Named Address Spaces ?

[Gobaan_Raveendran at amis dot com]

Hello,

I am a student working on a porting gcc to an embedded architecture, and
one of the key requirements for this is named address space support as
stated in the Embedded C standard.

Fortunately, Ben Elliston of IBM recently implemented these for the SPU in
two steps, as stated in the following two patch requests:
      http://gcc.gnu.org/ml/gcc-patches/2008-08/msg01361.html - Front-End
link [at the time of this writing attempting to view his original post
causes a file not found error]
      http://gcc.gnu.org/ml/gcc-patches/2008-08/msg01479.html  - Back end
link

I have three questions:
1. Using ViewVC we can look at the branch and most files are documented
with "?Merge with mainline.". So I checked out the trunk, but the changes
were not present. I am wondering if I did
something wrong, and there is a later snapshot (that may be stable?) that I
should have checked out that does have these changes integrated.

2. As I am new to developing/reading gcc code, and I have been reading a
lot of the documentation, I was wondering if there is any standard place
for documentation on retargetting gcc for different architectures,
specifically if there would be a place where it would tell me how to
implemented additional address spaces, as well as what conflicting
qualifier checking is implemented. Currently I am reading
http://en.wikibooks.org/wiki/GNU_C_Compiler_Internals and I do not expect
named address spaces to be documented much because they have been in gcc
for about three weeks, but any help will be useful.

3. Since the trunk did not work, I checked out the named_address_space
branch, built it, newlib, binutils etc. and eventually got it compiling SPU
code. However, it crashes on the following code
that I modified from an IBM tutorial.

/* pointer in PPU address space pointing to PPU address space */
extern __ea int * __ea p;
int main(unsigneng long speid, unsigned long long argp, unsigned long long
envp)
{?? ?int x = * ((int *) p);?? ?return x;
}

Causes:
spu-gcc -Wall -otest.o test.c
/usr/local/lib/gcc/spu/4.4.0/libgcc_cachemgr.a(cachemgr.o): In function
`__cache_miss':
PATH-TO-GCC-SOURCE/gcc/config/spu/cachemgr.c:301: undefined reference to
`__ea_local_store'
PATH-TO-GCC-SOURCE/gcc/config/spu/cachemgr.c:305: undefined reference to
`__ea_local_store'
collect2: ld returned 1 exit status

This is likely because it cannot determine if the __ea pointer is in SPU
memory, but I was just wondering if this reveals a deeper error, as it is
detecting that a variable in gcc's source is undefined, in which case maybe
I should submit this to the gcc-patch/test suite. Then again, it could be
that I am using an unstable version, I did something else wrong or as long
as the error is detected it doesn't matter.

Thank You For Your Time.

Gobaan Raveendran