[llvm-commits] [llvm] r169079 [1/2] - in /llvm/trunk: lib/Transforms/IPO/GlobalOpt.cpp test/Transforms/GlobalOpt/big-memory-footprint.ll

[Nick Lewycky]

Zhou Sheng wrote:
> Author: sheng
> Date: Fri Nov 30 22:38:53 2012
> New Revision: 169079
>
> URL: http://llvm.org/viewvc/llvm-project?rev=169079&view=rev
> Log:
> The patch is to improve the memory footprint of pass GlobalOpt.
> Also check in a case to repeat the issue, on which 'opt -globalopt' consumes 1.6GB memory.
> The big memory footprint cause is that current GlobalOpt one by one hoists and stores the leaf element constant into the global array, in each iteration, it recreates the global array initializer constant and leave the old initializer alone. This may result in many obsolete constants left.
> For example:  we have global array @rom = global [16 x i32] zeroinitializer
> After the first element value is hoisted and installed:   @rom = global [16 x i32] [ 1, 0, 0, ... ]
> After the second element value is installed:  @rom = global [16 x 32] [ 1, 2, 0, 0, ... ]        // here the previous initializer is obsolete
> ...
> When the transform is done, we have 15 obsolete initializers left useless.

It's great to have extra details like how you have a testcase on which 
globalopt takes 1.6GB, but you failed to explain the change you made. 
The change improves the memory footprint, but what is the change?

I think I understand it after reading the code, but please remember that 
about commit messages in the future.

> Added:
>      llvm/trunk/test/Transforms/GlobalOpt/big-memory-footprint.ll
> Modified:
>      llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp
>
> Modified: llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp?rev=169079&r1=169078&r2=169079&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp (original)
> +++ llvm/trunk/lib/Transforms/IPO/GlobalOpt.cpp Fri Nov 30 22:38:53 2012
> @@ -36,6 +36,7 @@
>   #include "llvm/ADT/DenseMap.h"
>   #include "llvm/ADT/SmallPtrSet.h"
>   #include "llvm/ADT/SmallVector.h"
> +#include "llvm/ADT/SetVector.h"
>   #include "llvm/ADT/Statistic.h"
>   #include "llvm/ADT/STLExtras.h"
>   #include<algorithm>
> @@ -2398,7 +2399,8 @@
>   /// initializer.  This returns 'Init' modified to reflect 'Val' stored into it.
>   /// At this point, the GEP operands of Addr [0, OpNo) have been stepped into.
>   static Constant *EvaluateStoreInto(Constant *Init, Constant *Val,
> -                                   ConstantExpr *Addr, unsigned OpNo) {
> +                                   ConstantExpr *Addr, unsigned OpNo,
> +                                   SetVector<Constant*>&  Obsolete) {

The & should go next to Obsolete.

>     // Base case of the recursion.
>     if (OpNo == Addr->getNumOperands()) {
>       assert(Val->getType() == Init->getType()&&  "Type mismatch!");
> @@ -2415,7 +2417,9 @@
>       ConstantInt *CU = cast<ConstantInt>(Addr->getOperand(OpNo));
>       unsigned Idx = CU->getZExtValue();
>       assert(Idx<  STy->getNumElements()&&  "Struct index out of range!");
> -    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1);
> +    if (Elts[Idx]->getType()->isAggregateType())
> +      Obsolete.insert(Elts[Idx]);
> +    Elts[Idx] = EvaluateStoreInto(Elts[Idx], Val, Addr, OpNo+1, Obsolete);
>
>       // Return the modified struct.
>       return ConstantStruct::get(STy, Elts);
> @@ -2435,8 +2439,11 @@
>       Elts.push_back(Init->getAggregateElement(i));
>
>     assert(CI->getZExtValue()<  NumElts);
> +  Constant *OrigElem = Elts[CI->getZExtValue()];
> +  if (OrigElem->getType()->isAggregateType())
> +    Obsolete.insert(OrigElem);
>     Elts[CI->getZExtValue()] =
> -    EvaluateStoreInto(Elts[CI->getZExtValue()], Val, Addr, OpNo+1);
> +    EvaluateStoreInto(OrigElem, Val, Addr, OpNo+1, Obsolete);
>
>     if (Init->getType()->isArrayTy())
>       return ConstantArray::get(cast<ArrayType>(InitTy), Elts);
> @@ -2452,9 +2459,20 @@
>       return;
>     }
>
> +  // Collect obsolete constants created in previous CommitValueTo() invoke.
> +  SetVector<Constant*>  Obsolete;
>     ConstantExpr *CE = cast<ConstantExpr>(Addr);
>     GlobalVariable *GV = cast<GlobalVariable>(CE->getOperand(0));
> -  GV->setInitializer(EvaluateStoreInto(GV->getInitializer(), Val, CE, 2));
> +  Constant *OrigInit = GV->getInitializer();
> +  if (OrigInit->getType()->isAggregateType())
> +    Obsolete.insert(OrigInit);
> +  Constant *Init = EvaluateStoreInto(OrigInit, Val, CE, 2, Obsolete);
> +  GV->setInitializer(Init);
> +
> +  for (unsigned i = 0; i<  Obsolete.size(); ++i) {
> +    if (Obsolete[i]->use_empty())
> +      Obsolete[i]->destroyConstant();
> +  }

This should use the form:

   for (unsigned i = 0, e = Obsolete.size(); i != e; ++i)

as explained in the style guide: 
http://llvm.org/docs/CodingStandards.html#don-t-evaluate-end-every-time-through-a-loop

Nick