The prior method was to generate one line of
`g_container.emplace_back(…)` per entry in the list & letting the
compiler chew on that. Each string argument in that call was done was
`u8"Some Name"s`, meaning as a std::string instance.
Drawbacks:
• takes the compiler ages to compile, even forcing me to drop all
optimizations for the ISO-639 language list file
• even smaller files such as the IANA language subtag registry lists
take more than 30s to compile
• due to no optimizations initialization is actually not as fast as
could be
The new method uses a plain C-style array of structs with `char
const *` entries for the initial list. The initialization method then
copies the entries from that list to the actual container, again using
`std::emplace_back(…)`.
This yields sub-1s compilation times even with the longest file, the
ISO-639 language list, and the runtime initialization is actually
faster.
Even `-O1` causes compilation time & memory usage to skyrocket,
possibly exponentially, with the number of entries to `emplace_back()`
into the vector.
This isn't so bad with the current number of entries (489). In that
case compilation with `-O3` only takes 7.2s.
However, extending the list to cover ISO 639-3 means that the list
will include 7160 entries. With that many entries things are much,
much more severe:
• with `-O1` alone compilation takes 11m 23s already.
• with `-O3` memory usage exceeded 20 GB after six minutes when I had
to abort due to other running applications getting killed.
Runtime cost is negligible. I ran a micro benchmark. With all 7160
entries and no optimizations (`-O0`) the initialization takes ~1.4
milliseconds for the one-time initialization on startup; with
optimizations (`-O1`) it still took ~570 microseconds.
Part of the implementation of #3007.
It's much faster than using the initializer lists. Here's the result
from a micro benchmark I ran:
2021-01-25T23:49:20+01:00
Running ./bench.g++
Run on (8 X 4500 MHz CPU s)
CPU Caches:
L1 Data 32 KiB (x4)
L1 Instruction 32 KiB (x4)
L2 Unified 256 KiB (x4)
L3 Unified 8192 KiB (x1)
Load Average: 1.08, 0.72, 0.60
-------------------------------------------------------------
Benchmark Time CPU Iterations
-------------------------------------------------------------
BM_InitializerList 59667 ns 59625 ns 70526
BM_EmplaceBack 24515 ns 24497 ns 176817
BM_EmplaceBack2 16970 ns 16961 ns 247652
BM_PushBack 52831 ns 52796 ns 79202
BM_PushBack2 52858 ns 52823 ns 79004
The five benchmarks were:
• BM_InitializerList — the old way with initializer lists. Basically
the same code currently being replaced.
• BM_EmplaceBack — Reserving space & adding each entry with
g_languages.emplace_back(). A constructor was added to language_t
struct taking the std::strings as const references (std::string
const &), assigning them to the member variables normally.
• BM_EmplaceBack2 — Reserving space & adding each entry with
g_languages.emplace_back(). A constructor was added to language_t
struct taking the std::strings as rvalue references (std::string &&)
assigning them to the member variables using std::move().
• BM_PushBack — Reserving space & adding each entry with
g_languages.push_back(). A constructor was added to language_t
struct taking the std::strings as const references (std::string
const &), assigning them to the member variables normally.
• BM_PushBack2 — Reserving space & adding each entry with
g_languages.push_back(). A constructor was added to language_t
struct taking the std::strings as rvalue references (std::string &&)
assigning them to the member variables using std::move().
