1 | //=== A15SDOptimizerPass.cpp - Optimize DPR and SPR register accesses on A15==// |
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 | // The Cortex-A15 processor employs a tracking scheme in its register renaming |
10 | // in order to process each instruction's micro-ops speculatively and |
11 | // out-of-order with appropriate forwarding. The ARM architecture allows VFP |
12 | // instructions to read and write 32-bit S-registers. Each S-register |
13 | // corresponds to one half (upper or lower) of an overlaid 64-bit D-register. |
14 | // |
15 | // There are several instruction patterns which can be used to provide this |
16 | // capability which can provide higher performance than other, potentially more |
17 | // direct patterns, specifically around when one micro-op reads a D-register |
18 | // operand that has recently been written as one or more S-register results. |
19 | // |
20 | // This file defines a pre-regalloc pass which looks for SPR producers which |
21 | // are going to be used by a DPR (or QPR) consumers and creates the more |
22 | // optimized access pattern. |
23 | // |
24 | //===----------------------------------------------------------------------===// |
25 | |
26 | #include "ARM.h" |
27 | #include "ARMBaseInstrInfo.h" |
28 | #include "ARMBaseRegisterInfo.h" |
29 | #include "ARMSubtarget.h" |
30 | #include "llvm/ADT/Statistic.h" |
31 | #include "llvm/CodeGen/MachineFunction.h" |
32 | #include "llvm/CodeGen/MachineFunctionPass.h" |
33 | #include "llvm/CodeGen/MachineInstr.h" |
34 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
35 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
36 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
37 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
38 | #include "llvm/Support/Debug.h" |
39 | #include "llvm/Support/raw_ostream.h" |
40 | #include <map> |
41 | #include <set> |
42 | |
43 | using namespace llvm; |
44 | |
45 | #define DEBUG_TYPE "a15-sd-optimizer" |
46 | |
47 | namespace { |
48 | struct A15SDOptimizer : public MachineFunctionPass { |
49 | static char ID; |
50 | A15SDOptimizer() : MachineFunctionPass(ID) {} |
51 | |
52 | bool runOnMachineFunction(MachineFunction &Fn) override; |
53 | |
54 | StringRef getPassName() const override { return "ARM A15 S->D optimizer" ; } |
55 | |
56 | private: |
57 | const ARMBaseInstrInfo *TII; |
58 | const TargetRegisterInfo *TRI; |
59 | MachineRegisterInfo *MRI; |
60 | |
61 | bool runOnInstruction(MachineInstr *MI); |
62 | |
63 | // |
64 | // Instruction builder helpers |
65 | // |
66 | unsigned createDupLane(MachineBasicBlock &MBB, |
67 | MachineBasicBlock::iterator InsertBefore, |
68 | const DebugLoc &DL, unsigned Reg, unsigned Lane, |
69 | bool QPR = false); |
70 | |
71 | unsigned createExtractSubreg(MachineBasicBlock &MBB, |
72 | MachineBasicBlock::iterator InsertBefore, |
73 | const DebugLoc &DL, unsigned DReg, |
74 | unsigned Lane, const TargetRegisterClass *TRC); |
75 | |
76 | unsigned createVExt(MachineBasicBlock &MBB, |
77 | MachineBasicBlock::iterator InsertBefore, |
78 | const DebugLoc &DL, unsigned Ssub0, unsigned Ssub1); |
79 | |
80 | unsigned createRegSequence(MachineBasicBlock &MBB, |
81 | MachineBasicBlock::iterator InsertBefore, |
82 | const DebugLoc &DL, unsigned Reg1, |
83 | unsigned Reg2); |
84 | |
85 | unsigned createInsertSubreg(MachineBasicBlock &MBB, |
