AMDGPU: Make sure to reserve superregisters

[arsenm]

Summary

I think this could potentially have broken if
one of the super registers were allocated
that contain v254/v255.

[• tstellarAMD]

LGTM.

[arsenm]

r246051

lib/Target/AMDGPU/SIRegisterInfo.h

 
 #include "AMDGPURegisterInfo.h"
 #include "AMDGPUSubtarget.h"
 #include "llvm/Support/Debug.h"
 
 namespace llvm {
 
 struct SIRegisterInfo : public AMDGPURegisterInfo {
+private:
+  void reserveRegisterTuples(BitVector &, unsigned Reg) const;
 
+public:
   SIRegisterInfo();
 
   BitVector getReservedRegs(const MachineFunction &MF) const override;
 
   unsigned getRegPressureSetLimit(const MachineFunction &MF,
                                   unsigned Idx) const override;
 
   bool requiresRegisterScavenging(const MachineFunction &Fn) const override;

lib/Target/AMDGPU/SIRegisterInfo.cpp

 #include "llvm/CodeGen/RegisterScavenging.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/LLVMContext.h"
 
 using namespace llvm;
 
 SIRegisterInfo::SIRegisterInfo() : AMDGPURegisterInfo() {}
 
-BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
-  BitVector Reserved(getNumRegs());
-  Reserved.set(AMDGPU::EXEC);
+void SIRegisterInfo::reserveRegisterTuples(BitVector &Reserved, unsigned Reg) const {
+  MCRegAliasIterator R(Reg, this, true);
 
-  // EXEC_LO and EXEC_HI could be allocated and used as regular register,
-  // but this seems likely to result in bugs, so I'm marking them as reserved.
-  Reserved.set(AMDGPU::EXEC_LO);
-  Reserved.set(AMDGPU::EXEC_HI);
+  for (; R.isValid(); ++R)
+    Reserved.set(*R);
+}
 
+BitVector SIRegisterInfo::getReservedRegs(const MachineFunction &MF) const {
+  BitVector Reserved(getNumRegs());
   Reserved.set(AMDGPU::INDIRECT_BASE_ADDR);
-  Reserved.set(AMDGPU::FLAT_SCR);
-  Reserved.set(AMDGPU::FLAT_SCR_LO);
-  Reserved.set(AMDGPU::FLAT_SCR_HI);
+
+  // EXEC_LO and EXEC_HI could be allocated and used as regular register, but
+  // this seems likely to result in bugs, so I'm marking them as reserved.
+  reserveRegisterTuples(Reserved, AMDGPU::EXEC);
+  reserveRegisterTuples(Reserved, AMDGPU::FLAT_SCR);
 
   // Reserve some VGPRs to use as temp registers in case we have to spill VGPRs
-  Reserved.set(AMDGPU::VGPR255);
-  Reserved.set(AMDGPU::VGPR254);
+  reserveRegisterTuples(Reserved, AMDGPU::VGPR254);
+  reserveRegisterTuples(Reserved, AMDGPU::VGPR255);
 
   // Tonga and Iceland can only allocate a fixed number of SGPRs due
   // to a hw bug.
   if (MF.getSubtarget<AMDGPUSubtarget>().hasSGPRInitBug()) {
     unsigned NumSGPRs = AMDGPU::SGPR_32RegClass.getNumRegs();
     // Reserve some SGPRs for FLAT_SCRATCH and VCC (4 SGPRs).
     // Assume XNACK_MASK is unused.
     unsigned Limit = AMDGPUSubtarget::FIXED_SGPR_COUNT_FOR_INIT_BUG - 4;
 
     for (unsigned i = Limit; i < NumSGPRs; ++i) {
       unsigned Reg = AMDGPU::SGPR_32RegClass.getRegister(i);
-      MCRegAliasIterator R = MCRegAliasIterator(Reg, this, true);
-
-      for (; R.isValid(); ++R)
-        Reserved.set(*R);
+      reserveRegisterTuples(Reserved, Reg);
     }
   }
 
   return Reserved;
 }
 
 unsigned SIRegisterInfo::getRegPressureSetLimit(const MachineFunction &MF,
                                                 unsigned Idx) const {