pypy's approach to construct domain-specific language runtime
TRANSCRIPT
Tag: virtual machine, compiler, performance
PyPy’s Approach to Construct Domain-specific Language Runtime
Tag: virtual machine, compiler, performance
Construct Domain-specific Language Runtimeusing
Speed
7.4 times faster than CPythonhttp://speed.pypy.org
antocuni (PyCon Otto) PyPy Status Update April 07 2017 4 / 19
Why is Python slow?
Interpretation overhead
Boxed arithmetic and automatic overflow handling
Dynamic dispatch of operations
Dynamic lookup of methods and attributes
Everything can change on runtime
Extreme introspective and reflective capabilities
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 8 / 51
Why is Python slow?Boxed arithmetic and automatic overflow handling
i = 0
while i < 10000000:
i = i +1
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 9 / 51
Why is Python slow?Dynamic dispatch of operations
# while i < 1000000
9 LOAD_FAST 0 (i)
12 LOAD_CONST 2 (10000000)
15 COMPARE_OP 0 (<)
18 POP_JUMP_IF_FALSE 34
# i = i + 1
21 LOAD_FAST 0 (i)
24 LOAD_CONST 3 (1)
27 BINARY_ADD
28 STORE_FAST 0 (i)
31 JUMP_ABSOLUTE 9
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 10 / 51
Why is Python slow?Dynamic lookup of methods and attributes
class MyExample(object ):
pass
def foo(target , flag):
if flag:
target.x = 42
obj = MyExample ()
foo(obj , True)
print obj.x #=> 42
print getattr(obj , "x") #=> 42
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 11 / 51
Why is Python slow?Everything can change on runtime
def fn():
return 42
def hello ():
return ’Hi! PyConEs!’
def change_the_world ():
global fn
fn = hello
print fn() #=> 42
change_the_world ()
print fn() => ’Hi! PyConEs!’
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 12 / 51
Why is Python slow?Everything can change on runtime
class Dog(object ):
def __init__(self):
self.name = ’Jandemor ’
def talk(self):
print "%s: guau!" % self.name
class Cat(object ):
def __init__(self):
self.name = ’CatInstance ’
def talk(self):
print "%s: miau!" % self.name
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 13 / 51
Why is Python slow?Everything can change on runtime
my_pet = Dog()
my_pet.talk() #=> ’Jandemor: guau!’
my_pet.__class__ = Cat
my_pet.talk() #=> ’Jandemor: miau!’
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 14 / 51
Why is Python slow?Extreme introspective and reflective capabilities
def fill_list(name):
frame = sys._getframe (). f_back
lst = frame.f_locals[name]
lst.append (42)
def foo():
things = []
fill_list(’things ’)
print things #=> 42
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 15 / 51
Why is Python slow?Everything can change on runtime
def fn():
return 42
def hello ():
return ’Hi! PyConEs!’
def change_the_world ():
global fn
fn = hello
print fn() #=> 42
change_the_world ()
print fn() => ’Hi! PyConEs!’
Francisco Fernandez Castano (@fcofdezc) PyPy November 8, 2014 12 / 51
PyPy Translation Toolchain
• Capable of compiling (R)Python!
• Garbage collection!
• Tracing just-in-time compiler generator!
• Software transactional memory?
PyPy Architecture
PyPy based interpreters• Topaz (Ruby)!
• HippyVM (PHP)!
• Pyrolog (Prolog)!
• pycket (Racket)!
• Various other interpreters for (Scheme, Javascript, io, Gameboy)
Compiler / Interpreter
Source: Compiler Construction, Prof. O. NierstraszSource: Compiler Construction, Prof. O. Nierstrasz
• intermediate representation (IR) • front end maps legal code into IR • back end maps IR onto target machine • simplify retargeting • allows multiple front ends • multiple passes better code →
Traditional 2 pass compiler
• analyzes and changes IR • goal is to reduce runtime • must preserve values
Traditional 3 pass compiler
• constant propagation and folding• code motion• reduction of operator strength • common sub-expression elimination• redundant store elimination • dead code elimination
Optimizer: middle end
Modern optimizers are usually built as a set of passes
• Preserve language semantics• Reflection, Introspection, Eval
• External APIs
• Interpreter consists of short sequences of code• Prevent global optimizations
• Typically implemented as a stack machine
• Dynamic, imprecise type information• Variables can change type
• Duck Typing: method works with any object that provides accessed interfaces
• Monkey Patching: add members to “class” after initialization
• Memory management and concurrency
• Function calls through packing of operands in fat object
Optimization Challenges
PyPy Functional Architecture
RPython• Python subset!
• Statically typed!
• Garbage collected!
• Standard library almost entirely unavailable!
• Some missing builtins (print, open(), …)!
• rpython.rlib!
• exceptions are (sometimes) ignored!
• Not a really a language, rather a "state"
22
PyPy Interpreter
def f(x):
return x + 1
>>> dis.dis(f)
2 0 LOAD_FAST 0 (x)
3 LOAD_CONST 1 (1)
6 BINARY_ADD
7 RETURN_VALUE
• written in Rpython• Stack-based bytecode interpreter (like JVM)
• bytecode compiler generates bytecode→
• bytecode evaluator interprets bytecode →
• object space handles operations on objects→
23
PyPy Bytecode Interpreter
31
CFG (Call Flow Graph)
• Consists of Blocks and Links
• Starting from entry_point
• “Single Static Information” form
def f(n):
return 3 * n + 2
Block(v1): # input argument
v2 = mul(Constant(3), v1)
v3 = add(v2, Constant(2))
33
CFG: Static Single Information
33
def test(a):if a > 0:
if a > 5:return 10
return 4if a < - 10:
return 3return 10
• SSI: “PHIs” for all used variables• Blocks as “functions without branches”
• High Level Language Implementation• to implement new features: lazily computed objects
and functions, plug-able garbage-collection, runtime replacement of live-objects, stackless concurrency
• JIT Generation• Object space• Stackless• infinite Recursion
• Microthreads: Coroutines, Tasklets and Channels, Greenlets
PyPy Advantages
PERCEPTION
http://abstrusegoose.com/secretarchives/under-the-hood - CC BY-NC 3.0 US
Assumptions
Pareto Principle (80-20 rule)I the 20% of the program accounts for the 80% of the
runtimeI hot-spots
Fast Path principleI optimize only what is necessaryI fall back for uncommon cases
Most of runtime spent in loopsAlways the same code paths (likely)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 9 / 32
Tracing JIT phases
Interpretation
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
Running
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
Running
cold guard failed
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
Running
cold guard failed
entering compiled loop
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
Running
cold guard failed
entering compiled loop
guard failure → hot
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Tracing JIT phases
Interpretation
Tracinghot loop detected
Compilation
Running
cold guard failed
entering compiled loop
guard failure → hot
hot guard failed
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 11 / 32
Trace trees (1)
tracetree.pydef foo():a = 0i = 0N = 100while i < N:
if i%2 == 0:a += 1
else:a *= 2;
i += 1return a
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 12 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
BLACKHOLE
COLD FAIL
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
BLACKHOLE
COLD FAIL
INTERPRETER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
BLACKHOLE
COLD FAIL
INTERPRETER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
a1 = int_mul(a0, 2)i1 = int_add(i0, 1)jump(start, i1, a1)
HOT FAIL
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
a1 = int_mul(a0, 2)i1 = int_add(i0, 1)jump(start, i1, a1)
HOT FAIL
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Trace trees (2)
label(start, i0, a0)v0 = int_lt(i0, 2000)guard_true(v0)v1 = int_mod(i0, 2)v2 = int_eq(v1, 0)guard_true(v1)a1 = int_add(a0, 10)i1 = int_add(i0, 1)jump(start, i1, a1)
a1 = int_mul(a0, 2)i1 = int_add(i0, 1)jump(start, i1, a1)
HOT FAIL
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 13 / 32
Part 3
The PyPy JIT
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 14 / 32
Terminology (1)
translation time: when you run "rpythontargetpypy.py" to get the pypy binaryruntime: everything which happens after you startpypy
interpretation, tracing, compilingassembler/machine code: the output of the JITcompilerexecution time: when your Python program is beingexecuted
I by the interpreterI by the machine code
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 15 / 32
Terminology (2)
interp-level: things written in RPython[PyPy] interpreter: the RPython program whichexecutes the final Python programsbytecode: "the output of dis.dis". It is executed by thePyPy interpreter.app-level: things written in Python, and executed bythe PyPy Interpreter
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 16 / 32
Terminology (3)
(the following is not 100% accurate but it’s enough tounderstand the general principle)low level op or ResOperation
I low-level instructions like "add two integers", "read a fieldout of a struct", "call this function"
I (more or less) the same level of C ("portable assembler")I knows about GC objects (e.g. you have getfield_gc
vs getfield_raw)
jitcodes: low-level representation of RPythonfunctions
I sequence of low level opsI generated at translation timeI 1 RPython function --> 1 C function --> 1 jitcode
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 17 / 32
Terminology (4)
JIT traces or loopsI a very specific sequence of llops as actually executed by
your Python programI generated at runtime (more specifically, during tracing)
JIT optimizer: takes JIT traces and emits JIT tracesJIT backend: takes JIT traces and emits machinecode
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 18 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
compile-time
runtime
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
compile-time
runtime
META-TRACER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
compile-time
runtime
META-TRACEROPTIMIZER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
compile-time
runtime
META-TRACEROPTIMIZERBACKEND
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
General architecture
def LOAD_GLOBAL(self): ...
