On 07/05/20 11:08 +0200, Jakub Jelinek wrote:
Hi!

This is my humble attempt to document the C++ ABI changes.
Or do you have something better?
Not sure e.g. how C++ calls the TREE_ADDRESSABLE types which are
passed/returned by invisible reference, perhaps it would be nice
to make it clear that those aren't affected.

As the release tarballs contain a copy of changes.html, I'm afraid
we need to resolve this before I can roll the tarballs.

--- htdocs/gcc-10/changes.html  2020-05-07 10:44:17.391465227 +0200
+++ htdocs/gcc-10/changes.html  2020-05-07 11:02:50.580691123 +0200
@@ -31,6 +31,27 @@ a work-in-progress.</p>
<h2>Caveats</h2>
<ul>
  <li>
+    <a name="empty_base">The ABI</a>
+    of passing and returning certain C++ classes by value changed
+    on several targets in GCC 10, including AArch64, ARM, PowerPC ELFv2,
+    S/390 and Itanium.  In <code>-std=c++17</code> and <code>-std=c++20</code>
+    modes for classes with empty bases which otherwise contain only a single
+    element or are handled as homogeneous aggregates in
+    <code>-std=c++14</code> and earlier modes the existence of the empty
+    bases resulted in those classes not to be considered as having a single
+    element or be homogeneous aggregate and so could be passed differently,
+    making <code>-std=c++17</code> and <code>-std=c++14</code> compiled code
+    ABI incompatible.  This has been corrected and the empty bases are
+    ignored in those ABI decisions, so <code>-std=c++14</code> and
+    <code>-std=c++17</code> compiled codes are now again ABI compatible.
+    Example: <code>struct empty {}; struct S : public empty { float f; 
}</code>.
+    Similarly, in classes containing non-static data members with empty
+    class types and with C++20 <code>[[no_unique_address]]</code> attribute
+    those members weren't ignored in the ABI argument passing decisions
+    as they should be.  Both of these ABI changes are now diagnosed with
+    <code>-Wpsabi</code>.
+  </li>
+  <li>
    The deprecated Profile Mode and <code>array_allocator</code> extensions
    have been removed from libstdc++.
  </li>

Here's what I've committed after discussion on IRC, combining Jakub's
text with what I had.


commit f1d2be6c9fcc52d676266e7ede123953d150aaf3
Author: Jonathan Wakely <jwak...@redhat.com>
Date:   Thu May 7 11:24:04 2020 +0100

    Document C++17 ABI changes in GCC 10

diff --git a/htdocs/gcc-10/changes.html b/htdocs/gcc-10/changes.html
index f5b70eb4..52fd6e6a 100644
--- a/htdocs/gcc-10/changes.html
+++ b/htdocs/gcc-10/changes.html
@@ -30,6 +30,12 @@ a work-in-progress.</p>
 <!-- .................................................................. -->
 <h2>Caveats</h2>
 <ul>
+  <li>
+    An ABI incompatibility between C++14 and C++17 has been fixed.  On some
+    targets a class with a zero-sized subobject would be passed incorrectly
+    when compiled as C++17 or C++20.
+    See the <a href="#empty_base">C++ notes below</a> for more details.
+  </li>
   <li>
     The deprecated Profile Mode and <code>array_allocator</code> extensions
     have been removed from libstdc++.
@@ -409,6 +415,39 @@ int get_na??ve_pi() {
     The attribute <code>deprecated</code> can now be used on
     <code>namespace</code>s too.
   </li>
+  <li>
+    <a name="empty_base">The ABI</a>
+    of passing and returning certain C++ classes by value changed
+    on several targets in GCC 10, including
+    <a href="https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94383";>AArch64</a>,
+    <a href="https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94711";>ARM</a>,
+    <a href="https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94383";>PowerPC ELFv2</a>,
+    <a href="https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94704";>S/390</a>
+    and
+    <a href="https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94706";>Itanium</a>.
+    These changes affect classes with a zero-sized
+    subobject (an empty base class, or data member with the
+    <code>[[no_unique_address]]</code> attribute) where all other non-static
+    data members have the same type (this is called a "homogeneous aggregate"
+    in some ABI specifications, or if there is only one such member,
+    a "single element").
+    In <code>-std=c++17</code> and <code>-std=c++20</code> modes, classes with
+    an empty base class were not considered to have a single element or
+    to be a homogeneous aggregate, and so could be passed differently
+    (in the wrong registers or at the wrong stack address). This could make
+    code compiled with <code>-std=c++17</code> and <code>-std=c++14</code>
+    ABI incompatible.  This has been corrected and the empty bases are
+    ignored in those ABI decisions, so functions compiled with
+    <code>-std=c++14</code> and <code>-std=c++17</code> are now ABI compatible
+    again.
+    Example:
+    <code>struct empty {}; struct S : empty { float f; }; void f(S);</code>.
+    Similarly, in classes containing non-static data members with empty
+    class types using the C++20 <code>[[no_unique_address]]</code> attribute,
+    those members weren't ignored in the ABI argument passing decisions
+    as they should be.  Both of these ABI changes are now diagnosed with
+    <code>-Wpsabi</code>.
+  </li>
 </ul>
 
 <h4 id="libstdcxx">Runtime Library (libstdc++)</h4>

Reply via email to