1 | //===- HexagonHardwareLoops.cpp - Identify and generate hardware loops ----===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | // |
9 | // This pass identifies loops where we can generate the Hexagon hardware |
10 | // loop instruction. The hardware loop can perform loop branches with a |
11 | // zero-cycle overhead. |
12 | // |
13 | // The pattern that defines the induction variable can changed depending on |
14 | // prior optimizations. For example, the IndVarSimplify phase run by 'opt' |
15 | // normalizes induction variables, and the Loop Strength Reduction pass |
16 | // run by 'llc' may also make changes to the induction variable. |
17 | // The pattern detected by this phase is due to running Strength Reduction. |
18 | // |
19 | // Criteria for hardware loops: |
20 | // - Countable loops (w/ ind. var for a trip count) |
21 | // - Assumes loops are normalized by IndVarSimplify |
22 | // - Try inner-most loops first |
23 | // - No function calls in loops. |
24 | // |
25 | //===----------------------------------------------------------------------===// |
26 | |
27 | #include "HexagonInstrInfo.h" |
28 | #include "HexagonSubtarget.h" |
29 | #include "llvm/ADT/ArrayRef.h" |
30 | #include "llvm/ADT/STLExtras.h" |
31 | #include "llvm/ADT/SmallSet.h" |
32 | #include "llvm/ADT/SmallVector.h" |
33 | #include "llvm/ADT/Statistic.h" |
34 | #include "llvm/ADT/StringRef.h" |
35 | #include "llvm/CodeGen/MachineBasicBlock.h" |
36 | #include "llvm/CodeGen/MachineDominators.h" |
37 | #include "llvm/CodeGen/MachineFunction.h" |
38 | #include "llvm/CodeGen/MachineFunctionPass.h" |
39 | #include "llvm/CodeGen/MachineInstr.h" |
40 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
41 | #include "llvm/CodeGen/MachineLoopInfo.h" |
42 | #include "llvm/CodeGen/MachineOperand.h" |
43 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
44 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
45 | #include "llvm/IR/Constants.h" |
46 | #include "llvm/IR/DebugLoc.h" |
47 | #include "llvm/InitializePasses.h" |
48 | #include "llvm/Pass.h" |
49 | #include "llvm/Support/CommandLine.h" |
50 | #include "llvm/Support/Debug.h" |
51 | #include "llvm/Support/ErrorHandling.h" |
52 | #include "llvm/Support/MathExtras.h" |
53 | #include "llvm/Support/raw_ostream.h" |
54 | #include <cassert> |
55 | #include <cstdint> |
56 | #include <cstdlib> |
57 | #include <iterator> |
58 | #include <map> |
59 | #include <set> |
60 | #include <string> |
61 | #include <utility> |
62 | #include <vector> |
63 | |
64 | using namespace llvm; |
65 | |
66 | #define DEBUG_TYPE "hwloops" |
67 | |
68 | #ifndef NDEBUG |
69 | static cl::opt<int> HWLoopLimit("hexagon-max-hwloop" , cl::Hidden, cl::init(-1)); |
70 | |
71 | // Option to create preheader only for a specific function. |
72 | static cl::opt<std::string> PHFn("hexagon-hwloop-phfn" , cl::Hidden, |
73 | cl::init("" )); |
74 | #endif |
75 | |
76 | // Option to create a preheader if one doesn't exist. |
77 | static cl::opt<bool> ("hexagon-hwloop-preheader" , |
78 | cl::Hidden, cl::init(Val: true), |
79 | cl::desc("Add a preheader to a hardware loop if one doesn't exist" )); |
80 | |
81 | // Turn it off by default. If a preheader block is not created here, the |
82 | // software pipeliner may be unable to find a block suitable to serve as |
83 | // a preheader. In that case SWP will not run. |
84 | static cl::opt<bool> ("hwloop-spec-preheader" , cl::Hidden, |
85 | cl::desc("Allow speculation of preheader " |
86 | "instructions" )); |
87 | |
88 | STATISTIC(NumHWLoops, "Number of loops converted to hardware loops" ); |
89 | |
90 | namespace llvm { |
91 | |
92 | FunctionPass *createHexagonHardwareLoops(); |
93 | void initializeHexagonHardwareLoopsPass(PassRegistry&); |
94 | |
95 | } // end namespace llvm |
96 | |
97 | namespace { |
98 | |
99 | class CountValue; |
100 | |
101 | struct HexagonHardwareLoops : public MachineFunctionPass { |
102 | MachineLoopInfo *MLI; |
103 | MachineRegisterInfo *MRI; |
104 | MachineDominatorTree *MDT; |
105 | const HexagonInstrInfo *TII; |
106 | const HexagonRegisterInfo *TRI; |
107 | #ifndef NDEBUG |
108 | static int Counter; |
109 | #endif |
110 | |
111 | public: |
112 | static char ID; |
113 | |
114 | HexagonHardwareLoops() : MachineFunctionPass(ID) {} |
115 | |
116 | bool runOnMachineFunction(MachineFunction &MF) override; |
117 | |
118 | StringRef getPassName() const override { return "Hexagon Hardware Loops" ; } |
119 | |
120 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
121 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
122 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
123 | MachineFunctionPass::getAnalysisUsage(AU); |
124 | } |
125 | |
126 | private: |
127 | using LoopFeederMap = std::map<Register, MachineInstr *>; |
128 | |
129 | /// Kinds of comparisons in the compare instructions. |
130 | struct Comparison { |
131 | enum Kind { |
132 | EQ = 0x01, |
133 | NE = 0x02, |
134 | L = 0x04, |
135 | G = 0x08, |
136 | U = 0x40, |
137 | LTs = L, |
138 | LEs = L | EQ, |
139 | GTs = G, |
140 | GEs = G | EQ, |
141 | LTu = L | U, |
142 | LEu = L | EQ | U, |
143 | GTu = G | U, |
144 | GEu = G | EQ | U |
145 | }; |
146 | |
147 | static Kind getSwappedComparison(Kind Cmp) { |
148 | assert ((!((Cmp & L) && (Cmp & G))) && "Malformed comparison operator" ); |
149 | if ((Cmp & L) || (Cmp & G)) |
150 | return (Kind)(Cmp ^ (L|G)); |
151 | return Cmp; |
152 | } |
153 | |
154 | static Kind getNegatedComparison(Kind Cmp) { |
155 | if ((Cmp & L) || (Cmp & G)) |
156 | return (Kind)((Cmp ^ (L | G)) ^ EQ); |
157 | if ((Cmp & NE) || (Cmp & EQ)) |
158 | return (Kind)(Cmp ^ (EQ | NE)); |
159 | return (Kind)0; |
160 | } |
161 | |
162 | static bool isSigned(Kind Cmp) { |
163 | return (Cmp & (L | G) && !(Cmp & U)); |
164 | } |
165 | |
166 | static bool isUnsigned(Kind Cmp) { |
167 | return (Cmp & U); |
168 | } |
169 | }; |
170 | |
171 | /// Find the register that contains the loop controlling |
172 | /// induction variable. |
173 | /// If successful, it will return true and set the \p Reg, \p IVBump |
174 | /// and \p IVOp arguments. Otherwise it will return false. |
175 | /// The returned induction register is the register R that follows the |
176 | /// following induction pattern: |
177 | /// loop: |
178 | /// R = phi ..., [ R.next, LatchBlock ] |
179 | /// R.next = R + #bump |
180 | /// if (R.next < #N) goto loop |
181 | /// IVBump is the immediate value added to R, and IVOp is the instruction |
182 | /// "R.next = R + #bump". |
183 | bool findInductionRegister(MachineLoop *L, Register &Reg, |
184 | int64_t &IVBump, MachineInstr *&IVOp) const; |
185 | |
186 | /// Return the comparison kind for the specified opcode. |
187 | Comparison::Kind getComparisonKind(unsigned CondOpc, |
188 | MachineOperand *InitialValue, |
189 | const MachineOperand *Endvalue, |
190 | int64_t IVBump) const; |
191 | |
192 | /// Analyze the statements in a loop to determine if the loop |
193 | /// has a computable trip count and, if so, return a value that represents |
194 | /// the trip count expression. |
195 | CountValue *getLoopTripCount(MachineLoop *L, |
196 | SmallVectorImpl<MachineInstr *> &OldInsts); |
197 | |
198 | /// Return the expression that represents the number of times |
199 | /// a loop iterates. The function takes the operands that represent the |
200 | /// loop start value, loop end value, and induction value. Based upon |
201 | /// these operands, the function attempts to compute the trip count. |
202 | /// If the trip count is not directly available (as an immediate value, |
203 | /// or a register), the function will attempt to insert computation of it |
204 | /// to the loop's preheader. |
205 | CountValue *computeCount(MachineLoop *Loop, const MachineOperand *Start, |
206 | const MachineOperand *End, Register IVReg, |
207 | int64_t IVBump, Comparison::Kind Cmp) const; |
208 | |
209 | /// Return true if the instruction is not valid within a hardware |
210 | /// loop. |
211 | bool isInvalidLoopOperation(const MachineInstr *MI, |
212 | bool IsInnerHWLoop) const; |
213 | |
214 | /// Return true if the loop contains an instruction that inhibits |
215 | /// using the hardware loop. |
216 | bool containsInvalidInstruction(MachineLoop *L, bool IsInnerHWLoop) const; |
217 | |
218 | /// Given a loop, check if we can convert it to a hardware loop. |
219 | /// If so, then perform the conversion and return true. |
220 | bool convertToHardwareLoop(MachineLoop *L, bool &L0used, bool &L1used); |
221 | |
222 | /// Return true if the instruction is now dead. |
223 | bool isDead(const MachineInstr *MI, |
224 | SmallVectorImpl<MachineInstr *> &DeadPhis) const; |
225 | |
226 | /// Remove the instruction if it is now dead. |
227 | void removeIfDead(MachineInstr *MI); |
228 | |
229 | /// Make sure that the "bump" instruction executes before the |
230 | /// compare. We need that for the IV fixup, so that the compare |
231 | /// instruction would not use a bumped value that has not yet been |
232 | /// defined. If the instructions are out of order, try to reorder them. |
233 | bool orderBumpCompare(MachineInstr *BumpI, MachineInstr *CmpI); |
234 | |
235 | /// Return true if MO and MI pair is visited only once. If visited |
236 | /// more than once, this indicates there is recursion. In such a case, |
237 | /// return false. |
238 | bool isLoopFeeder(MachineLoop *L, MachineBasicBlock *A, MachineInstr *MI, |
239 | const MachineOperand *MO, |
240 | LoopFeederMap &LoopFeederPhi) const; |
241 | |
242 | /// Return true if the Phi may generate a value that may underflow, |
243 | /// or may wrap. |
244 | bool phiMayWrapOrUnderflow(MachineInstr *Phi, const MachineOperand *EndVal, |
245 | MachineBasicBlock *MBB, MachineLoop *L, |
246 | LoopFeederMap &LoopFeederPhi) const; |
247 | |
248 | /// Return true if the induction variable may underflow an unsigned |
249 | /// value in the first iteration. |
250 | bool loopCountMayWrapOrUnderFlow(const MachineOperand *InitVal, |
251 | const MachineOperand *EndVal, |
252 | MachineBasicBlock *MBB, MachineLoop *L, |
253 | LoopFeederMap &LoopFeederPhi) const; |
254 | |
255 | /// Check if the given operand has a compile-time known constant |
256 | /// value. Return true if yes, and false otherwise. When returning true, set |
257 | /// Val to the corresponding constant value. |
258 | bool checkForImmediate(const MachineOperand &MO, int64_t &Val) const; |
259 | |
260 | /// Check if the operand has a compile-time known constant value. |
261 | bool isImmediate(const MachineOperand &MO) const { |
262 | int64_t V; |
263 | return checkForImmediate(MO, Val&: V); |
264 | } |
265 | |
266 | /// Return the immediate for the specified operand. |
267 | int64_t getImmediate(const MachineOperand &MO) const { |
268 | int64_t V; |
269 | if (!checkForImmediate(MO, Val&: V)) |
270 | llvm_unreachable("Invalid operand" ); |
271 | return V; |
272 | } |
273 | |
274 | /// Reset the given machine operand to now refer to a new immediate |
275 | /// value. Assumes that the operand was already referencing an immediate |
276 | /// value, either directly, or via a register. |
277 | void setImmediate(MachineOperand &MO, int64_t Val); |
278 | |
279 | /// Fix the data flow of the induction variable. |
280 | /// The desired flow is: phi ---> bump -+-> comparison-in-latch. |
281 | /// | |
282 | /// +-> back to phi |
283 | /// where "bump" is the increment of the induction variable: |
284 | /// iv = iv + #const. |
285 | /// Due to some prior code transformations, the actual flow may look |
286 | /// like this: |
287 | /// phi -+-> bump ---> back to phi |
288 | /// | |
289 | /// +-> comparison-in-latch (against upper_bound-bump), |
290 | /// i.e. the comparison that controls the loop execution may be using |
291 | /// the value of the induction variable from before the increment. |
292 | /// |
293 | /// Return true if the loop's flow is the desired one (i.e. it's |
294 | /// either been fixed, or no fixing was necessary). |
295 | /// Otherwise, return false. This can happen if the induction variable |
296 | /// couldn't be identified, or if the value in the latch's comparison |
297 | /// cannot be adjusted to reflect the post-bump value. |
298 | bool fixupInductionVariable(MachineLoop *L); |
299 | |
300 | /// Given a loop, if it does not have a preheader, create one. |
301 | /// Return the block that is the preheader. |
302 | MachineBasicBlock *createPreheaderForLoop(MachineLoop *L); |
303 | }; |
304 | |
305 | char HexagonHardwareLoops::ID = 0; |
306 | #ifndef NDEBUG |
307 | int HexagonHardwareLoops::Counter = 0; |
308 | #endif |
309 | |
310 | /// Abstraction for a trip count of a loop. A smaller version |
311 | /// of the MachineOperand class without the concerns of changing the |
312 | /// operand representation. |
313 | class CountValue { |
314 | public: |
315 | enum CountValueType { |
316 | CV_Register, |
317 | CV_Immediate |
318 | }; |
319 | |
320 | private: |
321 | CountValueType Kind; |
322 | union Values { |
323 | Values() : R{.Reg: Register(), .Sub: 0} {} |
324 | Values(const Values&) = default; |
325 | struct { |
326 | Register Reg; |
327 | unsigned Sub; |
328 | } R; |
329 | unsigned ImmVal; |
330 | } Contents; |
331 | |
332 | public: |
333 | explicit CountValue(CountValueType t, Register v, unsigned u = 0) { |
334 | Kind = t; |
335 | if (Kind == CV_Register) { |
336 | Contents.R.Reg = v; |
337 | Contents.R.Sub = u; |
338 | } else { |
339 | Contents.ImmVal = v; |
340 | } |
341 | } |
342 | |
343 | bool isReg() const { return Kind == CV_Register; } |
344 | bool isImm() const { return Kind == CV_Immediate; } |
345 | |
346 | Register getReg() const { |
347 | assert(isReg() && "Wrong CountValue accessor" ); |
348 | return Contents.R.Reg; |
349 | } |
350 | |
351 | unsigned getSubReg() const { |
352 | assert(isReg() && "Wrong CountValue accessor" ); |
353 | return Contents.R.Sub; |
354 | } |
355 | |
356 | unsigned getImm() const { |
357 | assert(isImm() && "Wrong CountValue accessor" ); |
358 | return Contents.ImmVal; |
359 | } |
360 | |
361 | void print(raw_ostream &OS, const TargetRegisterInfo *TRI = nullptr) const { |
362 | if (isReg()) { OS << printReg(Reg: Contents.R.Reg, TRI, SubIdx: Contents.R.Sub); } |
363 | if (isImm()) { OS << Contents.ImmVal; } |
364 | } |
365 | }; |
366 | |
367 | } // end anonymous namespace |
368 | |
369 | INITIALIZE_PASS_BEGIN(HexagonHardwareLoops, "hwloops" , |
370 | "Hexagon Hardware Loops" , false, false) |
371 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
372 | INITIALIZE_PASS_DEPENDENCY(MachineLoopInfoWrapperPass) |
373 | INITIALIZE_PASS_END(HexagonHardwareLoops, "hwloops" , |
374 | "Hexagon Hardware Loops" , false, false) |
375 | |
376 | FunctionPass *llvm::createHexagonHardwareLoops() { |
377 | return new HexagonHardwareLoops(); |
378 | } |
379 | |
380 | bool HexagonHardwareLoops::runOnMachineFunction(MachineFunction &MF) { |
381 | LLVM_DEBUG(dbgs() << "********* Hexagon Hardware Loops *********\n" ); |
382 | if (skipFunction(F: MF.getFunction())) |
383 | return false; |
384 | |
385 | bool Changed = false; |
386 | |
387 | MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
388 | MRI = &MF.getRegInfo(); |
389 | MDT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
390 | const HexagonSubtarget &HST = MF.getSubtarget<HexagonSubtarget>(); |
391 | TII = HST.getInstrInfo(); |
392 | TRI = HST.getRegisterInfo(); |
393 | |
394 | for (auto &L : *MLI) |
395 | if (L->isOutermost()) { |
396 | bool L0Used = false; |
397 | bool L1Used = false; |
398 | Changed |= convertToHardwareLoop(L, L0used&: L0Used, L1used&: L1Used); |
399 | } |
400 | |
401 | return Changed; |
402 | } |
403 | |
404 | bool HexagonHardwareLoops::findInductionRegister(MachineLoop *L, |
405 | Register &Reg, |
406 | int64_t &IVBump, |
407 | MachineInstr *&IVOp |
408 | ) const { |
409 | MachineBasicBlock * = L->getHeader(); |
410 | MachineBasicBlock * = MLI->findLoopPreheader(L, SpeculativePreheader: SpecPreheader); |
411 | MachineBasicBlock *Latch = L->getLoopLatch(); |
412 | MachineBasicBlock *ExitingBlock = L->findLoopControlBlock(); |
413 | if (!Header || !Preheader || !Latch || !ExitingBlock) |
414 | return false; |
415 | |
416 | // This pair represents an induction register together with an immediate |
417 | // value that will be added to it in each loop iteration. |
418 | using RegisterBump = std::pair<Register, int64_t>; |
419 | |
420 | // Mapping: R.next -> (R, bump), where R, R.next and bump are derived |
421 | // from an induction operation |
422 | // R.next = R + bump |
423 | // where bump is an immediate value. |
424 | using InductionMap = std::map<Register, RegisterBump>; |
425 | |
426 | InductionMap IndMap; |
427 | |
428 | using instr_iterator = MachineBasicBlock::instr_iterator; |
429 | |
430 | for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); |
431 | I != E && I->isPHI(); ++I) { |
432 | MachineInstr *Phi = &*I; |
433 | |
434 | // Have a PHI instruction. Get the operand that corresponds to the |
435 | // latch block, and see if is a result of an addition of form "reg+imm", |
436 | // where the "reg" is defined by the PHI node we are looking at. |
437 | for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) { |
438 | if (Phi->getOperand(i: i+1).getMBB() != Latch) |
439 | continue; |
440 | |
441 | Register PhiOpReg = Phi->getOperand(i).getReg(); |
442 | MachineInstr *DI = MRI->getVRegDef(Reg: PhiOpReg); |
443 | |
444 | if (DI->getDesc().isAdd()) { |
445 | // If the register operand to the add is the PHI we're looking at, this |
446 | // meets the induction pattern. |
447 | Register IndReg = DI->getOperand(i: 1).getReg(); |
448 | MachineOperand &Opnd2 = DI->getOperand(i: 2); |
449 | int64_t V; |
450 | if (MRI->getVRegDef(Reg: IndReg) == Phi && checkForImmediate(MO: Opnd2, Val&: V)) { |
451 | Register UpdReg = DI->getOperand(i: 0).getReg(); |
452 | IndMap.insert(x: std::make_pair(x&: UpdReg, y: std::make_pair(x&: IndReg, y&: V))); |
453 | } |
454 | } |
455 | } // for (i) |
456 | } // for (instr) |
457 | |
458 | SmallVector<MachineOperand,2> Cond; |
459 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
460 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *ExitingBlock, TBB&: TB, FBB&: FB, Cond, AllowModify: false); |
461 | if (NotAnalyzed) |
462 | return false; |
463 | |
464 | Register PredR; |
465 | unsigned PredPos, PredRegFlags; |
466 | if (!TII->getPredReg(Cond, PredReg&: PredR, PredRegPos&: PredPos, PredRegFlags)) |
467 | return false; |
468 | |
469 | MachineInstr *PredI = MRI->getVRegDef(Reg: PredR); |
470 | if (!PredI->isCompare()) |
471 | return false; |
472 | |
473 | Register CmpReg1, CmpReg2; |
474 | int64_t CmpImm = 0, CmpMask = 0; |
475 | bool CmpAnalyzed = |
476 | TII->analyzeCompare(MI: *PredI, SrcReg&: CmpReg1, SrcReg2&: CmpReg2, Mask&: CmpMask, Value&: CmpImm); |
477 | // Fail if the compare was not analyzed, or it's not comparing a register |
478 | // with an immediate value. Not checking the mask here, since we handle |
479 | // the individual compare opcodes (including A4_cmpb*) later on. |
480 | if (!CmpAnalyzed) |
481 | return false; |
482 | |
483 | // Exactly one of the input registers to the comparison should be among |
484 | // the induction registers. |
485 | InductionMap::iterator IndMapEnd = IndMap.end(); |
486 | InductionMap::iterator F = IndMapEnd; |
487 | if (CmpReg1 != 0) { |
488 | InductionMap::iterator F1 = IndMap.find(x: CmpReg1); |
489 | if (F1 != IndMapEnd) |
490 | F = F1; |
491 | } |
492 | if (CmpReg2 != 0) { |
493 | InductionMap::iterator F2 = IndMap.find(x: CmpReg2); |
494 | if (F2 != IndMapEnd) { |
495 | if (F != IndMapEnd) |
496 | return false; |
497 | F = F2; |
498 | } |
499 | } |
500 | if (F == IndMapEnd) |
501 | return false; |
502 | |
503 | Reg = F->second.first; |
504 | IVBump = F->second.second; |
505 | IVOp = MRI->getVRegDef(Reg: F->first); |
506 | return true; |
507 | } |
508 | |
509 | // Return the comparison kind for the specified opcode. |
510 | HexagonHardwareLoops::Comparison::Kind |
511 | HexagonHardwareLoops::getComparisonKind(unsigned CondOpc, |
512 | MachineOperand *InitialValue, |
513 | const MachineOperand *EndValue, |
514 | int64_t IVBump) const { |
515 | Comparison::Kind Cmp = (Comparison::Kind)0; |
516 | switch (CondOpc) { |
517 | case Hexagon::C2_cmpeq: |
