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 | MCRegister DReg = |
146 | TRI->getMatchingSuperReg(Reg: SReg, SubIdx: ARM::ssub_1, RC: &ARM::DPRRegClass); |
147 | if (DReg) |
148 | return ARM::ssub_1; |
149 | return ARM::ssub_0; |
150 | } |
151 | |
152 | // Get the subreg type that is most likely to be coalesced |
153 | // for an SPR register that will be used in VDUP32d pseudo. |
154 | unsigned A15SDOptimizer::getPrefSPRLane(unsigned SReg) { |
155 | if (!Register::isVirtualRegister(Reg: SReg)) |
156 | return getDPRLaneFromSPR(SReg); |
157 | |
158 | MachineInstr *MI = MRI->getVRegDef(Reg: SReg); |
159 | if (!MI) return ARM::ssub_0; |
160 | MachineOperand *MO = MI->findRegisterDefOperand(Reg: SReg, /*TRI=*/nullptr); |
161 | if (!MO) return ARM::ssub_0; |
162 | assert(MO->isReg() && "Non-register operand found!" ); |
163 | |
164 | if (MI->isCopy() && usesRegClass(MO&: MI->getOperand(i: 1), |
165 | TRC: &ARM::SPRRegClass)) { |
166 | SReg = MI->getOperand(i: 1).getReg(); |
167 | } |
168 | |
169 | if (Register::isVirtualRegister(Reg: SReg)) { |
170 | if (MO->getSubReg() == ARM::ssub_1) return ARM::ssub_1; |
171 | return ARM::ssub_0; |
172 | } |
173 | return getDPRLaneFromSPR(SReg); |
174 | } |
175 | |
176 | // MI is known to be dead. Figure out what instructions |
177 | // are also made dead by this and mark them for removal. |
178 | void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) { |
179 | SmallVector<MachineInstr *, 8> Front; |
180 | DeadInstr.insert(x: MI); |
181 | |
182 | LLVM_DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n" ); |
183 | Front.push_back(Elt: MI); |
184 | |
185 | while (Front.size() != 0) { |
186 | MI = Front.pop_back_val(); |
187 | |
188 | // MI is already known to be dead. We need to see |
189 | // if other instructions can also be removed. |
190 | for (MachineOperand &MO : MI->operands()) { |
191 | if ((!MO.isReg()) || (!MO.isUse())) |
192 | continue; |
193 | Register Reg = MO.getReg(); |
194 | if (!Reg.isVirtual()) |
195 | continue; |
196 | MachineOperand *Op = MI->findRegisterDefOperand(Reg, /*TRI=*/nullptr); |
197 | |
198 | if (!Op) |
199 | continue; |
200 | |
201 | MachineInstr *Def = Op->getParent(); |
202 | |
203 | // We don't need to do anything if we have already marked |
204 | // this instruction as being dead. |
205 | if (DeadInstr.find(x: Def) != DeadInstr.end()) |
206 | continue; |
207 | |
208 | // Check if all the uses of this instruction are marked as |
209 | // dead. If so, we can also mark this instruction as being |
210 | // dead. |
211 | bool IsDead = true; |
212 | for (MachineOperand &MODef : Def->operands()) { |
213 | if ((!MODef.isReg()) || (!MODef.isDef())) |
214 | continue; |
215 | Register DefReg = MODef.getReg(); |
216 | if (!DefReg.isVirtual()) { |
217 | IsDead = false; |
218 | break; |
219 | } |
220 | for (MachineInstr &Use : MRI->use_instructions(Reg)) { |
221 | // We don't care about self references. |
222 | if (&Use == Def) |
223 | continue; |
224 | if (DeadInstr.find(x: &Use) == DeadInstr.end()) { |
225 | IsDead = false; |
226 | break; |
227 | } |
228 | } |
229 | } |
230 | |
231 | if (!IsDead) continue; |
232 | |
233 | LLVM_DEBUG(dbgs() << "Deleting instruction " << *Def << "\n" ); |
234 | DeadInstr.insert(x: Def); |
