[PATCH] Bug 14070 - Deeply nested struct types cause opt run time to explode

[Oleg Ranevskyy]

Hi rafael, chandlerc,

**PROBLEM**

The problem appears on deeply nested structures due to the //getFoldedSizeOf// implementation (lib/IR/ConstantFold.cpp), which is very ineffective on such data types.

//getFoldedSizeOf// runs through the structure's fields to find out if all the fields are of the same size. If the function finds they are of different size, it calls //ConstantExpr::getSizeOf// with the structure type as an argument. //ConstantExpr::getSizeOf// internally leads to //getFoldedSizeOf// call that iterates the structure's fields again calculating structure size.

Since //getFoldedSizeOf// has recursive nature, the number of its calls grows immensely with structure nesting depth increase. E.g., for the attached structure having nesting depth=21 it is called 31719423 times. 

**PROPOSED SOLUTION**

//getFoldedSizeOf// is doing the same work twice: 
 1) when iterating the structure's fields to compare field sizes;
 2) when finally calling //ConstantExpr::getSizeOf// to calculate structure size.

The idea is to eliminate the 2nd step and calculate the whole size at the 1st step by summarizing sizes of separate fields.

This solution reduces //getFoldedSizeOf// calls from over 31M to about 80. The structure is processed in a fraction of a second.

Regression and performance tests show no regression.

http://reviews.llvm.org/D4928

Files:
  lib/IR/ConstantFold.cpp

Index: lib/IR/ConstantFold.cpp
===================================================================
--- lib/IR/ConstantFold.cpp
+++ lib/IR/ConstantFold.cpp
@@ -349,19 +349,29 @@
       if (NumElems == 0)
         return ConstantExpr::getNullValue(DestTy);
       // Check for a struct with all members having the same size.
-      Constant *MemberSize =
+      Constant *FirstMemberSize =
         getFoldedSizeOf(STy->getElementType(0), DestTy, true);
+      Constant *StructSize = FirstMemberSize;
+
       bool AllSame = true;
-      for (unsigned i = 1; i != NumElems; ++i)
-        if (MemberSize !=
-            getFoldedSizeOf(STy->getElementType(i), DestTy, true)) {
+      for (unsigned i = 1; i != NumElems; ++i){
+        Constant *MemberSize = getFoldedSizeOf(STy->getElementType(i), DestTy, true);
+        StructSize = ConstantExpr::getNUWAdd(StructSize, MemberSize);
+
+        if (FirstMemberSize != MemberSize){
+          if (!Folded)
+            return nullptr;
+
           AllSame = false;
-          break;
         }
+      }
+
       if (AllSame) {
         Constant *N = ConstantInt::get(DestTy, NumElems);
-        return ConstantExpr::getNUWMul(MemberSize, N);
+        return ConstantExpr::getNUWMul(FirstMemberSize, N);
       }
+
+      return ConstantExpr::getCast(CastInst::getCastOpcode(StructSize, false, DestTy, false), StructSize, DestTy);
     }
 
   // Pointer size doesn't depend on the pointee type, so canonicalize them
-------------- next part --------------
A non-text attachment was scrubbed...
Name: D4928.12552.patch
Type: text/x-patch
Size: 1461 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20140815/5df3d86a/attachment.bin>