Branch with unknown extendable field type");
4543   // Mark given operand as extended.
4544   MO.addTargetFlag(HexagonII::HMOTF_ConstExtended);
4545 }
4546 
invertAndChangeJumpTarget(MachineInstr & MI,MachineBasicBlock * NewTarget) const4547 bool HexagonInstrInfo::invertAndChangeJumpTarget(
4548       MachineInstr &MI, MachineBasicBlock *NewTarget) const {
4549   LLVM_DEBUG(dbgs() << "\n[invertAndChangeJumpTarget] to "
4550                     << printMBBReference(*NewTarget);
4551              MI.dump(););
4552   assert(MI.isBranch());
4553   unsigned NewOpcode = getInvertedPredicatedOpcode(MI.getOpcode());
4554   int TargetPos = MI.getNumOperands() - 1;
4555   // In general branch target is the last operand,
4556   // but some implicit defs added at the end might change it.
4557   while ((TargetPos > -1) && !MI.getOperand(TargetPos).isMBB())
4558     --TargetPos;
4559   assert((TargetPos >= 0) && MI.getOperand(TargetPos).isMBB());
4560   MI.getOperand(TargetPos).setMBB(NewTarget);
4561   if (EnableBranchPrediction && isPredicatedNew(MI)) {
4562     NewOpcode = reversePrediction(NewOpcode);
4563   }
4564   MI.setDesc(get(NewOpcode));
4565   return true;
4566 }
4567 
genAllInsnTimingClasses(MachineFunction & MF) const4568 void HexagonInstrInfo::genAllInsnTimingClasses(MachineFunction &MF) const {
4569   /* +++ The code below is used to generate complete set of Hexagon Insn +++ */
4570   MachineFunction::iterator A = MF.begin();
4571   MachineBasicBlock &B = *A;
4572   MachineBasicBlock::iterator I = B.begin();
4573   DebugLoc DL = I->getDebugLoc();
4574   MachineInstr *NewMI;
4575 
4576   for (unsigned insn = TargetOpcode::GENERIC_OP_END+1;
4577        insn < Hexagon::INSTRUCTION_LIST_END; ++insn) {
4578     NewMI = BuildMI(B, I, DL, get(insn));
4579     LLVM_DEBUG(dbgs() << "\n"
4580                       << getName(NewMI->getOpcode())
4581                       << "  Class: " << NewMI->getDesc().getSchedClass());
4582     NewMI->eraseFromParent();
4583   }
4584   /* --- The code above is used to generate complete set of Hexagon Insn --- */
4585 }
4586 
4587 // inverts the predication logic.
4588 // p -> NotP
4589 // NotP -> P
reversePredSense(MachineInstr & MI) const4590 bool HexagonInstrInfo::reversePredSense(MachineInstr &MI) const {
4591   LLVM_DEBUG(dbgs() << "\nTrying to reverse pred. sense of:"; MI.dump());
4592   MI.setDesc(get(getInvertedPredicatedOpcode(MI.getOpcode())));
4593   return true;
4594 }
4595 
4596 // Reverse the branch prediction.
reversePrediction(unsigned Opcode) const4597 unsigned HexagonInstrInfo::reversePrediction(unsigned Opcode) const {
4598   int PredRevOpcode = -1;
4599   if (isPredictedTaken(Opcode))
4600     PredRevOpcode = Hexagon::notTakenBranchPrediction(Opcode);
4601   else
4602     PredRevOpcode = Hexagon::takenBranchPrediction(Opcode);
4603   assert(PredRevOpcode > 0);
4604   return PredRevOpcode;
4605 }
4606 
4607 // TODO: Add more rigorous validation.
validateBranchCond(const ArrayRef<MachineOperand> & Cond) const4608 bool HexagonInstrInfo::validateBranchCond(const ArrayRef<MachineOperand> &Cond)
4609       const {
4610   return Cond.empty() || (Cond[0].isImm() && (Cond.size() != 1));
4611 }
4612 
4613 void HexagonInstrInfo::
setBundleNoShuf(MachineBasicBlock::instr_iterator MIB) const4614 setBundleNoShuf(MachineBasicBlock::instr_iterator MIB) const {
4615   assert(MIB->isBundle());
4616   MachineOperand &Operand = MIB->getOperand(0);
4617   if (Operand.isImm())
4618     Operand.setImm(Operand.getImm() | memShufDisabledMask);
4619   else
4620     MIB->addOperand(MachineOperand::CreateImm(memShufDisabledMask));
4621 }
4622 
getBundleNoShuf(const MachineInstr & MIB) const4623 bool HexagonInstrInfo::getBundleNoShuf(const MachineInstr &MIB) const {
4624   assert(MIB.isBundle());
4625   const MachineOperand &Operand = MIB.getOperand(0);
4626   return (Operand.isImm() && (Operand.getImm() & memShufDisabledMask) != 0);
4627 }
4628 
4629 // Addressing mode relations.
changeAddrMode_abs_io(short Opc) const4630 short HexagonInstrInfo::changeAddrMode_abs_io(short Opc) const {
4631   return Opc >= 0 ? Hexagon::changeAddrMode_abs_io(Opc) : Opc;
4632 }
4633 
changeAddrMode_io_abs(short Opc) const4634 short HexagonInstrInfo::changeAddrMode_io_abs(short Opc) const {
4635   return Opc >= 0 ? Hexagon::changeAddrMode_io_abs(Opc) : Opc;
4636 }
4637 
changeAddrMode_io_pi(short Opc) const4638 short HexagonInstrInfo::changeAddrMode_io_pi(short Opc) const {
4639   return Opc >= 0 ? Hexagon::changeAddrMode_io_pi(Opc) : Opc;
4640 }
4641 
changeAddrMode_io_rr(short Opc) const4642 short HexagonInstrInfo::changeAddrMode_io_rr(short Opc) const {
4643   return Opc >= 0 ? Hexagon::changeAddrMode_io_rr(Opc) : Opc;
4644 }
4645 
changeAddrMode_pi_io(short Opc) const4646 short HexagonInstrInfo::changeAddrMode_pi_io(short Opc) const {
4647   return Opc >= 0 ? Hexagon::changeAddrMode_pi_io(Opc) : Opc;
4648 }
4649 
changeAddrMode_rr_io(short Opc) const4650 short HexagonInstrInfo::changeAddrMode_rr_io(short Opc) const {
4651   return Opc >= 0 ? Hexagon::changeAddrMode_rr_io(Opc) : Opc;
4652 }
4653 
changeAddrMode_rr_ur(short Opc) const4654 short HexagonInstrInfo::changeAddrMode_rr_ur(short Opc) const {
4655   return Opc >= 0 ? Hexagon::changeAddrMode_rr_ur(Opc) : Opc;
4656 }
4657 
changeAddrMode_ur_rr(short Opc) const4658 short HexagonInstrInfo::changeAddrMode_ur_rr(short Opc) const {
4659   return Opc >= 0 ? Hexagon::changeAddrMode_ur_rr(Opc) : Opc;
4660 }
4661