Re: [PATCH] Move -mlong-double-128 away from platform-agnostic files

[Joseph Myers <joseph at codesourcery dot com>]

On Mon, 28 May 2018, Tulio Magno Quites Machado Filho wrote:

> The compiler flag -mlong-double-128 is platform-specific and shouldn't
> be used in platform-agnostic directories, such as ldbl-128ibm and
> ldbl-opt.

Theoretically, as a -m option, it's platform-specific.  But in practice, 
I'm not sure that's really the case - certainly all existing platforms 
using ldbl-opt support -mlong-double-128 and -mlong-double-64.  Indeed, 
sysdeps/ieee754/ldbl-opt/configure.ac tests that -mlong-double-128 is 
supported, and that __LONG_DOUBLE_128__ gets defined by the compiler; the 
ldbl-opt bits/long-double.h relies on that __LONG_DOUBLE_128__ macro; the 
ldbl-opt Makefile relies on -mlong-double-64 being available to build a 
test with that option (unfortunately there is only the one test for 
-mlong-double-64 support; new ieee128 interfaces should have more thorough 
test coverage added no later than when the interfaces are enabled for 
powerpc64le).

> This was not a problem because all the architectures implying these
> directories had to use -mlong-double-128 to set their long double type.
> As powerpc64le is changing its default long double type to be binary
> compatible with IEEE 128-bit floating point, this will have to change.

But as far as I know the option is still supported for powerpc64le.  And 
indeed it works fine together with -mabi=ieeelongdouble, and does not 
change the format away from IEEE, and -mlong-double-64 also works together 
with -mabi=ieeelongdouble (overriding it to make long double have the same 
ABI as double).

So what is the actual problem you're trying to fix here?  As far as I can 
tell, the existing code should work just fine for powerpc64le in the 
context of a changed long double default - you may need to *add* options 
somewhere else to ensure -mabi=ibmlongdouble or -mabi=ieeelongdouble are 
used to compile particular files (possibly together with 
-mno-gnu-attribute to avoid link errors for incompatible ABIs), but I 
don't see anything that would require -mlong-double-128 to be avoided, and 
thus don't see any need to duplicate the code using that option for 
different architectures.

-- 
Joseph S. Myers
joseph@codesourcery.com