xref: /freebsd/contrib/llvm-project/llvm/lib/Target/ARM/Thumb2InstrInfo.cpp (revision 5036d9652a5701d00e9e40ea942c278e9f77d33d)
1 //===- Thumb2InstrInfo.cpp - Thumb-2 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 Thumb-2 implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "Thumb2InstrInfo.h"
14 #include "ARMMachineFunctionInfo.h"
15 #include "ARMSubtarget.h"
16 #include "MCTargetDesc/ARMAddressingModes.h"
17 #include "llvm/CodeGen/MachineBasicBlock.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/MachineFunction.h"
20 #include "llvm/CodeGen/MachineInstr.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineMemOperand.h"
23 #include "llvm/CodeGen/MachineOperand.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/CodeGen/TargetRegisterInfo.h"
26 #include "llvm/IR/DebugLoc.h"
27 #include "llvm/IR/Module.h"
28 #include "llvm/MC/MCInst.h"
29 #include "llvm/MC/MCInstBuilder.h"
30 #include "llvm/MC/MCInstrDesc.h"
31 #include "llvm/Support/CommandLine.h"
32 #include "llvm/Support/ErrorHandling.h"
33 #include "llvm/Support/MathExtras.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include <cassert>
36 
37 using namespace llvm;
38 
39 static cl::opt<bool>
40 OldT2IfCvt("old-thumb2-ifcvt", cl::Hidden,
41            cl::desc("Use old-style Thumb2 if-conversion heuristics"),
42            cl::init(false));
43 
44 static cl::opt<bool>
45 PreferNoCSEL("prefer-no-csel", cl::Hidden,
46              cl::desc("Prefer predicated Move to CSEL"),
47              cl::init(false));
48 
49 Thumb2InstrInfo::Thumb2InstrInfo(const ARMSubtarget &STI)
50     : ARMBaseInstrInfo(STI) {}
51 
52 /// Return the noop instruction to use for a noop.
53 MCInst Thumb2InstrInfo::getNop() const {
54   return MCInstBuilder(ARM::tHINT).addImm(0).addImm(ARMCC::AL).addReg(0);
55 }
56 
57 unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const {
58   // FIXME
59   return 0;
60 }
61 
62 void
63 Thumb2InstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
64                                          MachineBasicBlock *NewDest) const {
65   MachineBasicBlock *MBB = Tail->getParent();
66   ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>();
67   if (!AFI->hasITBlocks() || Tail->isBranch()) {
68     TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest);
69     return;
70   }
71 
72   // If the first instruction of Tail is predicated, we may have to update
73   // the IT instruction.
74   Register PredReg;
75   ARMCC::CondCodes CC = getInstrPredicate(*Tail, PredReg);
76   MachineBasicBlock::iterator MBBI = Tail;
77   if (CC != ARMCC::AL)
78     // Expecting at least the t2IT instruction before it.
79     --MBBI;
80 
81   // Actually replace the tail.
82   TargetInstrInfo::ReplaceTailWithBranchTo(Tail, NewDest);
83 
84   // Fix up IT.
85   if (CC != ARMCC::AL) {
86     MachineBasicBlock::iterator E = MBB->begin();
87     unsigned Count = 4; // At most 4 instructions in an IT block.
88     while (Count && MBBI != E) {
89       if (MBBI->isDebugInstr()) {
90         --MBBI;
91         continue;
92       }
93       if (MBBI->getOpcode() == ARM::t2IT) {
94         unsigned Mask = MBBI->getOperand(1).getImm();
95         if (Count == 4)
96           MBBI->eraseFromParent();
97         else {
98           unsigned MaskOn = 1 << Count;
99           unsigned MaskOff = ~(MaskOn - 1);
100           MBBI->getOperand(1).setImm((Mask & MaskOff) | MaskOn);
101         }
102         return;
103       }
104       --MBBI;
105       --Count;
106     }
107 
108     // Ctrl flow can reach here if branch folding is run before IT block
109     // formation pass.
110   }
111 }
112 
113 bool
114 Thumb2InstrInfo::isLegalToSplitMBBAt(MachineBasicBlock &MBB,
115                                      MachineBasicBlock::iterator MBBI) const {
116   while (MBBI->isDebugInstr()) {
117     ++MBBI;
118     if (MBBI == MBB.end())
119       return false;
120   }
121 
122   Register PredReg;
123   return getITInstrPredicate(*MBBI, PredReg) == ARMCC::AL;
124 }
125 
126 MachineInstr *
127 Thumb2InstrInfo::optimizeSelect(MachineInstr &MI,
128                                 SmallPtrSetImpl<MachineInstr *> &SeenMIs,
129                                 bool PreferFalse) const {
130   // Try to use the base optimizeSelect, which uses canFoldIntoMOVCC to fold the
131   // MOVCC into another instruction. If that fails on 8.1-M fall back to using a
132   // CSEL.
133   MachineInstr *RV = ARMBaseInstrInfo::optimizeSelect(MI, SeenMIs, PreferFalse);
134   if (!RV && getSubtarget().hasV8_1MMainlineOps() && !PreferNoCSEL) {
135     Register DestReg = MI.getOperand(0).getReg();
136 
137     if (!DestReg.isVirtual())
138       return nullptr;
139 
140     MachineInstrBuilder NewMI = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
141                                         get(ARM::t2CSEL), DestReg)
142                                     .add(MI.getOperand(2))
143                                     .add(MI.getOperand(1))
144                                     .add(MI.getOperand(3));
145     SeenMIs.insert(NewMI);
146     return NewMI;
147   }
148   return RV;
149 }
150 
151 void Thumb2InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
152                                   MachineBasicBlock::iterator I,
153                                   const DebugLoc &DL, MCRegister DestReg,
154                                   MCRegister SrcReg, bool KillSrc) const {
155   // Handle SPR, DPR, and QPR copies.
156   if (!ARM::GPRRegClass.contains(DestReg, SrcReg))
157     return ARMBaseInstrInfo::copyPhysReg(MBB, I, DL, DestReg, SrcReg, KillSrc);
158 
159   BuildMI(MBB, I, DL, get(ARM::tMOVr), DestReg)
160       .addReg(SrcReg, getKillRegState(KillSrc))
161       .add(predOps(ARMCC::AL));
162 }
163 
164 void Thumb2InstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
165                                           MachineBasicBlock::iterator I,
166                                           Register SrcReg, bool isKill, int FI,
167                                           const TargetRegisterClass *RC,
168                                           const TargetRegisterInfo *TRI,
169                                           Register VReg) const {
170   DebugLoc DL;
171   if (I != MBB.end()) DL = I->getDebugLoc();
172 
173   MachineFunction &MF = *MBB.getParent();
174   MachineFrameInfo &MFI = MF.getFrameInfo();
175   MachineMemOperand *MMO = MF.getMachineMemOperand(
176       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
177       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
178 
179   if (ARM::GPRRegClass.hasSubClassEq(RC)) {
180     BuildMI(MBB, I, DL, get(ARM::t2STRi12))
181         .addReg(SrcReg, getKillRegState(isKill))
182         .addFrameIndex(FI)
183         .addImm(0)
184         .addMemOperand(MMO)
185         .add(predOps(ARMCC::AL));
186     return;
187   }
188 
189   if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
190     // Thumb2 STRD expects its dest-registers to be in rGPR. Not a problem for
191     // gsub_0, but needs an extra constraint for gsub_1 (which could be sp
192     // otherwise).
193     if (SrcReg.isVirtual()) {
194       MachineRegisterInfo *MRI = &MF.getRegInfo();
195       MRI->constrainRegClass(SrcReg, &ARM::GPRPairnospRegClass);
196     }
197 
198     MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2STRDi8));
199     AddDReg(MIB, SrcReg, ARM::gsub_0, getKillRegState(isKill), TRI);
200     AddDReg(MIB, SrcReg, ARM::gsub_1, 0, TRI);
201     MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL));
202     return;
203   }
204 
205   ARMBaseInstrInfo::storeRegToStackSlot(MBB, I, SrcReg, isKill, FI, RC, TRI,
206                                         Register());
207 }
208 
209 void Thumb2InstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
210                                            MachineBasicBlock::iterator I,
211                                            Register DestReg, int FI,
212                                            const TargetRegisterClass *RC,
213                                            const TargetRegisterInfo *TRI,
214                                            Register VReg) const {
215   MachineFunction &MF = *MBB.getParent();
216   MachineFrameInfo &MFI = MF.getFrameInfo();
217   MachineMemOperand *MMO = MF.getMachineMemOperand(
218       MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
219       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
220   DebugLoc DL;
221   if (I != MBB.end()) DL = I->getDebugLoc();
222 
223   if (ARM::GPRRegClass.hasSubClassEq(RC)) {
224     BuildMI(MBB, I, DL, get(ARM::t2LDRi12), DestReg)
225         .addFrameIndex(FI)
226         .addImm(0)
227         .addMemOperand(MMO)
228         .add(predOps(ARMCC::AL));
229     return;
230   }
231 
232   if (ARM::GPRPairRegClass.hasSubClassEq(RC)) {
233     // Thumb2 LDRD expects its dest-registers to be in rGPR. Not a problem for
234     // gsub_0, but needs an extra constraint for gsub_1 (which could be sp
235     // otherwise).
