64-bit long doubles on i386?

64-bit long doubles on i386?

[Paul N. Hilfinger]

What would need to be done to modify GCC so that it uses only <=64-bit
floating-point formats (so that long double and double have
essentially the same semantics, the extended precision is not used for
intermediate results)?  I've tried the obvious modifications to i386.h and
float-i386.h: setting

  #define LONG_DOUBLE_TYPE_SIZE 64
  #define MAX_LONG_DOUBLE_TYPE_SIZE 64
  #define INTEL_EXTENDED_IEEE_FORMAT 0
  #define LIBGCC2_LONG_DOUBLE_TYPE_SIZE 64
  #define INTEL_EXTENDED_IEEE_FORMAT 0

in i386.h and setting float-i386.h to take the LDBL constants from the 
corresponding DBL constants.  This, however, causes the bootstrap to 
segfault compiling __fixunsdfSI in libgcc2.c.  I was able to prevent the
segfault by kludging expand_float in optabs.c so that it doesn't try to
use XF and TF modes (which are still there, just not used for long double).
This, however, introduces test-suite regressions (unsurprisingly in 
gcc.c-torture/execute/conversion.c)

Why am I interested?  Well, although I've been experimenting with this on 
Linux, the eventual "client" is Interix, where there is sometimes a "legacy" 
problem of compatibility with (ugh) VC.  

If anyone has suggestions, I'd appreciate hearing from you.

Paul Hilfinger
ACT Inc.

Re: 64-bit long doubles on i386?

[Tim Prince]

On Wednesday 10 April 2002 04:00, Paul N. Hilfinger wrote:
> What would need to be done to modify GCC so that it uses only <=64-bit
> floating-point formats (so that long double and double have
> essentially the same semantics, the extended precision is not used for
> intermediate results)?  I've tried the obvious modifications to i386.h and
> float-i386.h: setting
>
>   #define LONG_DOUBLE_TYPE_SIZE 64
>   #define MAX_LONG_DOUBLE_TYPE_SIZE 64
>   #define INTEL_EXTENDED_IEEE_FORMAT 0
>   #define LIBGCC2_LONG_DOUBLE_TYPE_SIZE 64
>   #define INTEL_EXTENDED_IEEE_FORMAT 0
>
..............
> Why am I interested?  Well, although I've been experimenting with this on
> Linux, the eventual "client" is Interix, where there is sometimes a
> "legacy" problem of compatibility with (ugh) VC.
>
............
Why don't you simply set 53-bit precision mode at run time, as currentVC 
does, if you want similar operation?
-- 
Tim Prince

Re: 64-bit long doubles on i386?

[Paul N. Hilfinger]

>Why don't you simply set 53-bit precision mode at run time, as currentVC 
>does, if you want similar operation?

That's a reasonable suggestion, although it still leaves the possible issue of
data formats, which would be of interest when mixing legacy libraries with
compiled code.   Perhaps, though, as you say, it will suffice for the 
actual needs in this case.  

Paul 

Re: 64-bit long doubles on i386?

[Stephen L Moshier]

> in i386.h and setting float-i386.h to take the LDBL constants from the
> corresponding DBL constants.  This, however, causes the bootstrap to
> segfault compiling __fixunsdfSI in libgcc2.c.  I was able to prevent the
> segfault by kludging expand_float in optabs.c so that it doesn't try to
> use XF and TF modes (which are still there, just not used for long
> double).
> This, however, introduces test-suite regressions (unsurprisingly in
> gcc.c-torture/execute/conversion.c)


I think a better way to ensure the compiler generates no XFmode code
is to turn off all the XFmode patterns in the md file.

i386-osfrose and i386-isc have 64-bit long doubles.  This configuration
has been broken and fixed (or papered over) repeatedly in the past.
I guess no one cares about these targets any more.