Pass sret arguments through the stack instead of through registers in Sparc backend. It makes the code generated more compliant with the sparc32 ABI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@124030 91177308-0d34-0410-b5e6-96231b3b80d8

4 files changed, 111 insertions, 4 deletions

diff --git a/lib/Target/Sparc/SparcCallingConv.td b/lib/Target/Sparc/SparcCallingConv.td
index ed610a3ba6..856f87ad1d 100644
--- a/lib/Target/Sparc/SparcCallingConv.td
+++ b/lib/Target/Sparc/SparcCallingConv.td
@@ -24,6 +24,8 @@ def RetCC_Sparc32 : CallingConv<[
 
 // Sparc 32-bit C Calling convention.
 def CC_Sparc32 : CallingConv<[
+  //Custom assign SRet to [sp+64].
+  CCIfSRet<CCCustom<"CC_Sparc_Assign_SRet">>,
   // i32 f32 arguments get passed in integer registers if there is space.
   CCIfType<[i32, f32], CCAssignToReg<[I0, I1, I2, I3, I4, I5]>>,
   // f64 arguments are split and passed through registers or through stack.
diff --git a/lib/Target/Sparc/SparcISelLowering.cpp b/lib/Target/Sparc/SparcISelLowering.cpp
index 2767862dc3..515d9d7252 100644
--- a/lib/Target/Sparc/SparcISelLowering.cpp
+++ b/lib/Target/Sparc/SparcISelLowering.cpp
@@ -33,6 +33,19 @@ using namespace llvm;
 // Calling Convention Implementation
 //===----------------------------------------------------------------------===//
 
+static bool CC_Sparc_Assign_SRet(unsigned &ValNo, MVT &ValVT,
+                                 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
+                                 ISD::ArgFlagsTy &ArgFlags, CCState &State)
+{
+  assert (ArgFlags.isSRet());
+
+  //Assign SRet argument
+  State.addLoc(CCValAssign::getCustomMem(ValNo, ValVT,
+                                         0,
+                                         LocVT, LocInfo));
+  return true;
+}
+
 static bool CC_Sparc_Assign_f64(unsigned &ValNo, MVT &ValVT,
                                 MVT &LocVT, CCValAssign::LocInfo &LocInfo,
                                 ISD::ArgFlagsTy &ArgFlags, CCState &State)
@@ -70,6 +83,8 @@ SparcTargetLowering::LowerReturn(SDValue Chain,
                                  const SmallVectorImpl<SDValue> &OutVals,
                                  DebugLoc dl, SelectionDAG &DAG) const {
 
+  MachineFunction &MF = DAG.getMachineFunction();
+
   // CCValAssign - represent the assignment of the return value to locations.
   SmallVector<CCValAssign, 16> RVLocs;
 
@@ -82,10 +97,10 @@ SparcTargetLowering::LowerReturn(SDValue Chain,
 
   // If this is the first return lowered for this function, add the regs to the
   // liveout set for the function.
-  if (DAG.getMachineFunction().getRegInfo().liveout_empty()) {
+  if (MF.getRegInfo().liveout_empty()) {
     for (unsigned i = 0; i != RVLocs.size(); ++i)
       if (RVLocs[i].isRegLoc())
-        DAG.getMachineFunction().getRegInfo().addLiveOut(RVLocs[i].getLocReg());
+        MF.getRegInfo().addLiveOut(RVLocs[i].getLocReg());
   }
 
   SDValue Flag;
@@ -101,6 +116,18 @@ SparcTargetLowering::LowerReturn(SDValue Chain,
     // Guarantee that all emitted copies are stuck together with flags.
     Flag = Chain.getValue(1);
   }
+  // If the function returns a struct, copy the SRetReturnReg to I0
+  if (MF.getFunction()->hasStructRetAttr()) {
+    SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
+    unsigned Reg = SFI->getSRetReturnReg();
+    if (!Reg)
+      llvm_unreachable("sret virtual register not created in the entry block");
+    SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
+    Chain = DAG.getCopyToReg(Chain, dl, SP::I0, Val, Flag);
+    Flag = Chain.getValue(1);
+    if (MF.getRegInfo().liveout_empty())
+      MF.getRegInfo().addLiveOut(SP::I0);
+  }
 
   if (Flag.getNode())
     return DAG.getNode(SPISD::RET_FLAG, dl, MVT::Other, Chain, Flag);
@@ -134,6 +161,17 @@ SparcTargetLowering::LowerFormalArguments(SDValue Chain,
   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
     CCValAssign &VA = ArgLocs[i];
 
+    if (i == 0  && Ins[i].Flags.isSRet()) {
+      //Get SRet from [%fp+64]
+      int FrameIdx = MF.getFrameInfo()->CreateFixedObject(4, 64, true);
+      SDValue FIPtr = DAG.getFrameIndex(FrameIdx, MVT::i32);
+      SDValue Arg = DAG.getLoad(MVT::i32, dl, Chain, FIPtr,
+                                MachinePointerInfo(),
+                                false, false, 0);
+      InVals.push_back(Arg);
+      continue;
+    }
+
     if (VA.isRegLoc()) {
       EVT RegVT = VA.getLocVT();
 
@@ -244,6 +282,18 @@ SparcTargetLowering::LowerFormalArguments(SDValue Chain,
     InVals.push_back(Load);
   }
 
+  if (MF.getFunction()->hasStructRetAttr()) {
+    //Copy the SRet Argument to SRetReturnReg
+    SparcMachineFunctionInfo *SFI = MF.getInfo<SparcMachineFunctionInfo>();
+    unsigned Reg = SFI->getSRetReturnReg();
+    if (!Reg) {
+      Reg = MF.getRegInfo().createVirtualRegister(&SP::IntRegsRegClass);
+      SFI->setSRetReturnReg(Reg);
+    }
+    SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[0]);
+    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
+  }
+
   // Store remaining ArgRegs to the stack if this is a varargs function.
   if (isVarArg) {
     static const unsigned ArgRegs[] = {
@@ -374,6 +424,18 @@ SparcTargetLowering::LowerCall(SDValue Chain, SDValue Callee,
       break;
     }
 
+    if (Flags.isSRet()) {
+      assert(VA.needsCustom());
+      // store SRet argument in %sp+64
+      SDValue StackPtr = DAG.getRegister(SP::O6, MVT::i32);
+      SDValue PtrOff = DAG.getIntPtrConstant(64);
+      PtrOff = DAG.getNode(ISD::ADD, dl, MVT::i32, StackPtr, PtrOff);
+      MemOpChains.push_back(DAG.getStore(Chain, dl, Arg, PtrOff,
+                                         MachinePointerInfo(),
+                                         false, false, 0));
+      continue;
+    }
+
     if (VA.needsCustom()) {
       assert(VA.getLocVT() == MVT::f64);
 
diff --git a/lib/Target/Sparc/SparcMachineFunctionInfo.h b/lib/Target/Sparc/SparcMachineFunctionInfo.h
index e34c131281..0b74308eb0 100644
--- a/lib/Target/Sparc/SparcMachineFunctionInfo.h
+++ b/lib/Target/Sparc/SparcMachineFunctionInfo.h
@@ -24,16 +24,23 @@ namespace llvm {
     /// VarArgsFrameOffset - Frame offset to start of varargs area.
     int VarArgsFrameOffset;
 
+    /// SRetReturnReg - Holds the virtual register into which the sret
+    /// argument is passed.
+    unsigned SRetReturnReg;
   public:
-    SparcMachineFunctionInfo() : GlobalBaseReg(0), VarArgsFrameOffset(0) {}
+    SparcMachineFunctionInfo()
+      : GlobalBaseReg(0), VarArgsFrameOffset(0), SRetReturnReg(0) {}
     explicit SparcMachineFunctionInfo(MachineFunction &MF)
-      : GlobalBaseReg(0), VarArgsFrameOffset(0) {}
+      : GlobalBaseReg(0), VarArgsFrameOffset(0), SRetReturnReg(0) {}
 
     unsigned getGlobalBaseReg() const { return GlobalBaseReg; }
     void setGlobalBaseReg(unsigned Reg) { GlobalBaseReg = Reg; }
 
     int getVarArgsFrameOffset() const { return VarArgsFrameOffset; }
     void setVarArgsFrameOffset(int Offset) { VarArgsFrameOffset = Offset; }
+
+    unsigned getSRetReturnReg() const { return SRetReturnReg; }
+    void setSRetReturnReg(unsigned Reg) { SRetReturnReg = Reg; }
   };
 }
 
diff --git a/test/CodeGen/SPARC/2011-01-22-SRet.ll b/test/CodeGen/SPARC/2011-01-22-SRet.ll
new file mode 100644
index 0000000000..2f684b009c
--- /dev/null
+++ b/test/CodeGen/SPARC/2011-01-22-SRet.ll
@@ -0,0 +1,36 @@
+;RUN: llc -march=sparc < %s | FileCheck %s
+
+%struct.foo_t = type { i32, i32, i32 }
+
+define weak void @make_foo(%struct.foo_t* noalias sret %agg.result, i32 %a, i32 %b, i32 %c) nounwind {
+entry:
+;CHECK: make_foo
+;CHECK: ld [%fp+64], {{.+}}
+;CHECK: or {{.+}}, {{.+}}, %i0
+;CHECK: ret
+  %0 = getelementptr inbounds %struct.foo_t* %agg.result, i32 0, i32 0
+  store i32 %a, i32* %0, align 4
+  %1 = getelementptr inbounds %struct.foo_t* %agg.result, i32 0, i32 1
+  store i32 %b, i32* %1, align 4
+  %2 = getelementptr inbounds %struct.foo_t* %agg.result, i32 0, i32 2
+  store i32 %c, i32* %2, align 4
+  ret void
+}
+
+define i32 @test() nounwind {
+entry:
+;CHECK: test
+;CHECK: st {{.+}}, [%sp+64]
+;CHECK: make_foo
+  %f = alloca %struct.foo_t, align 8
+  call void @make_foo(%struct.foo_t* noalias sret %f, i32 10, i32 20, i32 30) nounwind
+  %0 = getelementptr inbounds %struct.foo_t* %f, i32 0, i32 0
+  %1 = load i32* %0, align 8
+  %2 = getelementptr inbounds %struct.foo_t* %f, i32 0, i32 1
+  %3 = load i32* %2, align 4
+  %4 = getelementptr inbounds %struct.foo_t* %f, i32 0, i32 2
+  %5 = load i32* %4, align 8
+  %6 = add nsw i32 %3, %1
+  %7 = add nsw i32 %6, %5
+  ret i32 %7
+}