Re: [PATCH] perf/x86/intel: Mark expected switch fall-throughs

[Joe Perches]

On Tue, 2019-06-25 at 15:57 -0700, Nick Desaulniers wrote:
> consider most other GNU C extensions.  How do I
> test whether they exist in my compiler or not?  Is it everything or
> nothing (do they all have to exist?).

Until such time as the linux source code supports alternate
mechanisms for existing gcc extension uses, I think yes.

> In those cases you either end
> up shelling out to something like autoconf (which is what I consider
> the current infra around CONFIG_CC_HAS_ASM_GOTO), or code filled with
> tons of version checks for specific compilers which are brittle.

Or just one...

> Of the two cases, now consider what happens when my compiler that
> previously did not support a particular feature now does.  In the
> first case, the guards were compiler agnostic, and I *don't have to
> change the source* to make use of the feature in the new compiler.  In
> the second case, I *need to modify the source* to update the version
> checks to be correct.
[]
> Back to your point about adding a minimal version of Clang to the
> kernel; I don't really want to do this.  For non-x86 architectures,
> people are happily compiling their kernels with versions of clang as
> old as clang-4.

Perhaps:

#if defined(CONFIG_X86_32) || defined(CONFIG_X86_64)
#define CLANG_MINIMUM_VERSION 90000
#elif defined(CONFIG_ARM) || defined(CONFIG_ARM64)
#define CLANG_MINIMUM_VERSION 40000
#else
etc...
#endif

#if CLANG_VERSION < CLANG_MINIMUM_VERSION
etc...
#endif

> and if it continues to work for them; I'm happy.  And
> if it doesn't, and they raise an alarm, we're happy to take a look.
> Old LTS distros may have ancient builds of clang, so maybe some kind
> of hint would be nice, but I'd also like to support older versions
> where we can and I think choosing clang-9 for x86's sake is too
> x86-centric.  A version check on CONFIG_JUMP_LABEL is maybe more
> appropriate, so it cannot be selected if you're using clang && your
> version of clang is not clang-9 or greater?

The now non-portable nature of .config files might be
improved.