1 /*
2 * Copyright (C) 2011 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 /* This file contains register alloction support. */
18
19 #include "mir_to_lir-inl.h"
20
21 #include "dex/compiler_ir.h"
22 #include "dex/dataflow_iterator-inl.h"
23 #include "dex/mir_graph.h"
24 #include "driver/compiler_driver.h"
25 #include "driver/dex_compilation_unit.h"
26 #include "utils/dex_cache_arrays_layout-inl.h"
27
28 namespace art {
29
30 /*
31 * Free all allocated temps in the temp pools. Note that this does
32 * not affect the "liveness" of a temp register, which will stay
33 * live until it is either explicitly killed or reallocated.
34 */
ResetRegPool()35 void Mir2Lir::ResetRegPool() {
36 for (RegisterInfo* info : tempreg_info_) {
37 info->MarkFree();
38 }
39 // Reset temp tracking sanity check.
40 if (kIsDebugBuild) {
41 live_sreg_ = INVALID_SREG;
42 }
43 }
44
RegisterInfo(RegStorage r,const ResourceMask & mask)45 Mir2Lir::RegisterInfo::RegisterInfo(RegStorage r, const ResourceMask& mask)
46 : reg_(r), is_temp_(false), wide_value_(false), dirty_(false), aliased_(false), partner_(r),
47 s_reg_(INVALID_SREG), def_use_mask_(mask), master_(this), def_start_(nullptr),
48 def_end_(nullptr), alias_chain_(nullptr) {
49 switch (r.StorageSize()) {
50 case 0: storage_mask_ = 0xffffffff; break;
51 case 4: storage_mask_ = 0x00000001; break;
52 case 8: storage_mask_ = 0x00000003; break;
53 case 16: storage_mask_ = 0x0000000f; break;
54 case 32: storage_mask_ = 0x000000ff; break;
55 case 64: storage_mask_ = 0x0000ffff; break;
56 case 128: storage_mask_ = 0xffffffff; break;
57 }
58 used_storage_ = r.Valid() ? ~storage_mask_ : storage_mask_;
59 liveness_ = used_storage_;
60 }
61
RegisterPool(Mir2Lir * m2l,ArenaAllocator * arena,const ArrayRef<const RegStorage> & core_regs,const ArrayRef<const RegStorage> & core64_regs,const ArrayRef<const RegStorage> & sp_regs,const ArrayRef<const RegStorage> & dp_regs,const ArrayRef<const RegStorage> & reserved_regs,const ArrayRef<const RegStorage> & reserved64_regs,const ArrayRef<const RegStorage> & core_temps,const ArrayRef<const RegStorage> & core64_temps,const ArrayRef<const RegStorage> & sp_temps,const ArrayRef<const RegStorage> & dp_temps)62 Mir2Lir::RegisterPool::RegisterPool(Mir2Lir* m2l, ArenaAllocator* arena,
63 const ArrayRef<const RegStorage>& core_regs,
64 const ArrayRef<const RegStorage>& core64_regs,
65 const ArrayRef<const RegStorage>& sp_regs,
66 const ArrayRef<const RegStorage>& dp_regs,
67 const ArrayRef<const RegStorage>& reserved_regs,
68 const ArrayRef<const RegStorage>& reserved64_regs,
69 const ArrayRef<const RegStorage>& core_temps,
70 const ArrayRef<const RegStorage>& core64_temps,
71 const ArrayRef<const RegStorage>& sp_temps,
72 const ArrayRef<const RegStorage>& dp_temps) :
73 core_regs_(arena->Adapter()), next_core_reg_(0),
74 core64_regs_(arena->Adapter()), next_core64_reg_(0),
75 sp_regs_(arena->Adapter()), next_sp_reg_(0),
76 dp_regs_(arena->Adapter()), next_dp_reg_(0), m2l_(m2l) {
77 // Initialize the fast lookup map.
78 m2l_->reginfo_map_.clear();
79 m2l_->reginfo_map_.resize(RegStorage::kMaxRegs, nullptr);
80
81 // Construct the register pool.
82 core_regs_.reserve(core_regs.size());
83 for (const RegStorage& reg : core_regs) {
84 RegisterInfo* info = new (arena) RegisterInfo(reg, m2l_->GetRegMaskCommon(reg));
85 m2l_->reginfo_map_[reg.GetReg()] = info;
86 core_regs_.push_back(info);
87 }
88 core64_regs_.reserve(core64_regs.size());
89 for (const RegStorage& reg : core64_regs) {
90 RegisterInfo* info = new (arena) RegisterInfo(reg, m2l_->GetRegMaskCommon(reg));
91 m2l_->reginfo_map_[reg.GetReg()] = info;
92 core64_regs_.push_back(info);
93 }
94 sp_regs_.reserve(sp_regs.size());
95 for (const RegStorage& reg : sp_regs) {
96 RegisterInfo* info = new (arena) RegisterInfo(reg, m2l_->GetRegMaskCommon(reg));
97 m2l_->reginfo_map_[reg.GetReg()] = info;
98 sp_regs_.push_back(info);
99 }
100 dp_regs_.reserve(dp_regs.size());
101 for (const RegStorage& reg : dp_regs) {
102 RegisterInfo* info = new (arena) RegisterInfo(reg, m2l_->GetRegMaskCommon(reg));
103 m2l_->reginfo_map_[reg.GetReg()] = info;
104 dp_regs_.push_back(info);
105 }
106
107 // Keep special registers from being allocated.
108 for (RegStorage reg : reserved_regs) {
109 m2l_->MarkInUse(reg);
110 }
111 for (RegStorage reg : reserved64_regs) {
112 m2l_->MarkInUse(reg);
113 }
114
115 // Mark temp regs - all others not in use can be used for promotion
116 for (RegStorage reg : core_temps) {
117 m2l_->MarkTemp(reg);
118 }
119 for (RegStorage reg : core64_temps) {
120 m2l_->MarkTemp(reg);
121 }
122 for (RegStorage reg : sp_temps) {
123 m2l_->MarkTemp(reg);
124 }
125 for (RegStorage reg : dp_temps) {
126 m2l_->MarkTemp(reg);
127 }
128
129 // Add an entry for InvalidReg with zero'd mask.
130 RegisterInfo* invalid_reg = new (arena) RegisterInfo(RegStorage::InvalidReg(), kEncodeNone);
131 m2l_->reginfo_map_[RegStorage::InvalidReg().GetReg()] = invalid_reg;
132
133 // Existence of core64 registers implies wide references.
134 if (core64_regs_.size() != 0) {
135 ref_regs_ = &core64_regs_;
136 next_ref_reg_ = &next_core64_reg_;
137 } else {
138 ref_regs_ = &core_regs_;
139 next_ref_reg_ = &next_core_reg_;
140 }
141 }
142
DumpRegPool(ArenaVector<RegisterInfo * > * regs)143 void Mir2Lir::DumpRegPool(ArenaVector<RegisterInfo*>* regs) {
144 LOG(INFO) << "================================================";
145 for (RegisterInfo* info : *regs) {
146 LOG(INFO) << StringPrintf(
147 "R[%d:%d:%c]: T:%d, U:%d, W:%d, p:%d, LV:%d, D:%d, SR:%d, DEF:%d",
148 info->GetReg().GetReg(), info->GetReg().GetRegNum(), info->GetReg().IsFloat() ? 'f' : 'c',
149 info->IsTemp(), info->InUse(), info->IsWide(), info->Partner().GetReg(), info->IsLive(),
150 info->IsDirty(), info->SReg(), info->DefStart() != nullptr);
151 }
152 LOG(INFO) << "================================================";
153 }
154
DumpCoreRegPool()155 void Mir2Lir::DumpCoreRegPool() {
156 DumpRegPool(®_pool_->core_regs_);
157 DumpRegPool(®_pool_->core64_regs_);
158 }
159
DumpFpRegPool()160 void Mir2Lir::DumpFpRegPool() {
161 DumpRegPool(®_pool_->sp_regs_);
162 DumpRegPool(®_pool_->dp_regs_);
163 }
164
DumpRegPools()165 void Mir2Lir::DumpRegPools() {
166 LOG(INFO) << "Core registers";
167 DumpCoreRegPool();
168 LOG(INFO) << "FP registers";
169 DumpFpRegPool();
170 }
171
Clobber(RegStorage reg)172 void Mir2Lir::Clobber(RegStorage reg) {
173 if (UNLIKELY(reg.IsPair())) {
174 DCHECK(!GetRegInfo(reg.GetLow())->IsAliased());
175 Clobber(reg.GetLow());
176 DCHECK(!GetRegInfo(reg.GetHigh())->IsAliased());
177 Clobber(reg.GetHigh());
178 } else {
179 RegisterInfo* info = GetRegInfo(reg);
180 if (info->IsTemp() && !info->IsDead()) {
181 if (info->GetReg().NotExactlyEquals(info->Partner())) {
182 ClobberBody(GetRegInfo(info->Partner()));
183 }
184 ClobberBody(info);
185 if (info->IsAliased()) {
186 ClobberAliases(info, info->StorageMask());
187 } else {
188 RegisterInfo* master = info->Master();
189 if (info != master) {
190 ClobberBody(info->Master());
191 ClobberAliases(info->Master(), info->StorageMask());
192 }
193 }
194 }
195 }
196 }
197
ClobberAliases(RegisterInfo * info,uint32_t clobber_mask)198 void Mir2Lir::ClobberAliases(RegisterInfo* info, uint32_t clobber_mask) {
199 for (RegisterInfo* alias = info->GetAliasChain(); alias != nullptr;
200 alias = alias->GetAliasChain()) {
201 DCHECK(!alias->IsAliased()); // Only the master should be marked as alised.
202 // Only clobber if we have overlap.