235 | } |
236 | } |
237 | } |
238 | |
239 | // Creates the more optimized patterns and generally does all the code |
240 | // transformations in this pass. |
241 | unsigned A15SDOptimizer::optimizeSDPattern(MachineInstr *MI) { |
242 | if (MI->isCopy()) { |
243 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 1).getReg()); |
244 | } |
245 | |
246 | if (MI->isInsertSubreg()) { |
247 | Register DPRReg = MI->getOperand(i: 1).getReg(); |
248 | Register SPRReg = MI->getOperand(i: 2).getReg(); |
249 | |
250 | if (DPRReg.isVirtual() && SPRReg.isVirtual()) { |
251 | MachineInstr *DPRMI = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
252 | MachineInstr *SPRMI = MRI->getVRegDef(Reg: MI->getOperand(i: 2).getReg()); |
253 | |
254 | if (DPRMI && SPRMI) { |
255 | // See if the first operand of this insert_subreg is IMPLICIT_DEF |
256 | MachineInstr *ECDef = elideCopies(MI: DPRMI); |
257 | if (ECDef && ECDef->isImplicitDef()) { |
258 | // Another corner case - if we're inserting something that is purely |
259 | // a subreg copy of a DPR, just use that DPR. |
260 | |
261 | MachineInstr *EC = elideCopies(MI: SPRMI); |
262 | // Is it a subreg copy of ssub_0? |
263 | if (EC && EC->isCopy() && |
264 | EC->getOperand(i: 1).getSubReg() == ARM::ssub_0) { |
265 | LLVM_DEBUG(dbgs() << "Found a subreg copy: " << *SPRMI); |
266 | |
267 | // Find the thing we're subreg copying out of - is it of the same |
268 | // regclass as DPRMI? (i.e. a DPR or QPR). |
269 | Register FullReg = SPRMI->getOperand(i: 1).getReg(); |
270 | const TargetRegisterClass *TRC = |
271 | MRI->getRegClass(Reg: MI->getOperand(i: 1).getReg()); |
272 | if (TRC->hasSuperClassEq(RC: MRI->getRegClass(Reg: FullReg))) { |
273 | LLVM_DEBUG(dbgs() << "Subreg copy is compatible - returning " ); |
274 | LLVM_DEBUG(dbgs() << printReg(FullReg) << "\n" ); |
275 | eraseInstrWithNoUses(MI); |
276 | return FullReg; |
277 | } |
278 | } |
279 | |
280 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 2).getReg()); |
281 | } |
282 | } |
283 | } |
284 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 0).getReg()); |
285 | } |
286 | |
287 | if (MI->isRegSequence() && usesRegClass(MO&: MI->getOperand(i: 1), |
288 | TRC: &ARM::SPRRegClass)) { |
289 | // See if all bar one of the operands are IMPLICIT_DEF and insert the |
290 | // optimizer pattern accordingly. |
291 | unsigned NumImplicit = 0, NumTotal = 0; |
292 | unsigned NonImplicitReg = ~0U; |
293 | |
294 | for (MachineOperand &MO : llvm::drop_begin(RangeOrContainer: MI->explicit_operands())) { |
295 | if (!MO.isReg()) |
296 | continue; |
297 | ++NumTotal; |
298 | Register OpReg = MO.getReg(); |
299 | |
300 | if (!OpReg.isVirtual()) |
301 | break; |
302 | |
303 | MachineInstr *Def = MRI->getVRegDef(Reg: OpReg); |
304 | if (!Def) |
305 | break; |
306 | if (Def->isImplicitDef()) |
307 | ++NumImplicit; |
308 | else |
309 | NonImplicitReg = MO.getReg(); |
310 | } |
311 | |
312 | if (NumImplicit == NumTotal - 1) |
313 | return optimizeAllLanesPattern(MI, Reg: NonImplicitReg); |
314 | else |
315 | return optimizeAllLanesPattern(MI, Reg: MI->getOperand(i: 0).getReg()); |
316 | } |
317 | |
318 | llvm_unreachable("Unhandled update pattern!" ); |
319 | } |
320 | |
