The main improvemens on fastruby v0.0.17 are the optimizations:
- Direct call to CFUNC methods on ruby1.9 (speed up for native methods like Fixnum#+, Fixnum#>, etc...)
- Refactored Ruby to C translation to use inlined code instead of anonymous functions for each expression, this avoids unnecesary C functions
- Improvement on non-local jumps return and next to avoid setjmp when possible using normal goto's. This affects performance on ruby1.9 and ruby1.8 as well
- Implemented scopes on C stack instead of heap when possible
And bug fixes:
- Fixed bug of setjmp and no volatile variables, by adding volatile keyword for all variables used on C functions calling setjmp
- Fixed encoding of native pointers for 32 bits systems, making use of PTR2NUM and NUM2PTR
- Fixed bug of re-executed default argument initialization when redo jump is used
The overall result of the optimi
zations (specially the one about relocating scopes on native stack) are the great time improvement on benchmark1.rb for which fastruby takes about 20-30% less time than ruby1.9
Install
You can clone the repository at github:
You can clone the repository at github:
git clone git://github.com/tario/fastruby.gitOr install it using gem install:
git checkout v0.0.17
gem install fastruby Placing Ruby scopes on native stack instead of heap
Initially, at the very beginning of the project, natural translation of ruby to C lead us to translate ruby local variables as C local variables, but this implementation is wrong since ruby stack can diverge in a tree while proc and continuation objects can potentially hold references to local variable scopes (e.g. lambdas can retain references to scopes, read and write while other scopes "at the same level" can be created but in another branch). Many releases ago I solved this problem by implement a structure to support this "branching" behavior of ruby scopes, the "stack chunk" (See Fastruby 0.0.8 released with support for lambdas and Callcc puzzle on fastruby IV: Execution stack as a DAG, lambda cases for more details)
This implementation, while allowing to create and use lambdas and continuation objects, imply a severe performance penalty forcing to all code to access local variables on the heap, access to the heap is slower than the stack for reasons not now be explained
The key issue here is: Why cannot local variables of the ruby scope live on the native stack? The reason for that is that references to the scope can be potentially created, example:
def foo(a)On the example, a proc is created with a reference to the scope created for the method call. This scope must be allocated on the heap.
proc do
a = a + 1
end
end
pr = foo(10)
p pr.call # 11
p pr.call # 12
What about this?:
def foo(a)
ret = 0
a.each do |x|
ret = ret + 1
end
ret
end
p foo([1,2,3]) # 6
A innocent iteration on a array, ok?
NO, imagine that:
Any block passed to any method can be used to create a lambda. So any method making block calls can potentially allow creation of lambdas referencing scopes created for these methods.
def foo(a)
ret = 0
a.each do |x|
ret = ret + x
end
ret
end
class X
def each(&blk)
$a = proc(&blk)
blk.call(1)
blk.call(1)
blk.call(1)
end
end
x = X.new
p foo(x) # 3
p $a.call(1) # 4
Any block passed to any method can be used to create a lambda. So any method making block calls can potentially allow creation of lambdas referencing scopes created for these methods.
Also, It's virtually impossible to create lambdas referencing scopes of methods which has no block calls. So, since these scopes can not be referenced by lambdas, methods without blocks can allocate their scopes on stack... if not for
Continuations
Continuations are the most macabre and wonderful feature of ruby. Continuations use setjmp/longjmp to work and also write directly to native stack to restore execution data stored on it.
C Local variables on native stack are overwritten each time a continuation is called with the values they had before at the moment that the continuation was created using callcc. This behavior is wrong for ruby variables, so, the use of ruby variables allocated on native stack will lead to unexpected behavior.
For example, this innocent loop would fail if the variables are allocated on stack:
For example, this innocent loop would fail if the variables are allocated on stack:
require "continuation"When the scope of local variables is allocated on heap, the method foo works as expected displaying a countdown on standard output. But, when the scope is allocated on native stack, the first call to callcc copy the entire stack including the variable a storing it; each time the continuation is called to iterate on the loop the value of a is restores to the value they had at the moment of create the continuation (10), the result is an inifinite loop displaying 9 all the time
def foo
a = 10
callcc {|cont| $cont = cont}
a = a - 1
p a
$cont.call if a > 0
end
The key problem here is reading a local variable after a continuation is created, after a continuation is created the value of all local variables allocated on stack will be the value they had at the moment of create the continuation. So, reading a local after creating a continuation may result on unexpected behavior. In summary. Any method with the possibility of following the sequence: continuation -> variable read can potentially work in unexpected way.
Converting Syntax trees to Graphs
The only way to detect potential continuation -> variable read sequences is to generate a graph from the syntax tree of the method being analyzed and then search the sequence on the graph
The graph is composed with the nodes of the AST as vertexes, and the edges are the possibility of execute a given destination node after another origin node. For example, the following source code:
def foo(a)
if a.to_i == 0
10
else
20
end
end
Corresponds to the following syntax tree:
And generates the following graph:
In that case, there is no read of local variables after call on any path, so the method of the example is able to allocate their local variables on native stack
In the other hand, the following method cannot use native stack:
def foo(a,b)Syntax tree:
while (a > 0)
b.foo
a = a - 1
end
end
Graph:
The use of while makes possible the existence of a cyclic path on the graph. Beyond that, exists many tours that make native stack storage for locals prohibitive, they are marked with blue, green and orange:
Note that all prohibitive tours on graph goes from a call node to a lvar node, of course, the method of the while example must be implemented with the local variable scope allocated on heap
Viewed in this way, almost any method with normal complexity will fall on "heap" category since the only requirement for this to be so is having at least one possible prohibitive tour (from a call node to a lvar node) on the graph, and for now this is true. But this algorithm will have greater relavance when the analyzer know a priori what call can potentially create a continuation and what call will never create a continuation (this calls will be ignored). Today, this improvement gains 20-30% of timed compared with MRI1.9
For example, in the while code sample the method calls are "foo", ">" and "-", if the analyzer can afford assume that these calls never will create a continuation... > and - on numeric dont create continuation objects, but foo?. Of course this will be a little more complicated and probably the optimizations around this will be about type inferences, and qualifiers
