On Mon, Dec 16, 2019 at 8:29 PM Martin Maechler <maech...@stat.math.ethz.ch> wrote: > > Thank you, Dirk, for keeping me in the loop. I'm happy if you test to > see if the change works. > > As a matter of fact, I think a slightly different patch should be used > rather than the one proposed. > Given what I've learned in the meantime, I think I'd rather use > > Index: src/main/arithmetic.c > =================================================================== > --- src/main/arithmetic.c (Revision 77566) > +++ src/main/arithmetic.c (Arbeitskopie) > @@ -176,8 +176,12 @@ > #endif > } > > + > #if HAVE_LONG_DOUBLE && (SIZEOF_LONG_DOUBLE > SIZEOF_DOUBLE) > -static LDOUBLE q_1_eps = 1 / LDBL_EPSILON; > +# ifndef __PPC64__ > +static // PowerPC 64 (when gcc has -mlong-double-128) breaks when > 'static' (gcc bug, there) > +# endif > + LDOUBLE q_1_eps = 1 / LDBL_EPSILON; > #else > static double q_1_eps = 1 / DBL_EPSILON; > #endif > > > which would keep PowerPC 64 to use the high precision epsilon here, > rather than the one to use when there is no long double present. > Martin
Oops, no! Tomas the expert has explained me why that would not be good enough, instead recommending a macro (which would be "constant folded" i.e., computed at compile time were available); If I conditionalize that .. in order to be back compatible outside of PPC_64, I'd get Index: src/main/arithmetic.c =================================================================== --- src/main/arithmetic.c (Revision 77583) +++ src/main/arithmetic.c (Arbeitskopie) @@ -176,8 +176,14 @@ #endif } + #if HAVE_LONG_DOUBLE && (SIZEOF_LONG_DOUBLE > SIZEOF_DOUBLE) +# ifdef __PPC64__ + // PowerPC 64 (when gcc has -mlong-double-128) breaks here ... +# define q_1_eps (1 / LDBL_EPSILON) +# else static LDOUBLE q_1_eps = 1 / LDBL_EPSILON; +# endif #else static double q_1_eps = 1 / DBL_EPSILON; #endif ------------- The above may work correctly (I tried a version where I replaced __PPC64__ by 1 such as to test with the gcc 9.2 (on my architecture). I'd be glad if you could confirm for powerPC 64 .. then I think we could commit it. Martin
