Failing 'make check' for non-Cygwin GMP-ECM package

Failing 'make check' for non-Cygwin GMP-ECM package

[Mark Geisert]

I'm seeing an odd issue after building the GMP-ECM 6.4.4 package from 
Inria.  I downloaded the .tar.gz file and unpacked, ran configure, then 
make and 'make check'.  I get a SIGSEGV from a test using ecm.exe as part 
of 'make check'.

Debugging with gdb, I see something odd and would appreciate any thoughts 
on where I should take this issue.  The faulting source line is 
mpmod.c:343 which reads:
   REDC2(rp, cp, np, nn, invm);
This macro expands to a call to __gmpn_redc_2().  That function is at an 
address above the 4GB line (this is on Win7 64-bit) as seen here:

(gdb) i func ^__gmpn_redc_2$
All functions matching regular expression "^__gmpn_redc_2$":

Non-debugging symbols:
0x00000003fcbee9e8  __gmpn_redc_2

But the instructions generated for that source line look like this:
(gdb) x/7i $rip
=> 0x1004166c0 <__ecm_mpres_get_z+240>: mov    0x10(%rdi),%rax
    0x1004166c4 <__ecm_mpres_get_z+244>: mov    %rbp,%r9
    0x1004166c7 <__ecm_mpres_get_z+247>: mov    %r13,%r8
    0x1004166ca <__ecm_mpres_get_z+250>: mov    %r12,%rcx
    0x1004166cd <__ecm_mpres_get_z+253>: mov    %rax,0x20(%rsp)
    0x1004166d2 <__ecm_mpres_get_z+258>: callq  0xfcbee9e8
    0x1004166d7 <__ecm_mpres_get_z+263>: test   %rax,%rax

Notice how the callq instruction has a truncated address.  I thought maybe 
there's an issue with displaying that instruction (unlikely I know) so I 
stepped by instruction, but no, the truncated address is actually hit:
(gdb) si
0x00000001004166c4      343       REDC2(rp, cp, np, nn, invm);
(gdb)
0x00000001004166c7      343       REDC2(rp, cp, np, nn, invm);
(gdb)
0x00000001004166ca      343       REDC2(rp, cp, np, nn, invm);
(gdb)
0x00000001004166cd      343       REDC2(rp, cp, np, nn, invm);
(gdb)
0x00000001004166d2      343       REDC2(rp, cp, np, nn, invm);
(gdb)
0x00000000fcbee9e8 in ?? ()
(gdb)
Program received signal SIGSEGV, Segmentation fault.
0x00000000fcbee9e8 in ?? ()
(gdb)

I'm current on all Cygwin packages such as binutils.  Does the above look 
like a linker issue or linker misuse (e.g. missing option)?  Could there 
be a mismatch of object file arch between my just-built GMP-ECM and 
Cygwin's libgmp-devel?  Or something else?  If it doesn't seem like a 
Cygwin issue I can take it upstream to Inria.
Thanks much,

..mark

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: Failing 'make check' for non-Cygwin GMP-ECM package

[Achim Gratz]

Mark Geisert writes:
> I'm seeing an odd issue after building the GMP-ECM 6.4.4 package from
> Inria.  I downloaded the .tar.gz file and unpacked, ran configure,
> then make and 'make check'.  I get a SIGSEGV from a test using ecm.exe
> as part of 'make check'.

It seems that this release is three years old.  Did you try autoreconf
or just configure?  If the latter, it may simply not recognize the
current Cygwin correctly in its configury.


Regards,
Achim.
-- 
+<[Q+ Matrix-12 WAVE#46+305 Neuron microQkb Andromeda XTk Blofeld]>+

SD adaptations for Waldorf Q V3.00R3 and Q+ V3.54R2:
http://Synth.Stromeko.net/Downloads.html#WaldorfSDada

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: Failing 'make check' for non-Cygwin GMP-ECM package

[Mark Geisert]

Achim Gratz wrote:
> Mark Geisert writes:
>> I'm seeing an odd issue after building the GMP-ECM 6.4.4 package from
>> Inria.  I downloaded the .tar.gz file and unpacked, ran configure,
>> then make and 'make check'.  I get a SIGSEGV from a test using ecm.exe
>> as part of 'make check'.
>
> It seems that this release is three years old.  Did you try autoreconf
> or just configure?  If the latter, it may simply not recognize the
> current Cygwin correctly in its configury.

Thanks for the idea but trying it didn't change the symptoms.  With more 
debugging this seems to me like some kind of dynamic relocation issue so I'm 
looking further in that direction within the Cygwin DLL.
Regards,

..mark

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: Failing 'make check' for non-Cygwin GMP-ECM package

[Mark Geisert]

Mark Geisert wrote:
> Achim Gratz wrote:
>> Mark Geisert writes:
>>> I'm seeing an odd issue after building the GMP-ECM 6.4.4 package from
>>> Inria.  I downloaded the .tar.gz file and unpacked, ran configure,
>>> then make and 'make check'.  I get a SIGSEGV from a test using ecm.exe
>>> as part of 'make check'.
>>
>> It seems that this release is three years old.  Did you try autoreconf
>> or just configure?  If the latter, it may simply not recognize the
>> current Cygwin correctly in its configury.
>
> Thanks for the idea but trying it didn't change the symptoms.  With more
> debugging this seems to me like some kind of dynamic relocation issue so I'm
> looking further in that direction within the Cygwin DLL.

