Extend SROA to handle arrays accessed as homogeneous structs and vice versa.

This is a minor extension of SROA to handle a special case that is
important for some ARM NEON operations.  Some of the NEON intrinsics
return multiple values, which are handled as struct types containing
multiple elements of the same vector type.  The corresponding return
types declared in the arm_neon.h header have equivalent arrays.  We
need SROA to recognize that it can split up those arrays and structs
into separate vectors, even though they are not always accessed with
the same type.  SROA already handles loads and stores of an entire
alloca by using insertvalue/extractvalue to access the individual
pieces, and that code works the same regardless of whether the type
is a struct or an array.  So, all that needs to be done is to check
for compatible arrays and homogeneous structs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123381 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Scalar/ScalarReplAggregates.cpp b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
index ec298e7316d80aabd3796a6022354a7e874447a7..0ec095749f9eac7232ccd6c0beaba56f0de48540 100644 (file)
--- a/lib/Transforms/Scalar/ScalarReplAggregates.cpp
+++ b/lib/Transforms/Scalar/ScalarReplAggregates.cpp
@@ -1076,6 +1076,46 @@ void SROA::isSafeGEP(GetElementPtrInst *GEPI, AllocaInst *AI,
     MarkUnsafe(Info);
 }
 
+/// isHomogeneousAggregate - Check if type T is a struct or array containing
+/// elements of the same type (which is always true for arrays).  If so,
+/// return true with NumElts and EltTy set to the number of elements and the
+/// element type, respectively.
+static bool isHomogeneousAggregate(const Type *T, unsigned &NumElts,
+                                   const Type *&EltTy) {
+  if (const ArrayType *AT = dyn_cast<ArrayType>(T)) {
+    NumElts = AT->getNumElements();
+    EltTy = AT->getElementType();
+    return true;
+  }
+  if (const StructType *ST = dyn_cast<StructType>(T)) {
+    NumElts = ST->getNumContainedTypes();
+    EltTy = ST->getContainedType(0);
+    for (unsigned n = 1; n < NumElts; ++n) {
+      if (ST->getContainedType(n) != EltTy)
+        return false;
+    }
+    return true;
+  }
+  return false;
+}
+
+/// isCompatibleAggregate - Check if T1 and T2 are either the same type or are
+/// "homogeneous" aggregates with the same element type and number of elements.
+static bool isCompatibleAggregate(const Type *T1, const Type *T2) {
+  if (T1 == T2)
+    return true;
+
+  unsigned NumElts1, NumElts2;
+  const Type *EltTy1, *EltTy2;
+  if (isHomogeneousAggregate(T1, NumElts1, EltTy1) &&
+      isHomogeneousAggregate(T2, NumElts2, EltTy2) &&
+      NumElts1 == NumElts2 &&
+      EltTy1 == EltTy2)
+    return true;
+
+  return false;
+}
+
 /// isSafeMemAccess - Check if a load/store/memcpy operates on the entire AI
 /// alloca or has an offset and size that corresponds to a component element
 /// within it.  The offset checked here may have been formed from a GEP with a
@@ -1085,20 +1125,23 @@ void SROA::isSafeMemAccess(AllocaInst *AI, uint64_t Offset, uint64_t MemSize,
                            AllocaInfo &Info) {
   // Check if this is a load/store of the entire alloca.
   if (Offset == 0 && MemSize == TD->getTypeAllocSize(AI->getAllocatedType())) {
-    bool UsesAggregateType = (MemOpType == AI->getAllocatedType());
-    // This is safe for MemIntrinsics (where MemOpType is 0), integer types
-    // (which are essentially the same as the MemIntrinsics, especially with
-    // regard to copying padding between elements), or references using the
-    // aggregate type of the alloca.
-    if (!MemOpType || MemOpType->isIntegerTy() || UsesAggregateType) {
-      if (!UsesAggregateType) {
-        if (isStore)
-          Info.isMemCpyDst = true;
-        else
-          Info.isMemCpySrc = true;
-      }
+    // This can be safe for MemIntrinsics (where MemOpType is 0) and integer
+    // loads/stores (which are essentially the same as the MemIntrinsics with
+    // regard to copying padding between elements).  But, if an alloca is
+    // flagged as both a source and destination of such operations, we'll need
+    // to check later for padding between elements.
+    if (!MemOpType || MemOpType->isIntegerTy()) {
+      if (isStore)
+        Info.isMemCpyDst = true;
+      else
+        Info.isMemCpySrc = true;
       return;
     }
+    // This is also safe for references using a type that is compatible with
+    // the type of the alloca, so that loads/stores can be rewritten using
+    // insertvalue/extractvalue.
+    if (isCompatibleAggregate(MemOpType, AI->getAllocatedType()))
+      return;
   }
   // Check if the offset/size correspond to a component within the alloca type.
   const Type *T = AI->getAllocatedType();
@@ -1159,7 +1202,7 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
       // address operand will be updated, so nothing else needs to be done.
     } else if (LoadInst *LI = dyn_cast<LoadInst>(User)) {
       const Type *LIType = LI->getType();
-      if (LIType == AI->getAllocatedType()) {
+      if (isCompatibleAggregate(LIType, AI->getAllocatedType())) {
         // Replace:
         //   %res = load { i32, i32 }* %alloc
         // with:
@@ -1184,7 +1227,7 @@ void SROA::RewriteForScalarRepl(Instruction *I, AllocaInst *AI, uint64_t Offset,
     } else if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
       Value *Val = SI->getOperand(0);
       const Type *SIType = Val->getType();
-      if (SIType == AI->getAllocatedType()) {
+      if (isCompatibleAggregate(SIType, AI->getAllocatedType())) {
         // Replace:
         //   store { i32, i32 } %val, { i32, i32 }* %alloc
         // with:

