Re: [RFC][Patch][middle-end/PR102359]Not add initialization for READONLY variables with -ftrivial-auto-var-init

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

> On Sep 30, 2021, at 1:54 AM, Richard Biener <rguenther@suse.de> wrote:
> 
> On Thu, 30 Sep 2021, Jason Merrill wrote:
> 
>> On 9/29/21 17:30, Qing Zhao wrote:
>>> Hi,
>>> 
>>> PR102359 (ICE gimplification failed since  r12-3433-ga25e0b5e6ac8a77a)
>>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102359
>>> 
>>> Is due to -ftrivial-auto-var-init adding initialization for READONLY
>>> variable “this” in the following routine: (t.cpp.005t.original)
>>> 
>>> =======
>>> 
>>> ;; Function A::foo()::<lambda()> (null)
>>> ;; enabled by -tree-original
>>> 
>>> {
>>>  const struct A * const this [value-expr: &__closure->__this];
>>>    const struct A * const this [value-expr: &__closure->__this];
>>>  return <retval> = (double) ((const struct A *) this)->a;
>>> }
>>> =======
>>> 
>>> However, in the above routine, “this” is NOT marked as READONLY, but its
>>> value-expr "&__closure->__this” is marked as READONLY.
>>> 
>>> There are two major issues:
>>> 
>>> 1. In the routine “is_var_need_auto_init”, we should exclude “decl” that is
>>> marked as READONLY;
>>> 2. In the C++ FE, “this” should be marked as READONLY.
>>> 
>>> The idea solution will be:
>>> 
>>> 1. Fix “is_var_need_auto_init” to exclude TREE_READONLY (decl);
>>> 2. Fix C++ FE to mark “this” as TREE_READONLY (decl)==true;
>>> 
>>> Not sure whether it’s hard for C++ FE to fix the 2nd issue or not?
>>> 
>>> In the case it’s not a quick fix in C++FE, I proposed the following fix in
>>> middle end:
>>> 
>>> Let me know your comments or suggestions on this.
>>> 
>>> Thanks a lot for the help.
>> 
>> I'd think is_var_need_auto_init should be false for any variable with
>> DECL_HAS_VALUE_EXPR_P, as they aren't really variables, just ways of naming
>> objects that are initialized elsewhere.
> 
> IIRC handing variables with DECL_HAS_VALUE_EXPR_P is necessary to
> auto-init VLAs, otherwise I tend to agree - would we handle those
> when we see a DECL_EXPR then?

The current implementation is:


gimplify_decl_expr:


For each DECL_EXPR “decl”

   If (VAR_P (decl) && !DECL_EXTERNAL (decl))
     {
	if (is_vla (decl))
          gimplify_vla_decl (decl, …);      /* existing handling: create a VALUE_EXPR for this vla decl*/

	…
	if (has_explicit_init (decl))
	  {
	   …;     /* existing handling.  */
	  }
	else if (is_var_need_auto_init (decl))      /*. New code. */
	  {
	    gimple_add_init_for_auto_var (….);   /*  new code.  */
	    ...
	  }
     }


Since the “DECL_VALUE_EXPR (decl)” is NOT a DECL_EXPR, it will not be scanned and added initialization. 

if we do not add initialization for a decl that has DECL_VALUE_EXPR, then the “DECL_VALUE_EXPR (decl)” will not be added an initialization either.  We will miss adding initializations for these decls. 

So, I think that the current implementation is correct. 

And if C++ FE will not mark “this” as READONLY, only mark DECL_VALUE_EXPR(this) as READONLY, the proposed fix is correct too.

Let me know your opinion on this.

Thanks.

Qing

> 
>> 
>>> Qing
>>> 
>>> ==============================
>>> From 0a5982cd61bc4610655d3df00ae8d2fbcb3c8e9b Mon Sep 17 00:00:00 2001
>>> From: Qing Zhao <qing.zhao@oracle.com>
>>> Date: Wed, 29 Sep 2021 20:49:59 +0000
>>> Subject: [PATCH] Fix PR102359
>>> 
>>> ---
>>> gcc/gimplify.c                  | 15 +++++++++++++++
>>> gcc/testsuite/g++.dg/pr102359.C | 13 +++++++++++++
>>> 2 files changed, 28 insertions(+)
>>> create mode 100644 gcc/testsuite/g++.dg/pr102359.C
>>> 
>>> diff --git a/gcc/gimplify.c b/gcc/gimplify.c
>>> index 1067113b1639..a2587869b35d 100644
>>> --- a/gcc/gimplify.c
>>> +++ b/gcc/gimplify.c
>>> @@ -1819,12 +1819,27 @@ gimple_add_padding_init_for_auto_var (tree decl,
>>> bool is_vla,
>>>   gimplify_seq_add_stmt (seq_p, call);
>>> }
>>> 
>>> +/* Return true if the DECL is READONLY.
>>> +   This is to workaround a C++ FE bug that only mark the value_expr of
>>> "this"
>>> +   as readonly but does not mark "this" as readonly.
>>> +   C++ FE should fix this issue before replacing this routine with
>>> +   TREE_READONLY (decl).  */
>>> +
>>> +static bool
>>> +is_decl_readonly (tree decl)
>>> +{
>>> +  return (TREE_READONLY (decl)
>>> +	   || (DECL_HAS_VALUE_EXPR_P (decl)
>>> +		 && TREE_READONLY (DECL_VALUE_EXPR (decl))));
>>> +}
>>> +
>>> /* Return true if the DECL need to be automaticly initialized by the
>>>    compiler.  */
>>> static bool
>>> is_var_need_auto_init (tree decl)
>>> {
>>>   if (auto_var_p (decl)
>>> +      && !is_decl_readonly (decl)
>>>       && (flag_auto_var_init > AUTO_INIT_UNINITIALIZED)
>>>       && (!lookup_attribute ("uninitialized", DECL_ATTRIBUTES (decl)))
>>>       && !is_empty_type (TREE_TYPE (decl)))
>>> diff --git a/gcc/testsuite/g++.dg/pr102359.C
>>> b/gcc/testsuite/g++.dg/pr102359.C
>>> new file mode 100644
>>> index 000000000000..da643cde7bed
>>> --- /dev/null
>>> +++ b/gcc/testsuite/g++.dg/pr102359.C
>>> @@ -0,0 +1,13 @@
>>> +/* PR middle-end/102359 ICE gimplification failed since
>>> +   r12-3433-ga25e0b5e6ac8a77a.  */
>>> +/* { dg-do compile } */
>>> +/* { dg-options "-ftrivial-auto-var-init=zero" } */
>>> +/* { dg-require-effective-target c++17 } */
>>> +
>>> +struct A {
>>> +  double a = 111;
>>> +  auto foo() {
>>> +    return [*this] { return a; };
>>> +  }
>>> +};
>>> +int X = A{}.foo()();
>>> 
>> 
>> 
> 
> -- 
> Richard Biener <rguenther@suse.de>
> SUSE Software Solutions Germany GmbH, Maxfeldstrasse 5, 90409 Nuernberg,
> Germany; GF: Felix Imendörffer; HRB 36809 (AG Nuernberg)