__attribute__((used)) and druntime

[Mike via Digitalmars-d]

On Saturday, 11 April 2015 at 15:35:47 UTC, w0rp wrote:
> I've been looking into compiling smaller executables, and the 
> trick I learned from watching sessions from the last DConf is 
> to use the KEEP symbol in a linker script with something like 
> gc-sections. So at the compilation stage, symbols are put in 
> different sections, a linker script marks some symbols to be 
> kept if they are not referenced, and then unreferenced symbols 
> are removed, resulting in sometimes very small executables.
>
> The problem is that if you don't KEEP some symbols for 
> druntime, or in your own program, you can run into errors 
> because symbols which weren't apparently referenced now get 
> used and your program breaks. I was just thinking, "There must 
> be a GCC extension for doing this without a linker script," I 
> *think* there might be.
>
> From what I just read, __attribute__((used)) in GCC should do 
> the job, and I think it should be usable from GDC with its 
> attribute pragmas. I believe there's also an attribute for 
> putting symbols in particular sections. I'm wondering, 
> supposing I'm not completely wrong and this actually works, 
> would it behoove us to add some version() blocks for GDC (Also 
> LDC?) in druntime and flag all of the symbols like this 
> appropriately, so that creating executables with gc-sections 
> will Just Work in some cases without a linker script?
>
> What do people think? Am I on to something, or I am just making 
> fanciful ideas which won't work?

I tried to elicit some direction on this in the past, but I 
wasn't able to get very far.  The following forum threads may 
interest you:

"Dead code removal"
http://forum.dlang.org/post/fqocnakibszciyptihpi@forum.dlang.org

"--gc-section and GDC]
http://forum.dlang.org/post/cqzazaqxpwezignixuds@forum.dlang.org

"Difference in perpective between LDC and GDC"
http://forum.dlang.org/post/xuosrmyumtitahinzrfi@forum.dlang.org

I get the impression from those discussions that the binutils 
implementation for D is incomplete but it's still not clear to me 
whether dead code removal, or retainining "used" code, is a job 
for the compiler or the linker.  It may require some 
orchestration between the two.

I proposed an idea to move declarations for certain runtime hooks 
out of the compiler and into druntime as *.di files.  That would 
allow for features of the runtime to be decorated in the 
platform's port with "used", "weak", "section", etc., but 
although I proposed it, I'm not convinced it's the way forward.  
You can see the thread here:  
http://forum.dlang.org/post/psssnzurlzeqeneagora@forum.dlang.org

Mike