clang-cl: #3285 dramatically increased compile times

[Kojoley]

#3285 switched from <intrin0.h> to <intrin.h> which is 50x slower on clang-cl.

msvc	clang was	clang now	slow down	header
0.068	0.082	0.369	351%	<limits>
0.093	0.101	0.387	284%	<atomic>
0.097	0.103	0.389	277%	<latch>
0.102	0.110	0.396	259%	<bit>
0.111	0.126	0.410	226%	<compare>
0.119	0.133	0.424	219%	<coroutine>
0.119	0.134	0.419	213%	<utility>
0.120	0.134	0.419	212%	<typeindex>
0.141	0.155	0.448	189%	<tuple>
0.179	0.206	0.491	139%	<array>
0.178	0.207	0.493	138%	<span>
0.180	0.207	0.491	137%	<numeric>
0.185	0.215	0.499	132%	<iterator>
0.195	0.222	0.511	131%	<semaphore>
0.191	0.219	0.504	130%	<expected>
0.193	0.229	0.513	124%	<cmath>
0.213	0.243	0.528	117%	<scoped_allocator>
0.217	0.245	0.530	116%	<any>
0.222	0.250	0.535	114%	<deque>
0.219	0.249	0.532	114%	<forward_list>
0.221	0.250	0.533	113%	<optional>
0.231	0.264	0.560	112%	<list>
0.226	0.258	0.539	109%	<set>
0.229	0.262	0.546	109%	<vector>
0.229	0.260	0.541	108%	<map>
0.245	0.283	0.576	104%	<barrier>
0.243	0.278	0.563	102%	<stop_token>
0.255	0.301	0.595	97%	<valarray>
0.258	0.299	0.582	95%	<charconv>
0.266	0.307	0.589	92%	<memory>
0.271	0.319	0.602	89%	<unordered_set>
0.287	0.328	0.620	89%	<variant>
0.272	0.319	0.603	89%	<unordered_map>
0.300	0.328	0.613	87%	<string_view>
0.302	0.330	0.614	86%	<stdexcept>
0.302	0.332	0.618	86%	<bitset>
0.296	0.343	0.626	83%	<functional>
0.300	0.345	0.629	82%	<algorithm>
0.319	0.351	0.634	81%	<string>
0.321	0.358	0.643	80%	<thread>
0.342	0.371	0.655	77%	<system_error>
0.359	0.397	0.685	72%	<stacktrace>
0.438	0.480	0.774	61%	<streambuf>
0.444	0.476	0.761	60%	<shared_mutex>
0.434	0.471	0.753	60%	<mutex>
0.441	0.476	0.759	59%	<condition_variable>
0.459	0.511	0.798	56%	<ios>
0.416	0.506	0.789	56%	<ranges>
0.424	0.518	0.799	54%	<stack>
0.505	0.528	0.811	54%	<memory_resource>
0.481	0.541	0.826	52%	<codecvt>
0.511	0.539	0.820	52%	<locale>
0.533	0.649	0.939	45%	<queue>
0.654	0.764	1.056	38%	<random>
0.666	0.762	1.048	38%	<format>
0.692	0.787	1.075	37%	<fstream>
0.691	0.788	1.075	37%	<syncstream>
0.724	0.824	1.121	36%	<iomanip>
0.689	0.787	1.070	36%	<istream>
0.658	0.794	1.079	36%	<regex>
0.693	0.790	1.073	36%	<spanstream>
0.714	0.803	1.089	36%	<sstream>
0.690	0.787	1.066	35%	<iostream>
0.731	0.833	1.125	35%	<strstream>
0.766	0.866	1.166	35%	<complex>
0.684	0.827	1.112	35%	<ostream>
0.807	0.853	1.143	34%	<print>
0.725	0.870	1.152	32%	<execution>
1.117	1.238	1.520	23%	<chrono>
1.388	1.512	1.795	19%	<filesystem>

[StephanTLavavej]

This was a known consequence of merging #3285. As Casey mentioned there:

The existence of intrin0.h is purely an optimization in that it allows us to avoid including the entirety of intrin.h. We could support this optimization with Clang in the future by adding an intrin0.h to Clang's set of override headers. If we were to do so we should synchronize with libc++ on the proper set of intrinsics to move into intrin0.h.

[CaseyCarter]

Related: LLVM-53520. Both hopefully to be fixed by llvm/llvm-project#75711.

[CaseyCarter]

llvm/llvm-project#75711 merged in 2024-03, this is fixed.