__sincosf_stret returns sinf / cosf in bits 0:31 and 32:63 of xmm0, not in

xmm0 / xmm1.

rdar://13599493


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179141 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 22 insertions, 6 deletions

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index c4f1176..95e7f85 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -12319,7 +12319,8 @@ SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
   assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
 
   // For MacOSX, we want to call an alternative entry point: __sincos_stret,
-  // which returns the values in two XMM registers.
+  // which returns the values as { float, float } (in XMM0) or
+  // { double, double } (which is returned in XMM0, XMM1).
   DebugLoc dl = Op.getDebugLoc();
   SDValue Arg = Op.getOperand(0);
   EVT ArgVT = Arg.getValueType();
@@ -12334,14 +12335,16 @@ SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
   Entry.isZExt = false;
   Args.push_back(Entry);
 
+  bool isF64 = ArgVT == MVT::f64;
   // Only optimize x86_64 for now. i386 is a bit messy. For f32,
   // the small struct {f32, f32} is returned in (eax, edx). For f64,
   // the results are returned via SRet in memory.
-  const char *LibcallName = (ArgVT == MVT::f64)
-    ? "__sincos_stret" : "__sincosf_stret";
+  const char *LibcallName =  isF64 ? "__sincos_stret" : "__sincosf_stret";
   SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
 
-  StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
+  Type *RetTy = isF64
+    ? (Type*)StructType::get(ArgTy, ArgTy, NULL)
+    : (Type*)VectorType::get(ArgTy, 4);
   TargetLowering::
     CallLoweringInfo CLI(DAG.getEntryNode(), RetTy,
                          false, false, false, false, 0,
@@ -12349,7 +12352,18 @@ SDValue X86TargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
                          /*doesNotRet=*/false, /*isReturnValueUsed*/true,
                          Callee, Args, DAG, dl);
   std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
-  return CallResult.first;
+
+  if (isF64)
+    // Returned in xmm0 and xmm1.
+    return CallResult.first;
+
+  // Returned in bits 0:31 and 32:64 xmm0.
+  SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
+                               CallResult.first, DAG.getIntPtrConstant(0));
+  SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
+                               CallResult.first, DAG.getIntPtrConstant(1));
+  SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
+  return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
 }
 
 /// LowerOperation - Provide custom lowering hooks for some operations.
diff --git a/test/CodeGen/X86/sincos-opt.ll b/test/CodeGen/X86/sincos-opt.ll
index f364d1f..333c466 100644
--- a/test/CodeGen/X86/sincos-opt.ll
+++ b/test/CodeGen/X86/sincos-opt.ll
@@ -4,6 +4,7 @@
 
 ; Combine sin / cos into a single call.
 ; rdar://13087969
+; rdar://13599493
 
 define float @test1(float %x) nounwind {
 entry:
@@ -14,7 +15,8 @@ entry:
 
 ; OSX_SINCOS: test1:
 ; OSX_SINCOS: callq ___sincosf_stret
-; OSX_SINCOS: addss %xmm1, %xmm0
+; OSX_SINCOS: pshufd $1, %xmm0, %xmm1
+; OSX_SINCOS: addss %xmm0, %xmm1
 
 ; OSX_NOOPT: test1
 ; OSX_NOOPT: callq _cosf