Issue on windows build (mingw64)

Vlad Gabriel clampi at yahoo.com
Tue Apr 17 07:23:06 CEST 2012


Hello Torbjörn,

I can confirm issue 2 was fixed with revision 14840:dc13f686713f. However, for mpn/x86_64/fastsse/lshiftc-movdqu2.asm, I strongly suggest to preserve %xmm7 across function call; it is required by the ABI. Without affecting the stack, you can use the register parameters shadow area that is required by the calling convention. The area is conveniently 4*8=32 bytes in size, and 16 byte alligned. Not saving %xmm7 across the call, will put a time-bomb in the code, as it will produce failure at some point in the caller who expects %xmm7 to be saved and will be hard to find the root cause.

Regarding issue 3, un-commenting those statements made a lot of test fail throughout with SIGSEV. I tried debugging some, but fail to understand why they crash, as I had trouble understanding the renaming macros purpose. As a note, although I do have the impression that some pop's are missing from the epilogue, the crash happens in function itself, not when the return is encountered. Please advise if you have more thoroughly investigations steps on how to further debug the cause.


Best Regards,
Gabriel


________________________________
 From: Torbjorn Granlund <tg at gmplib.org>
To: Vlad Gabriel <clampi at yahoo.com> 
Cc: "gmp-bugs at gmplib.org" <gmp-bugs at gmplib.org> 
Sent: Monday, April 16, 2012 9:23 PM
Subject: Re: Issue on windows build (mingw64)
 
Vlad Gabriel <clampi at yahoo.com> writes:

  Issue 1

Let's hope Marco has something saggio to say about this...
  
  Issue 2

  Is related with latest changes in mpn/x86_64/coreisbr/lshift.asm,
  mpn/x86_64/coreisbr/lshiftc.asm, etc

Please try revision 14839:a77538915df3 and let the list know if issue 2
is solved.

Let me add issue 3:

Could you perhaps also try uncommenting ABI_SUPPORT in
coreisbr/mul_1.asm and coreisbr/aorsmul_1.asm, and see if things still
work?

-- 
Torbjörn


More information about the gmp-bugs mailing list