86 | MachineBasicBlock::iterator InsertBefore, |
87 | const DebugLoc &DL, unsigned DReg, |
88 | unsigned Lane, unsigned ToInsert); |
89 | |
90 | unsigned createImplicitDef(MachineBasicBlock &MBB, |
91 | MachineBasicBlock::iterator InsertBefore, |
92 | const DebugLoc &DL); |
93 | |
94 | // |
95 | // Various property checkers |
96 | // |
97 | bool usesRegClass(MachineOperand &MO, const TargetRegisterClass *TRC); |
98 | bool hasPartialWrite(MachineInstr *MI); |
99 | SmallVector<unsigned, 8> getReadDPRs(MachineInstr *MI); |
100 | unsigned getDPRLaneFromSPR(unsigned SReg); |
101 | |
102 | // |
103 | // Methods used for getting the definitions of partial registers |
104 | // |
105 | |
106 | MachineInstr *elideCopies(MachineInstr *MI); |
107 | void elideCopiesAndPHIs(MachineInstr *MI, |
108 | SmallVectorImpl<MachineInstr*> &Outs); |
109 | |
110 | // |
111 | // Pattern optimization methods |
112 | // |
113 | unsigned optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg); |
114 | unsigned optimizeSDPattern(MachineInstr *MI); |
115 | unsigned getPrefSPRLane(unsigned SReg); |
116 | |
117 | // |
118 | // Sanitizing method - used to make sure if don't leave dead code around. |
119 | // |
120 | void eraseInstrWithNoUses(MachineInstr *MI); |
121 | |
122 | // |
123 | // A map used to track the changes done by this pass. |
124 | // |
125 | std::map<MachineInstr*, unsigned> Replacements; |
126 | std::set<MachineInstr *> DeadInstr; |
127 | }; |
128 | char A15SDOptimizer::ID = 0; |
129 | } // end anonymous namespace |
130 | |
131 | // Returns true if this is a use of a SPR register. |
132 | bool A15SDOptimizer::usesRegClass(MachineOperand &MO, |
133 | const TargetRegisterClass *TRC) { |
134 | if (!MO.isReg()) |
135 | return false; |
136 | Register Reg = MO.getReg(); |
137 | |
138 | if (Reg.isVirtual()) |
139 | return MRI->getRegClass(Reg)->hasSuperClassEq(RC: TRC); |
140 | else |
141 | return TRC->contains(Reg); |
142 | } |
143 | |
144 | unsigned A15SDOptimizer::getDPRLaneFromSPR(unsigned SReg) { |
145 | unsigned DReg = TRI->getMatchingSuperReg(Reg: SReg, SubIdx: ARM::ssub_1, |
146 | RC: &ARM::DPRRegClass); |
147 | if (DReg != ARM::NoRegister) return ARM::ssub_1; |
148 | return ARM::ssub_0; |
149 | } |
150 | |
151 | // Get the subreg type that is most likely to be coalesced |
152 | // for an SPR register that will be used in VDUP32d pseudo. |
153 | unsigned A15SDOptimizer::getPrefSPRLane(unsigned SReg) { |
154 | if (!Register::isVirtualRegister(Reg: SReg)) |
155 | return getDPRLaneFromSPR(SReg); |
156 | |
157 | MachineInstr *MI = MRI->getVRegDef(Reg: SReg); |
158 | if (!MI) return ARM::ssub_0; |
159 | MachineOperand *MO = MI->findRegisterDefOperand(Reg: SReg, /*TRI=*/nullptr); |
160 | if (!MO) return ARM::ssub_0; |
161 | assert(MO->isReg() && "Non-register operand found!" ); |
162 | |
163 | if (MI->isCopy() && usesRegClass(MO&: MI->getOperand(i: 1), |
164 | TRC: &ARM::SPRRegClass)) { |
165 | SReg = MI->getOperand(i: 1).getReg(); |
166 | } |
167 | |
168 | if (Register::isVirtualRegister(Reg: SReg)) { |
169 | if (MO->getSubReg() == ARM::ssub_1) return ARM::ssub_1; |
170 | return ARM::ssub_0; |
171 | } |
172 | return getDPRLaneFromSPR(SReg); |
173 | } |
174 | |
175 | // MI is known to be dead. Figure out what instructions |
176 | // are also made dead by this and mark them for removal. |
