Re: [PATCH 3/3] isystem: delete global -isystem compile option

[Alexey Dobriyan <adobriyan@gmail.com>]

On Mon, Aug 02, 2021 at 11:47:47AM -0500, Segher Boessenkool wrote:
> On Mon, Aug 02, 2021 at 09:42:45AM +0300, Alexey Dobriyan wrote:
> > On Sun, Aug 01, 2021 at 04:32:47PM -0500, Segher Boessenkool wrote:
> > > On Sun, Aug 01, 2021 at 11:13:36PM +0300, Alexey Dobriyan wrote:
> > > > In theory, it enables "leakage" of userspace headers into kernel which
> > > > may present licensing problem.
> > > 
> > > > -NOSTDINC_FLAGS += -nostdinc -isystem $(shell $(CC) -print-file-name=include)
> > > > +NOSTDINC_FLAGS += -nostdinc
> > > 
> > > This is removing the compiler's own include files.  These are required
> > > for all kinds of basic features, and required to be compliant to the C
> > > standard at all.
> > 
> > No they are not required.
> 
> This is false, they *are* required, whenever you want to use these
> features.  If you do not include the required headers you get undefined
> behaviour.
> 
> > Kernel uses its own bool, uintptr_t and
> > static_assert, memset(), CHAR_BIT.
> 
> Yes, and it occasionally gets it wrong.  Great fun.  See c46bbf5d2def
> for the latest episode in this saga.  (Yes I know this is uapi so maybe
> not the best example here, but it isn't like the kernel gets such things
> wrong so often these days ;-) )
> 
> The kernel *cannot* make up its own types for this.  It has to use the
> types it is required to use (by C, by the ABIs, etc.)  So why
> reimplement this?

Yes, it can. gcc headers have stuff like this:

	#define __PTRDIFF_TYPE__ long int
	#define __SIZE_TYPE__ long unsigned int

If gcc can defined standard types, kernel can too.

> > noreturn, alignas newest C standard
> > are next.
> 
> What is wrong with <stdalign.h> and <stdnoreturn.h>?

These two are actually quite nice.

Have you seen <stddef.h>? Loads of macrology crap.
Kernel can ship nicer one.

> > This version changelog didn't mention but kernel would use
> > -ffreestanding too if not other problems with the flag.
> 
> It is still true for freestanding C implementations, you just get a
> severely reduced standard library,
> 
> > > These are not "userspace headers", that is what
> > > -nostdinc takes care of already.
> > 
> > They are userspace headers in the sense they are external to the project
> > just like userspace programs are external to the kernel.
> 
> So you are going to rewrite all of the rest of GCC inside the kernel
> project as well?

What an argument. "the rest of GCC" is already there except for stdarg.h.

> > > In the case of GCC all these headers are GPL-with-runtime-exception, so
> > > claiming this can cause licensing problems is fearmongering.
> > 
> > I agree licensing problem doesn't really exist.
> > It would take gcc drop-in replacement with authors insane enough to not
> > license standard headers properly.
> 
> There does still not exist a drop-in replacement for GCC, not if you
> look closely and/or rely on details (like the kernel does).  Some of the
> differences are hidden by "linux/compiler-*.h", but hardly all.
> 
> > > I strongly advise against doing this.
> > 
> > Kernel chose to be self-contained.
> 
> That is largely historical, imo.  Nowadays this is less necessary.

I kind of agree as in kernel should use int8_t and stuff because they
are standard.

Also, -isystem removal disables <float.h> and <stdatomic.h> which is
desireable.

> Also, the kernel chose to *do* use the compiler include files.  It is
> you who wants to abolish that here.
> 
> > -isystem removal makes sense then.
> 
> -nostdinc -isystem $(shell $(CC) -print-file-name=include)  makes sense
> for that: you do indeed not want the userspace headers.  Maiming the
> compiler (by removing some of its functional parts, namely, its generic
> headers) does not make sense.
> 
> > It will be used for intrinsics where necessary.
> 
> Like, everywhere.

No, where necessary. Patch demostrates there are only a few places which
want -isystem back.