One last pertinent piece of information: ~/.fighome/runtime/gcc/5.2.0-1/bin/ld.gold --version GNU gold (GCC-explorer-build 2.24) 1.11 Copyright 2013 Free Software Foundation, Inc. This program is free software; you may redistribute it under the terms of the GNU General Public License version 3 or (at your option) a later version. This program has absolutely no warranty. Thanks, Matt On Thu, Mar 17, 2016 at 8:35 AM, Matt Godbolt <matt@xxxxxxxxxxx> wrote: > Hi again, > > This may be a GOLD bug, or at least an interaction between GOLD and > GCC. If I remove my "-fuse-ld=gold" line from the linker I get the > correct results too: > > Dump of assembler code for function std::_List_node<std::function<void > ()> >::~_List_node(): > 0x000000000077ca80 <+0>: push %rbp > 0x000000000077ca81 <+1>: mov %rsp,%rbp > 0x000000000077ca84 <+4>: sub $0x10,%rsp > 0x000000000077ca88 <+8>: mov %rdi,-0x8(%rbp) > 0x000000000077ca8c <+12>: mov -0x8(%rbp),%rax > 0x000000000077ca90 <+16>: add $0x10,%rax > 0x000000000077ca94 <+20>: mov %rax,%rdi > 0x000000000077ca97 <+23>: callq 0x4ff6f0 <std::function<void > ()>::~function()@plt> > 0x000000000077ca9c <+28>: nop > 0x000000000077ca9d <+29>: leaveq > 0x000000000077ca9e <+30>: retq > End of assembler dump. > > Can a gold expert (@ilt?) help me gather the information needed to file a bug? > > Thanks so much in advance, > > Matt > > On Wed, Mar 16, 2016 at 3:34 PM, Matt Godbolt <matt@xxxxxxxxxxx> wrote: >> Hi all, >> >> I've just lost a day to tracking down an unusual crash which I've >> traced back to what seems to be an LTO issue. >> >> I'm linking a large application against some open source libraries. >> All are built with the same version of GCC (5.2.0). We have been using >> LTO for a long time, since 4.9.x, without issue. >> >> Recently I upgraded one of our dependencies >> (https://github.com/mattgodbolt/seasocks). One change was that this >> dependency now uses std::function<void()> in its API. This library is >> built *without* LTO. >> >> In our application we now get a crash when the destructor for >> std::function<void()> is called. We have a std::list<> of such >> functions and the code generated now looks like this: >> >> gdb out/dist/lib-haswell/hippo -ex "disassemble >> 'std::_List_node<std::function<void ()> >::~_List_node()'" >> Dump of assembler code for function std::_List_node<std::function<void >> ()> >::~_List_node(): >> 0x0000000000763a40 <+0>: push %rbp >> 0x0000000000763a41 <+1>: mov %rsp,%rbp >> 0x0000000000763a44 <+4>: sub $0x10,%rsp >> 0x0000000000763a48 <+8>: mov %rdi,-0x8(%rbp) >> 0x0000000000763a4c <+12>: mov -0x8(%rbp),%rax >> 0x0000000000763a50 <+16>: add $0x10,%rax >> 0x0000000000763a54 <+20>: mov %rax,%rdi >> 0x0000000000763a57 <+23>: callq 0x0 >> 0x0000000000763a5c <+28>: nop >> 0x0000000000763a5d <+29>: leaveq >> 0x0000000000763a5e <+30>: retq >> End of assembler dump. >> >> The spot where the "call std::function::~function<void()>" should be >> has a "callq 0x0". If I build with LTO off globally, the callq is >> indeed correctly generated/linked. >> >> It's as if the linker has just plain not done its job there with LTO on. >> >> If I take the source of Seasocks and bring it into the project (and >> thus the files are built with LTO), all compiles fine. We run using >> -fuse-ld=gold -flto -fuse-linker-plugin, amongst many other options. >> >> Despite trying my hardest I was unable to reproduce this issue using >> two files calling each other (one with LTO on, one not). Can anyone >> give suggestions as to what I might try to figure out what's going on? >> I have just tried with GCC 5.3 too, and have had no success either. >> >> Many thanks in advance, >> >> Matt > > > > -- > Matt -- Matt
