Clang vs. GCC for building Emacs

Several days ago, I tried to play with Web Assembly and found out it requires Clang to run. Clang has become the default compiler for many BSD-based operating systems. And I was told from multiple sources that it is faster than GCC both in compilation time and binary speed. However, I had no personal experience with it and thus set out to gain some.

I am no expert in performance testing, so don’t blame me if I made some mistakes.

I used the Bash shell keyword time to obtain time consumed by the building process. I added the option -rtlib=compiler-rt to LDFLAGS, since Clang would complain undefined reference to `__muloti4' without it (see bug). Clang 5.0 also provides a linker called lld and claims that it is faster than the built-in linker. As per instruction, I added -fuse-ld=lld to LDFLAGS as a third group.

The following was the commands I used for each group.

# for GCC
time make bootstrap
# for Clang without lld
time make CC=clang LDFLAGS='-rtlib=compiler-rt' bootstrap
# for Clang with lld
time make CC=clang LDFLAGS='-fuse-ld=lld -rtlib=compiler-rt' bootstrap

The following was the results.

time make bootstrap

time make check

time make CC=clang LDFLAGS='-rtlib=compiler-rt' bootstrap

time make check

time make CC=clang LDFLAGS='-fuse-ld=lld -rtlib=compiler-rt' bootstrap

time make check

On the next day, I re-compiled Emacs using GCC to compare binary functionality but accidentally found out that GCC compiled much faster than last time. I don’t know what exactly the boost came from. I surmise that it was due to reboot. Therefore, I ran a follow-up and the following was results on the second day.

time make bootstrap

time make CC=clang LDFLAGS='-fuse-ld=lld -rtlib=compiler-rt' bootstrap

time make bootstrap

time make CC=clang LDFLAGS='-fuse-ld=lld -rtlib=compiler-rt' bootstrap

In conclusion, Clang, particularly with lld, is slightly faster.