LO A D_G LOB A L ( su m )
BU I LD_ LIS T
LO A D_F AST (x )
GET_ITER
10: FOR_ITER ( to 3 1 )
ST O RE_ FAS T ( xi )
LO A D_F AST ( xi )
LO A D_F AST (t )
CO M PAR E_O P ( <)
LI S T_A PPE N D
JU M P_A BSO L UTE 10
31: C A LL_ FUN C TIO N
RE T URN _VA L UE
Figure 4. Python stack machine bytecode
Handling control flow
For straight-line code this process is fairly easy; most Python
instructions have a fairly straightforward effect on the stack.
But what happens when we encounter a branch? We need
to properly simulate both execution paths. To handle this
situation, we must make a copy of our virtual stack, and
evaluate both sides of the branch.
With branches come merge points; places where two or
more branches of execution come together. Each thread of
control flow might have assigned different register names to
each stack position. To handle this situation Falcon inserts
rename instructions before merge points, ensuring that all
incoming register stacks are compatible with each other.
(This is the same mechanism employed by compilers which
use static single assignment form (SSA)[11] to resolve φ-
nodes.)
Example conversion
Let’s walk through how this works for the example stack
code above (figure 4).
First we find the value of the function “sum” using the
LOAD_GLOBAL instruction. In the CPython interpreter, LOAD_GLOBAL
looks up a particular name in the dictionary of global values
and pushes that value onto the stack. Since the set of literal
names used in a function is known at compile time, the in-
struction can simply reference the index of the string “sum”
in a table of constant names. The equivalent register machine
instruction assigns the global value to a fresh register (in this
case r4). For brevity, the “stack” column in the listings below
will show just the register number for each instruction.
Python Falcon Stack
LOAD_GLOBAL 0 r4 = LOAD_GLOBAL 0 hi → h4i
The effect of this operation on the virtual stack is to
push the register r4 on top. When a later operation consumes
inputs off the stack, it will be correctly wired to use r4 as an
argument.
BUILD_LIST constructs an empty list to contain the results.
We create a new register r5 and push it onto the stack.
Python has special operations to load and store local
variables and to load constants. Rather then implement these
Python Falcon Stack
BUILD_LIST 0 r5 = BUILD_LIST 0 h4i → h5, 4i
instructions directly, we can alias these variables to specially
designated register names, which simplifies our code and
reduces the number of instructions needed.
Python Falcon Stack
LOAD_FAST 0 (x) h5, 4i → h1, 5, 4i
Register r1 is aliased to the local variable x. Therefore
for the LOAD_FAST operation here, we don’t need to generate a
Falcon instruction, and can instead simply push r1 onto our
virtual stack.
GET_ITER pops a sequence off of the stack and pushes back
an iterator for the sequence.
Python Falcon Stack
GET_ITER r6 = GET_ITER(r1) h1, 5, 4i → h6, 5, 4i
FOR_ITER is a branch instruction. It either pushes the next
element in the iterator onto the stack and falls-through to the
next instruction, or pops the iterator off the stack and jumps
to the other side of the loop.
Python Falcon Stack
FOR_ITER r7 = FOR_ITER(r6) h6, 5, 4i → h7, 6, 5, 4i
or h5, 4i
One branch of the FOR_ITER instruction takes us into inner
loop, which continues until the iterator is exhausted:
Python Falcon Stack
STORE_FAST (xi) r3 = r7 h7, 6, 5, 4i → h6, 5, 4i
LOAD_FAST (xi) h6, 5, 4i → h3, 6, 5, 4i
LOAD_FAST (t) h3, 6, 5, 4i → h2, 3, 6, 5, 4i
COMPARE_OP r8 = r3 > r2 h2, 3, 6, 5, 4i → h8, 6, 5, 4i
LIST_APPEND APPEND(r5, r8) h8, 6, 5, 4i → h6, 5, 4i
JUMP_ABSOLUTE JUMP_ABSOLUTE h6, 5, 4i
The behavior of the LIST_APPEND instruction here might
look somewhat surprising; it appears to “peek into” the stack
to find r5. This special behavior is unique to the LIST_APPEND
instruction, and likely is a result of past performance tuning
in the CPython interpreter (building lists is a very common
operation in Python).
And the other branch takes us to our function’s epilogue:
Python Falcon Stack
CALL_FUNCTION (sum) r9 = sum(r4) h5, 4i → h6i
RETURN_VALUE RETURN_VALUE(r9) h6i → hi
剩余11页未读，继续阅读
weixin_38592332
- 粉丝: 7
- 资源: 888
我的内容管理
展开
最新资源
- 十种常见电感线圈电感量计算公式详解
- 军用车辆：CAN总线的集成与优势
- CAN总线在汽车智能换档系统中的作用与实现
- CAN总线数据超载问题及解决策略
- 汽车车身系统CAN总线设计与应用
- SAP企业需求深度剖析：财务会计与供应链的关键流程与改进策略
- CAN总线在发动机电控系统中的通信设计实践
- Spring与iBATIS整合：快速开发与比较分析
- CAN总线驱动的整车管理系统硬件设计详解
- CAN总线通讯智能节点设计与实现
- DSP实现电动汽车CAN总线通讯技术
- CAN协议网关设计：自动位速率检测与互连
- Xcode免证书调试iPad程序开发指南
- 分布式数据库查询优化算法探讨
- Win7安装VC++6.0完全指南：解决兼容性与Office冲突
- MFC实现学生信息管理系统：登录与数据库操作
资源上传下载、课程学习等过程中有任何疑问或建议，欢迎提出宝贵意见哦~我们会及时处理！
点击此处反馈
