Re: [PATCH 00/11] Various thread lists optimizations

[Simon Marchi via Gdb-patches <gdb-patches@sourceware.org>]

On 2021-06-22 12:56 p.m., Simon Marchi wrote:
> This series contains various optimizations written after profiling
> ROCm-GDB [1].  A typical ROCm program may have thousands of threads
> (waves), so everything that iterates often on the whole thread list pops
> up on the profile report.
> 
> Conceptually, the optimizations are:
> 
>  - maintain a ptid -> thread map in each inferior, to speed up the
>    thread lookups per ptid with find_thread_ptid
>  - make all_matching_threads_iterator a bit smarter, depending on the
>    filter_ptid passed, it can avoid iterating on all threads.  When the
>    filter_ptid points to a specific thread, it can use find_thread_ptid
>    to directly find the thread of interest, which is fast thanks to the
>    previous point
>  - maintain a per-process-target list of threads that are resumed and
>    have a pending event.  This helps speed up two hot path cases: when
>    we check if we want to re-enable commit-resumed, and when we want to
>    fetch a pending event.
> 
> Patches up to and including patch 6 are groundwork.  Notably, patch 2
> (thanks to Pedro) adds an intrusive_list type, which allows using
> intrusive linked lists in a very C++-y way without writing a ton of
> boilerplate.  This list type is useful for the patch that maintains a
> list of resumed threads with pending events, but we also converted the
> per-inferior thread list, the inferior list and the thread
> step-over-list to use it.  It helped iron out a few bugs and ensure the
> list works well for our purposes.
> 
> When debugging a ROCm test program whose threads continuously hit a
> condition breakpoint evaluation to false, the speedup observed with
> ROCm-GDB is around 5x (a run takes about 30 seconds before and 6 seconds
> after).
> 
> With x86-64 / Linux, it's less noticeable, because the time we spend in
> syscalls to fetch events and resume threads still dominates.  Still,
> when trying a program with 1000 threads hitting 100 times each a false
> conditional breakpoint, it goes from ~20 seconds to ~16 seconds.  But I
> don't really expect anything that noticeable in everyday use cases
> though.
> 
> [1] https://github.com/ROCm-Developer-Tools/ROCgdb
> 
> Pedro Alves (2):
>   gdb: introduce intrusive_list, make thread_info use it
>   gdb: make inferior_list use intrusive_list
> 
> Simon Marchi (9):
>   gdb: introduce iterator_range, remove next_adapter
>   gdb: use intrusive list for step-over chain
>   gdb: add setter / getter for thread_info resumed state
>   gdb: make thread_info::suspend private, add getters / setters
>   gdb: maintain per-process-target list of resumed threads with pending
>     wait status
>   gdb: optimize check for resumed threads with pending wait status in
>     maybe_set_commit_resumed_all_targets
>   gdb: optimize selection of resumed thread with pending event
>   gdb: maintain ptid -> thread map, optimize find_thread_ptid
>   gdb: optimize all_matching_threads_iterator
> 
>  gdb/Makefile.in                            |   1 +
>  gdb/ada-tasks.c                            |   4 +-
>  gdb/breakpoint.c                           |   7 +-
>  gdb/breakpoint.h                           |  10 +-
>  gdb/elf-none-tdep.c                        |   2 +-
>  gdb/fbsd-tdep.c                            |   6 +-
>  gdb/gcore.c                                |   4 +-
>  gdb/gdb-gdb.py.in                          |  91 ++-
>  gdb/gdb_bfd.h                              |   4 +-
>  gdb/gdbthread.h                            | 192 ++++-
>  gdb/infcmd.c                               |  33 +-
>  gdb/inferior-iter.h                        |  94 +--
>  gdb/inferior.c                             | 105 ++-
>  gdb/inferior.h                             |  34 +-
>  gdb/inflow.c                               |   2 +-
>  gdb/infrun.c                               | 476 ++++++------
>  gdb/infrun.h                               |   4 +-
>  gdb/linux-fork.c                           |   3 +-
>  gdb/linux-nat.c                            |  12 +-
>  gdb/linux-tdep.c                           |   2 +-
>  gdb/objfiles.h                             |   6 +-
>  gdb/process-stratum-target.c               |  77 ++
>  gdb/process-stratum-target.h               |  26 +
>  gdb/progspace.c                            |  11 +-
>  gdb/progspace.h                            |  45 +-
>  gdb/psymtab.h                              |   2 +-
>  gdb/python/py-inferior.c                   |   2 +-
>  gdb/record-btrace.c                        |   3 +-
>  gdb/record-full.c                          |   3 +-
>  gdb/regcache.c                             |   6 +-
>  gdb/remote.c                               |  68 +-
>  gdb/scoped-mock-context.h                  |  15 +-
>  gdb/solist.h                               |   2 +
>  gdb/symtab.h                               |  15 +-
>  gdb/thread-iter.c                          | 147 +++-
>  gdb/thread-iter.h                          |  61 +-
>  gdb/thread.c                               | 216 +++---
>  gdb/top.h                                  |   6 +-
>  gdb/unittests/intrusive_list-selftests.c   | 818 +++++++++++++++++++++
>  gdbsupport/intrusive_list.h                | 586 +++++++++++++++
>  gdbsupport/iterator-range.h                |  60 ++
>  gdbsupport/next-iterator.h                 |  32 +-
>  gdbsupport/reference-to-pointer-iterator.h |  82 +++
>  43 files changed, 2574 insertions(+), 801 deletions(-)
>  create mode 100644 gdb/unittests/intrusive_list-selftests.c
>  create mode 100644 gdbsupport/intrusive_list.h
>  create mode 100644 gdbsupport/iterator-range.h
>  create mode 100644 gdbsupport/reference-to-pointer-iterator.h
> 

I pushed this series after addressing Pedro's comments.

Simon