177 | void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) { |
178 | SmallVector<MachineInstr *, 8> Front; |
179 | DeadInstr.insert(x: MI); |
180 | |
181 | LLVM_DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n" ); |
182 | Front.push_back(Elt: MI); |
183 | |
184 | while (Front.size() != 0) { |
185 | MI = Front.pop_back_val(); |
186 | |
187 | // MI is already known to be dead. We need to see |
188 | // if other instructions can also be removed. |
189 | for (MachineOperand &MO : MI->operands()) { |
190 | if ((!MO.isReg()) || (!MO.isUse())) |
191 | continue; |
192 | Register Reg = MO.getReg(); |
193 | if (!Reg.isVirtual()) |
194 | continue; |
195 | MachineOperand *Op = MI->findRegisterDefOperand(Reg, /*TRI=*/nullptr); |
196 | |
197 | if (!Op) |
198 | continue; |
199 | |
200 | MachineInstr *Def = Op->getParent(); |
201 | |
202 | // We don't need to do anything if we have already marked |
203 | // this instruction as being dead. |
204 | if (DeadInstr.find(x: Def) != DeadInstr.end()) |
205 | continue; |
206 | |
207 | // Check if all the uses of this instruction are marked as |
208 | // dead. If so, we can also mark this instruction as being |
209 | // dead. |
210 | bool IsDead = true; |
211 | for (MachineOperand &MODef : Def->operands()) { |
212 | if ((!MODef.isReg()) || (!MODef.isDef())) |
213 | continue; |
214 | Register DefReg = MODef.getReg(); |
215 | if (!DefReg.isVirtual()) { |
216 | IsDead = false; |
217 | break; |
218 | } |
219 | for (MachineInstr &Use : MRI->use_instructions(Reg)) { |
220 | // We don't care about self references. |
221 | if (&Use == Def) |
222 | continue; |
223 | if (DeadInstr.find(x: &Use) == DeadInstr.end()) { |
224 | IsDead = false; |
225 | break; |
226 | } |
227 | } |
228 | } |
229 | |
230 | if (!IsDead) continue; |
231 | |
232 | LLVM_DEBUG(dbgs() << "Deleting instruction " << *Def << "\n" ); |
233 | DeadInstr.insert(x: Def); |
234 | } |
235 | } |
236 | } |
237 | |
238 | // Creates the more optimized patterns and generally does all the code |
239 | // transformations in this pass. |
240 | unsigned A15SDOptimizer::optimizeSDPattern(MachineInstr *MI) { |
241 | if (MI->isCopy()) { |
242 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 1).getReg()); |
243 | } |
244 | |
245 | if (MI->isInsertSubreg()) { |
246 | Register DPRReg = MI->getOperand(i: 1).getReg(); |
247 | Register SPRReg = MI->getOperand(i: 2).getReg(); |
248 | |
249 | if (DPRReg.isVirtual() && SPRReg.isVirtual()) { |
250 | MachineInstr *DPRMI = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
251 | MachineInstr *SPRMI = MRI->getVRegDef(Reg: MI->getOperand(i: 2).getReg()); |
252 | |
253 | if (DPRMI && SPRMI) { |
254 | // See if the first operand of this insert_subreg is IMPLICIT_DEF |
255 | MachineInstr *ECDef = elideCopies(MI: DPRMI); |
256 | if (ECDef && ECDef->isImplicitDef()) { |
257 | // Another corner case - if we're inserting something that is purely |
258 | // a subreg copy of a DPR, just use that DPR. |
259 | |
260 | MachineInstr *EC = elideCopies(MI: SPRMI); |
261 | // Is it a subreg copy of ssub_0? |
262 | if (EC && EC->isCopy() && |
263 | EC->getOperand(i: 1).getSubReg() == ARM::ssub_0) { |
264 | LLVM_DEBUG(dbgs() << "Found a subreg copy: " << *SPRMI); |
265 | |
266 | // Find the thing we're subreg copying out of - is it of the same |
267 | // regclass as DPRMI? (i.e. a DPR or QPR). |
