llvm dev - Jan 2012 - [LLVMdev] Fwd: How to force the creation of arrays with zeroes?

Sorry, forgot to reply-all

---------- Forwarded message ----------
From: Anton Korobeynikov <anton at korobeynikov.info>
Date: Sat, Jan 21, 2012 at 20:59
Subject: Re: [LLVMdev] How to force the creation of arrays with zeroes?
To: Stepan Dyatkovskiy <STPWORLD at narod.ru>

> Though, there is also cases when we really need to transform it to 10,000
zeroes (just look at CBackend.cpp, when we print constant arrays, string #1027).
There is also cases when we need to get either operand[i] or zero (SCCP.cpp, str
#413). Both for CBackend and for SCCP we can reduce code size keeping
performance on the same level.
In these 2 particular cases the changes of the code size can
definitely be neglected. Even more, in SCCP the use of CAZ is a
performance win, since we do not have to check at all operands are
zero, we already (!) know this.

--
With best regards, Anton Korobeynikov
Faculty of Mathematics and Mechanics, Saint Petersburg State University


-- 
With best regards, Anton Korobeynikov
Faculty of Mathematics and Mechanics, Saint Petersburg State University

To Anton:
>>  In these 2 particular cases the changes of the code size can
>>  definitely be neglected. Even more, in SCCP the use of CAZ is a
>>  performance win, since we do not have to check at all operands are
>>  zero, we already (!) know this.
Little fragment from SCCP.cpp, lines 413-418:
      ...
      else if (ConstantStruct *CS = dyn_cast<ConstantStruct>(C))
        LV.markConstant(CS->getOperand(i));      // Constants are constant.
      else if (isa<ConstantAggregateZero>(C)) {
        Type *FieldTy =
cast<StructType>(V->getType())->getElementType(i);
        LV.markConstant(Constant::getNullValue(FieldTy));
      } else
      ...
I couldn't found that performance wins here. We can just hide these branches
either inside the improved implementation of ConstantVector/Array/Struct or
inside the wraper.

I agree with Duncan reasoning. We can just to store single zero item and flag.
This changes also allows make code cleaner. 
> if a constant array has many repeated elements, maybe the
> array could be stored in a compressed form.
Did you mean reducing actual number of operands? As I found all repeated values
instead a pointers to the same Value in LLVMContext (or I missed something?).

-Stepan.

Yep check out PR1324.  Doing something like this would be a great improvement.

-Chris

On Jan 21, 2012, at 9:42 AM, Duncan Sands <baldrick at free.fr> wrote:
> Hi Anton, in a solution without CAZ, isNullValue can just return true when
it
> sees the special "this ConstantArray is all zero" flag.  So all
the places that
> now look for CAZ can just use isNullValue instead and there need be no
> performance loss.  That said, CAZ is more "in your force" so less
likely to be
> forgotten about.  Another interesting possibility is to handle more than
just
> all zero values: if a constant array has many repeated elements, maybe the
> array could be stored in a compressed form.  An all zeros array would just
be a
> special case of this.
> 
> Ciao, Duncan.
> 
> On 21/01/12 18:13, Anton Korobeynikov wrote:
>> Sorry, forgot to reply-all
>> 
>> ---------- Forwarded message ----------
>> From: Anton Korobeynikov<anton at korobeynikov.info>
>> Date: Sat, Jan 21, 2012 at 20:59
>> Subject: Re: [LLVMdev] How to force the creation of arrays with zeroes?
>> To: Stepan Dyatkovskiy<STPWORLD at narod.ru>
>> 
>> 
>>> Though, there is also cases when we really need to transform it to
10,000 zeroes (just look at CBackend.cpp, when we print constant arrays, string
#1027). There is also cases when we need to get either operand[i] or zero
(SCCP.cpp, str #413). Both for CBackend and for SCCP we can reduce code size
keeping performance on the same level.
>> In these 2 particular cases the changes of the code size can
>> definitely be neglected. Even more, in SCCP the use of CAZ is a
>> performance win, since we do not have to check at all operands are
>> zero, we already (!) know this.
>> 
>> --
>> With best regards, Anton Korobeynikov
>> Faculty of Mathematics and Mechanics, Saint Petersburg State University
>> 
>> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev