PR optimization/4490 (-m128bit-long-double on ia32)

[Jan Hubicka]

> Hi!
> 
> optimization/4490:
> 
> /* { dg-do compile } */
> /* { dg-options "-O2" } */
> /* { dg-options "-m128bit-long-double -O2" { target i?86-*-* } } */
> 
> double f (unsigned long long x)
> {
>   double d = x;
>   return d;
> }
> 
> ICEs because optabs for the d = x assignment (which cannot be done in
> DFmode) looks up a wider mode. Wider mode is XFmode and IA-32, if
> !TARGET_64BIT, has XFmode patterns, so optabs attempts to use them
> (which is what -m128bit-long-double should prevent IMHO).
> But as long double is TFmode in this case, type_for_mode cannot find any
> type for mode XFmode and soon after it segfaults in IA-32 CONSTANT_ALIGNMENT
> implementation.
> 
> Shouldn't all IA-32 XFmode patterns be banned if TARGET_128BIT_LONG_DOUBLE
> (ie. replace all !TARGET_64BIT in XFmode patterns with
> !TARGET_128BIT_LONG_DOUBLE)?

Probably yes, in the current shape.  Originally I've sent some patches letting
XFmode and TFmode floats to coexist, as my goal has been to let people use
__attribyte__ ((TFmode)) type and keep long double as it is when they need ABI
compatibility and fast floats, but they didn't get in.

I can make patch sometime at monday or tuesday, unless you will be faster.

Honza
> 
> 	Jakub