v8/src/wasm/wasm-code-manager.h
Clemens Hammacher ce2d01bca3 [wasm] Store protected instructions in an OwnedVector
We currently store the protected instructions per code object in a
{std::unique_ptr<std::vector<ProtectedInstructionData>>}. This wastes
memory, because it requires two heap allocations, plus the vector might
over-allocate (and it currently does, because it is filled dynamically
during compilation).
This CL changes that to store the protected instructions in an
{OwnedVector}. This requires one copy after generating the list of
{ProtectedInstructionData} in an {std::vector} during compilation, but
saves memory afterwards.

R=mstarzinger@chromium.org

Bug: chromium:856938
Change-Id: Ie290a17dc32f27fbbfe0c000a52297181c954550
Reviewed-on: https://chromium-review.googlesource.com/1116701
Reviewed-by: Michael Starzinger <mstarzinger@chromium.org>
Commit-Queue: Clemens Hammacher <clemensh@chromium.org>
Cr-Commit-Position: refs/heads/master@{#54052}
2018-06-27 12:22:10 +00:00

485 lines
18 KiB
C++

// Copyright 2017 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#ifndef V8_WASM_WASM_CODE_MANAGER_H_
#define V8_WASM_WASM_CODE_MANAGER_H_
#include <functional>
#include <list>
#include <map>
#include <unordered_map>
#include "src/base/macros.h"
#include "src/handles.h"
#include "src/trap-handler/trap-handler.h"
#include "src/vector.h"
#include "src/wasm/module-compiler.h"
namespace v8 {
namespace internal {
struct CodeDesc;
class Code;
class Histogram;
namespace wasm {
class NativeModule;
class WasmCodeManager;
struct WasmModule;
// Convenience macro listing all wasm runtime stubs. Note that the first few
// elements of the list coincide with {compiler::TrapId}, order matters.
#define WASM_RUNTIME_STUB_LIST(V, VTRAP) \
FOREACH_WASM_TRAPREASON(VTRAP) \
V(WasmAllocateHeapNumber) \
V(WasmArgumentsAdaptor) \
V(WasmCallJavaScript) \
V(WasmGrowMemory) \
V(WasmStackGuard) \
V(WasmToNumber) \
V(DoubleToI)
struct AddressRange {
Address start;
Address end;
AddressRange(Address s, Address e) : start(s), end(e) {
DCHECK_LE(start, end);
DCHECK_IMPLIES(start == kNullAddress, end == kNullAddress);
}
AddressRange() : AddressRange(kNullAddress, kNullAddress) {}
size_t size() const { return static_cast<size_t>(end - start); }
bool is_empty() const { return start == end; }
operator bool() const { return start == kNullAddress; }
};
// Sorted, disjoint and non-overlapping memory ranges. A range is of the
// form [start, end). So there's no [start, end), [end, other_end),
// because that should have been reduced to [start, other_end).
class V8_EXPORT_PRIVATE DisjointAllocationPool final {
public:
DisjointAllocationPool() = default;
explicit DisjointAllocationPool(AddressRange range) : ranges_({range}) {}
DisjointAllocationPool(DisjointAllocationPool&& other) = default;
DisjointAllocationPool& operator=(DisjointAllocationPool&& other) = default;
// Merge the parameter range into this object while preserving ordering of the
// ranges. The assumption is that the passed parameter is not intersecting
// this object - for example, it was obtained from a previous Allocate.
void Merge(AddressRange);
// Allocate a contiguous range of size {size}. Return an empty pool on
// failure.
AddressRange Allocate(size_t size);
bool IsEmpty() const { return ranges_.empty(); }
const std::list<AddressRange>& ranges() const { return ranges_; }
private:
std::list<AddressRange> ranges_;
DISALLOW_COPY_AND_ASSIGN(DisjointAllocationPool)
};
class V8_EXPORT_PRIVATE WasmCode final {
public:
enum Kind {
kFunction,
kWasmToJsWrapper,
kLazyStub,
kRuntimeStub,
kInterpreterEntry,
kJumpTable
};
// Each runtime stub is identified by an id. This id is used to reference the
// stub via {RelocInfo::WASM_STUB_CALL} and gets resolved during relocation.
enum RuntimeStubId {
#define DEF_ENUM(Name) k##Name,
#define DEF_ENUM_TRAP(Name) kThrowWasm##Name,
WASM_RUNTIME_STUB_LIST(DEF_ENUM, DEF_ENUM_TRAP)
#undef DEF_ENUM_TRAP
#undef DEF_ENUM
kRuntimeStubCount
};
// kOther is used if we have WasmCode that is neither
// liftoff- nor turbofan-compiled, i.e. if Kind is
// not a kFunction.
enum Tier : int8_t { kLiftoff, kTurbofan, kOther };
Vector<byte> instructions() const { return instructions_; }
Address instruction_start() const {
return reinterpret_cast<Address>(instructions_.start());
}
Vector<const byte> reloc_info() const {
return {reloc_info_.get(), reloc_size_};
}
Vector<const byte> source_positions() const {
return {source_position_table_.get(), source_position_size_};
}
uint32_t index() const { return index_.ToChecked(); }
// Anonymous functions are functions that don't carry an index.
bool IsAnonymous() const { return index_.IsNothing(); }
Kind kind() const { return kind_; }
NativeModule* native_module() const { return native_module_; }
Tier tier() const { return tier_; }
Address constant_pool() const;
size_t constant_pool_offset() const { return constant_pool_offset_; }
size_t safepoint_table_offset() const { return safepoint_table_offset_; }
size_t handler_table_offset() const { return handler_table_offset_; }
uint32_t stack_slots() const { return stack_slots_; }
bool is_liftoff() const { return tier_ == kLiftoff; }
bool contains(Address pc) const {
return reinterpret_cast<Address>(instructions_.start()) <= pc &&
pc < reinterpret_cast<Address>(instructions_.end());
}
Vector<trap_handler::ProtectedInstructionData> protected_instructions()
const {
return protected_instructions_.as_vector();
}
void Validate() const;
void Print(const char* name = nullptr) const;
void Disassemble(const char* name, std::ostream& os,
Address current_pc = kNullAddress) const;
static bool ShouldBeLogged(Isolate* isolate);
void LogCode(Isolate* isolate) const;
~WasmCode();
enum FlushICache : bool { kFlushICache = true, kNoFlushICache = false };
// Offset of {instructions_.start()}. It is used for tiering, when
// we check if optimized code is available during the prologue
// of Liftoff-compiled code.
static constexpr int kInstructionStartOffset = 0;
private:
friend class NativeModule;
WasmCode(Vector<byte> instructions, std::unique_ptr<const byte[]> reloc_info,
size_t reloc_size, std::unique_ptr<const byte[]> source_pos,
size_t source_pos_size, NativeModule* native_module,
Maybe<uint32_t> index, Kind kind, size_t constant_pool_offset,
uint32_t stack_slots, size_t safepoint_table_offset,
size_t handler_table_offset,
OwnedVector<trap_handler::ProtectedInstructionData>
protected_instructions,
Tier tier)
: instructions_(instructions),
reloc_info_(std::move(reloc_info)),
reloc_size_(reloc_size),
source_position_table_(std::move(source_pos)),
source_position_size_(source_pos_size),
native_module_(native_module),
index_(index),
kind_(kind),
constant_pool_offset_(constant_pool_offset),
stack_slots_(stack_slots),
safepoint_table_offset_(safepoint_table_offset),
handler_table_offset_(handler_table_offset),
protected_instructions_(std::move(protected_instructions)),
tier_(tier) {
DCHECK_LE(safepoint_table_offset, instructions.size());
DCHECK_LE(constant_pool_offset, instructions.size());
DCHECK_LE(handler_table_offset, instructions.size());
DCHECK_EQ(kInstructionStartOffset, OFFSET_OF(WasmCode, instructions_));
}
// Code objects that have been registered with the global trap handler within
// this process, will have a {trap_handler_index} associated with them.
size_t trap_handler_index() const;
void set_trap_handler_index(size_t);
bool HasTrapHandlerIndex() const;
// Register protected instruction information with the trap handler. Sets
// trap_handler_index.
void RegisterTrapHandlerData();
Vector<byte> instructions_;
std::unique_ptr<const byte[]> reloc_info_;
size_t reloc_size_ = 0;
std::unique_ptr<const byte[]> source_position_table_;
size_t source_position_size_ = 0;
NativeModule* native_module_ = nullptr;
Maybe<uint32_t> index_;
Kind kind_;
size_t constant_pool_offset_ = 0;
uint32_t stack_slots_ = 0;
// we care about safepoint data for wasm-to-js functions,
// since there may be stack/register tagged values for large number
// conversions.
size_t safepoint_table_offset_ = 0;
size_t handler_table_offset_ = 0;
intptr_t trap_handler_index_ = -1;
OwnedVector<trap_handler::ProtectedInstructionData> protected_instructions_;
Tier tier_;
DISALLOW_COPY_AND_ASSIGN(WasmCode);
};
// Return a textual description of the kind.
const char* GetWasmCodeKindAsString(WasmCode::Kind);
class V8_EXPORT_PRIVATE NativeModule final {
public:
WasmCode* AddCode(const CodeDesc& desc, uint32_t frame_count, uint32_t index,
size_t safepoint_table_offset, size_t handler_table_offset,
OwnedVector<trap_handler::ProtectedInstructionData>
protected_instructions,
OwnedVector<byte> source_position_table,
WasmCode::Tier tier);
// A way to copy over JS-allocated code. This is because we compile
// certain wrappers using a different pipeline.
WasmCode* AddCodeCopy(Handle<Code> code, WasmCode::Kind kind, uint32_t index);
// Add an interpreter entry. For the same reason as AddCodeCopy, we
// currently compile these using a different pipeline and we can't get a
// CodeDesc here. When adding interpreter wrappers, we do not insert them in
// the code_table, however, we let them self-identify as the {index} function
WasmCode* AddInterpreterEntry(Handle<Code> code, uint32_t index);
// When starting lazy compilation, provide the WasmLazyCompile builtin by
// calling SetLazyBuiltin. It will be copied into this NativeModule and the
// jump table will be populated with that copy.
void SetLazyBuiltin(Handle<Code> code);
// Initializes all runtime stubs by copying them over from the JS-allocated
// heap into this native module. It must be called exactly once per native
// module before adding other WasmCode so that runtime stub ids can be
// resolved during relocation.
void SetRuntimeStubs(Isolate* isolate);
WasmCode* code(uint32_t index) const {
DCHECK_LT(index, num_functions_);
DCHECK_LE(num_imported_functions_, index);
return code_table_[index - num_imported_functions_];
}
bool has_code(uint32_t index) const {
DCHECK_LT(index, num_functions_);
DCHECK_LE(num_imported_functions_, index);
return code_table_[index - num_imported_functions_] != nullptr;
}
WasmCode* runtime_stub(WasmCode::RuntimeStubId index) const {
DCHECK_LT(index, WasmCode::kRuntimeStubCount);
WasmCode* code = runtime_stub_table_[index];
DCHECK_NOT_NULL(code);
return code;
}
bool is_jump_table_slot(Address address) const {
return jump_table_->contains(address);
}
uint32_t GetFunctionIndexFromJumpTableSlot(Address slot_address);
// Transition this module from code relying on trap handlers (i.e. without
// explicit memory bounds checks) to code that does not require trap handlers
// (i.e. code with explicit bounds checks).
// This method must only be called if {use_trap_handler()} is true (it will be
// false afterwards). All code in this {NativeModule} needs to be re-added
// after calling this method.
void DisableTrapHandler();
// Returns the target to call for the given function (returns a jump table
// slot within {jump_table_}).
Address GetCallTargetForFunction(uint32_t func_index) const;
bool SetExecutable(bool executable);
// For cctests, where we build both WasmModule and the runtime objects
// on the fly, and bypass the instance builder pipeline.
void ReserveCodeTableForTesting(uint32_t max_functions);
void SetNumFunctionsForTesting(uint32_t num_functions);
void LogWasmCodes(Isolate* isolate);
CompilationState* compilation_state() { return compilation_state_.get(); }
uint32_t num_functions() const { return num_functions_; }
uint32_t num_imported_functions() const { return num_imported_functions_; }
Vector<WasmCode*> code_table() const {
return {code_table_.get(), num_functions_ - num_imported_functions_};
}
bool use_trap_handler() const { return use_trap_handler_; }
void set_lazy_compile_frozen(bool frozen) { lazy_compile_frozen_ = frozen; }
bool lazy_compile_frozen() const { return lazy_compile_frozen_; }
Vector<const byte> wire_bytes() const {
return {wire_bytes_.get(), wire_bytes_len_};
}
void set_wire_bytes(std::unique_ptr<const byte[]> wire_bytes, size_t length) {
wire_bytes_ = std::move(wire_bytes);
wire_bytes_len_ = length;
}
const WasmModule* module() const;
WasmCode* Lookup(Address) const;
const size_t instance_id = 0;
~NativeModule();
private:
friend class WasmCode;
friend class WasmCodeManager;
friend class NativeModuleSerializer;
friend class NativeModuleDeserializer;
friend class NativeModuleModificationScope;
static base::AtomicNumber<size_t> next_id_;
NativeModule(Isolate* isolate, uint32_t num_functions,
uint32_t num_imported_functions, bool can_request_more,
VirtualMemory* code_space, WasmCodeManager* code_manager,
ModuleEnv& env);
WasmCode* AddAnonymousCode(Handle<Code>, WasmCode::Kind kind);
Address AllocateForCode(size_t size);
// Primitive for adding code to the native module. All code added to a native
// module is owned by that module. Various callers get to decide on how the
// code is obtained (CodeDesc vs, as a point in time, Code*), the kind,
// whether it has an index or is anonymous, etc.
WasmCode* AddOwnedCode(Vector<const byte> orig_instructions,
std::unique_ptr<const byte[]> reloc_info,
size_t reloc_size,
std::unique_ptr<const byte[]> source_pos,
size_t source_pos_size, Maybe<uint32_t> index,
WasmCode::Kind kind, size_t constant_pool_offset,
uint32_t stack_slots, size_t safepoint_table_offset,
size_t handler_table_offset,
OwnedVector<trap_handler::ProtectedInstructionData>,
WasmCode::Tier, WasmCode::FlushICache);
WasmCode* CreateEmptyJumpTable(uint32_t num_wasm_functions);
void PatchJumpTable(uint32_t func_index, Address target,
WasmCode::FlushICache);
void set_code(uint32_t index, WasmCode* code) {
DCHECK_LT(index, num_functions_);
DCHECK_LE(num_imported_functions_, index);
DCHECK_EQ(code->index(), index);
code_table_[index - num_imported_functions_] = code;
}
// Holds all allocated code objects, is maintained to be in ascending order
// according to the codes instruction start address to allow lookups.
std::vector<std::unique_ptr<WasmCode>> owned_code_;
uint32_t num_functions_;
uint32_t num_imported_functions_;
std::unique_ptr<WasmCode* []> code_table_;
size_t wire_bytes_len_;
std::unique_ptr<const byte[]> wire_bytes_;
WasmCode* runtime_stub_table_[WasmCode::kRuntimeStubCount] = {nullptr};
// Jump table used to easily redirect wasm function calls.
WasmCode* jump_table_ = nullptr;
std::unique_ptr<CompilationState, CompilationStateDeleter> compilation_state_;
DisjointAllocationPool free_code_space_;
DisjointAllocationPool allocated_code_space_;
std::list<VirtualMemory> owned_code_space_;
WasmCodeManager* wasm_code_manager_;
base::Mutex allocation_mutex_;
size_t committed_code_space_ = 0;
int modification_scope_depth_ = 0;
bool can_request_more_memory_;
bool use_trap_handler_ = false;
bool is_executable_ = false;
bool lazy_compile_frozen_ = false;
DISALLOW_COPY_AND_ASSIGN(NativeModule);
};
class V8_EXPORT_PRIVATE WasmCodeManager final {
public:
explicit WasmCodeManager(size_t max_committed);
// Create a new NativeModule. The caller is responsible for its
// lifetime. The native module will be given some memory for code,
// which will be page size aligned. The size of the initial memory
// is determined with a heuristic based on the total size of wasm
// code. The native module may later request more memory.
// TODO(titzer): isolate is only required here for CompilationState.
std::unique_ptr<NativeModule> NewNativeModule(Isolate* isolate,
ModuleEnv& env);
// TODO(titzer): isolate is only required here for CompilationState.
std::unique_ptr<NativeModule> NewNativeModule(
Isolate* isolate, size_t memory_estimate, uint32_t num_functions,
uint32_t num_imported_functions, bool can_request_more, ModuleEnv& env);
NativeModule* LookupNativeModule(Address pc) const;
WasmCode* LookupCode(Address pc) const;
WasmCode* GetCodeFromStartAddress(Address pc) const;
size_t remaining_uncommitted_code_space() const;
void SetModuleCodeSizeHistogram(Histogram* histogram) {
module_code_size_mb_ = histogram;
}
static size_t EstimateNativeModuleSize(const WasmModule* module);
private:
friend class NativeModule;
void TryAllocate(size_t size, VirtualMemory*, void* hint = nullptr);
bool Commit(Address, size_t);
// Currently, we uncommit a whole module, so all we need is account
// for the freed memory size. We do that in FreeNativeModule.
// There's no separate Uncommit.
void FreeNativeModule(NativeModule*);
void Free(VirtualMemory* mem);
void AssignRanges(Address start, Address end, NativeModule*);
std::map<Address, std::pair<Address, NativeModule*>> lookup_map_;
// Count of NativeModules not yet collected. Helps determine if it's
// worth requesting a GC on memory pressure.
size_t active_ = 0;
std::atomic<size_t> remaining_uncommitted_code_space_;
// Histogram to update with the maximum used code space for each NativeModule.
Histogram* module_code_size_mb_ = nullptr;
DISALLOW_COPY_AND_ASSIGN(WasmCodeManager);
};
// Within the scope, the native_module is writable and not executable.
// At the scope's destruction, the native_module is executable and not writable.
// The states inside the scope and at the scope termination are irrespective of
// native_module's state when entering the scope.
// We currently mark the entire module's memory W^X:
// - for AOT, that's as efficient as it can be.
// - for Lazy, we don't have a heuristic for functions that may need patching,
// and even if we did, the resulting set of pages may be fragmented.
// Currently, we try and keep the number of syscalls low.
// - similar argument for debug time.
class NativeModuleModificationScope final {
public:
explicit NativeModuleModificationScope(NativeModule* native_module);
~NativeModuleModificationScope();
private:
NativeModule* native_module_;
};
} // namespace wasm
} // namespace internal
} // namespace v8
#endif // V8_WASM_WASM_CODE_MANAGER_H_