On Mon, 18 Mar 2019, Richard Biener wrote:
>
> The following adjusts release_ssa_name to not clear TREE_TYPE but set
> it to error_mark_node instead. This avoids ICEs when dumping stmts
> with released SSA names which can easily happen when doing -details
> dumps and blocks being removed during the PRE-order CFG cleanup
> phase which runs before SSA updating.
>
> It also removes restoring SSA_NAME_VAR in exchange. Not sure
> what that new incremental SSA update code was - the code is
> this way since forever, we now handle anonymous SSA names
> with NULL SSA_NAME_VAR just fine and SSA_NAME_VAR is set
> at allocation time (plus we delay releasing when a name is
> registered for update).
Needed to rip out some other old redundant-with-verify-SSA stuff.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied to trunk.
Richard.
2019-03-18 Richard Biener <rguent...@suse.de>
PR middle-end/88945
* tree-ssanames.c (release_ssa_name_fn): For released SSA names
use a TREE_TYPE of error_mark_node to avoid ICEs when dumping
basic-blocks that are removed. Remove restoring SSA_NAME_VAR.
* tree-outof-ssa.c (eliminate_useless_phis): Remove redundant checking.
Index: gcc/tree-ssanames.c
===================================================================
--- gcc/tree-ssanames.c (revision 269754)
+++ gcc/tree-ssanames.c (working copy)
@@ -595,7 +595,6 @@ release_ssa_name_fn (struct function *fn
defining statement. */
if (! SSA_NAME_IN_FREE_LIST (var))
{
- tree saved_ssa_name_var = SSA_NAME_VAR (var);
int saved_ssa_name_version = SSA_NAME_VERSION (var);
use_operand_p imm = &(SSA_NAME_IMM_USE_NODE (var));
@@ -621,13 +620,14 @@ release_ssa_name_fn (struct function *fn
/* Restore the version number. */
SSA_NAME_VERSION (var) = saved_ssa_name_version;
- /* Hopefully this can go away once we have the new incremental
- SSA updating code installed. */
- SET_SSA_NAME_VAR_OR_IDENTIFIER (var, saved_ssa_name_var);
-
/* Note this SSA_NAME is now in the first list. */
SSA_NAME_IN_FREE_LIST (var) = 1;
+ /* Put in a non-NULL TREE_TYPE so dumping code will not ICE
+ if it happens to come along a released SSA name and tries
+ to inspect its type. */
+ TREE_TYPE (var) = error_mark_node;
+
/* And finally queue it so that it will be put on the free list. */
vec_safe_push (FREE_SSANAMES_QUEUE (fn), var);
}
Index: gcc/tree-outof-ssa.c
===================================================================
--- gcc/tree-outof-ssa.c (revision 269754)
+++ gcc/tree-outof-ssa.c (working copy)
@@ -809,26 +809,7 @@ eliminate_useless_phis (void)
gphi *phi = gsi.phi ();
result = gimple_phi_result (phi);
if (virtual_operand_p (result))
- {
- /* There should be no arguments which are not virtual, or the
- results will be incorrect. */
- if (flag_checking)
- for (size_t i = 0; i < gimple_phi_num_args (phi); i++)
- {
- tree arg = PHI_ARG_DEF (phi, i);
- if (TREE_CODE (arg) == SSA_NAME
- && !virtual_operand_p (arg))
- {
- fprintf (stderr, "Argument of PHI is not virtual (");
- print_generic_expr (stderr, arg, TDF_SLIM);
- fprintf (stderr, "), but the result is :");
- print_gimple_stmt (stderr, phi, 0, TDF_SLIM);
- internal_error ("SSA corruption");
- }
- }
-
- remove_phi_node (&gsi, true);
- }
+ remove_phi_node (&gsi, true);
else
{
/* Also remove real PHIs with no uses. */
