[AMDGPU] Insert PS early exit at end of control flow

[critson]

Summary

Depends on D82641.
Exit early if the exec mask is zero at the end of control flow.
Mark the ends of control flow during control flow lowering and
convert these to exits during the insert skips pass.

[critson]

I appreciate there might be some resistance to this way of doing things, i.e. relying on control flow lowering to insert cleanup markers. As an alternative I can rewrite this to look directly at the exec mask processing during the insert skips pass.

[nhaehnle]

Can you explain the reasoning behind how you decide where to insert SI_KILL_CLEANUP?

llvm/lib/Target/AMDGPU/SIInstructions.td

382–383

Shouldn't this also Use EXEC?

[critson]

In D82737#2119798, @nhaehnle wrote:

Can you explain the reasoning behind how you decide where to insert SI_KILL_CLEANUP?

I will add this as a comment in the code (because apparently I forgot):
SI_KILL_CLEANUP is inserted immediately after SI_END_CF for control flow where a kill is present.
When SI_KILL_CLEANUP is uniformly reached it will become a test of the EXEC mask immediately after it has been updated by the SI_END_CF to reflect the true state of live lanes.

We could actually blindly insert SI_KILL_CLEANUP after every SI_END_CF, reducing the code complexity, but we'd end up with a lot of early exits that would never be taken.

llvm/lib/Target/AMDGPU/SIInstructions.td

382–383

It does, this entire block of instructions is Use EXEC.

[critson]

• Rebase
• Add comment

[nhaehnle]

Okay, thanks, this all makes sense to me. I do have some suggestions for tightening checks up a little, but either way LGTM.

llvm/lib/Target/AMDGPU/SILowerControlFlow.cpp

222

Would it make sense to add an assertion that UseMI->getOpcode() == AMDGPU::SI_ELSE?

228–232

When can the NextExec.isDead() case occur? Unreachable code or the like?

In any case, could you move this around a bit as:

  if (NextExec.isDead()) {
    assert(!SimpleIf);
    break;
  }
...
if (UseMI->getOpcode() == AMDGPU::SI_END_CF) {
  if (hasKill(...)) {
    NeedsKillCleanup.insert(&*UseMI);
    SimpleIf = false;
  }
}

llvm/lib/Target/AMDGPU/SIInsertSkips.cpp

for (I = MBB.begin(); I != MBB.end(); I = Next) {
           // change because it modifies the CFG.
           KillInstrs.push_back(&MI);
         } else {
           MI.eraseFromParent();
         }
         break;
       }
 
+      case AMDGPU::SI_KILL_CLEANUP:
+        if (MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS &&
+            dominatesAllReachable(MBB)) {
+          KillInstrs.push_back(&MI);
+        } else {
+          MI.eraseFromParent();
+        }
+        break;
+
       default:
         break;
       }
     }
   }
 
   for (MachineInstr *Kill : KillInstrs) {
     skipIfDead(*Kill->getParent(), std::next(Kill->getIterator()),
                Kill->getDebugLoc());
     Kill->eraseFromParent();
   }
   KillInstrs.clear();
   EarlyExitBlock = nullptr;
 
   return MadeChange;
 }

llvm/lib/Target/AMDGPU/SIInstructions.td

multiclass PseudoInstKill <dag ins> {
   def _TERMINATOR : SPseudoInstSI <(outs), ins> {
     let isTerminator = 1;
   }
 }
 
 defm SI_KILL_I1 : PseudoInstKill <(ins SCSrc_i1:$src, i1imm:$killvalue)>;
 defm SI_KILL_F32_COND_IMM : PseudoInstKill <(ins VSrc_b32:$src0, i32imm:$src1, i32imm:$cond)>;
 
+let Defs = [EXEC] in
+def SI_KILL_CLEANUP : SPseudoInstSI <(outs), (ins)>;
+
 let Defs = [EXEC,VCC] in
 def SI_ILLEGAL_COPY : SPseudoInstSI <
   (outs unknown:$dst), (ins unknown:$src),
   [], " ; illegal copy $src to $dst">;
 
 } // End Uses = [EXEC], Defs = [EXEC,VCC]
 