268 | Register FullReg = SPRMI->getOperand(i: 1).getReg(); |
269 | const TargetRegisterClass *TRC = |
270 | MRI->getRegClass(Reg: MI->getOperand(i: 1).getReg()); |
271 | if (TRC->hasSuperClassEq(RC: MRI->getRegClass(Reg: FullReg))) { |
272 | LLVM_DEBUG(dbgs() << "Subreg copy is compatible - returning " ); |
273 | LLVM_DEBUG(dbgs() << printReg(FullReg) << "\n" ); |
274 | eraseInstrWithNoUses(MI); |
275 | return FullReg; |
276 | } |
277 | } |
278 | |
279 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 2).getReg()); |
280 | } |
281 | } |
282 | } |
283 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 0).getReg()); |
284 | } |
285 | |
286 | if (MI->isRegSequence() && usesRegClass(MO&: MI->getOperand(i: 1), |
287 | TRC: &ARM::SPRRegClass)) { |
288 | // See if all bar one of the operands are IMPLICIT_DEF and insert the |
289 | // optimizer pattern accordingly. |
290 | unsigned NumImplicit = 0, NumTotal = 0; |
291 | unsigned NonImplicitReg = ~0U; |
292 | |
293 | for (MachineOperand &MO : llvm::drop_begin(RangeOrContainer: MI->explicit_operands())) { |
294 | if (!MO.isReg()) |
295 | continue; |
296 | ++NumTotal; |
297 | Register OpReg = MO.getReg(); |
298 | |
299 | if (!OpReg.isVirtual()) |
300 | break; |
301 | |
302 | MachineInstr *Def = MRI->getVRegDef(Reg: OpReg); |
303 | if (!Def) |
304 | break; |
305 | if (Def->isImplicitDef()) |
306 | ++NumImplicit; |
307 | else |
308 | NonImplicitReg = MO.getReg(); |
309 | } |
310 | |
311 | if (NumImplicit == NumTotal - 1) |
312 | return optimizeAllLanesPattern(MI, Reg: NonImplicitReg); |
313 | else |
314 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 0).getReg()); |
315 | } |
316 | |
317 | llvm_unreachable("Unhandled update pattern!" ); |
318 | } |
319 | |
320 | // Return true if this MachineInstr inserts a scalar (SPR) value into |
321 | // a D or Q register. |
322 | bool A15SDOptimizer::hasPartialWrite(MachineInstr *MI) { |
323 | // The only way we can do a partial register update is through a COPY, |
324 | // INSERT_SUBREG or REG_SEQUENCE. |
325 | if (MI->isCopy() && usesRegClass(MO&: MI->getOperand(i: 1), TRC: &ARM::SPRRegClass)) |
326 | return true; |
327 | |
328 | if (MI->isInsertSubreg() && usesRegClass(MO&: MI->getOperand(i: 2), |
329 | TRC: &ARM::SPRRegClass)) |
330 | return true; |
331 | |
332 | if (MI->isRegSequence() && usesRegClass(MO&: MI->getOperand(i: 1), TRC: &ARM::SPRRegClass)) |
333 | return true; |
334 | |
335 | return false; |
336 | } |
337 | |
338 | // Looks through full copies to get the instruction that defines the input |
339 | // operand for MI. |
340 | MachineInstr *A15SDOptimizer::elideCopies(MachineInstr *MI) { |
341 | if (!MI->isFullCopy()) |
342 | return MI; |
343 | if (!MI->getOperand(i: 1).getReg().isVirtual()) |
344 | return nullptr; |
345 | MachineInstr *Def = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
346 | if (!Def) |
347 | return nullptr; |
348 | return elideCopies(MI: Def); |
349 | } |
350 | |
351 | // Look through full copies and PHIs to get the set of non-copy MachineInstrs |
352 | // that can produce MI. |
353 | void A15SDOptimizer::elideCopiesAndPHIs(MachineInstr *MI, |
354 | SmallVectorImpl<MachineInstr*> &Outs) { |
355 | // Looking through PHIs may create loops so we need to track what |
356 | // instructions we have visited before. |
357 | std::set<MachineInstr *> Reached; |
