From: Arthur Norman <acn1@cam.ac.uk>
To: cygwin@cygwin.com
Subject: linker fails with multiple definitions after inline thread_local var within class
Date: Thu, 14 Nov 2019 18:41:00 -0000 [thread overview]
Message-ID: <alpine.WNT.2.00.1911141816490.14912@panamint> (raw)
[-- Attachment #1: Type: TEXT/PLAIN, Size: 2890 bytes --]
#! /bin/bash
# When I run this on Ubuntu-x86_64, Macintosh(clang) or a Raspberry pi
# the code links and when it runs it just prints 999 - which is the behaviour
# that I expect. On both Cygwin and using x86_64-w64-mingw32-g++ (and i686)
# I get a linker diagnostic of the form
# ./tltest.sh
# /usr/lib/gcc/x86_64-pc-cygwin/8.3.0/../../../../x86_64-pc-cygwin/bin/ld:
# /tmp/cczVTcZ1.o:t2.cpp:(.text+0x86): multiple definition of
# `TLS init function for Data::valref';
# /tmp/ccyLQlRb.o:t1.cpp:(.text+0x4a): first defined here
# collect2: error: ld returned 1 exit status
#
# I believe that given the specification of C++17 "inline" variables this
# is incorrect, but there are better experts who may be able to explain
# otherwise.
#
# When people raise issues here I often see other asking "Why are you doing
# that?". This is a cut-down version of my real code where rather than
# storing a reference to the thread_local Record that I am interested in
# in a simple array (tvec) at a fixed offset (1) I store and retrieve a
# referece to it using TlsAlloc(), TlsGetValue and TlsSetValue - those being
# the Microsoft API for thread-local access, and for my purposes my
# measurements suggest that using them gives me useful performance gains
# over the emutls code that g++ creates on the relevant platforms.
# And I am then (trying to) build a header-only library (for which t.h is
# the surrogate here) which can be included from several other compilation
# units but by virtue of "inline variables" its private (including thread-
# local) date can be defined within that header file so that the files that
# #include the header-only library do not need to contain anything beyond
# uses of it. To illustrate this I have two source files which each include
# t.h but the bad behaviour does not need any other code in the first one!
# The second just contains a tiny main program that inspects data from the
# library data.
# Have I misunderstood C++ and so am I doing something wrong or is this
# a g++/Windows bug?
cat <<XXX > t.h
#include <cstdint>
inline void *tvec[4];
class Record
{
public:
int val = 999;
};
class Data_Ref
{ static inline thread_local Record val;
public:
static Record* get() // Get reference to Record via tvec[]
{ return (Record*)tvec[1];
}
Data_Ref() // Stash it in tvec[] for later use
{ tvec[1] = (void *)&val;
}
};
class Data
{ static inline thread_local Data_Ref valref;
public:
static Record &get()
{ return *valref.get(); // note that get() is static in Data_Ref
}
};
XXX
cat <<XXX > t1.cpp
#include <iostream>
#include "t.h" // First copy
XXX
cat <<XXX > t2.cpp
#include <iostream>
#include "t.h" // Second copy
int main()
{ std::cout << Data::get().val << std::endl;
return 0;
}
XXX
${1:-g++} -std=c++17 -I. t1.cpp t2.cpp -o t
./t
# end of test script
[-- Attachment #2: Type: APPLICATION/octet-stream, Size: 2879 bytes --]
#! /bin/bash
# When I run this on Ubuntu-x86_64, Macintosh(clang) or a Raspberry pi
# the code links and when it runs it just prints 999 - which is the behaviour
# that I expect. On both Cygwin and using x86_64-w64-mingw32-g++ (and i686)
# I get a linker diagnostic of the form
# ./tltest.sh
# /usr/lib/gcc/x86_64-pc-cygwin/8.3.0/../../../../x86_64-pc-cygwin/bin/ld:
# /tmp/cczVTcZ1.o:t2.cpp:(.text+0x86): multiple definition of
# `TLS init function for Data::valref';
# /tmp/ccyLQlRb.o:t1.cpp:(.text+0x4a): first defined here
# collect2: error: ld returned 1 exit status
#
# I believe that given the specification of C++17 "inline" variables this
# is incorrect, but there are better experts who may be able to explain
# otherwise.
#
# When people raise issues here I often see other asking "Why are you doing
# that?". This is a cut-down version of my real code where rather than
# storing a reference to the thread_local Record that I am interested in
# in a simple array (tvec) at a fixed offset (1) I store and retrieve a
# referece to it using TlsAlloc(), TlsGetValue and TlsSetValue - those being
# the Microsoft API for thread-local access, and for my purposes my
# measurements suggest that using them gives me useful performance gains
# over the emutls code that g++ creates on the relevant platforms.
# And I am then (trying to) build a header-only library (for which t.h is
# the surrogate here) which can be included from several other compilation
# units but by virtue of "inline variables" its private (including thread-
# local) date can be defined within that header file so that the files that
# #include the header-only library do not need to contain anything beyond
# uses of it. To illustrate this I have two source files which each include
# t.h but the bad behaviour does not need any other code in the first one!
# The second just contains a tiny main program that inspects data from the
# library data.
# Have I misunderstood C++ and so am I doing something wrong or is this
# a g++/Windows bug?
cat <<XXX > t.h
#include <cstdint>
inline void *tvec[4];
class Record
{
public:
int val = 999;
};
class Data_Ref
{ static inline thread_local Record val;
public:
static Record* get() // Get reference to Record via tvec[]
{ return (Record*)tvec[1];
}
Data_Ref() // Stash it in tvec[] for later use
{ tvec[1] = (void *)&val;
}
};
class Data
{ static inline thread_local Data_Ref valref;
public:
static Record &get()
{ return *valref.get(); // note that get() is static in Data_Ref
}
};
XXX
cat <<XXX > t1.cpp
#include <iostream>
#include "t.h" // First copy
XXX
cat <<XXX > t2.cpp
#include <iostream>
#include "t.h" // Second copy
int main()
{ std::cout << Data::get().val << std::endl;
return 0;
}
XXX
${1:-g++} -std=c++17 -I. t1.cpp t2.cpp -o t
./t
# end of test script
[-- Attachment #3: Type: text/plain, Size: 219 bytes --]
--
Problem reports: http://cygwin.com/problems.html
FAQ: http://cygwin.com/faq/
Documentation: http://cygwin.com/docs.html
Unsubscribe info: http://cygwin.com/ml/#unsubscribe-simple
next reply other threads:[~2019-11-14 18:23 UTC|newest]
Thread overview: 8+ messages / expand[flat|nested] mbox.gz Atom feed top
2019-11-14 18:41 Arthur Norman [this message]
2019-11-14 23:45 ` Brian Inglis
2019-11-15 8:14 ` Arthur Norman
2019-11-15 22:07 ` Brian Inglis
2019-11-16 3:35 ` Arthur Norman
2019-11-16 3:54 ` Brian Inglis
2019-11-16 10:48 ` Brian Inglis
2019-11-17 21:05 ` Arthur Norman
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=alpine.WNT.2.00.1911141816490.14912@panamint \
--to=acn1@cam.ac.uk \
--cc=cygwin@cygwin.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).