Merge "Improve ParallelMoveResolver to work with pairs."
diff --git a/compiler/Android.mk b/compiler/Android.mk
index b87201a..61379fb 100644
--- a/compiler/Android.mk
+++ b/compiler/Android.mk
@@ -100,6 +100,7 @@
optimizing/inliner.cc \
optimizing/instruction_simplifier.cc \
optimizing/intrinsics.cc \
+ optimizing/intrinsics_arm.cc \
optimizing/intrinsics_arm64.cc \
optimizing/intrinsics_x86_64.cc \
optimizing/licm.cc \
diff --git a/compiler/dex/mir_graph.cc b/compiler/dex/mir_graph.cc
index 0f7d45d..93a31e9 100644
--- a/compiler/dex/mir_graph.cc
+++ b/compiler/dex/mir_graph.cc
@@ -113,7 +113,6 @@
entry_block_(NULL),
exit_block_(NULL),
current_code_item_(NULL),
- dex_pc_to_block_map_(arena->Adapter()),
m_units_(arena->Adapter()),
method_stack_(arena->Adapter()),
current_method_(kInvalidEntry),
@@ -268,31 +267,14 @@
DCHECK(insn != orig_block->first_mir_insn);
DCHECK(insn == bottom_block->first_mir_insn);
DCHECK_EQ(insn->offset, bottom_block->start_offset);
- DCHECK_EQ(dex_pc_to_block_map_[insn->offset], orig_block->id);
// Scan the "bottom" instructions, remapping them to the
// newly created "bottom" block.
MIR* p = insn;
p->bb = bottom_block->id;
- dex_pc_to_block_map_[p->offset] = bottom_block->id;
while (p != bottom_block->last_mir_insn) {
p = p->next;
DCHECK(p != nullptr);
p->bb = bottom_block->id;
- int opcode = p->dalvikInsn.opcode;
- /*
- * Some messiness here to ensure that we only enter real opcodes and only the
- * first half of a potentially throwing instruction that has been split into
- * CHECK and work portions. Since the 2nd half of a split operation is always
- * the first in a BasicBlock, we can't hit it here.
- */
- if ((opcode == kMirOpCheck) || !MIR::DecodedInstruction::IsPseudoMirOp(opcode)) {
- BasicBlockId mapped_id = dex_pc_to_block_map_[p->offset];
- // At first glance the instructions should all be mapped to orig_block.
- // However, multiple instructions may correspond to the same dex, hence an earlier
- // instruction may have already moved the mapping for dex to bottom_block.
- DCHECK((mapped_id == orig_block->id) || (mapped_id == bottom_block->id));
- dex_pc_to_block_map_[p->offset] = bottom_block->id;
- }
}
return bottom_block;
@@ -307,12 +289,13 @@
* Utilizes a map for fast lookup of the typical cases.
*/
BasicBlock* MIRGraph::FindBlock(DexOffset code_offset, bool create,
- BasicBlock** immed_pred_block_p) {
+ BasicBlock** immed_pred_block_p,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
if (code_offset >= current_code_item_->insns_size_in_code_units_) {
return nullptr;
}
- int block_id = dex_pc_to_block_map_[code_offset];
+ int block_id = (*dex_pc_to_block_map)[code_offset];
BasicBlock* bb = GetBasicBlock(block_id);
if ((bb != nullptr) && (bb->start_offset == code_offset)) {
@@ -327,19 +310,46 @@
if (bb != nullptr) {
// The target exists somewhere in an existing block.
- return SplitBlock(code_offset, bb, bb == *immed_pred_block_p ? immed_pred_block_p : nullptr);
+ BasicBlock* bottom_block = SplitBlock(code_offset, bb, bb == *immed_pred_block_p ? immed_pred_block_p : nullptr);
+ DCHECK(bottom_block != nullptr);
+ MIR* p = bottom_block->first_mir_insn;
+ BasicBlock* orig_block = bb;
+ DCHECK_EQ((*dex_pc_to_block_map)[p->offset], orig_block->id);
+ // Scan the "bottom" instructions, remapping them to the
+ // newly created "bottom" block.
+ (*dex_pc_to_block_map)[p->offset] = bottom_block->id;
+ while (p != bottom_block->last_mir_insn) {
+ p = p->next;
+ DCHECK(p != nullptr);
+ int opcode = p->dalvikInsn.opcode;
+ /*
+ * Some messiness here to ensure that we only enter real opcodes and only the
+ * first half of a potentially throwing instruction that has been split into
+ * CHECK and work portions. Since the 2nd half of a split operation is always
+ * the first in a BasicBlock, we can't hit it here.
+ */
+ if ((opcode == kMirOpCheck) || !MIR::DecodedInstruction::IsPseudoMirOp(opcode)) {
+ BasicBlockId mapped_id = (*dex_pc_to_block_map)[p->offset];
+ // At first glance the instructions should all be mapped to orig_block.
+ // However, multiple instructions may correspond to the same dex, hence an earlier
+ // instruction may have already moved the mapping for dex to bottom_block.
+ DCHECK((mapped_id == orig_block->id) || (mapped_id == bottom_block->id));
+ (*dex_pc_to_block_map)[p->offset] = bottom_block->id;
+ }
+ }
+ return bottom_block;
}
// Create a new block.
bb = CreateNewBB(kDalvikByteCode);
bb->start_offset = code_offset;
- dex_pc_to_block_map_[bb->start_offset] = bb->id;
+ (*dex_pc_to_block_map)[bb->start_offset] = bb->id;
return bb;
}
/* Identify code range in try blocks and set up the empty catch blocks */
-void MIRGraph::ProcessTryCatchBlocks() {
+void MIRGraph::ProcessTryCatchBlocks(ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
int tries_size = current_code_item_->tries_size_;
DexOffset offset;
@@ -364,7 +374,7 @@
CatchHandlerIterator iterator(handlers_ptr);
for (; iterator.HasNext(); iterator.Next()) {
uint32_t address = iterator.GetHandlerAddress();
- FindBlock(address, true /*create*/, /* immed_pred_block_p */ nullptr);
+ FindBlock(address, true /*create*/, /* immed_pred_block_p */ nullptr, dex_pc_to_block_map);
}
handlers_ptr = iterator.EndDataPointer();
}
@@ -439,7 +449,8 @@
/* Process instructions with the kBranch flag */
BasicBlock* MIRGraph::ProcessCanBranch(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset,
int width, int flags, const uint16_t* code_ptr,
- const uint16_t* code_end) {
+ const uint16_t* code_end,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
DexOffset target = cur_offset;
switch (insn->dalvikInsn.opcode) {
case Instruction::GOTO:
@@ -470,7 +481,8 @@
}
CountBranch(target);
BasicBlock* taken_block = FindBlock(target, /* create */ true,
- /* immed_pred_block_p */ &cur_block);
+ /* immed_pred_block_p */ &cur_block,
+ dex_pc_to_block_map);
cur_block->taken = taken_block->id;
taken_block->predecessors.push_back(cur_block->id);
@@ -480,18 +492,20 @@
/* create */
true,
/* immed_pred_block_p */
- &cur_block);
+ &cur_block,
+ dex_pc_to_block_map);
cur_block->fall_through = fallthrough_block->id;
fallthrough_block->predecessors.push_back(cur_block->id);
} else if (code_ptr < code_end) {
- FindBlock(cur_offset + width, /* create */ true, /* immed_pred_block_p */ nullptr);
+ FindBlock(cur_offset + width, /* create */ true, /* immed_pred_block_p */ nullptr, dex_pc_to_block_map);
}
return cur_block;
}
/* Process instructions with the kSwitch flag */
BasicBlock* MIRGraph::ProcessCanSwitch(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset,
- int width, int flags) {
+ int width, int flags,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
UNUSED(flags);
const uint16_t* switch_data =
reinterpret_cast<const uint16_t*>(GetCurrentInsns() + cur_offset + insn->dalvikInsn.vB);
@@ -545,7 +559,8 @@
for (i = 0; i < size; i++) {
BasicBlock* case_block = FindBlock(cur_offset + target_table[i], /* create */ true,
- /* immed_pred_block_p */ &cur_block);
+ /* immed_pred_block_p */ &cur_block,
+ dex_pc_to_block_map);
SuccessorBlockInfo* successor_block_info =
static_cast<SuccessorBlockInfo*>(arena_->Alloc(sizeof(SuccessorBlockInfo),
kArenaAllocSuccessor));
@@ -559,7 +574,8 @@
/* Fall-through case */
BasicBlock* fallthrough_block = FindBlock(cur_offset + width, /* create */ true,
- /* immed_pred_block_p */ nullptr);
+ /* immed_pred_block_p */ nullptr,
+ dex_pc_to_block_map);
cur_block->fall_through = fallthrough_block->id;
fallthrough_block->predecessors.push_back(cur_block->id);
return cur_block;
@@ -568,7 +584,8 @@
/* Process instructions with the kThrow flag */
BasicBlock* MIRGraph::ProcessCanThrow(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset,
int width, int flags, ArenaBitVector* try_block_addr,
- const uint16_t* code_ptr, const uint16_t* code_end) {
+ const uint16_t* code_ptr, const uint16_t* code_end,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
UNUSED(flags);
bool in_try_block = try_block_addr->IsBitSet(cur_offset);
bool is_throw = (insn->dalvikInsn.opcode == Instruction::THROW);
@@ -585,7 +602,8 @@
for (; iterator.HasNext(); iterator.Next()) {
BasicBlock* catch_block = FindBlock(iterator.GetHandlerAddress(), false /* create */,
- nullptr /* immed_pred_block_p */);
+ nullptr /* immed_pred_block_p */,
+ dex_pc_to_block_map);
if (insn->dalvikInsn.opcode == Instruction::MONITOR_EXIT &&
IsBadMonitorExitCatch(insn->offset, catch_block->start_offset)) {
// Don't allow monitor-exit to catch its own exception, http://b/15745363 .
@@ -620,7 +638,7 @@
cur_block->explicit_throw = true;
if (code_ptr < code_end) {
// Force creation of new block following THROW via side-effect.
- FindBlock(cur_offset + width, /* create */ true, /* immed_pred_block_p */ nullptr);
+ FindBlock(cur_offset + width, /* create */ true, /* immed_pred_block_p */ nullptr, dex_pc_to_block_map);
}
if (!in_try_block) {
// Don't split a THROW that can't rethrow - we're done.
@@ -652,7 +670,7 @@
* not automatically terminated after the work portion, and may
* contain following instructions.
*
- * Note also that the dex_pc_to_block_map_ entry for the potentially
+ * Note also that the dex_pc_to_block_map entry for the potentially
* throwing instruction will refer to the original basic block.
*/
BasicBlock* new_block = CreateNewBB(kDalvikByteCode);
@@ -687,7 +705,11 @@
// TODO: need to rework expansion of block list & try_block_addr when inlining activated.
// TUNING: use better estimate of basic blocks for following resize.
block_list_.reserve(block_list_.size() + current_code_item_->insns_size_in_code_units_);
- dex_pc_to_block_map_.resize(dex_pc_to_block_map_.size() + current_code_item_->insns_size_in_code_units_);
+ // FindBlock lookup cache.
+ ScopedArenaAllocator allocator(&cu_->arena_stack);
+ ScopedArenaVector<uint16_t> dex_pc_to_block_map(allocator.Adapter());
+ dex_pc_to_block_map.resize(dex_pc_to_block_map.size() +
+ current_code_item_->insns_size_in_code_units_);
// TODO: replace with explicit resize routine. Using automatic extension side effect for now.
try_block_addr_->SetBit(current_code_item_->insns_size_in_code_units_);
@@ -728,7 +750,7 @@
cur_block->predecessors.push_back(entry_block_->id);
/* Identify code range in try blocks and set up the empty catch blocks */
- ProcessTryCatchBlocks();
+ ProcessTryCatchBlocks(&dex_pc_to_block_map);
uint64_t merged_df_flags = 0u;
@@ -777,20 +799,21 @@
DCHECK(cur_block->taken == NullBasicBlockId);
// Unreachable instruction, mark for no continuation and end basic block.
flags &= ~Instruction::kContinue;
- FindBlock(current_offset_ + width, /* create */ true, /* immed_pred_block_p */ nullptr);
+ FindBlock(current_offset_ + width, /* create */ true,
+ /* immed_pred_block_p */ nullptr, &dex_pc_to_block_map);
}
} else {
cur_block->AppendMIR(insn);
}
// Associate the starting dex_pc for this opcode with its containing basic block.
- dex_pc_to_block_map_[insn->offset] = cur_block->id;
+ dex_pc_to_block_map[insn->offset] = cur_block->id;
code_ptr += width;
if (flags & Instruction::kBranch) {
cur_block = ProcessCanBranch(cur_block, insn, current_offset_,
- width, flags, code_ptr, code_end);
+ width, flags, code_ptr, code_end, &dex_pc_to_block_map);
} else if (flags & Instruction::kReturn) {
cur_block->terminated_by_return = true;
cur_block->fall_through = exit_block_->id;
@@ -804,13 +827,15 @@
* Create a fallthrough block for real instructions
* (incl. NOP).
*/
- FindBlock(current_offset_ + width, /* create */ true, /* immed_pred_block_p */ nullptr);
+ FindBlock(current_offset_ + width, /* create */ true,
+ /* immed_pred_block_p */ nullptr, &dex_pc_to_block_map);
}
} else if (flags & Instruction::kThrow) {
cur_block = ProcessCanThrow(cur_block, insn, current_offset_, width, flags, try_block_addr_,
- code_ptr, code_end);
+ code_ptr, code_end, &dex_pc_to_block_map);
} else if (flags & Instruction::kSwitch) {
- cur_block = ProcessCanSwitch(cur_block, insn, current_offset_, width, flags);
+ cur_block = ProcessCanSwitch(cur_block, insn, current_offset_, width,
+ flags, &dex_pc_to_block_map);
}
if (verify_flags & Instruction::kVerifyVarArgRange ||
verify_flags & Instruction::kVerifyVarArgRangeNonZero) {
@@ -828,7 +853,8 @@
}
current_offset_ += width;
BasicBlock* next_block = FindBlock(current_offset_, /* create */ false,
- /* immed_pred_block_p */ nullptr);
+ /* immed_pred_block_p */ nullptr,
+ &dex_pc_to_block_map);
if (next_block) {
/*
* The next instruction could be the target of a previously parsed
diff --git a/compiler/dex/mir_graph.h b/compiler/dex/mir_graph.h
index 5def191..5914245 100644
--- a/compiler/dex/mir_graph.h
+++ b/compiler/dex/mir_graph.h
@@ -542,8 +542,9 @@
uint32_t method_idx, jobject class_loader, const DexFile& dex_file);
/* Find existing block */
- BasicBlock* FindBlock(DexOffset code_offset) {
- return FindBlock(code_offset, false, NULL);
+ BasicBlock* FindBlock(DexOffset code_offset,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map) {
+ return FindBlock(code_offset, false, nullptr, dex_pc_to_block_map);
}
const uint16_t* GetCurrentInsns() const {
@@ -1249,16 +1250,20 @@
bool ContentIsInsn(const uint16_t* code_ptr);
BasicBlock* SplitBlock(DexOffset code_offset, BasicBlock* orig_block,
BasicBlock** immed_pred_block_p);
- BasicBlock* FindBlock(DexOffset code_offset, bool create, BasicBlock** immed_pred_block_p);
- void ProcessTryCatchBlocks();
+ BasicBlock* FindBlock(DexOffset code_offset, bool create, BasicBlock** immed_pred_block_p,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map);
+ void ProcessTryCatchBlocks(ScopedArenaVector<uint16_t>* dex_pc_to_block_map);
bool IsBadMonitorExitCatch(NarrowDexOffset monitor_exit_offset, NarrowDexOffset catch_offset);
BasicBlock* ProcessCanBranch(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset, int width,
- int flags, const uint16_t* code_ptr, const uint16_t* code_end);
+ int flags, const uint16_t* code_ptr, const uint16_t* code_end,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map);
BasicBlock* ProcessCanSwitch(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset, int width,
- int flags);
+ int flags,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map);
BasicBlock* ProcessCanThrow(BasicBlock* cur_block, MIR* insn, DexOffset cur_offset, int width,
int flags, ArenaBitVector* try_block_addr, const uint16_t* code_ptr,
- const uint16_t* code_end);
+ const uint16_t* code_end,
+ ScopedArenaVector<uint16_t>* dex_pc_to_block_map);
int AddNewSReg(int v_reg);
void HandleSSAUse(int* uses, int dalvik_reg, int reg_index);
void DataFlowSSAFormat35C(MIR* mir);
@@ -1391,7 +1396,6 @@
BasicBlock* entry_block_;
BasicBlock* exit_block_;
const DexFile::CodeItem* current_code_item_;
- ArenaVector<uint16_t> dex_pc_to_block_map_; // FindBlock lookup cache.
ArenaVector<DexCompilationUnit*> m_units_; // List of methods included in this graph
typedef std::pair<int, int> MIRLocation; // Insert point, (m_unit_ index, offset)
ArenaVector<MIRLocation> method_stack_; // Include stack
diff --git a/compiler/dex/quick/arm/call_arm.cc b/compiler/dex/quick/arm/call_arm.cc
index f15b727..1b5dde2 100644
--- a/compiler/dex/quick/arm/call_arm.cc
+++ b/compiler/dex/quick/arm/call_arm.cc
@@ -52,16 +52,13 @@
*/
void ArmMir2Lir::GenLargeSparseSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpSparseSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable *tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
uint32_t size = table[1];
- tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
@@ -100,17 +97,13 @@
void ArmMir2Lir::GenLargePackedSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpPackedSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable *tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
uint32_t size = table[1];
- tab_rec->targets =
- static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
diff --git a/compiler/dex/quick/arm64/call_arm64.cc b/compiler/dex/quick/arm64/call_arm64.cc
index 6492442..d1e4b7e 100644
--- a/compiler/dex/quick/arm64/call_arm64.cc
+++ b/compiler/dex/quick/arm64/call_arm64.cc
@@ -51,16 +51,13 @@
*/
void Arm64Mir2Lir::GenLargeSparseSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpSparseSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable *tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
uint32_t size = table[1];
- tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
@@ -103,17 +100,13 @@
void Arm64Mir2Lir::GenLargePackedSwitch(MIR* mir, uint32_t table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpPackedSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable *tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
uint32_t size = table[1];
- tab_rec->targets =
- static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*), kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
@@ -400,9 +393,8 @@
}
static bool Arm64UseRelativeCall(CompilationUnit* cu, const MethodReference& target_method) {
- UNUSED(cu, target_method);
- // Always emit relative calls.
- return true;
+ // Emit relative calls anywhere in the image or within a dex file otherwise.
+ return cu->compiler_driver->IsImage() || cu->dex_file == target_method.dex_file;
}
/*
diff --git a/compiler/dex/quick/codegen_util.cc b/compiler/dex/quick/codegen_util.cc
index 04113db..88a4605 100644
--- a/compiler/dex/quick/codegen_util.cc
+++ b/compiler/dex/quick/codegen_util.cc
@@ -557,29 +557,49 @@
LOG(INFO) << "Switch table for offset 0x" << std::hex << bx_offset;
}
if (tab_rec->table[0] == Instruction::kSparseSwitchSignature) {
- const int32_t* keys = reinterpret_cast<const int32_t*>(&(tab_rec->table[2]));
- for (int elems = 0; elems < tab_rec->table[1]; elems++) {
- int disp = tab_rec->targets[elems]->offset - bx_offset;
+ DCHECK(tab_rec->switch_mir != nullptr);
+ BasicBlock* bb = mir_graph_->GetBasicBlock(tab_rec->switch_mir->bb);
+ DCHECK(bb != nullptr);
+ int elems = 0;
+ for (SuccessorBlockInfo* successor_block_info : bb->successor_blocks) {
+ int key = successor_block_info->key;
+ int target = successor_block_info->block;
+ LIR* boundary_lir = InsertCaseLabel(target, key);
+ DCHECK(boundary_lir != nullptr);
+ int disp = boundary_lir->offset - bx_offset;
+ Push32(code_buffer_, key);
+ Push32(code_buffer_, disp);
if (cu_->verbose) {
LOG(INFO) << " Case[" << elems << "] key: 0x"
- << std::hex << keys[elems] << ", disp: 0x"
+ << std::hex << key << ", disp: 0x"
<< std::hex << disp;
}
- Push32(code_buffer_, keys[elems]);
- Push32(code_buffer_,
- tab_rec->targets[elems]->offset - bx_offset);
+ elems++;
}
+ DCHECK_EQ(elems, tab_rec->table[1]);
} else {
DCHECK_EQ(static_cast<int>(tab_rec->table[0]),
static_cast<int>(Instruction::kPackedSwitchSignature));
- for (int elems = 0; elems < tab_rec->table[1]; elems++) {
- int disp = tab_rec->targets[elems]->offset - bx_offset;
+ DCHECK(tab_rec->switch_mir != nullptr);
+ BasicBlock* bb = mir_graph_->GetBasicBlock(tab_rec->switch_mir->bb);
+ DCHECK(bb != nullptr);
+ int elems = 0;
+ int low_key = s4FromSwitchData(&tab_rec->table[2]);
+ for (SuccessorBlockInfo* successor_block_info : bb->successor_blocks) {
+ int key = successor_block_info->key;
+ DCHECK_EQ(elems + low_key, key);
+ int target = successor_block_info->block;
+ LIR* boundary_lir = InsertCaseLabel(target, key);
+ DCHECK(boundary_lir != nullptr);
+ int disp = boundary_lir->offset - bx_offset;
+ Push32(code_buffer_, disp);
if (cu_->verbose) {
LOG(INFO) << " Case[" << elems << "] disp: 0x"
<< std::hex << disp;
}
- Push32(code_buffer_, tab_rec->targets[elems]->offset - bx_offset);
+ elems++;
}
+ DCHECK_EQ(elems, tab_rec->table[1]);
}
}
}
@@ -830,13 +850,15 @@
* branch table during the assembly phase. All resource flags
* are set to prevent code motion. KeyVal is just there for debugging.
*/
-LIR* Mir2Lir::InsertCaseLabel(DexOffset vaddr, int keyVal) {
- LIR* boundary_lir = &block_label_list_[mir_graph_->FindBlock(vaddr)->id];
+LIR* Mir2Lir::InsertCaseLabel(uint32_t bbid, int keyVal) {
+ LIR* boundary_lir = &block_label_list_[bbid];
LIR* res = boundary_lir;
if (cu_->verbose) {
// Only pay the expense if we're pretty-printing.
LIR* new_label = static_cast<LIR*>(arena_->Alloc(sizeof(LIR), kArenaAllocLIR));
- new_label->dalvik_offset = vaddr;
+ BasicBlock* bb = mir_graph_->GetBasicBlock(bbid);
+ DCHECK(bb != nullptr);
+ new_label->dalvik_offset = bb->start_offset;
new_label->opcode = kPseudoCaseLabel;
new_label->operands[0] = keyVal;
new_label->flags.fixup = kFixupLabel;
@@ -848,40 +870,6 @@
return res;
}
-void Mir2Lir::MarkPackedCaseLabels(Mir2Lir::SwitchTable* tab_rec) {
- const uint16_t* table = tab_rec->table;
- DexOffset base_vaddr = tab_rec->vaddr;
- const int32_t *targets = reinterpret_cast<const int32_t*>(&table[4]);
- int entries = table[1];
- int low_key = s4FromSwitchData(&table[2]);
- for (int i = 0; i < entries; i++) {
- tab_rec->targets[i] = InsertCaseLabel(base_vaddr + targets[i], i + low_key);
- }
-}
-
-void Mir2Lir::MarkSparseCaseLabels(Mir2Lir::SwitchTable* tab_rec) {
- const uint16_t* table = tab_rec->table;
- DexOffset base_vaddr = tab_rec->vaddr;
- int entries = table[1];
- const int32_t* keys = reinterpret_cast<const int32_t*>(&table[2]);
- const int32_t* targets = &keys[entries];
- for (int i = 0; i < entries; i++) {
- tab_rec->targets[i] = InsertCaseLabel(base_vaddr + targets[i], keys[i]);
- }
-}
-
-void Mir2Lir::ProcessSwitchTables() {
- for (Mir2Lir::SwitchTable* tab_rec : switch_tables_) {
- if (tab_rec->table[0] == Instruction::kPackedSwitchSignature) {
- MarkPackedCaseLabels(tab_rec);
- } else if (tab_rec->table[0] == Instruction::kSparseSwitchSignature) {
- MarkSparseCaseLabels(tab_rec);
- } else {
- LOG(FATAL) << "Invalid switch table";
- }
- }
-}
-
void Mir2Lir::DumpSparseSwitchTable(const uint16_t* table) {
/*
* Sparse switch data format:
@@ -1032,9 +1020,6 @@
/* Method is not empty */
if (first_lir_insn_) {
- // mark the targets of switch statement case labels
- ProcessSwitchTables();
-
/* Convert LIR into machine code. */
AssembleLIR();
diff --git a/compiler/dex/quick/mips/call_mips.cc b/compiler/dex/quick/mips/call_mips.cc
index ccfdaf6..0719b52 100644
--- a/compiler/dex/quick/mips/call_mips.cc
+++ b/compiler/dex/quick/mips/call_mips.cc
@@ -68,17 +68,13 @@
*/
void MipsMir2Lir::GenLargeSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpSparseSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable* tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
int elements = table[1];
- tab_rec->targets =
- static_cast<LIR**>(arena_->Alloc(elements * sizeof(LIR*), kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// The table is composed of 8-byte key/disp pairs
@@ -145,17 +141,13 @@
*/
void MipsMir2Lir::GenLargePackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
- if (cu_->verbose) {
- DumpPackedSwitchTable(table);
- }
// Add the table to the list - we'll process it later
SwitchTable* tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
int size = table[1];
- tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*),
- kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
diff --git a/compiler/dex/quick/mir_to_lir.h b/compiler/dex/quick/mir_to_lir.h
index 888c34e..9f1a497 100644
--- a/compiler/dex/quick/mir_to_lir.h
+++ b/compiler/dex/quick/mir_to_lir.h
@@ -224,7 +224,7 @@
struct SwitchTable : EmbeddedData {
LIR* anchor; // Reference instruction for relative offsets.
- LIR** targets; // Array of case targets.
+ MIR* switch_mir; // The switch mir.
};
/* Static register use counts */
@@ -653,7 +653,6 @@
LIR* ScanLiteralPoolClass(LIR* data_target, const DexFile& dex_file, uint32_t type_idx);
LIR* AddWordData(LIR* *constant_list_p, int value);
LIR* AddWideData(LIR* *constant_list_p, int val_lo, int val_hi);
- void ProcessSwitchTables();
void DumpSparseSwitchTable(const uint16_t* table);
void DumpPackedSwitchTable(const uint16_t* table);
void MarkBoundary(DexOffset offset, const char* inst_str);
@@ -671,9 +670,7 @@
int AssignLiteralOffset(CodeOffset offset);
int AssignSwitchTablesOffset(CodeOffset offset);
int AssignFillArrayDataOffset(CodeOffset offset);
- virtual LIR* InsertCaseLabel(DexOffset vaddr, int keyVal);
- virtual void MarkPackedCaseLabels(Mir2Lir::SwitchTable* tab_rec);
- void MarkSparseCaseLabels(Mir2Lir::SwitchTable* tab_rec);
+ LIR* InsertCaseLabel(uint32_t bbid, int keyVal);
// Handle bookkeeping to convert a wide RegLocation to a narrow RegLocation. No code generated.
virtual RegLocation NarrowRegLoc(RegLocation loc);
diff --git a/compiler/dex/quick/x86/call_x86.cc b/compiler/dex/quick/x86/call_x86.cc
index aa0972f..284e8f6 100644
--- a/compiler/dex/quick/x86/call_x86.cc
+++ b/compiler/dex/quick/x86/call_x86.cc
@@ -37,84 +37,6 @@
}
/*
- * We override InsertCaseLabel, because the first parameter represents
- * a basic block id, instead of a dex offset.
- */
-LIR* X86Mir2Lir::InsertCaseLabel(DexOffset bbid, int keyVal) {
- LIR* boundary_lir = &block_label_list_[bbid];
- LIR* res = boundary_lir;
- if (cu_->verbose) {
- // Only pay the expense if we're pretty-printing.
- LIR* new_label = static_cast<LIR*>(arena_->Alloc(sizeof(LIR), kArenaAllocLIR));
- BasicBlock* bb = mir_graph_->GetBasicBlock(bbid);
- DCHECK(bb != nullptr);
- new_label->dalvik_offset = bb->start_offset;;
- new_label->opcode = kPseudoCaseLabel;
- new_label->operands[0] = keyVal;
- new_label->flags.fixup = kFixupLabel;
- DCHECK(!new_label->flags.use_def_invalid);
- new_label->u.m.def_mask = &kEncodeAll;
- InsertLIRAfter(boundary_lir, new_label);
- res = new_label;
- }
- return res;
-}
-
-void X86Mir2Lir::MarkPackedCaseLabels(Mir2Lir::SwitchTable* tab_rec) {
- const uint16_t* table = tab_rec->table;
- const int32_t *targets = reinterpret_cast<const int32_t*>(&table[4]);
- int entries = table[1];
- int low_key = s4FromSwitchData(&table[2]);
- for (int i = 0; i < entries; i++) {
- // The value at targets[i] is a basic block id, instead of a dex offset.
- tab_rec->targets[i] = InsertCaseLabel(targets[i], i + low_key);
- }
-}
-
-/*
- * We convert and create a new packed switch table that stores
- * basic block ids to targets[] by examining successor blocks.
- * Note that the original packed switch table stores dex offsets to targets[].
- */
-const uint16_t* X86Mir2Lir::ConvertPackedSwitchTable(MIR* mir, const uint16_t* table) {
- /*
- * The original packed switch data format:
- * ushort ident = 0x0100 magic value
- * ushort size number of entries in the table
- * int first_key first (and lowest) switch case value
- * int targets[size] branch targets, relative to switch opcode
- *
- * Total size is (4+size*2) 16-bit code units.
- *
- * Note that the new packed switch data format is the same as the original
- * format, except that targets[] are basic block ids.
- *
- */
- BasicBlock* bb = mir_graph_->GetBasicBlock(mir->bb);
- DCHECK(bb != nullptr);
- // Get the number of entries.
- int entries = table[1];
- const int32_t* as_int32 = reinterpret_cast<const int32_t*>(&table[2]);
- int32_t starting_key = as_int32[0];
- // Create a new table.
- int size = sizeof(uint16_t) * (4 + entries * 2);
- uint16_t* new_table = reinterpret_cast<uint16_t*>(arena_->Alloc(size, kArenaAllocMisc));
- // Copy ident, size, and first_key to the new table.
- memcpy(new_table, table, sizeof(uint16_t) * 4);
- // Get the new targets.
- int32_t* new_targets = reinterpret_cast<int32_t*>(&new_table[4]);
- // Find out targets for each entry.
- int i = 0;
- for (SuccessorBlockInfo* successor_block_info : bb->successor_blocks) {
- DCHECK_EQ(starting_key + i, successor_block_info->key);
- // Save target basic block id.
- new_targets[i++] = successor_block_info->block;
- }
- DCHECK_EQ(i, entries);
- return new_table;
-}
-
-/*
* Code pattern will look something like:
*
* mov r_val, ..
@@ -131,16 +53,14 @@
* done:
*/
void X86Mir2Lir::GenLargePackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) {
- const uint16_t* old_table = mir_graph_->GetTable(mir, table_offset);
- const uint16_t* table = ConvertPackedSwitchTable(mir, old_table);
+ const uint16_t* table = mir_graph_->GetTable(mir, table_offset);
// Add the table to the list - we'll process it later
SwitchTable* tab_rec =
static_cast<SwitchTable*>(arena_->Alloc(sizeof(SwitchTable), kArenaAllocData));
+ tab_rec->switch_mir = mir;
tab_rec->table = table;
tab_rec->vaddr = current_dalvik_offset_;
int size = table[1];
- tab_rec->targets = static_cast<LIR**>(arena_->Alloc(size * sizeof(LIR*),
- kArenaAllocLIR));
switch_tables_.push_back(tab_rec);
// Get the switch value
diff --git a/compiler/dex/quick/x86/codegen_x86.h b/compiler/dex/quick/x86/codegen_x86.h
index 811d4f5..ca60400 100644
--- a/compiler/dex/quick/x86/codegen_x86.h
+++ b/compiler/dex/quick/x86/codegen_x86.h
@@ -271,11 +271,8 @@
int first_bit, int second_bit) OVERRIDE;
void GenNegDouble(RegLocation rl_dest, RegLocation rl_src) OVERRIDE;
void GenNegFloat(RegLocation rl_dest, RegLocation rl_src) OVERRIDE;
- const uint16_t* ConvertPackedSwitchTable(MIR* mir, const uint16_t* table);
void GenLargePackedSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) OVERRIDE;
void GenLargeSparseSwitch(MIR* mir, DexOffset table_offset, RegLocation rl_src) OVERRIDE;
- LIR* InsertCaseLabel(DexOffset vaddr, int keyVal) OVERRIDE;
- void MarkPackedCaseLabels(Mir2Lir::SwitchTable* tab_rec) OVERRIDE;
/**
* @brief Implement instanceof a final class with x86 specific code.
diff --git a/compiler/driver/compiler_driver.cc b/compiler/driver/compiler_driver.cc
index 2d8c9d4..b8a8936 100644
--- a/compiler/driver/compiler_driver.cc
+++ b/compiler/driver/compiler_driver.cc
@@ -1285,7 +1285,15 @@
*stats_flags |= kFlagDirectCallToBoot | kFlagDirectMethodToBoot;
}
if (!use_dex_cache && force_relocations) {
- if (!IsImage() || !IsImageClass(method->GetDeclaringClassDescriptor())) {
+ bool is_in_image;
+ if (IsImage()) {
+ is_in_image = IsImageClass(method->GetDeclaringClassDescriptor());
+ } else {
+ is_in_image = instruction_set_ != kX86 && instruction_set_ != kX86_64 &&
+ Runtime::Current()->GetHeap()->FindSpaceFromObject(method->GetDeclaringClass(),
+ false)->IsImageSpace();
+ }
+ if (!is_in_image) {
// We can only branch directly to Methods that are resolved in the DexCache.
// Otherwise we won't invoke the resolution trampoline.
use_dex_cache = true;
diff --git a/compiler/image_writer.cc b/compiler/image_writer.cc
index b234249..670b76c 100644
--- a/compiler/image_writer.cc
+++ b/compiler/image_writer.cc
@@ -273,13 +273,7 @@
void ImageWriter::AssignImageBinSlot(mirror::Object* object) {
DCHECK(object != nullptr);
- size_t object_size;
- if (object->IsArtMethod()) {
- // Methods are sized based on the target pointer size.
- object_size = mirror::ArtMethod::InstanceSize(target_ptr_size_);
- } else {
- object_size = object->SizeOf();
- }
+ size_t object_size = object->SizeOf();
// The magic happens here. We segregate objects into different bins based
// on how likely they are to get dirty at runtime.
@@ -931,7 +925,7 @@
if (obj->IsArtMethod()) {
// Size without pointer fields since we don't want to overrun the buffer if target art method
// is 32 bits but source is 64 bits.
- n = mirror::ArtMethod::SizeWithoutPointerFields(sizeof(void*));
+ n = mirror::ArtMethod::SizeWithoutPointerFields(image_writer->target_ptr_size_);
} else {
n = obj->SizeOf();
}
@@ -1016,10 +1010,6 @@
}
if (orig->IsArtMethod<kVerifyNone>()) {
FixupMethod(orig->AsArtMethod<kVerifyNone>(), down_cast<ArtMethod*>(copy));
- } else if (orig->IsClass() && orig->AsClass()->IsArtMethodClass()) {
- // Set the right size for the target.
- size_t size = mirror::ArtMethod::InstanceSize(target_ptr_size_);
- down_cast<mirror::Class*>(copy)->SetObjectSizeWithoutChecks(size);
}
}
@@ -1031,7 +1021,9 @@
// trampoline.
// Quick entrypoint:
- const uint8_t* quick_code = GetOatAddress(method->GetQuickOatCodeOffset());
+ uint32_t quick_oat_code_offset = PointerToLowMemUInt32(
+ method->GetEntryPointFromQuickCompiledCodePtrSize(target_ptr_size_));
+ const uint8_t* quick_code = GetOatAddress(quick_oat_code_offset);
*quick_is_interpreted = false;
if (quick_code != nullptr &&
(!method->IsStatic() || method->IsConstructor() || method->GetDeclaringClass()->IsInitialized())) {
@@ -1082,11 +1074,12 @@
// locations.
// Copy all of the fields from the runtime methods to the target methods first since we did a
// bytewise copy earlier.
- copy->SetEntryPointFromInterpreterPtrSize<kVerifyNone>(orig->GetEntryPointFromInterpreter(),
- target_ptr_size_);
- copy->SetEntryPointFromJniPtrSize<kVerifyNone>(orig->GetEntryPointFromJni(), target_ptr_size_);
+ copy->SetEntryPointFromInterpreterPtrSize<kVerifyNone>(
+ orig->GetEntryPointFromInterpreterPtrSize(target_ptr_size_), target_ptr_size_);
+ copy->SetEntryPointFromJniPtrSize<kVerifyNone>(
+ orig->GetEntryPointFromJniPtrSize(target_ptr_size_), target_ptr_size_);
copy->SetEntryPointFromQuickCompiledCodePtrSize<kVerifyNone>(
- orig->GetEntryPointFromQuickCompiledCode(), target_ptr_size_);
+ orig->GetEntryPointFromQuickCompiledCodePtrSize(target_ptr_size_), target_ptr_size_);
// The resolution method has a special trampoline to call.
Runtime* runtime = Runtime::Current();
diff --git a/compiler/oat_writer.cc b/compiler/oat_writer.cc
index 3c36ffa..9c0157e 100644
--- a/compiler/oat_writer.cc
+++ b/compiler/oat_writer.cc
@@ -899,7 +899,8 @@
class OatWriter::InitImageMethodVisitor : public OatDexMethodVisitor {
public:
InitImageMethodVisitor(OatWriter* writer, size_t offset)
- : OatDexMethodVisitor(writer, offset) {
+ : OatDexMethodVisitor(writer, offset),
+ pointer_size_(GetInstructionSetPointerSize(writer_->compiler_driver_->GetInstructionSet())) {
}
bool VisitMethod(size_t class_def_method_index, const ClassDataItemIterator& it)
@@ -932,10 +933,14 @@
std::string dump = exc->Dump();
LOG(FATAL) << dump;
}
- method->SetQuickOatCodeOffset(offsets.code_offset_);
+ method->SetEntryPointFromQuickCompiledCodePtrSize(reinterpret_cast<void*>(offsets.code_offset_),
+ pointer_size_);
return true;
}
+
+ protected:
+ const size_t pointer_size_;
};
class OatWriter::WriteCodeMethodVisitor : public OatDexMethodVisitor {
@@ -1103,10 +1108,18 @@
if (UNLIKELY(target_offset == 0)) {
mirror::ArtMethod* target = GetTargetMethod(patch);
DCHECK(target != nullptr);
- DCHECK_EQ(target->GetQuickOatCodeOffset(), 0u);
- target_offset = target->IsNative()
- ? writer_->oat_header_->GetQuickGenericJniTrampolineOffset()
- : writer_->oat_header_->GetQuickToInterpreterBridgeOffset();
+ size_t size = GetInstructionSetPointerSize(writer_->compiler_driver_->GetInstructionSet());
+ const void* oat_code_offset = target->GetEntryPointFromQuickCompiledCodePtrSize(size);
+ if (oat_code_offset != 0) {
+ DCHECK(!Runtime::Current()->GetClassLinker()->IsQuickResolutionStub(oat_code_offset));
+ DCHECK(!Runtime::Current()->GetClassLinker()->IsQuickToInterpreterBridge(oat_code_offset));
+ DCHECK(!Runtime::Current()->GetClassLinker()->IsQuickGenericJniStub(oat_code_offset));
+ target_offset = PointerToLowMemUInt32(oat_code_offset);
+ } else {
+ target_offset = target->IsNative()
+ ? writer_->oat_header_->GetQuickGenericJniTrampolineOffset()
+ : writer_->oat_header_->GetQuickToInterpreterBridgeOffset();
+ }
}
return target_offset;
}
@@ -1138,10 +1151,9 @@
void PatchCodeAddress(std::vector<uint8_t>* code, uint32_t offset, uint32_t target_offset)
SHARED_LOCKS_REQUIRED(Locks::mutator_lock_) {
- // NOTE: Direct calls across oat files don't use linker patches.
- DCHECK(writer_->image_writer_ != nullptr);
- uint32_t address = PointerToLowMemUInt32(writer_->image_writer_->GetOatFileBegin() +
- writer_->oat_data_offset_ + target_offset);
+ uint32_t address = writer_->image_writer_ == nullptr ? target_offset :
+ PointerToLowMemUInt32(writer_->image_writer_->GetOatFileBegin() +
+ writer_->oat_data_offset_ + target_offset);
DCHECK_LE(offset + 4, code->size());
uint8_t* data = &(*code)[offset];
data[0] = address & 0xffu;
diff --git a/compiler/optimizing/bounds_check_elimination.cc b/compiler/optimizing/bounds_check_elimination.cc
index d6c3515..bcee563 100644
--- a/compiler/optimizing/bounds_check_elimination.cc
+++ b/compiler/optimizing/bounds_check_elimination.cc
@@ -28,10 +28,10 @@
*/
class ValueBound : public ValueObject {
public:
- ValueBound(HInstruction* instruction, int constant) {
+ ValueBound(HInstruction* instruction, int32_t constant) {
if (instruction != nullptr && instruction->IsIntConstant()) {
- // Normalizing ValueBound with constant instruction.
- int instr_const = instruction->AsIntConstant()->GetValue();
+ // Normalize ValueBound with constant instruction.
+ int32_t instr_const = instruction->AsIntConstant()->GetValue();
if (constant >= 0 && (instr_const <= INT_MAX - constant)) {
// No overflow.
instruction_ = nullptr;
@@ -49,6 +49,25 @@
constant_ = constant;
}
+ static bool IsAddOrSubAConstant(HInstruction* instruction,
+ HInstruction** left_instruction,
+ int* right_constant) {
+ if (instruction->IsAdd() || instruction->IsSub()) {
+ HBinaryOperation* bin_op = instruction->AsBinaryOperation();
+ HInstruction* left = bin_op->GetLeft();
+ HInstruction* right = bin_op->GetRight();
+ if (right->IsIntConstant()) {
+ *left_instruction = left;
+ int32_t c = right->AsIntConstant()->GetValue();
+ *right_constant = instruction->IsAdd() ? c : -c;
+ return true;
+ }
+ }
+ *left_instruction = nullptr;
+ *right_constant = 0;
+ return false;
+ }
+
// Try to detect useful value bound format from an instruction, e.g.
// a constant or array length related value.
static ValueBound DetectValueBoundFromValue(HInstruction* instruction, bool* found) {
@@ -63,13 +82,12 @@
return ValueBound(instruction, 0);
}
// Try to detect (array.length + c) format.
- if (instruction->IsAdd()) {
- HAdd* add = instruction->AsAdd();
- HInstruction* left = add->GetLeft();
- HInstruction* right = add->GetRight();
- if (left->IsArrayLength() && right->IsIntConstant()) {
+ HInstruction *left;
+ int32_t right;
+ if (IsAddOrSubAConstant(instruction, &left, &right)) {
+ if (left->IsArrayLength()) {
*found = true;
- return ValueBound(left, right->AsIntConstant()->GetValue());
+ return ValueBound(left, right);
}
}
@@ -79,10 +97,13 @@
}
HInstruction* GetInstruction() const { return instruction_; }
- int GetConstant() const { return constant_; }
+ int32_t GetConstant() const { return constant_; }
- bool IsRelativeToArrayLength() const {
- return instruction_ != nullptr && instruction_->IsArrayLength();
+ bool IsRelatedToArrayLength() const {
+ // Some bounds are created with HNewArray* as the instruction instead
+ // of HArrayLength*. They are treated the same.
+ return (instruction_ != nullptr) &&
+ (instruction_->IsArrayLength() || instruction_->IsNewArray());
}
bool IsConstant() const {
@@ -96,54 +117,45 @@
return instruction_ == bound.instruction_ && constant_ == bound.constant_;
}
- // Returns if it's certain bound1 >= bound2.
- bool GreaterThanOrEqual(ValueBound bound) const {
- if (instruction_ == bound.instruction_) {
- if (instruction_ == nullptr) {
- // Pure constant.
- return constant_ >= bound.constant_;
- }
- // There might be overflow/underflow. Be conservative for now.
+ static HInstruction* FromArrayLengthToNewArrayIfPossible(HInstruction* instruction) {
+ // Null check on the NewArray should have been eliminated by instruction
+ // simplifier already.
+ if (instruction->IsArrayLength() && instruction->InputAt(0)->IsNewArray()) {
+ return instruction->InputAt(0)->AsNewArray();
+ }
+ return instruction;
+ }
+
+ static bool Equal(HInstruction* instruction1, HInstruction* instruction2) {
+ if (instruction1 == instruction2) {
+ return true;
+ }
+
+ if (instruction1 == nullptr || instruction2 == nullptr) {
return false;
}
+
+ // Some bounds are created with HNewArray* as the instruction instead
+ // of HArrayLength*. They are treated the same.
+ instruction1 = FromArrayLengthToNewArrayIfPossible(instruction1);
+ instruction2 = FromArrayLengthToNewArrayIfPossible(instruction2);
+ return instruction1 == instruction2;
+ }
+
+ // Returns if it's certain this->bound >= `bound`.
+ bool GreaterThanOrEqualTo(ValueBound bound) const {
+ if (Equal(instruction_, bound.instruction_)) {
+ return constant_ >= bound.constant_;
+ }
// Not comparable. Just return false.
return false;
}
- // Returns if it's certain bound1 <= bound2.
- bool LessThanOrEqual(ValueBound bound) const {
- if (instruction_ == bound.instruction_) {
- if (instruction_ == nullptr) {
- // Pure constant.
- return constant_ <= bound.constant_;
- }
- if (IsRelativeToArrayLength()) {
- // Array length is guaranteed to be no less than 0.
- // No overflow/underflow can happen if both constants are negative.
- if (constant_ <= 0 && bound.constant_ <= 0) {
- return constant_ <= bound.constant_;
- }
- // There might be overflow/underflow. Be conservative for now.
- return false;
- }
+ // Returns if it's certain this->bound <= `bound`.
+ bool LessThanOrEqualTo(ValueBound bound) const {
+ if (Equal(instruction_, bound.instruction_)) {
+ return constant_ <= bound.constant_;
}
-
- // In case the array length is some constant, we can
- // still compare.
- if (IsConstant() && bound.IsRelativeToArrayLength()) {
- HInstruction* array = bound.GetInstruction()->AsArrayLength()->InputAt(0);
- if (array->IsNullCheck()) {
- array = array->AsNullCheck()->InputAt(0);
- }
- if (array->IsNewArray()) {
- HInstruction* len = array->InputAt(0);
- if (len->IsIntConstant()) {
- int len_const = len->AsIntConstant()->GetValue();
- return constant_ <= len_const + bound.GetConstant();
- }
- }
- }
-
// Not comparable. Just return false.
return false;
}
@@ -151,10 +163,11 @@
// Try to narrow lower bound. Returns the greatest of the two if possible.
// Pick one if they are not comparable.
static ValueBound NarrowLowerBound(ValueBound bound1, ValueBound bound2) {
- if (bound1.instruction_ == bound2.instruction_) {
- // Same instruction, compare the constant part.
- return ValueBound(bound1.instruction_,
- std::max(bound1.constant_, bound2.constant_));
+ if (bound1.GreaterThanOrEqualTo(bound2)) {
+ return bound1;
+ }
+ if (bound2.GreaterThanOrEqualTo(bound1)) {
+ return bound2;
}
// Not comparable. Just pick one. We may lose some info, but that's ok.
@@ -165,58 +178,71 @@
// Try to narrow upper bound. Returns the lowest of the two if possible.
// Pick one if they are not comparable.
static ValueBound NarrowUpperBound(ValueBound bound1, ValueBound bound2) {
- if (bound1.instruction_ == bound2.instruction_) {
- // Same instruction, compare the constant part.
- return ValueBound(bound1.instruction_,
- std::min(bound1.constant_, bound2.constant_));
+ if (bound1.LessThanOrEqualTo(bound2)) {
+ return bound1;
+ }
+ if (bound2.LessThanOrEqualTo(bound1)) {
+ return bound2;
}
// Not comparable. Just pick one. We may lose some info, but that's ok.
// Favor array length as upper bound.
- return bound1.IsRelativeToArrayLength() ? bound1 : bound2;
+ return bound1.IsRelatedToArrayLength() ? bound1 : bound2;
}
- // Add a constant to a ValueBound. If the constant part of the ValueBound
- // overflows/underflows, then we can't accurately represent it. For correctness,
- // just return Max/Min() depending on whether the returned ValueBound is used for
- // lower/upper bound.
- ValueBound Add(int c, bool* overflow_or_underflow) const {
- *overflow_or_underflow = false;
+ // Add a constant to a ValueBound.
+ // `overflow` or `underflow` will return whether the resulting bound may
+ // overflow or underflow an int.
+ ValueBound Add(int32_t c, bool* overflow, bool* underflow) const {
+ *overflow = *underflow = false;
if (c == 0) {
return *this;
}
- int new_constant;
+ int32_t new_constant;
if (c > 0) {
if (constant_ > INT_MAX - c) {
- // Constant part overflows.
- *overflow_or_underflow = true;
+ *overflow = true;
return Max();
- } else {
- new_constant = constant_ + c;
}
+
+ new_constant = constant_ + c;
+ // (array.length + non-positive-constant) won't overflow an int.
+ if (IsConstant() || (IsRelatedToArrayLength() && new_constant <= 0)) {
+ return ValueBound(instruction_, new_constant);
+ }
+ // Be conservative.
+ *overflow = true;
+ return Max();
} else {
if (constant_ < INT_MIN - c) {
- // Constant part underflows.
- *overflow_or_underflow = true;
- return Max();
- } else {
- new_constant = constant_ + c;
+ *underflow = true;
+ return Min();
}
+
+ new_constant = constant_ + c;
+ // Regardless of the value new_constant, (array.length+new_constant) will
+ // never underflow since array.length is no less than 0.
+ if (IsConstant() || IsRelatedToArrayLength()) {
+ return ValueBound(instruction_, new_constant);
+ }
+ // Be conservative.
+ *underflow = true;
+ return Min();
}
return ValueBound(instruction_, new_constant);
}
private:
HInstruction* instruction_;
- int constant_;
+ int32_t constant_;
};
/**
* Represent a range of lower bound and upper bound, both being inclusive.
* Currently a ValueRange may be generated as a result of the following:
* comparisons related to array bounds, array bounds check, add/sub on top
- * of an existing value range, or a loop phi corresponding to an
+ * of an existing value range, NewArray or a loop phi corresponding to an
* incrementing/decrementing array index (MonotonicValueRange).
*/
class ValueRange : public ArenaObject<kArenaAllocMisc> {
@@ -241,8 +267,8 @@
return true;
}
DCHECK(!other_range->IsMonotonicValueRange());
- return lower_.GreaterThanOrEqual(other_range->lower_) &&
- upper_.LessThanOrEqual(other_range->upper_);
+ return lower_.GreaterThanOrEqualTo(other_range->lower_) &&
+ upper_.LessThanOrEqualTo(other_range->upper_);
}
// Returns the intersection of this and range.
@@ -263,29 +289,24 @@
ValueBound::NarrowUpperBound(upper_, range->upper_));
}
- // Shift a range by a constant. If either bound can't be represented
- // as (instruction+c) format due to possible overflow/underflow,
- // return the full integer range.
- ValueRange* Add(int constant) const {
- bool overflow_or_underflow;
- ValueBound lower = lower_.Add(constant, &overflow_or_underflow);
- if (overflow_or_underflow) {
- // We can't accurately represent the bounds anymore.
- return FullIntRange();
+ // Shift a range by a constant.
+ ValueRange* Add(int32_t constant) const {
+ bool overflow, underflow;
+ ValueBound lower = lower_.Add(constant, &overflow, &underflow);
+ if (underflow) {
+ // Lower bound underflow will wrap around to positive values
+ // and invalidate the upper bound.
+ return nullptr;
}
- ValueBound upper = upper_.Add(constant, &overflow_or_underflow);
- if (overflow_or_underflow) {
- // We can't accurately represent the bounds anymore.
- return FullIntRange();
+ ValueBound upper = upper_.Add(constant, &overflow, &underflow);
+ if (overflow) {
+ // Upper bound overflow will wrap around to negative values
+ // and invalidate the lower bound.
+ return nullptr;
}
return new (allocator_) ValueRange(allocator_, lower, upper);
}
- // Return [INT_MIN, INT_MAX].
- ValueRange* FullIntRange() const {
- return new (allocator_) ValueRange(allocator_, ValueBound::Min(), ValueBound::Max());
- }
-
private:
ArenaAllocator* const allocator_;
const ValueBound lower_; // inclusive
@@ -304,7 +325,7 @@
public:
MonotonicValueRange(ArenaAllocator* allocator,
HInstruction* initial,
- int increment,
+ int32_t increment,
ValueBound bound)
// To be conservative, give it full range [INT_MIN, INT_MAX] in case it's
// used as a regular value range, due to possible overflow/underflow.
@@ -343,23 +364,17 @@
// make assumptions about the max array length, e.g. due to the max heap size,
// divided by the element size (such as 4 bytes for each integer array), we can
// lower this number and rule out some possible overflows.
- int max_array_len = INT_MAX;
+ int32_t max_array_len = INT_MAX;
- int upper = INT_MAX;
- if (range->GetUpper().IsConstant()) {
- upper = range->GetUpper().GetConstant();
- } else if (range->GetUpper().IsRelativeToArrayLength()) {
- int constant = range->GetUpper().GetConstant();
- if (constant <= 0) {
- // Normal case. e.g. <= array.length - 1, <= array.length - 2, etc.
- upper = max_array_len + constant;
- } else {
- // There might be overflow. Give up narrowing.
- return this;
- }
- } else {
- // There might be overflow. Give up narrowing.
- return this;
+ // max possible integer value of range's upper value.
+ int32_t upper = INT_MAX;
+ // Try to lower upper.
+ ValueBound upper_bound = range->GetUpper();
+ if (upper_bound.IsConstant()) {
+ upper = upper_bound.GetConstant();
+ } else if (upper_bound.IsRelatedToArrayLength() && upper_bound.GetConstant() <= 0) {
+ // Normal case. e.g. <= array.length - 1.
+ upper = max_array_len + upper_bound.GetConstant();
}
// If we can prove for the last number in sequence of initial_,
@@ -368,13 +383,13 @@
// then this MonoticValueRange is narrowed to a normal value range.
// Be conservative first, assume last number in the sequence hits upper.
- int last_num_in_sequence = upper;
+ int32_t last_num_in_sequence = upper;
if (initial_->IsIntConstant()) {
- int initial_constant = initial_->AsIntConstant()->GetValue();
+ int32_t initial_constant = initial_->AsIntConstant()->GetValue();
if (upper <= initial_constant) {
last_num_in_sequence = upper;
} else {
- // Cast to int64_t for the substraction part to avoid int overflow.
+ // Cast to int64_t for the substraction part to avoid int32_t overflow.
last_num_in_sequence = initial_constant +
((int64_t)upper - (int64_t)initial_constant) / increment_ * increment_;
}
@@ -392,23 +407,22 @@
ValueBound upper = ValueBound::NarrowUpperBound(bound_, range->GetUpper());
// Need to take care of underflow. Try to prove underflow won't happen
- // for common cases. Basically need to be able to prove for any value
- // that's >= range->GetLower(), it won't be positive with value+increment.
+ // for common cases.
if (range->GetLower().IsConstant()) {
- int constant = range->GetLower().GetConstant();
+ int32_t constant = range->GetLower().GetConstant();
if (constant >= INT_MIN - increment_) {
return new (GetAllocator()) ValueRange(GetAllocator(), range->GetLower(), upper);
}
}
- // There might be underflow. Give up narrowing.
+ // For non-constant lower bound, just assume might be underflow. Give up narrowing.
return this;
}
}
private:
HInstruction* const initial_;
- const int increment_;
+ const int32_t increment_;
ValueBound bound_; // Additional value bound info for initial_;
DISALLOW_COPY_AND_ASSIGN(MonotonicValueRange);
@@ -446,8 +460,8 @@
return nullptr;
}
- // Narrow the value range of 'instruction' at the end of 'basic_block' with 'range',
- // and push the narrowed value range to 'successor'.
+ // Narrow the value range of `instruction` at the end of `basic_block` with `range`,
+ // and push the narrowed value range to `successor`.
void ApplyRangeFromComparison(HInstruction* instruction, HBasicBlock* basic_block,
HBasicBlock* successor, ValueRange* range) {
ValueRange* existing_range = LookupValueRange(instruction, basic_block);
@@ -472,10 +486,12 @@
bool found;
ValueBound bound = ValueBound::DetectValueBoundFromValue(right, &found);
+ // Each comparison can establish a lower bound and an upper bound
+ // for the left hand side.
ValueBound lower = bound;
ValueBound upper = bound;
if (!found) {
- // No constant or array.length+c bound found.
+ // No constant or array.length+c format bound found.
// For i<j, we can still use j's upper bound as i's upper bound. Same for lower.
ValueRange* range = LookupValueRange(right, block);
if (range != nullptr) {
@@ -487,13 +503,13 @@
}
}
- bool overflow_or_underflow;
+ bool overflow, underflow;
if (cond == kCondLT || cond == kCondLE) {
if (!upper.Equals(ValueBound::Max())) {
- int compensation = (cond == kCondLT) ? -1 : 0; // upper bound is inclusive
- ValueBound new_upper = upper.Add(compensation, &overflow_or_underflow);
- if (overflow_or_underflow) {
- new_upper = ValueBound::Max();
+ int32_t compensation = (cond == kCondLT) ? -1 : 0; // upper bound is inclusive
+ ValueBound new_upper = upper.Add(compensation, &overflow, &underflow);
+ if (overflow || underflow) {
+ return;
}
ValueRange* new_range = new (GetGraph()->GetArena())
ValueRange(GetGraph()->GetArena(), ValueBound::Min(), new_upper);
@@ -501,11 +517,11 @@
}
// array.length as a lower bound isn't considered useful.
- if (!lower.Equals(ValueBound::Min()) && !lower.IsRelativeToArrayLength()) {
- int compensation = (cond == kCondLE) ? 1 : 0; // lower bound is inclusive
- ValueBound new_lower = lower.Add(compensation, &overflow_or_underflow);
- if (overflow_or_underflow) {
- new_lower = ValueBound::Min();
+ if (!lower.Equals(ValueBound::Min()) && !lower.IsRelatedToArrayLength()) {
+ int32_t compensation = (cond == kCondLE) ? 1 : 0; // lower bound is inclusive
+ ValueBound new_lower = lower.Add(compensation, &overflow, &underflow);
+ if (overflow || underflow) {
+ return;
}
ValueRange* new_range = new (GetGraph()->GetArena())
ValueRange(GetGraph()->GetArena(), new_lower, ValueBound::Max());
@@ -513,11 +529,11 @@
}
} else if (cond == kCondGT || cond == kCondGE) {
// array.length as a lower bound isn't considered useful.
- if (!lower.Equals(ValueBound::Min()) && !lower.IsRelativeToArrayLength()) {
- int compensation = (cond == kCondGT) ? 1 : 0; // lower bound is inclusive
- ValueBound new_lower = lower.Add(compensation, &overflow_or_underflow);
- if (overflow_or_underflow) {
- new_lower = ValueBound::Min();
+ if (!lower.Equals(ValueBound::Min()) && !lower.IsRelatedToArrayLength()) {
+ int32_t compensation = (cond == kCondGT) ? 1 : 0; // lower bound is inclusive
+ ValueBound new_lower = lower.Add(compensation, &overflow, &underflow);
+ if (overflow || underflow) {
+ return;
}
ValueRange* new_range = new (GetGraph()->GetArena())
ValueRange(GetGraph()->GetArena(), new_lower, ValueBound::Max());
@@ -525,10 +541,10 @@
}
if (!upper.Equals(ValueBound::Max())) {
- int compensation = (cond == kCondGE) ? -1 : 0; // upper bound is inclusive
- ValueBound new_upper = upper.Add(compensation, &overflow_or_underflow);
- if (overflow_or_underflow) {
- new_upper = ValueBound::Max();
+ int32_t compensation = (cond == kCondGE) ? -1 : 0; // upper bound is inclusive
+ ValueBound new_upper = upper.Add(compensation, &overflow, &underflow);
+ if (overflow || underflow) {
+ return;
}
ValueRange* new_range = new (GetGraph()->GetArena())
ValueRange(GetGraph()->GetArena(), ValueBound::Min(), new_upper);
@@ -541,41 +557,56 @@
HBasicBlock* block = bounds_check->GetBlock();
HInstruction* index = bounds_check->InputAt(0);
HInstruction* array_length = bounds_check->InputAt(1);
- ValueRange* index_range = LookupValueRange(index, block);
+ DCHECK(array_length->IsIntConstant() || array_length->IsArrayLength());
- if (index_range != nullptr) {
- ValueBound lower = ValueBound(nullptr, 0); // constant 0
- ValueBound upper = ValueBound(array_length, -1); // array_length - 1
- ValueRange* array_range = new (GetGraph()->GetArena())
- ValueRange(GetGraph()->GetArena(), lower, upper);
- if (index_range->FitsIn(array_range)) {
- ReplaceBoundsCheck(bounds_check, index);
- return;
- }
- }
-
- if (index->IsIntConstant()) {
- ValueRange* array_length_range = LookupValueRange(array_length, block);
- int constant = index->AsIntConstant()->GetValue();
- if (array_length_range != nullptr &&
- array_length_range->GetLower().IsConstant()) {
- if (constant < array_length_range->GetLower().GetConstant()) {
+ if (!index->IsIntConstant()) {
+ ValueRange* index_range = LookupValueRange(index, block);
+ if (index_range != nullptr) {
+ ValueBound lower = ValueBound(nullptr, 0); // constant 0
+ ValueBound upper = ValueBound(array_length, -1); // array_length - 1
+ ValueRange* array_range = new (GetGraph()->GetArena())
+ ValueRange(GetGraph()->GetArena(), lower, upper);
+ if (index_range->FitsIn(array_range)) {
ReplaceBoundsCheck(bounds_check, index);
return;
}
}
+ } else {
+ int32_t constant = index->AsIntConstant()->GetValue();
+ if (constant < 0) {
+ // Will always throw exception.
+ return;
+ }
+ if (array_length->IsIntConstant()) {
+ if (constant < array_length->AsIntConstant()->GetValue()) {
+ ReplaceBoundsCheck(bounds_check, index);
+ }
+ return;
+ }
+
+ DCHECK(array_length->IsArrayLength());
+ ValueRange* existing_range = LookupValueRange(array_length, block);
+ if (existing_range != nullptr) {
+ ValueBound lower = existing_range->GetLower();
+ DCHECK(lower.IsConstant());
+ if (constant < lower.GetConstant()) {
+ ReplaceBoundsCheck(bounds_check, index);
+ return;
+ } else {
+ // Existing range isn't strong enough to eliminate the bounds check.
+ // Fall through to update the array_length range with info from this
+ // bounds check.
+ }
+ }
// Once we have an array access like 'array[5] = 1', we record array.length >= 6.
+ // We currently don't do it for non-constant index since a valid array[i] can't prove
+ // a valid array[i-1] yet due to the lower bound side.
ValueBound lower = ValueBound(nullptr, constant + 1);
ValueBound upper = ValueBound::Max();
ValueRange* range = new (GetGraph()->GetArena())
ValueRange(GetGraph()->GetArena(), lower, upper);
- ValueRange* existing_range = LookupValueRange(array_length, block);
- ValueRange* new_range = range;
- if (existing_range != nullptr) {
- new_range = range->Narrow(existing_range);
- }
- GetValueRangeMap(block)->Overwrite(array_length->GetId(), new_range);
+ GetValueRangeMap(block)->Overwrite(array_length->GetId(), range);
}
}
@@ -588,14 +619,12 @@
if (phi->IsLoopHeaderPhi() && phi->GetType() == Primitive::kPrimInt) {
DCHECK_EQ(phi->InputCount(), 2U);
HInstruction* instruction = phi->InputAt(1);
- if (instruction->IsAdd()) {
- HAdd* add = instruction->AsAdd();
- HInstruction* left = add->GetLeft();
- HInstruction* right = add->GetRight();
- if (left == phi && right->IsIntConstant()) {
+ HInstruction *left;
+ int32_t increment;
+ if (ValueBound::IsAddOrSubAConstant(instruction, &left, &increment)) {
+ if (left == phi) {
HInstruction* initial_value = phi->InputAt(0);
ValueRange* range = nullptr;
- int increment = right->AsIntConstant()->GetValue();
if (increment == 0) {
// Add constant 0. It's really a fixed value.
range = new (GetGraph()->GetArena()) ValueRange(
@@ -682,10 +711,10 @@
if (right_range != nullptr) {
ValueBound lower = right_range->GetLower();
ValueBound upper = right_range->GetUpper();
- if (lower.IsConstant() && upper.IsRelativeToArrayLength()) {
+ if (lower.IsConstant() && upper.IsRelatedToArrayLength()) {
HInstruction* upper_inst = upper.GetInstruction();
- if (upper_inst->IsArrayLength() &&
- upper_inst->AsArrayLength() == array_length) {
+ // Make sure it's the same array.
+ if (ValueBound::Equal(array_length, upper_inst)) {
// (array.length - v) where v is in [c1, array.length + c2]
// gets [-c2, array.length - c1] as its value range.
ValueRange* range = new (GetGraph()->GetArena()) ValueRange(
@@ -699,6 +728,26 @@
}
}
+ void VisitNewArray(HNewArray* new_array) {
+ HInstruction* len = new_array->InputAt(0);
+ if (!len->IsIntConstant()) {
+ HInstruction *left;
+ int32_t right_const;
+ if (ValueBound::IsAddOrSubAConstant(len, &left, &right_const)) {
+ // (left + right_const) is used as size to new the array.
+ // We record "-right_const <= left <= new_array - right_const";
+ ValueBound lower = ValueBound(nullptr, -right_const);
+ // We use new_array for the bound instead of new_array.length,
+ // which isn't available as an instruction yet. new_array will
+ // be treated the same as new_array.length when it's used in a ValueBound.
+ ValueBound upper = ValueBound(new_array, -right_const);
+ ValueRange* range = new (GetGraph()->GetArena())
+ ValueRange(GetGraph()->GetArena(), lower, upper);
+ GetValueRangeMap(new_array->GetBlock())->Overwrite(left->GetId(), range);
+ }
+ }
+ }
+
std::vector<std::unique_ptr<ArenaSafeMap<int, ValueRange*>>> maps_;
DISALLOW_COPY_AND_ASSIGN(BCEVisitor);
diff --git a/compiler/optimizing/bounds_check_elimination_test.cc b/compiler/optimizing/bounds_check_elimination_test.cc
index 3dcb08d..662834a 100644
--- a/compiler/optimizing/bounds_check_elimination_test.cc
+++ b/compiler/optimizing/bounds_check_elimination_test.cc
@@ -17,6 +17,7 @@
#include "bounds_check_elimination.h"
#include "builder.h"
#include "gvn.h"
+#include "instruction_simplifier.h"
#include "nodes.h"
#include "optimizing_unit_test.h"
#include "side_effects_analysis.h"
@@ -26,7 +27,9 @@
namespace art {
-static void RunGvn(HGraph* graph) {
+static void RunSimplifierAndGvn(HGraph* graph) {
+ InstructionSimplifier simplify(graph);
+ simplify.Run();
SideEffectsAnalysis side_effects(graph);
side_effects.Run();
GVNOptimization(graph, side_effects).Run();
@@ -127,7 +130,7 @@
block3->AddSuccessor(block4); // False successor
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination(graph);
bounds_check_elimination.Run();
ASSERT_FALSE(IsRemoved(bounds_check2));
@@ -202,7 +205,7 @@
block3->AddSuccessor(exit);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination(graph);
bounds_check_elimination.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -277,7 +280,7 @@
block3->AddSuccessor(exit);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination(graph);
bounds_check_elimination.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -351,7 +354,7 @@
exit->AddInstruction(new (&allocator) HExit());
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination(graph);
bounds_check_elimination.Run();
ASSERT_FALSE(IsRemoved(bounds_check5));
@@ -450,7 +453,7 @@
// HArrayLength which uses the null check as its input.
graph = BuildSSAGraph1(&allocator, &bounds_check, 0, 1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_after_gvn(graph);
bounds_check_elimination_after_gvn.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -458,7 +461,7 @@
// for (int i=1; i<array.length; i++) { array[i] = 10; // Can eliminate. }
graph = BuildSSAGraph1(&allocator, &bounds_check, 1, 1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_1(graph);
bounds_check_elimination_with_initial_1.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -466,7 +469,7 @@
// for (int i=-1; i<array.length; i++) { array[i] = 10; // Can't eliminate. }
graph = BuildSSAGraph1(&allocator, &bounds_check, -1, 1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_minus_1(graph);
bounds_check_elimination_with_initial_minus_1.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -474,7 +477,7 @@
// for (int i=0; i<=array.length; i++) { array[i] = 10; // Can't eliminate. }
graph = BuildSSAGraph1(&allocator, &bounds_check, 0, 1, kCondGT);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_greater_than(graph);
bounds_check_elimination_with_greater_than.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -483,7 +486,7 @@
// array[i] = 10; // Can't eliminate due to overflow concern. }
graph = BuildSSAGraph1(&allocator, &bounds_check, 0, 2);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_increment_2(graph);
bounds_check_elimination_with_increment_2.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -491,7 +494,7 @@
// for (int i=1; i<array.length; i += 2) { array[i] = 10; // Can eliminate. }
graph = BuildSSAGraph1(&allocator, &bounds_check, 1, 2);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_increment_2_from_1(graph);
bounds_check_elimination_with_increment_2_from_1.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -591,7 +594,7 @@
// HArrayLength which uses the null check as its input.
graph = BuildSSAGraph2(&allocator, &bounds_check, 0);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_after_gvn(graph);
bounds_check_elimination_after_gvn.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -599,7 +602,7 @@
// for (int i=array.length; i>1; i--) { array[i-1] = 10; // Can eliminate. }
graph = BuildSSAGraph2(&allocator, &bounds_check, 1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_1(graph);
bounds_check_elimination_with_initial_1.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -607,7 +610,7 @@
// for (int i=array.length; i>-1; i--) { array[i-1] = 10; // Can't eliminate. }
graph = BuildSSAGraph2(&allocator, &bounds_check, -1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_minus_1(graph);
bounds_check_elimination_with_initial_minus_1.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -615,7 +618,7 @@
// for (int i=array.length; i>=0; i--) { array[i-1] = 10; // Can't eliminate. }
graph = BuildSSAGraph2(&allocator, &bounds_check, 0, -1, kCondLT);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_less_than(graph);
bounds_check_elimination_with_less_than.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -623,7 +626,7 @@
// for (int i=array.length; i>0; i-=2) { array[i-1] = 10; // Can eliminate. }
graph = BuildSSAGraph2(&allocator, &bounds_check, 0, -2);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_increment_minus_2(graph);
bounds_check_elimination_increment_minus_2.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -710,7 +713,7 @@
HInstruction* bounds_check = nullptr;
HGraph* graph = BuildSSAGraph3(&allocator, &bounds_check, 0, 1, kCondGE);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_after_gvn(graph);
bounds_check_elimination_after_gvn.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -719,7 +722,7 @@
// for (int i=1; i<10; i++) { array[i] = 10; // Can eliminate. }
graph = BuildSSAGraph3(&allocator, &bounds_check, 1, 1, kCondGE);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_1(graph);
bounds_check_elimination_with_initial_1.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -728,7 +731,7 @@
// for (int i=0; i<=10; i++) { array[i] = 10; // Can't eliminate. }
graph = BuildSSAGraph3(&allocator, &bounds_check, 0, 1, kCondGT);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_greater_than(graph);
bounds_check_elimination_with_greater_than.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -737,7 +740,7 @@
// for (int i=1; i<10; i+=8) { array[i] = 10; // Can eliminate. }
graph = BuildSSAGraph3(&allocator, &bounds_check, 1, 8, kCondGE);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_increment_8(graph);
bounds_check_elimination_increment_8.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -838,7 +841,7 @@
// HArrayLength which uses the null check as its input.
graph = BuildSSAGraph4(&allocator, &bounds_check, 0);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_after_gvn(graph);
bounds_check_elimination_after_gvn.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -846,7 +849,7 @@
// for (int i=1; i<array.length; i++) { array[array.length-i-1] = 10; // Can eliminate. }
graph = BuildSSAGraph4(&allocator, &bounds_check, 1);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_initial_1(graph);
bounds_check_elimination_with_initial_1.Run();
ASSERT_TRUE(IsRemoved(bounds_check));
@@ -854,7 +857,7 @@
// for (int i=0; i<=array.length; i++) { array[array.length-i] = 10; // Can't eliminate. }
graph = BuildSSAGraph4(&allocator, &bounds_check, 0, kCondGT);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
BoundsCheckElimination bounds_check_elimination_with_greater_than(graph);
bounds_check_elimination_with_greater_than.Run();
ASSERT_FALSE(IsRemoved(bounds_check));
@@ -1030,7 +1033,7 @@
outer_body_add->AddSuccessor(outer_header);
graph->BuildDominatorTree();
- RunGvn(graph);
+ RunSimplifierAndGvn(graph);
// gvn should remove the same bounds check.
ASSERT_FALSE(IsRemoved(bounds_check1));
ASSERT_FALSE(IsRemoved(bounds_check2));
diff --git a/compiler/optimizing/code_generator.cc b/compiler/optimizing/code_generator.cc
index fd4e391..d0739a6 100644
--- a/compiler/optimizing/code_generator.cc
+++ b/compiler/optimizing/code_generator.cc
@@ -41,8 +41,6 @@
}
void CodeGenerator::CompileBaseline(CodeAllocator* allocator, bool is_leaf) {
- DCHECK_EQ(frame_size_, kUninitializedFrameSize);
-
Initialize();
if (!is_leaf) {
MarkNotLeaf();
@@ -59,7 +57,6 @@
}
void CodeGenerator::CompileInternal(CodeAllocator* allocator, bool is_baseline) {
- HGraphVisitor* location_builder = GetLocationBuilder();
HGraphVisitor* instruction_visitor = GetInstructionVisitor();
DCHECK_EQ(current_block_index_, 0u);
GenerateFrameEntry();
@@ -69,8 +66,7 @@
for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
HInstruction* current = it.Current();
if (is_baseline) {
- current->Accept(location_builder);
- InitLocations(current);
+ InitLocationsBaseline(current);
}
current->Accept(instruction_visitor);
}
@@ -88,7 +84,6 @@
void CodeGenerator::CompileOptimized(CodeAllocator* allocator) {
// The register allocator already called `InitializeCodeGeneration`,
// where the frame size has been computed.
- DCHECK_NE(frame_size_, kUninitializedFrameSize);
DCHECK(block_order_ != nullptr);
Initialize();
CompileInternal(allocator, /* is_baseline */ false);
@@ -138,13 +133,22 @@
ComputeSpillMask();
first_register_slot_in_slow_path_ = (number_of_out_slots + number_of_spill_slots) * kVRegSize;
- SetFrameSize(RoundUp(
- number_of_spill_slots * kVRegSize
- + number_of_out_slots * kVRegSize
- + maximum_number_of_live_core_registers * GetWordSize()
- + maximum_number_of_live_fp_registers * GetFloatingPointSpillSlotSize()
- + FrameEntrySpillSize(),
- kStackAlignment));
+ if (number_of_spill_slots == 0
+ && !HasAllocatedCalleeSaveRegisters()
+ && IsLeafMethod()
+ && !RequiresCurrentMethod()) {
+ DCHECK_EQ(maximum_number_of_live_core_registers, 0u);
+ DCHECK_EQ(maximum_number_of_live_fp_registers, 0u);
+ SetFrameSize(CallPushesPC() ? GetWordSize() : 0);
+ } else {
+ SetFrameSize(RoundUp(
+ number_of_spill_slots * kVRegSize
+ + number_of_out_slots * kVRegSize
+ + maximum_number_of_live_core_registers * GetWordSize()
+ + maximum_number_of_live_fp_registers * GetFloatingPointSpillSlotSize()
+ + FrameEntrySpillSize(),
+ kStackAlignment));
+ }
}
Location CodeGenerator::GetTemporaryLocation(HTemporary* temp) const {
@@ -294,7 +298,8 @@
}
}
-void CodeGenerator::InitLocations(HInstruction* instruction) {
+void CodeGenerator::InitLocationsBaseline(HInstruction* instruction) {
+ AllocateLocations(instruction);
if (instruction->GetLocations() == nullptr) {
if (instruction->IsTemporary()) {
HInstruction* previous = instruction->GetPrevious();
@@ -320,6 +325,19 @@
}
}
+void CodeGenerator::AllocateLocations(HInstruction* instruction) {
+ instruction->Accept(GetLocationBuilder());
+ LocationSummary* locations = instruction->GetLocations();
+ if (!instruction->IsSuspendCheckEntry()) {
+ if (locations != nullptr && locations->CanCall()) {
+ MarkNotLeaf();
+ }
+ if (instruction->NeedsCurrentMethod()) {
+ SetRequiresCurrentMethod();
+ }
+ }
+}
+
bool CodeGenerator::GoesToNextBlock(HBasicBlock* current, HBasicBlock* next) const {
DCHECK_EQ(block_order_->Get(current_block_index_), current);
return (current_block_index_ < block_order_->Size() - 1)
diff --git a/compiler/optimizing/code_generator.h b/compiler/optimizing/code_generator.h
index ab63b91..efd0c84 100644
--- a/compiler/optimizing/code_generator.h
+++ b/compiler/optimizing/code_generator.h
@@ -30,7 +30,6 @@
namespace art {
static size_t constexpr kVRegSize = 4;
-static size_t constexpr kUninitializedFrameSize = 0;
// Binary encoding of 2^32 for type double.
static int64_t constexpr k2Pow32EncodingForDouble = INT64_C(0x41F0000000000000);
@@ -107,8 +106,6 @@
virtual void GenerateFrameExit() = 0;
virtual void Bind(HBasicBlock* block) = 0;
virtual void Move(HInstruction* instruction, Location location, HInstruction* move_for) = 0;
- virtual HGraphVisitor* GetLocationBuilder() = 0;
- virtual HGraphVisitor* GetInstructionVisitor() = 0;
virtual Assembler* GetAssembler() = 0;
virtual size_t GetWordSize() const = 0;
virtual size_t GetFloatingPointSpillSlotSize() const = 0;
@@ -196,6 +193,15 @@
void MarkNotLeaf() {
is_leaf_ = false;
+ requires_current_method_ = true;
+ }
+
+ void SetRequiresCurrentMethod() {
+ requires_current_method_ = true;
+ }
+
+ bool RequiresCurrentMethod() const {
+ return requires_current_method_;
}
// Clears the spill slots taken by loop phis in the `LocationSummary` of the
@@ -228,6 +234,8 @@
allocated_registers_.Add(location);
}
+ void AllocateLocations(HInstruction* instruction);
+
protected:
CodeGenerator(HGraph* graph,
size_t number_of_core_registers,
@@ -236,7 +244,7 @@
uint32_t core_callee_save_mask,
uint32_t fpu_callee_save_mask,
const CompilerOptions& compiler_options)
- : frame_size_(kUninitializedFrameSize),
+ : frame_size_(0),
core_spill_mask_(0),
fpu_spill_mask_(0),
first_register_slot_in_slow_path_(0),
@@ -255,6 +263,7 @@
block_order_(nullptr),
current_block_index_(0),
is_leaf_(true),
+ requires_current_method_(false),
stack_map_stream_(graph->GetArena()) {}
// Register allocation logic.
@@ -269,11 +278,12 @@
virtual Location GetStackLocation(HLoadLocal* load) const = 0;
virtual ParallelMoveResolver* GetMoveResolver() = 0;
+ virtual HGraphVisitor* GetLocationBuilder() = 0;
+ virtual HGraphVisitor* GetInstructionVisitor() = 0;
// Returns the location of the first spilled entry for floating point registers,
// relative to the stack pointer.
uint32_t GetFpuSpillStart() const {
- DCHECK_NE(frame_size_, kUninitializedFrameSize);
return GetFrameSize() - FrameEntrySpillSize();
}
@@ -289,6 +299,21 @@
return GetFpuSpillSize() + GetCoreSpillSize();
}
+ bool HasAllocatedCalleeSaveRegisters() const {
+ // We check the core registers against 1 because it always comprises the return PC.
+ return (POPCOUNT(allocated_registers_.GetCoreRegisters() & core_callee_save_mask_) != 1)
+ || (POPCOUNT(allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_) != 0);
+ }
+
+ bool CallPushesPC() const {
+ InstructionSet instruction_set = GetInstructionSet();
+ return instruction_set == kX86 || instruction_set == kX86_64;
+ }
+
+ bool HasEmptyFrame() const {
+ return GetFrameSize() == (CallPushesPC() ? GetWordSize() : 0);
+ }
+
// Frame size required for this method.
uint32_t frame_size_;
uint32_t core_spill_mask_;
@@ -311,7 +336,7 @@
const uint32_t fpu_callee_save_mask_;
private:
- void InitLocations(HInstruction* instruction);
+ void InitLocationsBaseline(HInstruction* instruction);
size_t GetStackOffsetOfSavedRegister(size_t index);
void CompileInternal(CodeAllocator* allocator, bool is_baseline);
@@ -328,8 +353,12 @@
// we are generating code for.
size_t current_block_index_;
+ // Whether the method is a leaf method.
bool is_leaf_;
+ // Whether an instruction in the graph accesses the current method.
+ bool requires_current_method_;
+
StackMapStream stack_map_stream_;
DISALLOW_COPY_AND_ASSIGN(CodeGenerator);
diff --git a/compiler/optimizing/code_generator_arm.cc b/compiler/optimizing/code_generator_arm.cc
index 164dce4..e5de2ab 100644
--- a/compiler/optimizing/code_generator_arm.cc
+++ b/compiler/optimizing/code_generator_arm.cc
@@ -19,6 +19,8 @@
#include "arch/arm/instruction_set_features_arm.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "gc/accounting/card_table.h"
+#include "intrinsics.h"
+#include "intrinsics_arm.h"
#include "mirror/array-inl.h"
#include "mirror/art_method.h"
#include "mirror/class.h"
@@ -32,11 +34,6 @@
namespace arm {
-static DRegister FromLowSToD(SRegister reg) {
- DCHECK_EQ(reg % 2, 0);
- return static_cast<DRegister>(reg / 2);
-}
-
static bool ExpectedPairLayout(Location location) {
// We expected this for both core and fpu register pairs.
return ((location.low() & 1) == 0) && (location.low() + 1 == location.high());
@@ -73,20 +70,6 @@
#define __ reinterpret_cast<ArmAssembler*>(codegen->GetAssembler())->
#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmWordSize, x).Int32Value()
-class SlowPathCodeARM : public SlowPathCode {
- public:
- SlowPathCodeARM() : entry_label_(), exit_label_() {}
-
- Label* GetEntryLabel() { return &entry_label_; }
- Label* GetExitLabel() { return &exit_label_; }
-
- private:
- Label entry_label_;
- Label exit_label_;
-
- DISALLOW_COPY_AND_ASSIGN(SlowPathCodeARM);
-};
-
class NullCheckSlowPathARM : public SlowPathCodeARM {
public:
explicit NullCheckSlowPathARM(HNullCheck* instruction) : instruction_(instruction) {}
@@ -396,10 +379,6 @@
move_resolver_(graph->GetArena(), this),
assembler_(true),
isa_features_(isa_features) {
- // Save one extra register for baseline. Note that on thumb2, there is no easy
- // instruction to restore just the PC, so this actually helps both baseline
- // and non-baseline to save and restore at least two registers at entry and exit.
- AddAllocatedRegister(Location::RegisterLocation(kCoreSavedRegisterForBaseline));
// Save the PC register to mimic Quick.
AddAllocatedRegister(Location::RegisterLocation(PC));
}
@@ -508,6 +487,10 @@
void CodeGeneratorARM::ComputeSpillMask() {
core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_;
+ // Save one extra register for baseline. Note that on thumb2, there is no easy
+ // instruction to restore just the PC, so this actually helps both baseline
+ // and non-baseline to save and restore at least two registers at entry and exit.
+ core_spill_mask_ |= (1 << kCoreSavedRegisterForBaseline);
DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved";
fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_;
// We use vpush and vpop for saving and restoring floating point registers, which take
@@ -529,6 +512,10 @@
DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks());
__ Bind(&frame_entry_label_);
+ if (HasEmptyFrame()) {
+ return;
+ }
+
if (!skip_overflow_check) {
__ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm)));
__ LoadFromOffset(kLoadWord, IP, IP, 0);
@@ -547,6 +534,10 @@
}
void CodeGeneratorARM::GenerateFrameExit() {
+ if (HasEmptyFrame()) {
+ __ bx(LR);
+ return;
+ }
__ AddConstant(SP, GetFrameSize() - FrameEntrySpillSize());
if (fpu_spill_mask_ != 0) {
SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_));
@@ -1168,44 +1159,37 @@
}
void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
+ IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(),
+ codegen_->GetInstructionSetFeatures());
+ if (intrinsic.TryDispatch(invoke)) {
+ return;
+ }
+
HandleInvoke(invoke);
}
void CodeGeneratorARM::LoadCurrentMethod(Register reg) {
+ DCHECK(RequiresCurrentMethod());
__ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset);
}
+static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) {
+ if (invoke->GetLocations()->Intrinsified()) {
+ IntrinsicCodeGeneratorARM intrinsic(codegen);
+ intrinsic.Dispatch(invoke);
+ return true;
+ }
+ return false;
+}
+
void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) {
- Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
-
- // TODO: Implement all kinds of calls:
- // 1) boot -> boot
- // 2) app -> boot
- // 3) app -> app
- //
- // Currently we implement the app -> app logic, which looks up in the resolve cache.
-
- // temp = method;
- codegen_->LoadCurrentMethod(temp);
- if (!invoke->IsRecursive()) {
- // temp = temp->dex_cache_resolved_methods_;
- __ LoadFromOffset(
- kLoadWord, temp, temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value());
- // temp = temp[index_in_cache]
- __ LoadFromOffset(
- kLoadWord, temp, temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex()));
- // LR = temp[offset_of_quick_compiled_code]
- __ LoadFromOffset(kLoadWord, LR, temp,
- mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
- kArmWordSize).Int32Value());
- // LR()
- __ blx(LR);
- } else {
- __ bl(codegen_->GetFrameEntryLabel());
+ if (TryGenerateIntrinsicCode(invoke, codegen_)) {
+ return;
}
- codegen_->RecordPcInfo(invoke, invoke->GetDexPc());
- DCHECK(!codegen_->IsLeafMethod());
+ Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
+
+ codegen_->GenerateStaticOrDirectCall(invoke, temp);
}
void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) {
@@ -1223,10 +1207,20 @@
}
void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) {
+ IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(),
+ codegen_->GetInstructionSetFeatures());
+ if (intrinsic.TryDispatch(invoke)) {
+ return;
+ }
+
HandleInvoke(invoke);
}
void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) {
+ if (TryGenerateIntrinsicCode(invoke, codegen_)) {
+ return;
+ }
+
Register temp = invoke->GetLocations()->GetTemp(0).AsRegister<Register>();
uint32_t method_offset = mirror::Class::EmbeddedVTableOffset().Uint32Value() +
invoke->GetVTableIndex() * sizeof(mirror::Class::VTableEntry);
@@ -3846,5 +3840,38 @@
}
}
+void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Register temp) {
+ DCHECK_EQ(temp, kArtMethodRegister);
+
+ // TODO: Implement all kinds of calls:
+ // 1) boot -> boot
+ // 2) app -> boot
+ // 3) app -> app
+ //
+ // Currently we implement the app -> app logic, which looks up in the resolve cache.
+
+ // temp = method;
+ LoadCurrentMethod(temp);
+ if (!invoke->IsRecursive()) {
+ // temp = temp->dex_cache_resolved_methods_;
+ __ LoadFromOffset(
+ kLoadWord, temp, temp, mirror::ArtMethod::DexCacheResolvedMethodsOffset().Int32Value());
+ // temp = temp[index_in_cache]
+ __ LoadFromOffset(
+ kLoadWord, temp, temp, CodeGenerator::GetCacheOffset(invoke->GetDexMethodIndex()));
+ // LR = temp[offset_of_quick_compiled_code]
+ __ LoadFromOffset(kLoadWord, LR, temp,
+ mirror::ArtMethod::EntryPointFromQuickCompiledCodeOffset(
+ kArmWordSize).Int32Value());
+ // LR()
+ __ blx(LR);
+ } else {
+ __ bl(GetFrameEntryLabel());
+ }
+
+ RecordPcInfo(invoke, invoke->GetDexPc());
+ DCHECK(!IsLeafMethod());
+}
+
} // namespace arm
} // namespace art
diff --git a/compiler/optimizing/code_generator_arm.h b/compiler/optimizing/code_generator_arm.h
index 4b03dff..47d81ff 100644
--- a/compiler/optimizing/code_generator_arm.h
+++ b/compiler/optimizing/code_generator_arm.h
@@ -39,6 +39,14 @@
{ S0, S1, S2, S3, S4, S5, S6, S7, S8, S9, S10, S11, S12, S13, S14, S15 };
static constexpr size_t kParameterFpuRegistersLength = arraysize(kParameterFpuRegisters);
+static constexpr Register kArtMethodRegister = R0;
+
+static constexpr DRegister FromLowSToD(SRegister reg) {
+ return DCHECK_CONSTEXPR(reg % 2 == 0, , D0)
+ static_cast<DRegister>(reg / 2);
+}
+
+
class InvokeDexCallingConvention : public CallingConvention<Register, SRegister> {
public:
InvokeDexCallingConvention()
@@ -90,6 +98,20 @@
DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverARM);
};
+class SlowPathCodeARM : public SlowPathCode {
+ public:
+ SlowPathCodeARM() : entry_label_(), exit_label_() {}
+
+ Label* GetEntryLabel() { return &entry_label_; }
+ Label* GetExitLabel() { return &exit_label_; }
+
+ private:
+ Label entry_label_;
+ Label exit_label_;
+
+ DISALLOW_COPY_AND_ASSIGN(SlowPathCodeARM);
+};
+
class LocationsBuilderARM : public HGraphVisitor {
public:
LocationsBuilderARM(HGraph* graph, CodeGeneratorARM* codegen)
@@ -249,6 +271,8 @@
Label* GetFrameEntryLabel() { return &frame_entry_label_; }
+ void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Register temp);
+
private:
// Labels for each block that will be compiled.
GrowableArray<Label> block_labels_;
diff --git a/compiler/optimizing/code_generator_arm64.cc b/compiler/optimizing/code_generator_arm64.cc
index 7588a29..46f1a9b 100644
--- a/compiler/optimizing/code_generator_arm64.cc
+++ b/compiler/optimizing/code_generator_arm64.cc
@@ -453,24 +453,27 @@
RecordPcInfo(nullptr, 0);
}
- int frame_size = GetFrameSize();
- __ Str(kArtMethodRegister, MemOperand(sp, -frame_size, PreIndex));
- __ PokeCPURegList(GetFramePreservedCoreRegisters(), frame_size - GetCoreSpillSize());
- __ PokeCPURegList(GetFramePreservedFPRegisters(), frame_size - FrameEntrySpillSize());
-
- // Stack layout:
- // sp[frame_size - 8] : lr.
- // ... : other preserved core registers.
- // ... : other preserved fp registers.
- // ... : reserved frame space.
- // sp[0] : current method.
+ if (!HasEmptyFrame()) {
+ int frame_size = GetFrameSize();
+ // Stack layout:
+ // sp[frame_size - 8] : lr.
+ // ... : other preserved core registers.
+ // ... : other preserved fp registers.
+ // ... : reserved frame space.
+ // sp[0] : current method.
+ __ Str(kArtMethodRegister, MemOperand(sp, -frame_size, PreIndex));
+ __ PokeCPURegList(GetFramePreservedCoreRegisters(), frame_size - GetCoreSpillSize());
+ __ PokeCPURegList(GetFramePreservedFPRegisters(), frame_size - FrameEntrySpillSize());
+ }
}
void CodeGeneratorARM64::GenerateFrameExit() {
- int frame_size = GetFrameSize();
- __ PeekCPURegList(GetFramePreservedFPRegisters(), frame_size - FrameEntrySpillSize());
- __ PeekCPURegList(GetFramePreservedCoreRegisters(), frame_size - GetCoreSpillSize());
- __ Drop(frame_size);
+ if (!HasEmptyFrame()) {
+ int frame_size = GetFrameSize();
+ __ PeekCPURegList(GetFramePreservedFPRegisters(), frame_size - FrameEntrySpillSize());
+ __ PeekCPURegList(GetFramePreservedCoreRegisters(), frame_size - GetCoreSpillSize());
+ __ Drop(frame_size);
+ }
}
void CodeGeneratorARM64::Bind(HBasicBlock* block) {
@@ -961,6 +964,7 @@
}
void CodeGeneratorARM64::LoadCurrentMethod(vixl::Register current_method) {
+ DCHECK(RequiresCurrentMethod());
DCHECK(current_method.IsW());
__ Ldr(current_method, MemOperand(sp, kCurrentMethodStackOffset));
}
diff --git a/compiler/optimizing/code_generator_x86.cc b/compiler/optimizing/code_generator_x86.cc
index 98f93a4..1a95f41 100644
--- a/compiler/optimizing/code_generator_x86.cc
+++ b/compiler/optimizing/code_generator_x86.cc
@@ -470,12 +470,16 @@
RecordPcInfo(nullptr, 0);
}
- __ subl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
- __ movl(Address(ESP, kCurrentMethodStackOffset), EAX);
+ if (!HasEmptyFrame()) {
+ __ subl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
+ __ movl(Address(ESP, kCurrentMethodStackOffset), EAX);
+ }
}
void CodeGeneratorX86::GenerateFrameExit() {
- __ addl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
+ if (!HasEmptyFrame()) {
+ __ addl(ESP, Immediate(GetFrameSize() - FrameEntrySpillSize()));
+ }
}
void CodeGeneratorX86::Bind(HBasicBlock* block) {
@@ -483,6 +487,7 @@
}
void CodeGeneratorX86::LoadCurrentMethod(Register reg) {
+ DCHECK(RequiresCurrentMethod());
__ movl(reg, Address(ESP, kCurrentMethodStackOffset));
}
diff --git a/compiler/optimizing/code_generator_x86_64.cc b/compiler/optimizing/code_generator_x86_64.cc
index 2ff53a0..88f1753 100644
--- a/compiler/optimizing/code_generator_x86_64.cc
+++ b/compiler/optimizing/code_generator_x86_64.cc
@@ -487,6 +487,10 @@
RecordPcInfo(nullptr, 0);
}
+ if (HasEmptyFrame()) {
+ return;
+ }
+
for (int i = arraysize(kCoreCalleeSaves) - 1; i >= 0; --i) {
Register reg = kCoreCalleeSaves[i];
if (allocated_registers_.ContainsCoreRegister(reg)) {
@@ -509,6 +513,9 @@
}
void CodeGeneratorX86_64::GenerateFrameExit() {
+ if (HasEmptyFrame()) {
+ return;
+ }
uint32_t xmm_spill_location = GetFpuSpillStart();
size_t xmm_spill_slot_size = GetFloatingPointSpillSlotSize();
for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) {
@@ -533,6 +540,7 @@
}
void CodeGeneratorX86_64::LoadCurrentMethod(CpuRegister reg) {
+ DCHECK(RequiresCurrentMethod());
__ movl(reg, Address(CpuRegister(RSP), kCurrentMethodStackOffset));
}
@@ -790,7 +798,7 @@
// Materialized condition, compare against 0.
Location lhs = if_instr->GetLocations()->InAt(0);
if (lhs.IsRegister()) {
- __ cmpl(lhs.AsRegister<CpuRegister>(), Immediate(0));
+ __ testl(lhs.AsRegister<CpuRegister>(), lhs.AsRegister<CpuRegister>());
} else {
__ cmpl(Address(CpuRegister(RSP), lhs.GetStackIndex()),
Immediate(0));
@@ -806,8 +814,12 @@
if (rhs.IsRegister()) {
__ cmpl(lhs.AsRegister<CpuRegister>(), rhs.AsRegister<CpuRegister>());
} else if (rhs.IsConstant()) {
- __ cmpl(lhs.AsRegister<CpuRegister>(),
- Immediate(rhs.GetConstant()->AsIntConstant()->GetValue()));
+ int32_t constant = rhs.GetConstant()->AsIntConstant()->GetValue();
+ if (constant == 0) {
+ __ testl(lhs.AsRegister<CpuRegister>(), lhs.AsRegister<CpuRegister>());
+ } else {
+ __ cmpl(lhs.AsRegister<CpuRegister>(), Immediate(constant));
+ }
} else {
__ cmpl(lhs.AsRegister<CpuRegister>(),
Address(CpuRegister(RSP), rhs.GetStackIndex()));
@@ -883,15 +895,19 @@
CpuRegister reg = locations->Out().AsRegister<CpuRegister>();
// Clear register: setcc only sets the low byte.
