1 //===-- HexagonInstrInfo.cpp - Hexagon Instruction Information ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains the Hexagon implementation of the TargetInstrInfo class.
12 //===----------------------------------------------------------------------===//
15 #include "HexagonHazardRecognizer.h"
16 #include "HexagonInstrInfo.h"
17 #include "HexagonRegisterInfo.h"
18 #include "HexagonSubtarget.h"
19 #include "llvm/ADT/SmallPtrSet.h"
20 #include "llvm/ADT/SmallVector.h"
21 #include "llvm/ADT/StringRef.h"
22 #include "llvm/CodeGen/DFAPacketizer.h"
23 #include "llvm/CodeGen/LivePhysRegs.h"
24 #include "llvm/CodeGen/MachineBasicBlock.h"
25 #include "llvm/CodeGen/MachineBranchProbabilityInfo.h"
26 #include "llvm/CodeGen/MachineFrameInfo.h"
27 #include "llvm/CodeGen/MachineFunction.h"
28 #include "llvm/CodeGen/MachineInstr.h"
29 #include "llvm/CodeGen/MachineInstrBuilder.h"
30 #include "llvm/CodeGen/MachineInstrBundle.h"
31 #include "llvm/CodeGen/MachineLoopInfo.h"
32 #include "llvm/CodeGen/MachineMemOperand.h"
33 #include "llvm/CodeGen/MachineOperand.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/ScheduleDAG.h"
36 #include "llvm/MC/MCAsmInfo.h"
37 #include "llvm/MC/MCInstrDesc.h"
38 #include "llvm/MC/MCInstrItineraries.h"
39 #include "llvm/MC/MCRegisterInfo.h"
40 #include "llvm/Support/BranchProbability.h"
41 #include "llvm/Support/CommandLine.h"
42 #include "llvm/Support/Debug.h"
43 #include "llvm/Support/ErrorHandling.h"
44 #include "llvm/Support/MathExtras.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/Target/TargetInstrInfo.h"
47 #include "llvm/Target/TargetSubtargetInfo.h"
56 #define DEBUG_TYPE "hexagon-instrinfo"
58 #define GET_INSTRINFO_CTOR_DTOR
59 #define GET_INSTRMAP_INFO
60 #include "HexagonGenInstrInfo.inc"
61 #include "HexagonGenDFAPacketizer.inc"
63 cl::opt<bool> ScheduleInlineAsm("hexagon-sched-inline-asm", cl::Hidden,
64 cl::init(false), cl::desc("Do not consider inline-asm a scheduling/"
65 "packetization boundary."));
67 static cl::opt<bool> EnableBranchPrediction("hexagon-enable-branch-prediction",
68 cl::Hidden, cl::init(true), cl::desc("Enable branch prediction"));
70 static cl::opt<bool> DisableNVSchedule("disable-hexagon-nv-schedule",
71 cl::Hidden, cl::ZeroOrMore, cl::init(false),
72 cl::desc("Disable schedule adjustment for new value stores."));
74 static cl::opt<bool> EnableTimingClassLatency(
75 "enable-timing-class-latency", cl::Hidden, cl::init(false),
76 cl::desc("Enable timing class latency"));
78 static cl::opt<bool> EnableALUForwarding(
79 "enable-alu-forwarding", cl::Hidden, cl::init(true),
80 cl::desc("Enable vec alu forwarding"));
82 static cl::opt<bool> EnableACCForwarding(
83 "enable-acc-forwarding", cl::Hidden, cl::init(true),
84 cl::desc("Enable vec acc forwarding"));
86 static cl::opt<bool> BranchRelaxAsmLarge("branch-relax-asm-large",
87 cl::init(true), cl::Hidden, cl::ZeroOrMore, cl::desc("branch relax asm"));
89 static cl::opt<bool> UseDFAHazardRec("dfa-hazard-rec",
90 cl::init(true), cl::Hidden, cl::ZeroOrMore,
91 cl::desc("Use the DFA based hazard recognizer."));
94 /// Constants for Hexagon instructions.
96 const int Hexagon_MEMV_OFFSET_MAX_128B = 896; // #s4: -8*128...7*128
97 const int Hexagon_MEMV_OFFSET_MIN_128B = -1024; // #s4
98 const int Hexagon_MEMV_OFFSET_MAX = 448; // #s4: -8*64...7*64
99 const int Hexagon_MEMV_OFFSET_MIN = -512; // #s4
100 const int Hexagon_MEMW_OFFSET_MAX = 4095;
101 const int Hexagon_MEMW_OFFSET_MIN = -4096;
102 const int Hexagon_MEMD_OFFSET_MAX = 8191;
103 const int Hexagon_MEMD_OFFSET_MIN = -8192;
104 const int Hexagon_MEMH_OFFSET_MAX = 2047;
105 const int Hexagon_MEMH_OFFSET_MIN = -2048;
106 const int Hexagon_MEMB_OFFSET_MAX = 1023;
107 const int Hexagon_MEMB_OFFSET_MIN = -1024;
108 const int Hexagon_ADDI_OFFSET_MAX = 32767;
109 const int Hexagon_ADDI_OFFSET_MIN = -32768;
110 const int Hexagon_MEMD_AUTOINC_MAX = 56;
111 const int Hexagon_MEMD_AUTOINC_MIN = -64;
112 const int Hexagon_MEMW_AUTOINC_MAX = 28;
113 const int Hexagon_MEMW_AUTOINC_MIN = -32;
114 const int Hexagon_MEMH_AUTOINC_MAX = 14;
115 const int Hexagon_MEMH_AUTOINC_MIN = -16;
116 const int Hexagon_MEMB_AUTOINC_MAX = 7;
117 const int Hexagon_MEMB_AUTOINC_MIN = -8;
118 const int Hexagon_MEMV_AUTOINC_MAX = 192; // #s3
119 const int Hexagon_MEMV_AUTOINC_MIN = -256; // #s3
120 const int Hexagon_MEMV_AUTOINC_MAX_128B = 384; // #s3
121 const int Hexagon_MEMV_AUTOINC_MIN_128B = -512; // #s3
123 // Pin the vtable to this file.
124 void HexagonInstrInfo::anchor() {}
126 HexagonInstrInfo::HexagonInstrInfo(HexagonSubtarget &ST)
127 : HexagonGenInstrInfo(Hexagon::ADJCALLSTACKDOWN, Hexagon::ADJCALLSTACKUP),
130 static bool isIntRegForSubInst(unsigned Reg) {
131 return (Reg >= Hexagon::R0 && Reg <= Hexagon::R7) ||
132 (Reg >= Hexagon::R16 && Reg <= Hexagon::R23);
135 static bool isDblRegForSubInst(unsigned Reg, const HexagonRegisterInfo &HRI) {
136 return isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_lo)) &&
137 isIntRegForSubInst(HRI.getSubReg(Reg, Hexagon::isub_hi));
140 /// Calculate number of instructions excluding the debug instructions.
141 static unsigned nonDbgMICount(MachineBasicBlock::const_instr_iterator MIB,
142 MachineBasicBlock::const_instr_iterator MIE) {
144 for (; MIB != MIE; ++MIB) {
145 if (!MIB->isDebugValue())
151 /// Find the hardware loop instruction used to set-up the specified loop.
152 /// On Hexagon, we have two instructions used to set-up the hardware loop
153 /// (LOOP0, LOOP1) with corresponding endloop (ENDLOOP0, ENDLOOP1) instructions
154 /// to indicate the end of a loop.
155 static MachineInstr *findLoopInstr(MachineBasicBlock *BB, int EndLoopOp,
156 SmallPtrSet<MachineBasicBlock *, 8> &Visited) {
159 if (EndLoopOp == Hexagon::ENDLOOP0) {
160 LOOPi = Hexagon::J2_loop0i;
161 LOOPr = Hexagon::J2_loop0r;
162 } else { // EndLoopOp == Hexagon::EndLOOP1
163 LOOPi = Hexagon::J2_loop1i;
164 LOOPr = Hexagon::J2_loop1r;
167 // The loop set-up instruction will be in a predecessor block
168 for (MachineBasicBlock::pred_iterator PB = BB->pred_begin(),
169 PE = BB->pred_end(); PB != PE; ++PB) {
170 // If this has been visited, already skip it.
171 if (!Visited.insert(*PB).second)
175 for (MachineBasicBlock::reverse_instr_iterator I = (*PB)->instr_rbegin(),
176 E = (*PB)->instr_rend(); I != E; ++I) {
177 int Opc = I->getOpcode();
178 if (Opc == LOOPi || Opc == LOOPr)
180 // We've reached a different loop, which means the loop0 has been removed.
181 if (Opc == EndLoopOp)
184 // Check the predecessors for the LOOP instruction.
185 MachineInstr *loop = findLoopInstr(*PB, EndLoopOp, Visited);
192 /// Gather register def/uses from MI.
193 /// This treats possible (predicated) defs as actually happening ones
194 /// (conservatively).
195 static inline void parseOperands(const MachineInstr &MI,
196 SmallVector<unsigned, 4> &Defs, SmallVector<unsigned, 8> &Uses) {
200 for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
201 const MachineOperand &MO = MI.getOperand(i);
206 unsigned Reg = MO.getReg();
211 Uses.push_back(MO.getReg());
214 Defs.push_back(MO.getReg());
218 // Position dependent, so check twice for swap.
219 static bool isDuplexPairMatch(unsigned Ga, unsigned Gb) {
221 case HexagonII::HSIG_None:
224 case HexagonII::HSIG_L1:
225 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_A);
226 case HexagonII::HSIG_L2:
227 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
228 Gb == HexagonII::HSIG_A);
229 case HexagonII::HSIG_S1:
230 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
231 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_A);
232 case HexagonII::HSIG_S2:
233 return (Gb == HexagonII::HSIG_L1 || Gb == HexagonII::HSIG_L2 ||
234 Gb == HexagonII::HSIG_S1 || Gb == HexagonII::HSIG_S2 ||
235 Gb == HexagonII::HSIG_A);
236 case HexagonII::HSIG_A:
237 return (Gb == HexagonII::HSIG_A);
238 case HexagonII::HSIG_Compound:
239 return (Gb == HexagonII::HSIG_Compound);
244 /// isLoadFromStackSlot - If the specified machine instruction is a direct
245 /// load from a stack slot, return the virtual or physical register number of
246 /// the destination along with the FrameIndex of the loaded stack slot. If
247 /// not, return 0. This predicate must return 0 if the instruction has
248 /// any side effects other than loading from the stack slot.
249 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
250 int &FrameIndex) const {
251 switch (MI.getOpcode()) {
254 case Hexagon::L2_loadri_io:
255 case Hexagon::L2_loadrd_io:
256 case Hexagon::V6_vL32b_ai:
257 case Hexagon::V6_vL32b_ai_128B:
258 case Hexagon::V6_vL32Ub_ai:
259 case Hexagon::V6_vL32Ub_ai_128B:
260 case Hexagon::LDriw_pred:
261 case Hexagon::LDriw_mod:
262 case Hexagon::PS_vloadrq_ai:
263 case Hexagon::PS_vloadrw_ai:
264 case Hexagon::PS_vloadrq_ai_128B:
265 case Hexagon::PS_vloadrw_ai_128B: {
266 const MachineOperand OpFI = MI.getOperand(1);
269 const MachineOperand OpOff = MI.getOperand(2);
270 if (!OpOff.isImm() || OpOff.getImm() != 0)
272 FrameIndex = OpFI.getIndex();
273 return MI.getOperand(0).getReg();
276 case Hexagon::L2_ploadrit_io:
277 case Hexagon::L2_ploadrif_io:
278 case Hexagon::L2_ploadrdt_io:
279 case Hexagon::L2_ploadrdf_io: {
280 const MachineOperand OpFI = MI.getOperand(2);
283 const MachineOperand OpOff = MI.getOperand(3);
284 if (!OpOff.isImm() || OpOff.getImm() != 0)
286 FrameIndex = OpFI.getIndex();
287 return MI.getOperand(0).getReg();
294 /// isStoreToStackSlot - If the specified machine instruction is a direct
295 /// store to a stack slot, return the virtual or physical register number of
296 /// the source reg along with the FrameIndex of the loaded stack slot. If
297 /// not, return 0. This predicate must return 0 if the instruction has
298 /// any side effects other than storing to the stack slot.
299 unsigned HexagonInstrInfo::isStoreToStackSlot(const MachineInstr &MI,
300 int &FrameIndex) const {
301 switch (MI.getOpcode()) {
304 case Hexagon::S2_storerb_io:
305 case Hexagon::S2_storerh_io:
306 case Hexagon::S2_storeri_io:
307 case Hexagon::S2_storerd_io:
308 case Hexagon::V6_vS32b_ai:
309 case Hexagon::V6_vS32b_ai_128B:
310 case Hexagon::V6_vS32Ub_ai:
311 case Hexagon::V6_vS32Ub_ai_128B:
312 case Hexagon::STriw_pred:
313 case Hexagon::STriw_mod:
314 case Hexagon::PS_vstorerq_ai:
315 case Hexagon::PS_vstorerw_ai:
316 case Hexagon::PS_vstorerq_ai_128B:
317 case Hexagon::PS_vstorerw_ai_128B: {
318 const MachineOperand &OpFI = MI.getOperand(0);
321 const MachineOperand &OpOff = MI.getOperand(1);
322 if (!OpOff.isImm() || OpOff.getImm() != 0)
324 FrameIndex = OpFI.getIndex();
325 return MI.getOperand(2).getReg();
328 case Hexagon::S2_pstorerbt_io:
329 case Hexagon::S2_pstorerbf_io:
330 case Hexagon::S2_pstorerht_io:
331 case Hexagon::S2_pstorerhf_io:
332 case Hexagon::S2_pstorerit_io:
333 case Hexagon::S2_pstorerif_io:
334 case Hexagon::S2_pstorerdt_io:
335 case Hexagon::S2_pstorerdf_io: {
336 const MachineOperand &OpFI = MI.getOperand(1);
339 const MachineOperand &OpOff = MI.getOperand(2);
340 if (!OpOff.isImm() || OpOff.getImm() != 0)
342 FrameIndex = OpFI.getIndex();
343 return MI.getOperand(3).getReg();
350 /// This function can analyze one/two way branching only and should (mostly) be
351 /// called by target independent side.
352 /// First entry is always the opcode of the branching instruction, except when
353 /// the Cond vector is supposed to be empty, e.g., when AnalyzeBranch fails, a
354 /// BB with only unconditional jump. Subsequent entries depend upon the opcode,
355 /// e.g. Jump_c p will have
359 /// Cond[0] = ENDLOOP
362 /// Cond[0] = Hexagon::CMPEQri_f_Jumpnv_t_V4 -- specific opcode
366 bool HexagonInstrInfo::analyzeBranch(MachineBasicBlock &MBB,
367 MachineBasicBlock *&TBB,
368 MachineBasicBlock *&FBB,
369 SmallVectorImpl<MachineOperand> &Cond,
370 bool AllowModify) const {
375 // If the block has no terminators, it just falls into the block after it.
376 MachineBasicBlock::instr_iterator I = MBB.instr_end();
377 if (I == MBB.instr_begin())
380 // A basic block may looks like this:
390 // It has two succs but does not have a terminator
391 // Don't know how to handle it.
395 // Don't analyze EH branches.
397 } while (I != MBB.instr_begin());
402 while (I->isDebugValue()) {
403 if (I == MBB.instr_begin())
408 bool JumpToBlock = I->getOpcode() == Hexagon::J2_jump &&
409 I->getOperand(0).isMBB();
410 // Delete the J2_jump if it's equivalent to a fall-through.
411 if (AllowModify && JumpToBlock &&
412 MBB.isLayoutSuccessor(I->getOperand(0).getMBB())) {
413 DEBUG(dbgs() << "\nErasing the jump to successor block\n";);
414 I->eraseFromParent();
416 if (I == MBB.instr_begin())
420 if (!isUnpredicatedTerminator(*I))
423 // Get the last instruction in the block.
424 MachineInstr *LastInst = &*I;
425 MachineInstr *SecondLastInst = nullptr;
426 // Find one more terminator if present.
428 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) {
430 SecondLastInst = &*I;
432 // This is a third branch.
435 if (I == MBB.instr_begin())
440 int LastOpcode = LastInst->getOpcode();
441 int SecLastOpcode = SecondLastInst ? SecondLastInst->getOpcode() : 0;
442 // If the branch target is not a basic block, it could be a tail call.
443 // (It is, if the target is a function.)
444 if (LastOpcode == Hexagon::J2_jump && !LastInst->getOperand(0).isMBB())
446 if (SecLastOpcode == Hexagon::J2_jump &&
447 !SecondLastInst->getOperand(0).isMBB())
450 bool LastOpcodeHasJMP_c = PredOpcodeHasJMP_c(LastOpcode);
451 bool LastOpcodeHasNVJump = isNewValueJump(*LastInst);
453 if (LastOpcodeHasJMP_c && !LastInst->getOperand(1).isMBB())
456 // If there is only one terminator instruction, process it.
457 if (LastInst && !SecondLastInst) {
458 if (LastOpcode == Hexagon::J2_jump) {
459 TBB = LastInst->getOperand(0).getMBB();
462 if (isEndLoopN(LastOpcode)) {
463 TBB = LastInst->getOperand(0).getMBB();
464 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
465 Cond.push_back(LastInst->getOperand(0));
468 if (LastOpcodeHasJMP_c) {
469 TBB = LastInst->getOperand(1).getMBB();
470 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
471 Cond.push_back(LastInst->getOperand(0));
474 // Only supporting rr/ri versions of new-value jumps.
475 if (LastOpcodeHasNVJump && (LastInst->getNumExplicitOperands() == 3)) {
476 TBB = LastInst->getOperand(2).getMBB();
477 Cond.push_back(MachineOperand::CreateImm(LastInst->getOpcode()));
478 Cond.push_back(LastInst->getOperand(0));
479 Cond.push_back(LastInst->getOperand(1));
482 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber()
483 << " with one jump\n";);
484 // Otherwise, don't know what this is.
488 bool SecLastOpcodeHasJMP_c = PredOpcodeHasJMP_c(SecLastOpcode);
489 bool SecLastOpcodeHasNVJump = isNewValueJump(*SecondLastInst);
490 if (SecLastOpcodeHasJMP_c && (LastOpcode == Hexagon::J2_jump)) {
491 if (!SecondLastInst->getOperand(1).isMBB())
493 TBB = SecondLastInst->getOperand(1).getMBB();
494 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
495 Cond.push_back(SecondLastInst->getOperand(0));
496 FBB = LastInst->getOperand(0).getMBB();
500 // Only supporting rr/ri versions of new-value jumps.
501 if (SecLastOpcodeHasNVJump &&
502 (SecondLastInst->getNumExplicitOperands() == 3) &&
503 (LastOpcode == Hexagon::J2_jump)) {
504 TBB = SecondLastInst->getOperand(2).getMBB();
505 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
506 Cond.push_back(SecondLastInst->getOperand(0));
507 Cond.push_back(SecondLastInst->getOperand(1));
508 FBB = LastInst->getOperand(0).getMBB();
512 // If the block ends with two Hexagon:JMPs, handle it. The second one is not
513 // executed, so remove it.
514 if (SecLastOpcode == Hexagon::J2_jump && LastOpcode == Hexagon::J2_jump) {
515 TBB = SecondLastInst->getOperand(0).getMBB();
516 I = LastInst->getIterator();
518 I->eraseFromParent();
522 // If the block ends with an ENDLOOP, and J2_jump, handle it.
523 if (isEndLoopN(SecLastOpcode) && LastOpcode == Hexagon::J2_jump) {
524 TBB = SecondLastInst->getOperand(0).getMBB();
525 Cond.push_back(MachineOperand::CreateImm(SecondLastInst->getOpcode()));
526 Cond.push_back(SecondLastInst->getOperand(0));
527 FBB = LastInst->getOperand(0).getMBB();
530 DEBUG(dbgs() << "\nCant analyze BB#" << MBB.getNumber()
531 << " with two jumps";);
532 // Otherwise, can't handle this.
536 unsigned HexagonInstrInfo::removeBranch(MachineBasicBlock &MBB,
537 int *BytesRemoved) const {
538 assert(!BytesRemoved && "code size not handled");
540 DEBUG(dbgs() << "\nRemoving branches out of BB#" << MBB.getNumber());
541 MachineBasicBlock::iterator I = MBB.end();
543 while (I != MBB.begin()) {
545 if (I->isDebugValue())
547 // Only removing branches from end of MBB.
550 if (Count && (I->getOpcode() == Hexagon::J2_jump))
551 llvm_unreachable("Malformed basic block: unconditional branch not last");
552 MBB.erase(&MBB.back());
559 unsigned HexagonInstrInfo::insertBranch(MachineBasicBlock &MBB,
560 MachineBasicBlock *TBB,
561 MachineBasicBlock *FBB,
562 ArrayRef<MachineOperand> Cond,
564 int *BytesAdded) const {
565 unsigned BOpc = Hexagon::J2_jump;
566 unsigned BccOpc = Hexagon::J2_jumpt;
567 assert(validateBranchCond(Cond) && "Invalid branching condition");
568 assert(TBB && "insertBranch must not be told to insert a fallthrough");
569 assert(!BytesAdded && "code size not handled");
571 // Check if reverseBranchCondition has asked to reverse this branch
572 // If we want to reverse the branch an odd number of times, we want
574 if (!Cond.empty() && Cond[0].isImm())
575 BccOpc = Cond[0].getImm();
579 // Due to a bug in TailMerging/CFG Optimization, we need to add a
580 // special case handling of a predicated jump followed by an
581 // unconditional jump. If not, Tail Merging and CFG Optimization go
582 // into an infinite loop.
