[AMDGPU] Implement -amdgpu-spill-cfi-saved-regs

These spills need special CFI anyway, so implementing them directly
where CFI is emitted avoids the need to invent a mechanism to track them
from ISel.

Co-authored-by: Scott Linder <scott.linder@amd.com>
Co-authored-by: Venkata Ramanaiah Nalamothu <VenkataRamanaiah.Nalamothu@amd.com>

Author: 3 people

llvm/lib/Target/AMDGPU/SIFrameLowering.cpp

@@ -663,12 +663,21 @@ void SIFrameLowering::emitEntryFunctionFlatScratchInit(
 }
 
 // Note SGPRSpill stack IDs should only be used for SGPR spilling to VGPRs, not
-// memory. They should have been removed by now.
-static bool allStackObjectsAreDead(const MachineFrameInfo &MFI) {
+// memory. They should have been removed by now, except CFI Saved Reg spills.
+static bool allStackObjectsAreDead(const MachineFunction &MF) {
+  const MachineFrameInfo &MFI = MF.getFrameInfo();
+  const SIMachineFunctionInfo *FuncInfo = MF.getInfo<SIMachineFunctionInfo>();
   for (int I = MFI.getObjectIndexBegin(), E = MFI.getObjectIndexEnd();
        I != E; ++I) {
-    if (!MFI.isDeadObjectIndex(I))
+    if (!MFI.isDeadObjectIndex(I)) {
+      // determineCalleeSaves() might have added the SGPRSpill stack IDs for
+      // CFI saves into scratch VGPR, ignore them
+      if (MFI.getStackID(I) == TargetStackID::SGPRSpill &&
+          FuncInfo->checkIndexInPrologEpilogSGPRSpills(I)) {
+        continue;
+      }
       return false;
+    }
   }
 
   return true;
@@ -688,8 +697,8 @@ Register SIFrameLowering::getEntryFunctionReservedScratchRsrcReg(
 
   Register ScratchRsrcReg = MFI->getScratchRSrcReg();
 
-  if (!ScratchRsrcReg || (!MRI.isPhysRegUsed(ScratchRsrcReg) &&
-                          allStackObjectsAreDead(MF.getFrameInfo())))
+  if (!ScratchRsrcReg ||
+      (!MRI.isPhysRegUsed(ScratchRsrcReg) && allStackObjectsAreDead(MF)))
     return Register();
 
   if (ST.hasSGPRInitBug() ||
@@ -916,7 +925,7 @@ void SIFrameLowering::emitEntryFunctionPrologue(MachineFunction &MF,
   bool NeedsFlatScratchInit =
       MFI->getUserSGPRInfo().hasFlatScratchInit() &&
       (MRI.isPhysRegUsed(AMDGPU::FLAT_SCR) || FrameInfo.hasCalls() ||
-       (!allStackObjectsAreDead(FrameInfo) && ST.enableFlatScratch()));
+       (!allStackObjectsAreDead(MF) && ST.enableFlatScratch()));
 
   if ((NeedsFlatScratchInit || ScratchRsrcReg) &&
       PreloadedScratchWaveOffsetReg && !ST.flatScratchIsArchitected()) {
@@ -1309,6 +1318,11 @@ void SIFrameLowering::emitCSRSpillStores(MachineFunction &MF,
         LiveUnits.addReg(Reg);
     }
   }
+
+  // Remove the spill entry created for EXEC. It is needed only for CFISaves in
+  // the prologue.
+  if (TRI.isCFISavedRegsSpillEnabled())
+    FuncInfo->removePrologEpilogSGPRSpillEntry(TRI.getExec());
 }
 
 void SIFrameLowering::emitCSRSpillRestores(
@@ -1790,14 +1804,14 @@ void SIFrameLowering::processFunctionBeforeFrameFinalized(
   // can. Any remaining SGPR spills will go to memory, so move them back to the
   // default stack.
   bool HaveSGPRToVMemSpill =
-      FuncInfo->removeDeadFrameIndices(MFI, /*ResetSGPRSpillStackIDs*/ true);
+      FuncInfo->removeDeadFrameIndices(MF, /*ResetSGPRSpillStackIDs*/ true);
   assert(allSGPRSpillsAreDead(MF) &&
          "SGPR spill should have been removed in SILowerSGPRSpills");
 
