Date: Fri, 28 Feb 92 13:57 +1100
From: Bill Metzenthen <APM233M AT vaxc DOT cc DOT monash DOT edu DOT au>
Subject: Re: help required
To: djgpp AT sun DOT soe DOT clarkson DOT edu, dj AT ctron DOT com
Status: O

Dirk Zabel writes:

>Hi,
>could somebody who is lucky enough to have a REAL 387 try to compile
>this short program with "gcc -c -S test.c":
>
>------- cut here -------- test.c ---------------
>double q = 0.3e10;
>------- cut here -------------------------------
>
>If the generated assembly-language output does not contain
>the following bad constant, the 387-emulator has most probably another bug.


This bug is due to problems with rounding in emu387 (basically, there is
none!).

With the patches which I present below, you will get the following:

--------------------- snip, snip ------------------------
	.file	"bug.c"
gcc_compiled.:
.globl _q
.data
	.align 2,144
_q:
	.double 0d3.00000000000000047684e+09
--------------------- snip, snip ------------------------


Here is my patch:

--------------------- snip, snip ------------------------
*** old/rmov.cc	Thu May  9 22:35:54 1991
--- rmov.cc	Fri Feb 21 08:08:58 1992
***************
*** 266,273 ****
    }
    else
    {
!     l[0] = (s.sigl >> 11) | (s.sigh << 21);
!     l[1] = ((s.sigh >> 11) & 0xfffff) | (((s.exp-EXP_BIAS+1023) & 0x7ff) << 20);
    }
    if (s.sign)
      l[1] |= 0x80000000;
--- 266,300 ----
    }
    else
    {
!     if ( s.sigl & 0x400 )
!       {
!       if ( s.sigl >= 0xfffff800 )
!         {
!         /* the sigl part will -> 0, with a carry */
!         if ( s.sigh == 0xffffffff )
!           {
!           /* The sigh part overflows */
!           l[0] = 0;
!           /* ***** Checking for exp overflow not yet implemented **** */
!           l[1] = (((s.exp-EXP_BIAS+1023+1) & 0x7ff) << 20);
!           }
!           else
!           {
!           l[0] = (s.sigh+1) << 21;
!           l[1] = (((s.sigh+1) >> 11) & 0xfffff) | (((s.exp-EXP_BIAS+1023) & 0x7ff) << 20);
!           }
!         }
!         else
!         {
!         l[0] = ((s.sigl+0x800) >> 11) | (s.sigh << 21);
!         l[1] = ((s.sigh >> 11) & 0xfffff) | (((s.exp-EXP_BIAS+1023) & 0x7ff) << 20);
!         }
!       }
!       else
!       {
!       l[0] = (s.sigl >> 11) | (s.sigh << 21);
!       l[1] = ((s.sigh >> 11) & 0xfffff) | (((s.exp-EXP_BIAS+1023) & 0x7ff) << 20);
!       }
    }
    if (s.sign)
      l[1] |= 0x80000000;
--------------------- snip, snip ------------------------


Note that this only handles conversion from the internal 80
bits or so to "double". I think that even for "double", it is
not complete because I seem to recall that the 80387 (or is it
the IEEE specification) has a means of turning rounding on and
off, etc.

A question to DJ Delorie: Is someone really working on a replacement
for emu387?



Bill Metzenthen
apm233m AT vaxc DOT cc DOT monash DOT edu DOT au