”ART’s Quick Compiler:an unofficial overview“
Matteo Franchin LCA15, February 2015
看了這個關於ART的PPT，記錄一下。
ART編譯器有兩種一種叫做Portable（後端好像是llvm），還一種就是常用quick，下文介紹基於quick。
ART不僅僅編譯執行，也會解釋執行（跟dalvik基本上一樣把）
Java method is the compilation unit for the Quick Compiler.
Quick編譯器的編譯單元是一個java方法
dalvikvm64 calls the compiler, dex2oat, to generate OAT file in /data/dalvik-cache/arm64.
64位虛擬機調用編譯 dex2oat方法 來生成一個oat文件，保存在/data/dalvik-cache/arm64目錄下
dex2oat uses several threads to compile all the methods.
DEX bytecode:
register-based variable length instructions (≥ 2 bytes).
dex指令是基於寄存器的，可變長度的（>2字節，16位或32位）
32-bit, untyped virtual registers (vregs): v0-v65535.
dex文件使用的是32位的虛擬寄存器V0-V65535 與機器碼使用的Rx物理寄存器不同
64-bit values stored in consecutive pairs of registers.
64位的值存儲於2個相鄰的32位虛擬寄存器中
See https://source.android.com/devices/tech/dalvik/dalvik-bytecode.html
MIR: Mid-level Intermediate Representation:
Methods as control-flow graphs made by sequences of MIR nodes
方法變成了由MIR序列組成的控制流程圖？
(organised in basic blocks).組織形式是BB基本塊，基本塊==方法？
Faithful to DEX bytecode: 1 DEX op → 1 MIR node.
dex操作與MIR一一對應
+ pseudo MIR nodes for annotation purposes.
僞MIR節點爲了註釋的目的？？？
+ extended MIR nodes for optimisation purposes.
擴展的MIR節點爲了優化的目的
MIR指令不過是把DEX指令給解釋一下 拆分一下 包裝一下，爲了LIR服務
LIR: Low-level Intermediate representation:
Methods as sequences of LIR nodes.
Faithful to instruction set: 1 instruction → 1 LIR node.
+ pseudo LIR nodes for annotation purposes.
Different LIR enums for ARM, ARM64, x86, MIPS.
E.g. NewLIR2(kA64Mov2rr, rw0, rw1) for mov w0, w1.
LIR指令跟機器指令似乎就比較接近了
OAT file:
Cached in /data/dalvik-cache/arm64.
Contain binary executable code.
Contain source DEX bytecode.
包括了完整DEX代碼，因爲可能還會需要到？有些方法要解釋執行。
Contain metadata for runtime (GC, deoptimization).
From DEX bytecode to MIR nodes:
DEX ops parsed and added as MIR nodes to blocks.
Blocks split at branches/labels: become basic blocks.
Middle-End passes: SSA, type inference, redundant checks elimination.
Pass engine to manage passes.
From MIR to LIR:
Physical regs are “assigned” at this stage.
virtual regs promoted to callee-saved physical regs.
虛擬寄存器變成了物理寄存器，Vx變成了Rx
MIR translated to LIR (big switch statement): caller-saved regs allocated for temps.
LIR optimisation passes:Load hoisting.Load-store elimination.
From LIR to native:
Fixups (branches resolved, instructions replaced if args out-of-range).
LIR translated to binary code.
Global (inter-method) fixups.
The good:
ART is typically 80% faster than Dalvik (Google I/O 2014 - The ART
Runtime).
Quick compiler is quick indeed (on target device must be able to produce
hundreds of MBs in minutes).
The bad:
LIR, MIR are sequences of nodes, rather than graphs:
Difficult to detect dead code and perform code motion.
Difficult to detect instructions sequences (and replace them with better ones).
MIR and LIR are quite different representations: difficult to share
optimisation passes among the two.
suspend-points
What are suspend-points?
1：Checks inserted in generated code to ensure Java threads stop executing
spontaneously when requested to do so.
2：Points of consistency between native-execution, runtime and
interpreted-execution.
好像就是主要用於GC和中斷/debug分析
機器碼實現：