203 if ((alias->StorageMask() & clobber_mask) != 0) {
204 ClobberBody(alias);
205 }
206 }
207 }
208
209 /*
210 * Break the association between a Dalvik vreg and a physical temp register of either register
211 * class.
212 * TODO: Ideally, the public version of this code should not exist. Besides its local usage
213 * in the register utilities, is is also used by code gen routines to work around a deficiency in
214 * local register allocation, which fails to distinguish between the "in" and "out" identities
215 * of Dalvik vregs. This can result in useless register copies when the same Dalvik vreg
216 * is used both as the source and destination register of an operation in which the type
217 * changes (for example: INT_TO_FLOAT v1, v1). Revisit when improved register allocation is
218 * addressed.
219 */
ClobberSReg(int s_reg)220 void Mir2Lir::ClobberSReg(int s_reg) {
221 if (s_reg != INVALID_SREG) {
222 if (kIsDebugBuild && s_reg == live_sreg_) {
223 live_sreg_ = INVALID_SREG;
224 }
225 for (RegisterInfo* info : tempreg_info_) {
226 if (info->SReg() == s_reg) {
227 if (info->GetReg().NotExactlyEquals(info->Partner())) {
228 // Dealing with a pair - clobber the other half.
229 DCHECK(!info->IsAliased());
230 ClobberBody(GetRegInfo(info->Partner()));
231 }
232 ClobberBody(info);
233 if (info->IsAliased()) {
234 ClobberAliases(info, info->StorageMask());
235 }
236 }
237 }
238 }
239 }
240
241 /*
242 * SSA names associated with the initial definitions of Dalvik
243 * registers are the same as the Dalvik register number (and
244 * thus take the same position in the promotion_map. However,
245 * the special Method* and compiler temp resisters use negative
246 * v_reg numbers to distinguish them and can have an arbitrary
247 * ssa name (above the last original Dalvik register). This function
248 * maps SSA names to positions in the promotion_map array.
249 */
SRegToPMap(int s_reg)250 int Mir2Lir::SRegToPMap(int s_reg) {
251 DCHECK_LT(s_reg, mir_graph_->GetNumSSARegs());
252 DCHECK_GE(s_reg, 0);
253 int v_reg = mir_graph_->SRegToVReg(s_reg);
254 return v_reg;
255 }
256
257 // TODO: refactor following Alloc/Record routines - much commonality.
RecordCorePromotion(RegStorage reg,int s_reg)258 void Mir2Lir::RecordCorePromotion(RegStorage reg, int s_reg) {
259 int p_map_idx = SRegToPMap(s_reg);
260 int v_reg = mir_graph_->SRegToVReg(s_reg);
261 int reg_num = reg.GetRegNum();
262 GetRegInfo(reg)->MarkInUse();
263 core_spill_mask_ |= (1 << reg_num);
264 // Include reg for later sort
265 core_vmap_table_.push_back(reg_num << VREG_NUM_WIDTH | (v_reg & ((1 << VREG_NUM_WIDTH) - 1)));
266 num_core_spills_++;
267 promotion_map_[p_map_idx].core_location = kLocPhysReg;
268 promotion_map_[p_map_idx].core_reg = reg_num;
269 }
270
271 /* Reserve a callee-save register. Return InvalidReg if none available */
AllocPreservedCoreReg(int s_reg)272 RegStorage Mir2Lir::AllocPreservedCoreReg(int s_reg) {
273 RegStorage res;
274 /*
275 * Note: it really doesn't matter much whether we allocate from the core or core64
276 * pool for 64-bit targets - but for some targets it does matter whether allocations
277 * happens from the single or double pool. This entire section of code could stand
278 * a good refactoring.
279 */
280 for (RegisterInfo* info : reg_pool_->core_regs_) {
281 if (!info->IsTemp() && !info->InUse()) {
282 res = info->GetReg();
283 RecordCorePromotion(res, s_reg);
284 break;
285 }
286 }
287 return res;
288 }
289
RecordFpPromotion(RegStorage reg,int s_reg)290 void Mir2Lir::RecordFpPromotion(RegStorage reg, int s_reg) {
291 DCHECK_NE(cu_->instruction_set, kThumb2);
292 int p_map_idx = SRegToPMap(s_reg);
293 int v_reg = mir_graph_->SRegToVReg(s_reg);
294 int reg_num = reg.GetRegNum();
295 GetRegInfo(reg)->MarkInUse();
296 fp_spill_mask_ |= (1 << reg_num);
297 // Include reg for later sort
298 fp_vmap_table_.push_back(reg_num << VREG_NUM_WIDTH | (v_reg & ((1 << VREG_NUM_WIDTH) - 1)));
299 num_fp_spills_++;
300 promotion_map_[p_map_idx].fp_location = kLocPhysReg;
301 promotion_map_[p_map_idx].fp_reg = reg.GetReg();
302 }
303
304 // Reserve a callee-save floating point.
AllocPreservedFpReg(int s_reg)305 RegStorage Mir2Lir::AllocPreservedFpReg(int s_reg) {
306 /*
307 * For targets other than Thumb2, it doesn't matter whether we allocate from
308 * the sp_regs_ or dp_regs_ pool. Some refactoring is in order here.
309 */
310 DCHECK_NE(cu_->instruction_set, kThumb2);
311 RegStorage res;
312 for (RegisterInfo* info : reg_pool_->sp_regs_) {
313 if (!info->IsTemp() && !info->InUse()) {
314 res = info->GetReg();
315 RecordFpPromotion(res, s_reg);
316 break;
317 }
318 }
319 return res;
320 }
321
322 // TODO: this is Thumb2 only. Remove when DoPromotion refactored.
AllocPreservedDouble(int s_reg)323 RegStorage Mir2Lir::AllocPreservedDouble(int s_reg) {
324 UNUSED(s_reg);
325 UNIMPLEMENTED(FATAL) << "Unexpected use of AllocPreservedDouble";
326 UNREACHABLE();
327 }
328
329 // TODO: this is Thumb2 only. Remove when DoPromotion refactored.
AllocPreservedSingle(int s_reg)330 RegStorage Mir2Lir::AllocPreservedSingle(int s_reg) {
331 UNUSED(s_reg);
332 UNIMPLEMENTED(FATAL) << "Unexpected use of AllocPreservedSingle";
333 UNREACHABLE();
334 }
335
336
AllocTempBody(ArenaVector<RegisterInfo * > & regs,int * next_temp,bool required)337 RegStorage Mir2Lir::AllocTempBody(ArenaVector<RegisterInfo*>& regs, int* next_temp, bool required) {
338 int num_regs = regs.size();
339 int next = *next_temp;
340 for (int i = 0; i< num_regs; i++) {
341 if (next >= num_regs) {
342 next = 0;
343 }
344 RegisterInfo* info = regs[next];
345 // Try to allocate a register that doesn't hold a live value.
346 if (info->IsTemp() && !info->InUse() && info->IsDead()) {
347 // If it's wide, split it up.
348 if (info->IsWide()) {
349 // If the pair was associated with a wide value, unmark the partner as well.
350 if (info->SReg() != INVALID_SREG) {
351 RegisterInfo* partner = GetRegInfo(info->Partner());
352 DCHECK_EQ(info->GetReg().GetRegNum(), partner->Partner().GetRegNum());
353 DCHECK(partner->IsWide());
354 partner->SetIsWide(false);
355 }
356 info->SetIsWide(false);
357 }
358 Clobber(info->GetReg());
359 info->MarkInUse();
360 *next_temp = next + 1;
361 return info->GetReg();
362 }
363 next++;
364 }
365 next = *next_temp;
366 // No free non-live regs. Anything we can kill?
367 for (int i = 0; i< num_regs; i++) {
368 if (next >= num_regs) {
369 next = 0;
370 }
371 RegisterInfo* info = regs[next];
372 if (info->IsTemp() && !info->InUse()) {
373 // Got one. Kill it.
