Re: linux-next: Tree for November 29 (aesni-intel)

[Randy Dunlap <randy.dunlap@oracle.com>]

On 11/29/10 11:45, Mathias Krause wrote:
> On 29.11.2010, 20:31 Randy Dunlap wrote:
>> On 11/29/10 11:21, Mathias Krause wrote:
>>> On 29.11.2010, 19:54 Randy Dunlap wrote:
>>>> On 11/29/10 10:26, Mathias Krause wrote:
>>>>> On 29.11.2010, 17:31 Randy Dunlap wrote:
>>>>>> On Mon, 29 Nov 2010 14:03:35 +1100 Stephen Rothwell wrote:
>>>>>>
>>>>>>> Hi all,
>>>>>>>
>>>>>>> Changes since 20101126:
>>>>>>
>>>>>>
>>>>>> on i386 builds, I get tons of these (and more) errors:
>>>>>>
>>>>>> arch/x86/crypto/aesni-intel_asm.S:841: Error: bad register name `%r12'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:842: Error: bad register name `%r13'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:843: Error: bad register name `%r14'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:844: Error: bad register name `%rsp'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:849: Error: bad register name `%rsp'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:850: Error: bad register name `%rsp'
>>>>>> arch/x86/crypto/aesni-intel_asm.S:851: Error: bad register name `%r9'
>>>>>>
>>>>>> even though the kernel .config file says:
>>>>>>
>>>>>> CONFIG_CRYPTO_AES=m
>>>>>> CONFIG_CRYPTO_AES_586=m
>>>>>> CONFIG_CRYPTO_AES_NI_INTEL=m
>>>>>>
>>>>>> Should arch/x86/crypto/aesni-intel_asm.S be testing
>>>>>> #ifdef CONFIG_X86_64
>>>>>> instead of
>>>>>> #ifdef __x86_64__
>>>>>> or does that not matter?
>>>>>>
>>>>>> or is this a toolchain issue?
>>>>>
>>>>> Well, __x86_64__ should be a build-in define of the compiler while
>>>>> CONFIG_X86_64 is defined for 64 bit builds in include/generated/autoconf.h.
>>>>> So by using the latter we should be on the safe side but if your compiler
>>>>> defines __x86_64__ for 32-bit builds it's simply broken. Also git grep
>>>>> showed quite a few more places using __x86_64__ so those would miscompile on
>>>>> your toolchain, too.
>>>>>
>>>>> But it looks like linux-next is just missing
>>>>> 559ad0ff1368baea14dbc3207d55b02bd69bda4b from Herbert's git repo at
>>>>> git://git.kernel.org/pub/scm/linux/kernel/git/herbert/cryptodev-2.6.git.
>>>>> That should fix the build issue.
>>>>
>>>> The build problem still happens when that patch is applied.
>>>
>>> That's weird. So it must be something with your toolchain.
>>> Can you please post the output of the following commands?:
>>>
>>> $ touch /tmp/null.c; cc -m32 -dD -E /tmp/null.c | grep -E 'x86|i.86'
>>
>> #define __i386 1
>> #define __i386__ 1
>> #define i386 1
>> #define __i586 1
>> #define __i586__ 1
>>
>>> $ touch /tmp/null.c; cc -m64 -dD -E /tmp/null.c | grep -E 'x86|i.86'
>>
>> #define __x86_64 1
>> #define __x86_64__ 1
>>
>> So that's not the problem... and the patch below didn't help.
> 
> That's odd. The output of the commands looks good so the x86-64 specific code
> should be left out for 32-bit builds. :/
> 
>> Sorry that I even asked about that.  What next?
> 
> Can you please post the full error message. Meanwhile I'm checking out a
> linux-next tree, trying to reproduce your problem.
> 

I just built with "make V=1" to see the full commands that are used, but
that didn't help me either:

  gcc -Wp,-MD,arch/x86/crypto/.aesni-intel_asm.o.d  -nostdinc -isystem /usr/lib/gcc/x86_64-redhat-linux/4.4.1/include -I/lnx/src/NEXT/linux-next-20101129/arch/x86/include -Iinclude  -I/lnx/src/NEXT/linux-next-20101129/include -include include/generated/autoconf.h -D__KERNEL__ -D__ASSEMBLY__ -m32 -DCONFIG_AS_CFI=1 -DCONFIG_AS_CFI_SIGNAL_FRAME=1 -DMODULE  -c -o arch/x86/crypto/aesni-intel_asm.o /lnx/src/NEXT/linux-next-20101129/arch/x86/crypto/aesni-intel_asm.S


There are 2945 lines like this:

linux-next-20101129/arch/x86/crypto/aesni-intel_asm.S:841: Error: bad register name `%r12'

It's around 311 KB, so I'll just put it here instead of emailing it:
http://oss.oracle.com/~rdunlap/doc/cry32.out

-- 
~Randy
*** Remember to use Documentation/SubmitChecklist when testing your code ***