[Bug tree-optimization/45421] [4.6 regression] Ada bootstrap failure on IRIX 6.5: SIGBUS in sem_aggr.sort_case_table

[ebotcazou at gcc dot gnu dot org]

------- Comment #4 from ebotcazou at gcc dot gnu dot org  2010-09-02 21:59 -------
It's IVOPTS so no way to get a reduced testcase.  The analysis is for the stock
sem_aggr.adb file compiled at -O2 -gnatpg for IA-64.  We have:

table.68_174 = .builtin_alloca (D.5147_173);

with pi->align = 16

  D.5252_266 = &*table.68_174;

with pi->align = 16 and then

 MEM[(struct sem_aggr__case_bounds[(size_type) <PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0:<PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->UB0 >= <PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0 ? (size_type) <PLACEHOLDER_EXPR
struct sem_aggr__case_table_type___XUP>.P_BOUNDS->UB0 : (size_type)
<PLACEHOLDER_EXPR struct sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0 + -1]
*)D.5252_266][D.6519_896]{lb: 1 sz: 12};

is replaced with

MEM[(struct sem_aggr__case_bounds[(size_type) <PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0:<PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->UB0 >= <PLACEHOLDER_EXPR struct
sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0 ? (size_type) <PLACEHOLDER_EXPR
struct sem_aggr__case_table_type___XUP>.P_BOUNDS->UB0 : (size_type)
<PLACEHOLDER_EXPR struct sem_aggr__case_table_type___XUP>.P_BOUNDS->LB0 + -1]
*)D.6658_890]

in IVOPTS.  And copy_ref_info reads:

  /* We can transfer points-to information from an old pointer
     or decl base to the new one.  */

But now duplicate_ssa_name_ptr_info duplicates more than the points-to info,
it also duplicates the alignment info, which is wrong for arrays since the
base decl (pointed to by D.5252_266) can be more aligned than the individual
elements (pointed to by D.6658_890).

Tentative patch (I cannot test it on IA-64 before next week):

Index: tree-ssa-loop-ivopts.c
===================================================================
--- tree-ssa-loop-ivopts.c      (revision 163745)
+++ tree-ssa-loop-ivopts.c      (working copy)
@@ -5891,8 +5891,8 @@ copy_ref_info (tree new_ref, tree old_re
   else if (TREE_CODE (new_ref) == MEM_REF)
     new_ptr_base = TREE_OPERAND (new_ref, 0);

-  /* We can transfer points-to information from an old pointer
-     or decl base to the new one.  */
+  /* We can transfer points-to information from an old pointer or base decl
+     to the new one, as well as some amount of alignment information.  */
   if (new_ptr_base
       && TREE_CODE (new_ptr_base) == SSA_NAME
       && POINTER_TYPE_P (TREE_TYPE (new_ptr_base))
@@ -5901,11 +5901,18 @@ copy_ref_info (tree new_ref, tree old_re
       tree base = get_base_address (old_ref);
       if (!base)
        ;
-      else if ((INDIRECT_REF_P (base)
-               || TREE_CODE (base) == MEM_REF)
-              && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME)
-       duplicate_ssa_name_ptr_info
-         (new_ptr_base, SSA_NAME_PTR_INFO (TREE_OPERAND (base, 0)));
+      else if ((INDIRECT_REF_P (base) || TREE_CODE (base) == MEM_REF)
+              && TREE_CODE (TREE_OPERAND (base, 0)) == SSA_NAME
+              && SSA_NAME_PTR_INFO (TREE_OPERAND (base, 0)))
+       {
+         struct ptr_info_def *pi;
+         duplicate_ssa_name_ptr_info
+           (new_ptr_base, SSA_NAME_PTR_INFO (TREE_OPERAND (base, 0)));
+         pi = SSA_NAME_PTR_INFO (new_ptr_base);
+         pi->misalign
+           = -(TYPE_ALIGN (TREE_TYPE (new_ref)) / BITS_PER_UNIT)
+             & (pi->align - 1);
+       }
       else if (TREE_CODE (base) == VAR_DECL
               || TREE_CODE (base) == PARM_DECL
               || TREE_CODE (base) == RESULT_DECL)


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45421