MPIR's configure fails with GCC 5.x

[leif]

...
checking whether gcc-5.0 is gcc... yes
checking compiler gcc-5.0 -m64 ... no, long long reliability test 1
configure: error: could not find a working compiler, see config.log for details
Error configuring MPIR (with CFLAGS unset).

This test already failed with Clang (cf. #13948); now that GCC (5.x) defaults to STDC inlining (as opposed to GNUC inlining) semantics, we have to adapt the test (which is supposed to detect an unrelated bug in older GCC versions) again.

[leif]

The following patch to configure fixes the issue:

src/configure

@@ -5663 +5663 @@
 
 #if defined(__GNUC__) && !defined(__clang__)
 typedef unsigned long long t1;typedef t1*t2;
+#if defined(__GNUC_STDC_INLINE__)  /* e.g. GCC 5.x default */
+extern
+#endif
 __inline__ t1 e(t2 rp,t2 up,int n,t1 v0)
 {t1 c,x,r;int i;if(v0){c=1;for(i=1;i<n;i++){x=up[i];r=x+1;rp[i]=r;}}return c;}
 f(){static const struct{t1 n;t1 src[9];t1 want[9];}d[]={{1,{0},{1}},};t1 got[9];int i;
@@ -6977 +6980 @@
 
 #if defined(__GNUC__) && !defined(__clang__)
 typedef unsigned long long t1;typedef t1*t2;
+#if defined(__GNUC_STDC_INLINE__)  /* e.g. GCC 5.x default */
+extern
+#endif
 __inline__ t1 e(t2 rp,t2 up,int n,t1 v0)
 {t1 c,x,r;int i;if(v0){c=1;for(i=1;i<n;i++){x=up[i];r=x+1;rp[i]=r;}}return c;}
 f(){static const struct{t1 n;t1 src[9];t1 want[9];}d[]={{1,{0},{1}},};t1 got[9];int i;

(Haven't messed with configure.ac, its source, yet.)

[leif]

P.S.: __GNUC_STDC_INLINE__ is implied by -std=gnu11 (and -std=gnu99), which is GCC 5.x's default, while all older versions defaulted to -std=gnu89 (implying __GNUC_GNU_INLINE__).

[leif]

Patch submitted upstream. (Only fixes acinclude.m4, not configure.)

[leif]

Just noticed there are bugs w.r.t. $SAGE_LOCAL/share/mp_config in both MPIR's spkg-install and $SAGE_ROOT/build/install, but that's for another ticket.

(E.g. if I install MPIR with sage -i ... before I've ever run make, MPIR is built without GMP compatibility, due to a syntax error caused by improper quoting. Also, MPIR should always get configured with --enable-gmpcompat if we're bootstrapping GCC. I'd also add an explicit message on whether GMP compatibility is enabled or not. Not sure whether we should also handle the whole stuff after a successful installation if GMP compatibility was enabled in MPIR's spkg-install, and also GMP's spkg-install, i.e., when the default MP library changed by manual installation of the packages.)

[leif]

New commits:

​764ad4e

Add patch fixing MPIR (2.7.0.alpha12) 'configure' failure with GCC 5.x (#18247)

[jdemeyer]

Why not unconditionally use static __inline__? I think that should work on all compilers.

[leif]

Replying to jdemeyer:

Why not unconditionally use static __inline__? I think that should work on all compilers.

Thought of that, too (already when we were fixing the test for clang), but this way it's more explicit, and more importantly doesn't change the test at all for older compilers.

After all, upstream will decide...

[leif]

P.S.: I didn't "bump" the patch level because MPIR 2.7.0-alpha12 didn't have one, although it was already patched before (meanwhile, in Sage 6.7.beta3, even twice)... ;-)

[jdemeyer]

On the other hand, with static __inline__ I will be more at ease regarding other (non-GCC) compilers. Here, you are fixing the test just for GCC but what about other compilers with potentially the same problem?

[jdemeyer]

Replying to leif:

P.S.: I didn't "bump" the patch level because MPIR 2.7.0-alpha12 didn't have one, although it was already patched before (meanwhile, in Sage 6.7.beta3, even twice)... ;-)

Since this is purely about a package not compiling, there is not even a need to bump the patchlevel:

• if a user already managed to compile mpir, there is no need for this fix.
• if a user did not manage to compile mpir, the version number doesn't matter anyway.

[leif]

Replying to jdemeyer:

On the other hand, with static __inline__ I will be more at ease regarding other (non-GCC) compilers. Here, you are fixing the test just for GCC but what about other compilers with potentially the same problem?

Bike-shedding?

This test is solely for GCC. Other compilers may pretend they're GCC, but then they should also define the respective macros. (We explicitly exclude clang, we could in principle exclude ICC as well, but that's something upstream should eventually decide, and not at all an issue for Sage. I'm not aware of any other relevant compiler identifying as GCC.)

Feel free to submit some other patch upstream.

[leif]

P.S.: Using static __inline__ would actually modify the test, so it wouldn't be guaranteed that it catches the situation it was written for (namely a bug in older GCC versions).

[jdemeyer]

Replying to leif:

This test is solely for GCC.

I missed this part. I still don't like your patch (in fact, I don't even like the test), but I can accept it.

[leif]

Replying to jdemeyer:

I still don't like your patch (in fact, I don't even like the test), but I can accept it.

I don't like the test either; I'd simply require some minimal version of GCC.

But as I said, it's IMHO up to upstream to clean up / decide on the tests...

but I can accept it.

Thanks. We can discuss again when finally a new upstream release gets out... ;-)

I just don't want to have a Sage 6.7 that cannot even bootstrap Sage's GCC 4.9.2 with GCC 5.x for no real reason; distros with GCC 5.1 will presumably pop up soon.