HeapAllocSRoA also needs to check if malloc array size can be computed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@84288 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 66 insertions, 36 deletions

diff --git a/lib/Transforms/IPO/GlobalOpt.cpp b/lib/Transforms/IPO/GlobalOpt.cpp
index 0d9818fdfa..3d10649c95 100644
--- a/lib/Transforms/IPO/GlobalOpt.cpp
+++ b/lib/Transforms/IPO/GlobalOpt.cpp
@@ -1586,6 +1586,8 @@ static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV,
                << " BITCAST = " << *BCI << '\n');
   const Type* MAT = getMallocAllocatedType(CI);
   const StructType *STy = cast<StructType>(MAT);
+  Value* ArraySize = getMallocArraySize(CI, Context, TD);
+  assert(ArraySize && "not a malloc whose array size can be determined");
 
   // There is guaranteed to be at least one use of the malloc (storing
   // it into GV).  If there are other uses, change them to be uses of
@@ -1610,8 +1612,8 @@ static GlobalVariable *PerformHeapAllocSRoA(GlobalVariable *GV,
                          GV->isThreadLocal());
     FieldGlobals.push_back(NGV);
     
-    Value *NMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context), FieldTy,
-                                        getMallocArraySize(CI, Context, TD),
+    Value *NMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
+                                        FieldTy, ArraySize,
                                         BCI->getName() + ".f" + Twine(FieldNo));
     FieldMallocs.push_back(NMI);
     new StoreInst(NMI, NGV, BCI);
@@ -1892,6 +1894,7 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
   // This eliminates dynamic allocation, avoids an indirection accessing the
   // data, and exposes the resultant global to further GlobalOpt.
   Value *NElems = getMallocArraySize(CI, Context, TD);
+  // We cannot optimize the malloc if we cannot determine malloc array size.
   if (NElems) {
     if (ConstantInt *NElements = dyn_cast<ConstantInt>(NElems))
       // Restrict this transformation to only working on small allocations
@@ -1902,42 +1905,43 @@ static bool TryToOptimizeStoreOfMallocToGlobal(GlobalVariable *GV,
         GVI = OptimizeGlobalAddressOfMalloc(GV, CI, BCI, Context, TD);
         return true;
       }
-  }
   
-  // If the allocation is an array of structures, consider transforming this
-  // into multiple malloc'd arrays, one for each field.  This is basically
-  // SRoA for malloc'd memory.
-
-  // If this is an allocation of a fixed size array of structs, analyze as a
-  // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
-  if (!isArrayMalloc(CI, Context, TD))
-    if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
-      AllocTy = AT->getElementType();
-  
-  if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
-    // This the structure has an unreasonable number of fields, leave it
-    // alone.
-    if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
-        AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, BCI)) {
-
-      // If this is a fixed size array, transform the Malloc to be an alloc of
-      // structs.  malloc [100 x struct],1 -> malloc struct, 100
-      if (const ArrayType *AT = dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
-        Value* NumElements = ConstantInt::get(Type::getInt32Ty(Context),
-                                              AT->getNumElements());
-        Value* NewMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
-                                              AllocSTy, NumElements,
-                                              BCI->getName());
-        Value *Cast = new BitCastInst(NewMI, getMallocType(CI), "tmp", CI);
-        BCI->replaceAllUsesWith(Cast);
-        BCI->eraseFromParent();
-        CI->eraseFromParent();
-        BCI = cast<BitCastInst>(NewMI);
-        CI = extractMallocCallFromBitCast(NewMI);
-      }
+    // If the allocation is an array of structures, consider transforming this
+    // into multiple malloc'd arrays, one for each field.  This is basically
+    // SRoA for malloc'd memory.
+
+    // If this is an allocation of a fixed size array of structs, analyze as a
+    // variable size array.  malloc [100 x struct],1 -> malloc struct, 100
+    if (!isArrayMalloc(CI, Context, TD))
+      if (const ArrayType *AT = dyn_cast<ArrayType>(AllocTy))
+        AllocTy = AT->getElementType();
+  
+    if (const StructType *AllocSTy = dyn_cast<StructType>(AllocTy)) {
+      // This the structure has an unreasonable number of fields, leave it
+      // alone.
+      if (AllocSTy->getNumElements() <= 16 && AllocSTy->getNumElements() != 0 &&
+          AllGlobalLoadUsesSimpleEnoughForHeapSRA(GV, BCI)) {
+
+        // If this is a fixed size array, transform the Malloc to be an alloc of
+        // structs.  malloc [100 x struct],1 -> malloc struct, 100
+        if (const ArrayType *AT =
+                              dyn_cast<ArrayType>(getMallocAllocatedType(CI))) {
+          Value* NumElements = ConstantInt::get(Type::getInt32Ty(Context),
+                                                AT->getNumElements());
+          Value* NewMI = CallInst::CreateMalloc(CI, TD->getIntPtrType(Context),
+                                                AllocSTy, NumElements,
+                                                BCI->getName());
+          Value *Cast = new BitCastInst(NewMI, getMallocType(CI), "tmp", CI);
+          BCI->replaceAllUsesWith(Cast);
+          BCI->eraseFromParent();
+          CI->eraseFromParent();
+          BCI = cast<BitCastInst>(NewMI);
+          CI = extractMallocCallFromBitCast(NewMI);
+        }
       