583 MachineBasicBlock *NewTBB, *NewFBB;
584 SmallVector<MachineOperand, 4> Cond;
585 auto Term = MBB.getFirstTerminator();
586 if (Term != MBB.end() && isPredicated(*Term) &&
587 !analyzeBranch(MBB, NewTBB, NewFBB, Cond, false) &&
588 MachineFunction::iterator(NewTBB) == ++MBB.getIterator()) {
589 reverseBranchCondition(Cond);
591 return insertBranch(MBB, TBB, nullptr, Cond, DL);
593 BuildMI(&MBB, DL, get(BOpc)).addMBB(TBB);
594 } else if (isEndLoopN(Cond[0].getImm())) {
595 int EndLoopOp = Cond[0].getImm();
596 assert(Cond[1].isMBB());
597 // Since we're adding an ENDLOOP, there better be a LOOP instruction.
598 // Check for it, and change the BB target if needed.
599 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
600 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, VisitedBBs);
601 assert(Loop != 0 && "Inserting an ENDLOOP without a LOOP");
602 Loop->getOperand(0).setMBB(TBB);
603 // Add the ENDLOOP after the finding the LOOP0.
604 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
605 } else if (isNewValueJump(Cond[0].getImm())) {
606 assert((Cond.size() == 3) && "Only supporting rr/ri version of nvjump");
608 // (ins IntRegs:$src1, IntRegs:$src2, brtarget:$offset)
609 // (ins IntRegs:$src1, u5Imm:$src2, brtarget:$offset)
610 unsigned Flags1 = getUndefRegState(Cond[1].isUndef());
611 DEBUG(dbgs() << "\nInserting NVJump for BB#" << MBB.getNumber(););
612 if (Cond[2].isReg()) {
613 unsigned Flags2 = getUndefRegState(Cond[2].isUndef());
614 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
615 addReg(Cond[2].getReg(), Flags2).addMBB(TBB);
616 } else if(Cond[2].isImm()) {
617 BuildMI(&MBB, DL, get(BccOpc)).addReg(Cond[1].getReg(), Flags1).
618 addImm(Cond[2].getImm()).addMBB(TBB);
620 llvm_unreachable("Invalid condition for branching");
622 assert((Cond.size() == 2) && "Malformed cond vector");
623 const MachineOperand &RO = Cond[1];
624 unsigned Flags = getUndefRegState(RO.isUndef());
625 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
629 assert((!Cond.empty()) &&
630 "Cond. cannot be empty when multiple branchings are required");
631 assert((!isNewValueJump(Cond[0].getImm())) &&
632 "NV-jump cannot be inserted with another branch");
633 // Special case for hardware loops. The condition is a basic block.
634 if (isEndLoopN(Cond[0].getImm())) {
635 int EndLoopOp = Cond[0].getImm();
636 assert(Cond[1].isMBB());
637 // Since we're adding an ENDLOOP, there better be a LOOP instruction.
638 // Check for it, and change the BB target if needed.
639 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
640 MachineInstr *Loop = findLoopInstr(TBB, EndLoopOp, VisitedBBs);
641 assert(Loop != 0 && "Inserting an ENDLOOP without a LOOP");
642 Loop->getOperand(0).setMBB(TBB);
643 // Add the ENDLOOP after the finding the LOOP0.
644 BuildMI(&MBB, DL, get(EndLoopOp)).addMBB(TBB);
646 const MachineOperand &RO = Cond[1];
647 unsigned Flags = getUndefRegState(RO.isUndef());
648 BuildMI(&MBB, DL, get(BccOpc)).addReg(RO.getReg(), Flags).addMBB(TBB);
650 BuildMI(&MBB, DL, get(BOpc)).addMBB(FBB);
655 /// Analyze the loop code to find the loop induction variable and compare used
656 /// to compute the number of iterations. Currently, we analyze loop that are
657 /// controlled using hardware loops. In this case, the induction variable
658 /// instruction is null. For all other cases, this function returns true, which
659 /// means we're unable to analyze it.
660 bool HexagonInstrInfo::analyzeLoop(MachineLoop &L,
661 MachineInstr *&IndVarInst,
662 MachineInstr *&CmpInst) const {
664 MachineBasicBlock *LoopEnd = L.getBottomBlock();
665 MachineBasicBlock::iterator I = LoopEnd->getFirstTerminator();
666 // We really "analyze" only hardware loops right now.
667 if (I != LoopEnd->end() && isEndLoopN(I->getOpcode())) {
668 IndVarInst = nullptr;
675 /// Generate code to reduce the loop iteration by one and check if the loop is
676 /// finished. Return the value/register of the new loop count. this function
677 /// assumes the nth iteration is peeled first.
678 unsigned HexagonInstrInfo::reduceLoopCount(MachineBasicBlock &MBB,
679 MachineInstr *IndVar, MachineInstr &Cmp,
680 SmallVectorImpl<MachineOperand> &Cond,
681 SmallVectorImpl<MachineInstr *> &PrevInsts,
682 unsigned Iter, unsigned MaxIter) const {
683 // We expect a hardware loop currently. This means that IndVar is set
684 // to null, and the compare is the ENDLOOP instruction.
685 assert((!IndVar) && isEndLoopN(Cmp.getOpcode())
686 && "Expecting a hardware loop");
687 MachineFunction *MF = MBB.getParent();
688 DebugLoc DL = Cmp.getDebugLoc();
689 SmallPtrSet<MachineBasicBlock *, 8> VisitedBBs;
690 MachineInstr *Loop = findLoopInstr(&MBB, Cmp.getOpcode(), VisitedBBs);
693 // If the loop trip count is a compile-time value, then just change the
695 if (Loop->getOpcode() == Hexagon::J2_loop0i ||
696 Loop->getOpcode() == Hexagon::J2_loop1i) {
697 int64_t Offset = Loop->getOperand(1).getImm();
699 Loop->eraseFromParent();
701 Loop->getOperand(1).setImm(Offset - 1);
704 // The loop trip count is a run-time value. We generate code to subtract
705 // one from the trip count, and update the loop instruction.
706 assert(Loop->getOpcode() == Hexagon::J2_loop0r && "Unexpected instruction");
707 unsigned LoopCount = Loop->getOperand(1).getReg();
708 // Check if we're done with the loop.
709 unsigned LoopEnd = createVR(MF, MVT::i1);
710 MachineInstr *NewCmp = BuildMI(&MBB, DL, get(Hexagon::C2_cmpgtui), LoopEnd).
711 addReg(LoopCount).addImm(1);
712 unsigned NewLoopCount = createVR(MF, MVT::i32);
713 MachineInstr *NewAdd = BuildMI(&MBB, DL, get(Hexagon::A2_addi), NewLoopCount).
714 addReg(LoopCount).addImm(-1);
715 // Update the previously generated instructions with the new loop counter.
716 for (SmallVectorImpl<MachineInstr *>::iterator I = PrevInsts.begin(),
717 E = PrevInsts.end(); I != E; ++I)
718 (*I)->substituteRegister(LoopCount, NewLoopCount, 0, getRegisterInfo());
720 PrevInsts.push_back(NewCmp);
721 PrevInsts.push_back(NewAdd);
722 // Insert the new loop instruction if this is the last time the loop is
725 BuildMI(&MBB, DL, get(Hexagon::J2_loop0r)).
726 addMBB(Loop->getOperand(0).getMBB()).addReg(NewLoopCount);
727 // Delete the old loop instruction.
729 Loop->eraseFromParent();
730 Cond.push_back(MachineOperand::CreateImm(Hexagon::J2_jumpf));
731 Cond.push_back(NewCmp->getOperand(0));
735 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &MBB,
736 unsigned NumCycles, unsigned ExtraPredCycles,
737 BranchProbability Probability) const {
738 return nonDbgBBSize(&MBB) <= 3;
741 bool HexagonInstrInfo::isProfitableToIfCvt(MachineBasicBlock &TMBB,
742 unsigned NumTCycles, unsigned ExtraTCycles, MachineBasicBlock &FMBB,
743 unsigned NumFCycles, unsigned ExtraFCycles, BranchProbability Probability)
745 return nonDbgBBSize(&TMBB) <= 3 && nonDbgBBSize(&FMBB) <= 3;
748 bool HexagonInstrInfo::isProfitableToDupForIfCvt(MachineBasicBlock &MBB,
749 unsigned NumInstrs, BranchProbability Probability) const {
750 return NumInstrs <= 4;
753 void HexagonInstrInfo::copyPhysReg(MachineBasicBlock &MBB,
754 MachineBasicBlock::iterator I,
755 const DebugLoc &DL, unsigned DestReg,
756 unsigned SrcReg, bool KillSrc) const {
757 auto &HRI = getRegisterInfo();
758 unsigned KillFlag = getKillRegState(KillSrc);
760 if (Hexagon::IntRegsRegClass.contains(SrcReg, DestReg)) {
761 BuildMI(MBB, I, DL, get(Hexagon::A2_tfr), DestReg)
762 .addReg(SrcReg, KillFlag);
765 if (Hexagon::DoubleRegsRegClass.contains(SrcReg, DestReg)) {
766 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrp), DestReg)
767 .addReg(SrcReg, KillFlag);
770 if (Hexagon::PredRegsRegClass.contains(SrcReg, DestReg)) {
771 // Map Pd = Ps to Pd = or(Ps, Ps).
772 BuildMI(MBB, I, DL, get(Hexagon::C2_or), DestReg)
773 .addReg(SrcReg).addReg(SrcReg, KillFlag);
776 if (Hexagon::CtrRegsRegClass.contains(DestReg) &&
777 Hexagon::IntRegsRegClass.contains(SrcReg)) {
778 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
779 .addReg(SrcReg, KillFlag);
782 if (Hexagon::IntRegsRegClass.contains(DestReg) &&
783 Hexagon::CtrRegsRegClass.contains(SrcReg)) {
784 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrcrr), DestReg)
785 .addReg(SrcReg, KillFlag);
788 if (Hexagon::ModRegsRegClass.contains(DestReg) &&
789 Hexagon::IntRegsRegClass.contains(SrcReg)) {
790 BuildMI(MBB, I, DL, get(Hexagon::A2_tfrrcr), DestReg)
791 .addReg(SrcReg, KillFlag);
794 if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
795 Hexagon::IntRegsRegClass.contains(DestReg)) {
796 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
797 .addReg(SrcReg, KillFlag);
800 if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
801 Hexagon::PredRegsRegClass.contains(DestReg)) {
802 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrrp), DestReg)
803 .addReg(SrcReg, KillFlag);
806 if (Hexagon::PredRegsRegClass.contains(SrcReg) &&
807 Hexagon::IntRegsRegClass.contains(DestReg)) {
808 BuildMI(MBB, I, DL, get(Hexagon::C2_tfrpr), DestReg)
809 .addReg(SrcReg, KillFlag);
812 if (Hexagon::VectorRegsRegClass.contains(SrcReg, DestReg)) {
813 BuildMI(MBB, I, DL, get(Hexagon::V6_vassign), DestReg).
814 addReg(SrcReg, KillFlag);
817 if (Hexagon::VecDblRegsRegClass.contains(SrcReg, DestReg)) {
818 unsigned LoSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
819 unsigned HiSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
820 BuildMI(MBB, I, DL, get(Hexagon::V6_vcombine), DestReg)
821 .addReg(HiSrc, KillFlag)
822 .addReg(LoSrc, KillFlag);
825 if (Hexagon::VecPredRegsRegClass.contains(SrcReg, DestReg)) {
826 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), DestReg)
828 .addReg(SrcReg, KillFlag);
831 if (Hexagon::VecPredRegsRegClass.contains(SrcReg) &&
832 Hexagon::VectorRegsRegClass.contains(DestReg)) {
833 llvm_unreachable("Unimplemented pred to vec");
836 if (Hexagon::VecPredRegsRegClass.contains(DestReg) &&
837 Hexagon::VectorRegsRegClass.contains(SrcReg)) {
838 llvm_unreachable("Unimplemented vec to pred");
841 if (Hexagon::VecPredRegs128BRegClass.contains(SrcReg, DestReg)) {
842 unsigned HiDst = HRI.getSubReg(DestReg, Hexagon::vsub_hi);
843 unsigned LoDst = HRI.getSubReg(DestReg, Hexagon::vsub_lo);
844 unsigned HiSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
845 unsigned LoSrc = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
846 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), HiDst)
847 .addReg(HiSrc, KillFlag);
848 BuildMI(MBB, I, DL, get(Hexagon::V6_pred_and), LoDst)
849 .addReg(LoSrc, KillFlag);
854 // Show the invalid registers to ease debugging.
855 dbgs() << "Invalid registers for copy in BB#" << MBB.getNumber()
856 << ": " << PrintReg(DestReg, &HRI)
857 << " = " << PrintReg(SrcReg, &HRI) << '\n';
859 llvm_unreachable("Unimplemented");
862 void HexagonInstrInfo::storeRegToStackSlot(MachineBasicBlock &MBB,
863 MachineBasicBlock::iterator I, unsigned SrcReg, bool isKill, int FI,
864 const TargetRegisterClass *RC, const TargetRegisterInfo *TRI) const {
865 DebugLoc DL = MBB.findDebugLoc(I);
866 MachineFunction &MF = *MBB.getParent();
867 MachineFrameInfo &MFI = MF.getFrameInfo();
868 unsigned Align = MFI.getObjectAlignment(FI);
869 unsigned KillFlag = getKillRegState(isKill);
871 MachineMemOperand *MMO = MF.getMachineMemOperand(
872 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOStore,
873 MFI.getObjectSize(FI), Align);
875 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
876 BuildMI(MBB, I, DL, get(Hexagon::S2_storeri_io))
877 .addFrameIndex(FI).addImm(0)
878 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
879 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
880 BuildMI(MBB, I, DL, get(Hexagon::S2_storerd_io))
881 .addFrameIndex(FI).addImm(0)
882 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
883 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
884 BuildMI(MBB, I, DL, get(Hexagon::STriw_pred))
885 .addFrameIndex(FI).addImm(0)
886 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
887 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
888 BuildMI(MBB, I, DL, get(Hexagon::STriw_mod))
889 .addFrameIndex(FI).addImm(0)
890 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
891 } else if (Hexagon::VecPredRegs128BRegClass.hasSubClassEq(RC)) {
892 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai_128B))
893 .addFrameIndex(FI).addImm(0)
894 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
895 } else if (Hexagon::VecPredRegsRegClass.hasSubClassEq(RC)) {
896 BuildMI(MBB, I, DL, get(Hexagon::PS_vstorerq_ai))
897 .addFrameIndex(FI).addImm(0)
898 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
899 } else if (Hexagon::VectorRegs128BRegClass.hasSubClassEq(RC)) {
900 unsigned Opc = Align < 128 ? Hexagon::V6_vS32Ub_ai_128B
901 : Hexagon::V6_vS32b_ai_128B;
902 BuildMI(MBB, I, DL, get(Opc))
903 .addFrameIndex(FI).addImm(0)
904 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
905 } else if (Hexagon::VectorRegsRegClass.hasSubClassEq(RC)) {
906 unsigned Opc = Align < 64 ? Hexagon::V6_vS32Ub_ai
907 : Hexagon::V6_vS32b_ai;
908 BuildMI(MBB, I, DL, get(Opc))
909 .addFrameIndex(FI).addImm(0)
910 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
911 } else if (Hexagon::VecDblRegsRegClass.hasSubClassEq(RC)) {
912 unsigned Opc = Align < 64 ? Hexagon::PS_vstorerwu_ai
913 : Hexagon::PS_vstorerw_ai;
914 BuildMI(MBB, I, DL, get(Opc))
915 .addFrameIndex(FI).addImm(0)
916 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
917 } else if (Hexagon::VecDblRegs128BRegClass.hasSubClassEq(RC)) {
918 unsigned Opc = Align < 128 ? Hexagon::PS_vstorerwu_ai_128B
919 : Hexagon::PS_vstorerw_ai_128B;
920 BuildMI(MBB, I, DL, get(Opc))
921 .addFrameIndex(FI).addImm(0)
922 .addReg(SrcReg, KillFlag).addMemOperand(MMO);
924 llvm_unreachable("Unimplemented");
928 void HexagonInstrInfo::loadRegFromStackSlot(
929 MachineBasicBlock &MBB, MachineBasicBlock::iterator I, unsigned DestReg,
930 int FI, const TargetRegisterClass *RC,
931 const TargetRegisterInfo *TRI) const {
932 DebugLoc DL = MBB.findDebugLoc(I);
933 MachineFunction &MF = *MBB.getParent();
934 MachineFrameInfo &MFI = MF.getFrameInfo();
935 unsigned Align = MFI.getObjectAlignment(FI);
937 MachineMemOperand *MMO = MF.getMachineMemOperand(
938 MachinePointerInfo::getFixedStack(MF, FI), MachineMemOperand::MOLoad,
939 MFI.getObjectSize(FI), Align);
941 if (Hexagon::IntRegsRegClass.hasSubClassEq(RC)) {
942 BuildMI(MBB, I, DL, get(Hexagon::L2_loadri_io), DestReg)
943 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
944 } else if (Hexagon::DoubleRegsRegClass.hasSubClassEq(RC)) {
945 BuildMI(MBB, I, DL, get(Hexagon::L2_loadrd_io), DestReg)
946 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
947 } else if (Hexagon::PredRegsRegClass.hasSubClassEq(RC)) {
948 BuildMI(MBB, I, DL, get(Hexagon::LDriw_pred), DestReg)
949 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
950 } else if (Hexagon::ModRegsRegClass.hasSubClassEq(RC)) {
951 BuildMI(MBB, I, DL, get(Hexagon::LDriw_mod), DestReg)
952 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
953 } else if (Hexagon::VecPredRegs128BRegClass.hasSubClassEq(RC)) {
954 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai_128B), DestReg)
955 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
956 } else if (Hexagon::VecPredRegsRegClass.hasSubClassEq(RC)) {
957 BuildMI(MBB, I, DL, get(Hexagon::PS_vloadrq_ai), DestReg)
958 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
959 } else if (Hexagon::VecDblRegs128BRegClass.hasSubClassEq(RC)) {
960 unsigned Opc = Align < 128 ? Hexagon::PS_vloadrwu_ai_128B
961 : Hexagon::PS_vloadrw_ai_128B;
962 BuildMI(MBB, I, DL, get(Opc), DestReg)
963 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
964 } else if (Hexagon::VectorRegs128BRegClass.hasSubClassEq(RC)) {
965 unsigned Opc = Align < 128 ? Hexagon::V6_vL32Ub_ai_128B
966 : Hexagon::V6_vL32b_ai_128B;
967 BuildMI(MBB, I, DL, get(Opc), DestReg)
968 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
969 } else if (Hexagon::VectorRegsRegClass.hasSubClassEq(RC)) {
970 unsigned Opc = Align < 64 ? Hexagon::V6_vL32Ub_ai
971 : Hexagon::V6_vL32b_ai;
972 BuildMI(MBB, I, DL, get(Opc), DestReg)
973 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
974 } else if (Hexagon::VecDblRegsRegClass.hasSubClassEq(RC)) {
975 unsigned Opc = Align < 64 ? Hexagon::PS_vloadrwu_ai
976 : Hexagon::PS_vloadrw_ai;
977 BuildMI(MBB, I, DL, get(Opc), DestReg)
978 .addFrameIndex(FI).addImm(0).addMemOperand(MMO);
980 llvm_unreachable("Can't store this register to stack slot");
984 static void getLiveRegsAt(LivePhysRegs &Regs, const MachineInstr &MI) {
985 const MachineBasicBlock &B = *MI.getParent();
987 auto E = ++MachineBasicBlock::const_iterator(MI.getIterator()).getReverse();
988 for (auto I = B.rbegin(); I != E; ++I)
989 Regs.stepBackward(*I);
992 /// expandPostRAPseudo - This function is called for all pseudo instructions
993 /// that remain after register allocation. Many pseudo instructions are
994 /// created to help register allocation. This is the place to convert them
995 /// into real instructions. The target can edit MI in place, or it can insert
996 /// new instructions and erase MI. The function should return true if
997 /// anything was changed.
