Grow BumpPtrAllocator's slab size dynamically if we allocated many slabs. This
reduces the amount of malloc calls and may reduce memory overhead.

Some numbers:
ASTContext stats, clang -cc1 -disable-free -fsyntax-only Cocoa_h.m
without dynamic growth                          |  with dynamic growth
Number of memory regions: 3158                  |  Number of memory regions: 432
Bytes used: 12333185                            |  Bytes used: 12333185
Bytes allocated: 12935168                       |  Bytes allocated: 12800000
Bytes wasted: 601983 (includes alignment, etc)  |  Bytes wasted: 466815 (includes alignment, etc)

ASTContext stats, clang -cc1 -disable-free -fsyntax-only on clang's ASTReader.cpp
without dynamic growth                          |  with dynamic growth
Number of memory regions: 10987                 |  Number of memory regions: 551
Bytes used: 42910356                            |  Bytes used: 42910356
Bytes allocated: 45002752                       |  Bytes allocated: 44711936
Bytes wasted: 2092396 (includes alignment, etc) |  Bytes wasted: 1801580 (includes alignment, etc)


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@115151 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Support/Allocator.cpp b/lib/Support/Allocator.cpp
index 90df262..02b45d8 100644
--- a/lib/Support/Allocator.cpp
+++ b/lib/Support/Allocator.cpp
@@ -44,6 +44,12 @@
 /// StartNewSlab - Allocate a new slab and move the bump pointers over into
 /// the new slab.  Modifies CurPtr and End.
 void BumpPtrAllocator::StartNewSlab() {
+  // If we allocated a big number of slabs already it's likely that we're going
+  // to allocate more. Increase slab size to reduce mallocs and possibly memory
+  // overhead. The factors are chosen conservatively to avoid overallocation.
+  if (BytesAllocated >= SlabSize * 128)
+    SlabSize *= 2;
+
   MemSlab *NewSlab = Allocator.Allocate(SlabSize);
   NewSlab->NextPtr = CurSlab;
   CurSlab = NewSlab;