[LLVMdev] Emulating an infinite register file in the backend

[Justin Holewinski]

On Mon, Feb 11, 2013 at 6:32 PM, <nkavv at physics.auth.gr> wrote:

> Hi Justin and all,
>
> you've mentioned that you used an easy "trick" for defining an infinite
> register file in the backend.
>

The original PTX back-end just didn't perform register allocation.  All
registers emitted in the assembly were virtual registers, mapped to a
consecutive range.  E.g., %vreg7, %vreg13, and %vreg72 might be emitted as
%r0, %r1, %r2.  Just before machine code emission, we scanned the function
to determine this range.  Unfortunately, during the switch to the NVPTX
back-end, some changes were made in LLVM that prevented this from working
in some cases.  Basically, some post-regalloc passes were requiring
physical registers and choked when they saw virtual registers.  Because of
this, the newer NVPTX back-end uses a large physical register file and
performs register allocation.  It's not ideal, but I haven't had the time
to identify and fix every case where virtual registers won't work yet.


>
> Does this involve defining a single dummy register, and then adding this
> dummy reg to each register class? Is anything more needed? I refer to the
> RegisterInfo.td file, as e.g: llvm-3.0.src/lib/Target/PTX/**
> PTXRegisterInfo.td
>

You can define one dummy register, or one for each register class.  You
only need to define the register classes that you support, and the way
tablegen generates code leads to compiler warnings about empty arrays if
you do not add at least one register to each register class.


>
> I'm (just starting) implementing an LLVM target that bares some
> similarities to PTX. One of these is the notion of infinite registers,
> since my target too is a kind of virtual/universal machine.
>

It's a bit of an uphill battle since there are places in LLVM that make
assumptions about targeting a "real" architecture, to varying degrees.  One
thing to keep in mind is that the PTX back-end from 3.1 may be a better
example to use than the NVPTX back-end in 3.2 and TOT.  The original PTX
back-end supported less (and is therefore a bit simpler), and it also
doesn't suffer from as much backwards-compatibility cruft.  NVPTX has
evolved over many LLVM revisions, and sometimes uses old techniques.  I'm
working on getting it up to date.


>
> My target will not emit object files, so there i would prefer not to
> define a very larget register file and mess with formats. A bit-level
> format will at some point be defined but i don't want to fix it from now.
>
> So your approach seems highly relevant.
>
> Best regards
> Nikolaos Kavvadias
>
>
> Quoting Justin Holewinski <justin.holewinski at gmail.com>:
>
>  configure:12131: $? = 0
>> configure:12145: result: yes
>> configure:12157: checking tool compatibility
>> configure:12180: error: g++|clang++|icc required but not found
>>
>> configure can't find a c++ compiler.  On Ubuntu, please install g++ (sudo
>> apt-get install g++).
>>
>> This error should have been written to your console when you ran
>> configure.
>>
>>
>>
>> On Sun, Feb 10, 2013 at 8:35 AM, Manoj C <manoj.chinthala at gmail.com>
>> wrote:
>>
>>  hello sir,
>>>           i am using llvm compiler for my project.i an doing llvm
>>> installation.i followed all the steps in the llvm website but after
>>> running
>>> this command
>>>
>>> ../llvm/configure --enable-targets=host-only  it executed but it didnt
>>> create any make file in build directory ,only config.log file is appeared
>>> after running this command.
>>>
>>> after running the later command it is showing like this
>>> manoj at ubuntu:~/Desktop/LLVM/
>>> build$ make -j 4
>>> make: *** No targets specified and no makefile found.  Stop
>>>
>>> please help me to finish installation. i am running installation in
>>> ubuntu os.
>>>
>>> here is the config.log file ...
>>>
>>> ______________________________**_________________
>>> LLVM Developers mailing list
>>> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
>>> http://lists.cs.uiuc.edu/**mailman/listinfo/llvmdev<http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev>
>>>
>>>
>>>
>>
>> --
>>
>> Thanks,
>>
>> Justin Holewinski
>>
>>
>
>


-- 

Thanks,

Justin Holewinski
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130211/b889d383/attachment.html>