__ xorq(reg, reg);
- if (locations->InAt(1).IsRegister()) {
- __ cmpl(locations->InAt(0).AsRegister<CpuRegister>(),
- locations->InAt(1).AsRegister<CpuRegister>());
- } else if (locations->InAt(1).IsConstant()) {
- __ cmpl(locations->InAt(0).AsRegister<CpuRegister>(),
- Immediate(locations->InAt(1).GetConstant()->AsIntConstant()->GetValue()));
+ Location lhs = locations->InAt(0);
+ Location rhs = locations->InAt(1);
+ if (rhs.IsRegister()) {
+ __ cmpl(lhs.AsRegister<CpuRegister>(), rhs.AsRegister<CpuRegister>());
+ } else if (rhs.IsConstant()) {
+ int32_t constant = rhs.GetConstant()->AsIntConstant()->GetValue();
+ if (constant == 0) {
+ __ testl(lhs.AsRegister<CpuRegister>(), lhs.AsRegister<CpuRegister>());
+ } else {
+ __ cmpl(lhs.AsRegister<CpuRegister>(), Immediate(constant));
+ }
} else {
- __ cmpl(locations->InAt(0).AsRegister<CpuRegister>(),
- Address(CpuRegister(RSP), locations->InAt(1).GetStackIndex()));
+ __ cmpl(lhs.AsRegister<CpuRegister>(), Address(CpuRegister(RSP), rhs.GetStackIndex()));
}
__ setcc(X86_64Condition(comp->GetCondition()), reg);
}
@@ -1840,8 +1856,8 @@
switch (add->GetResultType()) {
case Primitive::kPrimInt: {
locations->SetInAt(0, Location::RequiresRegister());
- locations->SetInAt(1, Location::Any());
- locations->SetOut(Location::SameAsFirstInput());
+ locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1)));
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
break;
}
@@ -1869,16 +1885,27 @@
LocationSummary* locations = add->GetLocations();
Location first = locations->InAt(0);
Location second = locations->InAt(1);
- DCHECK(first.Equals(locations->Out()));
+ Location out = locations->Out();
switch (add->GetResultType()) {
case Primitive::kPrimInt: {
if (second.IsRegister()) {
- __ addl(first.AsRegister<CpuRegister>(), second.AsRegister<CpuRegister>());
+ if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
+ __ addl(out.AsRegister<CpuRegister>(), second.AsRegister<CpuRegister>());
+ } else {
+ __ leal(out.AsRegister<CpuRegister>(), Address(
+ first.AsRegister<CpuRegister>(), second.AsRegister<CpuRegister>(), TIMES_1, 0));
+ }
} else if (second.IsConstant()) {
- Immediate imm(second.GetConstant()->AsIntConstant()->GetValue());
- __ addl(first.AsRegister<CpuRegister>(), imm);
+ if (out.AsRegister<Register>() == first.AsRegister<Register>()) {
+ __ addl(out.AsRegister<CpuRegister>(),
+ Immediate(second.GetConstant()->AsIntConstant()->GetValue()));
+ } else {
+ __ leal(out.AsRegister<CpuRegister>(), Address(
+ first.AsRegister<CpuRegister>(), second.GetConstant()->AsIntConstant()->GetValue()));
+ }
} else {
+ DCHECK(first.Equals(locations->Out()));
__ addl(first.AsRegister<CpuRegister>(), Address(CpuRegister(RSP), second.GetStackIndex()));
}
break;
@@ -2754,7 +2781,7 @@
Location obj = locations->InAt(0);
if (obj.IsRegister()) {
- __ cmpl(obj.AsRegister<CpuRegister>(), Immediate(0));
+ __ testl(obj.AsRegister<CpuRegister>(), obj.AsRegister<CpuRegister>());
} else if (obj.IsStackSlot()) {
__ cmpl(Address(CpuRegister(RSP), obj.GetStackIndex()), Immediate(0));
} else {
@@ -3237,12 +3264,16 @@
} else if (source.IsConstant()) {
HConstant* constant = source.GetConstant();
if (constant->IsIntConstant()) {
- Immediate imm(constant->AsIntConstant()->GetValue());
+ int32_t value = constant->AsIntConstant()->GetValue();
if (destination.IsRegister()) {
- __ movl(destination.AsRegister<CpuRegister>(), imm);
+ if (value == 0) {
+ __ xorl(destination.AsRegister<CpuRegister>(), destination.AsRegister<CpuRegister>());
+ } else {
+ __ movl(destination.AsRegister<CpuRegister>(), Immediate(value));
+ }
} else {
DCHECK(destination.IsStackSlot()) << destination;
- __ movl(Address(CpuRegister(RSP), destination.GetStackIndex()), imm);
+ __ movl(Address(CpuRegister(RSP), destination.GetStackIndex()), Immediate(value));
}
} else if (constant->IsLongConstant()) {
int64_t value = constant->AsLongConstant()->GetValue();
diff --git a/compiler/optimizing/instruction_simplifier.cc b/compiler/optimizing/instruction_simplifier.cc
index 17c8f33..44dbb9d 100644
--- a/compiler/optimizing/instruction_simplifier.cc
+++ b/compiler/optimizing/instruction_simplifier.cc
@@ -28,6 +28,7 @@
void VisitArraySet(HArraySet* equal) OVERRIDE;
void VisitTypeConversion(HTypeConversion* instruction) OVERRIDE;
void VisitNullCheck(HNullCheck* instruction) OVERRIDE;
+ void VisitArrayLength(HArrayLength* instruction) OVERRIDE;
};
void InstructionSimplifier::Run() {
@@ -75,6 +76,18 @@
}
}
+void InstructionSimplifierVisitor::VisitArrayLength(HArrayLength* instruction) {
+ HInstruction* input = instruction->InputAt(0);
+ // If the array is a NewArray with constant size, replace the array length
+ // with the constant instruction. This helps the bounds check elimination phase.
+ if (input->IsNewArray()) {
+ input = input->InputAt(0);
+ if (input->IsIntConstant()) {
+ instruction->ReplaceWith(input);
+ }
+ }
+}
+
void InstructionSimplifierVisitor::VisitArraySet(HArraySet* instruction) {
HInstruction* value = instruction->GetValue();
if (value->GetType() != Primitive::kPrimNot) return;
diff --git a/compiler/optimizing/intrinsics_arm.cc b/compiler/optimizing/intrinsics_arm.cc
new file mode 100644
index 0000000..a82d80a
--- /dev/null
+++ b/compiler/optimizing/intrinsics_arm.cc
@@ -0,0 +1,883 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "intrinsics_arm.h"
+
+#include "arch/arm/instruction_set_features_arm.h"
+#include "code_generator_arm.h"
+#include "entrypoints/quick/quick_entrypoints.h"
+#include "intrinsics.h"
+#include "mirror/array-inl.h"
+#include "mirror/art_method.h"
+#include "mirror/string.h"
+#include "thread.h"
+#include "utils/arm/assembler_arm.h"
+
+namespace art {
+
+namespace arm {
+
+ArmAssembler* IntrinsicCodeGeneratorARM::GetAssembler() {
+ return codegen_->GetAssembler();
+}
+
+ArenaAllocator* IntrinsicCodeGeneratorARM::GetAllocator() {
+ return codegen_->GetGraph()->GetArena();
+}
+
+#define __ codegen->GetAssembler()->
+
+static void MoveFromReturnRegister(Location trg, Primitive::Type type, CodeGeneratorARM* codegen) {
+ if (!trg.IsValid()) {
+ DCHECK(type == Primitive::kPrimVoid);
+ return;
+ }
+
+ DCHECK_NE(type, Primitive::kPrimVoid);
+
+ if (Primitive::IsIntegralType(type)) {
+ if (type == Primitive::kPrimLong) {
+ Register trg_reg_lo = trg.AsRegisterPairLow<Register>();
+ Register trg_reg_hi = trg.AsRegisterPairHigh<Register>();
+ Register res_reg_lo = R0;
+ Register res_reg_hi = R1;
+ if (trg_reg_lo != res_reg_hi) {
+ if (trg_reg_lo != res_reg_lo) {
+ __ mov(trg_reg_lo, ShifterOperand(res_reg_lo));
+ __ mov(trg_reg_hi, ShifterOperand(res_reg_hi));
+ } else {
+ DCHECK_EQ(trg_reg_lo + 1, trg_reg_hi);
+ }
+ } else {
+ __ mov(trg_reg_hi, ShifterOperand(res_reg_hi));
+ __ mov(trg_reg_lo, ShifterOperand(res_reg_lo));
+ }
+ } else {
+ Register trg_reg = trg.AsRegister<Register>();
+ Register res_reg = R0;
+ if (trg_reg != res_reg) {
+ __ mov(trg_reg, ShifterOperand(res_reg));
+ }
+ }
+ } else {
+ UNIMPLEMENTED(FATAL) << "Floating-point return.";
+ }
+}
+
+static void MoveArguments(HInvoke* invoke, ArenaAllocator* arena, CodeGeneratorARM* codegen) {
+ if (invoke->InputCount() == 0) {
+ return;
+ }
+
+ LocationSummary* locations = invoke->GetLocations();
+ InvokeDexCallingConventionVisitor calling_convention_visitor;
+
+ // We're moving potentially two or more locations to locations that could overlap, so we need
+ // a parallel move resolver.
+ HParallelMove parallel_move(arena);
+
+ for (size_t i = 0; i < invoke->InputCount(); i++) {
+ HInstruction* input = invoke->InputAt(i);
+ Location cc_loc = calling_convention_visitor.GetNextLocation(input->GetType());
+ Location actual_loc = locations->InAt(i);
+
+ parallel_move.AddMove(actual_loc, cc_loc, nullptr);
+ }
+
+ codegen->GetMoveResolver()->EmitNativeCode(¶llel_move);
+}
+
+// Slow-path for fallback (calling the managed code to handle the intrinsic) in an intrinsified
+// call. This will copy the arguments into the positions for a regular call.
+//
+// Note: The actual parameters are required to be in the locations given by the invoke's location
+// summary. If an intrinsic modifies those locations before a slowpath call, they must be
+// restored!
+class IntrinsicSlowPathARM : public SlowPathCodeARM {
+ public:
+ explicit IntrinsicSlowPathARM(HInvoke* invoke) : invoke_(invoke) { }
+
+ void EmitNativeCode(CodeGenerator* codegen_in) OVERRIDE {
+ CodeGeneratorARM* codegen = down_cast<CodeGeneratorARM*>(codegen_in);
+ __ Bind(GetEntryLabel());
+
+ codegen->SaveLiveRegisters(invoke_->GetLocations());
+
+ MoveArguments(invoke_, codegen->GetGraph()->GetArena(), codegen);
+
+ if (invoke_->IsInvokeStaticOrDirect()) {
+ codegen->GenerateStaticOrDirectCall(invoke_->AsInvokeStaticOrDirect(), kArtMethodRegister);
+ } else {
+ UNIMPLEMENTED(FATAL) << "Non-direct intrinsic slow-path not yet implemented";
+ UNREACHABLE();
+ }
+
+ // Copy the result back to the expected output.
+ Location out = invoke_->GetLocations()->Out();
+ if (out.IsValid()) {
+ DCHECK(out.IsRegister()); // TODO: Replace this when we support output in memory.
+ DCHECK(!invoke_->GetLocations()->GetLiveRegisters()->ContainsCoreRegister(out.reg()));
+ MoveFromReturnRegister(out, invoke_->GetType(), codegen);
+ }
+
+ codegen->RestoreLiveRegisters(invoke_->GetLocations());
+ __ b(GetExitLabel());
+ }
+
+ private:
+ // The instruction where this slow path is happening.
+ HInvoke* const invoke_;
+
+ DISALLOW_COPY_AND_ASSIGN(IntrinsicSlowPathARM);
+};
+
+#undef __
+
+bool IntrinsicLocationsBuilderARM::TryDispatch(HInvoke* invoke) {
+ Dispatch(invoke);
+ LocationSummary* res = invoke->GetLocations();
+ return res != nullptr && res->Intrinsified();
+}
+
+#define __ assembler->
+
+static void CreateFPToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresRegister());
+}
+
+static void CreateIntToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresFpuRegister());
+}
+
+static void MoveFPToInt(LocationSummary* locations, bool is64bit, ArmAssembler* assembler) {
+ Location input = locations->InAt(0);
+ Location output = locations->Out();
+ if (is64bit) {
+ __ vmovrrd(output.AsRegisterPairLow<Register>(),
+ output.AsRegisterPairHigh<Register>(),
+ FromLowSToD(input.AsFpuRegisterPairLow<SRegister>()));
+ } else {
+ __ vmovrs(output.AsRegister<Register>(), input.AsFpuRegister<SRegister>());
+ }
+}
+
+static void MoveIntToFP(LocationSummary* locations, bool is64bit, ArmAssembler* assembler) {
+ Location input = locations->InAt(0);
+ Location output = locations->Out();
+ if (is64bit) {
+ __ vmovdrr(FromLowSToD(output.AsFpuRegisterPairLow<SRegister>()),
+ input.AsRegisterPairLow<Register>(),
+ input.AsRegisterPairHigh<Register>());
+ } else {
+ __ vmovsr(output.AsFpuRegister<SRegister>(), input.AsRegister<Register>());
+ }
+}
+
+void IntrinsicLocationsBuilderARM::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+ CreateFPToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+ CreateIntToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitDoubleDoubleToRawLongBits(HInvoke* invoke) {
+ MoveFPToInt(invoke->GetLocations(), true, GetAssembler());
+}
+void IntrinsicCodeGeneratorARM::VisitDoubleLongBitsToDouble(HInvoke* invoke) {
+ MoveIntToFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderARM::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+ CreateFPToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitFloatIntBitsToFloat(HInvoke* invoke) {
+ CreateIntToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitFloatFloatToRawIntBits(HInvoke* invoke) {
+ MoveFPToInt(invoke->GetLocations(), false, GetAssembler());
+}
+void IntrinsicCodeGeneratorARM::VisitFloatIntBitsToFloat(HInvoke* invoke) {
+ MoveIntToFP(invoke->GetLocations(), false, GetAssembler());
+}
+
+static void CreateIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+}
+
+static void CreateFPToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresFpuRegister());
+ locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap);
+}
+
+static void MathAbsFP(LocationSummary* locations, bool is64bit, ArmAssembler* assembler) {
+ Location in = locations->InAt(0);
+ Location out = locations->Out();
+
+ if (is64bit) {
+ __ vabsd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()),
+ FromLowSToD(in.AsFpuRegisterPairLow<SRegister>()));
+ } else {
+ __ vabss(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>());
+ }
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathAbsDouble(HInvoke* invoke) {
+ CreateFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathAbsDouble(HInvoke* invoke) {
+ MathAbsFP(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathAbsFloat(HInvoke* invoke) {
+ CreateFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathAbsFloat(HInvoke* invoke) {
+ MathAbsFP(invoke->GetLocations(), false, GetAssembler());
+}
+
+static void CreateIntToIntPlusTemp(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+
+ locations->AddTemp(Location::RequiresRegister());
+}
+
+static void GenAbsInteger(LocationSummary* locations,
+ bool is64bit,
+ ArmAssembler* assembler) {
+ Location in = locations->InAt(0);
+ Location output = locations->Out();
+
+ Register mask = locations->GetTemp(0).AsRegister<Register>();
+
+ if (is64bit) {
+ Register in_reg_lo = in.AsRegisterPairLow<Register>();
+ Register in_reg_hi = in.AsRegisterPairHigh<Register>();
+ Register out_reg_lo = output.AsRegisterPairLow<Register>();
+ Register out_reg_hi = output.AsRegisterPairHigh<Register>();
+
+ DCHECK_NE(out_reg_lo, in_reg_hi) << "Diagonal overlap unexpected.";
+
+ __ Asr(mask, in_reg_hi, 31);
+ __ adds(out_reg_lo, in_reg_lo, ShifterOperand(mask));
+ __ adc(out_reg_hi, in_reg_hi, ShifterOperand(mask));
+ __ eor(out_reg_lo, mask, ShifterOperand(out_reg_lo));
+ __ eor(out_reg_hi, mask, ShifterOperand(out_reg_hi));
+ } else {
+ Register in_reg = in.AsRegister<Register>();
+ Register out_reg = output.AsRegister<Register>();
+
+ __ Asr(mask, in_reg, 31);
+ __ add(out_reg, in_reg, ShifterOperand(mask));
+ __ eor(out_reg, mask, ShifterOperand(out_reg));
+ }
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathAbsInt(HInvoke* invoke) {
+ CreateIntToIntPlusTemp(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathAbsInt(HInvoke* invoke) {
+ GenAbsInteger(invoke->GetLocations(), false, GetAssembler());
+}
+
+
+void IntrinsicLocationsBuilderARM::VisitMathAbsLong(HInvoke* invoke) {
+ CreateIntToIntPlusTemp(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathAbsLong(HInvoke* invoke) {
+ GenAbsInteger(invoke->GetLocations(), true, GetAssembler());
+}
+
+static void GenMinMax(LocationSummary* locations,
+ bool is_min,
+ ArmAssembler* assembler) {
+ Register op1 = locations->InAt(0).AsRegister<Register>();
+ Register op2 = locations->InAt(1).AsRegister<Register>();
+ Register out = locations->Out().AsRegister<Register>();
+
+ __ cmp(op1, ShifterOperand(op2));
+
+ __ it((is_min) ? Condition::LT : Condition::GT, kItElse);
+ __ mov(out, ShifterOperand(op1), is_min ? Condition::LT : Condition::GT);
+ __ mov(out, ShifterOperand(op2), is_min ? Condition::GE : Condition::LE);
+}
+
+static void CreateIntIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathMinIntInt(HInvoke* invoke) {
+ CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathMinIntInt(HInvoke* invoke) {
+ GenMinMax(invoke->GetLocations(), true, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathMaxIntInt(HInvoke* invoke) {
+ CreateIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathMaxIntInt(HInvoke* invoke) {
+ GenMinMax(invoke->GetLocations(), false, GetAssembler());
+}
+
+void IntrinsicLocationsBuilderARM::VisitMathSqrt(HInvoke* invoke) {
+ CreateFPToFPLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMathSqrt(HInvoke* invoke) {
+ LocationSummary* locations = invoke->GetLocations();
+ ArmAssembler* assembler = GetAssembler();
+ __ vsqrtd(FromLowSToD(locations->Out().AsFpuRegisterPairLow<SRegister>()),
+ FromLowSToD(locations->InAt(0).AsFpuRegisterPairLow<SRegister>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPeekByte(HInvoke* invoke) {
+ CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPeekByte(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ // Ignore upper 4B of long address.
+ __ ldrsb(invoke->GetLocations()->Out().AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPeekIntNative(HInvoke* invoke) {
+ CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPeekIntNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ // Ignore upper 4B of long address.
+ __ ldr(invoke->GetLocations()->Out().AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPeekLongNative(HInvoke* invoke) {
+ CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPeekLongNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ // Ignore upper 4B of long address.
+ Register addr = invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>();
+ // Worst case: Control register bit SCTLR.A = 0. Then unaligned accesses throw a processor
+ // exception. So we can't use ldrd as addr may be unaligned.
+ Register lo = invoke->GetLocations()->Out().AsRegisterPairLow<Register>();
+ Register hi = invoke->GetLocations()->Out().AsRegisterPairHigh<Register>();
+ if (addr == lo) {
+ __ ldr(hi, Address(addr, 4));
+ __ ldr(lo, Address(addr, 0));
+ } else {
+ __ ldr(lo, Address(addr, 0));
+ __ ldr(hi, Address(addr, 4));
+ }
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPeekShortNative(HInvoke* invoke) {
+ CreateIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPeekShortNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ // Ignore upper 4B of long address.
+ __ ldrsh(invoke->GetLocations()->Out().AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+static void CreateIntIntToVoidLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RequiresRegister());
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPokeByte(HInvoke* invoke) {
+ CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPokeByte(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ __ strb(invoke->GetLocations()->InAt(1).AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPokeIntNative(HInvoke* invoke) {
+ CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPokeIntNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ __ str(invoke->GetLocations()->InAt(1).AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPokeLongNative(HInvoke* invoke) {
+ CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPokeLongNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ // Ignore upper 4B of long address.
+ Register addr = invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>();
+ // Worst case: Control register bit SCTLR.A = 0. Then unaligned accesses throw a processor
+ // exception. So we can't use ldrd as addr may be unaligned.
+ __ str(invoke->GetLocations()->InAt(1).AsRegisterPairLow<Register>(), Address(addr, 0));
+ __ str(invoke->GetLocations()->InAt(1).AsRegisterPairHigh<Register>(), Address(addr, 4));
+}
+
+void IntrinsicLocationsBuilderARM::VisitMemoryPokeShortNative(HInvoke* invoke) {
+ CreateIntIntToVoidLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitMemoryPokeShortNative(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ __ strh(invoke->GetLocations()->InAt(1).AsRegister<Register>(),
+ Address(invoke->GetLocations()->InAt(0).AsRegisterPairLow<Register>()));
+}
+
+void IntrinsicLocationsBuilderARM::VisitThreadCurrentThread(HInvoke* invoke) {
+ LocationSummary* locations = new (arena_) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetOut(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorARM::VisitThreadCurrentThread(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ __ LoadFromOffset(kLoadWord,
+ invoke->GetLocations()->Out().AsRegister<Register>(),
+ TR,
+ Thread::PeerOffset<kArmPointerSize>().Int32Value());
+}
+
+static void GenUnsafeGet(HInvoke* invoke,
+ Primitive::Type type,
+ bool is_volatile,
+ CodeGeneratorARM* codegen) {
+ LocationSummary* locations = invoke->GetLocations();
+ DCHECK((type == Primitive::kPrimInt) ||
+ (type == Primitive::kPrimLong) ||
+ (type == Primitive::kPrimNot));
+ ArmAssembler* assembler = codegen->GetAssembler();
+ Register base = locations->InAt(1).AsRegister<Register>(); // Object pointer.
+ Register offset = locations->InAt(2).AsRegisterPairLow<Register>(); // Long offset, lo part only.
+
+ if (type == Primitive::kPrimLong) {
+ Register trg_lo = locations->Out().AsRegisterPairLow<Register>();
+ __ add(IP, base, ShifterOperand(offset));
+ if (is_volatile && !codegen->GetInstructionSetFeatures().HasAtomicLdrdAndStrd()) {
+ Register trg_hi = locations->Out().AsRegisterPairHigh<Register>();
+ __ ldrexd(trg_lo, trg_hi, IP);
+ } else {
+ __ ldrd(trg_lo, Address(IP));
+ }
+ } else {
+ Register trg = locations->Out().AsRegister<Register>();
+ __ ldr(trg, Address(base, offset));
+ }
+
+ if (is_volatile) {
+ __ dmb(ISH);
+ }
+}
+
+static void CreateIntIntIntToIntLocations(ArenaAllocator* arena, HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::NoLocation()); // Unused receiver.
