commit e5a7979d36a84831652abf3138c4d76797c002b5
author Abseil Team <absl-team@google.com> Mon Nov 28 12:27:06 2022 -0800
committer Copybara-Service <copybara-worker@google.com> Mon Nov 28 12:27:43 2022 -0800
tree e7dc92ce03b1b04428b588f33bddf657ea97f554
parent e3158086978de4ff15a2058f49ba525c89e14ce6

Reduce flat_hash_{set,map} generated code size.

This CL makes a bunch of changes (mostly to raw_hash_set which
underlies flat_hash_set and flat_hash_map). Techniques used:

* Extract code that does not depend on the specific hash table type
  into common (non-inlined) functions.
* Place ABSL_ATTRIBUTE_NOINLINE directives judiciously.
* Out-of-line some slow paths.

Reduces sizes of some large binaries by ~0.5%.

Has no significant performance impact on a few performance critical
binaries.

## Speed of fleetbench micro-benchmarks

Following is a histogram of %-age changes in
[fleetbench](https://github.com/google/fleetbench)
hot_swissmap_benchmark results. Negative numbers indicate a speedup
caused by this change. Statistically insignificant changes are mapped
to zero.

XXX Also run and merge in cold_swissmap_benchmark

Across all 351 benchmarks, the average speedup is 0.38%.
The best speedup was -25%, worst slowdown was +6.81%.

```
Count: 351  Average: -0.382764  StdDev: 3.77807
Min: -25  Median: 0.435135  Max: 6.81
---------------------------------------------
[ -25, -10)  16  4.558%   4.558% #
[  -9,  -8)   2  0.570%   5.128%
[  -8,  -7)   1  0.285%   5.413%
[  -7,  -6)   1  0.285%   5.698%
[  -6,  -5)   2  0.570%   6.268%
[  -5,  -4)   5  1.425%   7.692%
[  -4,  -3)  13  3.704%  11.396% #
[  -3,  -2)  15  4.274%  15.670% #
[  -2,  -1)  26  7.407%  23.077% ##
[  -1,   0)  14  3.989%  27.066% #
[   0,   1) 185 52.707%  79.772% ############
[   1,   2)  14  3.989%  83.761% #
[   2,   3)   8  2.279%  86.040% #
[   3,   4)   7  1.994%  88.034%
[   4,   5)  32  9.117%  97.151% ##
[   5,   6)   6  1.709%  98.860%
[   6,   7)   4  1.140% 100.000%
```

We looked at the slowdowns and they do not seem worth worrying
about. E.g., the worst one was:

```
BM_FindHit_Hot<::absl::node_hash_set,64>/set_size:4096/density:0
  2.61ns ± 1%  2.79ns ± 1%   +6.81%  (p=0.008 n=5+5)
```

## Detailed changes

* Out-of-line slow paths in hash table sampler methods.
* Explicitly unregister from sampler instead of from destructor.
* Introduced a non-templated CommonFields struct that holds some of
  the hash table fields (infoz, ctrl, slots, size, capacity). This
  struct can be passed to new non-templated helpers. The struct is
  a private base class of raw_hash_set.
* Made non-inlined InitializeSlots<> that is only templated on
  allocator and size/alignment of the slot type so that we can share
  instantiations across types that have the same size/alignment.
* Moved some infrequently called code paths into non-inlined type-erased.
  functions. Pass a suite of type-specific function pointers to these
  routines for when they need to operate on slots.
* Marked some methods as non-inlined.
* Avoid unnecessary reinitialization in destructor.
* Introduce UpdateSpine type-erased helper that is called from
  clear() and rehash().

PiperOrigin-RevId: 491413386
Change-Id: Ia5495c5a6ec73622a785a0d260e406ddb9085a7c