Replying to myself...  Yes, there appear to be two separate bugs colluding to 
break things the way I'm seeing.  First, the application's object file has 
R_X86_64_PC32 relocation markers for functions it wants to pull in from 
cyggmp-10.dll.  These relocation entries apparently allow for only 4 bytes of 
address.

Second, the relocation that actually needs doing in this testcase results in 
addresses above 4GB, so they need more than 4 bytes to be stored.  There is code 
in Cygwin's pseudo-reloc.cc to check for this situation but it is being skipped 
because __OPTIMIZE__ has been #define'd somewhere by the procedure that builds 
Cygwin.  Instead, the relocated address gets truncated before being written back 
to memory, rather than provoking the error report "Invalid relocation. [...] 
doesn't fit into 32 bits".

I don't know if the first bug could be caused by improper compiler options or 
whatnot.  Here's the gcc invocation for that one object file:

/bin/sh ./libtool  --tag=CC   --mode=compile gcc -DHAVE_CONFIG_H -I.  -I./x86_64 
   -pipe -g -O2 -march=haswell -MT libecm_la-mpmod.lo -MD -MP -MF 
deps/libecm_la-mpmod.Tpo -c -o libecm_la-mpmod.lo `test -f 'mpmod.c' || echo 
'./'`mpmod.c
libtool: compile:  gcc -DHAVE_CONFIG_H -I. -I./x86_64 -pipe -g -O2 
-march=haswell -MT libecm_la-mpmod.lo -MD -MP -MF .deps/libecm_la-mpmod.Tpo -c 
mpmod.c -o libecm_la-mpmod.o

Configure does know it's building for x86_64.  I am up to date on binutils and 
this is on Cygwin 2.4.1, Win7 64-bit.

..mark

--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple

Re: Failing 'make check' for non-Cygwin GMP-ECM package -- workaround

[Mark Geisert]

Mark Geisert wrote:
> Mark Geisert wrote:
>> Achim Gratz wrote:
>>> Mark Geisert writes:
>>>> I'm seeing an odd issue after building the GMP-ECM 6.4.4 package from
>>>> Inria.  I downloaded the .tar.gz file and unpacked, ran configure,
>>>> then make and 'make check'.  I get a SIGSEGV from a test using ecm.exe
>>>> as part of 'make check'.
>>>
>>> It seems that this release is three years old.  Did you try autoreconf
>>> or just configure?  If the latter, it may simply not recognize the
>>> current Cygwin correctly in its configury.
>>
>> Thanks for the idea but trying it didn't change the symptoms.  With more
>> debugging this seems to me like some kind of dynamic relocation issue so I'm
>> looking further in that direction within the Cygwin DLL.
>
> Replying to myself...  Yes, there appear to be two separate bugs colluding to
> break things the way I'm seeing.  First, the application's object file has
> R_X86_64_PC32 relocation markers for functions it wants to pull in from
> cyggmp-10.dll.  These relocation entries apparently allow for only 4 bytes of
> address.
>
> Second, the relocation that actually needs doing in this testcase results in
> addresses above 4GB, so they need more than 4 bytes to be stored.  There is code
> in Cygwin's pseudo-reloc.cc to check for this situation but it is being skipped
> because __OPTIMIZE__ has been #define'd somewhere by the procedure that builds
> Cygwin.  Instead, the relocated address gets truncated before being written back
> to memory, rather than provoking the error report "Invalid relocation. [...]
> doesn't fit into 32 bits".

I've worked around the first "bug" by rebasing the Cygwin gmp DLLs.  It seems 
those R_X86_64_PC32 relocation markers I had thought problematic are ubiquitous 
and work without issue the vast majority of the time.  They are used to relocate 
addresses up to 2GB above or below the site of the relocation.  It's not working 
here but I can't see what's so special about this case.  The relocation sites 
are in a static library FWIW.  But in any case, on my system the original DLLs 
were located here:

/usr/bin/cyggmpxx-4.dll                   base 0x0003fcee0000 size 0x0000e000
/usr/bin/cyggmp-10.dll                    base 0x0003fcef0000 size 0x0008c000

and I rebased them to here:

/usr/bin/cyggmpxx-4.dll                   base 0x0000fcee0000 size 0x0000e000
/usr/bin/cyggmp-10.dll                    base 0x0000fcef0000 size 0x0008c000

and that did the trick.  I'll take the second bug to the developers list.
Cheers,

..mark


--
Problem reports:       http://cygwin.com/problems.html
FAQ:                   http://cygwin.com/faq/
Documentation:         http://cygwin.com/docs.html
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple