]> git.proxmox.com Git - rustc.git/blob - src/llvm/lib/Target/Hexagon/HexagonHardwareLoops.cpp
Imported Upstream version 1.0.0+dfsg1
[rustc.git] / src / llvm / lib / Target / Hexagon / HexagonHardwareLoops.cpp
1 //===-- HexagonHardwareLoops.cpp - Identify and generate hardware loops ---===//
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 pass identifies loops where we can generate the Hexagon hardware
11 // loop instruction. The hardware loop can perform loop branches with a
12 // zero-cycle overhead.
13 //
14 // The pattern that defines the induction variable can changed depending on
15 // prior optimizations. For example, the IndVarSimplify phase run by 'opt'
16 // normalizes induction variables, and the Loop Strength Reduction pass
17 // run by 'llc' may also make changes to the induction variable.
18 // The pattern detected by this phase is due to running Strength Reduction.
19 //
20 // Criteria for hardware loops:
21 // - Countable loops (w/ ind. var for a trip count)
22 // - Assumes loops are normalized by IndVarSimplify
23 // - Try inner-most loops first
24 // - No nested hardware loops.
25 // - No function calls in loops.
26 //
27 //===----------------------------------------------------------------------===//
28
29 #include "llvm/ADT/SmallSet.h"
30 #include "Hexagon.h"
31 #include "HexagonTargetMachine.h"
32 #include "llvm/ADT/Statistic.h"
33 #include "llvm/CodeGen/MachineDominators.h"
34 #include "llvm/CodeGen/MachineFunction.h"
35 #include "llvm/CodeGen/MachineFunctionPass.h"
36 #include "llvm/CodeGen/MachineInstrBuilder.h"
37 #include "llvm/CodeGen/MachineLoopInfo.h"
38 #include "llvm/CodeGen/MachineRegisterInfo.h"
39 #include "llvm/PassSupport.h"
40 #include "llvm/Support/CommandLine.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
43 #include "llvm/Target/TargetInstrInfo.h"
44 #include <algorithm>
45 #include <vector>
46
47 using namespace llvm;
48
49 #define DEBUG_TYPE "hwloops"
50
51 #ifndef NDEBUG
52 static cl::opt<int> HWLoopLimit("max-hwloop", cl::Hidden, cl::init(-1));
53 #endif
54
55 STATISTIC(NumHWLoops, "Number of loops converted to hardware loops");
56
57 namespace llvm {
58 void initializeHexagonHardwareLoopsPass(PassRegistry&);
59 }
60
61 namespace {
62 class CountValue;
63 struct HexagonHardwareLoops : public MachineFunctionPass {
64 MachineLoopInfo *MLI;
65 MachineRegisterInfo *MRI;
66 MachineDominatorTree *MDT;
67 const HexagonTargetMachine *TM;
68 const HexagonInstrInfo *TII;
69 const HexagonRegisterInfo *TRI;
70 #ifndef NDEBUG
71 static int Counter;
72 #endif
73
74 public:
75 static char ID;
76
77 HexagonHardwareLoops() : MachineFunctionPass(ID) {
78 initializeHexagonHardwareLoopsPass(*PassRegistry::getPassRegistry());
79 }
80
81 bool runOnMachineFunction(MachineFunction &MF) override;
82
83 const char *getPassName() const override { return "Hexagon Hardware Loops"; }
84
85 void getAnalysisUsage(AnalysisUsage &AU) const override {
86 AU.addRequired<MachineDominatorTree>();
87 AU.addRequired<MachineLoopInfo>();
88 MachineFunctionPass::getAnalysisUsage(AU);
89 }
90
91 private:
92 /// Kinds of comparisons in the compare instructions.
93 struct Comparison {
94 enum Kind {
95 EQ = 0x01,
96 NE = 0x02,
97 L = 0x04, // Less-than property.
98 G = 0x08, // Greater-than property.
99 U = 0x40, // Unsigned property.
100 LTs = L,
101 LEs = L | EQ,
102 GTs = G,
103 GEs = G | EQ,
104 LTu = L | U,
105 LEu = L | EQ | U,
106 GTu = G | U,
107 GEu = G | EQ | U
108 };
109
110 static Kind getSwappedComparison(Kind Cmp) {
111 assert ((!((Cmp & L) && (Cmp & G))) && "Malformed comparison operator");
112 if ((Cmp & L) || (Cmp & G))
113 return (Kind)(Cmp ^ (L|G));
114 return Cmp;
115 }
116 };
117
118 /// \brief Find the register that contains the loop controlling
119 /// induction variable.
120 /// If successful, it will return true and set the \p Reg, \p IVBump
121 /// and \p IVOp arguments. Otherwise it will return false.
122 /// The returned induction register is the register R that follows the
123 /// following induction pattern:
124 /// loop:
125 /// R = phi ..., [ R.next, LatchBlock ]
126 /// R.next = R + #bump
127 /// if (R.next < #N) goto loop
128 /// IVBump is the immediate value added to R, and IVOp is the instruction
129 /// "R.next = R + #bump".
130 bool findInductionRegister(MachineLoop *L, unsigned &Reg,
131 int64_t &IVBump, MachineInstr *&IVOp) const;
132
133 /// \brief Analyze the statements in a loop to determine if the loop
134 /// has a computable trip count and, if so, return a value that represents
135 /// the trip count expression.
136 CountValue *getLoopTripCount(MachineLoop *L,
137 SmallVectorImpl<MachineInstr *> &OldInsts);
138
139 /// \brief Return the expression that represents the number of times
140 /// a loop iterates. The function takes the operands that represent the
141 /// loop start value, loop end value, and induction value. Based upon
142 /// these operands, the function attempts to compute the trip count.
143 /// If the trip count is not directly available (as an immediate value,
144 /// or a register), the function will attempt to insert computation of it
145 /// to the loop's preheader.
146 CountValue *computeCount(MachineLoop *Loop,
147 const MachineOperand *Start,
148 const MachineOperand *End,
149 unsigned IVReg,
150 int64_t IVBump,
151 Comparison::Kind Cmp) const;
152
153 /// \brief Return true if the instruction is not valid within a hardware
154 /// loop.
155 bool isInvalidLoopOperation(const MachineInstr *MI) const;
156
157 /// \brief Return true if the loop contains an instruction that inhibits
158 /// using the hardware loop.
159 bool containsInvalidInstruction(MachineLoop *L) const;
160
161 /// \brief Given a loop, check if we can convert it to a hardware loop.
162 /// If so, then perform the conversion and return true.
163 bool convertToHardwareLoop(MachineLoop *L);
164
165 /// \brief Return true if the instruction is now dead.
166 bool isDead(const MachineInstr *MI,
167 SmallVectorImpl<MachineInstr *> &DeadPhis) const;
168
169 /// \brief Remove the instruction if it is now dead.
170 void removeIfDead(MachineInstr *MI);
171
172 /// \brief Make sure that the "bump" instruction executes before the
173 /// compare. We need that for the IV fixup, so that the compare
174 /// instruction would not use a bumped value that has not yet been
175 /// defined. If the instructions are out of order, try to reorder them.
176 bool orderBumpCompare(MachineInstr *BumpI, MachineInstr *CmpI);
177
178 /// \brief Get the instruction that loads an immediate value into \p R,
179 /// or 0 if such an instruction does not exist.
180 MachineInstr *defWithImmediate(unsigned R);
181
182 /// \brief Get the immediate value referenced to by \p MO, either for
183 /// immediate operands, or for register operands, where the register
184 /// was defined with an immediate value.
185 int64_t getImmediate(MachineOperand &MO);
186
187 /// \brief Reset the given machine operand to now refer to a new immediate
188 /// value. Assumes that the operand was already referencing an immediate
189 /// value, either directly, or via a register.
190 void setImmediate(MachineOperand &MO, int64_t Val);
191
192 /// \brief Fix the data flow of the induction varible.
193 /// The desired flow is: phi ---> bump -+-> comparison-in-latch.
194 /// |
195 /// +-> back to phi
196 /// where "bump" is the increment of the induction variable:
197 /// iv = iv + #const.
198 /// Due to some prior code transformations, the actual flow may look
199 /// like this:
200 /// phi -+-> bump ---> back to phi
201 /// |
202 /// +-> comparison-in-latch (against upper_bound-bump),
203 /// i.e. the comparison that controls the loop execution may be using
204 /// the value of the induction variable from before the increment.
205 ///
206 /// Return true if the loop's flow is the desired one (i.e. it's
207 /// either been fixed, or no fixing was necessary).
208 /// Otherwise, return false. This can happen if the induction variable
209 /// couldn't be identified, or if the value in the latch's comparison
210 /// cannot be adjusted to reflect the post-bump value.
211 bool fixupInductionVariable(MachineLoop *L);
212
213 /// \brief Given a loop, if it does not have a preheader, create one.
214 /// Return the block that is the preheader.
215 MachineBasicBlock *createPreheaderForLoop(MachineLoop *L);
216 };
217
218 char HexagonHardwareLoops::ID = 0;
219 #ifndef NDEBUG
220 int HexagonHardwareLoops::Counter = 0;
221 #endif
222
223 /// \brief Abstraction for a trip count of a loop. A smaller version
224 /// of the MachineOperand class without the concerns of changing the
225 /// operand representation.
226 class CountValue {
227 public:
228 enum CountValueType {
229 CV_Register,
230 CV_Immediate
231 };
232 private:
233 CountValueType Kind;
234 union Values {
235 struct {
236 unsigned Reg;
237 unsigned Sub;
238 } R;
239 unsigned ImmVal;
240 } Contents;
241
242 public:
243 explicit CountValue(CountValueType t, unsigned v, unsigned u = 0) {
244 Kind = t;
245 if (Kind == CV_Register) {
246 Contents.R.Reg = v;
247 Contents.R.Sub = u;
248 } else {
249 Contents.ImmVal = v;
250 }
251 }
252 bool isReg() const { return Kind == CV_Register; }
253 bool isImm() const { return Kind == CV_Immediate; }
254
255 unsigned getReg() const {
256 assert(isReg() && "Wrong CountValue accessor");
257 return Contents.R.Reg;
258 }
259 unsigned getSubReg() const {
260 assert(isReg() && "Wrong CountValue accessor");
261 return Contents.R.Sub;
262 }
263 unsigned getImm() const {
264 assert(isImm() && "Wrong CountValue accessor");
265 return Contents.ImmVal;
266 }
267
268 void print(raw_ostream &OS, const TargetMachine *TM = nullptr) const {
269 const TargetRegisterInfo *TRI =
270 TM ? TM->getSubtargetImpl()->getRegisterInfo() : nullptr;
271 if (isReg()) { OS << PrintReg(Contents.R.Reg, TRI, Contents.R.Sub); }
272 if (isImm()) { OS << Contents.ImmVal; }
273 }
274 };
275 } // end anonymous namespace
276
277
278 INITIALIZE_PASS_BEGIN(HexagonHardwareLoops, "hwloops",
279 "Hexagon Hardware Loops", false, false)
280 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
281 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
282 INITIALIZE_PASS_END(HexagonHardwareLoops, "hwloops",
283 "Hexagon Hardware Loops", false, false)
284
285
286 /// \brief Returns true if the instruction is a hardware loop instruction.
287 static bool isHardwareLoop(const MachineInstr *MI) {
288 return MI->getOpcode() == Hexagon::J2_loop0r ||
289 MI->getOpcode() == Hexagon::J2_loop0i;
290 }
291
292 FunctionPass *llvm::createHexagonHardwareLoops() {
293 return new HexagonHardwareLoops();
294 }
295
296
297 bool HexagonHardwareLoops::runOnMachineFunction(MachineFunction &MF) {
298 DEBUG(dbgs() << "********* Hexagon Hardware Loops *********\n");
299
300 bool Changed = false;
301
302 MLI = &getAnalysis<MachineLoopInfo>();
303 MRI = &MF.getRegInfo();
304 MDT = &getAnalysis<MachineDominatorTree>();
305 TM = static_cast<const HexagonTargetMachine*>(&MF.getTarget());
306 TII = static_cast<const HexagonInstrInfo *>(
307 TM->getSubtargetImpl()->getInstrInfo());
308 TRI = static_cast<const HexagonRegisterInfo *>(
309 TM->getSubtargetImpl()->getRegisterInfo());
310
311 for (MachineLoopInfo::iterator I = MLI->begin(), E = MLI->end();
312 I != E; ++I) {
313 MachineLoop *L = *I;
314 if (!L->getParentLoop())
315 Changed |= convertToHardwareLoop(L);
316 }
317
318 return Changed;
319 }
320
321
322 bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L,
323 unsigned &Reg,
324 int64_t &IVBump,
325 MachineInstr *&IVOp
326 ) const {
327 MachineBasicBlock *Header = L->getHeader();
328 MachineBasicBlock *Preheader = L->getLoopPreheader();
329 MachineBasicBlock *Latch = L->getLoopLatch();
330 if (!Header || !Preheader || !Latch)
331 return false;
332
333 // This pair represents an induction register together with an immediate
334 // value that will be added to it in each loop iteration.
335 typedef std::pair<unsigned,int64_t> RegisterBump;
336
337 // Mapping: R.next -> (R, bump), where R, R.next and bump are derived
338 // from an induction operation
339 // R.next = R + bump
340 // where bump is an immediate value.
341 typedef std::map<unsigned,RegisterBump> InductionMap;
342
343 InductionMap IndMap;
344
345 typedef MachineBasicBlock::instr_iterator instr_iterator;
346 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
347 I != E && I->isPHI(); ++I) {
348 MachineInstr *Phi = &*I;
349
350 // Have a PHI instruction. Get the operand that corresponds to the
351 // latch block, and see if is a result of an addition of form "reg+imm",
352 // where the "reg" is defined by the PHI node we are looking at.
353 for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) {
354 if (Phi->getOperand(i+1).getMBB() != Latch)
355 continue;
356
357 unsigned PhiOpReg = Phi->getOperand(i).getReg();
358 MachineInstr *DI = MRI->getVRegDef(PhiOpReg);
359 unsigned UpdOpc = DI->getOpcode();
360 bool isAdd = (UpdOpc == Hexagon::ADD_ri);
361
362 if (isAdd) {
363 // If the register operand to the add is the PHI we're
364 // looking at, this meets the induction pattern.
365 unsigned IndReg = DI->getOperand(1).getReg();
366 if (MRI->getVRegDef(IndReg) == Phi) {
367 unsigned UpdReg = DI->getOperand(0).getReg();
368 int64_t V = DI->getOperand(2).getImm();
369 IndMap.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V)));
370 }
371 }
372 } // for (i)
373 } // for (instr)
374
375 SmallVector<MachineOperand,2> Cond;
376 MachineBasicBlock *TB = nullptr, *FB = nullptr;
377 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false);
378 if (NotAnalyzed)
379 return false;
380
381 unsigned CSz = Cond.size();
382 assert (CSz == 1 || CSz == 2);
383 unsigned PredR = Cond[CSz-1].getReg();
384
385 MachineInstr *PredI = MRI->getVRegDef(PredR);
386 if (!PredI->isCompare())
387 return false;
388
389 unsigned CmpReg1 = 0, CmpReg2 = 0;
390 int CmpImm = 0, CmpMask = 0;
391 bool CmpAnalyzed = TII->analyzeCompare(PredI, CmpReg1, CmpReg2,
392 CmpMask, CmpImm);
393 // Fail if the compare was not analyzed, or it's not comparing a register
394 // with an immediate value. Not checking the mask here, since we handle
395 // the individual compare opcodes (including CMPb) later on.
396 if (!CmpAnalyzed)
397 return false;
398
399 // Exactly one of the input registers to the comparison should be among
400 // the induction registers.
401 InductionMap::iterator IndMapEnd = IndMap.end();
402 InductionMap::iterator F = IndMapEnd;
403 if (CmpReg1 != 0) {
404 InductionMap::iterator F1 = IndMap.find(CmpReg1);
405 if (F1 != IndMapEnd)
406 F = F1;
407 }
408 if (CmpReg2 != 0) {
409 InductionMap::iterator F2 = IndMap.find(CmpReg2);
410 if (F2 != IndMapEnd) {
411 if (F != IndMapEnd)
412 return false;
413 F = F2;
414 }
415 }
416 if (F == IndMapEnd)
417 return false;
418
419 Reg = F->second.first;
420 IVBump = F->second.second;
421 IVOp = MRI->getVRegDef(F->first);
422 return true;
423 }
424
425
426 /// \brief Analyze the statements in a loop to determine if the loop has
427 /// a computable trip count and, if so, return a value that represents
428 /// the trip count expression.
429 ///
430 /// This function iterates over the phi nodes in the loop to check for
431 /// induction variable patterns that are used in the calculation for
432 /// the number of time the loop is executed.
433 CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L,
434 SmallVectorImpl<MachineInstr *> &OldInsts) {
435 MachineBasicBlock *TopMBB = L->getTopBlock();
436 MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin();
437 assert(PI != TopMBB->pred_end() &&
438 "Loop must have more than one incoming edge!");
439 MachineBasicBlock *Backedge = *PI++;
440 if (PI == TopMBB->pred_end()) // dead loop?
441 return nullptr;
442 MachineBasicBlock *Incoming = *PI++;
443 if (PI != TopMBB->pred_end()) // multiple backedges?
444 return nullptr;
445
446 // Make sure there is one incoming and one backedge and determine which
447 // is which.
448 if (L->contains(Incoming)) {
449 if (L->contains(Backedge))
450 return nullptr;
451 std::swap(Incoming, Backedge);
452 } else if (!L->contains(Backedge))
453 return nullptr;
454
455 // Look for the cmp instruction to determine if we can get a useful trip
456 // count. The trip count can be either a register or an immediate. The
457 // location of the value depends upon the type (reg or imm).
458 MachineBasicBlock *Latch = L->getLoopLatch();
459 if (!Latch)
460 return nullptr;
461
462 unsigned IVReg = 0;
463 int64_t IVBump = 0;
464 MachineInstr *IVOp;
465 bool FoundIV = findInductionRegister(L, IVReg, IVBump, IVOp);
466 if (!FoundIV)
467 return nullptr;
468
469 MachineBasicBlock *Preheader = L->getLoopPreheader();
470
471 MachineOperand *InitialValue = nullptr;
472 MachineInstr *IV_Phi = MRI->getVRegDef(IVReg);
473 for (unsigned i = 1, n = IV_Phi->getNumOperands(); i < n; i += 2) {
474 MachineBasicBlock *MBB = IV_Phi->getOperand(i+1).getMBB();
475 if (MBB == Preheader)
476 InitialValue = &IV_Phi->getOperand(i);
477 else if (MBB == Latch)
478 IVReg = IV_Phi->getOperand(i).getReg(); // Want IV reg after bump.
479 }
480 if (!InitialValue)
481 return nullptr;
482
483 SmallVector<MachineOperand,2> Cond;
484 MachineBasicBlock *TB = nullptr, *FB = nullptr;
485 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false);
486 if (NotAnalyzed)
487 return nullptr;
488
489 MachineBasicBlock *Header = L->getHeader();
490 // TB must be non-null. If FB is also non-null, one of them must be
491 // the header. Otherwise, branch to TB could be exiting the loop, and
492 // the fall through can go to the header.
493 assert (TB && "Latch block without a branch?");
494 assert ((!FB || TB == Header || FB == Header) && "Branches not to header?");
495 if (!TB || (FB && TB != Header && FB != Header))
496 return nullptr;
497
498 // Branches of form "if (!P) ..." cause HexagonInstrInfo::AnalyzeBranch
499 // to put imm(0), followed by P in the vector Cond.
500 // If TB is not the header, it means that the "not-taken" path must lead
501 // to the header.
502 bool Negated = (Cond.size() > 1) ^ (TB != Header);
503 unsigned PredReg = Cond[Cond.size()-1].getReg();
504 MachineInstr *CondI = MRI->getVRegDef(PredReg);
505 unsigned CondOpc = CondI->getOpcode();
506
507 unsigned CmpReg1 = 0, CmpReg2 = 0;
508 int Mask = 0, ImmValue = 0;
509 bool AnalyzedCmp = TII->analyzeCompare(CondI, CmpReg1, CmpReg2,
510 Mask, ImmValue);
511 if (!AnalyzedCmp)
512 return nullptr;
513
514 // The comparison operator type determines how we compute the loop
515 // trip count.
516 OldInsts.push_back(CondI);
517 OldInsts.push_back(IVOp);
518
519 // Sadly, the following code gets information based on the position
520 // of the operands in the compare instruction. This has to be done
521 // this way, because the comparisons check for a specific relationship
522 // between the operands (e.g. is-less-than), rather than to find out
523 // what relationship the operands are in (as on PPC).
524 Comparison::Kind Cmp;
525 bool isSwapped = false;
526 const MachineOperand &Op1 = CondI->getOperand(1);
527 const MachineOperand &Op2 = CondI->getOperand(2);
528 const MachineOperand *EndValue = nullptr;
529
530 if (Op1.isReg()) {
531 if (Op2.isImm() || Op1.getReg() == IVReg)
532 EndValue = &Op2;
533 else {
534 EndValue = &Op1;
535 isSwapped = true;
536 }
537 }
538
539 if (!EndValue)
540 return nullptr;
541
542 switch (CondOpc) {
543 case Hexagon::C2_cmpeqi:
544 case Hexagon::C2_cmpeq:
545 Cmp = !Negated ? Comparison::EQ : Comparison::NE;
546 break;
547 case Hexagon::C2_cmpgtui:
548 case Hexagon::C2_cmpgtu:
549 Cmp = !Negated ? Comparison::GTu : Comparison::LEu;
550 break;
551 case Hexagon::C2_cmpgti:
552 case Hexagon::C2_cmpgt:
553 Cmp = !Negated ? Comparison::GTs : Comparison::LEs;
554 break;
555 // Very limited support for byte/halfword compares.
556 case Hexagon::CMPbEQri_V4:
557 case Hexagon::CMPhEQri_V4: {
558 if (IVBump != 1)
559 return nullptr;
560
561 int64_t InitV, EndV;
562 // Since the comparisons are "ri", the EndValue should be an
563 // immediate. Check it just in case.
564 assert(EndValue->isImm() && "Unrecognized latch comparison");
565 EndV = EndValue->getImm();
566 // Allow InitialValue to be a register defined with an immediate.
567 if (InitialValue->isReg()) {
568 if (!defWithImmediate(InitialValue->getReg()))
569 return nullptr;
570 InitV = getImmediate(*InitialValue);
571 } else {
572 assert(InitialValue->isImm());
573 InitV = InitialValue->getImm();
574 }
575 if (InitV >= EndV)
576 return nullptr;
577 if (CondOpc == Hexagon::CMPbEQri_V4) {
578 if (!isInt<8>(InitV) || !isInt<8>(EndV))
579 return nullptr;
580 } else { // Hexagon::CMPhEQri_V4
581 if (!isInt<16>(InitV) || !isInt<16>(EndV))
582 return nullptr;
583 }
584 Cmp = !Negated ? Comparison::EQ : Comparison::NE;
585 break;
586 }
587 default:
588 return nullptr;
589 }
590
591 if (isSwapped)
592 Cmp = Comparison::getSwappedComparison(Cmp);
593
594 if (InitialValue->isReg()) {
595 unsigned R = InitialValue->getReg();
596 MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
597 if (!MDT->properlyDominates(DefBB, Header))
598 return nullptr;
599 OldInsts.push_back(MRI->getVRegDef(R));
600 }
601 if (EndValue->isReg()) {
602 unsigned R = EndValue->getReg();
603 MachineBasicBlock *DefBB = MRI->getVRegDef(R)->getParent();
604 if (!MDT->properlyDominates(DefBB, Header))
605 return nullptr;
606 }
607
608 return computeCount(L, InitialValue, EndValue, IVReg, IVBump, Cmp);
609 }
610
611 /// \brief Helper function that returns the expression that represents the
612 /// number of times a loop iterates. The function takes the operands that
613 /// represent the loop start value, loop end value, and induction value.
614 /// Based upon these operands, the function attempts to compute the trip count.
615 CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop,
616 const MachineOperand *Start,
617 const MachineOperand *End,
618 unsigned IVReg,
619 int64_t IVBump,
620 Comparison::Kind Cmp) const {
621 // Cannot handle comparison EQ, i.e. while (A == B).
622 if (Cmp == Comparison::EQ)
623 return nullptr;
624
625 // Check if either the start or end values are an assignment of an immediate.
626 // If so, use the immediate value rather than the register.
627 if (Start->isReg()) {
628 const MachineInstr *StartValInstr = MRI->getVRegDef(Start->getReg());
629 if (StartValInstr && StartValInstr->getOpcode() == Hexagon::A2_tfrsi)
630 Start = &StartValInstr->getOperand(1);
631 }
632 if (End->isReg()) {
633 const MachineInstr *EndValInstr = MRI->getVRegDef(End->getReg());
634 if (EndValInstr && EndValInstr->getOpcode() == Hexagon::A2_tfrsi)
635 End = &EndValInstr->getOperand(1);
636 }
637
638 assert (Start->isReg() || Start->isImm());
639 assert (End->isReg() || End->isImm());
640
641 bool CmpLess = Cmp & Comparison::L;
642 bool CmpGreater = Cmp & Comparison::G;
643 bool CmpHasEqual = Cmp & Comparison::EQ;
644
645 // Avoid certain wrap-arounds. This doesn't detect all wrap-arounds.
646 // If loop executes while iv is "less" with the iv value going down, then
647 // the iv must wrap.
648 if (CmpLess && IVBump < 0)
649 return nullptr;
650 // If loop executes while iv is "greater" with the iv value going up, then
651 // the iv must wrap.
652 if (CmpGreater && IVBump > 0)
653 return nullptr;
654
655 if (Start->isImm() && End->isImm()) {
656 // Both, start and end are immediates.
657 int64_t StartV = Start->getImm();
658 int64_t EndV = End->getImm();
659 int64_t Dist = EndV - StartV;
660 if (Dist == 0)
661 return nullptr;
662
663 bool Exact = (Dist % IVBump) == 0;
664
665 if (Cmp == Comparison::NE) {
666 if (!Exact)
667 return nullptr;
668 if ((Dist < 0) ^ (IVBump < 0))
669 return nullptr;
670 }
671
672 // For comparisons that include the final value (i.e. include equality
673 // with the final value), we need to increase the distance by 1.
674 if (CmpHasEqual)
675 Dist = Dist > 0 ? Dist+1 : Dist-1;
676
677 // assert (CmpLess => Dist > 0);
678 assert ((!CmpLess || Dist > 0) && "Loop should never iterate!");
679 // assert (CmpGreater => Dist < 0);
680 assert ((!CmpGreater || Dist < 0) && "Loop should never iterate!");
681
682 // "Normalized" distance, i.e. with the bump set to +-1.
683 int64_t Dist1 = (IVBump > 0) ? (Dist + (IVBump-1)) / IVBump
684 : (-Dist + (-IVBump-1)) / (-IVBump);
685 assert (Dist1 > 0 && "Fishy thing. Both operands have the same sign.");
686
687 uint64_t Count = Dist1;
688
689 if (Count > 0xFFFFFFFFULL)
690 return nullptr;
691
692 return new CountValue(CountValue::CV_Immediate, Count);
693 }
694
695 // A general case: Start and End are some values, but the actual
696 // iteration count may not be available. If it is not, insert
697 // a computation of it into the preheader.
698
699 // If the induction variable bump is not a power of 2, quit.
700 // Othwerise we'd need a general integer division.
701 if (!isPowerOf2_64(abs64(IVBump)))
702 return nullptr;
703
704 MachineBasicBlock *PH = Loop->getLoopPreheader();
705 assert (PH && "Should have a preheader by now");
706 MachineBasicBlock::iterator InsertPos = PH->getFirstTerminator();
707 DebugLoc DL = (InsertPos != PH->end()) ? InsertPos->getDebugLoc()
708 : DebugLoc();
709
710 // If Start is an immediate and End is a register, the trip count
711 // will be "reg - imm". Hexagon's "subtract immediate" instruction
712 // is actually "reg + -imm".
713
714 // If the loop IV is going downwards, i.e. if the bump is negative,
715 // then the iteration count (computed as End-Start) will need to be
716 // negated. To avoid the negation, just swap Start and End.
717 if (IVBump < 0) {
718 std::swap(Start, End);
719 IVBump = -IVBump;
720 }
721 // Cmp may now have a wrong direction, e.g. LEs may now be GEs.
722 // Signedness, and "including equality" are preserved.
723
724 bool RegToImm = Start->isReg() && End->isImm(); // for (reg..imm)
725 bool RegToReg = Start->isReg() && End->isReg(); // for (reg..reg)
726
727 int64_t StartV = 0, EndV = 0;
728 if (Start->isImm())
729 StartV = Start->getImm();
730 if (End->isImm())
731 EndV = End->getImm();
732
733 int64_t AdjV = 0;
734 // To compute the iteration count, we would need this computation:
735 // Count = (End - Start + (IVBump-1)) / IVBump
736 // or, when CmpHasEqual:
737 // Count = (End - Start + (IVBump-1)+1) / IVBump
738 // The "IVBump-1" part is the adjustment (AdjV). We can avoid
739 // generating an instruction specifically to add it if we can adjust
740 // the immediate values for Start or End.
741
742 if (CmpHasEqual) {
743 // Need to add 1 to the total iteration count.
744 if (Start->isImm())
745 StartV--;
746 else if (End->isImm())
747 EndV++;
748 else
749 AdjV += 1;
750 }
751
752 if (Cmp != Comparison::NE) {
753 if (Start->isImm())
754 StartV -= (IVBump-1);
755 else if (End->isImm())
756 EndV += (IVBump-1);
757 else
758 AdjV += (IVBump-1);
759 }
760
761 unsigned R = 0, SR = 0;
762 if (Start->isReg()) {
763 R = Start->getReg();
764 SR = Start->getSubReg();
765 } else {
766 R = End->getReg();
767 SR = End->getSubReg();
768 }
769 const TargetRegisterClass *RC = MRI->getRegClass(R);
770 // Hardware loops cannot handle 64-bit registers. If it's a double
771 // register, it has to have a subregister.
772 if (!SR && RC == &Hexagon::DoubleRegsRegClass)
773 return nullptr;
774 const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass;
775
776 // Compute DistR (register with the distance between Start and End).
777 unsigned DistR, DistSR;
778
779 // Avoid special case, where the start value is an imm(0).
780 if (Start->isImm() && StartV == 0) {
781 DistR = End->getReg();
782 DistSR = End->getSubReg();
783 } else {
784 const MCInstrDesc &SubD = RegToReg ? TII->get(Hexagon::A2_sub) :
785 (RegToImm ? TII->get(Hexagon::SUB_ri) :
786 TII->get(Hexagon::ADD_ri));
787 unsigned SubR = MRI->createVirtualRegister(IntRC);
788 MachineInstrBuilder SubIB =
789 BuildMI(*PH, InsertPos, DL, SubD, SubR);
790
791 if (RegToReg) {
792 SubIB.addReg(End->getReg(), 0, End->getSubReg())
793 .addReg(Start->getReg(), 0, Start->getSubReg());
794 } else if (RegToImm) {
795 SubIB.addImm(EndV)
796 .addReg(Start->getReg(), 0, Start->getSubReg());
797 } else { // ImmToReg
798 SubIB.addReg(End->getReg(), 0, End->getSubReg())
799 .addImm(-StartV);
800 }
801 DistR = SubR;
802 DistSR = 0;
803 }
804
805 // From DistR, compute AdjR (register with the adjusted distance).
