In dynamic languages, how is dynamically typed code JIT compiled into machine code? More specifically: does the compiler infer the types at some point? Or is it strictly interpreted in these cases?
For example, if I have something like the following pseuocode
def func(arg)
if (arg)
return 6
else
return "Hi"
How can the execution platform know before running the code what the return type of the function is?
In general, it doesn't. However, it can assume either type, and optimize for that. The details depend on what kind of JIT it is.
The so-called tracing JIT compilers interpret and observe the program, and record types, branches, etc. for a single run (e.g. loop iteration). They record these observations, insert a (quite fast) check that these assumptions are still true when the code is executed, and then optimize the heck out of the following code based on these assuptions. For example, if your function is called in a loop with a constantly true argument and adds one to it, the JIT compiler first records instructions like this (we'll ignore call frame management, memory allocation, variable indirection, etc. not because those aren't important, but because they take a lot of code and are optimized away too):
; calculate arg
guard_true(arg)
boxed_object o1 = box_int(6)
guard_is_boxed_int(o1)
int i1 = unbox_int(o1)
int i2 = 1
i3 = add_int(res2, res3)
and then optimizes it like this:
; calculate arg
; may even be elided, arg may be constant without you realizing it
guard_true(arg)
; guard_is_boxed_int constant-folded away
; unbox_int constant-folded away
; add_int constant-folded away
int i3 = 7
Guards can also be moved to allow optimizing earlier code, combined to have fewer guards, elided if they are redundant, strengthened to allow more optimizations, etc.
If guards fail too frequently, or some code is otherwise rendered useless, it can be discarded, or at least patched to jump to a different version on guard failure.
Other JITs take a more static approach. For instance, you can do quick, inaccurate type inference to at least recognize a few operations. Some JIT compilers only operate on function scope (they are thus called method JIT compilers by some), so they probably can't make much of your code snippet (one reason tracing JIT compilers are very popular for). Nevertheless, they exist -- an example is the latest revision of Mozilla's JavaScript engine, Ion Monkey, although it apparently takes inspiration from tracing JITs as well. You can also insert add not-always-valid optimizations (e.g. inline a function that may be changed later) and remove them when they become wrong.
When all else fails, you can do what interpreters do, box objects, use pointers to them, tag the data, and select code based on the tag. But this is extremely inefficient, the whole purpose of JIT compilers is getting rid of that overhead, so they will only do that when there is no reasonable alternative (or when they are still warming up).
Related
Coming from Wolfram Mathematica, I like the idea that whenever I pass a variable to a function I am effectively creating a copy of that variable. On the other hand, I am learning that in Julia there are the notions of mutable and immutable types, with the former passed by reference and the latter passed by value. Can somebody explain me the advantage of such a distinction? why arrays are passed by reference? Naively I see this as a bad aspect, since it creates side effects and ruins the possibility to write purely functional code. Where I am wrong in my reasoning? is there a way to make immutable an array, such that when it is passed to a function it is effectively passed by value?
here an example of code
#x is an in INT and so is immutable: it is passed by value
x = 10
function change_value(x)
x = 17
end
change_value(x)
println(x)
#arrays are mutable: they are passed by reference
arr = [1, 2, 3]
function change_array!(A)
A[1] = 20
end
change_array!(arr)
println(arr)
which indeed modifies the array arr
There is a fair bit to respond to here.
First, Julia does not pass-by-reference or pass-by-value. Rather it employs a paradigm known as pass-by-sharing. Quoting the docs:
Function arguments themselves act as new variable bindings (new
locations that can refer to values), but the values they refer to are
identical to the passed values.
Second, you appear to be asking why Julia does not copy arrays when passing them into functions. This is a simple one to answer: Performance. Julia is a performance oriented language. Making a copy every time you pass an array into a function is bad for performance. Every copy operation takes time.
This has some interesting side-effects. For example, you'll notice that a lot of the mature Julia packages (as well as the Base code) consists of many short functions. This code structure is a direct consequence of near-zero overhead to function calls. Languages like Mathematica and MatLab on the other hand tend towards long functions. I have no desire to start a flame war here, so I'll merely state that personally I prefer the Julia style of many short functions.
Third, you are wondering about the potential negative implications of pass-by-sharing. In theory you are correct that this can result in problems when users are unsure whether a function will modify its inputs. There were long discussions about this in the early days of the language, and based on your question, you appear to have worked out that the convention is that functions that modify their arguments have a trailing ! in the function name. Interestingly, this standard is not compulsory so yes, it is in theory possible to end up with a wild-west type scenario where users live in a constant state of uncertainty. In practice this has never been a problem (to my knowledge). The convention of using ! is enforced in Base Julia, and in fact I have never encountered a package that does not adhere to this convention. In summary, yes, it is possible to run into issues when pass-by-sharing, but in practice it has never been a problem, and the performance benefits far outweigh the cost.
Fourth (and finally), you ask whether there is a way to make an array immutable. First things first, I would strongly recommend against hacks to attempt to make native arrays immutable. For example, you could attempt to disable the setindex! function for arrays... but please don't do this. It will break so many things.
As was mentioned in the comments on the question, you could use StaticArrays. However, as Simeon notes in the comments on this answer, there are performance penalties for using static arrays for really big datasets. More than 100 elements and you can run into compilation issues. The main benefit of static arrays really is the optimizations that can be implemented for smaller static arrays.
Another package-based options suggested by phipsgabler in the comments below is FunctionalCollections. This appears to do what you want, although it looks to be only sporadically maintained. Of course, that isn't always a bad thing.
A simpler approach is just to copy arrays in your own code whenever you want to implement pass-by-value. For example:
f!(copy(x))
Just be sure you understand the difference between copy and deepcopy, and when you may need to use the latter. If you're only working with arrays of numbers, you'll never need the latter, and in fact using it will probably drastically slow down your code.
If you wanted to do a bit of work then you could also build your own array type in the spirit of static arrays, but without all the bells and whistles that static arrays entails. For example:
struct MyImmutableArray{T,N}
x::Array{T,N}
end
Base.getindex(y::MyImmutableArray, inds...) = getindex(y.x, inds...)
and similarly you could add any other functions you wanted to this type, while excluding functions like setindex!.
Sorry if this is detailed somewhere, I tried searching in the different documentations of Frama-C without luck.
I'm trying to do dead code elimination in my code, but I don't understand the results of the tool. Is there any paper / documentation that explains how this plugin works? I only know that it uses the results of the Value analysis.
Admittedly, the sparecode page on Frama-C's website is a bit terse. However, this is partly due to the fact that there's not much to parameterize in this plug-in. Mainly, it is a specialized form of the slicing plug-in, where the criterion is "preserve the state at the end of the program".
More generally, slicing consists in removing instructions that do not contribute to a user given criterion (e.g. the whole program state at a given point, the validity status of an ACSL annotation, or simply the fact that the program reaches a particular instruction).
In order to compute such slice, slicing, hence sparecode, indeed relies on the results of Eva, mainly to obtain an over-approximation (as always with Eva) of the dependencies between the various memory locations involved at each point in the program (you might want to have a look at Chapter 7 of the Eva manual which deals with dependencies. Very roughly speaking the slice will consist in the transitive closure of the dependencies for the memory locations involved in the criterion (in presence of pointers and branches, this notion of "transitive closure" becomes a bit complicated to define formally, but the essence is there).
Now, with respect to dead code, there are two points worth noting:
As mentioned before, Eva provides an over-approximation of the behavior of the program. For slicing, this means that some statements might be kept even though they do not contribute to the slicing criterion in any concrete trace, but appear to do so in the abstract trace due to over-approximation. On the other hand, if a statement is not included, then it definitely does not contribute to the criterion.
For sparecode, not contributing to the final state of the program does not mean that the code is dead, but merely that all its side-effects are shadowed by further instructions. The simplest example of that would be the following:
int x;
int main() {
x = 1;
x = 2;
}
Here, the x=1 has no influence over the final state of the program, and only x=2 will be kept.
