[llvm-dev] Performance of large llvm::ConstantDataArrays

[Chris Lattner via llvm-dev]

> On Sep 7, 2017, at 11:06 PM, Chris Lovett via llvm-dev <llvm-dev at lists.llvm.org> wrote:
> 
> I'm running into some pretty bad performance in llc.exe when compiling some large neural networks into code that contains some very large llvm::ConstantDataArrays, some are { size=102,760,448 }. There's a small about of actual code for processing the network, but the assembly is mostly global data.

Yes, llvm’s representation of constant arrays is insanity for cases like this.  Your case is bad, but just imagine the cost of a large char[] initialization: even though each byte is stored as a ConstantInt, the bloat isn’t huge because they are uniqued.  The real problem comes from each entry in the ConstantArray being stored as an operand list.  An operand in the operand list consumes something like 3-4 words per operand to maintain the uselist and a bunch of other nonsense that isn’t right for this.

IMO, there is a relatively easy solution for this.  Introduce a new subclass of ConstantData which represents a blob of data that gets emitted to the .o file, stored in a reasonable native format.  I think it would be fine to limit this to only representing arrays of primitive types (e.g. array of float, array of bytes, etc) since this keeps the API to the type simple (the type models an array, so it should have array element members only), and things that want to get the elements of the array out can have them returned as ConstantInt’s (or whatever).  I’d name this something like “ConstantArrayBlob”.

There are cases this wouldn’t cover well, e.g. an array of small structs, but I think that is ok, and it could be feature crept to support that over time.   The next trick is adding the corresponding special case to Clang to not generate the ConstantArray and the ConstantFP/Int members when given a candidate initialization.  This can be done as a secondary optimization after the basic mechanics are in place.

-Chris
 

> 
> I'm finding that llc.exe memory spikes up around 30 gigabytes and the job takes 20-30 minutes compiling from bitcode.  When I looked into it I found that every single floating point number is loaded into ConstantFP object where the float is parsed into exponent, mantissa and stored in an integer part is stored in a heap allocated array, then these are emitted into MCDataFragments where again more heap allocated data, the float appears to be stored in SmallVectorImpl<char>.  On top of this I see a lot of MCFillFragments added because of long double padding.
> 
> All up the code I'm compiling ends up with 276 million MCFragments, which just take a super long time in each phase of compiling (loading from bitcode, emitting, layout and writing).  With a peak working set of 30 gigabytes each float is taking around 108 bytes!  
> 
> Is there a more efficient way to do this? Or is there any plan in the works to handle global data more efficiently in llc ?
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20170909/356c05ef/attachment.html>