[LLVMdev] Address space information dropped

[Joerg Sonnenberger]

On Tue, May 08, 2012 at 12:57:15AM +0200, Ivan Llopard wrote:
> Le 07/05/2012 17:15, Joerg Sonnenberger a écrit :
> >On Mon, May 07, 2012 at 05:15:59PM +0200, Ivan Llopard wrote:
> >>Thanks for your quick response Joerg.
> >>
> >>We have a very small test case where there is global array and its
> >>address space attribute specified like in the following code
> >>
> >>const int __attribute__((address_space(256))) fangle[13] =
> >>   {2341, 4681, 7022, 9362, 11703, 1403,16384,
> >>    18725, 21065, 23406, 25746, 28087, 30427};
> >>
> >>I need to put its initializer in another memory because it has a
> >>different address space and I wanted to pass this information
> >>through a target-depend asm directive.
> >>At the moment, I've overridden EmitGlobalVariable() from AsmPrinter
> >>because I didn't see any other spot to put this special directive.
> >Have you thought about always tagging such global variables with
> >__attribute__((section("fancyaddresspace"))) as well? That seems to be
> >the sanest way I can think of.
> 
> Yes, but it seems a little odd to me. Is it correct to assume that
> different sections imply different address spaces ?
> I rather wanted to have a special asm directive on the top of
> section definitions to switch between memories following an
> identical section layout for both of them.

Let's take a step back. "address space" is a type modifier like __thread.
>From codegen perspective, primarily changes the way access is done.
Variable declaration requires obtaining some pointer value and possibly
initialising it. For __thread, this is done by putting the object into
.tdata or .tbss sections and letting the linker do the rest. That is:
- merging corresponding sections
- ensuring that (in cooperation with the runtime linker and thread
library) initialisation happens as needed
- correct relocation records are created.

What needs to happen in your use case is essentially the same. You don't
need some of the complications of __thread, but the idea is:
(1) Put the objects into a separate section.
(2) Make sure that the section ends up in your memory image by using an
appropiate linker script.
(3) Make sure that the section is relocated to a fixed location,
representing the offset in the address space.
(4) Add startup code to copy the content of the executable to the
address space.

At least for an ELF target or comparable object format, you don't need
special assembler support.

Joerg