358 | SmallVector<MachineInstr *, 8> Front; |
359 | Front.push_back(Elt: MI); |
360 | while (Front.size() != 0) { |
361 | MI = Front.pop_back_val(); |
362 | |
363 | // If we have already explored this MachineInstr, ignore it. |
364 | if (!Reached.insert(x: MI).second) |
365 | continue; |
366 | if (MI->isPHI()) { |
367 | for (unsigned I = 1, E = MI->getNumOperands(); I != E; I += 2) { |
368 | Register Reg = MI->getOperand(i: I).getReg(); |
369 | if (!Reg.isVirtual()) { |
370 | continue; |
371 | } |
372 | MachineInstr *NewMI = MRI->getVRegDef(Reg); |
373 | if (!NewMI) |
374 | continue; |
375 | Front.push_back(Elt: NewMI); |
376 | } |
377 | } else if (MI->isFullCopy()) { |
378 | if (!MI->getOperand(i: 1).getReg().isVirtual()) |
379 | continue; |
380 | MachineInstr *NewMI = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
381 | if (!NewMI) |
382 | continue; |
383 | Front.push_back(Elt: NewMI); |
384 | } else { |
385 | LLVM_DEBUG(dbgs() << "Found partial copy" << *MI << "\n" ); |
386 | Outs.push_back(Elt: MI); |
387 | } |
388 | } |
389 | } |
390 | |
391 | // Return the DPR virtual registers that are read by this machine instruction |
392 | // (if any). |
393 | SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) { |
394 | if (MI->isCopyLike() || MI->isInsertSubreg() || MI->isRegSequence() || |
395 | MI->isKill()) |
396 | return SmallVector<unsigned, 8>(); |
397 | |
398 | SmallVector<unsigned, 8> Defs; |
399 | for (MachineOperand &MO : MI->operands()) { |
400 | if (!MO.isReg() || !MO.isUse()) |
401 | continue; |
402 | if (!usesRegClass(MO, TRC: &ARM::DPRRegClass) && |
403 | !usesRegClass(MO, TRC: &ARM::QPRRegClass) && |
404 | !usesRegClass(MO, TRC: &ARM::DPairRegClass)) // Treat DPair as QPR |
405 | continue; |
406 | |
407 | Defs.push_back(Elt: MO.getReg()); |
408 | } |
409 | return Defs; |
410 | } |
411 | |
412 | // Creates a DPR register from an SPR one by using a VDUP. |
413 | unsigned A15SDOptimizer::createDupLane(MachineBasicBlock &MBB, |
414 | MachineBasicBlock::iterator InsertBefore, |
415 | const DebugLoc &DL, unsigned Reg, |
416 | unsigned Lane, bool QPR) { |
417 | Register Out = |
418 | MRI->createVirtualRegister(RegClass: QPR ? &ARM::QPRRegClass : &ARM::DPRRegClass); |
419 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DL, |
420 | MCID: TII->get(Opcode: QPR ? ARM::VDUPLN32q : ARM::VDUPLN32d), DestReg: Out) |
421 | .addReg(RegNo: Reg) |
422 | .addImm(Val: Lane) |
423 | .add(MOs: predOps(Pred: ARMCC::AL)); |
424 | |
425 | return Out; |
426 | } |
427 | |
428 | // Creates a SPR register from a DPR by copying the value in lane 0. |
429 | unsigned A15SDOptimizer::( |
430 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
431 | const DebugLoc &DL, unsigned DReg, unsigned Lane, |
432 | const TargetRegisterClass *TRC) { |
433 | Register Out = MRI->createVirtualRegister(RegClass: TRC); |
434 | BuildMI(BB&: MBB, |
435 | I: InsertBefore, |
436 | MIMD: DL, |
437 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: Out) |
438 | .addReg(RegNo: DReg, flags: 0, SubReg: Lane); |
439 | |
440 | return Out; |
441 | } |
442 | |
443 | // Takes two SPR registers and creates a DPR by using a REG_SEQUENCE. |
444 | unsigned A15SDOptimizer::createRegSequence( |
445 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
446 | const DebugLoc &DL, unsigned Reg1, unsigned Reg2) { |
