llvm dev - Feb 2012 - [LLVMdev] llc is taking a long time at -O1 to convert *.bc to *.s

Hi All,

I am using llvm-3.0 release version for compiling some files. I have a cpp file
on which I did a few custom passes. After that I used llc to convert the *.bc
file to *.s file at -O1 optimization level. llc took ~2630 seconds to perform
this conversion. I timed llc and these are the top three passes that is taking
most of the time.

 ---User Time---   --System Time--   --User+System--   ---Wall Time---  --- Name
---
  1816.1735 ( 77.8%)  23.3215 ( 69.7%)  1839.4950 ( 77.6%)  1839.4132 ( 77.6%) 
Simple Register Coalescing
  493.5948 ( 21.1%)   9.1846 ( 27.4%)  502.7794 ( 21.2%)  502.7681 ( 21.2%)  X86
DAG->DAG Instruction Selection
   5.7244 (  0.2%)   0.0000 (  0.0%)   5.7244 (  0.2%)   5.7204 (  0.2%) 
Machine Instruction LICM

I am running this in a fairly new Xeon server. Furthermore, at -O0, llc takes
less than 20 seconds to complete. Is this behavior expected and is there a way
to disable Simple Register Coalescing pass without switching to -O0 optimization
level to get around this problem?

Thanks
Sreekumar


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120203/5bf979ab/attachment.html>

Sreekumar,
Without a test case, we really can't determine if this is expected behavior.
Certainly, we'd prefer -O1 compile-times to be relatively fast.  Have you
tried this with a ToT clang?  If you're still seeing such poor performance
then I suggest filing a bug (http://llvm.org/bugs).  Please include a test case
and the command line arguments to reproduce the problem.  I believe ToT is
compatible with 3.0 bitcode files, so you could just submit your bitcode file.

 Chad



On Feb 2, 2012, at 11:03 PM, "Kodakara, Sreekumar V"
<sreekumar.v.kodakara at intel.com> wrote:
> Hi All,
>  
> I am using llvm-3.0 release version for compiling some files. I have a cpp
file on which I did a few custom passes. After that I used llc to convert the
*.bc file to *.s file at –O1 optimization level. llc took ~2630 seconds to
perform this conversion. I timed llc and these are the top three passes that is
taking most of the time.
>  
>  ---User Time---   --System Time--   --User+System--   ---Wall Time---  ---
Name ---
>   1816.1735 ( 77.8%)  23.3215 ( 69.7%)  1839.4950 ( 77.6%)  1839.4132 (
77.6%)  Simple Register Coalescing
>   493.5948 ( 21.1%)   9.1846 ( 27.4%)  502.7794 ( 21.2%)  502.7681 ( 21.2%)
X86 DAG->DAG Instruction Selection
>    5.7244 (  0.2%)   0.0000 (  0.0%)   5.7244 (  0.2%)   5.7204 (  0.2%) 
Machine Instruction LICM
>  
> I am running this in a fairly new Xeon server. Furthermore, at –O0, llc
takes less than 20 seconds to complete. Is this behavior expected and is there a
way to disable Simple Register Coalescing pass without switching to –O0
optimization level to get around this problem?
>  
> Thanks
> Sreekumar
>  
>  
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120203/064f68ea/attachment.html>

On Feb 2, 2012, at 11:03 PM, Kodakara, Sreekumar V wrote:
> Hi All,
>  
> I am using llvm-3.0 release version for compiling some files. I have a cpp
file on which I did a few custom passes. After that I used llc to convert the
*.bc file to *.s file at –O1 optimization level. llc took ~2630 seconds to
perform this conversion. I timed llc and these are the top three passes that is
taking most of the time.
>  
>  ---User Time---   --System Time--   --User+System--   ---Wall Time---  ---
Name ---
>   1816.1735 ( 77.8%)  23.3215 ( 69.7%)  1839.4950 ( 77.6%)  1839.4132 (
77.6%)  Simple Register Coalescing
>   493.5948 ( 21.1%)   9.1846 ( 27.4%)  502.7794 ( 21.2%)  502.7681 ( 21.2%)
X86 DAG->DAG Instruction Selection
>    5.7244 (  0.2%)   0.0000 (  0.0%)   5.7244 (  0.2%)   5.7204 (  0.2%) 
Machine Instruction LICM
>  
> I am running this in a fairly new Xeon server. Furthermore, at –O0, llc
takes less than 20 seconds to complete. Is this behavior expected and is there a
way to disable Simple Register Coalescing pass without switching to –O0
optimization level to get around this problem?
Without coalescing you will get really bad code, but the llc option to turn it
off is -join-intervals=0.

There is a good chance the performance problem has already been fixed on trunk.

/jakob

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20120203/42a04a99/attachment.html>