AMDGPU: Support commuting a FrameIndex operand

[nhaehnle]

LGTM, just one small point.

lib/Target/AMDGPU/SIInstrInfo.cpp

932

ImmOp should be renamed now.

[arsenm]

r281369

include/llvm/CodeGen/MachineOperand.h

public:
   void ChangeToFPImmediate(const ConstantFP *FPImm);
 
   /// ChangeToES - Replace this operand with a new external symbol operand.
   void ChangeToES(const char *SymName, unsigned char TargetFlags = 0);
 
   /// ChangeToMCSymbol - Replace this operand with a new MC symbol operand.
   void ChangeToMCSymbol(MCSymbol *Sym);
 
+  /// Replace this operand with a frame index.
+  void ChangeToFrameIndex(int Idx);
+
   /// ChangeToRegister - Replace this operand with a new register operand of
   /// the specified value.  If an operand is known to be an register already,
   /// the setReg method should be used.
   void ChangeToRegister(unsigned Reg, bool isDef, bool isImp = false,
                         bool isKill = false, bool isDead = false,
                         bool isUndef = false, bool isDebug = false);
 
   //===--------------------------------------------------------------------===//

lib/CodeGen/MachineInstr.cpp

assert((!isReg() || !isTied()) &&
          "Cannot change a tied operand into an MCSymbol");
 
   removeRegFromUses();
 
   OpKind = MO_MCSymbol;
   Contents.Sym = Sym;
 }
 
+void MachineOperand::ChangeToFrameIndex(int Idx) {
+  assert((!isReg() || !isTied()) &&
+         "Cannot change a tied operand into a FrameIndex");
+
+  removeRegFromUses();
+
+  OpKind = MO_FrameIndex;
+  setIndex(Idx);
+}
+
 /// ChangeToRegister - Replace this operand with a new register operand of
 /// the specified value.  If an operand is known to be an register already,
 /// the setReg method should be used.
 void MachineOperand::ChangeToRegister(unsigned Reg, bool isDef, bool isImp,
                                       bool isKill, bool isDead, bool isUndef,
                                       bool isDebug) {
   MachineRegisterInfo *RegInfo = nullptr;
   if (MachineInstr *MI = getParent())

lib/Target/AMDGPU/SIInstrInfo.cpp

bool SIInstrInfo::swapSourceModifiers(MachineInstr &MI,
 
   Src1Mods->setImm(Src0ModsVal);
   Src0Mods->setImm(Src1ModsVal);
   return true;
 }
 
 static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
                                              MachineOperand &RegOp,
                                              MachineOperand &ImmOp) {
-  // TODO: Handle other immediate like types.
-  if (!ImmOp.isImm())
+  unsigned Reg = RegOp.getReg();
+  unsigned SubReg = RegOp.getSubReg();
+  bool IsKill = RegOp.isKill();
+  bool IsDead = RegOp.isDead();
+  bool IsUndef = RegOp.isUndef();
+  bool IsDebug = RegOp.isDebug();
+
+  if (ImmOp.isImm())
+    RegOp.ChangeToImmediate(ImmOp.getImm());
+  else if (ImmOp.isFI())
+    RegOp.ChangeToFrameIndex(ImmOp.getIndex());
+  else
     return nullptr;
 
-  int64_t ImmVal = ImmOp.getImm();
-  ImmOp.ChangeToRegister(RegOp.getReg(), false, false,
-                         RegOp.isKill(), RegOp.isDead(), RegOp.isUndef(),
-                         RegOp.isDebug());
-  ImmOp.setSubReg(RegOp.getSubReg());
-  RegOp.ChangeToImmediate(ImmVal);
+  ImmOp.ChangeToRegister(Reg, false, false, IsKill, IsDead, IsUndef, IsDebug);
+  ImmOp.setSubReg(SubReg);
+
   return &MI;
 }
 
 MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
                                                   unsigned Src0Idx,
                                                   unsigned Src1Idx) const {
   assert(!NewMI && "this should never be used");
 

test/CodeGen/AMDGPU/commute-compares.ll

define void @commute_uno_2.0_f64(i32 addrspace(1)* %out, double addrspace(1)* %in) #1 {
   %gep.out = getelementptr i32, i32 addrspace(1)* %out, i32 %tid
   %val = load double, double addrspace(1)* %gep.in
   %cmp = fcmp uno double %val, 2.0
   %ext = sext i1 %cmp to i32
   store i32 %ext, i32 addrspace(1)* %gep.out
   ret void
 }
 
+; Without commuting the frame index in the pre-regalloc run of
+; SIShrinkInstructions, this was using the VOP3 compare.
+
+; GCN-LABEL: {{^}}commute_frameindex:
+; GCN: v_cmp_eq_i32_e32 vcc, 0, v{{[0-9]+}}
+define void @commute_frameindex(i32 addrspace(1)* nocapture %out) #0 {
+entry:
+  %stack0 = alloca i32
+  %ptr0 = load volatile i32*, i32* addrspace(1)* undef
+  %eq = icmp eq i32* %ptr0, %stack0
+  %ext = zext i1 %eq to i32
+  store volatile i32 %ext, i32 addrspace(1)* %out
+  ret void
+}
+
 attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind }