------- Additional Comments From steven at gcc dot gnu dot org 2005-05-12
10:23 -------
The failure I mentioned in comment #7 is caused by remove_dead_inserted_code,
which removes the statement setting that SSA name:
In the .pre dump file:
Removing unnecessary insertion:pretmp.308_35 = (unsigned char *)pretmp.307_36;
And the ICE message:
error: Missing definition for SSA_NAME: pretmp.308_35 in statement:
That remove_dead_inserted_code passlet doesn't work very well anyway, for
the original test case we also have dead code in the .pre dump:
do_locator1 (call)
{
int prephitmp.4;
int pretmp.3;
char prephitmp.2;
char pretmp.1;
int type;
char D.1566;
int D.1565;
<bb 0>:
if (call_3 != 0B) goto <L0>; else goto <L4>;
<L4>:;
goto <bb 2> (<L1>);
<L0>:;
type_7 = *call_3;
pretmp.1_2 = (char) type_7;
pretmp.3_10 = (int) pretmp.1_2;
# prephitmp.4_11 = PHI <1(3), pretmp.3_10(1)>;
# prephitmp.2_9 = PHI <1(3), pretmp.1_2(1)>;
# type_1 = PHI <1(3), type_7(1)>;
<L1>:;
D.1566_4 = prephitmp.2_9;
D.1565_5 = prephitmp.4_11;
return D.1565_5;
}
It may be better to just not have this ad-hoc DCE and move a real DCE pass
up in the pass schedule -- something like this:
Index: tree-optimize.c
===================================================================
RCS file: /cvs/gcc/gcc/gcc/tree-optimize.c,v
retrieving revision 2.90
diff -u -3 -p -r2.90 tree-optimize.c
--- tree-optimize.c 11 May 2005 02:24:42 -0000 2.90
+++ tree-optimize.c 12 May 2005 09:09:53 -0000
@@ -405,10 +405,10 @@ init_tree_optimization_passes (void)
NEXT_PASS (pass_split_crit_edges);
NEXT_PASS (pass_pre);
NEXT_PASS (pass_sink_code);
+ NEXT_PASS (pass_dce);
NEXT_PASS (pass_loop);
NEXT_PASS (pass_dominator);
NEXT_PASS (pass_copy_prop);
- NEXT_PASS (pass_dce);
/* FIXME: If DCE is not run before checking for uninitialized uses,
we may get false warnings (e.g., testsuite/gcc.dg/uninit-5.c).
However, this also causes us to misdiagnose cases that should be
@@ -418,8 +418,8 @@ init_tree_optimization_passes (void)
account for the predicates protecting the set and the use of each
variable. Using a representation like Gated Single Assignment
may help. */
- NEXT_PASS (pass_late_warn_uninitialized);
NEXT_PASS (pass_cd_dce);
+ NEXT_PASS (pass_late_warn_uninitialized);
NEXT_PASS (pass_dse);
NEXT_PASS (pass_forwprop);
NEXT_PASS (pass_phiopt);
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=21520
