Re: Data representation and cross compiles

[Gabriel Dos Reis <gdr at integrable-solutions dot net>]

Daniel Jacobowitz <drow@mvista.com> writes:

| On Wed, Aug 28, 2002 at 02:25:21PM +0200, Gabriel Dos Reis wrote:
| > 
| > I'm investigating the V3 build failure reported by the regression
| > tester.  I think it doesn't occur on natives, it occurs on
| > cross-compilers only.  
| > 
| > Suppose I configure for a cross-compiler with
| > 
| >    --build=i686-pc-linux
| >    --host=i686-pc-linux
| >    --target=mips-elf
| > 
| > Furthermore, assume:
| > 
| >   sizeof (host_int) == 4            | sizeof (host_long_double) == 12
| >   sizeof (target_int) == 4          | sizeof (target_long_double) == 16
| > 
| > 
| > In the *library* source code, I have something like :
| > 
| >    sizeof (long double) / sizeof (int)
| > 
| > I think the above expression should evaluate to 4 -- since the library
| > is being built for mips-elf -- but the compiler seems to evaluate
| > it to 3 (apparently using i686-pc-linux information).
| > 
| > Thoughts?
| 
| Well, I don't like your assumptions very much :)  mips.h:
| 
| /* A C expression for the size in bits of the type `long double' on
|    the target machine.  If you don't define this, the default is two
|    words.  */
| #define LONG_DOUBLE_TYPE_SIZE 64
| 
| Long doubles are twelve bytes on i386 and eight bytes on MIPS.  If
| you're seeing some sort of size mismatch, I'd bet that was it.

Hmm, I tried a cross-compile and the size I got from the preprocessed
output are those I used in the above description :-)  

More specifically, all "long double" settings I got from the
preprocessed source are as if long double was 128 bits wide on
misp-elf.
That means, as you say, that somewhere some (duplicated) configuration
settings don't match.

Wait.  Indeed, I don't see any thing under libstdc++-v3/config/cpu/mips/
that sets the sizes.  Therefore, it defaults -- according to the
comment

  // These values can be overridden in the target configuration file.
  // The default values are appropriate for many 32-bit targets.

to 128 as per libstdc++-v3/include/std/std_limits.h:115:

   #ifndef __glibcpp_long_double_bits
   #  define __glibcpp_long_double_bits 128
   #endif

Bummer.

OK, thanks for your comment.  It was very helfpul. I'll try to
minimise the configuration settings duplication by having
gcc/c-common.c:cpp_define_data_format() define the appropriate bits.

Thanks,

-- Gaby