321 | // Return true if this MachineInstr inserts a scalar (SPR) value into |
322 | // a D or Q register. |
323 | bool A15SDOptimizer::hasPartialWrite(MachineInstr *MI) { |
324 | // The only way we can do a partial register update is through a COPY, |
325 | // INSERT_SUBREG or REG_SEQUENCE. |
326 | if (MI->isCopy() && usesRegClass(MO&: MI->getOperand(i: 1), TRC: &ARM::SPRRegClass)) |
327 | return true; |
328 | |
329 | if (MI->isInsertSubreg() && usesRegClass(MO&: MI->getOperand(i: 2), |
330 | TRC: &ARM::SPRRegClass)) |
331 | return true; |
332 | |
333 | if (MI->isRegSequence() && usesRegClass(MO&: MI->getOperand(i: 1), TRC: &ARM::SPRRegClass)) |
334 | return true; |
335 | |
336 | return false; |
337 | } |
338 | |
339 | // Looks through full copies to get the instruction that defines the input |
340 | // operand for MI. |
341 | MachineInstr *A15SDOptimizer::elideCopies(MachineInstr *MI) { |
342 | if (!MI->isFullCopy()) |
343 | return MI; |
344 | if (!MI->getOperand(i: 1).getReg().isVirtual()) |
345 | return nullptr; |
346 | MachineInstr *Def = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
347 | if (!Def) |
348 | return nullptr; |
349 | return elideCopies(MI: Def); |
350 | } |
351 | |
352 | // Look through full copies and PHIs to get the set of non-copy MachineInstrs |
353 | // that can produce MI. |
354 | void A15SDOptimizer::elideCopiesAndPHIs(MachineInstr *MI, |
355 | SmallVectorImpl<MachineInstr*> &Outs) { |
356 | // Looking through PHIs may create loops so we need to track what |
357 | // instructions we have visited before. |
358 | std::set<MachineInstr *> Reached; |
359 | SmallVector<MachineInstr *, 8> Front; |
360 | Front.push_back(Elt: MI); |
361 | while (Front.size() != 0) { |
362 | MI = Front.pop_back_val(); |
363 | |
364 | // If we have already explored this MachineInstr, ignore it. |
365 | if (!Reached.insert(x: MI).second) |
366 | continue; |
367 | if (MI->isPHI()) { |
368 | for (unsigned I = 1, E = MI->getNumOperands(); I != E; I += 2) { |
369 | Register Reg = MI->getOperand(i: I).getReg(); |
370 | if (!Reg.isVirtual()) { |
371 | continue; |
372 | } |
373 | MachineInstr *NewMI = MRI->getVRegDef(Reg); |
374 | if (!NewMI) |
375 | continue; |
376 | Front.push_back(Elt: NewMI); |
377 | } |
378 | } else if (MI->isFullCopy()) { |
379 | if (!MI->getOperand(i: 1).getReg().isVirtual()) |
380 | continue; |
381 | MachineInstr *NewMI = MRI->getVRegDef(Reg: MI->getOperand(i: 1).getReg()); |
382 | if (!NewMI) |
383 | continue; |
384 | Front.push_back(Elt: NewMI); |
385 | } else { |
386 | LLVM_DEBUG(dbgs() << "Found partial copy" << *MI << "\n" ); |
387 | Outs.push_back(Elt: MI); |
388 | } |
389 | } |
390 | } |
391 | |
392 | // Return the DPR virtual registers that are read by this machine instruction |
393 | // (if any). |
394 | SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) { |
395 | if (MI->isCopyLike() || MI->isInsertSubreg() || MI->isRegSequence() || |
396 | MI->isKill()) |
397 | return SmallVector<unsigned, 8>(); |
398 | |
399 | SmallVector<unsigned, 8> Defs; |
400 | for (MachineOperand &MO : MI->operands()) { |
401 | if (!MO.isReg() || !MO.isUse()) |
402 | continue; |
403 | if (!usesRegClass(MO, TRC: &ARM::DPRRegClass) && |
404 | !usesRegClass(MO, TRC: &ARM::QPRRegClass) && |
