Re: mes wip-arm state

[Jan Nieuwenhuizen]

Danny Milosavljevic writes:

Hello Danny,

> On Sun, 31 May 2020 23:08:13 +0200
> Jan Nieuwenhuizen <janneke@gnu.org> wrote:
>
>> > FAIL: lib/tests/scaffold/7l-struct-any-size-array-simple.c
>> > FAIL: lib/tests/scaffold/7r-sign-extend.c
>> 
>> So this is a test-suite bug, right?  Are the tests broken or should they
>> be marked XFAIL (for ARM?)
>
> I think that these tests are broken--or at least not applicable on ARM.
>
> Both tests basically do
>
>   {
>     char c = -1;
>     int i = c;
>   }
>
> and char is unsigned on ARM, so i will NEVER contain a negative value (reason:
> because of C integer promotion rules).

Ouch, yes I see.

> If those are from tcc, let's check whether they force signed char by compiler
> option.

I checked and I don't believe they do.  tcc does have the -fsigned-char,
-funsigned-char options though.  So, some client code (in the Guix
bootstrap) just might use that.  Unlikely, right?  We'll see.

In arm-gen.c I found

    #define CHAR_IS_UNSIGNED

but they also use

    typedef signed char int8_t;
    ...
    int8_t  reg; /* register, -1 if none */
    ...
    signed char avail[3]; /* 3 holes max with only float and double alignments 
*/

so signed chars should work.  How about just adding 'signed' to those
tests above?  We're really testing if sign extension works -- testing
for signdness of char could/should be another test.  WDYT?

>> > # FAIL:  0
>> 
>> Congratulations, that's just terriffic \o/
>
> Yeah.  I like it!
>
> What's the next step?  Should we merge this part to master?
>
> After that, make tcc work?

Yes.  And then: the Reduced binary bootstrap for Guix!

As we cannot rewrite master, I'd like to check if we didn't break
anything for mes-x86 :-)

We can also make tcc work before merging, but we could also start a new
wip-tcc / wip-arm-tcc later if the need arises.  Whatever works for you!

As you propbably have seen, in the bootstrap we use a forked tcc:
somewhat older (0.9.27), patched.

    https://gitlab.com/janneke/tinycc/-/tree/mes-0.21

I'm hoping to get rid of that fork, some time.  That needs work...
Functionality-wise, the fork supports a gradual introduction of
bitfield, float/double and long long, by using #define switches.

It also has this

--8<---------------cut here---------------start------------->8---
17df781 * Revert "lotsa debug printing."
ba1c47d * lotsa debug printing.
--8<---------------cut here---------------end--------------->8---

debugging aid, which was created ad-hoc but is a pretty good start to
help bisecting a piece of wrongly compiled code; by ascii-diffing
gcc-compiled and mescc-compiled runs of tcc.

Greetings,
Janneke

-- 
Jan Nieuwenhuizen <janneke@gnu.org> | GNU LilyPond http://lilypond.org
Freelance IT http://JoyofSource.com | Avatar® http://AvatarAcademy.com