518 | case Hexagon::C2_cmpeqi: |
519 | case Hexagon::C2_cmpeqp: |
520 | Cmp = Comparison::EQ; |
521 | break; |
522 | case Hexagon::C4_cmpneq: |
523 | case Hexagon::C4_cmpneqi: |
524 | Cmp = Comparison::NE; |
525 | break; |
526 | case Hexagon::C2_cmplt: |
527 | Cmp = Comparison::LTs; |
528 | break; |
529 | case Hexagon::C2_cmpltu: |
530 | Cmp = Comparison::LTu; |
531 | break; |
532 | case Hexagon::C4_cmplte: |
533 | case Hexagon::C4_cmpltei: |
534 | Cmp = Comparison::LEs; |
535 | break; |
536 | case Hexagon::C4_cmplteu: |
537 | case Hexagon::C4_cmplteui: |
538 | Cmp = Comparison::LEu; |
539 | break; |
540 | case Hexagon::C2_cmpgt: |
541 | case Hexagon::C2_cmpgti: |
542 | case Hexagon::C2_cmpgtp: |
543 | Cmp = Comparison::GTs; |
544 | break; |
545 | case Hexagon::C2_cmpgtu: |
546 | case Hexagon::C2_cmpgtui: |
547 | case Hexagon::C2_cmpgtup: |
548 | Cmp = Comparison::GTu; |
549 | break; |
550 | case Hexagon::C2_cmpgei: |
551 | Cmp = Comparison::GEs; |
552 | break; |
553 | case Hexagon::C2_cmpgeui: |
554 | Cmp = Comparison::GEs; |
555 | break; |
556 | default: |
557 | return (Comparison::Kind)0; |
558 | } |
559 | return Cmp; |
560 | } |
561 | |
562 | /// Analyze the statements in a loop to determine if the loop has |
563 | /// a computable trip count and, if so, return a value that represents |
564 | /// the trip count expression. |
565 | /// |
566 | /// This function iterates over the phi nodes in the loop to check for |
567 | /// induction variable patterns that are used in the calculation for |
568 | /// the number of time the loop is executed. |
569 | CountValue *HexagonHardwareLoops::getLoopTripCount(MachineLoop *L, |
570 | SmallVectorImpl<MachineInstr *> &OldInsts) { |
571 | MachineBasicBlock *TopMBB = L->getTopBlock(); |
572 | MachineBasicBlock::pred_iterator PI = TopMBB->pred_begin(); |
573 | assert(PI != TopMBB->pred_end() && |
574 | "Loop must have more than one incoming edge!" ); |
575 | MachineBasicBlock *Backedge = *PI++; |
576 | if (PI == TopMBB->pred_end()) // dead loop? |
577 | return nullptr; |
578 | MachineBasicBlock *Incoming = *PI++; |
579 | if (PI != TopMBB->pred_end()) // multiple backedges? |
580 | return nullptr; |
581 | |
582 | // Make sure there is one incoming and one backedge and determine which |
583 | // is which. |
584 | if (L->contains(BB: Incoming)) { |
585 | if (L->contains(BB: Backedge)) |
586 | return nullptr; |
587 | std::swap(a&: Incoming, b&: Backedge); |
588 | } else if (!L->contains(BB: Backedge)) |
589 | return nullptr; |
590 | |
591 | // Look for the cmp instruction to determine if we can get a useful trip |
592 | // count. The trip count can be either a register or an immediate. The |
593 | // location of the value depends upon the type (reg or imm). |
594 | MachineBasicBlock *ExitingBlock = L->findLoopControlBlock(); |
595 | if (!ExitingBlock) |
596 | return nullptr; |
597 | |
598 | Register IVReg = 0; |
599 | int64_t IVBump = 0; |
600 | MachineInstr *IVOp; |
601 | bool FoundIV = findInductionRegister(L, Reg&: IVReg, IVBump, IVOp); |
602 | if (!FoundIV) |
603 | return nullptr; |
604 | |
605 | MachineBasicBlock * = MLI->findLoopPreheader(L, SpeculativePreheader: SpecPreheader); |
606 | |
607 | MachineOperand *InitialValue = nullptr; |
608 | MachineInstr *IV_Phi = MRI->getVRegDef(Reg: IVReg); |
609 | MachineBasicBlock *Latch = L->getLoopLatch(); |
610 | for (unsigned i = 1, n = IV_Phi->getNumOperands(); i < n; i += 2) { |
611 | MachineBasicBlock *MBB = IV_Phi->getOperand(i: i+1).getMBB(); |
612 | if (MBB == Preheader) |
613 | InitialValue = &IV_Phi->getOperand(i); |
614 | else if (MBB == Latch) |
615 | IVReg = IV_Phi->getOperand(i).getReg(); // Want IV reg after bump. |
616 | } |
617 | if (!InitialValue) |
618 | return nullptr; |
619 | |
620 | SmallVector<MachineOperand,2> Cond; |
621 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
622 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *ExitingBlock, TBB&: TB, FBB&: FB, Cond, AllowModify: false); |
623 | if (NotAnalyzed) |
624 | return nullptr; |
625 | |
626 | MachineBasicBlock * = L->getHeader(); |
627 | // TB must be non-null. If FB is also non-null, one of them must be |
628 | // the header. Otherwise, branch to TB could be exiting the loop, and |
629 | // the fall through can go to the header. |
630 | assert (TB && "Exit block without a branch?" ); |
631 | if (ExitingBlock != Latch && (TB == Latch || FB == Latch)) { |
632 | MachineBasicBlock *LTB = nullptr, *LFB = nullptr; |
633 | SmallVector<MachineOperand,2> LCond; |
634 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *Latch, TBB&: LTB, FBB&: LFB, Cond&: LCond, AllowModify: false); |
635 | if (NotAnalyzed) |
636 | return nullptr; |
637 | if (TB == Latch) |
638 | TB = (LTB == Header) ? LTB : LFB; |
639 | else |
640 | FB = (LTB == Header) ? LTB: LFB; |
641 | } |
642 | assert ((!FB || TB == Header || FB == Header) && "Branches not to header?" ); |
643 | if (!TB || (FB && TB != Header && FB != Header)) |
644 | return nullptr; |
645 | |
646 | // Branches of form "if (!P) ..." cause HexagonInstrInfo::analyzeBranch |
647 | // to put imm(0), followed by P in the vector Cond. |
648 | // If TB is not the header, it means that the "not-taken" path must lead |
649 | // to the header. |
650 | bool Negated = TII->predOpcodeHasNot(Cond) ^ (TB != Header); |
651 | Register PredReg; |
652 | unsigned PredPos, PredRegFlags; |
653 | if (!TII->getPredReg(Cond, PredReg, PredRegPos&: PredPos, PredRegFlags)) |
654 | return nullptr; |
655 | MachineInstr *CondI = MRI->getVRegDef(Reg: PredReg); |
656 | unsigned CondOpc = CondI->getOpcode(); |
657 | |
658 | Register CmpReg1, CmpReg2; |
659 | int64_t Mask = 0, ImmValue = 0; |
660 | bool AnalyzedCmp = |
661 | TII->analyzeCompare(MI: *CondI, SrcReg&: CmpReg1, SrcReg2&: CmpReg2, Mask, Value&: ImmValue); |
662 | if (!AnalyzedCmp) |
663 | return nullptr; |
664 | |
665 | // The comparison operator type determines how we compute the loop |
666 | // trip count. |
667 | OldInsts.push_back(Elt: CondI); |
668 | OldInsts.push_back(Elt: IVOp); |
669 | |
670 | // Sadly, the following code gets information based on the position |
671 | // of the operands in the compare instruction. This has to be done |
672 | // this way, because the comparisons check for a specific relationship |
673 | // between the operands (e.g. is-less-than), rather than to find out |
674 | // what relationship the operands are in (as on PPC). |
675 | Comparison::Kind Cmp; |
676 | bool isSwapped = false; |
677 | const MachineOperand &Op1 = CondI->getOperand(i: 1); |
678 | const MachineOperand &Op2 = CondI->getOperand(i: 2); |
679 | const MachineOperand *EndValue = nullptr; |
680 | |
681 | if (Op1.isReg()) { |
682 | if (Op2.isImm() || Op1.getReg() == IVReg) |
683 | EndValue = &Op2; |
684 | else { |
685 | EndValue = &Op1; |
686 | isSwapped = true; |
687 | } |
688 | } |
689 | |
690 | if (!EndValue) |
691 | return nullptr; |
692 | |
693 | Cmp = getComparisonKind(CondOpc, InitialValue, EndValue, IVBump); |
694 | if (!Cmp) |
695 | return nullptr; |
696 | if (Negated) |
697 | Cmp = Comparison::getNegatedComparison(Cmp); |
698 | if (isSwapped) |
699 | Cmp = Comparison::getSwappedComparison(Cmp); |
700 | |
701 | if (InitialValue->isReg()) { |
702 | Register R = InitialValue->getReg(); |
703 | MachineBasicBlock *DefBB = MRI->getVRegDef(Reg: R)->getParent(); |
704 | if (!MDT->properlyDominates(A: DefBB, B: Header)) { |
705 | int64_t V; |
706 | if (!checkForImmediate(MO: *InitialValue, Val&: V)) |
707 | return nullptr; |
708 | } |
709 | OldInsts.push_back(Elt: MRI->getVRegDef(Reg: R)); |
710 | } |
711 | if (EndValue->isReg()) { |
712 | Register R = EndValue->getReg(); |
713 | MachineBasicBlock *DefBB = MRI->getVRegDef(Reg: R)->getParent(); |
714 | if (!MDT->properlyDominates(A: DefBB, B: Header)) { |
715 | int64_t V; |
716 | if (!checkForImmediate(MO: *EndValue, Val&: V)) |
717 | return nullptr; |
718 | } |
719 | OldInsts.push_back(Elt: MRI->getVRegDef(Reg: R)); |
720 | } |
721 | |
722 | return computeCount(Loop: L, Start: InitialValue, End: EndValue, IVReg, IVBump, Cmp); |
723 | } |
724 | |
725 | /// Helper function that returns the expression that represents the |
726 | /// number of times a loop iterates. The function takes the operands that |
727 | /// represent the loop start value, loop end value, and induction value. |
728 | /// Based upon these operands, the function attempts to compute the trip count. |
729 | CountValue *HexagonHardwareLoops::computeCount(MachineLoop *Loop, |
730 | const MachineOperand *Start, |
731 | const MachineOperand *End, |
732 | Register IVReg, |
733 | int64_t IVBump, |
734 | Comparison::Kind Cmp) const { |
735 | // Cannot handle comparison EQ, i.e. while (A == B). |
736 | if (Cmp == Comparison::EQ) |
737 | return nullptr; |
738 | |
739 | // Check if either the start or end values are an assignment of an immediate. |
740 | // If so, use the immediate value rather than the register. |
741 | if (Start->isReg()) { |
742 | const MachineInstr *StartValInstr = MRI->getVRegDef(Reg: Start->getReg()); |
743 | if (StartValInstr && (StartValInstr->getOpcode() == Hexagon::A2_tfrsi || |
744 | StartValInstr->getOpcode() == Hexagon::A2_tfrpi)) |
745 | Start = &StartValInstr->getOperand(i: 1); |
746 | } |
747 | if (End->isReg()) { |
748 | const MachineInstr *EndValInstr = MRI->getVRegDef(Reg: End->getReg()); |
749 | if (EndValInstr && (EndValInstr->getOpcode() == Hexagon::A2_tfrsi || |
750 | EndValInstr->getOpcode() == Hexagon::A2_tfrpi)) |
751 | End = &EndValInstr->getOperand(i: 1); |
752 | } |
753 | |
754 | if (!Start->isReg() && !Start->isImm()) |
755 | return nullptr; |
756 | if (!End->isReg() && !End->isImm()) |
757 | return nullptr; |
758 | |
759 | bool CmpLess = Cmp & Comparison::L; |
760 | bool CmpGreater = Cmp & Comparison::G; |
761 | bool CmpHasEqual = Cmp & Comparison::EQ; |
762 | |
763 | // Avoid certain wrap-arounds. This doesn't detect all wrap-arounds. |
764 | if (CmpLess && IVBump < 0) |
765 | // Loop going while iv is "less" with the iv value going down. Must wrap. |
766 | return nullptr; |
767 | |
768 | if (CmpGreater && IVBump > 0) |
769 | // Loop going while iv is "greater" with the iv value going up. Must wrap. |
770 | return nullptr; |
771 | |
772 | // Phis that may feed into the loop. |
773 | LoopFeederMap LoopFeederPhi; |
774 | |
775 | // Check if the initial value may be zero and can be decremented in the first |
776 | // iteration. If the value is zero, the endloop instruction will not decrement |
777 | // the loop counter, so we shouldn't generate a hardware loop in this case. |
778 | if (loopCountMayWrapOrUnderFlow(InitVal: Start, EndVal: End, MBB: Loop->getLoopPreheader(), L: Loop, |
779 | LoopFeederPhi)) |
780 | return nullptr; |
781 | |
782 | if (Start->isImm() && End->isImm()) { |
783 | // Both, start and end are immediates. |
784 | int64_t StartV = Start->getImm(); |
785 | int64_t EndV = End->getImm(); |
786 | int64_t Dist = EndV - StartV; |
787 | if (Dist == 0) |
788 | return nullptr; |
789 | |
790 | bool Exact = (Dist % IVBump) == 0; |
791 | |
792 | if (Cmp == Comparison::NE) { |
793 | if (!Exact) |
794 | return nullptr; |
795 | if ((Dist < 0) ^ (IVBump < 0)) |
796 | return nullptr; |
797 | } |
798 | |
799 | // For comparisons that include the final value (i.e. include equality |
800 | // with the final value), we need to increase the distance by 1. |
801 | if (CmpHasEqual) |
802 | Dist = Dist > 0 ? Dist+1 : Dist-1; |
803 | |
804 | // For the loop to iterate, CmpLess should imply Dist > 0. Similarly, |
805 | // CmpGreater should imply Dist < 0. These conditions could actually |
806 | // fail, for example, in unreachable code (which may still appear to be |
807 | // reachable in the CFG). |
808 | if ((CmpLess && Dist < 0) || (CmpGreater && Dist > 0)) |
809 | return nullptr; |
810 | |
811 | // "Normalized" distance, i.e. with the bump set to +-1. |
812 | int64_t Dist1 = (IVBump > 0) ? (Dist + (IVBump - 1)) / IVBump |
813 | : (-Dist + (-IVBump - 1)) / (-IVBump); |
814 | assert (Dist1 > 0 && "Fishy thing. Both operands have the same sign." ); |
815 | |
816 | uint64_t Count = Dist1; |
817 | |
818 | if (Count > 0xFFFFFFFFULL) |
819 | return nullptr; |
820 | |
821 | return new CountValue(CountValue::CV_Immediate, Count); |
822 | } |
823 | |
824 | // A general case: Start and End are some values, but the actual |
825 | // iteration count may not be available. If it is not, insert |
826 | // a computation of it into the preheader. |
827 | |
828 | // If the induction variable bump is not a power of 2, quit. |
829 | // Othwerise we'd need a general integer division. |
830 | if (!isPowerOf2_64(Value: std::abs(i: IVBump))) |
831 | return nullptr; |
832 | |
833 | MachineBasicBlock *PH = MLI->findLoopPreheader(L: Loop, SpeculativePreheader: SpecPreheader); |
834 | assert (PH && "Should have a preheader by now" ); |
835 | MachineBasicBlock::iterator InsertPos = PH->getFirstTerminator(); |
836 | DebugLoc DL; |
837 | if (InsertPos != PH->end()) |
838 | DL = InsertPos->getDebugLoc(); |
839 | |
840 | // If Start is an immediate and End is a register, the trip count |
841 | // will be "reg - imm". Hexagon's "subtract immediate" instruction |
842 | // is actually "reg + -imm". |
843 | |
844 | // If the loop IV is going downwards, i.e. if the bump is negative, |
845 | // then the iteration count (computed as End-Start) will need to be |
846 | // negated. To avoid the negation, just swap Start and End. |
847 | if (IVBump < 0) { |
848 | std::swap(a&: Start, b&: End); |
849 | IVBump = -IVBump; |
850 | } |
851 | // Cmp may now have a wrong direction, e.g. LEs may now be GEs. |
852 | // Signedness, and "including equality" are preserved. |
853 | |
854 | bool RegToImm = Start->isReg() && End->isImm(); // for (reg..imm) |
855 | bool RegToReg = Start->isReg() && End->isReg(); // for (reg..reg) |
856 | |
857 | int64_t StartV = 0, EndV = 0; |
858 | if (Start->isImm()) |
859 | StartV = Start->getImm(); |
860 | if (End->isImm()) |
861 | EndV = End->getImm(); |
862 | |
863 | int64_t AdjV = 0; |
864 | // To compute the iteration count, we would need this computation: |
865 | // Count = (End - Start + (IVBump-1)) / IVBump |
866 | // or, when CmpHasEqual: |
867 | // Count = (End - Start + (IVBump-1)+1) / IVBump |
868 | // The "IVBump-1" part is the adjustment (AdjV). We can avoid |
869 | // generating an instruction specifically to add it if we can adjust |
870 | // the immediate values for Start or End. |
871 | |
872 | if (CmpHasEqual) { |
873 | // Need to add 1 to the total iteration count. |
874 | if (Start->isImm()) |
875 | StartV--; |
876 | else if (End->isImm()) |
877 | EndV++; |
878 | else |
879 | AdjV += 1; |
880 | } |
881 | |
882 | if (Cmp != Comparison::NE) { |
883 | if (Start->isImm()) |
884 | StartV -= (IVBump-1); |
885 | else if (End->isImm()) |
886 | EndV += (IVBump-1); |
887 | else |
888 | AdjV += (IVBump-1); |
889 | } |
890 | |
891 | Register R = 0; |
892 | unsigned SR = 0; |
893 | if (Start->isReg()) { |
894 | R = Start->getReg(); |
895 | SR = Start->getSubReg(); |
896 | } else { |
897 | R = End->getReg(); |
898 | SR = End->getSubReg(); |
899 | } |
900 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: R); |
901 | // Hardware loops cannot handle 64-bit registers. If it's a double |
902 | // register, it has to have a subregister. |
903 | if (!SR && RC == &Hexagon::DoubleRegsRegClass) |
904 | return nullptr; |
905 | const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass; |
906 | |
907 | // Compute DistR (register with the distance between Start and End). |
908 | Register DistR; |
909 | unsigned DistSR; |
910 | |
911 | // Avoid special case, where the start value is an imm(0). |
912 | if (Start->isImm() && StartV == 0) { |
913 | DistR = End->getReg(); |
914 | DistSR = End->getSubReg(); |
915 | } else { |
916 | const MCInstrDesc &SubD = RegToReg ? TII->get(Opcode: Hexagon::A2_sub) : |
917 | (RegToImm ? TII->get(Opcode: Hexagon::A2_subri) : |
918 | TII->get(Opcode: Hexagon::A2_addi)); |
919 | if (RegToReg || RegToImm) { |
920 | Register SubR = MRI->createVirtualRegister(RegClass: IntRC); |
921 | MachineInstrBuilder SubIB = |
922 | BuildMI(BB&: *PH, I: InsertPos, MIMD: DL, MCID: SubD, DestReg: SubR); |
923 | |
924 | if (RegToReg) |
925 | SubIB.addReg(RegNo: End->getReg(), flags: 0, SubReg: End->getSubReg()) |
926 | .addReg(RegNo: Start->getReg(), flags: 0, SubReg: Start->getSubReg()); |
927 | else |
928 | SubIB.addImm(Val: EndV) |
929 | .addReg(RegNo: Start->getReg(), flags: 0, SubReg: Start->getSubReg()); |
930 | DistR = SubR; |
931 | } else { |
932 | // If the loop has been unrolled, we should use the original loop count |
933 | // instead of recalculating the value. This will avoid additional |
934 | // 'Add' instruction. |
935 | const MachineInstr *EndValInstr = MRI->getVRegDef(Reg: End->getReg()); |
936 | if (EndValInstr->getOpcode() == Hexagon::A2_addi && |
937 | EndValInstr->getOperand(i: 1).getSubReg() == 0 && |
938 | EndValInstr->getOperand(i: 2).getImm() == StartV) { |
939 | DistR = EndValInstr->getOperand(i: 1).getReg(); |
940 | } else { |
941 | Register SubR = MRI->createVirtualRegister(RegClass: IntRC); |
942 | MachineInstrBuilder SubIB = |
943 | BuildMI(BB&: *PH, I: InsertPos, MIMD: DL, MCID: SubD, DestReg: SubR); |
944 | SubIB.addReg(RegNo: End->getReg(), flags: 0, SubReg: End->getSubReg()) |
945 | .addImm(Val: -StartV); |
946 | DistR = SubR; |
947 | } |
948 | } |
949 | DistSR = 0; |
950 | } |
951 | |
952 | // From DistR, compute AdjR (register with the adjusted distance). |
953 | Register AdjR; |
954 | unsigned AdjSR; |
955 | |
956 | if (AdjV == 0) { |
957 | AdjR = DistR; |
958 | AdjSR = DistSR; |
959 | } else { |
960 | // Generate CountR = ADD DistR, AdjVal |
961 | Register AddR = MRI->createVirtualRegister(RegClass: IntRC); |
962 | MCInstrDesc const &AddD = TII->get(Opcode: Hexagon::A2_addi); |
963 | BuildMI(BB&: *PH, I: InsertPos, MIMD: DL, MCID: AddD, DestReg: AddR) |
964 | .addReg(RegNo: DistR, flags: 0, SubReg: DistSR) |
965 | .addImm(Val: AdjV); |
966 | |
967 | AdjR = AddR; |
968 | AdjSR = 0; |
969 | } |
970 | |
971 | // From AdjR, compute CountR (register with the final count). |
972 | Register CountR; |
973 | unsigned CountSR; |
974 | |
975 | if (IVBump == 1) { |
976 | CountR = AdjR; |
977 | CountSR = AdjSR; |
978 | } else { |
979 | // The IV bump is a power of two. Log_2(IV bump) is the shift amount. |
980 | unsigned Shift = Log2_32(Value: IVBump); |
981 | |
982 | // Generate NormR = LSR DistR, Shift. |
983 | Register LsrR = MRI->createVirtualRegister(RegClass: IntRC); |
984 | const MCInstrDesc &LsrD = TII->get(Opcode: Hexagon::S2_lsr_i_r); |
985 | BuildMI(BB&: *PH, I: InsertPos, MIMD: DL, MCID: LsrD, DestReg: LsrR) |
986 | .addReg(RegNo: AdjR, flags: 0, SubReg: AdjSR) |
987 | .addImm(Val: Shift); |
988 | |
989 | CountR = LsrR; |
990 | CountSR = 0; |
991 | } |
992 | |
993 | return new CountValue(CountValue::CV_Register, CountR, CountSR); |
994 | } |
995 | |
996 | /// Return true if the operation is invalid within hardware loop. |
997 | bool HexagonHardwareLoops::isInvalidLoopOperation(const MachineInstr *MI, |
998 | bool IsInnerHWLoop) const { |
999 | // Call is not allowed because the callee may use a hardware loop except for |
1000 | // the case when the call never returns. |
1001 | if (MI->getDesc().isCall()) |
1002 | return !TII->doesNotReturn(CallMI: *MI); |
1003 | |
1004 | // Check if the instruction defines a hardware loop register. |
1005 | using namespace Hexagon; |
1006 | |
1007 | static const Register Regs01[] = { LC0, SA0, LC1, SA1 }; |
1008 | static const Register Regs1[] = { LC1, SA1 }; |
1009 | auto CheckRegs = IsInnerHWLoop ? ArrayRef(Regs01) : ArrayRef(Regs1); |
1010 | for (Register R : CheckRegs) |
1011 | if (MI->modifiesRegister(Reg: R, TRI)) |
1012 | return true; |
1013 | |
1014 | return false; |
1015 | } |
1016 | |
1017 | /// Return true if the loop contains an instruction that inhibits |
1018 | /// the use of the hardware loop instruction. |
1019 | bool HexagonHardwareLoops::containsInvalidInstruction(MachineLoop *L, |
1020 | bool IsInnerHWLoop) const { |
1021 | LLVM_DEBUG(dbgs() << "\nhw_loop head, " |
1022 | << printMBBReference(**L->block_begin())); |
1023 | for (MachineBasicBlock *MBB : L->getBlocks()) { |
1024 | for (const MachineInstr &MI : *MBB) { |
1025 | if (isInvalidLoopOperation(MI: &MI, IsInnerHWLoop)) { |
1026 | LLVM_DEBUG(dbgs() << "\nCannot convert to hw_loop due to:" ; |
1027 | MI.dump();); |
1028 | return true; |
1029 | } |
1030 | } |
1031 | } |
1032 | return false; |
1033 | } |
1034 | |
1035 | /// Returns true if the instruction is dead. This was essentially |
1036 | /// copied from DeadMachineInstructionElim::isDead, but with special cases |
1037 | /// for inline asm, physical registers and instructions with side effects |
1038 | /// removed. |
1039 | bool HexagonHardwareLoops::isDead(const MachineInstr *MI, |
1040 | SmallVectorImpl<MachineInstr *> &DeadPhis) const { |
1041 | // Examine each operand. |
1042 | for (const MachineOperand &MO : MI->operands()) { |
1043 | if (!MO.isReg() || !MO.isDef()) |
1044 | continue; |
1045 | |
1046 | Register Reg = MO.getReg(); |
1047 | if (MRI->use_nodbg_empty(RegNo: Reg)) |
1048 | continue; |
1049 | |
1050 | using use_nodbg_iterator = MachineRegisterInfo::use_nodbg_iterator; |
1051 | |
1052 | // This instruction has users, but if the only user is the phi node for the |
1053 | // parent block, and the only use of that phi node is this instruction, then |
1054 | // this instruction is dead: both it (and the phi node) can be removed. |
1055 | use_nodbg_iterator I = MRI->use_nodbg_begin(RegNo: Reg); |
1056 | use_nodbg_iterator End = MRI->use_nodbg_end(); |
1057 | if (std::next(x: I) != End || !I->getParent()->isPHI()) |
1058 | return false; |
1059 | |
1060 | MachineInstr *OnePhi = I->getParent(); |
1061 | for (const MachineOperand &OPO : OnePhi->operands()) { |
1062 | if (!OPO.isReg() || !OPO.isDef()) |
1063 | continue; |
1064 | |
1065 | Register OPReg = OPO.getReg(); |
1066 | use_nodbg_iterator nextJ; |
1067 | for (use_nodbg_iterator J = MRI->use_nodbg_begin(RegNo: OPReg); |
1068 | J != End; J = nextJ) { |
1069 | nextJ = std::next(x: J); |
1070 | MachineOperand &Use = *J; |
1071 | MachineInstr *UseMI = Use.getParent(); |
1072 | |
1073 | // If the phi node has a user that is not MI, bail. |
1074 | if (MI != UseMI) |
1075 | return false; |
1076 | } |
1077 | } |
1078 | DeadPhis.push_back(Elt: OnePhi); |
1079 | } |
1080 | |
1081 | // If there are no defs with uses, the instruction is dead. |
1082 | return true; |
1083 | } |
1084 | |
1085 | void HexagonHardwareLoops::removeIfDead(MachineInstr *MI) { |
1086 | // This procedure was essentially copied from DeadMachineInstructionElim. |
1087 | |
1088 | SmallVector<MachineInstr*, 1> DeadPhis; |
1089 | if (isDead(MI, DeadPhis)) { |
1090 | LLVM_DEBUG(dbgs() << "HW looping will remove: " << *MI); |
1091 | |
1092 | // It is possible that some DBG_VALUE instructions refer to this |
1093 | // instruction. Examine each def operand for such references; |
1094 | // if found, mark the DBG_VALUE as undef (but don't delete it). |
1095 | for (const MachineOperand &MO : MI->operands()) { |
1096 | if (!MO.isReg() || !MO.isDef()) |
1097 | continue; |
1098 | Register Reg = MO.getReg(); |
1099 | // We use make_early_inc_range here because setReg below invalidates the |
1100 | // iterator. |
1101 | for (MachineOperand &MO : |
1102 | llvm::make_early_inc_range(Range: MRI->use_operands(Reg))) { |
1103 | MachineInstr *UseMI = MO.getParent(); |
1104 | if (UseMI == MI) |
1105 | continue; |
1106 | if (MO.isDebug()) |
1107 | MO.setReg(0U); |
1108 | } |
1109 | } |
1110 | |
1111 | MI->eraseFromParent(); |
1112 | for (unsigned i = 0; i < DeadPhis.size(); ++i) |
1113 | DeadPhis[i]->eraseFromParent(); |
1114 | } |
1115 | } |
1116 | |
1117 | /// Check if the loop is a candidate for converting to a hardware |
1118 | /// loop. If so, then perform the transformation. |
1119 | /// |
1120 | /// This function works on innermost loops first. A loop can be converted |
1121 | /// if it is a counting loop; either a register value or an immediate. |
1122 | /// |
1123 | /// The code makes several assumptions about the representation of the loop |
1124 | /// in llvm. |
1125 | bool HexagonHardwareLoops::convertToHardwareLoop(MachineLoop *L, |
1126 | bool &RecL0used, |
1127 | bool &RecL1used) { |
1128 | // This is just to confirm basic correctness. |
1129 | assert(L->getHeader() && "Loop without a header?" ); |
1130 | |
1131 | bool Changed = false; |
1132 | bool L0Used = false; |
1133 | bool L1Used = false; |
1134 | |
1135 | // Process nested loops first. |
1136 | for (MachineLoop *I : *L) { |
1137 | Changed |= convertToHardwareLoop(L: I, RecL0used, RecL1used); |
1138 | L0Used |= RecL0used; |
1139 | L1Used |= RecL1used; |
1140 | } |
1141 | |
1142 | // If a nested loop has been converted, then we can't convert this loop. |
1143 | if (Changed && L0Used && L1Used) |
1144 | return Changed; |
1145 | |
1146 | unsigned LOOP_i; |
1147 | unsigned LOOP_r; |
1148 | unsigned ENDLOOP; |
1149 | |
1150 | // Flag used to track loopN instruction: |
1151 | // 1 - Hardware loop is being generated for the inner most loop. |
1152 | // 0 - Hardware loop is being generated for the outer loop. |
1153 | unsigned IsInnerHWLoop = 1; |
1154 | |
1155 | if (L0Used) { |
1156 | LOOP_i = Hexagon::J2_loop1i; |
1157 | LOOP_r = Hexagon::J2_loop1r; |
1158 | ENDLOOP = Hexagon::ENDLOOP1; |
1159 | IsInnerHWLoop = 0; |
1160 | } else { |
1161 | LOOP_i = Hexagon::J2_loop0i; |
1162 | LOOP_r = Hexagon::J2_loop0r; |
1163 | ENDLOOP = Hexagon::ENDLOOP0; |
1164 | } |
1165 | |
1166 | #ifndef NDEBUG |
1167 | // Stop trying after reaching the limit (if any). |
1168 | int Limit = HWLoopLimit; |
1169 | if (Limit >= 0) { |
1170 | if (Counter >= HWLoopLimit) |
1171 | return false; |
1172 | Counter++; |
1173 | } |
1174 | #endif |
1175 | |
1176 | // Does the loop contain any invalid instructions? |
1177 | if (containsInvalidInstruction(L, IsInnerHWLoop)) |
1178 | return false; |
1179 | |
1180 | MachineBasicBlock *LastMBB = L->findLoopControlBlock(); |
1181 | // Don't generate hw loop if the loop has more than one exit. |
1182 | if (!LastMBB) |
1183 | return false; |
1184 | |
1185 | MachineBasicBlock::iterator LastI = LastMBB->getFirstTerminator(); |
1186 | if (LastI == LastMBB->end()) |
1187 | return false; |
1188 | |
1189 | // Is the induction variable bump feeding the latch condition? |
1190 | if (!fixupInductionVariable(L)) |
1191 | return false; |
1192 | |
1193 | // Ensure the loop has a preheader: the loop instruction will be |
1194 | // placed there. |
1195 | MachineBasicBlock * = MLI->findLoopPreheader(L, SpeculativePreheader: SpecPreheader); |
1196 | if (!Preheader) { |
1197 | Preheader = createPreheaderForLoop(L); |
1198 | if (!Preheader) |
1199 | return false; |
1200 | } |
1201 | |
1202 | MachineBasicBlock::iterator InsertPos = Preheader->getFirstTerminator(); |
1203 | |
1204 | SmallVector<MachineInstr*, 2> OldInsts; |
1205 | // Are we able to determine the trip count for the loop? |
1206 | CountValue *TripCount = getLoopTripCount(L, OldInsts); |
1207 | if (!TripCount) |
1208 | return false; |
1209 | |
1210 | // Is the trip count available in the preheader? |
1211 | if (TripCount->isReg()) { |
1212 | // There will be a use of the register inserted into the preheader, |
1213 | // so make sure that the register is actually defined at that point. |
1214 | MachineInstr *TCDef = MRI->getVRegDef(Reg: TripCount->getReg()); |
1215 | MachineBasicBlock *BBDef = TCDef->getParent(); |
1216 | if (!MDT->dominates(A: BBDef, B: Preheader)) |
1217 | return false; |
1218 | } |
1219 | |
1220 | // Determine the loop start. |
1221 | MachineBasicBlock *TopBlock = L->getTopBlock(); |
1222 | MachineBasicBlock *ExitingBlock = L->findLoopControlBlock(); |
1223 | MachineBasicBlock *LoopStart = nullptr; |
1224 | if (ExitingBlock != L->getLoopLatch()) { |
1225 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
1226 | SmallVector<MachineOperand, 2> Cond; |
1227 | |
1228 | if (TII->analyzeBranch(MBB&: *ExitingBlock, TBB&: TB, FBB&: FB, Cond, AllowModify: false)) |
1229 | return false; |
1230 | |
1231 | if (L->contains(BB: TB)) |
1232 | LoopStart = TB; |
1233 | else if (L->contains(BB: FB)) |
1234 | LoopStart = FB; |
1235 | else |
1236 | return false; |
1237 | } |
1238 | else |
1239 | LoopStart = TopBlock; |
1240 | |
1241 | // Convert the loop to a hardware loop. |
1242 | LLVM_DEBUG(dbgs() << "Change to hardware loop at " ; L->dump()); |
1243 | DebugLoc DL; |
1244 | if (InsertPos != Preheader->end()) |
1245 | DL = InsertPos->getDebugLoc(); |
1246 | |
1247 | if (TripCount->isReg()) { |
1248 | // Create a copy of the loop count register. |
1249 | Register CountReg = MRI->createVirtualRegister(RegClass: &Hexagon::IntRegsRegClass); |
1250 | BuildMI(BB&: *Preheader, I: InsertPos, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: CountReg) |
1251 | .addReg(RegNo: TripCount->getReg(), flags: 0, SubReg: TripCount->getSubReg()); |
1252 | // Add the Loop instruction to the beginning of the loop. |
1253 | BuildMI(BB&: *Preheader, I: InsertPos, MIMD: DL, MCID: TII->get(Opcode: LOOP_r)).addMBB(MBB: LoopStart) |
1254 | .addReg(RegNo: CountReg); |
1255 | } else { |
1256 | assert(TripCount->isImm() && "Expecting immediate value for trip count" ); |
1257 | // Add the Loop immediate instruction to the beginning of the loop, |
1258 | // if the immediate fits in the instructions. Otherwise, we need to |
1259 | // create a new virtual register. |
1260 | int64_t CountImm = TripCount->getImm(); |
1261 | if (!TII->isValidOffset(Opcode: LOOP_i, Offset: CountImm, TRI)) { |
1262 | Register CountReg = MRI->createVirtualRegister(RegClass: &Hexagon::IntRegsRegClass); |
1263 | BuildMI(BB&: *Preheader, I: InsertPos, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_tfrsi), DestReg: CountReg) |
1264 | .addImm(Val: CountImm); |
1265 | BuildMI(BB&: *Preheader, I: InsertPos, MIMD: DL, MCID: TII->get(Opcode: LOOP_r)) |
1266 | .addMBB(MBB: LoopStart).addReg(RegNo: CountReg); |
1267 | } else |
1268 | BuildMI(BB&: *Preheader, I: InsertPos, MIMD: DL, MCID: TII->get(Opcode: LOOP_i)) |
1269 | .addMBB(MBB: LoopStart).addImm(Val: CountImm); |
1270 | } |
1271 | |
1272 | // Make sure the loop start always has a reference in the CFG. |
1273 | LoopStart->setMachineBlockAddressTaken(); |
1274 | |
1275 | // Replace the loop branch with an endloop instruction. |
1276 | DebugLoc LastIDL = LastI->getDebugLoc(); |
1277 | BuildMI(BB&: *LastMBB, I: LastI, MIMD: LastIDL, MCID: TII->get(Opcode: ENDLOOP)).addMBB(MBB: LoopStart); |
1278 | |
1279 | // The loop ends with either: |
1280 | // - a conditional branch followed by an unconditional branch, or |
1281 | // - a conditional branch to the loop start. |
1282 | if (LastI->getOpcode() == Hexagon::J2_jumpt || |
1283 | LastI->getOpcode() == Hexagon::J2_jumpf) { |
1284 | // Delete one and change/add an uncond. branch to out of the loop. |
1285 | MachineBasicBlock *BranchTarget = LastI->getOperand(i: 1).getMBB(); |
1286 | LastI = LastMBB->erase(I: LastI); |
1287 | if (!L->contains(BB: BranchTarget)) { |
1288 | if (LastI != LastMBB->end()) |
1289 | LastI = LastMBB->erase(I: LastI); |
1290 | SmallVector<MachineOperand, 0> Cond; |
1291 | TII->insertBranch(MBB&: *LastMBB, TBB: BranchTarget, FBB: nullptr, Cond, DL: LastIDL); |
1292 | } |
1293 | } else { |
1294 | // Conditional branch to loop start; just delete it. |
1295 | LastMBB->erase(I: LastI); |
1296 | } |
1297 | delete TripCount; |
1298 | |
1299 | // The induction operation and the comparison may now be |
1300 | // unneeded. If these are unneeded, then remove them. |
1301 | for (unsigned i = 0; i < OldInsts.size(); ++i) |
1302 | removeIfDead(MI: OldInsts[i]); |
1303 | |
1304 | ++NumHWLoops; |
1305 | |
1306 | // Set RecL1used and RecL0used only after hardware loop has been |
1307 | // successfully generated. Doing it earlier can cause wrong loop instruction |
1308 | // to be used. |
1309 | if (L0Used) // Loop0 was already used. So, the correct loop must be loop1. |
1310 | RecL1used = true; |
1311 | else |
1312 | RecL0used = true; |
1313 | |
1314 | return true; |
1315 | } |
1316 | |
1317 | bool HexagonHardwareLoops::orderBumpCompare(MachineInstr *BumpI, |
1318 | MachineInstr *CmpI) { |
1319 | assert (BumpI != CmpI && "Bump and compare in the same instruction?" ); |
1320 | |
1321 | MachineBasicBlock *BB = BumpI->getParent(); |
1322 | if (CmpI->getParent() != BB) |
1323 | return false; |
1324 | |
1325 | using instr_iterator = MachineBasicBlock::instr_iterator; |
1326 | |
1327 | // Check if things are in order to begin with. |
1328 | for (instr_iterator I(BumpI), E = BB->instr_end(); I != E; ++I) |
1329 | if (&*I == CmpI) |
1330 | return true; |
1331 | |
1332 | // Out of order. |
1333 | Register PredR = CmpI->getOperand(i: 0).getReg(); |
1334 | bool FoundBump = false; |
1335 | instr_iterator CmpIt = CmpI->getIterator(), NextIt = std::next(x: CmpIt); |
1336 | for (instr_iterator I = NextIt, E = BB->instr_end(); I != E; ++I) { |
1337 | MachineInstr *In = &*I; |
1338 | for (unsigned i = 0, n = In->getNumOperands(); i < n; ++i) { |
1339 | MachineOperand &MO = In->getOperand(i); |
1340 | if (MO.isReg() && MO.isUse()) { |
1341 | if (MO.getReg() == PredR) // Found an intervening use of PredR. |
1342 | return false; |
1343 | } |
1344 | } |
1345 | |
1346 | if (In == BumpI) { |
1347 | BB->splice(Where: ++BumpI->getIterator(), Other: BB, From: CmpI->getIterator()); |
1348 | FoundBump = true; |
1349 | break; |
1350 | } |
1351 | } |
1352 | assert (FoundBump && "Cannot determine instruction order" ); |
1353 | return FoundBump; |
1354 | } |
1355 | |
1356 | /// This function is required to break recursion. Visiting phis in a loop may |
1357 | /// result in recursion during compilation. We break the recursion by making |
1358 | /// sure that we visit a MachineOperand and its definition in a |
1359 | /// MachineInstruction only once. If we attempt to visit more than once, then |
1360 | /// there is recursion, and will return false. |
1361 | bool HexagonHardwareLoops::isLoopFeeder(MachineLoop *L, MachineBasicBlock *A, |
1362 | MachineInstr *MI, |
1363 | const MachineOperand *MO, |
1364 | LoopFeederMap &LoopFeederPhi) const { |
1365 | if (LoopFeederPhi.find(x: MO->getReg()) == LoopFeederPhi.end()) { |
1366 | LLVM_DEBUG(dbgs() << "\nhw_loop head, " |
1367 | << printMBBReference(**L->block_begin())); |
1368 | // Ignore all BBs that form Loop. |
1369 | if (llvm::is_contained(Range: L->getBlocks(), Element: A)) |
1370 | return false; |
1371 | MachineInstr *Def = MRI->getVRegDef(Reg: MO->getReg()); |
1372 | LoopFeederPhi.insert(x: std::make_pair(x: MO->getReg(), y&: Def)); |
1373 | return true; |
1374 | } else |
1375 | // Already visited node. |
1376 | return false; |
1377 | } |
1378 | |
1379 | /// Return true if a Phi may generate a value that can underflow. |
1380 | /// This function calls loopCountMayWrapOrUnderFlow for each Phi operand. |
1381 | bool HexagonHardwareLoops::phiMayWrapOrUnderflow( |
1382 | MachineInstr *Phi, const MachineOperand *EndVal, MachineBasicBlock *MBB, |
1383 | MachineLoop *L, LoopFeederMap &LoopFeederPhi) const { |
1384 | assert(Phi->isPHI() && "Expecting a Phi." ); |
1385 | // Walk through each Phi, and its used operands. Make sure that |
1386 | // if there is recursion in Phi, we won't generate hardware loops. |
1387 | for (int i = 1, n = Phi->getNumOperands(); i < n; i += 2) |
1388 | if (isLoopFeeder(L, A: MBB, MI: Phi, MO: &(Phi->getOperand(i)), LoopFeederPhi)) |
1389 | if (loopCountMayWrapOrUnderFlow(InitVal: &(Phi->getOperand(i)), EndVal, |
1390 | MBB: Phi->getParent(), L, LoopFeederPhi)) |
1391 | return true; |
1392 | return false; |
1393 | } |
1394 | |
1395 | /// Return true if the induction variable can underflow in the first iteration. |
1396 | /// An example, is an initial unsigned value that is 0 and is decrement in the |
1397 | /// first itertion of a do-while loop. In this case, we cannot generate a |
1398 | /// hardware loop because the endloop instruction does not decrement the loop |
1399 | /// counter if it is <= 1. We only need to perform this analysis if the |
1400 | /// initial value is a register. |
1401 | /// |
1402 | /// This function assumes the initial value may underfow unless proven |
1403 | /// otherwise. If the type is signed, then we don't care because signed |
1404 | /// underflow is undefined. We attempt to prove the initial value is not |
1405 | /// zero by perfoming a crude analysis of the loop counter. This function |
1406 | /// checks if the initial value is used in any comparison prior to the loop |
1407 | /// and, if so, assumes the comparison is a range check. This is inexact, |
1408 | /// but will catch the simple cases. |
1409 | bool HexagonHardwareLoops::loopCountMayWrapOrUnderFlow( |
1410 | const MachineOperand *InitVal, const MachineOperand *EndVal, |
1411 | MachineBasicBlock *MBB, MachineLoop *L, |
1412 | LoopFeederMap &LoopFeederPhi) const { |
1413 | // Only check register values since they are unknown. |
1414 | if (!InitVal->isReg()) |
1415 | return false; |
1416 | |
1417 | if (!EndVal->isImm()) |
1418 | return false; |
1419 | |
1420 | // A register value that is assigned an immediate is a known value, and it |
1421 | // won't underflow in the first iteration. |
1422 | int64_t Imm; |
1423 | if (checkForImmediate(MO: *InitVal, Val&: Imm)) |
1424 | return (EndVal->getImm() == Imm); |
1425 | |
1426 | Register Reg = InitVal->getReg(); |
1427 | |
1428 | // We don't know the value of a physical register. |
1429 | if (!Reg.isVirtual()) |
1430 | return true; |
1431 | |
1432 | MachineInstr *Def = MRI->getVRegDef(Reg); |
1433 | if (!Def) |
1434 | return true; |
1435 | |
1436 | // If the initial value is a Phi or copy and the operands may not underflow, |
1437 | // then the definition cannot be underflow either. |
1438 | if (Def->isPHI() && !phiMayWrapOrUnderflow(Phi: Def, EndVal, MBB: Def->getParent(), |
1439 | L, LoopFeederPhi)) |
1440 | return false; |
1441 | if (Def->isCopy() && !loopCountMayWrapOrUnderFlow(InitVal: &(Def->getOperand(i: 1)), |
1442 | EndVal, MBB: Def->getParent(), |
1443 | L, LoopFeederPhi)) |
1444 | return false; |
1445 | |
1446 | // Iterate over the uses of the initial value. If the initial value is used |
1447 | // in a compare, then we assume this is a range check that ensures the loop |
1448 | // doesn't underflow. This is not an exact test and should be improved. |
1449 | for (MachineRegisterInfo::use_instr_nodbg_iterator I = MRI->use_instr_nodbg_begin(RegNo: Reg), |
1450 | E = MRI->use_instr_nodbg_end(); I != E; ++I) { |
1451 | MachineInstr *MI = &*I; |
1452 | Register CmpReg1, CmpReg2; |
1453 | int64_t CmpMask = 0, CmpValue = 0; |
1454 | |
1455 | if (!TII->analyzeCompare(MI: *MI, SrcReg&: CmpReg1, SrcReg2&: CmpReg2, Mask&: CmpMask, Value&: CmpValue)) |
1456 | continue; |
1457 | |
1458 | MachineBasicBlock *TBB = nullptr, *FBB = nullptr; |
1459 | SmallVector<MachineOperand, 2> Cond; |
1460 | if (TII->analyzeBranch(MBB&: *MI->getParent(), TBB, FBB, Cond, AllowModify: false)) |
1461 | continue; |
1462 | |
1463 | Comparison::Kind Cmp = |
1464 | getComparisonKind(CondOpc: MI->getOpcode(), InitialValue: nullptr, EndValue: nullptr, IVBump: 0); |
1465 | if (Cmp == 0) |
1466 | continue; |
1467 | if (TII->predOpcodeHasNot(Cond) ^ (TBB != MBB)) |
1468 | Cmp = Comparison::getNegatedComparison(Cmp); |
1469 | if (CmpReg2 != 0 && CmpReg2 == Reg) |
1470 | Cmp = Comparison::getSwappedComparison(Cmp); |
1471 | |
1472 | // Signed underflow is undefined. |
1473 | if (Comparison::isSigned(Cmp)) |
1474 | return false; |
1475 | |
1476 | // Check if there is a comparison of the initial value. If the initial value |
1477 | // is greater than or not equal to another value, then assume this is a |
1478 | // range check. |
1479 | if ((Cmp & Comparison::G) || Cmp == Comparison::NE) |
1480 | return false; |
1481 | } |
1482 | |
1483 | // OK - this is a hack that needs to be improved. We really need to analyze |
1484 | // the instructions performed on the initial value. This works on the simplest |
1485 | // cases only. |
1486 | if (!Def->isCopy() && !Def->isPHI()) |
1487 | return false; |
1488 | |
1489 | return true; |
1490 | } |
1491 | |
1492 | bool HexagonHardwareLoops::checkForImmediate(const MachineOperand &MO, |
1493 | int64_t &Val) const { |
1494 | if (MO.isImm()) { |
1495 | Val = MO.getImm(); |
1496 | return true; |
1497 | } |
1498 | if (!MO.isReg()) |
1499 | return false; |
1500 | |
1501 | // MO is a register. Check whether it is defined as an immediate value, |
1502 | // and if so, get the value of it in TV. That value will then need to be |
1503 | // processed to handle potential subregisters in MO. |
1504 | int64_t TV; |
1505 | |
1506 | Register R = MO.getReg(); |
1507 | if (!R.isVirtual()) |
1508 | return false; |
1509 | MachineInstr *DI = MRI->getVRegDef(Reg: R); |
1510 | unsigned DOpc = DI->getOpcode(); |
1511 | switch (DOpc) { |
1512 | case TargetOpcode::COPY: |
1513 | case Hexagon::A2_tfrsi: |
1514 | case Hexagon::A2_tfrpi: |
1515 | case Hexagon::CONST32: |
1516 | case Hexagon::CONST64: |
1517 | // Call recursively to avoid an extra check whether operand(1) is |
1518 | // indeed an immediate (it could be a global address, for example), |
1519 | // plus we can handle COPY at the same time. |
1520 | if (!checkForImmediate(MO: DI->getOperand(i: 1), Val&: TV)) |
1521 | return false; |
1522 | break; |
1523 | case Hexagon::A2_combineii: |
1524 | case Hexagon::A4_combineir: |
1525 | case Hexagon::A4_combineii: |
1526 | case Hexagon::A4_combineri: |
1527 | case Hexagon::A2_combinew: { |
1528 | const MachineOperand &S1 = DI->getOperand(i: 1); |
1529 | const MachineOperand &S2 = DI->getOperand(i: 2); |
1530 | int64_t V1, V2; |
1531 | if (!checkForImmediate(MO: S1, Val&: V1) || !checkForImmediate(MO: S2, Val&: V2)) |
1532 | return false; |
1533 | TV = V2 | (static_cast<uint64_t>(V1) << 32); |
1534 | break; |
1535 | } |
1536 | case TargetOpcode::REG_SEQUENCE: { |
1537 | const MachineOperand &S1 = DI->getOperand(i: 1); |
1538 | const MachineOperand &S3 = DI->getOperand(i: 3); |
1539 | int64_t V1, V3; |
1540 | if (!checkForImmediate(MO: S1, Val&: V1) || !checkForImmediate(MO: S3, Val&: V3)) |
1541 | return false; |
1542 | unsigned Sub2 = DI->getOperand(i: 2).getImm(); |
1543 | unsigned Sub4 = DI->getOperand(i: 4).getImm(); |
1544 | if (Sub2 == Hexagon::isub_lo && Sub4 == Hexagon::isub_hi) |
1545 | TV = V1 | (V3 << 32); |
1546 | else if (Sub2 == Hexagon::isub_hi && Sub4 == Hexagon::isub_lo) |
1547 | TV = V3 | (V1 << 32); |
1548 | else |
1549 | llvm_unreachable("Unexpected form of REG_SEQUENCE" ); |
1550 | break; |
1551 | } |
1552 | |
1553 | default: |
1554 | return false; |
1555 | } |
1556 | |
1557 | // By now, we should have successfully obtained the immediate value defining |
1558 | // the register referenced in MO. Handle a potential use of a subregister. |
1559 | switch (MO.getSubReg()) { |
1560 | case Hexagon::isub_lo: |
1561 | Val = TV & 0xFFFFFFFFULL; |
1562 | break; |
1563 | case Hexagon::isub_hi: |
1564 | Val = (TV >> 32) & 0xFFFFFFFFULL; |
1565 | break; |
1566 | default: |
1567 | Val = TV; |
1568 | break; |
1569 | } |
1570 | return true; |
1571 | } |
1572 | |
1573 | void HexagonHardwareLoops::setImmediate(MachineOperand &MO, int64_t Val) { |
1574 | if (MO.isImm()) { |
1575 | MO.setImm(Val); |
1576 | return; |
1577 | } |
1578 | |
1579 | assert(MO.isReg()); |
1580 | Register R = MO.getReg(); |
1581 | MachineInstr *DI = MRI->getVRegDef(Reg: R); |
1582 | |
1583 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: R); |
1584 | Register NewR = MRI->createVirtualRegister(RegClass: RC); |
1585 | MachineBasicBlock &B = *DI->getParent(); |
1586 | DebugLoc DL = DI->getDebugLoc(); |
1587 | BuildMI(BB&: B, I: DI, MIMD: DL, MCID: TII->get(Opcode: DI->getOpcode()), DestReg: NewR).addImm(Val); |
1588 | MO.setReg(NewR); |
1589 | } |
1590 | |
1591 | bool HexagonHardwareLoops::fixupInductionVariable(MachineLoop *L) { |
1592 | MachineBasicBlock * = L->getHeader(); |
1593 | MachineBasicBlock *Latch = L->getLoopLatch(); |
1594 | MachineBasicBlock *ExitingBlock = L->findLoopControlBlock(); |
1595 | |
1596 | if (!(Header && Latch && ExitingBlock)) |
1597 | return false; |
1598 | |
1599 | // These data structures follow the same concept as the corresponding |
1600 | // ones in findInductionRegister (where some comments are). |
1601 | using RegisterBump = std::pair<Register, int64_t>; |
1602 | using RegisterInduction = std::pair<Register, RegisterBump>; |
1603 | using RegisterInductionSet = std::set<RegisterInduction>; |
1604 | |
1605 | // Register candidates for induction variables, with their associated bumps. |
1606 | RegisterInductionSet IndRegs; |
1607 | |
1608 | // Look for induction patterns: |
1609 | // %1 = PHI ..., [ latch, %2 ] |
1610 | // %2 = ADD %1, imm |
1611 | using instr_iterator = MachineBasicBlock::instr_iterator; |
1612 | |
1613 | for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); |
1614 | I != E && I->isPHI(); ++I) { |
1615 | MachineInstr *Phi = &*I; |
1616 | |
1617 | // Have a PHI instruction. |
1618 | for (unsigned i = 1, n = Phi->getNumOperands(); i < n; i += 2) { |
1619 | if (Phi->getOperand(i: i+1).getMBB() != Latch) |
1620 | continue; |
1621 | |
1622 | Register PhiReg = Phi->getOperand(i).getReg(); |
1623 | MachineInstr *DI = MRI->getVRegDef(Reg: PhiReg); |
1624 | |
1625 | if (DI->getDesc().isAdd()) { |
1626 | // If the register operand to the add/sub is the PHI we are looking |
1627 | // at, this meets the induction pattern. |
1628 | Register IndReg = DI->getOperand(i: 1).getReg(); |
1629 | MachineOperand &Opnd2 = DI->getOperand(i: 2); |
1630 | int64_t V; |
1631 | if (MRI->getVRegDef(Reg: IndReg) == Phi && checkForImmediate(MO: Opnd2, Val&: V)) { |
1632 | Register UpdReg = DI->getOperand(i: 0).getReg(); |
1633 | IndRegs.insert(x: std::make_pair(x&: UpdReg, y: std::make_pair(x&: IndReg, y&: V))); |
1634 | } |
1635 | } |
1636 | } // for (i) |
1637 | } // for (instr) |
1638 | |
1639 | if (IndRegs.empty()) |
1640 | return false; |
1641 | |
1642 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
1643 | SmallVector<MachineOperand,2> Cond; |
1644 | // analyzeBranch returns true if it fails to analyze branch. |
1645 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *ExitingBlock, TBB&: TB, FBB&: FB, Cond, AllowModify: false); |
1646 | if (NotAnalyzed || Cond.empty()) |
1647 | return false; |
1648 | |
1649 | if (ExitingBlock != Latch && (TB == Latch || FB == Latch)) { |
1650 | MachineBasicBlock *LTB = nullptr, *LFB = nullptr; |
1651 | SmallVector<MachineOperand,2> LCond; |
1652 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *Latch, TBB&: LTB, FBB&: LFB, Cond&: LCond, AllowModify: false); |
1653 | if (NotAnalyzed) |
1654 | return false; |
1655 | |
1656 | // Since latch is not the exiting block, the latch branch should be an |
1657 | // unconditional branch to the loop header. |
1658 | if (TB == Latch) |
1659 | TB = (LTB == Header) ? LTB : LFB; |
1660 | else |
1661 | FB = (LTB == Header) ? LTB : LFB; |
1662 | } |
1663 | if (TB != Header) { |
1664 | if (FB != Header) { |
1665 | // The latch/exit block does not go back to the header. |
1666 | return false; |
1667 | } |
1668 | // FB is the header (i.e., uncond. jump to branch header) |
1669 | // In this case, the LoopBody -> TB should not be a back edge otherwise |
1670 | // it could result in an infinite loop after conversion to hw_loop. |
1671 | // This case can happen when the Latch has two jumps like this: |
1672 | // Jmp_c OuterLoopHeader <-- TB |
1673 | // Jmp InnerLoopHeader <-- FB |
1674 | if (MDT->dominates(A: TB, B: FB)) |
1675 | return false; |
1676 | } |
1677 | |
1678 | // Expecting a predicate register as a condition. It won't be a hardware |
1679 | // predicate register at this point yet, just a vreg. |
1680 | // HexagonInstrInfo::analyzeBranch for negated branches inserts imm(0) |
1681 | // into Cond, followed by the predicate register. For non-negated branches |
1682 | // it's just the register. |
1683 | unsigned CSz = Cond.size(); |
1684 | if (CSz != 1 && CSz != 2) |
1685 | return false; |
1686 | |
1687 | if (!Cond[CSz-1].isReg()) |
1688 | return false; |
1689 | |
1690 | Register P = Cond[CSz - 1].getReg(); |
1691 | MachineInstr *PredDef = MRI->getVRegDef(Reg: P); |
1692 | |
1693 | if (!PredDef->isCompare()) |
1694 | return false; |
1695 | |
1696 | SmallSet<Register,2> CmpRegs; |
1697 | MachineOperand *CmpImmOp = nullptr; |
1698 | |
1699 | // Go over all operands to the compare and look for immediate and register |
1700 | // operands. Assume that if the compare has a single register use and a |
1701 | // single immediate operand, then the register is being compared with the |
1702 | // immediate value. |
1703 | for (MachineOperand &MO : PredDef->operands()) { |
1704 | if (MO.isReg()) { |
1705 | // Skip all implicit references. In one case there was: |
1706 | // %140 = FCMPUGT32_rr %138, %139, implicit %usr |
1707 | if (MO.isImplicit()) |
1708 | continue; |
1709 | if (MO.isUse()) { |
1710 | if (!isImmediate(MO)) { |
1711 | CmpRegs.insert(V: MO.getReg()); |
1712 | continue; |
1713 | } |
1714 | // Consider the register to be the "immediate" operand. |
1715 | if (CmpImmOp) |
1716 | return false; |
1717 | CmpImmOp = &MO; |
1718 | } |
1719 | } else if (MO.isImm()) { |
1720 | if (CmpImmOp) // A second immediate argument? Confusing. Bail out. |
1721 | return false; |
1722 | CmpImmOp = &MO; |
1723 | } |
1724 | } |
1725 | |
1726 | if (CmpRegs.empty()) |
1727 | return false; |
1728 | |
1729 | // Check if the compared register follows the order we want. Fix if needed. |
1730 | for (RegisterInductionSet::iterator I = IndRegs.begin(), E = IndRegs.end(); |
1731 | I != E; ++I) { |
1732 | // This is a success. If the register used in the comparison is one that |
1733 | // we have identified as a bumped (updated) induction register, there is |
1734 | // nothing to do. |
1735 | if (CmpRegs.count(V: I->first)) |
1736 | return true; |
1737 | |
1738 | // Otherwise, if the register being compared comes out of a PHI node, |
1739 | // and has been recognized as following the induction pattern, and is |
1740 | // compared against an immediate, we can fix it. |
1741 | const RegisterBump &RB = I->second; |
1742 | if (CmpRegs.count(V: RB.first)) { |
1743 | if (!CmpImmOp) { |
1744 | // If both operands to the compare instruction are registers, see if |
1745 | // it can be changed to use induction register as one of the operands. |
1746 | MachineInstr *IndI = nullptr; |
1747 | MachineInstr *nonIndI = nullptr; |
1748 | MachineOperand *IndMO = nullptr; |
1749 | MachineOperand *nonIndMO = nullptr; |
1750 | |
1751 | for (unsigned i = 1, n = PredDef->getNumOperands(); i < n; ++i) { |
1752 | MachineOperand &MO = PredDef->getOperand(i); |
1753 | if (MO.isReg() && MO.getReg() == RB.first) { |
1754 | LLVM_DEBUG(dbgs() << "\n DefMI(" << i |
1755 | << ") = " << *(MRI->getVRegDef(I->first))); |
1756 | if (IndI) |
1757 | return false; |
1758 | |
1759 | IndI = MRI->getVRegDef(Reg: I->first); |
1760 | IndMO = &MO; |
1761 | } else if (MO.isReg()) { |
1762 | LLVM_DEBUG(dbgs() << "\n DefMI(" << i |
1763 | << ") = " << *(MRI->getVRegDef(MO.getReg()))); |
1764 | if (nonIndI) |
1765 | return false; |
1766 | |
1767 | nonIndI = MRI->getVRegDef(Reg: MO.getReg()); |
1768 | nonIndMO = &MO; |
1769 | } |
1770 | } |
1771 | if (IndI && nonIndI && |
1772 | nonIndI->getOpcode() == Hexagon::A2_addi && |
1773 | nonIndI->getOperand(i: 2).isImm() && |
1774 | nonIndI->getOperand(i: 2).getImm() == - RB.second) { |
1775 | bool Order = orderBumpCompare(BumpI: IndI, CmpI: PredDef); |
1776 | if (Order) { |
1777 | IndMO->setReg(I->first); |
1778 | nonIndMO->setReg(nonIndI->getOperand(i: 1).getReg()); |
1779 | return true; |
1780 | } |
1781 | } |
1782 | return false; |
1783 | } |
1784 | |
1785 | // It is not valid to do this transformation on an unsigned comparison |
1786 | // because it may underflow. |
1787 | Comparison::Kind Cmp = |
1788 | getComparisonKind(CondOpc: PredDef->getOpcode(), InitialValue: nullptr, EndValue: nullptr, IVBump: 0); |
1789 | if (!Cmp || Comparison::isUnsigned(Cmp)) |
1790 | return false; |
1791 | |
1792 | // If the register is being compared against an immediate, try changing |
1793 | // the compare instruction to use induction register and adjust the |
1794 | // immediate operand. |
1795 | int64_t CmpImm = getImmediate(MO: *CmpImmOp); |
1796 | int64_t V = RB.second; |
1797 | // Handle Overflow (64-bit). |
1798 | if (((V > 0) && (CmpImm > INT64_MAX - V)) || |
1799 | ((V < 0) && (CmpImm < INT64_MIN - V))) |
1800 | return false; |
1801 | CmpImm += V; |
1802 | // Most comparisons of register against an immediate value allow |
1803 | // the immediate to be constant-extended. There are some exceptions |
1804 | // though. Make sure the new combination will work. |
1805 | if (CmpImmOp->isImm() && !TII->isExtendable(MI: *PredDef) && |
1806 | !TII->isValidOffset(Opcode: PredDef->getOpcode(), Offset: CmpImm, TRI, Extend: false)) |
1807 | return false; |
1808 | |
1809 | // Make sure that the compare happens after the bump. Otherwise, |
1810 | // after the fixup, the compare would use a yet-undefined register. |
1811 | MachineInstr *BumpI = MRI->getVRegDef(Reg: I->first); |
1812 | bool Order = orderBumpCompare(BumpI, CmpI: PredDef); |
1813 | if (!Order) |
1814 | return false; |
1815 | |
1816 | // Finally, fix the compare instruction. |
1817 | setImmediate(MO&: *CmpImmOp, Val: CmpImm); |
1818 | for (MachineOperand &MO : PredDef->operands()) { |
1819 | if (MO.isReg() && MO.getReg() == RB.first) { |
1820 | MO.setReg(I->first); |
1821 | return true; |
1822 | } |
1823 | } |
1824 | } |
1825 | } |
1826 | |
1827 | return false; |
1828 | } |
1829 | |
1830 | /// createPreheaderForLoop - Create a preheader for a given loop. |
1831 | MachineBasicBlock *HexagonHardwareLoops::( |
1832 | MachineLoop *L) { |
1833 | if (MachineBasicBlock *TmpPH = MLI->findLoopPreheader(L, SpeculativePreheader: SpecPreheader)) |
1834 | return TmpPH; |
1835 | if (!HWCreatePreheader) |
1836 | return nullptr; |
1837 | |
1838 | MachineBasicBlock * = L->getHeader(); |
1839 | MachineBasicBlock *Latch = L->getLoopLatch(); |
1840 | MachineBasicBlock *ExitingBlock = L->findLoopControlBlock(); |
1841 | MachineFunction *MF = Header->getParent(); |
1842 | DebugLoc DL; |
1843 | |
1844 | #ifndef NDEBUG |
1845 | if ((!PHFn.empty()) && (PHFn != MF->getName())) |
1846 | return nullptr; |
1847 | #endif |
1848 | |
1849 | if (!Latch || !ExitingBlock || Header->hasAddressTaken()) |
1850 | return nullptr; |
1851 | |
1852 | using instr_iterator = MachineBasicBlock::instr_iterator; |
1853 | |
1854 | // Verify that all existing predecessors have analyzable branches |
1855 | // (or no branches at all). |
1856 | using MBBVector = std::vector<MachineBasicBlock *>; |
1857 | |
1858 | MBBVector Preds(Header->pred_begin(), Header->pred_end()); |
1859 | SmallVector<MachineOperand,2> Tmp1; |
1860 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
1861 | |
1862 | if (TII->analyzeBranch(MBB&: *ExitingBlock, TBB&: TB, FBB&: FB, Cond&: Tmp1, AllowModify: false)) |
1863 | return nullptr; |
1864 | |
1865 | for (MachineBasicBlock *PB : Preds) { |
1866 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *PB, TBB&: TB, FBB&: FB, Cond&: Tmp1, AllowModify: false); |
1867 | if (NotAnalyzed) |
1868 | return nullptr; |
1869 | } |
1870 | |
1871 | MachineBasicBlock *NewPH = MF->CreateMachineBasicBlock(); |
1872 | MF->insert(MBBI: Header->getIterator(), MBB: NewPH); |
1873 | |
1874 | if (Header->pred_size() > 2) { |
1875 | // Ensure that the header has only two predecessors: the preheader and |
1876 | // the loop latch. Any additional predecessors of the header should |
1877 | // join at the newly created preheader. Inspect all PHI nodes from the |
1878 | // header and create appropriate corresponding PHI nodes in the preheader. |
1879 | |
1880 | for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); |
1881 | I != E && I->isPHI(); ++I) { |
1882 | MachineInstr *PN = &*I; |
1883 | |
1884 | const MCInstrDesc &PD = TII->get(Opcode: TargetOpcode::PHI); |
1885 | MachineInstr *NewPN = MF->CreateMachineInstr(MCID: PD, DL); |
1886 | NewPH->insert(I: NewPH->end(), MI: NewPN); |
1887 | |
1888 | Register PR = PN->getOperand(i: 0).getReg(); |
1889 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: PR); |
1890 | Register NewPR = MRI->createVirtualRegister(RegClass: RC); |
1891 | NewPN->addOperand(Op: MachineOperand::CreateReg(Reg: NewPR, isDef: true)); |
1892 | |
1893 | // Copy all non-latch operands of a header's PHI node to the newly |
1894 | // created PHI node in the preheader. |
1895 | for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) { |
1896 | Register PredR = PN->getOperand(i).getReg(); |
1897 | unsigned PredRSub = PN->getOperand(i).getSubReg(); |
1898 | MachineBasicBlock *PredB = PN->getOperand(i: i+1).getMBB(); |
1899 | if (PredB == Latch) |
1900 | continue; |
1901 | |
1902 | MachineOperand MO = MachineOperand::CreateReg(Reg: PredR, isDef: false); |
1903 | MO.setSubReg(PredRSub); |
1904 | NewPN->addOperand(Op: MO); |
1905 | NewPN->addOperand(Op: MachineOperand::CreateMBB(MBB: PredB)); |
1906 | } |
1907 | |
1908 | // Remove copied operands from the old PHI node and add the value |
1909 | // coming from the preheader's PHI. |
1910 | for (int i = PN->getNumOperands()-2; i > 0; i -= 2) { |
1911 | MachineBasicBlock *PredB = PN->getOperand(i: i+1).getMBB(); |
1912 | if (PredB != Latch) { |
1913 | PN->removeOperand(OpNo: i+1); |
1914 | PN->removeOperand(OpNo: i); |
1915 | } |
1916 | } |
1917 | PN->addOperand(Op: MachineOperand::CreateReg(Reg: NewPR, isDef: false)); |
1918 | PN->addOperand(Op: MachineOperand::CreateMBB(MBB: NewPH)); |
1919 | } |
1920 | } else { |
1921 | assert(Header->pred_size() == 2); |
1922 | |
1923 | // The header has only two predecessors, but the non-latch predecessor |
1924 | // is not a preheader (e.g. it has other successors, etc.) |
1925 | // In such a case we don't need any extra PHI nodes in the new preheader, |
1926 | // all we need is to adjust existing PHIs in the header to now refer to |
1927 | // the new preheader. |
1928 | for (instr_iterator I = Header->instr_begin(), E = Header->instr_end(); |
1929 | I != E && I->isPHI(); ++I) { |
1930 | MachineInstr *PN = &*I; |
1931 | for (unsigned i = 1, n = PN->getNumOperands(); i < n; i += 2) { |
1932 | MachineOperand &MO = PN->getOperand(i: i+1); |
1933 | if (MO.getMBB() != Latch) |
1934 | MO.setMBB(NewPH); |
1935 | } |
1936 | } |
1937 | } |
1938 | |
1939 | // "Reroute" the CFG edges to link in the new preheader. |
1940 | // If any of the predecessors falls through to the header, insert a branch |
1941 | // to the new preheader in that place. |
1942 | SmallVector<MachineOperand,1> Tmp2; |
1943 | SmallVector<MachineOperand,1> EmptyCond; |
1944 | |
1945 | TB = FB = nullptr; |
1946 | |
1947 | for (MachineBasicBlock *PB : Preds) { |
1948 | if (PB != Latch) { |
1949 | Tmp2.clear(); |
1950 | bool NotAnalyzed = TII->analyzeBranch(MBB&: *PB, TBB&: TB, FBB&: FB, Cond&: Tmp2, AllowModify: false); |
1951 | (void)NotAnalyzed; // suppress compiler warning |
1952 | assert (!NotAnalyzed && "Should be analyzable!" ); |
1953 | if (TB != Header && (Tmp2.empty() || FB != Header)) |
1954 | TII->insertBranch(MBB&: *PB, TBB: NewPH, FBB: nullptr, Cond: EmptyCond, DL); |
1955 | PB->ReplaceUsesOfBlockWith(Old: Header, New: NewPH); |
1956 | } |
1957 | } |
1958 | |
1959 | // It can happen that the latch block will fall through into the header. |
1960 | // Insert an unconditional branch to the header. |
1961 | TB = FB = nullptr; |
1962 | bool LatchNotAnalyzed = TII->analyzeBranch(MBB&: *Latch, TBB&: TB, FBB&: FB, Cond&: Tmp2, AllowModify: false); |
1963 | (void)LatchNotAnalyzed; // suppress compiler warning |
1964 | assert (!LatchNotAnalyzed && "Should be analyzable!" ); |
1965 | if (!TB && !FB) |
1966 | TII->insertBranch(MBB&: *Latch, TBB: Header, FBB: nullptr, Cond: EmptyCond, DL); |
1967 | |
1968 | // Finally, the branch from the preheader to the header. |
1969 | TII->insertBranch(MBB&: *NewPH, TBB: Header, FBB: nullptr, Cond: EmptyCond, DL); |
1970 | NewPH->addSuccessor(Succ: Header); |
1971 | |
1972 | MachineLoop *ParentLoop = L->getParentLoop(); |
1973 | if (ParentLoop) |
1974 | ParentLoop->addBasicBlockToLoop(NewBB: NewPH, LI&: *MLI); |
1975 | |
1976 | // Update the dominator information with the new preheader. |
1977 | if (MDT) { |
1978 | if (MachineDomTreeNode *HN = MDT->getNode(BB: Header)) { |
1979 | if (MachineDomTreeNode *DHN = HN->getIDom()) { |
1980 | MDT->addNewBlock(BB: NewPH, DomBB: DHN->getBlock()); |
1981 | MDT->changeImmediateDominator(N: Header, NewIDom: NewPH); |
1982 | } |
1983 | } |
1984 | } |
1985 | |
1986 | return NewPH; |
1987 | } |
1988 | |