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetInAt(2, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+}
+
+void IntrinsicLocationsBuilderARM::VisitUnsafeGet(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeGetVolatile(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeGetLong(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeGetObject(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeGetObjectVolatile(HInvoke* invoke) {
+ CreateIntIntIntToIntLocations(arena_, invoke);
+}
+
+void IntrinsicCodeGeneratorARM::VisitUnsafeGet(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimInt, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeGetVolatile(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimInt, true, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeGetLong(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimLong, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeGetLongVolatile(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimLong, true, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeGetObject(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimNot, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeGetObjectVolatile(HInvoke* invoke) {
+ GenUnsafeGet(invoke, Primitive::kPrimNot, true, codegen_);
+}
+
+static void CreateIntIntIntIntToVoid(ArenaAllocator* arena,
+ const ArmInstructionSetFeatures& features,
+ Primitive::Type type,
+ bool is_volatile,
+ HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::NoLocation()); // Unused receiver.
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetInAt(2, Location::RequiresRegister());
+ locations->SetInAt(3, Location::RequiresRegister());
+
+ if (type == Primitive::kPrimLong) {
+ // Potentially need temps for ldrexd-strexd loop.
+ if (is_volatile && !features.HasAtomicLdrdAndStrd()) {
+ locations->AddTemp(Location::RequiresRegister()); // Temp_lo.
+ locations->AddTemp(Location::RequiresRegister()); // Temp_hi.
+ }
+ } else if (type == Primitive::kPrimNot) {
+ // Temps for card-marking.
+ locations->AddTemp(Location::RequiresRegister()); // Temp.
+ locations->AddTemp(Location::RequiresRegister()); // Card.
+ }
+}
+
+void IntrinsicLocationsBuilderARM::VisitUnsafePut(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimInt, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutOrdered(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimInt, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutVolatile(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimInt, true, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutObject(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimNot, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimNot, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimNot, true, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutLong(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimLong, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimLong, false, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+ CreateIntIntIntIntToVoid(arena_, features_, Primitive::kPrimLong, true, invoke);
+}
+
+static void GenUnsafePut(LocationSummary* locations,
+ Primitive::Type type,
+ bool is_volatile,
+ bool is_ordered,
+ CodeGeneratorARM* codegen) {
+ ArmAssembler* assembler = codegen->GetAssembler();
+
+ Register base = locations->InAt(1).AsRegister<Register>(); // Object pointer.
+ Register offset = locations->InAt(2).AsRegisterPairLow<Register>(); // Long offset, lo part only.
+ Register value;
+
+ if (is_volatile || is_ordered) {
+ __ dmb(ISH);
+ }
+
+ if (type == Primitive::kPrimLong) {
+ Register value_lo = locations->InAt(3).AsRegisterPairLow<Register>();
+ value = value_lo;
+ if (is_volatile && !codegen->GetInstructionSetFeatures().HasAtomicLdrdAndStrd()) {
+ Register temp_lo = locations->GetTemp(0).AsRegister<Register>();
+ Register temp_hi = locations->GetTemp(1).AsRegister<Register>();
+ Register value_hi = locations->InAt(3).AsRegisterPairHigh<Register>();
+
+ __ add(IP, base, ShifterOperand(offset));
+ Label loop_head;
+ __ Bind(&loop_head);
+ __ ldrexd(temp_lo, temp_hi, IP);
+ __ strexd(temp_lo, value_lo, value_hi, IP);
+ __ cmp(temp_lo, ShifterOperand(0));
+ __ b(&loop_head, NE);
+ } else {
+ __ add(IP, base, ShifterOperand(offset));
+ __ strd(value_lo, Address(IP));
+ }
+ } else {
+ value = locations->InAt(3).AsRegister<Register>();
+ __ str(value, Address(base, offset));
+ }
+
+ if (is_volatile) {
+ __ dmb(ISH);
+ }
+
+ if (type == Primitive::kPrimNot) {
+ Register temp = locations->GetTemp(0).AsRegister<Register>();
+ Register card = locations->GetTemp(1).AsRegister<Register>();
+ codegen->MarkGCCard(temp, card, base, value);
+ }
+}
+
+void IntrinsicCodeGeneratorARM::VisitUnsafePut(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutOrdered(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, false, true, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutVolatile(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimInt, true, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutObject(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutObjectOrdered(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, false, true, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutObjectVolatile(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimNot, true, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutLong(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, false, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutLongOrdered(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, false, true, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafePutLongVolatile(HInvoke* invoke) {
+ GenUnsafePut(invoke->GetLocations(), Primitive::kPrimLong, true, false, codegen_);
+}
+
+static void CreateIntIntIntIntIntToIntPlusTemps(ArenaAllocator* arena,
+ HInvoke* invoke) {
+ LocationSummary* locations = new (arena) LocationSummary(invoke,
+ LocationSummary::kNoCall,
+ kIntrinsified);
+ locations->SetInAt(0, Location::NoLocation()); // Unused receiver.
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetInAt(2, Location::RequiresRegister());
+ locations->SetInAt(3, Location::RequiresRegister());
+ locations->SetInAt(4, Location::RequiresRegister());
+
+ locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap);
+
+ locations->AddTemp(Location::RequiresRegister()); // Pointer.
+ locations->AddTemp(Location::RequiresRegister()); // Temp 1.
+ locations->AddTemp(Location::RequiresRegister()); // Temp 2.
+}
+
+static void GenCas(LocationSummary* locations, Primitive::Type type, CodeGeneratorARM* codegen) {
+ DCHECK_NE(type, Primitive::kPrimLong);
+
+ ArmAssembler* assembler = codegen->GetAssembler();
+
+ Register out = locations->Out().AsRegister<Register>(); // Boolean result.
+
+ Register base = locations->InAt(1).AsRegister<Register>(); // Object pointer.
+ Register offset = locations->InAt(2).AsRegisterPairLow<Register>(); // Offset (discard high 4B).
+ Register expected_lo = locations->InAt(3).AsRegister<Register>(); // Expected.
+ Register value_lo = locations->InAt(4).AsRegister<Register>(); // Value.
+
+ Register tmp_ptr = locations->GetTemp(0).AsRegister<Register>(); // Pointer to actual memory.
+ Register tmp_lo = locations->GetTemp(1).AsRegister<Register>(); // Value in memory.
+
+ if (type == Primitive::kPrimNot) {
+ // Mark card for object assuming new value is stored. Worst case we will mark an unchanged
+ // object and scan the receiver at the next GC for nothing.
+ codegen->MarkGCCard(tmp_ptr, tmp_lo, base, value_lo);
+ }
+
+ // Prevent reordering with prior memory operations.
+ __ dmb(ISH);
+
+ __ add(tmp_ptr, base, ShifterOperand(offset));
+
+ // do {
+ // tmp = [r_ptr] - expected;
+ // } while (tmp == 0 && failure([r_ptr] <- r_new_value));
+ // result = tmp != 0;
+
+ Label loop_head;
+ __ Bind(&loop_head);
+
+ __ ldrex(tmp_lo, tmp_ptr);
+
+ __ subs(tmp_lo, tmp_lo, ShifterOperand(expected_lo));
+
+ __ it(EQ, ItState::kItT);
+ __ strex(tmp_lo, value_lo, tmp_ptr, EQ);
+ __ cmp(tmp_lo, ShifterOperand(1), EQ);
+
+ __ b(&loop_head, EQ);
+
+ __ dmb(ISH);
+
+ __ rsbs(out, tmp_lo, ShifterOperand(1));
+ __ it(CC);
+ __ mov(out, ShifterOperand(0), CC);
+}
+
+void IntrinsicLocationsBuilderARM::VisitUnsafeCASInt(HInvoke* invoke ATTRIBUTE_UNUSED) {
+ CreateIntIntIntIntIntToIntPlusTemps(arena_, invoke);
+}
+void IntrinsicLocationsBuilderARM::VisitUnsafeCASObject(HInvoke* invoke ATTRIBUTE_UNUSED) {
+ CreateIntIntIntIntIntToIntPlusTemps(arena_, invoke);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeCASInt(HInvoke* invoke) {
+ GenCas(invoke->GetLocations(), Primitive::kPrimInt, codegen_);
+}
+void IntrinsicCodeGeneratorARM::VisitUnsafeCASObject(HInvoke* invoke) {
+ GenCas(invoke->GetLocations(), Primitive::kPrimNot, codegen_);
+}
+
+void IntrinsicLocationsBuilderARM::VisitStringCharAt(HInvoke* invoke) {
+ LocationSummary* locations = new (arena_) LocationSummary(invoke,
+ LocationSummary::kCallOnSlowPath,
+ kIntrinsified);
+ locations->SetInAt(0, Location::RequiresRegister());
+ locations->SetInAt(1, Location::RequiresRegister());
+ locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap);
+
+ locations->AddTemp(Location::RequiresRegister());
+ locations->AddTemp(Location::RequiresRegister());
+}
+
+void IntrinsicCodeGeneratorARM::VisitStringCharAt(HInvoke* invoke) {
+ ArmAssembler* assembler = GetAssembler();
+ LocationSummary* locations = invoke->GetLocations();
+
+ // Location of reference to data array
+ const MemberOffset value_offset = mirror::String::ValueOffset();
+ // Location of count
+ const MemberOffset count_offset = mirror::String::CountOffset();
+ // Starting offset within data array
+ const MemberOffset offset_offset = mirror::String::OffsetOffset();
+ // Start of char data with array_
+ const MemberOffset data_offset = mirror::Array::DataOffset(sizeof(uint16_t));
+
+ Register obj = locations->InAt(0).AsRegister<Register>(); // String object pointer.
+ Register idx = locations->InAt(1).AsRegister<Register>(); // Index of character.
+ Register out = locations->Out().AsRegister<Register>(); // Result character.
+
+ Register temp = locations->GetTemp(0).AsRegister<Register>();
+ Register array_temp = locations->GetTemp(1).AsRegister<Register>();
+
+ // TODO: Maybe we can support range check elimination. Overall, though, I think it's not worth
+ // the cost.
+ // TODO: For simplicity, the index parameter is requested in a register, so different from Quick
+ // we will not optimize the code for constants (which would save a register).
+
+ SlowPathCodeARM* slow_path = new (GetAllocator()) IntrinsicSlowPathARM(invoke);
+ codegen_->AddSlowPath(slow_path);
+
+ __ ldr(temp, Address(obj, count_offset.Int32Value())); // temp = str.length.
+ codegen_->MaybeRecordImplicitNullCheck(invoke);
+ __ cmp(idx, ShifterOperand(temp));
+ __ b(slow_path->GetEntryLabel(), CS);
+
+ // Index computation.
+ __ ldr(temp, Address(obj, offset_offset.Int32Value())); // temp := str.offset.
+ __ ldr(array_temp, Address(obj, value_offset.Int32Value())); // array_temp := str.offset.
+ __ add(temp, temp, ShifterOperand(idx));
+ DCHECK_EQ(data_offset.Int32Value() % 2, 0); // We'll compensate by shifting.
+ __ add(temp, temp, ShifterOperand(data_offset.Int32Value() / 2));
+
+ // Load the value.
+ __ ldrh(out, Address(array_temp, temp, LSL, 1)); // out := array_temp[temp].
+
+ __ Bind(slow_path->GetExitLabel());
+}
+
+// Unimplemented intrinsics.
+
+#define UNIMPLEMENTED_INTRINSIC(Name) \
+void IntrinsicLocationsBuilderARM::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) { \
+} \
+void IntrinsicCodeGeneratorARM::Visit ## Name(HInvoke* invoke ATTRIBUTE_UNUSED) { \
+}
+
+UNIMPLEMENTED_INTRINSIC(IntegerReverse)
+UNIMPLEMENTED_INTRINSIC(IntegerReverseBytes)
+UNIMPLEMENTED_INTRINSIC(LongReverse)
+UNIMPLEMENTED_INTRINSIC(LongReverseBytes)
+UNIMPLEMENTED_INTRINSIC(ShortReverseBytes)
+UNIMPLEMENTED_INTRINSIC(MathMinDoubleDouble)
+UNIMPLEMENTED_INTRINSIC(MathMinFloatFloat)
+UNIMPLEMENTED_INTRINSIC(MathMaxDoubleDouble)
+UNIMPLEMENTED_INTRINSIC(MathMaxFloatFloat)
+UNIMPLEMENTED_INTRINSIC(MathMinLongLong)
+UNIMPLEMENTED_INTRINSIC(MathMaxLongLong)
+UNIMPLEMENTED_INTRINSIC(MathCeil) // Could be done by changing rounding mode, maybe?
+UNIMPLEMENTED_INTRINSIC(MathFloor) // Could be done by changing rounding mode, maybe?
+UNIMPLEMENTED_INTRINSIC(MathRint)
+UNIMPLEMENTED_INTRINSIC(MathRoundDouble) // Could be done by changing rounding mode, maybe?
+UNIMPLEMENTED_INTRINSIC(MathRoundFloat) // Could be done by changing rounding mode, maybe?
+UNIMPLEMENTED_INTRINSIC(UnsafeCASLong) // High register pressure.
+UNIMPLEMENTED_INTRINSIC(SystemArrayCopyChar)
+UNIMPLEMENTED_INTRINSIC(StringCompareTo)
+UNIMPLEMENTED_INTRINSIC(StringIsEmpty) // Might not want to do these two anyways, inlining should
+UNIMPLEMENTED_INTRINSIC(StringLength) // be good enough here.
+UNIMPLEMENTED_INTRINSIC(StringIndexOf)
+UNIMPLEMENTED_INTRINSIC(StringIndexOfAfter)
+UNIMPLEMENTED_INTRINSIC(ReferenceGetReferent)
+
+} // namespace arm
+} // namespace art
diff --git a/compiler/optimizing/intrinsics_arm.h b/compiler/optimizing/intrinsics_arm.h
new file mode 100644
index 0000000..8bfb7d4
--- /dev/null
+++ b/compiler/optimizing/intrinsics_arm.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (C) 2015 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#ifndef ART_COMPILER_OPTIMIZING_INTRINSICS_ARM_H_
+#define ART_COMPILER_OPTIMIZING_INTRINSICS_ARM_H_
+
+#include "intrinsics.h"
+
+namespace art {
+
+class ArenaAllocator;
+class ArmInstructionSetFeatures;
+class HInvokeStaticOrDirect;
+class HInvokeVirtual;
+
+namespace arm {
+
+class ArmAssembler;
+class CodeGeneratorARM;
+
+class IntrinsicLocationsBuilderARM FINAL : public IntrinsicVisitor {
+ public:
+ explicit IntrinsicLocationsBuilderARM(ArenaAllocator* arena,
+ const ArmInstructionSetFeatures& features)
+ : arena_(arena), features_(features) {}
+
+ // Define visitor methods.
+
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) \
+ void Visit ## Name(HInvoke* invoke) OVERRIDE;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+ // Check whether an invoke is an intrinsic, and if so, create a location summary. Returns whether
+ // a corresponding LocationSummary with the intrinsified_ flag set was generated and attached to
+ // the invoke.
+ bool TryDispatch(HInvoke* invoke);
+
+ private:
+ ArenaAllocator* arena_;
+
+ const ArmInstructionSetFeatures& features_;
+
+ DISALLOW_COPY_AND_ASSIGN(IntrinsicLocationsBuilderARM);
+};
+
+class IntrinsicCodeGeneratorARM FINAL : public IntrinsicVisitor {
+ public:
+ explicit IntrinsicCodeGeneratorARM(CodeGeneratorARM* codegen) : codegen_(codegen) {}
+
+ // Define visitor methods.
+
+#define OPTIMIZING_INTRINSICS(Name, IsStatic) \
+ void Visit ## Name(HInvoke* invoke) OVERRIDE;
+#include "intrinsics_list.h"
+INTRINSICS_LIST(OPTIMIZING_INTRINSICS)
+#undef INTRINSICS_LIST
+#undef OPTIMIZING_INTRINSICS
+
+ private:
+ ArmAssembler* GetAssembler();
+
+ ArenaAllocator* GetAllocator();
+
+ CodeGeneratorARM* codegen_;
+
+ DISALLOW_COPY_AND_ASSIGN(IntrinsicCodeGeneratorARM);
+};
+
+} // namespace arm
+} // namespace art
+
+#endif // ART_COMPILER_OPTIMIZING_INTRINSICS_ARM_H_
diff --git a/compiler/optimizing/intrinsics_arm64.cc b/compiler/optimizing/intrinsics_arm64.cc
index 7a3d7d8..8874edc 100644
--- a/compiler/optimizing/intrinsics_arm64.cc
+++ b/compiler/optimizing/intrinsics_arm64.cc
@@ -300,7 +300,6 @@
}
static void CreateFPToFPLocations(ArenaAllocator* arena, HInvoke* invoke) {
- // We only support FP registers here.
LocationSummary* locations = new (arena) LocationSummary(invoke,
LocationSummary::kNoCall,
kIntrinsified);
@@ -924,7 +923,6 @@
}
void IntrinsicLocationsBuilderARM64::VisitStringCharAt(HInvoke* invoke) {
- // The inputs plus one temp.
LocationSummary* locations = new (arena_) LocationSummary(invoke,
LocationSummary::kCallOnSlowPath,
kIntrinsified);
diff --git a/compiler/optimizing/nodes.h b/compiler/optimizing/nodes.h
index ee7701b..fd88e42 100644
--- a/compiler/optimizing/nodes.h
+++ b/compiler/optimizing/nodes.h
@@ -1015,6 +1015,18 @@
void SetLiveInterval(LiveInterval* interval) { live_interval_ = interval; }
bool HasLiveInterval() const { return live_interval_ != nullptr; }
+ bool IsSuspendCheckEntry() const { return IsSuspendCheck() && GetBlock()->IsEntryBlock(); }
+
+ // Returns whether the code generation of the instruction will require to have access
+ // to the current method. Such instructions are:
+ // (1): Instructions that require an environment, as calling the runtime requires
+ // to walk the stack and have the current method stored at a specific stack address.
+ // (2): Object literals like classes and strings, that are loaded from the dex cache
+ // fields of the current method.
+ bool NeedsCurrentMethod() const {
+ return NeedsEnvironment() || IsLoadClass() || IsLoadString();
+ }
+
private:
HInstruction* previous_;
HInstruction* next_;
diff --git a/compiler/optimizing/register_allocator.cc b/compiler/optimizing/register_allocator.cc
index 3809720..bfbe63f 100644
--- a/compiler/optimizing/register_allocator.cc
+++ b/compiler/optimizing/register_allocator.cc
@@ -252,8 +252,13 @@
&& (instruction->GetType() != Primitive::kPrimFloat);
if (locations->CanCall()) {
- if (!instruction->IsSuspendCheck()) {
- codegen_->MarkNotLeaf();
+ if (codegen_->IsLeafMethod()) {
+ // TODO: We do this here because we do not want the suspend check to artificially
+ // create live registers. We should find another place, but this is currently the
+ // simplest.
+ DCHECK(instruction->IsSuspendCheckEntry());
+ instruction->GetBlock()->RemoveInstruction(instruction);
+ return;
}
safepoints_.Add(instruction);
if (locations->OnlyCallsOnSlowPath()) {
diff --git a/compiler/optimizing/ssa_liveness_analysis.cc b/compiler/optimizing/ssa_liveness_analysis.cc
index 1b06315..2a84735 100644
--- a/compiler/optimizing/ssa_liveness_analysis.cc
+++ b/compiler/optimizing/ssa_liveness_analysis.cc
@@ -115,14 +115,13 @@
// to differentiate between the start and end of an instruction. Adding 2 to
// the lifetime position for each instruction ensures the start of an
// instruction is different than the end of the previous instruction.
- HGraphVisitor* location_builder = codegen_->GetLocationBuilder();
for (HLinearOrderIterator it(*this); !it.Done(); it.Advance()) {
HBasicBlock* block = it.Current();
block->SetLifetimeStart(lifetime_position);
for (HInstructionIterator inst_it(block->GetPhis()); !inst_it.Done(); inst_it.Advance()) {
HInstruction* current = inst_it.Current();
- current->Accept(location_builder);
+ codegen_->AllocateLocations(current);
LocationSummary* locations = current->GetLocations();
if (locations != nullptr && locations->Out().IsValid()) {
instructions_from_ssa_index_.Add(current);
@@ -140,7 +139,7 @@
for (HInstructionIterator inst_it(block->GetInstructions()); !inst_it.Done();
inst_it.Advance()) {
HInstruction* current = inst_it.Current();
- current->Accept(codegen_->GetLocationBuilder());
+ codegen_->AllocateLocations(current);
LocationSummary* locations = current->GetLocations();
if (locations != nullptr && locations->Out().IsValid()) {
instructions_from_ssa_index_.Add(current);
diff --git a/compiler/utils/arm/assembler_arm.cc b/compiler/utils/arm/assembler_arm.cc
index 0528773..1f44f19b23 100644
--- a/compiler/utils/arm/assembler_arm.cc
+++ b/compiler/utils/arm/assembler_arm.cc
@@ -245,6 +245,7 @@
// This is very like the ARM encoding except the offset is 10 bits.
uint32_t Address::encodingThumbLdrdStrd() const {
+ DCHECK(IsImmediate());
uint32_t encoding;
uint32_t am = am_;
// If P is 0 then W must be 1 (Different from ARM).
diff --git a/compiler/utils/arm/assembler_thumb2_test.cc b/compiler/utils/arm/assembler_thumb2_test.cc
index e571e72..ebea9d4 100644
--- a/compiler/utils/arm/assembler_thumb2_test.cc
+++ b/compiler/utils/arm/assembler_thumb2_test.cc
@@ -198,6 +198,18 @@
DriverStr(expected, "strexd");
}
+TEST_F(AssemblerThumb2Test, LdrdStrd) {
+ GetAssembler()->ldrd(arm::R0, arm::Address(arm::R2, 8));
+ GetAssembler()->ldrd(arm::R0, arm::Address(arm::R12));
+ GetAssembler()->strd(arm::R0, arm::Address(arm::R2, 8));
+
+ const char* expected =
+ "ldrd r0, r1, [r2, #8]\n"
+ "ldrd r0, r1, [r12]\n"
+ "strd r0, r1, [r2, #8]\n";
+ DriverStr(expected, "ldrdstrd");
+}
+
TEST_F(AssemblerThumb2Test, eor) {
#define __ GetAssembler()->
__ eor(arm::R1, arm::R1, arm::ShifterOperand(arm::R0));
diff --git a/compiler/utils/x86_64/assembler_x86_64.cc b/compiler/utils/x86_64/assembler_x86_64.cc
index 7e8e769..556fa9b 100644
--- a/compiler/utils/x86_64/assembler_x86_64.cc
+++ b/compiler/utils/x86_64/assembler_x86_64.cc
@@ -328,6 +328,14 @@
}
+void X86_64Assembler::leal(CpuRegister dst, const Address& src) {
+ AssemblerBuffer::EnsureCapacity ensured(&buffer_);
+ EmitOptionalRex32(dst, src);
+ EmitUint8(0x8D);
+ EmitOperand(dst.LowBits(), src);
+}
+
+
void X86_64Assembler::movaps(XmmRegister dst, XmmRegister src) {
AssemblerBuffer::EnsureCapacity ensured(&buffer_);
EmitOptionalRex32(dst, src);
diff --git a/compiler/utils/x86_64/assembler_x86_64.h b/compiler/utils/x86_64/assembler_x86_64.h
index 2fc251b..a1c704e 100644
--- a/compiler/utils/x86_64/assembler_x86_64.h
+++ b/compiler/utils/x86_64/assembler_x86_64.h
@@ -296,6 +296,7 @@
void movw(const Address& dst, const Immediate& imm);
void leaq(CpuRegister dst, const Address& src);
+ void leal(CpuRegister dst, const Address& src);
void movaps(XmmRegister dst, XmmRegister src);
diff --git a/runtime/class_linker.cc b/runtime/class_linker.cc
index e0df6db..25750bb 100644
--- a/runtime/class_linker.cc
+++ b/runtime/class_linker.cc
@@ -377,7 +377,8 @@
Handle<mirror::Class> java_lang_reflect_ArtMethod(hs.NewHandle(
AllocClass(self, java_lang_Class.Get(), mirror::ArtMethod::ClassSize())));
CHECK(java_lang_reflect_ArtMethod.Get() != nullptr);
- java_lang_reflect_ArtMethod->SetObjectSize(mirror::ArtMethod::InstanceSize(sizeof(void*)));
+ size_t pointer_size = GetInstructionSetPointerSize(Runtime::Current()->GetInstructionSet());
+ java_lang_reflect_ArtMethod->SetObjectSize(mirror::ArtMethod::InstanceSize(pointer_size));
SetClassRoot(kJavaLangReflectArtMethod, java_lang_reflect_ArtMethod.Get());
java_lang_reflect_ArtMethod->SetStatus(mirror::Class::kStatusResolved, self);
mirror::ArtMethod::SetClass(java_lang_reflect_ArtMethod.Get());
@@ -4297,7 +4298,9 @@
{
StackHandleScope<1> hs(self);
Handle<mirror::Class> return_type(hs.NewHandle(method1->GetReturnType()));
- if (UNLIKELY(method2->GetReturnType() != return_type.Get())) {
+ mirror::Class* other_return_type = method2->GetReturnType();
+ // NOTE: return_type.Get() must be sequenced after method2->GetReturnType().
+ if (UNLIKELY(other_return_type != return_type.Get())) {
return false;
}
}
@@ -4313,11 +4316,13 @@
return false;
}
for (uint32_t i = 0; i < num_types; ++i) {
- mirror::Class* param_type =
- method1->GetClassFromTypeIndex(types1->GetTypeItem(i).type_idx_, true);
+ StackHandleScope<1> hs(self);
+ Handle<mirror::Class> param_type(hs.NewHandle(
+ method1->GetClassFromTypeIndex(types1->GetTypeItem(i).type_idx_, true)));
mirror::Class* other_param_type =
method2->GetClassFromTypeIndex(types2->GetTypeItem(i).type_idx_, true);
- if (UNLIKELY(param_type != other_param_type)) {
+ // NOTE: param_type.Get() must be sequenced after method2->GetClassFromTypeIndex(...).
+ if (UNLIKELY(param_type.Get() != other_param_type)) {
return false;
}
}
@@ -5481,7 +5486,8 @@
klass->SetNumReferenceInstanceFields(num_reference_fields);
if (!klass->IsVariableSize()) {
if (klass->DescriptorEquals("Ljava/lang/reflect/ArtMethod;")) {
- klass->SetObjectSize(mirror::ArtMethod::InstanceSize(sizeof(void*)));
+ size_t pointer_size = GetInstructionSetPointerSize(Runtime::Current()->GetInstructionSet());
+ klass->SetObjectSize(mirror::ArtMethod::InstanceSize(pointer_size));
} else {
std::string temp;
DCHECK_GE(size, sizeof(mirror::Object)) << klass->GetDescriptor(&temp);
diff --git a/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc b/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
index 98f1684..00251ff 100644
--- a/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
+++ b/runtime/entrypoints/quick/quick_trampoline_entrypoints.cc
@@ -171,6 +171,7 @@
// | | padding
// | A0/Method* | <- sp
// NOTE: for Mip64, when A0 is skipped, F0 is also skipped.
+ static constexpr bool kSplitPairAcrossRegisterAndStack = false;
static constexpr bool kAlignPairRegister = false;
static constexpr bool kQuickSoftFloatAbi = false;
static constexpr bool kQuickDoubleRegAlignedFloatBackFilled = false;
diff --git a/runtime/parsed_options.cc b/runtime/parsed_options.cc
index b27bd6b..a5df892 100644
--- a/runtime/parsed_options.cc
+++ b/runtime/parsed_options.cc
@@ -464,13 +464,6 @@
}
args.Set(M::BackgroundGc, BackgroundGcOption { background_collector_type_ });
- {
- XGcOption* xgc = args.Get(M::GcOption);
- if (xgc != nullptr) {
- xgc->collector_type_ = collector_type_;
- args.Set(M::GcOption, *xgc);
- }
- }
}
// If a reference to the dalvik core.jar snuck in, replace it with
diff --git a/runtime/runtime.cc b/runtime/runtime.cc
index e868e75..6704963 100644
--- a/runtime/runtime.cc
+++ b/runtime/runtime.cc
@@ -909,10 +909,11 @@
dex_locations,
runtime_options.GetOrDefault(Opt::Image),
&boot_class_path);
+ instruction_set_ = runtime_options.GetOrDefault(Opt::ImageInstructionSet);
class_linker_->InitWithoutImage(std::move(boot_class_path));
// TODO: Should we move the following to InitWithoutImage?
- SetInstructionSet(kRuntimeISA);
+ SetInstructionSet(instruction_set_);
for (int i = 0; i < Runtime::kLastCalleeSaveType; i++) {
Runtime::CalleeSaveType type = Runtime::CalleeSaveType(i);
if (!HasCalleeSaveMethod(type)) {
@@ -1332,7 +1333,8 @@
method->SetDexMethodIndex(DexFile::kDexNoIndex);
// When compiling, the code pointer will get set later when the image is loaded.
if (runtime->IsCompiler()) {
- method->SetEntryPointFromQuickCompiledCode(nullptr);
+ size_t pointer_size = GetInstructionSetPointerSize(instruction_set_);
+ method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
} else {
method->SetEntryPointFromQuickCompiledCode(GetQuickImtConflictStub());
}
@@ -1350,7 +1352,8 @@
method->SetDexMethodIndex(DexFile::kDexNoIndex);
// When compiling, the code pointer will get set later when the image is loaded.
if (runtime->IsCompiler()) {
- method->SetEntryPointFromQuickCompiledCode(nullptr);
+ size_t pointer_size = GetInstructionSetPointerSize(instruction_set_);
+ method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
} else {
method->SetEntryPointFromQuickCompiledCode(GetQuickResolutionStub());
}
@@ -1366,7 +1369,8 @@
method->SetDeclaringClass(mirror::ArtMethod::GetJavaLangReflectArtMethod());
// TODO: use a special method for callee saves
method->SetDexMethodIndex(DexFile::kDexNoIndex);
- method->SetEntryPointFromQuickCompiledCode(nullptr);
+ size_t pointer_size = GetInstructionSetPointerSize(instruction_set_);
+ method->SetEntryPointFromQuickCompiledCodePtrSize(nullptr, pointer_size);
DCHECK_NE(instruction_set_, kNone);
return method.Get();
}
diff --git a/runtime/stack.cc b/runtime/stack.cc
index b771aa7..b39aebf 100644
--- a/runtime/stack.cc
+++ b/runtime/stack.cc
@@ -134,7 +134,8 @@
} else {
return cur_shadow_frame_->GetVRegReference(0);
}
- } else if (m->IsOptimized(sizeof(void*))) {
+ } else if (m->IsOptimized(GetInstructionSetPointerSize(
+ Runtime::Current()->GetInstructionSet()))) {
// TODO: Implement, currently only used for exceptions when jdwp is enabled.
UNIMPLEMENTED(WARNING)
<< "StackVisitor::GetThisObject is unimplemented with the optimizing compiler";
diff --git a/runtime/utils.cc b/runtime/utils.cc
index 8347b7e..af16d7e 100644
--- a/runtime/utils.cc
+++ b/runtime/utils.cc
@@ -1225,11 +1225,11 @@
// #XX. <RELATIVE_ADDR> has to be a hex number but with no 0x prefix.
os << prefix << StringPrintf("#%02zu pc ", it->num);
bool try_addr2line = false;
- if (!it->map) {
+ if (!BacktraceMap::IsValid(it->map)) {
os << StringPrintf("%08" PRIxPTR " ???", it->pc);
} else {
- os << StringPrintf("%08" PRIxPTR " ", it->pc - it->map->start);
- os << it->map->name;
+ os << StringPrintf("%08" PRIxPTR " ", it->pc - it->map.start);
+ os << it->map.name;
os << " (";
if (!it->func_name.empty()) {
os << it->func_name;
@@ -1250,7 +1250,7 @@
}
os << "\n";
if (try_addr2line && use_addr2line) {
- Addr2line(it->map->name, it->pc - it->map->start, os, prefix);
+ Addr2line(it->map.name, it->pc - it->map.start, os, prefix);
}
}
#else
diff --git a/test/004-UnsafeTest/src/Main.java b/test/004-UnsafeTest/src/Main.java
index 743d62c..3d0f074 100644
--- a/test/004-UnsafeTest/src/Main.java
+++ b/test/004-UnsafeTest/src/Main.java
@@ -94,6 +94,16 @@
unsafe.putLong(t, longOffset, longValue);
check(t.longVar, longValue, "Unsafe.putLong(Object, long, long)");
check(unsafe.getLong(t, longOffset), longValue, "Unsafe.getLong(Object, long)");
+
+ if (unsafe.compareAndSwapInt(t, intOffset, 0, 1)) {
+ System.out.println("Unexpectedly succeeding compareAndSwap...");
+ }
+ if (!unsafe.compareAndSwapInt(t, intOffset, intValue, 0)) {
+ System.out.println("Unexpectedly not succeeding compareAndSwap...");
+ }
+ if (!unsafe.compareAndSwapInt(t, intOffset, 0, 1)) {
+ System.out.println("Unexpectedly not succeeding compareAndSwap...");
+ }
}
private static class TestClass {
diff --git a/test/082-inline-execute/src/Main.java b/test/082-inline-execute/src/Main.java
index bf6c802..0e90c4d 100644
--- a/test/082-inline-execute/src/Main.java
+++ b/test/082-inline-execute/src/Main.java
@@ -375,6 +375,7 @@
Assert.assertEquals(Math.abs(Long.MAX_VALUE), Long.MAX_VALUE);
Assert.assertEquals(Math.abs(Long.MIN_VALUE), Long.MIN_VALUE);
Assert.assertEquals(Math.abs(Long.MIN_VALUE - 1), Long.MAX_VALUE);
+ Assert.assertEquals(Math.abs(2147483648L), 2147483648L);
}
public static void test_Math_min_I() {
diff --git a/test/444-checker-nce/src/Main.java b/test/444-checker-nce/src/Main.java
index cdca332..9fb9c46 100644
--- a/test/444-checker-nce/src/Main.java
+++ b/test/444-checker-nce/src/Main.java
@@ -39,11 +39,11 @@
}
// CHECK-START: Main Main.newInstanceRemoveTest() instruction_simplifier (before)
- // CHECK-DAG: NewInstance
- // CHECK-DAG: NullCheck
- // CHECK-DAG: InvokeStaticOrDirect
- // CHECK-DAG: NullCheck
- // CHECK-DAG: InvokeStaticOrDirect
+ // CHECK: NewInstance
+ // CHECK: NullCheck
+ // CHECK: InvokeStaticOrDirect
+ // CHECK: NullCheck
+ // CHECK: InvokeStaticOrDirect
// CHECK-START: Main Main.newInstanceRemoveTest() instruction_simplifier (after)
// CHECK-NOT: NullCheck
@@ -53,25 +53,25 @@
}
// CHECK-START: Main Main.newArrayRemoveTest() instruction_simplifier (before)
- // CHECK-DAG: NewArray
- // CHECK-DAG: NullCheck
- // CHECK-DAG: ArrayGet
+ // CHECK: NewArray
+ // CHECK: NullCheck
+ // CHECK: ArrayGet
// CHECK-START: Main Main.newArrayRemoveTest() instruction_simplifier (after)
- // CHECK-DAG: NewArray
+ // CHECK: NewArray
// CHECK-NOT: NullCheck
- // CHECK-DAG: ArrayGet
+ // CHECK: ArrayGet
public Main newArrayRemoveTest() {
Main[] ms = new Main[1];
return ms[0];
}
// CHECK-START: Main Main.ifRemoveTest(boolean) instruction_simplifier_after_types (before)
- // CHECK-DAG: NewInstance
- // CHECK-DAG: NullCheck
+ // CHECK: NewInstance
+ // CHECK: NullCheck
// CHECK-START: Main Main.ifRemoveTest(boolean) instruction_simplifier_after_types (after)
- // CHECK-DAG: NewInstance
+ // CHECK: NewInstance
// CHECK-NOT: NullCheck
public Main ifRemoveTest(boolean flag) {
Main m = null;
@@ -84,12 +84,12 @@
}
// CHECK-START: Main Main.ifKeepTest(boolean) instruction_simplifier_after_types (before)
- // CHECK-DAG: NewInstance
- // CHECK-DAG: NullCheck
+ // CHECK: NewInstance
+ // CHECK: NullCheck
// CHECK-START: Main Main.ifKeepTest(boolean) instruction_simplifier_after_types (after)
- // CHECK-DAG: NewInstance
- // CHECK-DAG: NullCheck
+ // CHECK: NewInstance
+ // CHECK: NullCheck
public Main ifKeepTest(boolean flag) {
Main m = null;
if (flag) {
diff --git a/test/449-checker-bce/expected.txt b/test/449-checker-bce/expected.txt
new file mode 100644
index 0000000..e69de29
--- /dev/null
+++ b/test/449-checker-bce/expected.txt
diff --git a/test/449-checker-bce/info.txt b/test/449-checker-bce/info.txt
new file mode 100644
index 0000000..0a08808
--- /dev/null
+++ b/test/449-checker-bce/info.txt
@@ -0,0 +1 @@
+Checker test for testing array bounds check elimination.
diff --git a/test/449-checker-bce/src/Main.java b/test/449-checker-bce/src/Main.java
new file mode 100644
index 0000000..5a0e13b
--- /dev/null
+++ b/test/449-checker-bce/src/Main.java
@@ -0,0 +1,96 @@
+/*
+* Copyright (C) 2015 The Android Open Source Project
+*
+* Licensed under the Apache License, Version 2.0 (the "License");
+* you may not use this file except in compliance with the License.
+* You may obtain a copy of the License at
+*
+* http://www.apache.org/licenses/LICENSE-2.0
+*
+* Unless required by applicable law or agreed to in writing, software
+* distributed under the License is distributed on an "AS IS" BASIS,
+* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+* See the License for the specific language governing permissions and
+* limitations under the License.
+*/
+
+public class Main {
+
+ // CHECK-START: int Main.sieve(int) BCE (before)
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK: BoundsCheck
+ // CHECK: ArrayGet
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+
+ // CHECK-START: int Main.sieve(int) BCE (after)
+ // CHECK-NOT: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK-NOT: BoundsCheck
+ // CHECK: ArrayGet
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+
+ static int sieve(int size) {
+ int primeCount = 0;
+ boolean[] flags = new boolean[size + 1];
+ for (int i = 1; i < size; i++) flags[i] = true; // Can eliminate.
+ for (int i = 2; i < size; i++) {
+ if (flags[i]) { // Can eliminate.
+ primeCount++;
+ for (int k = i + 1; k <= size; k += i)
+ flags[k - 1] = false; // Can't eliminate yet due to (k+i) may overflow.
+ }
+ }
+ return primeCount;
+ }
+
+ // CHECK-START: void Main.narrow(int[], int) BCE (before)
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+
+ // CHECK-START: void Main.narrow(int[], int) BCE (after)
+ // CHECK-NOT: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK-NOT: BoundsCheck
+ // CHECK: ArraySet
+ // CHECK: BoundsCheck
+ // CHECK: ArraySet
+
+ static void narrow(int array[], int offset) {
+ if (offset < 0) {
+ return;
+ }
+ if (offset < array.length) {
+ // offset is in range [0, array.length-1].
+ // Bounds check can be eliminated.
+ array[offset] = 1;
+
+ int biased_offset1 = offset + 1;
+ // biased_offset1 is in range [1, array.length].
+ if (biased_offset1 < array.length) {
+ // biased_offset1 is in range [1, array.length-1].
+ // Bounds check can be eliminated.
+ array[biased_offset1] = 1;
+ }
+
+ int biased_offset2 = offset + 0x70000000;
+ // biased_offset2 is in range [0x70000000, array.length-1+0x70000000].
+ // It may overflow and be negative.
+ if (biased_offset2 < array.length) {
+ // Even with this test, biased_offset2 can be negative so we can't
+ // eliminate this bounds check.
+ array[biased_offset2] = 1;
+ }
+ }
+ }
+
+ public static void main(String[] args) {
+ sieve(20);
+ }
+}