[clang][PPC] support for -mprofile-kernel

[nickdesaulniers]

From an LKML thread, it seems like the Linux kernel makes use of -mprofile-kernel.
https://lore.kernel.org/llvm/[email protected]/
cc @nemanjai @nemanja-ibm

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

[nickdesaulniers]

This script is what's used to feature test for this flag.

[nemanjai]

So this is specific to PowerPC? I suppose only the kernel build on PowerPC uses it so that is why this is specific to PowerPC but Clang doesn't support it on any architecture.
Do you happen to have the documentation from GCC for what this option does? It doesn't seem to come up in my searches in the GCC 11.2.0 manual nor in my searches on Google.
If this is an undocumented GCC option, then we can't really implement it in Clang and we need to find a documented alternative for the kernel build.

[nickdesaulniers]

In the Linux kernel sources, I only see references to -mprofile-kernel in the PPC port.

If this is an undocumented GCC option
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65372

From that bug, the command line help text prints:

Call mcount for profiling before a function prologue

https://kamaleshbabulal.wordpress.com/2016/02/22/linux-kernel-livepatch-mprofile-kernel/ has some more info.

I'm guessing it's easy to find what it's doing in godbolt. Seems like it needs -pg do call _mcount, and is only supported on 64b targets. https://godbolt.org/z/37ooxnh3P

[rnav]

It is unfortunate that -mprofile-kernel is still not documented, but we rely on it heavily for a lot of the ftrace features, including kernel live patching. I doubt the kernel gets much coverage for the older mcount ABI for ppc64le.

From the kernel perspective, -mprofile-kernel largely boils down to the two instruction sequence mflr r0; bl _mcount emitted at function entry, after the optional two-instruction sequence for TOC setup.

Not to replace documentation, but another good reference for this is the writeup from Michael Ellerman:
https://mpe.github.io/posts/2016/05/23/kernel-live-patching-for-ppc64le/

[MaskRay]

I am unsure whether we need to add more variants for mcount. GCC's PowerPC port has some support for -fpatchable-function-entry= now: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99888 This is the new way for live patching IIUC.

[chenzheng1030]

#92997 is created for this issue. I don't have a ppc/ppc64 server to check the end-to-end functionality of the patchable function entry. For now, the code generation is like other platforms, like X86/RISCV on Linux. If any issue found in the compiler code generation, please let me know.