Re: [PATCH v4 0/4] tools/nolibc: Adding stdint.h, more integer types and tests

[Willy Tarreau <w@1wt.eu>]

Hi Vincent,

On Wed, Feb 15, 2023 at 07:11:09PM -0500, Vincent Dagonneau wrote:
> > So after investigation, __WORDSIZE is not defined, hence the failures!
> > It proves the importance of the tests you've added ;-)
> >
> > However we have the size of these types defined by the compiler itself
> > at least since gcc-4.4 and clang-3.8 which are the oldest I could test:
> >
> >   $ mips-gcc44_glibc214-linux-gnu-gcc -xc -dM -E - </dev/null |grep SIZE
> >   #define __SIZEOF_POINTER__ 4
> >   #define __SIZEOF_LONG__ 4
> >   #define __SIZEOF_LONG_DOUBLE__ 8
> >   #define __SIZEOF_SIZE_T__ 4
> >   #define __SIZEOF_WINT_T__ 4
> >   #define __SIZE_TYPE__ unsigned int
> >   #define __SIZEOF_PTRDIFF_T__ 4
> >   #define __SIZEOF_INT__ 4
> >   #define __SIZEOF_FLOAT__ 4
> >   #define __SIZEOF_SHORT__ 2
> >   #define __SIZEOF_WCHAR_T__ 4
> >   #define __SIZEOF_DOUBLE__ 8
> >   #define __SIZEOF_LONG_LONG__ 8
> >
> > In addition both provide __SIZE_TYPE__ which is defined either as
> > unsigned int or long unsigned int, so that can simplify quite some
> > parts (and other types are defined for other types in more recent
> > versions). Both also define __LONG_MAX__ that you could possibly
> > use to conveniently create INTPTR_MAX, INTPTR_MIN, UINTPTR_MAX and
> > so on.
> >
> 
> Mmmh, interesting, I hadn't thought about verifying what defined the
> __WORDSIZE. I assumed wrongly that it was set by standard but it seems not. 

No problem, that's exactly why I wanted to retry every combination.

> I replicated your example on my machine to see the intersection of what is
> defined by both GCC and clang. Do you know if we would need to check any
> other compilers?

I don't think it's particularly needed to go further for now. For example
I know that tcc doesn't support some of the asm constraints that we use
in register alllocation. Supporting the most commonly encountered
compilers is sufficient for our use case.

> > And finally we should set the __WORDSIZE ourselves as 8*__SIZEOF_LONG__
> > and that would do the job.
> >
> > I tested the following patch which passes all the tests successfully
> > on all supported archs. Let me know if you agree with such a change
> > and how you want us to proceed. It would require a small change in the
> > commit message though, to explain that our pointers are the size of a
> > long on supported platforms.
> >
> 
> I can integrate the changes and write an explanation as a commit message and
> submit a new version tomorrow, would that work for you?

Sure that would be great! I'll look at it this week-end anyway.

Thank you!
Willy