diff --git a/test/Transforms/ScalarRepl/copy-aggregate.ll b/test/Transforms/ScalarRepl/copy-aggregate.ll
index 6e01aa3c51840662a5beba95f2c343bbf2daa55d..997da4bdb2fc049fc1c164a88cdee37b231704a8 100644 (file)
--- a/test/Transforms/ScalarRepl/copy-aggregate.ll
+++ b/test/Transforms/ScalarRepl/copy-aggregate.ll
@@ -76,10 +76,33 @@ entry:
   %var = alloca %arr, align 4
   %vari8 = bitcast %arr* %var to i8*
   %pi8 = bitcast %arr* %p to i8*
-  call void @llvm.memcpy.i32(i8* %vari8, i8* %pi8, i32 16, i32 4)
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %vari8, i8* %pi8, i32 16, i32 4, i1 false)
   %qi8 = bitcast %arr* %q to i8*
-  call void @llvm.memcpy.i32(i8* %qi8, i8* %vari8, i32 16, i32 4)
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %qi8, i8* %vari8, i32 16, i32 4, i1 false)
   ret void
 }
 
-declare void @llvm.memcpy.i32(i8* nocapture, i8* nocapture, i32, i32) nounwind
+;; Check that an array alloca can be split up when it is also accessed with
+;; a load or store as a homogeneous structure with the same element type and
+;; number of elements as the array.
+%homogeneous = type { <8 x i16>, <8 x i16>, <8 x i16> }
+%wrapped_array = type { [3 x <8 x i16>] }
+define void @test6(i8* %p, %wrapped_array* %arr) {
+entry:
+; CHECK: test6
+; CHECK: store <8 x i16>
+; CHECK: store <8 x i16>
+; CHECK: store <8 x i16>
+  %var = alloca %wrapped_array, align 16
+  %res = call %homogeneous @test6callee(i8* %p)
+  %varcast = bitcast %wrapped_array* %var to %homogeneous*
+  store %homogeneous %res, %homogeneous* %varcast
+  %tmp1 = bitcast %wrapped_array* %arr to i8*
+  %tmp2 = bitcast %wrapped_array* %var to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %tmp1, i8* %tmp2, i32 48, i32 16, i1 false)
+  ret void
+}
+
+declare %homogeneous @test6callee(i8* nocapture) nounwind
+
+declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind