1 | //===- HexagonSplitDouble.cpp ---------------------------------------------===// |
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 | #include "HexagonInstrInfo.h" |
10 | #include "HexagonRegisterInfo.h" |
11 | #include "HexagonSubtarget.h" |
12 | #include "llvm/ADT/BitVector.h" |
13 | #include "llvm/ADT/STLExtras.h" |
14 | #include "llvm/ADT/SmallVector.h" |
15 | #include "llvm/ADT/StringRef.h" |
16 | #include "llvm/CodeGen/MachineBasicBlock.h" |
17 | #include "llvm/CodeGen/MachineFunction.h" |
18 | #include "llvm/CodeGen/MachineFunctionPass.h" |
19 | #include "llvm/CodeGen/MachineInstr.h" |
20 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
21 | #include "llvm/CodeGen/MachineLoopInfo.h" |
22 | #include "llvm/CodeGen/MachineMemOperand.h" |
23 | #include "llvm/CodeGen/MachineOperand.h" |
24 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
25 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
26 | #include "llvm/Config/llvm-config.h" |
27 | #include "llvm/IR/DebugLoc.h" |
28 | #include "llvm/Pass.h" |
29 | #include "llvm/Support/CommandLine.h" |
30 | #include "llvm/Support/Compiler.h" |
31 | #include "llvm/Support/Debug.h" |
32 | #include "llvm/Support/ErrorHandling.h" |
33 | #include "llvm/Support/raw_ostream.h" |
34 | #include <algorithm> |
35 | #include <cassert> |
36 | #include <cstdint> |
37 | #include <limits> |
38 | #include <map> |
39 | #include <set> |
40 | #include <utility> |
41 | #include <vector> |
42 | |
43 | #define DEBUG_TYPE "hsdr" |
44 | |
45 | using namespace llvm; |
46 | |
47 | namespace llvm { |
48 | |
49 | FunctionPass *createHexagonSplitDoubleRegs(); |
50 | void initializeHexagonSplitDoubleRegsPass(PassRegistry&); |
51 | |
52 | } // end namespace llvm |
53 | |
54 | static cl::opt<int> MaxHSDR("max-hsdr" , cl::Hidden, cl::init(Val: -1), |
55 | cl::desc("Maximum number of split partitions" )); |
56 | static cl::opt<bool> MemRefsFixed("hsdr-no-mem" , cl::Hidden, cl::init(Val: true), |
57 | cl::desc("Do not split loads or stores" )); |
58 | static cl::opt<bool> SplitAll("hsdr-split-all" , cl::Hidden, cl::init(Val: false), |
59 | cl::desc("Split all partitions" )); |
60 | |
61 | namespace { |
62 | |
63 | class HexagonSplitDoubleRegs : public MachineFunctionPass { |
64 | public: |
65 | static char ID; |
66 | |
67 | HexagonSplitDoubleRegs() : MachineFunctionPass(ID) {} |
68 | |
69 | StringRef getPassName() const override { |
70 | return "Hexagon Split Double Registers" ; |
71 | } |
72 | |
73 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
74 | AU.addRequired<MachineLoopInfoWrapperPass>(); |
75 | AU.addPreserved<MachineLoopInfoWrapperPass>(); |
76 | MachineFunctionPass::getAnalysisUsage(AU); |
77 | } |
78 | |
79 | bool runOnMachineFunction(MachineFunction &MF) override; |
80 | |
81 | private: |
82 | static const TargetRegisterClass *const DoubleRC; |
83 | |
84 | const HexagonRegisterInfo *TRI = nullptr; |
85 | const HexagonInstrInfo *TII = nullptr; |
86 | const MachineLoopInfo *MLI; |
87 | MachineRegisterInfo *MRI; |
88 | |
89 | using USet = std::set<unsigned>; |
90 | using UUSetMap = std::map<unsigned, USet>; |
91 | using UUPair = std::pair<unsigned, unsigned>; |
92 | using UUPairMap = std::map<unsigned, UUPair>; |
93 | using LoopRegMap = std::map<const MachineLoop *, USet>; |
94 | |
95 | bool isInduction(unsigned Reg, LoopRegMap &IRM) const; |
96 | bool isVolatileInstr(const MachineInstr *MI) const; |
97 | bool isFixedInstr(const MachineInstr *MI) const; |
98 | void partitionRegisters(UUSetMap &P2Rs); |
99 | int32_t profit(const MachineInstr *MI) const; |
100 | int32_t profit(Register Reg) const; |
101 | bool isProfitable(const USet &Part, LoopRegMap &IRM) const; |
102 | |
103 | void collectIndRegsForLoop(const MachineLoop *L, USet &Rs); |
104 | void collectIndRegs(LoopRegMap &IRM); |
105 | |
106 | void createHalfInstr(unsigned Opc, MachineInstr *MI, |
107 | const UUPairMap &PairMap, unsigned SubR); |
108 | void splitMemRef(MachineInstr *MI, const UUPairMap &PairMap); |
109 | void splitImmediate(MachineInstr *MI, const UUPairMap &PairMap); |
110 | void splitCombine(MachineInstr *MI, const UUPairMap &PairMap); |
111 | void splitExt(MachineInstr *MI, const UUPairMap &PairMap); |
112 | void splitShift(MachineInstr *MI, const UUPairMap &PairMap); |
113 | void splitAslOr(MachineInstr *MI, const UUPairMap &PairMap); |
114 | bool splitInstr(MachineInstr *MI, const UUPairMap &PairMap); |
115 | void replaceSubregUses(MachineInstr *MI, const UUPairMap &PairMap); |
116 | void collapseRegPairs(MachineInstr *MI, const UUPairMap &PairMap); |
117 | bool splitPartition(const USet &Part); |
118 | |
119 | static int Counter; |
120 | |
121 | static void dump_partition(raw_ostream&, const USet&, |
122 | const TargetRegisterInfo&); |
123 | }; |
124 | |
125 | } // end anonymous namespace |
126 | |
127 | char HexagonSplitDoubleRegs::ID; |
128 | int HexagonSplitDoubleRegs::Counter = 0; |
129 | const TargetRegisterClass *const HexagonSplitDoubleRegs::DoubleRC = |
130 | &Hexagon::DoubleRegsRegClass; |
131 | |
132 | INITIALIZE_PASS(HexagonSplitDoubleRegs, "hexagon-split-double" , |
133 | "Hexagon Split Double Registers" , false, false) |
134 | |
135 | #if !defined(NDEBUG) || defined(LLVM_ENABLE_DUMP) |
136 | LLVM_DUMP_METHOD void HexagonSplitDoubleRegs::dump_partition(raw_ostream &os, |
137 | const USet &Part, const TargetRegisterInfo &TRI) { |
138 | dbgs() << '{'; |
139 | for (auto I : Part) |
140 | dbgs() << ' ' << printReg(I, &TRI); |
141 | dbgs() << " }" ; |
142 | } |
143 | #endif |
144 | |
145 | bool HexagonSplitDoubleRegs::isInduction(unsigned Reg, LoopRegMap &IRM) const { |
146 | for (auto I : IRM) { |
147 | const USet &Rs = I.second; |
148 | if (Rs.find(x: Reg) != Rs.end()) |
149 | return true; |
150 | } |
151 | return false; |
152 | } |
153 | |
154 | bool HexagonSplitDoubleRegs::isVolatileInstr(const MachineInstr *MI) const { |
155 | for (auto &MO : MI->memoperands()) |
156 | if (MO->isVolatile() || MO->isAtomic()) |
157 | return true; |
158 | return false; |
159 | } |
160 | |
161 | bool HexagonSplitDoubleRegs::isFixedInstr(const MachineInstr *MI) const { |
162 | if (MI->mayLoadOrStore()) |
163 | if (MemRefsFixed || isVolatileInstr(MI)) |
164 | return true; |
165 | if (MI->isDebugInstr()) |
166 | return false; |
167 | |
168 | unsigned Opc = MI->getOpcode(); |
169 | switch (Opc) { |
170 | default: |
171 | return true; |
172 | |
173 | case TargetOpcode::PHI: |
174 | case TargetOpcode::COPY: |
175 | break; |
176 | |
177 | case Hexagon::L2_loadrd_io: |
178 | // Not handling stack stores (only reg-based addresses). |
179 | if (MI->getOperand(i: 1).isReg()) |
180 | break; |
181 | return true; |
182 | case Hexagon::S2_storerd_io: |
183 | // Not handling stack stores (only reg-based addresses). |
184 | if (MI->getOperand(i: 0).isReg()) |
185 | break; |
186 | return true; |
187 | case Hexagon::L2_loadrd_pi: |
188 | case Hexagon::S2_storerd_pi: |
189 | |
190 | case Hexagon::A2_tfrpi: |
191 | case Hexagon::A2_combineii: |
192 | case Hexagon::A4_combineir: |
193 | case Hexagon::A4_combineii: |
194 | case Hexagon::A4_combineri: |
195 | case Hexagon::A2_combinew: |
196 | case Hexagon::CONST64: |
197 | |
198 | case Hexagon::A2_sxtw: |
199 | |
200 | case Hexagon::A2_andp: |
201 | case Hexagon::A2_orp: |
202 | case Hexagon::A2_xorp: |
203 | case Hexagon::S2_asl_i_p_or: |
204 | case Hexagon::S2_asl_i_p: |
205 | case Hexagon::S2_asr_i_p: |
206 | case Hexagon::S2_lsr_i_p: |
207 | break; |
208 | } |
209 | |
210 | for (auto &Op : MI->operands()) { |
211 | if (!Op.isReg()) |
212 | continue; |
213 | Register R = Op.getReg(); |
214 | if (!R.isVirtual()) |
215 | return true; |
216 | } |
217 | return false; |
218 | } |
219 | |
220 | void HexagonSplitDoubleRegs::partitionRegisters(UUSetMap &P2Rs) { |
221 | using UUMap = std::map<unsigned, unsigned>; |
222 | using UVect = std::vector<unsigned>; |
223 | |
224 | unsigned NumRegs = MRI->getNumVirtRegs(); |
225 | BitVector DoubleRegs(NumRegs); |
226 | for (unsigned i = 0; i < NumRegs; ++i) { |
227 | Register R = Register::index2VirtReg(Index: i); |
228 | if (MRI->getRegClass(Reg: R) == DoubleRC) |
229 | DoubleRegs.set(i); |
230 | } |
231 | |
232 | BitVector FixedRegs(NumRegs); |
233 | for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(Prev: x)) { |
234 | Register R = Register::index2VirtReg(Index: x); |
235 | MachineInstr *DefI = MRI->getVRegDef(Reg: R); |
236 | // In some cases a register may exist, but never be defined or used. |
237 | // It should never appear anywhere, but mark it as "fixed", just to be |
238 | // safe. |
239 | if (!DefI || isFixedInstr(MI: DefI)) |
240 | FixedRegs.set(x); |
241 | } |
242 | |
243 | UUSetMap AssocMap; |
244 | for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(Prev: x)) { |
245 | if (FixedRegs[x]) |
246 | continue; |
247 | Register R = Register::index2VirtReg(Index: x); |
248 | LLVM_DEBUG(dbgs() << printReg(R, TRI) << " ~~" ); |
249 | USet &Asc = AssocMap[R]; |
250 | for (auto U = MRI->use_nodbg_begin(RegNo: R), Z = MRI->use_nodbg_end(); |
251 | U != Z; ++U) { |
252 | MachineOperand &Op = *U; |
253 | MachineInstr *UseI = Op.getParent(); |
254 | if (isFixedInstr(MI: UseI)) |
255 | continue; |
256 | for (MachineOperand &MO : UseI->operands()) { |
257 | // Skip non-registers or registers with subregisters. |
258 | if (&MO == &Op || !MO.isReg() || MO.getSubReg()) |
259 | continue; |
260 | Register T = MO.getReg(); |
261 | if (!T.isVirtual()) { |
262 | FixedRegs.set(x); |
263 | continue; |
264 | } |
265 | if (MRI->getRegClass(Reg: T) != DoubleRC) |
266 | continue; |
267 | unsigned u = Register::virtReg2Index(Reg: T); |
268 | if (FixedRegs[u]) |
269 | continue; |
270 | LLVM_DEBUG(dbgs() << ' ' << printReg(T, TRI)); |
271 | Asc.insert(x: T); |
272 | // Make it symmetric. |
273 | AssocMap[T].insert(x: R); |
274 | } |
275 | } |
276 | LLVM_DEBUG(dbgs() << '\n'); |
277 | } |
278 | |
279 | UUMap R2P; |
280 | unsigned NextP = 1; |
281 | USet Visited; |
282 | for (int x = DoubleRegs.find_first(); x >= 0; x = DoubleRegs.find_next(Prev: x)) { |
283 | Register R = Register::index2VirtReg(Index: x); |
284 | if (Visited.count(x: R)) |
285 | continue; |
286 | // Create a new partition for R. |
287 | unsigned ThisP = FixedRegs[x] ? 0 : NextP++; |
288 | UVect WorkQ; |
289 | WorkQ.push_back(x: R); |
290 | for (unsigned i = 0; i < WorkQ.size(); ++i) { |
291 | unsigned T = WorkQ[i]; |
292 | if (Visited.count(x: T)) |
293 | continue; |
294 | R2P[T] = ThisP; |
295 | Visited.insert(x: T); |
296 | // Add all registers associated with T. |
297 | USet &Asc = AssocMap[T]; |
298 | append_range(C&: WorkQ, R&: Asc); |
299 | } |
300 | } |
301 | |
302 | for (auto I : R2P) |
303 | P2Rs[I.second].insert(x: I.first); |
304 | } |
305 | |
306 | static inline int32_t profitImm(unsigned Imm) { |
307 | int32_t P = 0; |
308 | if (Imm == 0 || Imm == 0xFFFFFFFF) |
309 | P += 10; |
310 | return P; |
311 | } |
312 | |
313 | int32_t HexagonSplitDoubleRegs::profit(const MachineInstr *MI) const { |
314 | unsigned ImmX = 0; |
315 | unsigned Opc = MI->getOpcode(); |
316 | switch (Opc) { |
317 | case TargetOpcode::PHI: |
318 | for (const auto &Op : MI->operands()) |
319 | if (!Op.getSubReg()) |
320 | return 0; |
321 | return 10; |
322 | case TargetOpcode::COPY: |
323 | if (MI->getOperand(i: 1).getSubReg() != 0) |
324 | return 10; |
325 | return 0; |
326 | |
327 | case Hexagon::L2_loadrd_io: |
328 | case Hexagon::S2_storerd_io: |
329 | return -1; |
330 | case Hexagon::L2_loadrd_pi: |
331 | case Hexagon::S2_storerd_pi: |
332 | return 2; |
333 | |
334 | case Hexagon::A2_tfrpi: |
335 | case Hexagon::CONST64: { |
336 | uint64_t D = MI->getOperand(i: 1).getImm(); |
337 | unsigned Lo = D & 0xFFFFFFFFULL; |
338 | unsigned Hi = D >> 32; |
339 | return profitImm(Imm: Lo) + profitImm(Imm: Hi); |
340 | } |
341 | case Hexagon::A2_combineii: |
342 | case Hexagon::A4_combineii: { |
343 | const MachineOperand &Op1 = MI->getOperand(i: 1); |
344 | const MachineOperand &Op2 = MI->getOperand(i: 2); |
345 | int32_t Prof1 = Op1.isImm() ? profitImm(Imm: Op1.getImm()) : 0; |
346 | int32_t Prof2 = Op2.isImm() ? profitImm(Imm: Op2.getImm()) : 0; |
347 | return Prof1 + Prof2; |
348 | } |
349 | case Hexagon::A4_combineri: |
350 | ImmX++; |
351 | // Fall through into A4_combineir. |
352 | [[fallthrough]]; |
353 | case Hexagon::A4_combineir: { |
354 | ImmX++; |
355 | const MachineOperand &OpX = MI->getOperand(i: ImmX); |
356 | if (OpX.isImm()) { |
357 | int64_t V = OpX.getImm(); |
358 | if (V == 0 || V == -1) |
359 | return 10; |
360 | } |
361 | // Fall through into A2_combinew. |
362 | [[fallthrough]]; |
363 | } |
364 | case Hexagon::A2_combinew: |
365 | return 2; |
366 | |
367 | case Hexagon::A2_sxtw: |
368 | return 3; |
369 | |
370 | case Hexagon::A2_andp: |
371 | case Hexagon::A2_orp: |
372 | case Hexagon::A2_xorp: { |
373 | Register Rs = MI->getOperand(i: 1).getReg(); |
374 | Register Rt = MI->getOperand(i: 2).getReg(); |
375 | return profit(Reg: Rs) + profit(Reg: Rt); |
376 | } |
377 | |
378 | case Hexagon::S2_asl_i_p_or: { |
379 | unsigned S = MI->getOperand(i: 3).getImm(); |
380 | if (S == 0 || S == 32) |
381 | return 10; |
382 | return -1; |
383 | } |
384 | case Hexagon::S2_asl_i_p: |
385 | case Hexagon::S2_asr_i_p: |
386 | case Hexagon::S2_lsr_i_p: |
387 | unsigned S = MI->getOperand(i: 2).getImm(); |
388 | if (S == 0 || S == 32) |
389 | return 10; |
390 | if (S == 16) |
391 | return 5; |
392 | if (S == 48) |
393 | return 7; |
394 | return -10; |
395 | } |
396 | |
397 | return 0; |
398 | } |
399 | |
400 | int32_t HexagonSplitDoubleRegs::profit(Register Reg) const { |
401 | assert(Reg.isVirtual()); |
402 | |
403 | const MachineInstr *DefI = MRI->getVRegDef(Reg); |
404 | switch (DefI->getOpcode()) { |
405 | case Hexagon::A2_tfrpi: |
406 | case Hexagon::CONST64: |
407 | case Hexagon::A2_combineii: |
408 | case Hexagon::A4_combineii: |
409 | case Hexagon::A4_combineri: |
410 | case Hexagon::A4_combineir: |
411 | case Hexagon::A2_combinew: |
412 | return profit(MI: DefI); |
413 | default: |
414 | break; |
415 | } |
416 | return 0; |
417 | } |
418 | |
419 | bool HexagonSplitDoubleRegs::isProfitable(const USet &Part, LoopRegMap &IRM) |
420 | const { |
421 | unsigned FixedNum = 0, LoopPhiNum = 0; |
422 | int32_t TotalP = 0; |
423 | |
424 | for (unsigned DR : Part) { |
425 | MachineInstr *DefI = MRI->getVRegDef(Reg: DR); |
426 | int32_t P = profit(MI: DefI); |
427 | if (P == std::numeric_limits<int>::min()) |
428 | return false; |
429 | TotalP += P; |
430 | // Reduce the profitability of splitting induction registers. |
431 | if (isInduction(Reg: DR, IRM)) |
432 | TotalP -= 30; |
433 | |
434 | for (auto U = MRI->use_nodbg_begin(RegNo: DR), W = MRI->use_nodbg_end(); |
435 | U != W; ++U) { |
436 | MachineInstr *UseI = U->getParent(); |
437 | if (isFixedInstr(MI: UseI)) { |
438 | FixedNum++; |
439 | // Calculate the cost of generating REG_SEQUENCE instructions. |
440 | for (auto &Op : UseI->operands()) { |
441 | if (Op.isReg() && Part.count(x: Op.getReg())) |
442 | if (Op.getSubReg()) |
443 | TotalP -= 2; |
444 | } |
445 | continue; |
446 | } |
447 | // If a register from this partition is used in a fixed instruction, |
448 | // and there is also a register in this partition that is used in |
449 | // a loop phi node, then decrease the splitting profit as this can |
450 | // confuse the modulo scheduler. |
451 | if (UseI->isPHI()) { |
452 | const MachineBasicBlock *PB = UseI->getParent(); |
453 | const MachineLoop *L = MLI->getLoopFor(BB: PB); |
454 | if (L && L->getHeader() == PB) |
455 | LoopPhiNum++; |
456 | } |
457 | // Splittable instruction. |
458 | int32_t P = profit(MI: UseI); |
459 | if (P == std::numeric_limits<int>::min()) |
460 | return false; |
461 | TotalP += P; |
462 | } |
463 | } |
464 | |
465 | if (FixedNum > 0 && LoopPhiNum > 0) |
466 | TotalP -= 20*LoopPhiNum; |
467 | |
468 | LLVM_DEBUG(dbgs() << "Partition profit: " << TotalP << '\n'); |
469 | if (SplitAll) |
470 | return true; |
471 | return TotalP > 0; |
472 | } |
473 | |
474 | void HexagonSplitDoubleRegs::collectIndRegsForLoop(const MachineLoop *L, |
475 | USet &Rs) { |
476 | const MachineBasicBlock *HB = L->getHeader(); |
477 | const MachineBasicBlock *LB = L->getLoopLatch(); |
478 | if (!HB || !LB) |
479 | return; |
480 | |
481 | // Examine the latch branch. Expect it to be a conditional branch to |
482 | // the header (either "br-cond header" or "br-cond exit; br header"). |
483 | MachineBasicBlock *TB = nullptr, *FB = nullptr; |
484 | MachineBasicBlock *TmpLB = const_cast<MachineBasicBlock*>(LB); |
485 | SmallVector<MachineOperand,2> Cond; |
486 | bool BadLB = TII->analyzeBranch(MBB&: *TmpLB, TBB&: TB, FBB&: FB, Cond, AllowModify: false); |
487 | // Only analyzable conditional branches. HII::analyzeBranch will put |
488 | // the branch opcode as the first element of Cond, and the predicate |
489 | // operand as the second. |
490 | if (BadLB || Cond.size() != 2) |
491 | return; |
492 | // Only simple jump-conditional (with or without negation). |
493 | if (!TII->PredOpcodeHasJMP_c(Opcode: Cond[0].getImm())) |
494 | return; |
495 | // Must go to the header. |
496 | if (TB != HB && FB != HB) |
497 | return; |
498 | assert(Cond[1].isReg() && "Unexpected Cond vector from analyzeBranch" ); |
499 | // Expect a predicate register. |
500 | Register PR = Cond[1].getReg(); |
501 | assert(MRI->getRegClass(PR) == &Hexagon::PredRegsRegClass); |
502 | |
