Confirmed on powerpc-apple-darwin* (see http://gcc.gnu.org/regtest/HEAD/native-lastbuild.txt.gzip).
I'll attach the preprocessed file strtod.i. The backtrace of

/opt/gcc/darwin_buildw/gcc/cc1 -O2 -g strtod.i

(-m32 default) is

#0  fancy_abort (file=0x9a4664 "../../work/gcc/tree-flow-inline.h", line=745, function=0xb886e4 "op_iter_init") at ../../work/gcc/diagnostic.c:893
#1  0x007a4570 in op_iter_init () at tree-flow-inline.h:743
#2  0x007ac974 in convert_mult_to_fma (mul_stmt=0xcf08bc, op1=0xd12e70, op2=0x41fa8bd0) at ../../work/gcc/tree-ssa-math-opts.c:2132
#3  0x007acd4c in execute_optimize_widening_mul () at ../../work/gcc/tree-ssa-math-opts.c:2319
#4  0x005b3d24 in execute_one_pass (pass=0xbbefd4) at ../../work/gcc/passes.c:1899
#5  0x005b40b8 in execute_pass_list (pass=0xc00000) at ../../work/gcc/passes.c:1953
#6  execute_pass_list (pass=0xbbe850) at ../../work/gcc/passes.c:1955
#7  0x00703088 in tree_rest_of_compilation (fndecl=0x41fc5480) at ../../work/gcc/tree-optimize.c:417
#8  0x0029dde0 in cgraph_expand_function (node=0x41fab1f8) at ../../work/gcc/cgraphunit.c:1635
#9  0x0029fe08 in cgraph_optimize () at ../../work/gcc/cgraphunit.c:1694
#10 0x002a052c in cgraph_finalize_compilation_unit () at ../../work/gcc/cgraphunit.c:1131
#11 0x000200ec in c_write_global_declarations () at ../../work/gcc/c-decl.c:9844
#12 0x0067ae54 in toplev_main (argc=3, argv=0x0) at ../../work/gcc/toplev.c:586
#13 0x00001ca0 in start ()

According http://gcc.gnu.org/ml/gcc-regression/2011-06/msg00099.html, the failure occurred between revisions 174833 and 174836.

Created attachment 24476 [details]
preprocessed file strtod.i

Sounds a bit strange.  It must be (my tree doesn't match the lines exactly):

          /* Make sure the negate statement becomes dead with this
             single transformation.  */
          if (!single_imm_use (gimple_assign_lhs (use_stmt),
                               &use_p, &neguse_stmt))
            return false;

          /* Make sure the multiplication isn't also used on that stmt.  */
          FOR_EACH_SSA_TREE_OPERAND (use, neguse_stmt, iter, SSA_OP_USE)
            if (use == mul_result)
              return false;

"guessed patch":

Index: gcc/tree-ssa-math-opts.c
===================================================================
--- gcc/tree-ssa-math-opts.c    (revision 174845)
+++ gcc/tree-ssa-math-opts.c    (working copy)
@@ -2174,7 +2174,7 @@ convert_mult_to_fma (gimple mul_stmt, tr
       if (use_code == NEGATE_EXPR)
        {
          ssa_op_iter iter;
-         tree use;
+         use_operand_p usep;
 
          result = gimple_assign_lhs (use_stmt);
 
@@ -2185,8 +2185,8 @@ convert_mult_to_fma (gimple mul_stmt, tr
            return false;
 
          /* Make sure the multiplication isn't also used on that stmt.  */
-         FOR_EACH_SSA_TREE_OPERAND (use, neguse_stmt, iter, SSA_OP_USE)
-           if (use == mul_result)
+         FOR_EACH_PHI_OR_STMT_USE (usep, neguse_stmt, iter, SSA_OP_USE)
+           if (USE_FROM_PTR (usep) == mul_result)
              return false;
 
          /* Re-validate.  */

CCed Andi Kleen (BTW Andi could you update your email address in MAINTAINERS? TIA).

Hmm, it's hard to see how my patch could have caused this.
It doesn't really change any RTL.

Does the test case even use global registers?

I don't see any in native/fdlibm/strtod.c

The only thing that could possibly have a side effect is storing
the decl in a global array, but even that should be benign.

Are you sure you bisected right?

> Are you sure you bisected right?

I should have looked at http://gcc.gnu.org/ml/gcc-regression/2011-06/msg00092.html where the range is given by

> With your recent patch, GCC HEAD revision 174833 had problems on:
> native: build (NEW build failure)
> Attached is build output for those targets.
> The previous build was of revision 174812.

Sorry about that.

Four cpu2000 benchmarks (eon,fma3d,sixtrack,apsi) also fail to build with the same ICE starting with the following revision.

------------------------------------------------------------------------
r174815 | aoliva | 2011-06-08 14:39:12 -0500 (Wed, 08 Jun 2011) | 4 lines

* tree-flow-inline.h (op_iter_init): Reject GIMPLE_PHI stmts.
(num_ssa_operands): Likewise.
(op_iter_init_phiuse): Forward-declare.
(delink_stmt_imm_use): Iterate with FOR_EACH_PHI_OR_STMT_USE.

With the patch in comment #3 applied on top of revision 174846 I have been able to update powerpc-apple-darwin9 (long update, but not a full bootstrap).

Well, I'm going to test and install it.

Author: rguenth
Date: Fri Jun 10 12:18:55 2011
New Revision: 174902

URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=174902
Log:
2011-06-10  Richard Guenther  <rguenther@suse.de>

	PR bootstrap/49344
	* tree-ssa-math-opts.c (convert_mult_to_fma): Use
	FOR_EACH_PHI_OR_STMT_USE.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/tree-ssa-math-opts.c

Fixed.