unaligned testcases fail on arm

[juliantaylor]

I have disabled the blosc compression which does not work on architectures without unaligned memory access.
Unfortunately the testsuite still fails at "Comparing written unaligned time data with read data in a Table" tests:
https://launchpadlibrarian.net/94731766/buildlog_ubuntu-precise-armel.pytables_2.3.1-2ubuntu2_FAILEDTOBUILD.txt.gz

do these tests require unaligned memory access too?

[FrancescAlted]

On Feb 28, 2012, at 11:33 PM, Julian Taylor wrote:

I have disabled the blosc compression which does not work on architectures without unaligned memory access.

That's good. Thanks.

Unfortunately the testsuite still fails at "Comparing written unaligned time data with read data in a Table" tests:
https://launchpadlibrarian.net/94731766/buildlog_ubuntu-precise-armel.pytables_2.3.1-2ubuntu2_FAILEDTOBUILD.txt.gz

do these tests require unaligned memory access too?

Yes, they do. What I don't quite understand is that this test is there since 2006, and it always passed tests on many different platforms. But anyway, feel free to produce a patch so that this test can be skipped.

-- Francesc Alted

[juliantaylor]

I don't see how that test could ever have succeeded, except for random chance. There is even a comment in the problematic file that it will break.

in the mean time I have applied this stupid patch to the problematic file to avoid the unaligned access.
With it and disabled blosc now the whole testsuite succeeds on a sparc64 machine and it will hopefully also work on other arches like arm.

--- pytables-2.3.1.orig/src/typeconv.c  2011-10-28 22:11:14.000000000 +0200
+++ pytables-2.3.1/src/typeconv.c       2012-03-01 00:52:34.000000000 +0200
@@ -66,18 +66,21 @@

   for (record = 0;  record < nrecords;  record++) {
     for (element = 0;  element < nelements;  element++) {
+      double fb;
+      memcpy(&fb, fieldbase, sizeof(*fieldbase));
       if (sense == 0) {
        /* Convert from float64 to timeval32. */
-       tv.i64 = (((PY_LONG_LONG)(*fieldbase) << 32)
-                 | (lround((*fieldbase - (int)(*fieldbase)) * 1e+6)
+       tv.i64 = (((PY_LONG_LONG)(fb) << 32)
+                 | (lround((fb - (int)(fb)) * 1e+6)
                     & 0x0ffffffff));
-       *fieldbase = tv.f64;
+       fb = tv.f64;
       } else {
        /* Convert from timeval32 to float64. */
-       tv.f64 = *fieldbase;
+       tv.f64 = fb;
        /* the next computation is 64 bit-platforms aware */
-       *fieldbase = 1e-6 * (int)tv.i64 + (tv.i64 >> 32);
+       fb = 1e-6 * (int)tv.i64 + (tv.i64 >> 32);
       }   
+      memcpy(fieldbase, &fb, sizeof(*fieldbase));
       fieldbase++;
     }   

[FrancescAlted]

On Feb 29, 2012, at 4:06 PM, Julian Taylor wrote:

I don't see how that test could ever have succeeded, except for random chance. There is even a comment in the problematic file that it will break.

Yeah, it should have been probably just random chance, but still...

in the mean time I have applied this stupid patch to the problematic file to avoid the unaligned access.

That's great. I'm fine with the patch, and should be definitely applied to upstream too.

With it and disabled blosc now the whole testsuite succeeds on a sparc64 machine and it will hopefully also work on other arches like arm.

Excellent! Thanks for your contribution!

-- Francesc Alted