Re: Strange crashes with std::vector push_back() and emplace_back()

Le sam. 18 juil. 2026 =C3=A0 13:18, Undisclosed via Gcc-help
<[email protected]> a =C3=A9crit :
>
> Hi,
> some premises first:
> -I have no proper debugging skills, plus I have some strict work timeline=
s (being into dsp coding) and cannot materially afford to open a "sub-chapt=
er" and drift to investigate this problem too much. Plus, I cannot material=
ly provide any final code to replicate the crashes, since the code I am pro=
viding is part of a music plugin; therefore, I am unable to open any offici=
al ticket (in case it is about a bug at all).
> -For the reasons above, my message has to be taken as an inquiry to know =
if anybody is aware of possible bugs involving relocations in std::vector, =
or if anyone has an idea of why I am getting crashes.
>
> I apologize in advance in case it's me who simply overlooked some deep as=
pect of C++ of course...
>
> -- Illustration of the issue --
>
> I am using winlibs msvcrt (gcc 16.4.0 but gcc 14.2.0 had the same issue),=
 windows 10 x64, i7 8770, codeblocks.
> I am using -O3 but I am getting crashes even without any -O option.
> The plugin is technically assembled as a DLL and uses static linking for =
the C++ libs.
>
> I have this (simplified) code:
>
>
> class MsgRow // Made as a class so it gets initialized properly on instan=
tiation
> {
> public:
>
>         bool clear =3D false;
>         bool dsp =3D false;
>         wchar_t chars[50] =3D {};
> };
>
>
>
> class MyPlugin
> {
> public:
>
>         void enqueueMessage(void *event)
>         {
>                 if(!event) return;
>
>                 MsgRow new_row;
>    // filling new_row fields here
>
>                 messages_.push_back(new_row); // this crashes
>         }
>
>    // more methods
>
> private:
>
>         std::vector<MsgRow> messages_;
>
>    // more stuff
>
> }
>
>
> What happens is that when enqueueMessage() is called with a valid event p=
tr, I get random crashes when push_back() is invoked.
>
> If I rearrange the code to do the same using emplace_back():
>
> auto& new_row =3D messages_.emplace_back();
> // set up new_row fields
>
> ...I still get random crashes when emplace_back() is called.
>
> On the contrary, if I do (that's stupid but very illustrative):
>
> MsgRow new_row;
> // set up new_row fields here
> messages_.resize(messages_.size()+1);
> messages_.back() =3D new_row;
>
> ...it apparently works without crashing !!
>
> More clues:
>
> If in MyPlugin::ctor (not shown) I do:
>
> messages_.reserve(10); // 10 seems enough
>
> ...both push_back() and emplace_back() work without apparently crashing.
>
> Also, if before calling push_back() or emplace_back() I call messages_.ca=
pacity() and I intercept its return value for debugging, I see a bogus (ran=
dom, often even negative) huge value
>
> I am very puzzled...
> Thanks for any hint.

Pretty sure that the bug is outside your usage of std::vector...

In any case you can (portably) use _GLIBCXX_DEBUG /
_GLIBCXX_DEBUG_PEDANTIC if you think this is related to the usage of
std::vector.

>
>
>
>


--=20
Mathieu