On Mon, 23 Jan 2023, Jakub Jelinek wrote:
> On Mon, Jan 23, 2023 at 09:25:50AM +0100, Jakub Jelinek via Gcc-patches wrote:
> > On Sun, Jan 22, 2023 at 07:19:07PM -0500, Jason Merrill wrote:
> > > > I was just trying to be careful, because unfortunately this spot
> > > > doesn't mean it really is only expanded in static var DECL_INITIAL,
> > > > it can make it into dynamic initializers, and most of DECL_VALUE_EXPRs
> > > > appear only in runtime code, otherwise we'd have much more of these
> > > > issues.
> > >
> > > But it shouldn't be harmful anywhere, right?
> > >
> > > > But if you think it is ok, I'll test tonight a version just with
> > > > if (!data->genericize && DECL_HAS_VALUE_EXPR_P (stmt)
> > >
> > > OK with that change.
> > >
> > > Though, actually, why not instead fix expand_expr_real_1 (and staticp) to
> > > look through DECL_VALUE_EXPR?
> >
> > I'm afraid that is way too late. GIMPLE optimizations don't try to
> > regimplify expressions they take from DECL_INITIAL of static vars, they
> > just unshare them and use them. So, if this is to be done in the generic
> > code, it would need to be done by cgraph around the time when it gimplifies
> > function if there is any such point for variables.
>
> I believe the right spot is record_reference, which is called through
> walk_tree from record_references_in_initializer which is called from
> varpool_node::analyze.
>
> The new tests as well as g++.dg/init/pr53932.C pass with it, will do full
> bootstrap/regtest soon.
>
> What do you think about this?
Guess it should work. Do we (accidentially?) do anything special
to static var initializers in nested functions? Can you think of
any other C/C++ construct that would have DECL_VALUE_EXPR in them?
Richard.
> 2023-01-23 Jakub Jelinek <ja...@redhat.com>
>
> PR c++/108474
> * cgraphbuild.cc: Include gimplify.h.
> (record_reference): Replace VAR_DECLs with DECL_HAS_VALUE_EXPR_P with
> their corresponding DECL_VALUE_EXPR expressions after unsharing.
>
> * cp-gimplify.cc (cp_fold_r): Revert 2023-01-19 changes.
>
> * g++.dg/cpp1z/decomp57.C: New test.
> * g++.dg/cpp1z/decomp58.C: New test.
>
> --- gcc/cgraphbuild.cc.jj 2023-01-02 09:32:34.889104620 +0100
> +++ gcc/cgraphbuild.cc 2023-01-23 12:10:35.042067571 +0100
> @@ -31,6 +31,7 @@ along with GCC; see the file COPYING3.
> #include "gimple-walk.h"
> #include "ipa-utils.h"
> #include "except.h"
> +#include "gimplify.h"
>
> /* Context of record_reference. */
> struct record_reference_ctx
> @@ -79,6 +80,17 @@ record_reference (tree *tp, int *walk_su
>
> if (VAR_P (decl))
> {
> + /* Replace vars with their DECL_VALUE_EXPR if any.
> + This is normally done during gimplification, but
> + static var initializers are never gimplified. */
> + if (DECL_HAS_VALUE_EXPR_P (decl))
> + {
> + tree *p;
> + for (p = tp; *p != decl; p = &TREE_OPERAND (*p, 0))
> + ;
> + *p = unshare_expr (DECL_VALUE_EXPR (decl));
> + return record_reference (tp, walk_subtrees, data);
> + }
> varpool_node *vnode = varpool_node::get_create (decl);
> ctx->varpool_node->create_reference (vnode, IPA_REF_ADDR);
> }
> --- gcc/cp/cp-gimplify.cc.jj 2023-01-23 11:47:49.889875804 +0100
> +++ gcc/cp/cp-gimplify.cc 2023-01-23 11:49:01.227841759 +0100
> @@ -1010,16 +1010,6 @@ cp_fold_r (tree *stmt_p, int *walk_subtr
> }
> break;
>
> - case VAR_DECL:
> - /* In initializers replace anon union artificial VAR_DECLs
> - with their DECL_VALUE_EXPRs, as nothing will do it later. */
> - if (DECL_ANON_UNION_VAR_P (stmt) && !data->genericize)
> - {
> - *stmt_p = stmt = unshare_expr (DECL_VALUE_EXPR (stmt));
> - break;
> - }
> - break;
> -
> default:
> break;
> }
> --- gcc/testsuite/g++.dg/cpp1z/decomp57.C.jj 2023-01-23 11:48:36.367202107
> +0100
> +++ gcc/testsuite/g++.dg/cpp1z/decomp57.C 2023-01-23 11:48:36.367202107
> +0100
> @@ -0,0 +1,27 @@
> +// PR c++/108474
> +// { dg-do link { target c++17 } }
> +
> +struct T { int i, j; };
> +T h;
> +auto [i, j] = h;
> +int &r = i;
> +int s = i;
> +int *t = &i;
> +
> +void
> +foo (int **p, int *q)
> +{
> + static int &u = i;
> + static int v = i;
> + static int *w = &i;
> + int &x = i;
> + int y = i;
> + int *z = &i;
> + *p = &i;
> + *q = i;
> +}
> +
> +int
> +main ()
> +{
> +}
> --- gcc/testsuite/g++.dg/cpp1z/decomp58.C.jj 2023-01-23 11:48:36.367202107
> +0100
> +++ gcc/testsuite/g++.dg/cpp1z/decomp58.C 2023-01-23 11:48:36.367202107
> +0100
> @@ -0,0 +1,39 @@
> +// PR c++/108474
> +// { dg-do link { target c++17 } }
> +
> +namespace std {
> + template <typename T> struct tuple_size;
> + template <int, typename> struct tuple_element;
> +}
> +
> +struct A {
> + int i;
> + template <int I> int& get() { return i; }
> +};
> +
> +template <> struct std::tuple_size <A> { static const int value = 2; };
> +template <int I> struct std::tuple_element <I, A> { using type = int; };
> +
> +struct A a;
> +auto [i, j] = a;
> +int &r = i;
> +int s = i;
> +int *t = &i;
> +
> +void
> +foo (int **p, int *q)
> +{
> + static int &u = i;
> + static int v = i;
> + static int *w = &i;
> + int &x = i;
> + int y = i;
> + int *z = &i;
> + *p = &i;
> + *q = i;
> +}
> +
> +int
> +main ()
> +{
> +}
>
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE Software Solutions Germany GmbH, Frankenstrasse 146, 90461 Nuernberg,
Germany; GF: Ivo Totev, Andrew Myers, Andrew McDonald, Boudien Moerman;
HRB 36809 (AG Nuernberg)
