Date: Wed, 9 Sep 1998 18:41:17 +0300 (IDT)
From: Eli Zaretskii <eliz AT is DOT elta DOT co DOT il>
To: Tal Lavi <ranla AT post DOT tau DOT ac DOT il>
cc: djgpp AT delorie DOT com
Subject: Re: MAJOR slowdowns in translating TP7 gfx code to DJGPP2: Suplement 2
In-Reply-To: <35F6CB8F.1772@post.tau.ac.il>
Message-ID: <Pine.SUN.3.91.980909182846.2901F-100000@is>
MIME-Version: 1.0
Content-Type: TEXT/PLAIN; charset=US-ASCII
Precedence: bulk


On Wed, 9 Sep 1998, Tal Lavi wrote:

> 	short_div_t short_div(unsigned short numberator, 
>                               unsigned short denomonator)
> 	{
> 	  short_div_t short_div;
> 	  __asm__ __volatile__("
> 	    divw %%bx"
> 	  :"=a"(short_div.quot),"=d"(short_div.rem)
> 	  :"a"(numberator),"d"((unsigned short)0),"b"(denomonator)
> 	  :"ax","bx","dx"
> 	  );
> 	  return(short_div);
> 	}

Assuming this works (sorry, don't have enough time to have a good look), 
here are some comments:

	1) Why do you use word-size arguments and DIVW opcode?  Going to 
shorts will slow down your program considerably, since switching to 
16-bit operands in an otherwise 32-bit program requires a size-override 
prefix which causes the code to run at half the speed.  Maybe the fact 
that you use shorts in your program is the real reason behind your 
slow-downs?

	I suggest to work with int's, not short's.

	2) The above code doesn't handle the case where the second 
argument is negative.  I don't know whether your program needs that.

> Can I #define the inline asm command, instead of calling a real
> function?

Just define it with the __inline__ qualifier, and make sure its 
definition is *before* it is called in the program, and GCC will inline 
it for you thus avoiding the overhead of the call.

> Bear in mind that I'm using _farns* functions on a double page LFB,
> it should be FASSST!

Did you try to use _farnspokel instead of _farnspokew?

> But when I try to compile the file in order to profile it:
> 
>  	gcc -o ray.exe ray.c vesa64.c -O2 -ffast-math -pg
> 
> The running of the .exe is hanging my system!

This is explained in the FAQ (section 13.2): stock v2.01 library has a 
bug in one of its functions that is linked into your program when you use 
-pg.  Get a patched library from the URL below and relink, and the 
problem should go away:

   WWW: http://www.cartsys.com/eldredge/djgpp-patches.html
   ftp: ftp://www-leedr.arme.cornell.edu/pub/djgpp-patches/).  Be sure to 
read the file `ftp.message.txt' when using the FTP server above.  The 
file `README' explains how the patched library can be downloaded and used. 

> create a gmon.out file.
> here it is, does it make sense?!?

No, it doesn't.  The FAQ explains that gmon.out is written by a function 
registered via `atexit'.  If you abort the profiled program with Ctrl-C, 
the normal exit code is bypassed, and gmon.out info is not written.