Bug 120776 - [C++26] P1306R5 - Expansion statements
Summary: [C++26] P1306R5 - Expansion statements
Status: RESOLVED FIXED
Alias: None
Product: gcc
Classification: Unclassified
Component: c++ (show other bugs)
Version: 16.0
: P3 enhancement
Target Milestone: 16.0
Assignee: Jakub Jelinek
URL:
Keywords: c++26
Depends on:
Blocks: c++26-core
  Show dependency treegraph
 
Reported: 2025-06-23 09:04 UTC by Jakub Jelinek
Modified: 2025-08-23 17:40 UTC (History)
3 users (show)

See Also:
Host:
Target:
Build:
Known to work:
Known to fail:
Last reconfirmed: 2025-06-23 00:00:00


Attachments
gcc16-pr120776-wip.patch (3.66 KB, patch)
2025-07-02 12:26 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (9.90 KB, patch)
2025-07-08 10:25 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (12.30 KB, patch)
2025-07-08 16:58 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (14.34 KB, patch)
2025-07-11 16:02 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (16.03 KB, patch)
2025-07-16 15:38 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (16.18 KB, patch)
2025-07-17 12:05 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (19.24 KB, patch)
2025-07-18 11:53 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776-wip.patch (21.23 KB, patch)
2025-07-18 17:23 UTC, Jakub Jelinek
Details | Diff
gcc16-pr120776.patch (24.02 KB, patch)
2025-07-24 16:06 UTC, Jakub Jelinek
Details | Diff

Note You need to log in before you can comment on or make changes to this bug.
Description Jakub Jelinek 2025-06-23 09:04:21 UTC
See <https://wg21.link/P1306R5>.
Comment 1 Andrew Pinski 2025-06-23 14:30:59 UTC
.
Comment 2 Jakub Jelinek 2025-07-02 12:26:58 UTC
Created attachment 61786 [details]
gcc16-pr120776-wip.patch

Very early WIP, this can just parse stuff, on say
struct S { int a, b, c; };
struct T { long long a, b, c; };
constexpr S d = { 1, 2, 3 }, e = { 4, 5, 6 }, f = { 7, 8, 9 };
constexpr T j = { 10LL, 11LL, 12LL };

void
foo ()
{
  template for (auto &[g, h, i] : { d, e, f, j })
//    __builtin_printf ("%d %d %d\n", g, h, i);
    ;
}
it can parse it (and throws it away right now).  With the printf uncommented (where the right behavior is then to warn about %d vs. %lld 3 times for the 4th iteration) it errors because decl_dependent_p doesn't have special case for expansion-statement for-range-declaration vars/structured bindings.  Guess even
before parsing the body I'll need to figure out which of the cases it is and depending on that whether the vars/sbs are dependent or not.
Comment 3 Jakub Jelinek 2025-07-08 10:25:57 UTC
Created attachment 61817 [details]
gcc16-pr120776-wip.patch

Some further progress.  Current main problem is instantiation of expansion stmt body.
And really nothing is done so far about destructuring expansion stmts, break/continue isn't handled etc.
Comment 4 Jakub Jelinek 2025-07-08 16:58:30 UTC
Created attachment 61819 [details]
gcc16-pr120776-wip.patch

This version can do destructuring expansion stmts as well (but for none of the kinds if for-range-declaration is a structured binding).
Comment 5 Jakub Jelinek 2025-07-11 16:02:22 UTC
Created attachment 61842 [details]
gcc16-pr120776-wip.patch

To my big surprise, the instantiation seems to work even for variables defined
outside of the construct, will need to figure out why.
Anyway, this is the first version that can handle some examples of expanion
stmts at runtime (including enumerating, iterating and destructuring); for now
only when defined outside of templates.
Comment 6 Jakub Jelinek 2025-07-16 15:38:03 UTC
Created attachment 61885 [details]
gcc16-pr120776-wip.patch

Updated WIP patch.

As the expansion-stmt2.C testcase shows, if expansion-init is dependent and during first tsubst_stmt of it not dependent, those cases seem to work quite well (will need to work on diagnostics for control-flow-limited statement, testcase for no labels in the body etc.

