[openssl-users] Unhandled exception at 0x005904dc (libeay32.dll) (Windows x86)

[Jakob Bohm]

On 22/08/2016 22:33, Scott Ware wrote:
>
> On Mon, Aug 22, 2016 at 3:04 PM, Jakob Bohm <jb-openssl at wisemo.com 
> <mailto:jb-openssl at wisemo.com>>wrote:
>
>     On 22/08/2016 20:09, Scott Ware wrote:
>
>         We use libeay32.dll and ssleay32.dll from
>         https://slproweb.com/products/Win32OpenSSL.htmlin
>         <https://slproweb.com/products/Win32OpenSSL.htmlin>our
>         applications and we recently moved from version 1.0.2a to
>         1.0.2g and now on a few machines running a AMD Geode processor
>         we are getting "Unhandled exception at 0x005904dc
>         (libeay32.dll) in Test.exe: 0xC000001D: Illegal Instruction".
>         We ended up building OpennSSL so we could debug into it and
>         found it is failing on "movsd xmm0,mmword" (see below) which
>         the AMD Geode does not seem to support. I have tried "SET
>         OPENSSL_ia32cap=~0x1000000", "SET OPENSSL_ia32cap=~0x2000000",
>         and "SET OPENSSL_ia32cap=~0x7000000"; and nothing seems to
>         change. I may not be using OPENSSL_ia32cap correctly. This
>         happens when calling SSL_CTX_new which then calls RAND_add.
>
>         Any ideas on the best thing to do? We don't want to have to
>         manage different compiled versions of libeay32.dll and
>         ssleay32.dll if we can help it.
>
>     Your disassembly looks like the C compiler was invoked with
>     options that caused regular C floating point code (in this
>     case, the passing of 45.0 as an argument to RAND_add()) to
>     be compiled into MMX/SSE instructions instead of backwards
>     compatible 80x87 floating point instructions or (for simple
>     cases like this) regular integer unit data movement
>     instructions (such as two pushes of 32 bit constants that
>     contain the halves of the 64 bit double constant, which
>     would have been more efficient on every x86 CPU).
>
>     Did the build scripts or other source code contain any
>     differences from the official source code that can be
>     downloaded from openssl.org <http://openssl.org>?
>
>     How did you invoke the build scripts (command sequence,
>     special build environment, special environment variables
>     etc.)?
>
>     Which compiler and compiler version/edition did you use?
>
>     It would be interesting to know if one of the common Windows
>     compilers does this unconditionally, making it unsuitable
>     for use in programs that need to be backwards compatible.
>
>
>
> I compiled using this process and seem to be getting the same result 
> as the .dll I downloaded from slproweb.com <http://slproweb.com>
> I downloaded the 1.0.2g source from openssl.com 
> <http://openssl.com>and didn't change anything.
>
> From the "Developer Command Promt for VS2013"
> perl Configure debug-VC-WIN32 no-asm --prefix=C:\OpenSSL-VC-32-dbg
> ms\do_ms
> nmake -f ms\ntdll.mak
> nmake -f ms\ntdll.mak install
According to the following page

https://msdn.microsoft.com/en-us/library/7t5yh4fd%28v=vs.120%29.aspx

Visual Studio 2012 and later requires the following compiler
option to generate code compatible with older CPUs (this is the
default in Visual Studio 2010, and VS2010 does not support the
option):

/arch:IA32


This compiler gotcha is specific to the 32 bit x86 architecture,
the default looks like it is still sane for x86_64.

Note to the FIPS team: Please check if this affects the FIPS
module building procedure.


Enjoy

Jakob
-- 
Jakob Bohm, CIO, Partner, WiseMo A/S.  https://www.wisemo.com
Transformervej 29, 2860 Søborg, Denmark.  Direct +45 31 13 16 10
This public discussion message is non-binding and may contain errors.
WiseMo - Remote Service Management for PCs, Phones and Embedded