Thanks, nice result, I'll try to run the performance benchmarks that are
coming with libstdc++ to see if they spot anything.
That's tests in testsuite/performance folder in case you want to have a
try yourself.
François
On 18/01/2024 10:26, Huanghui Nie wrote:
Yes, I have. I did a benchmark today.
The conclusion is: the time consumption can be reduced by 0.4% ~ 1.2%
when unordered_set erase(begin()), and 1.2% ~ 2.4% when erase(begin(),
end()).
My test environment:
CPU: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GHz, 2393.365 MHz, 56 CPUs
MEM: 256G
OS: CentOS-8.2
g++: gcc version 8.3.1 20191121 (Red Hat 8.3.1-5) (GCC)
Compile flags: -O3 -std=c++17
Test conclusion data (time taken to delete every 100 million elements):
erase(begin()):
|size of unordered_set |100 |1,000 |10,000|100,000 |1,000,000|10,000,000|
|base time consuming (ms)|3827.736|3807.725|3830.168|3807.373|3798.713
|3854.168|
|test time consuming (ms)|3783.406|3789.460|3791.146|3778.033|3783.494
|3808.137|
|Time-consuming reduction|1.16% |0.48% |1.02% |0.77% |0.40%|1.19% |
erase(begin(),end()):
|size of unordered_set |100 |1,000 |10,000|100,000 |1,000,000|10,000,000|
|base time consuming (ms)|2779.229|2768.550|2795.778|2767.385|2761.521
|2804.099|
|test time consuming (ms)|2712.759|2726.578|2752.224|2732.140|2718.953
|2739.727|
|Time-consuming reduction|2.39% |1.52% |1.56% |1.27% |1.54%|2.30% |
Please see the attachment for test code and detailed test result.
2024年1月18日(木) 4:04 François Dumont <frs.dum...@gmail.com>:
Hi
Looks like a great finding to me, this is indeed a useless check,
thanks!
Have you any figures on the performance enhancement ? It might
help to get proper approval as gcc is currently in dev stage 4
that is to say only bug fixes normally.
François
On 17/01/2024 09:11, Huanghui Nie wrote:
Hi.
When I implemented a hash table with reference to the C++ STL, I
found that when the hash table in the C++ STL deletes elements,
if the first element deleted is the begin element, the before
begin node is repeatedly assigned. This creates unnecessary
performance overhead.
First, let’s see the code implementation:
In _M_remove_bucket_begin, _M_before_begin._M_nxt is assigned
when &_M_before_begin == _M_buckets[__bkt]. That also means
_M_buckets[__bkt]->_M_nxt is assigned under some conditions.
_M_remove_bucket_begin is called by _M_erase and _M_extract_node:
1. Case _M_erase a range: _M_remove_bucket_begin is called in a
for loop when __is_bucket_begin is true. And if
__is_bucket_begin is true and &_M_before_begin ==
_M_buckets[__bkt], __prev_n must be &_M_before_begin.
__prev_n->_M_nxt is always assigned in _M_erase. That means
_M_before_begin._M_nxt is always assigned, if
_M_remove_bucket_begin is called and &_M_before_begin ==
_M_buckets[__bkt]. So there’s no need to assign
_M_before_begin._M_nxt in _M_remove_bucket_begin.
2. Other cases: _M_remove_bucket_begin is called when __prev_n
== _M_buckets[__bkt]. And __prev_n->_M_nxt is always assigned
in _M_erase and _M_before_begin. That means
_M_buckets[__bkt]->_M_nxt is always assigned. So there's no
need to assign _M_buckets[__bkt]->_M_nxt in
_M_remove_bucket_begin.
In summary, there’s no need to check &_M_before_begin ==
_M_buckets[__bkt] and assign _M_before_begin._M_nxt in
_M_remove_bucket_begin.
Then let’s see the responsibility of each method:
The hash table in the C++ STL is composed of hash buckets and a
node list. The update of the node list is responsible for
_M_erase and _M_extract_node method. _M_remove_bucket_begin
method only needs to update the hash buckets. The update of
_M_before_begin belongs to the update of the node list. So
_M_remove_bucket_begin doesn’t need to update _M_before_begin.
Existing tests listed below cover this change:
23_containers/unordered_set/allocator/copy.cc
23_containers/unordered_set/allocator/copy_assign.cc
23_containers/unordered_set/allocator/move.cc
23_containers/unordered_set/allocator/move_assign.cc
23_containers/unordered_set/allocator/swap.cc
23_containers/unordered_set/erase/1.cc
23_containers/unordered_set/erase/24061-set.cc
23_containers/unordered_set/modifiers/extract.cc
23_containers/unordered_set/operations/count.cc
23_containers/unordered_set/requirements/exception/basic.cc
23_containers/unordered_map/allocator/copy.cc
23_containers/unordered_map/allocator/copy_assign.cc
23_containers/unordered_map/allocator/move.cc
23_containers/unordered_map/allocator/move_assign.cc
23_containers/unordered_map/allocator/swap.cc
23_containers/unordered_map/erase/1.cc
23_containers/unordered_map/erase/24061-map.cc
23_containers/unordered_map/modifiers/extract.cc
23_containers/unordered_map/modifiers/move_assign.cc
23_containers/unordered_map/operations/count.cc
23_containers/unordered_map/requirements/exception/basic.cc
Regression tested on x86_64-pc-linux-gnu. Is it OK to commit?
---
ChangeLog:
libstdc++: hashtable: No need to update before begin node in
_M_remove_bucket_begin
2024-01-16Huanghui Nie<nnn...@gmail.com>
gcc/
* libstdc++-v3/include/bits/hashtable.h
---
diff --git a/libstdc++-v3/include/bits/hashtable.h
b/libstdc++-v3/include/bits/hashtable.h
index b48610036fa..6056639e663 100644
--- a/libstdc++-v3/include/bits/hashtable.h
+++ b/libstdc++-v3/include/bits/hashtable.h
@@ -872,13 +872,10 @@ _GLIBCXX_BEGIN_NAMESPACE_VERSION
if (!__next_n || __next_bkt != __bkt)
{
// Bucket is now empty
- // First update next bucket if any
+ // Update next bucket if any
if (__next_n)
_M_buckets[__next_bkt] = _M_buckets[__bkt];
- // Second update before begin node if necessary
- if (&_M_before_begin == _M_buckets[__bkt])
- _M_before_begin._M_nxt = __next_n;
_M_buckets[__bkt] = nullptr;
}
}
