Most of the precise GCs I know of are in JIT'd or interpreted languages
GC is orthogonal to JITing. OCaml and Haskell are two examples of languages with non-JIT AOT compilers (ocamlopt and ghc, respectively) that do precise GC. I'm sure there are plenty of others.
Yep, you're right, it is. And I totally forgot about those compiled functional languages. Thanks! I mentioned JITs and interpreters because those are the obvious cases where type information still exists at runtime, but the compiler can emit that metadata too (as I mentioned).
GCing has nothing to do with a language being functional.
All you need is access to the roots and the ability always to determine which data is another pointer. The main 'enemy' of this is the ability to cast integers to pointers (as you may in C/C++).
Interpreters don't necessarily do GC: if the language that the interpreter is written in has GC, then the interpreted language is automatically GCed. If the interpreter is written in a non-GC language you have to manage the interpreted program's memory manually too (which you may do using a GC for the interpreted language).
As an aside: this is comparable to evaluation order: if the interpreter is written in a call-by-value language, then the interpreted language is CBV, if the interpreter is written in a lazy language, the interpreted language has lazy evaluation. (Assuming you are not doing something fancy like translation to continuation-passing-style.)
GCing has nothing to do with a language being functional
Yep, I didn't say it did! You mentioned Haskell and OCaml as counterexamples. I referred to them as "the compiled functional languages". This isn't a claim that only compiled functional languages do GC.
Interpreters don't do GC
Sure they do! One example: Ruby (the original interpreter in C). It has a full mark/sweep collector that traverses user data structures. Custom object types implement their own 'mark' routines, so it's fully precise. (I've implemented a C extension that uses this API.)
Another example: often, Java runtimes interpret Java bytecode on the first pass, before JIT'ing, to give fast startup on cold code. The JRE would be written in C or C++ (non-GC'd), and Java is GC'd.
In general an interpreter can add any additional or different semantics it needs to on top of the host language's semantics. I don't think your call-by-value claim is correct either, because function calls in the interpreted program do not map one-to-one to function calls in the interpreter. (The interpreter would generally manage a data structure representing the interpreted program's stack.)
My goodness, this is going off the rails. Sorry to OP for derailing. Just trying to fix things that are Wrong On The Internet...
One example: Ruby (the original interpreter in C).
OK, I should have said, and have now modified my original statement to "Interpreters don't necessarily do GC".
In general an interpreter can add any additional or different semantics
Yes, and that falls under "something fancy"! The naive way of interpreting object-language application by meta-language application transfers the latter's calling discipline to the former.
Of course, but no one claimed that interpreters necessarily do GC either, so this statement is not very useful. You keep arguing against claims that aren't there. The only claim ever made in response to OP was that precise GC requires runtime information, and that this can sometimes be easier when it exists anyway (as in JITs and real-world interpreters), but can also be implemented when the compiler produces it AOT. I hope that's clear enough :-)
Comments
Yep, you're right, it is. And I totally forgot about those compiled functional languages. Thanks! I mentioned JITs and interpreters because those are the obvious cases where type information still exists at runtime, but the compiler can emit that metadata too (as I mentioned).
GCing has nothing to do with a language being functional. All you need is access to the roots and the ability always to determine which data is another pointer. The main 'enemy' of this is the ability to cast integers to pointers (as you may in C/C++).
Interpreters don't necessarily do GC: if the language that the interpreter is written in has GC, then the interpreted language is automatically GCed. If the interpreter is written in a non-GC language you have to manage the interpreted program's memory manually too (which you may do using a GC for the interpreted language). As an aside: this is comparable to evaluation order: if the interpreter is written in a call-by-value language, then the interpreted language is CBV, if the interpreter is written in a lazy language, the interpreted language has lazy evaluation. (Assuming you are not doing something fancy like translation to continuation-passing-style.)
Yep, I didn't say it did! You mentioned Haskell and OCaml as counterexamples. I referred to them as "the compiled functional languages". This isn't a claim that only compiled functional languages do GC.
Sure they do! One example: Ruby (the original interpreter in C). It has a full mark/sweep collector that traverses user data structures. Custom object types implement their own 'mark' routines, so it's fully precise. (I've implemented a C extension that uses this API.)
Another example: often, Java runtimes interpret Java bytecode on the first pass, before JIT'ing, to give fast startup on cold code. The JRE would be written in C or C++ (non-GC'd), and Java is GC'd.
In general an interpreter can add any additional or different semantics it needs to on top of the host language's semantics. I don't think your call-by-value claim is correct either, because function calls in the interpreted program do not map one-to-one to function calls in the interpreter. (The interpreter would generally manage a data structure representing the interpreted program's stack.)
My goodness, this is going off the rails. Sorry to OP for derailing. Just trying to fix things that are Wrong On The Internet...
Of course, but no one claimed that interpreters necessarily do GC either, so this statement is not very useful. You keep arguing against claims that aren't there. The only claim ever made in response to OP was that precise GC requires runtime information, and that this can sometimes be easier when it exists anyway (as in JITs and real-world interpreters), but can also be implemented when the compiler produces it AOT. I hope that's clear enough :-)