374 ClobberSReg(info->SReg());
375 Clobber(info->GetReg());
376 info->MarkInUse();
377 if (info->IsWide()) {
378 RegisterInfo* partner = GetRegInfo(info->Partner());
379 DCHECK_EQ(info->GetReg().GetRegNum(), partner->Partner().GetRegNum());
380 DCHECK(partner->IsWide());
381 info->SetIsWide(false);
382 partner->SetIsWide(false);
383 }
384 *next_temp = next + 1;
385 return info->GetReg();
386 }
387 next++;
388 }
389 if (required) {
390 CodegenDump();
391 DumpRegPools();
392 LOG(FATAL) << "No free temp registers";
393 }
394 return RegStorage::InvalidReg(); // No register available
395 }
396
AllocTemp(bool required)397 RegStorage Mir2Lir::AllocTemp(bool required) {
398 return AllocTempBody(reg_pool_->core_regs_, ®_pool_->next_core_reg_, required);
399 }
400
AllocTempWide(bool required)401 RegStorage Mir2Lir::AllocTempWide(bool required) {
402 RegStorage res;
403 if (reg_pool_->core64_regs_.size() != 0) {
404 res = AllocTempBody(reg_pool_->core64_regs_, ®_pool_->next_core64_reg_, required);
405 } else {
406 RegStorage low_reg = AllocTemp();
407 RegStorage high_reg = AllocTemp();
408 res = RegStorage::MakeRegPair(low_reg, high_reg);
409 }
410 if (required) {
411 CheckRegStorage(res, WidenessCheck::kCheckWide, RefCheck::kIgnoreRef, FPCheck::kCheckNotFP);
412 }
413 return res;
414 }
415
AllocTempRef(bool required)416 RegStorage Mir2Lir::AllocTempRef(bool required) {
417 RegStorage res = AllocTempBody(*reg_pool_->ref_regs_, reg_pool_->next_ref_reg_, required);
418 if (required) {
419 DCHECK(!res.IsPair());
420 CheckRegStorage(res, WidenessCheck::kCheckNotWide, RefCheck::kCheckRef, FPCheck::kCheckNotFP);
421 }
422 return res;
423 }
424
AllocTempSingle(bool required)425 RegStorage Mir2Lir::AllocTempSingle(bool required) {
426 RegStorage res = AllocTempBody(reg_pool_->sp_regs_, ®_pool_->next_sp_reg_, required);
427 if (required) {
428 DCHECK(res.IsSingle()) << "Reg: 0x" << std::hex << res.GetRawBits();
429 CheckRegStorage(res, WidenessCheck::kCheckNotWide, RefCheck::kCheckNotRef, FPCheck::kIgnoreFP);
430 }
431 return res;
432 }
433
AllocTempDouble(bool required)434 RegStorage Mir2Lir::AllocTempDouble(bool required) {
435 RegStorage res = AllocTempBody(reg_pool_->dp_regs_, ®_pool_->next_dp_reg_, required);
436 if (required) {
437 DCHECK(res.IsDouble()) << "Reg: 0x" << std::hex << res.GetRawBits();
438 CheckRegStorage(res, WidenessCheck::kCheckWide, RefCheck::kCheckNotRef, FPCheck::kIgnoreFP);
439 }
440 return res;
441 }
442
AllocTypedTempWide(bool fp_hint,int reg_class,bool required)443 RegStorage Mir2Lir::AllocTypedTempWide(bool fp_hint, int reg_class, bool required) {
444 DCHECK_NE(reg_class, kRefReg); // NOTE: the Dalvik width of a reference is always 32 bits.
445 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) {
446 return AllocTempDouble(required);
447 }
448 return AllocTempWide(required);
449 }
450
AllocTypedTemp(bool fp_hint,int reg_class,bool required)451 RegStorage Mir2Lir::AllocTypedTemp(bool fp_hint, int reg_class, bool required) {
452 if (((reg_class == kAnyReg) && fp_hint) || (reg_class == kFPReg)) {
453 return AllocTempSingle(required);
454 } else if (reg_class == kRefReg) {
455 return AllocTempRef(required);
456 }
457 return AllocTemp(required);
458 }
459
FindLiveReg(ArenaVector<RegisterInfo * > & regs,int s_reg)460 RegStorage Mir2Lir::FindLiveReg(ArenaVector<RegisterInfo*>& regs, int s_reg) {
461 RegStorage res;
462 for (RegisterInfo* info : regs) {
463 if ((info->SReg() == s_reg) && info->IsLive()) {
464 res = info->GetReg();
465 break;
466 }
467 }
468 return res;
469 }
470
AllocLiveReg(int s_reg,int reg_class,bool wide)471 RegStorage Mir2Lir::AllocLiveReg(int s_reg, int reg_class, bool wide) {
472 RegStorage reg;
473 if (reg_class == kRefReg) {
474 reg = FindLiveReg(*reg_pool_->ref_regs_, s_reg);
475 CheckRegStorage(reg, WidenessCheck::kCheckNotWide, RefCheck::kCheckRef, FPCheck::kCheckNotFP);
476 }
477 if (!reg.Valid() && ((reg_class == kAnyReg) || (reg_class == kFPReg))) {
478 reg = FindLiveReg(wide ? reg_pool_->dp_regs_ : reg_pool_->sp_regs_, s_reg);
479 }
480 if (!reg.Valid() && (reg_class != kFPReg)) {
481 if (cu_->target64) {
482 reg = FindLiveReg(wide || reg_class == kRefReg ? reg_pool_->core64_regs_ :
483 reg_pool_->core_regs_, s_reg);
484 } else {
485 reg = FindLiveReg(reg_pool_->core_regs_, s_reg);
486 }
487 }
488 if (reg.Valid()) {
489 if (wide && !reg.IsFloat() && !cu_->target64) {
490 // Only allow reg pairs for core regs on 32-bit targets.
491 RegStorage high_reg = FindLiveReg(reg_pool_->core_regs_, s_reg + 1);
492 if (high_reg.Valid()) {
493 reg = RegStorage::MakeRegPair(reg, high_reg);
494 MarkWide(reg);
495 } else {
496 // Only half available.
497 reg = RegStorage::InvalidReg();
498 }
499 }
500 if (reg.Valid() && (wide != GetRegInfo(reg)->IsWide())) {
501 // Width mismatch - don't try to reuse.
502 reg = RegStorage::InvalidReg();
503 }
504 }
505 if (reg.Valid()) {
506 if (reg.IsPair()) {
507 RegisterInfo* info_low = GetRegInfo(reg.GetLow());
508 RegisterInfo* info_high = GetRegInfo(reg.GetHigh());
509 if (info_low->IsTemp()) {
510 info_low->MarkInUse();
511 }
512 if (info_high->IsTemp()) {
513 info_high->MarkInUse();
514 }
515 } else {
516 RegisterInfo* info = GetRegInfo(reg);
517 if (info->IsTemp()) {
518 info->MarkInUse();
519 }
520 }
521 } else {
522 // Either not found, or something didn't match up. Clobber to prevent any stale instances.
523 ClobberSReg(s_reg);
524 if (wide) {
525 ClobberSReg(s_reg + 1);
526 }
527 }
528 CheckRegStorage(reg, WidenessCheck::kIgnoreWide,
529 reg_class == kRefReg ? RefCheck::kCheckRef : RefCheck::kIgnoreRef,
530 FPCheck::kIgnoreFP);
531 return reg;
532 }
533
FreeTemp(RegStorage reg)534 void Mir2Lir::FreeTemp(RegStorage reg) {
535 if (reg.IsPair()) {
536 FreeTemp(reg.GetLow());
537 FreeTemp(reg.GetHigh());
538 } else {
539 RegisterInfo* p = GetRegInfo(reg);
540 if (p->IsTemp()) {
541 p->MarkFree();
542 p->SetIsWide(false);
543 p->SetPartner(reg);
544 }
545 }
546 }
547
FreeRegLocTemps(RegLocation rl_keep,RegLocation rl_free)548 void Mir2Lir::FreeRegLocTemps(RegLocation rl_keep, RegLocation rl_free) {
549 DCHECK(rl_keep.wide);
550 DCHECK(rl_free.wide);
551 int free_low = rl_free.reg.GetLowReg();
552 int free_high = rl_free.reg.GetHighReg();
553 int keep_low = rl_keep.reg.GetLowReg();
554 int keep_high = rl_keep.reg.GetHighReg();
555 if ((free_low != keep_low) && (free_low != keep_high) &&
556 (free_high != keep_low) && (free_high != keep_high)) {
557 // No overlap, free both
558 FreeTemp(rl_free.reg);
559 }
560 }
561
IsLive(RegStorage reg)562 bool Mir2Lir::IsLive(RegStorage reg) {
563 bool res;
564 if (reg.IsPair()) {
565 RegisterInfo* p_lo = GetRegInfo(reg.GetLow());
566 RegisterInfo* p_hi = GetRegInfo(reg.GetHigh());
567 DCHECK_EQ(p_lo->IsLive(), p_hi->IsLive());
568 res = p_lo->IsLive() || p_hi->IsLive();
569 } else {
570 RegisterInfo* p = GetRegInfo(reg);
571 res = p->IsLive();
572 }
573 return res;
574 }
575
IsTemp(RegStorage reg)576 bool Mir2Lir::IsTemp(RegStorage reg) {
577 bool res;
578 if (reg.IsPair()) {
579 RegisterInfo* p_lo = GetRegInfo(reg.GetLow());
580 RegisterInfo* p_hi = GetRegInfo(reg.GetHigh());
581 res = p_lo->IsTemp() || p_hi->IsTemp();
582 } else {
583 RegisterInfo* p = GetRegInfo(reg);
584 res = p->IsTemp();
585 }
586 return res;
587 }
588
IsPromoted(RegStorage reg)589 bool Mir2Lir::IsPromoted(RegStorage reg) {
590 bool res;
591 if (reg.IsPair()) {
592 RegisterInfo* p_lo = GetRegInfo(reg.GetLow());
593 RegisterInfo* p_hi = GetRegInfo(reg.GetHigh());
594 res = !p_lo->IsTemp() || !p_hi->IsTemp();
595 } else {
596 RegisterInfo* p = GetRegInfo(reg);
597 res = !p->IsTemp();
598 }
599 return res;
600 }
601
IsDirty(RegStorage reg)602 bool Mir2Lir::IsDirty(RegStorage reg) {
603 bool res;
604 if (reg.IsPair()) {
605 RegisterInfo* p_lo = GetRegInfo(reg.GetLow());
606 RegisterInfo* p_hi = GetRegInfo(reg.GetHigh());
607 res = p_lo->IsDirty() || p_hi->IsDirty();
608 } else {
609 RegisterInfo* p = GetRegInfo(reg);
610 res = p->IsDirty();
611 }
612 return res;
613 }
614
615 /*
616 * Similar to AllocTemp(), but forces the allocation of a specific
617 * register. No check is made to see if the register was previously
618 * allocated. Use with caution.