998 bool HexagonInstrInfo::expandPostRAPseudo(MachineInstr &MI) const {
999 const HexagonRegisterInfo &HRI = getRegisterInfo();
1000 MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
1001 MachineBasicBlock &MBB = *MI.getParent();
1002 DebugLoc DL = MI.getDebugLoc();
1003 unsigned Opc = MI.getOpcode();
1004 const unsigned VecOffset = 1;
1007 case TargetOpcode::COPY: {
1008 MachineOperand &MD = MI.getOperand(0);
1009 MachineOperand &MS = MI.getOperand(1);
1010 MachineBasicBlock::iterator MBBI = MI.getIterator();
1011 if (MD.getReg() != MS.getReg() && !MS.isUndef()) {
1012 copyPhysReg(MBB, MI, DL, MD.getReg(), MS.getReg(), MS.isKill());
1013 std::prev(MBBI)->copyImplicitOps(*MBB.getParent(), MI);
1018 case Hexagon::PS_aligna:
1019 BuildMI(MBB, MI, DL, get(Hexagon::A2_andir), MI.getOperand(0).getReg())
1020 .addReg(HRI.getFrameRegister())
1021 .addImm(-MI.getOperand(1).getImm());
1024 case Hexagon::V6_vassignp_128B:
1025 case Hexagon::V6_vassignp: {
1026 unsigned SrcReg = MI.getOperand(1).getReg();
1027 unsigned DstReg = MI.getOperand(0).getReg();
1028 unsigned Kill = getKillRegState(MI.getOperand(1).isKill());
1029 BuildMI(MBB, MI, DL, get(Hexagon::V6_vcombine), DstReg)
1030 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_hi), Kill)
1031 .addReg(HRI.getSubReg(SrcReg, Hexagon::vsub_lo), Kill);
1035 case Hexagon::V6_lo_128B:
1036 case Hexagon::V6_lo: {
1037 unsigned SrcReg = MI.getOperand(1).getReg();
1038 unsigned DstReg = MI.getOperand(0).getReg();
1039 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
1040 copyPhysReg(MBB, MI, DL, DstReg, SrcSubLo, MI.getOperand(1).isKill());
1042 MRI.clearKillFlags(SrcSubLo);
1045 case Hexagon::V6_hi_128B:
1046 case Hexagon::V6_hi: {
1047 unsigned SrcReg = MI.getOperand(1).getReg();
1048 unsigned DstReg = MI.getOperand(0).getReg();
1049 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
1050 copyPhysReg(MBB, MI, DL, DstReg, SrcSubHi, MI.getOperand(1).isKill());
1052 MRI.clearKillFlags(SrcSubHi);
1055 case Hexagon::PS_vstorerw_ai:
1056 case Hexagon::PS_vstorerwu_ai:
1057 case Hexagon::PS_vstorerw_ai_128B:
1058 case Hexagon::PS_vstorerwu_ai_128B: {
1059 bool Is128B = (Opc == Hexagon::PS_vstorerw_ai_128B ||
1060 Opc == Hexagon::PS_vstorerwu_ai_128B);
1061 bool Aligned = (Opc == Hexagon::PS_vstorerw_ai ||
1062 Opc == Hexagon::PS_vstorerw_ai_128B);
1063 unsigned SrcReg = MI.getOperand(2).getReg();
1064 unsigned SrcSubHi = HRI.getSubReg(SrcReg, Hexagon::vsub_hi);
1065 unsigned SrcSubLo = HRI.getSubReg(SrcReg, Hexagon::vsub_lo);
1068 NewOpc = Is128B ? Hexagon::V6_vS32b_ai_128B
1069 : Hexagon::V6_vS32b_ai;
1071 NewOpc = Is128B ? Hexagon::V6_vS32Ub_ai_128B
1072 : Hexagon::V6_vS32Ub_ai;
1074 unsigned Offset = Is128B ? VecOffset << 7 : VecOffset << 6;
1075 MachineInstr *MI1New =
1076 BuildMI(MBB, MI, DL, get(NewOpc))
1077 .addOperand(MI.getOperand(0))
1078 .addImm(MI.getOperand(1).getImm())
1080 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
1081 MI1New->getOperand(0).setIsKill(false);
1082 BuildMI(MBB, MI, DL, get(NewOpc))
1083 .addOperand(MI.getOperand(0))
1084 // The Vectors are indexed in multiples of vector size.
1085 .addImm(MI.getOperand(1).getImm() + Offset)
1087 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
1091 case Hexagon::PS_vloadrw_ai:
1092 case Hexagon::PS_vloadrwu_ai:
1093 case Hexagon::PS_vloadrw_ai_128B:
1094 case Hexagon::PS_vloadrwu_ai_128B: {
1095 bool Is128B = (Opc == Hexagon::PS_vloadrw_ai_128B ||
1096 Opc == Hexagon::PS_vloadrwu_ai_128B);
1097 bool Aligned = (Opc == Hexagon::PS_vloadrw_ai ||
1098 Opc == Hexagon::PS_vloadrw_ai_128B);
1101 NewOpc = Is128B ? Hexagon::V6_vL32b_ai_128B
1102 : Hexagon::V6_vL32b_ai;
1104 NewOpc = Is128B ? Hexagon::V6_vL32Ub_ai_128B
1105 : Hexagon::V6_vL32Ub_ai;
1107 unsigned DstReg = MI.getOperand(0).getReg();
1108 unsigned Offset = Is128B ? VecOffset << 7 : VecOffset << 6;
1109 MachineInstr *MI1New =
1110 BuildMI(MBB, MI, DL, get(NewOpc),
1111 HRI.getSubReg(DstReg, Hexagon::vsub_lo))
1112 .addOperand(MI.getOperand(1))
1113 .addImm(MI.getOperand(2).getImm());
1114 MI1New->getOperand(1).setIsKill(false);
1115 BuildMI(MBB, MI, DL, get(NewOpc),
1116 HRI.getSubReg(DstReg, Hexagon::vsub_hi))
1117 .addOperand(MI.getOperand(1))
1118 // The Vectors are indexed in multiples of vector size.
1119 .addImm(MI.getOperand(2).getImm() + Offset)
1120 .setMemRefs(MI.memoperands_begin(), MI.memoperands_end());
1124 case Hexagon::PS_true: {
1125 unsigned Reg = MI.getOperand(0).getReg();
1126 BuildMI(MBB, MI, DL, get(Hexagon::C2_orn), Reg)
1127 .addReg(Reg, RegState::Undef)
1128 .addReg(Reg, RegState::Undef);
1132 case Hexagon::PS_false: {
1133 unsigned Reg = MI.getOperand(0).getReg();
1134 BuildMI(MBB, MI, DL, get(Hexagon::C2_andn), Reg)
1135 .addReg(Reg, RegState::Undef)
1136 .addReg(Reg, RegState::Undef);
1140 case Hexagon::PS_vmulw: {
1141 // Expand a 64-bit vector multiply into 2 32-bit scalar multiplies.
1142 unsigned DstReg = MI.getOperand(0).getReg();
1143 unsigned Src1Reg = MI.getOperand(1).getReg();
1144 unsigned Src2Reg = MI.getOperand(2).getReg();
1145 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
1146 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
1147 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
1148 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
1149 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
1150 HRI.getSubReg(DstReg, Hexagon::isub_hi))
1153 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_mpyi),
1154 HRI.getSubReg(DstReg, Hexagon::isub_lo))
1158 MRI.clearKillFlags(Src1SubHi);
1159 MRI.clearKillFlags(Src1SubLo);
1160 MRI.clearKillFlags(Src2SubHi);
1161 MRI.clearKillFlags(Src2SubLo);
1164 case Hexagon::PS_vmulw_acc: {
1165 // Expand 64-bit vector multiply with addition into 2 scalar multiplies.
1166 unsigned DstReg = MI.getOperand(0).getReg();
1167 unsigned Src1Reg = MI.getOperand(1).getReg();
1168 unsigned Src2Reg = MI.getOperand(2).getReg();
1169 unsigned Src3Reg = MI.getOperand(3).getReg();
1170 unsigned Src1SubHi = HRI.getSubReg(Src1Reg, Hexagon::isub_hi);
1171 unsigned Src1SubLo = HRI.getSubReg(Src1Reg, Hexagon::isub_lo);
1172 unsigned Src2SubHi = HRI.getSubReg(Src2Reg, Hexagon::isub_hi);
1173 unsigned Src2SubLo = HRI.getSubReg(Src2Reg, Hexagon::isub_lo);
1174 unsigned Src3SubHi = HRI.getSubReg(Src3Reg, Hexagon::isub_hi);
1175 unsigned Src3SubLo = HRI.getSubReg(Src3Reg, Hexagon::isub_lo);
1176 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
1177 HRI.getSubReg(DstReg, Hexagon::isub_hi))
1181 BuildMI(MBB, MI, MI.getDebugLoc(), get(Hexagon::M2_maci),
1182 HRI.getSubReg(DstReg, Hexagon::isub_lo))
1187 MRI.clearKillFlags(Src1SubHi);
1188 MRI.clearKillFlags(Src1SubLo);
1189 MRI.clearKillFlags(Src2SubHi);
1190 MRI.clearKillFlags(Src2SubLo);
1191 MRI.clearKillFlags(Src3SubHi);
1192 MRI.clearKillFlags(Src3SubLo);
1195 case Hexagon::PS_pselect: {
1196 const MachineOperand &Op0 = MI.getOperand(0);
1197 const MachineOperand &Op1 = MI.getOperand(1);
1198 const MachineOperand &Op2 = MI.getOperand(2);
1199 const MachineOperand &Op3 = MI.getOperand(3);
1200 unsigned Rd = Op0.getReg();
1201 unsigned Pu = Op1.getReg();
1202 unsigned Rs = Op2.getReg();
1203 unsigned Rt = Op3.getReg();
1204 DebugLoc DL = MI.getDebugLoc();
1205 unsigned K1 = getKillRegState(Op1.isKill());
1206 unsigned K2 = getKillRegState(Op2.isKill());
1207 unsigned K3 = getKillRegState(Op3.isKill());
1209 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpt), Rd)
1210 .addReg(Pu, (Rd == Rt) ? K1 : 0)
1213 BuildMI(MBB, MI, DL, get(Hexagon::A2_tfrpf), Rd)
1219 case Hexagon::PS_vselect:
1220 case Hexagon::PS_vselect_128B: {
1221 const MachineOperand &Op0 = MI.getOperand(0);
1222 const MachineOperand &Op1 = MI.getOperand(1);
1223 const MachineOperand &Op2 = MI.getOperand(2);
1224 const MachineOperand &Op3 = MI.getOperand(3);
1225 LivePhysRegs LiveAtMI(&HRI);
1226 getLiveRegsAt(LiveAtMI, MI);
1227 bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg());
1228 if (Op0.getReg() != Op2.getReg()) {
1229 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vcmov))
1234 T.addReg(Op0.getReg(), RegState::Implicit);
1237 if (Op0.getReg() != Op3.getReg()) {
1238 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vncmov))
1243 T.addReg(Op0.getReg(), RegState::Implicit);
1248 case Hexagon::PS_wselect:
1249 case Hexagon::PS_wselect_128B: {
1250 MachineOperand &Op0 = MI.getOperand(0);
1251 MachineOperand &Op1 = MI.getOperand(1);
1252 MachineOperand &Op2 = MI.getOperand(2);
1253 MachineOperand &Op3 = MI.getOperand(3);
1254 LivePhysRegs LiveAtMI(&HRI);
1255 getLiveRegsAt(LiveAtMI, MI);
1256 bool IsDestLive = !LiveAtMI.available(MRI, Op0.getReg());
1258 if (Op0.getReg() != Op2.getReg()) {
1259 unsigned SrcLo = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_lo);
1260 unsigned SrcHi = HRI.getSubReg(Op2.getReg(), Hexagon::vsub_hi);
1261 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vccombine))
1267 T.addReg(Op0.getReg(), RegState::Implicit);
1270 if (Op0.getReg() != Op3.getReg()) {
1271 unsigned SrcLo = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_lo);
1272 unsigned SrcHi = HRI.getSubReg(Op3.getReg(), Hexagon::vsub_hi);
1273 auto T = BuildMI(MBB, MI, DL, get(Hexagon::V6_vnccombine))
1279 T.addReg(Op0.getReg(), RegState::Implicit);
1284 case Hexagon::PS_tailcall_i:
1285 MI.setDesc(get(Hexagon::J2_jump));
1287 case Hexagon::PS_tailcall_r:
1288 case Hexagon::PS_jmpret:
1289 MI.setDesc(get(Hexagon::J2_jumpr));
1291 case Hexagon::PS_jmprett:
1292 MI.setDesc(get(Hexagon::J2_jumprt));
1294 case Hexagon::PS_jmpretf:
1295 MI.setDesc(get(Hexagon::J2_jumprf));
1297 case Hexagon::PS_jmprettnewpt:
1298 MI.setDesc(get(Hexagon::J2_jumprtnewpt));
1300 case Hexagon::PS_jmpretfnewpt:
1301 MI.setDesc(get(Hexagon::J2_jumprfnewpt));
1303 case Hexagon::PS_jmprettnew:
1304 MI.setDesc(get(Hexagon::J2_jumprtnew));
1306 case Hexagon::PS_jmpretfnew:
1307 MI.setDesc(get(Hexagon::J2_jumprfnew));
1314 // We indicate that we want to reverse the branch by
1315 // inserting the reversed branching opcode.
1316 bool HexagonInstrInfo::reverseBranchCondition(
1317 SmallVectorImpl<MachineOperand> &Cond) const {
1320 assert(Cond[0].isImm() && "First entry in the cond vector not imm-val");
1321 unsigned opcode = Cond[0].getImm();
1323 assert(get(opcode).isBranch() && "Should be a branching condition.");
1324 if (isEndLoopN(opcode))
1326 unsigned NewOpcode = getInvertedPredicatedOpcode(opcode);
1327 Cond[0].setImm(NewOpcode);
1331 void HexagonInstrInfo::insertNoop(MachineBasicBlock &MBB,
1332 MachineBasicBlock::iterator MI) const {
1334 BuildMI(MBB, MI, DL, get(Hexagon::A2_nop));
1337 bool HexagonInstrInfo::isPostIncrement(const MachineInstr &MI) const {
1338 return getAddrMode(MI) == HexagonII::PostInc;
1341 // Returns true if an instruction is predicated irrespective of the predicate
1342 // sense. For example, all of the following will return true.
1343 // if (p0) R1 = add(R2, R3)
1344 // if (!p0) R1 = add(R2, R3)
1345 // if (p0.new) R1 = add(R2, R3)
1346 // if (!p0.new) R1 = add(R2, R3)
1347 // Note: New-value stores are not included here as in the current
1348 // implementation, we don't need to check their predicate sense.
1349 bool HexagonInstrInfo::isPredicated(const MachineInstr &MI) const {
1350 const uint64_t F = MI.getDesc().TSFlags;
1351 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
1354 bool HexagonInstrInfo::PredicateInstruction(
1355 MachineInstr &MI, ArrayRef<MachineOperand> Cond) const {
1356 if (Cond.empty() || isNewValueJump(Cond[0].getImm()) ||
1357 isEndLoopN(Cond[0].getImm())) {
1358 DEBUG(dbgs() << "\nCannot predicate:"; MI.dump(););
1361 int Opc = MI.getOpcode();
1362 assert (isPredicable(MI) && "Expected predicable instruction");
1363 bool invertJump = predOpcodeHasNot(Cond);
1365 // We have to predicate MI "in place", i.e. after this function returns,
1366 // MI will need to be transformed into a predicated form. To avoid com-
1367 // plicated manipulations with the operands (handling tied operands,
1368 // etc.), build a new temporary instruction, then overwrite MI with it.
1370 MachineBasicBlock &B = *MI.getParent();
1371 DebugLoc DL = MI.getDebugLoc();
1372 unsigned PredOpc = getCondOpcode(Opc, invertJump);
1373 MachineInstrBuilder T = BuildMI(B, MI, DL, get(PredOpc));
1374 unsigned NOp = 0, NumOps = MI.getNumOperands();
1375 while (NOp < NumOps) {
1376 MachineOperand &Op = MI.getOperand(NOp);
1377 if (!Op.isReg() || !Op.isDef() || Op.isImplicit())
1383 unsigned PredReg, PredRegPos, PredRegFlags;
1384 bool GotPredReg = getPredReg(Cond, PredReg, PredRegPos, PredRegFlags);
1387 T.addReg(PredReg, PredRegFlags);
1388 while (NOp < NumOps)
1389 T.addOperand(MI.getOperand(NOp++));
1391 MI.setDesc(get(PredOpc));
1392 while (unsigned n = MI.getNumOperands())
1393 MI.RemoveOperand(n-1);
1394 for (unsigned i = 0, n = T->getNumOperands(); i < n; ++i)
1395 MI.addOperand(T->getOperand(i));
1397 MachineBasicBlock::instr_iterator TI = T->getIterator();
1400 MachineRegisterInfo &MRI = B.getParent()->getRegInfo();
1401 MRI.clearKillFlags(PredReg);
1405 bool HexagonInstrInfo::SubsumesPredicate(ArrayRef<MachineOperand> Pred1,
1406 ArrayRef<MachineOperand> Pred2) const {
1411 bool HexagonInstrInfo::DefinesPredicate(
1412 MachineInstr &MI, std::vector<MachineOperand> &Pred) const {
1413 auto &HRI = getRegisterInfo();
1414 for (unsigned oper = 0; oper < MI.getNumOperands(); ++oper) {
1415 MachineOperand MO = MI.getOperand(oper);
1416 if (MO.isReg() && MO.isDef()) {
1417 const TargetRegisterClass* RC = HRI.getMinimalPhysRegClass(MO.getReg());
1418 if (RC == &Hexagon::PredRegsRegClass) {
1427 bool HexagonInstrInfo::isPredicable(MachineInstr &MI) const {
1428 return MI.getDesc().isPredicable();
1431 bool HexagonInstrInfo::isSchedulingBoundary(const MachineInstr &MI,
1432 const MachineBasicBlock *MBB,
1433 const MachineFunction &MF) const {
1434 // Debug info is never a scheduling boundary. It's necessary to be explicit
1435 // due to the special treatment of IT instructions below, otherwise a
1436 // dbg_value followed by an IT will result in the IT instruction being
1437 // considered a scheduling hazard, which is wrong. It should be the actual
1438 // instruction preceding the dbg_value instruction(s), just like it is
1439 // when debug info is not present.
1440 if (MI.isDebugValue())
1443 // Throwing call is a boundary.
1445 // Don't mess around with no return calls.
1446 if (doesNotReturn(MI))
1448 // If any of the block's successors is a landing pad, this could be a
1450 for (auto I : MBB->successors())
1455 // Terminators and labels can't be scheduled around.
1456 if (MI.getDesc().isTerminator() || MI.isPosition())
1459 if (MI.isInlineAsm() && !ScheduleInlineAsm)
1465 /// Measure the specified inline asm to determine an approximation of its
1467 /// Comments (which run till the next SeparatorString or newline) do not
1468 /// count as an instruction.
1469 /// Any other non-whitespace text is considered an instruction, with
1470 /// multiple instructions separated by SeparatorString or newlines.
1471 /// Variable-length instructions are not handled here; this function
1472 /// may be overloaded in the target code to do that.
1473 /// Hexagon counts the number of ##'s and adjust for that many
1474 /// constant exenders.
1475 unsigned HexagonInstrInfo::getInlineAsmLength(const char *Str,
1476 const MCAsmInfo &MAI) const {
1477 StringRef AStr(Str);
1478 // Count the number of instructions in the asm.
1479 bool atInsnStart = true;
1480 unsigned Length = 0;
1481 for (; *Str; ++Str) {
1482 if (*Str == '\n' || strncmp(Str, MAI.getSeparatorString(),
1483 strlen(MAI.getSeparatorString())) == 0)
1485 if (atInsnStart && !std::isspace(static_cast<unsigned char>(*Str))) {
1486 Length += MAI.getMaxInstLength();
1487 atInsnStart = false;
1489 if (atInsnStart && strncmp(Str, MAI.getCommentString().data(),
1490 MAI.getCommentString().size()) == 0)
1491 atInsnStart = false;
1494 // Add to size number of constant extenders seen * 4.
1495 StringRef Occ("##");
1496 Length += AStr.count(Occ)*4;
1500 ScheduleHazardRecognizer*
1501 HexagonInstrInfo::CreateTargetPostRAHazardRecognizer(
1502 const InstrItineraryData *II, const ScheduleDAG *DAG) const {
1503 if (UseDFAHazardRec) {
1504 auto &HST = DAG->MF.getSubtarget<HexagonSubtarget>();
1505 return new HexagonHazardRecognizer(II, this, HST);
1507 return TargetInstrInfo::CreateTargetPostRAHazardRecognizer(II, DAG);
1510 /// \brief For a comparison instruction, return the source registers in
1511 /// \p SrcReg and \p SrcReg2 if having two register operands, and the value it
1512 /// compares against in CmpValue. Return true if the comparison instruction
1513 /// can be analyzed.
1514 bool HexagonInstrInfo::analyzeCompare(const MachineInstr &MI, unsigned &SrcReg,
1515 unsigned &SrcReg2, int &Mask,
1517 unsigned Opc = MI.getOpcode();
1519 // Set mask and the first source register.
1521 case Hexagon::C2_cmpeq:
1522 case Hexagon::C2_cmpeqp:
1523 case Hexagon::C2_cmpgt:
1524 case Hexagon::C2_cmpgtp:
1525 case Hexagon::C2_cmpgtu:
1526 case Hexagon::C2_cmpgtup:
1527 case Hexagon::C4_cmpneq:
1528 case Hexagon::C4_cmplte:
1529 case Hexagon::C4_cmplteu:
1530 case Hexagon::C2_cmpeqi:
1531 case Hexagon::C2_cmpgti:
1532 case Hexagon::C2_cmpgtui:
1533 case Hexagon::C4_cmpneqi:
1534 case Hexagon::C4_cmplteui:
1535 case Hexagon::C4_cmpltei:
1536 SrcReg = MI.getOperand(1).getReg();
1539 case Hexagon::A4_cmpbeq:
1540 case Hexagon::A4_cmpbgt:
1541 case Hexagon::A4_cmpbgtu:
1542 case Hexagon::A4_cmpbeqi:
1543 case Hexagon::A4_cmpbgti:
1544 case Hexagon::A4_cmpbgtui:
1545 SrcReg = MI.getOperand(1).getReg();
1548 case Hexagon::A4_cmpheq:
1549 case Hexagon::A4_cmphgt:
1550 case Hexagon::A4_cmphgtu:
1551 case Hexagon::A4_cmpheqi:
1552 case Hexagon::A4_cmphgti:
1553 case Hexagon::A4_cmphgtui:
1554 SrcReg = MI.getOperand(1).getReg();
1559 // Set the value/second source register.