447 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::QPRRegClass); |
448 | BuildMI(BB&: MBB, |
449 | I: InsertBefore, |
450 | MIMD: DL, |
451 | MCID: TII->get(Opcode: TargetOpcode::REG_SEQUENCE), DestReg: Out) |
452 | .addReg(RegNo: Reg1) |
453 | .addImm(Val: ARM::dsub_0) |
454 | .addReg(RegNo: Reg2) |
455 | .addImm(Val: ARM::dsub_1); |
456 | return Out; |
457 | } |
458 | |
459 | // Takes two DPR registers that have previously been VDUPed (Ssub0 and Ssub1) |
460 | // and merges them into one DPR register. |
461 | unsigned A15SDOptimizer::createVExt(MachineBasicBlock &MBB, |
462 | MachineBasicBlock::iterator InsertBefore, |
463 | const DebugLoc &DL, unsigned Ssub0, |
464 | unsigned Ssub1) { |
465 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPRRegClass); |
466 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DL, MCID: TII->get(Opcode: ARM::VEXTd32), DestReg: Out) |
467 | .addReg(RegNo: Ssub0) |
468 | .addReg(RegNo: Ssub1) |
469 | .addImm(Val: 1) |
470 | .add(MOs: predOps(Pred: ARMCC::AL)); |
471 | return Out; |
472 | } |
473 | |
474 | unsigned A15SDOptimizer::createInsertSubreg( |
475 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
476 | const DebugLoc &DL, unsigned DReg, unsigned Lane, unsigned ToInsert) { |
477 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPR_VFP2RegClass); |
478 | BuildMI(BB&: MBB, |
479 | I: InsertBefore, |
480 | MIMD: DL, |
481 | MCID: TII->get(Opcode: TargetOpcode::INSERT_SUBREG), DestReg: Out) |
482 | .addReg(RegNo: DReg) |
483 | .addReg(RegNo: ToInsert) |
484 | .addImm(Val: Lane); |
485 | |
486 | return Out; |
487 | } |
488 | |
489 | unsigned |
490 | A15SDOptimizer::createImplicitDef(MachineBasicBlock &MBB, |
491 | MachineBasicBlock::iterator InsertBefore, |
492 | const DebugLoc &DL) { |
493 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPRRegClass); |
494 | BuildMI(BB&: MBB, |
495 | I: InsertBefore, |
496 | MIMD: DL, |
497 | MCID: TII->get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: Out); |
498 | return Out; |
499 | } |
500 | |
501 | // This function inserts instructions in order to optimize interactions between |
502 | // SPR registers and DPR/QPR registers. It does so by performing VDUPs on all |
503 | // lanes, and the using VEXT instructions to recompose the result. |
504 | unsigned |
505 | A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) { |
506 | MachineBasicBlock::iterator InsertPt(MI); |
507 | DebugLoc DL = MI->getDebugLoc(); |
508 | MachineBasicBlock &MBB = *MI->getParent(); |
509 | InsertPt++; |
510 | unsigned Out; |
511 | |
512 | // DPair has the same length as QPR and also has two DPRs as subreg. |
513 | // Treat DPair as QPR. |
514 | if (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::QPRRegClass) || |
515 | MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::DPairRegClass)) { |
516 | unsigned DSub0 = createExtractSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Reg, |
517 | Lane: ARM::dsub_0, TRC: &ARM::DPRRegClass); |
518 | unsigned DSub1 = createExtractSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Reg, |
519 | Lane: ARM::dsub_1, TRC: &ARM::DPRRegClass); |
520 | |
521 | unsigned Out1 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub0, Lane: 0); |
522 | unsigned Out2 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub0, Lane: 1); |