806 unsigned AdjR, AdjSR;
807
808 if (AdjV == 0) {
809 AdjR = DistR;
810 AdjSR = DistSR;
811 } else {
812 // Generate CountR = ADD DistR, AdjVal
813 unsigned AddR = MRI->createVirtualRegister(IntRC);
814 const MCInstrDesc &AddD = TII->get(Hexagon::ADD_ri);
815 BuildMI(*PH, InsertPos, DL, AddD, AddR)
816 .addReg(DistR, 0, DistSR)
817 .addImm(AdjV);
818
819 AdjR = AddR;
820 AdjSR = 0;
821 }
822
823 // From AdjR, compute CountR (register with the final count).
824 unsigned CountR, CountSR;
825
826 if (IVBump == 1) {
827 CountR = AdjR;
828 CountSR = AdjSR;
829 } else {
830 // The IV bump is a power of two. Log_2(IV bump) is the shift amount.
831 unsigned Shift = Log2_32(IVBump);
832
833 // Generate NormR = LSR DistR, Shift.
834 unsigned LsrR = MRI->createVirtualRegister(IntRC);
835 const MCInstrDesc &LsrD = TII->get(Hexagon::S2_lsr_i_r);
836 BuildMI(*PH, InsertPos, DL, LsrD, LsrR)
837 .addReg(AdjR, 0, AdjSR)
838 .addImm(Shift);
839
840 CountR = LsrR;
841 CountSR = 0;
842 }
843
844 return new CountValue(CountValue::CV_Register, CountR, CountSR);
845 }
846
847
848 /// \brief Return true if the operation is invalid within hardware loop.
849 bool HexagonHardwareLoops::isInvalidLoopOperation(
850 const MachineInstr *MI) const {
851
852 // call is not allowed because the callee may use a hardware loop
853 if (MI->getDesc().isCall())
854 return true;
855
856 // do not allow nested hardware loops
857 if (isHardwareLoop(MI))
858 return true;
859
860 // check if the instruction defines a hardware loop register
861 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
862 const MachineOperand &MO = MI->getOperand(i);
863 if (!MO.isReg() || !MO.isDef())
864 continue;
865 unsigned R = MO.getReg();
866 if (R == Hexagon::LC0 || R == Hexagon::LC1 ||
867 R == Hexagon::SA0 || R == Hexagon::SA1)
868 return true;
869 }
870 return false;
871 }
872
873
874 /// \brief - Return true if the loop contains an instruction that inhibits
875 /// the use of the hardware loop function.
876 bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L) const {
877 const std::vector<MachineBasicBlock *> &Blocks = L->getBlocks();
878 for (unsigned i = 0, e = Blocks.size(); i != e; ++i) {
879 MachineBasicBlock *MBB = Blocks[i];
880 for (MachineBasicBlock::iterator
881 MII = MBB->begin(), E = MBB->end(); MII != E; ++MII) {
882 const MachineInstr *MI = &*MII;
883 if (isInvalidLoopOperation(MI))
884 return true;
885 }
886 }
887 return false;
888 }
889
890
891 /// \brief Returns true if the instruction is dead. This was essentially
892 /// copied from DeadMachineInstructionElim::isDead, but with special cases
893 /// for inline asm, physical registers and instructions with side effects
894 /// removed.
895 bool HexagonHardwareLoops::isDead(const MachineInstr *MI,
896 SmallVectorImpl<MachineInstr *> &DeadPhis) const {
897 // Examine each operand.
898 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
899 const MachineOperand &MO = MI->getOperand(i);
900 if (!MO.isReg() || !MO.isDef())
901 continue;
902
903 unsigned Reg = MO.getReg();
904 if (MRI->use_nodbg_empty(Reg))
905 continue;
906
907 typedef MachineRegisterInfo::use_nodbg_iterator use_nodbg_iterator;
908
909 // This instruction has users, but if the only user is the phi node for the
910 // parent block, and the only use of that phi node is this instruction, then
911 // this instruction is dead: both it (and the phi node) can be removed.
912 use_nodbg_iterator I = MRI->use_nodbg_begin(Reg);
913 use_nodbg_iterator End = MRI->use_nodbg_end();
914 if (std::next(I) != End || !I->getParent()->isPHI())
915 return false;
916
917 MachineInstr *OnePhi = I->getParent();
918 for (unsigned j = 0, f = OnePhi->getNumOperands(); j != f; ++j) {
919 const MachineOperand &OPO = OnePhi->getOperand(j);
920 if (!OPO.isReg() || !OPO.isDef())
921 continue;
922
923 unsigned OPReg = OPO.getReg();
924 use_nodbg_iterator nextJ;
925 for (use_nodbg_iterator J = MRI->use_nodbg_begin(OPReg);
926 J != End; J = nextJ) {
927 nextJ = std::next(J);
928 MachineOperand &Use = *J;
929 MachineInstr *UseMI = Use.getParent();
930
931 // If the phi node has a user that is not MI, bail...
932 if (MI != UseMI)
933 return false;
934 }
935 }
936 DeadPhis.push_back(OnePhi);
937 }
938
939 // If there are no defs with uses, the instruction is dead.
940 return true;
941 }
942
943 void HexagonHardwareLoops::removeIfDead(MachineInstr *MI) {
944 // This procedure was essentially copied from DeadMachineInstructionElim.
945
946 SmallVector<MachineInstr*, 1> DeadPhis;
947 if (isDead(MI, DeadPhis)) {
948 DEBUG(dbgs() << "HW looping will remove: " << *MI);
949
950 // It is possible that some DBG_VALUE instructions refer to this
951 // instruction. Examine each def operand for such references;
952 // if found, mark the DBG_VALUE as undef (but don't delete it).
953 for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
954 const MachineOperand &MO = MI->getOperand(i);
955 if (!MO.isReg() || !MO.isDef())
956 continue;
957 unsigned Reg = MO.getReg();
958 MachineRegisterInfo::use_iterator nextI;
959 for (MachineRegisterInfo::use_iterator I = MRI->use_begin(Reg),
960 E = MRI->use_end(); I != E; I = nextI) {
961 nextI = std::next(I); // I is invalidated by the setReg
962 MachineOperand &Use = *I;
963 MachineInstr *UseMI = I->getParent();
964 if (UseMI == MI)
965 continue;
966 if (Use.isDebug())
967 UseMI->getOperand(0).setReg(0U);
968 // This may also be a "instr -> phi -> instr" case which can
969 // be removed too.
970 }
971 }
972
973 MI->eraseFromParent();
974 for (unsigned i = 0; i < DeadPhis.size(); ++i)
975 DeadPhis[i]->eraseFromParent();
976 }
977 }
978
979 /// \brief Check if the loop is a candidate for converting to a hardware
980 /// loop. If so, then perform the transformation.
981 ///
982 /// This function works on innermost loops first. A loop can be converted
983 /// if it is a counting loop; either a register value or an immediate.
984 ///
985 /// The code makes several assumptions about the representation of the loop
986 /// in llvm.
987 bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L) {
988 // This is just for sanity.
989 assert(L->getHeader() && "Loop without a header?");
990
991 bool Changed = false;
992 // Process nested loops first.
993 for (MachineLoop::iterator I = L->begin(), E = L->end(); I != E; ++I)
994 Changed |= convertToHardwareLoop(*I);
995
996 // If a nested loop has been converted, then we can't convert this loop.