 // Branch on undef scc. Used to avoid intermediate copy from

llvm/lib/Target/AMDGPU/SILowerControlFlow.cpp

 class SILowerControlFlow : public MachineFunctionPass {
 private:
   const SIRegisterInfo *TRI = nullptr;
   const SIInstrInfo *TII = nullptr;
   LiveIntervals *LIS = nullptr;
   MachineRegisterInfo *MRI = nullptr;
   SetVector<MachineInstr*> LoweredEndCf;
   DenseSet<Register> LoweredIf;
+  SmallSet<MachineInstr *, 16> NeedsKillCleanup;
 
   const TargetRegisterClass *BoolRC = nullptr;
+  bool InsertKillCleanups;
   unsigned AndOpc;
   unsigned OrOpc;
   unsigned XorOpc;
   unsigned MovTermOpc;
   unsigned Andn2TermOpc;
   unsigned XorTermrOpc;
   unsigned OrSaveExecOpc;
   unsigned Exec;
 
   void emitIf(MachineInstr &MI);
   void emitElse(MachineInstr &MI);
   void emitIfBreak(MachineInstr &MI);
   void emitLoop(MachineInstr &MI);
   void emitEndCf(MachineInstr &MI);
 
   void findMaskOperands(MachineInstr &MI, unsigned OpNo,
                         SmallVectorImpl<MachineOperand> &Src) const;
 
   void combineMasks(MachineInstr &MI);
 
+  void process(MachineInstr &MI);
+
   // Skip to the next instruction, ignoring debug instructions, and trivial
   // block boundaries (blocks that have one (typically fallthrough) successor,
   // and the successor has one predecessor.
   MachineBasicBlock::iterator
   skipIgnoreExecInstsTrivialSucc(MachineBasicBlock &MBB,
                                  MachineBasicBlock::iterator It) const;
 
   // Remove redundant SI_END_CF instructions.
static void setImpSCCDefDead(MachineInstr &MI, bool IsDead) {
   MachineOperand &ImpDefSCC = MI.getOperand(3);
   assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef());
 
   ImpDefSCC.setIsDead(IsDead);
 }
 
 char &llvm::SILowerControlFlowID = SILowerControlFlow::ID;
 
-static bool isSimpleIf(const MachineInstr &MI, const MachineRegisterInfo *MRI,
-                       const SIInstrInfo *TII) {
-  Register SaveExecReg = MI.getOperand(0).getReg();
-  auto U = MRI->use_instr_nodbg_begin(SaveExecReg);
-
-  if (U == MRI->use_instr_nodbg_end() ||
-      std::next(U) != MRI->use_instr_nodbg_end() ||
-      U->getOpcode() != AMDGPU::SI_END_CF)
-    return false;
-
-  // Check for SI_KILL_*_TERMINATOR on path from if to endif.
-  // if there is any such terminator simplififcations are not safe.
-  auto SMBB = MI.getParent();
-  auto EMBB = U->getParent();
+static bool hasKill(const MachineBasicBlock *Begin,
+                    const MachineBasicBlock *End, const SIInstrInfo *TII) {
   DenseSet<const MachineBasicBlock*> Visited;
-  SmallVector<MachineBasicBlock*, 4> Worklist(SMBB->succ_begin(),
-                                              SMBB->succ_end());
+  SmallVector<MachineBasicBlock *, 4> Worklist(Begin->succ_begin(),
+                                               Begin->succ_end());
 
   while (!Worklist.empty()) {
     MachineBasicBlock *MBB = Worklist.pop_back_val();
 
-    if (MBB == EMBB || !Visited.insert(MBB).second)
+    if (MBB == End || !Visited.insert(MBB).second)
       continue;
-    for(auto &Term : MBB->terminators())
+    for (auto &Term : MBB->terminators())
       if (TII->isKillTerminator(Term.getOpcode()))
-        return false;
+        return true;
 