503 | // Get the registers on which the loop controlling compare instruction |
504 | // depends. |
505 | Register CmpR1, CmpR2; |
506 | const MachineInstr *CmpI = MRI->getVRegDef(Reg: PR); |
507 | while (CmpI->getOpcode() == Hexagon::C2_not) |
508 | CmpI = MRI->getVRegDef(Reg: CmpI->getOperand(i: 1).getReg()); |
509 | |
510 | int64_t Mask = 0, Val = 0; |
511 | bool OkCI = TII->analyzeCompare(MI: *CmpI, SrcReg&: CmpR1, SrcReg2&: CmpR2, Mask, Value&: Val); |
512 | if (!OkCI) |
513 | return; |
514 | // Eliminate non-double input registers. |
515 | if (CmpR1 && MRI->getRegClass(Reg: CmpR1) != DoubleRC) |
516 | CmpR1 = 0; |
517 | if (CmpR2 && MRI->getRegClass(Reg: CmpR2) != DoubleRC) |
518 | CmpR2 = 0; |
519 | if (!CmpR1 && !CmpR2) |
520 | return; |
521 | |
522 | // Now examine the top of the loop: the phi nodes that could poten- |
523 | // tially define loop induction registers. The registers defined by |
524 | // such a phi node would be used in a 64-bit add, which then would |
525 | // be used in the loop compare instruction. |
526 | |
527 | // Get the set of all double registers defined by phi nodes in the |
528 | // loop header. |
529 | using UVect = std::vector<unsigned>; |
530 | |
531 | UVect DP; |
532 | for (auto &MI : *HB) { |
533 | if (!MI.isPHI()) |
534 | break; |
535 | const MachineOperand &MD = MI.getOperand(i: 0); |
536 | Register R = MD.getReg(); |
537 | if (MRI->getRegClass(Reg: R) == DoubleRC) |
538 | DP.push_back(x: R); |
539 | } |
540 | if (DP.empty()) |
541 | return; |
542 | |
543 | auto NoIndOp = [this, CmpR1, CmpR2] (unsigned R) -> bool { |
544 | for (auto I = MRI->use_nodbg_begin(RegNo: R), E = MRI->use_nodbg_end(); |
545 | I != E; ++I) { |
546 | const MachineInstr *UseI = I->getParent(); |
547 | if (UseI->getOpcode() != Hexagon::A2_addp) |
548 | continue; |
549 | // Get the output from the add. If it is one of the inputs to the |
550 | // loop-controlling compare instruction, then R is likely an induc- |
551 | // tion register. |
552 | Register T = UseI->getOperand(i: 0).getReg(); |
553 | if (T == CmpR1 || T == CmpR2) |
554 | return false; |
555 | } |
556 | return true; |
557 | }; |
558 | UVect::iterator End = llvm::remove_if(Range&: DP, P: NoIndOp); |
559 | Rs.insert(first: DP.begin(), last: End); |
560 | Rs.insert(x: CmpR1); |
561 | Rs.insert(x: CmpR2); |
562 | |
563 | LLVM_DEBUG({ |
564 | dbgs() << "For loop at " << printMBBReference(*HB) << " ind regs: " ; |
565 | dump_partition(dbgs(), Rs, *TRI); |
566 | dbgs() << '\n'; |
567 | }); |
568 | } |
569 | |
570 | void HexagonSplitDoubleRegs::collectIndRegs(LoopRegMap &IRM) { |
571 | using LoopVector = std::vector<MachineLoop *>; |
572 | |
573 | LoopVector WorkQ; |
574 | |
575 | append_range(C&: WorkQ, R: *MLI); |
576 | for (unsigned i = 0; i < WorkQ.size(); ++i) |
577 | append_range(C&: WorkQ, R&: *WorkQ[i]); |
578 | |
579 | USet Rs; |
580 | for (MachineLoop *L : WorkQ) { |
581 | Rs.clear(); |
582 | collectIndRegsForLoop(L, Rs); |
583 | if (!Rs.empty()) |
584 | IRM.insert(x: std::make_pair(x&: L, y&: Rs)); |
585 | } |
586 | } |
587 | |
588 | void HexagonSplitDoubleRegs::createHalfInstr(unsigned Opc, MachineInstr *MI, |
589 | const UUPairMap &PairMap, unsigned SubR) { |
590 | MachineBasicBlock &B = *MI->getParent(); |
591 | DebugLoc DL = MI->getDebugLoc(); |
592 | MachineInstr *NewI = BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Opc)); |
593 | |
594 | for (auto &Op : MI->operands()) { |
595 | if (!Op.isReg()) { |
596 | NewI->addOperand(Op); |
597 | continue; |
598 | } |
599 | // For register operands, set the subregister. |
600 | Register R = Op.getReg(); |
601 | unsigned SR = Op.getSubReg(); |
602 | bool isVirtReg = R.isVirtual(); |
603 | bool isKill = Op.isKill(); |
604 | if (isVirtReg && MRI->getRegClass(Reg: R) == DoubleRC) { |
605 | isKill = false; |
606 | UUPairMap::const_iterator F = PairMap.find(x: R); |
607 | if (F == PairMap.end()) { |
608 | SR = SubR; |
609 | } else { |
610 | const UUPair &P = F->second; |
611 | R = (SubR == Hexagon::isub_lo) ? P.first : P.second; |
612 | SR = 0; |
613 | } |
614 | } |
615 | auto CO = MachineOperand::CreateReg(Reg: R, isDef: Op.isDef(), isImp: Op.isImplicit(), isKill, |
616 | isDead: Op.isDead(), isUndef: Op.isUndef(), isEarlyClobber: Op.isEarlyClobber(), SubReg: SR, isDebug: Op.isDebug(), |
617 | isInternalRead: Op.isInternalRead()); |
618 | NewI->addOperand(Op: CO); |
619 | } |
620 | } |
621 | |
622 | void HexagonSplitDoubleRegs::splitMemRef(MachineInstr *MI, |
623 | const UUPairMap &PairMap) { |
624 | bool Load = MI->mayLoad(); |
625 | unsigned OrigOpc = MI->getOpcode(); |
626 | bool PostInc = (OrigOpc == Hexagon::L2_loadrd_pi || |
627 | OrigOpc == Hexagon::S2_storerd_pi); |
628 | MachineInstr *LowI, *HighI; |
629 | MachineBasicBlock &B = *MI->getParent(); |
630 | DebugLoc DL = MI->getDebugLoc(); |
631 | |
632 | // Index of the base-address-register operand. |
633 | unsigned AdrX = PostInc ? (Load ? 2 : 1) |
634 | : (Load ? 1 : 0); |
635 | MachineOperand &AdrOp = MI->getOperand(i: AdrX); |
636 | unsigned RSA = getRegState(RegOp: AdrOp); |
637 | MachineOperand &ValOp = Load ? MI->getOperand(i: 0) |
638 | : (PostInc ? MI->getOperand(i: 3) |
639 | : MI->getOperand(i: 2)); |
640 | UUPairMap::const_iterator F = PairMap.find(x: ValOp.getReg()); |
641 | assert(F != PairMap.end()); |
642 | |
643 | if (Load) { |
644 | const UUPair &P = F->second; |
645 | int64_t Off = PostInc ? 0 : MI->getOperand(i: 2).getImm(); |
646 | LowI = BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::L2_loadri_io), DestReg: P.first) |
647 | .addReg(RegNo: AdrOp.getReg(), flags: RSA & ~RegState::Kill, SubReg: AdrOp.getSubReg()) |
648 | .addImm(Val: Off); |
649 | HighI = BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::L2_loadri_io), DestReg: P.second) |
650 | .addReg(RegNo: AdrOp.getReg(), flags: RSA & ~RegState::Kill, SubReg: AdrOp.getSubReg()) |
