Well, for a start there's the problem that LLVM itself is architecture-dependent, last I checked. It's not as arhictecture-dependent as assembly but the last time I looked you couldn't generate the same LLVM on x86 and x86-64, say. Has that changed?
It looks like PNaCl aims to "solve" this problem by using 32-bit-targeted LLVM... including 32-bit pointer alignment. It's not immediately obvious how portable this would be to some hardware, and more importantly I'm not aware of any plans to support anything other than x86 and ARM initially (well, and x86-64 using its ability to run 32-bit code as far as I can tell). Am I just missing something on this front?
If I'm not missing anything, then this project's main impact if it caught on would be to restrict the set of hardware on which you can access web content, which is NOT a good thing.
> Well, for a start there's the problem that LLVM itself is architecture-dependent, last I checked.
The idea is not to produce a single cross-platform binary using LLVM - let alone using GCC! - but to produce a set of NaCl binaries for common archs from your portable HLL code using a cross-platform compiler - like LLVM, or GCC - plus PNaCl (or whatever) as the wildcard. Content negotiation is your friend.
> It looks like PNaCl aims to "solve" this problem by using 32-bit-targeted LLVM... including 32-bit pointer alignment. It's not immediately obvious how portable this would be to some hardware,
shrug I haven't heard of any major problems in converting LLVM IR to decently-performing assembler in various archs, so I presume that PNaCl will be able to do okay, especially since the bar is not set very high: consistently world-beating performance is not a necessity in a backstop solution. And if you're on a new or obscure platform (or, er, IE/Windows/x86-64) you're not guaranteed the swiftest Javascript speeds either. The likely worst is that PNaCl will need a redesign: see below.
> and more importantly I'm not aware of any plans to support anything other than x86 and ARM initially (well, and x86-64 using its ability to run 32-bit code as far as I can tell). Am I just missing something on this front?
As I said in my previous post "Obviously the implementation is incomplete and immature, but Mozilla and Opera aren't even pretending that their opposition to NaCl is based on the immaturity of the implementation." Seriously, if the problem with NaCl is that it's not mature and widely ported enough yet, then there's an obvious solution to that.
> I haven't heard of any major problems in converting LLVM IR to decently-performing assembler
But normally when you create LLVM IR you can choose what your pointer-size representation is, as I understand. Again, I could be completely off base here; I've just read some of the docs and talked to people who have worked with LLVM, not worked with it myself. Please tell me if I'm wrong!
The problem with PNaCl is not that it's not widely ported enough "yet", but rather that there are no plans for a way to run it on platforms what PNaCl hasn't been ported to. Compare this to JS, where you can get JS working on a new platform in fairly short order (e.g. by just compiling Spidermonkey, which is largely fairly portable C code and has a platform-independent interpreter). Now your new platform won't have a JIT yet at that point, so performance may not be great, but at least you'll have a fighting chance at using the web. You won't have that without a pretty large porting job if NaCl is in wide use.
> Well, for a start there's the problem that LLVM itself is architecture-dependent, last I checked. It's not as arhictecture-dependent as assembly but the last time I looked you couldn't generate the same LLVM on x86 and x86-64, say. Has that changed?
LLVM bytecode is 100% architecture and platform independent, unless you call into platform-specific bits or into blobs of machine code.
And if I read http://llvm.org/docs/LangRef.html correctly, some bitcode is "not supported by all targets". How is that reconciled with "100% platform and architecture independent"?
It looks like PNaCl aims to "solve" this problem by using 32-bit-targeted LLVM... including 32-bit pointer alignment. It's not immediately obvious how portable this would be to some hardware, and more importantly I'm not aware of any plans to support anything other than x86 and ARM initially (well, and x86-64 using its ability to run 32-bit code as far as I can tell). Am I just missing something on this front?
If I'm not missing anything, then this project's main impact if it caught on would be to restrict the set of hardware on which you can access web content, which is NOT a good thing.