-      GVI = PerformHeapAllocSRoA(GV, CI, BCI, Context, TD);
-      return true;
+        GVI = PerformHeapAllocSRoA(GV, CI, BCI, Context, TD);
+        return true;
+      }
     }
   }
   
diff --git a/test/Transforms/GlobalOpt/2009-11-16-BrokenPerformHeapAllocSRoA.ll b/test/Transforms/GlobalOpt/2009-11-16-BrokenPerformHeapAllocSRoA.ll
new file mode 100644
index 0000000000..54e8f90979
--- /dev/null
+++ b/test/Transforms/GlobalOpt/2009-11-16-BrokenPerformHeapAllocSRoA.ll
@@ -0,0 +1,26 @@
+; RUN: opt < %s -globalopt -S | FileCheck %s
+
+target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128"
+target triple = "x86_64-apple-darwin10.0"
+
+%struct.hashheader = type { i16, i16, i16, i16, i16, i16, i32, i32, i32, i32, i32, i32, i32, i32, i32, [5 x i8], [13 x i8], i8, i8, i8, [228 x i16], [228 x i8], [228 x i8], [228 x i8], [228 x i8], [228 x i8], [228 x i8], [128 x i8], [100 x [11 x i8]], [100 x i32], [100 x i32], i16 }
+%struct.strchartype = type { i8*, i8*, i8* }
+
+@hashheader = internal global %struct.hashheader zeroinitializer, align 32 ; <%struct.hashheader*> [#uses=1]
+@chartypes = internal global %struct.strchartype* null ; <%struct.strchartype**> [#uses=1]
+; CHECK-NOT: @hashheader
+; CHECK-NOT: @chartypes
+
+; based on linit in office-ispell
+define void @test() nounwind ssp {
+  %1 = load i32* getelementptr inbounds (%struct.hashheader* @hashheader, i64 0, i32 13), align 8 ; <i32> [#uses=1]
+  %2 = sext i32 %1 to i64                         ; <i64> [#uses=1]
+  %3 = mul i64 %2, ptrtoint (%struct.strchartype* getelementptr (%struct.strchartype* null, i64 1) to i64) ; <i64> [#uses=1]
+  %4 = tail call i8* @malloc(i64 %3)              ; <i8*> [#uses=1]
+; CHECK: call i8* @malloc(i64
+  %5 = bitcast i8* %4 to %struct.strchartype*     ; <%struct.strchartype*> [#uses=1]
+  store %struct.strchartype* %5, %struct.strchartype** @chartypes, align 8
+  ret void
+}
+
+declare noalias i8* @malloc(i64)