llvm dev - Jul 2014 - [LLVMdev] LLVM Code Generation on flattened IR code

Hello,

I made two flattening transformation filters, one use a switch instruction for
the dispatcher, and the other an indirectbranch instruction. Both work well but
llc is very time consuming in the second case : 10 minutes for a 500 basic block
function, while it takes 10 seconds for the switchcase implementation . The
instrumentation of llc shows that most of the time is passed in  the Machine
Block Frequency Analysis pass


I don't understand why there is so huge difference. Is there a way to speed
up llc ?

Charles Salmon

> On Jul 17, 2014, at 11:41 AM, charlessl at free.fr wrote:
> 
> Hello,
> 
> I made two flattening transformation filters, one use a switch instruction
for the dispatcher, and the other an indirectbranch instruction. Both work well
but llc is very time consuming in the second case : 10 minutes for a 500 basic
block function, while it takes 10 seconds for the switchcase implementation .
The instrumentation of llc shows that most of the time is passed in  the Machine
Block Frequency Analysis pass
> 
> 
> I don't understand why there is so huge difference. Is there a way to
speed up llc ?
That’s a bug. Can you try with trunk and provide the LLVM IR test case? You can
file a PR it at llvm.org/bugs.

There’s been some work on the block frequency algorithm recently, which is
shared between LLVM IR and Machine IR, so it would be interesting to get that
test case.

-Andy


-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20140718/9d2f97e8/attachment.html>

Thanks, i will do that
how can i attach a LLVM IR source file  to  a bug report?

Charles


----- Mail original -----
De: "Andrew Trick" <atrick at apple.com>
À: charlessl at free.fr
Cc: llvmdev at cs.uiuc.edu
Envoyé: Vendredi 18 Juillet 2014 09:18:41
Objet: Re: [LLVMdev] LLVM Code Generation  on flattened IR code






On Jul 17, 2014, at 11:41 AM, charlessl at free.fr wrote: 

Hello, 

I made two flattening transformation filters, one use a switch instruction for
the dispatcher, and the other an indirectbranch instruction. Both work well but
llc is very time consuming in the second case : 10 minutes for a 500 basic block
function, while it takes 10 seconds for the switchcase implementation . The
instrumentation of llc shows that most of the time is passed in the Machine
Block Frequency Analysis pass


I don't understand why there is so huge difference. Is there a way to speed
up llc ?


That’s a bug. Can you try with trunk and provide the LLVM IR test case? You can
file a PR it at llvm.org/bugs .


There’s been some work on the block frequency algorithm recently, which is
shared between LLVM IR and Machine IR, so it would be interesting to get that
test case.


-Andy