On Fri, Jan 27, 2017 at 08:41:59AM +0100, Rainer Orth wrote:
> > On 01/24/2017 02:37 AM, Markus Trippelsdorf wrote:
> >> MPFR_RNDx was introduced in MPFR 3.0.0. Since the minimal version that
> >> gcc checks for is 2.4.0, this leads to a build failure.
> >>
> >> The fix is straightforward.
> >>
> >> Tested on x86_64-pc-linux-gnu. Committed to trunk as obvious.
> >>
> >> * gimple-ssa-sprintf.c (format_floating): Change MPFR_RNDx to
> >> GMP_RNDx for compatibility.
> >
> > This was changed once before for this reason (in r240350). I forgot
> > all about it and put it back in my latest patch for some reason. I
> > don't remember why exactly but I suspect I might have been trying to
> > overcome some MPFR oddity. It might help to #undef the MPFR_RNDx
> > macros after including <mpfr.h> to avoid this in the future. In any
> > event, thanks for the fix.
>
> we certainly should do something like this: your latest patch introduced
> the same failure again:
>
> /vol/gcc/src/hg/trunk/local/gcc/gimple-ssa-sprintf.c: In function 'long long
> unsigned int {anonymous}::format_floating_max(tree, char, long long int)':
> /vol/gcc/src/hg/trunk/local/gcc/gimple-ssa-sprintf.c:1501:27: error:
> 'MPFR_RNDN' was not declared in this scope
> mpfr_from_real (x, &rv, MPFR_RNDN);
> ^
> Makefile:1098: recipe for target 'gimple-ssa-sprintf.o' failed
> make[3]: *** [gimple-ssa-sprintf.o] Error 1
So what about poisoning MPFR_RND{N,D,Z,U} then?
With mpfr 3.1.5 it seems to work, both when using system gcc (6.3.1)
and when building in stage3. I don't have mpfr 2.4.* around and apparently
it is incompatible with gmp 6.1.x, so I'd have to build both gmp and mpfr
(and likely mpc too). Can you just quickly try this if it works for you?
Newer mpfr.h has:
typedef enum {
MPFR_RNDN=0, /* round to nearest, with ties to even */
MPFR_RNDZ, /* round toward zero */
MPFR_RNDU, /* round toward +Inf */
MPFR_RNDD, /* round toward -Inf */
MPFR_RNDA, /* round away from zero */
MPFR_RNDF, /* faithful rounding (not implemented yet) */
MPFR_RNDNA=-1 /* round to nearest, with ties away from zero (mpfr_round) */
} mpfr_rnd_t;
/* kept for compatibility with MPFR 2.4.x and before */
#define GMP_RNDN MPFR_RNDN
#define GMP_RNDZ MPFR_RNDZ
#define GMP_RNDU MPFR_RNDU
#define GMP_RNDD MPFR_RNDD
so the question is if it will work with compilers that pretend to be GCC
(clang++, icc). I've just tried it with both, clang++ 3.8.0 compiles it
fine (with noisy warnings about various non-issues), icpc 16.0.2 does error
out on a couple of unrelated things, but doesn't mention the *RND* stuff.
icpc claims to be GCC 6.3.1 (maybe it always pretends to be the same version
as system gcc?) and errors e.g. on:
../../gcc/hash-table.h(956): error: declaration is incompatible with function
template "void hash_table<Descriptor,
Allocator>::traverse_noresize<Argument,Callback>(Argument)" (declared at line
444)
hash_table<Descriptor, Allocator>::traverse_noresize (Argument argument)
../../gcc/gimple.h(62): error: identifier "__builtin_FILE" is undefined
../../gcc/gimple.h(63): error: identifier "__builtin_LINE" is undefined
../../gcc/gimple.h(64): error: identifier "__builtin_FUNCTION" is undefined
2017-01-27 Jakub Jelinek <ja...@redhat.com>
* realmpfr.h: Poison MPFR_RND{N,Z,U,D}.
* gimple-ssa-sprintf.c (format_floating_max): Use GMP_RNDN
instead of MPFR_RNDN.
--- gcc/realmpfr.h.jj 2017-01-01 12:45:39.000000000 +0100
+++ gcc/realmpfr.h 2017-01-27 09:04:56.082042138 +0100
@@ -32,5 +32,10 @@ extern void real_from_mpfr (REAL_VALUE_T
const real_format *, mp_rnd_t);
extern void mpfr_from_real (mpfr_ptr, const REAL_VALUE_TYPE *, mp_rnd_t);
-#endif /* ! GCC_REALGMP_H */
+#if (GCC_VERSION >= 3000)
+ /* For compatibility with mpfr 2.4 and earlier, we want to only use
+ GMP_RND*. */
+ #pragma GCC poison MPFR_RNDN MPFR_RNDZ MPFR_RNDU MPFR_RNDD
+#endif
+#endif /* ! GCC_REALGMP_H */
--- gcc/gimple-ssa-sprintf.c.jj 2017-01-27 08:49:03.000000000 +0100
+++ gcc/gimple-ssa-sprintf.c 2017-01-27 08:53:52.911721864 +0100
@@ -1498,7 +1498,7 @@ format_floating_max (tree type, char spe
round-to-nearest mode. */
mpfr_t x;
mpfr_init2 (x, rfmt->p);
- mpfr_from_real (x, &rv, MPFR_RNDN);
+ mpfr_from_real (x, &rv, GMP_RNDN);
/* Return a value one greater to account for the leading minus sign. */
return 1 + get_mpfr_format_length (x, "", prec, spec, 'D');
Jakub