Anyway, still have a couple of major issues:
1) the instantiation when not inside of a template works weirdly; this is because
   tsubst_expr has:
      else if (local_variable_p (t)
               && ((r = retrieve_local_specialization (t))
                   || TREE_CODE (t) == PARM_DECL
                   || uses_template_parms (DECL_CONTEXT (t))))
   for non-DECL_LOCAL_DECL_P VAR_DECLs without DECL_TEMPLATE_INFO.
   The code right now registers local specialization of the for-range-declaration
   (and structured bindings in there if any), but nothing else.  As in non-templates
   uses_template_parms (DECL_CONTEXT (t)) is false, it means only for-range-declaration
   is locally specialized and the rest is used directly (except PARM_DECLs).
   Now, we want PARM_DECLs and VAR_DECLs defined outside of the TEMPLATE_FOR_STMT
   to map to themselves, but VAR_DECLs defined inside of the TEMPLATE_FOR_STMT to be
   specialized.
   Shall I use some flag next to this || uses_template_parms (e.g. instead of 1-bit
   scope_chain->expansion_stmt meant primarily for the control-flow-limited diagnostics
   use 2 bits to differentiate between non-template instantiation during parsing
   vs. the rest?
2) from the paper, it is unclear when the lowering of expansion statement needs to
   happen; it says that for-range-declaration is not type dependent if
   expansion-init of iterating expansion stmt is not type dependent and I think I
   handle that already doing auto deduction in that case.  But is there some strong
   reason why the expansion statement would need to be lowered still when
   processing_template_decl?  Not really sure how to (partially) instantiate it
   in that case
3) the TEMPLATE_FOR_STMT part doesn't have the if (!finish_expansion_stmt (...)) case
   implemented, I guess I can handle it easily, but I'm afraid I don't understand
   how to construct a testcase, where something is only partially instantiated
   and the expansion-initializer is still type-dependent during that instantiation
   and it will need another instantiation later on
Comment 7 Jakub Jelinek 2025-07-17 08:46:48 UTC
Regarding 1), I think the problem is that if there are automatic vars defined in the expansion stmt body, there is no DECL_EXPR created for them.
The problem is start_decl doing:
6214	  if (processing_template_decl)
6215	    decl = push_template_decl (decl);
There is ++processing_template_decl around parsing of the expansion stmt bodies, but
push_template_decl then does:
5893	  if (decl == error_mark_node || !current_template_parms)
5894	    return error_mark_node;
and because I haven't created current_template_parms if expansion stmt is not inside of a template, this returns error_mark_node and a lot of things break because of that obviously.
So probably I need to arrange for current_template_parms to be set and restored back as well.
Comment 8 Jakub Jelinek 2025-07-17 12:05:39 UTC
Created attachment 61896 [details]
gcc16-pr120776-wip.patch

I think I've 1) and 2) resolved and for 3) testcase which currently FAILs, so working on that next.
Comment 9 Jakub Jelinek 2025-07-18 11:53:18 UTC
Created attachment 61911 [details]
gcc16-pr120776-wip.patch

Further progress.
Comment 10 Jakub Jelinek 2025-07-18 17:23:08 UTC
Created attachment 61912 [details]
gcc16-pr120776-wip.patch

Latest state.
Still need more testcases (though admittedly something will just not work until P2686R4 is implemented fully), figure out what to do about range extension (see https://github.com/cplusplus/CWG/issues/724 for some issues), whether using ptrdiff_t(i) that the patch currently uses for iterating expansion statements is ok (see https://github.com/cplusplus/CWG/issues/725 ) and at least add a testcase for redeclarations (both something declared in init-stmt with for-range-declaration and possibly (see https://github.com/cplusplus/CWG/issues/729 ) also in the body.  I think the code currently does diagnose it even in the body, but will need testcases.
Also tests for tuple based structured bindings (both in destructuring expansion statements and when structured bindings are used in for-range-declaration).
Also tests for what happens when it is not iterating and is not valid destructuring either.
Comment 11 Jakub Jelinek 2025-07-24 16:06:22 UTC
Created attachment 61958 [details]
gcc16-pr120776.patch

Full untested patch.
Comment 12 GCC Commits 2025-08-13 20:10:27 UTC
The master branch has been updated by Jakub Jelinek <jakub@gcc.gnu.org>:

https://gcc.gnu.org/g:458773ac7bca792db044ad6c241f566c9ba87cb7

commit r16-3192-g458773ac7bca792db044ad6c241f566c9ba87cb7
Author: Jakub Jelinek <jakub@redhat.com>
Date:   Wed Aug 13 22:07:27 2025 +0200

    c++: Implement C++26 P1306R5 - Expansion statements [PR120776]
    
    The following patch implements the C++26 P1306R5 - Expansion statements
    paper.
    When expansion statements are used outside of templates, the lowering
    of the statement uses push_tinst_level_loc and instantiates the body
    multiple times, otherwise when the new TEMPLATE_FOR_STMT statement is
    being instantiated and !processing_template_decl, it instantiates the
    body several times with just local_specialization_stack around each
    iteration but with the original args.
    Because the lowering of these statements is mostly about instantiation,
    I've put the lowering code into pt.cc rather than semantics.cc.
    Only destructuring expansion statements currently use in the patch
    temporary lifetime extension which matches the proposed resolution of
    https://cplusplus.github.io/CWG/issues/3043.html
    I'm not sure what will CWG decide about that if there will be some
    temporary lifetime extension for enumerating expansion statements and if
    yes, under what exact rules (e.g. whether it extends all the temporaries
    across one iteration of the body, or only if a reference is initialized
    or nothing at all).  And for iterating expansion statements, I think I
    don't understand the P2686R4 rules well yet, I think if the
    expansion-initializer is used in static constexpr rvalue reference, then
    it isn't needed, but not sure if it won't be needed if static would be
    dropped (whether
    struct S { constexpr S () : s (0) {} constexpr ~S () {} int s; };
    struct T { const S &t, &u; };
    void foo () { constexpr T t = { S {}, S {} }; use (t.t, t.u); }
    is ok under P2686R4; though without constexpr before T I see S::~S () being
    called after use, not at the end of the t declaration, so maybe it is
    fine also without static).
    As per
    https://cplusplus.github.io/CWG/issues/3044.html
    the patch uses build_int_cst (ptrdiff_type_node, i) to create second
    operand of begin + i and doesn't lookup overloaded comma operator (note, I'm
    actually not even creating a lambda there, just using TARGET_EXPRs).
    I guess my preference would be dropping those 4 static keywords from
    [stmt.expand] but the patch does use those for now and it won't be possible
    to change that until the rest of P2686R4 is implemented.
    As per
    https://cplusplus.github.io/CWG/issues/3045.html
    it treats sk_template_for like sk_for for the purpose of redeclaration of
    vars in the body but doesn't yet reject [[fallthrough]]; in the expansion
    stmt body (when not nested in another switch).
    I'm not sure if cp_perform_range_for_lookup used in the patch is exactly
    what we want for the https://eel.is/c++draft/stmt.expand#3.2
    - it does finish_call_expr on the perform_koenig_lookup as well, shouldn't
    for the decision whether it is iterating or destructing (i.e. tf_none)
    just call perform_koenig_lookup and check if it found some
    FUNCTION_DECL/OVERLOAD/TEMPLATE_DECL?
    cp_decomp_size in the patch has tsubst_flags_t argument and attempts to be
    SFINAE friendly, even when it isn't needed strictly for this patch.
    This is with PR96185 __builtin_structured_binding_size implementation in
    mind (to follow clang).
    The new TEMPLATE_FOR_STMT statement is expected to be lowered to something
    that doesn't use the statement at all, I've implemented break/continue
    discovery in the body, so all I needed was to punt on TEMPLATE_FOR_STMT
    in potential_constant_expression_1 so that we don't try to constant evaluate
    it when it is still dependent (and cxx_eval_constant_expression rejects
    it without any extra code).
    I think only enumerating and iterating expansion statements can have zero
    iteration, because for destructuring ones it doesn't use a structured
    binding pack and so valid structured binding has to have at least one
    iteration.  Though
    https://cplusplus.github.io/CWG/issues/3048.html
    could change that, this patch currently rejects it though.
    
    2025-08-13  Jakub Jelinek  <jakub@redhat.com>
    
            PR c++/120776
    gcc/c-family/
            * c-cppbuiltin.cc (c_cpp_builtins): Predefine
            __cpp_expansion_statements=202506L for C++26.
    gcc/cp/
            * cp-tree.def: Implement C++26 P1306R5 - Expansion statements.
            (TEMPLATE_FOR_STMT): New tree code.
            * cp-tree.h (struct saved_scope): Add expansion_stmt.
            (in_expansion_stmt): Define.
            (TEMPLATE_FOR_DECL, TEMPLATE_FOR_EXPR, TEMPLATE_FOR_BODY,
            TEMPLATE_FOR_SCOPE, TEMPLATE_FOR_INIT_STMT): Define.
            (struct tinst_level): Adjust comment.
            (cp_decomp_size, finish_expansion_stmt, do_pushlevel,
            cp_build_range_for_decls, build_range_temp,
            cp_perform_range_for_lookup, begin_template_for_scope): Declare.
            (finish_range_for_stmt): Remove declaration.
            * cp-objcp-common.cc (cp_common_init_ts): Handle TEMPLATE_FOR_STMT.
            * name-lookup.h (enum scope_kind): Add sk_template_for enumerator.
            (struct cp_binding_level): Enlarge kind bitfield from 4 to 5 bits.
            Adjust comment with remaining space bits.
            * name-lookup.cc (check_local_shadow): Handle sk_template_for like
            sk_for.
            (cp_binding_level_descriptor): Add entry for sk_template_for.
            (begin_scope): Handle sk_template_for.
            * parser.h (IN_EXPANSION_STMT): Define.
            * parser.cc (cp_debug_parser): Print IN_EXPANSION_STMT bit.
            (cp_parser_lambda_expression): Temporarily clear in_expansion_stmt.
            (cp_parser_statement): Handle RID_TEMPLATE followed by RID_FOR for
            C++11.
            (cp_parser_label_for_labeled_statement): Complain about named labels
            inside of expansion stmt body.
            (cp_hide_range_decl): New function.
            (cp_parser_range_for): Use it.  Adjust do_range_for_auto_deduction
            caller.  Remove second template argument from auto_vecs bindings and
            names.
            (build_range_temp): No longer static.
            (do_range_for_auto_deduction): Add expansion_stmt argument.
            (cp_build_range_for_decls): New function.
            (cp_convert_range_for): Use it.  Call cp_perform_range_for_lookup
            rather than cp_parser_perform_range_for_lookup.
            (cp_parser_perform_range_for_lookup): Rename to ...
            (cp_perform_range_for_lookup): ... this.  No longer static.  Add
            complain argument and handle it.
            (cp_parser_range_for_member_function): Rename to ...
            (cp_range_for_member_function): ... this.
            (cp_parser_expansion_statement): New function.
            (cp_parser_jump_statement): Handle IN_EXPANSION_STMT.
            (cp_convert_omp_range_for): Adjust do_range_for_auto_deduction caller.
            Call cp_perform_range_for_lookup rather than
            cp_parser_perform_range_for_lookup.
            * error.cc (print_instantiation_full_context): Handle tldcl being
            TEMPLATE_FOR_STMT.
            (print_instantiation_partial_context_line): Likewise.
            * constexpr.cc (potential_constant_expression_1): Handle
            TEMPLATE_FOR_STMT.
            * decl.cc (poplevel_named_label_1): Use obl instead of bl->level_chain.
            (finish_case_label): Diagnose case labels inside of template for.
            (find_decomp_class_base): Add complain argument, don't diagnose
            anything and just return error_mark_node if tf_none, adjust recursive
            call.
            (cp_decomp_size): New function.
            (cp_finish_decomp): Adjust find_decomp_class_base caller.
            * semantics.cc (do_pushlevel): No longer static.
            (begin_template_for_scope): New function.
            * pt.cc (push_tinst_level_loc): Handle TEMPLATE_FOR_STMT.
            (reopen_tinst_level): Likewise.
            (tsubst_stmt): Handle TEMPLATE_FOR_STMT.
            (struct expansion_stmt_bc): New type.
            (expansion_stmt_find_bc_r, finish_expansion_stmt): New functions.
            * decl2.cc (decl_dependent_p): Return true for current function's decl
            if in_expansion_stmt.
            * call.cc (extend_ref_init_temps): Don't extend_all_temps if
            TREE_STATIC (decl).
            * cxx-pretty-print.cc (cxx_pretty_printer::statement): Handle
            TEMPLATE_FOR_STMT.
    gcc/testsuite/
            * g++.dg/cpp1z/decomp64.C: New test.
            * g++.dg/cpp26/expansion-stmt1.C: New test.
            * g++.dg/cpp26/expansion-stmt2.C: New test.
            * g++.dg/cpp26/expansion-stmt3.C: New test.
            * g++.dg/cpp26/expansion-stmt4.C: New test.
            * g++.dg/cpp26/expansion-stmt5.C: New test.
            * g++.dg/cpp26/expansion-stmt6.C: New test.
            * g++.dg/cpp26/expansion-stmt7.C: New test.
            * g++.dg/cpp26/expansion-stmt8.C: New test.
            * g++.dg/cpp26/expansion-stmt9.C: New test.
            * g++.dg/cpp26/expansion-stmt10.C: New test.
            * g++.dg/cpp26/expansion-stmt11.C: New test.
            * g++.dg/cpp26/expansion-stmt12.C: New test.
            * g++.dg/cpp26/expansion-stmt13.C: New test.
            * g++.dg/cpp26/expansion-stmt14.C: New test.
            * g++.dg/cpp26/expansion-stmt15.C: New test.
            * g++.dg/cpp26/expansion-stmt16.C: New test.
            * g++.dg/cpp26/expansion-stmt17.C: New test.
            * g++.dg/cpp26/expansion-stmt18.C: New test.
            * g++.dg/cpp26/expansion-stmt19.C: New test.
            * g++.dg/cpp26/feat-cxx26.C: Add __cpp_expansion_statements
            tests.
Comment 13 Jakub Jelinek 2025-08-13 20:12:16 UTC
Implemented for GCC 16.