[J-core] Linux SH build issues with 4.14

[Rich Felker]

On Sat, Dec 02, 2017 at 06:48:59PM -0500, Rich Felker wrote:
> On Sun, Dec 03, 2017 at 12:25:10AM +0100, John Paul Adrian Glaubitz wrote:
> > On 12/02/2017 09:49 PM, John Paul Adrian Glaubitz wrote:
> > > On 12/02/2017 09:23 PM, John Paul Adrian Glaubitz wrote:
> > >> Compiler is gcc-7.
> > >>
> > >> I'll try a cross-compile with gcc-7 myself.
> > > 
> > > I can reproduce it with gcc-7 as a cross-compiler and the attached
> > > configuration which is derived from Debian's kernel configuration:
> > 
> > Switching to gcc-6 seems to avoid this issue and the kernel builds fine.
> 
> OK, I don't have gcc 7. But since this works with gcc 6, it's almost
> surely a gcc bug. I can try to get a gcc 7 build for sh4 and test
> myself, but if you could send me a dump of gcc -S (or objdump -dr for
> the produced .o file, not quite as good but should be easier to make
> and should suffice) I can probably tell you right away what the
> reference to abort seems to have been generated from.

This is reproducible with a trivial test case:

int main(int argc, char **argv) { return argc/0; }

or what gcc7 optimizes it to:

int main(int argc, char **argv) { __builtin_trap(); }

The only change seems to be this new optimization; AFAICT gcc's sh
backend has always lacked a "trap" implementation, causing gcc to use
the (awful) default call to abort() to implement it.

I think this should be filed as a bug with gcc. The right fix is
probably adding an implementation for trap using a
permanently-undefined opcode. IIRC that's fffe or feff or similar but
I'd have to check the official ISA manual again to make sure.

Adding gcc target maintainers on CC.

Rich