On Fri, Dec 13, 2024 at 04:52:58PM +1300, Campbell Suter wrote:
> Each lambda that can be converted to a plain function pointer has a
> thunk generated for it, which invokes the body of the lambda function.
>
> When a section attribute is added to a lambda function, it only applies
> to the body of the lambda function, and not the thunk. When a lambda is
> only ever used by converting it to a function pointer, the body of the
> lambda is inlined into this thunk. As a result, the section attribute
> is effectively ignored: the function it applied to is gone, and the thunk
> does not have the section attribute applied to it either.
>
> This patch checks if a section attribute is present on a lambda, and
> applies it to the thunk.
>
> The motivation for this change is embedded devices where most code is
> executed from flash, but code which must execute while the device is
> being reprogrammed can be moved to RAM by placing it in a different
> section.
>
> This patch was tested with bootstrapping on x86-64 under WSL, and
> the newly added test was also run on 32-bit ARM.
Hi, thanks for the patch.
> gcc/cp/ChangeLog:
>
> * lambda.cc (maybe_add_lambda_conv_op): Don't ignore section
> attributes on lambda functions which are converted to plain
> function pointers.
>
> gcc/testsuite/ChangeLog:
>
> * g++.dg/ext/attr-section-lambda.C: New test.
>
> Signed-off-by: Campbell Suter <z...@znix.xyz>
> ---
> gcc/cp/lambda.cc | 8 ++++
> .../g++.dg/ext/attr-section-lambda.C | 42 +++++++++++++++++++
> 2 files changed, 50 insertions(+)
> create mode 100644 gcc/testsuite/g++.dg/ext/attr-section-lambda.C
>
> diff --git a/gcc/cp/lambda.cc b/gcc/cp/lambda.cc
> index d8a15d97d..e8937cc0d 100644
> --- a/gcc/cp/lambda.cc
> +++ b/gcc/cp/lambda.cc
> @@ -32,6 +32,7 @@ along with GCC; see the file COPYING3. If not see
> #include "gimplify.h"
> #include "target.h"
> #include "decl.h"
> +#include "attribs.h"
>
> /* Constructor for a lambda expression. */
>
> @@ -1376,6 +1377,13 @@ maybe_add_lambda_conv_op (tree type)
> if (generic_lambda_p)
> fn = add_inherited_template_parms (fn, DECL_TI_TEMPLATE (callop));
>
> + if (lookup_attribute ("section", DECL_ATTRIBUTES (callop)))
> + {
> + duplicate_one_attribute(&DECL_ATTRIBUTES (fn),
> + DECL_ATTRIBUTES (callop), "section");
> + set_decl_section_name (statfn, callop);
> + }
duplicate_one_attribute does two lookups, but we just looked up the
callop attr, and we just built up fn so it shouldn't have any attrs.
Thus I wonder if writing it like this would be a little neater:
if (tree a = lookup_attribute ("section", DECL_ATTRIBUTES (callop)))
{
DECL_ATTRIBUTES (fn) = attr_chainon (DECL_ATTRIBUTES (fn), a);
set_decl_section_name (fn, callop);
}
Sadly I see that just the set_decl_section_name call wouldn't be enough...
Marek
