xref: /freebsd/contrib/llvm-project/llvm/lib/Target/VE/VEInstrInfo.cpp (revision 370e009188ba90c3290b1479aa06ec98b66e140a)
1 //===-- VEInstrInfo.cpp - VE 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 VE implementation of the TargetInstrInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "VEInstrInfo.h"
14 #include "VE.h"
15 #include "VEMachineFunctionInfo.h"
16 #include "VESubtarget.h"
17 #include "llvm/ADT/STLExtras.h"
18 #include "llvm/ADT/SmallVector.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineInstrBuilder.h"
21 #include "llvm/CodeGen/MachineMemOperand.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/MC/TargetRegistry.h"
24 #include "llvm/Support/CommandLine.h"
25 #include "llvm/Support/Debug.h"
26 #include "llvm/Support/ErrorHandling.h"
27 
28 #define DEBUG_TYPE "ve-instr-info"
29 
30 using namespace llvm;
31 
32 #define GET_INSTRINFO_CTOR_DTOR
33 #include "VEGenInstrInfo.inc"
34 
35 // Pin the vtable to this file.
36 void VEInstrInfo::anchor() {}
37 
38 VEInstrInfo::VEInstrInfo(VESubtarget &ST)
39     : VEGenInstrInfo(VE::ADJCALLSTACKDOWN, VE::ADJCALLSTACKUP), RI() {}
40 
41 static bool IsIntegerCC(unsigned CC) { return (CC < VECC::CC_AF); }
42 
43 static VECC::CondCode GetOppositeBranchCondition(VECC::CondCode CC) {
44   switch (CC) {
45   case VECC::CC_IG:
46     return VECC::CC_ILE;
47   case VECC::CC_IL:
48     return VECC::CC_IGE;
49   case VECC::CC_INE:
50     return VECC::CC_IEQ;
51   case VECC::CC_IEQ:
52     return VECC::CC_INE;
53   case VECC::CC_IGE:
54     return VECC::CC_IL;
55   case VECC::CC_ILE:
56     return VECC::CC_IG;
57   case VECC::CC_AF:
58     return VECC::CC_AT;
59   case VECC::CC_G:
60     return VECC::CC_LENAN;
61   case VECC::CC_L:
62     return VECC::CC_GENAN;
63   case VECC::CC_NE:
64     return VECC::CC_EQNAN;
65   case VECC::CC_EQ:
66     return VECC::CC_NENAN;
67   case VECC::CC_GE:
68     return VECC::CC_LNAN;
69   case VECC::CC_LE:
70     return VECC::CC_GNAN;
71   case VECC::CC_NUM:
72     return VECC::CC_NAN;
73   case VECC::CC_NAN:
74     return VECC::CC_NUM;
75   case VECC::CC_GNAN:
76     return VECC::CC_LE;
77   case VECC::CC_LNAN:
78     return VECC::CC_GE;
79   case VECC::CC_NENAN:
80     return VECC::CC_EQ;
81   case VECC::CC_EQNAN:
82     return VECC::CC_NE;
83   case VECC::CC_GENAN:
84     return VECC::CC_L;
85   case VECC::CC_LENAN:
86     return VECC::CC_G;
87   case VECC::CC_AT:
88     return VECC::CC_AF;
89   case VECC::UNKNOWN:
90     return VECC::UNKNOWN;
91   }
92   llvm_unreachable("Invalid cond code");
93 }
94 
95 // Treat a branch relative long always instruction as unconditional branch.
96 // For example, br.l.t and br.l.
97 static bool isUncondBranchOpcode(int Opc) {
98   using namespace llvm::VE;
99 
100 #define BRKIND(NAME) (Opc == NAME##a || Opc == NAME##a_nt || Opc == NAME##a_t)
101   // VE has other branch relative always instructions for word/double/float,
102   // but we use only long branches in our lower.  So, check it here.
103   assert(!BRKIND(BRCFW) && !BRKIND(BRCFD) && !BRKIND(BRCFS) &&
104          "Branch relative word/double/float always instructions should not be "
105          "used!");
106   return BRKIND(BRCFL);
107 #undef BRKIND
108 }
109 
110 // Treat branch relative conditional as conditional branch instructions.
111 // For example, brgt.l.t and brle.s.nt.
112 static bool isCondBranchOpcode(int Opc) {
113   using namespace llvm::VE;
114 
115 #define BRKIND(NAME)                                                           \
116   (Opc == NAME##rr || Opc == NAME##rr_nt || Opc == NAME##rr_t ||               \
117    Opc == NAME##ir || Opc == NAME##ir_nt || Opc == NAME##ir_t)
118   return BRKIND(BRCFL) || BRKIND(BRCFW) || BRKIND(BRCFD) || BRKIND(BRCFS);
119 #undef BRKIND
120 }
121 
122 // Treat branch long always instructions as indirect branch.
123 // For example, b.l.t and b.l.
124 static bool isIndirectBranchOpcode(int Opc) {
125   using namespace llvm::VE;
126 
127 #define BRKIND(NAME)                                                           \
128   (Opc == NAME##ari || Opc == NAME##ari_nt || Opc == NAME##ari_t)
129   // VE has other branch always instructions for word/double/float, but
130   // we use only long branches in our lower.  So, check it here.
131   assert(!BRKIND(BCFW) && !BRKIND(BCFD) && !BRKIND(BCFS) &&
132          "Branch word/double/float always instructions should not be used!");
133   return BRKIND(BCFL);
134 #undef BRKIND
135 }
136 
137 static void parseCondBranch(MachineInstr *LastInst, MachineBasicBlock *&Target,
138                             SmallVectorImpl<MachineOperand> &Cond) {
139   Cond.push_back(MachineOperand::CreateImm(LastInst->getOperand(0).getImm()));
140   Cond.push_back(LastInst->getOperand(1));
141   Cond.push_back(LastInst->getOperand(2));
142   Target = LastInst->getOperand(3).getMBB();
143 }
144 
145 bool VEInstrInfo::analyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB,
146                                 MachineBasicBlock *&FBB,
147                                 SmallVectorImpl<MachineOperand> &Cond,
148                                 bool AllowModify) const {
149   MachineBasicBlock::iterator I = MBB.getLastNonDebugInstr();
150   if (I == MBB.end())
151     return false;
152 
153   if (!isUnpredicatedTerminator(*I))
154     return false;
155 
156   // Get the last instruction in the block.
