xref: /freebsd/contrib/llvm-project/llvm/lib/Target/VE/VEInstrInfo.cpp (revision 8ddb146abcdf061be9f2c0db7e391697dafad85c)
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   ) {
423     if (MI.getOperand(1).isFI() && MI.getOperand(2).isImm() &&
424         MI.getOperand(2).getImm() == 0 && MI.getOperand(3).isImm() &&
425         MI.getOperand(3).getImm() == 0) {
426       FrameIndex = MI.getOperand(1).getIndex();
427       return MI.getOperand(0).getReg();
428     }
429   }
430   return 0;
431 }
432 
433 /// isStoreToStackSlot - If the specified machine instruction is a direct
434 /// store to a stack slot, return the virtual or physical register number of
435 /// the source reg along with the FrameIndex of the loaded stack slot.  If
436 /// not, return 0.  This predicate must return 0 if the instruction has
437 /// any side effects other than storing to the stack slot.
438 unsigned VEInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
439                                          int &FrameIndex) const {
440   if (MI.getOpcode() == VE::STrii ||  // I64
441       MI.getOpcode() == VE::STLrii || // I32
442       MI.getOpcode() == VE::STUrii || // F32
443       MI.getOpcode() == VE::STQrii    // F128 (pseudo)
444   ) {
445     if (MI.getOperand(0).isFI() && MI.getOperand(1).isImm() &&
446         MI.getOperand(1).getImm() == 0 && MI.getOperand(2).isImm() &&
447         MI.getOperand(2).getImm() == 0) {
448       FrameIndex = MI.getOperand(0).getIndex();
449       return MI.getOperand(3).getReg();
450     }
451   }
452   return 0;
453 }
454 
455 void VEInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
456                                       MachineBasicBlock::iterator I,
457                                       Register SrcReg, bool isKill, int FI,
458                                       const TargetRegisterClass *RC,
459                                       const TargetRegisterInfo *TRI) const {
460   DebugLoc DL;
461   if (I != MBB.end())
462     DL = I->getDebugLoc();
463 
464   MachineFunction *MF = MBB.getParent();
465   const MachineFrameInfo &MFI = MF->getFrameInfo();
466   MachineMemOperand *MMO = MF->getMachineMemOperand(
467       MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOStore,
468       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
469 
470   // On the order of operands here: think "[FrameIdx + 0] = SrcReg".
471   if (RC == &VE::I64RegClass) {
472     BuildMI(MBB, I, DL, get(VE::STrii))
473         .addFrameIndex(FI)
474         .addImm(0)
475         .addImm(0)
476         .addReg(SrcReg, getKillRegState(isKill))
477         .addMemOperand(MMO);
478   } else if (RC == &VE::I32RegClass) {
479     BuildMI(MBB, I, DL, get(VE::STLrii))
480         .addFrameIndex(FI)
481         .addImm(0)
482         .addImm(0)
483         .addReg(SrcReg, getKillRegState(isKill))
484         .addMemOperand(MMO);
485   } else if (RC == &VE::F32RegClass) {
486     BuildMI(MBB, I, DL, get(VE::STUrii))
487         .addFrameIndex(FI)
488         .addImm(0)
489         .addImm(0)
490         .addReg(SrcReg, getKillRegState(isKill))
491         .addMemOperand(MMO);
492   } else if (VE::F128RegClass.hasSubClassEq(RC)) {
493     BuildMI(MBB, I, DL, get(VE::STQrii))
494         .addFrameIndex(FI)
495         .addImm(0)
496         .addImm(0)
497         .addReg(SrcReg, getKillRegState(isKill))
498         .addMemOperand(MMO);
499   } else
500     report_fatal_error("Can't store this register to stack slot");
501 }
502 
503 void VEInstrInfo::loadRegFromStackSlot(MachineBasicBlock &MBB,
504                                        MachineBasicBlock::iterator I,
505                                        Register DestReg, int FI,