405 | !usesRegClass(MO, TRC: &ARM::DPairRegClass)) // Treat DPair as QPR |
406 | continue; |
407 | |
408 | Defs.push_back(Elt: MO.getReg()); |
409 | } |
410 | return Defs; |
411 | } |
412 | |
413 | // Creates a DPR register from an SPR one by using a VDUP. |
414 | unsigned A15SDOptimizer::createDupLane(MachineBasicBlock &MBB, |
415 | MachineBasicBlock::iterator InsertBefore, |
416 | const DebugLoc &DL, unsigned Reg, |
417 | unsigned Lane, bool QPR) { |
418 | Register Out = |
419 | MRI->createVirtualRegister(RegClass: QPR ? &ARM::QPRRegClass : &ARM::DPRRegClass); |
420 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DL, |
421 | MCID: TII->get(Opcode: QPR ? ARM::VDUPLN32q : ARM::VDUPLN32d), DestReg: Out) |
422 | .addReg(RegNo: Reg) |
423 | .addImm(Val: Lane) |
424 | .add(MOs: predOps(Pred: ARMCC::AL)); |
425 | |
426 | return Out; |
427 | } |
428 | |
429 | // Creates a SPR register from a DPR by copying the value in lane 0. |
430 | unsigned A15SDOptimizer::( |
431 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
432 | const DebugLoc &DL, unsigned DReg, unsigned Lane, |
433 | const TargetRegisterClass *TRC) { |
434 | Register Out = MRI->createVirtualRegister(RegClass: TRC); |
435 | BuildMI(BB&: MBB, |
436 | I: InsertBefore, |
437 | MIMD: DL, |
438 | MCID: TII->get(Opcode: TargetOpcode::COPY), DestReg: Out) |
439 | .addReg(RegNo: DReg, flags: 0, SubReg: Lane); |
440 | |
441 | return Out; |
442 | } |
443 | |
444 | // Takes two SPR registers and creates a DPR by using a REG_SEQUENCE. |
445 | unsigned A15SDOptimizer::createRegSequence( |
446 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
447 | const DebugLoc &DL, unsigned Reg1, unsigned Reg2) { |
448 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::QPRRegClass); |
449 | BuildMI(BB&: MBB, |
450 | I: InsertBefore, |
451 | MIMD: DL, |
452 | MCID: TII->get(Opcode: TargetOpcode::REG_SEQUENCE), DestReg: Out) |
453 | .addReg(RegNo: Reg1) |
454 | .addImm(Val: ARM::dsub_0) |
455 | .addReg(RegNo: Reg2) |
456 | .addImm(Val: ARM::dsub_1); |
457 | return Out; |
458 | } |
459 | |
460 | // Takes two DPR registers that have previously been VDUPed (Ssub0 and Ssub1) |
461 | // and merges them into one DPR register. |
462 | unsigned A15SDOptimizer::createVExt(MachineBasicBlock &MBB, |
463 | MachineBasicBlock::iterator InsertBefore, |
464 | const DebugLoc &DL, unsigned Ssub0, |
465 | unsigned Ssub1) { |
466 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPRRegClass); |
467 | BuildMI(BB&: MBB, I: InsertBefore, MIMD: DL, MCID: TII->get(Opcode: ARM::VEXTd32), DestReg: Out) |
468 | .addReg(RegNo: Ssub0) |
469 | .addReg(RegNo: Ssub1) |
470 | .addImm(Val: 1) |
471 | .add(MOs: predOps(Pred: ARMCC::AL)); |
472 | return Out; |
473 | } |
474 | |
475 | unsigned A15SDOptimizer::createInsertSubreg( |
476 | MachineBasicBlock &MBB, MachineBasicBlock::iterator InsertBefore, |
477 | const DebugLoc &DL, unsigned DReg, unsigned Lane, unsigned ToInsert) { |
478 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPR_VFP2RegClass); |
479 | BuildMI(BB&: MBB, |
480 | I: InsertBefore, |
481 | MIMD: DL, |
482 | MCID: TII->get(Opcode: TargetOpcode::INSERT_SUBREG), DestReg: Out) |
483 | .addReg(RegNo: DReg) |
484 | .addReg(RegNo: ToInsert) |
485 | .addImm(Val: Lane); |