I'm working on an Ada based project that I'm not terribly familiar with, and I've just seen something that at first glance seems inefficient, but of course that all depends on what the compiler might do.
if Ada.Strings.Fixed.Trim
(Source => Test_String,
Side => Ada.Strings.Both) =
String_1 then
--Do something here
elsif Ada.Strings.Fixed.Trim
(Source => Test_String,
Side => Ada.Strings.Both) =
String_2 then
--Do something else here
end if;
I feel that it would be more efficient to call the Trim procedure and store the result in a String variable, then test against different Strings in each condition of the if statement, especially if there are many conditions to check (never mind that using a binary search might be even better). Of course, I may be wrong, so my question is, is there something about compile time optimisation in Ada that I do not know about, that might cause the Trim function to only be called once, and only have the result tested in each condition of the if statement?
That would be compiler-dependent, not language-dependent. Certainly, GNAT GPL 2013 calls Trim twice both at -O2 and at -O3.
Your (and my) intuition seems to be right: do the trim once and store the result ...
Trimmed : constant String :=
Ada.Strings.Fixed.Trim (Source => Test_String, Side => Ada.Strings.Both);
... though personally I’d write
Trimmed : constant String :=
Ada.Strings.Fixed.Trim (Test_String, Side => Ada.Strings.Both);
on the grounds that in this case no one should need the named parameter association to clarify the programmer’s intention!
I don't think we need to worry about efficiency for this case. But the coding style that do need to be improved. To define a variable to hold the result of the Trim function is a better practice in my opinion. The program logic is clearer and easier for maintenance.
Let's say you want to change the Trim function to another one or change the parameter passing in, you only need to change one place. Although for this case there are only two places, you still have a chance to error. If there are more cases to test, then there definitely will be missing case.
While I have no disagreement whatsoever with your assessment or Simon's answer, it's worth bearing in mind what optimisation actually means...
If each call to "trim" takes (conservatively) 0.1ms of CPU time, and the rewrite takes 2 minutes, the breakeven point in time saved is over 1 million executions of that particular statement!
This of course ignores (a) on the positive side, the value of gaining experience, and (b) on the negative side, the fact that CPU time is nowadays less valuable than your time. And (c) the time we have spent discussing the optimisation too!
With respect to "compile time optimisation in Ada" , the Ada language doesn't say anything about this. All it would say here is that the program must behave as if the function were called twice, i.e. it must produce the same results. But if the compiler "knows" that the function would produce the exact same result the second time, it can generate code that calls it only once. It can't do this for a function call in general, because a function could include side effects or use global variables that whose values could have changed. In this case, since the effects of Ada.Strings.Fixed.Trim are defined by the language and since those effects do guarantee that the effects would be the same, a compiler could, in theory, mark this function as one for which a call with the exact same parameters could be optimized. (A function in a Pure package would definitely be one where a second call could be eliminated, but unfortunately Ada.Strings.Fixed isn't defined by the Ada language as being Pure.) To find out whether a compiler actually does this particular optimization, though, you'd have to try it and check the code. I believe that if Test_String is not marked as Volatile, then compilers are allowed to assume that its value will be the same for each Ada.Strings.Fixed.Trim call (i.e. it can't be changed by another task in between the calls), but I'm not 100% sure about this.
I'd declare a constant to hold the result (like Simon), but for me this is more out of a desire to avoid duplicated code than it is out of a concern for efficiency.
My Frama-C plug-in creates some varinfos with makeGlobalVar ~logic:true name type. These varinfos do not exist in the AST (they are placeholders for the results of calls to allocating functions in the target program, created “dynamically” during the analysis). If my plug-in takes care not to keep any strong pointer onto these varinfos, will they have a chance to be garbage-collected? Or are they registered in a data structure with strong pointers? If so, would it be possible to make that data structure weak? OCaml does not have the variety of weak data structure found in the literature for other languages, but there is nothing a periodical explicit pass to clean up empty stubs cannot fix.
Now that I think about it, I may not even have to create a varinfo. But it is a bit late to change my plug-in now. What I use of the varinfo is a name and a representation of a C type. Function makeGlobalVar offers a guarantee of unicity for the name, which is nice, I guess, as long as it does not create a strong pointer to it or to part of it in the process.
Context:
Say that you are writing a C interpreter to execute C programs that call malloc() and free(). If the target program does not have a memory leak (it frees everything it allocates and never holds too much memory), you would like the interpreter to behave the same.
