Optimize LimitedSet and LimitedMap (#10)

* In `LimitedSet` and `LimitedMap` use new optimized code for `std::less` and make that the default.

* Update format

* Fix comment

* Implement separate benchmark for limited_set (tests limited_ordered).
* Add `LimitedOptionsFlag::kNoOptimizeIndexOf` (disallows unrolling, etc).
* The benchmark tests all versions.
* The normal set test is expanded to verify all versions.
* Further optimizations for `index_of`.
  * Extend unrolling to 32 cases (might need to be even 34).
  * Explain unlikely cases in compiler-hints.

Performance analysis on a x86_64 system shows better or same performance
for `LimitedSet` compared to `std::set`. This has to be further analyzed
and tweaked.

* Improve benhchmark

* In `LimitedSet` and `LimitedMap` use new optimized code for `std::less` and make that the default.

* Fix comment

* Implement separate benchmark for limited_set (tests limited_ordered).
* Add `LimitedOptionsFlag::kNoOptimizeIndexOf` (disallows unrolling, etc).
* The benchmark tests all versions.
* The normal set test is expanded to verify all versions.
* Further optimizations for `index_of`.
  * Extend unrolling to 32 cases (might need to be even 34).
  * Explain unlikely cases in compiler-hints.

Performance analysis on a x86_64 system shows better or same performance
for `LimitedSet` compared to `std::set`. This has to be further analyzed
and tweaked.

* Add missing private marker

* Mark benchmark as manual

* Test iterator difference

* A few more micro optimizations:
* Mark additional wrapper functions as force inline
* Provide additional branch guidance

* Fix after sync.

* Optimize according to performance measurements:
* By default do not optimize past loop-unroll.
* Change loop unroll to 24.

* Add missing new line

* Cleanup and move LimitedOrdered unrolling config into separate file.
That file should be configurable outside - maybe via a bazel flag,
maybe by generating the file, possibly a configuration.

* Added custom bazel flag `--//mbo/container:limited_ordered_max_unroll_capacity=N`.

* Add explanation to static_assert.
* Add bzl library `//mbo/container:limited_ordered_config_bzl`.

.bazelrc

@@ -14,6 +14,8 @@ build --define absl=1
 build --cxxopt=-std=c++20 --host_cxxopt=-std=c++20 --cxxopt=-Werror
 test --test_output=errors
 
+build --copt=-fdiagnostics-color=always
+
 # Custom --config=asan mode:
 build:asan --copt -fsanitize=address,undefined
 build:asan --linkopt -fsanitize=address,undefined

.clang-format

@@ -24,6 +24,7 @@ BreakInheritanceList: BeforeComma
 DerivePointerAlignment: false
 EmptyLineBeforeAccessModifier: Always
 FixNamespaceComments: true
+IndentPPDirectives: AfterHash
 IndentRequiresClause: false
 InsertBraces: true
 InsertNewlineAtEOF: true
@@ -36,6 +37,7 @@ MaxEmptyLinesToKeep: 1
 PackConstructorInitializers: NextLine
 PenaltyReturnTypeOnItsOwnLine: 1000
 PointerAlignment: Left
+PPIndentWidth: 1
 QualifierAlignment: Left
 ReferenceAlignment: Left
 RemoveSemicolon: true

.gitignore

@@ -8,4 +8,4 @@
 /compile_commands.json
 # Ignore the directory in which `clangd` stores its local index.
 /.cache/
-/.vscode/
+/.vscode/

CHANGELOG.md

@@ -1,3 +1,9 @@
+# 0.2.18
+
+* Changed `LimitedSet` and `LimitedMap`:
+  * Changed to new optimized code for `std::less` and make that the default.
+  * Added custom bazel flag `--//mbo/container:limited_ordered_max_unroll_capacity=N`. This controls the maximum capacity `N` for which `index_of` will be unrolled (defaults to 16, range [4...32], see `mbo::container::container_internal::kUnrollMaxCapacityDefault`.
+
 # 0.2.17
 
 * Added template-type `RefWrap<T>`: similar to `std::reference_wrapper` but supports operators `->` and `*`.

mbo/container/BUILD.bazel

@@ -12,7 +12,30 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-package(default_visibility = ["//visibility:private"])
+load("@bazel_skylib//:bzl_library.bzl", "bzl_library")
+load("@bazel_skylib//rules:common_settings.bzl", "int_flag")
+load(":internal/limited_ordered_config.bzl", "limited_ordered_config_gen")
+
+# Create custom bazel flag `--//mbo/container:limited_ordered_max_unroll_capacity`.
+int_flag(
+    name = "limited_ordered_max_unroll_capacity",
+    build_setting_default = 16,
+    visibility = ["//visibility:private"],
+)
+
+bzl_library(
+    name = "limited_ordered_config_bzl",
+    srcs = [":internal/limited_ordered_config.bzl"],
+    visibility = ["//visibility:private"],
+    deps = ["@bazel_skylib//rules:common_settings"],
+)
+
+limited_ordered_config_gen(
+    name = "limited_ordered_config_gen",
+    visibility = ["//visibility:private"],
+    output = "internal/limited_ordered_config.h",
+    template = "internal/limited_ordered_config.h.in"
+)
 
 cc_library(
     name = "any_scan_cc",
@@ -94,7 +117,10 @@ cc_library(
 
 cc_library(
     name = "limited_ordered_cc",
-    hdrs = ["internal/limited_ordered.h"],
+    hdrs = [
+      "internal/limited_ordered_config.h",
+      "internal/limited_ordered.h",
+    ],
     visibility = ["//visibility:private"],
     deps = [
         ":limited_options_cc",
@@ -141,6 +167,21 @@ cc_test(
     ],
 )
 
+cc_binary(
+    name = "limited_set_benchmark",
+    visibility = ["//visibility:private"],
+    srcs = ["limited_set_benchmark.cc"],
+    deps = [
+        ":limited_set_cc",
+        ":limited_vector_cc",
+        "//mbo/testing:matchers_cc",
+        "@com_github_google_benchmark//:benchmark",
+        "@com_google_absl//absl/log:initialize",
+        "@com_google_absl//absl/strings",
+    ],
+    tags = ["manual"],
+)
+
 cc_library(
     name = "limited_vector_cc",
     hdrs = ["limited_vector.h"],