From Kalle Raiskila:

"the bigstack patch for SPU, with testcase. It is essentially the patch committed as 97091, and reverted as 97099, but with the following additions:
-in vararg handling, registers are marked to be live, to not confuse the register scavenger
-function prologue and epilogue are not emitted, if the stack size is 16. 16 means it is empty - there is only the register scavenger emergency spill slot, which is not used as there is no stack."



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@99819 91177308-0d34-0410-b5e6-96231b3b80d8

5 files changed, 175 insertions, 14 deletions

diff --git a/lib/Target/CellSPU/SPU.h b/lib/Target/CellSPU/SPU.h
index c960974..c620812 100644
--- a/lib/Target/CellSPU/SPU.h
+++ b/lib/Target/CellSPU/SPU.h
@@ -66,9 +66,6 @@ namespace llvm {
   //! Predicate test for an unsigned 10-bit value
   /*!
     \param Value The input value to be tested
-
-    This predicate tests for an unsigned 10-bit value, returning the 10-bit value
-    as a short if true.
    */
   inline bool isU10Constant(short Value) {
     return (Value == (Value & 0x3ff));
@@ -90,6 +87,70 @@ namespace llvm {
     return (Value == (Value & 0x3ff));
   }
 
+  //! Predicate test for a signed 14-bit value
+  /*!
+    \param Value The input value to be tested
+   */
+  template<typename T>
+  inline bool isS14Constant(T Value);
+
+  template<>
+  inline bool isS14Constant<short>(short Value) {
+    return (Value >= -(1 << 13) && Value <= (1 << 13) - 1);
+  }
+
+  template<>
+  inline bool isS14Constant<int>(int Value) {
+    return (Value >= -(1 << 13) && Value <= (1 << 13) - 1);
+  }
+
+  template<>
+  inline bool isS14Constant<uint32_t>(uint32_t Value) {
+    return (Value <= ((1 << 13) - 1));
+  }
+
+  template<>
+  inline bool isS14Constant<int64_t>(int64_t Value) {
+    return (Value >= -(1 << 13) && Value <= (1 << 13) - 1);
+  }
+
+  template<>
+  inline bool isS14Constant<uint64_t>(uint64_t Value) {
+    return (Value <= ((1 << 13) - 1));
+  }
+
+  //! Predicate test for a signed 16-bit value
+  /*!
+    \param Value The input value to be tested
+   */
+  template<typename T>
+  inline bool isS16Constant(T Value);
+
+  template<>
+  inline bool isS16Constant<short>(short Value) {
+    return true;
+  }
+
+  template<>
+  inline bool isS16Constant<int>(int Value) {
+    return (Value >= -(1 << 15) && Value <= (1 << 15) - 1);
+  }
+
+  template<>
+  inline bool isS16Constant<uint32_t>(uint32_t Value) {
+    return (Value <= ((1 << 15) - 1));
+  }
+
+  template<>
+  inline bool isS16Constant<int64_t>(int64_t Value) {
+    return (Value >= -(1 << 15) && Value <= (1 << 15) - 1);
+  }
+
+  template<>
+  inline bool isS16Constant<uint64_t>(uint64_t Value) {
+    return (Value <= ((1 << 15) - 1));
+  }
+
   extern Target TheCellSPUTarget;
 
 }
diff --git a/lib/Target/CellSPU/SPUISelLowering.cpp b/lib/Target/CellSPU/SPUISelLowering.cpp
index e863ee3..e7c7881 100644
--- a/lib/Target/CellSPU/SPUISelLowering.cpp
+++ b/lib/Target/CellSPU/SPUISelLowering.cpp
@@ -1107,7 +1107,8 @@ SPUTargetLowering::LowerFormalArguments(SDValue Chain,
       VarArgsFrameIndex = MFI->CreateFixedObject(StackSlotSize, ArgOffset,
                                                  true, false);
       SDValue FIN = DAG.getFrameIndex(VarArgsFrameIndex, PtrVT);
-      SDValue ArgVal = DAG.getRegister(ArgRegs[ArgRegIdx], MVT::v16i8);
+      unsigned VReg = MF.addLiveIn(ArgRegs[ArgRegIdx], &SPU::R32CRegClass);
+      SDValue ArgVal = DAG.getRegister(VReg, MVT::v16i8);
       SDValue Store = DAG.getStore(Chain, dl, ArgVal, FIN, NULL, 0,
                                    false, false, 0);
       Chain = Store.getOperand(0);
diff --git a/lib/Target/CellSPU/SPURegisterInfo.cpp b/lib/Target/CellSPU/SPURegisterInfo.cpp
index 8c78bab..916ebf0 100644
--- a/lib/Target/CellSPU/SPURegisterInfo.cpp
+++ b/lib/Target/CellSPU/SPURegisterInfo.cpp
@@ -28,6 +28,7 @@
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineLocation.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/RegisterScavenging.h"
 #include "llvm/CodeGen/ValueTypes.h"
 #include "llvm/Target/TargetFrameInfo.h"
 #include "llvm/Target/TargetInstrInfo.h"
@@ -336,6 +337,7 @@ SPURegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj,
   MachineBasicBlock &MBB = *MI.getParent();
   MachineFunction &MF = *MBB.getParent();
   MachineFrameInfo *MFI = MF.getFrameInfo();
+  DebugLoc dl = II->getDebugLoc();
 
   while (!MI.getOperand(i).isFI()) {
     ++i;
@@ -364,11 +366,22 @@ SPURegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II, int SPAdj,
 
   // Replace the FrameIndex with base register with $sp (aka $r1)
   SPOp.ChangeToRegister(SPU::R1, false);
-  if (Offset > SPUFrameInfo::maxFrameOffset()
-      || Offset < SPUFrameInfo::minFrameOffset()) {
-    errs() << "Large stack adjustment ("
-         << Offset
-         << ") in SPURegisterInfo::eliminateFrameIndex.";
+
+  // if 'Offset' doesn't fit to the D-form instruction's
+  // immediate, convert the instruction to X-form
+  // if the instruction is not an AI (which takes a s10 immediate), assume
+  // it is a load/store that can take a s14 immediate
+  if ( (MI.getOpcode() == SPU::AIr32 && !isS10Constant(Offset))
+       || !isS14Constant(Offset) ) {
+    int newOpcode = convertDFormToXForm(MI.getOpcode());
+    unsigned tmpReg = findScratchRegister(II, RS, &SPU::R32CRegClass, SPAdj);
+    BuildMI(MBB, II, dl, TII.get(SPU::ILr32), tmpReg )
+        .addImm(Offset);
+    BuildMI(MBB, II, dl, TII.get(newOpcode), MI.getOperand(0).getReg())
+        .addReg(tmpReg, RegState::Kill)
+        .addReg(SPU::R1);
+    // remove the replaced D-form instruction
+    MBB.erase(II);
   } else {
     MO.ChangeToImmediate(Offset);
   }
@@ -423,6 +436,14 @@ void SPURegisterInfo::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
   MF.getRegInfo().setPhysRegUnused(SPU::R0);
   MF.getRegInfo().setPhysRegUnused(SPU::R1);
   MF.getRegInfo().setPhysRegUnused(SPU::R2);
+
+  MachineFrameInfo *MFI = MF.getFrameInfo(); 
+  const TargetRegisterClass *RC = &SPU::R32CRegClass;
+  RS->setScavengingFrameIndex(MFI->CreateStackObject(RC->getSize(),
+                                                     RC->getAlignment(),
+                                                     false));
+  
+  
 }
 
 void SPURegisterInfo::emitPrologue(MachineFunction &MF) const
@@ -448,7 +469,8 @@ void SPURegisterInfo::emitPrologue(MachineFunction &MF) const
   assert((FrameSize & 0xf) == 0
          && "SPURegisterInfo::emitPrologue: FrameSize not aligned");
 
-  if (FrameSize > 0 || MFI->hasCalls()) {
+  // the "empty" frame size is 16 - just the register scavenger spill slot
+  if (FrameSize > 16 || MFI->hasCalls()) {
     FrameSize = -(FrameSize + SPUFrameInfo::minStackSize());
     if (hasDebugInfo) {
       // Mark effective beginning of when frame pointer becomes valid.
@@ -467,7 +489,7 @@ void SPURegisterInfo::emitPrologue(MachineFunction &MF) const
       // Adjust $sp by required amout
       BuildMI(MBB, MBBI, dl, TII.get(SPU::AIr32), SPU::R1).addReg(SPU::R1)
         .addImm(FrameSize);
-    } else if (FrameSize <= (1 << 16) - 1 && FrameSize >= -(1 << 16)) {
+    } else if (isS16Constant(FrameSize)) {
       // Frame size can be loaded into ILr32n, so temporarily spill $r2 and use
       // $r2 to adjust $sp:
       BuildMI(MBB, MBBI, dl, TII.get(SPU::STQDr128), SPU::R2)
@@ -475,7 +497,7 @@ void SPURegisterInfo::emitPrologue(MachineFunction &MF) const
         .addReg(SPU::R1);
       BuildMI(MBB, MBBI, dl, TII.get(SPU::ILr32), SPU::R2)
         .addImm(FrameSize);
-      BuildMI(MBB, MBBI, dl, TII.get(SPU::STQDr32), SPU::R1)
+      BuildMI(MBB, MBBI, dl, TII.get(SPU::STQXr32), SPU::R1)
         .addReg(SPU::R2)
         .addReg(SPU::R1);
       BuildMI(MBB, MBBI, dl, TII.get(SPU::Ar32), SPU::R1)
@@ -549,7 +571,9 @@ SPURegisterInfo::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const
          "Can only insert epilog into returning blocks");
   assert((FrameSize & 0xf) == 0
          && "SPURegisterInfo::emitEpilogue: FrameSize not aligned");
-  if (FrameSize > 0 || MFI->hasCalls()) {
+
+  // the "empty" frame size is 16 - just the register scavenger spill slot
+  if (FrameSize > 16 || MFI->hasCalls()) {
     FrameSize = FrameSize + SPUFrameInfo::minStackSize();
     if (isS10Constant(FrameSize + LinkSlotOffset)) {
       // Reload $lr, adjust $sp by required amount
@@ -574,7 +598,7 @@ SPURegisterInfo::emitEpilogue(MachineFunction &MF, MachineBasicBlock &MBB) const
         .addReg(SPU::R2);
       BuildMI(MBB, MBBI, dl, TII.get(SPU::LQDr128), SPU::R0)
         .addImm(16)
-        .addReg(SPU::R2);
+        .addReg(SPU::R1);
       BuildMI(MBB, MBBI, dl, TII.get(SPU::SFIr32), SPU::R2).
         addReg(SPU::R2)
         .addImm(16);
@@ -618,4 +642,43 @@ SPURegisterInfo::getDwarfRegNum(unsigned RegNum, bool isEH) const {
   return SPUGenRegisterInfo::getDwarfRegNumFull(RegNum, 0);
 }
 
