[AMDGPU] Use std::variant in ArgDescriptor. (llvm#167992)

This replaces the 2 bool flags and the anonymous union. This also
removes an implicit conversion from Register to unsigned and a call to
MCRegister::id().

The ArgDescriptor constructor was always assigning the union through the
MCRegister field even for stack offsets.

The change to SIMachineFunctionInfo.h fixes a case where getRegister was
being called on an unset ArgDescriptor. Since it was only this case, it
seemed cleaner to fix it at the caller. The other option would be to
make getRegister() return MCRegister() for an unset ArgDescriptor.

Author: topperc

llvm/lib/Target/AMDGPU/AMDGPUArgumentUsageInfo.h

@@ -13,6 +13,7 @@
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/CodeGen/Register.h"
 #include "llvm/Pass.h"
+#include <variant>
 
 namespace llvm {
 
@@ -27,55 +28,44 @@ struct ArgDescriptor {
   friend struct AMDGPUFunctionArgInfo;
   friend class AMDGPUArgumentUsageInfo;
 
-  union {
-    MCRegister Reg;
-    unsigned StackOffset;
-  };
+  std::variant<std::monostate, MCRegister, unsigned> Val;
 
   // Bitmask to locate argument within the register.
   unsigned Mask;
 
-  bool IsStack : 1;
-  bool IsSet : 1;
-
 public:
-  ArgDescriptor(unsigned Val = 0, unsigned Mask = ~0u, bool IsStack = false,
-                bool IsSet = false)
-      : Reg(Val), Mask(Mask), IsStack(IsStack), IsSet(IsSet) {}
+  ArgDescriptor(unsigned Mask = ~0u) : Mask(Mask) {}
 
   static ArgDescriptor createRegister(Register Reg, unsigned Mask = ~0u) {
-    return ArgDescriptor(Reg, Mask, false, true);
+    ArgDescriptor Ret(Mask);
+    Ret.Val = Reg.asMCReg();
+    return Ret;
   }
 
   static ArgDescriptor createStack(unsigned Offset, unsigned Mask = ~0u) {
-    return ArgDescriptor(Offset, Mask, true, true);
+    ArgDescriptor Ret(Mask);
+    Ret.Val = Offset;
+    return Ret;
   }
 
   static ArgDescriptor createArg(const ArgDescriptor &Arg, unsigned Mask) {
-    return ArgDescriptor(Arg.Reg.id(), Mask, Arg.IsStack, Arg.IsSet);
+    // Copy the descriptor, then change the mask.
+    ArgDescriptor Ret(Arg);
+    Ret.Mask = Mask;
+    return Ret;
   }
 
-  bool isSet() const {
-    return IsSet;
-  }
+  bool isSet() const { return !std::holds_alternative<std::monostate>(Val); }
 
   explicit operator bool() const {
     return isSet();
   }
 
-  bool isRegister() const {
-    return !IsStack;
-  }
+  bool isRegister() const { return std::holds_alternative<MCRegister>(Val); }
 
-  MCRegister getRegister() const {
-    assert(!IsStack);
-    return Reg;
-  }
+  MCRegister getRegister() const { return std::get<MCRegister>(Val); }
 
-  unsigned getStackOffset() const {
-    assert(IsStack);
-    return StackOffset;
-  }
+  unsigned getStackOffset() const { return std::get<unsigned>(Val); }
 
   unsigned getMask() const {
     // None of the target SGPRs or VGPRs are expected to have a 'zero' mask.

llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h

@@ -1014,7 +1014,9 @@ class SIMachineFunctionInfo final : public AMDGPUMachineFunction,
   void setNumWaveDispatchVGPRs(unsigned Count) { NumWaveDispatchVGPRs = Count; }
 
   Register getPrivateSegmentWaveByteOffsetSystemSGPR() const {
-    return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
+    if (ArgInfo.PrivateSegmentWaveByteOffset)
+      return ArgInfo.PrivateSegmentWaveByteOffset.getRegister();
+    return MCRegister();
   }
 
   /// Returns the physical register reserved for use as the resource