def STORE_FAST(self): ...
def BINARY_ADD(self): ...
RPYTHON
CODEWRITER
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)....
...p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
...promote_class(p0)i0 = getfield_gc(p0, 'intval')promote_class(p1)i1 = getfield_gc(p1, 'intval')i2 = int_add(i0, i1)if (overflowed) goto ...p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')....
JITCODE
compile-time
runtime
META-TRACEROPTIMIZERBACKENDASSEMBLER
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 19 / 32
PyPy trace example
def fn(): c = a+b ...
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy trace example
def fn(): c = a+b ...
LOAD_GLOBAL ALOAD_GLOBAL BBINARY_ADDSTORE_FAST C
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy trace example
def fn(): c = a+b ...
LOAD_GLOBAL ALOAD_GLOBAL BBINARY_ADDSTORE_FAST C
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)...
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy trace example
def fn(): c = a+b ...
LOAD_GLOBAL ALOAD_GLOBAL BBINARY_ADDSTORE_FAST C
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)......p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)...
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy trace example
def fn(): c = a+b ...
LOAD_GLOBAL ALOAD_GLOBAL BBINARY_ADDSTORE_FAST C
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)......p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)...
...guard_class(p0, W_IntObject)i0 = getfield_gc(p0, 'intval')guard_class(p1, W_IntObject)i1 = getfield_gc(p1, 'intval')i2 = int_add(00, i1)guard_not_overflow()p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')...
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy trace example
def fn(): c = a+b ...
LOAD_GLOBAL ALOAD_GLOBAL BBINARY_ADDSTORE_FAST C
...p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)......p0 = getfield_gc(p0, 'func_globals')p2 = getfield_gc(p1, 'strval')call(dict_lookup, p0, p2)...
...guard_class(p0, W_IntObject)i0 = getfield_gc(p0, 'intval')guard_class(p1, W_IntObject)i1 = getfield_gc(p1, 'intval')i2 = int_add(00, i1)guard_not_overflow()p2 = new_with_vtable('W_IntObject')setfield_gc(p2, i2, 'intval')......p0 = getfield_gc(p0, 'locals_w')setarrayitem_gc(p0, i0, p1)....
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 20 / 32
PyPy optimizer
intboundsconstant folding / pure operationsvirtualsstring optimizationsheap (multiple get/setfield, etc)unroll
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 21 / 32
Intbound optimization (1)
intbound.py
def fn():i = 0while i < 5000:
i += 2return i
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 22 / 32
Intbound optimization (2)
unoptimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add_ovf(i15, 2)guard_no_overflow()...
optimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add(i15, 2)...
It works oftenarray bound checkingintbound info propagates all over the trace
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 23 / 32
Intbound optimization (2)
unoptimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add_ovf(i15, 2)guard_no_overflow()...
optimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add(i15, 2)...
It works oftenarray bound checkingintbound info propagates all over the trace
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 23 / 32
Intbound optimization (2)
unoptimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add_ovf(i15, 2)guard_no_overflow()...
optimized...i17 = int_lt(i15, 5000)guard_true(i17)i19 = int_add(i15, 2)...
It works oftenarray bound checkingintbound info propagates all over the trace
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 23 / 32
Virtuals (1)
virtuals.py
def fn():i = 0while i < 5000:
i += 2return i
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 24 / 32
Virtuals (2)
unoptimized...guard_class(p0, W_IntObject)i1 = getfield_pure(p0, ’intval’)i2 = int_add(i1, 2)p3 = new(W_IntObject)setfield_gc(p3, i2, ’intval’)...
optimized...i2 = int_add(i1, 2)...
The most important optimization (TM)It works both inside the trace and across the loopIt works for tons of cases
I e.g. function frames
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 25 / 32
Virtuals (2)
unoptimized...guard_class(p0, W_IntObject)i1 = getfield_pure(p0, ’intval’)i2 = int_add(i1, 2)p3 = new(W_IntObject)setfield_gc(p3, i2, ’intval’)...
optimized...i2 = int_add(i1, 2)...
The most important optimization (TM)It works both inside the trace and across the loopIt works for tons of cases
I e.g. function frames
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 25 / 32
Virtuals (2)
unoptimized...guard_class(p0, W_IntObject)i1 = getfield_pure(p0, ’intval’)i2 = int_add(i1, 2)p3 = new(W_IntObject)setfield_gc(p3, i2, ’intval’)...
optimized...i2 = int_add(i1, 2)...
The most important optimization (TM)It works both inside the trace and across the loopIt works for tons of cases
I e.g. function frames
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 25 / 32
Constant folding (1)
constfold.py
def fn():i = 0while i < 5000:
i += 2return i
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 26 / 32
Constant folding (2)
unoptimized...i1 = getfield_pure(p0, ’intval’)i2 = getfield_pure(<W_Int(2)>,
’intval’)i3 = int_add(i1, i2)...
optimized...i1 = getfield_pure(p0, ’intval’)i3 = int_add(i1, 2)...
It "finishes the job"Works well together with other optimizations (e.g.virtuals)It also does "normal, boring, static" constant-folding
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 27 / 32
Constant folding (2)
unoptimized...i1 = getfield_pure(p0, ’intval’)i2 = getfield_pure(<W_Int(2)>,
’intval’)i3 = int_add(i1, i2)...
optimized...i1 = getfield_pure(p0, ’intval’)i3 = int_add(i1, 2)...
It "finishes the job"Works well together with other optimizations (e.g.virtuals)It also does "normal, boring, static" constant-folding
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 27 / 32
Constant folding (2)
unoptimized...i1 = getfield_pure(p0, ’intval’)i2 = getfield_pure(<W_Int(2)>,
’intval’)i3 = int_add(i1, i2)...
optimized...i1 = getfield_pure(p0, ’intval’)i3 = int_add(i1, 2)...
It "finishes the job"Works well together with other optimizations (e.g.virtuals)It also does "normal, boring, static" constant-folding
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 27 / 32
Out of line guards (1)
outoflineguards.py
N = 2def fn():
i = 0while i < 5000:
i += Nreturn i
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 28 / 32
Out of line guards (2)
unoptimized...quasiimmut_field(<Cell>, ’val’)guard_not_invalidated()p0 = getfield_gc(<Cell>, ’val’)...i2 = getfield_pure(p0, ’intval’)i3 = int_add(i1, i2)
optimized...guard_not_invalidated()...i3 = int_add(i1, 2)...
Python is too dynamic, but we don’t care :-)No overhead in assembler codeUsed a bit "everywhere"
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 29 / 32
Out of line guards (2)
unoptimized...quasiimmut_field(<Cell>, ’val’)guard_not_invalidated()p0 = getfield_gc(<Cell>, ’val’)...i2 = getfield_pure(p0, ’intval’)i3 = int_add(i1, i2)
optimized...guard_not_invalidated()...i3 = int_add(i1, 2)...
Python is too dynamic, but we don’t care :-)No overhead in assembler codeUsed a bit "everywhere"
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 29 / 32
Out of line guards (2)
unoptimized...quasiimmut_field(<Cell>, ’val’)guard_not_invalidated()p0 = getfield_gc(<Cell>, ’val’)...i2 = getfield_pure(p0, ’intval’)i3 = int_add(i1, i2)
optimized...guard_not_invalidated()...i3 = int_add(i1, 2)...
Python is too dynamic, but we don’t care :-)No overhead in assembler codeUsed a bit "everywhere"
antocuni (Intel@Bucharest) PyPy Intro April 4 2016 29 / 32
Hello RPython# hello_rpython.pyimport os!
def entry_point(argv): os.write(2, “Hello, World!\n”) return 0!
def target(driver, argv): return entry_point, None
$ rpython hello_rpython.py…$ ./hello_python-cHello, RPython!
Goal
• BASIC interpreter capable of running Hamurabi!
• Bytecode based!
• Garbage Collection!
• Just-In-Time Compilation
Live play session
Architecture
Parser
Compiler
Virtual Machine
AST
Bytecode
Source
10 PRINT TAB(32);"HAMURABI"20 PRINT TAB(15);"CREATIVE COMPUTING MORRISTOWN, NEW JERSEY"30 PRINT:PRINT:PRINT80 PRINT "TRY YOUR HAND AT GOVERNING ANCIENT SUMERIA"90 PRINT "FOR A TEN-YEAR TERM OF OFFICE.":PRINT95 D1=0: P1=0100 Z=0: P=95:S=2800: H=3000: E=H-S110 Y=3: A=H/Y: I=5: Q=1210 D=0215 PRINT:PRINT:PRINT "HAMURABI: I BEG TO REPORT TO YOU,": Z=Z+1217 PRINT "IN YEAR";Z;",";D;"PEOPLE STARVED,";I;"CAME TO THE CITY,"218 P=P+I227 IF Q>0 THEN 230228 P=INT(P/2)229 PRINT "A HORRIBLE PLAGUE STRUCK! HALF THE PEOPLE DIED."230 PRINT "POPULATION IS NOW";P232 PRINT "THE CITY NOW OWNS ";A;"ACRES."235 PRINT "YOU HARVESTED";Y;"BUSHELS PER ACRE."250 PRINT "THE RATS ATE";E;"BUSHELS."260 PRINT "YOU NOW HAVE ";S;"BUSHELS IN STORE.": PRINT270 REM *** MORE CODE THAT DID NOT FIT INTO THE SLIDE FOLLOWS
Parser
Parser
Abstract Syntax Tree (AST)
Source
Parser
Parser
AST
SourceLexer
Tokens
Source
Parser
AST
RPLY
• Based on PLY, which is based on Lex and Yacc!
• Lexer generator!
• LALR parser generator
Lexerfrom rply import LexerGenerator!
lg = LexerGenerator()!
lg.add(“NUMBER”, “[0-9]+”)# …lg.ignore(“ +”) # whitespace!
lexer = lg.build().lex
lg.add('NUMBER', r'[0-9]*\.[0-9]+')lg.add('PRINT', r'PRINT')lg.add('IF', r'IF')lg.add('THEN', r'THEN')lg.add('GOSUB', r'GOSUB')lg.add('GOTO', r'GOTO')lg.add('INPUT', r'INPUT')lg.add('REM', r'REM')lg.add('RETURN', r'RETURN')lg.add('END', r'END')lg.add('FOR', r'FOR')lg.add('TO', r'TO')lg.add('NEXT', r'NEXT')lg.add('NAME', r'[A-Z][A-Z0-9$]*')lg.add('(', r'\(')lg.add(')', r'\)')lg.add(';', r';')lg.add('STRING', r'"[^"]*"')
lg.add(':', r'\r?\n')lg.add(':', r':')lg.add('=', r'=')lg.add('<>', r'<>')lg.add('-', r'-')lg.add('/', r'/')lg.add('+', r'\+')lg.add('>=', r'>=')lg.add('>', r'>')lg.add('***', r'\*\*\*.*')lg.add('*', r'\*')lg.add('<=', r'<=')lg.add('<', r'<')
>>> from basic.lexer import lex>>> source = open("hello.bas").read()>>> for token in lex(source):... print tokenToken("NUMBER", "10")Token("PRINT", "PRINT")Token("STRING",'"HELLO BASIC!"')Token(":", "\n")
Grammar
• A set of formal rules that defines the syntax!
• terminals = tokens!
• nonterminals = rules defining a sequence of one or more (non)terminals
10 PRINT TAB(32);"HAMURABI"20 PRINT TAB(15);"CREATIVE COMPUTING MORRISTOWN, NEW JERSEY"30 PRINT:PRINT:PRINT80 PRINT "TRY YOUR HAND AT GOVERNING ANCIENT SUMERIA"90 PRINT "FOR A TEN-YEAR TERM OF OFFICE.":PRINT95 D1=0: P1=0100 Z=0: P=95:S=2800: H=3000: E=H-S110 Y=3: A=H/Y: I=5: Q=1210 D=0215 PRINT:PRINT:PRINT "HAMURABI: I BEG TO REPORT TO YOU,": Z=Z+1217 PRINT "IN YEAR";Z;",";D;"PEOPLE STARVED,";I;"CAME TO THE CITY,"218 P=P+I227 IF Q>0 THEN 230228 P=INT(P/2)229 PRINT "A HORRIBLE PLAGUE STRUCK! HALF THE PEOPLE DIED."230 PRINT "POPULATION IS NOW";P232 PRINT "THE CITY NOW OWNS ";A;"ACRES."235 PRINT "YOU HARVESTED";Y;"BUSHELS PER ACRE."250 PRINT "THE RATS ATE";E;"BUSHELS."260 PRINT "YOU NOW HAVE ";S;"BUSHELS IN STORE.": PRINT270 REM *** MORE CODE THAT DID NOT FIT INTO THE SLIDE FOLLOWS
program :program : lineprogram : line program
line : NUMBER statements
statements : statementstatements : statement statements
statement : PRINT :statement : PRINT expressions :expressions : expressionexpressions : expression ;expressions : expression ; expressions
statement : NAME = expression :
statement : IF expression THEN number :
statement : INPUT name :
statement : GOTO NUMBER :statement : GOSUB NUMBER :statement : RETURN :
statement : REM *** :
statement : FOR NAME = NUMBER TO NUMBER :statement : NEXT NAME :
statement : END :
expression : NUMBERexpression : NAMEexpression : STRINGexpression : operationexpression : ( expression )expression : NAME ( expression )
operation : expression + expressionoperation : expression - expressionoperation : expression * expressionoperation : expression / expressionoperation : expression <= expressionoperation : expression < expressionoperation : expression = expressionoperation : expression <> expressionoperation : expression > expressionoperation : expression >= expression
from rply.token import BaseBox!class Program(BaseBox): def __init__(self, lines): self.lines = lines
AST
class Line(BaseBox): def __init__(self, lineno, statements): self.lineno = lineno self.statements = statements
class Statements(BaseBox): def __init__(self, statements): self.statements = statements
class Print(BaseBox): def __init__(self, expressions, newline=True): self.expressions = expressions self.newline = newline
…
from rply import ParserGenerator!pg = ParserGenerator(["NUMBER", "PRINT", …])
Parser
@pg.production("program : ")@pg.production("program : line")@pg.production("program : line program")def program(p): if len(p) == 2: return Program([p[0]] + p[1].get_lines()) return Program(p)
@pg.production("line : number statements")def line(p): return Line(p[0], p[1].get_statements())
@pg.production("op : expression + expression")@pg.production("op : expression * expression")def op(p): if p[1].gettokentype() == "+": return Add(p[0], p[2]) elif p[1].gettokentype() == "*": return Mul(p[0], p[2])
pg = ParserGenerator([…], precedence=[ ("left", ["+", "-"]), ("left", ["*", "/"])])
parse = pg.build().parse
Compiler/Virtual Machine
Compiler
Virtual Machine
AST
Bytecode
class VM(object): def __init__(self, program): self.program = program
class VM(object): def __init__(self, program): self.program = program self.pc = 0
class VM(object): def __init__(self, program): self.program = program self.pc = 0 self.frames = []
class VM(object): def __init__(self, program): self.program = program self.pc = 0 self.frames = [] self.iterators = []
class VM(object): def __init__(self, program): self.program = program self.pc = 0 self.frames = [] self.iterators = [] self.stack = []
class VM(object): def __init__(self, program): self.program = program self.pc = 0 self.frames = [] self.iterators = {} self.stack = [] self.variables = {}
class VM(object): … def execute(self): while self.pc < len(self.program.instructions): self.execute_bytecode(self.program.instructions[self.pc])
class VM(object): … def execute_bytecode(self, code): raise NotImplementedError(code)
class VM(object): ... def execute_bytecode(self): if isinstance(code, TYPE): self.execute_TYPE(code) ... else: raise NotImplementedError(code)
class Program(object): def __init__(self): self.instructions = []
Bytecode
class Instruction(object): pass
class Number(Instruction): def __init__(self, value): self.value = value!class String(Instructions): def __init__(self, value): self.value = value
class Print(Instruction): def __init__(self, expressions, newline): self.expressions = expressions self.newline = newline
class Call(Instruction): def __init__(self, function_name): self.function_name = function_name
class Let(Instruction): def __init__(self, name): self.name = name
class Lookup(Instruction): def __init__(self, name): self.name = name
class Add(Instruction): pass!class Sub(Instruction): pass!class Mul(Instruction): pass!class Equal(Instruction): pass!...
class GotoIfTrue(Instruction): def __init__(self, target): self.target = target!class Goto(Instruction): def __init__(self, target, with_frame=False): self.target = target self.with_frame = with_frame!class Return(Instruction): pass
class Input(object): def __init__(self, name): self.name = name
class For(Instruction): def __init__(self, variable): self.variable = variable!class Next(Instruction): def __init__(self, variable): self.variable = variable
class Program(object): def __init__(self): self.instructions = [] self.lineno2instruction = {}! def __enter__(self): return self! def __exit__(self, exc_type, exc_value, tb): if exc_type is None: for i, instruction in enumerate(self.instructions): instruction.finalize(self, i)
def finalize(self, program, index): self.target = program.lineno2instruction[self.target]
class Program(BaseBox): … def compile(self): with bytecode.Program() as program: for line in self.lines: line.compile(program) return program
class Line(BaseBox): ... def compile(self, program): program.lineno2instruction[self.lineno] = len(program.instructions) for statement in self.statements: statement.compile(program)
class Line(BaseBox): ... def compile(self, program): program.lineno2instruction[self.lineno] = len(program.instructions) for statement in self.statements: statement.compile(program)
class Print(Statement): def compile(self, program): for expression in self.expressions: expression.compile(program) program.instructions.append( bytecode.Print( len(self.expressions), self.newline ) )
class Print(Statement): ... def compile(self, program): for expression in self.expressions: expression.compile(program) program.instructions.append( bytecode.Print( len(self.expressions), self.newline ) )
class Let(Statement): ... def compile(self, program): self.value.compile(program) program.instructions.append( bytecode.Let(self.name) )
class Input(Statement): ... def compile(self, program): program.instructions.append( bytecode.Input(self.variable) )
class Goto(Statement): ... def compile(self, program): program.instructions.append( bytecode.Goto(self.target) )!class Gosub(Statement): ... def compile(self, program): program.instructions.append( bytecode.Goto( self.target, with_frame=True ) )!class Return(Statement): ... def compile(self, program): program.instructions.append( bytecode.Return() )
class For(Statement): ... def compile(self, program): self.start.compile(program) program.instructions.append( bytecode.Let(self.variable) ) self.end.compile(program) program.instructions.append( bytecode.For(self.variable) )
class WrappedObject(object): pass!class WrappedString(WrappedObject): def __init__(self, value): self.value = value!class WrappedFloat(WrappedObject): def __init__(self, value): self.value = value
class VM(object): … def execute_number(self, code): self.stack.append(WrappedFloat(code.value)) self.pc += 1! def execute_string(self, code): self.stack.append(WrappedString(code.value)) self.pc += 1
class VM(object): … def execute_call(self, code): argument = self.stack.pop() if code.function_name == "TAB": self.stack.append(WrappedString(" " * int(argument))) elif code.function_name == "RND": self.stack.append(WrappedFloat(random.random())) ... self.pc += 1
class VM(object): … def execute_let(self, code): value = self.stack.pop() self.variables[code.name] = value self.pc += 1! def execute_lookup(self, code): value = self.variables[code.name] self.stack.append(value) self.pc += 1
class VM(object): … def execute_add(self, code): right = self.stack.pop() left = self.stack.pop() self.stack.append(WrappedFloat(left + right)) self.pc += 1