157   MachineInstr *LastInst = &*I;
158   unsigned LastOpc = LastInst->getOpcode();
159 
160   // If there is only one terminator instruction, process it.
161   if (I == MBB.begin() || !isUnpredicatedTerminator(*--I)) {
162     if (isUncondBranchOpcode(LastOpc)) {
163       TBB = LastInst->getOperand(0).getMBB();
164       return false;
165     }
166     if (isCondBranchOpcode(LastOpc)) {
167       // Block ends with fall-through condbranch.
168       parseCondBranch(LastInst, TBB, Cond);
169       return false;
170     }
171     return true; // Can't handle indirect branch.
172   }
173 
174   // Get the instruction before it if it is a terminator.
175   MachineInstr *SecondLastInst = &*I;
176   unsigned SecondLastOpc = SecondLastInst->getOpcode();
177 
178   // If AllowModify is true and the block ends with two or more unconditional
179   // branches, delete all but the first unconditional branch.
180   if (AllowModify && isUncondBranchOpcode(LastOpc)) {
181     while (isUncondBranchOpcode(SecondLastOpc)) {
182       LastInst->eraseFromParent();
183       LastInst = SecondLastInst;
184       LastOpc = LastInst->getOpcode();
185       if (I == MBB.begin() || !isUnpredicatedTerminator(*--I)) {
186         // Return now the only terminator is an unconditional branch.
187         TBB = LastInst->getOperand(0).getMBB();
188         return false;
189       }
190       SecondLastInst = &*I;
191       SecondLastOpc = SecondLastInst->getOpcode();
192     }
193   }
194 
195   // If there are three terminators, we don't know what sort of block this is.
196   if (SecondLastInst && I != MBB.begin() && isUnpredicatedTerminator(*--I))
197     return true;
198 
199   // If the block ends with a B and a Bcc, handle it.
200   if (isCondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
201     parseCondBranch(SecondLastInst, TBB, Cond);
202     FBB = LastInst->getOperand(0).getMBB();
203     return false;
204   }
205 
206   // If the block ends with two unconditional branches, handle it.  The second
207   // one is not executed.
208   if (isUncondBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
209     TBB = SecondLastInst->getOperand(0).getMBB();
210     return false;
211   }
212 
213   // ...likewise if it ends with an indirect branch followed by an unconditional
214   // branch.
215   if (isIndirectBranchOpcode(SecondLastOpc) && isUncondBranchOpcode(LastOpc)) {
216     I = LastInst;
217     if (AllowModify)
218       I->eraseFromParent();
219     return true;
220   }
221 
222   // Otherwise, can't handle this.
223   return true;
224 }
225 
226 unsigned VEInstrInfo::insertBranch(MachineBasicBlock &MBB,
227                                    MachineBasicBlock *TBB,
228                                    MachineBasicBlock *FBB,
229                                    ArrayRef<MachineOperand> Cond,
230                                    const DebugLoc &DL, int *BytesAdded) const {
231   assert(TBB && "insertBranch must not be told to insert a fallthrough");
232   assert((Cond.size() == 3 || Cond.size() == 0) &&
233          "VE branch conditions should have three component!");
234   assert(!BytesAdded && "code size not handled");
235   if (Cond.empty()) {
236     // Uncondition branch
237     assert(!FBB && "Unconditional branch with multiple successors!");
238     BuildMI(&MBB, DL, get(VE::BRCFLa_t))
239         .addMBB(TBB);
240     return 1;
241   }
242 
243   // Conditional branch
244   //   (BRCFir CC sy sz addr)
245   assert(Cond[0].isImm() && Cond[2].isReg() && "not implemented");
246 
247   unsigned opc[2];
248   const TargetRegisterInfo *TRI = &getRegisterInfo();
249   MachineFunction *MF = MBB.getParent();
250   const MachineRegisterInfo &MRI = MF->getRegInfo();
251   Register Reg = Cond[2].getReg();
252   if (IsIntegerCC(Cond[0].getImm())) {
253     if (TRI->getRegSizeInBits(Reg, MRI) == 32) {
254       opc[0] = VE::BRCFWir;
255       opc[1] = VE::BRCFWrr;
256     } else {
257       opc[0] = VE::BRCFLir;
258       opc[1] = VE::BRCFLrr;
259     }
260   } else {
261     if (TRI->getRegSizeInBits(Reg, MRI) == 32) {
262       opc[0] = VE::BRCFSir;
263       opc[1] = VE::BRCFSrr;
264     } else {
265       opc[0] = VE::BRCFDir;
266       opc[1] = VE::BRCFDrr;
267     }
268   }
269   if (Cond[1].isImm()) {
270       BuildMI(&MBB, DL, get(opc[0]))
271           .add(Cond[0]) // condition code
272           .add(Cond[1]) // lhs
273           .add(Cond[2]) // rhs
274           .addMBB(TBB);
275   } else {
276       BuildMI(&MBB, DL, get(opc[1]))
277           .add(Cond[0])
278           .add(Cond[1])
279           .add(Cond[2])
280           .addMBB(TBB);
281   }
282 
283   if (!FBB)
284     return 1;
285 
286   BuildMI(&MBB, DL, get(VE::BRCFLa_t))
287       .addMBB(FBB);
288   return 2;
289 }
290 
291 unsigned VEInstrInfo::removeBranch(MachineBasicBlock &MBB,
292                                    int *BytesRemoved) const {
293   assert(!BytesRemoved && "code size not handled");
294 
295   MachineBasicBlock::iterator I = MBB.end();
296   unsigned Count = 0;
297   while (I != MBB.begin()) {
298     --I;
299 
300     if (I->isDebugValue())
301       continue;
302 
303     if (!isUncondBranchOpcode(I->getOpcode()) &&
304         !isCondBranchOpcode(I->getOpcode()))
305       break; // Not a branch
306 
307     I->eraseFromParent();
308     I = MBB.end();
309     ++Count;
310   }
311   return Count;
312 }
313 
314 bool VEInstrInfo::reverseBranchCondition(
315     SmallVectorImpl<MachineOperand> &Cond) const {
316   VECC::CondCode CC = static_cast<VECC::CondCode>(Cond[0].getImm());
317   Cond[0].setImm(GetOppositeBranchCondition(CC));
318   return false;
319 }
320 
321 static bool IsAliasOfSX(Register Reg) {
322   return VE::I32RegClass.contains(Reg) || VE::I64RegClass.contains(Reg) ||
323          VE::F32RegClass.contains(Reg);
324 }
325 
326 static void copyPhysSubRegs(MachineBasicBlock &MBB,
327                             MachineBasicBlock::iterator I, const DebugLoc &DL,
328                             MCRegister DestReg, MCRegister SrcReg, bool KillSrc,
329                             const MCInstrDesc &MCID, unsigned int NumSubRegs,
330                             const unsigned *SubRegIdx,
331                             const TargetRegisterInfo *TRI) {
332   MachineInstr *MovMI = nullptr;
333 
334   for (unsigned Idx = 0; Idx != NumSubRegs; ++Idx) {
335     Register SubDest = TRI->getSubReg(DestReg, SubRegIdx[Idx]);
336     Register SubSrc = TRI->getSubReg(SrcReg, SubRegIdx[Idx]);
337     assert(SubDest && SubSrc && "Bad sub-register");
338 
339     if (MCID.getOpcode() == VE::ORri) {
340       // generate "ORri, dest, src, 0" instruction.
341       MachineInstrBuilder MIB =
342           BuildMI(MBB, I, DL, MCID, SubDest).addReg(SubSrc).addImm(0);
343       MovMI = MIB.getInstr();
344     } else if (MCID.getOpcode() == VE::ANDMmm) {
345       // generate "ANDM, dest, vm0, src" instruction.
346       MachineInstrBuilder MIB =
347           BuildMI(MBB, I, DL, MCID, SubDest).addReg(VE::VM0).addReg(SubSrc);
348       MovMI = MIB.getInstr();
349     } else {
350       llvm_unreachable("Unexpected reg-to-reg copy instruction");
351     }
352   }
353   // Add implicit super-register defs and kills to the last MovMI.
354   MovMI->addRegisterDefined(DestReg, TRI);
355   if (KillSrc)
356     MovMI->addRegisterKilled(SrcReg, TRI, true);
357 }
358 
359 void VEInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
360                               MachineBasicBlock::iterator I, const DebugLoc &DL,
361                               MCRegister DestReg, MCRegister SrcReg,
362                               bool KillSrc) const {
363 
364   if (IsAliasOfSX(SrcReg) && IsAliasOfSX(DestReg)) {
365     BuildMI(MBB, I, DL, get(VE::ORri), DestReg)
366         .addReg(SrcReg, getKillRegState(KillSrc))
367         .addImm(0);
368   } else if (VE::V64RegClass.contains(DestReg, SrcReg)) {
369     // Generate following instructions
370     //   %sw16 = LEA32zii 256
371     //   VORmvl %dest, (0)1, %src, %sw16
372     // TODO: reuse a register if vl is already assigned to a register
373     // FIXME: it would be better to scavenge a register here instead of
374     // reserving SX16 all of the time.
375     const TargetRegisterInfo *TRI = &getRegisterInfo();
376     Register TmpReg = VE::SX16;
377     Register SubTmp = TRI->getSubReg(TmpReg, VE::sub_i32);
378     BuildMI(MBB, I, DL, get(VE::LEAzii), TmpReg)
379         .addImm(0)
380         .addImm(0)
381         .addImm(256);
382     MachineInstrBuilder MIB = BuildMI(MBB, I, DL, get(VE::VORmvl), DestReg)
383                                   .addImm(M1(0)) // Represent (0)1.
384                                   .addReg(SrcReg, getKillRegState(KillSrc))
385                                   .addReg(SubTmp, getKillRegState(true));
386     MIB.getInstr()->addRegisterKilled(TmpReg, TRI, true);
387   } else if (VE::VMRegClass.contains(DestReg, SrcReg)) {
388     BuildMI(MBB, I, DL, get(VE::ANDMmm), DestReg)
389         .addReg(VE::VM0)
390         .addReg(SrcReg, getKillRegState(KillSrc));
391   } else if (VE::VM512RegClass.contains(DestReg, SrcReg)) {
392     // Use two instructions.
393     const unsigned SubRegIdx[] = {VE::sub_vm_even, VE::sub_vm_odd};
394     unsigned int NumSubRegs = 2;
395     copyPhysSubRegs(MBB, I, DL, DestReg, SrcReg, KillSrc, get(VE::ANDMmm),
396                     NumSubRegs, SubRegIdx, &getRegisterInfo());
397   } else if (VE::F128RegClass.contains(DestReg, SrcReg)) {
398     // Use two instructions.
399     const unsigned SubRegIdx[] = {VE::sub_even, VE::sub_odd};
400     unsigned int NumSubRegs = 2;
401     copyPhysSubRegs(MBB, I, DL, DestReg, SrcReg, KillSrc, get(VE::ORri),
402                     NumSubRegs, SubRegIdx, &getRegisterInfo());
403   } else {
404     const TargetRegisterInfo *TRI = &getRegisterInfo();
405     dbgs() << "Impossible reg-to-reg copy from " << printReg(SrcReg, TRI)
406            << " to " << printReg(DestReg, TRI) << "\n";
407     llvm_unreachable("Impossible reg-to-reg copy");
408   }
409 }
410 
411 /// isLoadFromStackSlot - If the specified machine instruction is a direct
412 /// load from a stack slot, return the virtual or physical register number of
413 /// the destination along with the FrameIndex of the loaded stack slot.  If
414 /// not, return 0.  This predicate must return 0 if the instruction has
415 /// any side effects other than loading from the stack slot.