997 if (Changed)
998 return Changed;
999
1000 #ifndef NDEBUG
1001 // Stop trying after reaching the limit (if any).
1002 int Limit = HWLoopLimit;
1003 if (Limit >= 0) {
1004 if (Counter >= HWLoopLimit)
1005 return false;
1006 Counter++;
1007 }
1008 #endif
1009
1010 // Does the loop contain any invalid instructions?
1011 if (containsInvalidInstruction(L))
1012 return false;
1013
1014 // Is the induction variable bump feeding the latch condition?
1015 if (!fixupInductionVariable(L))
1016 return false;
1017
1018 MachineBasicBlock *LastMBB = L->getExitingBlock();
1019 // Don't generate hw loop if the loop has more than one exit.
1020 if (!LastMBB)
1021 return false;
1022
1023 MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator();
1024 if (LastI == LastMBB->end())
1025 return false;
1026
1027 // Ensure the loop has a preheader: the loop instruction will be
1028 // placed there.
1029 bool NewPreheader = false;
1030 MachineBasicBlock *Preheader = L->getLoopPreheader();
1031 if (!Preheader) {
1032 Preheader = createPreheaderForLoop(L);
1033 if (!Preheader)
1034 return false;
1035 NewPreheader = true;
1036 }
1037 MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator();
1038
1039 SmallVector<MachineInstr*, 2> OldInsts;
1040 // Are we able to determine the trip count for the loop?
1041 CountValue *TripCount = getLoopTripCount(L, OldInsts);
1042 if (!TripCount)
1043 return false;
1044
1045 // Is the trip count available in the preheader?
1046 if (TripCount->isReg()) {
1047 // There will be a use of the register inserted into the preheader,
1048 // so make sure that the register is actually defined at that point.
1049 MachineInstr *TCDef = MRI->getVRegDef(TripCount->getReg());
1050 MachineBasicBlock *BBDef = TCDef->getParent();
1051 if (!NewPreheader) {
1052 if (!MDT->dominates(BBDef, Preheader))
1053 return false;
1054 } else {
1055 // If we have just created a preheader, the dominator tree won't be
1056 // aware of it. Check if the definition of the register dominates
1057 // the header, but is not the header itself.
1058 if (!MDT->properlyDominates(BBDef, L->getHeader()))
1059 return false;
1060 }
1061 }
1062
1063 // Determine the loop start.
1064 MachineBasicBlock *LoopStart = L->getTopBlock();
1065 if (L->getLoopLatch() != LastMBB) {
1066 // When the exit and latch are not the same, use the latch block as the
1067 // start.
1068 // The loop start address is used only after the 1st iteration, and the
1069 // loop latch may contains instrs. that need to be executed after the
1070 // first iteration.
1071 LoopStart = L->getLoopLatch();
1072 // Make sure the latch is a successor of the exit, otherwise it won't work.
1073 if (!LastMBB->isSuccessor(LoopStart))
1074 return false;
1075 }
1076
1077 // Convert the loop to a hardware loop.
1078 DEBUG(dbgs() << "Change to hardware loop at "; L->dump());
1079 DebugLoc DL;
1080 if (InsertPos != Preheader->end())
1081 DL = InsertPos->getDebugLoc();
1082
1083 if (TripCount->isReg()) {
1084 // Create a copy of the loop count register.
1085 unsigned CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass);
1086 BuildMI(*Preheader, InsertPos, DL, TII->get(TargetOpcode::COPY), CountReg)
1087 .addReg(TripCount->getReg(), 0, TripCount->getSubReg());
1088 // Add the Loop instruction to the beginning of the loop.
1089 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::J2_loop0r))
1090 .addMBB(LoopStart)
1091 .addReg(CountReg);
1092 } else {
1093 assert(TripCount->isImm() && "Expecting immediate value for trip count");
1094 // Add the Loop immediate instruction to the beginning of the loop,
1095 // if the immediate fits in the instructions. Otherwise, we need to
1096 // create a new virtual register.
1097 int64_t CountImm = TripCount->getImm();
1098 if (!TII->isValidOffset(Hexagon::J2_loop0i, CountImm)) {
1099 unsigned CountReg = MRI->createVirtualRegister(&Hexagon::IntRegsRegClass);
1100 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::A2_tfrsi), CountReg)
1101 .addImm(CountImm);
1102 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::J2_loop0r))
1103 .addMBB(LoopStart).addReg(CountReg);
1104 } else
1105 BuildMI(*Preheader, InsertPos, DL, TII->get(Hexagon::J2_loop0i))
1106 .addMBB(LoopStart).addImm(CountImm);
1107 }
1108
1109 // Make sure the loop start always has a reference in the CFG. We need
1110 // to create a BlockAddress operand to get this mechanism to work both the
1111 // MachineBasicBlock and BasicBlock objects need the flag set.
1112 LoopStart->setHasAddressTaken();
1113 // This line is needed to set the hasAddressTaken flag on the BasicBlock
1114 // object.
1115 BlockAddress::get(const_cast<BasicBlock *>(LoopStart->getBasicBlock()));
1116
1117 // Replace the loop branch with an endloop instruction.
1118 DebugLoc LastIDL = LastI->getDebugLoc();
1119 BuildMI(*LastMBB, LastI, LastIDL,
1120 TII->get(Hexagon::ENDLOOP0)).addMBB(LoopStart);
1121
1122 // The loop ends with either:
1123 // - a conditional branch followed by an unconditional branch, or
1124 // - a conditional branch to the loop start.
1125 if (LastI->getOpcode() == Hexagon::J2_jumpt ||
1126 LastI->getOpcode() == Hexagon::J2_jumpf) {
1127 // Delete one and change/add an uncond. branch to out of the loop.
1128 MachineBasicBlock *BranchTarget = LastI->getOperand(1).getMBB();
1129 LastI = LastMBB->erase(LastI);
1130 if (!L->contains(BranchTarget)) {
1131 if (LastI != LastMBB->end())
1132 LastI = LastMBB->erase(LastI);
1133 SmallVector<MachineOperand, 0> Cond;
1134 TII->InsertBranch(*LastMBB, BranchTarget, nullptr, Cond, LastIDL);
1135 }
1136 } else {
1137 // Conditional branch to loop start; just delete it.
1138 LastMBB->erase(LastI);
1139 }
1140 delete TripCount;
1141
1142 // The induction operation and the comparison may now be
1143 // unneeded. If these are unneeded, then remove them.
1144 for (unsigned i = 0; i < OldInsts.size(); ++i)
1145 removeIfDead(OldInsts[i]);
1146
1147 ++NumHWLoops;
1148 return true;
1149 }
1150
1151
1152 bool HexagonHardwareLoops::orderBumpCompare(MachineInstr *BumpI,
1153 MachineInstr *CmpI) {
1154 assert (BumpI != CmpI && "Bump and compare in the same instruction?");
1155
1156 MachineBasicBlock *BB = BumpI->getParent();
1157 if (CmpI->getParent() != BB)
1158 return false;
1159
1160 typedef MachineBasicBlock::instr_iterator instr_iterator;
1161 // Check if things are in order to begin with.
1162 for (instr_iterator I = BumpI, E = BB->instr_end(); I != E; ++I)
1163 if (&*I == CmpI)
1164 return true;
1165
1166 // Out of order.
1167 unsigned PredR = CmpI->getOperand(0).getReg();
1168 bool FoundBump = false;
1169 instr_iterator CmpIt = CmpI, NextIt = std::next(CmpIt);
1170 for (instr_iterator I = NextIt, E = BB->instr_end(); I != E; ++I) {
1171 MachineInstr *In = &*I;
1172 for (unsigned i = 0, n = In->getNumOperands(); i < n; ++i) {
1173 MachineOperand &MO = In->getOperand(i);
1174 if (MO.isReg() && MO.isUse()) {
1175 if (MO.getReg() == PredR) // Found an intervening use of PredR.
