Re: [PATCH v3] c++: Move consteval folding to cp_fold_r

[Jason Merrill <jason@redhat.com>]

On 9/8/23 14:24, Marek Polacek wrote:
> On Thu, Sep 07, 2023 at 02:32:51PM -0400, Jason Merrill wrote:
>> On 9/7/23 11:23, Marek Polacek wrote:
>>> On Tue, Sep 05, 2023 at 04:36:34PM -0400, Jason Merrill wrote:
>>>> On 9/5/23 15:59, Marek Polacek wrote:
>>>>> On Tue, Sep 05, 2023 at 10:52:04AM -0400, Jason Merrill wrote:
>>>>>> On 9/1/23 13:23, Marek Polacek wrote:
>>>>>>> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
>>>>>>>
>>>>>>> -- >8 --
>>>>>>>
>>>>>>> In the review of P2564:
>>>>>>> <https://gcc.gnu.org/pipermail/gcc-patches/2023-August/628747.html>
>>>>>>> it turned out that in order to correctly handle an example in the paper,
>>>>>>> we should stop doing immediate evaluation in build_over_call and
>>>>>>> bot_replace, and instead do it in cp_fold_r.  This patch does that.
>>>>>>>
>>>>>>> Another benefit is that this is a pretty significant simplification, at
>>>>>>> least in my opinion.  Also, this fixes the c++/110997 ICE (but the test
>>>>>>> doesn't compile yet).
>>>>>>>
>>>>>>> The main drawback seems to be that cp_fold_r doesn't process as much
>>>>>>> code as we did before: uninstantiated templates
>>>>>>
>>>>>> That's acceptable, it's an optional diagnostic.
>>>>>>
>>>>>>> and things like "false ? foo () : 1".
>>>>>>
>>>>>> This is a problem.  Maybe we want cp_fold_r to recurse into the arms of a
>>>>>> COND_EXPR before folding them away?  Maybe only if we know we've seen an
>>>>>> immediate function?
>>>>>
>>>>> Unfortunately we had already thrown the dead branch away when we got to
>>>>> cp_fold_r.  I wonder if we have to adjust cxx_eval_conditional_expression
>>>>> to call cp_fold_r on the dead branch too,
>>>>
>>>> Hmm, I guess so.
>>>>
>>>>> perhaps with a new ff_ flag to skip the whole second switch in cp_fold_r?
>>>>
>>>> Or factor out the immediate function handling to a separate walk function
>>>> that cp_fold_r also calls?
>>>
>>> I did that.
>>>>> But then it's possible that the in_immediate_context checks have to stay.
>>>>
>>>> We can just not do the walk in immediate (or mce_true) context, like we
>>>> currently avoid calling cp_fold_function.
>>>
>>> Right.  Unfortunately I have to check even when mce_true, consider
>>>
>>>     consteval int bar (int i) { if (i != 1) throw 1; return 0; }
>>>     constexpr int a = 0 ? bar(3) : 3;
>>
>> I disagree; the call is in a manifestly constant-evaluated expression, and
>> so is now considered an immediate function context, and we should accept
>> that example.
> 
> Ack.  I was still living in pre-P2564 world.
>   
>>>> For mce_unknown I guess we'd want
>>>> to set *non_constant_p instead of giving an error.
>>>
>>> I did not do this because I haven't found a case where it would make
>>> a difference.
>>
>> I think it will given the above comment.
> 
> Correct.  For instance, in:
> 
>    consteval int bar (int i) { if (i != 1) throw 1; return 0; }
> 
>    constexpr int
>    foo (bool b)
>    {
>      return b ? bar (3) : 2;
>    }
> 
>    static_assert (foo (false) == 2);
> 
> we should complain only once.  I've implemented your suggestion to set
> *non_constant_p instead of giving an error for mce_unknown.
> 
>>> diff --git a/gcc/cp/constexpr.cc b/gcc/cp/constexpr.cc
>>> index 0ca4370deab..397d5c7ec3f 100644
>>> --- a/gcc/cp/constexpr.cc
>>> +++ b/gcc/cp/constexpr.cc
>>> @@ -2311,6 +2311,29 @@ cxx_dynamic_cast_fn_p (tree fndecl)
>>>    	  && CP_DECL_CONTEXT (fndecl) == abi_node);
>>>    }
>>> +/* Return true if we are in the body of a consteval function. > +   This is in addition to in_immediate_context because that
>>> +   uses current_function_decl which may not be available.  CTX is
>>> +   the current constexpr context.  */
>>> +
>>> +static bool
>>> +in_immediate_context (const constexpr_ctx *ctx)
>>> +{
>>> +  if (in_immediate_context ())
>>> +    return true;
>>
>> Can't we check for mce_true here instead of looking at the call chain?
> 
> Yes.
>   
>>> +/* A wrapper around cp_fold_immediate_r.  */
>>> +
>>> +void
>>> +cp_fold_immediate (tree *tp)
>>> +{
>>
>> Maybe return early if consteval isn't supported in the active standard?
> 
> Absolutely.
> 
> Bootstrapped/regtested on x86_64-pc-linux-gnu, ok for trunk?
> 
> -- >8 --
> In the review of P2564:
> <https://gcc.gnu.org/pipermail/gcc-patches/2023-August/628747.html>
> it turned out that in order to correctly handle an example in the paper,
> we should stop doing immediate evaluation in build_over_call and
> bot_replace, and instead do it in cp_fold_r.  This patch does that.
> 
> Another benefit is that this is a pretty significant simplification, at
> least in my opinion.  Also, this fixes the c++/110997 ICE (but the test
> doesn't compile yet).
> 
> The main drawback seems to be that cp_fold_r doesn't process
> uninstantiated templates.  We still have to handle things like
> "false ? foo () : 1".  To that end, I've added cp_fold_immediate, called
> on dead branches in cxx_eval_conditional_expression.
> 
> You'll see that I've reintroduced ADDR_EXPR_DENOTES_CALL_P here.  This
> is to detect
> 
>    *(&foo)) ()
>    (s.*&S::foo) ()
> 
> which were deemed ill-formed.
> 
> gcc/cp/ChangeLog:
> 
> 	* call.cc (build_over_call): Set ADDR_EXPR_DENOTES_CALL_P.  Don't handle
> 	immediate_invocation_p here.
> 	* constexpr.cc (in_immediate_context): New overload.
> 	(cxx_eval_call_expression): Use mce_true for DECL_IMMEDIATE_FUNCTION_P.
> 	(cxx_eval_conditional_expression): Call cp_fold_immediate.
> 	* cp-gimplify.cc (maybe_replace_decl): Make static.
> 	(cp_fold_r): Expand immediate invocations.
> 	(cp_fold_immediate_r): New.
> 	(cp_fold_immediate): New.
> 	* cp-tree.h (ADDR_EXPR_DENOTES_CALL_P): Define.
> 	(cp_fold_immediate): Declare.
> 	* tree.cc (bot_replace): Don't handle immediate invocations here.
> 
> libstdc++-v3/ChangeLog:
> 
> 	* testsuite/20_util/allocator/105975.cc: Add dg-error.
> 
> gcc/testsuite/ChangeLog:
> 
> 	* g++.dg/cpp23/consteval-if2.C: Add xfail.
> 	* g++.dg/cpp2a/consteval-memfn1.C: Adjust.
> 	* g++.dg/cpp2a/consteval11.C: Remove dg-message.
> 	* g++.dg/cpp2a/consteval3.C: Remove dg-message and dg-error.
> 	* g++.dg/cpp2a/consteval9.C: Remove dg-message.
> 	* g++.dg/cpp2a/consteval32.C: New test.
> 	* g++.dg/cpp2a/consteval33.C: New test.
> 	* g++.dg/cpp2a/consteval34.C: New test.
> 	* g++.dg/cpp2a/consteval35.C: New test.
> ---
>   gcc/cp/call.cc                                |  40 +------
>   gcc/cp/constexpr.cc                           |  23 +++-
>   gcc/cp/cp-gimplify.cc                         | 108 ++++++++++++++----
>   gcc/cp/cp-tree.h                              |  42 ++++---
>   gcc/cp/tree.cc                                |  23 +---
>   gcc/testsuite/g++.dg/cpp23/consteval-if2.C    |   3 +-
>   gcc/testsuite/g++.dg/cpp2a/consteval-memfn1.C |   7 ++
>   gcc/testsuite/g++.dg/cpp2a/consteval11.C      |  33 +++---
>   gcc/testsuite/g++.dg/cpp2a/consteval3.C       |   3 +-
>   gcc/testsuite/g++.dg/cpp2a/consteval32.C      |   4 +
>   gcc/testsuite/g++.dg/cpp2a/consteval33.C      |  34 ++++++
>   gcc/testsuite/g++.dg/cpp2a/consteval34.C      |  18 +++
>   gcc/testsuite/g++.dg/cpp2a/consteval35.C      |  10 ++
>   gcc/testsuite/g++.dg/cpp2a/consteval9.C       |   3 +-
>   .../testsuite/20_util/allocator/105975.cc     |   2 +-
>   15 files changed, 234 insertions(+), 119 deletions(-)
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/consteval32.C
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/consteval33.C
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/consteval34.C
>   create mode 100644 gcc/testsuite/g++.dg/cpp2a/consteval35.C
> 
> diff --git a/gcc/cp/constexpr.cc b/gcc/cp/constexpr.cc
> index 0ca4370deab..8c077aa22fe 100644
> --- a/gcc/cp/constexpr.cc
> +++ b/gcc/cp/constexpr.cc
> @@ -3837,12 +3841,25 @@ cxx_eval_conditional_expression (const constexpr_ctx *ctx, tree t,
>   				   boolean_type_node);
>       }
>     /* Don't VERIFY_CONSTANT the other operands.  */
> -  if (integer_zerop (val))
> +  const bool zero_p = integer_zerop (val);
> +  if (zero_p)
>       val = TREE_OPERAND (t, 2);
>     else
>       val = TREE_OPERAND (t, 1);
>     if (TREE_CODE (t) == IF_STMT && !val)
>       val = void_node;
> +
> +  if (!in_immediate_context ()
> +      /* P2564: a subexpression of a manifestly constant-evaluated expression
> +	 or conversion is an immediate function context.  */
> +      && ctx->manifestly_const_eval != mce_true

I might check this first as a tiny optimization.

> +      && cp_fold_immediate (&TREE_OPERAND (t, zero_p ? 1 : 2),
> +			    ctx->manifestly_const_eval))
> +    {
> +      *non_constant_p = true;
> +      return t;
> +    }
> +
>     /* A TARGET_EXPR may be nested inside another TARGET_EXPR, but still
>        serve as the initializer for the same object as the outer TARGET_EXPR,
>        as in
> diff --git a/gcc/cp/cp-gimplify.cc b/gcc/cp/cp-gimplify.cc
> index 206e791fcfd..9901513461c 100644
> --- a/gcc/cp/cp-gimplify.cc
> +++ b/gcc/cp/cp-gimplify.cc
> @@ -1000,7 +1000,7 @@ cp_genericize_target_expr (tree *stmt_p)
>      replacement when cp_folding TARGET_EXPR to preserve the invariant that
>      AGGR_INIT_EXPR_SLOT agrees with the enclosing TARGET_EXPR_SLOT.  */
>   
> -bool
> +static bool
>   maybe_replace_decl (tree *tp, tree decl, tree replacement)
>   {
>     if (!*tp || !VOID_TYPE_P (TREE_TYPE (*tp)))
> @@ -1029,44 +1029,73 @@ struct cp_genericize_data
>     bool handle_invisiref_parm_p;
>   };
>   
> -/* Perform any pre-gimplification folding of C++ front end trees to
> -   GENERIC.
> -   Note:  The folding of non-omp cases is something to move into
> -     the middle-end.  As for now we have most foldings only on GENERIC
> -     in fold-const, we need to perform this before transformation to
> -     GIMPLE-form.  */
> +/* A subroutine of cp_fold_r to handle immediate functions.  */
>   
>   static tree
> -cp_fold_r (tree *stmt_p, int *walk_subtrees, void *data_)
> +cp_fold_immediate_r (tree *stmt_p, int *walk_subtrees, void *data_)
>   {
> -  cp_fold_data *data = (cp_fold_data*)data_;
> +  auto data = static_cast<cp_fold_data *>(data_);
>     tree stmt = *stmt_p;
> -  enum tree_code code = TREE_CODE (stmt);
> +  const tsubst_flags_t complain = (data->flags == ff_none ? tf_none : tf_error);
>   
> -  switch (code)
> +  /* No need to look into types or unevaluated operands.
> +     NB: This affects cp_fold_r as well.  */
> +  if (TYPE_P (stmt) || unevaluated_p (TREE_CODE (stmt)))
>       {
> +      *walk_subtrees = 0;
> +      return NULL_TREE;
> +    }
> +
> +  switch (TREE_CODE (stmt))
> +    {
> +    /* Unfortunately we must handle code like
> +	 false ? bar () : 42
> +       where we have to check bar too.  */
> +    case COND_EXPR:
> +      if (cp_fold_immediate_r (&TREE_OPERAND (stmt, 1), walk_subtrees, data))
> +	return error_mark_node;
> +      if (TREE_OPERAND (stmt, 2)
> +	  && cp_fold_immediate_r (&TREE_OPERAND (stmt, 2), walk_subtrees, data))
> +	return error_mark_node;

Is this necessary?  Doesn't walk_tree already walk into the arms of 
COND_EXPR?

> +      break;
> +
>       case PTRMEM_CST:
>         if (TREE_CODE (PTRMEM_CST_MEMBER (stmt)) == FUNCTION_DECL
>   	  && DECL_IMMEDIATE_FUNCTION_P (PTRMEM_CST_MEMBER (stmt)))
>   	{
> -	  if (!data->pset.add (stmt))
> +	  if (!data->pset.add (stmt) && (complain & tf_error))
>   	    error_at (PTRMEM_CST_LOCATION (stmt),
>   		      "taking address of an immediate function %qD",
>   		      PTRMEM_CST_MEMBER (stmt));
>   	  stmt = *stmt_p = build_zero_cst (TREE_TYPE (stmt));

It looks like this will overwrite *stmt_p even if we didn't give an error.

> -	  break;
> +	  return error_mark_node;
>   	}
>         break;
>   
> +    /* Expand immediate invocations.  */
> +    case CALL_EXPR:
> +    case AGGR_INIT_EXPR:
> +      if (tree fn = cp_get_callee (stmt))
> +	if (TREE_CODE (fn) != ADDR_EXPR || ADDR_EXPR_DENOTES_CALL_P (fn))
> +	  if (tree fndecl = cp_get_fndecl_from_callee (fn, /*fold*/false))
> +	    if (DECL_IMMEDIATE_FUNCTION_P (fndecl))
> +	      {
> +		*stmt_p = stmt = cxx_constant_value (stmt, complain);

Likewise.

> +		if (stmt == error_mark_node)
> +		  return error_mark_node;
> +	      }
> +      break;
> +
>       case ADDR_EXPR:
>         if (TREE_CODE (TREE_OPERAND (stmt, 0)) == FUNCTION_DECL
>   	  && DECL_IMMEDIATE_FUNCTION_P (TREE_OPERAND (stmt, 0)))
>   	{
> -	  error_at (EXPR_LOCATION (stmt),
> -		    "taking address of an immediate function %qD",
> -		    TREE_OPERAND (stmt, 0));
> +	  if (complain & tf_error)
> +	    error_at (EXPR_LOCATION (stmt),
> +		      "taking address of an immediate function %qD",
> +		      TREE_OPERAND (stmt, 0));
>   	  stmt = *stmt_p = build_zero_cst (TREE_TYPE (stmt));

Likewise.

Jason