[clang][offload] Different optimization level for compiling host code and offload target code

[littlewu2508]

Currently when I tried to compile a hip program, clang accepts one optimization level -O<N> and use it in both compiling CPU code and GPU code. Is there any reason to use the same optimization level, or it can be set differently?

X problem: Linux distro commonly used -O2 flags (for CPU, apparently), while for GPGPU programs -O3 is more widely adopted. See gentoo/gentoo#33400 (comment)

[llvmbot]

@llvm/issue-subscribers-clang-driver

Author: None (littlewu2508)

Currently when I tried to compile a hip program, `clang` accepts one optimization level `-O<N>` and use it in both compiling CPU code and GPU code. Is there any reason to use the same optimization level, or it can be set differently?

X problem: Linux distro commonly used -O2 flags (for CPU, apparently), while for GPGPU programs -O3 is more widely adopted. See gentoo/gentoo#33400 (comment)

[AngryLoki]

As mentioned in #70799 , it is possible with https://clang.llvm.org/docs/ClangCommandLineReference.html#cmdoption-clang-Xarch_-arg1 -family flags.

Here is an example: https://godbolt.org/z/ahGfnMqqW

But I noticed, that one can not pass any optimization flag to -Xarch_device.
What works in -Xarch_device <arg>: -O0, -O3, -ffast-math (enables relaxed math), -fno-stack-protector (but I see no difference in assembly).
What does not work as arg: -O1, -O2, -Ox, -Og...
As you can see in example link, unroll works slightly different for -O3 and -O1, and it is possible to see it if -O1 is set globally, but it is impossible to specify -Xarch_device -O1:

clang++: error: invalid Xarch argument: '-Xarch_device -O1', not all driver options can be forwared via Xarch argument

Another note: -Xarch_device -O0 -O3 evaluates as -O3 for device.

Upd: looks like https://reviews.llvm.org/D152061 fixes -Xarch_device -O1 flag.