call_arm.cc revision 03c9785a8a6d712775cf406c4371d0227c44148f
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 codegen for the Thumb2 ISA. */
18
19#include "arm_lir.h"
20#include "codegen_arm.h"
21#include "dex/quick/mir_to_lir-inl.h"
22#include "gc/accounting/card_table.h"
23#include "entrypoints/quick/quick_entrypoints.h"
24
25namespace art {
26
27/*
28 * The sparse table in the literal pool is an array of <key,displacement>
29 * pairs.  For each set, we'll load them as a pair using ldmia.
30 * This means that the register number of the temp we use for the key
31 * must be lower than the reg for the displacement.
32 *
33 * The test loop will look something like:
34 *
35 *   adr   r_base, <table>
36 *   ldr   r_val, [rARM_SP, v_reg_off]
37 *   mov   r_idx, #table_size
38 * lp:
39 *   ldmia r_base!, {r_key, r_disp}
40 *   sub   r_idx, #1
41 *   cmp   r_val, r_key
42 *   ifeq
43 *   add   rARM_PC, r_disp   ; This is the branch from which we compute displacement
44 *   cbnz  r_idx, lp
45 */
46void ArmMir2Lir::GenLargeSparseSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
47  const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
48  if (cu_->verbose) {
49    DumpSparseSwitchTable(table);
50  }
51  // Add the table to the list - we'll process it later
52  SwitchTable *tab_rec =
53      static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
54  tab_rec->table = table;
55  tab_rec->vaddr = current_dalvik_offset_;
56  uint32_t size = table[1];
57  tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
58  switch_tables_.Insert(tab_rec);
59
60  // Get the switch value
61  rl_src = LoadValue(rl_src, kCoreReg);
62  RegStorage r_base = AllocTemp();
63  /* Allocate key and disp temps */
64  RegStorage r_key = AllocTemp();
65  RegStorage r_disp = AllocTemp();
66  // Make sure r_key's register number is less than r_disp's number for ldmia
67  if (r_key.GetReg() > r_disp.GetReg()) {
68    RegStorage tmp = r_disp;
69    r_disp = r_key;
70    r_key = tmp;
71  }
72  // Materialize a pointer to the switch table
73  NewLIR3(kThumb2Adr, r_base.GetReg(), 0, WrapPointer(tab_rec));
74  // Set up r_idx
75  RegStorage r_idx = AllocTemp();
76  LoadConstant(r_idx, size);
77  // Establish loop branch target
78  LIR* target = NewLIR0(kPseudoTargetLabel);
79  // Load next key/disp
80  NewLIR2(kThumb2LdmiaWB, r_base.GetReg(), (1 << r_key.GetRegNum()) | (1 << r_disp.GetRegNum()));
81  OpRegReg(kOpCmp, r_key, rl_src.reg);
82  // Go if match. NOTE: No instruction set switch here - must stay Thumb2
83  LIR* it = OpIT(kCondEq, "");
84  LIR* switch_branch = NewLIR1(kThumb2AddPCR, r_disp.GetReg());
85  OpEndIT(it);
86  tab_rec->anchor = switch_branch;