651 | .addImm(Val: Off+4); |
652 | } else { |
653 | const UUPair &P = F->second; |
654 | int64_t Off = PostInc ? 0 : MI->getOperand(i: 1).getImm(); |
655 | LowI = BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::S2_storeri_io)) |
656 | .addReg(RegNo: AdrOp.getReg(), flags: RSA & ~RegState::Kill, SubReg: AdrOp.getSubReg()) |
657 | .addImm(Val: Off) |
658 | .addReg(RegNo: P.first); |
659 | HighI = BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::S2_storeri_io)) |
660 | .addReg(RegNo: AdrOp.getReg(), flags: RSA & ~RegState::Kill, SubReg: AdrOp.getSubReg()) |
661 | .addImm(Val: Off+4) |
662 | .addReg(RegNo: P.second); |
663 | } |
664 | |
665 | if (PostInc) { |
666 | // Create the increment of the address register. |
667 | int64_t Inc = Load ? MI->getOperand(i: 3).getImm() |
668 | : MI->getOperand(i: 2).getImm(); |
669 | MachineOperand &UpdOp = Load ? MI->getOperand(i: 1) : MI->getOperand(i: 0); |
670 | const TargetRegisterClass *RC = MRI->getRegClass(Reg: UpdOp.getReg()); |
671 | Register NewR = MRI->createVirtualRegister(RegClass: RC); |
672 | assert(!UpdOp.getSubReg() && "Def operand with subreg" ); |
673 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_addi), DestReg: NewR) |
674 | .addReg(RegNo: AdrOp.getReg(), flags: RSA) |
675 | .addImm(Val: Inc); |
676 | MRI->replaceRegWith(FromReg: UpdOp.getReg(), ToReg: NewR); |
677 | // The original instruction will be deleted later. |
678 | } |
679 | |
680 | // Generate a new pair of memory-operands. |
681 | MachineFunction &MF = *B.getParent(); |
682 | for (auto &MO : MI->memoperands()) { |
683 | const MachinePointerInfo &Ptr = MO->getPointerInfo(); |
684 | MachineMemOperand::Flags F = MO->getFlags(); |
685 | Align A = MO->getAlign(); |
686 | |
687 | auto *Tmp1 = MF.getMachineMemOperand(PtrInfo: Ptr, F, Size: 4 /*size*/, BaseAlignment: A); |
688 | LowI->addMemOperand(MF, MO: Tmp1); |
689 | auto *Tmp2 = |
690 | MF.getMachineMemOperand(PtrInfo: Ptr, F, Size: 4 /*size*/, BaseAlignment: std::min(a: A, b: Align(4))); |
691 | HighI->addMemOperand(MF, MO: Tmp2); |
692 | } |
693 | } |
694 | |
695 | void HexagonSplitDoubleRegs::splitImmediate(MachineInstr *MI, |
696 | const UUPairMap &PairMap) { |
697 | MachineOperand &Op0 = MI->getOperand(i: 0); |
698 | MachineOperand &Op1 = MI->getOperand(i: 1); |
699 | assert(Op0.isReg() && Op1.isImm()); |
700 | uint64_t V = Op1.getImm(); |
701 | |
702 | MachineBasicBlock &B = *MI->getParent(); |
703 | DebugLoc DL = MI->getDebugLoc(); |
704 | UUPairMap::const_iterator F = PairMap.find(x: Op0.getReg()); |
705 | assert(F != PairMap.end()); |
706 | const UUPair &P = F->second; |
707 | |
708 | // The operand to A2_tfrsi can only have 32 significant bits. Immediate |
709 | // values in MachineOperand are stored as 64-bit integers, and so the |
710 | // value -1 may be represented either as 64-bit -1, or 4294967295. Both |
711 | // will have the 32 higher bits truncated in the end, but -1 will remain |
712 | // as -1, while the latter may appear to be a large unsigned value |
713 | // requiring a constant extender. The casting to int32_t will select the |
714 | // former representation. (The same reasoning applies to all 32-bit |
715 | // values.) |
716 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_tfrsi), DestReg: P.first) |
717 | .addImm(Val: int32_t(V & 0xFFFFFFFFULL)); |
718 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_tfrsi), DestReg: P.second) |
719 | .addImm(Val: int32_t(V >> 32)); |
720 | } |
721 | |
722 | void HexagonSplitDoubleRegs::splitCombine(MachineInstr *MI, |
723 | const UUPairMap &PairMap) { |
724 | MachineOperand &Op0 = MI->getOperand(i: 0); |
725 | MachineOperand &Op1 = MI->getOperand(i: 1); |
726 | MachineOperand &Op2 = MI->getOperand(i: 2); |
727 | assert(Op0.isReg()); |
728 | |
729 | MachineBasicBlock &B = *MI->getParent(); |
730 | DebugLoc DL = MI->getDebugLoc(); |
731 | UUPairMap::const_iterator F = PairMap.find(x: Op0.getReg()); |
732 | assert(F != PairMap.end()); |
733 | const UUPair &P = F->second; |
734 | |
735 | if (!Op1.isReg()) { |
736 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_tfrsi), DestReg: P.second) |
737 | .add(MO: Op1); |
738 | } else { |
739 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: P.second) |
740 | .addReg(RegNo: Op1.getReg(), flags: getRegState(RegOp: Op1), SubReg: Op1.getSubReg()); |
741 | } |
742 | |
743 | if (!Op2.isReg()) { |
744 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::A2_tfrsi), DestReg: P.first) |
745 | .add(MO: Op2); |
746 | } else { |
747 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: P.first) |
748 | .addReg(RegNo: Op2.getReg(), flags: getRegState(RegOp: Op2), SubReg: Op2.getSubReg()); |
749 | } |
750 | } |
751 | |
752 | void HexagonSplitDoubleRegs::splitExt(MachineInstr *MI, |
753 | const UUPairMap &PairMap) { |
754 | MachineOperand &Op0 = MI->getOperand(i: 0); |
755 | MachineOperand &Op1 = MI->getOperand(i: 1); |
756 | assert(Op0.isReg() && Op1.isReg()); |
757 | |
758 | MachineBasicBlock &B = *MI->getParent(); |
759 | DebugLoc DL = MI->getDebugLoc(); |
760 | UUPairMap::const_iterator F = PairMap.find(x: Op0.getReg()); |
761 | assert(F != PairMap.end()); |
762 | const UUPair &P = F->second; |
763 | unsigned RS = getRegState(RegOp: Op1); |
764 | |
765 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: P.first) |
766 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: Op1.getSubReg()); |
767 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: Hexagon::S2_asr_i_r), DestReg: P.second) |
768 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: Op1.getSubReg()) |
769 | .addImm(Val: 31); |
770 | } |
771 | |
772 | void HexagonSplitDoubleRegs::splitShift(MachineInstr *MI, |
773 | const UUPairMap &PairMap) { |
774 | using namespace Hexagon; |
775 | |