1561 case Hexagon::C2_cmpeq:
1562 case Hexagon::C2_cmpeqp:
1563 case Hexagon::C2_cmpgt:
1564 case Hexagon::C2_cmpgtp:
1565 case Hexagon::C2_cmpgtu:
1566 case Hexagon::C2_cmpgtup:
1567 case Hexagon::A4_cmpbeq:
1568 case Hexagon::A4_cmpbgt:
1569 case Hexagon::A4_cmpbgtu:
1570 case Hexagon::A4_cmpheq:
1571 case Hexagon::A4_cmphgt:
1572 case Hexagon::A4_cmphgtu:
1573 case Hexagon::C4_cmpneq:
1574 case Hexagon::C4_cmplte:
1575 case Hexagon::C4_cmplteu:
1576 SrcReg2 = MI.getOperand(2).getReg();
1579 case Hexagon::C2_cmpeqi:
1580 case Hexagon::C2_cmpgtui:
1581 case Hexagon::C2_cmpgti:
1582 case Hexagon::C4_cmpneqi:
1583 case Hexagon::C4_cmplteui:
1584 case Hexagon::C4_cmpltei:
1585 case Hexagon::A4_cmpbeqi:
1586 case Hexagon::A4_cmpbgti:
1587 case Hexagon::A4_cmpbgtui:
1588 case Hexagon::A4_cmpheqi:
1589 case Hexagon::A4_cmphgti:
1590 case Hexagon::A4_cmphgtui:
1592 Value = MI.getOperand(2).getImm();
1599 unsigned HexagonInstrInfo::getInstrLatency(const InstrItineraryData *ItinData,
1600 const MachineInstr &MI,
1601 unsigned *PredCost) const {
1602 return getInstrTimingClassLatency(ItinData, MI);
1605 DFAPacketizer *HexagonInstrInfo::CreateTargetScheduleState(
1606 const TargetSubtargetInfo &STI) const {
1607 const InstrItineraryData *II = STI.getInstrItineraryData();
1608 return static_cast<const HexagonSubtarget&>(STI).createDFAPacketizer(II);
1611 // Inspired by this pair:
1612 // %R13<def> = L2_loadri_io %R29, 136; mem:LD4[FixedStack0]
1613 // S2_storeri_io %R29, 132, %R1<kill>; flags: mem:ST4[FixedStack1]
1614 // Currently AA considers the addresses in these instructions to be aliasing.
1615 bool HexagonInstrInfo::areMemAccessesTriviallyDisjoint(
1616 MachineInstr &MIa, MachineInstr &MIb, AliasAnalysis *AA) const {
1617 int OffsetA = 0, OffsetB = 0;
1618 unsigned SizeA = 0, SizeB = 0;
1620 if (MIa.hasUnmodeledSideEffects() || MIb.hasUnmodeledSideEffects() ||
1621 MIa.hasOrderedMemoryRef() || MIb.hasOrderedMemoryRef())
1624 // Instructions that are pure loads, not loads and stores like memops are not
1626 if (MIa.mayLoad() && !isMemOp(MIa) && MIb.mayLoad() && !isMemOp(MIb))
1629 // Get base, offset, and access size in MIa.
1630 unsigned BaseRegA = getBaseAndOffset(MIa, OffsetA, SizeA);
1631 if (!BaseRegA || !SizeA)
1634 // Get base, offset, and access size in MIb.
1635 unsigned BaseRegB = getBaseAndOffset(MIb, OffsetB, SizeB);
1636 if (!BaseRegB || !SizeB)
1639 if (BaseRegA != BaseRegB)
1642 // This is a mem access with the same base register and known offsets from it.
1644 if (OffsetA > OffsetB) {
1645 uint64_t offDiff = (uint64_t)((int64_t)OffsetA - (int64_t)OffsetB);
1646 return (SizeB <= offDiff);
1647 } else if (OffsetA < OffsetB) {
1648 uint64_t offDiff = (uint64_t)((int64_t)OffsetB - (int64_t)OffsetA);
1649 return (SizeA <= offDiff);
1655 /// If the instruction is an increment of a constant value, return the amount.
1656 bool HexagonInstrInfo::getIncrementValue(const MachineInstr &MI,
1658 if (isPostIncrement(MI)) {
1659 unsigned AccessSize;
1660 return getBaseAndOffset(MI, Value, AccessSize);
1662 if (MI.getOpcode() == Hexagon::A2_addi) {
1663 Value = MI.getOperand(2).getImm();
1670 unsigned HexagonInstrInfo::createVR(MachineFunction *MF, MVT VT) const {
1671 MachineRegisterInfo &MRI = MF->getRegInfo();
1672 const TargetRegisterClass *TRC;
1673 if (VT == MVT::i1) {
1674 TRC = &Hexagon::PredRegsRegClass;
1675 } else if (VT == MVT::i32 || VT == MVT::f32) {
1676 TRC = &Hexagon::IntRegsRegClass;
1677 } else if (VT == MVT::i64 || VT == MVT::f64) {
1678 TRC = &Hexagon::DoubleRegsRegClass;
1680 llvm_unreachable("Cannot handle this register class");
1683 unsigned NewReg = MRI.createVirtualRegister(TRC);
1687 bool HexagonInstrInfo::isAbsoluteSet(const MachineInstr &MI) const {
1688 return (getAddrMode(MI) == HexagonII::AbsoluteSet);
1691 bool HexagonInstrInfo::isAccumulator(const MachineInstr &MI) const {
1692 const uint64_t F = MI.getDesc().TSFlags;
1693 return((F >> HexagonII::AccumulatorPos) & HexagonII::AccumulatorMask);
1696 bool HexagonInstrInfo::isComplex(const MachineInstr &MI) const {
1697 const MachineFunction *MF = MI.getParent()->getParent();
1698 const TargetInstrInfo *TII = MF->getSubtarget().getInstrInfo();
1699 const HexagonInstrInfo *QII = (const HexagonInstrInfo *) TII;
1702 && !(QII->isTC2Early(MI))
1703 && !(MI.getDesc().mayLoad())
1704 && !(MI.getDesc().mayStore())
1705 && (MI.getDesc().getOpcode() != Hexagon::S2_allocframe)
1706 && (MI.getDesc().getOpcode() != Hexagon::L2_deallocframe)
1707 && !(QII->isMemOp(MI))
1716 // Return true if the instruction is a compund branch instruction.
1717 bool HexagonInstrInfo::isCompoundBranchInstr(const MachineInstr &MI) const {
1718 return (getType(MI) == HexagonII::TypeCOMPOUND && MI.isBranch());
1721 bool HexagonInstrInfo::isCondInst(const MachineInstr &MI) const {
1722 return (MI.isBranch() && isPredicated(MI)) ||
1723 isConditionalTransfer(MI) ||
1724 isConditionalALU32(MI) ||
1725 isConditionalLoad(MI) ||
1726 // Predicated stores which don't have a .new on any operands.
1727 (MI.mayStore() && isPredicated(MI) && !isNewValueStore(MI) &&
1728 !isPredicatedNew(MI));
1731 bool HexagonInstrInfo::isConditionalALU32(const MachineInstr &MI) const {
1732 switch (MI.getOpcode()) {
1733 case Hexagon::A2_paddf:
1734 case Hexagon::A2_paddfnew:
1735 case Hexagon::A2_paddif:
1736 case Hexagon::A2_paddifnew:
1737 case Hexagon::A2_paddit:
1738 case Hexagon::A2_padditnew:
1739 case Hexagon::A2_paddt:
1740 case Hexagon::A2_paddtnew:
1741 case Hexagon::A2_pandf:
1742 case Hexagon::A2_pandfnew:
1743 case Hexagon::A2_pandt:
1744 case Hexagon::A2_pandtnew:
1745 case Hexagon::A2_porf:
1746 case Hexagon::A2_porfnew:
1747 case Hexagon::A2_port:
1748 case Hexagon::A2_portnew:
1749 case Hexagon::A2_psubf:
1750 case Hexagon::A2_psubfnew:
1751 case Hexagon::A2_psubt:
1752 case Hexagon::A2_psubtnew:
1753 case Hexagon::A2_pxorf:
1754 case Hexagon::A2_pxorfnew:
1755 case Hexagon::A2_pxort:
1756 case Hexagon::A2_pxortnew:
1757 case Hexagon::A4_paslhf:
1758 case Hexagon::A4_paslhfnew:
1759 case Hexagon::A4_paslht:
1760 case Hexagon::A4_paslhtnew:
1761 case Hexagon::A4_pasrhf:
1762 case Hexagon::A4_pasrhfnew:
1763 case Hexagon::A4_pasrht:
1764 case Hexagon::A4_pasrhtnew:
1765 case Hexagon::A4_psxtbf:
1766 case Hexagon::A4_psxtbfnew:
1767 case Hexagon::A4_psxtbt:
1768 case Hexagon::A4_psxtbtnew:
1769 case Hexagon::A4_psxthf:
1770 case Hexagon::A4_psxthfnew:
1771 case Hexagon::A4_psxtht:
1772 case Hexagon::A4_psxthtnew:
1773 case Hexagon::A4_pzxtbf:
1774 case Hexagon::A4_pzxtbfnew:
1775 case Hexagon::A4_pzxtbt:
1776 case Hexagon::A4_pzxtbtnew:
1777 case Hexagon::A4_pzxthf:
1778 case Hexagon::A4_pzxthfnew:
1779 case Hexagon::A4_pzxtht:
1780 case Hexagon::A4_pzxthtnew:
1781 case Hexagon::C2_ccombinewf:
1782 case Hexagon::C2_ccombinewt:
1788 // FIXME - Function name and it's functionality don't match.
1789 // It should be renamed to hasPredNewOpcode()
1790 bool HexagonInstrInfo::isConditionalLoad(const MachineInstr &MI) const {
1791 if (!MI.getDesc().mayLoad() || !isPredicated(MI))
1794 int PNewOpcode = Hexagon::getPredNewOpcode(MI.getOpcode());
1795 // Instruction with valid predicated-new opcode can be promoted to .new.
1796 return PNewOpcode >= 0;
1799 // Returns true if an instruction is a conditional store.
1801 // Note: It doesn't include conditional new-value stores as they can't be
1802 // converted to .new predicate.
1803 bool HexagonInstrInfo::isConditionalStore(const MachineInstr &MI) const {
1804 switch (MI.getOpcode()) {
1805 default: return false;
1806 case Hexagon::S4_storeirbt_io:
1807 case Hexagon::S4_storeirbf_io:
1808 case Hexagon::S4_pstorerbt_rr:
1809 case Hexagon::S4_pstorerbf_rr:
1810 case Hexagon::S2_pstorerbt_io:
1811 case Hexagon::S2_pstorerbf_io:
1812 case Hexagon::S2_pstorerbt_pi:
1813 case Hexagon::S2_pstorerbf_pi:
1814 case Hexagon::S2_pstorerdt_io:
1815 case Hexagon::S2_pstorerdf_io:
1816 case Hexagon::S4_pstorerdt_rr:
1817 case Hexagon::S4_pstorerdf_rr:
1818 case Hexagon::S2_pstorerdt_pi:
1819 case Hexagon::S2_pstorerdf_pi:
1820 case Hexagon::S2_pstorerht_io:
1821 case Hexagon::S2_pstorerhf_io:
1822 case Hexagon::S4_storeirht_io:
1823 case Hexagon::S4_storeirhf_io:
1824 case Hexagon::S4_pstorerht_rr:
1825 case Hexagon::S4_pstorerhf_rr:
1826 case Hexagon::S2_pstorerht_pi:
1827 case Hexagon::S2_pstorerhf_pi:
1828 case Hexagon::S2_pstorerit_io:
1829 case Hexagon::S2_pstorerif_io:
1830 case Hexagon::S4_storeirit_io:
1831 case Hexagon::S4_storeirif_io:
1832 case Hexagon::S4_pstorerit_rr:
1833 case Hexagon::S4_pstorerif_rr:
1834 case Hexagon::S2_pstorerit_pi:
1835 case Hexagon::S2_pstorerif_pi:
1837 // V4 global address store before promoting to dot new.
1838 case Hexagon::S4_pstorerdt_abs:
1839 case Hexagon::S4_pstorerdf_abs:
1840 case Hexagon::S4_pstorerbt_abs:
1841 case Hexagon::S4_pstorerbf_abs:
1842 case Hexagon::S4_pstorerht_abs:
1843 case Hexagon::S4_pstorerhf_abs:
1844 case Hexagon::S4_pstorerit_abs:
1845 case Hexagon::S4_pstorerif_abs:
1848 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded
1849 // from the "Conditional Store" list. Because a predicated new value store
1850 // would NOT be promoted to a double dot new store.
1851 // This function returns yes for those stores that are predicated but not
1852 // yet promoted to predicate dot new instructions.
1856 bool HexagonInstrInfo::isConditionalTransfer(const MachineInstr &MI) const {
1857 switch (MI.getOpcode()) {
1858 case Hexagon::A2_tfrt:
1859 case Hexagon::A2_tfrf:
1860 case Hexagon::C2_cmoveit:
1861 case Hexagon::C2_cmoveif:
1862 case Hexagon::A2_tfrtnew:
1863 case Hexagon::A2_tfrfnew:
1864 case Hexagon::C2_cmovenewit:
1865 case Hexagon::C2_cmovenewif:
1866 case Hexagon::A2_tfrpt:
1867 case Hexagon::A2_tfrpf:
1876 // TODO: In order to have isExtendable for fpimm/f32Ext, we need to handle
1877 // isFPImm and later getFPImm as well.
1878 bool HexagonInstrInfo::isConstExtended(const MachineInstr &MI) const {
1879 const uint64_t F = MI.getDesc().TSFlags;
1880 unsigned isExtended = (F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask;
1881 if (isExtended) // Instruction must be extended.
1884 unsigned isExtendable =
1885 (F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask;
1892 short ExtOpNum = getCExtOpNum(MI);
1893 const MachineOperand &MO = MI.getOperand(ExtOpNum);
1894 // Use MO operand flags to determine if MO
1895 // has the HMOTF_ConstExtended flag set.
1896 if (MO.getTargetFlags() && HexagonII::HMOTF_ConstExtended)
1898 // If this is a Machine BB address we are talking about, and it is
1899 // not marked as extended, say so.
1903 // We could be using an instruction with an extendable immediate and shoehorn
1904 // a global address into it. If it is a global address it will be constant
1905 // extended. We do this for COMBINE.
1906 // We currently only handle isGlobal() because it is the only kind of
1907 // object we are going to end up with here for now.
1908 // In the future we probably should add isSymbol(), etc.
1909 if (MO.isGlobal() || MO.isSymbol() || MO.isBlockAddress() ||
1910 MO.isJTI() || MO.isCPI() || MO.isFPImm())
1913 // If the extendable operand is not 'Immediate' type, the instruction should
1914 // have 'isExtended' flag set.
1915 assert(MO.isImm() && "Extendable operand must be Immediate type");
1917 int MinValue = getMinValue(MI);
1918 int MaxValue = getMaxValue(MI);
1919 int ImmValue = MO.getImm();
1921 return (ImmValue < MinValue || ImmValue > MaxValue);
1924 bool HexagonInstrInfo::isDeallocRet(const MachineInstr &MI) const {
1925 switch (MI.getOpcode()) {
1926 case Hexagon::L4_return :
1927 case Hexagon::L4_return_t :
1928 case Hexagon::L4_return_f :
1929 case Hexagon::L4_return_tnew_pnt :
1930 case Hexagon::L4_return_fnew_pnt :
1931 case Hexagon::L4_return_tnew_pt :
1932 case Hexagon::L4_return_fnew_pt :
1938 // Return true when ConsMI uses a register defined by ProdMI.
1939 bool HexagonInstrInfo::isDependent(const MachineInstr &ProdMI,
1940 const MachineInstr &ConsMI) const {
1941 if (!ProdMI.getDesc().getNumDefs())
1944 auto &HRI = getRegisterInfo();
1946 SmallVector<unsigned, 4> DefsA;
1947 SmallVector<unsigned, 4> DefsB;
1948 SmallVector<unsigned, 8> UsesA;
1949 SmallVector<unsigned, 8> UsesB;
1951 parseOperands(ProdMI, DefsA, UsesA);
1952 parseOperands(ConsMI, DefsB, UsesB);
1954 for (auto &RegA : DefsA)
1955 for (auto &RegB : UsesB) {
1956 // True data dependency.
1960 if (Hexagon::DoubleRegsRegClass.contains(RegA))
1961 for (MCSubRegIterator SubRegs(RegA, &HRI); SubRegs.isValid(); ++SubRegs)
1962 if (RegB == *SubRegs)
1965 if (Hexagon::DoubleRegsRegClass.contains(RegB))
1966 for (MCSubRegIterator SubRegs(RegB, &HRI); SubRegs.isValid(); ++SubRegs)
1967 if (RegA == *SubRegs)
1974 // Returns true if the instruction is alread a .cur.
1975 bool HexagonInstrInfo::isDotCurInst(const MachineInstr &MI) const {
1976 switch (MI.getOpcode()) {
1977 case Hexagon::V6_vL32b_cur_pi:
1978 case Hexagon::V6_vL32b_cur_ai:
1979 case Hexagon::V6_vL32b_cur_pi_128B:
1980 case Hexagon::V6_vL32b_cur_ai_128B:
1986 // Returns true, if any one of the operands is a dot new
1987 // insn, whether it is predicated dot new or register dot new.
1988 bool HexagonInstrInfo::isDotNewInst(const MachineInstr &MI) const {
1989 if (isNewValueInst(MI) || (isPredicated(MI) && isPredicatedNew(MI)))
1995 /// Symmetrical. See if these two instructions are fit for duplex pair.
1996 bool HexagonInstrInfo::isDuplexPair(const MachineInstr &MIa,
1997 const MachineInstr &MIb) const {
1998 HexagonII::SubInstructionGroup MIaG = getDuplexCandidateGroup(MIa);
1999 HexagonII::SubInstructionGroup MIbG = getDuplexCandidateGroup(MIb);
2000 return (isDuplexPairMatch(MIaG, MIbG) || isDuplexPairMatch(MIbG, MIaG));
2003 bool HexagonInstrInfo::isEarlySourceInstr(const MachineInstr &MI) const {
2004 if (MI.mayLoad() || MI.mayStore() || MI.isCompare())
2008 unsigned SchedClass = MI.getDesc().getSchedClass();
2009 if (SchedClass == Hexagon::Sched::M_tc_3or4x_SLOT23)
2014 bool HexagonInstrInfo::isEndLoopN(unsigned Opcode) const {
2015 return (Opcode == Hexagon::ENDLOOP0 ||
2016 Opcode == Hexagon::ENDLOOP1);
2019 bool HexagonInstrInfo::isExpr(unsigned OpType) const {
2021 case MachineOperand::MO_MachineBasicBlock:
2022 case MachineOperand::MO_GlobalAddress:
2023 case MachineOperand::MO_ExternalSymbol:
2024 case MachineOperand::MO_JumpTableIndex:
2025 case MachineOperand::MO_ConstantPoolIndex:
2026 case MachineOperand::MO_BlockAddress:
2033 bool HexagonInstrInfo::isExtendable(const MachineInstr &MI) const {
2034 const MCInstrDesc &MID = MI.getDesc();
2035 const uint64_t F = MID.TSFlags;
2036 if ((F >> HexagonII::ExtendablePos) & HexagonII::ExtendableMask)
2039 // TODO: This is largely obsolete now. Will need to be removed
2040 // in consecutive patches.
2041 switch (MI.getOpcode()) {
2042 // PS_fi and PS_fia remain special cases.
2043 case Hexagon::PS_fi:
2044 case Hexagon::PS_fia:
2052 // This returns true in two cases:
2053 // - The OP code itself indicates that this is an extended instruction.
2054 // - One of MOs has been marked with HMOTF_ConstExtended flag.
2055 bool HexagonInstrInfo::isExtended(const MachineInstr &MI) const {
2056 // First check if this is permanently extended op code.
2057 const uint64_t F = MI.getDesc().TSFlags;
2058 if ((F >> HexagonII::ExtendedPos) & HexagonII::ExtendedMask)
2060 // Use MO operand flags to determine if one of MI's operands
2061 // has HMOTF_ConstExtended flag set.
2062 for (MachineInstr::const_mop_iterator I = MI.operands_begin(),
2063 E = MI.operands_end(); I != E; ++I) {
2064 if (I->getTargetFlags() && HexagonII::HMOTF_ConstExtended)
2070 bool HexagonInstrInfo::isFloat(const MachineInstr &MI) const {
2071 unsigned Opcode = MI.getOpcode();
2072 const uint64_t F = get(Opcode).TSFlags;
2073 return (F >> HexagonII::FPPos) & HexagonII::FPMask;
2076 // No V60 HVX VMEM with A_INDIRECT.
2077 bool HexagonInstrInfo::isHVXMemWithAIndirect(const MachineInstr &I,
2078 const MachineInstr &J) const {
2079 if (!isV60VectorInstruction(I))
2081 if (!I.mayLoad() && !I.mayStore())
2083 return J.isIndirectBranch() || isIndirectCall(J) || isIndirectL4Return(J);
2086 bool HexagonInstrInfo::isIndirectCall(const MachineInstr &MI) const {
2087 switch (MI.getOpcode()) {
2088 case Hexagon::J2_callr :
2089 case Hexagon::J2_callrf :
2090 case Hexagon::J2_callrt :
2091 case Hexagon::PS_call_nr :
2097 bool HexagonInstrInfo::isIndirectL4Return(const MachineInstr &MI) const {
2098 switch (MI.getOpcode()) {
2099 case Hexagon::L4_return :
2100 case Hexagon::L4_return_t :
2101 case Hexagon::L4_return_f :
2102 case Hexagon::L4_return_fnew_pnt :
2103 case Hexagon::L4_return_fnew_pt :
2104 case Hexagon::L4_return_tnew_pnt :
2105 case Hexagon::L4_return_tnew_pt :
2111 bool HexagonInstrInfo::isJumpR(const MachineInstr &MI) const {
2112 switch (MI.getOpcode()) {
2113 case Hexagon::J2_jumpr :
2114 case Hexagon::J2_jumprt :
2115 case Hexagon::J2_jumprf :
2116 case Hexagon::J2_jumprtnewpt :
2117 case Hexagon::J2_jumprfnewpt :
2118 case Hexagon::J2_jumprtnew :
2119 case Hexagon::J2_jumprfnew :
2125 // Return true if a given MI can accommodate given offset.
2126 // Use abs estimate as oppose to the exact number.
2127 // TODO: This will need to be changed to use MC level
2128 // definition of instruction extendable field size.
2129 bool HexagonInstrInfo::isJumpWithinBranchRange(const MachineInstr &MI,
2130 unsigned offset) const {
2131 // This selection of jump instructions matches to that what
2132 // AnalyzeBranch can parse, plus NVJ.
2133 if (isNewValueJump(MI)) // r9:2
2134 return isInt<11>(offset);
2136 switch (MI.getOpcode()) {
2137 // Still missing Jump to address condition on register value.
2140 case Hexagon::J2_jump: // bits<24> dst; // r22:2
2141 case Hexagon::J2_call:
2142 case Hexagon::PS_call_nr:
2143 return isInt<24>(offset);
2144 case Hexagon::J2_jumpt: //bits<17> dst; // r15:2
2145 case Hexagon::J2_jumpf:
2146 case Hexagon::J2_jumptnew:
2147 case Hexagon::J2_jumptnewpt:
2148 case Hexagon::J2_jumpfnew:
2149 case Hexagon::J2_jumpfnewpt:
2150 case Hexagon::J2_callt:
2151 case Hexagon::J2_callf:
2152 return isInt<17>(offset);
2153 case Hexagon::J2_loop0i:
2154 case Hexagon::J2_loop0iext:
2155 case Hexagon::J2_loop0r:
2156 case Hexagon::J2_loop0rext:
2157 case Hexagon::J2_loop1i:
2158 case Hexagon::J2_loop1iext:
2159 case Hexagon::J2_loop1r:
2160 case Hexagon::J2_loop1rext:
2161 return isInt<9>(offset);
2162 // TODO: Add all the compound branches here. Can we do this in Relation model?
2163 case Hexagon::J4_cmpeqi_tp0_jump_nt:
2164 case Hexagon::J4_cmpeqi_tp1_jump_nt:
2165 return isInt<11>(offset);
2169 bool HexagonInstrInfo::isLateInstrFeedsEarlyInstr(const MachineInstr &LRMI,
2170 const MachineInstr &ESMI) const {
2171 bool isLate = isLateResultInstr(LRMI);
2172 bool isEarly = isEarlySourceInstr(ESMI);
2174 DEBUG(dbgs() << "V60" << (isLate ? "-LR " : " -- "));
2176 DEBUG(dbgs() << "V60" << (isEarly ? "-ES " : " -- "));
2179 if (isLate && isEarly) {
2180 DEBUG(dbgs() << "++Is Late Result feeding Early Source\n");
2187 bool HexagonInstrInfo::isLateResultInstr(const MachineInstr &MI) const {
2188 switch (MI.getOpcode()) {
2189 case TargetOpcode::EXTRACT_SUBREG:
2190 case TargetOpcode::INSERT_SUBREG:
2191 case TargetOpcode::SUBREG_TO_REG:
2192 case TargetOpcode::REG_SEQUENCE:
2193 case TargetOpcode::IMPLICIT_DEF:
2194 case TargetOpcode::COPY:
2195 case TargetOpcode::INLINEASM:
2196 case TargetOpcode::PHI:
2202 unsigned SchedClass = MI.getDesc().getSchedClass();
2204 switch (SchedClass) {
2205 case Hexagon::Sched::ALU32_2op_tc_1_SLOT0123:
2206 case Hexagon::Sched::ALU32_3op_tc_1_SLOT0123:
2207 case Hexagon::Sched::ALU32_ADDI_tc_1_SLOT0123:
2208 case Hexagon::Sched::ALU64_tc_1_SLOT23:
2209 case Hexagon::Sched::EXTENDER_tc_1_SLOT0123:
2210 case Hexagon::Sched::S_2op_tc_1_SLOT23:
2211 case Hexagon::Sched::S_3op_tc_1_SLOT23:
2212 case Hexagon::Sched::V2LDST_tc_ld_SLOT01:
2213 case Hexagon::Sched::V2LDST_tc_st_SLOT0:
2214 case Hexagon::Sched::V2LDST_tc_st_SLOT01:
2215 case Hexagon::Sched::V4LDST_tc_ld_SLOT01:
2216 case Hexagon::Sched::V4LDST_tc_st_SLOT0:
2217 case Hexagon::Sched::V4LDST_tc_st_SLOT01:
2223 bool HexagonInstrInfo::isLateSourceInstr(const MachineInstr &MI) const {
2224 // Instructions with iclass A_CVI_VX and attribute A_CVI_LATE uses a multiply
2225 // resource, but all operands can be received late like an ALU instruction.
2226 return MI.getDesc().getSchedClass() == Hexagon::Sched::CVI_VX_LATE;
2229 bool HexagonInstrInfo::isLoopN(const MachineInstr &MI) const {
2230 unsigned Opcode = MI.getOpcode();
2231 return Opcode == Hexagon::J2_loop0i ||
2232 Opcode == Hexagon::J2_loop0r ||
2233 Opcode == Hexagon::J2_loop0iext ||
2234 Opcode == Hexagon::J2_loop0rext ||
2235 Opcode == Hexagon::J2_loop1i ||
2236 Opcode == Hexagon::J2_loop1r ||
2237 Opcode == Hexagon::J2_loop1iext ||
2238 Opcode == Hexagon::J2_loop1rext;
2241 bool HexagonInstrInfo::isMemOp(const MachineInstr &MI) const {
2242 switch (MI.getOpcode()) {
2243 default: return false;
2244 case Hexagon::L4_iadd_memopw_io :
2245 case Hexagon::L4_isub_memopw_io :
2246 case Hexagon::L4_add_memopw_io :
2247 case Hexagon::L4_sub_memopw_io :
2248 case Hexagon::L4_and_memopw_io :
2249 case Hexagon::L4_or_memopw_io :
2250 case Hexagon::L4_iadd_memoph_io :
2251 case Hexagon::L4_isub_memoph_io :
2252 case Hexagon::L4_add_memoph_io :
2253 case Hexagon::L4_sub_memoph_io :
2254 case Hexagon::L4_and_memoph_io :
2255 case Hexagon::L4_or_memoph_io :
2256 case Hexagon::L4_iadd_memopb_io :
2257 case Hexagon::L4_isub_memopb_io :
2258 case Hexagon::L4_add_memopb_io :
2259 case Hexagon::L4_sub_memopb_io :
2260 case Hexagon::L4_and_memopb_io :
2261 case Hexagon::L4_or_memopb_io :
2262 case Hexagon::L4_ior_memopb_io:
2263 case Hexagon::L4_ior_memoph_io:
2264 case Hexagon::L4_ior_memopw_io:
2265 case Hexagon::L4_iand_memopb_io:
2266 case Hexagon::L4_iand_memoph_io:
2267 case Hexagon::L4_iand_memopw_io:
2273 bool HexagonInstrInfo::isNewValue(const MachineInstr &MI) const {
2274 const uint64_t F = MI.getDesc().TSFlags;
2275 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
2278 bool HexagonInstrInfo::isNewValue(unsigned Opcode) const {
2279 const uint64_t F = get(Opcode).TSFlags;
2280 return (F >> HexagonII::NewValuePos) & HexagonII::NewValueMask;
2283 bool HexagonInstrInfo::isNewValueInst(const MachineInstr &MI) const {
2284 return isNewValueJump(MI) || isNewValueStore(MI);
2287 bool HexagonInstrInfo::isNewValueJump(const MachineInstr &MI) const {
2288 return isNewValue(MI) && MI.isBranch();
2291 bool HexagonInstrInfo::isNewValueJump(unsigned Opcode) const {
2292 return isNewValue(Opcode) && get(Opcode).isBranch() && isPredicated(Opcode);
2295 bool HexagonInstrInfo::isNewValueStore(const MachineInstr &MI) const {
2296 const uint64_t F = MI.getDesc().TSFlags;
2297 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
2300 bool HexagonInstrInfo::isNewValueStore(unsigned Opcode) const {
2301 const uint64_t F = get(Opcode).TSFlags;
2302 return (F >> HexagonII::NVStorePos) & HexagonII::NVStoreMask;
2305 // Returns true if a particular operand is extendable for an instruction.
2306 bool HexagonInstrInfo::isOperandExtended(const MachineInstr &MI,
2307 unsigned OperandNum) const {
2308 const uint64_t F = MI.getDesc().TSFlags;
2309 return ((F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask)
2313 bool HexagonInstrInfo::isPredicatedNew(const MachineInstr &MI) const {
2314 const uint64_t F = MI.getDesc().TSFlags;
2315 assert(isPredicated(MI));
2316 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
2319 bool HexagonInstrInfo::isPredicatedNew(unsigned Opcode) const {
2320 const uint64_t F = get(Opcode).TSFlags;
2321 assert(isPredicated(Opcode));
2322 return (F >> HexagonII::PredicatedNewPos) & HexagonII::PredicatedNewMask;
2325 bool HexagonInstrInfo::isPredicatedTrue(const MachineInstr &MI) const {
2326 const uint64_t F = MI.getDesc().TSFlags;
2327 return !((F >> HexagonII::PredicatedFalsePos) &
2328 HexagonII::PredicatedFalseMask);
2331 bool HexagonInstrInfo::isPredicatedTrue(unsigned Opcode) const {
2332 const uint64_t F = get(Opcode).TSFlags;
2333 // Make sure that the instruction is predicated.
2334 assert((F>> HexagonII::PredicatedPos) & HexagonII::PredicatedMask);
2335 return !((F >> HexagonII::PredicatedFalsePos) &
2336 HexagonII::PredicatedFalseMask);
2339 bool HexagonInstrInfo::isPredicated(unsigned Opcode) const {
2340 const uint64_t F = get(Opcode).TSFlags;
2341 return (F >> HexagonII::PredicatedPos) & HexagonII::PredicatedMask;
2344 bool HexagonInstrInfo::isPredicateLate(unsigned Opcode) const {
2345 const uint64_t F = get(Opcode).TSFlags;
2346 return ~(F >> HexagonII::PredicateLatePos) & HexagonII::PredicateLateMask;
2349 bool HexagonInstrInfo::isPredictedTaken(unsigned Opcode) const {
2350 const uint64_t F = get(Opcode).TSFlags;
2351 assert(get(Opcode).isBranch() &&
2352 (isPredicatedNew(Opcode) || isNewValue(Opcode)));
2353 return (F >> HexagonII::TakenPos) & HexagonII::TakenMask;
2356 bool HexagonInstrInfo::isSaveCalleeSavedRegsCall(const MachineInstr &MI) const {
2357 return MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4 ||
2358 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT ||
2359 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_PIC ||
2360 MI.getOpcode() == Hexagon::SAVE_REGISTERS_CALL_V4_EXT_PIC;
2363 bool HexagonInstrInfo::isSignExtendingLoad(const MachineInstr &MI) const {
2364 switch (MI.getOpcode()) {
2366 case Hexagon::L2_loadrb_io:
2367 case Hexagon::L4_loadrb_ur:
2368 case Hexagon::L4_loadrb_ap:
2369 case Hexagon::L2_loadrb_pr:
2370 case Hexagon::L2_loadrb_pbr:
2371 case Hexagon::L2_loadrb_pi:
2372 case Hexagon::L2_loadrb_pci:
2373 case Hexagon::L2_loadrb_pcr:
2374 case Hexagon::L2_loadbsw2_io:
2375 case Hexagon::L4_loadbsw2_ur:
2376 case Hexagon::L4_loadbsw2_ap:
2377 case Hexagon::L2_loadbsw2_pr:
2378 case Hexagon::L2_loadbsw2_pbr:
2379 case Hexagon::L2_loadbsw2_pi:
2380 case Hexagon::L2_loadbsw2_pci:
2381 case Hexagon::L2_loadbsw2_pcr:
2382 case Hexagon::L2_loadbsw4_io:
2383 case Hexagon::L4_loadbsw4_ur:
2384 case Hexagon::L4_loadbsw4_ap:
2385 case Hexagon::L2_loadbsw4_pr:
2386 case Hexagon::L2_loadbsw4_pbr:
2387 case Hexagon::L2_loadbsw4_pi:
2388 case Hexagon::L2_loadbsw4_pci:
2389 case Hexagon::L2_loadbsw4_pcr:
2390 case Hexagon::L4_loadrb_rr:
2391 case Hexagon::L2_ploadrbt_io:
2392 case Hexagon::L2_ploadrbt_pi:
2393 case Hexagon::L2_ploadrbf_io:
2394 case Hexagon::L2_ploadrbf_pi:
2395 case Hexagon::L2_ploadrbtnew_io:
2396 case Hexagon::L2_ploadrbfnew_io:
2397 case Hexagon::L4_ploadrbt_rr:
2398 case Hexagon::L4_ploadrbf_rr:
2399 case Hexagon::L4_ploadrbtnew_rr:
2400 case Hexagon::L4_ploadrbfnew_rr:
2401 case Hexagon::L2_ploadrbtnew_pi:
2402 case Hexagon::L2_ploadrbfnew_pi:
2403 case Hexagon::L4_ploadrbt_abs:
2404 case Hexagon::L4_ploadrbf_abs:
2405 case Hexagon::L4_ploadrbtnew_abs:
2406 case Hexagon::L4_ploadrbfnew_abs:
2407 case Hexagon::L2_loadrbgp:
2409 case Hexagon::L2_loadrh_io:
2410 case Hexagon::L4_loadrh_ur:
2411 case Hexagon::L4_loadrh_ap:
2412 case Hexagon::L2_loadrh_pr:
2413 case Hexagon::L2_loadrh_pbr:
2414 case Hexagon::L2_loadrh_pi:
2415 case Hexagon::L2_loadrh_pci:
2416 case Hexagon::L2_loadrh_pcr:
2417 case Hexagon::L4_loadrh_rr:
2418 case Hexagon::L2_ploadrht_io:
2419 case Hexagon::L2_ploadrht_pi:
2420 case Hexagon::L2_ploadrhf_io:
2421 case Hexagon::L2_ploadrhf_pi:
2422 case Hexagon::L2_ploadrhtnew_io:
2423 case Hexagon::L2_ploadrhfnew_io:
2424 case Hexagon::L4_ploadrht_rr:
2425 case Hexagon::L4_ploadrhf_rr:
2426 case Hexagon::L4_ploadrhtnew_rr:
2427 case Hexagon::L4_ploadrhfnew_rr:
2428 case Hexagon::L2_ploadrhtnew_pi:
2429 case Hexagon::L2_ploadrhfnew_pi:
2430 case Hexagon::L4_ploadrht_abs:
2431 case Hexagon::L4_ploadrhf_abs:
2432 case Hexagon::L4_ploadrhtnew_abs:
2433 case Hexagon::L4_ploadrhfnew_abs:
2434 case Hexagon::L2_loadrhgp:
2441 bool HexagonInstrInfo::isSolo(const MachineInstr &MI) const {
2442 const uint64_t F = MI.getDesc().TSFlags;
2443 return (F >> HexagonII::SoloPos) & HexagonII::SoloMask;
2446 bool HexagonInstrInfo::isSpillPredRegOp(const MachineInstr &MI) const {
2447 switch (MI.getOpcode()) {
2448 case Hexagon::STriw_pred :
2449 case Hexagon::LDriw_pred :
2456 bool HexagonInstrInfo::isTailCall(const MachineInstr &MI) const {
2460 for (auto &Op : MI.operands())
2461 if (Op.isGlobal() || Op.isSymbol())
2466 // Returns true when SU has a timing class TC1.
2467 bool HexagonInstrInfo::isTC1(const MachineInstr &MI) const {
2468 unsigned SchedClass = MI.getDesc().getSchedClass();
2469 switch (SchedClass) {
2470 case Hexagon::Sched::ALU32_2op_tc_1_SLOT0123:
2471 case Hexagon::Sched::ALU32_3op_tc_1_SLOT0123:
2472 case Hexagon::Sched::ALU32_ADDI_tc_1_SLOT0123:
2473 case Hexagon::Sched::ALU64_tc_1_SLOT23:
2474 case Hexagon::Sched::EXTENDER_tc_1_SLOT0123:
2475 //case Hexagon::Sched::M_tc_1_SLOT23:
2476 case Hexagon::Sched::S_2op_tc_1_SLOT23:
2477 case Hexagon::Sched::S_3op_tc_1_SLOT23:
2485 bool HexagonInstrInfo::isTC2(const MachineInstr &MI) const {
2486 unsigned SchedClass = MI.getDesc().getSchedClass();
2487 switch (SchedClass) {
2488 case Hexagon::Sched::ALU32_3op_tc_2_SLOT0123:
2489 case Hexagon::Sched::ALU64_tc_2_SLOT23:
2490 case Hexagon::Sched::CR_tc_2_SLOT3:
2491 case Hexagon::Sched::M_tc_2_SLOT23:
2492 case Hexagon::Sched::S_2op_tc_2_SLOT23:
2493 case Hexagon::Sched::S_3op_tc_2_SLOT23:
2501 bool HexagonInstrInfo::isTC2Early(const MachineInstr &MI) const {
2502 unsigned SchedClass = MI.getDesc().getSchedClass();
2503 switch (SchedClass) {
2504 case Hexagon::Sched::ALU32_2op_tc_2early_SLOT0123:
2505 case Hexagon::Sched::ALU32_3op_tc_2early_SLOT0123:
2506 case Hexagon::Sched::ALU64_tc_2early_SLOT23:
2507 case Hexagon::Sched::CR_tc_2early_SLOT23:
2508 case Hexagon::Sched::CR_tc_2early_SLOT3:
2509 case Hexagon::Sched::J_tc_2early_SLOT0123:
2510 case Hexagon::Sched::J_tc_2early_SLOT2:
2511 case Hexagon::Sched::J_tc_2early_SLOT23:
2512 case Hexagon::Sched::S_2op_tc_2early_SLOT23:
2513 case Hexagon::Sched::S_3op_tc_2early_SLOT23:
2521 bool HexagonInstrInfo::isTC4x(const MachineInstr &MI) const {
2522 unsigned SchedClass = MI.getDesc().getSchedClass();
2523 return SchedClass == Hexagon::Sched::M_tc_3or4x_SLOT23;
2526 // Schedule this ASAP.
2527 bool HexagonInstrInfo::isToBeScheduledASAP(const MachineInstr &MI1,
2528 const MachineInstr &MI2) const {
2529 if (mayBeCurLoad(MI1)) {
2530 // if (result of SU is used in Next) return true;
2531 unsigned DstReg = MI1.getOperand(0).getReg();
2532 int N = MI2.getNumOperands();
2533 for (int I = 0; I < N; I++)
2534 if (MI2.getOperand(I).isReg() && DstReg == MI2.getOperand(I).getReg())
2537 if (mayBeNewStore(MI2))
2538 if (MI2.getOpcode() == Hexagon::V6_vS32b_pi)
2539 if (MI1.getOperand(0).isReg() && MI2.getOperand(3).isReg() &&
2540 MI1.getOperand(0).getReg() == MI2.getOperand(3).getReg())
2545 bool HexagonInstrInfo::isV60VectorInstruction(const MachineInstr &MI) const {
2546 const uint64_t V = getType(MI);
2547 return HexagonII::TypeCVI_FIRST <= V && V <= HexagonII::TypeCVI_LAST;
2550 // Check if the Offset is a valid auto-inc imm by Load/Store Type.
2552 bool HexagonInstrInfo::isValidAutoIncImm(const EVT VT, const int Offset) const {
2553 if (VT == MVT::v16i32 || VT == MVT::v8i64 ||
2554 VT == MVT::v32i16 || VT == MVT::v64i8) {
2555 return (Offset >= Hexagon_MEMV_AUTOINC_MIN &&
2556 Offset <= Hexagon_MEMV_AUTOINC_MAX &&
2557 (Offset & 0x3f) == 0);
2560 if (VT == MVT::v32i32 || VT == MVT::v16i64 ||
2561 VT == MVT::v64i16 || VT == MVT::v128i8) {
2562 return (Offset >= Hexagon_MEMV_AUTOINC_MIN_128B &&
2563 Offset <= Hexagon_MEMV_AUTOINC_MAX_128B &&
2564 (Offset & 0x7f) == 0);
2566 if (VT == MVT::i64) {
2567 return (Offset >= Hexagon_MEMD_AUTOINC_MIN &&
2568 Offset <= Hexagon_MEMD_AUTOINC_MAX &&
2569 (Offset & 0x7) == 0);
2571 if (VT == MVT::i32) {
2572 return (Offset >= Hexagon_MEMW_AUTOINC_MIN &&
2573 Offset <= Hexagon_MEMW_AUTOINC_MAX &&
2574 (Offset & 0x3) == 0);
2576 if (VT == MVT::i16) {
2577 return (Offset >= Hexagon_MEMH_AUTOINC_MIN &&
2578 Offset <= Hexagon_MEMH_AUTOINC_MAX &&
2579 (Offset & 0x1) == 0);
2581 if (VT == MVT::i8) {
2582 return (Offset >= Hexagon_MEMB_AUTOINC_MIN &&
2583 Offset <= Hexagon_MEMB_AUTOINC_MAX);
2585 llvm_unreachable("Not an auto-inc opc!");
2588 bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset,
2589 bool Extend) const {
2590 // This function is to check whether the "Offset" is in the correct range of
2591 // the given "Opcode". If "Offset" is not in the correct range, "A2_addi" is
2592 // inserted to calculate the final address. Due to this reason, the function
2593 // assumes that the "Offset" has correct alignment.
2594 // We used to assert if the offset was not properly aligned, however,
2595 // there are cases where a misaligned pointer recast can cause this
2596 // problem, and we need to allow for it. The front end warns of such
2597 // misaligns with respect to load size.
2600 case Hexagon::PS_vstorerq_ai:
2601 case Hexagon::PS_vstorerw_ai:
2602 case Hexagon::PS_vloadrq_ai:
2603 case Hexagon::PS_vloadrw_ai:
2604 case Hexagon::V6_vL32b_ai:
2605 case Hexagon::V6_vS32b_ai:
2606 case Hexagon::V6_vL32Ub_ai:
2607 case Hexagon::V6_vS32Ub_ai:
2608 return (Offset >= Hexagon_MEMV_OFFSET_MIN) &&
2609 (Offset <= Hexagon_MEMV_OFFSET_MAX);
2611 case Hexagon::PS_vstorerq_ai_128B:
2612 case Hexagon::PS_vstorerw_ai_128B:
2613 case Hexagon::PS_vloadrq_ai_128B:
2614 case Hexagon::PS_vloadrw_ai_128B:
2615 case Hexagon::V6_vL32b_ai_128B:
2616 case Hexagon::V6_vS32b_ai_128B:
2617 case Hexagon::V6_vL32Ub_ai_128B:
2618 case Hexagon::V6_vS32Ub_ai_128B:
2619 return (Offset >= Hexagon_MEMV_OFFSET_MIN_128B) &&
2620 (Offset <= Hexagon_MEMV_OFFSET_MAX_128B);
2622 case Hexagon::J2_loop0i:
2623 case Hexagon::J2_loop1i:
2624 return isUInt<10>(Offset);
2626 case Hexagon::S4_storeirb_io:
2627 case Hexagon::S4_storeirbt_io:
2628 case Hexagon::S4_storeirbf_io:
2629 return isUInt<6>(Offset);
2631 case Hexagon::S4_storeirh_io:
2632 case Hexagon::S4_storeirht_io:
2633 case Hexagon::S4_storeirhf_io:
2634 return isShiftedUInt<6,1>(Offset);
2636 case Hexagon::S4_storeiri_io:
2637 case Hexagon::S4_storeirit_io:
2638 case Hexagon::S4_storeirif_io:
2639 return isShiftedUInt<6,2>(Offset);
2646 case Hexagon::L2_loadri_io:
2647 case Hexagon::S2_storeri_io:
2648 return (Offset >= Hexagon_MEMW_OFFSET_MIN) &&
2649 (Offset <= Hexagon_MEMW_OFFSET_MAX);
2651 case Hexagon::L2_loadrd_io:
2652 case Hexagon::S2_storerd_io:
2653 return (Offset >= Hexagon_MEMD_OFFSET_MIN) &&
2654 (Offset <= Hexagon_MEMD_OFFSET_MAX);
2656 case Hexagon::L2_loadrh_io:
2657 case Hexagon::L2_loadruh_io:
2658 case Hexagon::S2_storerh_io:
2659 return (Offset >= Hexagon_MEMH_OFFSET_MIN) &&
2660 (Offset <= Hexagon_MEMH_OFFSET_MAX);
2662 case Hexagon::L2_loadrb_io:
2663 case Hexagon::L2_loadrub_io:
2664 case Hexagon::S2_storerb_io:
2665 return (Offset >= Hexagon_MEMB_OFFSET_MIN) &&
2666 (Offset <= Hexagon_MEMB_OFFSET_MAX);
2668 case Hexagon::A2_addi:
2669 return (Offset >= Hexagon_ADDI_OFFSET_MIN) &&
2670 (Offset <= Hexagon_ADDI_OFFSET_MAX);
2672 case Hexagon::L4_iadd_memopw_io :
2673 case Hexagon::L4_isub_memopw_io :
2674 case Hexagon::L4_add_memopw_io :
2675 case Hexagon::L4_sub_memopw_io :
2676 case Hexagon::L4_and_memopw_io :
2677 case Hexagon::L4_or_memopw_io :
2678 return (0 <= Offset && Offset <= 255);
2680 case Hexagon::L4_iadd_memoph_io :
2681 case Hexagon::L4_isub_memoph_io :
2682 case Hexagon::L4_add_memoph_io :
2683 case Hexagon::L4_sub_memoph_io :
2684 case Hexagon::L4_and_memoph_io :
2685 case Hexagon::L4_or_memoph_io :
2686 return (0 <= Offset && Offset <= 127);
2688 case Hexagon::L4_iadd_memopb_io :
2689 case Hexagon::L4_isub_memopb_io :
2690 case Hexagon::L4_add_memopb_io :
2691 case Hexagon::L4_sub_memopb_io :
2692 case Hexagon::L4_and_memopb_io :
2693 case Hexagon::L4_or_memopb_io :
2694 return (0 <= Offset && Offset <= 63);
2696 // LDriw_xxx and STriw_xxx are pseudo operations, so it has to take offset of
2697 // any size. Later pass knows how to handle it.
2698 case Hexagon::STriw_pred:
2699 case Hexagon::LDriw_pred:
2700 case Hexagon::STriw_mod:
2701 case Hexagon::LDriw_mod:
2704 case Hexagon::PS_fi:
2705 case Hexagon::PS_fia:
2706 case Hexagon::INLINEASM:
2709 case Hexagon::L2_ploadrbt_io:
2710 case Hexagon::L2_ploadrbf_io:
2711 case Hexagon::L2_ploadrubt_io:
2712 case Hexagon::L2_ploadrubf_io:
2713 case Hexagon::S2_pstorerbt_io:
2714 case Hexagon::S2_pstorerbf_io:
2715 return isUInt<6>(Offset);
2717 case Hexagon::L2_ploadrht_io:
2718 case Hexagon::L2_ploadrhf_io:
2719 case Hexagon::L2_ploadruht_io:
2720 case Hexagon::L2_ploadruhf_io:
2721 case Hexagon::S2_pstorerht_io:
2722 case Hexagon::S2_pstorerhf_io:
2723 return isShiftedUInt<6,1>(Offset);
2725 case Hexagon::L2_ploadrit_io:
2726 case Hexagon::L2_ploadrif_io:
2727 case Hexagon::S2_pstorerit_io:
2728 case Hexagon::S2_pstorerif_io:
2729 return isShiftedUInt<6,2>(Offset);
2731 case Hexagon::L2_ploadrdt_io:
2732 case Hexagon::L2_ploadrdf_io:
2733 case Hexagon::S2_pstorerdt_io:
2734 case Hexagon::S2_pstorerdf_io:
2735 return isShiftedUInt<6,3>(Offset);
2738 llvm_unreachable("No offset range is defined for this opcode. "
2739 "Please define it in the above switch statement!");
2742 bool HexagonInstrInfo::isVecAcc(const MachineInstr &MI) const {
2743 return isV60VectorInstruction(MI) && isAccumulator(MI);
2746 bool HexagonInstrInfo::isVecALU(const MachineInstr &MI) const {
2747 const uint64_t F = get(MI.getOpcode()).TSFlags;
2748 const uint64_t V = ((F >> HexagonII::TypePos) & HexagonII::TypeMask);
2750 V == HexagonII::TypeCVI_VA ||
2751 V == HexagonII::TypeCVI_VA_DV;
2754 bool HexagonInstrInfo::isVecUsableNextPacket(const MachineInstr &ProdMI,
2755 const MachineInstr &ConsMI) const {
2756 if (EnableACCForwarding && isVecAcc(ProdMI) && isVecAcc(ConsMI))
2759 if (EnableALUForwarding && (isVecALU(ConsMI) || isLateSourceInstr(ConsMI)))
2762 if (mayBeNewStore(ConsMI))
2768 bool HexagonInstrInfo::isZeroExtendingLoad(const MachineInstr &MI) const {
2769 switch (MI.getOpcode()) {
2771 case Hexagon::L2_loadrub_io:
2772 case Hexagon::L4_loadrub_ur:
2773 case Hexagon::L4_loadrub_ap:
2774 case Hexagon::L2_loadrub_pr:
2775 case Hexagon::L2_loadrub_pbr:
2776 case Hexagon::L2_loadrub_pi:
2777 case Hexagon::L2_loadrub_pci:
2778 case Hexagon::L2_loadrub_pcr:
2779 case Hexagon::L2_loadbzw2_io:
2780 case Hexagon::L4_loadbzw2_ur:
2781 case Hexagon::L4_loadbzw2_ap:
2782 case Hexagon::L2_loadbzw2_pr:
2783 case Hexagon::L2_loadbzw2_pbr:
2784 case Hexagon::L2_loadbzw2_pi:
2785 case Hexagon::L2_loadbzw2_pci:
2786 case Hexagon::L2_loadbzw2_pcr:
2787 case Hexagon::L2_loadbzw4_io:
2788 case Hexagon::L4_loadbzw4_ur:
2789 case Hexagon::L4_loadbzw4_ap:
2790 case Hexagon::L2_loadbzw4_pr:
2791 case Hexagon::L2_loadbzw4_pbr:
2792 case Hexagon::L2_loadbzw4_pi:
2793 case Hexagon::L2_loadbzw4_pci:
2794 case Hexagon::L2_loadbzw4_pcr:
2795 case Hexagon::L4_loadrub_rr:
2796 case Hexagon::L2_ploadrubt_io:
2797 case Hexagon::L2_ploadrubt_pi:
2798 case Hexagon::L2_ploadrubf_io:
2799 case Hexagon::L2_ploadrubf_pi:
2800 case Hexagon::L2_ploadrubtnew_io:
2801 case Hexagon::L2_ploadrubfnew_io:
2802 case Hexagon::L4_ploadrubt_rr:
2803 case Hexagon::L4_ploadrubf_rr:
2804 case Hexagon::L4_ploadrubtnew_rr:
2805 case Hexagon::L4_ploadrubfnew_rr:
2806 case Hexagon::L2_ploadrubtnew_pi:
2807 case Hexagon::L2_ploadrubfnew_pi:
2808 case Hexagon::L4_ploadrubt_abs:
2809 case Hexagon::L4_ploadrubf_abs:
2810 case Hexagon::L4_ploadrubtnew_abs:
2811 case Hexagon::L4_ploadrubfnew_abs:
2812 case Hexagon::L2_loadrubgp:
2814 case Hexagon::L2_loadruh_io:
2815 case Hexagon::L4_loadruh_ur:
2816 case Hexagon::L4_loadruh_ap:
2817 case Hexagon::L2_loadruh_pr:
2818 case Hexagon::L2_loadruh_pbr:
2819 case Hexagon::L2_loadruh_pi:
2820 case Hexagon::L2_loadruh_pci:
2821 case Hexagon::L2_loadruh_pcr:
2822 case Hexagon::L4_loadruh_rr:
2823 case Hexagon::L2_ploadruht_io:
2824 case Hexagon::L2_ploadruht_pi:
2825 case Hexagon::L2_ploadruhf_io:
2826 case Hexagon::L2_ploadruhf_pi:
2827 case Hexagon::L2_ploadruhtnew_io:
2828 case Hexagon::L2_ploadruhfnew_io:
2829 case Hexagon::L4_ploadruht_rr:
2830 case Hexagon::L4_ploadruhf_rr:
2831 case Hexagon::L4_ploadruhtnew_rr:
2832 case Hexagon::L4_ploadruhfnew_rr:
2833 case Hexagon::L2_ploadruhtnew_pi:
2834 case Hexagon::L2_ploadruhfnew_pi:
2835 case Hexagon::L4_ploadruht_abs:
2836 case Hexagon::L4_ploadruhf_abs:
2837 case Hexagon::L4_ploadruhtnew_abs:
2838 case Hexagon::L4_ploadruhfnew_abs:
2839 case Hexagon::L2_loadruhgp:
2846 // Add latency to instruction.
2847 bool HexagonInstrInfo::addLatencyToSchedule(const MachineInstr &MI1,
2848 const MachineInstr &MI2) const {
2849 if (isV60VectorInstruction(MI1) && isV60VectorInstruction(MI2))
2850 if (!isVecUsableNextPacket(MI1, MI2))
2855 /// \brief Get the base register and byte offset of a load/store instr.
2856 bool HexagonInstrInfo::getMemOpBaseRegImmOfs(MachineInstr &LdSt,
2857 unsigned &BaseReg, int64_t &Offset, const TargetRegisterInfo *TRI)
2859 unsigned AccessSize = 0;
2861 BaseReg = getBaseAndOffset(LdSt, OffsetVal, AccessSize);
2863 return BaseReg != 0;
2866 /// \brief Can these instructions execute at the same time in a bundle.
2867 bool HexagonInstrInfo::canExecuteInBundle(const MachineInstr &First,
2868 const MachineInstr &Second) const {
2869 if (DisableNVSchedule)
2871 if (mayBeNewStore(Second)) {
2872 // Make sure the definition of the first instruction is the value being
2874 const MachineOperand &Stored =
2875 Second.getOperand(Second.getNumOperands() - 1);
2876 if (!Stored.isReg())
2878 for (unsigned i = 0, e = First.getNumOperands(); i < e; ++i) {
2879 const MachineOperand &Op = First.getOperand(i);
2880 if (Op.isReg() && Op.isDef() && Op.getReg() == Stored.getReg())
2887 bool HexagonInstrInfo::doesNotReturn(const MachineInstr &CallMI) const {
2888 unsigned Opc = CallMI.getOpcode();
2889 return Opc == Hexagon::PS_call_nr || Opc == Hexagon::PS_callr_nr;
2892 bool HexagonInstrInfo::hasEHLabel(const MachineBasicBlock *B) const {
2899 // Returns true if an instruction can be converted into a non-extended
2900 // equivalent instruction.
2901 bool HexagonInstrInfo::hasNonExtEquivalent(const MachineInstr &MI) const {
2903 // Check if the instruction has a register form that uses register in place
2904 // of the extended operand, if so return that as the non-extended form.
2905 if (Hexagon::getRegForm(MI.getOpcode()) >= 0)
2908 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) {
2909 // Check addressing mode and retrieve non-ext equivalent instruction.
2911 switch (getAddrMode(MI)) {
2912 case HexagonII::Absolute :
2913 // Load/store with absolute addressing mode can be converted into
2914 // base+offset mode.
2915 NonExtOpcode = Hexagon::getBaseWithImmOffset(MI.getOpcode());
2917 case HexagonII::BaseImmOffset :
2918 // Load/store with base+offset addressing mode can be converted into
2919 // base+register offset addressing mode. However left shift operand should
2921 NonExtOpcode = Hexagon::getBaseWithRegOffset(MI.getOpcode());
2923 case HexagonII::BaseLongOffset:
2924 NonExtOpcode = Hexagon::getRegShlForm(MI.getOpcode());
2929 if (NonExtOpcode < 0)
2936 bool HexagonInstrInfo::hasPseudoInstrPair(const MachineInstr &MI) const {
2937 return Hexagon::getRealHWInstr(MI.getOpcode(),
2938 Hexagon::InstrType_Pseudo) >= 0;
2941 bool HexagonInstrInfo::hasUncondBranch(const MachineBasicBlock *B)
2943 MachineBasicBlock::const_iterator I = B->getFirstTerminator(), E = B->end();
2952 // Returns true, if a LD insn can be promoted to a cur load.
2953 bool HexagonInstrInfo::mayBeCurLoad(const MachineInstr &MI) const {
2954 auto &HST = MI.getParent()->getParent()->getSubtarget<HexagonSubtarget>();
2955 const uint64_t F = MI.getDesc().TSFlags;
2956 return ((F >> HexagonII::mayCVLoadPos) & HexagonII::mayCVLoadMask) &&
2960 // Returns true, if a ST insn can be promoted to a new-value store.
2961 bool HexagonInstrInfo::mayBeNewStore(const MachineInstr &MI) const {
2962 const uint64_t F = MI.getDesc().TSFlags;
2963 return (F >> HexagonII::mayNVStorePos) & HexagonII::mayNVStoreMask;
2966 bool HexagonInstrInfo::producesStall(const MachineInstr &ProdMI,
2967 const MachineInstr &ConsMI) const {
2968 // There is no stall when ProdMI is not a V60 vector.
2969 if (!isV60VectorInstruction(ProdMI))
2972 // There is no stall when ProdMI and ConsMI are not dependent.
2973 if (!isDependent(ProdMI, ConsMI))
2976 // When Forward Scheduling is enabled, there is no stall if ProdMI and ConsMI
2977 // are scheduled in consecutive packets.
2978 if (isVecUsableNextPacket(ProdMI, ConsMI))
2984 bool HexagonInstrInfo::producesStall(const MachineInstr &MI,
2985 MachineBasicBlock::const_instr_iterator BII) const {
2986 // There is no stall when I is not a V60 vector.
2987 if (!isV60VectorInstruction(MI))
2990 MachineBasicBlock::const_instr_iterator MII = BII;
2991 MachineBasicBlock::const_instr_iterator MIE = MII->getParent()->instr_end();
2993 if (!MII->isBundle()) {
2994 const MachineInstr &J = *MII;
2995 if (!isV60VectorInstruction(J))
2997 else if (isVecUsableNextPacket(J, MI))
3002 for (++MII; MII != MIE && MII->isInsideBundle(); ++MII) {
3003 const MachineInstr &J = *MII;
3004 if (producesStall(J, MI))
3010 bool HexagonInstrInfo::predCanBeUsedAsDotNew(const MachineInstr &MI,
3011 unsigned PredReg) const {
3012 for (unsigned opNum = 0; opNum < MI.getNumOperands(); opNum++) {
3013 const MachineOperand &MO = MI.getOperand(opNum);
3014 if (MO.isReg() && MO.isDef() && MO.isImplicit() && (MO.getReg() == PredReg))
3015 return false; // Predicate register must be explicitly defined.
3018 // Hexagon Programmer's Reference says that decbin, memw_locked, and
3019 // memd_locked cannot be used as .new as well,
3020 // but we don't seem to have these instructions defined.
3021 return MI.getOpcode() != Hexagon::A4_tlbmatch;
3024 bool HexagonInstrInfo::PredOpcodeHasJMP_c(unsigned Opcode) const {
3025 return (Opcode == Hexagon::J2_jumpt) ||
3026 (Opcode == Hexagon::J2_jumpf) ||
3027 (Opcode == Hexagon::J2_jumptnew) ||
3028 (Opcode == Hexagon::J2_jumpfnew) ||
3029 (Opcode == Hexagon::J2_jumptnewpt) ||
3030 (Opcode == Hexagon::J2_jumpfnewpt);
3033 bool HexagonInstrInfo::predOpcodeHasNot(ArrayRef<MachineOperand> Cond) const {
3034 if (Cond.empty() || !isPredicated(Cond[0].getImm()))
3036 return !isPredicatedTrue(Cond[0].getImm());
3039 short HexagonInstrInfo::getAbsoluteForm(const MachineInstr &MI) const {
3040 return Hexagon::getAbsoluteForm(MI.getOpcode());
3043 unsigned HexagonInstrInfo::getAddrMode(const MachineInstr &MI) const {
3044 const uint64_t F = MI.getDesc().TSFlags;
3045 return (F >> HexagonII::AddrModePos) & HexagonII::AddrModeMask;
3048 // Returns the base register in a memory access (load/store). The offset is
3049 // returned in Offset and the access size is returned in AccessSize.
3050 unsigned HexagonInstrInfo::getBaseAndOffset(const MachineInstr &MI,
3051 int &Offset, unsigned &AccessSize) const {
3052 // Return if it is not a base+offset type instruction or a MemOp.
3053 if (getAddrMode(MI) != HexagonII::BaseImmOffset &&
3054 getAddrMode(MI) != HexagonII::BaseLongOffset &&
3055 !isMemOp(MI) && !isPostIncrement(MI))
3058 // Since it is a memory access instruction, getMemAccessSize() should never
3060 assert (getMemAccessSize(MI) &&
3061 "BaseImmOffset or BaseLongOffset or MemOp without accessSize");
3063 // Return Values of getMemAccessSize() are
3064 // 0 - Checked in the assert above.
3065 // 1, 2, 3, 4 & 7, 8 - The statement below is correct for all these.
3066 // MemAccessSize is represented as 1+log2(N) where N is size in bits.
3067 AccessSize = (1U << (getMemAccessSize(MI) - 1));
3069 unsigned basePos = 0, offsetPos = 0;
3070 if (!getBaseAndOffsetPosition(MI, basePos, offsetPos))
3073 // Post increment updates its EA after the mem access,
3074 // so we need to treat its offset as zero.
3075 if (isPostIncrement(MI))
3078 Offset = MI.getOperand(offsetPos).getImm();
3081 return MI.getOperand(basePos).getReg();
3084 /// Return the position of the base and offset operands for this instruction.
3085 bool HexagonInstrInfo::getBaseAndOffsetPosition(const MachineInstr &MI,
3086 unsigned &BasePos, unsigned &OffsetPos) const {
3087 // Deal with memops first.
