Re: [PATCH] build-many-glibcs: Remove no_isolate from SH config

[Adhemerval Zanella <adhemerval dot zanella at linaro dot org>]

On 13/03/2017 20:44, Adhemerval Zanella wrote:
> 
> 
> On 13/03/2017 20:17, Joseph Myers wrote:
>> On Mon, 13 Mar 2017, Adhemerval Zanella wrote:
>>
>>> On 13/03/2017 13:40, Joseph Myers wrote:
>>>> On Mon, 13 Mar 2017, Adhemerval Zanella wrote:
>>>>
>>>>> Now with d40dbe7 SH build does not require more the no_isolate gcc
>>>>> options to correct build glibc (since SH build now does not generate
>>>>> a trap anymore).  This patch removes the unrequired options from
>>>>> SH config.
>>>>
>>>> no_isolate is *only* used for SH, so should be removed completely if 
>>>> removed from use for SH.
>>>>
>>>
>>> Right, I thought about letting the options for possible other future
>>> architectures that might face the same issue.  I will remove it.
>>
>> It seems this broke the build for SH with GCC mainline (same 
>> multiple-definitions errors that are symptoms of needing these options); 
>> please revert.
>>
>> https://sourceware.org/ml/libc-testresults/2017-q1/msg00248.html
>>
> 
> Right, I will revert and check what is not failing with default
> version.
> 

Joseph, I tracked down the issue and it is due the snippet:

sysdeps/wordsize-32/divdi3.c:

133       else
134         {
135           /* qq = NN / 0d */
136 
137           if (d0 == 0)
138             d0 = 1 / d0;        /* Divide intentionally by zero.  */

GCC 6.3 and older lowers it to a software division call (__sdivsi3_i4i)
while GCC 7.0 with -fisolate-erroneous-paths-dereference found the
undefined behaviour and transform to a trap and subsequent abort call.

So I think we have some options:

  1. Revert the patch and make SH toolchain compile with
     -fno-isolate-erroneous-paths-dereference (there is no need for
     add -fno-isolate-erroneous-paths-attribute).

  2. Build divdi3 for SH explicit with -fno-isolate-erroneous-paths-attribute.

  3. Port libgcc r205444 with add an __udivmoddi4 implementation for
     architectures that do not have division instruction (which does not
     generate a trap for division by 0).

I would prefer either 2 or 3.