RobRich999
commented
2 years ago AVX is a superset of SSE4, so there is no need to specify SSE4 or any other lower SIMD instruction set.
Hardware AES has been supported since Westmere and Bulldozer, so it should be fine to enable it as well for AVX and higher builds.
I did pull back rust targets optimized for size to -Os from the previous -O3 config, as I doubt we have PGO profiling for any rust code at this time, so we might as well do some manual tuning of code bloat.
if (optimize_for_size) { rustflags = [ "-Copt-level=s", "-Ctarget-feature=+avx,+aes", ]
Additionally, you might try passing in VEAS when enabliing AES with AVX or higher SIMD. It can promote some AES instructions to AVX registers. That should be -mvaes for cflags and +vaes for rust flags. YMMV.
Hello. My rust lines in the main compiler build.gn all have copt level = 3 but they also have arch specific lines afterwards. However the difference is that mine has sse4,avx,aes, while yours just have avx or avx2
See this line (yours) >
rustflags += [ "-Copt-level=3", "-Ctarget-feature=+avx", ]versus mine >
rustflags += [ "-Ccodegen-units=1", "-Copt-level=3", "-Ctarget-feature=+sse4,+aes,+avx", ]By using mine, will this cause it to use sse4 in places where it would otherwise use avx? Should I make it have just avx and aes?
Also, regarding aes, I know we added it as a workaround for windows, but I think you should add it for all platforms (what Im doing), since it makes sense because any CPU with avx will also have aes, and compiling with aes allows speedup for certain things that use it, namely libaes which can explicitly use this compiler option if present (kinda like how, for example, libffmpeg has specific code to use avx intrinsics only if it detects that it is compiling for it). I imagine this will speed up parts of chromium that use libaes, for example encrypt/decrypt of AES encrypted certificates.