506                                        const TargetRegisterClass *RC,
507                                        const TargetRegisterInfo *TRI) const {
508   DebugLoc DL;
509   if (I != MBB.end())
510     DL = I->getDebugLoc();
511 
512   MachineFunction *MF = MBB.getParent();
513   const MachineFrameInfo &MFI = MF->getFrameInfo();
514   MachineMemOperand *MMO = MF->getMachineMemOperand(
515       MachinePointerInfo::getFixedStack(*MF, FI), MachineMemOperand::MOLoad,
516       MFI.getObjectSize(FI), MFI.getObjectAlign(FI));
517 
518   if (RC == &VE::I64RegClass) {
519     BuildMI(MBB, I, DL, get(VE::LDrii), DestReg)
520         .addFrameIndex(FI)
521         .addImm(0)
522         .addImm(0)
523         .addMemOperand(MMO);
524   } else if (RC == &VE::I32RegClass) {
525     BuildMI(MBB, I, DL, get(VE::LDLSXrii), DestReg)
526         .addFrameIndex(FI)
527         .addImm(0)
528         .addImm(0)
529         .addMemOperand(MMO);
530   } else if (RC == &VE::F32RegClass) {
531     BuildMI(MBB, I, DL, get(VE::LDUrii), DestReg)
532         .addFrameIndex(FI)
533         .addImm(0)
534         .addImm(0)
535         .addMemOperand(MMO);
536   } else if (VE::F128RegClass.hasSubClassEq(RC)) {
537     BuildMI(MBB, I, DL, get(VE::LDQrii), DestReg)
538         .addFrameIndex(FI)
539         .addImm(0)
540         .addImm(0)
541         .addMemOperand(MMO);
542   } else
543     report_fatal_error("Can't load this register from stack slot");
544 }
545 
546 bool VEInstrInfo::FoldImmediate(MachineInstr &UseMI, MachineInstr &DefMI,
547                                 Register Reg, MachineRegisterInfo *MRI) const {
548   LLVM_DEBUG(dbgs() << "FoldImmediate\n");
549 
550   LLVM_DEBUG(dbgs() << "checking DefMI\n");
551   int64_t ImmVal;
552   switch (DefMI.getOpcode()) {
553   default:
554     return false;
555   case VE::ORim:
556     // General move small immediate instruction on VE.
557     LLVM_DEBUG(dbgs() << "checking ORim\n");
558     LLVM_DEBUG(DefMI.dump());
559     // FIXME: We may need to support FPImm too.
560     assert(DefMI.getOperand(1).isImm());
561     assert(DefMI.getOperand(2).isImm());
562     ImmVal =
563         DefMI.getOperand(1).getImm() + mimm2Val(DefMI.getOperand(2).getImm());
564     LLVM_DEBUG(dbgs() << "ImmVal is " << ImmVal << "\n");
565     break;
566   case VE::LEAzii:
567     // General move immediate instruction on VE.
568     LLVM_DEBUG(dbgs() << "checking LEAzii\n");
569     LLVM_DEBUG(DefMI.dump());
570     // FIXME: We may need to support FPImm too.
571     assert(DefMI.getOperand(2).isImm());
572     if (!DefMI.getOperand(3).isImm())
573       // LEAzii may refer label
574       return false;
575     ImmVal = DefMI.getOperand(2).getImm() + DefMI.getOperand(3).getImm();
576     LLVM_DEBUG(dbgs() << "ImmVal is " << ImmVal << "\n");
577     break;
578   }
579 
580   // Try to fold like below:
581   //   %1:i64 = ORim 0, 0(1)
582   //   %2:i64 = CMPSLrr %0, %1
583   // To
584   //   %2:i64 = CMPSLrm %0, 0(1)
585   //
586   // Another example:
587   //   %1:i64 = ORim 6, 0(1)
588   //   %2:i64 = CMPSLrr %1, %0
589   // To
590   //   %2:i64 = CMPSLir 6, %0
591   //
592   // Support commutable instructions like below:
593   //   %1:i64 = ORim 6, 0(1)
594   //   %2:i64 = ADDSLrr %1, %0
595   // To
596   //   %2:i64 = ADDSLri %0, 6
597   //
598   // FIXME: Need to support i32.  Current implementtation requires
599   //        EXTRACT_SUBREG, so input has following COPY and it avoids folding:
600   //   %1:i64 = ORim 6, 0(1)
601   //   %2:i32 = COPY %1.sub_i32
602   //   %3:i32 = ADDSWSXrr %0, %2
603   // FIXME: Need to support shift, cmov, and more instructions.
604   // FIXME: Need to support lvl too, but LVLGen runs after peephole-opt.
605 
606   LLVM_DEBUG(dbgs() << "checking UseMI\n");
607   LLVM_DEBUG(UseMI.dump());
608   unsigned NewUseOpcSImm7;
609   unsigned NewUseOpcMImm;
610   enum InstType {
611     rr2ri_rm, // rr -> ri or rm, commutable
612     rr2ir_rm, // rr -> ir or rm
613   } InstType;
614 
615   using namespace llvm::VE;
616 #define INSTRKIND(NAME)                                                        \
617   case NAME##rr:                                                               \
618     NewUseOpcSImm7 = NAME##ri;                                                 \
619     NewUseOpcMImm = NAME##rm;                                                  \
620     InstType = rr2ri_rm;                                                       \
621     break
622 #define NCINSTRKIND(NAME)                                                      \
623   case NAME##rr:                                                               \
624     NewUseOpcSImm7 = NAME##ir;                                                 \
625     NewUseOpcMImm = NAME##rm;                                                  \
626     InstType = rr2ir_rm;                                                       \
627     break
628 
629   switch (UseMI.getOpcode()) {
630   default:
631     return false;
632 
633     INSTRKIND(ADDUL);
634     INSTRKIND(ADDSWSX);
635     INSTRKIND(ADDSWZX);
636     INSTRKIND(ADDSL);
637     NCINSTRKIND(SUBUL);
638     NCINSTRKIND(SUBSWSX);
639     NCINSTRKIND(SUBSWZX);
640     NCINSTRKIND(SUBSL);
641     INSTRKIND(MULUL);
642     INSTRKIND(MULSWSX);
643     INSTRKIND(MULSWZX);
644     INSTRKIND(MULSL);
645     NCINSTRKIND(DIVUL);
646     NCINSTRKIND(DIVSWSX);
647     NCINSTRKIND(DIVSWZX);
648     NCINSTRKIND(DIVSL);
649     NCINSTRKIND(CMPUL);
650     NCINSTRKIND(CMPSWSX);
651     NCINSTRKIND(CMPSWZX);
652     NCINSTRKIND(CMPSL);
653     INSTRKIND(MAXSWSX);
654     INSTRKIND(MAXSWZX);
655     INSTRKIND(MAXSL);
656     INSTRKIND(MINSWSX);
657     INSTRKIND(MINSWZX);
658     INSTRKIND(MINSL);
659     INSTRKIND(AND);
660     INSTRKIND(OR);
661     INSTRKIND(XOR);
662     INSTRKIND(EQV);
663     NCINSTRKIND(NND);
664     NCINSTRKIND(MRG);
665   }
666 
667 #undef INSTRKIND
668 
669   unsigned NewUseOpc;
670   unsigned UseIdx;
671   bool Commute = false;
672   LLVM_DEBUG(dbgs() << "checking UseMI operands\n");
673   switch (InstType) {
674   case rr2ri_rm:
675     UseIdx = 2;
676     if (UseMI.getOperand(1).getReg() == Reg) {
677       Commute = true;
678     } else {
679       assert(UseMI.getOperand(2).getReg() == Reg);
680     }
681     if (isInt<7>(ImmVal)) {
682       // This ImmVal matches to SImm7 slot, so change UseOpc to an instruction
683       // holds a simm7 slot.
684       NewUseOpc = NewUseOpcSImm7;
685     } else if (isMImmVal(ImmVal)) {
686       // Similarly, change UseOpc to an instruction holds a mimm slot.
687       NewUseOpc = NewUseOpcMImm;
688       ImmVal = val2MImm(ImmVal);
689     } else
690       return false;
691     break;
692   case rr2ir_rm:
693     if (UseMI.getOperand(1).getReg() == Reg) {
694       // Check immediate value whether it matchs to the UseMI instruction.
695       if (!isInt<7>(ImmVal))
696         return false;
697       NewUseOpc = NewUseOpcSImm7;
698       UseIdx = 1;
699     } else {
700       assert(UseMI.getOperand(2).getReg() == Reg);
701       // Check immediate value whether it matchs to the UseMI instruction.
702       if (!isMImmVal(ImmVal))
703         return false;
704       NewUseOpc = NewUseOpcMImm;
705       ImmVal = val2MImm(ImmVal);
706       UseIdx = 2;
707     }
708     break;
709   }
710 
711   LLVM_DEBUG(dbgs() << "modifying UseMI\n");
712   bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
713   UseMI.setDesc(get(NewUseOpc));
714   if (Commute) {
715     UseMI.getOperand(1).setReg(UseMI.getOperand(UseIdx).getReg());
716   }
717   UseMI.getOperand(UseIdx).ChangeToImmediate(ImmVal);
718   if (DeleteDef)
719     DefMI.eraseFromParent();
720 
721   return true;
722 }
723 
724 Register VEInstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
725   VEMachineFunctionInfo *VEFI = MF->getInfo<VEMachineFunctionInfo>();
726   Register GlobalBaseReg = VEFI->getGlobalBaseReg();
727   if (GlobalBaseReg != 0)
728     return GlobalBaseReg;
729 
730   // We use %s15 (%got) as a global base register
731   GlobalBaseReg = VE::SX15;
732 
733   // Insert a pseudo instruction to set the GlobalBaseReg into the first
734   // MBB of the function
735   MachineBasicBlock &FirstMBB = MF->front();
736   MachineBasicBlock::iterator MBBI = FirstMBB.begin();
737   DebugLoc dl;
738   BuildMI(FirstMBB, MBBI, dl, get(VE::GETGOT), GlobalBaseReg);
739   VEFI->setGlobalBaseReg(GlobalBaseReg);
740   return GlobalBaseReg;
741 }
742 
743 static Register getVM512Upper(Register reg) {
744   return (reg - VE::VMP0) * 2 + VE::VM0;
745 }
746 
747 static Register getVM512Lower(Register reg) { return getVM512Upper(reg) + 1; }
748 
749 // Expand pseudo logical vector instructions for VM512 registers.
750 static void expandPseudoLogM(MachineInstr &MI, const MCInstrDesc &MCID) {
751   MachineBasicBlock *MBB = MI.getParent();
752   DebugLoc DL = MI.getDebugLoc();
753 
754   Register VMXu = getVM512Upper(MI.getOperand(0).getReg());
755   Register VMXl = getVM512Lower(MI.getOperand(0).getReg());
756   Register VMYu = getVM512Upper(MI.getOperand(1).getReg());
757   Register VMYl = getVM512Lower(MI.getOperand(1).getReg());
758 
759   switch (MI.getOpcode()) {
760   default: {
761     Register VMZu = getVM512Upper(MI.getOperand(2).getReg());
762     Register VMZl = getVM512Lower(MI.getOperand(2).getReg());
763     BuildMI(*MBB, MI, DL, MCID).addDef(VMXu).addUse(VMYu).addUse(VMZu);
764     BuildMI(*MBB, MI, DL, MCID).addDef(VMXl).addUse(VMYl).addUse(VMZl);
765     break;
766   }
767   case VE::NEGMy:
768     BuildMI(*MBB, MI, DL, MCID).addDef(VMXu).addUse(VMYu);
769     BuildMI(*MBB, MI, DL, MCID).addDef(VMXl).addUse(VMYl);
770     break;
771   }
772   MI.eraseFromParent();
773 }
774 
775 static void addOperandsForVFMK(MachineInstrBuilder &MIB, MachineInstr &MI,
776                                bool Upper) {
777   // VM512
778   MIB.addReg(Upper ? getVM512Upper(MI.getOperand(0).getReg())
779                    : getVM512Lower(MI.getOperand(0).getReg()));
780 
781   switch (MI.getNumExplicitOperands()) {
782   default:
783     report_fatal_error("unexpected number of operands for pvfmk");
784   case 2: // _Ml: VM512, VL
785     // VL
786     MIB.addReg(MI.getOperand(1).getReg());
787     break;
788   case 4: // _Mvl: VM512, CC, VR, VL
789     // CC
790     MIB.addImm(MI.getOperand(1).getImm());
791     // VR
792     MIB.addReg(MI.getOperand(2).getReg());
793     // VL
794     MIB.addReg(MI.getOperand(3).getReg());
795     break;
796   case 5: // _MvMl: VM512, CC, VR, VM512, VL
797     // CC
798     MIB.addImm(MI.getOperand(1).getImm());
799     // VR
800     MIB.addReg(MI.getOperand(2).getReg());
801     // VM512
802     MIB.addReg(Upper ? getVM512Upper(MI.getOperand(3).getReg())
803                      : getVM512Lower(MI.getOperand(3).getReg()));
804     // VL
805     MIB.addReg(MI.getOperand(4).getReg());
806     break;
807   }
808 }
809 
810 static void expandPseudoVFMK(const TargetInstrInfo &TI, MachineInstr &MI) {
811   // replace to pvfmk.w.up and pvfmk.w.lo
812   // replace to pvfmk.s.up and pvfmk.s.lo
813 
814   static std::map<unsigned, std::pair<unsigned, unsigned>> VFMKMap = {
815       {VE::VFMKyal, {VE::VFMKLal, VE::VFMKLal}},
816       {VE::VFMKynal, {VE::VFMKLnal, VE::VFMKLnal}},
817       {VE::VFMKWyvl, {VE::PVFMKWUPvl, VE::PVFMKWLOvl}},
818       {VE::VFMKWyvyl, {VE::PVFMKWUPvml, VE::PVFMKWLOvml}},
819       {VE::VFMKSyvl, {VE::PVFMKSUPvl, VE::PVFMKSLOvl}},
820       {VE::VFMKSyvyl, {VE::PVFMKSUPvml, VE::PVFMKSLOvml}},
821   };
822 
823   unsigned Opcode = MI.getOpcode();
824 
825   auto Found = VFMKMap.find(Opcode);
826   if (Found == VFMKMap.end())
827     report_fatal_error("unexpected opcode for pseudo vfmk");
828 
829   unsigned OpcodeUpper = (*Found).second.first;
830   unsigned OpcodeLower = (*Found).second.second;
831 
832   MachineBasicBlock *MBB = MI.getParent();
833   DebugLoc DL = MI.getDebugLoc();
834 
835   MachineInstrBuilder Bu = BuildMI(*MBB, MI, DL, TI.get(OpcodeUpper));
836   addOperandsForVFMK(Bu, MI, /* Upper */ true);
837   MachineInstrBuilder Bl = BuildMI(*MBB, MI, DL, TI.get(OpcodeLower));
838   addOperandsForVFMK(Bl, MI, /* Upper */ false);
839 
840   MI.eraseFromParent();
841 }
842 
843 bool VEInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
844   switch (MI.getOpcode()) {
845   case VE::EXTEND_STACK: {
846     return expandExtendStackPseudo(MI);
847   }
848   case VE::EXTEND_STACK_GUARD: {
849     MI.eraseFromParent(); // The pseudo instruction is gone now.
850     return true;
851   }
852   case VE::GETSTACKTOP: {
853     return expandGetStackTopPseudo(MI);
854   }
855 
856   case VE::ANDMyy:
857     expandPseudoLogM(MI, get(VE::ANDMmm));
858     return true;
859   case VE::ORMyy:
860     expandPseudoLogM(MI, get(VE::ORMmm));
861     return true;
862   case VE::XORMyy:
863     expandPseudoLogM(MI, get(VE::XORMmm));
864     return true;
865   case VE::EQVMyy:
866     expandPseudoLogM(MI, get(VE::EQVMmm));
867     return true;
868   case VE::NNDMyy:
869     expandPseudoLogM(MI, get(VE::NNDMmm));
870     return true;
871   case VE::NEGMy:
872     expandPseudoLogM(MI, get(VE::NEGMm));
873     return true;
874 
875   case VE::LVMyir:
876   case VE::LVMyim:
877   case VE::LVMyir_y:
878   case VE::LVMyim_y: {
879     Register VMXu = getVM512Upper(MI.getOperand(0).getReg());
880     Register VMXl = getVM512Lower(MI.getOperand(0).getReg());
881     int64_t Imm = MI.getOperand(1).getImm();
882     bool IsSrcReg =
883         MI.getOpcode() == VE::LVMyir || MI.getOpcode() == VE::LVMyir_y;
884     Register Src = IsSrcReg ? MI.getOperand(2).getReg() : VE::NoRegister;
885     int64_t MImm = IsSrcReg ? 0 : MI.getOperand(2).getImm();
886     bool KillSrc = IsSrcReg ? MI.getOperand(2).isKill() : false;
887     Register VMX = VMXl;
888     if (Imm >= 4) {
889       VMX = VMXu;
890       Imm -= 4;
891     }
892     MachineBasicBlock *MBB = MI.getParent();
893     DebugLoc DL = MI.getDebugLoc();
894     switch (MI.getOpcode()) {
895     case VE::LVMyir:
896       BuildMI(*MBB, MI, DL, get(VE::LVMir))
897           .addDef(VMX)
898           .addImm(Imm)
899           .addReg(Src, getKillRegState(KillSrc));
900       break;
901     case VE::LVMyim:
902       BuildMI(*MBB, MI, DL, get(VE::LVMim))
