aboutsummaryrefslogtreecommitdiff
path: root/gdb/regcache.h
diff options
context:
space:
mode:
authorSimon Marchi <simon.marchi@polymtl.ca>2020-08-07 11:28:52 -0400
committerSimon Marchi <simon.marchi@polymtl.ca>2020-08-07 11:29:00 -0400
commit888bdb2b744515d111953a0a3aa674ea822416f9 (patch)
tree24407e8f85a294f5adad480644a0461255d0bbe8 /gdb/regcache.h
parentb161a60d1fe2a7383c7940815687c6100b97204e (diff)
downloadgdb-888bdb2b744515d111953a0a3aa674ea822416f9.zip
gdb-888bdb2b744515d111953a0a3aa674ea822416f9.tar.gz
gdb-888bdb2b744515d111953a0a3aa674ea822416f9.tar.bz2
gdb: change regcache list to be a map
One regcache object is created for each stopped thread and is stored in the regcache::regcaches linked list. Looking up a regcache for a given thread is therefore in O(number of threads). Stopping all threads then becomes O((number of threads) ^ 2). Same goes for resuming a thread (need to delete the regcache of a given ptid) and resuming all threads. It becomes noticeable when debugging thousands of threads, which is typical with GPU targets. This patch replaces the linked list with some maps to reduce that complexity. The first design was using an std::unordered_map with (target, ptid, arch) as the key, because that's how lookups are done (in get_thread_arch_aspace_regcache). However, the registers_changed_ptid function, also somewhat on the hot path (it is used when resuming threads), needs to delete all regcaches associated to a given (target, ptid) tuple. If the key of the map is (target, ptid, arch), we have to walk all items of the map, not good. The second design was therefore using an std::unordered_multimap with (target, ptid) as the key. One key could be associated to multiple regcaches, all with different gdbarches. When looking up, we would have to walk all these regcaches. This would be ok, because there will usually be actually one matching regcache. In the exceptional multi-arch thread cases, there will be maybe two. However, in registers_changed_ptid, we sometimes need to remove all regcaches matching a given target. We would then have to talk all items of the map again, not good. The design as implemented in this patch therefore uses two levels of map. One std::unordered_map uses the target as the key. The value type is an std::unordered_multimap that itself uses the ptid as the key. The values of the multimap are the regcaches themselves. Again, we expect to have one or very few regcaches per (target, ptid). So, in summary: * The lookups (in get_thread_arch_aspace_regcache), become faster when the number of threads grows, compared to the linked list. With a small number of threads, it will probably be a bit slower to do map lookups than to walk a few linked list nodes, but I don't think it will be noticeable in practice. * The function registers_changed_ptid deletes all regcaches related to a given (target, ptid). It must now handle the different cases separately: - NULL target and minus_one_ptid: we delete all the entries - NULL target and non-minus_one_ptid: invalid (checked by assert) - non-NULL target and non-minus_one_ptid: we delete all the entries associated to that tuple - a non-NULL target and minus_one_ptid: we delete all the entries associated to that target * The function regcache_thread_ptid_changed is called when a thread changes ptid. It is implemented efficiently using the map, although that's not very important: it is not called often, mostly when creating an inferior, on some specific platforms. This patch is a tiny bit from ROCm GDB [1] we would like to merge upstream. Laurent Morichetti gave be these performance numbers: The benchmark used is: time ./gdb --data-directory=data-directory /extra/lmoriche/hip/samples/0_Intro/bit_extract/bit_extract -ex "set pagination off" -ex "set breakpoint pending on" -ex "b bit_extract_kernel if \$_thread == 5" -ex run -ex c -batch It measures the time it takes to continue from a conditional breakpoint with 2048 threads at that breakpoint, one of them reporting the breakpoint. baseline: real 0m10.227s real 0m10.177s real 0m10.362s with patch: real 0m8.356s real 0m8.424s real 0m8.494s [1] https://github.com/ROCm-Developer-Tools/ROCgdb gdb/ChangeLog: * regcache.c (ptid_regcache_map): New type. (target_ptid_regcache_map): New type. (regcaches): Change type to target_ptid_regcache_map. (get_thread_arch_aspace_regcache): Update to regcaches' new type. (regcache_thread_ptid_changed): Likewise. (registers_changed_ptid): Likewise. (regcaches_size): Likewise. (regcaches_test): Update. (regcache_thread_ptid_changed): Update. * regcache.h (regcache_up): New type. * gdbsupport/ptid.h (hash_ptid): New struct. Change-Id: Iabb0a1111707936ca111ddb13f3b09efa83d3402
Diffstat (limited to 'gdb/regcache.h')
-rw-r--r--gdb/regcache.h2
1 files changed, 2 insertions, 0 deletions
diff --git a/gdb/regcache.h b/gdb/regcache.h
index dd0c2f2..9390f57 100644
--- a/gdb/regcache.h
+++ b/gdb/regcache.h
@@ -435,6 +435,8 @@ private:
struct address_space *aspace);
};
+using regcache_up = std::unique_ptr<regcache>;
+
class readonly_detached_regcache : public readable_regcache
{
public: