Re: [patch][version 6] add -ftrivial-auto-var-init and variable attribute "uninitialized" to gcc

[Qing Zhao <qing.zhao@oracle.com>]

Hi,

> On Aug 10, 2021, at 9:16 AM, Richard Biener <rguenther@suse.de> wrote:
> 
> On Tue, 10 Aug 2021, Qing Zhao wrote:
> 
>>>>> 
>>>>> +static void
>>>>> +expand_DEFERRED_INIT (internal_fn, gcall *stmt)
>>>>> +{
>>>>> +  tree var = gimple_call_lhs (stmt);
>>>>> +  tree size_of_var = gimple_call_arg (stmt, 0);
>>>>> +  tree vlaaddr = NULL_TREE;
>>>>> +  tree var_type = TREE_TYPE (var);
>>>>> +  bool is_vla = (bool) TREE_INT_CST_LOW (gimple_call_arg (stmt, 2));
>>>>> +  enum auto_init_type init_type
>>>>> +    = (enum auto_init_type) TREE_INT_CST_LOW (gimple_call_arg (stmt, 1));
>>>>> +
>>>>> +  gcc_assert (init_type > AUTO_INIT_UNINITIALIZED);
>>>>> +
>>>>> +  /* if this variable is a VLA, get its SIZE and ADDR first.  */
>>>>> +  if (is_vla)
>>>>> +    {
>>>>> +      /* The temporary address variable for this vla should have been
>>>>> +        created during gimplification phase.  Refer to gimplify_vla_decl
>>>>> +        for details.  */
>>>>> +      tree var_decl = (TREE_CODE (var) == SSA_NAME) ?
>>>>> +                      SSA_NAME_VAR (var) : var;
>>>>> +      gcc_assert (DECL_HAS_VALUE_EXPR_P (var_decl));
>>>>> +      gcc_assert (TREE_CODE (DECL_VALUE_EXPR (var_decl)) == 
>>>>> INDIRECT_REF);
>>>>> +      /* Get the address of this vla variable.  */
>>>>> +      vlaaddr = TREE_OPERAND (DECL_VALUE_EXPR (var_decl), 0);
>>>>> 
>>>>> err - isn't the address of the decl represented by the LHS 
>>>>> regardless whether this is a VLA or not?
>>>> 
>>>> The LHS of the call to .DEFERRED_INIT is the DECL itself whatever it’s a VLA or not. 
>>>> 
>>>> In order to create a memset call, we need the Address of this DECL as the first argument. 
>>>> If the DECL is not a VLA, we just simply apply “build_fold_addr_expr” on this DECL to get its address,
>>>> However, for VLA, during gimplification phase “gimplify_vla_decl”, we have already created a temporary
>>>> address variable for this DECL, and recorded this address variable with “DECL_VALUE_EXPR(DECL), 
>>>> We should use this already created address variable  for VLAs. 
>>> 
>>> So the issue is that the LHS of the .DEFERRED_INIT call is not properly
>>> gimplified.  We should not have such decl there but I see we do not
>>> have IL verification that covers this.
>> 
>> Don’t quite understand here:  do you mean all the LHS of .DEFERRED_INIT call are not properly gimplified, or
>> Only the LHS of .DEFERRED_INIT call for VLA are not properly gimplified?
> 
> Especially in the VLA case but likely also in general (though unlikely
> since usually the receiver of initializations are simple enough).  I'd
> expect the VLA case end up as
> 
> *ptr_to_decl = .DEFERRED_INIT (...);
> 
> where *ptr_to_decl is the DECL_VALUE_EXPR of the decl.

So, for the following small testing case:

====
extern void bar (int);

void foo(int n)
{
  int arr[n];
  bar (arr[2]);
  return;
}
=====

If I compile it with -ftrivial-auto-var-init=zero -fdump-tree-gimple -S -o auto-init-11.s -fdump-rtl-expand, the *.gimple dump is:

=====
void foo (int n)
{
  int n.0;
  sizetype D.1950;
  bitsizetype D.1951;
  sizetype D.1952;
  bitsizetype D.1953;
  sizetype D.1954;
  int[0:D.1950] * arr.1;
  void * saved_stack.2;
  int arr[0:D.1950] [value-expr: *arr.1];

  saved_stack.2 = __builtin_stack_save ();
  try
    {
      n.0 = n;
      _1 = (long int) n.0;
      _2 = _1 + -1;
      _3 = (sizetype) _2;
      D.1950 = _3;
      _4 = (sizetype) n.0;
      _5 = (bitsizetype) _4;
      _6 = _5 * 32;
      D.1951 = _6;
      _7 = (sizetype) n.0;
      _8 = _7 * 4;
      D.1952 = _8;
      _9 = (sizetype) n.0;
      _10 = (bitsizetype) _9;
      _11 = _10 * 32;
      D.1953 = _11;
      _12 = (sizetype) n.0;
      _13 = _12 * 4;
      D.1954 = _13;
      arr.1 = __builtin_alloca_with_align (D.1954, 32);
      arr = .DEFERRED_INIT (D.1952, 2, 1);
      _14 = (*arr.1)[2];
      bar (_14);
      return;
    }
  finally
    {
      __builtin_stack_restore (saved_stack.2);
    }
}

====

You think that the above .DEFEERED_INIT is not correct?
It should be:

*arr.1 = .DEFERRED_INIT (D.1952. 2, 1);

?

> 
>> What do you mean by “such” decl? A decl whole “DECL_VALUE_EXPR(DECL)” is valid?
> 
> A 'decl' that has a DECL_VALUE_EXPR should not appear in the IL, it should
> always be refered to as its DECL_VALUE_EXPR.

Okay.

Qing
> 
> Richard.
> 
>> Qing
>>>