903           .addDef(VMX)
904           .addImm(Imm)
905           .addImm(MImm);
906       break;
907     case VE::LVMyir_y:
908       assert(MI.getOperand(0).getReg() == MI.getOperand(3).getReg() &&
909              "LVMyir_y has different register in 3rd operand");
910       BuildMI(*MBB, MI, DL, get(VE::LVMir_m))
911           .addDef(VMX)
912           .addImm(Imm)
913           .addReg(Src, getKillRegState(KillSrc))
914           .addReg(VMX);
915       break;
916     case VE::LVMyim_y:
917       assert(MI.getOperand(0).getReg() == MI.getOperand(3).getReg() &&
918              "LVMyim_y has different register in 3rd operand");
919       BuildMI(*MBB, MI, DL, get(VE::LVMim_m))
920           .addDef(VMX)
921           .addImm(Imm)
922           .addImm(MImm)
923           .addReg(VMX);
924       break;
925     }
926     MI.eraseFromParent();
927     return true;
928   }
929   case VE::SVMyi: {
930     Register Dest = MI.getOperand(0).getReg();
931     Register VMZu = getVM512Upper(MI.getOperand(1).getReg());
932     Register VMZl = getVM512Lower(MI.getOperand(1).getReg());
933     bool KillSrc = MI.getOperand(1).isKill();
934     int64_t Imm = MI.getOperand(2).getImm();
935     Register VMZ = VMZl;
936     if (Imm >= 4) {
937       VMZ = VMZu;
938       Imm -= 4;
939     }
940     MachineBasicBlock *MBB = MI.getParent();
941     DebugLoc DL = MI.getDebugLoc();
942     MachineInstrBuilder MIB =
943         BuildMI(*MBB, MI, DL, get(VE::SVMmi), Dest).addReg(VMZ).addImm(Imm);
944     MachineInstr *Inst = MIB.getInstr();
945     if (KillSrc) {
946       const TargetRegisterInfo *TRI = &getRegisterInfo();
947       Inst->addRegisterKilled(MI.getOperand(1).getReg(), TRI, true);
948     }
949     MI.eraseFromParent();
950     return true;
951   }
952   case VE::VFMKyal:
953   case VE::VFMKynal:
954   case VE::VFMKWyvl:
955   case VE::VFMKWyvyl:
956   case VE::VFMKSyvl:
957   case VE::VFMKSyvyl:
958     expandPseudoVFMK(*this, MI);
959     return true;
960   }
961   return false;
962 }
963 
964 bool VEInstrInfo::expandExtendStackPseudo(MachineInstr &MI) const {
965   MachineBasicBlock &MBB = *MI.getParent();
966   MachineFunction &MF = *MBB.getParent();
967   const VESubtarget &STI = MF.getSubtarget<VESubtarget>();
968   const VEInstrInfo &TII = *STI.getInstrInfo();
969   DebugLoc dl = MBB.findDebugLoc(MI);
970 
971   // Create following instructions and multiple basic blocks.
972   //
973   // thisBB:
974   //   brge.l.t %sp, %sl, sinkBB
975   // syscallBB:
976   //   ld      %s61, 0x18(, %tp)        // load param area
977   //   or      %s62, 0, %s0             // spill the value of %s0
978   //   lea     %s63, 0x13b              // syscall # of grow
979   //   shm.l   %s63, 0x0(%s61)          // store syscall # at addr:0
980   //   shm.l   %sl, 0x8(%s61)           // store old limit at addr:8
981   //   shm.l   %sp, 0x10(%s61)          // store new limit at addr:16
982   //   monc                             // call monitor
983   //   or      %s0, 0, %s62             // restore the value of %s0
984   // sinkBB:
985 
986   // Create new MBB
987   MachineBasicBlock *BB = &MBB;
988   const BasicBlock *LLVM_BB = BB->getBasicBlock();
989   MachineBasicBlock *syscallMBB = MF.CreateMachineBasicBlock(LLVM_BB);
990   MachineBasicBlock *sinkMBB = MF.CreateMachineBasicBlock(LLVM_BB);
991   MachineFunction::iterator It = ++(BB->getIterator());
992   MF.insert(It, syscallMBB);
993   MF.insert(It, sinkMBB);
994 
995   // Transfer the remainder of BB and its successor edges to sinkMBB.
