Hi Andrey,

Thanks so much for your reply. Yes, the ISA definitely does not support
these instructions, I guess I will have to modify some code to make
instruction selector to pattern match the existing instructions, and I
think the modification should be on  the or1k backend side. It would be
great if you can send me your modifications because I think I might need to
modify some of the toolchains too. Right now when I try to compile
SPEC2006, none of them pass the linking stages. I'm getting a lot of
"undefined reference to ..." error, for example, "stderr" which should be
in libc, but turns out not. I wonder if you have encountered similar
situations.

I really appreciate your help.
Patrick

2015-01-11 7:48 GMT-05:00 BAndViG <[email protected]>:

>   I checked the OpenRISC architecture spec. v.1.1 and it looks like
> OpenRISC ISA doesn’t include such instructions. So they have to be
> implemented somewhere in LLVM by appropriate routines.
> Personally, I’m working on FPU32 port now but I use GCC/NewLIB tool chain
> for SoftFloat (as testing tool), Whetstone (as FPU benchmark) and Coremark
> (as integer pipeline benchmark).
> I made some (really not significant)  modifications to port these tools
> under GCC/NewLIB, but I haven’t shared the source yet. If you are
> interested in source code of the tools I’m ready to e-mail them to you (the
> several KB in zip).
> BR
> Andrey
>
>  *From:* Ziqiang Patrick Huang <[email protected]>
> *Sent:* Friday, January 09, 2015 5:44 PM
> *To:* [email protected] ; [email protected]
> *Subject:* [OpenRISC] Instruction selection failed when compiling SPEC
> 2006
>
>  Hi,
>
> I'm trying to use llvm-or1k to compile SPEC 2006, I'm getting the
> following errors when compiling some of the benchmarks:
>
>
> LLVM ERROR: Cannot select: 0x5a96d80: f32 = fabs 0x5a95498 [ORD=48]
> [ID=122]
> LLVM ERROR: Cannot select: 0x39443f0: f32 = bitcast 0x38ffca8 [ORD=43]
> [ID=488]
> LLVM ERROR: Cannot select: 0x5f13298: f32 = fneg 0x5f2e6d0 [ORD=2] [ID=8]
>
> It seems like the ISA does not support a bunch of floating point
> instructions ... Is anyone working on this ?
>
> Also what benchmarks have been successfully compiled so far using
> llvm-or1k ?
>
> Thanks,
> Ziqiang
>
> ------------------------------
> _______________________________________________
> OpenRISC mailing list
> [email protected]
> http://lists.openrisc.net/listinfo/openrisc
>
>


-- 
*Ziqiang Huang*
*Electrical and Computer Engineering*
*Hudson 213c, Duke University*
*Tel: 919-491-3677*
*Email: [email protected] <[email protected]>*
_______________________________________________
OpenRISC mailing list
[email protected]
http://lists.openrisc.net/listinfo/openrisc

Reply via email to