619 */
LockTemp(RegStorage reg)620 void Mir2Lir::LockTemp(RegStorage reg) {
621 DCHECK(IsTemp(reg));
622 if (reg.IsPair()) {
623 RegisterInfo* p_lo = GetRegInfo(reg.GetLow());
624 RegisterInfo* p_hi = GetRegInfo(reg.GetHigh());
625 p_lo->MarkInUse();
626 p_lo->MarkDead();
627 p_hi->MarkInUse();
628 p_hi->MarkDead();
629 } else {
630 RegisterInfo* p = GetRegInfo(reg);
631 p->MarkInUse();
632 p->MarkDead();
633 }
634 }
635
ResetDef(RegStorage reg)636 void Mir2Lir::ResetDef(RegStorage reg) {
637 if (reg.IsPair()) {
638 GetRegInfo(reg.GetLow())->ResetDefBody();
639 GetRegInfo(reg.GetHigh())->ResetDefBody();
640 } else {
641 GetRegInfo(reg)->ResetDefBody();
642 }
643 }
644
NullifyRange(RegStorage reg,int s_reg)645 void Mir2Lir::NullifyRange(RegStorage reg, int s_reg) {
646 RegisterInfo* info = nullptr;
647 RegStorage rs = reg.IsPair() ? reg.GetLow() : reg;
648 if (IsTemp(rs)) {
649 info = GetRegInfo(reg);
650 }
651 if ((info != nullptr) && (info->DefStart() != nullptr) && (info->DefEnd() != nullptr)) {
652 DCHECK_EQ(info->SReg(), s_reg); // Make sure we're on the same page.
653 for (LIR* p = info->DefStart();; p = p->next) {
654 NopLIR(p);
655 if (p == info->DefEnd()) {
656 break;
657 }
658 }
659 }
660 }
661
662 /*
663 * Mark the beginning and end LIR of a def sequence. Note that
664 * on entry start points to the LIR prior to the beginning of the
665 * sequence.
666 */
MarkDef(RegLocation rl,LIR * start,LIR * finish)667 void Mir2Lir::MarkDef(RegLocation rl, LIR *start, LIR *finish) {
668 DCHECK(!rl.wide);
669 DCHECK(start && start->next);
670 DCHECK(finish);
671 RegisterInfo* p = GetRegInfo(rl.reg);
672 p->SetDefStart(start->next);
673 p->SetDefEnd(finish);
674 }
675
676 /*
677 * Mark the beginning and end LIR of a def sequence. Note that
678 * on entry start points to the LIR prior to the beginning of the
679 * sequence.
680 */
MarkDefWide(RegLocation rl,LIR * start,LIR * finish)681 void Mir2Lir::MarkDefWide(RegLocation rl, LIR *start, LIR *finish) {
682 DCHECK(rl.wide);
683 DCHECK(start && start->next);
684 DCHECK(finish);
685 RegisterInfo* p;
686 if (rl.reg.IsPair()) {
687 p = GetRegInfo(rl.reg.GetLow());
688 ResetDef(rl.reg.GetHigh()); // Only track low of pair
689 } else {
690 p = GetRegInfo(rl.reg);
691 }
692 p->SetDefStart(start->next);
693 p->SetDefEnd(finish);
694 }
695
ResetDefLoc(RegLocation rl)696 void Mir2Lir::ResetDefLoc(RegLocation rl) {
697 DCHECK(!rl.wide);
698 if (IsTemp(rl.reg) && !(cu_->disable_opt & (1 << kSuppressLoads))) {
699 NullifyRange(rl.reg, rl.s_reg_low);
700 }
701 ResetDef(rl.reg);
702 }
703
ResetDefLocWide(RegLocation rl)704 void Mir2Lir::ResetDefLocWide(RegLocation rl) {
705 DCHECK(rl.wide);
706 // If pair, only track low reg of pair.
707 RegStorage rs = rl.reg.IsPair() ? rl.reg.GetLow() : rl.reg;
708 if (IsTemp(rs) && !(cu_->disable_opt & (1 << kSuppressLoads))) {
709 NullifyRange(rs, rl.s_reg_low);
710 }
711 ResetDef(rs);
712 }
713
ResetDefTracking()714 void Mir2Lir::ResetDefTracking() {
715 for (RegisterInfo* info : tempreg_info_) {
716 info->ResetDefBody();
717 }
718 }
719
ClobberAllTemps()720 void Mir2Lir::ClobberAllTemps() {
721 for (RegisterInfo* info : tempreg_info_) {
722 ClobberBody(info);
723 }
724 }
725
FlushRegWide(RegStorage reg)726 void Mir2Lir::FlushRegWide(RegStorage reg) {
727 if (reg.IsPair()) {
728 RegisterInfo* info1 = GetRegInfo(reg.GetLow());
729 RegisterInfo* info2 = GetRegInfo(reg.GetHigh());
730 DCHECK(info1 && info2 && info1->IsWide() && info2->IsWide() &&
731 (info1->Partner().ExactlyEquals(info2->GetReg())) &&
732 (info2->Partner().ExactlyEquals(info1->GetReg())));
733 if ((info1->IsLive() && info1->IsDirty()) || (info2->IsLive() && info2->IsDirty())) {
734 if (!(info1->IsTemp() && info2->IsTemp())) {
735 /* Should not happen. If it does, there's a problem in eval_loc */
736 LOG(FATAL) << "Long half-temp, half-promoted";
737 }
738
739 info1->SetIsDirty(false);
740 info2->SetIsDirty(false);
741 if (mir_graph_->SRegToVReg(info2->SReg()) < mir_graph_->SRegToVReg(info1->SReg())) {
742 info1 = info2;
743 }
744 int v_reg = mir_graph_->SRegToVReg(info1->SReg());
745 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
746 StoreBaseDisp(TargetPtrReg(kSp), VRegOffset(v_reg), reg, k64, kNotVolatile);
747 }
748 } else {
749 RegisterInfo* info = GetRegInfo(reg);
750 if (info->IsLive() && info->IsDirty()) {
751 info->SetIsDirty(false);
752 int v_reg = mir_graph_->SRegToVReg(info->SReg());
753 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
754 StoreBaseDisp(TargetPtrReg(kSp), VRegOffset(v_reg), reg, k64, kNotVolatile);
755 }
756 }
757 }
758
FlushReg(RegStorage reg)759 void Mir2Lir::FlushReg(RegStorage reg) {
760 DCHECK(!reg.IsPair());
761 RegisterInfo* info = GetRegInfo(reg);
762 if (info->IsLive() && info->IsDirty()) {
763 info->SetIsDirty(false);
764 int v_reg = mir_graph_->SRegToVReg(info->SReg());
765 ScopedMemRefType mem_ref_type(this, ResourceMask::kDalvikReg);
766 StoreBaseDisp(TargetPtrReg(kSp), VRegOffset(v_reg), reg, kWord, kNotVolatile);
767 }
768 }
769
FlushSpecificReg(RegisterInfo * info)770 void Mir2Lir::FlushSpecificReg(RegisterInfo* info) {
771 if (info->IsWide()) {
772 FlushRegWide(info->GetReg());
773 } else {
774 FlushReg(info->GetReg());
775 }
776 }
777
FlushAllRegs()778 void Mir2Lir::FlushAllRegs() {
779 for (RegisterInfo* info : tempreg_info_) {
780 if (info->IsDirty() && info->IsLive()) {
781 FlushSpecificReg(info);
782 }
783 info->MarkDead();
784 info->SetIsWide(false);
785 }
786 }
787
788
RegClassMatches(int reg_class,RegStorage reg)789 bool Mir2Lir::RegClassMatches(int reg_class, RegStorage reg) {
790 if (reg_class == kAnyReg) {
791 return true;
792 } else if ((reg_class == kCoreReg) || (reg_class == kRefReg)) {
793 /*
794 * For this purpose, consider Core and Ref to be the same class. We aren't dealing
795 * with width here - that should be checked at a higher level (if needed).
