As a student of compilers myself, it's my understanding that producing bytecode is as easy as--if not vastly easier than--producing code in another language, and that this is particularly true when the target language is high-level. The only exception to this I can envision is when the source language maps easily and completely to every single target language, in which case the source language must be the least common denominator, and the compiler may be a glorified awk script.
Significantly, a major feature of the JVM and CLR bytecodes is that targeting them makes it extremely easy to permit interoperation with other code native to the same VM. I'm not sure generating source would be any improvement on this whatsoever. I also question the value of being able to debug generated code; assuming that HaXe is indeed more than a glorified awk script, the process might be compared to using an assembly debugger on C++-generated code. Not entirely useless, certainly, but neither is it precisely desirable.
VMs can have bugs and nuances that aren't captured in their specification - and a compiler has to evolve awareness of this in an iterative fashion. Rendering to source languages constitutes another way to sanity check by going through a compiler that has already dealt with those implementation concerns. So Haxe is taking on a lot of extra work, as you say, but it also gains more assurances that output is "as good as" target-native code.
> VMs can have bugs and nuances that aren't captured in their specification
So can compilers. Adding the additional translation layer of javac or equivalent increases the potential of being affected by bugs in third-party code. Let P be the probability of encountering a bug in the JVM, and Q be the probability of encountering a bug in javac. If we multiply the complements, (1-Q)x(1-P) we have the probability of final execution being in keeping with the HaXe implementation's intentions. Assuming all values are nonzero, (1-P)x(1-Q) < (1-P), therefore the additional translation layer increases the probability of being affected by a bug. Not to mention that the vastly more complex output format increases the possibility of introducing bugs into HaXe itself.
> output is "as good as" target-native code.
But it isn't. Translating code between languages is lossy. Unless HaXe is trivial, the HaXe implementation has information that could be used to generate better bytecode that is lost when bytecode generation is performed by a tool that knows nothing about HaXe, e.g. javac.
> Adding the additional translation layer increases the potential of bugs
> the implementation has information that could be used to generate better bytecode
These are theoreticals. Let me analogize. Some algorithms have a worst-case big O that is significantly worse than the average case. So even though the algorithm could have very poor performance in certain situations, it's better for the "real-world" cases that it gets used in.
Most of the existing Haxe targets have similar external semantics(Algol-derived, GC, dynamic types), so we're in an average-case situation of both reliability and performance; the Q probability is close to zero because the code we input isn't that drastically different from human source input, and the input we pass in from Haxe can be a bit more optimized in most respects, because we don't have to make it maintainable. If we were targeting a really bad compiler, it would blow up. But in practice, it doesn't and we gain more than we lose, even when considering areas where there's an impedance mismatch and we have to, for example, add dynamic types on top.
Bytecode output is more sensible within a short-term view - a single project with known specifications. But it's ultimately the need for flexibility that drives Haxe, and you aren't gaining additional flexibility from bytecode.
> it's my understanding that producing bytecode is as easy as--if not vastly easier than--producing code in another language
Maybe true for unoptimized code. However, Java and C# compilers have received years of work to make efficient use of their respective VMs; standing on their shoulders makes some sense.
Really? What kind of optimizations does the C# compiler do when emitting IL? As far as I know, they are fairly straightfoward optimizations - there's nothing really that complex going on there. The C# team has said they aim for a straightfoward mapping to IL.
You won't see the C# compiler inlining functions (even though the CLR does _way_ better with large functions and doesn't handle inlining well). It won't propagate constant expressions. You won't see it transforming recursive functions into loops. Does it even remove unused variables?
All the optimizations that C# really relies on for performance are handled by the JIT, and any compiler following the same patterns will get the same enhancements, plus the ability to emit better IL than C#.
That said, I'm still unconvinced it's easier to emit MSIL than C#, and I'm rather well versed with MSIL.
Maybe things are different on Dalvik, but javac for the Oracle JVM is about as braindead as a java compiler can get; all of the optimizing is done during JIT.
Source to source solves the problem where the only way to ship for certain platform is to compile only with the toolchain provided.
The Unity3D engine for example uses Mono's AOT (C#) to produce a very big assembly file (.S) which later goes through apple's own (or is it gnu's) as (assembler).
Once again, there's nothing wrong with jvm output. It's "better" for the reasons you say. And, there may be a jvm target down the road.
That said, there's a number of situations where providing source code in java/c++ is critical for non-technical reasons... Say, if you want to get your project accepted in one of the various mobile app stores.
> Once again, there's nothing wrong with jvm output. It's "better" for the reasons you say.
So if my understanding, including that bytecode generation requires a comparable amount of effort, why didn't they do that? I don't intend to critique their decisions; I'm interested in why they made them. Their reasoning might well teach me something.
> providing source code in java/c++ is critical for ... various mobile app stores.
Really? I wasn't aware of that. Are we talking Apple's and Google's? What app store's catalog is predominately C++?
> So if my understanding, including that bytecode generation requires a comparable amount of effort, why didn't they do that?
They might. It's just not what they've chosen to do first.
> Really? I wasn't aware of that. Are we talking Apple's and Google's? What app store's catalog is predominately C++?
Well, we're talking the iOS app store here. Objective C is the main language, but Apple also supports c++ xcode projects, which Haxe can produce. The reason why Haxe targets c++ is because it can provide garbage collection through the Boehm libs, and (in general) can be made to fit better with the Ecmascript nature of the Haxe language.
Google is much less restrictive in terms of supported languages, but their java toolkit is very polished.
I also think that source code generation also lets you understand better what the compiler is doing, and how to best take advantage of it. Right now, the java target is very early, but it already has a clever way of handling reflection that is much faster than the standard method. It was interesting to me to read through the generated output, even if it was a little ugly.
Why should it be predominately C++, in order to have an easy time being reviewed?
There's plenty of C++ code in the iOS App Store. I know a couple of people who've developed most of the code for their iOS apps in Visual Studio. They were computer vision apps with just a little bit of GUI toolkit code.
Significantly, a major feature of the JVM and CLR bytecodes is that targeting them makes it extremely easy to permit interoperation with other code native to the same VM. I'm not sure generating source would be any improvement on this whatsoever. I also question the value of being able to debug generated code; assuming that HaXe is indeed more than a glorified awk script, the process might be compared to using an assembly debugger on C++-generated code. Not entirely useless, certainly, but neither is it precisely desirable.