236     if (DestReg.isVirtual()) {
237       MachineRegisterInfo *MRI = &MF.getRegInfo();
238       MRI->constrainRegClass(DestReg, &ARM::GPRPairnospRegClass);
239     }
240 
241     MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(ARM::t2LDRDi8));
242     AddDReg(MIB, DestReg, ARM::gsub_0, RegState::DefineNoRead, TRI);
243     AddDReg(MIB, DestReg, ARM::gsub_1, RegState::DefineNoRead, TRI);
244     MIB.addFrameIndex(FI).addImm(0).addMemOperand(MMO).add(predOps(ARMCC::AL));
245 
246     if (DestReg.isPhysical())
247       MIB.addReg(DestReg, RegState::ImplicitDefine);
248     return;
249   }
250 
251   ARMBaseInstrInfo::loadRegFromStackSlot(MBB, I, DestReg, FI, RC, TRI,
252                                          Register());
253 }
254 
255 void Thumb2InstrInfo::expandLoadStackGuard(
256     MachineBasicBlock::iterator MI) const {
257   MachineFunction &MF = *MI->getParent()->getParent();
258   Module &M = *MF.getFunction().getParent();
259 
260   if (M.getStackProtectorGuard() == "tls") {
261     expandLoadStackGuardBase(MI, ARM::t2MRC, ARM::t2LDRi12);
262     return;
263   }
264 
265   const auto *GV = cast<GlobalValue>((*MI->memoperands_begin())->getValue());
266   if (MF.getSubtarget<ARMSubtarget>().isTargetELF() && !GV->isDSOLocal())
267     expandLoadStackGuardBase(MI, ARM::t2LDRLIT_ga_pcrel, ARM::t2LDRi12);
268   else if (MF.getTarget().isPositionIndependent())
269     expandLoadStackGuardBase(MI, ARM::t2MOV_ga_pcrel, ARM::t2LDRi12);
270   else
271     expandLoadStackGuardBase(MI, ARM::t2MOVi32imm, ARM::t2LDRi12);
272 }
273 
274 MachineInstr *Thumb2InstrInfo::commuteInstructionImpl(MachineInstr &MI,
275                                                       bool NewMI,
276                                                       unsigned OpIdx1,
277                                                       unsigned OpIdx2) const {
278   switch (MI.getOpcode()) {
279   case ARM::MVE_VMAXNMAf16:
280   case ARM::MVE_VMAXNMAf32:
281   case ARM::MVE_VMINNMAf16:
282   case ARM::MVE_VMINNMAf32:
283     // Don't allow predicated instructions to be commuted.
284     if (getVPTInstrPredicate(MI) != ARMVCC::None)
285       return nullptr;
286   }
287   return ARMBaseInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
288 }
289 
290 bool Thumb2InstrInfo::isSchedulingBoundary(const MachineInstr &MI,
291                                            const MachineBasicBlock *MBB,
292                                            const MachineFunction &MF) const {
293   // BTI clearing instructions shall not take part in scheduling regions as
294   // they must stay in their intended place. Although PAC isn't BTI clearing,
295   // it can be transformed into PACBTI after the pre-RA Machine Scheduling
296   // has taken place, so its movement must also be restricted.
297   switch (MI.getOpcode()) {
298   case ARM::t2BTI:
299   case ARM::t2PAC:
300   case ARM::t2PACBTI:
301   case ARM::t2SG:
302     return true;
303   default:
304     break;
305   }
306   return ARMBaseInstrInfo::isSchedulingBoundary(MI, MBB, MF);
307 }
308 
309 void llvm::emitT2RegPlusImmediate(MachineBasicBlock &MBB,
310                                   MachineBasicBlock::iterator &MBBI,
311                                   const DebugLoc &dl, Register DestReg,
312                                   Register BaseReg, int NumBytes,
313                                   ARMCC::CondCodes Pred, Register PredReg,
314                                   const ARMBaseInstrInfo &TII,
315                                   unsigned MIFlags) {
316   if (NumBytes == 0 && DestReg != BaseReg) {
317     BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg)
318       .addReg(BaseReg, RegState::Kill)
319       .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
320     return;
321   }
322 
323   bool isSub = NumBytes < 0;
324   if (isSub) NumBytes = -NumBytes;
325 
326   // If profitable, use a movw or movt to materialize the offset.
327   // FIXME: Use the scavenger to grab a scratch register.
328   if (DestReg != ARM::SP && DestReg != BaseReg &&
329       NumBytes >= 4096 &&
330       ARM_AM::getT2SOImmVal(NumBytes) == -1) {
331     bool Fits = false;
332     if (NumBytes < 65536) {
333       // Use a movw to materialize the 16-bit constant.
334       BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVi16), DestReg)
335         .addImm(NumBytes)
336         .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
337       Fits = true;
338     } else if ((NumBytes & 0xffff) == 0) {
339       // Use a movt to materialize the 32-bit constant.
