Re: [Gm2] A burst of internal compiler issues from gm2-6.4.0, when linking a stand-alone application

[John Dubery]

On 22/11/17 17:00, address@hidden wrote:
> ------------------------------
> Message: 2
> Date: Wed, 22 Nov 2017 10:19:07 +0000
> From: Gaius Mulley <address@hidden>
> To: John Dubery <address@hidden>
> Cc: "address@hidden" <address@hidden>
> Subject: Re: [Gm2] A burst of internal compiler issues from gm2-6.4.0
>       when linking a stand-alone application
> Message-ID: <address@hidden>
> Content-Type: text/plain
>
> John Dubery <address@hidden> writes:
>
>
> Hi John,
>
>>      gm2 -static -nostdlib -nostartfiles -fno-exceptions -fno-pth -
>>      fm2-whole-program -flibs=min -fonlylink -TSimple.ld ScreenTest.o
>>      ScreenTest-start.o -o ScreenTest.elf
> gm2 needs the source file of the top program module as well on the
> command line.  It uses this to generate the scaffold of module calls for
> all dependant modules (the order can be overwritten by you if
> necessary).
>
> It might be worth trying:
>
> $ gm2 -static -nostdlib -nostartfiles -fno-exceptions -fno-pth -
>    fm2-whole-program -flibs=min -fonlylink -TSimple.ld ScreenTest.o
>    ScreenTest-start.o -o ScreenTest.elf -v  ScreenTest.mod
>
> regards,
> Gaius
>
Gaius,
Thanks.
Adding the source file didn't make any difference to the list of 
errors.  I've looked at the issues more and it seems to me that in both 
cases the compiler is not putting in code that it has determined the 
program needs.  It's relying on the libraries to have the missing pieces.

I link with -nostdlib so there are no libraries to be linked in. When I 
remove this option (and the simple load file) there is no error 
reported.  That sounds great but, unfortunately, the binary size jumps 
from about 5Kb to 500Kb!

1 MOD & DIV
I'm fairly sure that the missing function - __aeabi_ldivmod - is needed 
to get the correct answer for the complicated +ve/-ve cases. On this 
occasion I don't need that extra evaluation but even when I change to 
all CARDINALs (I've even tried CARDINAL(10) as a divisor) an extra 
function is still needed: __uidiv.

I've solved the problem on this occasion by using repeated subtraction.  
However, I guess I feel that the compiler should fetch this code as soon 
as it is seen to be needed, rather than rely on a library to be linked.  
Alternatively, could there be a very small library that provides just 
these 2 functions?

2. unwind-arm-common.inc - can't find abort, etc.
These cases seem to be caused by the inclusion of code for unwinding the 
stack but not the functions that this extra code calls - __exidx_end, 
__exidx_start, memcpy, abort.  I don't need this stack unwinding 
facility at the moment, as far as I know.  One of my compiler options is 
"-fno-exceptions" and I have also tried "-fno-unwind-tables" but that 
made no difference.  What can I do to prevent the inclusion of any 
unwinding code?

Is it reasonable to say that code should be generated only for language 
features used, e.g. MOD & DIV, or listed in IMPORT statements?

Regards,
John