class VM(object): … def execute_goto_if_true(self, code): condition = self.stack.pop() if condition: self.pc = code.target else: self.pc += 1
class VM(object): … def execute_goto(self, code): if code.with_frame: self.frames.append(self.pc + 1) self.pc = code.target
class VM(object): … def execute_return(self, code): self.pc = self.frames.pop()
class VM(object): … def execute_input(self, code): value = WrappedFloat(float(raw_input() or “0.0”)) self.variables[code.name] = value self.pc += 1
class VM(object): … def execute_for(code): self.pc += 1 self.iterators[code.variable] = ( self.pc, self.stack.pop() )
class VM(object): … def execute_next(self, code): loop_begin, end = self.iterators[code.variable] current_value = self.variables[code.variable].value next_value = current_value + 1.0 if next_value <= end: self.variables[code.variable] = \ WrappedFloat(next_value) self.pc = loop_begin else: del self.iterators[code.variable] self.pc += 1
def entry_point(argv): try: filename = argv[1] except IndexError: print(“You must supply a filename”) return 1 content = read_file(filename) tokens = lex(content) ast = parse(tokens) program = ast.compile() vm = VM(program) vm.execute() return 0
Entry Point
JIT (in PyPy)1. Identify “hot" loops!
2. Create trace inserting guards based on observed values!
3. Optimize trace!
4. Compile trace!
5. Execute machine code instead of interpreter
from rpython.rlib.jit import JitDriver!jitdriver = JitDriver( greens=[“pc”, “vm”, “program”, “frames”, “iterators”], reds=[“stack”, “variables"])
class VM(object): … def execute(self): while self.pc < len(self.program.instructions): jitdriver.merge_point( vm=self, pc=self.pc, … )
Benchmark10 N = 120 IF N <= 10000 THEN 4030 END40 GOSUB 10050 IF R = 0 THEN 7060 PRINT "PRIME"; N70 N = N + 1: GOTO 20100 REM *** ISPRIME N -> R110 IF N <= 2 THEN 170120 FOR I = 2 TO (N - 1)130 A = N: B = I: GOSUB 200140 IF R <> 0 THEN 160150 R = 0: RETURN160 NEXT I170 R = 1: RETURN200 REM *** MOD A -> B -> R210 R = A - (B * INT(A / B))220 RETURN
cbmbasic 58.22s
basic-c 5.06s
basic-c-jit 2.34s
Python implementation (CPython) 2.83s
Python implementation (PyPy) 0.11s
C implementation 0.03s
Project milestones
2008 Django support
2010 First JIT-compiler
2011 Compatibility with CPython 2.7
2014 Basic ARM support
CPython 3 support
Improve compatibility with C extensions
NumPyPy
Multi-threading support
PyPy STM
PyPy STM
10 loops, best of 3: 1.2 sec per loop10 loops, best of 3: 822 msec per loop
from threading import Thread
def count(n): while n > 0: n -= 1
def run(): t1 = Thread(target=count, args=(10000000,)) t1.start() t2 = Thread(target=count, args=(10000000,)) t2.start() t1.join(); t2.join()
def count(n): while n > 0: n -= 1
def run(): count(10000000) count(10000000)
Inside the Python GIL - David Beazley
PyPy in the real world (1)
High frequency trading platform for sports betsI low latency is a must
PyPy used in production since 2012~100 PyPy processes running 24/7up to 10x speedups
I after careful tuning and optimizing for PyPy
antocuni (PyCon Otto) PyPy Status Update April 07 2017 6 / 19
PyPy in the real world (2)
Real-time online advertising auctionsI tight latency requirement (<100ms)I high throughput (hundreds of thousands of requests per
second)
30% speedup
We run PyPy basically everywhere
Julian Berman
antocuni (PyCon Otto) PyPy Status Update April 07 2017 7 / 19
PyPy in the real world (3)
IoT on the cloud5-10x fasterWe do not even run benchmarks on CPython
because we just know that PyPy is way faster
Tobias Oberstein
antocuni (PyCon Otto) PyPy Status Update April 07 2017 8 / 19