On 6 February 2012 09:40, Johannes Pfau <nospam@example.com> wrote:
Am Mon, 6 Feb 2012 03:07:31 +0200
schrieb Manu <turkeyman@gmail.com>:

> On 6 February 2012 02:25, Manu <turkeyman@gmail.com> wrote:
>
> > On 6 February 2012 00:04, Johannes Pfau <nospam@example.com> wrote:
> >
> >> Am Sun, 5 Feb 2012 18:04:12 +0100
> >> schrieb Johannes Pfau <nospam@example.com>:
> >>
> >> > I will probably need some more time to get this working...
> >> >
> >>
> >> I have some good news:
> >> http://www.mediafire.com/?107we120sh3xx
> >>
> >> I fixed that problem and then the whole build worked fine. I'll
> >> post build instructions soon, but the binaries are ready. I only
> >> did a simple gdc -c test.d to check the compiler, but it seems to
> >> work.
> >>
> >> Linking against druntime fails, as it uses functions which are not
> >> available on Android (backtrace, signal stuff).
> >>
> >> I also built a simple hello world on linux (printf, no runtime)
> >> and ran it on my android phone, and it worked!
> >>
> >> In case you haven't used GDC without runtime before, a short
> >> introduction:
> >>
> >> * use gdc -nophoboslib to make gdc not link against phobos (and
> >> afaik, druntime)
> >> * theres also -nostdlib in case you need it
> >> * complex code may require -fno-section-anchors because of bug #120
> >> * You'll get an error about a missing _Dmodule_ref symbol. That
> >> symbol is used by module constructors and not generated by gdc if
> >>  -nophoboslib was passed. As long as you don't run the module
> >>  constructors, you can add a fake _Dmodule_ref in a .c file:
> >>
> >> ------------
> >> void* _Dmodule_ref;
> >> ------------
> >>
> >> * The compiler defines version(Android)
> >>
> >> Here's my hello world:
> >> ------------
> >> version(Android)
> >> {
> >>    pragma(msg, "Hello Android!");
> >> }
> >>
> >> extern(C)
> >> {
> >>    int printf(in char* format, ...);
> >> }
> >>
> >> extern(C) void main()
> >> {
> >>    printf("Hello, %s!\n".ptr, "Android".ptr);
> >> }
> >> ------------
> >>
> >> compile the _Dmodule_ref into hack.o, then use
> >> gdc -nophoboslib hello.d hack.o
> >>
> >
> > Amazing! You sir, are a genius!
> > I'm gonna have to have some fun with that after work tomorrow :)
> >
> > I reckon this binary distro should be put on d-p-l somewhere
> > obvious.
> >
>
> I just tried with -mfpu=neon, which should allow GCC to generate neon
> opcodes for the simd stuff, but it didn't seem to want to do that. In
> fact it generates really horrible code where it CALLS an intrinsic
> for each float in the vector...
> Any idea why this wouldn't work out of the box?

The android toolchain configures gcc with: += --with-float=soft
--with-fpu=vfp --with-arch=armv5te

I could try to build an optimized build, but then you'll have to tell
me what processor you're using. Cortex A8?

I'm not sure if you can override that at runtime. You could probably
try:
-mfloat-abi=hard -mcpu=cortex-a8 -march=armv7-a -mfpu=neon

That's interesting. I seem to recall in a recent NDK noticing that they had changed to armv7 by default... but looking at the R6 NDK, it looks like it's configured basically identical to your toolchain.
I must just be missing a whole function of compile options, but I can't see any option in the standard build scripts to enable hardware fpu :/
The options I see that are supplied are: -ffast-math -mfloat-abi=softfp -march=armv7-a

I imagine all android 2+ devices have at least armv7's with a hardware fpu...

Interesting that NEON is not supported on all hardware. Annoying to do runtime detection to take advantage of stuff like that >_<
 
> I wanted to add ARM support to my std.simd work here as a better
> proof of concept.
>
> Iain: was there anything particularly special you needed to do to
> hook the x86 SSE stuff to GDC which would need to be duplicated for
> ARM? gcc.builtins doesn't seen to have any ARM intrinsics in there
> either...
>