[PowerPC] call to _mcount emitted by -pg needs to be emitted BEFORE function prolog

[nickdesaulniers]

From an LKML thread: https://lore.kernel.org/llvm/20230609034501.407971-1-naveen@kernel.org/

Ftrace on ppc32 expects a three instruction sequence at the beginning of
each function when specifying -pg:
	mflr	r0
	stw	r0,4(r1)
	bl	_mcount

This is the case with all supported versions of gcc. Clang however emits
a branch to _mcount after the function prologue

Example: https://godbolt.org/z/8r83PKorK

As a result, support for function tracing is being disabled in the linux kernel for ppc32 when built with clang.

cc @nemanjai @nemanja-ibm

[llvmbot]

@llvm/issue-subscribers-backend-powerpc

[chenzheng1030]

@nickdesaulniers Hi thanks for reporting this issue. Does this issue still matter? If so, maybe I can make a fix.

[chenzheng1030]

With latest clang, bl _mcount is still after the whole prologue.

[nickdesaulniers]

Probably. cc @nathanchance who can follow up.

[chenzheng1030]

On PPC, there is a codegen improvement that moves stw away from its feeder mflr because mflr might have long latency. This will break the expected instructions.

Must these three instructions be in the sequence in the description? What's the relationship between this task and the -fpatchable-function-entry support on ppc32? In https://lore.kernel.org/llvm/20230609034501.407971-1-naveen@kernel.org/, it says "This can be re-enabled
later if clang picks up support for -fpatchable-function-entry on ppc32.", so we can also fix this by supporting -fpatchable-function-entry on ppc32?

@nickdesaulniers @nathanchance

[rnav]

Yes, adding support for -fpatchable-function-entry would be the right way to address this. Support for that was introduced in the kernel with the below commit:
https://lore.kernel.org/linuxppc-dev/68586d22981a2c3bb45f27a2b621173d10a7d092.1687166935.git.naveen@kernel.org/

Support for that is being discussed in #57031

[chenzheng1030]

#92997 is created for this issue.