< Chronological >      < Thread >

Not sure if it's the same compiler version, but at least "back then"
when I last played with icc on IA64, the compiler always died as soon as
you has a a reference to an sse-type (__m128 &).

Cheers,

Ingo



Abe Stephens wrote:

> It is clear that no one has built in IA64 for quite some time:
>
> The problem in CPUTime.cc (appearing in both icc and gcc) is due to
> non-portable assembly code being used.  I'm not sure where CPUTime is
> used, possibly in the deadline image traverser.  I also discovered
> compilation problems in TesselatedCylinder (vla/pod error) and other
> new code.
>
> It appears that some code in RayPacket.h causes the IA64 icc to
> segfault. It appears to be SSE related as the files compile if SSE is
> disabled.  This might take some time to track down (by enabling
> MANTA_SSE and then explicitly disabling it for different sections of
> the header).
>
> I'd like very much for Manta to continue to build on Altix/SSE, but I
> don't have a large number of free cycles right now to find the
> offending code. It's likely that there is some intrinsic which is
> causing the issue and could be inclosed in a #if __ia64__ block.
>
> If you could help isolate where it is, I will certainly help fix it
> where it occurs.
>
> Abe
>
> On Aug 23, 2007, at 12:47 PM, Christian Odom wrote:
>
>>
>>
>> On 8/23/07, *Abe Stephens* <abe@sci.utah.edu
>> <mailto:abe@sci.utah.edu>> wrote:
>>
>>
>>     Which source file causes the icc internal error? It looks like
>>     icpc 9.0 segfaults when compiling CheapRNG.cc. (Unfortunately I
>>     don't have icc ia64 10.0 here...)
>>
>> Here is the output from make->
>> //////////////////////////////////////////////////////////////////////////////
>> [ 13%] Building CXX object Core/CMakeFiles/Manta_Core.dir/Math/CheapRNG.o
>> (0): internal error: backend signals
>>
>> compilation aborted for /store/toolkit/Manta/Core/Math/CheapRNG.cc
>> (code 4)
>> make[2]: *** [Core/CMakeFiles/Manta_Core.dir/Math/CheapRNG.o] Error 4
>> make[1]: *** [Core/CMakeFiles/Manta_Core.dir/all] Error 2
>> make: *** [all] Error 2
>> //////////////////////////////////////////////////////////////////////////////
>> It still seems to be CheapRNG
>>
>>     The gcc problem is certainly fixable by using separate code on
>>     IA64.  Back in the day icc was much, much faster than gcc--I'd
>>     imagine that is still the case today. Let's try to figure out
>>     what code is causing problems with icc before switching to gcc.
>>
>>     Abe
>>
>>     On Aug 23, 2007, at 12:10 PM, Christian Odom wrote:
>>
>>>     Hi guys
>>>
>>>     Im running into a couple of compile errors on our ia64 platforms
>>>     when trying to compile  Rev: 1678.
>>>
>>>     With icc 10.0 we are still getting the dreaded->
>>>     (0):  internal error: backend signals
>>>     //////////////////////////////////////////////////////////////////////////////
>>>
>>>
>>>     When using gcc 4.1.2 Im getting this mess->
>>>
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc: In static member
>>>     function `static
>>>     long long unsigned int Manta::CPUTime::currentTicks()':
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc:90: error: impossible
>>>     register
>>>     constraint in `asm'
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc:90: error: impossible
>>>     register
>>>     constraint in `asm'
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc:90: error: impossible
>>>     register
>>>     constraint in `asm'
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc:93: error:
>>>     unrecognizable insn:
>>>     (insn 28 40 55 2 0x2000000001631860 (parallel [
>>>     (set (reg:SI 2 r2)
>>>     (asm_operands/v:SI ("rdtsc") ("=a") 0 []
>>>     [] ("/store/toolkit/Manta/Core/Util/CPUTime.cc") 90))
>>>     (set (reg:SI 15 r15)
>>>     (asm_operands/v:SI ("rdtsc") ("=d") 1 []
>>>     [] ("/store/toolkit/Manta/Core/Util/CPUTime.cc") 90))
>>>     ]) -1 (insn_list:REG_DEP_ANTI 15 (insn_list 14 (insn_list 12
>>>     (insn_list 10 (insn_list 20 (nil))))))
>>>     (nil))
>>>     /store/toolkit/Manta/Core/Util/CPUTime.cc:93: confused by
>>>     earlier errors, bailing out
>>>     gmake[2]: *** [Core/CMakeFiles/Manta_Core.dir/Util/CPUTime.o]
>>>     Error 1
>>>     gmake[1]: *** [Core/CMakeFiles/Manta_Core.dir/all] Error 2
>>>     gmake: *** [all] Error 2
>>>     //////////////////////////////////////////////////////////////////////////////
>>>
>>>     I dont know if there is any hope for the icc compiler, but any
>>>     thoughts on the gcc issue would be awesome.
>>>     --
>>>     Christian Odom
>>>     Graduate Student
>>>     Center for Advanced Computer Studies
>>>     http://www.cacs.louisiana.edu < http://www.cacs.louisiana.edu>
>>>     Louisiana Immersive Technologies Enterprise
>>>     http://www.lite3d.com
>>
>>
>>
>>
>>
>> --
>> Christian Odom
>> Graduate Student
>> Center for Advanced Computer Studies
>> http://www.cacs.louisiana.edu
>> Louisiana Immersive Technologies Enterprise
>> http://www.lite3d.com
>
>