Home | History | Annotate | Download | only in x86
      1 /*
      2  * Copyright (C) 2012 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 X86 ISA */
     18 
     19 #include "codegen_x86.h"
     20 #include "dex/quick/mir_to_lir-inl.h"
     21 #include "gc/accounting/card_table.h"
     22 #include "x86_lir.h"
     23 
     24 namespace art {
     25 
     26 /*
     27  * The sparse table in the literal pool is an array of <key,displacement>
     28  * pairs.
     29  */
     30 void X86Mir2Lir::GenLargeSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
     31   const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
     32   if (cu_->verbose) {
     33     DumpSparseSwitchTable(table);
     34   }
     35   int entries = table[1];
     36   const int32_t* keys = reinterpret_cast<const int32_t*>(&table[2]);
     37   const int32_t* targets = &keys[entries];
     38   rl_src = LoadValue(rl_src, kCoreReg);
     39   for (int i = 0; i < entries; i++) {
     40     int key = keys[i];
     41     BasicBlock* case_block =
     42         mir_graph_->FindBlock(current_dalvik_offset_ + targets[i]);
     43     OpCmpImmBranch(kCondEq, rl_src.reg, key, &block_label_list_[case_block->id]);
     44   }
     45 }
     46 
     47 /*
     48  * Code pattern will look something like:
     49  *
     50  * mov  r_val, ..
     51  * call 0
     52  * pop  r_start_of_method
     53  * sub  r_start_of_method, ..
     54  * mov  r_key_reg, r_val
     55  * sub  r_key_reg, low_key
     56  * cmp  r_key_reg, size-1  ; bound check
     57  * ja   done
     58  * mov  r_disp, [r_start_of_method + r_key_reg * 4 + table_offset]
     59  * add  r_start_of_method, r_disp
     60  * jmp  r_start_of_method
     61  * done:
     62  */
     63 void X86Mir2Lir::GenLargePackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
     64   const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
     65   if (cu_->verbose) {
     66     DumpPackedSwitchTable(table);
     67   }
     68   // Add the table to the list - we'll process it later
     69   SwitchTable* tab_rec =
     70       static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
     71   tab_rec->table = table;
     72   tab_rec->vaddr = current_dalvik_offset_;
     73   int size = table[1];
     74   tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*),
     75                                                       kArenaAllocLIR));
     76   switch_tables_.Insert(tab_rec);
     77 
     78   // Get the switch value
     79   rl_src = LoadValue(rl_src, kCoreReg);
     80   // NewLIR0(kX86Bkpt);
     81 
     82   // Materialize a pointer to the switch table
     83   RegStorage start_of_method_reg;
     84   if (base_of_code_ != nullptr) {
     85     // We can use the saved value.
     86     RegLocation rl_method = mir_graph_->GetRegLocation(base_of_code_->s_reg_low);
     87     if (rl_method.wide) {
     88       rl_method = LoadValueWide(rl_method, kCoreReg);
     89     } else {
     90       rl_method = LoadValue(rl_method, kCoreReg);
     91     }
     92     start_of_method_reg = rl_method.reg;
     93     store_method_addr_used_ = true;
     94   } else {
     95     start_of_method_reg = AllocTempRef();
     96     NewLIR1(kX86StartOfMethod, start_of_method_reg.GetReg());
     97   }
     98   DCHECK_EQ(start_of_method_reg.Is64Bit(), cu_->target64);
     99   int low_key = s4FromSwitchData(&table[2]);
    100   RegStorage keyReg;
    101   // Remove the bias, if necessary
    102   if (low_key == 0) {
    103     keyReg = rl_src.reg;
    104   } else {
    105     keyReg = AllocTemp();
    106     OpRegRegImm(kOpSub, keyReg, rl_src.reg, low_key);
    107   }
    108   // Bounds check - if < 0 or >= size continue following switch
    109   OpRegImm(kOpCmp, keyReg, size - 1);
    110   LIR* branch_over = OpCondBranch(kCondHi, NULL);
    111 
    112   // Load the displacement from the switch table
    113   RegStorage disp_reg = AllocTemp();
    114   NewLIR5(kX86PcRelLoadRA, disp_reg.GetReg(), start_of_method_reg.GetReg(), keyReg.GetReg(),
    115           2, WrapPointer(tab_rec));
    116   // Add displacement to start of method
    117   OpRegReg(kOpAdd, start_of_method_reg, cu_->target64 ? As64BitReg(disp_reg) : disp_reg);
    118   // ..and go!
