Re: [bug-mes] mescc arm progress

[Danny Milosavljevic]

Hi Janneke,

On Sun, 07 Apr 2019 23:59:37 +0200
Jan Nieuwenhuizen <address@hidden> wrote:

>I found that something
> weird happens an EOF on arm, not exactly sure what's going on there but
> I found a workaround for now.

Hmm, EOF is famously (int) -1 and char is unsigned on ARM (if not changed by 
compiler flag).  Might have something to do with it.

Also not sure whether we are officially allowed to unread EOF--although the 
ungetter is our implementation, right?  That one seems to allow it.

> So, I added two ugly commits on top that now enable mescc to compile run
> on mes, with pretty good results:
> 
> --8<---------------cut here---------------start------------->8---
> $ make check recheck=true
> GNUmakefile:165: skipping man: no help2man
> GNUmakefile:285: update-hash: no guix
> /bin/bash check.sh
> FAIL: lib/tests/scaffold/t.c
> FAIL: lib/tests/scaffold/06-not-call-1.c
> FAIL: lib/tests/scaffold/06-call-variable.c
> FAIL: lib/tests/scaffold/08-assign-global.c
> FAIL: lib/tests/scaffold/12-if-eq.c
> FAIL: lib/tests/scaffold/13-if-neq.c
> FAIL: lib/tests/scaffold/14-if-goto.c
> FAIL: lib/tests/scaffold/17-compare-char.c
> FAIL: lib/tests/scaffold/17-compare-ge.c
> FAIL: lib/tests/scaffold/17-compare-gt.c
> FAIL: lib/tests/scaffold/17-compare-lt.c
> ...
> --8<---------------cut here---------------end--------------->8---

Nice short failure list!  (some of them ridiculous ones that I apparently broke 
again--oops).

I've taken a look at your commits and saw:

>+#if 0 ///Error: thumb conditional instruction should be in IT block -- `moveq 
>r1,#1'

Hmm, that means that thumb mode is enabled which means it will use a different 
frame pointer (r7).  That's very bad because that frame pointer is not set up 
by our runtime.  (it might be also why you had to comment out r7 from the 
syscall clobber list.  Though I think the clobber list is much too paranoid 
anyway--but better than debugging forever :) )

Do you pass "-marm" to eventual gcc invocations (if any) and to the assembler?  
That should disable Thumb mode.

As for syscalls, for some reason, the following construct doesn't do what I 
think it should do in gcc:

unsigned int foo(a,b,c)
  unsigned int a asm("r7");
  unsigned int b asm("r0");
  unsigned int c asm("r1");
{
        asm("swi $0");
}

int main() {
        foo(1,2,3);
}

That's too bad.

pgp7YwyNluGZf.pgp
Description: OpenPGP digital signature