GCC builds are ~30% faster than Clang/MSVC builds

[dwatteau]

Observed on macOS Mojave x86-64 with Clang 11.0.0, and OpenBSD/macppc 6.8 with Clang 10.0.1.

scummtr is approximately 30% faster with GCC than with Clang, when importing a translation.

GCC 10.2.0 and GCC 4.2.1 at -O2 are both faster than Clang at -O2 or -O3, so I don't think that it's a recent GCC improvement, since GCC 4.2.1 is from 2007.

There's no need for this set of tools to be extremely performant, but a reproducible 30% difference seems strange.

Possible reasons:

• Clang really is slower than GCC
• GCC enables a useful optimisation that Clang doesn't enable (but Clang was tested with -O3, too)
• we have some undefined behaviour somewhere, and GCC chooses to optimise this away, while Clang is more strict.

[dwatteau]

And, on the same machine, it seems that a release MSVC build is even a bit slower than the Clang build.

Roughly, importing a translation in the floppy version of ATLANTIS takes:

• 5.6 seconds with a GCC release build
• 8.2 seconds with a Clang release build
• 9.9 seconds with an MSVC release build

I'll try playing with a bit of profiling to see where it seems to be slower.