80 bit floating point support for i960.
Mon Jan 25 08:52:00 GMT 1999
Stephen L Moshier wrote:
> > The patch permits 80-bit floating point numbers for i960.
> I think it may be impractical to select this at run time (i.e. not worth
> the amount of effort it would require). The size of long double is a
> compile time constant macro, and the definition of a REAL_VALUE_TYPE
> depends on that size. You can build the compiler one way or the other but
> not both ways at once.
> To permit building either a 96-bit or 128-bit long double version for i960
> you would need to check the size in i960_output_long_double.
There was one reason of this. Usage of plumhall for testing the compiler.
Generally plumhall test contains many subtests and any error in one subtest
prevents running the remaining subtests. The plumhall does not work with 80
bit long double.
I think that such infrastructure (possibility to change long double
implementation in a compiler) would be useful not only for i960. But you are
right it would require efforts for tailoring real.[ch] for this.
I found inaccurate description of LONG_DOUBLE_TYPE_SIZE in tm.texi as simply
C expression. May be therefore we missed that it is used as preprocessor
constant in real.[ch]
More information about the Gcc-patches