templates and assembler

[Don Clugston]

manu wrote:
> James Dunne says...
>> manu wrote:
>>> Hi there!
>>> I´m new to D and quite surprised about all the nice features it provides. So I
>>> started to write some test apps and also tried to write some templates. It all
>>> worked perfect until I wanted to use template identifiers in asm blocks.
>>>
>>> consider the following:
>>>
>>> /* file - main.d **************************************/
>>> 01: template sine(REAL) // where REAL should be one of float,double,real 
>>> 02: {
>>> 03:     float sine(REAL x)
>>> 04:	{
>>> 05:	    asm 
>>> 06:	    {
>>> 07:	        naked		    ; // <- make it naked to get maximum speed
>>> 08:		fld REAL ptr[ESP+4] ; // <- critical opcode
>>> 09:		fsin		    ;
>>> 10:		ret		    ;
>>> 11:	    }
>>> 12:     }
>>> 13: }
>>> 14: alias sine!(float) sine_float; // instantiation
>>> / *****************************************************/
>>> An error occurs when typing the last line, due to the instantiation. It says:
>>>
>>> main.d(8): end of instruction
>>>
>>> I´ve figured out that it comes because of the REAL in the asm block. I also
>>> tried this for line 08:
>>>
>>> 08:   fld x
>>>
>>> Though this compiles there remains an error at runtime: the program assumes x to
>>> be at a different position on the stack and so just a dump of stack memory is
>>> loaded into the FPU. May this an unfixed bug of the compiler(I´m using dmd)? 
>>>
>>> I would be very thankful to everybody who helps me fixing that problem!
>>>
>>>
>> Removing 'naked' and 'ret' instructions fixes the problem.  'real 
>> ptr[ESP+4]' also didn't work, so use 'fld x' instead.  I think you were 
>> trying to load the pointer value into the floating point register, which 
>> doesn't quite work. :)
>>
>> -- 
>> Regards,
>> James Dunne
> 
> First thank´s for your help, but 'real ptr[ESP+4]' works, there is no pointer
> value
> loaded into the FPU. 'REAL ptr[ESP+4]' is the code that doesn´t work(CAPITAL
> letters!) 
> cause the compiler(dmd) can´t handle template arguments in asm blocks.

Actually REAL is an undocumented reserved word in D assembler! It means 
the same as TBYTE in other assemblers. Use a different name!

Also, I don't think that 'ret' on its own works correctly.

Here's (untemplated) code that works.
real sin(real x)
{
     asm {
         naked;
         fld real ptr [ESP+4];
         fsin;
         ret x.sizeof + x.alignof;
     }
}


> 
> And sure using 'fld x' works but only if I don´t use 'naked', as you
> said. Otherwise the compiler doesn´t load 'x' from the proper location. Probably
> 
> a stack frame is assumed and so it sure fails.
> 
> So I think there should be two new features implemented:
> 
> 1. compiler knows the parameters offsets even if 'naked' is used
> 2. template arguments can be used inside asm blocks
> 
> (Please correct me if one of them is already available.)