Re: [Mingw-w64-public] Problem with gmp-ecm and mingw64...

[David C. <wr...@mo...>]

Hello,

The file that I think may be causing problems does not contain inline assembler, 
it is a straight assembly file (redc.asm).  I do not see any inline assembly in 
gmp-ecm.

One strange thing I see at the top of that file is the following:
# void ecm_redc3(mp_limb_t * z, const mp_limb_t * x, size_t n, mp_limb_t m)
#                     %rdi           %rsi              %rdx      %rcx
#  MinGW w64          %rcx           %rdx              %r8       %r9
#  save them in       %r8            %r9               %r10      %r11

Then there are a few push and pop statements in the file.  Maybe the wrong 
registers are being push-ed and pop-ed for the way mingw64 works?  Hmmm, but 
then why would some printf statements make it work?

I'm not very good at tracking down some of these problems.  I wish I could 
create a small test case to reproduce the problem, but I'm not sure that this 
file is the problem.  Actually, I'll try to create a test case with this file 
later today to see what happens.  But in the mean time, can someone else try the 
steps I have outlined below and let me know if you run into the same problem? 
Thanks for your time.

-David C.

On 2/19/2012 11:48 PM, Kai Tietz wrote:
> Hi,
>
> such failures a commonly caused by inline-assembler not treating
> memory-ref proper.  I assume that inline assembler doesn't specify
> explicit memory-clobber/modify.  Newer gcc versions are here more
> strict then older versions.
>
> Kai
>
> 2012/2/20 David Cleaver:
>> Hello everyone,
>>
>> I have run into a very strange problem and I am not sure what might be causing
>> it.  I am compiling the latest svn of gmp-ecm (right now it is 1746) and
>> depending on whether I insert a few extra print statements or not, one certain
>> test in gmp-ecm will either run to completion or will seg_fault.  Can someone
>> here help me test whether this might be a problem with MingW64, Win XP x64, MPIR
>> 2.5.0, or maybe something else?  I have run into the problem with two different
>> sezero builds (20101003 and 20111101) as my MingW64 compiler.  I am compiling
>> inside an MSYS shell (1.0.16).  I am on a Windows XP x64 computer.
>>
>> Steps to reproduce the problem I am seeing:
>> 0) Use MPIR as your GMP library and MingW64 as your compiler
>> 1) Download latest gmp-ecm svn (currently 1746)
>> 2) run 'autoreconf -i'
>> 3) I configure with the following options:
>> ./configure CC=gcc CFLAGS="-O2 -pedantic -m64 -std=gnu99 -mtune=core2
>> -march=core2" --enable-asm-redc --build=x86_64-w64-mingw32 --disable-assert
>> 4) run 'make'
>>
>> At this point we can just run the test that fails for me, it is (sorry for any
>> wrap around):
>> echo
>> 449590253344339769860648131841615148645295989319968106906219761704350259884936939123964073775456979170209297434164627098624602597663490109944575251386017
>> | ecm -sigma 63844855 -go 172969 61843 20658299
>>
>> The test should produce a 27 digit factor of the 153 digit input.  However, in
>> my setup it seg_faults, actually just a silent crash with no screen output.
>>
>> Now, I have been using some printf's (actually outputf in gmp-ecm code) to try
>> to track down where the crash was happening.  I have found that if I put three
>> outputf statements in one part of the code, the program runs successfully and
>> produces the factorization.  Can anyone else here reproduce this problem?  I'm
>> curious to see if this problem happens for anyone else on WinXP x64, or maybe on
>> Win7 x64.
>>
>> If anyone can reproduce the crash, can you see if the following change fixes the
>> problem in your environment?  The diff just applies to one file, mpmod.c.
>> $ diff -u ../svn/mpmod.c mpmod.c
>> --- ../svn/mpmod.c      2012-02-12 19:59:32 -0600
>> +++ mpmod.c     2012-02-19 14:48:36 -0600
>> @@ -719,8 +719,11 @@
>>   #endif /* otherwise go through to the next available mode */
>>           case MPMOD_MUL_REDC3: /* mpn_mul_n + ecm_redc3 */
>>   #if defined(HAVE_ASM_REDC3)
>> +outputf(OUTPUT_VERBOSE, "tmp0=0x%I64x tmp1=0x%I64x tmp2=0x%I64x\n", tmp[0],
>> tmp[1], tmp[2]);
>>             mpn_sqr (tmp, s1p, nn);
>> +outputf(OUTPUT_VERBOSE, "tmp0=0x%I64x tmp1=0x%I64x tmp2=0x%I64x\n", tmp[0],
>> tmp[1], tmp[2]);
>>             ecm_redc3 (tmp, np, nn, invm[0]);
>> +outputf(OUTPUT_VERBOSE, "tmp0=0x%I64x tmp1=0x%I64x tmp2=0x%I64x\n", tmp[0],
>> tmp[1], tmp[2]);
>>             cy = mpn_add_n (rp, tmp + nn, tmp, nn);
>>             if (cy != 0)
>>               mpn_sub_n (rp, rp, np, nn); /* a borrow should always occur here */
>>
>> At this point, if you run 'make clean' and 'make' again, then run the test case
>> from up above, it should produce a factor very quickly.
>>
>> For some strange reason, all three of the 'outputf' statements above are
>> necessary to make gmp-ecm run to completion.  If any of those 'outputf'
>> statements are left out, then the program will silently crash.
>>
>> My guess is that there is some assembly code in gmp-ecm (the ecm_redc3 function,
>> found in ecm_svn_1746/x86_64/redc.asm) that MingW64 might not be processing
>> correctly.  This is just a guess based on the fact that the crash happens with
>> both MPIR and GMP (it only happens to GMP if it goes through the same code path
>> as MPIR, ask me for the change if you would like to see it with GMP).
>>
>> Is anyone willing to help me track this problem down?  Is there any more
>> information that anyone needs to look into this problem?  Thanks for your time
>> and I'll talk with you later.
>>
>> -David C.