1 //===----- HexagonPacketizer.cpp - vliw packetizer ---------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This implements a simple VLIW packetizer using DFA. The packetizer works on
11 // machine basic blocks. For each instruction I in BB, the packetizer consults
12 // the DFA to see if machine resources are available to execute I. If so, the
13 // packetizer checks if I depends on any instruction J in the current packet.
14 // If no dependency is found, I is added to current packet and machine resource
15 // is marked as taken. If any dependency is found, a target API call is made to
16 // prune the dependence.
17 //
18 //===----------------------------------------------------------------------===//
19 #include "HexagonRegisterInfo.h"
20 #include "HexagonSubtarget.h"
21 #include "HexagonTargetMachine.h"
22 #include "HexagonVLIWPacketizer.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/CodeGen/MachineDominators.h"
25 #include "llvm/CodeGen/MachineFunctionAnalysis.h"
26 #include "llvm/CodeGen/MachineFunctionPass.h"
27 #include "llvm/CodeGen/MachineLoopInfo.h"
28 #include "llvm/CodeGen/MachineRegisterInfo.h"
29 #include "llvm/CodeGen/Passes.h"
30 #include "llvm/Support/CommandLine.h"
31 #include "llvm/Support/Debug.h"
32 #include <map>
33 #include <vector>
34
35 using namespace llvm;
36
37 #define DEBUG_TYPE "packets"
38
39 static cl::opt<bool> DisablePacketizer("disable-packetizer", cl::Hidden,
40 cl::ZeroOrMore, cl::init(false),
41 cl::desc("Disable Hexagon packetizer pass"));
42
43 static cl::opt<bool> PacketizeVolatiles("hexagon-packetize-volatiles",
44 cl::ZeroOrMore, cl::Hidden, cl::init(true),
45 cl::desc("Allow non-solo packetization of volatile memory references"));
46
47 static cl::opt<bool> EnableGenAllInsnClass("enable-gen-insn", cl::init(false),
48 cl::Hidden, cl::ZeroOrMore, cl::desc("Generate all instruction with TC"));
49
50 static cl::opt<bool> DisableVecDblNVStores("disable-vecdbl-nv-stores",
51 cl::init(false), cl::Hidden, cl::ZeroOrMore,
52 cl::desc("Disable vector double new-value-stores"));
53
54 extern cl::opt<bool> ScheduleInlineAsm;
55
56 namespace llvm {
57 FunctionPass *createHexagonPacketizer();
58 void initializeHexagonPacketizerPass(PassRegistry&);
59 }
60
61
62 namespace {
63 class HexagonPacketizer : public MachineFunctionPass {
64 public:
65 static char ID;
HexagonPacketizer()66 HexagonPacketizer() : MachineFunctionPass(ID) {
67 initializeHexagonPacketizerPass(*PassRegistry::getPassRegistry());
68 }
69
getAnalysisUsage(AnalysisUsage & AU) const70 void getAnalysisUsage(AnalysisUsage &AU) const override {
71 AU.setPreservesCFG();
72 AU.addRequired<AAResultsWrapperPass>();
73 AU.addRequired<MachineBranchProbabilityInfo>();
74 AU.addRequired<MachineDominatorTree>();
75 AU.addRequired<MachineLoopInfo>();
76 AU.addPreserved<MachineDominatorTree>();
77 AU.addPreserved<MachineLoopInfo>();
78 MachineFunctionPass::getAnalysisUsage(AU);
79 }
getPassName() const80 const char *getPassName() const override {
81 return "Hexagon Packetizer";
82 }
83 bool runOnMachineFunction(MachineFunction &Fn) override;
84
85 private:
86 const HexagonInstrInfo *HII;
87 const HexagonRegisterInfo *HRI;
88 };
89
90 char HexagonPacketizer::ID = 0;
91 }
92
93 INITIALIZE_PASS_BEGIN(HexagonPacketizer, "packets", "Hexagon Packetizer",
94 false, false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)95 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
96 INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfo)
97 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
98 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
99 INITIALIZE_PASS_END(HexagonPacketizer, "packets", "Hexagon Packetizer",
100 false, false)
101
102
103 HexagonPacketizerList::HexagonPacketizerList(MachineFunction &MF,
104 MachineLoopInfo &MLI, AliasAnalysis *AA,
105 const MachineBranchProbabilityInfo *MBPI)
106 : VLIWPacketizerList(MF, MLI, AA), MBPI(MBPI), MLI(&MLI) {
107 HII = MF.getSubtarget<HexagonSubtarget>().getInstrInfo();
108 HRI = MF.getSubtarget<HexagonSubtarget>().getRegisterInfo();
109 }
110
111 // Check if FirstI modifies a register that SecondI reads.
hasWriteToReadDep(const MachineInstr * FirstI,const MachineInstr * SecondI,const TargetRegisterInfo * TRI)112 static bool hasWriteToReadDep(const MachineInstr *FirstI,
113 const MachineInstr *SecondI, const TargetRegisterInfo *TRI) {
114 for (auto &MO : FirstI->operands()) {
115 if (!MO.isReg() || !MO.isDef())
116 continue;
117 unsigned R = MO.getReg();
118 if (SecondI->readsRegister(R, TRI))
119 return true;
120 }
121 return false;
122 }
123
124
moveInstrOut(MachineInstr * MI,MachineBasicBlock::iterator BundleIt,bool Before)125 static MachineBasicBlock::iterator moveInstrOut(MachineInstr *MI,
126 MachineBasicBlock::iterator BundleIt, bool Before) {
127 MachineBasicBlock::instr_iterator InsertPt;
128 if (Before)
129 InsertPt = BundleIt.getInstrIterator();
130 else
131 InsertPt = std::next(BundleIt).getInstrIterator();
132
133 MachineBasicBlock &B = *MI->getParent();
134 // The instruction should at least be bundled with the preceding instruction
135 // (there will always be one, i.e. BUNDLE, if nothing else).
136 assert(MI->isBundledWithPred());
137 if (MI->isBundledWithSucc()) {
138 MI->clearFlag(MachineInstr::BundledSucc);
139 MI->clearFlag(MachineInstr::BundledPred);
140 } else {
141 // If it's not bundled with the successor (i.e. it is the last one
142 // in the bundle), then we can simply unbundle it from the predecessor,
143 // which will take care of updating the predecessor's flag.
144 MI->unbundleFromPred();
145 }
146 B.splice(InsertPt, &B, MI);
147
148 // Get the size of the bundle without asserting.
149 MachineBasicBlock::const_instr_iterator I(BundleIt);
150 MachineBasicBlock::const_instr_iterator E = B.instr_end();
151 unsigned Size = 0;
152 for (++I; I != E && I->isBundledWithPred(); ++I)
153 ++Size;
154
155 // If there are still two or more instructions, then there is nothing
156 // else to be done.
157 if (Size > 1)
158 return BundleIt;
159
160 // Otherwise, extract the single instruction out and delete the bundle.
161 MachineBasicBlock::iterator NextIt = std::next(BundleIt);
162 MachineInstr *SingleI = BundleIt->getNextNode();
163 SingleI->unbundleFromPred();
164 assert(!SingleI->isBundledWithSucc());
165 BundleIt->eraseFromParent();
166 return NextIt;
167 }
168
169
runOnMachineFunction(MachineFunction & MF)170 bool HexagonPacketizer::runOnMachineFunction(MachineFunction &MF) {
171 if (DisablePacketizer)
172 return false;
173
174 HII = MF.getSubtarget<HexagonSubtarget>().getInstrInfo();
175 HRI = MF.getSubtarget<HexagonSubtarget>().getRegisterInfo();
176 auto &MLI = getAnalysis<MachineLoopInfo>();
177 auto *AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
178 auto *MBPI = &getAnalysis<MachineBranchProbabilityInfo>();
179
180 if (EnableGenAllInsnClass)
181 HII->genAllInsnTimingClasses(MF);
182
183 // Instantiate the packetizer.
184 HexagonPacketizerList Packetizer(MF, MLI, AA, MBPI);
185
186 // DFA state table should not be empty.
187 assert(Packetizer.getResourceTracker() && "Empty DFA table!");
188
189 //
190 // Loop over all basic blocks and remove KILL pseudo-instructions
191 // These instructions confuse the dependence analysis. Consider:
192 // D0 = ... (Insn 0)
193 // R0 = KILL R0, D0 (Insn 1)
194 // R0 = ... (Insn 2)
195 // Here, Insn 1 will result in the dependence graph not emitting an output
196 // dependence between Insn 0 and Insn 2. This can lead to incorrect
197 // packetization
198 //
199 for (auto &MB : MF) {
200 auto End = MB.end();
201 auto MI = MB.begin();
202 while (MI != End) {
203 auto NextI = std::next(MI);
204 if (MI->isKill()) {
205 MB.erase(MI);
206 End = MB.end();
207 }
208 MI = NextI;
209 }
210 }
211
212 // Loop over all of the basic blocks.
213 for (auto &MB : MF) {
214 auto Begin = MB.begin(), End = MB.end();
215 while (Begin != End) {
216 // First the first non-boundary starting from the end of the last
217 // scheduling region.
218 MachineBasicBlock::iterator RB = Begin;
219 while (RB != End && HII->isSchedulingBoundary(RB, &MB, MF))
220 ++RB;
221 // First the first boundary starting from the beginning of the new
222 // region.
223 MachineBasicBlock::iterator RE = RB;
224 while (RE != End && !HII->isSchedulingBoundary(RE, &MB, MF))
225 ++RE;
226 // Add the scheduling boundary if it's not block end.
227 if (RE != End)
228 ++RE;
229 // If RB == End, then RE == End.
230 if (RB != End)
231 Packetizer.PacketizeMIs(&MB, RB, RE);
232
233 Begin = RE;
234 }
235 }
236
237 Packetizer.unpacketizeSoloInstrs(MF);
238 return true;
239 }
240
241
242 // Reserve resources for a constant extender. Trigger an assertion if the
243 // reservation fails.
reserveResourcesForConstExt()244 void HexagonPacketizerList::reserveResourcesForConstExt() {
245 if (!tryAllocateResourcesForConstExt(true))
246 llvm_unreachable("Resources not available");
247 }
248
canReserveResourcesForConstExt()249 bool HexagonPacketizerList::canReserveResourcesForConstExt() {
250 return tryAllocateResourcesForConstExt(false);
251 }
252
253 // Allocate resources (i.e. 4 bytes) for constant extender. If succeeded,
254 // return true, otherwise, return false.
tryAllocateResourcesForConstExt(bool Reserve)255 bool HexagonPacketizerList::tryAllocateResourcesForConstExt(bool Reserve) {
256 auto *ExtMI = MF.CreateMachineInstr(HII->get(Hexagon::A4_ext), DebugLoc());
257 bool Avail = ResourceTracker->canReserveResources(ExtMI);
258 if (Reserve && Avail)
259 ResourceTracker->reserveResources(ExtMI);
260 MF.DeleteMachineInstr(ExtMI);
261 return Avail;
262 }
263
264
isCallDependent(const MachineInstr * MI,SDep::Kind DepType,unsigned DepReg)265 bool HexagonPacketizerList::isCallDependent(const MachineInstr* MI,
266 SDep::Kind DepType, unsigned DepReg) {
267 // Check for LR dependence.
268 if (DepReg == HRI->getRARegister())
269 return true;
270
271 if (HII->isDeallocRet(MI))
272 if (DepReg == HRI->getFrameRegister() || DepReg == HRI->getStackRegister())
273 return true;
274
275 // Check if this is a predicate dependence.
276 const TargetRegisterClass* RC = HRI->getMinimalPhysRegClass(DepReg);
277 if (RC == &Hexagon::PredRegsRegClass)
278 return true;
279
280 // Assumes that the first operand of the CALLr is the function address.
281 if (HII->isIndirectCall(MI) && (DepType == SDep::Data)) {
282 MachineOperand MO = MI->getOperand(0);
283 if (MO.isReg() && MO.isUse() && (MO.getReg() == DepReg))
284 return true;
285 }
286
287 return false;
288 }
289
isRegDependence(const SDep::Kind DepType)290 static bool isRegDependence(const SDep::Kind DepType) {
291 return DepType == SDep::Data || DepType == SDep::Anti ||
292 DepType == SDep::Output;
293 }
294
isDirectJump(const MachineInstr * MI)295 static bool isDirectJump(const MachineInstr* MI) {
296 return MI->getOpcode() == Hexagon::J2_jump;
297 }
298
isSchedBarrier(const MachineInstr * MI)299 static bool isSchedBarrier(const MachineInstr* MI) {
300 switch (MI->getOpcode()) {
301 case Hexagon::Y2_barrier:
302 return true;
303 }
304 return false;
305 }
306
isControlFlow(const MachineInstr * MI)307 static bool isControlFlow(const MachineInstr* MI) {
308 return (MI->getDesc().isTerminator() || MI->getDesc().isCall());
309 }
310
311
312 /// Returns true if the instruction modifies a callee-saved register.
doesModifyCalleeSavedReg(const MachineInstr * MI,const TargetRegisterInfo * TRI)313 static bool doesModifyCalleeSavedReg(const MachineInstr *MI,
314 const TargetRegisterInfo *TRI) {
315 const MachineFunction &MF = *MI->getParent()->getParent();
316 for (auto *CSR = TRI->getCalleeSavedRegs(&MF); CSR && *CSR; ++CSR)
317 if (MI->modifiesRegister(*CSR, TRI))
318 return true;
319 return false;
320 }
321
322 // TODO: MI->isIndirectBranch() and IsRegisterJump(MI)
323 // Returns true if an instruction can be promoted to .new predicate or
324 // new-value store.
isNewifiable(const MachineInstr * MI)325 bool HexagonPacketizerList::isNewifiable(const MachineInstr* MI) {
326 return HII->isCondInst(MI) || MI->isReturn() || HII->mayBeNewStore(MI);
327 }
328
329 // Promote an instructiont to its .cur form.
330 // At this time, we have already made a call to canPromoteToDotCur and made
331 // sure that it can *indeed* be promoted.
promoteToDotCur(MachineInstr * MI,SDep::Kind DepType,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)332 bool HexagonPacketizerList::promoteToDotCur(MachineInstr* MI,
333 SDep::Kind DepType, MachineBasicBlock::iterator &MII,
334 const TargetRegisterClass* RC) {
335 assert(DepType == SDep::Data);
336 int CurOpcode = HII->getDotCurOp(MI);
337 MI->setDesc(HII->get(CurOpcode));
338 return true;
339 }
340
cleanUpDotCur()341 void HexagonPacketizerList::cleanUpDotCur() {
342 MachineInstr *MI = NULL;
343 for (auto BI : CurrentPacketMIs) {
344 DEBUG(dbgs() << "Cleanup packet has "; BI->dump(););
345 if (BI->getOpcode() == Hexagon::V6_vL32b_cur_ai) {
346 MI = BI;
347 continue;
348 }
349 if (MI) {
350 for (auto &MO : BI->operands())
351 if (MO.isReg() && MO.getReg() == MI->getOperand(0).getReg())
352 return;
353 }
354 }
355 if (!MI)
356 return;
357 // We did not find a use of the CUR, so de-cur it.
358 MI->setDesc(HII->get(Hexagon::V6_vL32b_ai));
359 DEBUG(dbgs() << "Demoted CUR "; MI->dump(););
360 }
361
362 // Check to see if an instruction can be dot cur.
canPromoteToDotCur(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)363 bool HexagonPacketizerList::canPromoteToDotCur(const MachineInstr *MI,
364 const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
365 const TargetRegisterClass *RC) {
366 if (!HII->isV60VectorInstruction(MI))
367 return false;
368 if (!HII->isV60VectorInstruction(MII))
369 return false;
370
371 // Already a dot new instruction.
372 if (HII->isDotCurInst(MI) && !HII->mayBeCurLoad(MI))
373 return false;
374
375 if (!HII->mayBeCurLoad(MI))
376 return false;
377
378 // The "cur value" cannot come from inline asm.
379 if (PacketSU->getInstr()->isInlineAsm())
380 return false;
381
382 // Make sure candidate instruction uses cur.
383 DEBUG(dbgs() << "Can we DOT Cur Vector MI\n";
384 MI->dump();
385 dbgs() << "in packet\n";);
386 MachineInstr *MJ = MII;
387 DEBUG(dbgs() << "Checking CUR against "; MJ->dump(););
388 unsigned DestReg = MI->getOperand(0).getReg();
389 bool FoundMatch = false;
390 for (auto &MO : MJ->operands())
391 if (MO.isReg() && MO.getReg() == DestReg)
392 FoundMatch = true;
393 if (!FoundMatch)
394 return false;
395
396 // Check for existing uses of a vector register within the packet which
397 // would be affected by converting a vector load into .cur formt.
398 for (auto BI : CurrentPacketMIs) {
399 DEBUG(dbgs() << "packet has "; BI->dump(););
400 if (BI->readsRegister(DepReg, MF.getSubtarget().getRegisterInfo()))
401 return false;
402 }
403
404 DEBUG(dbgs() << "Can Dot CUR MI\n"; MI->dump(););
405 // We can convert the opcode into a .cur.
406 return true;
407 }
408
409 // Promote an instruction to its .new form. At this time, we have already
410 // made a call to canPromoteToDotNew and made sure that it can *indeed* be
411 // promoted.
promoteToDotNew(MachineInstr * MI,SDep::Kind DepType,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)412 bool HexagonPacketizerList::promoteToDotNew(MachineInstr* MI,
413 SDep::Kind DepType, MachineBasicBlock::iterator &MII,
414 const TargetRegisterClass* RC) {
415 assert (DepType == SDep::Data);
416 int NewOpcode;
417 if (RC == &Hexagon::PredRegsRegClass)
418 NewOpcode = HII->getDotNewPredOp(MI, MBPI);
419 else
420 NewOpcode = HII->getDotNewOp(MI);
421 MI->setDesc(HII->get(NewOpcode));
422 return true;
423 }
424
demoteToDotOld(MachineInstr * MI)425 bool HexagonPacketizerList::demoteToDotOld(MachineInstr* MI) {
426 int NewOpcode = HII->getDotOldOp(MI->getOpcode());
427 MI->setDesc(HII->get(NewOpcode));
428 return true;
429 }
430
431 enum PredicateKind {
432 PK_False,
433 PK_True,
434 PK_Unknown
435 };
436
437 /// Returns true if an instruction is predicated on p0 and false if it's
438 /// predicated on !p0.
getPredicateSense(const MachineInstr * MI,const HexagonInstrInfo * HII)439 static PredicateKind getPredicateSense(const MachineInstr *MI,
440 const HexagonInstrInfo *HII) {
441 if (!HII->isPredicated(MI))
442 return PK_Unknown;
443 if (HII->isPredicatedTrue(MI))
444 return PK_True;
445 return PK_False;
446 }
447
getPostIncrementOperand(const MachineInstr * MI,const HexagonInstrInfo * HII)448 static const MachineOperand &getPostIncrementOperand(const MachineInstr *MI,
449 const HexagonInstrInfo *HII) {
450 assert(HII->isPostIncrement(MI) && "Not a post increment operation.");
451 #ifndef NDEBUG
452 // Post Increment means duplicates. Use dense map to find duplicates in the
453 // list. Caution: Densemap initializes with the minimum of 64 buckets,
454 // whereas there are at most 5 operands in the post increment.
455 DenseSet<unsigned> DefRegsSet;
456 for (auto &MO : MI->operands())
457 if (MO.isReg() && MO.isDef())
458 DefRegsSet.insert(MO.getReg());
459
460 for (auto &MO : MI->operands())
461 if (MO.isReg() && MO.isUse() && DefRegsSet.count(MO.getReg()))
462 return MO;
463 #else
464 if (MI->mayLoad()) {
465 const MachineOperand &Op1 = MI->getOperand(1);
466 // The 2nd operand is always the post increment operand in load.
467 assert(Op1.isReg() && "Post increment operand has be to a register.");
468 return Op1;
469 }
470 if (MI->getDesc().mayStore()) {
471 const MachineOperand &Op0 = MI->getOperand(0);
472 // The 1st operand is always the post increment operand in store.
473 assert(Op0.isReg() && "Post increment operand has be to a register.");
474 return Op0;
475 }
476 #endif
477 // we should never come here.
478 llvm_unreachable("mayLoad or mayStore not set for Post Increment operation");
479 }
480
481 // Get the value being stored.
getStoreValueOperand(const MachineInstr * MI)482 static const MachineOperand& getStoreValueOperand(const MachineInstr *MI) {
483 // value being stored is always the last operand.
484 return MI->getOperand(MI->getNumOperands()-1);
485 }
486
isLoadAbsSet(const MachineInstr * MI)487 static bool isLoadAbsSet(const MachineInstr *MI) {
488 unsigned Opc = MI->getOpcode();
489 switch (Opc) {
490 case Hexagon::L4_loadrd_ap:
491 case Hexagon::L4_loadrb_ap:
492 case Hexagon::L4_loadrh_ap:
493 case Hexagon::L4_loadrub_ap:
494 case Hexagon::L4_loadruh_ap:
495 case Hexagon::L4_loadri_ap:
496 return true;
497 }
498 return false;
499 }
500
getAbsSetOperand(const MachineInstr * MI)501 static const MachineOperand &getAbsSetOperand(const MachineInstr *MI) {
502 assert(isLoadAbsSet(MI));
503 return MI->getOperand(1);
504 }
505
506
507 // Can be new value store?
508 // Following restrictions are to be respected in convert a store into
509 // a new value store.
510 // 1. If an instruction uses auto-increment, its address register cannot
511 // be a new-value register. Arch Spec 5.4.2.1
512 // 2. If an instruction uses absolute-set addressing mode, its address
513 // register cannot be a new-value register. Arch Spec 5.4.2.1.
514 // 3. If an instruction produces a 64-bit result, its registers cannot be used
515 // as new-value registers. Arch Spec 5.4.2.2.
516 // 4. If the instruction that sets the new-value register is conditional, then
517 // the instruction that uses the new-value register must also be conditional,
518 // and both must always have their predicates evaluate identically.
519 // Arch Spec 5.4.2.3.
520 // 5. There is an implied restriction that a packet cannot have another store,
521 // if there is a new value store in the packet. Corollary: if there is
522 // already a store in a packet, there can not be a new value store.
523 // Arch Spec: 3.4.4.2
canPromoteToNewValueStore(const MachineInstr * MI,const MachineInstr * PacketMI,unsigned DepReg)524 bool HexagonPacketizerList::canPromoteToNewValueStore(const MachineInstr *MI,
525 const MachineInstr *PacketMI, unsigned DepReg) {
526 // Make sure we are looking at the store, that can be promoted.
527 if (!HII->mayBeNewStore(MI))
528 return false;
529
530 // Make sure there is dependency and can be new value'd.
531 const MachineOperand &Val = getStoreValueOperand(MI);
532 if (Val.isReg() && Val.getReg() != DepReg)
533 return false;
534
535 const MCInstrDesc& MCID = PacketMI->getDesc();
536
537 // First operand is always the result.
538 const TargetRegisterClass *PacketRC = HII->getRegClass(MCID, 0, HRI, MF);
539 // Double regs can not feed into new value store: PRM section: 5.4.2.2.
540 if (PacketRC == &Hexagon::DoubleRegsRegClass)
541 return false;
542
543 // New-value stores are of class NV (slot 0), dual stores require class ST
544 // in slot 0 (PRM 5.5).
545 for (auto I : CurrentPacketMIs) {
546 SUnit *PacketSU = MIToSUnit.find(I)->second;
547 if (PacketSU->getInstr()->mayStore())
548 return false;
549 }
550
551 // Make sure it's NOT the post increment register that we are going to
552 // new value.
553 if (HII->isPostIncrement(MI) &&
554 getPostIncrementOperand(MI, HII).getReg() == DepReg) {
555 return false;
556 }
557
558 if (HII->isPostIncrement(PacketMI) && PacketMI->mayLoad() &&
559 getPostIncrementOperand(PacketMI, HII).getReg() == DepReg) {
560 // If source is post_inc, or absolute-set addressing, it can not feed
561 // into new value store
562 // r3 = memw(r2++#4)
563 // memw(r30 + #-1404) = r2.new -> can not be new value store
564 // arch spec section: 5.4.2.1.
565 return false;
566 }
567
568 if (isLoadAbsSet(PacketMI) && getAbsSetOperand(PacketMI).getReg() == DepReg)
569 return false;
570
571 // If the source that feeds the store is predicated, new value store must
572 // also be predicated.
573 if (HII->isPredicated(PacketMI)) {
574 if (!HII->isPredicated(MI))
575 return false;
576
577 // Check to make sure that they both will have their predicates
578 // evaluate identically.
579 unsigned predRegNumSrc = 0;
580 unsigned predRegNumDst = 0;
581 const TargetRegisterClass* predRegClass = nullptr;
582
583 // Get predicate register used in the source instruction.
584 for (auto &MO : PacketMI->operands()) {
585 if (!MO.isReg())
586 continue;
587 predRegNumSrc = MO.getReg();
588 predRegClass = HRI->getMinimalPhysRegClass(predRegNumSrc);
589 if (predRegClass == &Hexagon::PredRegsRegClass)
590 break;
591 }
592 assert((predRegClass == &Hexagon::PredRegsRegClass) &&
593 "predicate register not found in a predicated PacketMI instruction");
594
595 // Get predicate register used in new-value store instruction.
596 for (auto &MO : MI->operands()) {
597 if (!MO.isReg())
598 continue;
599 predRegNumDst = MO.getReg();
600 predRegClass = HRI->getMinimalPhysRegClass(predRegNumDst);
601 if (predRegClass == &Hexagon::PredRegsRegClass)
602 break;
603 }
604 assert((predRegClass == &Hexagon::PredRegsRegClass) &&
605 "predicate register not found in a predicated MI instruction");
606
607 // New-value register producer and user (store) need to satisfy these
608 // constraints:
609 // 1) Both instructions should be predicated on the same register.
610 // 2) If producer of the new-value register is .new predicated then store
611 // should also be .new predicated and if producer is not .new predicated
612 // then store should not be .new predicated.
613 // 3) Both new-value register producer and user should have same predicate
614 // sense, i.e, either both should be negated or both should be non-negated.
615 if (predRegNumDst != predRegNumSrc ||
616 HII->isDotNewInst(PacketMI) != HII->isDotNewInst(MI) ||
617 getPredicateSense(MI, HII) != getPredicateSense(PacketMI, HII))
618 return false;
619 }
620
621 // Make sure that other than the new-value register no other store instruction
622 // register has been modified in the same packet. Predicate registers can be
623 // modified by they should not be modified between the producer and the store
624 // instruction as it will make them both conditional on different values.
625 // We already know this to be true for all the instructions before and
626 // including PacketMI. Howerver, we need to perform the check for the
627 // remaining instructions in the packet.
628
629 unsigned StartCheck = 0;
630
631 for (auto I : CurrentPacketMIs) {
632 SUnit *TempSU = MIToSUnit.find(I)->second;
633 MachineInstr* TempMI = TempSU->getInstr();
634
635 // Following condition is true for all the instructions until PacketMI is
636 // reached (StartCheck is set to 0 before the for loop).
637 // StartCheck flag is 1 for all the instructions after PacketMI.
638 if (TempMI != PacketMI && !StartCheck) // Start processing only after
639 continue; // encountering PacketMI.
640
641 StartCheck = 1;
642 if (TempMI == PacketMI) // We don't want to check PacketMI for dependence.
643 continue;
644
645 for (auto &MO : MI->operands())
646 if (MO.isReg() && TempSU->getInstr()->modifiesRegister(MO.getReg(), HRI))
647 return false;
648 }
649
650 // Make sure that for non-POST_INC stores:
651 // 1. The only use of reg is DepReg and no other registers.
652 // This handles V4 base+index registers.
653 // The following store can not be dot new.
654 // Eg. r0 = add(r0, #3)
655 // memw(r1+r0<<#2) = r0
656 if (!HII->isPostIncrement(MI)) {
657 for (unsigned opNum = 0; opNum < MI->getNumOperands()-1; opNum++) {
658 const MachineOperand &MO = MI->getOperand(opNum);
659 if (MO.isReg() && MO.getReg() == DepReg)
660 return false;
661 }
662 }
663
664 // If data definition is because of implicit definition of the register,
665 // do not newify the store. Eg.
666 // %R9<def> = ZXTH %R12, %D6<imp-use>, %R12<imp-def>
667 // S2_storerh_io %R8, 2, %R12<kill>; mem:ST2[%scevgep343]
668 for (auto &MO : PacketMI->operands()) {
669 if (!MO.isReg() || !MO.isDef() || !MO.isImplicit())
670 continue;
671 unsigned R = MO.getReg();
672 if (R == DepReg || HRI->isSuperRegister(DepReg, R))
673 return false;
674 }
675
676 // Handle imp-use of super reg case. There is a target independent side
677 // change that should prevent this situation but I am handling it for
678 // just-in-case. For example, we cannot newify R2 in the following case:
679 // %R3<def> = A2_tfrsi 0;
680 // S2_storeri_io %R0<kill>, 0, %R2<kill>, %D1<imp-use,kill>;
681 for (auto &MO : MI->operands()) {
682 if (MO.isReg() && MO.isUse() && MO.isImplicit() && MO.getReg() == DepReg)
683 return false;
684 }
685
686 // Can be dot new store.
687 return true;
688 }
689
690 // Can this MI to promoted to either new value store or new value jump.
canPromoteToNewValue(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII)691 bool HexagonPacketizerList::canPromoteToNewValue(const MachineInstr *MI,
692 const SUnit *PacketSU, unsigned DepReg,
693 MachineBasicBlock::iterator &MII) {
694 if (!HII->mayBeNewStore(MI))
695 return false;
696
697 // Check to see the store can be new value'ed.
698 MachineInstr *PacketMI = PacketSU->getInstr();
699 if (canPromoteToNewValueStore(MI, PacketMI, DepReg))
700 return true;
701
702 // Check to see the compare/jump can be new value'ed.
703 // This is done as a pass on its own. Don't need to check it here.
704 return false;
705 }
706
isImplicitDependency(const MachineInstr * I,unsigned DepReg)707 static bool isImplicitDependency(const MachineInstr *I, unsigned DepReg) {
708 for (auto &MO : I->operands())
709 if (MO.isReg() && MO.isDef() && (MO.getReg() == DepReg) && MO.isImplicit())
710 return true;
711 return false;
712 }
713
714 // Check to see if an instruction can be dot new
715 // There are three kinds.
716 // 1. dot new on predicate - V2/V3/V4
717 // 2. dot new on stores NV/ST - V4
718 // 3. dot new on jump NV/J - V4 -- This is generated in a pass.
canPromoteToDotNew(const MachineInstr * MI,const SUnit * PacketSU,unsigned DepReg,MachineBasicBlock::iterator & MII,const TargetRegisterClass * RC)719 bool HexagonPacketizerList::canPromoteToDotNew(const MachineInstr *MI,
720 const SUnit *PacketSU, unsigned DepReg, MachineBasicBlock::iterator &MII,
721 const TargetRegisterClass* RC) {
722 // Already a dot new instruction.
723 if (HII->isDotNewInst(MI) && !HII->mayBeNewStore(MI))
724 return false;
725
726 if (!isNewifiable(MI))
727 return false;
728
729 const MachineInstr *PI = PacketSU->getInstr();
730
731 // The "new value" cannot come from inline asm.
732 if (PI->isInlineAsm())
733 return false;
734
735 // IMPLICIT_DEFs won't materialize as real instructions, so .new makes no
736 // sense.
737 if (PI->isImplicitDef())
738 return false;
739
740 // If dependency is trough an implicitly defined register, we should not
741 // newify the use.
742 if (isImplicitDependency(PI, DepReg))
743 return false;
744
745 const MCInstrDesc& MCID = PI->getDesc();
746 const TargetRegisterClass *VecRC = HII->getRegClass(MCID, 0, HRI, MF);
747 if (DisableVecDblNVStores && VecRC == &Hexagon::VecDblRegsRegClass)
748 return false;
749
750 // predicate .new
751 // bug 5670: until that is fixed
752 // TODO: MI->isIndirectBranch() and IsRegisterJump(MI)
753 if (RC == &Hexagon::PredRegsRegClass)
754 if (HII->isCondInst(MI) || MI->isReturn())
755 return HII->predCanBeUsedAsDotNew(PI, DepReg);
756
757 if (RC != &Hexagon::PredRegsRegClass && !HII->mayBeNewStore(MI))
758 return false;
759
760 // Create a dot new machine instruction to see if resources can be
761 // allocated. If not, bail out now.
762 int NewOpcode = HII->getDotNewOp(MI);
763 const MCInstrDesc &D = HII->get(NewOpcode);
764 MachineInstr *NewMI = MF.CreateMachineInstr(D, DebugLoc());
765 bool ResourcesAvailable = ResourceTracker->canReserveResources(NewMI);
766 MF.DeleteMachineInstr(NewMI);
767 if (!ResourcesAvailable)
768 return false;
769
770 // New Value Store only. New Value Jump generated as a separate pass.
771 if (!canPromoteToNewValue(MI, PacketSU, DepReg, MII))
772 return false;
773
774 return true;
775 }
776
777 // Go through the packet instructions and search for an anti dependency between
778 // them and DepReg from MI. Consider this case:
779 // Trying to add
780 // a) %R1<def> = TFRI_cdNotPt %P3, 2
781 // to this packet:
782 // {
783 // b) %P0<def> = C2_or %P3<kill>, %P0<kill>
784 // c) %P3<def> = C2_tfrrp %R23
785 // d) %R1<def> = C2_cmovenewit %P3, 4
786 // }
787 // The P3 from a) and d) will be complements after
788 // a)'s P3 is converted to .new form
789 // Anti-dep between c) and b) is irrelevant for this case
restrictingDepExistInPacket(MachineInstr * MI,unsigned DepReg)790 bool HexagonPacketizerList::restrictingDepExistInPacket(MachineInstr* MI,
791 unsigned DepReg) {
792 SUnit *PacketSUDep = MIToSUnit.find(MI)->second;
793
794 for (auto I : CurrentPacketMIs) {
795 // We only care for dependencies to predicated instructions
796 if (!HII->isPredicated(I))
797 continue;
798
799 // Scheduling Unit for current insn in the packet
800 SUnit *PacketSU = MIToSUnit.find(I)->second;
801
802 // Look at dependencies between current members of the packet and
803 // predicate defining instruction MI. Make sure that dependency is
804 // on the exact register we care about.
805 if (PacketSU->isSucc(PacketSUDep)) {
806 for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
807 auto &Dep = PacketSU->Succs[i];
808 if (Dep.getSUnit() == PacketSUDep && Dep.getKind() == SDep::Anti &&
809 Dep.getReg() == DepReg)
810 return true;
811 }
812 }
813 }
814
815 return false;
816 }
817
818
819 /// Gets the predicate register of a predicated instruction.
getPredicatedRegister(MachineInstr * MI,const HexagonInstrInfo * QII)820 static unsigned getPredicatedRegister(MachineInstr *MI,
821 const HexagonInstrInfo *QII) {
822 /// We use the following rule: The first predicate register that is a use is
823 /// the predicate register of a predicated instruction.
824 assert(QII->isPredicated(MI) && "Must be predicated instruction");
825
826 for (auto &Op : MI->operands()) {
827 if (Op.isReg() && Op.getReg() && Op.isUse() &&
828 Hexagon::PredRegsRegClass.contains(Op.getReg()))
829 return Op.getReg();
830 }
831
832 llvm_unreachable("Unknown instruction operand layout");
833 return 0;
834 }
835
836 // Given two predicated instructions, this function detects whether
837 // the predicates are complements.
arePredicatesComplements(MachineInstr * MI1,MachineInstr * MI2)838 bool HexagonPacketizerList::arePredicatesComplements(MachineInstr *MI1,
839 MachineInstr *MI2) {
840 // If we don't know the predicate sense of the instructions bail out early, we
841 // need it later.
842 if (getPredicateSense(MI1, HII) == PK_Unknown ||
843 getPredicateSense(MI2, HII) == PK_Unknown)
844 return false;
845
846 // Scheduling unit for candidate.
847 SUnit *SU = MIToSUnit[MI1];
848
849 // One corner case deals with the following scenario:
850 // Trying to add
851 // a) %R24<def> = A2_tfrt %P0, %R25
852 // to this packet:
853 // {
854 // b) %R25<def> = A2_tfrf %P0, %R24
855 // c) %P0<def> = C2_cmpeqi %R26, 1
856 // }
857 //
858 // On general check a) and b) are complements, but presence of c) will
859 // convert a) to .new form, and then it is not a complement.
860 // We attempt to detect it by analyzing existing dependencies in the packet.
861
862 // Analyze relationships between all existing members of the packet.
863 // Look for Anti dependecy on the same predicate reg as used in the
864 // candidate.
865 for (auto I : CurrentPacketMIs) {
866 // Scheduling Unit for current insn in the packet.
867 SUnit *PacketSU = MIToSUnit.find(I)->second;
868
869 // If this instruction in the packet is succeeded by the candidate...
870 if (PacketSU->isSucc(SU)) {
871 for (unsigned i = 0; i < PacketSU->Succs.size(); ++i) {
872 auto Dep = PacketSU->Succs[i];
873 // The corner case exist when there is true data dependency between
874 // candidate and one of current packet members, this dep is on
875 // predicate reg, and there already exist anti dep on the same pred in
876 // the packet.
877 if (Dep.getSUnit() == SU && Dep.getKind() == SDep::Data &&
878 Hexagon::PredRegsRegClass.contains(Dep.getReg())) {
879 // Here I know that I is predicate setting instruction with true
880 // data dep to candidate on the register we care about - c) in the
881 // above example. Now I need to see if there is an anti dependency
882 // from c) to any other instruction in the same packet on the pred
883 // reg of interest.
884 if (restrictingDepExistInPacket(I, Dep.getReg()))
885 return false;
886 }
887 }
888 }
889 }
890
891 // If the above case does not apply, check regular complement condition.
892 // Check that the predicate register is the same and that the predicate
893 // sense is different We also need to differentiate .old vs. .new: !p0
894 // is not complementary to p0.new.
895 unsigned PReg1 = getPredicatedRegister(MI1, HII);
896 unsigned PReg2 = getPredicatedRegister(MI2, HII);
897 return PReg1 == PReg2 &&
898 Hexagon::PredRegsRegClass.contains(PReg1) &&
899 Hexagon::PredRegsRegClass.contains(PReg2) &&
900 getPredicateSense(MI1, HII) != getPredicateSense(MI2, HII) &&
901 HII->isDotNewInst(MI1) == HII->isDotNewInst(MI2);
902 }
903
904 // Initialize packetizer flags.
initPacketizerState()905 void HexagonPacketizerList::initPacketizerState() {
906 Dependence = false;
907 PromotedToDotNew = false;
908 GlueToNewValueJump = false;
909 GlueAllocframeStore = false;
910 FoundSequentialDependence = false;
911 }
912
913 // Ignore bundling of pseudo instructions.
ignorePseudoInstruction(const MachineInstr * MI,const MachineBasicBlock *)914 bool HexagonPacketizerList::ignorePseudoInstruction(const MachineInstr *MI,
915 const MachineBasicBlock*) {
916 if (MI->isDebugValue())
917 return true;
918
919 if (MI->isCFIInstruction())
920 return false;
921
922 // We must print out inline assembly.
923 if (MI->isInlineAsm())
924 return false;
925
926 if (MI->isImplicitDef())
927 return false;
928
929 // We check if MI has any functional units mapped to it. If it doesn't,
930 // we ignore the instruction.
931 const MCInstrDesc& TID = MI->getDesc();
932 auto *IS = ResourceTracker->getInstrItins()->beginStage(TID.getSchedClass());
933 unsigned FuncUnits = IS->getUnits();
934 return !FuncUnits;
935 }
936
isSoloInstruction(const MachineInstr * MI)937 bool HexagonPacketizerList::isSoloInstruction(const MachineInstr *MI) {
938 if (MI->isEHLabel() || MI->isCFIInstruction())
939 return true;
940
941 // Consider inline asm to not be a solo instruction by default.
942 // Inline asm will be put in a packet temporarily, but then it will be
943 // removed, and placed outside of the packet (before or after, depending
944 // on dependencies). This is to reduce the impact of inline asm as a
945 // "packet splitting" instruction.
946 if (MI->isInlineAsm() && !ScheduleInlineAsm)
947 return true;
948
949 // From Hexagon V4 Programmer's Reference Manual 3.4.4 Grouping constraints:
950 // trap, pause, barrier, icinva, isync, and syncht are solo instructions.
951 // They must not be grouped with other instructions in a packet.
952 if (isSchedBarrier(MI))
953 return true;
954
955 if (HII->isSolo(MI))
956 return true;
957
958 if (MI->getOpcode() == Hexagon::A2_nop)
959 return true;
960
961 return false;
962 }
963
964
965 // Quick check if instructions MI and MJ cannot coexist in the same packet.
966 // Limit the tests to be "one-way", e.g. "if MI->isBranch and MJ->isInlineAsm",
967 // but not the symmetric case: "if MJ->isBranch and MI->isInlineAsm".
968 // For full test call this function twice:
969 // cannotCoexistAsymm(MI, MJ) || cannotCoexistAsymm(MJ, MI)
970 // Doing the test only one way saves the amount of code in this function,
971 // since every test would need to be repeated with the MI and MJ reversed.
cannotCoexistAsymm(const MachineInstr * MI,const MachineInstr * MJ,const HexagonInstrInfo & HII)972 static bool cannotCoexistAsymm(const MachineInstr *MI, const MachineInstr *MJ,
973 const HexagonInstrInfo &HII) {
974 const MachineFunction *MF = MI->getParent()->getParent();
975 if (MF->getSubtarget<HexagonSubtarget>().hasV60TOpsOnly() &&
976 HII.isHVXMemWithAIndirect(MI, MJ))
977 return true;
978
979 // An inline asm cannot be together with a branch, because we may not be
980 // able to remove the asm out after packetizing (i.e. if the asm must be
981 // moved past the bundle). Similarly, two asms cannot be together to avoid
982 // complications when determining their relative order outside of a bundle.
983 if (MI->isInlineAsm())
984 return MJ->isInlineAsm() || MJ->isBranch() || MJ->isBarrier() ||
985 MJ->isCall() || MJ->isTerminator();
986
987 // "False" really means that the quick check failed to determine if
988 // I and J cannot coexist.
989 return false;
990 }
991
992
993 // Full, symmetric check.
cannotCoexist(const MachineInstr * MI,const MachineInstr * MJ)994 bool HexagonPacketizerList::cannotCoexist(const MachineInstr *MI,
995 const MachineInstr *MJ) {
996 return cannotCoexistAsymm(MI, MJ, *HII) || cannotCoexistAsymm(MJ, MI, *HII);
997 }
998
unpacketizeSoloInstrs(MachineFunction & MF)999 void HexagonPacketizerList::unpacketizeSoloInstrs(MachineFunction &MF) {
1000 for (auto &B : MF) {
1001 MachineBasicBlock::iterator BundleIt;
1002 MachineBasicBlock::instr_iterator NextI;
1003 for (auto I = B.instr_begin(), E = B.instr_end(); I != E; I = NextI) {
1004 NextI = std::next(I);
1005 MachineInstr *MI = &*I;
1006 if (MI->isBundle())
1007 BundleIt = I;
1008 if (!MI->isInsideBundle())
1009 continue;
1010
1011 // Decide on where to insert the instruction that we are pulling out.
1012 // Debug instructions always go before the bundle, but the placement of
1013 // INLINE_ASM depends on potential dependencies. By default, try to
1014 // put it before the bundle, but if the asm writes to a register that
1015 // other instructions in the bundle read, then we need to place it
1016 // after the bundle (to preserve the bundle semantics).
1017 bool InsertBeforeBundle;
1018 if (MI->isInlineAsm())
1019 InsertBeforeBundle = !hasWriteToReadDep(MI, BundleIt, HRI);
1020 else if (MI->isDebugValue())
1021 InsertBeforeBundle = true;
1022 else
1023 continue;
1024
1025 BundleIt = moveInstrOut(MI, BundleIt, InsertBeforeBundle);
1026 }
1027 }
1028 }
1029
1030 // Check if a given instruction is of class "system".
isSystemInstr(const MachineInstr * MI)1031 static bool isSystemInstr(const MachineInstr *MI) {
1032 unsigned Opc = MI->getOpcode();
1033 switch (Opc) {
1034 case Hexagon::Y2_barrier:
1035 case Hexagon::Y2_dcfetchbo:
1036 return true;
1037 }
1038 return false;
1039 }
1040
hasDeadDependence(const MachineInstr * I,const MachineInstr * J)1041 bool HexagonPacketizerList::hasDeadDependence(const MachineInstr *I,
1042 const MachineInstr *J) {
1043 // The dependence graph may not include edges between dead definitions,
1044 // so without extra checks, we could end up packetizing two instruction
1045 // defining the same (dead) register.
1046 if (I->isCall() || J->isCall())
1047 return false;
1048 if (HII->isPredicated(I) || HII->isPredicated(J))
1049 return false;
1050
1051 BitVector DeadDefs(Hexagon::NUM_TARGET_REGS);
1052 for (auto &MO : I->operands()) {
1053 if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1054 continue;
1055 DeadDefs[MO.getReg()] = true;
1056 }
1057
1058 for (auto &MO : J->operands()) {
1059 if (!MO.isReg() || !MO.isDef() || !MO.isDead())
1060 continue;
1061 unsigned R = MO.getReg();
1062 if (R != Hexagon::USR_OVF && DeadDefs[R])
1063 return true;
1064 }
1065 return false;
1066 }
1067
hasControlDependence(const MachineInstr * I,const MachineInstr * J)1068 bool HexagonPacketizerList::hasControlDependence(const MachineInstr *I,
1069 const MachineInstr *J) {
1070 // A save callee-save register function call can only be in a packet
1071 // with instructions that don't write to the callee-save registers.
1072 if ((HII->isSaveCalleeSavedRegsCall(I) &&
1073 doesModifyCalleeSavedReg(J, HRI)) ||
1074 (HII->isSaveCalleeSavedRegsCall(J) &&
1075 doesModifyCalleeSavedReg(I, HRI)))
1076 return true;
1077
1078 // Two control flow instructions cannot go in the same packet.
1079 if (isControlFlow(I) && isControlFlow(J))
1080 return true;
1081
1082 // \ref-manual (7.3.4) A loop setup packet in loopN or spNloop0 cannot
1083 // contain a speculative indirect jump,
1084 // a new-value compare jump or a dealloc_return.
1085 auto isBadForLoopN = [this] (const MachineInstr *MI) -> bool {
1086 if (MI->isCall() || HII->isDeallocRet(MI) || HII->isNewValueJump(MI))
1087 return true;
1088 if (HII->isPredicated(MI) && HII->isPredicatedNew(MI) && HII->isJumpR(MI))
1089 return true;
1090 return false;
1091 };
1092
1093 if (HII->isLoopN(I) && isBadForLoopN(J))
1094 return true;
1095 if (HII->isLoopN(J) && isBadForLoopN(I))
1096 return true;
1097
1098 // dealloc_return cannot appear in the same packet as a conditional or
1099 // unconditional jump.
1100 return HII->isDeallocRet(I) &&
1101 (J->isBranch() || J->isCall() || J->isBarrier());
1102 }
1103
hasV4SpecificDependence(const MachineInstr * I,const MachineInstr * J)1104 bool HexagonPacketizerList::hasV4SpecificDependence(const MachineInstr *I,
1105 const MachineInstr *J) {
1106 bool SysI = isSystemInstr(I), SysJ = isSystemInstr(J);
1107 bool StoreI = I->mayStore(), StoreJ = J->mayStore();
1108 if ((SysI && StoreJ) || (SysJ && StoreI))
1109 return true;
1110
1111 if (StoreI && StoreJ) {
1112 if (HII->isNewValueInst(J) || HII->isMemOp(J) || HII->isMemOp(I))
1113 return true;
1114 } else {
1115 // A memop cannot be in the same packet with another memop or a store.
1116 // Two stores can be together, but here I and J cannot both be stores.
1117 bool MopStI = HII->isMemOp(I) || StoreI;
1118 bool MopStJ = HII->isMemOp(J) || StoreJ;
1119 if (MopStI && MopStJ)
1120 return true;
1121 }
1122
1123 return (StoreJ && HII->isDeallocRet(I)) || (StoreI && HII->isDeallocRet(J));
1124 }
1125
1126 // SUI is the current instruction that is out side of the current packet.
1127 // SUJ is the current instruction inside the current packet against which that
1128 // SUI will be packetized.
isLegalToPacketizeTogether(SUnit * SUI,SUnit * SUJ)1129 bool HexagonPacketizerList::isLegalToPacketizeTogether(SUnit *SUI, SUnit *SUJ) {
1130 MachineInstr *I = SUI->getInstr();
1131 MachineInstr *J = SUJ->getInstr();
1132 assert(I && J && "Unable to packetize null instruction!");
1133
1134 // Clear IgnoreDepMIs when Packet starts.
1135 if (CurrentPacketMIs.size() == 1)
1136 IgnoreDepMIs.clear();
1137
1138 MachineBasicBlock::iterator II = I;
1139 const unsigned FrameSize = MF.getFrameInfo()->getStackSize();
1140
1141 // Solo instructions cannot go in the packet.
1142 assert(!isSoloInstruction(I) && "Unexpected solo instr!");
1143
1144 if (cannotCoexist(I, J))
1145 return false;
1146
1147 Dependence = hasDeadDependence(I, J) || hasControlDependence(I, J);
1148 if (Dependence)
1149 return false;
1150
1151 // V4 allows dual stores. It does not allow second store, if the first
1152 // store is not in SLOT0. New value store, new value jump, dealloc_return
1153 // and memop always take SLOT0. Arch spec 3.4.4.2.
1154 Dependence = hasV4SpecificDependence(I, J);
1155 if (Dependence)
1156 return false;
1157
1158 // If an instruction feeds new value jump, glue it.
1159 MachineBasicBlock::iterator NextMII = I;
1160 ++NextMII;
1161 if (NextMII != I->getParent()->end() && HII->isNewValueJump(NextMII)) {
1162 MachineInstr *NextMI = NextMII;
1163
1164 bool secondRegMatch = false;
1165 const MachineOperand &NOp0 = NextMI->getOperand(0);
1166 const MachineOperand &NOp1 = NextMI->getOperand(1);
1167
1168 if (NOp1.isReg() && I->getOperand(0).getReg() == NOp1.getReg())
1169 secondRegMatch = true;
1170
1171 for (auto I : CurrentPacketMIs) {
1172 SUnit *PacketSU = MIToSUnit.find(I)->second;
1173 MachineInstr *PI = PacketSU->getInstr();
1174 // NVJ can not be part of the dual jump - Arch Spec: section 7.8.
1175 if (PI->isCall()) {
1176 Dependence = true;
1177 break;
1178 }
1179 // Validate:
1180 // 1. Packet does not have a store in it.
1181 // 2. If the first operand of the nvj is newified, and the second
1182 // operand is also a reg, it (second reg) is not defined in
1183 // the same packet.
1184 // 3. If the second operand of the nvj is newified, (which means
1185 // first operand is also a reg), first reg is not defined in
1186 // the same packet.
1187 if (PI->getOpcode() == Hexagon::S2_allocframe || PI->mayStore() ||
1188 HII->isLoopN(PI)) {
1189 Dependence = true;
1190 break;
1191 }
1192 // Check #2/#3.
1193 const MachineOperand &OpR = secondRegMatch ? NOp0 : NOp1;
1194 if (OpR.isReg() && PI->modifiesRegister(OpR.getReg(), HRI)) {
1195 Dependence = true;
1196 break;
1197 }
1198 }
1199
1200 if (Dependence)
1201 return false;
1202 GlueToNewValueJump = true;
1203 }
1204
1205 // There no dependency between a prolog instruction and its successor.
1206 if (!SUJ->isSucc(SUI))
1207 return true;
1208
1209 for (unsigned i = 0; i < SUJ->Succs.size(); ++i) {
1210 if (FoundSequentialDependence)
1211 break;
1212
1213 if (SUJ->Succs[i].getSUnit() != SUI)
1214 continue;
1215
1216 SDep::Kind DepType = SUJ->Succs[i].getKind();
1217 // For direct calls:
1218 // Ignore register dependences for call instructions for packetization
1219 // purposes except for those due to r31 and predicate registers.
1220 //
1221 // For indirect calls:
1222 // Same as direct calls + check for true dependences to the register
1223 // used in the indirect call.
1224 //
1225 // We completely ignore Order dependences for call instructions.
1226 //
1227 // For returns:
1228 // Ignore register dependences for return instructions like jumpr,
1229 // dealloc return unless we have dependencies on the explicit uses
1230 // of the registers used by jumpr (like r31) or dealloc return
1231 // (like r29 or r30).
1232 //
1233 // TODO: Currently, jumpr is handling only return of r31. So, the
1234 // following logic (specificaly isCallDependent) is working fine.
1235 // We need to enable jumpr for register other than r31 and then,
1236 // we need to rework the last part, where it handles indirect call
1237 // of that (isCallDependent) function. Bug 6216 is opened for this.
1238 unsigned DepReg = 0;
1239 const TargetRegisterClass *RC = nullptr;
1240 if (DepType == SDep::Data) {
1241 DepReg = SUJ->Succs[i].getReg();
1242 RC = HRI->getMinimalPhysRegClass(DepReg);
1243 }
1244
1245 if (I->isCall() || I->isReturn()) {
1246 if (!isRegDependence(DepType))
1247 continue;
1248 if (!isCallDependent(I, DepType, SUJ->Succs[i].getReg()))
1249 continue;
1250 }
1251
1252 if (DepType == SDep::Data) {
1253 if (canPromoteToDotCur(J, SUJ, DepReg, II, RC))
1254 if (promoteToDotCur(J, DepType, II, RC))
1255 continue;
1256 }
1257
1258 // Data dpendence ok if we have load.cur.
1259 if (DepType == SDep::Data && HII->isDotCurInst(J)) {
1260 if (HII->isV60VectorInstruction(I))
1261 continue;
1262 }
1263
1264 // For instructions that can be promoted to dot-new, try to promote.
1265 if (DepType == SDep::Data) {
1266 if (canPromoteToDotNew(I, SUJ, DepReg, II, RC)) {
1267 if (promoteToDotNew(I, DepType, II, RC)) {
1268 PromotedToDotNew = true;
1269 continue;
1270 }
1271 }
1272 if (HII->isNewValueJump(I))
1273 continue;
1274 }
1275
1276 // For predicated instructions, if the predicates are complements then
1277 // there can be no dependence.
1278 if (HII->isPredicated(I) && HII->isPredicated(J) &&
1279 arePredicatesComplements(I, J)) {
1280 // Not always safe to do this translation.
1281 // DAG Builder attempts to reduce dependence edges using transitive
1282 // nature of dependencies. Here is an example:
1283 //
1284 // r0 = tfr_pt ... (1)
1285 // r0 = tfr_pf ... (2)
1286 // r0 = tfr_pt ... (3)
1287 //
1288 // There will be an output dependence between (1)->(2) and (2)->(3).
1289 // However, there is no dependence edge between (1)->(3). This results
1290 // in all 3 instructions going in the same packet. We ignore dependce
1291 // only once to avoid this situation.
1292 auto Itr = std::find(IgnoreDepMIs.begin(), IgnoreDepMIs.end(), J);
1293 if (Itr != IgnoreDepMIs.end()) {
1294 Dependence = true;
1295 return false;
1296 }
1297 IgnoreDepMIs.push_back(I);
1298 continue;
1299 }
1300
1301 // Ignore Order dependences between unconditional direct branches
1302 // and non-control-flow instructions.
1303 if (isDirectJump(I) && !J->isBranch() && !J->isCall() &&
1304 DepType == SDep::Order)
1305 continue;
1306
1307 // Ignore all dependences for jumps except for true and output
1308 // dependences.
1309 if (I->isConditionalBranch() && DepType != SDep::Data &&
1310 DepType != SDep::Output)
1311 continue;
1312
1313 // Ignore output dependences due to superregs. We can write to two
1314 // different subregisters of R1:0 for instance in the same cycle.
1315
1316 // If neither I nor J defines DepReg, then this is a superfluous output
1317 // dependence. The dependence must be of the form:
1318 // R0 = ...
1319 // R1 = ...
1320 // and there is an output dependence between the two instructions with
1321 // DepReg = D0.
1322 // We want to ignore these dependences. Ideally, the dependence
1323 // constructor should annotate such dependences. We can then avoid this
1324 // relatively expensive check.
1325 //
1326 if (DepType == SDep::Output) {
1327 // DepReg is the register that's responsible for the dependence.
1328 unsigned DepReg = SUJ->Succs[i].getReg();
1329
1330 // Check if I and J really defines DepReg.
1331 if (!I->definesRegister(DepReg) && !J->definesRegister(DepReg))
1332 continue;
1333 FoundSequentialDependence = true;
1334 break;
1335 }
1336
1337 // For Order dependences:
1338 // 1. On V4 or later, volatile loads/stores can be packetized together,
1339 // unless other rules prevent is.
1340 // 2. Store followed by a load is not allowed.
1341 // 3. Store followed by a store is only valid on V4 or later.
1342 // 4. Load followed by any memory operation is allowed.
1343 if (DepType == SDep::Order) {
1344 if (!PacketizeVolatiles) {
1345 bool OrdRefs = I->hasOrderedMemoryRef() || J->hasOrderedMemoryRef();
1346 if (OrdRefs) {
1347 FoundSequentialDependence = true;
1348 break;
1349 }
1350 }
1351 // J is first, I is second.
1352 bool LoadJ = J->mayLoad(), StoreJ = J->mayStore();
1353 bool LoadI = I->mayLoad(), StoreI = I->mayStore();
1354 if (StoreJ) {
1355 // Two stores are only allowed on V4+. Load following store is never
1356 // allowed.
1357 if (LoadI) {
1358 FoundSequentialDependence = true;
1359 break;
1360 }
1361 } else if (!LoadJ || (!LoadI && !StoreI)) {
1362 // If J is neither load nor store, assume a dependency.
1363 // If J is a load, but I is neither, also assume a dependency.
1364 FoundSequentialDependence = true;
1365 break;
1366 }
1367 // Store followed by store: not OK on V2.
1368 // Store followed by load: not OK on all.
1369 // Load followed by store: OK on all.
1370 // Load followed by load: OK on all.
1371 continue;
1372 }
1373
1374 // For V4, special case ALLOCFRAME. Even though there is dependency
1375 // between ALLOCFRAME and subsequent store, allow it to be packetized
1376 // in a same packet. This implies that the store is using the caller's
1377 // SP. Hence, offset needs to be updated accordingly.
1378 if (DepType == SDep::Data && J->getOpcode() == Hexagon::S2_allocframe) {
1379 unsigned Opc = I->getOpcode();
1380 switch (Opc) {
1381 case Hexagon::S2_storerd_io:
1382 case Hexagon::S2_storeri_io:
1383 case Hexagon::S2_storerh_io:
1384 case Hexagon::S2_storerb_io:
1385 if (I->getOperand(0).getReg() == HRI->getStackRegister()) {
1386 int64_t Imm = I->getOperand(1).getImm();
1387 int64_t NewOff = Imm - (FrameSize + HEXAGON_LRFP_SIZE);
1388 if (HII->isValidOffset(Opc, NewOff)) {
1389 GlueAllocframeStore = true;
1390 // Since this store is to be glued with allocframe in the same
1391 // packet, it will use SP of the previous stack frame, i.e.
1392 // caller's SP. Therefore, we need to recalculate offset
1393 // according to this change.
1394 I->getOperand(1).setImm(NewOff);
1395 continue;
1396 }
1397 }
1398 default:
1399 break;
1400 }
1401 }
1402
1403 // Skip over anti-dependences. Two instructions that are anti-dependent
1404 // can share a packet.
1405 if (DepType != SDep::Anti) {
1406 FoundSequentialDependence = true;
1407 break;
1408 }
1409 }
1410
1411 if (FoundSequentialDependence) {
1412 Dependence = true;
1413 return false;
1414 }
1415
1416 return true;
1417 }
1418
isLegalToPruneDependencies(SUnit * SUI,SUnit * SUJ)1419 bool HexagonPacketizerList::isLegalToPruneDependencies(SUnit *SUI, SUnit *SUJ) {
1420 MachineInstr *I = SUI->getInstr();
1421 MachineInstr *J = SUJ->getInstr();
1422 assert(I && J && "Unable to packetize null instruction!");
1423
1424 if (cannotCoexist(I, J))
1425 return false;
1426
1427 if (!Dependence)
1428 return true;
1429
1430 // Check if the instruction was promoted to a dot-new. If so, demote it
1431 // back into a dot-old.
1432 if (PromotedToDotNew)
1433 demoteToDotOld(I);
1434
1435 cleanUpDotCur();
1436 // Check if the instruction (must be a store) was glued with an allocframe
1437 // instruction. If so, restore its offset to its original value, i.e. use
1438 // current SP instead of caller's SP.
1439 if (GlueAllocframeStore) {
1440 unsigned FrameSize = MF.getFrameInfo()->getStackSize();
1441 MachineOperand &MOff = I->getOperand(1);
1442 MOff.setImm(MOff.getImm() + FrameSize + HEXAGON_LRFP_SIZE);
1443 }
1444 return false;
1445 }
1446
1447
1448 MachineBasicBlock::iterator
addToPacket(MachineInstr * MI)1449 HexagonPacketizerList::addToPacket(MachineInstr *MI) {
1450 MachineBasicBlock::iterator MII = MI;
1451 MachineBasicBlock *MBB = MI->getParent();
1452 if (MI->isImplicitDef()) {
1453 unsigned R = MI->getOperand(0).getReg();
1454 if (Hexagon::IntRegsRegClass.contains(R)) {
1455 MCSuperRegIterator S(R, HRI, false);
1456 MI->addOperand(MachineOperand::CreateReg(*S, true, true));
1457 }
1458 return MII;
1459 }
1460 assert(ResourceTracker->canReserveResources(MI));
1461
1462 bool ExtMI = HII->isExtended(MI) || HII->isConstExtended(MI);
1463 bool Good = true;
1464
1465 if (GlueToNewValueJump) {
1466 MachineInstr *NvjMI = ++MII;
1467 // We need to put both instructions in the same packet: MI and NvjMI.
1468 // Either of them can require a constant extender. Try to add both to
1469 // the current packet, and if that fails, end the packet and start a
1470 // new one.
1471 ResourceTracker->reserveResources(MI);
1472 if (ExtMI)
1473 Good = tryAllocateResourcesForConstExt(true);
1474
1475 bool ExtNvjMI = HII->isExtended(NvjMI) || HII->isConstExtended(NvjMI);
1476 if (Good) {
1477 if (ResourceTracker->canReserveResources(NvjMI))
1478 ResourceTracker->reserveResources(NvjMI);
1479 else
1480 Good = false;
1481 }
1482 if (Good && ExtNvjMI)
1483 Good = tryAllocateResourcesForConstExt(true);
1484
1485 if (!Good) {
1486 endPacket(MBB, MI);
1487 assert(ResourceTracker->canReserveResources(MI));
1488 ResourceTracker->reserveResources(MI);
1489 if (ExtMI) {
1490 assert(canReserveResourcesForConstExt());
1491 tryAllocateResourcesForConstExt(true);
1492 }
1493 assert(ResourceTracker->canReserveResources(NvjMI));
1494 ResourceTracker->reserveResources(NvjMI);
1495 if (ExtNvjMI) {
1496 assert(canReserveResourcesForConstExt());
1497 reserveResourcesForConstExt();
1498 }
1499 }
1500 CurrentPacketMIs.push_back(MI);
1501 CurrentPacketMIs.push_back(NvjMI);
1502 return MII;
1503 }
1504
1505 ResourceTracker->reserveResources(MI);
1506 if (ExtMI && !tryAllocateResourcesForConstExt(true)) {
1507 endPacket(MBB, MI);
1508 if (PromotedToDotNew)
1509 demoteToDotOld(MI);
1510 ResourceTracker->reserveResources(MI);
1511 reserveResourcesForConstExt();
1512 }
1513
1514 CurrentPacketMIs.push_back(MI);
1515 return MII;
1516 }
1517
endPacket(MachineBasicBlock * MBB,MachineInstr * MI)1518 void HexagonPacketizerList::endPacket(MachineBasicBlock *MBB,
1519 MachineInstr *MI) {
1520 OldPacketMIs = CurrentPacketMIs;
1521 VLIWPacketizerList::endPacket(MBB, MI);
1522 }
1523
shouldAddToPacket(const MachineInstr * MI)1524 bool HexagonPacketizerList::shouldAddToPacket(const MachineInstr *MI) {
1525 return !producesStall(MI);
1526 }
1527
1528
1529 // Return true when ConsMI uses a register defined by ProdMI.
isDependent(const MachineInstr * ProdMI,const MachineInstr * ConsMI)1530 static bool isDependent(const MachineInstr *ProdMI,
1531 const MachineInstr *ConsMI) {
1532 if (!ProdMI->getOperand(0).isReg())
1533 return false;
1534 unsigned DstReg = ProdMI->getOperand(0).getReg();
1535
1536 for (auto &Op : ConsMI->operands())
1537 if (Op.isReg() && Op.isUse() && Op.getReg() == DstReg)
1538 // The MIs depend on each other.
1539 return true;
1540
1541 return false;
1542 }
1543
1544 // V60 forward scheduling.
producesStall(const MachineInstr * I)1545 bool HexagonPacketizerList::producesStall(const MachineInstr *I) {
1546 // Check whether the previous packet is in a different loop. If this is the
1547 // case, there is little point in trying to avoid a stall because that would
1548 // favor the rare case (loop entry) over the common case (loop iteration).
1549 //
1550 // TODO: We should really be able to check all the incoming edges if this is
1551 // the first packet in a basic block, so we can avoid stalls from the loop
1552 // backedge.
1553 if (!OldPacketMIs.empty()) {
1554 auto *OldBB = OldPacketMIs.front()->getParent();
1555 auto *ThisBB = I->getParent();
1556 if (MLI->getLoopFor(OldBB) != MLI->getLoopFor(ThisBB))
1557 return false;
1558 }
1559
1560 // Check for stall between two vector instructions.
1561 if (HII->isV60VectorInstruction(I)) {
1562 for (auto J : OldPacketMIs) {
1563 if (!HII->isV60VectorInstruction(J))
1564 continue;
1565 if (isDependent(J, I) && !HII->isVecUsableNextPacket(J, I))
1566 return true;
1567 }
1568 return false;
1569 }
1570
1571 // Check for stall between two scalar instructions. First, check that
1572 // there is no definition of a use in the current packet, because it
1573 // may be a candidate for .new.
1574 for (auto J : CurrentPacketMIs)
1575 if (!HII->isV60VectorInstruction(J) && isDependent(J, I))
1576 return false;
1577
1578 // Check for stall between I and instructions in the previous packet.
1579 if (MF.getSubtarget<HexagonSubtarget>().useBSBScheduling()) {
1580 for (auto J : OldPacketMIs) {
1581 if (HII->isV60VectorInstruction(J))
1582 continue;
1583 if (!HII->isLateInstrFeedsEarlyInstr(J, I))
1584 continue;
1585 if (isDependent(J, I) && !HII->canExecuteInBundle(J, I))
1586 return true;
1587 }
1588 }
1589
1590 return false;
1591 }
1592
1593
1594 //===----------------------------------------------------------------------===//
1595 // Public Constructor Functions
1596 //===----------------------------------------------------------------------===//
1597
createHexagonPacketizer()1598 FunctionPass *llvm::createHexagonPacketizer() {
1599 return new HexagonPacketizer();
1600 }
1601
1602