796 */
797 return !reg.IsFloat();
798 } else {
799 return reg.IsFloat();
800 }
801 }
802
MarkLive(RegLocation loc)803 void Mir2Lir::MarkLive(RegLocation loc) {
804 RegStorage reg = loc.reg;
805 if (!IsTemp(reg)) {
806 return;
807 }
808 int s_reg = loc.s_reg_low;
809 if (s_reg == INVALID_SREG) {
810 // Can't be live if no associated sreg.
811 if (reg.IsPair()) {
812 GetRegInfo(reg.GetLow())->MarkDead();
813 GetRegInfo(reg.GetHigh())->MarkDead();
814 } else {
815 GetRegInfo(reg)->MarkDead();
816 }
817 } else {
818 if (reg.IsPair()) {
819 RegisterInfo* info_lo = GetRegInfo(reg.GetLow());
820 RegisterInfo* info_hi = GetRegInfo(reg.GetHigh());
821 if (info_lo->IsLive() && (info_lo->SReg() == s_reg) && info_hi->IsLive() &&
822 (info_hi->SReg() == s_reg)) {
823 return; // Already live.
824 }
825 ClobberSReg(s_reg);
826 ClobberSReg(s_reg + 1);
827 info_lo->MarkLive(s_reg);
828 info_hi->MarkLive(s_reg + 1);
829 } else {
830 RegisterInfo* info = GetRegInfo(reg);
831 if (info->IsLive() && (info->SReg() == s_reg)) {
832 return; // Already live.
833 }
834 ClobberSReg(s_reg);
835 if (loc.wide) {
836 ClobberSReg(s_reg + 1);
837 }
838 info->MarkLive(s_reg);
839 }
840 if (loc.wide) {
841 MarkWide(reg);
842 } else {
843 MarkNarrow(reg);
844 }
845 }
846 }
847
MarkTemp(RegStorage reg)848 void Mir2Lir::MarkTemp(RegStorage reg) {
849 DCHECK(!reg.IsPair());
850 RegisterInfo* info = GetRegInfo(reg);
851 tempreg_info_.push_back(info);
852 info->SetIsTemp(true);
853 }
854
UnmarkTemp(RegStorage reg)855 void Mir2Lir::UnmarkTemp(RegStorage reg) {
856 DCHECK(!reg.IsPair());
857 RegisterInfo* info = GetRegInfo(reg);
858 auto pos = std::find(tempreg_info_.begin(), tempreg_info_.end(), info);
859 DCHECK(pos != tempreg_info_.end());
860 tempreg_info_.erase(pos);
861 info->SetIsTemp(false);
862 }
863
MarkWide(RegStorage reg)864 void Mir2Lir::MarkWide(RegStorage reg) {
865 if (reg.IsPair()) {
866 RegisterInfo* info_lo = GetRegInfo(reg.GetLow());
867 RegisterInfo* info_hi = GetRegInfo(reg.GetHigh());
868 // Unpair any old partners.
869 if (info_lo->IsWide() && info_lo->Partner().NotExactlyEquals(info_hi->GetReg())) {
870 GetRegInfo(info_lo->Partner())->SetIsWide(false);
871 }
872 if (info_hi->IsWide() && info_hi->Partner().NotExactlyEquals(info_lo->GetReg())) {
873 GetRegInfo(info_hi->Partner())->SetIsWide(false);
874 }
875 info_lo->SetIsWide(true);
876 info_hi->SetIsWide(true);
877 info_lo->SetPartner(reg.GetHigh());
878 info_hi->SetPartner(reg.GetLow());
879 } else {
880 RegisterInfo* info = GetRegInfo(reg);
881 info->SetIsWide(true);
882 info->SetPartner(reg);
883 }
884 }
885
MarkNarrow(RegStorage reg)886 void Mir2Lir::MarkNarrow(RegStorage reg) {
887 DCHECK(!reg.IsPair());
888 RegisterInfo* info = GetRegInfo(reg);
889 info->SetIsWide(false);
890 info->SetPartner(reg);
891 }
892
MarkClean(RegLocation loc)893 void Mir2Lir::MarkClean(RegLocation loc) {
894 if (loc.reg.IsPair()) {
895 RegisterInfo* info = GetRegInfo(loc.reg.GetLow());
896 info->SetIsDirty(false);
897 info = GetRegInfo(loc.reg.GetHigh());
898 info->SetIsDirty(false);
899 } else {
900 RegisterInfo* info = GetRegInfo(loc.reg);
901 info->SetIsDirty(false);
902 }
903 }
904
905 // FIXME: need to verify rules/assumptions about how wide values are treated in 64BitSolos.
MarkDirty(RegLocation loc)906 void Mir2Lir::MarkDirty(RegLocation loc) {
907 if (loc.home) {
908 // If already home, can't be dirty
909 return;
910 }
911 if (loc.reg.IsPair()) {
912 RegisterInfo* info = GetRegInfo(loc.reg.GetLow());
913 info->SetIsDirty(true);
914 info = GetRegInfo(loc.reg.GetHigh());
915 info->SetIsDirty(true);
916 } else {
917 RegisterInfo* info = GetRegInfo(loc.reg);
918 info->SetIsDirty(true);
919 }
920 }
921
MarkInUse(RegStorage reg)922 void Mir2Lir::MarkInUse(RegStorage reg) {
923 if (reg.IsPair()) {
924 GetRegInfo(reg.GetLow())->MarkInUse();
925 GetRegInfo(reg.GetHigh())->MarkInUse();
926 } else {
927 GetRegInfo(reg)->MarkInUse();
928 }
929 }
930
CheckCorePoolSanity()931 bool Mir2Lir::CheckCorePoolSanity() {
932 for (RegisterInfo* info : tempreg_info_) {
933 int my_sreg = info->SReg();
934 if (info->IsTemp() && info->IsLive() && info->IsWide() && my_sreg != INVALID_SREG) {
935 RegStorage my_reg = info->GetReg();
936 RegStorage partner_reg = info->Partner();
937 RegisterInfo* partner = GetRegInfo(partner_reg);
938 DCHECK(partner != nullptr);
939 DCHECK(partner->IsWide());
940 DCHECK_EQ(my_reg.GetReg(), partner->Partner().GetReg());
941 DCHECK(partner->IsLive());
942 int partner_sreg = partner->SReg();
943 int diff = my_sreg - partner_sreg;
944 DCHECK((diff == 0) || (diff == -1) || (diff == 1));
945 }
946 if (info->Master() != info) {
947 // Aliased.
948 if (info->IsLive() && (info->SReg() != INVALID_SREG)) {
949 // If I'm live, master should not be live, but should show liveness in alias set.
950 DCHECK_EQ(info->Master()->SReg(), INVALID_SREG);
951 DCHECK(!info->Master()->IsDead());
952 }
953 // TODO: Add checks in !info->IsDead() case to ensure every live bit is owned by exactly 1 reg.
954 }
955 if (info->IsAliased()) {
956 // Has child aliases.
957 DCHECK_EQ(info->Master(), info);
958 if (info->IsLive() && (info->SReg() != INVALID_SREG)) {
959 // Master live, no child should be dead - all should show liveness in set.
960 for (RegisterInfo* p = info->GetAliasChain(); p != nullptr; p = p->GetAliasChain()) {
961 DCHECK(!p->IsDead());
962 DCHECK_EQ(p->SReg(), INVALID_SREG);
963 }
964 } else if (!info->IsDead()) {
965 // Master not live, one or more aliases must be.
966 bool live_alias = false;
967 for (RegisterInfo* p = info->GetAliasChain(); p != nullptr; p = p->GetAliasChain()) {
968 live_alias |= p->IsLive();
969 }
970 DCHECK(live_alias);
971 }
972 }
973 if (info->IsLive() && (info->SReg() == INVALID_SREG)) {
974 // If not fully live, should have INVALID_SREG and def's should be null.
975 DCHECK(info->DefStart() == nullptr);
976 DCHECK(info->DefEnd() == nullptr);
977 }
978 }
979 return true;
980 }
981
982 /*
983 * Return an updated location record with current in-register status.
984 * If the value lives in live temps, reflect that fact. No code
985 * is generated. If the live value is part of an older pair,
986 * clobber both low and high.
987 * TUNING: clobbering both is a bit heavy-handed, but the alternative
988 * is a bit complex when dealing with FP regs. Examine code to see
989 * if it's worthwhile trying to be more clever here.