1176 return false;
1177 }
1178 }
1179
1180 if (In == BumpI) {
1181 instr_iterator After = BumpI;
1182 instr_iterator From = CmpI;
1183 BB->splice(std::next(After), BB, From);
1184 FoundBump = true;
1185 break;
1186 }
1187 }
1188 assert (FoundBump && "Cannot determine instruction order");
1189 return FoundBump;
1190 }
1191
1192
1193 MachineInstr *HexagonHardwareLoops::defWithImmediate(unsigned R) {
1194 MachineInstr *DI = MRI->getVRegDef(R);
1195 unsigned DOpc = DI->getOpcode();
1196 switch (DOpc) {
1197 case Hexagon::A2_tfrsi:
1198 case Hexagon::A2_tfrpi:
1199 case Hexagon::CONST32_Int_Real:
1200 case Hexagon::CONST64_Int_Real:
1201 return DI;
1202 }
1203 return nullptr;
1204 }
1205
1206
1207 int64_t HexagonHardwareLoops::getImmediate(MachineOperand &MO) {
1208 if (MO.isImm())
1209 return MO.getImm();
1210 assert(MO.isReg());
1211 unsigned R = MO.getReg();
1212 MachineInstr *DI = defWithImmediate(R);
1213 assert(DI && "Need an immediate operand");
1214 // All currently supported "define-with-immediate" instructions have the
1215 // actual immediate value in the operand(1).
1216 int64_t v = DI->getOperand(1).getImm();
1217 return v;
1218 }
1219
1220
1221 void HexagonHardwareLoops::setImmediate(MachineOperand &MO, int64_t Val) {
1222 if (MO.isImm()) {
1223 MO.setImm(Val);
1224 return;
1225 }
1226
1227 assert(MO.isReg());
1228 unsigned R = MO.getReg();
1229 MachineInstr *DI = defWithImmediate(R);
1230 if (MRI->hasOneNonDBGUse(R)) {
1231 // If R has only one use, then just change its defining instruction to
1232 // the new immediate value.
1233 DI->getOperand(1).setImm(Val);
1234 return;
1235 }
1236
1237 const TargetRegisterClass *RC = MRI->getRegClass(R);
1238 unsigned NewR = MRI->createVirtualRegister(RC);
1239 MachineBasicBlock &B = *DI->getParent();
1240 DebugLoc DL = DI->getDebugLoc();
1241 BuildMI(B, DI, DL, TII->get(DI->getOpcode()), NewR)
1242 .addImm(Val);
1243 MO.setReg(NewR);
1244 }
1245
1246
1247 bool HexagonHardwareLoops::fixupInductionVariable(MachineLoop *L) {
1248 MachineBasicBlock *Header = L->getHeader();
1249 MachineBasicBlock *Preheader = L->getLoopPreheader();
1250 MachineBasicBlock *Latch = L->getLoopLatch();
1251
1252 if (!Header || !Preheader || !Latch)
1253 return false;
1254
1255 // These data structures follow the same concept as the corresponding
1256 // ones in findInductionRegister (where some comments are).
1257 typedef std::pair<unsigned,int64_t> RegisterBump;
1258 typedef std::pair<unsigned,RegisterBump> RegisterInduction;
1259 typedef std::set<RegisterInduction> RegisterInductionSet;
1260
1261 // Register candidates for induction variables, with their associated bumps.
1262 RegisterInductionSet IndRegs;
1263
1264 // Look for induction patterns:
1265 // vreg1 = PHI ..., [ latch, vreg2 ]
1266 // vreg2 = ADD vreg1, imm
1267 typedef MachineBasicBlock::instr_iterator instr_iterator;
1268 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1269 I != E && I->isPHI(); ++I) {
1270 MachineInstr *Phi = &*I;
1271
1272 // Have a PHI instruction.
1273 for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) {
1274 if (Phi->getOperand(i+1).getMBB() != Latch)
1275 continue;
1276
1277 unsigned PhiReg = Phi->getOperand(i).getReg();
1278 MachineInstr *DI = MRI->getVRegDef(PhiReg);
1279 unsigned UpdOpc = DI->getOpcode();
1280 bool isAdd = (UpdOpc == Hexagon::ADD_ri);
1281
1282 if (isAdd) {
1283 // If the register operand to the add/sub is the PHI we are looking
1284 // at, this meets the induction pattern.
1285 unsigned IndReg = DI->getOperand(1).getReg();
1286 if (MRI->getVRegDef(IndReg) == Phi) {
1287 unsigned UpdReg = DI->getOperand(0).getReg();
1288 int64_t V = DI->getOperand(2).getImm();
1289 IndRegs.insert(std::make_pair(UpdReg, std::make_pair(IndReg, V)));
1290 }
1291 }
1292 } // for (i)
1293 } // for (instr)
1294
1295 if (IndRegs.empty())
1296 return false;
1297
1298 MachineBasicBlock *TB = nullptr, *FB = nullptr;
1299 SmallVector<MachineOperand,2> Cond;
1300 // AnalyzeBranch returns true if it fails to analyze branch.
1301 bool NotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Cond, false);
1302 if (NotAnalyzed)
1303 return false;
1304
1305 // Check if the latch branch is unconditional.
1306 if (Cond.empty())
1307 return false;
1308
1309 if (TB != Header && FB != Header)
1310 // The latch does not go back to the header. Not a latch we know and love.
1311 return false;
1312
1313 // Expecting a predicate register as a condition. It won't be a hardware
1314 // predicate register at this point yet, just a vreg.
1315 // HexagonInstrInfo::AnalyzeBranch for negated branches inserts imm(0)
1316 // into Cond, followed by the predicate register. For non-negated branches
1317 // it's just the register.
1318 unsigned CSz = Cond.size();
1319 if (CSz != 1 && CSz != 2)
1320 return false;
1321
1322 unsigned P = Cond[CSz-1].getReg();
1323 MachineInstr *PredDef = MRI->getVRegDef(P);
1324
1325 if (!PredDef->isCompare())
1326 return false;
1327
1328 SmallSet<unsigned,2> CmpRegs;
1329 MachineOperand *CmpImmOp = nullptr;
1330
1331 // Go over all operands to the compare and look for immediate and register
1332 // operands. Assume that if the compare has a single register use and a
1333 // single immediate operand, then the register is being compared with the
1334 // immediate value.
1335 for (unsigned i = 0, n = PredDef->getNumOperands(); i < n; ++i) {
1336 MachineOperand &MO = PredDef->getOperand(i);
1337 if (MO.isReg()) {
1338 // Skip all implicit references. In one case there was:
1339 // %vreg140<def> = FCMPUGT32_rr %vreg138, %vreg139, %USR<imp-use>
1340 if (MO.isImplicit())
1341 continue;
1342 if (MO.isUse()) {
1343 unsigned R = MO.getReg();
1344 if (!defWithImmediate(R)) {
1345 CmpRegs.insert(MO.getReg());
1346 continue;
1347 }
1348 // Consider the register to be the "immediate" operand.
1349 if (CmpImmOp)
1350 return false;
1351 CmpImmOp = &MO;
1352 }
1353 } else if (MO.isImm()) {
1354 if (CmpImmOp) // A second immediate argument? Confusing. Bail out.
1355 return false;
1356 CmpImmOp = &MO;
1357 }
1358 }
1359
1360 if (CmpRegs.empty())
1361 return false;
1362
1363 // Check if the compared register follows the order we want. Fix if needed.
1364 for (RegisterInductionSet::iterator I = IndRegs.begin(), E = IndRegs.end();
1365 I != E; ++I) {
1366 // This is a success. If the register used in the comparison is one that
1367 // we have identified as a bumped (updated) induction register, there is
1368 // nothing to do.
1369 if (CmpRegs.count(I->first))
1370 return true;
1371
1372 // Otherwise, if the register being compared comes out of a PHI node,
1373 // and has been recognized as following the induction pattern, and is
1374 // compared against an immediate, we can fix it.
1375 const RegisterBump &RB = I->second;
1376 if (CmpRegs.count(RB.first)) {
1377 if (!CmpImmOp)
1378 return false;
1379
1380 int64_t CmpImm = getImmediate(*CmpImmOp);
1381 int64_t V = RB.second;
1382 if (V > 0 && CmpImm+V < CmpImm) // Overflow (64-bit).
1383 return false;
1384 if (V < 0 && CmpImm+V > CmpImm) // Overflow (64-bit).
1385 return false;
1386 CmpImm += V;
1387 // Some forms of cmp-immediate allow u9 and s10. Assume the worst case
1388 // scenario, i.e. an 8-bit value.
1389 if (CmpImmOp->isImm() && !isInt<8>(CmpImm))
1390 return false;
1391
1392 // Make sure that the compare happens after the bump. Otherwise,
1393 // after the fixup, the compare would use a yet-undefined register.
1394 MachineInstr *BumpI = MRI->getVRegDef(I->first);
1395 bool Order = orderBumpCompare(BumpI, PredDef);
1396 if (!Order)
1397 return false;
1398
1399 // Finally, fix the compare instruction.
1400 setImmediate(*CmpImmOp, CmpImm);
1401 for (unsigned i = 0, n = PredDef->getNumOperands(); i < n; ++i) {
1402 MachineOperand &MO = PredDef->getOperand(i);
1403 if (MO.isReg() && MO.getReg() == RB.first) {
1404 MO.setReg(I->first);
1405 return true;
1406 }
1407 }
1408 }
1409 }
1410
1411 return false;
1412 }
1413
1414
1415 /// \brief Create a preheader for a given loop.
1416 MachineBasicBlock *HexagonHardwareLoops::createPreheaderForLoop(
1417 MachineLoop *L) {
1418 if (MachineBasicBlock *TmpPH = L->getLoopPreheader())
1419 return TmpPH;
1420
1421 MachineBasicBlock *Header = L->getHeader();
1422 MachineBasicBlock *Latch = L->getLoopLatch();
1423 MachineFunction *MF = Header->getParent();
1424 DebugLoc DL;
1425
1426 if (!Latch || Header->hasAddressTaken())
1427 return nullptr;
1428
1429 typedef MachineBasicBlock::instr_iterator instr_iterator;
1430
1431 // Verify that all existing predecessors have analyzable branches
1432 // (or no branches at all).
1433 typedef std::vector<MachineBasicBlock*> MBBVector;
1434 MBBVector Preds(Header->pred_begin(), Header->pred_end());
1435 SmallVector<MachineOperand,2> Tmp1;
1436 MachineBasicBlock *TB = nullptr, *FB = nullptr;
1437
1438 if (TII->AnalyzeBranch(*Latch, TB, FB, Tmp1, false))
1439 return nullptr;
1440
1441 for (MBBVector::iterator I = Preds.begin(), E = Preds.end(); I != E; ++I) {
1442 MachineBasicBlock *PB = *I;
1443 if (PB != Latch) {
1444 bool NotAnalyzed = TII->AnalyzeBranch(*PB, TB, FB, Tmp1, false);
1445 if (NotAnalyzed)
1446 return nullptr;
1447 }
1448 }
1449
1450 MachineBasicBlock *NewPH = MF->CreateMachineBasicBlock();
1451 MF->insert(Header, NewPH);
1452
1453 if (Header->pred_size() > 2) {
1454 // Ensure that the header has only two predecessors: the preheader and
1455 // the loop latch. Any additional predecessors of the header should
1456 // join at the newly created preheader. Inspect all PHI nodes from the
1457 // header and create appropriate corresponding PHI nodes in the preheader.
1458
1459 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1460 I != E && I->isPHI(); ++I) {
1461 MachineInstr *PN = &*I;
1462
1463 const MCInstrDesc &PD = TII->get(TargetOpcode::PHI);
1464 MachineInstr *NewPN = MF->CreateMachineInstr(PD, DL);
1465 NewPH->insert(NewPH->end(), NewPN);
1466
1467 unsigned PR = PN->getOperand(0).getReg();
1468 const TargetRegisterClass *RC = MRI->getRegClass(PR);
1469 unsigned NewPR = MRI->createVirtualRegister(RC);
1470 NewPN->addOperand(MachineOperand::CreateReg(NewPR, true));
1471
1472 // Copy all non-latch operands of a header's PHI node to the newly
1473 // created PHI node in the preheader.
1474 for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) {
1475 unsigned PredR = PN->getOperand(i).getReg();
1476 MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB();
1477 if (PredB == Latch)
1478 continue;
1479
1480 NewPN->addOperand(MachineOperand::CreateReg(PredR, false));
1481 NewPN->addOperand(MachineOperand::CreateMBB(PredB));
1482 }
1483
1484 // Remove copied operands from the old PHI node and add the value
1485 // coming from the preheader's PHI.
1486 for (int i = PN->getNumOperands()-2; i > 0; i -= 2) {
1487 MachineBasicBlock *PredB = PN->getOperand(i+1).getMBB();
1488 if (PredB != Latch) {
1489 PN->RemoveOperand(i+1);
1490 PN->RemoveOperand(i);
1491 }
1492 }
1493 PN->addOperand(MachineOperand::CreateReg(NewPR, false));
1494 PN->addOperand(MachineOperand::CreateMBB(NewPH));
1495 }
1496
1497 } else {
1498 assert(Header->pred_size() == 2);
1499
1500 // The header has only two predecessors, but the non-latch predecessor
1501 // is not a preheader (e.g. it has other successors, etc.)
1502 // In such a case we don't need any extra PHI nodes in the new preheader,
1503 // all we need is to adjust existing PHIs in the header to now refer to
1504 // the new preheader.
1505 for (instr_iterator I = Header->instr_begin(), E = Header->instr_end();
1506 I != E && I->isPHI(); ++I) {
1507 MachineInstr *PN = &*I;
1508 for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) {
1509 MachineOperand &MO = PN->getOperand(i+1);
1510 if (MO.getMBB() != Latch)
1511 MO.setMBB(NewPH);
1512 }
1513 }
1514 }
1515
1516 // "Reroute" the CFG edges to link in the new preheader.
1517 // If any of the predecessors falls through to the header, insert a branch
1518 // to the new preheader in that place.
1519 SmallVector<MachineOperand,1> Tmp2;
1520 SmallVector<MachineOperand,1> EmptyCond;
1521
1522 TB = FB = nullptr;
1523
1524 for (MBBVector::iterator I = Preds.begin(), E = Preds.end(); I != E; ++I) {
1525 MachineBasicBlock *PB = *I;
1526 if (PB != Latch) {
1527 Tmp2.clear();
1528 bool NotAnalyzed = TII->AnalyzeBranch(*PB, TB, FB, Tmp2, false);
1529 (void)NotAnalyzed; // suppress compiler warning
1530 assert (!NotAnalyzed && "Should be analyzable!");
1531 if (TB != Header && (Tmp2.empty() || FB != Header))
1532 TII->InsertBranch(*PB, NewPH, nullptr, EmptyCond, DL);
1533 PB->ReplaceUsesOfBlockWith(Header, NewPH);
1534 }
1535 }
1536
1537 // It can happen that the latch block will fall through into the header.
1538 // Insert an unconditional branch to the header.
1539 TB = FB = nullptr;
1540 bool LatchNotAnalyzed = TII->AnalyzeBranch(*Latch, TB, FB, Tmp2, false);
1541 (void)LatchNotAnalyzed; // suppress compiler warning
1542 assert (!LatchNotAnalyzed && "Should be analyzable!");
1543 if (!TB && !FB)
1544 TII->InsertBranch(*Latch, Header, nullptr, EmptyCond, DL);
1545
1546 // Finally, the branch from the preheader to the header.
1547 TII->InsertBranch(*NewPH, Header, nullptr, EmptyCond, DL);
1548 NewPH->addSuccessor(Header);
1549
1550 return NewPH;
1551 }