If you don't explicitely register the varinfos into one of the Globals table, Frama-C won't do it for you (and in fact, if you do, you're supposed to add their declaration in the AST and vice-versa), so I guess that you are safe here. The only visible side-effect as far as the kernel is concerned should be the incrementation of the Vid counter. Note however that makeGlobalVar itself does not guarantee the unicity of the vname, but only of the vid field.
Let's say we define a function c sum(a, b), functional programming -style, that returns the sum of its arguments. So far so good; all the nice things of FP without any problems.
Now let's say we run this in an environment with dynamic typing and a singleton, stateful error stream. Then let's say we pass a value of a and/or b that sum isn't designed to handle (i.e. not numbers), and it needs to indicate an error somehow.
But how? This function is supposed to be pure and side-effect-less. How does it insert an error into the global error stream without violating that?
No programming language that I know of has anything like a "singleton stateful error stream" built in, so you'd have to make one. And you simply wouldn't make such a thing if you were trying to write your program in a pure functional style.
You could, however, have a sum function that returns either the sum or an indication of an error. The type used to do this is in fact often known by the name Either. Then you could easily make a function that invokes a whole bunch of computations that could possibly return an error, and returns a list of all the errors that were encountered in the other computations. That's pretty close to what you were talking about; it's just explicitly returned rather than being global.
Remember, the question when you're writing a functional program is "how do I make a program that has the behavior I want?" not, "how would I duplicate one particular approach taken in another programming style?". A "global stateful error stream" is a means not an end. You can't have a global stateful error stream in pure function style, no. But ask yourself what you're using the global stateful error stream to achieve; whatever it is, you can achieve that in functional programming, just not with the same mechanism.
Asking whether pure functional programming can implement a particular technique that depends on side effects is like asking how you use techniques from assembly in object-oriented programming. OO provides different tools for you to use to solve problems; limiting yourself to using those tools to emulate a different toolset is not going to be an effective way to work with them.
In response to comments: If what you want to achieve with your error stream is logging error messages to a terminal, then yes, at some level the code is going to have to do IO to do that.1
Printing to terminal is just like any other IO, there's nothing particularly special about it that makes it worthy of singling out as a case where state seems especially unavoidable. So if this turns your question into "How do pure functional programs handle IO?", then there are no doubt many duplicate questions on SO, not to mention many many blog posts and tutorials speaking precisely to that issue. It's not like it's a sudden surprise to implementors and users of pure programming languages, the question has been around for decades, and there have been some quite sophisticated thought put into the answers.
There are different approaches taken in different languages (IO monad in Haskell, unique modes in Mercury, lazy streams of requests and responses in historical versions of Haskell, and more). The basic idea is to come up with a model which can be manipulated by pure code, and hook up manipulations of the model to actual impure operations within the language implementation. This allows you to keep the benefits of purity (the proofs that apply to pure code but not to general impure code will still apply to code using the pure IO model).
The pure model has to be carefully designed so that you can't actually do anything with it that doesn't make sense in terms of actual IO. For example, Mercury does IO by having you write programs as if you're passing around the current state of the universe as an extra parameter. This pure model accurately represents the behaviour of operations that depend on and affect the universe outside the program, but only when there is exactly one state of the universe in the system at any one time, which is threaded through the entire program from start to finish. So some restrictions are put in
The type io is made abstract so that there's no way to construct a value of that type; the only way you can get one is to be passed one from your caller. An io value is passed into the main predicate by the language implementation to kick the whole thing off.
The mode of the io value passed in to main is declared such that it is unique. This means you can't do things that might cause it to be duplicated, such as putting it in a container or passing the same io value to multiple different invocations. The unique mode ensures that you can only ass the io value to a predicate that also uses the unique mode, and as soon as you pass it once the value is "dead" and can't be passed anywhere else.
1 Note that even in imperative programs, you gain a lot of flexibility if you have your error logging system return a stream of error messages and then only actually make the decision to print them close to the outermost layer of the program. If your log calls are directly writing the output immediately, here's just a few things I can think of off the top of my head that become much harder to do with such a system:
Speculatively execute a computation and see whether it failed by checking whether it emitted any errors
Combine multiple high level systems into a single system, adding tags to the logs to distinguish each system
Emit debug and info log messages only if there is also an error message (so the output is clean when there are no errors to debug, and rich in detail when there are)