Are precompiled headers worthwhile?

[slipher]

I timed building the client, and cgame+sgame DLL targets with the following commands.

for x in 1 2 3 4 5; do make clean; time(make -j15 client  > /dev/null 2> /dev/null); done

for x in 1 2 3 4 5; do make clean; time(make -j15 cgame-native-dll sgame-native-dll > /dev/null 2> /dev/null); done

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	16.6	17.0	32.0	32.4
PCH on	12.9	13.0	24.8	25.0

So a 20-25% speedup. Seems not really worth it?

[illwieckz]

I don't have much opinion on that, if it can make the build faster it's welcome, to me 25% looks very good when the computer running isn't a very fast computer.

Of course with my ThreadRipper I probably never felt the difference, but not anyone has such beast. 15~25 minutes for 5 build means 3~5 minutes per build, which is very fast. People achieving 3~5 minutes per build (like I do) are likely not the target users of PCH.

[illwieckz]

Ah no I misread the table, it's not 25 minutes for 5 builds but 25 seconds per build, 5 times… Well. 25 seconds looks very fast.

Also I don't know why I wrongly remembered a full build was lasting 5 minutes on my end, maybe I counted all engines + nacl games (which is not what is benchmarked here).

[illwieckz]

I did some more benchmarks, with a large diversity of hardware, I used those commands:

for x in {1..5}; do make clean; time (make -j"$(nproc)" client >/dev/null 2>&1); done
for x in {1..5}; do make clean; time (make -j"$(nproc)" cgame-native-dll sgame-native-dll >/dev/null 2>&1); done

• AMD Ryzen Threadripper PRO 3955WX, (amd64, 16c32t, Zen2):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	19.9	21.3	32.9	34.4
PCH on	18.0	19.1	26.7	27.3

• AMD Ryzen Z1 Extreme (amd64, 8c16t, Zen4):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	30.4	32.6	1m2.0	1m10.0
PCH on	25.7	30.5	48.2	1m2.3

• AMD Ryzen 3 3200G (amd64, 4c4t, Zen+):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	1m24.7	1m35.7	2m47.7	2m56.0
PCH on	1m3.3	1m8.8	2m4.3	2m15.1

• Intel Core i7-6500U (amd64, 2c4t, Skylake):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	2m33.0	2m39.9	5m7.0	5m26.2
PCH on	1m44.6	1m45.6	3m36.8	3m40.8

• Intel Core i7-4810MQ (amd64, 4c8t, Haswell):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	1m43.8	1m56.1	3m43.2	3m59.7
PCH on	56.5	1m1.6	2m16.2	2m24.8

• Intel Core 2 Duo L7500 (amd64, 1c2t, Merom):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	7m33.7	7m47.8	15m20.8	18m11.0
PCH on	5m9.9	5m31.8	10m49.9	11m9.8

• Amlogic S905X3 (arm64, 4c8t, Cortex-A55):

PCH status	Client min time (s)	Client max time (s)	Gamelogic min time (s)	Gamelogic max time (s)
PCH off	1m49,2	1m56.3	3m34.3	3m39.8
PCH on	1m17.4	1m19.8	2m33.2	2m40.8

[illwieckz]

So, with hardware allowing less than 16 jobs per CPU, we start to see a significant difference between PCH disabled or enabled.