[WIP] ARM: Add some inline assembly for bigint

[redstar]

Implements the internal interface in ARM assembly.

[redstar]

@smolt @joakim-noah Is this useful? For sure I need to add loop unrolling.

[joakim-noah]

Looks like it might be, but its release tests assert at line 90 on Android/ARM and it causes test failures in other dependent modules like std.bigint too (at line 236 in ltsmaster).

[redstar]

@joakim-noah Thanks for trying. I try to correct the failures. (Didn't see them yesterday on my arm box but I made some more changes this morning.)

[JohanEngelen]

@redstar Just a suggestion. Isn't it possible to do this in IR? Perhaps then you get the optimization for free. Or can LLVM also optimize the asm?

[redstar]

IR is no better than D. With D and IR you have to use the overflow intrinsics which leads to horrible codegen. I tried it with x86_64 and it is only marginable better than the D version.

[smolt]

Any opportunity to write assembly code should be taken advantage of 😄

I am on vacation, would like to spend time on on this when I am worn out.

[JohanEngelen]

Any opportunity to write assembly code should be taken advantage of

Haha, indeed!
Thanks @redstar for increasing my ammunition in discussions about asm :)

Enjoy the snow @smolt .

[redstar]

The implemented functions now pass all tests, including -O3 -release. Currently measured speedup: multibyteAddSub!('+'): 2.2, multibyteMul: 1.4 and multibyteMulAdd!('+'): 1.2. There is still room for improvement.
To measure the timings I compiled the binary with: ldc2 -release -O3 -singleobj -unittest -d-version=timings biguintarm.d biguintnoasm.d.

[joakim-noah]

All tests now pass on Android/ARM when compiled in release mode, the new module's tests take 6 seconds to run.

[redstar]

@jpf91 Maybe this is interesting for gdc, too?

[jpf91]

That's certainly interesting, thanks for the ping :-)

Is the LDC inline ASM string and the constraint string GCC compatible? Then we could probably implement the __asm interface in gdc to provide a common inline ASM syntax for LDC and GDC. ping @ibuclaw

[redstar]

@jpf91 Unfortunately the constraint string is not GCC compatible. (The content is but not the syntax.) Maybe we can agree on some standard here?

[ibuclaw]

Does the constraints string match what the LLVM backend expects? Or is it specially handled in LDC?

[dnadlinger]

It is a backend thing.

[ibuclaw]

Then we should continue doing what is most natural for our respective backends. :-)

[redstar]

Merged in master now.