3091 } else if (MI.mayStore()) {
3094 } else if (MI.mayLoad()) {
3100 if (isPredicated(MI)) {
3104 if (isPostIncrement(MI)) {
3109 if (!MI.getOperand(BasePos).isReg() || !MI.getOperand(OffsetPos).isImm())
3115 // Inserts branching instructions in reverse order of their occurrence.
3116 // e.g. jump_t t1 (i1)
3118 // Jumpers = {i2, i1}
3119 SmallVector<MachineInstr*, 2> HexagonInstrInfo::getBranchingInstrs(
3120 MachineBasicBlock& MBB) const {
3121 SmallVector<MachineInstr*, 2> Jumpers;
3122 // If the block has no terminators, it just falls into the block after it.
3123 MachineBasicBlock::instr_iterator I = MBB.instr_end();
3124 if (I == MBB.instr_begin())
3127 // A basic block may looks like this:
3137 // It has two succs but does not have a terminator
3138 // Don't know how to handle it.
3143 } while (I != MBB.instr_begin());
3145 I = MBB.instr_end();
3148 while (I->isDebugValue()) {
3149 if (I == MBB.instr_begin())
3153 if (!isUnpredicatedTerminator(*I))
3156 // Get the last instruction in the block.
3157 MachineInstr *LastInst = &*I;
3158 Jumpers.push_back(LastInst);
3159 MachineInstr *SecondLastInst = nullptr;
3160 // Find one more terminator if present.
3162 if (&*I != LastInst && !I->isBundle() && isUnpredicatedTerminator(*I)) {
3163 if (!SecondLastInst) {
3164 SecondLastInst = &*I;
3165 Jumpers.push_back(SecondLastInst);
3166 } else // This is a third branch.
3169 if (I == MBB.instr_begin())
3176 short HexagonInstrInfo::getBaseWithLongOffset(short Opcode) const {
3179 return Hexagon::getBaseWithLongOffset(Opcode);
3182 short HexagonInstrInfo::getBaseWithLongOffset(const MachineInstr &MI) const {
3183 return Hexagon::getBaseWithLongOffset(MI.getOpcode());
3186 short HexagonInstrInfo::getBaseWithRegOffset(const MachineInstr &MI) const {
3187 return Hexagon::getBaseWithRegOffset(MI.getOpcode());
3190 // Returns Operand Index for the constant extended instruction.
3191 unsigned HexagonInstrInfo::getCExtOpNum(const MachineInstr &MI) const {
3192 const uint64_t F = MI.getDesc().TSFlags;
3193 return (F >> HexagonII::ExtendableOpPos) & HexagonII::ExtendableOpMask;
3196 // See if instruction could potentially be a duplex candidate.
3197 // If so, return its group. Zero otherwise.
3198 HexagonII::CompoundGroup HexagonInstrInfo::getCompoundCandidateGroup(
3199 const MachineInstr &MI) const {
3200 unsigned DstReg, SrcReg, Src1Reg, Src2Reg;
3202 switch (MI.getOpcode()) {
3204 return HexagonII::HCG_None;
3207 // "p0=cmp.eq(Rs16,Rt16); if (p0.new) jump:nt #r9:2"
3208 // "Rd16=#U6 ; jump #r9:2"
3209 // "Rd16=Rs16 ; jump #r9:2"
3211 case Hexagon::C2_cmpeq:
3212 case Hexagon::C2_cmpgt:
3213 case Hexagon::C2_cmpgtu:
3214 DstReg = MI.getOperand(0).getReg();
3215 Src1Reg = MI.getOperand(1).getReg();
3216 Src2Reg = MI.getOperand(2).getReg();
3217 if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3218 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3219 isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg))
3220 return HexagonII::HCG_A;
3222 case Hexagon::C2_cmpeqi:
3223 case Hexagon::C2_cmpgti:
3224 case Hexagon::C2_cmpgtui:
3225 // P0 = cmp.eq(Rs,#u2)
3226 DstReg = MI.getOperand(0).getReg();
3227 SrcReg = MI.getOperand(1).getReg();
3228 if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3229 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3230 isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() &&
3231 ((isUInt<5>(MI.getOperand(2).getImm())) ||
3232 (MI.getOperand(2).getImm() == -1)))
3233 return HexagonII::HCG_A;
3235 case Hexagon::A2_tfr:
3237 DstReg = MI.getOperand(0).getReg();
3238 SrcReg = MI.getOperand(1).getReg();
3239 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
3240 return HexagonII::HCG_A;
3242 case Hexagon::A2_tfrsi:
3244 // Do not test for #u6 size since the const is getting extended
3245 // regardless and compound could be formed.
3246 DstReg = MI.getOperand(0).getReg();
3247 if (isIntRegForSubInst(DstReg))
3248 return HexagonII::HCG_A;
3250 case Hexagon::S2_tstbit_i:
3251 DstReg = MI.getOperand(0).getReg();
3252 Src1Reg = MI.getOperand(1).getReg();
3253 if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3254 (Hexagon::P0 == DstReg || Hexagon::P1 == DstReg) &&
3255 MI.getOperand(2).isImm() &&
3256 isIntRegForSubInst(Src1Reg) && (MI.getOperand(2).getImm() == 0))
3257 return HexagonII::HCG_A;
3259 // The fact that .new form is used pretty much guarantees
3260 // that predicate register will match. Nevertheless,
3261 // there could be some false positives without additional
3263 case Hexagon::J2_jumptnew:
3264 case Hexagon::J2_jumpfnew:
3265 case Hexagon::J2_jumptnewpt:
3266 case Hexagon::J2_jumpfnewpt:
3267 Src1Reg = MI.getOperand(0).getReg();
3268 if (Hexagon::PredRegsRegClass.contains(Src1Reg) &&
3269 (Hexagon::P0 == Src1Reg || Hexagon::P1 == Src1Reg))
3270 return HexagonII::HCG_B;
3272 // Transfer and jump:
3273 // Rd=#U6 ; jump #r9:2
3274 // Rd=Rs ; jump #r9:2
3275 // Do not test for jump range here.
3276 case Hexagon::J2_jump:
3277 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4:
3278 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
3279 return HexagonII::HCG_C;
3283 return HexagonII::HCG_None;
3286 // Returns -1 when there is no opcode found.
3287 unsigned HexagonInstrInfo::getCompoundOpcode(const MachineInstr &GA,
3288 const MachineInstr &GB) const {
3289 assert(getCompoundCandidateGroup(GA) == HexagonII::HCG_A);
3290 assert(getCompoundCandidateGroup(GB) == HexagonII::HCG_B);
3291 if ((GA.getOpcode() != Hexagon::C2_cmpeqi) ||
3292 (GB.getOpcode() != Hexagon::J2_jumptnew))
3294 unsigned DestReg = GA.getOperand(0).getReg();
3295 if (!GB.readsRegister(DestReg))
3297 if (DestReg == Hexagon::P0)
3298 return Hexagon::J4_cmpeqi_tp0_jump_nt;
3299 if (DestReg == Hexagon::P1)
3300 return Hexagon::J4_cmpeqi_tp1_jump_nt;
3304 int HexagonInstrInfo::getCondOpcode(int Opc, bool invertPredicate) const {
3305 enum Hexagon::PredSense inPredSense;
3306 inPredSense = invertPredicate ? Hexagon::PredSense_false :
3307 Hexagon::PredSense_true;
3308 int CondOpcode = Hexagon::getPredOpcode(Opc, inPredSense);
3309 if (CondOpcode >= 0) // Valid Conditional opcode/instruction
3312 llvm_unreachable("Unexpected predicable instruction");
3315 // Return the cur value instruction for a given store.
3316 int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const {
3317 switch (MI.getOpcode()) {
3318 default: llvm_unreachable("Unknown .cur type");
3319 case Hexagon::V6_vL32b_pi:
3320 return Hexagon::V6_vL32b_cur_pi;
3321 case Hexagon::V6_vL32b_ai:
3322 return Hexagon::V6_vL32b_cur_ai;
3324 case Hexagon::V6_vL32b_pi_128B:
3325 return Hexagon::V6_vL32b_cur_pi_128B;
3326 case Hexagon::V6_vL32b_ai_128B:
3327 return Hexagon::V6_vL32b_cur_ai_128B;
3332 // The diagram below shows the steps involved in the conversion of a predicated
3333 // store instruction to its .new predicated new-value form.
3335 // p.new NV store [ if(p0.new)memw(R0+#0)=R2.new ]
3337 // / \ (not OK. it will cause new-value store to be
3338 // / X conditional on p0.new while R2 producer is
3341 // p.new store p.old NV store
3342 // [if(p0.new)memw(R0+#0)=R2] [if(p0)memw(R0+#0)=R2.new]
3348 // [if (p0)memw(R0+#0)=R2]
3351 // The following set of instructions further explains the scenario where
3352 // conditional new-value store becomes invalid when promoted to .new predicate
3355 // { 1) if (p0) r0 = add(r1, r2)
3356 // 2) p0 = cmp.eq(r3, #0) }
3358 // 3) if (p0) memb(r1+#0) = r0 --> this instruction can't be grouped with
3359 // the first two instructions because in instr 1, r0 is conditional on old value
3360 // of p0 but its use in instr 3 is conditional on p0 modified by instr 2 which
3361 // is not valid for new-value stores.
3362 // Predicated new value stores (i.e. if (p0) memw(..)=r0.new) are excluded
3363 // from the "Conditional Store" list. Because a predicated new value store
3364 // would NOT be promoted to a double dot new store. See diagram below:
3365 // This function returns yes for those stores that are predicated but not
3366 // yet promoted to predicate dot new instructions.
3368 // +---------------------+
3369 // /-----| if (p0) memw(..)=r0 |---------\~
3370 // || +---------------------+ ||
3371 // promote || /\ /\ || promote
3373 // \||/ demote || \||/
3375 // +-------------------------+ || +-------------------------+
3376 // | if (p0.new) memw(..)=r0 | || | if (p0) memw(..)=r0.new |
3377 // +-------------------------+ || +-------------------------+
3380 // promote || \/ NOT possible
3384 // +-----------------------------+
3385 // | if (p0.new) memw(..)=r0.new |
3386 // +-----------------------------+
3387 // Double Dot New Store
3389 // Returns the most basic instruction for the .new predicated instructions and
3390 // new-value stores.
3391 // For example, all of the following instructions will be converted back to the
3392 // same instruction:
3393 // 1) if (p0.new) memw(R0+#0) = R1.new --->
3394 // 2) if (p0) memw(R0+#0)= R1.new -------> if (p0) memw(R0+#0) = R1
3395 // 3) if (p0.new) memw(R0+#0) = R1 --->
3397 // To understand the translation of instruction 1 to its original form, consider
3398 // a packet with 3 instructions.
3399 // { p0 = cmp.eq(R0,R1)
3400 // if (p0.new) R2 = add(R3, R4)
3401 // R5 = add (R3, R1)
3403 // if (p0) memw(R5+#0) = R2 <--- trying to include it in the previous packet
3405 // This instruction can be part of the previous packet only if both p0 and R2
3406 // are promoted to .new values. This promotion happens in steps, first
3407 // predicate register is promoted to .new and in the next iteration R2 is
3408 // promoted. Therefore, in case of dependence check failure (due to R5) during
3409 // next iteration, it should be converted back to its most basic form.
3411 // Return the new value instruction for a given store.
3412 int HexagonInstrInfo::getDotNewOp(const MachineInstr &MI) const {
3413 int NVOpcode = Hexagon::getNewValueOpcode(MI.getOpcode());
3414 if (NVOpcode >= 0) // Valid new-value store instruction.
3417 switch (MI.getOpcode()) {
3418 default: llvm_unreachable("Unknown .new type");
3419 case Hexagon::S4_storerb_ur:
3420 return Hexagon::S4_storerbnew_ur;
3422 case Hexagon::S2_storerb_pci:
3423 return Hexagon::S2_storerb_pci;
3425 case Hexagon::S2_storeri_pci:
3426 return Hexagon::S2_storeri_pci;
3428 case Hexagon::S2_storerh_pci:
3429 return Hexagon::S2_storerh_pci;
3431 case Hexagon::S2_storerd_pci:
3432 return Hexagon::S2_storerd_pci;
3434 case Hexagon::S2_storerf_pci:
3435 return Hexagon::S2_storerf_pci;
3437 case Hexagon::V6_vS32b_ai:
3438 return Hexagon::V6_vS32b_new_ai;
3440 case Hexagon::V6_vS32b_pi:
3441 return Hexagon::V6_vS32b_new_pi;
3444 case Hexagon::V6_vS32b_ai_128B:
3445 return Hexagon::V6_vS32b_new_ai_128B;
3447 case Hexagon::V6_vS32b_pi_128B:
3448 return Hexagon::V6_vS32b_new_pi_128B;
3453 // Returns the opcode to use when converting MI, which is a conditional jump,
3454 // into a conditional instruction which uses the .new value of the predicate.
3455 // We also use branch probabilities to add a hint to the jump.
3456 int HexagonInstrInfo::getDotNewPredJumpOp(const MachineInstr &MI,
3457 const MachineBranchProbabilityInfo *MBPI) const {
3458 // We assume that block can have at most two successors.
3460 const MachineBasicBlock *Src = MI.getParent();
3461 const MachineOperand &BrTarget = MI.getOperand(1);
3462 const MachineBasicBlock *Dst = BrTarget.getMBB();
3464 const BranchProbability Prediction = MBPI->getEdgeProbability(Src, Dst);
3465 if (Prediction >= BranchProbability(1,2))
3468 switch (MI.getOpcode()) {
3469 case Hexagon::J2_jumpt:
3470 return taken ? Hexagon::J2_jumptnewpt : Hexagon::J2_jumptnew;
3471 case Hexagon::J2_jumpf:
3472 return taken ? Hexagon::J2_jumpfnewpt : Hexagon::J2_jumpfnew;
3475 llvm_unreachable("Unexpected jump instruction.");
3479 // Return .new predicate version for an instruction.
3480 int HexagonInstrInfo::getDotNewPredOp(const MachineInstr &MI,
3481 const MachineBranchProbabilityInfo *MBPI) const {
3482 int NewOpcode = Hexagon::getPredNewOpcode(MI.getOpcode());
3483 if (NewOpcode >= 0) // Valid predicate new instruction
3486 switch (MI.getOpcode()) {
3488 case Hexagon::J2_jumpt:
3489 case Hexagon::J2_jumpf:
3490 return getDotNewPredJumpOp(MI, MBPI);
3493 assert(0 && "Unknown .new type");
3498 int HexagonInstrInfo::getDotOldOp(const int opc) const {
3500 if (isPredicated(NewOp) && isPredicatedNew(NewOp)) { // Get predicate old form
3501 NewOp = Hexagon::getPredOldOpcode(NewOp);
3502 assert(NewOp >= 0 &&
3503 "Couldn't change predicate new instruction to its old form.");
3506 if (isNewValueStore(NewOp)) { // Convert into non-new-value format
3507 NewOp = Hexagon::getNonNVStore(NewOp);
3508 assert(NewOp >= 0 && "Couldn't change new-value store to its old form.");
3513 // See if instruction could potentially be a duplex candidate.
3514 // If so, return its group. Zero otherwise.
3515 HexagonII::SubInstructionGroup HexagonInstrInfo::getDuplexCandidateGroup(
3516 const MachineInstr &MI) const {
3517 unsigned DstReg, SrcReg, Src1Reg, Src2Reg;
3518 auto &HRI = getRegisterInfo();
3520 switch (MI.getOpcode()) {
3522 return HexagonII::HSIG_None;
3526 // Rd = memw(Rs+#u4:2)
3527 // Rd = memub(Rs+#u4:0)
3528 case Hexagon::L2_loadri_io:
3529 DstReg = MI.getOperand(0).getReg();
3530 SrcReg = MI.getOperand(1).getReg();
3531 // Special case this one from Group L2.
3532 // Rd = memw(r29+#u5:2)
3533 if (isIntRegForSubInst(DstReg)) {
3534 if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
3535 HRI.getStackRegister() == SrcReg &&
3536 MI.getOperand(2).isImm() &&
3537 isShiftedUInt<5,2>(MI.getOperand(2).getImm()))
3538 return HexagonII::HSIG_L2;
3539 // Rd = memw(Rs+#u4:2)
3540 if (isIntRegForSubInst(SrcReg) &&
3541 (MI.getOperand(2).isImm() &&
3542 isShiftedUInt<4,2>(MI.getOperand(2).getImm())))
3543 return HexagonII::HSIG_L1;
3546 case Hexagon::L2_loadrub_io:
3547 // Rd = memub(Rs+#u4:0)
3548 DstReg = MI.getOperand(0).getReg();
3549 SrcReg = MI.getOperand(1).getReg();
3550 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3551 MI.getOperand(2).isImm() && isUInt<4>(MI.getOperand(2).getImm()))
3552 return HexagonII::HSIG_L1;
3557 // Rd = memh/memuh(Rs+#u3:1)
3558 // Rd = memb(Rs+#u3:0)
3559 // Rd = memw(r29+#u5:2) - Handled above.
3560 // Rdd = memd(r29+#u5:3)
3562 // [if ([!]p0[.new])] dealloc_return
3563 // [if ([!]p0[.new])] jumpr r31
3564 case Hexagon::L2_loadrh_io:
3565 case Hexagon::L2_loadruh_io:
3566 // Rd = memh/memuh(Rs+#u3:1)
3567 DstReg = MI.getOperand(0).getReg();
3568 SrcReg = MI.getOperand(1).getReg();
3569 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3570 MI.getOperand(2).isImm() &&
3571 isShiftedUInt<3,1>(MI.getOperand(2).getImm()))
3572 return HexagonII::HSIG_L2;
3574 case Hexagon::L2_loadrb_io:
3575 // Rd = memb(Rs+#u3:0)
3576 DstReg = MI.getOperand(0).getReg();
3577 SrcReg = MI.getOperand(1).getReg();
3578 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3579 MI.getOperand(2).isImm() &&
3580 isUInt<3>(MI.getOperand(2).getImm()))
3581 return HexagonII::HSIG_L2;
3583 case Hexagon::L2_loadrd_io:
3584 // Rdd = memd(r29+#u5:3)
3585 DstReg = MI.getOperand(0).getReg();
3586 SrcReg = MI.getOperand(1).getReg();
3587 if (isDblRegForSubInst(DstReg, HRI) &&
3588 Hexagon::IntRegsRegClass.contains(SrcReg) &&
3589 HRI.getStackRegister() == SrcReg &&
3590 MI.getOperand(2).isImm() &&
3591 isShiftedUInt<5,3>(MI.getOperand(2).getImm()))
3592 return HexagonII::HSIG_L2;
3594 // dealloc_return is not documented in Hexagon Manual, but marked
3595 // with A_SUBINSN attribute in iset_v4classic.py.
3596 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4:
3597 case Hexagon::RESTORE_DEALLOC_RET_JMP_V4_PIC:
3598 case Hexagon::L4_return:
3599 case Hexagon::L2_deallocframe:
3600 return HexagonII::HSIG_L2;
3601 case Hexagon::EH_RETURN_JMPR:
3602 case Hexagon::PS_jmpret:
3604 // Actual form JMPR %PC<imp-def>, %R31<imp-use>, %R0<imp-use,internal>.
3605 DstReg = MI.getOperand(0).getReg();
3606 if (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg))
3607 return HexagonII::HSIG_L2;
3609 case Hexagon::PS_jmprett:
3610 case Hexagon::PS_jmpretf:
3611 case Hexagon::PS_jmprettnewpt:
3612 case Hexagon::PS_jmpretfnewpt:
3613 case Hexagon::PS_jmprettnew:
3614 case Hexagon::PS_jmpretfnew:
3615 DstReg = MI.getOperand(1).getReg();
3616 SrcReg = MI.getOperand(0).getReg();
3617 // [if ([!]p0[.new])] jumpr r31
3618 if ((Hexagon::PredRegsRegClass.contains(SrcReg) &&
3619 (Hexagon::P0 == SrcReg)) &&
3620 (Hexagon::IntRegsRegClass.contains(DstReg) && (Hexagon::R31 == DstReg)))
3621 return HexagonII::HSIG_L2;
3623 case Hexagon::L4_return_t :
3624 case Hexagon::L4_return_f :
3625 case Hexagon::L4_return_tnew_pnt :
3626 case Hexagon::L4_return_fnew_pnt :
3627 case Hexagon::L4_return_tnew_pt :
3628 case Hexagon::L4_return_fnew_pt :
3629 // [if ([!]p0[.new])] dealloc_return
3630 SrcReg = MI.getOperand(0).getReg();
3631 if (Hexagon::PredRegsRegClass.contains(SrcReg) && (Hexagon::P0 == SrcReg))
3632 return HexagonII::HSIG_L2;
3637 // memw(Rs+#u4:2) = Rt
3638 // memb(Rs+#u4:0) = Rt
3639 case Hexagon::S2_storeri_io:
3640 // Special case this one from Group S2.
3641 // memw(r29+#u5:2) = Rt
3642 Src1Reg = MI.getOperand(0).getReg();
3643 Src2Reg = MI.getOperand(2).getReg();
3644 if (Hexagon::IntRegsRegClass.contains(Src1Reg) &&
3645 isIntRegForSubInst(Src2Reg) &&
3646 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() &&
3647 isShiftedUInt<5,2>(MI.getOperand(1).getImm()))
3648 return HexagonII::HSIG_S2;
3649 // memw(Rs+#u4:2) = Rt
3650 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3651 MI.getOperand(1).isImm() &&
3652 isShiftedUInt<4,2>(MI.getOperand(1).getImm()))
3653 return HexagonII::HSIG_S1;
3655 case Hexagon::S2_storerb_io:
3656 // memb(Rs+#u4:0) = Rt
3657 Src1Reg = MI.getOperand(0).getReg();
3658 Src2Reg = MI.getOperand(2).getReg();
3659 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3660 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()))
3661 return HexagonII::HSIG_S1;
3666 // memh(Rs+#u3:1) = Rt
3667 // memw(r29+#u5:2) = Rt
3668 // memd(r29+#s6:3) = Rtt
3669 // memw(Rs+#u4:2) = #U1
3670 // memb(Rs+#u4) = #U1
3671 // allocframe(#u5:3)
3672 case Hexagon::S2_storerh_io:
3673 // memh(Rs+#u3:1) = Rt
3674 Src1Reg = MI.getOperand(0).getReg();
3675 Src2Reg = MI.getOperand(2).getReg();
3676 if (isIntRegForSubInst(Src1Reg) && isIntRegForSubInst(Src2Reg) &&
3677 MI.getOperand(1).isImm() &&
3678 isShiftedUInt<3,1>(MI.getOperand(1).getImm()))
3679 return HexagonII::HSIG_S1;
3681 case Hexagon::S2_storerd_io:
3682 // memd(r29+#s6:3) = Rtt
3683 Src1Reg = MI.getOperand(0).getReg();
3684 Src2Reg = MI.getOperand(2).getReg();
3685 if (isDblRegForSubInst(Src2Reg, HRI) &&
3686 Hexagon::IntRegsRegClass.contains(Src1Reg) &&
3687 HRI.getStackRegister() == Src1Reg && MI.getOperand(1).isImm() &&
3688 isShiftedInt<6,3>(MI.getOperand(1).getImm()))
3689 return HexagonII::HSIG_S2;
3691 case Hexagon::S4_storeiri_io:
3692 // memw(Rs+#u4:2) = #U1
3693 Src1Reg = MI.getOperand(0).getReg();
3694 if (isIntRegForSubInst(Src1Reg) && MI.getOperand(1).isImm() &&
3695 isShiftedUInt<4,2>(MI.getOperand(1).getImm()) &&
3696 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm()))
3697 return HexagonII::HSIG_S2;
3699 case Hexagon::S4_storeirb_io:
3700 // memb(Rs+#u4) = #U1
3701 Src1Reg = MI.getOperand(0).getReg();
3702 if (isIntRegForSubInst(Src1Reg) &&
3703 MI.getOperand(1).isImm() && isUInt<4>(MI.getOperand(1).getImm()) &&
3704 MI.getOperand(2).isImm() && isUInt<1>(MI.getOperand(2).getImm()))
3705 return HexagonII::HSIG_S2;
3707 case Hexagon::S2_allocframe:
3708 if (MI.getOperand(0).isImm() &&
3709 isShiftedUInt<5,3>(MI.getOperand(0).getImm()))
3710 return HexagonII::HSIG_S1;
3719 // if ([!]P0[.new]) Rd = #0
3720 // Rd = add(r29,#u6:2)
3722 // P0 = cmp.eq(Rs,#u2)
3723 // Rdd = combine(#0,Rs)
3724 // Rdd = combine(Rs,#0)
3725 // Rdd = combine(#u2,#U2)
3728 // Rd = sxth/sxtb/zxtb/zxth(Rs)
3730 case Hexagon::A2_addi:
3731 DstReg = MI.getOperand(0).getReg();
3732 SrcReg = MI.getOperand(1).getReg();
3733 if (isIntRegForSubInst(DstReg)) {
3734 // Rd = add(r29,#u6:2)
3735 if (Hexagon::IntRegsRegClass.contains(SrcReg) &&
3736 HRI.getStackRegister() == SrcReg && MI.getOperand(2).isImm() &&
3737 isShiftedUInt<6,2>(MI.getOperand(2).getImm()))
3738 return HexagonII::HSIG_A;
3740 if ((DstReg == SrcReg) && MI.getOperand(2).isImm() &&
3741 isInt<7>(MI.getOperand(2).getImm()))
3742 return HexagonII::HSIG_A;
3745 if (isIntRegForSubInst(SrcReg) && MI.getOperand(2).isImm() &&
3746 ((MI.getOperand(2).getImm() == 1) ||
3747 (MI.getOperand(2).getImm() == -1)))
3748 return HexagonII::HSIG_A;
3751 case Hexagon::A2_add:
3753 DstReg = MI.getOperand(0).getReg();
3754 Src1Reg = MI.getOperand(1).getReg();
3755 Src2Reg = MI.getOperand(2).getReg();
3756 if (isIntRegForSubInst(DstReg) && (DstReg == Src1Reg) &&
3757 isIntRegForSubInst(Src2Reg))
3758 return HexagonII::HSIG_A;
3760 case Hexagon::A2_andir:
3762 // Rd16=and(Rs16,#255)
3763 // Rd16=and(Rs16,#1)
3764 DstReg = MI.getOperand(0).getReg();
3765 SrcReg = MI.getOperand(1).getReg();
3766 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg) &&
3767 MI.getOperand(2).isImm() &&
3768 ((MI.getOperand(2).getImm() == 1) ||
3769 (MI.getOperand(2).getImm() == 255)))
3770 return HexagonII::HSIG_A;
3772 case Hexagon::A2_tfr:
3774 DstReg = MI.getOperand(0).getReg();
3775 SrcReg = MI.getOperand(1).getReg();
3776 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
3777 return HexagonII::HSIG_A;
3779 case Hexagon::A2_tfrsi:
3781 // Do not test for #u6 size since the const is getting extended
3782 // regardless and compound could be formed.
3784 DstReg = MI.getOperand(0).getReg();
3785 if (isIntRegForSubInst(DstReg))
3786 return HexagonII::HSIG_A;
3788 case Hexagon::C2_cmoveit:
3789 case Hexagon::C2_cmovenewit:
3790 case Hexagon::C2_cmoveif:
3791 case Hexagon::C2_cmovenewif:
3792 // if ([!]P0[.new]) Rd = #0
3794 // %R16<def> = C2_cmovenewit %P0<internal>, 0, %R16<imp-use,undef>;
3795 DstReg = MI.getOperand(0).getReg();
3796 SrcReg = MI.getOperand(1).getReg();
3797 if (isIntRegForSubInst(DstReg) &&
3798 Hexagon::PredRegsRegClass.contains(SrcReg) && Hexagon::P0 == SrcReg &&
3799 MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0)
3800 return HexagonII::HSIG_A;
3802 case Hexagon::C2_cmpeqi:
3803 // P0 = cmp.eq(Rs,#u2)
3804 DstReg = MI.getOperand(0).getReg();
3805 SrcReg = MI.getOperand(1).getReg();
3806 if (Hexagon::PredRegsRegClass.contains(DstReg) &&
3807 Hexagon::P0 == DstReg && isIntRegForSubInst(SrcReg) &&
3808 MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm()))
3809 return HexagonII::HSIG_A;
3811 case Hexagon::A2_combineii:
3812 case Hexagon::A4_combineii:
3813 // Rdd = combine(#u2,#U2)
3814 DstReg = MI.getOperand(0).getReg();
3815 if (isDblRegForSubInst(DstReg, HRI) &&
3816 ((MI.getOperand(1).isImm() && isUInt<2>(MI.getOperand(1).getImm())) ||
3817 (MI.getOperand(1).isGlobal() &&
3818 isUInt<2>(MI.getOperand(1).getOffset()))) &&
3819 ((MI.getOperand(2).isImm() && isUInt<2>(MI.getOperand(2).getImm())) ||
3820 (MI.getOperand(2).isGlobal() &&
3821 isUInt<2>(MI.getOperand(2).getOffset()))))
3822 return HexagonII::HSIG_A;
3824 case Hexagon::A4_combineri:
3825 // Rdd = combine(Rs,#0)
3826 DstReg = MI.getOperand(0).getReg();
3827 SrcReg = MI.getOperand(1).getReg();
3828 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) &&
3829 ((MI.getOperand(2).isImm() && MI.getOperand(2).getImm() == 0) ||
3830 (MI.getOperand(2).isGlobal() && MI.getOperand(2).getOffset() == 0)))
3831 return HexagonII::HSIG_A;
3833 case Hexagon::A4_combineir:
3834 // Rdd = combine(#0,Rs)
3835 DstReg = MI.getOperand(0).getReg();
3836 SrcReg = MI.getOperand(2).getReg();
3837 if (isDblRegForSubInst(DstReg, HRI) && isIntRegForSubInst(SrcReg) &&
3838 ((MI.getOperand(1).isImm() && MI.getOperand(1).getImm() == 0) ||
3839 (MI.getOperand(1).isGlobal() && MI.getOperand(1).getOffset() == 0)))
3840 return HexagonII::HSIG_A;
3842 case Hexagon::A2_sxtb:
3843 case Hexagon::A2_sxth:
3844 case Hexagon::A2_zxtb:
3845 case Hexagon::A2_zxth:
3846 // Rd = sxth/sxtb/zxtb/zxth(Rs)
3847 DstReg = MI.getOperand(0).getReg();
3848 SrcReg = MI.getOperand(1).getReg();
3849 if (isIntRegForSubInst(DstReg) && isIntRegForSubInst(SrcReg))
3850 return HexagonII::HSIG_A;
3854 return HexagonII::HSIG_None;
3857 short HexagonInstrInfo::getEquivalentHWInstr(const MachineInstr &MI) const {
3858 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Real);
3861 // Return first non-debug instruction in the basic block.
3862 MachineInstr *HexagonInstrInfo::getFirstNonDbgInst(MachineBasicBlock *BB)
3864 for (auto MII = BB->instr_begin(), End = BB->instr_end(); MII != End; MII++) {
3865 MachineInstr &MI = *MII;
3866 if (MI.isDebugValue())
3873 unsigned HexagonInstrInfo::getInstrTimingClassLatency(
3874 const InstrItineraryData *ItinData, const MachineInstr &MI) const {
3875 // Default to one cycle for no itinerary. However, an "empty" itinerary may
3876 // still have a MinLatency property, which getStageLatency checks.
3878 return getInstrLatency(ItinData, MI);
3880 // Get the latency embedded in the itinerary. If we're not using timing class
3881 // latencies or if we using BSB scheduling, then restrict the maximum latency
3882 // to 1 (that is, either 0 or 1).
3883 if (MI.isTransient())
3885 unsigned Latency = ItinData->getStageLatency(MI.getDesc().getSchedClass());
3886 if (!EnableTimingClassLatency ||
3887 MI.getParent()->getParent()->getSubtarget<HexagonSubtarget>().
3894 // inverts the predication logic.
3897 bool HexagonInstrInfo::getInvertedPredSense(
3898 SmallVectorImpl<MachineOperand> &Cond) const {
3901 unsigned Opc = getInvertedPredicatedOpcode(Cond[0].getImm());
3902 Cond[0].setImm(Opc);
3906 unsigned HexagonInstrInfo::getInvertedPredicatedOpcode(const int Opc) const {
3908 InvPredOpcode = isPredicatedTrue(Opc) ? Hexagon::getFalsePredOpcode(Opc)
3909 : Hexagon::getTruePredOpcode(Opc);
3910 if (InvPredOpcode >= 0) // Valid instruction with the inverted predicate.
3911 return InvPredOpcode;
3913 llvm_unreachable("Unexpected predicated instruction");
3916 // Returns the max value that doesn't need to be extended.
3917 int HexagonInstrInfo::getMaxValue(const MachineInstr &MI) const {
3918 const uint64_t F = MI.getDesc().TSFlags;
3919 unsigned isSigned = (F >> HexagonII::ExtentSignedPos)
3920 & HexagonII::ExtentSignedMask;
3921 unsigned bits = (F >> HexagonII::ExtentBitsPos)
3922 & HexagonII::ExtentBitsMask;
3924 if (isSigned) // if value is signed
3925 return ~(-1U << (bits - 1));
3927 return ~(-1U << bits);
3930 unsigned HexagonInstrInfo::getMemAccessSize(const MachineInstr &MI) const {
3931 const uint64_t F = MI.getDesc().TSFlags;
3932 return (F >> HexagonII::MemAccessSizePos) & HexagonII::MemAccesSizeMask;
3935 // Returns the min value that doesn't need to be extended.
3936 int HexagonInstrInfo::getMinValue(const MachineInstr &MI) const {
3937 const uint64_t F = MI.getDesc().TSFlags;
3938 unsigned isSigned = (F >> HexagonII::ExtentSignedPos)
3939 & HexagonII::ExtentSignedMask;
3940 unsigned bits = (F >> HexagonII::ExtentBitsPos)
3941 & HexagonII::ExtentBitsMask;
3943 if (isSigned) // if value is signed
3944 return -1U << (bits - 1);
3949 // Returns opcode of the non-extended equivalent instruction.
3950 short HexagonInstrInfo::getNonExtOpcode(const MachineInstr &MI) const {
3951 // Check if the instruction has a register form that uses register in place
3952 // of the extended operand, if so return that as the non-extended form.
3953 short NonExtOpcode = Hexagon::getRegForm(MI.getOpcode());
3954 if (NonExtOpcode >= 0)
3955 return NonExtOpcode;
3957 if (MI.getDesc().mayLoad() || MI.getDesc().mayStore()) {
3958 // Check addressing mode and retrieve non-ext equivalent instruction.
3959 switch (getAddrMode(MI)) {
3960 case HexagonII::Absolute :
3961 return Hexagon::getBaseWithImmOffset(MI.getOpcode());
3962 case HexagonII::BaseImmOffset :
3963 return Hexagon::getBaseWithRegOffset(MI.getOpcode());
3964 case HexagonII::BaseLongOffset:
3965 return Hexagon::getRegShlForm(MI.getOpcode());
3974 bool HexagonInstrInfo::getPredReg(ArrayRef<MachineOperand> Cond,
3975 unsigned &PredReg, unsigned &PredRegPos, unsigned &PredRegFlags) const {
3978 assert(Cond.size() == 2);
3979 if (isNewValueJump(Cond[0].getImm()) || Cond[1].isMBB()) {
3980 DEBUG(dbgs() << "No predregs for new-value jumps/endloop");
3983 PredReg = Cond[1].getReg();
3985 // See IfConversion.cpp why we add RegState::Implicit | RegState::Undef
3987 if (Cond[1].isImplicit())
3988 PredRegFlags = RegState::Implicit;
3989 if (Cond[1].isUndef())
3990 PredRegFlags |= RegState::Undef;
3994 short HexagonInstrInfo::getPseudoInstrPair(const MachineInstr &MI) const {
3995 return Hexagon::getRealHWInstr(MI.getOpcode(), Hexagon::InstrType_Pseudo);
3998 short HexagonInstrInfo::getRegForm(const MachineInstr &MI) const {
3999 return Hexagon::getRegForm(MI.getOpcode());
4002 // Return the number of bytes required to encode the instruction.
4003 // Hexagon instructions are fixed length, 4 bytes, unless they
4004 // use a constant extender, which requires another 4 bytes.
4005 // For debug instructions and prolog labels, return 0.
4006 unsigned HexagonInstrInfo::getSize(const MachineInstr &MI) const {
4007 if (MI.isDebugValue() || MI.isPosition())
4010 unsigned Size = MI.getDesc().getSize();
4012 // Assume the default insn size in case it cannot be determined
4013 // for whatever reason.
4014 Size = HEXAGON_INSTR_SIZE;
4016 if (isConstExtended(MI) || isExtended(MI))
4017 Size += HEXAGON_INSTR_SIZE;
4019 // Try and compute number of instructions in asm.
4020 if (BranchRelaxAsmLarge && MI.getOpcode() == Hexagon::INLINEASM) {
4021 const MachineBasicBlock &MBB = *MI.getParent();
4022 const MachineFunction *MF = MBB.getParent();
4023 const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
4025 // Count the number of register definitions to find the asm string.
4026 unsigned NumDefs = 0;
4027 for (; MI.getOperand(NumDefs).isReg() && MI.getOperand(NumDefs).isDef();
4029 assert(NumDefs != MI.getNumOperands()-2 && "No asm string?");
4031 assert(MI.getOperand(NumDefs).isSymbol() && "No asm string?");
4032 // Disassemble the AsmStr and approximate number of instructions.
4033 const char *AsmStr = MI.getOperand(NumDefs).getSymbolName();
4034 Size = getInlineAsmLength(AsmStr, *MAI);
4040 uint64_t HexagonInstrInfo::getType(const MachineInstr &MI) const {
4041 const uint64_t F = MI.getDesc().TSFlags;
4042 return (F >> HexagonII::TypePos) & HexagonII::TypeMask;
4045 unsigned HexagonInstrInfo::getUnits(const MachineInstr &MI) const {
4046 const TargetSubtargetInfo &ST = MI.getParent()->getParent()->getSubtarget();
4047 const InstrItineraryData &II = *ST.getInstrItineraryData();
4048 const InstrStage &IS = *II.beginStage(MI.getDesc().getSchedClass());
4050 return IS.getUnits();
4053 unsigned HexagonInstrInfo::getValidSubTargets(const unsigned Opcode) const {
4054 const uint64_t F = get(Opcode).TSFlags;
4055 return (F >> HexagonII::validSubTargetPos) & HexagonII::validSubTargetMask;
4058 // Calculate size of the basic block without debug instructions.
4059 unsigned HexagonInstrInfo::nonDbgBBSize(const MachineBasicBlock *BB) const {
4060 return nonDbgMICount(BB->instr_begin(), BB->instr_end());
4063 unsigned HexagonInstrInfo::nonDbgBundleSize(
4064 MachineBasicBlock::const_iterator BundleHead) const {
4065 assert(BundleHead->isBundle() && "Not a bundle header");
4066 auto MII = BundleHead.getInstrIterator();
4067 // Skip the bundle header.
4068 return nonDbgMICount(++MII, getBundleEnd(BundleHead.getInstrIterator()));
4071 /// immediateExtend - Changes the instruction in place to one using an immediate
4073 void HexagonInstrInfo::immediateExtend(MachineInstr &MI) const {
4074 assert((isExtendable(MI)||isConstExtended(MI)) &&
4075 "Instruction must be extendable");
4076 // Find which operand is extendable.
4077 short ExtOpNum = getCExtOpNum(MI);
4078 MachineOperand &MO = MI.getOperand(ExtOpNum);
4079 // This needs to be something we understand.
4080 assert((MO.isMBB() || MO.isImm()) &&
4081 "Branch with unknown extendable field type");
4082 // Mark given operand as extended.
4083 MO.addTargetFlag(HexagonII::HMOTF_ConstExtended);
4086 bool HexagonInstrInfo::invertAndChangeJumpTarget(
4087 MachineInstr &MI, MachineBasicBlock *NewTarget) const {
4088 DEBUG(dbgs() << "\n[invertAndChangeJumpTarget] to BB#"
4089 << NewTarget->getNumber(); MI.dump(););
4090 assert(MI.isBranch());
4091 unsigned NewOpcode = getInvertedPredicatedOpcode(MI.getOpcode());
4092 int TargetPos = MI.getNumOperands() - 1;
4093 // In general branch target is the last operand,
4094 // but some implicit defs added at the end might change it.
4095 while ((TargetPos > -1) && !MI.getOperand(TargetPos).isMBB())
4097 assert((TargetPos >= 0) && MI.getOperand(TargetPos).isMBB());
4098 MI.getOperand(TargetPos).setMBB(NewTarget);
4099 if (EnableBranchPrediction && isPredicatedNew(MI)) {
4100 NewOpcode = reversePrediction(NewOpcode);
4102 MI.setDesc(get(NewOpcode));
4106 void HexagonInstrInfo::genAllInsnTimingClasses(MachineFunction &MF) const {
4107 /* +++ The code below is used to generate complete set of Hexagon Insn +++ */
4108 MachineFunction::iterator A = MF.begin();
4109 MachineBasicBlock &B = *A;
4110 MachineBasicBlock::iterator I = B.begin();
4111 DebugLoc DL = I->getDebugLoc();
4112 MachineInstr *NewMI;
4114 for (unsigned insn = TargetOpcode::GENERIC_OP_END+1;
4115 insn < Hexagon::INSTRUCTION_LIST_END; ++insn) {
4116 NewMI = BuildMI(B, I, DL, get(insn));
4117 DEBUG(dbgs() << "\n" << getName(NewMI->getOpcode()) <<
4118 " Class: " << NewMI->getDesc().getSchedClass());
4119 NewMI->eraseFromParent();
4121 /* --- The code above is used to generate complete set of Hexagon Insn --- */
4124 // inverts the predication logic.
4127 bool HexagonInstrInfo::reversePredSense(MachineInstr &MI) const {
4128 DEBUG(dbgs() << "\nTrying to reverse pred. sense of:"; MI.dump());
4129 MI.setDesc(get(getInvertedPredicatedOpcode(MI.getOpcode())));
4133 // Reverse the branch prediction.
4134 unsigned HexagonInstrInfo::reversePrediction(unsigned Opcode) const {
4135 int PredRevOpcode = -1;
4136 if (isPredictedTaken(Opcode))
4137 PredRevOpcode = Hexagon::notTakenBranchPrediction(Opcode);
4139 PredRevOpcode = Hexagon::takenBranchPrediction(Opcode);
4140 assert(PredRevOpcode > 0);
4141 return PredRevOpcode;
4144 // TODO: Add more rigorous validation.
4145 bool HexagonInstrInfo::validateBranchCond(const ArrayRef<MachineOperand> &Cond)
4147 return Cond.empty() || (Cond[0].isImm() && (Cond.size() != 1));
4150 short HexagonInstrInfo::xformRegToImmOffset(const MachineInstr &MI) const {
4151 return Hexagon::xformRegToImmOffset(MI.getOpcode());