    119   LIR* switch_branch = NewLIR1(kX86JmpR, start_of_method_reg.GetReg());
    120   tab_rec->anchor = switch_branch;
    121 
    122   /* branch_over target here */
    123   LIR* target = NewLIR0(kPseudoTargetLabel);
    124   branch_over->target = target;
    125 }
    126 
    127 /*
    128  * Array data table format:
    129  *  ushort ident = 0x0300   magic value
    130  *  ushort width            width of each element in the table
    131  *  uint   size             number of elements in the table
    132  *  ubyte  data[size*width] table of data values (may contain a single-byte
    133  *                          padding at the end)
    134  *
    135  * Total size is 4+(width * size + 1)/2 16-bit code units.
    136  */
    137 void X86Mir2Lir::GenFillArrayData(DexOffset table_offset, RegLocation rl_src) {
    138   const uint16_t* table = cu_->insns + current_dalvik_offset_ + table_offset;
    139   // Add the table to the list - we'll process it later
    140   FillArrayData* tab_rec =
    141       static_cast<FillArrayData*>(arena_->Alloc(sizeof(FillArrayData), kArenaAllocData));
    142   tab_rec->table = table;
    143   tab_rec->vaddr = current_dalvik_offset_;
    144   uint16_t width = tab_rec->table[1];
    145   uint32_t size = tab_rec->table[2] | ((static_cast<uint32_t>(tab_rec->table[3])) << 16);
    146   tab_rec->size = (size * width) + 8;
    147 
    148   fill_array_data_.Insert(tab_rec);
    149 
    150   // Making a call - use explicit registers
    151   FlushAllRegs();   /* Everything to home location */
    152   RegStorage array_ptr = TargetReg(kArg0, kRef);
    153   RegStorage payload = TargetPtrReg(kArg1);
    154   RegStorage method_start = TargetPtrReg(kArg2);
    155 
    156   LoadValueDirectFixed(rl_src, array_ptr);
    157   // Materialize a pointer to the fill data image
    158   if (base_of_code_ != nullptr) {
    159     // We can use the saved value.
    160     RegLocation rl_method = mir_graph_->GetRegLocation(base_of_code_->s_reg_low);
    161     if (rl_method.wide) {
    162       LoadValueDirectWide(rl_method, method_start);
    163     } else {
    164       LoadValueDirect(rl_method, method_start);
    165     }
    166     store_method_addr_used_ = true;
    167   } else {
    168     NewLIR1(kX86StartOfMethod, method_start.GetReg());
    169   }
    170   NewLIR2(kX86PcRelAdr, payload.GetReg(), WrapPointer(tab_rec));
    171   OpRegReg(kOpAdd, payload, method_start);
    172   CallRuntimeHelperRegReg(kQuickHandleFillArrayData, array_ptr, payload, true);
    173 }
    174 
    175 void X86Mir2Lir::GenMoveException(RegLocation rl_dest) {
    176   int ex_offset = cu_->target64 ?
    177       Thread::ExceptionOffset<8>().Int32Value() :
    178       Thread::ExceptionOffset<4>().Int32Value();
    179   RegLocation rl_result = EvalLoc(rl_dest, kRefReg, true);
    180   NewLIR2(cu_->target64 ? kX86Mov64RT : kX86Mov32RT, rl_result.reg.GetReg(), ex_offset);
    181   NewLIR2(cu_->target64 ? kX86Mov64TI : kX86Mov32TI, ex_offset, 0);
    182   StoreValue(rl_dest, rl_result);
    183 }
    184 
    185 /*
    186  * Mark garbage collection card. Skip if the value we're storing is null.