340       BuildMI(MBB, MBBI, dl, TII.get(ARM::t2MOVTi16), DestReg)
341         .addReg(DestReg)
342         .addImm(NumBytes >> 16)
343         .addImm((unsigned)Pred).addReg(PredReg).setMIFlags(MIFlags);
344       Fits = true;
345     }
346 
347     if (Fits) {
348       if (isSub) {
349         BuildMI(MBB, MBBI, dl, TII.get(ARM::t2SUBrr), DestReg)
350             .addReg(BaseReg)
351             .addReg(DestReg, RegState::Kill)
352             .add(predOps(Pred, PredReg))
353             .add(condCodeOp())
354             .setMIFlags(MIFlags);
355       } else {
356         // Here we know that DestReg is not SP but we do not
357         // know anything about BaseReg. t2ADDrr is an invalid
358         // instruction is SP is used as the second argument, but
359         // is fine if SP is the first argument. To be sure we
360         // do not generate invalid encoding, put BaseReg first.
361         BuildMI(MBB, MBBI, dl, TII.get(ARM::t2ADDrr), DestReg)
362             .addReg(BaseReg)
363             .addReg(DestReg, RegState::Kill)
364             .add(predOps(Pred, PredReg))
365             .add(condCodeOp())
366             .setMIFlags(MIFlags);
367       }
368       return;
369     }
370   }
371 
372   while (NumBytes) {
373     unsigned ThisVal = NumBytes;
374     unsigned Opc = 0;
375     if (DestReg == ARM::SP && BaseReg != ARM::SP) {
376       // mov sp, rn. Note t2MOVr cannot be used.
377       BuildMI(MBB, MBBI, dl, TII.get(ARM::tMOVr), DestReg)
378           .addReg(BaseReg)
379           .setMIFlags(MIFlags)
380           .add(predOps(ARMCC::AL));
381       BaseReg = ARM::SP;
382       continue;
383     }
384 
385     assert((DestReg != ARM::SP || BaseReg == ARM::SP) &&
386            "Writing to SP, from other register.");
387 
388     // Try to use T1, as it smaller
389     if ((DestReg == ARM::SP) && (ThisVal < ((1 << 7) - 1) * 4)) {
390       assert((ThisVal & 3) == 0 && "Stack update is not multiple of 4?");
391       Opc = isSub ? ARM::tSUBspi : ARM::tADDspi;
392       BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
393           .addReg(BaseReg)
394           .addImm(ThisVal / 4)
395           .setMIFlags(MIFlags)
396           .add(predOps(ARMCC::AL));
397       break;
398     }
399     bool HasCCOut = true;
400     int ImmIsT2SO = ARM_AM::getT2SOImmVal(ThisVal);
401     bool ToSP = DestReg == ARM::SP;
402     unsigned t2SUB = ToSP ? ARM::t2SUBspImm : ARM::t2SUBri;
403     unsigned t2ADD = ToSP ? ARM::t2ADDspImm : ARM::t2ADDri;
404     unsigned t2SUBi12 = ToSP ? ARM::t2SUBspImm12 : ARM::t2SUBri12;
405     unsigned t2ADDi12 = ToSP ? ARM::t2ADDspImm12 : ARM::t2ADDri12;
406     Opc = isSub ? t2SUB : t2ADD;
407     // Prefer T2: sub rd, rn, so_imm | sub sp, sp, so_imm
408     if (ImmIsT2SO != -1) {
409       NumBytes = 0;
410     } else if (ThisVal < 4096) {
411       // Prefer T3 if can make it in a single go: subw rd, rn, imm12 | subw sp,
412       // sp, imm12
413       Opc = isSub ? t2SUBi12 : t2ADDi12;
414       HasCCOut = false;
415       NumBytes = 0;
416     } else {
417       // Use one T2 instruction to reduce NumBytes
418       // FIXME: Move this to ARMAddressingModes.h?
419       unsigned RotAmt = llvm::countl_zero(ThisVal);
420       ThisVal = ThisVal & llvm::rotr<uint32_t>(0xff000000U, RotAmt);
421       NumBytes &= ~ThisVal;
422       assert(ARM_AM::getT2SOImmVal(ThisVal) != -1 &&
423              "Bit extraction didn't work?");
424     }
425 
426     // Build the new ADD / SUB.
