Compiler performance, do we want to build the release with a specific compiler?

[illwieckz]

On the same hardware and operating system:

• CPU: AMD FX 9590
  performance profile, 8 cores, 4.4GHz (a bit underclocked from stock 4.7GHz for stability), Piledriver architecture
• RAM: 16 GB, 1600 MHz, DDR3
• GPU: AMD R9 390X
  high performance profile, 44 compute units, 1050MHz, GCN2/Grenada XT architecture, PCIe 2.0
• VRAM: 8GB, 1500 MHz, GDDR5
• OS: Ubuntu 20.04.2 LTS “Focal Fossa” with Mesa radeonsi 21.0.3

I tested the /devmap plat23 ; delay 100f setviewpos 1893 1920 0 0 0 scene (a scene with many animated models) with 640×480 resolution and lowest preset to minimize GPU computation and r_vboVertexSkinning 0 to use the CPU code path to animate models and then stress test a lot the CPU.

I built the game with RelWithDebInfo build profile and ran the game with .nexe.

Here is the performance I get:

Compiler	FPS	Note
GCC 9.3.0	42	Ubuntu 20.04.2 default GCC
GCC 9.3.0 + LTO	46	Ubuntu 20.04.2 default GCC
GCC 10.3.0	10	Ubuntu 20.04.2 alternate GCC
GCC 10.3.0 + LTO	10	Ubuntu 20.04.2 alternate GCC
Clang 10.0.0	47	Ubuntu 20.04.2 default Clang
Clang 12.0.0	48	Ubuntu 20.04.2 alternate Clang
Clang 12.0.0 + LTO	49	Ubuntu 20.04.2 alternate Clang

Here are details for 20 second capture (1 line is the average FPS in a second) for those compilers, such line print is produced by setting LIBGL_SHOW_FPS=1 environment variable on Mesa:

GCC 9

libGL: FPS = 43.00
libGL: FPS = 42.00
libGL: FPS = 43.00
libGL: FPS = 42.00
libGL: FPS = 42.00
libGL: FPS = 42.29
libGL: FPS = 42.00
libGL: FPS = 42.29
libGL: FPS = 42.00
libGL: FPS = 42.29
libGL: FPS = 43.00
libGL: FPS = 42.29
libGL: FPS = 42.29
libGL: FPS = 43.00
libGL: FPS = 42.00
libGL: FPS = 42.00
libGL: FPS = 42.29
libGL: FPS = 43.00
libGL: FPS = 42.00
libGL: FPS = 42.00
libGL: FPS = 42.29

GCC 9 + LTO

libGL: FPS = 46.00
libGL: FPS = 46.00
libGL: FPS = 46.00
libGL: FPS = 46.00
libGL: FPS = 45.24
libGL: FPS = 46.23
libGL: FPS = 46.00
libGL: FPS = 45.00
libGL: FPS = 46.00
libGL: FPS = 45.24
libGL: FPS = 46.00
libGL: FPS = 45.00
libGL: FPS = 45.00
libGL: FPS = 46.00
libGL: FPS = 46.00
libGL: FPS = 45.24
libGL: FPS = 46.00
libGL: FPS = 45.24
libGL: FPS = 46.00
libGL: FPS = 46.00

GCC 10

libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 11.00
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 10.64
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.64
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 10.64

GCC 10 + LTO

libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 10.48
libGL: FPS = 10.82
libGL: FPS = 11.00
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.64
libGL: FPS = 10.82
libGL: FPS = 10.64
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.64
libGL: FPS = 10.82
libGL: FPS = 10.82
libGL: FPS = 10.82

Clang 10

libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.21
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 46.00
libGL: FPS = 47.00
libGL: FPS = 46.23
libGL: FPS = 46.23
libGL: FPS = 47.00
libGL: FPS = 46.23
libGL: FPS = 47.00
libGL: FPS = 47.00
libGL: FPS = 46.23
libGL: FPS = 47.00

Clang 12

libGL: FPS = 48.00
libGL: FPS = 49.00
libGL: FPS = 48.19
libGL: FPS = 48.19
libGL: FPS = 48.19
libGL: FPS = 48.19
libGL: FPS = 48.19
libGL: FPS = 48.19
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 47.00
libGL: FPS = 48.00
libGL: FPS = 47.21
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 48.00
libGL: FPS = 47.21

Clang 12 + LTO

libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 48.19
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 48.00
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 48.19
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 48.00
libGL: FPS = 49.00
libGL: FPS = 49.00
libGL: FPS = 48.00
libGL: FPS = 49.18
libGL: FPS = 49.00

[illwieckz]

I updated times with some LTO builds.

With GCC 9, LTO is a big win, it jumps from 42 fps to 46 fps.

With Clang 12, LTO adds one more FPS, it jumps from solid 48 fps to 49 fps.

Note: for Clang 12, the LTO build was done with lld12, the non-LTO Clang 12 build performs the same with ld and lld12.

I was not able to produce LTO builds with Clang 10, and LTO builds with GCC 10 just adds nothing. I don't know what happens with GCC 10.