Enable optimization of sin / cos pair into call to __sincos_stret for iOS7+.

rdar://12856873
Patch by Evan Cheng, with a fix for rdar://13209539 by Tilmann Scheller

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@193942 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index 2de2dfa0ca78ad42b9b67fb398a51edfbfcac2ec..9dee9314f68ce35ce7ffb521105468da192b9dee 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -869,6 +869,18 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
       setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
     }
   }
+      
+  // Combine sin / cos into one node or libcall if possible.
+  if (Subtarget->hasSinCos()) {
+    setLibcallName(RTLIB::SINCOS_F32, "sincosf");
+    setLibcallName(RTLIB::SINCOS_F64, "sincos");
+    if (Subtarget->getTargetTriple().getOS() == Triple::IOS) {
+      // For iOS, we don't want to the normal expansion of a libcall to
+      // sincos. We want to issue a libcall to __sincos_stret.
+      setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
+      setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
+    }
+  }
 
   // We have target-specific dag combine patterns for the following nodes:
   // ARMISD::VMOVRRD  - No need to call setTargetDAGCombine
@@ -5950,6 +5962,70 @@ static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
                      Op.getOperand(1), Op.getOperand(2));
 }
 
+SDValue ARMTargetLowering::LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const {
+  assert(Subtarget->isTargetDarwin());
+
+  // For iOS, we want to call an alternative entry point: __sincos_stret,
+  // return values are passed via sret.
+  SDLoc dl(Op);
+  SDValue Arg = Op.getOperand(0);
+  EVT ArgVT = Arg.getValueType();
+  Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+
+  MachineFrameInfo *FrameInfo = DAG.getMachineFunction().getFrameInfo();
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+  // Pair of floats / doubles used to pass the result.
+  StructType *RetTy = StructType::get(ArgTy, ArgTy, NULL);
+
+  // Create stack object for sret.
+  const uint64_t ByteSize = TLI.getDataLayout()->getTypeAllocSize(RetTy);
+  const unsigned StackAlign = TLI.getDataLayout()->getPrefTypeAlignment(RetTy);
+  int FrameIdx = FrameInfo->CreateStackObject(ByteSize, StackAlign, false);
+  SDValue SRet = DAG.getFrameIndex(FrameIdx, TLI.getPointerTy());
+
+  ArgListTy Args;
+  ArgListEntry Entry;
+
+  Entry.Node = SRet;
+  Entry.Ty = RetTy->getPointerTo();
+  Entry.isSExt = false;
+  Entry.isZExt = false;
+  Entry.isSRet = true;
+  Args.push_back(Entry);
+
+  Entry.Node = Arg;
+  Entry.Ty = ArgTy;
+  Entry.isSExt = false;
+  Entry.isZExt = false;
+  Args.push_back(Entry);
+
+  const char *LibcallName  = (ArgVT == MVT::f64)
+  ? "__sincos_stret" : "__sincosf_stret";
+  SDValue Callee = DAG.getExternalSymbol(LibcallName, getPointerTy());
+
+  TargetLowering::
+  CallLoweringInfo CLI(DAG.getEntryNode(), Type::getVoidTy(*DAG.getContext()),
+                       false, false, false, false, 0,
+                       CallingConv::C, /*isTaillCall=*/false,
+                       /*doesNotRet=*/false, /*isReturnValueUsed*/false,
+                       Callee, Args, DAG, dl);
+  std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
+
+  SDValue LoadSin = DAG.getLoad(ArgVT, dl, CallResult.second, SRet,
+                                MachinePointerInfo(), false, false, false, 0);
+
+  // Address of cos field.
+  SDValue Add = DAG.getNode(ISD::ADD, dl, getPointerTy(), SRet,
+                            DAG.getIntPtrConstant(ArgVT.getStoreSize()));
+  SDValue LoadCos = DAG.getLoad(ArgVT, dl, LoadSin.getValue(1), Add,
+                                MachinePointerInfo(), false, false, false, 0);
+
+  SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
+  return DAG.getNode(ISD::MERGE_VALUES, dl, Tys,
+                     LoadSin.getValue(0), LoadCos.getValue(0));
+}
+
 static SDValue LowerAtomicLoadStore(SDValue Op, SelectionDAG &DAG) {
   // Monotonic load/store is legal for all targets
   if (cast<AtomicSDNode>(Op)->getOrdering() <= Monotonic)
@@ -6081,6 +6157,7 @@ SDValue ARMTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::SUBE:          return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
   case ISD::ATOMIC_LOAD:
   case ISD::ATOMIC_STORE:  return LowerAtomicLoadStore(Op, DAG);
+  case ISD::FSINCOS:       return LowerFSINCOS(Op, DAG);
   case ISD::SDIVREM:
   case ISD::UDIVREM:       return LowerDivRem(Op, DAG);
   }

diff --git a/lib/Target/ARM/ARMISelLowering.h b/lib/Target/ARM/ARMISelLowering.h
index 3c80334dc4a20b824f785b1981b286547181f5ac..f195381538be28c3067c89cf6468d81c61f72e10 100644 (file)
--- a/lib/Target/ARM/ARMISelLowering.h
+++ b/lib/Target/ARM/ARMISelLowering.h
@@ -448,6 +448,7 @@ namespace llvm {
                             const ARMSubtarget *ST) const;
     SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                               const ARMSubtarget *ST) const;
+    SDValue LowerFSINCOS(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerDivRem(SDValue Op, SelectionDAG &DAG) const;
 
     /// isFMAFasterThanFMulAndFAdd - Return true if an FMA operation is faster

diff --git a/lib/Target/ARM/ARMSubtarget.cpp b/lib/Target/ARM/ARMSubtarget.cpp
index 392b6f590eb5cdeb61724bb007d0cf2bdc671c54..f59e522195c2193e121b5641a68fdb6050da087f 100644 (file)
--- a/lib/Target/ARM/ARMSubtarget.cpp
+++ b/lib/Target/ARM/ARMSubtarget.cpp
@@ -281,6 +281,11 @@ unsigned ARMSubtarget::getMispredictionPenalty() const {
   return SchedModel->MispredictPenalty;
 }
 
+bool ARMSubtarget::hasSinCos() const {
+  return getTargetTriple().getOS() == Triple::IOS &&
+    !getTargetTriple().isOSVersionLT(7, 0);
+}
+
 bool ARMSubtarget::enablePostRAScheduler(
            CodeGenOpt::Level OptLevel,
            TargetSubtargetInfo::AntiDepBreakMode& Mode,

diff --git a/lib/Target/ARM/ARMSubtarget.h b/lib/Target/ARM/ARMSubtarget.h
index 0eb0a4463e45e67545f8a25e0fb4bfd83891850d..aab93f1d6c0cc924bd07f68ec303f7853f5509a2 100644 (file)
--- a/lib/Target/ARM/ARMSubtarget.h
+++ b/lib/Target/ARM/ARMSubtarget.h
@@ -330,6 +330,10 @@ public:
   const std::string & getCPUString() const { return CPUString; }
 
   unsigned getMispredictionPenalty() const;
+  
+  /// This function returns true if the target has sincos() routine in its
+  /// compiler runtime or math libraries.
+  bool hasSinCos() const;
 
   /// enablePostRAScheduler - True at 'More' optimization.
   bool enablePostRAScheduler(CodeGenOpt::Level OptLevel,

diff --git a/test/CodeGen/ARM/sincos.ll b/test/CodeGen/ARM/sincos.ll
new file mode 100644 (file)
index 0000000..30b2664
--- /dev/null
+++ b/test/CodeGen/ARM/sincos.ll
@@ -0,0 +1,38 @@
+; RUN: llc < %s -mtriple=armv7-apple-ios6 -mcpu=cortex-a8 | FileCheck %s --check-prefix=NOOPT
+; RUN: llc < %s -mtriple=armv7-apple-ios7 -mcpu=cortex-a8 | FileCheck %s --check-prefix=SINCOS
+
+; Combine sin / cos into a single call.
+; rdar://12856873
+
+define float @test1(float %x) nounwind {
+entry:
+; SINCOS-LABEL: test1:
+; SINCOS: bl ___sincosf_stret
+
+; NOOPT-LABEL: test1:
+; NOOPT: bl _sinf
+; NOOPT: bl _cosf
+  %call = tail call float @sinf(float %x) nounwind readnone
+  %call1 = tail call float @cosf(float %x) nounwind readnone
+  %add = fadd float %call, %call1
+  ret float %add
+}
+
+define double @test2(double %x) nounwind {
+entry:
+; SINCOS-LABEL: test2:
+; SINCOS: bl ___sincos_stret
+
+; NOOPT-LABEL: test2:
+; NOOPT: bl _sin
+; NOOPT: bl _cos
+  %call = tail call double @sin(double %x) nounwind readnone
+  %call1 = tail call double @cos(double %x) nounwind readnone
+  %add = fadd double %call, %call1
+  ret double %add
+}
+
+declare float  @sinf(float) readonly
+declare double @sin(double) readonly
+declare float @cosf(float) readonly
+declare double @cos(double) readonly