Bug#869373: Bug#869122: perl: 5.26 FTBFS on hppa: broken miniperl?

[John David Anglin]

On 2017-07-24, at 4:36 PM, Niko Tyni wrote:

> Do you know why the GCC optimization issue only seems to affect hppa+sh4?

No.

This hunk of code is wrong:

.loc 1 14875 0 is_stmt 0 discriminator 2
	bl Perl_sv_2iv_flags,%r2
	copy %r4,%r26
.LVL1245:
	copy %r29,%r3

The above line is wrong.  r29 is potentially clobbered by the call to Perl_sv_2iv_flags.
It's a call clobbered register.  It sets %r3 which is responsible for the garbage return value.

I think this probably occurs as a result of the merging of the switch statements but I'm
not 100% certain.

> 
> I wonder if lowering the optimization of the affected
> code would be an easier workaround. On that note, looking
> at the build logs makes me think the cflags fiddling in
> debian/patches/debian/hppa_op_optimize_workaround.diff is still working,
> but the problem here is that op.c is linked to opmini.c for the miniperl
> build and that still gets compiled at -O2.
> 
> Does it work for you without the Perl_custom_op_get_field() code
> reorganization if you patch cflags.SH further like this?
> 
> -    op) : work around http://bugs.debian.org/838613
> +    op|opmini) : work around http://bugs.debian.org/838613
> 
> This should make it build both op.c and opmini.c at -O0.

Might work.

> 
> Or does this hit the hppa "-O0 -fPIC" problem again?

Again, I don't know.  I had thought the R_PARISC_DPREL21L relocation problem was
an optimization issue.  I have installed perl 5.26+b1 on a couple of systems.  Building
libjavascript-minifier-xs-perl with it doesn't trigger the issue nor does building it with
DEB_BUILD_OPTIONS=noop.  Something is wrong with Adrian's build that doesn't
occur with my patched build.  buildd still has Adrian's version.

Dave
--
John David Anglin	dave.anglin at bell.net