416 unsigned VEInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
417                                           int &FrameIndex) const {
418   if (MI.getOpcode() == VE::LDrii ||    // I64
419       MI.getOpcode() == VE::LDLSXrii || // I32
420       MI.getOpcode() == VE::LDUrii ||   // F32
421       MI.getOpcode() == VE::LDQrii ||   // F128 (pseudo)
422       MI.getOpcode() == VE::LDVMrii ||  // VM (pseudo)
423       MI.getOpcode() == VE::LDVM512rii  // VM512 (pseudo)
424   ) {
425     if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() &&
426         MI.getOperand(2).getImm() == 0 && MI.getOperand(3).isImm() &&
427         MI.getOperand(3).getImm() == 0) {
428       FrameIndex = MI.getOperand(1).getIndex();
429       return MI.getOperand(0).getReg();
430     }
431   }
432   return 0;
433 }
434 
435 /// isStoreToStackSlot - If the specified machine instruction is a direct
436 /// store to a stack slot, return the virtual or physical register number of
437 /// the source reg along with the FrameIndex of the loaded stack slot.  If
438 /// not, return 0.  This predicate must return 0 if the instruction has
439 /// any side effects other than storing to the stack slot.
440 unsigned VEInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
441                                          int &FrameIndex) const {
442   if (MI.getOpcode() == VE::STrii ||   // I64
443       MI.getOpcode() == VE::STLrii ||  // I32
444       MI.getOpcode() == VE::STUrii ||  // F32
445       MI.getOpcode() == VE::STQrii ||  // F128 (pseudo)
446       MI.getOpcode() == VE::STVMrii || // VM (pseudo)
447       MI.getOpcode() == VE::STVM512rii // VM512 (pseudo)
448   ) {
449     if (MI.getOperand(0).isFI() && MI.getOperand(1).isImm() &&
450         MI.getOperand(1).getImm() == 0 && MI.getOperand(2).isImm() &&
451         MI.getOperand(2).getImm() == 0) {
452       FrameIndex = MI.getOperand(0).getIndex();
453       return MI.getOperand(3).getReg();
454     }
455   }
456   return 0;
457 }
458 
459 void VEInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
460                                       MachineBasicBlock::iterator I,
461                                       Register SrcReg, bool isKill, int FI,
462                                       const TargetRegisterClass *RC,
463                                       const TargetRegisterInfo *TRI) const {
464   DebugLoc DL;
465   if (I != MBB.end())
466     DL = I->getDebugLoc();
467 
468   MachineFunction *MF = MBB.getParent();
469   const MachineFrameInfo &MFI = MF->getFrameInfo();
470   MachineMemOperand *MMO = MF->getMachineMemOperand(
471       MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore,
472       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
473 
474   // On the order of operands here: think "[FrameIdx + 0] = SrcReg".
475   if (RC == &VE::I64RegClass) {
476     BuildMI(MBB, I, DL, get(VE::STrii))
477         .addFrameIndex(FI)
478         .addImm(0)
479         .addImm(0)
480         .addReg(SrcReg, getKillRegState(isKill))
481         .addMemOperand(MMO);
482   } else if (RC == &VE::I32RegClass) {
483     BuildMI(MBB, I, DL, get(VE::STLrii))
484         .addFrameIndex(FI)
485         .addImm(0)
486         .addImm(0)
487         .addReg(SrcReg, getKillRegState(isKill))
488         .addMemOperand(MMO);
489   } else if (RC == &VE::F32RegClass) {
490     BuildMI(MBB, I, DL, get(VE::STUrii))
491         .addFrameIndex(FI)
492         .addImm(0)
493         .addImm(0)
494         .addReg(SrcReg, getKillRegState(isKill))
495         .addMemOperand(MMO);
496   } else if (VE::F128RegClass.hasSubClassEq(RC)) {
497     BuildMI(MBB, I, DL, get(VE::STQrii))
498         .addFrameIndex(FI)
499         .addImm(0)
500         .addImm(0)
501         .addReg(SrcReg, getKillRegState(isKill))
502         .addMemOperand(MMO);
503   } else if (RC == &VE::VMRegClass) {
504     BuildMI(MBB, I, DL, get(VE::STVMrii))
505         .addFrameIndex(FI)
506         .addImm(0)
507         .addImm(0)
508         .addReg(SrcReg, getKillRegState(isKill))
509         .addMemOperand(MMO);
510   } else if (VE::VM512RegClass.hasSubClassEq(RC)) {
511     BuildMI(MBB, I, DL, get(VE::STVM512rii))
512         .addFrameIndex(FI)
513         .addImm(0)
514         .addImm(0)
515         .addReg(SrcReg, getKillRegState(isKill))
516         .addMemOperand(MMO);
517   } else
518     report_fatal_error("Can't store this register to stack slot");
519 }
520 
521 void VEInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
522                                        MachineBasicBlock::iterator I,
523                                        Register DestReg, int FI,
524                                        const TargetRegisterClass *RC,
525                                        const TargetRegisterInfo *TRI) const {
526   DebugLoc DL;
527   if (I != MBB.end())
528     DL = I->getDebugLoc();
529 
530   MachineFunction *MF = MBB.getParent();
531   const MachineFrameInfo &MFI = MF->getFrameInfo();
532   MachineMemOperand *MMO = MF->getMachineMemOperand(
533       MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad,
534       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
535 
536   if (RC == &VE::I64RegClass) {
537     BuildMI(MBB, I, DL, get(VE::LDrii), DestReg)
538         .addFrameIndex(FI)
539         .addImm(0)
540         .addImm(0)
541         .addMemOperand(MMO);
542   } else if (RC == &VE::I32RegClass) {
543     BuildMI(MBB, I, DL, get(VE::LDLSXrii), DestReg)
544         .addFrameIndex(FI)
545         .addImm(0)
546         .addImm(0)
547         .addMemOperand(MMO);
548   } else if (RC == &VE::F32RegClass) {
549     BuildMI(MBB, I, DL, get(VE::LDUrii), DestReg)
550         .addFrameIndex(FI)
551         .addImm(0)
552         .addImm(0)
553         .addMemOperand(MMO);
554   } else if (VE::F128RegClass.hasSubClassEq(RC)) {
555     BuildMI(MBB, I, DL, get(VE::LDQrii), DestReg)
556         .addFrameIndex(FI)
557         .addImm(0)
558         .addImm(0)
559         .addMemOperand(MMO);
560   } else if (RC == &VE::VMRegClass) {
561     BuildMI(MBB, I, DL, get(VE::LDVMrii), DestReg)
562         .addFrameIndex(FI)
563         .addImm(0)
564         .addImm(0)
565         .addMemOperand(MMO);
566   } else if (VE::VM512RegClass.hasSubClassEq(RC)) {
567     BuildMI(MBB, I, DL, get(VE::LDVM512rii), DestReg)
568         .addFrameIndex(FI)
569         .addImm(0)
570         .addImm(0)
571         .addMemOperand(MMO);
572   } else
573     report_fatal_error("Can't load this register from stack slot");
574 }
575 
576 bool VEInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
577                                 Register Reg, MachineRegisterInfo *MRI) const {
578   LLVM_DEBUG(dbgs() << "FoldImmediate\n");
579 
580   LLVM_DEBUG(dbgs() << "checking DefMI\n");
581   int64_t ImmVal;
582   switch (DefMI.getOpcode()) {
583   default:
584     return false;
585   case VE::ORim:
586     // General move small immediate instruction on VE.
