v8/test/cctest/wasm/test-jump-table-assembler.cc
Clemens Backes 89e0902d57 [codegen] Reduce kMinimalBufferSize and add kDefaultBufferSize
In Liftoff, we have a good estimate about how big the generated code
might get. Also, we often compile hundreds of functions which each hold
an assembler buffer alive until we finally add that code to the wasm
module.
In order to reduce memory consumption in Liftoff, this CL reduces
{AssemblerBase::kMinimalBufferSize} from 4096 to 128, and adds
{AssemblerBase::kDefaultBufferSize} to be used instead.

R=jkummerow@chromium.org

Change-Id: I7029bf501244770f4824a86b233d7f99c4b7910b
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/1914559
Reviewed-by: Jakob Kummerow <jkummerow@chromium.org>
Commit-Queue: Clemens Backes <clemensb@chromium.org>
Cr-Commit-Position: refs/heads/master@{#64958}
2019-11-14 15:05:31 +00:00

291 lines
11 KiB
C++

// Copyright 2018 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.
#include <bitset>
#include "src/codegen/assembler-inl.h"
#include "src/codegen/macro-assembler-inl.h"
#include "src/execution/simulator.h"
#include "src/utils/utils.h"
#include "src/wasm/jump-table-assembler.h"
#include "test/cctest/cctest.h"
#include "test/common/assembler-tester.h"
namespace v8 {
namespace internal {
namespace wasm {
#if 0
#define TRACE(...) PrintF(__VA_ARGS__)
#else
#define TRACE(...)
#endif
#define __ masm.
namespace {
static volatile int global_stop_bit = 0;
constexpr int kJumpTableSlotCount = 128;
constexpr uint32_t kJumpTableSize =
JumpTableAssembler::SizeForNumberOfSlots(kJumpTableSlotCount);
// Must be a safe commit page size.
constexpr size_t kThunkBufferSize = 4 * KB;
#if V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_X64
constexpr uint32_t kAvailableBufferSlots =
(kMaxWasmCodeSpaceSize - kJumpTableSize) / kThunkBufferSize;
constexpr uint32_t kBufferSlotStartOffset =
RoundUp<kThunkBufferSize>(kJumpTableSize);
#else
constexpr uint32_t kAvailableBufferSlots = 0;
#endif
Address AllocateJumpTableThunk(
Address jump_target, byte* thunk_slot_buffer,
std::bitset<kAvailableBufferSlots>* used_slots,
std::vector<std::unique_ptr<TestingAssemblerBuffer>>* thunk_buffers) {
#if V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_X64
// To guarantee that the branch range lies within the near-call range,
// generate the thunk in the same (kMaxWasmCodeSpaceSize-sized) buffer as the
// jump_target itself.
//
// Allocate a slot that we haven't already used. This is necessary because
// each test iteration expects to generate two unique addresses and we leave
// each slot executable (and not writable).
base::RandomNumberGenerator* rng =
CcTest::i_isolate()->random_number_generator();
// Ensure a chance of completion without too much thrashing.
DCHECK(used_slots->count() < (used_slots->size() / 2));
int buffer_index;
do {
buffer_index = rng->NextInt(kAvailableBufferSlots);
} while (used_slots->test(buffer_index));
used_slots->set(buffer_index);
return reinterpret_cast<Address>(thunk_slot_buffer +
buffer_index * kThunkBufferSize);
#else
USE(thunk_slot_buffer);
USE(used_slots);
thunk_buffers->emplace_back(
AllocateAssemblerBuffer(kThunkBufferSize, GetRandomMmapAddr()));
return reinterpret_cast<Address>(thunk_buffers->back()->start());
#endif
}
void CompileJumpTableThunk(Address thunk, Address jump_target) {
MacroAssembler masm(nullptr, AssemblerOptions{}, CodeObjectRequired::kNo,
ExternalAssemblerBuffer(reinterpret_cast<void*>(thunk),
kThunkBufferSize));
Label exit;
Register scratch = kReturnRegister0;
Address stop_bit_address = reinterpret_cast<Address>(&global_stop_bit);
#if V8_TARGET_ARCH_X64
__ Move(scratch, stop_bit_address, RelocInfo::NONE);
__ testl(MemOperand(scratch, 0), Immediate(1));
__ j(not_zero, &exit);
__ Jump(jump_target, RelocInfo::NONE);
#elif V8_TARGET_ARCH_IA32
__ Move(scratch, Immediate(stop_bit_address, RelocInfo::NONE));
__ test(MemOperand(scratch, 0), Immediate(1));
__ j(not_zero, &exit);
__ jmp(jump_target, RelocInfo::NONE);
#elif V8_TARGET_ARCH_ARM
__ mov(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ ldr(scratch, MemOperand(scratch, 0));
__ tst(scratch, Operand(1));
__ b(ne, &exit);
__ Jump(jump_target, RelocInfo::NONE);
#elif V8_TARGET_ARCH_ARM64
__ Mov(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ Ldr(scratch, MemOperand(scratch, 0));
__ Tbnz(scratch, 0, &exit);
__ Mov(scratch, Immediate(jump_target, RelocInfo::NONE));
__ Br(scratch);
#elif V8_TARGET_ARCH_PPC64
__ mov(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ LoadP(scratch, MemOperand(scratch));
__ cmpi(scratch, Operand::Zero());
__ bne(&exit);
__ mov(scratch, Operand(jump_target, RelocInfo::NONE));
__ Jump(scratch);
#elif V8_TARGET_ARCH_S390X
__ mov(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ LoadP(scratch, MemOperand(scratch));
__ CmpP(scratch, Operand(0));
__ bne(&exit);
__ mov(scratch, Operand(jump_target, RelocInfo::NONE));
__ Jump(scratch);
#elif V8_TARGET_ARCH_MIPS64
__ li(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ Lw(scratch, MemOperand(scratch, 0));
__ Branch(&exit, ne, scratch, Operand(zero_reg));
__ Jump(jump_target, RelocInfo::NONE);
#elif V8_TARGET_ARCH_MIPS
__ li(scratch, Operand(stop_bit_address, RelocInfo::NONE));
__ lw(scratch, MemOperand(scratch, 0));
__ Branch(&exit, ne, scratch, Operand(zero_reg));
__ Jump(jump_target, RelocInfo::NONE);
#else
#error Unsupported architecture
#endif
__ bind(&exit);
__ Ret();
FlushInstructionCache(thunk, kThunkBufferSize);
CHECK(SetPermissions(GetPlatformPageAllocator(), thunk, kThunkBufferSize,
v8::PageAllocator::kReadExecute));
}
class JumpTableRunner : public v8::base::Thread {
public:
JumpTableRunner(Address slot_address, int runner_id)
: Thread(Options("JumpTableRunner")),
slot_address_(slot_address),
runner_id_(runner_id) {}
void Run() override {
TRACE("Runner #%d is starting ...\n", runner_id_);
GeneratedCode<void>::FromAddress(CcTest::i_isolate(), slot_address_).Call();
TRACE("Runner #%d is stopping ...\n", runner_id_);
USE(runner_id_);
}
private:
Address slot_address_;
int runner_id_;
};
class JumpTablePatcher : public v8::base::Thread {
public:
JumpTablePatcher(Address slot_start, uint32_t slot_index, Address thunk1,
Address thunk2, base::Mutex* jump_table_mutex)
: Thread(Options("JumpTablePatcher")),
slot_start_(slot_start),
slot_index_(slot_index),
thunks_{thunk1, thunk2},
jump_table_mutex_(jump_table_mutex) {}
void Run() override {
TRACE("Patcher %p is starting ...\n", this);
Address slot_address =
slot_start_ + JumpTableAssembler::JumpSlotIndexToOffset(slot_index_);
// First, emit code to the two thunks.
for (Address thunk : thunks_) {
CompileJumpTableThunk(thunk, slot_address);
}
// Then, repeatedly patch the jump table to jump to one of the two thunks.
constexpr int kNumberOfPatchIterations = 64;
for (int i = 0; i < kNumberOfPatchIterations; ++i) {
TRACE(" patcher %p patch slot " V8PRIxPTR_FMT
" to thunk #%d (" V8PRIxPTR_FMT ")\n",
this, slot_address, i % 2, thunks_[i % 2]);
base::MutexGuard jump_table_guard(jump_table_mutex_);
JumpTableAssembler::PatchJumpTableSlot(
slot_start_ + JumpTableAssembler::JumpSlotIndexToOffset(slot_index_),
kNullAddress, thunks_[i % 2]);
}
TRACE("Patcher %p is stopping ...\n", this);
}
private:
Address slot_start_;
uint32_t slot_index_;
Address thunks_[2];
base::Mutex* jump_table_mutex_;
};
} // namespace
// This test is intended to stress concurrent patching of jump-table slots. It
// uses the following setup:
// 1) Picks a particular slot of the jump-table. Slots are iterated over to
// ensure multiple entries (at different offset alignments) are tested.
// 2) Starts multiple runners that spin through the above slot. The runners
// use thunk code that will jump to the same jump-table slot repeatedly
// until the {global_stop_bit} indicates a test-end condition.
// 3) Start a patcher that repeatedly patches the jump-table slot back and
// forth between two thunk. If there is a race then chances are high that
// one of the runners is currently executing the jump-table slot.
TEST(JumpTablePatchingStress) {
constexpr int kNumberOfRunnerThreads = 5;
constexpr int kNumberOfPatcherThreads = 3;
#if V8_TARGET_ARCH_ARM64 || V8_TARGET_ARCH_X64
// We need the branches (from CompileJumpTableThunk) to be within near-call
// range of the jump table slots. The address hint to AllocateAssemblerBuffer
// is not reliable enough to guarantee that we can always achieve this with
// separate allocations, so for Arm64 we generate all code in a single
// kMaxMasmCodeMemory-sized chunk.
STATIC_ASSERT(kMaxWasmCodeSpaceSize >= kJumpTableSize);
auto buffer = AllocateAssemblerBuffer(kMaxWasmCodeSpaceSize);
byte* thunk_slot_buffer = buffer->start() + kBufferSlotStartOffset;
#else
auto buffer = AllocateAssemblerBuffer(kJumpTableSize);
byte* thunk_slot_buffer = nullptr;
#endif
std::bitset<kAvailableBufferSlots> used_thunk_slots;
buffer->MakeWritableAndExecutable();
// Iterate through jump-table slots to hammer at different alignments within
// the jump-table, thereby increasing stress for variable-length ISAs.
Address slot_start = reinterpret_cast<Address>(buffer->start());
for (int slot = 0; slot < kJumpTableSlotCount; ++slot) {
TRACE("Hammering on jump table slot #%d ...\n", slot);
uint32_t slot_offset = JumpTableAssembler::JumpSlotIndexToOffset(slot);
std::vector<std::unique_ptr<TestingAssemblerBuffer>> thunk_buffers;
// Patch the jump table slot to jump to itself. This will later be patched
// by the patchers.
Address slot_addr =
slot_start + JumpTableAssembler::JumpSlotIndexToOffset(slot);
JumpTableAssembler::PatchJumpTableSlot(slot_addr, kNullAddress, slot_addr);
// For each patcher, generate two thunks where this patcher can emit code
// which finally jumps back to {slot} in the jump table.
std::vector<Address> patcher_thunks;
for (int i = 0; i < 2 * kNumberOfPatcherThreads; ++i) {
Address thunk =
AllocateJumpTableThunk(slot_start + slot_offset, thunk_slot_buffer,
&used_thunk_slots, &thunk_buffers);
ZapCode(thunk, kThunkBufferSize);
patcher_thunks.push_back(thunk);
TRACE(" generated jump thunk: " V8PRIxPTR_FMT "\n",
patcher_thunks.back());
}
// Start multiple runner threads that execute the jump table slot
// concurrently.
std::list<JumpTableRunner> runners;
for (int runner = 0; runner < kNumberOfRunnerThreads; ++runner) {
runners.emplace_back(slot_start + slot_offset, runner);
}
// Start multiple patcher thread that concurrently generate code and insert
// jumps to that into the jump table slot.
std::list<JumpTablePatcher> patchers;
// Only one patcher should modify the jump table at a time.
base::Mutex jump_table_mutex;
for (int i = 0; i < kNumberOfPatcherThreads; ++i) {
patchers.emplace_back(slot_start, slot, patcher_thunks[2 * i],
patcher_thunks[2 * i + 1], &jump_table_mutex);
}
global_stop_bit = 0; // Signal runners to keep going.
for (auto& runner : runners) CHECK(runner.Start());
for (auto& patcher : patchers) CHECK(patcher.Start());
for (auto& patcher : patchers) patcher.Join();
global_stop_bit = -1; // Signal runners to stop.
for (auto& runner : runners) runner.Join();
}
}
#undef __
#undef TRACE
} // namespace wasm
} // namespace internal
} // namespace v8