87  // Needs to use setflags encoding here
88  OpRegRegImm(kOpSub, r_idx, r_idx, 1);  // For value == 1, this should set flags.
89  DCHECK(last_lir_insn_->u.m.def_mask->HasBit(ResourceMask::kCCode));
90  OpCondBranch(kCondNe, target);
91}
92
93
94void ArmMir2Lir::GenLargePackedSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
95  const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
96  if (cu_->verbose) {
97    DumpPackedSwitchTable(table);
98  }
99  // Add the table to the list - we'll process it later
100  SwitchTable *tab_rec =
101      static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable),  kArenaAllocData));
102  tab_rec->table = table;
103  tab_rec->vaddr = current_dalvik_offset_;
104  uint32_t size = table[1];
105  tab_rec->targets =
106      static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
107  switch_tables_.Insert(tab_rec);
108
109  // Get the switch value
110  rl_src = LoadValue(rl_src, kCoreReg);
111  RegStorage table_base = AllocTemp();
112  // Materialize a pointer to the switch table
113  NewLIR3(kThumb2Adr, table_base.GetReg(), 0, WrapPointer(tab_rec));
114  int low_key = s4FromSwitchData(&table[2]);
115  RegStorage keyReg;
116  // Remove the bias, if necessary
117  if (low_key == 0) {
118    keyReg = rl_src.reg;
119  } else {
120    keyReg = AllocTemp();
121    OpRegRegImm(kOpSub, keyReg, rl_src.reg, low_key);
122  }
123  // Bounds check - if < 0 or >= size continue following switch
124  OpRegImm(kOpCmp, keyReg, size-1);
125  LIR* branch_over = OpCondBranch(kCondHi, NULL);
126
127  // Load the displacement from the switch table
128  RegStorage disp_reg = AllocTemp();
129  LoadBaseIndexed(table_base, keyReg, disp_reg, 2, k32);
130
131  // ..and go! NOTE: No instruction set switch here - must stay Thumb2
132  LIR* switch_branch = NewLIR1(kThumb2AddPCR, disp_reg.GetReg());
133  tab_rec->anchor = switch_branch;
134
135  /* branch_over target here */
136  LIR* target = NewLIR0(kPseudoTargetLabel);
137  branch_over->target = target;
138}
139
140/*
141 * Array data table format:
142 *  ushort ident = 0x0300   magic value
143 *  ushort width            width of each element in the table
144 *  uint   size             number of elements in the table
145 *  ubyte  data[size*width] table of data values (may contain a single-byte
146 *                          padding at the end)
147 *
148 * Total size is 4+(width * size + 1)/2 16-bit code units.
149 */
150void ArmMir2Lir::GenFillArrayData(uint32_t table_offset, RegLocation rl_src) {
151  const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
152  // Add the table to the list - we'll process it later
153  FillArrayData *tab_rec =
154      static_cast<FillArrayData*>(arena_->Alloc(sizeof(FillArrayData), kArenaAllocData));
155  tab_rec->table = table;
156  tab_rec->vaddr = current_dalvik_offset_;
157  uint16_t width = tab_rec->table[1];
158  uint32_t size = tab_rec->table[2] | ((static_cast<uint32_t>(tab_rec->table[3])) << 16);
159  tab_rec->size = (size * width) + 8;
160
161  fill_array_data_.Insert(tab_rec);
162
163  // Making a call - use explicit registers
164  FlushAllRegs();   /* Everything to home location */
165  LoadValueDirectFixed(rl_src, rs_r0);
166  LoadWordDisp(rs_rARM_SELF, QUICK_ENTRYPOINT_OFFSET(4, pHandleFillArrayData).Int32Value(),
167               rs_rARM_LR);
168  // Materialize a pointer to the fill data image
169  NewLIR3(kThumb2Adr, rs_r1.GetReg(), 0, WrapPointer(tab_rec));
170  ClobberCallerSave();
171  LIR* call_inst = OpReg(kOpBlx, rs_rARM_LR);
172  MarkSafepointPC(call_inst);
173}
174
175/*
176 * Handle unlocked -> thin locked transition inline or else call out to quick entrypoint. For more
177 * details see monitor.cc.
178 */
179void ArmMir2Lir::GenMonitorEnter(int opt_flags, RegLocation rl_src) {
180  FlushAllRegs();
181  // FIXME: need separate LoadValues for object references.
182  LoadValueDirectFixed(rl_src, rs_r0);  // Get obj
183  LockCallTemps();  // Prepare for explicit register usage
184  constexpr bool kArchVariantHasGoodBranchPredictor = false;  // TODO: true if cortex-A15.
185  if (kArchVariantHasGoodBranchPredictor) {
186    LIR* null_check_branch = nullptr;
187    if ((opt_flags & MIR_IGNORE_NULL_CHECK) && !(cu_->disable_opt & (1 << kNullCheckElimination))) {
188      null_check_branch = nullptr;  // No null check.
189    } else {
190      // If the null-check fails its handled by the slow-path to reduce exception related meta-data.