523 | Out = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out1, Ssub1: Out2); |
524 | |
525 | unsigned Out3 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub1, Lane: 0); |
526 | unsigned Out4 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub1, Lane: 1); |
527 | Out2 = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out3, Ssub1: Out4); |
528 | |
529 | Out = createRegSequence(MBB, InsertBefore: InsertPt, DL, Reg1: Out, Reg2: Out2); |
530 | |
531 | } else if (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::DPRRegClass)) { |
532 | unsigned Out1 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg, Lane: 0); |
533 | unsigned Out2 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg, Lane: 1); |
534 | Out = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out1, Ssub1: Out2); |
535 | |
536 | } else { |
537 | assert(MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::SPRRegClass) && |
538 | "Found unexpected regclass!" ); |
539 | |
540 | unsigned PrefLane = getPrefSPRLane(SReg: Reg); |
541 | unsigned Lane; |
542 | switch (PrefLane) { |
543 | case ARM::ssub_0: Lane = 0; break; |
544 | case ARM::ssub_1: Lane = 1; break; |
545 | default: llvm_unreachable("Unknown preferred lane!" ); |
546 | } |
547 | |
548 | // Treat DPair as QPR |
549 | bool UsesQPR = usesRegClass(MO&: MI->getOperand(i: 0), TRC: &ARM::QPRRegClass) || |
550 | usesRegClass(MO&: MI->getOperand(i: 0), TRC: &ARM::DPairRegClass); |
551 | |
552 | Out = createImplicitDef(MBB, InsertBefore: InsertPt, DL); |
553 | Out = createInsertSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Out, Lane: PrefLane, ToInsert: Reg); |
554 | Out = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: Out, Lane, QPR: UsesQPR); |
555 | eraseInstrWithNoUses(MI); |
556 | } |
557 | return Out; |
558 | } |
559 | |
560 | bool A15SDOptimizer::runOnInstruction(MachineInstr *MI) { |
561 | // We look for instructions that write S registers that are then read as |
562 | // D/Q registers. These can only be caused by COPY, INSERT_SUBREG and |
563 | // REG_SEQUENCE pseudos that insert an SPR value into a DPR register or |
564 | // merge two SPR values to form a DPR register. In order avoid false |
565 | // positives we make sure that there is an SPR producer so we look past |
566 | // COPY and PHI nodes to find it. |
567 | // |
568 | // The best code pattern for when an SPR producer is going to be used by a |
569 | // DPR or QPR consumer depends on whether the other lanes of the |
570 | // corresponding DPR/QPR are currently defined. |
571 | // |
572 | // We can handle these efficiently, depending on the type of |
573 | // pseudo-instruction that is producing the pattern |
574 | // |
575 | // * COPY: * VDUP all lanes and merge the results together |
576 | // using VEXTs. |
577 | // |
578 | // * INSERT_SUBREG: * If the SPR value was originally in another DPR/QPR |
579 | // lane, and the other lane(s) of the DPR/QPR register |
580 | // that we are inserting in are undefined, use the |
581 | // original DPR/QPR value. |
582 | // * Otherwise, fall back on the same stategy as COPY. |
583 | // |
584 | // * REG_SEQUENCE: * If all except one of the input operands are |
585 | // IMPLICIT_DEFs, insert the VDUP pattern for just the |
586 | // defined input operand |
587 | // * Otherwise, fall back on the same stategy as COPY. |
588 | // |
589 | |
590 | // First, get all the reads of D-registers done by this instruction. |
591 | SmallVector<unsigned, 8> Defs = getReadDPRs(MI); |
592 | bool Modified = false; |
593 | |
594 | for (unsigned I : Defs) { |
595 | // Follow the def-use chain for this DPR through COPYs, and also through |
596 | // PHIs (which are essentially multi-way COPYs). It is because of PHIs that |
597 | // we can end up with multiple defs of this DPR. |
598 | |
599 | SmallVector<MachineInstr *, 8> DefSrcs; |
600 | if (!Register::isVirtualRegister(Reg: I)) |
601 | continue; |
602 | MachineInstr *Def = MRI->getVRegDef(Reg: I); |
603 | if (!Def) |
604 | continue; |
605 | |
606 | elideCopiesAndPHIs(MI: Def, Outs&: DefSrcs); |
607 | |
608 | for (MachineInstr *MI : DefSrcs) { |
609 | // If we've already analyzed and replaced this operand, don't do |
610 | // anything. |
611 | if (Replacements.find(x: MI) != Replacements.end()) |
612 | continue; |
613 | |
614 | // Now, work out if the instruction causes a SPR->DPR dependency. |
615 | if (!hasPartialWrite(MI)) |
616 | continue; |
617 | |
618 | // Collect all the uses of this MI's DPR def for updating later. |
619 | SmallVector<MachineOperand*, 8> Uses; |
620 | Register DPRDefReg = MI->getOperand(i: 0).getReg(); |
621 | for (MachineOperand &MO : MRI->use_operands(Reg: DPRDefReg)) |
622 | Uses.push_back(Elt: &MO); |
623 | |
624 | // We can optimize this. |
625 | unsigned NewReg = optimizeSDPattern(MI); |
626 | |
627 | if (NewReg != 0) { |
628 | Modified = true; |
629 | for (MachineOperand *Use : Uses) { |
630 | // Make sure to constrain the register class of the new register to |
631 | // match what we're replacing. Otherwise we can optimize a DPR_VFP2 |
632 | // reference into a plain DPR, and that will end poorly. NewReg is |
633 | // always virtual here, so there will always be a matching subclass |
634 | // to find. |
635 | MRI->constrainRegClass(Reg: NewReg, RC: MRI->getRegClass(Reg: Use->getReg())); |
636 | |
637 | LLVM_DEBUG(dbgs() << "Replacing operand " << *Use << " with " |
638 | << printReg(NewReg) << "\n" ); |
639 | Use->substVirtReg(Reg: NewReg, SubIdx: 0, *TRI); |
640 | } |
641 | } |
642 | Replacements[MI] = NewReg; |
643 | } |
644 | } |
645 | return Modified; |
646 | } |
647 | |
648 | bool A15SDOptimizer::runOnMachineFunction(MachineFunction &Fn) { |
649 | if (skipFunction(F: Fn.getFunction())) |
650 | return false; |
651 | |
652 | const ARMSubtarget &STI = Fn.getSubtarget<ARMSubtarget>(); |
653 | // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be |
654 | // enabled when NEON is available. |
655 | if (!(STI.useSplatVFPToNeon() && STI.hasNEON())) |
656 | return false; |
657 | |
658 | TII = STI.getInstrInfo(); |
659 | TRI = STI.getRegisterInfo(); |
660 | MRI = &Fn.getRegInfo(); |
661 | bool Modified = false; |
662 | |
663 | LLVM_DEBUG(dbgs() << "Running on function " << Fn.getName() << "\n" ); |
664 | |
665 | DeadInstr.clear(); |
666 | Replacements.clear(); |
667 | |
668 | for (MachineBasicBlock &MBB : Fn) { |
669 | for (MachineInstr &MI : MBB) { |
670 | Modified |= runOnInstruction(MI: &MI); |
671 | } |
672 | } |
673 | |
674 | for (MachineInstr *MI : DeadInstr) { |
675 | MI->eraseFromParent(); |
676 | } |
677 | |
678 | return Modified; |
679 | } |
680 | |
681 | FunctionPass *llvm::createA15SDOptimizerPass() { |
682 | return new A15SDOptimizer(); |
683 | } |
684 | |