Re: [Tinycc-devel] uint64_t/int64_t => double broken on ARM

[Thomas Preud'homme]

Le samedi 1 février 2014, 14:41:48 Daniel Glöckner a écrit :
> Hello Thomas,
> 
> On Sat, Feb 01, 2014 at 07:04:24PM +0800, Thomas Preud'homme wrote:
> > Got a bit of time (and more importantly internet access) so it's now
> > fixed in mob. Thanks for reporting.
> 
> your patch makes
> 
> float f(unsigned long long x)
> {
>       return x;
> }
> 
> compile to
> 
>   ...
>   bl      __aeabi_ul2f
>   vmov    s0, r0
>   vmov.32 d0[0], r0
>   vmov.32 d0[1], r1
>   ...
> 
> with softfp ABI and it generates the superfluous move from r1 with hard
> float ABI. Maybe the cleanest solution is to temporarily change the ABI
> when calling these functions.

Right, I confused the size of the int with the size of the float the int is 
converted to. It's easy to add a switch to test between float and double.

As to changing the ABI, there was actually something for it (see 
floats_in_core_regs function). Strange indeed that it doesn't work since the 
long long -> float conversion function are indeed considered for the special 
ABI. I'll check more closely. Thanks for your attention.

Thomas