427     MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII.get(Opc), DestReg)
428                                   .addReg(BaseReg, RegState::Kill)
429                                   .addImm(ThisVal)
430                                   .add(predOps(ARMCC::AL))
431                                   .setMIFlags(MIFlags);
432     if (HasCCOut)
433       MIB.add(condCodeOp());
434 
435     BaseReg = DestReg;
436   }
437 }
438 
439 static unsigned
440 negativeOffsetOpcode(unsigned opcode)
441 {
442   switch (opcode) {
443   case ARM::t2LDRi12:   return ARM::t2LDRi8;
444   case ARM::t2LDRHi12:  return ARM::t2LDRHi8;
445   case ARM::t2LDRBi12:  return ARM::t2LDRBi8;
446   case ARM::t2LDRSHi12: return ARM::t2LDRSHi8;
447   case ARM::t2LDRSBi12: return ARM::t2LDRSBi8;
448   case ARM::t2STRi12:   return ARM::t2STRi8;
449   case ARM::t2STRBi12:  return ARM::t2STRBi8;
450   case ARM::t2STRHi12:  return ARM::t2STRHi8;
451   case ARM::t2PLDi12:   return ARM::t2PLDi8;
452   case ARM::t2PLDWi12:  return ARM::t2PLDWi8;
453   case ARM::t2PLIi12:   return ARM::t2PLIi8;
454 
455   case ARM::t2LDRi8:
456   case ARM::t2LDRHi8:
457   case ARM::t2LDRBi8:
458   case ARM::t2LDRSHi8:
459   case ARM::t2LDRSBi8:
460   case ARM::t2STRi8:
461   case ARM::t2STRBi8:
462   case ARM::t2STRHi8:
463   case ARM::t2PLDi8:
464   case ARM::t2PLDWi8:
465   case ARM::t2PLIi8:
466     return opcode;
467 
468   default:
469     llvm_unreachable("unknown thumb2 opcode.");
470   }
471 }
472 
473 static unsigned
474 positiveOffsetOpcode(unsigned opcode)
475 {
476   switch (opcode) {
477   case ARM::t2LDRi8:   return ARM::t2LDRi12;
478   case ARM::t2LDRHi8:  return ARM::t2LDRHi12;
479   case ARM::t2LDRBi8:  return ARM::t2LDRBi12;
480   case ARM::t2LDRSHi8: return ARM::t2LDRSHi12;
481   case ARM::t2LDRSBi8: return ARM::t2LDRSBi12;
482   case ARM::t2STRi8:   return ARM::t2STRi12;
483   case ARM::t2STRBi8:  return ARM::t2STRBi12;
484   case ARM::t2STRHi8:  return ARM::t2STRHi12;
485   case ARM::t2PLDi8:   return ARM::t2PLDi12;
486   case ARM::t2PLDWi8:  return ARM::t2PLDWi12;
487   case ARM::t2PLIi8:   return ARM::t2PLIi12;
488 
489   case ARM::t2LDRi12:
490   case ARM::t2LDRHi12:
491   case ARM::t2LDRBi12:
492   case ARM::t2LDRSHi12:
493   case ARM::t2LDRSBi12:
494   case ARM::t2STRi12:
495   case ARM::t2STRBi12:
496   case ARM::t2STRHi12:
497   case ARM::t2PLDi12:
498   case ARM::t2PLDWi12:
499   case ARM::t2PLIi12:
500     return opcode;
501 
502   default:
503     llvm_unreachable("unknown thumb2 opcode.");
504   }
505 }
506 
507 static unsigned
508 immediateOffsetOpcode(unsigned opcode)
509 {
510   switch (opcode) {
511   case ARM::t2LDRs:   return ARM::t2LDRi12;
512   case ARM::t2LDRHs:  return ARM::t2LDRHi12;
513   case ARM::t2LDRBs:  return ARM::t2LDRBi12;
514   case ARM::t2LDRSHs: return ARM::t2LDRSHi12;
515   case ARM::t2LDRSBs: return ARM::t2LDRSBi12;
516   case ARM::t2STRs:   return ARM::t2STRi12;
517   case ARM::t2STRBs:  return ARM::t2STRBi12;
518   case ARM::t2STRHs:  return ARM::t2STRHi12;
519   case ARM::t2PLDs:   return ARM::t2PLDi12;
520   case ARM::t2PLDWs:  return ARM::t2PLDWi12;
521   case ARM::t2PLIs:   return ARM::t2PLIi12;
522 
523   case ARM::t2LDRi12:
524   case ARM::t2LDRHi12:
525   case ARM::t2LDRBi12:
526   case ARM::t2LDRSHi12:
527   case ARM::t2LDRSBi12:
528   case ARM::t2STRi12:
529   case ARM::t2STRBi12:
530   case ARM::t2STRHi12:
531   case ARM::t2PLDi12:
532   case ARM::t2PLDWi12:
533   case ARM::t2PLIi12:
534   case ARM::t2LDRi8:
535   case ARM::t2LDRHi8:
536   case ARM::t2LDRBi8:
537   case ARM::t2LDRSHi8:
538   case ARM::t2LDRSBi8:
539   case ARM::t2STRi8:
540   case ARM::t2STRBi8:
541   case ARM::t2STRHi8:
542   case ARM::t2PLDi8:
543   case ARM::t2PLDWi8:
544   case ARM::t2PLIi8:
545     return opcode;
546 
547   default:
548     llvm_unreachable("unknown thumb2 opcode.");
549   }
550 }
551 
552 bool llvm::rewriteT2FrameIndex(MachineInstr &MI, unsigned FrameRegIdx,
553                                Register FrameReg, int &Offset,
554                                const ARMBaseInstrInfo &TII,
555                                const TargetRegisterInfo *TRI) {
556   unsigned Opcode = MI.getOpcode();
557   const MCInstrDesc &Desc = MI.getDesc();
558   unsigned AddrMode = (Desc.TSFlags & ARMII::AddrModeMask);
559   bool isSub = false;
560 
561   MachineFunction &MF = *MI.getParent()->getParent();
562   const TargetRegisterClass *RegClass =
563       TII.getRegClass(Desc, FrameRegIdx, TRI, MF);
564 
565   // Memory operands in inline assembly always use AddrModeT2_i12.
566   if (Opcode == ARM::INLINEASM || Opcode == ARM::INLINEASM_BR)
567     AddrMode = ARMII::AddrModeT2_i12; // FIXME. mode for thumb2?
568 
569   const bool IsSP = Opcode == ARM::t2ADDspImm12 || Opcode == ARM::t2ADDspImm;
570   if (IsSP || Opcode == ARM::t2ADDri || Opcode == ARM::t2ADDri12) {
571     Offset += MI.getOperand(FrameRegIdx+1).getImm();
572 
573     Register PredReg;
574     if (Offset == 0 && getInstrPredicate(MI, PredReg) == ARMCC::AL &&
575         !MI.definesRegister(ARM::CPSR, /*TRI=*/nullptr)) {
576       // Turn it into a move.
577       MI.setDesc(TII.get(ARM::tMOVr));
578       MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
579       // Remove offset and remaining explicit predicate operands.
580       do MI.removeOperand(FrameRegIdx+1);
581       while (MI.getNumOperands() > FrameRegIdx+1);
582       MachineInstrBuilder MIB(*MI.getParent()->getParent(), &MI);
583       MIB.add(predOps(ARMCC::AL));
584       return true;
585     }
586 
587     bool HasCCOut = (Opcode != ARM::t2ADDspImm12 && Opcode != ARM::t2ADDri12);
588 
589     if (Offset < 0) {
590       Offset = -Offset;
591       isSub = true;
592       MI.setDesc(IsSP ? TII.get(ARM::t2SUBspImm) : TII.get(ARM::t2SUBri));
593     } else {
594       MI.setDesc(IsSP ? TII.get(ARM::t2ADDspImm) : TII.get(ARM::t2ADDri));
595     }
596 
597     // Common case: small offset, fits into instruction.
598     if (ARM_AM::getT2SOImmVal(Offset) != -1) {
599       MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
600       MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
601       // Add cc_out operand if the original instruction did not have one.
602       if (!HasCCOut)
603         MI.addOperand(MachineOperand::CreateReg(0, false));
604       Offset = 0;
605       return true;
606     }
607     // Another common case: imm12.
608     if (Offset < 4096 &&
609         (!HasCCOut || MI.getOperand(MI.getNumOperands()-1).getReg() == 0)) {
610       unsigned NewOpc = isSub ? IsSP ? ARM::t2SUBspImm12 : ARM::t2SUBri12
611                               : IsSP ? ARM::t2ADDspImm12 : ARM::t2ADDri12;
612       MI.setDesc(TII.get(NewOpc));
613       MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
614       MI.getOperand(FrameRegIdx+1).ChangeToImmediate(Offset);
615       // Remove the cc_out operand.
616       if (HasCCOut)
617         MI.removeOperand(MI.getNumOperands()-1);
618       Offset = 0;
619       return true;
620     }
621 
622     // Otherwise, extract 8 adjacent bits from the immediate into this
623     // t2ADDri/t2SUBri.
624     unsigned RotAmt = llvm::countl_zero<unsigned>(Offset);
625     unsigned ThisImmVal = Offset & llvm::rotr<uint32_t>(0xff000000U, RotAmt);
626 
627     // We will handle these bits from offset, clear them.
628     Offset &= ~ThisImmVal;
629 
630     assert(ARM_AM::getT2SOImmVal(ThisImmVal) != -1 &&
631            "Bit extraction didn't work?");
632     MI.getOperand(FrameRegIdx+1).ChangeToImmediate(ThisImmVal);
633     // Add cc_out operand if the original instruction did not have one.
634     if (!HasCCOut)
635       MI.addOperand(MachineOperand::CreateReg(0, false));
636   } else {
637     // AddrMode4 and AddrMode6 cannot handle any offset.
638     if (AddrMode == ARMII::AddrMode4 || AddrMode == ARMII::AddrMode6)
639       return false;
640 
641     // AddrModeT2_so cannot handle any offset. If there is no offset
642     // register then we change to an immediate version.
643     unsigned NewOpc = Opcode;
644     if (AddrMode == ARMII::AddrModeT2_so) {
645       Register OffsetReg = MI.getOperand(FrameRegIdx + 1).getReg();
646       if (OffsetReg != 0) {
647         MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
648         return Offset == 0;
649       }
650 
651       MI.removeOperand(FrameRegIdx+1);
652       MI.getOperand(FrameRegIdx+1).ChangeToImmediate(0);
653       NewOpc = immediateOffsetOpcode(Opcode);
654       AddrMode = ARMII::AddrModeT2_i12;
655     }
656 
657     unsigned NumBits = 0;
658     unsigned Scale = 1;
659     if (AddrMode == ARMII::AddrModeT2_i8neg ||
660         AddrMode == ARMII::AddrModeT2_i12) {
661       // i8 supports only negative, and i12 supports only positive, so
662       // based on Offset sign convert Opcode to the appropriate
663       // instruction
664       Offset += MI.getOperand(FrameRegIdx+1).getImm();
665       if (Offset < 0) {
666         NewOpc = negativeOffsetOpcode(Opcode);
667         NumBits = 8;
668         isSub = true;
669         Offset = -Offset;
670       } else {
671         NewOpc = positiveOffsetOpcode(Opcode);
672         NumBits = 12;
673       }
674     } else if (AddrMode == ARMII::AddrMode5) {
675       // VFP address mode.
676       const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1);
677       int InstrOffs = ARM_AM::getAM5Offset(OffOp.getImm());
678       if (ARM_AM::getAM5Op(OffOp.getImm()) == ARM_AM::sub)
679         InstrOffs *= -1;
680       NumBits = 8;
681       Scale = 4;
682       Offset += InstrOffs * 4;
683       assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
684       if (Offset < 0) {
685         Offset = -Offset;
686         isSub = true;
687       }
688     } else if (AddrMode == ARMII::AddrMode5FP16) {
689       // VFP address mode.
690       const MachineOperand &OffOp = MI.getOperand(FrameRegIdx+1);
691       int InstrOffs = ARM_AM::getAM5FP16Offset(OffOp.getImm());
692       if (ARM_AM::getAM5FP16Op(OffOp.getImm()) == ARM_AM::sub)
693         InstrOffs *= -1;
694       NumBits = 8;
695       Scale = 2;
696       Offset += InstrOffs * 2;
697       assert((Offset & (Scale-1)) == 0 && "Can't encode this offset!");
698       if (Offset < 0) {
699         Offset = -Offset;
700         isSub = true;
701       }
702     } else if (AddrMode == ARMII::AddrModeT2_i7s4 ||
703                AddrMode == ARMII::AddrModeT2_i7s2 ||
704                AddrMode == ARMII::AddrModeT2_i7) {
705       Offset += MI.getOperand(FrameRegIdx + 1).getImm();
706       unsigned OffsetMask;
707       switch (AddrMode) {
708       case ARMII::AddrModeT2_i7s4: NumBits = 9; OffsetMask = 0x3; break;
709       case ARMII::AddrModeT2_i7s2: NumBits = 8; OffsetMask = 0x1; break;
710       default:                     NumBits = 7; OffsetMask = 0x0; break;
711       }
712       // MCInst operand expects already scaled value.
713       Scale = 1;
714       assert((Offset & OffsetMask) == 0 && "Can't encode this offset!");
715       (void)OffsetMask; // squash unused-variable warning at -NDEBUG
716     } else if (AddrMode == ARMII::AddrModeT2_i8s4) {
717       Offset += MI.getOperand(FrameRegIdx + 1).getImm();
718       NumBits = 8 + 2;
719       // MCInst operand expects already scaled value.
720       Scale = 1;
721       assert((Offset & 3) == 0 && "Can't encode this offset!");
722     } else if (AddrMode == ARMII::AddrModeT2_ldrex) {
723       Offset += MI.getOperand(FrameRegIdx + 1).getImm() * 4;
724       NumBits = 8; // 8 bits scaled by 4
725       Scale = 4;
726       assert((Offset & 3) == 0 && "Can't encode this offset!");
727     } else {
728       llvm_unreachable("Unsupported addressing mode!");
729     }
730 
731     if (NewOpc != Opcode)
732       MI.setDesc(TII.get(NewOpc));
733 
734     MachineOperand &ImmOp = MI.getOperand(FrameRegIdx+1);
735 
736     // Attempt to fold address computation
737     // Common case: small offset, fits into instruction. We need to make sure
738     // the register class is correct too, for instructions like the MVE
739     // VLDRH.32, which only accepts low tGPR registers.
740     int ImmedOffset = Offset / Scale;
741     unsigned Mask = (1 << NumBits) - 1;
742     if ((unsigned)Offset <= Mask * Scale &&
743         (FrameReg.isVirtual() || RegClass->contains(FrameReg))) {
744       if (FrameReg.isVirtual()) {
745         // Make sure the register class for the virtual register is correct
746         MachineRegisterInfo *MRI = &MF.getRegInfo();
747         if (!MRI->constrainRegClass(FrameReg, RegClass))
748           llvm_unreachable("Unable to constrain virtual register class.");
749       }
750 
751       // Replace the FrameIndex with fp/sp
752       MI.getOperand(FrameRegIdx).ChangeToRegister(FrameReg, false);
753       if (isSub) {
754         if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
755           // FIXME: Not consistent.
756           ImmedOffset |= 1 << NumBits;
757         else
758           ImmedOffset = -ImmedOffset;
759       }
760       ImmOp.ChangeToImmediate(ImmedOffset);
761       Offset = 0;
762       return true;
763     }
764 
765     // Otherwise, offset doesn't fit. Pull in what we can to simplify
766     ImmedOffset = ImmedOffset & Mask;
767     if (isSub) {
768       if (AddrMode == ARMII::AddrMode5 || AddrMode == ARMII::AddrMode5FP16)
769         // FIXME: Not consistent.
770         ImmedOffset |= 1 << NumBits;
771       else {
772         ImmedOffset = -ImmedOffset;
773         if (ImmedOffset == 0)
774           // Change the opcode back if the encoded offset is zero.
775           MI.setDesc(TII.get(positiveOffsetOpcode(NewOpc)));
776       }
777     }
778     ImmOp.ChangeToImmediate(ImmedOffset);
779     Offset &= ~(Mask*Scale);
780   }
781 
782   Offset = (isSub) ? -Offset : Offset;
783   return Offset == 0 && (FrameReg.isVirtual() || RegClass->contains(FrameReg));
784 }
785 
786 ARMCC::CondCodes llvm::getITInstrPredicate(const MachineInstr &MI,
787                                            Register &PredReg) {
788   unsigned Opc = MI.getOpcode();
789   if (Opc == ARM::tBcc || Opc == ARM::t2Bcc)
790     return ARMCC::AL;
791   return getInstrPredicate(MI, PredReg);
792 }
793 
794 int llvm::findFirstVPTPredOperandIdx(const MachineInstr &MI) {
795   const MCInstrDesc &MCID = MI.getDesc();
796 
797   for (unsigned i = 0, e = MCID.getNumOperands(); i != e; ++i)
798     if (ARM::isVpred(MCID.operands()[i].OperandType))
799       return i;
800 
801   return -1;
802 }
803 
804 ARMVCC::VPTCodes llvm::getVPTInstrPredicate(const MachineInstr &MI,
805                                             Register &PredReg) {
806   int PIdx = findFirstVPTPredOperandIdx(MI);
807   if (PIdx == -1) {
808     PredReg = 0;
809     return ARMVCC::None;
810   }
811 
812   PredReg = MI.getOperand(PIdx+1).getReg();
813   return (ARMVCC::VPTCodes)MI.getOperand(PIdx).getImm();
814 }
815 
816 void llvm::recomputeVPTBlockMask(MachineInstr &Instr) {
817   assert(isVPTOpcode(Instr.getOpcode()) && "Not a VPST or VPT Instruction!");
818 
819   MachineOperand &MaskOp = Instr.getOperand(0);
820   assert(MaskOp.isImm() && "Operand 0 is not the block mask of the VPT/VPST?!");
821 
822   MachineBasicBlock::iterator Iter = ++Instr.getIterator(),
823                               End = Instr.getParent()->end();
824 
825   while (Iter != End && Iter->isDebugInstr())
826     ++Iter;
827 
828   // Verify that the instruction after the VPT/VPST is predicated (it should
829   // be), and skip it.
830   assert(Iter != End && "Expected some instructions in any VPT block");
831   assert(
832       getVPTInstrPredicate(*Iter) == ARMVCC::Then &&
833       "VPT/VPST should be followed by an instruction with a 'then' predicate!");
834   ++Iter;
835 
836   // Iterate over the predicated instructions, updating the BlockMask as we go.
837   ARM::PredBlockMask BlockMask = ARM::PredBlockMask::T;
838   while (Iter != End) {
839     if (Iter->isDebugInstr()) {
840       ++Iter;
841       continue;
842     }
843     ARMVCC::VPTCodes Pred = getVPTInstrPredicate(*Iter);
844     if (Pred == ARMVCC::None)
845       break;
846     BlockMask = expandPredBlockMask(BlockMask, Pred);
847     ++Iter;
848   }
849 
850   // Rewrite the BlockMask.
851   MaskOp.setImm((int64_t)(BlockMask));
852 }
853