https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111588
Bug ID: 111588
Summary: Provide opt-out of shared_ptr single-threaded
optimization
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: libstdc++
Assignee: unassigned at gcc dot gnu.org
Reporter: redbeard0531 at gmail dot com
Target Milestone: ---
Right now there is a fast-path for single-threaded programs to avoid the
overhead of atomics in shared_ptr, but there is no equivalent for a program the
knows it is multi-threaded to remove the check and branch. If __GTHREADS is not
defined then no atomic code is emitted.
There are two issues with this: 1) for programs that know they are effectively
always multithreaded they pay for a runtime branch and .text segment bloat for
an optimization that never applies. This may have knock-on effects of making
functions that use shared_ptr less likely to be inlined by pushing them
slightly over the complexity threshold. 2) It invalidates singlethreaded
microbenchmarks of code that uses shared_ptr because the performance of the
code may be very different from when run in the real multithreaded program.
I understand the value of making a fast mode for single-threaded code, and I
can even except having the runtime branch by default, rather than as an opt-in,
when it is unknown if the program will be run with multiple threads. But an
opt-out would be nice to have. If it had to be a gcc-build time option rather
than a #define, that would be acceptable for us since we always use our own
build of gcc, but it seems like a worse option for other users.
FWIW, neither llvm libc++
(https://github.com/llvm/llvm-project/blob/0bfaed8c612705cfa8c5382d26d8089a0a26386b/libcxx/include/__memory/shared_ptr.h#L103-L110)
nor MS-STL
(https://github.com/microsoft/STL/blob/main/stl/inc/memory#L1171-L1173) ever
use runtime branching to detect multithreading.
