[PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Jakub Jelinek]

Hi!

As the testcase shows, the C FE can call relayout_decl even on FIELD_DECLs
in certain cases.  Trying to call only layout_decl on FIELD_DECL and
relayout_decl on other decls would be insufficient, we'd need to repeat
there most of the relayout_decl code (except for SET_DECL_RTL, which
FIELD_DECLs don't have).

So I think it is better to allow relayout_decl also on FIELD_DECLs.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2016-08-06  Jakub Jelinek  <jakub@redhat.com>

	PR c/72816
	* stor-layout.c (layout_decl): Fix up formatting.
	(relayout_decl): Allow DECL to be FIELD_DECL.

	* gcc.dg/pr72816.c: New test.

--- gcc/stor-layout.c.jj	2016-08-06 12:11:56.000000000 +0200
+++ gcc/stor-layout.c	2016-08-06 13:04:43.662532852 +0200
@@ -596,7 +596,7 @@ layout_decl (tree decl, unsigned int kno
     return;
 
   gcc_assert (code == VAR_DECL || code == PARM_DECL || code == RESULT_DECL
-	      || code == TYPE_DECL ||code == FIELD_DECL);
+	      || code == TYPE_DECL || code == FIELD_DECL);
 
   rtl = DECL_RTL_IF_SET (decl);
 
@@ -768,8 +768,8 @@ layout_decl (tree decl, unsigned int kno
     }
 }
 
-/* Given a VAR_DECL, PARM_DECL or RESULT_DECL, clears the results of
-   a previous call to layout_decl and calls it again.  */
+/* Given a VAR_DECL, PARM_DECL, RESULT_DECL, or FIELD_DECL, clears the
+   results of a previous call to layout_decl and calls it again.  */
 
 void
 relayout_decl (tree decl)
@@ -778,7 +778,8 @@ relayout_decl (tree decl)
   DECL_MODE (decl) = VOIDmode;
   if (!DECL_USER_ALIGN (decl))
     SET_DECL_ALIGN (decl, 0);
-  SET_DECL_RTL (decl, 0);
+  if (DECL_RTL_SET_P (decl))
+    SET_DECL_RTL (decl, 0);
 
   layout_decl (decl, 0);
 }
--- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-06 13:06:45.046003282 +0200
+++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-06 13:07:57.217093845 +0200
@@ -0,0 +1,9 @@
+/* PR c/72816 */
+/* { dg-do compile } */
+/* { dg-options "-std=gnu11" } */
+
+typedef const int A[];
+struct S {
+  int a;
+  A b;	/* { dg-error "array size missing" } */
+};

	Jakub

Re: [PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Richard Biener]

On August 6, 2016 9:09:08 PM GMT+02:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>As the testcase shows, the C FE can call relayout_decl even on
>FIELD_DECLs
>in certain cases.  Trying to call only layout_decl on FIELD_DECL and
>relayout_decl on other decls would be insufficient, we'd need to repeat
>there most of the relayout_decl code (except for SET_DECL_RTL, which
>FIELD_DECLs don't have).
>
>So I think it is better to allow relayout_decl also on FIELD_DECLs.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK.

Thanks,
Richard.

>2016-08-06  Jakub Jelinek  <jakub@redhat.com>
>
>	PR c/72816
>	* stor-layout.c (layout_decl): Fix up formatting.
>	(relayout_decl): Allow DECL to be FIELD_DECL.
>
>	* gcc.dg/pr72816.c: New test.
>
>--- gcc/stor-layout.c.jj	2016-08-06 12:11:56.000000000 +0200
>+++ gcc/stor-layout.c	2016-08-06 13:04:43.662532852 +0200
>@@ -596,7 +596,7 @@ layout_decl (tree decl, unsigned int kno
>     return;
> 
>gcc_assert (code == VAR_DECL || code == PARM_DECL || code ==
>RESULT_DECL
>-	      || code == TYPE_DECL ||code == FIELD_DECL);
>+	      || code == TYPE_DECL || code == FIELD_DECL);
> 
>   rtl = DECL_RTL_IF_SET (decl);
> 
>@@ -768,8 +768,8 @@ layout_decl (tree decl, unsigned int kno
>     }
> }
> 
>-/* Given a VAR_DECL, PARM_DECL or RESULT_DECL, clears the results of
>-   a previous call to layout_decl and calls it again.  */
>+/* Given a VAR_DECL, PARM_DECL, RESULT_DECL, or FIELD_DECL, clears the
>+   results of a previous call to layout_decl and calls it again.  */
> 
> void
> relayout_decl (tree decl)
>@@ -778,7 +778,8 @@ relayout_decl (tree decl)
>   DECL_MODE (decl) = VOIDmode;
>   if (!DECL_USER_ALIGN (decl))
>     SET_DECL_ALIGN (decl, 0);
>-  SET_DECL_RTL (decl, 0);
>+  if (DECL_RTL_SET_P (decl))
>+    SET_DECL_RTL (decl, 0);
> 
>   layout_decl (decl, 0);
> }
>--- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-06 13:06:45.046003282
>+0200
>+++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-06 13:07:57.217093845 +0200
>@@ -0,0 +1,9 @@
>+/* PR c/72816 */
>+/* { dg-do compile } */
>+/* { dg-options "-std=gnu11" } */
>+
>+typedef const int A[];
>+struct S {
>+  int a;
>+  A b;	/* { dg-error "array size missing" } */
>+};
>
>	Jakub

Re: [PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Joseph Myers]

On Sat, 6 Aug 2016, Jakub Jelinek wrote:

> --- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-06 13:06:45.046003282 +0200
> +++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-06 13:07:57.217093845 +0200
> @@ -0,0 +1,9 @@
> +/* PR c/72816 */
> +/* { dg-do compile } */
> +/* { dg-options "-std=gnu11" } */
> +
> +typedef const int A[];
> +struct S {
> +  int a;
> +  A b;	/* { dg-error "array size missing" } */
> +};

As far as I can tell, this is actually valid code that should not produce 
an error; the type of a flexible array member can be given by a typedef, 
and I see nothing to disallow it being given by a typedef for an array of 
qualified type.  Note that both the version of this test without const, 
and the version with const but not using a typedef, are accepted.

-- 
Joseph S. Myers
joseph@codesourcery.com

Re: [PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Marek Polacek]

On Mon, Aug 08, 2016 at 11:04:32AM +0000, Joseph Myers wrote:
> On Sat, 6 Aug 2016, Jakub Jelinek wrote:
> 
> > --- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-06 13:06:45.046003282 +0200
> > +++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-06 13:07:57.217093845 +0200
> > @@ -0,0 +1,9 @@
> > +/* PR c/72816 */
> > +/* { dg-do compile } */
> > +/* { dg-options "-std=gnu11" } */
> > +
> > +typedef const int A[];
> > +struct S {
> > +  int a;
> > +  A b;	/* { dg-error "array size missing" } */
> > +};
> 
> As far as I can tell, this is actually valid code that should not produce 
> an error; the type of a flexible array member can be given by a typedef, 
> and I see nothing to disallow it being given by a typedef for an array of 
> qualified type.  Note that both the version of this test without const, 
> and the version with const but not using a typedef, are accepted.

I was looking into this last week a bot and the problem seems to be that for
typedefs the array 'b' doesn't have TYPE_DOMAIN, so its type is in fact
const int b[<unknown>] (so the code tries to deduce the size of the array
which fails) whereas without typedef the type is const int b[0:].  I thought
this discrepancy was the core issue here.

	Marek

Re: [PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Jakub Jelinek]

On Mon, Aug 08, 2016 at 11:04:32AM +0000, Joseph Myers wrote:
> On Sat, 6 Aug 2016, Jakub Jelinek wrote:
> 
> > --- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-06 13:06:45.046003282 +0200
> > +++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-06 13:07:57.217093845 +0200
> > @@ -0,0 +1,9 @@
> > +/* PR c/72816 */
> > +/* { dg-do compile } */
> > +/* { dg-options "-std=gnu11" } */
> > +
> > +typedef const int A[];
> > +struct S {
> > +  int a;
> > +  A b;	/* { dg-error "array size missing" } */
> > +};
> 
> As far as I can tell, this is actually valid code that should not produce 
> an error; the type of a flexible array member can be given by a typedef, 
> and I see nothing to disallow it being given by a typedef for an array of 
> qualified type.  Note that both the version of this test without const, 
> and the version with const but not using a typedef, are accepted.

Indeed.
The following untested patch fixes the issue for me.  The problem was that
when we create the distinct type with TYPE_DOMAIN [0:], if
TYPE_MAIN_VARIANT (orig_qual_type) is TYPE_MAIN_VARIANT (type) before we
do this, then c_build_qualified_type will return that orig_qual_type and
the TYPE_DOMAIN [0:] is gone again.  I think for orig_qual_indirect > 0
we are ok, in that case c_build_qualified_type will never return that and
as the flexible array member is the outermost ARRAY_TYPE, we should be fine.
But perhaps if (orig_qual_indirect > 0 && orig_qual_type), we never reach
the flexible array member through typedef else if, so perhaps
just unconditional orig_qual_type = NULL_TREE; in that else if would work too.
Which of these two versions do you prefer (or something else)?
Shall I revert the stor-layout.c change which might be unnecessary (though
shouldn't hurt)?

2016-08-10  Jakub Jelinek  <jakub@redhat.com>

	PR c/72816
	* c-decl.c (grokdeclarator): When adding TYPE_DOMAIN for flexible
	array member through typedef, for orig_qual_indirect == 0 clear
	orig_qual_type.

	* gcc.dg/pr72816.c: New test.

--- gcc/c/c-decl.c.jj	2016-08-06 12:11:47.000000000 +0200
+++ gcc/c/c-decl.c	2016-08-10 17:07:10.675144603 +0200
@@ -6710,6 +6710,8 @@ grokdeclarator (const struct c_declarato
 	    type = build_distinct_type_copy (TYPE_MAIN_VARIANT (type));
 	    TYPE_DOMAIN (type) = build_range_type (sizetype, size_zero_node,
 						   NULL_TREE);
+	    if (orig_qual_indirect == 0)
+	      orig_qual_type = NULL_TREE;
 	  }
 	type = c_build_qualified_type (type, type_quals, orig_qual_type,
 				       orig_qual_indirect);
--- gcc/testsuite/gcc.dg/pr72816.c.jj	2016-08-07 11:48:34.000000000 +0200
+++ gcc/testsuite/gcc.dg/pr72816.c	2016-08-10 17:08:21.153266419 +0200
@@ -5,5 +5,5 @@
 typedef const int A[];
 struct S {
   int a;
-  A b;	/* { dg-error "array size missing" } */
+  A b;
 };


	Jakub

Re: [PATCH] Allow relayout_decl on FIELD_DECLs (PR c/72816)

[Joseph Myers]

On Wed, 10 Aug 2016, Jakub Jelinek wrote:

> 2016-08-10  Jakub Jelinek  <jakub@redhat.com>
> 
> 	PR c/72816
> 	* c-decl.c (grokdeclarator): When adding TYPE_DOMAIN for flexible
> 	array member through typedef, for orig_qual_indirect == 0 clear
> 	orig_qual_type.
> 
> 	* gcc.dg/pr72816.c: New test.

OK (subject to testing).

-- 
Joseph S. Myers
joseph@codesourcery.com