[Top] [All Lists]

Re: Re: SNaN & QNaN on mips

To: "H . J . Lu" <>
Subject: Re: Re: SNaN & QNaN on mips
From: Zhang Fuxin <>
Date: Tue, 5 Feb 2002 17:48:18 +0800
Cc: "" <>

  I am sorry but it seems i can't fix this without ugly changes.
since i am not familiar with gcc code, i decide to leave it to you,
but provide some information instead.

  In gcc there are 3 spaces where the NaN handling is assumed the 
Intel way.

   1. gcc/real.c (the most important one)
       here the author seems to have known the NaN pattern problem,so
     he leaves a interface macro for defining non intel NaN patterns:
     (comment of function "make_nan()",at about line 6219)

/* Output a binary NaN bit pattern in the target machine's format.  */

/* If special NaN bit patterns are required, define them in tm.h
   as arrays of unsigned 16-bit shorts.  Otherwise, use the default
   patterns here.  */

  I have read through this file and decided that the follow defined should
be enough for mips:
/* NaN pattern,mips QNAN & SNAN is different from intel's 
 * DFMODE_NAN and SFMODE_NAN is used in real.c */
#define DFMODE_NAN \
        unsigned short DFbignan[4] = {0x7ff7, 0xffff, 0xffff, 0xffff}; \
        unsigned short DFlittlenan[4] = {0xffff, 0xffff, 0xffff, 0xfff7}
#define SFMODE_NAN \
        unsigned short SFbignan[2] = {0x7fbf, 0xffff}; \
        unsigned short SFlittlenan[2] = {0xffff, 0xffbf}

   But the problem is where to put them:(. Obviously it is target specified
definitions and should be in config/mips/. Documents say tm.h is a symbol
link and included in config.h,but it is no long true.If i add them to xm-mips.h
then for native compilation it is ok but it fails for cross-compile.

     There is a hardcoded QNaN used around line 477:
      nan = gen_lowpart (SFmode, GEN_INT (0x7fc00000));
     I sugest defining a macro QNAN_HAS_1ST_FRACBIT_CLEARED for mips and change
     it to,just don't know where to put it:
         nan = gen_lowpart (SFmode, GEN_INT (0x7fc00000));
         nan = gen_lowpart (SFmode, GEN_INT (0x7fbfffff));

   3. config/fp-bit.c
      this is for machine having no fpu hardware.
      again i susgest define QNAN_HAS_1ST_FRACBIT_CLEARED and then apply this 

< #else
<         fraction &= ~QUIET_NAN;
< #endif
< #else
<         if (!(fraction & QUIET_NAN))
< #endif
在 2002-02-03 22:54:00 you wrote:
>On Mon, Feb 04, 2002 at 02:22:48PM +0800, Zhang Fuxin wrote:
>> hi,
>> Gcc (2.96 20000731,H.J.LU's rh port for mips) think 0x7fc00000 is QNaN and 
>> optimize 0.0/0.0 as 0x7fc00000 for single precision ops,while for my cpu
>> (maybe most mips cpu) is a SNaN. R4k user's manual and "See Mips Run" both
>>  say so.And experiments confirm this.
>> Should we correct it?
>Yes. Do you have a patch?

            Zhang Fuxin

<Prev in Thread] Current Thread [Next in Thread>