Need patch for Fedora package for non x86 architectures

[pcpa]

Arm apparently built fine, but s390 failed:
https://bugzilla.redhat.com/show_bug.cgi?id=1186162

But I am afraid there will be issues on other architectures.

The patch I made, http://pkgs.fedoraproject.org/cgit/mp.git/diff/mp-arm.patch?id=95f0b59ae66eb4c870f4a3110a3362cd83fb4344
is not enough.

[vitaut]

Thanks for the bug report.

Here's a patch that should fix the problem:

From 986a91a46e206e268eeda7c3ff614f4cd982a4dc Mon Sep 17 00:00:00 2001
From: Victor Zverovich <victor.zverovich@gmail.com>
Date: Tue, 27 Jan 2015 14:22:08 -0800
Subject: [PATCH] Improve portability of fpinit

---
 src/asl/solvers/fpinit.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/src/asl/solvers/fpinit.c b/src/asl/solvers/fpinit.c
index ae06cf0..1afc23a 100644
--- a/src/asl/solvers/fpinit.c
+++ b/src/asl/solvers/fpinit.c
@@ -125,14 +125,18 @@ fpinit_ASL(Void)
    _FPU_GETCW(__fpu_control);
    __fpu_control &= ~_FPU_EXTENDED;    /* clear rounding precision bits */
    __fpu_control |= _FPU_DOUBLE;       /* set the ones we want set */
+   _FPU_SETCW(__fpu_control);
 #else
 #if defined(_FPU_IEEE) && defined(_FPU_EXTENDED) && defined(_FPU_DOUBLE)
    __fpu_control = _FPU_IEEE - _FPU_EXTENDED + _FPU_DOUBLE;
-#else
+   _FPU_SETCW(__fpu_control);
+#elif defined(__i386__) || defined(__x86_64__)
    __fpu_control = 0x27f;
+   _FPU_SETCW(__fpu_control);
+#elif defined(FE_ALL_EXCEPT)
+   fedisableexcept(FE_ALL_EXCEPT);
 #endif
 #endif /* ASL_FPINIT_KEEP_TRAPBITS */
-   _FPU_SETCW(__fpu_control);
 #endif
    }
 #endif /*} NO_fpu_control */
-- 
1.9.1

However, I haven't tried it as I don't have access to a s390 machine.

Please let me know if it helps and I'll merge it into master.

[pcpa]

I do not have a shell to see exactly what happened, but submitting a scratch build failed.
http://s390.koji.fedoraproject.org/kojifiles/work/tasks/6258/1716258/build.log
(temporary link) I will ask someone with shell access to see if it was a coredump.
I can also try to create a vm, but that would take "forever" under emulation...
For reference, the bug report that showed the values were x86 specific
https://bugzilla.redhat.com/show_bug.cgi?id=1186162

[pcpa]

This should give a hint of what is the problem. If you need some further testing please let me know.

[ RUN      ] ASLBuilderTest.InitASLFull
/root/rpmbuild/BUILD/mp-35060ba2a59f2b0f0fd622ed9df678f142f846ed/test/asl/aslbuilder-test.cc:187: Failure
Value of: actual.i.xscanf_
  Actual: 0x800b64c0
Expected: expected.i.xscanf_
Which is: 0x800b6100
/root/rpmbuild/BUILD/mp-35060ba2a59f2b0f0fd622ed9df678f142f846ed/test/asl/aslbuilder-test.cc:277: Failure
Value of: actual.i.binary_nl_
  Actual: 2
Expected: expected.i.binary_nl_
Which is: 0
/root/rpmbuild/BUILD/mp-35060ba2a59f2b0f0fd622ed9df678f142f846ed/test/asl/aslbuilder-test.cc:384: Failure
Value of: actual.i.iadjfcn
  Actual: 0x800c5390
Expected: expected.i.iadjfcn
Which is: NULL
/root/rpmbuild/BUILD/mp-35060ba2a59f2b0f0fd622ed9df678f142f846ed/test/asl/aslbuilder-test.cc:385: Failure
Value of: actual.i.dadjfcn
  Actual: 0x800c5390
Expected: expected.i.dadjfcn
Which is: NULL
[  FAILED  ] ASLBuilderTest.InitASLFull (1 ms)
[ RUN      ] ASLBuilderTest.ASLBuilderAdjFcn

[vitaut]

Well, at least it gets to the tests now. The test failure looks related to the detection of floating-point arithmetic, I'll look into it in more details shortly. Thanks for providing the test log BTW, this is very useful.

[vitaut]

This should (hopefully) be fixed in 0b53c90.

[pcpa]

Confirmed! With that patch it pass all tests :)
I believe something like this may also be good, but should not make difference
currently for Fedora:

 #elif defined(__i386__) || defined(__x86_64__)
    __fpu_control = 0x27f;
    _FPU_SETCW(__fpu_control);
+#elif defined(_FPU_DEFAULT)
+   __fpu_control = _FPU_DEFAULT;
+   _FPU_SETCW(__fpu_control);
 #elif defined(FE_ALL_EXCEPT)
    fedisableexcept(FE_ALL_EXCEPT);

as _FPU_DEFAULT appears to be defined in most, if not all glibc ports,
and defaults to double precision IEEE floating math.

[vitaut]

Great, thanks for the confirmation!

I'm not sure about _FPU_DEFAULT though, as it doesn't mask interrupts. I think it's better to fall back on fedisableexcept which does. Besides ASL should work with non-IEEE floating point too.