+int 
+SPURegisterInfo::convertDFormToXForm(int dFormOpcode) const
+{
+  switch(dFormOpcode) 
+  {
+    case SPU::AIr32:     return SPU::Ar32;
+    case SPU::LQDr32:    return SPU::LQXr32;
+    case SPU::LQDr128:   return SPU::LQXr128;
+    case SPU::LQDv16i8:  return SPU::LQXv16i8;
+    case SPU::LQDv4f32:  return SPU::LQXv4f32;
+    case SPU::STQDr32:   return SPU::STQXr32;
+    case SPU::STQDr128:  return SPU::STQXr128;
+    case SPU::STQDv16i8: return SPU::STQXv16i8;
+    case SPU::STQDv4i32: return SPU::STQXv4i32;
+    case SPU::STQDv4f32: return SPU::STQXv4f32;
+
+    default: assert( false && "Unhandled D to X-form conversion");
+  }
+  // default will assert, but need to return something to keep the
+  // compiler happy.
+  return dFormOpcode;
+}
+
+// TODO this is already copied from PPC. Could this convenience function
+// be moved to the RegScavenger class?
+unsigned  
+SPURegisterInfo::findScratchRegister(MachineBasicBlock::iterator II, 
+                                     RegScavenger *RS,
+                                     const TargetRegisterClass *RC, 
+                                     int SPAdj) const
+{
+  assert(RS && "Register scavenging must be on");
+  unsigned Reg = RS->FindUnusedReg(RC);
+  if (Reg == 0)
+    Reg = RS->scavengeRegister(RC, II, SPAdj);
+  assert( Reg && "Register scavenger failed");
+  return Reg;
+}
+
 #include "SPUGenRegisterInfo.inc"
diff --git a/lib/Target/CellSPU/SPURegisterInfo.h b/lib/Target/CellSPU/SPURegisterInfo.h
index 48feb5c..0a70318 100644
--- a/lib/Target/CellSPU/SPURegisterInfo.h
+++ b/lib/Target/CellSPU/SPURegisterInfo.h
@@ -53,6 +53,10 @@ namespace llvm {
     virtual const TargetRegisterClass* const *
       getCalleeSavedRegClasses(const MachineFunction *MF) const;
 
+    //! Allow for scavenging, so we can get scratch registers when needed.
+    virtual bool requiresRegisterScavenging(const MachineFunction &MF) const
+    { return true; }
+
     //! Return the reserved registers
     BitVector getReservedRegs(const MachineFunction &MF) const;
 
@@ -97,6 +101,21 @@ namespace llvm {
 
     //! Get DWARF debugging register number
     int getDwarfRegNum(unsigned RegNum, bool isEH) const;
+
+    //! Convert D-form load/store to X-form load/store
+    /*!
+      Converts a regiser displacement load/store into a register-indexed
+      load/store for large stack frames, when the stack frame exceeds the
+      range of a s10 displacement.
+     */
+    int convertDFormToXForm(int dFormOpcode) const;
+
+    //! Acquire an unused register in an emergency.
+    unsigned findScratchRegister(MachineBasicBlock::iterator II,
+                                 RegScavenger *RS,
+                                 const TargetRegisterClass *RC, 
+                                 int SPAdj) const;
+    
   };
 } // end namespace llvm
 
diff --git a/test/CodeGen/CellSPU/bigstack.ll b/test/CodeGen/CellSPU/bigstack.ll
new file mode 100644
index 0000000..5483f46
--- /dev/null
+++ b/test/CodeGen/CellSPU/bigstack.ll
@@ -0,0 +1,17 @@
+; RUN: llc < %s -march=cellspu -o %t1.s
+; RUN: grep lqx   %t1.s | count 4
+; RUN: grep il    %t1.s | grep -v file | count 7
+; RUN: grep stqx  %t1.s | count 2
+
+define i32 @bigstack() nounwind {
+entry:
+  %avar = alloca i32                            
+  %big_data = alloca [2048 x i32]                
+  store i32 3840, i32* %avar, align 4
+  br label %return
+
+return:                                          
+  %retval = load i32* %avar                
+  ret i32 %retval
+}
+