Hi!

The testcase in the PR (GC related, so not easily reduceable) ICEs
because two TEMPLATE_TYPE_PARMs that comptypes returns true on have
different TYPE_CANONICAL (each has itself).
The problem is that 
comptypes->structural_comptypes->comp_template_parms_position
compares not just idx and level, but also the
TEMPLATE_PARM_PARAMETER_PACK bit, and rewrite_template_parm first
computes the TYPE_CANONICAL (with the bit still unset) and then copies
over the TEMPLATE_PARM_PARAMETER_PACK bit.  This means that it can for
TYPE_CANONICAL comparison compare unequal when it will compare equal later
or maybe vice versa.

Fixed thusly (the rewrite_template_parm changes), additionally while
debugging I've noticed an inefficiency, vec_safe_grow_cleared is certainly
more efficient than vec_safe_push of NULL in a loop - e.g. the latter
might reallocate multiple times and really can't beat one growing and
one memset, and two minor formatting nits.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2017-04-05  Jakub Jelinek  <ja...@redhat.com>

        PR c++/80309
        * pt.c (canonical_type_parameter): Use vec_safe_grow_cleared instead
        of a loop doing vec_safe_push of NULL.  Formatting fixes.
        (rewrite_template_parm): Copy TEMPLATE_PARM_PARAMETER_PACK from oldidx
        to newidx before calling canonical_type_parameter on newtype.

--- gcc/cp/pt.c.jj      2017-04-04 07:32:53.000000000 +0200
+++ gcc/cp/pt.c 2017-04-05 13:43:18.883690743 +0200
@@ -3997,10 +3997,10 @@ canonical_type_parameter (tree type)
   tree list;
   int idx = TEMPLATE_TYPE_IDX (type);
   if (!canonical_template_parms)
-    vec_alloc (canonical_template_parms, idx+1);
+    vec_alloc (canonical_template_parms, idx + 1);
 
-  while (canonical_template_parms->length () <= (unsigned)idx)
-    vec_safe_push (canonical_template_parms, NULL_TREE);
+  if (canonical_template_parms->length () <= (unsigned) idx)
+    vec_safe_grow_cleared (canonical_template_parms, idx + 1);
 
   list = (*canonical_template_parms)[idx];
   while (list && !comptypes (type, TREE_VALUE (list), COMPARE_STRUCTURAL))
@@ -4011,8 +4011,7 @@ canonical_type_parameter (tree type)
   else
     {
       (*canonical_template_parms)[idx]
-               = tree_cons (NULL_TREE, type,
-                            (*canonical_template_parms)[idx]);
+       = tree_cons (NULL_TREE, type, (*canonical_template_parms)[idx]);
       return type;
     }
 }
@@ -24955,6 +24954,8 @@ rewrite_template_parm (tree olddecl, uns
       newidx = TEMPLATE_TYPE_PARM_INDEX (newtype)
        = build_template_parm_index (index, level, level,
                                     newdecl, newtype);
+      TEMPLATE_PARM_PARAMETER_PACK (newidx)
+       = TEMPLATE_PARM_PARAMETER_PACK (oldidx);
       TYPE_STUB_DECL (newtype) = TYPE_NAME (newtype) = newdecl;
       TYPE_CANONICAL (newtype) = canonical_type_parameter (newtype);
 
@@ -25002,11 +25003,11 @@ rewrite_template_parm (tree olddecl, uns
       SET_DECL_TEMPLATE_PARM_P (newconst);
       newidx = build_template_parm_index (index, level, level,
                                          newconst, newtype);
+      TEMPLATE_PARM_PARAMETER_PACK (newidx)
+       = TEMPLATE_PARM_PARAMETER_PACK (oldidx);
       DECL_INITIAL (newdecl) = DECL_INITIAL (newconst) = newidx;
     }
 
-  TEMPLATE_PARM_PARAMETER_PACK (newidx)
-    = TEMPLATE_PARM_PARAMETER_PACK (oldidx);
   return newdecl;
 }
 

        Jakub

Reply via email to