[MANUAL SQUASH] Precompiled headers support

[Desour]

Recent CMake has a target_precompile_headers function that makes using precompiled headers pretty trivial.
(For comparison, see #8643. (cc @numberZero))

Goal: Faster compilation, climate protection, less turbine noise.

1st half of commit:
Moves sha256.c to lib/. This is needed because precompiled headers don't work (assume C) if we're mixing C and C++.

2nd half of commit:
Precompiled header option. (Requires new enough cmake.)

Which headers are precompiled:

• We only want to precompile headers that seldom or never change. Otherwise we'd have to recompile the precompiled header and recompile everything else when the header changes. And ccache can't know if the precompiled header is different, so even just switching between branches would be horribly painful.
• => Only stdlib and lib headers are precompiled (excluding irrlicht).
• I've taken the sdlib headers from Add precompiled header support #8643, and added jsoncpp. (Could add more.)
• The list of headers is read from a file (src/precompiled_headers.txt). The user can configure this.

To do

This PR is a Ready for Review.

How to test

• If you're using ccache, add to your ccache.conf: sloppiness = pch_defines,time_macros
• $ cmake . -DPRECOMPILE_HEADERS=1
• $ make -j<you choose>

Measure speed:

• $ cmake . -DCMAKE_CXX_COMPILER=/usr/bin/g++
• $ make -B -j5
• Results: (may be inaccurate)
  • No precompile:

    real	7m31,836s
    user	34m38,999s
    sys	1m53,541s
    

  • Precompile:

    real	6m16,476s
    user	28m39,059s
    sys	1m32,657s
    

  • => ca. 17% faster

[lhofhansl]

I'd be worried the additional fragility in the process is not warranted by a 17% compilation time decrease.

[numberZero]

I tested this (8812b79) with 3 build directories (warm-up, PCH disabled, and PCH enabled) on my system (Ryzen 7 5800X). Here is time comparison:

Target	Real, no PCH	Real, PCH	Real, diff	CPU, no PCH	CPU, PCH	CPU, diff
Clean build	56	48	−15%	785	629	−20%
Rebuild after touching client/content_mapblock.cpp	6.0	5.6	−6%	6.0	5.6	−6%
Rebuild after touching client/tile.h	41	35	−16%	542	437	−19%

Time is in seconds. Time diff is percentage of “no PCH.”

17% faster

Corresponds to −17% “Real, diff” (Clean build) in my format. Which is just a bit better than what I get.

Here are some details:

Without PCH

$ cmake -GNinja -DCMAKE_BUILD_TYPE=Debug -DRUN_IN_PLACE=ON -DPRECOMPILE_HEADERS=OFF
<...>
[482/482] Linking CXX executable <...>/bin/minetest
real    0m56,241s
user    12m13,479s
sys     0m51,037s

$ touch src/client/content_mapblock.cpp
<...>
[4/4] Linking CXX executable <...>/bin/minetest
real    0m5,991s
user    0m5,480s
sys     0m0,510s

$ touch src/client/tile.h
<...>
[178/178] Linking CXX executable <...>/bin/minetest
real    0m41,414s
user    8m29,035s
sys     0m32,512s

With PCH

$ cmake -GNinja -DCMAKE_BUILD_TYPE=Debug -DRUN_IN_PLACE=ON -DPRECOMPILE_HEADERS=ON
<...>
[483/483] Linking CXX executable <...>/bin/minetest
real    0m48,005s
user    9m41,919s
sys     0m46,652s

$ touch src/client/content_mapblock.cpp
<...>
[6/6] Linking CXX executable <...>/bin/minetest
real    0m5,616s
user    0m5,125s
sys     0m0,508s

$ touch src/client/tile.h
<...>
[177/177] Linking CXX executable <...>/bin/minetest
real    0m34,724s
user    6m47,845s
sys     0m29,482s

Conclusion

While single source file updates are dominated by link time, there is a clear speedup for header updates which tend to require massive recompilation, and for full rebuilds too.

[Desour]

I'd be worried the additional fragility in the process is not warranted by a 17% compilation time decrease.

It's optional and marked as experimental. And CI doesn't use it, so missing includes and similar should usually be found.

Thanks for testing, @numberZero!

[numberZero]

missing includes

Which may very well happen without any PCH because e.g. today, <vector> may include <cstdint> so your code will compile fine... until STL upgrade.

[Zughy]

@Desour rebase needed

[Desour]

Rebased and comments addressed.

[sfan5]

Works for me but needs a rebase.