This is the mail archive of the gdb-patches@sourceware.org mailing list for the GDB project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 2/5] Don't memcpy non-trivially-copyable types: Make enum_flags triv. copyable


On 04/20/2017 04:34 AM, Simon Marchi wrote:
> On 2017-04-12 22:27, Pedro Alves wrote:

>> +  /* Define copy/move ctor/op= as defaulted so that enum_flags is
>> +     trivially copyable.  */
>> +  enum_flags (const enum_flags &other) = default;
>> +  enum_flags (enum_flags &&) noexcept = default;
>> +  enum_flags &operator= (const enum_flags &other) = default;
>> +  enum_flags &operator= (enum_flags &&) = default;
> 
> What's the difference between defining these as default, and not
> defining them at all (which I assume will still make the compiler emit
> the default versions)?

Yeah, there's no difference.  I'll just remove them, since it clearly
confuses things otherwise.  Thanks.

> Do you want to add a static_assert(is_pod...) for that class?  For now
> we have the VECs that will warn us if it becomes non-POD, but eventually
> they'll be gone.

I do, but enum_flags is a template, and we need to add the assertion to
some template instantiation.  I'd like to add it in the enum
flags unit tests file, currently here [1]:

  https://github.com/palves/gdb/blob/palves/cxx11-enum-flags/gdb/enum-flags-selftests.c

which I need to rebase and repost.  I should be reposting that soon.

[1] - I'll move it to unittests/

Thanks,
Pedro Alves

>From 23bcc18f470ee4364bd362a8b78c6c1415a9dadb Mon Sep 17 00:00:00 2001
From: Pedro Alves <palves@redhat.com>
Date: Tue, 25 Apr 2017 01:27:42 +0100
Subject: [PATCH] Don't memcpy non-trivially-copyable types: Make enum_flags
 triv. copyable
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

The delete-memcpy-with-non-trivial-types patch exposed many instances
of this problem:

  src/gdb/btrace.h: In function ‘btrace_insn_s* VEC_btrace_insn_s_quick_insert(VEC_btrace_insn_s*, unsigned int, const btrace_insn_s*, const char*, unsigned int)’:
  src/gdb/common/vec.h:948:62: error: use of deleted function ‘void* memmove(T*, const U*, size_t) [with T = btrace_insn; U = btrace_insn; <template-parameter-1-3> = void; size_t = long unsigned int]’
     memmove (slot_ + 1, slot_, (vec_->num++ - ix_) * sizeof (T));    \
								^
  src/gdb/common/vec.h:436:1: note: in expansion of macro ‘DEF_VEC_FUNC_O’
   DEF_VEC_FUNC_O(T)         \
   ^
  src/gdb/btrace.h:84:1: note: in expansion of macro ‘DEF_VEC_O’
   DEF_VEC_O (btrace_insn_s);
   ^
[...]
  src/gdb/common/vec.h:1060:31: error: use of deleted function ‘void* memcpy(T*, const U*, size_t) [with T = btrace_insn; U = btrace_insn; <template-parameter-1-3> = void; size_t = long unsigned int]’
	  sizeof (T) * vec2_->num);       \
				 ^
  src/gdb/common/vec.h:437:1: note: in expansion of macro ‘DEF_VEC_ALLOC_FUNC_O’
   DEF_VEC_ALLOC_FUNC_O(T)         \
   ^
  src/gdb/btrace.h:84:1: note: in expansion of macro ‘DEF_VEC_O’
   DEF_VEC_O (btrace_insn_s);
   ^

So, VECs (given it's C roots) rely on memcpy/memcpy of VEC elements to
be well defined, in order to grow/reallocate its internal elements
array.  This means that we can only put trivially copyable types in
VECs.  E.g., if a type requires using a custom copy/move ctor to
relocate, then we can't put it in a VEC (so we use std::vector
instead).  But, as shown above, we're violating that requirement.

btrace_insn is currently not trivially copyable, because it contains
an enum_flags field, and that is itself not trivially copyable.  This
patch corrects that, by simply removing the user-provided copy
constructor and assignment operator.  The compiler-generated versions
work just fine.

Note that std::vector relies on std::is_trivially_copyable too to know
whether it can reallocate its elements with memcpy/memmove instead of
having to call copy/move ctors and dtors, so if we have types in
std::vectors that weren't trivially copyable because of enum_flags,
this will make such vectors more efficient.

gdb/ChangeLog:
2017-04-25  Pedro Alves  <palves@redhat.com>

	* common/enum-flags.h (enum_flags): Don't implement copy ctor and
	assignment operator.
---
 gdb/ChangeLog           |  5 +++++
 gdb/common/enum-flags.h | 10 ----------
 2 files changed, 5 insertions(+), 10 deletions(-)

diff --git a/gdb/ChangeLog b/gdb/ChangeLog
index 33dd53b..f5419db 100644
--- a/gdb/ChangeLog
+++ b/gdb/ChangeLog
@@ -1,3 +1,8 @@
+2017-04-25  Pedro Alves  <palves@redhat.com>
+
+	* common/enum-flags.h (enum_flags): Don't implement copy ctor and
+	assignment operator.
+
 2017-04-24  Yao Qi  <yao.qi@linaro.org>
 
 	* doublest.c (convert_doublest_to_floatformat): Call
diff --git a/gdb/common/enum-flags.h b/gdb/common/enum-flags.h
index e63c8a4..ddfcddf 100644
--- a/gdb/common/enum-flags.h
+++ b/gdb/common/enum-flags.h
@@ -120,16 +120,6 @@ public:
     : m_enum_value ((enum_type) 0)
   {}
 
-  enum_flags (const enum_flags &other)
-    : m_enum_value (other.m_enum_value)
-  {}
-
-  enum_flags &operator= (const enum_flags &other)
-  {
-    m_enum_value = other.m_enum_value;
-    return *this;
-  }
-
   /* If you get an error saying these two overloads are ambiguous,
      then you tried to mix values of different enum types.  */
   enum_flags (enum_type e)
-- 
2.5.5



Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]