    187  */
    188 void X86Mir2Lir::MarkGCCard(RegStorage val_reg, RegStorage tgt_addr_reg) {
    189   DCHECK_EQ(tgt_addr_reg.Is64Bit(), cu_->target64);
    190   DCHECK_EQ(val_reg.Is64Bit(), cu_->target64);
    191   RegStorage reg_card_base = AllocTempRef();
    192   RegStorage reg_card_no = AllocTempRef();
    193   LIR* branch_over = OpCmpImmBranch(kCondEq, val_reg, 0, NULL);
    194   int ct_offset = cu_->target64 ?
    195       Thread::CardTableOffset<8>().Int32Value() :
    196       Thread::CardTableOffset<4>().Int32Value();
    197   NewLIR2(cu_->target64 ? kX86Mov64RT : kX86Mov32RT, reg_card_base.GetReg(), ct_offset);
    198   OpRegRegImm(kOpLsr, reg_card_no, tgt_addr_reg, gc::accounting::CardTable::kCardShift);
    199   StoreBaseIndexed(reg_card_base, reg_card_no, reg_card_base, 0, kUnsignedByte);
    200   LIR* target = NewLIR0(kPseudoTargetLabel);
    201   branch_over->target = target;
    202   FreeTemp(reg_card_base);
    203   FreeTemp(reg_card_no);
    204 }
    205 
    206 void X86Mir2Lir::GenEntrySequence(RegLocation* ArgLocs, RegLocation rl_method) {
    207   /*
    208    * On entry, rX86_ARG0, rX86_ARG1, rX86_ARG2 are live.  Let the register
    209    * allocation mechanism know so it doesn't try to use any of them when
    210    * expanding the frame or flushing.  This leaves the utility
    211    * code with no spare temps.
    212    */
    213   LockTemp(rs_rX86_ARG0);
    214   LockTemp(rs_rX86_ARG1);
    215   LockTemp(rs_rX86_ARG2);
    216 
    217   /*
    218    * We can safely skip the stack overflow check if we're
    219    * a leaf *and* our frame size < fudge factor.
    220    */
    221   InstructionSet isa =  cu_->target64 ? kX86_64 : kX86;
    222   bool skip_overflow_check = mir_graph_->MethodIsLeaf() && !FrameNeedsStackCheck(frame_size_, isa);
    223 
    224   // If we doing an implicit stack overflow check, perform the load immediately
    225   // before the stack pointer is decremented and anything is saved.
    226   if (!skip_overflow_check &&
    227       cu_->compiler_driver->GetCompilerOptions().GetImplicitStackOverflowChecks()) {
    228     // Implicit stack overflow check.
    229     // test eax,[esp + -overflow]
    230     int overflow = GetStackOverflowReservedBytes(isa);
    231     NewLIR3(kX86Test32RM, rs_rAX.GetReg(), rs_rX86_SP.GetReg(), -overflow);
    232     MarkPossibleStackOverflowException();
    233   }
    234 
    235   /* Build frame, return address already on stack */
    236   stack_decrement_ = OpRegImm(kOpSub, rs_rX86_SP, frame_size_ -
    237                               GetInstructionSetPointerSize(cu_->instruction_set));
    238 
    239   NewLIR0(kPseudoMethodEntry);
    240   /* Spill core callee saves */
    241   SpillCoreRegs();
    242   SpillFPRegs();
    243   if (!skip_overflow_check) {
    244     class StackOverflowSlowPath : public LIRSlowPath {
    245      public:
    246       StackOverflowSlowPath(Mir2Lir* m2l, LIR* branch, size_t sp_displace)
    247           : LIRSlowPath(m2l, m2l->GetCurrentDexPc(), branch, nullptr), sp_displace_(sp_displace) {
    248       }
    249       void Compile() OVERRIDE {
    250         m2l_->ResetRegPool();
    251         m2l_->ResetDefTracking();
    252         GenerateTargetLabel(kPseudoThrowTarget);
    253         m2l_->OpRegImm(kOpAdd, rs_rX86_SP, sp_displace_);
    254         m2l_->ClobberCallerSave();
    255         // Assumes codegen and target are in thumb2 mode.
    256         m2l_->CallHelper(RegStorage::InvalidReg(), kQuickThrowStackOverflow,
    257                          false /* MarkSafepointPC */, false /* UseLink */);
    258       }
    259 
    260      private:
    261       const size_t sp_displace_;
    262     };
    263     if (!cu_->compiler_driver->GetCompilerOptions().GetImplicitStackOverflowChecks()) {
    264       // TODO: for large frames we should do something like:
    265       // spill ebp
    266       // lea ebp, [esp + frame_size]
    267       // cmp ebp, fs:[stack_end_]
    268       // jcc stack_overflow_exception
    269       // mov esp, ebp
    270       // in case a signal comes in that's not using an alternate signal stack and the large frame
    271       // may have moved us outside of the reserved area at the end of the stack.
    272       // cmp rs_rX86_SP, fs:[stack_end_]; jcc throw_slowpath
    273       if (cu_->target64) {
    274         OpRegThreadMem(kOpCmp, rs_rX86_SP, Thread::StackEndOffset<8>());
    275       } else {
    276         OpRegThreadMem(kOpCmp, rs_rX86_SP, Thread::StackEndOffset<4>());
    277       }
    278       LIR* branch = OpCondBranch(kCondUlt, nullptr);
    279       AddSlowPath(
    280         new(arena_)StackOverflowSlowPath(this, branch,
    281                                          frame_size_ -
    282                                          GetInstructionSetPointerSize(cu_->instruction_set)));
    283     }
    284   }
    285 
    286   FlushIns(ArgLocs, rl_method);
    287 
    288   if (base_of_code_ != nullptr) {
    289     RegStorage method_start = TargetPtrReg(kArg0);
    290     // We have been asked to save the address of the method start for later use.
    291     setup_method_address_[0] = NewLIR1(kX86StartOfMethod, method_start.GetReg());
    292     int displacement = SRegOffset(base_of_code_->s_reg_low);
    293     // Native pointer - must be natural word size.
    294     setup_method_address_[1] = StoreBaseDisp(rs_rX86_SP, displacement, method_start,
    295                                              cu_->target64 ? k64 : k32, kNotVolatile);
    296   }
    297 
    298   FreeTemp(rs_rX86_ARG0);
    299   FreeTemp(rs_rX86_ARG1);
    300   FreeTemp(rs_rX86_ARG2);
    301 }
    302 
    303 void X86Mir2Lir::GenExitSequence() {
    304   /*
    305    * In the exit path, rX86_RET0/rX86_RET1 are live - make sure they aren't
    306    * allocated by the register utilities as temps.
    307    */
    308   LockTemp(rs_rX86_RET0);
    309   LockTemp(rs_rX86_RET1);
    310 
    311   NewLIR0(kPseudoMethodExit);
    312   UnSpillCoreRegs();
    313   UnSpillFPRegs();
    314   /* Remove frame except for return address */
    315   stack_increment_ = OpRegImm(kOpAdd, rs_rX86_SP, frame_size_ - GetInstructionSetPointerSize(cu_->instruction_set));
    316   NewLIR0(kX86Ret);
    317 }
    318 
    319 void X86Mir2Lir::GenSpecialExitSequence() {
    320   NewLIR0(kX86Ret);
    321 }
    322 
    323 void X86Mir2Lir::GenImplicitNullCheck(RegStorage reg, int opt_flags) {
    324   if (!(cu_->disable_opt & (1 << kNullCheckElimination)) && (opt_flags & MIR_IGNORE_NULL_CHECK)) {
    325     return;
    326   }
    327   // Implicit null pointer check.
    328   // test eax,[arg1+0]
    329   NewLIR3(kX86Test32RM, rs_rAX.GetReg(), reg.GetReg(), 0);
    330   MarkPossibleNullPointerException(opt_flags);
    331 }
    332 
    333 }  // namespace art
    334