587     LLVM_DEBUG(dbgs() << "checking ORim\n");
588     LLVM_DEBUG(DefMI.dump());
589     // FIXME: We may need to support FPImm too.
590     assert(DefMI.getOperand(1).isImm());
591     assert(DefMI.getOperand(2).isImm());
592     ImmVal =
593         DefMI.getOperand(1).getImm() + mimm2Val(DefMI.getOperand(2).getImm());
594     LLVM_DEBUG(dbgs() << "ImmVal is " << ImmVal << "\n");
595     break;
596   case VE::LEAzii:
597     // General move immediate instruction on VE.
598     LLVM_DEBUG(dbgs() << "checking LEAzii\n");
599     LLVM_DEBUG(DefMI.dump());
600     // FIXME: We may need to support FPImm too.
601     assert(DefMI.getOperand(2).isImm());
602     if (!DefMI.getOperand(3).isImm())
603       // LEAzii may refer label
604       return false;
605     ImmVal = DefMI.getOperand(2).getImm() + DefMI.getOperand(3).getImm();
606     LLVM_DEBUG(dbgs() << "ImmVal is " << ImmVal << "\n");
607     break;
608   }
609 
610   // Try to fold like below:
611   //   %1:i64 = ORim 0, 0(1)
612   //   %2:i64 = CMPSLrr %0, %1
613   // To
614   //   %2:i64 = CMPSLrm %0, 0(1)
615   //
616   // Another example:
617   //   %1:i64 = ORim 6, 0(1)
618   //   %2:i64 = CMPSLrr %1, %0
619   // To
620   //   %2:i64 = CMPSLir 6, %0
621   //
622   // Support commutable instructions like below:
623   //   %1:i64 = ORim 6, 0(1)
624   //   %2:i64 = ADDSLrr %1, %0
625   // To
626   //   %2:i64 = ADDSLri %0, 6
627   //
628   // FIXME: Need to support i32.  Current implementtation requires
629   //        EXTRACT_SUBREG, so input has following COPY and it avoids folding:
630   //   %1:i64 = ORim 6, 0(1)
631   //   %2:i32 = COPY %1.sub_i32
632   //   %3:i32 = ADDSWSXrr %0, %2
633   // FIXME: Need to support shift, cmov, and more instructions.
634   // FIXME: Need to support lvl too, but LVLGen runs after peephole-opt.
635 
636   LLVM_DEBUG(dbgs() << "checking UseMI\n");
637   LLVM_DEBUG(UseMI.dump());
638   unsigned NewUseOpcSImm7;
639   unsigned NewUseOpcMImm;
640   enum InstType {
641     rr2ri_rm, // rr -> ri or rm, commutable
642     rr2ir_rm, // rr -> ir or rm
643   } InstType;
644 
645   using namespace llvm::VE;
646 #define INSTRKIND(NAME)                                                        \
647   case NAME##rr:                                                               \
648     NewUseOpcSImm7 = NAME##ri;                                                 \
649     NewUseOpcMImm = NAME##rm;                                                  \
650     InstType = rr2ri_rm;                                                       \
651     break
652 #define NCINSTRKIND(NAME)                                                      \
653   case NAME##rr:                                                               \
654     NewUseOpcSImm7 = NAME##ir;                                                 \
655     NewUseOpcMImm = NAME##rm;                                                  \
656     InstType = rr2ir_rm;                                                       \
657     break
658 
659   switch (UseMI.getOpcode()) {
660   default:
661     return false;
662 
663     INSTRKIND(ADDUL);
664     INSTRKIND(ADDSWSX);
665     INSTRKIND(ADDSWZX);
666     INSTRKIND(ADDSL);
667     NCINSTRKIND(SUBUL);
668     NCINSTRKIND(SUBSWSX);
669     NCINSTRKIND(SUBSWZX);
670     NCINSTRKIND(SUBSL);
671     INSTRKIND(MULUL);
672     INSTRKIND(MULSWSX);
673     INSTRKIND(MULSWZX);
674     INSTRKIND(MULSL);
675     NCINSTRKIND(DIVUL);
676     NCINSTRKIND(DIVSWSX);
677     NCINSTRKIND(DIVSWZX);
678     NCINSTRKIND(DIVSL);
679     NCINSTRKIND(CMPUL);
680     NCINSTRKIND(CMPSWSX);
681     NCINSTRKIND(CMPSWZX);
682     NCINSTRKIND(CMPSL);
683     INSTRKIND(MAXSWSX);
684     INSTRKIND(MAXSWZX);
685     INSTRKIND(MAXSL);
686     INSTRKIND(MINSWSX);
687     INSTRKIND(MINSWZX);
688     INSTRKIND(MINSL);
689     INSTRKIND(AND);
690     INSTRKIND(OR);
691     INSTRKIND(XOR);
692     INSTRKIND(EQV);
693     NCINSTRKIND(NND);
694     NCINSTRKIND(MRG);
695   }
696 
697 #undef INSTRKIND
698 
699   unsigned NewUseOpc;
700   unsigned UseIdx;
701   bool Commute = false;
702   LLVM_DEBUG(dbgs() << "checking UseMI operands\n");
703   switch (InstType) {
704   case rr2ri_rm:
705     UseIdx = 2;
706     if (UseMI.getOperand(1).getReg() == Reg) {
707       Commute = true;
708     } else {
709       assert(UseMI.getOperand(2).getReg() == Reg);
710     }
711     if (isInt<7>(ImmVal)) {
712       // This ImmVal matches to SImm7 slot, so change UseOpc to an instruction
713       // holds a simm7 slot.
714       NewUseOpc = NewUseOpcSImm7;
715     } else if (isMImmVal(ImmVal)) {
716       // Similarly, change UseOpc to an instruction holds a mimm slot.
717       NewUseOpc = NewUseOpcMImm;
718       ImmVal = val2MImm(ImmVal);
719     } else
720       return false;
721     break;
722   case rr2ir_rm:
723     if (UseMI.getOperand(1).getReg() == Reg) {
724       // Check immediate value whether it matchs to the UseMI instruction.
725       if (!isInt<7>(ImmVal))
726         return false;
727       NewUseOpc = NewUseOpcSImm7;
728       UseIdx = 1;
729     } else {
730       assert(UseMI.getOperand(2).getReg() == Reg);
731       // Check immediate value whether it matchs to the UseMI instruction.
732       if (!isMImmVal(ImmVal))
733         return false;
734       NewUseOpc = NewUseOpcMImm;
735       ImmVal = val2MImm(ImmVal);
736       UseIdx = 2;
737     }
738     break;
739   }
740 
741   LLVM_DEBUG(dbgs() << "modifying UseMI\n");
742   bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
743   UseMI.setDesc(get(NewUseOpc));
744   if (Commute) {
745     UseMI.getOperand(1).setReg(UseMI.getOperand(UseIdx).getReg());
746   }
747   UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal);
748   if (DeleteDef)
749     DefMI.eraseFromParent();
750 
751   return true;
752 }
753 
754 Register VEInstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
755   VEMachineFunctionInfo *VEFI = MF->getInfo<VEMachineFunctionInfo>();
756   Register GlobalBaseReg = VEFI->getGlobalBaseReg();
757   if (GlobalBaseReg != 0)
758     return GlobalBaseReg;
759 
760   // We use %s15 (%got) as a global base register
761   GlobalBaseReg = VE::SX15;
762 
763   // Insert a pseudo instruction to set the GlobalBaseReg into the first
764   // MBB of the function
765   MachineBasicBlock &FirstMBB = MF->front();
766   MachineBasicBlock::iterator MBBI = FirstMBB.begin();
767   DebugLoc dl;
768   BuildMI(FirstMBB, MBBI, dl, get(VE::GETGOT), GlobalBaseReg);
769   VEFI->setGlobalBaseReg(GlobalBaseReg);
770   return GlobalBaseReg;
771 }
772 
773 static Register getVM512Upper(Register reg) {
774   return (reg - VE::VMP0) * 2 + VE::VM0;
775 }
776 
777 static Register getVM512Lower(Register reg) { return getVM512Upper(reg) + 1; }
778 
779 // Expand pseudo logical vector instructions for VM512 registers.
780 static void expandPseudoLogM(MachineInstr &MI, const MCInstrDesc &MCID) {
781   MachineBasicBlock *MBB = MI.getParent();
782   DebugLoc DL = MI.getDebugLoc();
783 
784   Register VMXu = getVM512Upper(MI.getOperand(0).getReg());
785   Register VMXl = getVM512Lower(MI.getOperand(0).getReg());
786   Register VMYu = getVM512Upper(MI.getOperand(1).getReg());
787   Register VMYl = getVM512Lower(MI.getOperand(1).getReg());
788 
789   switch (MI.getOpcode()) {
790   default: {
791     Register VMZu = getVM512Upper(MI.getOperand(2).getReg());
792     Register VMZl = getVM512Lower(MI.getOperand(2).getReg());
793     BuildMI(*MBB, MI, DL, MCID).addDef(VMXu).addUse(VMYu).addUse(VMZu);
794     BuildMI(*MBB, MI, DL, MCID).addDef(VMXl).addUse(VMYl).addUse(VMZl);
795     break;
796   }
797   case VE::NEGMy:
798     BuildMI(*MBB, MI, DL, MCID).addDef(VMXu).addUse(VMYu);
799     BuildMI(*MBB, MI, DL, MCID).addDef(VMXl).addUse(VMYl);
800     break;
801   }
802   MI.eraseFromParent();
803 }
804 
805 static void addOperandsForVFMK(MachineInstrBuilder &MIB, MachineInstr &MI,
806                                bool Upper) {
807   // VM512
808   MIB.addReg(Upper ? getVM512Upper(MI.getOperand(0).getReg())
809                    : getVM512Lower(MI.getOperand(0).getReg()));
810 
811   switch (MI.getNumExplicitOperands()) {
812   default:
813     report_fatal_error("unexpected number of operands for pvfmk");
814   case 2: // _Ml: VM512, VL
815     // VL
816     MIB.addReg(MI.getOperand(1).getReg());
817     break;
818   case 4: // _Mvl: VM512, CC, VR, VL
819     // CC
820     MIB.addImm(MI.getOperand(1).getImm());
821     // VR
822     MIB.addReg(MI.getOperand(2).getReg());
823     // VL
824     MIB.addReg(MI.getOperand(3).getReg());
825     break;
826   case 5: // _MvMl: VM512, CC, VR, VM512, VL
827     // CC
828     MIB.addImm(MI.getOperand(1).getImm());
829     // VR
830     MIB.addReg(MI.getOperand(2).getReg());
831     // VM512
832     MIB.addReg(Upper ? getVM512Upper(MI.getOperand(3).getReg())
833                      : getVM512Lower(MI.getOperand(3).getReg()));
834     // VL
835     MIB.addReg(MI.getOperand(4).getReg());
836     break;
837   }
838 }
839 
840 static void expandPseudoVFMK(const TargetInstrInfo &TI, MachineInstr &MI) {
841   // replace to pvfmk.w.up and pvfmk.w.lo
842   // replace to pvfmk.s.up and pvfmk.s.lo
843 
844   static const std::pair<unsigned, std::pair<unsigned, unsigned>> VFMKMap[] = {
845       {VE::VFMKyal, {VE::VFMKLal, VE::VFMKLal}},
846       {VE::VFMKynal, {VE::VFMKLnal, VE::VFMKLnal}},
847       {VE::VFMKWyvl, {VE::PVFMKWUPvl, VE::PVFMKWLOvl}},
848       {VE::VFMKWyvyl, {VE::PVFMKWUPvml, VE::PVFMKWLOvml}},
849       {VE::VFMKSyvl, {VE::PVFMKSUPvl, VE::PVFMKSLOvl}},
850       {VE::VFMKSyvyl, {VE::PVFMKSUPvml, VE::PVFMKSLOvml}},
851   };
852 
853   unsigned Opcode = MI.getOpcode();
854 