996   sinkMBB->splice(sinkMBB->begin(), BB,
997                   std::next(std::next(MachineBasicBlock::iterator(MI))),
998                   BB->end());
999   sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
1000 
1001   // Next, add the true and fallthrough blocks as its successors.
1002   BB->addSuccessor(syscallMBB);
1003   BB->addSuccessor(sinkMBB);
1004   BuildMI(BB, dl, TII.get(VE::BRCFLrr_t))
1005       .addImm(VECC::CC_IGE)
1006       .addReg(VE::SX11) // %sp
1007       .addReg(VE::SX8)  // %sl
1008       .addMBB(sinkMBB);
1009 
1010   BB = syscallMBB;
1011 
1012   // Update machine-CFG edges
1013   BB->addSuccessor(sinkMBB);
1014 
1015   BuildMI(BB, dl, TII.get(VE::LDrii), VE::SX61)
1016       .addReg(VE::SX14)
1017       .addImm(0)
1018       .addImm(0x18);
1019   BuildMI(BB, dl, TII.get(VE::ORri), VE::SX62)
1020       .addReg(VE::SX0)
1021       .addImm(0);
1022   BuildMI(BB, dl, TII.get(VE::LEAzii), VE::SX63)
1023       .addImm(0)
1024       .addImm(0)
1025       .addImm(0x13b);
1026   BuildMI(BB, dl, TII.get(VE::SHMLri))
1027       .addReg(VE::SX61)
1028       .addImm(0)
1029       .addReg(VE::SX63);
1030   BuildMI(BB, dl, TII.get(VE::SHMLri))
1031       .addReg(VE::SX61)
1032       .addImm(8)
1033       .addReg(VE::SX8);
1034   BuildMI(BB, dl, TII.get(VE::SHMLri))
1035       .addReg(VE::SX61)
1036       .addImm(16)
1037       .addReg(VE::SX11);
1038   BuildMI(BB, dl, TII.get(VE::MONC));
1039 
1040   BuildMI(BB, dl, TII.get(VE::ORri), VE::SX0)
1041       .addReg(VE::SX62)
1042       .addImm(0);
1043 
1044   MI.eraseFromParent(); // The pseudo instruction is gone now.
1045   return true;
1046 }
1047 
1048 bool VEInstrInfo::expandGetStackTopPseudo(MachineInstr &MI) const {
1049   MachineBasicBlock *MBB = MI.getParent();
1050   MachineFunction &MF = *MBB->getParent();
1051   const VESubtarget &STI = MF.getSubtarget<VESubtarget>();
1052   const VEInstrInfo &TII = *STI.getInstrInfo();
1053   DebugLoc DL = MBB->findDebugLoc(MI);
1054 
1055   // Create following instruction
1056   //
1057   //   dst = %sp + target specific frame + the size of parameter area
1058 
1059   const MachineFrameInfo &MFI = MF.getFrameInfo();
1060   const VEFrameLowering &TFL = *STI.getFrameLowering();
1061 
1062   // The VE ABI requires a reserved area at the top of stack as described
1063   // in VEFrameLowering.cpp.  So, we adjust it here.
1064   unsigned NumBytes = STI.getAdjustedFrameSize(0);
1065 
1066   // Also adds the size of parameter area.
1067   if (MFI.adjustsStack() && TFL.hasReservedCallFrame(MF))
1068     NumBytes += MFI.getMaxCallFrameSize();
1069 
1070   BuildMI(*MBB, MI, DL, TII.get(VE::LEArii))
1071       .addDef(MI.getOperand(0).getReg())
1072       .addReg(VE::SX11)
1073       .addImm(0)
1074       .addImm(NumBytes);
1075 
1076   MI.eraseFromParent(); // The pseudo instruction is gone now.
1077   return true;
1078 }
1079