[LLVMdev] Polly test and example

[MORIYAMA Tomohiro]

(2011/06/10 14:01), Tobias Grosser wrote:
> On 06/10/2011 01:52 AM, MORIYAMA Tomohiro wrote:
>> (2011/06/10 13:25), Tobias Grosser wrote:
>>> On 06/10/2011 12:00 AM, MORIYAMA Tomohiro wrote:
>>>> H, Tobias
>>>> Thank you for your reply.
>>>>
>>>>
>>>> Tobias Grosser wrote:
>>>>> Interesting. On what kind of Platform are you running this? Is it
>>>>> still
>>>>> Ubuntu 11.04? I assume it's an intel platform, but is it a 32-bit or
>>>>> 64bit?
>>>>
>>>> It is a little bit complicated.
>>>>
>>>> Intel Core i5 M520.
>>>> Ubuntu 11.04-i386 using VMWare Player on Windows 7 Professional(64bit).
>>>
>>> So it's 32bit.
>>
>> Certainly.
>>
>>>>>
>>>>> The test cases that are failing are the ones that use 'lli' to execute
>>>>> the generated program. There seems to be something wrong in respect to
>>>>> lli. Can you run e.g. single_loop.ll with llvm-lit (-v) and show the
>>>>> error message.
>>>>
>>>> OK. This is the message.
>>>>
>>>> --------
>>>> tomohiro at ubuntu:~/llvm/tools/polly/test/CodeGen$ llvm-lit -v
>>>> single_loop.ll
>>>> llvm-lit: lit.cfg:94: note: using out-of-tree build at
>>>> '/home/tomohiro/build/
>>>> tools/polly'
>>>> -- Testing: 1 tests, 2 threads --
>>>> FAIL: Polly :: CodeGen/single_loop.ll (1 of 1)
>>>> ******************** TEST 'Polly :: CodeGen/single_loop.ll' FAILED
>>>> *************
>>>> *******
>>>> Script:
>>>> --
>>>> opt -load /home/tomohiro/build/lib/LLVMPolly.so -basicaa
>>>> -polly-prepare -polly
>>>> -region-simplify -scev-aa -polly-cloog -analyze<
>>>> /home/tomohiro/llvm/tools/
>>>> polly/test/CodeGen/single_loop.ll | FileCheck
>>>> /home/tomohiro/llvm/tools/polly/
>>>> test/CodeGen/single_loop.ll
>>>> opt -load /home/tomohiro/build/lib/LLVMPolly.so -basicaa
>>>> -polly-prepare -polly
>>>> -region-simplify -scev-aa -polly-codegen<
>>>> /home/tomohiro/llvm/tools/polly/test
>>>> /CodeGen/single_loop.ll | lli -
>>>> --
>>>> Exit Code: 139
>>>> Command Output (stderr):
>>>> --
>>>> Stack dump:
>>>> 0. Program arguments: lli -
>>>> /home/tomohiro/build/tools/polly/test/CodeGen/Output/single_loop.ll.script:
>>>>
>>>>
>>>> line
>>>> 2: 3250 Done opt -load /home/tomohiro/build/lib/LLVMPolly.so
>>>> -basicaa -polly-prepare -polly-region-simplify -scev-aa
>>>> -polly-codegen< /home/
>>>> tomohiro/llvm/tools/polly/test/CodeGen/single_loop.ll
>>>> 3251 Segmentation fault | lli -
>>>
>>>
>>> Interesting. It seems lli is segfaulting. Can you try to use the
>>> following command?
>>>
>>> opt -load /home/tomohiro/build/lib/LLVMPolly.so \
>>> -basicaa -polly-prepare -polly-region-simplify -scev-aa \
>>> -polly-codegen \
>>> /home/tomohiro/llvm/tools/polly/test/CodeGen/single_loop.ll \
>>> | llc - out.s && gcc out.s && ./a.out
>>>
>>> This means we do not run the created LLVM-IR file with the lli JIT
>>> compiler, but use the llc compiler and than gcc to statically compile
>>> the binary.
>>
>> OK. This result is as follows.
>>
>> tomohiro at ubuntu:~/llvm/tools/polly/test/CodeGen$ opt -load
>> /home/tomohiro/build/lib/LLVMPolly.so \
>> > -basicaa -polly-prepare -polly-region-simplify -scev-aa \
>> > -polly-codegen \
>> > /home/tomohiro/llvm/tools/polly/test/CodeGen/single_loop.ll \
>> > | llc - out.s && gcc out.s && ./a.out
>> llc: Too many positional arguments specified!
>> Can specify at most 1 positional arguments: See: llc -help
>
> Sorry, the last line should be:
> | llc -o out.s && gcc out.s && ./a.out

This is the corrected result.

tomohiro at ubuntu:~$ opt -load /home/tomohiro/build/lib/LLVMPolly.so \
 >  -basicaa -polly-prepare -polly-region-simplify -scev-aa \
 >  -polly-codegen \
 >  /home/tomohiro/llvm/tools/polly/test/CodeGen/single_loop.ll \
 >  | llc -o out.s && gcc out.s && ./a.out
out.s: Assembler messages:
out.s:8: Error: bad register name `%rsp'
out.s:9: Error: bad register name `%rsp)'
out.s:12: Error: invalid instruction suffix for `call'
out.s:14: Error: bad register name `%rax'
out.s:18: Error: bad register name `%rcx'
out.s:22: Error: bad register name `%rcx)'
out.s:23: Error: bad register name `%rsp,%rcx,4)'
out.s:24: Error: bad register name `%rdx'
out.s:29: Error: bad register name `%rax'
out.s:36: Error: bad register name `%rsp,%rax,4)'
out.s:39: Error: bad register name `%rsp)'
out.s:42: Error: bad register name `%rsp)'
out.s:44: Error: bad register name `%rsp)'
out.s:45: Error: bad register name `%rsp'
tomohiro at ubuntu:~$


>>> In case this does not work, I will investigate this on an i386 machine.
>>> All our development machines are amd64 machines and I must admit we did
>>> not test too much on i386. So issues on those machines are not too
>>> unexpected. Though Raghesh runs i386. He may have seen such a problem
>>> before.
>>
>> I see. I will try it amd64 version but now on VMWare.
>> I'm glad if you tell me your environment in detail.
>
> Basically I have an Ubuntu amd64. So if you just want to go ahead with
> some experiments I would us an 64bit operating system. Running Polly on
> 64bit is possible (as raghesh did some work on this), but most probably
> our implementation is not yet 100% bug free on 64bit.

Thanks. I'll try it.

> I just looked again at the lli test cases and they specifically contain
> amd64 bit target triples. So most probably executing them on a 32bit
> system will never work. If you want to stay with 32bit, I would try the
> example on the webpage (from scratch) without using the pregenerated
> LLVM-IR files. Let me know if/how this works.
>
> Cheers and thanks for your testing
> Tobi
>
>

-- 
Tomohiro
moriyama at isit.or.jp