Solaris sparc build problems

Status: resolved Severity: failure Category: Vlerq SPARC Solaris Created: 2006-12-10 Updated: 2008-02-18

Using vlerqtcl-2006-12-06.tar.gz, I'm had problems compiling on sparc solaris.

I had a problem with int64_t being undefined in the column.h portion of vlerq.c. I found that adding '#include <unistd.h>' to line 12 of vlerq.c fixed that.

I also had a problem with _BIG_ENDIAN on my system defined to something that isn't valid in a #if expression (looks like it is defined as empty string). I fixed this by adding:

 #if WORDS_BIGENDIAN
 #define _BIG_ENDIAN 1
 #endif

Compiling succeeded, but tests failed due to alignment issues (I'm pretty sure). So I added:

 #ifdef __sparc__
 #define VALUES_MUST_BE_ALIGNED 1
 #endif

Now I get compile errors in the VALUES_MUST_BE_ALIGNED versions of the getter functions. I started fixing these, but got stuck on one in Getter_i32 and Getter_i64:

 item->c[i] = ptr[i];

The c field is not an array or pointer and I'm not sure what this should be.

Summary

Several portability bugs fixed w.r.t. Solaris/SPARC, which is big-endian and which requires all word accesses to be aligned to their natural widths.

Fixes now in darcs, some tests added to basic.test to detect alignment problems earlier on.

^2006-12-10 btheado: Imported

^2006-12-10 jcw: Thanks, that catches a lot of issues which regressed from vlerq3 - I see I forgot to check several details. Here's a copy of vlerq.c which should fix things for Solaris: <a title="File uploaded at Dec 11/2006 04:24PM" href="files/3223.gz"><img border=0 alt="Document Icon" src="styles/doc.gif">vlerq.c.gz</a>

^2006-12-10 btheado

Great, it compiles. Now I get a crash in test 26 of emit.test:

 Program received signal SIGSEGV, Segmentation fault.
 #0  0xff15e594 in CoerceColType ()
 #1  0xff15ea10 in CoerceColumn ()
 #2  0xff15c774 in DefCmd_OO ()
 #3  0xff1612a0 in VlerqObjCmd ()

Is that enough clues to figure out the problem?

^2006-12-10 jcw

Hm, I suspect an alignment issue - saves were not implemented in C in v3. Please try the following: in vlerq.c locate the CoerceColType line. Then inside, there's a switch case which contains:

    case IT_wide:
      result = InitColumn(count, PickIntGetter(64), 8 * count);
      for (i = 0; i &lt; count; ++i) {
        item.c = column;
        obj = ItemAsObj(&item, GetItem(i, &item));
        if (!CastObjToItem('L', obj, &item))
          Tcl_Panic("cct-L?");
        ((int64_t*) result.seq-&gt;data[0].p)[i] = item.w;
      }
      break;

Before that last assignment, add:

  Assert(((int) result.seq-&gt;data[0].p) % 8 == 0);

I suspect that it will fail, which is a problem when mis-alignment is not handled by hardware.

I'll look for a portable alternative if this is indeed the problem.

^2006-12-10 jcw

Yep, that's most likely it. You can work around this for now by adding a dummy field in the struct Sequence definition - just add "void *dummy;" after the line:

  Cleaner  cleaner;   /* called during cleanup */

^2006-12-10 btheado: The "void *dummy" workaround fixes the crash. Now I only get a few "asview" failures, but those can be ignored because the vlerq.c is now slightly out of synce with the test files. So it looks good, thanks.

^2006-12-10 jcw: FYI, I've changed the portability code a bit more, am now using "stdint.h" - if WORDS_BIGENDIAN is a built-in define on SPARC then this should still all work.

^2006-12-11 btheado: WORDS_BIGENDIAN comes from ./configure, so it should work for all TEA based builds.