776 | MachineOperand &Op0 = MI->getOperand(i: 0); |
777 | MachineOperand &Op1 = MI->getOperand(i: 1); |
778 | MachineOperand &Op2 = MI->getOperand(i: 2); |
779 | assert(Op0.isReg() && Op1.isReg() && Op2.isImm()); |
780 | int64_t Sh64 = Op2.getImm(); |
781 | assert(Sh64 >= 0 && Sh64 < 64); |
782 | unsigned S = Sh64; |
783 | |
784 | UUPairMap::const_iterator F = PairMap.find(x: Op0.getReg()); |
785 | assert(F != PairMap.end()); |
786 | const UUPair &P = F->second; |
787 | Register LoR = P.first; |
788 | Register HiR = P.second; |
789 | |
790 | unsigned Opc = MI->getOpcode(); |
791 | bool Right = (Opc == S2_lsr_i_p || Opc == S2_asr_i_p); |
792 | bool Left = !Right; |
793 | bool Signed = (Opc == S2_asr_i_p); |
794 | |
795 | MachineBasicBlock &B = *MI->getParent(); |
796 | DebugLoc DL = MI->getDebugLoc(); |
797 | unsigned RS = getRegState(RegOp: Op1); |
798 | unsigned ShiftOpc = Left ? S2_asl_i_r |
799 | : (Signed ? S2_asr_i_r : S2_lsr_i_r); |
800 | unsigned LoSR = isub_lo; |
801 | unsigned HiSR = isub_hi; |
802 | |
803 | if (S == 0) { |
804 | // No shift, subregister copy. |
805 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: LoR) |
806 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR); |
807 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: HiR) |
808 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: HiSR); |
809 | } else if (S < 32) { |
810 | const TargetRegisterClass *IntRC = &IntRegsRegClass; |
811 | Register TmpR = MRI->createVirtualRegister(RegClass: IntRC); |
812 | // Expansion: |
813 | // Shift left: DR = shl R, #s |
814 | // LoR = shl R.lo, #s |
815 | // TmpR = extractu R.lo, #s, #32-s |
816 | // HiR = or (TmpR, asl(R.hi, #s)) |
817 | // Shift right: DR = shr R, #s |
818 | // HiR = shr R.hi, #s |
819 | // TmpR = shr R.lo, #s |
820 | // LoR = insert TmpR, R.hi, #s, #32-s |
821 | |
822 | // Shift left: |
823 | // LoR = shl R.lo, #s |
824 | // Shift right: |
825 | // TmpR = shr R.lo, #s |
826 | |
827 | // Make a special case for A2_aslh and A2_asrh (they are predicable as |
828 | // opposed to S2_asl_i_r/S2_asr_i_r). |
829 | if (S == 16 && Left) |
830 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_aslh), DestReg: LoR) |
831 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR); |
832 | else if (S == 16 && Signed) |
833 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_asrh), DestReg: TmpR) |
834 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR); |
835 | else |
836 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: ShiftOpc), DestReg: (Left ? LoR : TmpR)) |
837 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR) |
838 | .addImm(Val: S); |
839 | |
840 | if (Left) { |
841 | // TmpR = extractu R.lo, #s, #32-s |
842 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_extractu), DestReg: TmpR) |
843 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR) |
844 | .addImm(Val: S) |
845 | .addImm(Val: 32-S); |
846 | // HiR = or (TmpR, asl(R.hi, #s)) |
847 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asl_i_r_or), DestReg: HiR) |
848 | .addReg(RegNo: TmpR) |
849 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: HiSR) |
850 | .addImm(Val: S); |
851 | } else { |
852 | // HiR = shr R.hi, #s |
853 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: ShiftOpc), DestReg: HiR) |
854 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: HiSR) |
855 | .addImm(Val: S); |
856 | // LoR = insert TmpR, R.hi, #s, #32-s |
857 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_insert), DestReg: LoR) |
858 | .addReg(RegNo: TmpR) |
859 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: HiSR) |
860 | .addImm(Val: S) |
861 | .addImm(Val: 32-S); |
862 | } |
863 | } else if (S == 32) { |
864 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: (Left ? HiR : LoR)) |
865 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: (Left ? LoSR : HiSR)); |
866 | if (!Signed) |
867 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_tfrsi), DestReg: (Left ? LoR : HiR)) |
868 | .addImm(Val: 0); |
869 | else // Must be right shift. |
870 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asr_i_r), DestReg: HiR) |
871 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: HiSR) |
872 | .addImm(Val: 31); |
873 | } else if (S < 64) { |
874 | S -= 32; |
875 | if (S == 16 && Left) |
876 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_aslh), DestReg: HiR) |
877 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: LoSR); |
878 | else if (S == 16 && Signed) |
879 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_asrh), DestReg: LoR) |
880 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: HiSR); |
881 | else |
882 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: ShiftOpc), DestReg: (Left ? HiR : LoR)) |
883 | .addReg(RegNo: Op1.getReg(), flags: RS & ~RegState::Kill, SubReg: (Left ? LoSR : HiSR)) |
884 | .addImm(Val: S); |
885 | |
886 | if (Signed) |
887 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asr_i_r), DestReg: HiR) |
888 | .addReg(RegNo: Op1.getReg(), flags: RS, SubReg: HiSR) |
889 | .addImm(Val: 31); |
890 | else |
891 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_tfrsi), DestReg: (Left ? LoR : HiR)) |
892 | .addImm(Val: 0); |
893 | } |
894 | } |
895 | |
896 | void HexagonSplitDoubleRegs::splitAslOr(MachineInstr *MI, |
897 | const UUPairMap &PairMap) { |
898 | using namespace Hexagon; |
899 | |
900 | MachineOperand &Op0 = MI->getOperand(i: 0); |
901 | MachineOperand &Op1 = MI->getOperand(i: 1); |
902 | MachineOperand &Op2 = MI->getOperand(i: 2); |
903 | MachineOperand &Op3 = MI->getOperand(i: 3); |
904 | assert(Op0.isReg() && Op1.isReg() && Op2.isReg() && Op3.isImm()); |
905 | int64_t Sh64 = Op3.getImm(); |
906 | assert(Sh64 >= 0 && Sh64 < 64); |
907 | unsigned S = Sh64; |
908 | |
909 | UUPairMap::const_iterator F = PairMap.find(x: Op0.getReg()); |
910 | assert(F != PairMap.end()); |
911 | const UUPair &P = F->second; |
912 | unsigned LoR = P.first; |
913 | unsigned HiR = P.second; |
914 | |
915 | MachineBasicBlock &B = *MI->getParent(); |
916 | DebugLoc DL = MI->getDebugLoc(); |
917 | unsigned RS1 = getRegState(RegOp: Op1); |
918 | unsigned RS2 = getRegState(RegOp: Op2); |
919 | const TargetRegisterClass *IntRC = &IntRegsRegClass; |
920 | |
921 | unsigned LoSR = isub_lo; |
922 | unsigned HiSR = isub_hi; |
923 | |
924 | // Op0 = S2_asl_i_p_or Op1, Op2, Op3 |
925 | // means: Op0 = or (Op1, asl(Op2, Op3)) |
926 | |
927 | // Expansion of |
928 | // DR = or (R1, asl(R2, #s)) |
929 | // |
930 | // LoR = or (R1.lo, asl(R2.lo, #s)) |
931 | // Tmp1 = extractu R2.lo, #s, #32-s |
932 | // Tmp2 = or R1.hi, Tmp1 |
933 | // HiR = or (Tmp2, asl(R2.hi, #s)) |
934 | |
935 | if (S == 0) { |
936 | // DR = or (R1, asl(R2, #0)) |
937 | // -> or (R1, R2) |
938 | // i.e. LoR = or R1.lo, R2.lo |
939 | // HiR = or R1.hi, R2.hi |
940 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_or), DestReg: LoR) |
941 | .addReg(RegNo: Op1.getReg(), flags: RS1 & ~RegState::Kill, SubReg: LoSR) |
942 | .addReg(RegNo: Op2.getReg(), flags: RS2 & ~RegState::Kill, SubReg: LoSR); |
943 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_or), DestReg: HiR) |
944 | .addReg(RegNo: Op1.getReg(), flags: RS1, SubReg: HiSR) |
945 | .addReg(RegNo: Op2.getReg(), flags: RS2, SubReg: HiSR); |
946 | } else if (S < 32) { |
947 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asl_i_r_or), DestReg: LoR) |
948 | .addReg(RegNo: Op1.getReg(), flags: RS1 & ~RegState::Kill, SubReg: LoSR) |
949 | .addReg(RegNo: Op2.getReg(), flags: RS2 & ~RegState::Kill, SubReg: LoSR) |
950 | .addImm(Val: S); |
951 | Register TmpR1 = MRI->createVirtualRegister(RegClass: IntRC); |
952 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_extractu), DestReg: TmpR1) |
953 | .addReg(RegNo: Op2.getReg(), flags: RS2 & ~RegState::Kill, SubReg: LoSR) |
954 | .addImm(Val: S) |
955 | .addImm(Val: 32-S); |
956 | Register TmpR2 = MRI->createVirtualRegister(RegClass: IntRC); |
957 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_or), DestReg: TmpR2) |
958 | .addReg(RegNo: Op1.getReg(), flags: RS1, SubReg: HiSR) |
959 | .addReg(RegNo: TmpR1); |
960 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asl_i_r_or), DestReg: HiR) |
961 | .addReg(RegNo: TmpR2) |
962 | .addReg(RegNo: Op2.getReg(), flags: RS2, SubReg: HiSR) |
963 | .addImm(Val: S); |
964 | } else if (S == 32) { |
965 | // DR = or (R1, asl(R2, #32)) |
966 | // -> or R1, R2.lo |
967 | // LoR = R1.lo |
968 | // HiR = or R1.hi, R2.lo |
969 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: LoR) |
970 | .addReg(RegNo: Op1.getReg(), flags: RS1 & ~RegState::Kill, SubReg: LoSR); |
971 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: A2_or), DestReg: HiR) |
972 | .addReg(RegNo: Op1.getReg(), flags: RS1, SubReg: HiSR) |
973 | .addReg(RegNo: Op2.getReg(), flags: RS2, SubReg: LoSR); |
974 | } else if (S < 64) { |
975 | // DR = or (R1, asl(R2, #s)) |
976 | // |
977 | // LoR = R1:lo |
978 | // HiR = or (R1:hi, asl(R2:lo, #s-32)) |
979 | S -= 32; |
980 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: LoR) |
981 | .addReg(RegNo: Op1.getReg(), flags: RS1 & ~RegState::Kill, SubReg: LoSR); |
982 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: S2_asl_i_r_or), DestReg: HiR) |
983 | .addReg(RegNo: Op1.getReg(), flags: RS1, SubReg: HiSR) |
984 | .addReg(RegNo: Op2.getReg(), flags: RS2, SubReg: LoSR) |
985 | .addImm(Val: S); |
986 | } |
987 | } |
988 | |
989 | bool HexagonSplitDoubleRegs::splitInstr(MachineInstr *MI, |
990 | const UUPairMap &PairMap) { |
991 | using namespace Hexagon; |
992 | |
993 | LLVM_DEBUG(dbgs() << "Splitting: " << *MI); |
994 | bool Split = false; |
995 | unsigned Opc = MI->getOpcode(); |
996 | |
997 | switch (Opc) { |
998 | case TargetOpcode::PHI: |
999 | case TargetOpcode::COPY: { |
1000 | Register DstR = MI->getOperand(i: 0).getReg(); |
1001 | if (MRI->getRegClass(Reg: DstR) == DoubleRC) { |
1002 | createHalfInstr(Opc, MI, PairMap, SubR: isub_lo); |
1003 | createHalfInstr(Opc, MI, PairMap, SubR: isub_hi); |
1004 | Split = true; |
1005 | } |
1006 | break; |
1007 | } |
1008 | case A2_andp: |
1009 | createHalfInstr(Opc: A2_and, MI, PairMap, SubR: isub_lo); |
1010 | createHalfInstr(Opc: A2_and, MI, PairMap, SubR: isub_hi); |
1011 | Split = true; |
1012 | break; |
1013 | case A2_orp: |
1014 | createHalfInstr(Opc: A2_or, MI, PairMap, SubR: isub_lo); |
1015 | createHalfInstr(Opc: A2_or, MI, PairMap, SubR: isub_hi); |
1016 | Split = true; |
1017 | break; |
1018 | case A2_xorp: |
1019 | createHalfInstr(Opc: A2_xor, MI, PairMap, SubR: isub_lo); |
1020 | createHalfInstr(Opc: A2_xor, MI, PairMap, SubR: isub_hi); |
1021 | Split = true; |
1022 | break; |
1023 | |
1024 | case L2_loadrd_io: |
1025 | case L2_loadrd_pi: |
1026 | case S2_storerd_io: |
1027 | case S2_storerd_pi: |
1028 | splitMemRef(MI, PairMap); |
1029 | Split = true; |
1030 | break; |
1031 | |
1032 | case A2_tfrpi: |
1033 | case CONST64: |
1034 | splitImmediate(MI, PairMap); |
1035 | Split = true; |
1036 | break; |
1037 | |
1038 | case A2_combineii: |
1039 | case A4_combineir: |
1040 | case A4_combineii: |
1041 | case A4_combineri: |
1042 | case A2_combinew: |
1043 | splitCombine(MI, PairMap); |
1044 | Split = true; |
1045 | break; |
1046 | |
1047 | case A2_sxtw: |
1048 | splitExt(MI, PairMap); |
1049 | Split = true; |
1050 | break; |
1051 | |
1052 | case S2_asl_i_p: |
1053 | case S2_asr_i_p: |
1054 | case S2_lsr_i_p: |
1055 | splitShift(MI, PairMap); |
1056 | Split = true; |
1057 | break; |
1058 | |
1059 | case S2_asl_i_p_or: |
1060 | splitAslOr(MI, PairMap); |
1061 | Split = true; |
1062 | break; |
1063 | |
1064 | default: |
1065 | llvm_unreachable("Instruction not splitable" ); |
1066 | return false; |
1067 | } |
1068 | |
1069 | return Split; |
1070 | } |
1071 | |
1072 | void HexagonSplitDoubleRegs::replaceSubregUses(MachineInstr *MI, |
1073 | const UUPairMap &PairMap) { |
1074 | for (auto &Op : MI->operands()) { |
1075 | if (!Op.isReg() || !Op.isUse() || !Op.getSubReg()) |
1076 | continue; |
1077 | Register R = Op.getReg(); |
1078 | UUPairMap::const_iterator F = PairMap.find(x: R); |
1079 | if (F == PairMap.end()) |
1080 | continue; |
1081 | const UUPair &P = F->second; |
1082 | switch (Op.getSubReg()) { |
1083 | case Hexagon::isub_lo: |
1084 | Op.setReg(P.first); |
1085 | break; |
1086 | case Hexagon::isub_hi: |
1087 | Op.setReg(P.second); |
1088 | break; |
1089 | } |
1090 | Op.setSubReg(0); |
1091 | } |
1092 | } |
1093 | |
1094 | void HexagonSplitDoubleRegs::collapseRegPairs(MachineInstr *MI, |
1095 | const UUPairMap &PairMap) { |
1096 | MachineBasicBlock &B = *MI->getParent(); |
1097 | DebugLoc DL = MI->getDebugLoc(); |
1098 | |
1099 | for (auto &Op : MI->operands()) { |
1100 | if (!Op.isReg() || !Op.isUse()) |
1101 | continue; |
1102 | Register R = Op.getReg(); |
1103 | if (!R.isVirtual()) |
1104 | continue; |
1105 | if (MRI->getRegClass(Reg: R) != DoubleRC || Op.getSubReg()) |
1106 | continue; |
1107 | UUPairMap::const_iterator F = PairMap.find(x: R); |
1108 | if (F == PairMap.end()) |
1109 | continue; |
1110 | const UUPair &Pr = F->second; |
1111 | Register NewDR = MRI->createVirtualRegister(RegClass: DoubleRC); |
1112 | BuildMI(BB&: B, I: MI, MIMD: DL, MCID: TII->get(Opcode: TargetOpcode::REG_SEQUENCE), DestReg: NewDR) |
1113 | .addReg(RegNo: Pr.first) |
1114 | .addImm(Val: Hexagon::isub_lo) |
1115 | .addReg(RegNo: Pr.second) |
1116 | .addImm(Val: Hexagon::isub_hi); |
1117 | Op.setReg(NewDR); |
1118 | } |
1119 | } |
1120 | |
1121 | bool HexagonSplitDoubleRegs::splitPartition(const USet &Part) { |
1122 | using MISet = std::set<MachineInstr *>; |
1123 | |
1124 | const TargetRegisterClass *IntRC = &Hexagon::IntRegsRegClass; |
1125 | bool Changed = false; |
1126 | |
1127 | LLVM_DEBUG(dbgs() << "Splitting partition: " ; |
1128 | dump_partition(dbgs(), Part, *TRI); dbgs() << '\n'); |
1129 | |
1130 | UUPairMap PairMap; |
1131 | |
1132 | MISet SplitIns; |
1133 | for (unsigned DR : Part) { |
1134 | MachineInstr *DefI = MRI->getVRegDef(Reg: DR); |
1135 | SplitIns.insert(x: DefI); |
1136 | |
1137 | // Collect all instructions, including fixed ones. We won't split them, |
1138 | // but we need to visit them again to insert the REG_SEQUENCE instructions. |
1139 | for (auto U = MRI->use_nodbg_begin(RegNo: DR), W = MRI->use_nodbg_end(); |
1140 | U != W; ++U) |
1141 | SplitIns.insert(x: U->getParent()); |
1142 | |
1143 | Register LoR = MRI->createVirtualRegister(RegClass: IntRC); |
1144 | Register HiR = MRI->createVirtualRegister(RegClass: IntRC); |
1145 | LLVM_DEBUG(dbgs() << "Created mapping: " << printReg(DR, TRI) << " -> " |
1146 | << printReg(HiR, TRI) << ':' << printReg(LoR, TRI) |
1147 | << '\n'); |
1148 | PairMap.insert(x: std::make_pair(x&: DR, y: UUPair(LoR, HiR))); |
1149 | } |
1150 | |
1151 | MISet Erase; |
1152 | for (auto *MI : SplitIns) { |
1153 | if (isFixedInstr(MI)) { |
1154 | collapseRegPairs(MI, PairMap); |
1155 | } else { |
1156 | bool Done = splitInstr(MI, PairMap); |
1157 | if (Done) |
1158 | Erase.insert(x: MI); |
1159 | Changed |= Done; |
1160 | } |
1161 | } |
1162 | |
1163 | for (unsigned DR : Part) { |
1164 | // Before erasing "double" instructions, revisit all uses of the double |
1165 | // registers in this partition, and replace all uses of them with subre- |
1166 | // gisters, with the corresponding single registers. |
1167 | MISet Uses; |
1168 | for (auto U = MRI->use_nodbg_begin(RegNo: DR), W = MRI->use_nodbg_end(); |
1169 | U != W; ++U) |
1170 | Uses.insert(x: U->getParent()); |
1171 | for (auto *M : Uses) |
1172 | replaceSubregUses(MI: M, PairMap); |
1173 | } |
1174 | |
1175 | for (auto *MI : Erase) { |
1176 | MachineBasicBlock *B = MI->getParent(); |
1177 | B->erase(I: MI); |
1178 | } |
1179 | |
1180 | return Changed; |
1181 | } |
1182 | |
1183 | bool HexagonSplitDoubleRegs::runOnMachineFunction(MachineFunction &MF) { |
1184 | if (skipFunction(F: MF.getFunction())) |
1185 | return false; |
1186 | |
1187 | LLVM_DEBUG(dbgs() << "Splitting double registers in function: " |
1188 | << MF.getName() << '\n'); |
1189 | |
1190 | auto &ST = MF.getSubtarget<HexagonSubtarget>(); |
1191 | TRI = ST.getRegisterInfo(); |
1192 | TII = ST.getInstrInfo(); |
1193 | MRI = &MF.getRegInfo(); |
1194 | MLI = &getAnalysis<MachineLoopInfoWrapperPass>().getLI(); |
1195 | |
1196 | UUSetMap P2Rs; |
1197 | LoopRegMap IRM; |
1198 | |
1199 | collectIndRegs(IRM); |
1200 | partitionRegisters(P2Rs); |
1201 | |
1202 | LLVM_DEBUG({ |
1203 | dbgs() << "Register partitioning: (partition #0 is fixed)\n" ; |
1204 | for (UUSetMap::iterator I = P2Rs.begin(), E = P2Rs.end(); I != E; ++I) { |
1205 | dbgs() << '#' << I->first << " -> " ; |
1206 | dump_partition(dbgs(), I->second, *TRI); |
1207 | dbgs() << '\n'; |
1208 | } |
1209 | }); |
1210 | |
1211 | bool Changed = false; |
1212 | int Limit = MaxHSDR; |
1213 | |
1214 | for (UUSetMap::iterator I = P2Rs.begin(), E = P2Rs.end(); I != E; ++I) { |
1215 | if (I->first == 0) |
1216 | continue; |
1217 | if (Limit >= 0 && Counter >= Limit) |
1218 | break; |
1219 | USet &Part = I->second; |
1220 | LLVM_DEBUG(dbgs() << "Calculating profit for partition #" << I->first |
1221 | << '\n'); |
1222 | if (!isProfitable(Part, IRM)) |
1223 | continue; |
1224 | Counter++; |
1225 | Changed |= splitPartition(Part); |
1226 | } |
1227 | |
1228 | return Changed; |
1229 | } |
1230 | |
1231 | FunctionPass *llvm::createHexagonSplitDoubleRegs() { |
1232 | return new HexagonSplitDoubleRegs(); |
1233 | } |
1234 | |