https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68659
vries at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |vries at gcc dot gnu.org
--- Comment #17 from vries at gcc dot gnu.org ---
(In reply to John David Anglin from comment #11)
> Seeing this on hppa with -mschedule=7100LC. This is with r231795.
>
> There is null pointer exception here:
>
> Program received signal SIGSEGV, Segmentation fault.
> translate_isl_ast_to_gimple::collect_all_ssa_names (
> this=this@entry=0xfd703908, new_expr=0x0,
> vec_ssa=vec_ssa@entry=0xfd7044c8)
> at ../../gcc/gcc/graphite-isl-ast-to-gimple.c:1603
> 1603 if (TREE_CODE (new_expr) == SSA_NAME)
> (gdb) p new_expr
> $1 = (tree) 0x0
I wonder if this patch is correct (at least fixes the ICE with x86_64 -m32):
...
index d3614e4..e5b03ee 100644
--- a/gcc/graphite-isl-ast-to-gimple.c
+++ b/gcc/graphite-isl-ast-to-gimple.c
@@ -1404,6 +1404,9 @@ void translate_isl_ast_to_gimple::
collect_all_ssa_names (tree new_expr, vec<tree> *vec_ssa)
{
+ if (new_expr == NULL_TREE)
+ return;
+
/* Rename all uses in new_expr. */
if (TREE_CODE (new_expr) == SSA_NAME)
{
@@ -1804,6 +1807,9 @@ get_new_name (basic_block new_bb, tree op,
if (is_constant (op))
return op;
+ if (TREE_CODE (op) == ADDR_EXPR)
+ return op;
+
return get_rename (new_bb, op, old_bb, phi_kind);
}
...
We're doing collect_all_ssa_names for component_ref u2.buf. Operand 2 is
NULL_TREE, which AFAIU according to the definition of component_ref is allowed
(tree.def: Operand 2, if present, ...). This code in collect_all_ssa_names just
looks at TREE_CODE_LENGTH, which is 3 for component_ref, and for i == 2, op ==
NULL_TREE:
...
/* Iterate over SSA_NAMES in NEW_EXPR. */
for (int i = 0; i < (TREE_CODE_LENGTH (TREE_CODE (new_expr))); i++)
{
tree op = TREE_OPERAND (new_expr, i);
collect_all_ssa_names (op, vec_ssa);
}
...
The first bit of the patch allows the NULL_TREE operand in
collect_all_ssa_names. But subsequently, we run into handling op == &u2.buf in
get_new_name, which results in an assert that op is not an ssa-name in
get_rename. The last bit of the patch handles that.