990 */
UpdateLoc(RegLocation loc)991 RegLocation Mir2Lir::UpdateLoc(RegLocation loc) {
992 DCHECK(!loc.wide);
993 DCHECK(CheckCorePoolSanity());
994 if (loc.location != kLocPhysReg) {
995 DCHECK((loc.location == kLocDalvikFrame) ||
996 (loc.location == kLocCompilerTemp));
997 RegStorage reg = AllocLiveReg(loc.s_reg_low, loc.ref ? kRefReg : kAnyReg, false);
998 if (reg.Valid()) {
999 bool match = true;
1000 RegisterInfo* info = GetRegInfo(reg);
1001 match &= !reg.IsPair();
1002 match &= !info->IsWide();
1003 if (match) {
1004 loc.location = kLocPhysReg;
1005 loc.reg = reg;
1006 } else {
1007 Clobber(reg);
1008 FreeTemp(reg);
1009 }
1010 }
1011 CheckRegLocation(loc);
1012 }
1013 return loc;
1014 }
1015
UpdateLocWide(RegLocation loc)1016 RegLocation Mir2Lir::UpdateLocWide(RegLocation loc) {
1017 DCHECK(loc.wide);
1018 DCHECK(CheckCorePoolSanity());
1019 if (loc.location != kLocPhysReg) {
1020 DCHECK((loc.location == kLocDalvikFrame) ||
1021 (loc.location == kLocCompilerTemp));
1022 RegStorage reg = AllocLiveReg(loc.s_reg_low, kAnyReg, true);
1023 if (reg.Valid()) {
1024 bool match = true;
1025 if (reg.IsPair()) {
1026 // If we've got a register pair, make sure that it was last used as the same pair.
1027 RegisterInfo* info_lo = GetRegInfo(reg.GetLow());
1028 RegisterInfo* info_hi = GetRegInfo(reg.GetHigh());
1029 match &= info_lo->IsWide();
1030 match &= info_hi->IsWide();
1031 match &= (info_lo->Partner().ExactlyEquals(info_hi->GetReg()));
1032 match &= (info_hi->Partner().ExactlyEquals(info_lo->GetReg()));
1033 } else {
1034 RegisterInfo* info = GetRegInfo(reg);
1035 match &= info->IsWide();
1036 match &= (info->GetReg().ExactlyEquals(info->Partner()));
1037 }
1038 if (match) {
1039 loc.location = kLocPhysReg;
1040 loc.reg = reg;
1041 } else {
1042 Clobber(reg);
1043 FreeTemp(reg);
1044 }
1045 }
1046 CheckRegLocation(loc);
1047 }
1048 return loc;
1049 }
1050
1051 /* For use in cases we don't know (or care) width */
UpdateRawLoc(RegLocation loc)1052 RegLocation Mir2Lir::UpdateRawLoc(RegLocation loc) {
1053 if (loc.wide)
1054 return UpdateLocWide(loc);
1055 else
1056 return UpdateLoc(loc);
1057 }
1058
EvalLocWide(RegLocation loc,int reg_class,bool update)1059 RegLocation Mir2Lir::EvalLocWide(RegLocation loc, int reg_class, bool update) {
1060 DCHECK(loc.wide);
1061
1062 loc = UpdateLocWide(loc);
1063
1064 /* If already in registers, we can assume proper form. Right reg class? */
1065 if (loc.location == kLocPhysReg) {
1066 if (!RegClassMatches(reg_class, loc.reg)) {
1067 // Wrong register class. Reallocate and transfer ownership.
1068 RegStorage new_regs = AllocTypedTempWide(loc.fp, reg_class);
1069 // Clobber the old regs.
1070 Clobber(loc.reg);
1071 // ...and mark the new ones live.
1072 loc.reg = new_regs;
1073 MarkWide(loc.reg);
1074 MarkLive(loc);
1075 }
1076 CheckRegLocation(loc);
1077 return loc;
1078 }
1079
1080 DCHECK_NE(loc.s_reg_low, INVALID_SREG);
1081 DCHECK_NE(GetSRegHi(loc.s_reg_low), INVALID_SREG);
1082
1083 loc.reg = AllocTypedTempWide(loc.fp, reg_class);
1084 MarkWide(loc.reg);
1085
1086 if (update) {
1087 loc.location = kLocPhysReg;
1088 MarkLive(loc);
1089 }
1090 CheckRegLocation(loc);
1091 return loc;
1092 }
1093
EvalLoc(RegLocation loc,int reg_class,bool update)1094 RegLocation Mir2Lir::EvalLoc(RegLocation loc, int reg_class, bool update) {
1095 // Narrow reg_class if the loc is a ref.
1096 if (loc.ref && reg_class == kAnyReg) {
1097 reg_class = kRefReg;
1098 }
1099
1100 if (loc.wide) {
1101 return EvalLocWide(loc, reg_class, update);
1102 }
1103
1104 loc = UpdateLoc(loc);
1105
1106 if (loc.location == kLocPhysReg) {
1107 if (!RegClassMatches(reg_class, loc.reg)) {
1108 // Wrong register class. Reallocate and transfer ownership.
1109 RegStorage new_reg = AllocTypedTemp(loc.fp, reg_class);
1110 // Clobber the old reg.
1111 Clobber(loc.reg);
1112 // ...and mark the new one live.
1113 loc.reg = new_reg;
1114 MarkLive(loc);
1115 }
1116 CheckRegLocation(loc);
1117 return loc;
1118 }
1119
1120 DCHECK_NE(loc.s_reg_low, INVALID_SREG);
1121
1122 loc.reg = AllocTypedTemp(loc.fp, reg_class);
1123 CheckRegLocation(loc);
1124
1125 if (update) {
1126 loc.location = kLocPhysReg;
1127 MarkLive(loc);
1128 }
1129 CheckRegLocation(loc);
1130 return loc;
1131 }
1132
AnalyzeMIR(RefCounts * core_counts,MIR * mir,uint32_t weight)1133 void Mir2Lir::AnalyzeMIR(RefCounts* core_counts, MIR* mir, uint32_t weight) {
1134 // NOTE: This should be in sync with functions that actually generate code for
1135 // the opcodes below. However, if we get this wrong, the generated code will
1136 // still be correct even if it may be sub-optimal.
1137 int opcode = mir->dalvikInsn.opcode;
1138 bool uses_method = false;
1139 bool uses_pc_rel_load = false;
1140 uint32_t dex_cache_array_offset = std::numeric_limits<uint32_t>::max();
1141 switch (opcode) {
1142 case Instruction::CHECK_CAST:
1143 case Instruction::INSTANCE_OF: {
1144 if ((opcode == Instruction::CHECK_CAST) &&
1145 (mir->optimization_flags & MIR_IGNORE_CHECK_CAST) != 0) {
1146 break; // No code generated.
1147 }
1148 uint32_t type_idx =
1149 (opcode == Instruction::CHECK_CAST) ? mir->dalvikInsn.vB : mir->dalvikInsn.vC;
1150 bool type_known_final, type_known_abstract, use_declaring_class;
1151 bool needs_access_check = !cu_->compiler_driver->CanAccessTypeWithoutChecks(
1152 cu_->method_idx, *cu_->dex_file, type_idx,
1153 &type_known_final, &type_known_abstract, &use_declaring_class);
1154 if (opcode == Instruction::CHECK_CAST && !needs_access_check &&
1155 cu_->compiler_driver->IsSafeCast(
1156 mir_graph_->GetCurrentDexCompilationUnit(), mir->offset)) {
1157 break; // No code generated.
1158 }
1159 if (!needs_access_check && !use_declaring_class && CanUseOpPcRelDexCacheArrayLoad()) {
1160 uses_pc_rel_load = true; // And ignore method use in slow path.
1161 dex_cache_array_offset = dex_cache_arrays_layout_.TypeOffset(type_idx);
1162 } else {
1163 uses_method = true;
1164 }
1165 break;
1166 }
1167
1168 case Instruction::CONST_CLASS:
1169 if (CanUseOpPcRelDexCacheArrayLoad() &&
1170 cu_->compiler_driver->CanAccessTypeWithoutChecks(cu_->method_idx, *cu_->dex_file,
1171 mir->dalvikInsn.vB)) {
1172 uses_pc_rel_load = true; // And ignore method use in slow path.
1173 dex_cache_array_offset = dex_cache_arrays_layout_.TypeOffset(mir->dalvikInsn.vB);
1174 } else {
1175 uses_method = true;
1176 }
1177 break;
1178
1179 case Instruction::CONST_STRING:
1180 case Instruction::CONST_STRING_JUMBO:
1181 if (CanUseOpPcRelDexCacheArrayLoad()) {
1182 uses_pc_rel_load = true; // And ignore method use in slow path.
1183 dex_cache_array_offset = dex_cache_arrays_layout_.StringOffset(mir->dalvikInsn.vB);
1184 } else {
1185 uses_method = true;
1186 }
1187 break;
1188
1189 case Instruction::INVOKE_VIRTUAL:
1190 case Instruction::INVOKE_SUPER:
1191 case Instruction::INVOKE_DIRECT:
1192 case Instruction::INVOKE_STATIC:
1193 case Instruction::INVOKE_INTERFACE:
1194 case Instruction::INVOKE_VIRTUAL_RANGE:
1195 case Instruction::INVOKE_SUPER_RANGE:
1196 case Instruction::INVOKE_DIRECT_RANGE:
1197 case Instruction::INVOKE_STATIC_RANGE:
1198 case Instruction::INVOKE_INTERFACE_RANGE:
1199 case Instruction::INVOKE_VIRTUAL_QUICK:
1200 case Instruction::INVOKE_VIRTUAL_RANGE_QUICK: {
1201 const MirMethodLoweringInfo& info = mir_graph_->GetMethodLoweringInfo(mir);
1202 InvokeType sharp_type = info.GetSharpType();
1203 if (info.IsIntrinsic()) {
1204 // Nothing to do, if an intrinsic uses ArtMethod* it's in the slow-path - don't count it.