191      if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
192        null_check_branch = OpCmpImmBranch(kCondEq, rs_r0, 0, NULL);
193      }
194    }
195    Load32Disp(rs_rARM_SELF, Thread::ThinLockIdOffset<4>().Int32Value(), rs_r2);
196    NewLIR3(kThumb2Ldrex, rs_r1.GetReg(), rs_r0.GetReg(),
197        mirror::Object::MonitorOffset().Int32Value() >> 2);
198    MarkPossibleNullPointerException(opt_flags);
199    LIR* not_unlocked_branch = OpCmpImmBranch(kCondNe, rs_r1, 0, NULL);
200    NewLIR4(kThumb2Strex, rs_r1.GetReg(), rs_r2.GetReg(), rs_r0.GetReg(),
201        mirror::Object::MonitorOffset().Int32Value() >> 2);
202    LIR* lock_success_branch = OpCmpImmBranch(kCondEq, rs_r1, 0, NULL);
203
204
205    LIR* slow_path_target = NewLIR0(kPseudoTargetLabel);
206    not_unlocked_branch->target = slow_path_target;
207    if (null_check_branch != nullptr) {
208      null_check_branch->target = slow_path_target;
209    }
210    // TODO: move to a slow path.
211    // Go expensive route - artLockObjectFromCode(obj);
212    LoadWordDisp(rs_rARM_SELF, QUICK_ENTRYPOINT_OFFSET(4, pLockObject).Int32Value(), rs_rARM_LR);
213    ClobberCallerSave();
214    LIR* call_inst = OpReg(kOpBlx, rs_rARM_LR);
215    MarkSafepointPC(call_inst);
216
217    LIR* success_target = NewLIR0(kPseudoTargetLabel);
218    lock_success_branch->target = success_target;
219    GenMemBarrier(kLoadAny);
220  } else {
221    // Explicit null-check as slow-path is entered using an IT.
222    GenNullCheck(rs_r0, opt_flags);
223    Load32Disp(rs_rARM_SELF, Thread::ThinLockIdOffset<4>().Int32Value(), rs_r2);
224    NewLIR3(kThumb2Ldrex, rs_r1.GetReg(), rs_r0.GetReg(),
225        mirror::Object::MonitorOffset().Int32Value() >> 2);
226    MarkPossibleNullPointerException(opt_flags);
227    OpRegImm(kOpCmp, rs_r1, 0);
228    LIR* it = OpIT(kCondEq, "");
229    NewLIR4(kThumb2Strex/*eq*/, rs_r1.GetReg(), rs_r2.GetReg(), rs_r0.GetReg(),
230        mirror::Object::MonitorOffset().Int32Value() >> 2);
231    OpEndIT(it);
232    OpRegImm(kOpCmp, rs_r1, 0);
233    it = OpIT(kCondNe, "T");
234    // Go expensive route - artLockObjectFromCode(self, obj);
235    LoadWordDisp/*ne*/(rs_rARM_SELF, QUICK_ENTRYPOINT_OFFSET(4, pLockObject).Int32Value(),
236                       rs_rARM_LR);
237    ClobberCallerSave();
238    LIR* call_inst = OpReg(kOpBlx/*ne*/, rs_rARM_LR);
239    OpEndIT(it);
240    MarkSafepointPC(call_inst);
241    GenMemBarrier(kLoadAny);
242  }
243}
244
245/*
246 * Handle thin locked -> unlocked transition inline or else call out to quick entrypoint. For more
247 * details see monitor.cc. Note the code below doesn't use ldrex/strex as the code holds the lock
248 * and can only give away ownership if its suspended.
249 */
250void ArmMir2Lir::GenMonitorExit(int opt_flags, RegLocation rl_src) {
251  FlushAllRegs();
252  LoadValueDirectFixed(rl_src, rs_r0);  // Get obj
253  LockCallTemps();  // Prepare for explicit register usage
254  LIR* null_check_branch = nullptr;
255  Load32Disp(rs_rARM_SELF, Thread::ThinLockIdOffset<4>().Int32Value(), rs_r2);
256  constexpr bool kArchVariantHasGoodBranchPredictor = false;  // TODO: true if cortex-A15.
257  if (kArchVariantHasGoodBranchPredictor) {
258    if ((opt_flags & MIR_IGNORE_NULL_CHECK) && !(cu_->disable_opt & (1 << kNullCheckElimination))) {
259      null_check_branch = nullptr;  // No null check.
260    } else {
261      // If the null-check fails its handled by the slow-path to reduce exception related meta-data.
262      if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitNullChecks()) {
263        null_check_branch = OpCmpImmBranch(kCondEq, rs_r0, 0, NULL);
264      }
265    }
266    Load32Disp(rs_r0, mirror::Object::MonitorOffset().Int32Value(), rs_r1);
267    MarkPossibleNullPointerException(opt_flags);
268    LoadConstantNoClobber(rs_r3, 0);
269    LIR* slow_unlock_branch = OpCmpBranch(kCondNe, rs_r1, rs_r2, NULL);
270    GenMemBarrier(kAnyStore);
271    Store32Disp(rs_r0, mirror::Object::MonitorOffset().Int32Value(), rs_r3);
272    LIR* unlock_success_branch = OpUnconditionalBranch(NULL);
273
274    LIR* slow_path_target = NewLIR0(kPseudoTargetLabel);
275    slow_unlock_branch->target = slow_path_target;
276    if (null_check_branch != nullptr) {
277      null_check_branch->target = slow_path_target;
278    }
279    // TODO: move to a slow path.
280    // Go expensive route - artUnlockObjectFromCode(obj);
281    LoadWordDisp(rs_rARM_SELF, QUICK_ENTRYPOINT_OFFSET(4, pUnlockObject).Int32Value(), rs_rARM_LR);
282    ClobberCallerSave();
283    LIR* call_inst = OpReg(kOpBlx, rs_rARM_LR);
284    MarkSafepointPC(call_inst);
285
286    LIR* success_target = NewLIR0(kPseudoTargetLabel);
287    unlock_success_branch->target = success_target;
288  } else {
289    // Explicit null-check as slow-path is entered using an IT.
290    GenNullCheck(rs_r0, opt_flags);
291    Load32Disp(rs_r0, mirror::Object::MonitorOffset().Int32Value(), rs_r1);  // Get lock
292    MarkPossibleNullPointerException(opt_flags);
293    Load32Disp(rs_rARM_SELF, Thread::ThinLockIdOffset<4>().Int32Value(), rs_r2);
294    LoadConstantNoClobber(rs_r3, 0);
295    // Is lock unheld on lock or held by us (==thread_id) on unlock?
296    OpRegReg(kOpCmp, rs_r1, rs_r2);
297
298    LIR* it = OpIT(kCondEq, "EE");
299    if (GenMemBarrier(kAnyStore)) {
300      UpdateIT(it, "TEE");
301    }
302    Store32Disp/*eq*/(rs_r0, mirror::Object::MonitorOffset().Int32Value(), rs_r3);
303    // Go expensive route - UnlockObjectFromCode(obj);
304    LoadWordDisp/*ne*/(rs_rARM_SELF, QUICK_ENTRYPOINT_OFFSET(4, pUnlockObject).Int32Value(),
305                       rs_rARM_LR);
306    ClobberCallerSave();
307    LIR* call_inst = OpReg(kOpBlx/*ne*/, rs_rARM_LR);
308    OpEndIT(it);
309    MarkSafepointPC(call_inst);
310  }
311}
312
313void ArmMir2Lir::GenMoveException(RegLocation rl_dest) {
314  int ex_offset = Thread::ExceptionOffset<4>().Int32Value();
315  RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
316  RegStorage reset_reg = AllocTempRef();
317  LoadRefDisp(rs_rARM_SELF, ex_offset, rl_result.reg, kNotVolatile);
318  LoadConstant(reset_reg, 0);
319  StoreRefDisp(rs_rARM_SELF, ex_offset, reset_reg, kNotVolatile);
320  FreeTemp(reset_reg);
321  StoreValue(rl_dest, rl_result);
322}
323
324/*
325 * Mark garbage collection card. Skip if the value we're storing is null.
