[Bug tree-optimization/64111] [5.0 regression] ICE: conversion of register to a different size

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64111

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |ASSIGNED
           Assignee|unassigned at gcc dot gnu.org      |rguenth at gcc dot gnu.org
   Target Milestone|---                         |5.0

--- Comment #17 from Richard Biener <rguenth at gcc dot gnu.org> ---
 <view_convert_expr 0x7ffff6836660
    type <vector_type 0x10001e8608
        type <integer_type 0x10001dedc8 long long int DI
            size <integer_cst 0x100000eb00 constant 64>
            unit size <integer_cst 0x100000eb20 constant 8>
            align 64 symtab 0 alias set -1 canonical type 0x10001dedc8
precision 64 min <integer_cst 0x1000010b40 -9223372036854775808> max
<integer_cst 0x1000010b60 9223372036854775807>
            pointer_to_this <pointer_type 0x10001e5a10>>
        V4DI
        size <integer_cst 0x1000010d40 constant 256>
        unit size <integer_cst 0x1000010d80 constant 32>
        align 256 symtab 0 alias set -1 canonical type 0x10001e86b0 nunits 4
        pointer_to_this <pointer_type 0x10001e8758>>

    arg 0 <ssa_name 0x7ffff68830d8
        type <vector_type 0x7ffff686a000 type <integer_type 0x10001ded20 long
unsigned int>
            unsigned V4DI
            size <integer_cst 0x7ffff6851de0 constant 256> unit size
<integer_cst 0x1000010d80 32>
            align 256 symtab 0 alias set 6 canonical type 0x7ffff686a000 nunits
4
            pointer_to_this <pointer_type 0x7ffff6870690>>
        var <var_decl 0x7ffff68825f0 vect_var_.22>def_stmt vect_var_.22_70 =
MEM[(struct QAbstractAnimationTimer * const &)vect_ppretmp.18_68];

        version 70>>

It looks like PCH causes integer constant sharing to break down.

(gdb) p debug_tree ((tree)0x7ffff6851de0)
 <integer_cst 0x7ffff6851de0 type <integer_type 0x10001de690 bitsizetype>
constant 256>
$2 = void
(gdb) p debug_tree ((tree)0x1000010d40)
 <integer_cst 0x1000010d40 type <integer_type 0x10001de690 bitsizetype>
constant 256>

the constant is above the integer share limit thus it only gets shared
via the int_cst_hash_table hashtable which is collectible:

static GTY ((if_marked ("ggc_marked_p"), param_is (union tree_node)))
     htab_t int_cst_hash_table;

but we are supposed to save that into the PCH file and restore it.

tree-cfg.c relies on this:

          else if (TREE_CODE (op) == SSA_NAME
                   && TYPE_SIZE (TREE_TYPE (expr)) != TYPE_SIZE (TREE_TYPE
(op)))
            {
              error ("conversion of register to a different size");

though from a quick grep it doesn't seem to be the only one doing so:

fold-const.c:     return ((TYPE_SIZE (TREE_TYPE (arg0)) == TYPE_SIZE (TREE_TYPE
(arg1))
tree-inline.c:      gcc_checking_assert (TYPE_SIZE (type) == TYPE_SIZE
(TYPE_MAIN_VARIANT (type)));
tree-inline.c:      gcc_checking_assert (TYPE_SIZE_UNIT (type) ==
TYPE_SIZE_UNIT (TYPE_MAIN_VARIANT (type)));
tree-inline.c:        || TYPE_SIZE (TREE_TYPE (p)) == TYPE_SIZE (TREE_TYPE
(value)))
tree-inline.c:              else if (TYPE_SIZE (req_type) == TYPE_SIZE
(TREE_TYPE (replace_info->new_tree)))

Ah, the issue might be that we have (on the 4.8 branch):

/* Return the hash code code X, an INTEGER_CST.  */

static hashval_t
int_cst_hash_hash (const void *x)
{
  const_tree const t = (const_tree) x;

  return (TREE_INT_CST_HIGH (t) ^ TREE_INT_CST_LOW (t)
          ^ htab_hash_pointer (TREE_TYPE (t)));
}

thus we hash in the pointer value ... (oops).  Similar on trunk:

/* Return the hash code code X, an INTEGER_CST.  */

hashval_t
int_cst_hasher::hash (tree x)
{
  const_tree const t = x;
  hashval_t code = htab_hash_pointer (TREE_TYPE (t));
  int i;

  for (i = 0; i < TREE_INT_CST_NUNITS (t); i++)
    code ^= TREE_INT_CST_ELT (t, i);

  return code;
}

I'm quite sure that doesn't survive pointer-swizzling with PCH.

I suggest to use a better hash than a simple XOR and rather than hashing
in TREE_TYPE hash in TYPE_UID.