Re: Does/will gcc support REAL(16) in Fortran?

[Jim Wilson <wilson at specifixinc dot com>]

H. J. Lu wrote:
> BTW, why is __float128 enabled only for MMX on ia32? Does __float128
> use MMX?

I suspect practical considerations.  You can't fit __float128 in the x86 
FP stack regs, though you can fit them into SSE regs and AMD64 64-bit 
integer regs.

So suppose you allow __float128 on an ordinary x86, you have to put them 
them in the integer regs, and you need 2 integer regs for each value, 
and you will need at least 3 register pairs to do anything interesting, 
and you immediately have a register allocation problem, as you don't 
have enough regs.  We can get around this by keeping them in memory, but 
that will complicate the implementation, and this support would need to 
be written by someone.

There may also be AMD64 issues here.  This code was originally added for 
AMD64 which uses 16-byte 80-bit precision FP values, and then was 
cleaned up a bit after Zack's patch.  AMD64 doesn't have the register 
allocation problem that x86 does.
--
Jim Wilson, GNU Tools Support, http://www.SpecifixInc.com