   // FIXME: The other checks should be redundant with allStackObjectsAreDead,
   // but currently hasNonSpillStackObjects is set only from source
   // allocas. Stack temps produced from legalization are not counted currently.
-  if (!allStackObjectsAreDead(MFI)) {
+  if (!allStackObjectsAreDead(MF)) {
     assert(RS && "RegScavenger required if spilling");
 
     // Add an emergency spill slot
@@ -1897,6 +1911,18 @@ void SIFrameLowering::determinePrologEpilogSGPRSaves(
     MFI->setSGPRForEXECCopy(AMDGPU::NoRegister);
   }
 
+  if (TRI->isCFISavedRegsSpillEnabled()) {
+    Register Exec = TRI->getExec();
+    assert(!MFI->hasPrologEpilogSGPRSpillEntry(Exec) &&
+           "Re-reserving spill slot for EXEC");
+    // FIXME: Machine Copy Propagation currently optimizes away the EXEC copy to
+    // the scratch as we emit it only in the prolog. This optimization should
+    // not happen for frame related instructions. Until this is fixed ignore
+    // copy to scratch SGPR.
+    getVGPRSpillLaneOrTempRegister(MF, LiveUnits, Exec, RC,
+                                   /*IncludeScratchCopy=*/false);
+  }
+
   // hasFP only knows about stack objects that already exist. We're now
   // determining the stack slots that will be created, so we have to predict
   // them. Stack objects force FP usage with calls.
@@ -1906,8 +1932,7 @@ void SIFrameLowering::determinePrologEpilogSGPRSaves(
   //
   // FIXME: Is this really hasReservedCallFrame?
   const bool WillHaveFP =
-      FrameInfo.hasCalls() &&
-      (SavedVGPRs.any() || !allStackObjectsAreDead(FrameInfo));
+      FrameInfo.hasCalls() && (SavedVGPRs.any() || !allStackObjectsAreDead(MF));
 
   if (WillHaveFP || hasFP(MF)) {
     Register FramePtrReg = MFI->getFrameOffsetReg();

llvm/lib/Target/AMDGPU/SIFrameLowering.h

@@ -114,6 +114,13 @@ class SIFrameLowering final : public AMDGPUFrameLowering {
 public:
   bool requiresStackPointerReference(const MachineFunction &MF) const;
 
+  /// If '-amdgpu-spill-cfi-saved-regs' is enabled, emit RA/EXEC spills to
+  /// a free VGPR (lanes) or memory and corresponding CFI rules.
+  void emitCFISavedRegSpills(MachineFunction &MF, MachineBasicBlock &MBB,
+                             MachineBasicBlock::iterator MBBI,
+                             LiveRegUnits &LiveRegs,
+                             bool emitSpillsToMem) const;
+
   /// Create a CFI index for CFIInst and build a MachineInstr around it.
   MachineInstr *
   buildCFI(MachineBasicBlock &MBB, MachineBasicBlock::iterator MBBI,

llvm/lib/Target/AMDGPU/SILowerSGPRSpills.cpp

@@ -531,7 +531,7 @@ bool SILowerSGPRSpills::run(MachineFunction &MF) {
     // free frame index ids by the later pass(es) like "stack slot coloring"
     // which in turn could mess-up with the book keeping of "frame index to VGPR
     // lane".
-    FuncInfo->removeDeadFrameIndices(MFI, /*ResetSGPRSpillStackIDs*/ false);
+    FuncInfo->removeDeadFrameIndices(MF, /*ResetSGPRSpillStackIDs*/ false);
 
     MadeChange = true;
   }

llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp

@@ -566,7 +566,8 @@ bool SIMachineFunctionInfo::allocateVGPRSpillToAGPR(MachineFunction &MF,
 }
 
 bool SIMachineFunctionInfo::removeDeadFrameIndices(
-    MachineFrameInfo &MFI, bool ResetSGPRSpillStackIDs) {
+    MachineFunction &MF, bool ResetSGPRSpillStackIDs) {
+  MachineFrameInfo &MFI = MF.getFrameInfo();
   // Remove dead frame indices from function frame, however keep FP & BP since
   // spills for them haven't been inserted yet. And also make sure to remove the
   // frame indices from `SGPRSpillsToVirtualVGPRLanes` data structure,

llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.h

@@ -757,6 +757,16 @@ class SIMachineFunctionInfo final : public AMDGPUMachineFunction,
                    }) != PrologEpilogSGPRSpills.end();
   }
 
+  // Remove if an entry created for \p Reg.
+  void removePrologEpilogSGPRSpillEntry(Register Reg) {
+    auto I = find_if(PrologEpilogSGPRSpills,
+                     [&Reg](const auto &Spill) { return Spill.first == Reg; });
+    if (I == PrologEpilogSGPRSpills.end())
+      return;
+
+    PrologEpilogSGPRSpills.erase(I);
+  }
+
   const PrologEpilogSGPRSaveRestoreInfo &
   getPrologEpilogSGPRSaveRestoreInfo(Register Reg) const {
     const auto *I = find_if(PrologEpilogSGPRSpills, [&Reg](const auto &Spill) {
@@ -835,8 +845,7 @@ class SIMachineFunctionInfo final : public AMDGPUMachineFunction,
 
   /// If \p ResetSGPRSpillStackIDs is true, reset the stack ID from sgpr-spill
   /// to the default stack.
-  bool removeDeadFrameIndices(MachineFrameInfo &MFI,
-                              bool ResetSGPRSpillStackIDs);
+  bool removeDeadFrameIndices(MachineFunction &MF, bool ResetSGPRSpillStackIDs);
 
   int getScavengeFI(MachineFrameInfo &MFI, const SIRegisterInfo &TRI);
   std::optional<int> getOptionalScavengeFI() const { return ScavengeFI; }

llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp

@@ -35,6 +35,11 @@ static cl::opt<bool> EnableSpillSGPRToVGPR(
   cl::ReallyHidden,
   cl::init(true));
 
+static cl::opt<bool> EnableSpillCFISavedRegs(
+    "amdgpu-spill-cfi-saved-regs",
+    cl::desc("Enable spilling the registers required for CFI emission"),
+    cl::ReallyHidden, cl::init(false), cl::ZeroOrMore);
+
 std::array<std::vector<int16_t>, 32> SIRegisterInfo::RegSplitParts;
 std::array<std::array<uint16_t, 32>, 9> SIRegisterInfo::SubRegFromChannelTable;
 
@@ -561,6 +566,10 @@ unsigned SIRegisterInfo::getSubRegFromChannel(unsigned Channel,
   return SubRegFromChannelTable[NumRegIndex - 1][Channel];
 }
 
+bool SIRegisterInfo::isCFISavedRegsSpillEnabled() const {
+  return EnableSpillCFISavedRegs;
+}
+
 MCRegister
 SIRegisterInfo::getAlignedHighSGPRForRC(const MachineFunction &MF,
                                         const unsigned Align,

llvm/lib/Target/AMDGPU/SIRegisterInfo.h

@@ -80,6 +80,8 @@ class SIRegisterInfo final : public AMDGPUGenRegisterInfo {
     return SpillSGPRToVGPR;
   }
 
+  bool isCFISavedRegsSpillEnabled() const;
+
   /// Return the largest available SGPR aligned to \p Align for the register
   /// class \p RC.
   MCRegister getAlignedHighSGPRForRC(const MachineFunction &MF,