简单介绍一下Art里的jit和aot.
本文分成三个部分:
- JIT Introduction
- AOT Introduction
- Relation between JIT&AOT
JIT Introduction
JIT这里没有看太多的内容,先占个位:D
Profile文件
Art会在app执行过程中对它做profiling, 记录hot method等优化信息, profiling的结果会存在storage当中, 以便下次执行的时候供art参考.
profile的文件格式如下:
324 /**
325 * Serialization format:
326 * [profile_header, zipped[[profile_line_header1, profile_line_header2...],[profile_line_data1,
327 * profile_line_data2...]]]
328 * profile_header:
329 * magic,version,number_of_dex_files,uncompressed_size_of_zipped_data,compressed_data_size
330 * profile_line_header:
331 * dex_location,number_of_classes,methods_region_size,dex_location_checksum,num_method_ids
332 * profile_line_data:
333 * method_encoding_1,method_encoding_2...,class_id1,class_id2...,startup/post startup bitmap
334 * The method_encoding is:
335 * method_id,number_of_inline_caches,inline_cache1,inline_cache2...
336 * The inline_cache is:
337 * dex_pc,[M|dex_map_size], dex_profile_index,class_id1,class_id2...,dex_profile_index2,...
338 * dex_map_size is the number of dex_indeces that follows.
339 * Classes are grouped per their dex files and the line
340 * `dex_profile_index,class_id1,class_id2...,dex_profile_index2,...` encodes the
341 * mapping from `dex_profile_index` to the set of classes `class_id1,class_id2...`
342 * M stands for megamorphic or missing types and it's encoded as either
343 * the byte kIsMegamorphicEncoding or kIsMissingTypesEncoding.
344 * When present, there will be no class ids following.
345 **/
在art目录下面, google还提供了profman来作为profile相关操作的接口.
使用profman可以轻松地将profile打印成人类友好模式.
Example:
1. adb shell profman --generate-test-profile=/data/testprofile --generate-test-profile-seed=0 --apk=/system/framework/am.jar --dex-location=/system/framework/am.jar
2. adb shell profman --dump-only --profile-file=/data/testprofile
以下内容来自am.jar的profile, just an example:
=== Dex files ===
=== profile ===
ProfileInfo:
am.jar [index=0] [checksum=ccdf4a91]
hot methods: 40[], 50[], 64[], 89[], 91[], 99[], 135[], 144[],
startup methods: 40, 50, 64, 144,
post startup methods: 89, 91, 99, 135,
classes:
根据profile文件的优化信息, art会在下面这个函数里面去判断是否要将method编译成native code.
986 bool CompilerDriver::ShouldCompileBasedOnProfile(const MethodReference& method_ref) const {
987 // Profile compilation info may be null if no profile is passed.
988 if (!CompilerFilter::DependsOnProfile(compiler_options_->GetCompilerFilter())) {
989 // Use the compiler filter instead of the presence of profile_compilation_info_ since
990 // we may want to have full speed compilation along with profile based layout optimizations.
991 return true;
992 }
993 // If we are using a profile filter but do not have a profile compilation info, compile nothing.
994 if (profile_compilation_info_ == nullptr) {
995 return false;
996 }
997 // Compile only hot methods, it is the profile saver's job to decide what startup methods to mark
998 // as hot.
999 bool result = profile_compilation_info_->GetMethodHotness(method_ref).IsHot();
1000
1001 if (kDebugProfileGuidedCompilation) {
1002 LOG(INFO) << "[ProfileGuidedCompilation] "
1003 << (result ? "Compiled" : "Skipped") << " method:" << method_ref.PrettyMethod(true);
1004 }
1005 return result;
1006}
AOT Introduction
AOT的编译策略
dex2oat有个选项叫--compiler-filter, 它的参数如下.
32 enum Filter {
33 kAssumeVerified, // Skip verification but mark all classes as verified anyway.
34 kExtract, // Delay verication to runtime, do not compile anything.
35 kVerify, // Only verify classes.
36 kQuicken, // Verify, quicken, and compile JNI stubs.
37 kSpaceProfile, // Maximize space savings based on profile.
38 kSpace, // Maximize space savings.
39 kSpeedProfile, // Maximize runtime performance based on profile.
40 kSpeed, // Maximize runtime performance.
41 kEverythingProfile, // Compile everything capable of being compiled based on profile.
42 kEverything, // Compile everything capable of being compiled.
// 注意kEverything这个参数
43 };
相关逻辑在这里
986 bool CompilerDriver::ShouldCompileBasedOnProfile(const MethodReference& method_ref) const {
987 // Profile compilation info may be null if no profile is passed.
// 当选择了everthing之后, DependsOnProfile返回false,所以ShouldCompileBasedOnProfile返回true.
988 if (!CompilerFilter::DependsOnProfile(compiler_options_->GetCompilerFilter())) {
989 // Use the compiler filter instead of the presence of profile_compilation_info_ since
990 // we may want to have full speed compilation along with profile based layout optimizations.
991 return true;
992 }
993 // If we are using a profile filter but do not have a profile compilation info, compile nothing.
994 if (profile_compilation_info_ == nullptr) {
995 return false;
996 }
997 // Compile only hot methods, it is the profile saver's job to decide what startup methods to mark
998 // as hot.
// 假设选了kSpeedProfile, 会走到这里, 然后会去查询profile, 如果是hot的话, 才会返回true
999 bool result = profile_compilation_info_->GetMethodHotness(method_ref).IsHot();
1000
1001 if (kDebugProfileGuidedCompilation) {
1002 LOG(INFO) << "[ProfileGuidedCompilation] "
1003 << (result ? "Compiled" : "Skipped") << " method:" << method_ref.PrettyMethod(true);
1004 }
1005 return result;
1006 }
如果想要在安装apk的时候,将app完全编译成native, 可以这么做:
//在install apk的时候就会默认全部aot
setprop pm.dexopt.install everything
或者重编译指定的apk:
// everything是编译所有能编译的code
pm compile -m everything -f com.kiloo.subwaysurf
// speed-profile会根据profile文件来编译
pm compile -m speed-profile -f com.kiloo.subwaysurf
我尝试编译了一个apk. 下面是编译后的文件大小.
speed-profile (default mode) | everything | |
---|---|---|
base.art | 32K | 0 |
base.odex | 224K | 21M |
base.vdex | 8M | 7.4M |
base.odex是一个elf, aot之后的native code就存在里面.
因为是首次安装, profile信息几乎没有, 所以speed-profile模式下的base.odex很小.
PS: profile的信息都保存在/data/misc/profiles, 查看profile的内容可用以下命令
profman --profile-file=./primary.prof --dump-only
everything模式的缺点就是(1) install的时候慢一些 (2) 占用空间大一些
dex2oat的参数解析
--instruction-set-variant= (default | atom | sandybridge | silvermont)
--instruction-set-features = (default | sse3 | sse4.1 | avx…)
这两个参数共同决定了编译时候的cpu feature.
instruction-set-variant和instruction-set-features的值是由system property来的:
- dalvik.vm.isa.x86.variant/ dalvik.vm.isa.x86_64.variant
- dalvik.vm.isa.x86.features/ dalvik.vm.isa.x86_64.features
--huge-method-max/--large-method-max/--small-method-max/--tiny-method-max
定义包含多少条指令的method是huge/large/small/tiny method. 编译器对于它们有不同的处理方法.
- compiler filter!= everything && huge, 会跳过编译步骤.
- compiler filter!= everything && large && no branch, 会跳过编译步骤.
76 bool HGraphBuilder::SkipCompilation(size_t number_of_branches) {
- small/tiny method在P的code里没有任何特殊处理.
--inline-max-code-units
定义一个函数最多包含多少code unit, 可以被优化成inline函数.
这个值等于0的话, 就不进行inline优化.
--compiler-backend=(Quick|Optimizing)
这个选项其实没什么用, 就算选quick也会跳转到optimizing.
31 case kQuick:
32 // TODO: Remove Quick in options.
33 case kOptimizing:
34 return CreateOptimizingCompiler(driver);
.art/.odex/.vdex的具体含义
.vdex: verified dex,其中包含 APK 的 DEX 代码,另外还有一些旨在加快验证速度的元数据.
验证主要指的是, 当系统OTA等原因, 会需要重新验证dex的合法性.
而vdex存放的是已经验证过的dex. 如果vdex存在, 就可以跳过验证这个步骤.
.odex: elf, native code存放的位置
.art: .art就是art虚拟机内部数据的一个snapshot.
因为有一些数据是通用的, 是每次vm起来都会需要的, 所以可以将它们提前存储起来.
而vm起来的时候通过mmap .art文件来达到加速目的.
Relation between JIT&AOT
在jit运行的时候, 它会收集hot method/class的数据, 然后把profile持久化到disk上面. 在android空闲的时候, BackgroundDexOptService会根据这个profile信息, 把apk重新aot.
jit-arch.png jit-profile-comp.png
BackgroundDexOptService的调用大致是:
runIdleOptimization(BackgroundDexOptService.java)
-> performDexOptWithStatus(PackageManagerService.java)
-> dexopt(InstalldNativeService.cpp)
-> run_dex2oat(dexopt.cpp) //这里调用了dex2oat这个binary去做aot
BackgroundDexOptService会在特定条件下对apk进行aot.
它被调用的条件需要同时满足以下三条:
- setRequiresDeviceIdle(true) // 设备处于idle状态, idle的定义是: 息屏71mins or 睡眠状态71mins or无线充电71mins
- setRequiresCharging(true) // 设备处于充电状态
- setPeriodic(IDLE_OPTIMIZATION_PERIOD) // 一天仅触发一次
BackgroundDexOptService的触发条件:
123 // Schedule a daily job which scans installed packages and compiles
124 // those with fresh profiling data.
125 js.schedule(new JobInfo.Builder(JOB_IDLE_OPTIMIZE, sDexoptServiceName)
126 .setRequiresDeviceIdle(true)
127 .setRequiresCharging(true)
128 .setPeriodic(IDLE_OPTIMIZATION_PERIOD)
129 .build());
关于Idle的定义:
44 // Policy: we decide that we're "idle" if the device has been unused /
45 // screen off or dreaming or wireless charging dock idle for at least this long
46 private long mInactivityIdleThreshold; // 这个值等于71mins
网友评论