326 */
327void ArmMir2Lir::MarkGCCard(RegStorage val_reg, RegStorage tgt_addr_reg) {
328  RegStorage reg_card_base = AllocTemp();
329  RegStorage reg_card_no = AllocTemp();
330  LIR* branch_over = OpCmpImmBranch(kCondEq, val_reg, 0, NULL);
331  LoadWordDisp(rs_rARM_SELF, Thread::CardTableOffset<4>().Int32Value(), reg_card_base);
332  OpRegRegImm(kOpLsr, reg_card_no, tgt_addr_reg, gc::accounting::CardTable::kCardShift);
333  StoreBaseIndexed(reg_card_base, reg_card_no, reg_card_base, 0, kUnsignedByte);
334  LIR* target = NewLIR0(kPseudoTargetLabel);
335  branch_over->target = target;
336  FreeTemp(reg_card_base);
337  FreeTemp(reg_card_no);
338}
339
340void ArmMir2Lir::GenEntrySequence(RegLocation* ArgLocs, RegLocation rl_method) {
341  int spill_count = num_core_spills_ + num_fp_spills_;
342  /*
343   * On entry, r0, r1, r2 & r3 are live.  Let the register allocation
344   * mechanism know so it doesn't try to use any of them when
345   * expanding the frame or flushing.  This leaves the utility
346   * code with a single temp: r12.  This should be enough.
347   */
348  LockTemp(rs_r0);
349  LockTemp(rs_r1);
350  LockTemp(rs_r2);
351  LockTemp(rs_r3);
352
353  /*
354   * We can safely skip the stack overflow check if we're
355   * a leaf *and* our frame size < fudge factor.
356   */
357  bool skip_overflow_check = mir_graph_->MethodIsLeaf() && !FrameNeedsStackCheck(frame_size_, kArm);
358  NewLIR0(kPseudoMethodEntry);
359  const size_t kStackOverflowReservedUsableBytes = GetStackOverflowReservedBytes(kArm);
360  bool large_frame = (static_cast<size_t>(frame_size_) > kStackOverflowReservedUsableBytes);
361  bool generate_explicit_stack_overflow_check = large_frame ||
362    !cu_->compiler_driver->GetCompilerOptions().GetImplicitStackOverflowChecks();
363  if (!skip_overflow_check) {
364    if (generate_explicit_stack_overflow_check) {
365      if (!large_frame) {
366        /* Load stack limit */
367        LockTemp(rs_r12);
368        Load32Disp(rs_rARM_SELF, Thread::StackEndOffset<4>().Int32Value(), rs_r12);
369      }
370    } else {
371      // Implicit stack overflow check.
372      // Generate a load from [sp, #-overflowsize].  If this is in the stack
373      // redzone we will get a segmentation fault.
374      //
375      // Caveat coder: if someone changes the kStackOverflowReservedBytes value
376      // we need to make sure that it's loadable in an immediate field of
377      // a sub instruction.  Otherwise we will get a temp allocation and the
378      // code size will increase.
379      //
380      // This is done before the callee save instructions to avoid any possibility
381      // of these overflowing.  This uses r12 and that's never saved in a callee
382      // save.
383      OpRegRegImm(kOpSub, rs_r12, rs_rARM_SP, GetStackOverflowReservedBytes(kArm));
384      Load32Disp(rs_r12, 0, rs_r12);
385      MarkPossibleStackOverflowException();
386    }
387  }
388  /* Spill core callee saves */
389  NewLIR1(kThumb2Push, core_spill_mask_);
390  /* Need to spill any FP regs? */
391  if (num_fp_spills_) {
392    /*
393     * NOTE: fp spills are a little different from core spills in that
394     * they are pushed as a contiguous block.  When promoting from
395     * the fp set, we must allocate all singles from s16..highest-promoted
396     */
397    NewLIR1(kThumb2VPushCS, num_fp_spills_);
398  }
399
400  const int spill_size = spill_count * 4;
401  const int frame_size_without_spills = frame_size_ - spill_size;
402  if (!skip_overflow_check) {
403    if (generate_explicit_stack_overflow_check) {
404      class StackOverflowSlowPath : public LIRSlowPath {
405       public:
406        StackOverflowSlowPath(Mir2Lir* m2l, LIR* branch, bool restore_lr, size_t sp_displace)
407            : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch, nullptr), restore_lr_(restore_lr),
408              sp_displace_(sp_displace) {
409        }
410        void Compile() OVERRIDE {
411          m2l_->ResetRegPool();
412          m2l_->ResetDefTracking();
413          GenerateTargetLabel(kPseudoThrowTarget);
414          if (restore_lr_) {
415            m2l_->LoadWordDisp(rs_rARM_SP, sp_displace_ - 4, rs_rARM_LR);
416          }
417          m2l_->OpRegImm(kOpAdd, rs_rARM_SP, sp_displace_);
418          m2l_->ClobberCallerSave();
419          ThreadOffset<4> func_offset = QUICK_ENTRYPOINT_OFFSET(4, pThrowStackOverflow);
420          // Load the entrypoint directly into the pc instead of doing a load + branch. Assumes
421          // codegen and target are in thumb2 mode.
422          // NOTE: native pointer.
423          m2l_->LoadWordDisp(rs_rARM_SELF, func_offset.Int32Value(), rs_rARM_PC);
424        }
425
426       private:
427        const bool restore_lr_;
428        const size_t sp_displace_;
429      };
430      if (large_frame) {
431        // Note: may need a temp reg, and we only have r12 free at this point.
432        OpRegRegImm(kOpSub, rs_rARM_LR, rs_rARM_SP, frame_size_without_spills);
433        Load32Disp(rs_rARM_SELF, Thread::StackEndOffset<4>().Int32Value(), rs_r12);
434        LIR* branch = OpCmpBranch(kCondUlt, rs_rARM_LR, rs_r12, nullptr);
435        // Need to restore LR since we used it as a temp.
436        AddSlowPath(new(arena_)StackOverflowSlowPath(this, branch, true, spill_size));
437        OpRegCopy(rs_rARM_SP, rs_rARM_LR);     // Establish stack
438      } else {
439        /*
440         * If the frame is small enough we are guaranteed to have enough space that remains to
441         * handle signals on the user stack.  However, we may not have any free temp
442         * registers at this point, so we'll temporarily add LR to the temp pool.
443         */
444        DCHECK(!GetRegInfo(rs_rARM_LR)->IsTemp());
445        MarkTemp(rs_rARM_LR);
446        FreeTemp(rs_rARM_LR);
447        OpRegRegImm(kOpSub, rs_rARM_SP, rs_rARM_SP, frame_size_without_spills);
448        Clobber(rs_rARM_LR);
449        UnmarkTemp(rs_rARM_LR);
450        LIR* branch = OpCmpBranch(kCondUlt, rs_rARM_SP, rs_r12, nullptr);
451        AddSlowPath(new(arena_)StackOverflowSlowPath(this, branch, false, frame_size_));
452      }
453    } else {
454      // Implicit stack overflow check has already been done.  Just make room on the
455      // stack for the frame now.
456      OpRegImm(kOpSub, rs_rARM_SP, frame_size_without_spills);
457    }
458  } else {
459    OpRegImm(kOpSub, rs_rARM_SP, frame_size_without_spills);
460  }
461
462  FlushIns(ArgLocs, rl_method);
463
464  FreeTemp(rs_r0);
465  FreeTemp(rs_r1);
466  FreeTemp(rs_r2);
467  FreeTemp(rs_r3);
468  FreeTemp(rs_r12);
469}
470
471void ArmMir2Lir::GenExitSequence() {
472  int spill_count = num_core_spills_ + num_fp_spills_;
473  /*
474   * In the exit path, r0/r1 are live - make sure they aren't
475   * allocated by the register utilities as temps.
476   */
477  LockTemp(rs_r0);
478  LockTemp(rs_r1);
479
480  NewLIR0(kPseudoMethodExit);
481  OpRegImm(kOpAdd, rs_rARM_SP, frame_size_ - (spill_count * 4));
482  /* Need to restore any FP callee saves? */
483  if (num_fp_spills_) {
484    NewLIR1(kThumb2VPopCS, num_fp_spills_);
485  }
486  if (core_spill_mask_ & (1 << rs_rARM_LR.GetRegNum())) {
487    /* Unspill rARM_LR to rARM_PC */
488    core_spill_mask_ &= ~(1 << rs_rARM_LR.GetRegNum());
489    core_spill_mask_ |= (1 << rs_rARM_PC.GetRegNum());
490  }
491  NewLIR1(kThumb2Pop, core_spill_mask_);
492  if (!(core_spill_mask_ & (1 << rs_rARM_PC.GetRegNum()))) {
493    /* We didn't pop to rARM_PC, so must do a bv rARM_LR */
494    NewLIR1(kThumbBx, rs_rARM_LR.GetReg());
495  }
496}
497
498void ArmMir2Lir::GenSpecialExitSequence() {
499  NewLIR1(kThumbBx, rs_rARM_LR.GetReg());
500}
501
502}  // namespace art
503