Re: [PATCH] frv: fix build failure

Re: [PATCH] frv: fix build failure

[Vineet Gupta]

+CC linux-arch, Arnd

On 11/23/2017 09:17 AM, Alexey Brodkin wrote:
> Hi Sudip,
> 
> On Tue, 2017-11-21 at 22:10 +0000, Sudip Mukherjee wrote:
>> The frv defconfig build is failing with the error:
>> lib/mpi/mpih-div.o: In function `mpihelp_divrem':
>> mpih-div.c:(.text+0x30c): undefined reference to `abort'
>>
>> The function 'abort' was never defined for the frv architecture.
>> Create 'abort' as is done in other arch like 'arm' and 'unicore32'.
>>
>> Signed-off-by: Sudip Mukherjee <sudipm.mukherjee@gmail.com>
>> ---
> 
> I'm seeing the same issue building for ARC from today's Linus' tree.

With which compiler ? Claudiu implemented __builtin_trap recently so ARC builds 
should not get abort() but a trap 5 instruction instead.

> Maybe it worth implementing abort() as a weak function for every arch/platform
> that doesn't have it explicitly defined? Otherwise we'll end-up with
> useless code duplication.

I understand the case/need for adding a weak/common abort() as a quick fix for 
handling such cases, but perhaps we should not and instead fix the rootcause. In 
this specific case, Claudiu mentioned that gcc was generating abort due to 
something like this (flagging a possible divide by zero due to 
-fno-isolate-erroneous-paths-dereference.

a;

fn1() {
   switch (a)
   case 0:
   return 1 / a;
}

Sudeep can you confirm that removing this hacking FRV build to not have this 
toggle fixes the abort issue.

So the offending code needs to be fixed as in the end when that code path is taken 
system is hosed.

-Vineet

Re: [PATCH] frv: fix build failure

[Sudip Mukherjee]

On Mon, Nov 27, 2017 at 10:25:16AM -0800, Vineet Gupta wrote:
> +CC linux-arch, Arnd
> 
> On 11/23/2017 09:17 AM, Alexey Brodkin wrote:
> >Hi Sudip,
> >
> >On Tue, 2017-11-21 at 22:10 +0000, Sudip Mukherjee wrote:

<snip>

> I understand the case/need for adding a weak/common abort() as a
> quick fix for handling such cases, but perhaps we should not and
> instead fix the rootcause. In this specific case, Claudiu mentioned
> that gcc was generating abort due to something like this (flagging a
> possible divide by zero due to
> -fno-isolate-erroneous-paths-dereference.
> 
> a;
> 
> fn1() {
>   switch (a)
>   case 0:
>   return 1 / a;
> }
> 
> Sudeep can you confirm that removing this hacking FRV build to not
> have this toggle fixes the abort issue.
> 
> So the offending code needs to be fixed as in the end when that code
> path is taken system is hosed.

The above case that you mentioned is indeed the case here.
mpihelp_divrem() in lib/mpi/mpih-div.c has this same divide by zero case.

So, what should be the actual fix here?

Re: [PATCH] frv: fix build failure

[Sudip Mukherjee]

On Mon, Nov 27, 2017 at 10:25:16AM -0800, Vineet Gupta wrote:
> +CC linux-arch, Arnd
> 
> On 11/23/2017 09:17 AM, Alexey Brodkin wrote:
> >Hi Sudip,
> >
> >On Tue, 2017-11-21 at 22:10 +0000, Sudip Mukherjee wrote:

<snip>

> I understand the case/need for adding a weak/common abort() as a
> quick fix for handling such cases, but perhaps we should not and
> instead fix the rootcause. In this specific case, Claudiu mentioned
> that gcc was generating abort due to something like this (flagging a
> possible divide by zero due to
> -fno-isolate-erroneous-paths-dereference.
> 
> a;
> 
> fn1() {
>   switch (a)
>   case 0:
>   return 1 / a;
> }
> 
> Sudeep can you confirm that removing this hacking FRV build to not
> have this toggle fixes the abort issue.
> 
> So the offending code needs to be fixed as in the end when that code
> path is taken system is hosed.

The above case that you mentioned is indeed the case here.
mpihelp_divrem() in lib/mpi/mpih-div.c has this same divide by zero case.

So, what should be the actual fix here?

--
Regards
Sudip

Re: [PATCH] frv: fix build failure

[Vineet Gupta]

On 12/03/2017 02:32 PM, Sudip Mukherjee wrote:
> On Mon, Nov 27, 2017 at 10:25:16AM -0800, Vineet Gupta wrote:
>> +CC linux-arch, Arnd
>>
>> On 11/23/2017 09:17 AM, Alexey Brodkin wrote:
>>> Hi Sudip,
>>>
>>> On Tue, 2017-11-21 at 22:10 +0000, Sudip Mukherjee wrote:
> 
> <snip>
> 
>> I understand the case/need for adding a weak/common abort() as a
>> quick fix for handling such cases, but perhaps we should not and
>> instead fix the rootcause. In this specific case, Claudiu mentioned
>> that gcc was generating abort due to something like this (flagging a
>> possible divide by zero due to
>> -fno-isolate-erroneous-paths-dereference.
>>
>> a;
>>
>> fn1() {
>>    switch (a)
>>    case 0:
>>    return 1 / a;
>> }
>>
>> Sudeep can you confirm that removing this hacking FRV build to not
>> have this toggle fixes the abort issue.
>>
>> So the offending code needs to be fixed as in the end when that code
>> path is taken system is hosed.
> 
> The above case that you mentioned is indeed the case here.
> mpihelp_divrem() in lib/mpi/mpih-div.c has this same divide by zero case.
> 
> So, what should be the actual fix here?

To remove the divide by zero.

Or atleast have a __weak abort() defined in common code so all arches benefit from 
it !

-Vineet