The test didn't fail in revision 133125.

Same failure is also present on hppa2.0w-hp-hpux11.11 (4.3.1) and
hppa64-hp-hpux11.11 (4.4.0).  There are quite a few other libgomp
fails that are probably the same bug.

Same on sh4-unknown-linux-gnu.  It seems that they started
to fail after

r133162 | jakub | 2008-03-13 18:26:25 +0900 (Thu, 13 Mar 2008) | 7 lines

        PR middle-end/35185
        * omp-low.c (lower_regimplify, init_tmp_var, save_tmp_var): Removed.
        (lower_omp_2): New function.
        (lower_omp_1, lower_omp): Rewritten.

applied.

There is a libgomp failure on x86-linux with -march=i386

  FAIL: libgomp.c/pr26943-4.c execution test

which doesn't fail on r133161, though I'm not sure it's
related with this PR.

Seems OMP_ATOMIC_LOAD isn't properly regimplified and so &i in TREE_OPERAND (omp_atomic_load, 1) isn't replaced with .omp_data_i->i.  For targets with
sync builtins this doesn't matter, as expand_omp_atomic for them runs all
addr references through force_gimple_operand_bsi and thus it is gimplified at
expand_omp time.  But the expand_omp_atomic_mutex fallback just assumes the
address is gimple value.

Here is a fix I'll be testing (though for hppa-linux I can only eyeball the
assembly).

2008-03-18  Jakub Jelinek  <jakub@redhat.com>

        PR middle-end/35611
        * gimplify.c (gimplify_expr): Gimplify second operand of
        OMP_ATOMIC_LOAD.

--- gcc/gimplify.c.jj   2008-03-06 18:45:59.000000000 +0100
+++ gcc/gimplify.c      2008-03-18 11:55:04.000000000 +0100
@@ -6022,12 +6022,18 @@ gimplify_expr (tree *expr_p, tree *pre_p
 
        case OMP_RETURN:
        case OMP_CONTINUE:
-        case OMP_ATOMIC_LOAD:
-        case OMP_ATOMIC_STORE:
-
+       case OMP_ATOMIC_STORE:
          ret = GS_ALL_DONE;
          break;
 
+       case OMP_ATOMIC_LOAD:
+         if (gimplify_expr (&TREE_OPERAND (*expr_p, 1), pre_p, NULL,
+             is_gimple_val, fb_rvalue) != GS_ALL_DONE)
+           ret = GS_ERROR;
+         else
+           ret = GS_ALL_DONE;
+         break;
+
        case POINTER_PLUS_EXPR:
           /* Convert ((type *)A)+offset into &A->field_of_type_and_offset.
             The second is gimple immediate saving a need for extra statement.

libgomp.c/pr26943-4.c used to fail intermitently before http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=132977
but I certainly can't reproduce it with current trunk or 4.3, and I don't see
how this could be related to this - even -m32 -march=i386 has the needed sync
builtins - both QImode and SImode lock; add{b,l}.

Subject: Bug 35611

Author: jakub
Date: Tue Mar 18 12:21:02 2008
New Revision: 133309

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133309
Log:
	PR middle-end/35611
	* gimplify.c (gimplify_expr): Gimplify second operand of
	OMP_ATOMIC_LOAD.

	* testsuite/libgomp.c/atomic-4.c: New test.

Added:
    trunk/libgomp/testsuite/libgomp.c/atomic-4.c
Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/gimplify.c
    trunk/libgomp/ChangeLog

Subject: Bug 35611

Author: jakub
Date: Tue Mar 18 12:31:28 2008
New Revision: 133310

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=133310
Log:
	PR middle-end/35611
	* gimplify.c (gimplify_expr): Gimplify second operand of
	OMP_ATOMIC_LOAD.

	* testsuite/libgomp.c/atomic-4.c: New test.

Added:
    branches/gcc-4_3-branch/libgomp/testsuite/libgomp.c/atomic-4.c
Modified:
    branches/gcc-4_3-branch/gcc/ChangeLog
    branches/gcc-4_3-branch/gcc/gimplify.c
    branches/gcc-4_3-branch/libgomp/ChangeLog

Fixed.