Re: [PATCH v4 02/35] compiler-context-analysis: Add infrastructure for Context Analysis with Clang

[Linus Torvalds <torvalds@linux-foundation.org>]

On Thu, 20 Nov 2025 at 07:13, Marco Elver <elver@google.com> wrote:
>
> --- a/include/linux/compiler-context-analysis.h
> +++ b/include/linux/compiler-context-analysis.h
> @@ -6,27 +6,465 @@
>  #ifndef _LINUX_COMPILER_CONTEXT_ANALYSIS_H
>  #define _LINUX_COMPILER_CONTEXT_ANALYSIS_H
>
> +#if defined(WARN_CONTEXT_ANALYSIS)

Note the 400+ added lines to this header...

And then note how the header gets used:

> +++ b/scripts/Makefile.context-analysis
> @@ -0,0 +1,7 @@
> +# SPDX-License-Identifier: GPL-2.0
> +
> +context-analysis-cflags := -DWARN_CONTEXT_ANALYSIS             \
> +       -fexperimental-late-parse-attributes -Wthread-safety    \
> +       -Wthread-safety-pointer -Wthread-safety-beta
> +
> +export CFLAGS_CONTEXT_ANALYSIS := $(context-analysis-cflags)

Please let's *not* do it this way, where the header contents basically
get enabled or not based on a compiler flag, but then everybody
includes this 400+ line file whether they need it or not.

Can we please just make the header file *itself* not have any
conditionals, and what happens is that the header file is included (or
not) using a pattern something like

   -include $(srctree)/include/linux/$(context-analysis-header)

instead.

IOW, we'd have three different header files entirely: the "no context
analysis", the "sparse" and the "clang context analysis" header, and
instead of having a "-DWARN_CONTEXT_ANALYSIS" define, we'd just
include the appropriate header automatically.

We already use that "-include" pattern for <linux/kconfig.h> and
<linux/compiler-version.h>. It's probably what we should have done for
<linux/compiler.h> and friends too.

The reason I react to things like this is that I've actually seen just
the parsing of header files being a surprisingly big cost in build
times. People think that optimizations are expensive, and yes, some of
them really are, but when a lot of the code we parse is never actually
*used*, but just hangs out in header files that gets included by
everybody, the parsing overhead tends to be noticeable. There's a
reason why most C compilers end up integrating the C pre-processor: it
avoids parsing and tokenizing things multiple times.

The other reason is that I often use "git grep" for looking up
definitions of things, and when there are multiple definitions of the
same thing, I actually find it much more informative when they are in
two different files than when I see two different definitions (or
declarations) in the same file and then I have to go look at what the
#ifdef condition is. In contrast, when it's something where there are
per-architecture definitions, you *see* that, because the grep results
come from different header files.

I dunno. This is not a huge deal, but I do think that it would seem to
be much simpler and more straightforward to treat this as a kind of "N
different baseline header files" than as "include this one header file
in everything, and then we'll have #ifdef's for the configuration".

Particularly when that config is not even a global config, but a per-file one.

Hmm? Maybe there's some reason why this suggestion is very
inconvenient, but please at least consider it.

              Linus