1205 } else if (!info.FastPath() || (sharp_type != kStatic && sharp_type != kDirect)) {
1206 // Nothing to do, the generated code or entrypoint uses method from the stack.
1207 } else if (info.DirectCode() != 0 && info.DirectMethod() != 0) {
1208 // Nothing to do, the generated code uses method from the stack.
1209 } else if (CanUseOpPcRelDexCacheArrayLoad()) {
1210 uses_pc_rel_load = true;
1211 dex_cache_array_offset = dex_cache_arrays_layout_.MethodOffset(mir->dalvikInsn.vB);
1212 } else {
1213 uses_method = true;
1214 }
1215 break;
1216 }
1217
1218 case Instruction::NEW_INSTANCE:
1219 case Instruction::NEW_ARRAY:
1220 case Instruction::FILLED_NEW_ARRAY:
1221 case Instruction::FILLED_NEW_ARRAY_RANGE:
1222 uses_method = true;
1223 break;
1224 case Instruction::FILL_ARRAY_DATA:
1225 // Nothing to do, the entrypoint uses method from the stack.
1226 break;
1227 case Instruction::THROW:
1228 // Nothing to do, the entrypoint uses method from the stack.
1229 break;
1230
1231 case Instruction::SGET:
1232 case Instruction::SGET_WIDE:
1233 case Instruction::SGET_OBJECT:
1234 case Instruction::SGET_BOOLEAN:
1235 case Instruction::SGET_BYTE:
1236 case Instruction::SGET_CHAR:
1237 case Instruction::SGET_SHORT:
1238 case Instruction::SPUT:
1239 case Instruction::SPUT_WIDE:
1240 case Instruction::SPUT_OBJECT:
1241 case Instruction::SPUT_BOOLEAN:
1242 case Instruction::SPUT_BYTE:
1243 case Instruction::SPUT_CHAR:
1244 case Instruction::SPUT_SHORT: {
1245 const MirSFieldLoweringInfo& field_info = mir_graph_->GetSFieldLoweringInfo(mir);
1246 bool fast = IsInstructionSGet(static_cast<Instruction::Code>(opcode))
1247 ? field_info.FastGet()
1248 : field_info.FastPut();
1249 if (fast && (cu_->enable_debug & (1 << kDebugSlowFieldPath)) == 0) {
1250 if (!field_info.IsReferrersClass() && CanUseOpPcRelDexCacheArrayLoad()) {
1251 uses_pc_rel_load = true; // And ignore method use in slow path.
1252 dex_cache_array_offset = dex_cache_arrays_layout_.TypeOffset(field_info.StorageIndex());
1253 } else {
1254 uses_method = true;
1255 }
1256 } else {
1257 // Nothing to do, the entrypoint uses method from the stack.
1258 }
1259 break;
1260 }
1261
1262 default:
1263 break;
1264 }
1265 if (uses_method) {
1266 core_counts[SRegToPMap(mir_graph_->GetMethodLoc().s_reg_low)].count += weight;
1267 }
1268 if (uses_pc_rel_load) {
1269 if (pc_rel_temp_ != nullptr) {
1270 core_counts[SRegToPMap(pc_rel_temp_->s_reg_low)].count += weight;
1271 DCHECK_NE(dex_cache_array_offset, std::numeric_limits<uint32_t>::max());
1272 dex_cache_arrays_min_offset_ = std::min(dex_cache_arrays_min_offset_, dex_cache_array_offset);
1273 } else {
1274 // Nothing to do, using PC-relative addressing without promoting base PC to register.
1275 }
1276 }
1277 }
1278
1279 /* USE SSA names to count references of base Dalvik v_regs. */
CountRefs(RefCounts * core_counts,RefCounts * fp_counts,size_t num_regs)1280 void Mir2Lir::CountRefs(RefCounts* core_counts, RefCounts* fp_counts, size_t num_regs) {
1281 for (int i = 0; i < mir_graph_->GetNumSSARegs(); i++) {
1282 RegLocation loc = mir_graph_->reg_location_[i];
1283 RefCounts* counts = loc.fp ? fp_counts : core_counts;
1284 int p_map_idx = SRegToPMap(loc.s_reg_low);
1285 int use_count = mir_graph_->GetUseCount(i);
1286 if (loc.fp) {
1287 if (loc.wide) {
1288 if (WideFPRsAreAliases()) {
1289 // Floats and doubles can be counted together.
1290 counts[p_map_idx].count += use_count;
1291 } else {
1292 // Treat doubles as a unit, using upper half of fp_counts array.
1293 counts[p_map_idx + num_regs].count += use_count;
1294 }
1295 i++;
1296 } else {
1297 counts[p_map_idx].count += use_count;
1298 }
1299 } else {
1300 if (loc.wide && WideGPRsAreAliases()) {
1301 i++;
1302 }
1303 if (!IsInexpensiveConstant(loc)) {
1304 counts[p_map_idx].count += use_count;
1305 }
1306 }
1307 }
1308
1309 // Now analyze the ArtMethod* and pc_rel_temp_ uses.
1310 DCHECK_EQ(core_counts[SRegToPMap(mir_graph_->GetMethodLoc().s_reg_low)].count, 0);
1311 if (pc_rel_temp_ != nullptr) {
1312 DCHECK_EQ(core_counts[SRegToPMap(pc_rel_temp_->s_reg_low)].count, 0);
1313 }
1314 PreOrderDfsIterator iter(mir_graph_);
1315 for (BasicBlock* bb = iter.Next(); bb != nullptr; bb = iter.Next()) {
1316 if (bb->block_type == kDead) {
1317 continue;
1318 }
1319 uint32_t weight = mir_graph_->GetUseCountWeight(bb);
1320 for (MIR* mir = bb->first_mir_insn; mir != nullptr; mir = mir->next) {
1321 AnalyzeMIR(core_counts, mir, weight);
1322 }
1323 }
1324 }
1325
1326 /* qsort callback function, sort descending */
SortCounts(const void * val1,const void * val2)1327 static int SortCounts(const void *val1, const void *val2) {
1328 const Mir2Lir::RefCounts* op1 = reinterpret_cast<const Mir2Lir::RefCounts*>(val1);
1329 const Mir2Lir::RefCounts* op2 = reinterpret_cast<const Mir2Lir::RefCounts*>(val2);
1330 // Note that we fall back to sorting on reg so we get stable output on differing qsort
1331 // implementations (such as on host and target or between local host and build servers).
1332 // Note also that if a wide val1 and a non-wide val2 have the same count, then val1 always
1333 // ``loses'' (as STARTING_WIDE_SREG is or-ed in val1->s_reg).
1334 return (op1->count == op2->count)
1335 ? (op1->s_reg - op2->s_reg)
1336 : (op1->count < op2->count ? 1 : -1);
1337 }
1338
DumpCounts(const RefCounts * arr,int size,const char * msg)1339 void Mir2Lir::DumpCounts(const RefCounts* arr, int size, const char* msg) {
1340 LOG(INFO) << msg;
1341 for (int i = 0; i < size; i++) {
1342 if ((arr[i].s_reg & STARTING_WIDE_SREG) != 0) {
1343 LOG(INFO) << "s_reg[64_" << (arr[i].s_reg & ~STARTING_WIDE_SREG) << "]: " << arr[i].count;
1344 } else {
1345 LOG(INFO) << "s_reg[32_" << arr[i].s_reg << "]: " << arr[i].count;
1346 }
1347 }
1348 }
1349
1350 /*
1351 * Note: some portions of this code required even if the kPromoteRegs
1352 * optimization is disabled.
1353 */
DoPromotion()1354 void Mir2Lir::DoPromotion() {
1355 int num_regs = mir_graph_->GetNumOfCodeAndTempVRs();
1356 const int promotion_threshold = 1;
1357 // Allocate the promotion map - one entry for each Dalvik vReg or compiler temp
1358 promotion_map_ = arena_->AllocArray<PromotionMap>(num_regs, kArenaAllocRegAlloc);
1359
1360 // Allow target code to add any special registers
1361 AdjustSpillMask();
1362
1363 /*
1364 * Simple register promotion. Just do a static count of the uses
1365 * of Dalvik registers. Note that we examine the SSA names, but
1366 * count based on original Dalvik register name. Count refs
1367 * separately based on type in order to give allocation
1368 * preference to fp doubles - which must be allocated sequential
1369 * physical single fp registers starting with an even-numbered
1370 * reg.
1371 * TUNING: replace with linear scan once we have the ability
1372 * to describe register live ranges for GC.
1373 */
1374 size_t core_reg_count_size = WideGPRsAreAliases() ? num_regs : num_regs * 2;
1375 size_t fp_reg_count_size = WideFPRsAreAliases() ? num_regs : num_regs * 2;
1376 RefCounts *core_regs = arena_->AllocArray<RefCounts>(core_reg_count_size, kArenaAllocRegAlloc);
1377 RefCounts *fp_regs = arena_->AllocArray<RefCounts>(fp_reg_count_size, kArenaAllocRegAlloc);
1378 // Set ssa names for original Dalvik registers
1379 for (int i = 0; i < num_regs; i++) {
1380 core_regs[i].s_reg = fp_regs[i].s_reg = i;
1381 }
1382
1383 // Duplicate in upper half to represent possible wide starting sregs.
1384 for (size_t i = num_regs; i < fp_reg_count_size; i++) {
1385 fp_regs[i].s_reg = fp_regs[i - num_regs].s_reg | STARTING_WIDE_SREG;
1386 }
1387 for (size_t i = num_regs; i < core_reg_count_size; i++) {
1388 core_regs[i].s_reg = core_regs[i - num_regs].s_reg | STARTING_WIDE_SREG;
1389 }
1390
1391 // Sum use counts of SSA regs by original Dalvik vreg.
1392 CountRefs(core_regs, fp_regs, num_regs);
1393
1394 // Sort the count arrays
1395 qsort(core_regs, core_reg_count_size, sizeof(RefCounts), SortCounts);
1396 qsort(fp_regs, fp_reg_count_size, sizeof(RefCounts), SortCounts);
1397
1398 if (cu_->verbose) {
1399 DumpCounts(core_regs, core_reg_count_size, "Core regs after sort");
1400 DumpCounts(fp_regs, fp_reg_count_size, "Fp regs after sort");
1401 }
1402
1403 if (!(cu_->disable_opt & (1 << kPromoteRegs))) {
1404 // Promote fp regs
1405 for (size_t i = 0; (i < fp_reg_count_size) && (fp_regs[i].count >= promotion_threshold); i++) {
1406 int low_sreg = fp_regs[i].s_reg & ~STARTING_WIDE_SREG;
1407 size_t p_map_idx = SRegToPMap(low_sreg);
1408 RegStorage reg = RegStorage::InvalidReg();
1409 if (promotion_map_[p_map_idx].fp_location != kLocPhysReg) {
1410 // TODO: break out the Thumb2-specific code.
1411 if (cu_->instruction_set == kThumb2) {
1412 bool wide = fp_regs[i].s_reg & STARTING_WIDE_SREG;
1413 if (wide) {
1414 if (promotion_map_[p_map_idx + 1].fp_location != kLocPhysReg) {
1415 // Ignore result - if can't alloc double may still be able to alloc singles.
1416 AllocPreservedDouble(low_sreg);
1417 }
1418 // Continue regardless of success - might still be able to grab a single.
1419 continue;
1420 } else {
1421 reg = AllocPreservedSingle(low_sreg);
1422 }
1423 } else {
1424 reg = AllocPreservedFpReg(low_sreg);
1425 }
1426 if (!reg.Valid()) {
1427 break; // No more left
1428 }
1429 }
1430 }
1431
1432 // Promote core regs
1433 for (size_t i = 0; (i < core_reg_count_size) &&
1434 (core_regs[i].count >= promotion_threshold); i++) {
1435 int low_sreg = core_regs[i].s_reg & ~STARTING_WIDE_SREG;
1436 size_t p_map_idx = SRegToPMap(low_sreg);
1437 if (promotion_map_[p_map_idx].core_location != kLocPhysReg) {
1438 RegStorage reg = AllocPreservedCoreReg(low_sreg);
1439 if (!reg.Valid()) {
1440 break; // No more left
1441 }
1442 }
1443 }
1444 }
1445
1446 // Now, update SSA names to new home locations
1447 for (int i = 0; i < mir_graph_->GetNumSSARegs(); i++) {
1448 RegLocation *curr = &mir_graph_->reg_location_[i];
1449 int p_map_idx = SRegToPMap(curr->s_reg_low);
1450 int reg_num = curr->fp ? promotion_map_[p_map_idx].fp_reg : promotion_map_[p_map_idx].core_reg;
1451 bool wide = curr->wide || (cu_->target64 && curr->ref);
1452 RegStorage reg = RegStorage::InvalidReg();
1453 if (curr->fp && promotion_map_[p_map_idx].fp_location == kLocPhysReg) {
1454 if (wide && cu_->instruction_set == kThumb2) {
1455 if (promotion_map_[p_map_idx + 1].fp_location == kLocPhysReg) {
1456 int high_reg = promotion_map_[p_map_idx+1].fp_reg;
1457 // TODO: move target-specific restrictions out of here.
1458 if (((reg_num & 0x1) == 0) && ((reg_num + 1) == high_reg)) {
1459 reg = RegStorage::FloatSolo64(RegStorage::RegNum(reg_num) >> 1);
1460 }
1461 }
1462 } else {
1463 reg = wide ? RegStorage::FloatSolo64(reg_num) : RegStorage::FloatSolo32(reg_num);
1464 }
1465 } else if (!curr->fp && promotion_map_[p_map_idx].core_location == kLocPhysReg) {
1466 if (wide && !cu_->target64) {
1467 if (promotion_map_[p_map_idx + 1].core_location == kLocPhysReg) {
1468 int high_reg = promotion_map_[p_map_idx+1].core_reg;
1469 reg = RegStorage(RegStorage::k64BitPair, reg_num, high_reg);
1470 }
1471 } else {
1472 reg = wide ? RegStorage::Solo64(reg_num) : RegStorage::Solo32(reg_num);
1473 }
1474 }
1475 if (reg.Valid()) {
1476 curr->reg = reg;
1477 curr->location = kLocPhysReg;
1478 curr->home = true;
1479 }
1480 }
1481 if (cu_->verbose) {
1482 DumpPromotionMap();
1483 }
1484 }
1485
1486 /* Returns sp-relative offset in bytes for a VReg */
VRegOffset(int v_reg)1487 int Mir2Lir::VRegOffset(int v_reg) {
1488 const DexFile::CodeItem* code_item = mir_graph_->GetCurrentDexCompilationUnit()->GetCodeItem();
1489 return StackVisitor::GetVRegOffsetFromQuickCode(code_item, core_spill_mask_,
1490 fp_spill_mask_, frame_size_, v_reg,
1491 cu_->instruction_set);
1492 }
1493
1494 /* Returns sp-relative offset in bytes for a SReg */
SRegOffset(int s_reg)1495 int Mir2Lir::SRegOffset(int s_reg) {
1496 return VRegOffset(mir_graph_->SRegToVReg(s_reg));
1497 }
1498
1499 /* Mark register usage state and return long retloc */
GetReturnWide(RegisterClass reg_class)1500 RegLocation Mir2Lir::GetReturnWide(RegisterClass reg_class) {
1501 RegLocation res;
1502 switch (reg_class) {
1503 case kRefReg: LOG(FATAL); break;
1504 case kFPReg: res = LocCReturnDouble(); break;
1505 default: res = LocCReturnWide(); break;
1506 }
1507 Clobber(res.reg);
1508 LockTemp(res.reg);
1509 MarkWide(res.reg);
1510 CheckRegLocation(res);
1511 return res;
1512 }
1513
GetReturn(RegisterClass reg_class)1514 RegLocation Mir2Lir::GetReturn(RegisterClass reg_class) {
1515 RegLocation res;
1516 switch (reg_class) {
1517 case kRefReg: res = LocCReturnRef(); break;
1518 case kFPReg: res = LocCReturnFloat(); break;
1519 default: res = LocCReturn(); break;
1520 }
1521 Clobber(res.reg);
1522 if (cu_->instruction_set == kMips || cu_->instruction_set == kMips64) {
1523 MarkInUse(res.reg);
1524 } else {
1525 LockTemp(res.reg);
1526 }
1527 CheckRegLocation(res);
1528 return res;
1529 }
1530
SimpleRegAlloc()1531 void Mir2Lir::SimpleRegAlloc() {
1532 DoPromotion();
1533
1534 if (cu_->verbose && !(cu_->disable_opt & (1 << kPromoteRegs))) {
1535 LOG(INFO) << "After Promotion";
1536 mir_graph_->DumpRegLocTable(mir_graph_->reg_location_, mir_graph_->GetNumSSARegs());
1537 }
1538
1539 /* Set the frame size */
1540 frame_size_ = ComputeFrameSize();
1541 }
1542
1543 /*
1544 * Get the "real" sreg number associated with an s_reg slot. In general,
1545 * s_reg values passed through codegen are the SSA names created by
1546 * dataflow analysis and refer to slot numbers in the mir_graph_->reg_location
1547 * array. However, renaming is accomplished by simply replacing RegLocation
1548 * entries in the reglocation[] array. Therefore, when location
1549 * records for operands are first created, we need to ask the locRecord
1550 * identified by the dataflow pass what it's new name is.
1551 */
GetSRegHi(int lowSreg)1552 int Mir2Lir::GetSRegHi(int lowSreg) {
1553 return (lowSreg == INVALID_SREG) ? INVALID_SREG : lowSreg + 1;
1554 }
1555
LiveOut(int s_reg)1556 bool Mir2Lir::LiveOut(int s_reg) {
1557 UNUSED(s_reg);
1558 // For now.
1559 return true;
1560 }
1561
1562 } // namespace art
1563