855   const auto *Found =
856       llvm::find_if(VFMKMap, [&](auto P) { return P.first == Opcode; });
857   if (Found == std::end(VFMKMap))
858     report_fatal_error("unexpected opcode for pseudo vfmk");
859 
860   unsigned OpcodeUpper = (*Found).second.first;
861   unsigned OpcodeLower = (*Found).second.second;
862 
863   MachineBasicBlock *MBB = MI.getParent();
864   DebugLoc DL = MI.getDebugLoc();
865 
866   MachineInstrBuilder Bu = BuildMI(*MBB, MI, DL, TI.get(OpcodeUpper));
867   addOperandsForVFMK(Bu, MI, /* Upper */ true);
868   MachineInstrBuilder Bl = BuildMI(*MBB, MI, DL, TI.get(OpcodeLower));
869   addOperandsForVFMK(Bl, MI, /* Upper */ false);
870 
871   MI.eraseFromParent();
872 }
873 
874 bool VEInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
875   switch (MI.getOpcode()) {
876   case VE::EXTEND_STACK: {
877     return expandExtendStackPseudo(MI);
878   }
879   case VE::EXTEND_STACK_GUARD: {
880     MI.eraseFromParent(); // The pseudo instruction is gone now.
881     return true;
882   }
883   case VE::GETSTACKTOP: {
884     return expandGetStackTopPseudo(MI);
885   }
886 
887   case VE::ANDMyy:
888     expandPseudoLogM(MI, get(VE::ANDMmm));
889     return true;
890   case VE::ORMyy:
891     expandPseudoLogM(MI, get(VE::ORMmm));
892     return true;
893   case VE::XORMyy:
894     expandPseudoLogM(MI, get(VE::XORMmm));
895     return true;
896   case VE::EQVMyy:
897     expandPseudoLogM(MI, get(VE::EQVMmm));
898     return true;
899   case VE::NNDMyy:
900     expandPseudoLogM(MI, get(VE::NNDMmm));
901     return true;
902   case VE::NEGMy:
903     expandPseudoLogM(MI, get(VE::NEGMm));
904     return true;
905 
906   case VE::LVMyir:
907   case VE::LVMyim:
908   case VE::LVMyir_y:
909   case VE::LVMyim_y: {
910     Register VMXu = getVM512Upper(MI.getOperand(0).getReg());
911     Register VMXl = getVM512Lower(MI.getOperand(0).getReg());
912     int64_t Imm = MI.getOperand(1).getImm();
913     bool IsSrcReg =
914         MI.getOpcode() == VE::LVMyir || MI.getOpcode() == VE::LVMyir_y;
915     Register Src = IsSrcReg ? MI.getOperand(2).getReg() : VE::NoRegister;
916     int64_t MImm = IsSrcReg ? 0 : MI.getOperand(2).getImm();
917     bool KillSrc = IsSrcReg ? MI.getOperand(2).isKill() : false;
918     Register VMX = VMXl;
919     if (Imm >= 4) {
920       VMX = VMXu;
921       Imm -= 4;
922     }
923     MachineBasicBlock *MBB = MI.getParent();
924     DebugLoc DL = MI.getDebugLoc();
925     switch (MI.getOpcode()) {
926     case VE::LVMyir:
927       BuildMI(*MBB, MI, DL, get(VE::LVMir))
928           .addDef(VMX)
929           .addImm(Imm)
930           .addReg(Src, getKillRegState(KillSrc));
931       break;
932     case VE::LVMyim:
933       BuildMI(*MBB, MI, DL, get(VE::LVMim))
934           .addDef(VMX)
935           .addImm(Imm)
936           .addImm(MImm);
937       break;
938     case VE::LVMyir_y:
939       assert(MI.getOperand(0).getReg() == MI.getOperand(3).getReg() &&
940              "LVMyir_y has different register in 3rd operand");
941       BuildMI(*MBB, MI, DL, get(VE::LVMir_m))
942           .addDef(VMX)
943           .addImm(Imm)
944           .addReg(Src, getKillRegState(KillSrc))
945           .addReg(VMX);
946       break;
947     case VE::LVMyim_y:
948       assert(MI.getOperand(0).getReg() == MI.getOperand(3).getReg() &&
949              "LVMyim_y has different register in 3rd operand");
950       BuildMI(*MBB, MI, DL, get(VE::LVMim_m))
951           .addDef(VMX)
952           .addImm(Imm)
953           .addImm(MImm)
954           .addReg(VMX);
955       break;
956     }
957     MI.eraseFromParent();
958     return true;
959   }
960   case VE::SVMyi: {
961     Register Dest = MI.getOperand(0).getReg();
962     Register VMZu = getVM512Upper(MI.getOperand(1).getReg());
963     Register VMZl = getVM512Lower(MI.getOperand(1).getReg());
964     bool KillSrc = MI.getOperand(1).isKill();
965     int64_t Imm = MI.getOperand(2).getImm();
966     Register VMZ = VMZl;
967     if (Imm >= 4) {
968       VMZ = VMZu;
969       Imm -= 4;
970     }
971     MachineBasicBlock *MBB = MI.getParent();
972     DebugLoc DL = MI.getDebugLoc();
973     MachineInstrBuilder MIB =
974         BuildMI(*MBB, MI, DL, get(VE::SVMmi), Dest).addReg(VMZ).addImm(Imm);
975     MachineInstr *Inst = MIB.getInstr();
976     if (KillSrc) {
977       const TargetRegisterInfo *TRI = &getRegisterInfo();
978       Inst->addRegisterKilled(MI.getOperand(1).getReg(), TRI, true);
979     }
980     MI.eraseFromParent();
981     return true;
982   }
983   case VE::VFMKyal:
984   case VE::VFMKynal:
985   case VE::VFMKWyvl:
986   case VE::VFMKWyvyl:
987   case VE::VFMKSyvl:
988   case VE::VFMKSyvyl:
989     expandPseudoVFMK(*this, MI);
990     return true;
991   }
992   return false;
993 }
994 
995 bool VEInstrInfo::expandExtendStackPseudo(MachineInstr &MI) const {
996   MachineBasicBlock &MBB = *MI.getParent();
997   MachineFunction &MF = *MBB.getParent();
998   const VESubtarget &STI = MF.getSubtarget<VESubtarget>();
999   const VEInstrInfo &TII = *STI.getInstrInfo();
1000   DebugLoc dl = MBB.findDebugLoc(MI);
1001 
1002   // Create following instructions and multiple basic blocks.
1003   //
1004   // thisBB:
1005   //   brge.l.t %sp, %sl, sinkBB
1006   // syscallBB:
1007   //   ld      %s61, 0x18(, %tp)        // load param area
1008   //   or      %s62, 0, %s0             // spill the value of %s0
1009   //   lea     %s63, 0x13b              // syscall # of grow
1010   //   shm.l   %s63, 0x0(%s61)          // store syscall # at addr:0
1011   //   shm.l   %sl, 0x8(%s61)           // store old limit at addr:8
1012   //   shm.l   %sp, 0x10(%s61)          // store new limit at addr:16
1013   //   monc                             // call monitor
1014   //   or      %s0, 0, %s62             // restore the value of %s0
1015   // sinkBB:
1016 
1017   // Create new MBB
1018   MachineBasicBlock *BB = &MBB;
1019   const BasicBlock *LLVM_BB = BB->getBasicBlock();
1020   MachineBasicBlock *syscallMBB = MF.CreateMachineBasicBlock(LLVM_BB);
1021   MachineBasicBlock *sinkMBB = MF.CreateMachineBasicBlock(LLVM_BB);
1022   MachineFunction::iterator It = ++(BB->getIterator());
1023   MF.insert(It, syscallMBB);
1024   MF.insert(It, sinkMBB);
1025 
1026   // Transfer the remainder of BB and its successor edges to sinkMBB.
1027   sinkMBB->splice(sinkMBB->begin(), BB,
1028                   std::next(std::next(MachineBasicBlock::iterator(MI))),
1029                   BB->end());
1030   sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1031 
1032   // Next, add the true and fallthrough blocks as its successors.
1033   BB->addSuccessor(syscallMBB);
1034   BB->addSuccessor(sinkMBB);
1035   BuildMI(BB, dl, TII.get(VE::BRCFLrr_t))
1036       .addImm(VECC::CC_IGE)
1037       .addReg(VE::SX11) // %sp
1038       .addReg(VE::SX8)  // %sl
1039       .addMBB(sinkMBB);
1040 
1041   BB = syscallMBB;
1042 
1043   // Update machine-CFG edges
1044   BB->addSuccessor(sinkMBB);
1045 
1046   BuildMI(BB, dl, TII.get(VE::LDrii), VE::SX61)
1047       .addReg(VE::SX14)
1048       .addImm(0)
1049       .addImm(0x18);
1050   BuildMI(BB, dl, TII.get(VE::ORri), VE::SX62)
1051       .addReg(VE::SX0)
1052       .addImm(0);
1053   BuildMI(BB, dl, TII.get(VE::LEAzii), VE::SX63)
1054       .addImm(0)
1055       .addImm(0)
1056       .addImm(0x13b);
1057   BuildMI(BB, dl, TII.get(VE::SHMLri))
1058       .addReg(VE::SX61)
1059       .addImm(0)
1060       .addReg(VE::SX63);
1061   BuildMI(BB, dl, TII.get(VE::SHMLri))
1062       .addReg(VE::SX61)
1063       .addImm(8)
1064       .addReg(VE::SX8);
1065   BuildMI(BB, dl, TII.get(VE::SHMLri))
1066       .addReg(VE::SX61)
1067       .addImm(16)
1068       .addReg(VE::SX11);
1069   BuildMI(BB, dl, TII.get(VE::MONC));
1070 
1071   BuildMI(BB, dl, TII.get(VE::ORri), VE::SX0)
1072       .addReg(VE::SX62)
1073       .addImm(0);
1074 
1075   MI.eraseFromParent(); // The pseudo instruction is gone now.
1076   return true;
1077 }
1078 
1079 bool VEInstrInfo::expandGetStackTopPseudo(MachineInstr &MI) const {
1080   MachineBasicBlock *MBB = MI.getParent();
1081   MachineFunction &MF = *MBB->getParent();
1082   const VESubtarget &STI = MF.getSubtarget<VESubtarget>();
1083   const VEInstrInfo &TII = *STI.getInstrInfo();
1084   DebugLoc DL = MBB->findDebugLoc(MI);
1085 
1086   // Create following instruction
1087   //
1088   //   dst = %sp + target specific frame + the size of parameter area
1089 
1090   const MachineFrameInfo &MFI = MF.getFrameInfo();
1091   const VEFrameLowering &TFL = *STI.getFrameLowering();
1092 
1093   // The VE ABI requires a reserved area at the top of stack as described
1094   // in VEFrameLowering.cpp.  So, we adjust it here.
1095   unsigned NumBytes = STI.getAdjustedFrameSize(0);
1096 
1097   // Also adds the size of parameter area.
1098   if (MFI.adjustsStack() && TFL.hasReservedCallFrame(MF))
1099     NumBytes += MFI.getMaxCallFrameSize();
1100 
1101   BuildMI(*MBB, MI, DL, TII.get(VE::LEArii))
1102       .addDef(MI.getOperand(0).getReg())
1103       .addReg(VE::SX11)
1104       .addImm(0)
1105       .addImm(NumBytes);
1106 
1107   MI.eraseFromParent(); // The pseudo instruction is gone now.
1108   return true;
1109 }
1110