486 | |
487 | return Out; |
488 | } |
489 | |
490 | unsigned |
491 | A15SDOptimizer::createImplicitDef(MachineBasicBlock &MBB, |
492 | MachineBasicBlock::iterator InsertBefore, |
493 | const DebugLoc &DL) { |
494 | Register Out = MRI->createVirtualRegister(RegClass: &ARM::DPRRegClass); |
495 | BuildMI(BB&: MBB, |
496 | I: InsertBefore, |
497 | MIMD: DL, |
498 | MCID: TII->get(Opcode: TargetOpcode::IMPLICIT_DEF), DestReg: Out); |
499 | return Out; |
500 | } |
501 | |
502 | // This function inserts instructions in order to optimize interactions between |
503 | // SPR registers and DPR/QPR registers. It does so by performing VDUPs on all |
504 | // lanes, and the using VEXT instructions to recompose the result. |
505 | unsigned |
506 | A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) { |
507 | MachineBasicBlock::iterator InsertPt(MI); |
508 | DebugLoc DL = MI->getDebugLoc(); |
509 | MachineBasicBlock &MBB = *MI->getParent(); |
510 | InsertPt++; |
511 | unsigned Out; |
512 | |
513 | // DPair has the same length as QPR and also has two DPRs as subreg. |
514 | // Treat DPair as QPR. |
515 | if (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::QPRRegClass) || |
516 | MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::DPairRegClass)) { |
517 | unsigned DSub0 = createExtractSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Reg, |
518 | Lane: ARM::dsub_0, TRC: &ARM::DPRRegClass); |
519 | unsigned DSub1 = createExtractSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Reg, |
520 | Lane: ARM::dsub_1, TRC: &ARM::DPRRegClass); |
521 | |
522 | unsigned Out1 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub0, Lane: 0); |
523 | unsigned Out2 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub0, Lane: 1); |
524 | Out = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out1, Ssub1: Out2); |
525 | |
526 | unsigned Out3 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub1, Lane: 0); |
527 | unsigned Out4 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: DSub1, Lane: 1); |
528 | Out2 = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out3, Ssub1: Out4); |
529 | |
530 | Out = createRegSequence(MBB, InsertBefore: InsertPt, DL, Reg1: Out, Reg2: Out2); |
531 | |
532 | } else if (MRI->getRegClass(Reg)->hasSuperClassEq(RC: &ARM::DPRRegClass)) { |
533 | unsigned Out1 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg, Lane: 0); |
534 | unsigned Out2 = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg, Lane: 1); |
535 | Out = createVExt(MBB, InsertBefore: InsertPt, DL, Ssub0: Out1, Ssub1: Out2); |
536 | |
537 | } else { |
538 | assert(MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::SPRRegClass) && |
539 | "Found unexpected regclass!" ); |
540 | |
541 | unsigned PrefLane = getPrefSPRLane(SReg: Reg); |
542 | unsigned Lane; |
543 | switch (PrefLane) { |
544 | case ARM::ssub_0: Lane = 0; break; |
545 | case ARM::ssub_1: Lane = 1; break; |
546 | default: llvm_unreachable("Unknown preferred lane!" ); |
547 | } |
548 | |
549 | // Treat DPair as QPR |
550 | bool UsesQPR = usesRegClass(MO&: MI->getOperand(i: 0), TRC: &ARM::QPRRegClass) || |
551 | usesRegClass(MO&: MI->getOperand(i: 0), TRC: &ARM::DPairRegClass); |
552 | |
553 | Out = createImplicitDef(MBB, InsertBefore: InsertPt, DL); |
554 | Out = createInsertSubreg(MBB, InsertBefore: InsertPt, DL, DReg: Out, Lane: PrefLane, ToInsert: Reg); |
555 | Out = createDupLane(MBB, InsertBefore: InsertPt, DL, Reg: Out, Lane, QPR: UsesQPR); |
556 | eraseInstrWithNoUses(MI); |
557 | } |
558 | return Out; |
559 | } |
560 | |
561 | bool A15SDOptimizer::runOnInstruction(MachineInstr *MI) { |
562 | // We look for instructions that write S registers that are then read as |
563 | // D/Q registers. These can only be caused by COPY, INSERT_SUBREG and |
564 | // REG_SEQUENCE pseudos that insert an SPR value into a DPR register or |
565 | // merge two SPR values to form a DPR register. In order avoid false |
566 | // positives we make sure that there is an SPR producer so we look past |
567 | // COPY and PHI nodes to find it. |
568 | // |
569 | // The best code pattern for when an SPR producer is going to be used by a |
570 | // DPR or QPR consumer depends on whether the other lanes of the |
571 | // corresponding DPR/QPR are currently defined. |
572 | // |
573 | // We can handle these efficiently, depending on the type of |
574 | // pseudo-instruction that is producing the pattern |
575 | // |
576 | // * COPY: * VDUP all lanes and merge the results together |
577 | // using VEXTs. |
578 | // |
579 | // * INSERT_SUBREG: * If the SPR value was originally in another DPR/QPR |
580 | // lane, and the other lane(s) of the DPR/QPR register |
581 | // that we are inserting in are undefined, use the |
582 | // original DPR/QPR value. |
583 | // * Otherwise, fall back on the same stategy as COPY. |
584 | // |
585 | // * REG_SEQUENCE: * If all except one of the input operands are |
586 | // IMPLICIT_DEFs, insert the VDUP pattern for just the |
587 | // defined input operand |
588 | // * Otherwise, fall back on the same stategy as COPY. |
589 | // |
590 | |
591 | // First, get all the reads of D-registers done by this instruction. |
592 | SmallVector<unsigned, 8> Defs = getReadDPRs(MI); |
593 | bool Modified = false; |
594 | |
595 | for (unsigned I : Defs) { |
596 | // Follow the def-use chain for this DPR through COPYs, and also through |
597 | // PHIs (which are essentially multi-way COPYs). It is because of PHIs that |
598 | // we can end up with multiple defs of this DPR. |
599 | |
600 | SmallVector<MachineInstr *, 8> DefSrcs; |
601 | if (!Register::isVirtualRegister(Reg: I)) |
602 | continue; |
603 | MachineInstr *Def = MRI->getVRegDef(Reg: I); |
604 | if (!Def) |
605 | continue; |
606 | |
607 | elideCopiesAndPHIs(MI: Def, Outs&: DefSrcs); |
608 | |
609 | for (MachineInstr *MI : DefSrcs) { |
610 | // If we've already analyzed and replaced this operand, don't do |
611 | // anything. |
612 | if (Replacements.find(x: MI) != Replacements.end()) |
613 | continue; |
614 | |
615 | // Now, work out if the instruction causes a SPR->DPR dependency. |
616 | if (!hasPartialWrite(MI)) |
617 | continue; |
618 | |
619 | // Collect all the uses of this MI's DPR def for updating later. |
620 | Register DPRDefReg = MI->getOperand(i: 0).getReg(); |
621 | SmallVector<MachineOperand *, 8> Uses( |
622 | llvm::make_pointer_range(Range: MRI->use_operands(Reg: DPRDefReg))); |
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 | |