     Worklist.append(MBB->succ_begin(), MBB->succ_end());
   }
 
+  return false;
+}
+
+static bool isSimpleIf(const MachineInstr &MI, const MachineRegisterInfo *MRI) {
+  Register SaveExecReg = MI.getOperand(0).getReg();
+  auto U = MRI->use_instr_nodbg_begin(SaveExecReg);
+
+  if (U == MRI->use_instr_nodbg_end() ||
+      std::next(U) != MRI->use_instr_nodbg_end() ||
+      U->getOpcode() != AMDGPU::SI_END_CF)
+    return false;
+
   return true;
 }
 
 void SILowerControlFlow::emitIf(MachineInstr &MI) {
   MachineBasicBlock &MBB = *MI.getParent();
   const DebugLoc &DL = MI.getDebugLoc();
   MachineBasicBlock::iterator I(&MI);
   Register SaveExecReg = MI.getOperand(0).getReg();
   MachineOperand& Cond = MI.getOperand(1);
   assert(Cond.getSubReg() == AMDGPU::NoSubRegister);
 
   MachineOperand &ImpDefSCC = MI.getOperand(4);
   assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef());
 
   // If there is only one use of save exec register and that use is SI_END_CF,
   // we can optimize SI_IF by returning the full saved exec mask instead of
   // just cleared bits.
-  bool SimpleIf = isSimpleIf(MI, MRI, TII);
+  bool SimpleIf = isSimpleIf(MI, MRI);
+
+  if (InsertKillCleanups) {
+    // Check for SI_KILL_*_TERMINATOR on full path of control flow and
+    // flag the associated SI_END_CF for insertion of a kill cleanup.
+    auto UseMI = MRI->use_instr_nodbg_begin(SaveExecReg);
+    while (UseMI->getOpcode() != AMDGPU::SI_END_CF) {
+      assert(std::next(UseMI) == MRI->use_instr_nodbg_end());
+      MachineOperand &NextExec = UseMI->getOperand(0);
+      Register NextExecReg = NextExec.getReg();
+      if (NextExec.isDead())
+        break;
+      UseMI = MRI->use_instr_nodbg_begin(NextExecReg);
+    }
+    if (hasKill(MI.getParent(), UseMI->getParent(), TII)) {
+      if (UseMI->getOpcode() == AMDGPU::SI_END_CF)
+        NeedsKillCleanup.insert(&*UseMI);
+      SimpleIf = false;
+    }
+  } else if (SimpleIf) {
+    // Check for SI_KILL_*_TERMINATOR on path from if to endif.
+    // if there is any such terminator simplifications are not safe.
+    auto UseMI = MRI->use_instr_nodbg_begin(SaveExecReg);
+    SimpleIf = !hasKill(MI.getParent(), UseMI->getParent(), TII);
+  }
 
   // Add an implicit def of exec to discourage scheduling VALU after this which
   // will interfere with trying to form s_and_saveexec_b64 later.
   Register CopyReg = SimpleIf ? SaveExecReg
                        : MRI->createVirtualRegister(BoolRC);
   MachineInstr *CopyExec =
     BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), CopyReg)
     .addReg(Exec)
SILowerControlFlow::skipIgnoreExecInstsTrivialSucc(
   SmallSet<const MachineBasicBlock *, 4> Visited;
   MachineBasicBlock *B = &MBB;
   do {
     if (!Visited.insert(B).second)
       return MBB.end();
 
     auto E = B->end();
     for ( ; It != E; ++It) {
+      if (It->getOpcode() == AMDGPU::SI_KILL_CLEANUP)
+        continue;
       if (TII->mayReadEXEC(*MRI, *It))
         break;
     }
 
     if (It != E)
       return It;
 
     if (B->succ_size() != 1)
MachineBasicBlock::iterator InsPt =
       Def && Def->getParent() == &MBB ? std::next(MachineBasicBlock::iterator(Def))
                                : MBB.begin();
   MachineInstr *NewMI = BuildMI(MBB, InsPt, DL, TII->get(OrOpc), Exec)
                             .addReg(Exec)
                             .add(MI.getOperand(0));
 
   LoweredEndCf.insert(NewMI);
 
-  if (LIS)
+  MachineInstr *CleanUpMI = nullptr;
+  if (NeedsKillCleanup.count(&MI))
+    CleanUpMI = BuildMI(MBB, InsPt, DL, TII->get(AMDGPU::SI_KILL_CLEANUP));
+
+  if (LIS) {
     LIS->ReplaceMachineInstrInMaps(MI, *NewMI);
+    if (CleanUpMI)
+      LIS->InsertMachineInstrInMaps(*CleanUpMI);
+  }
 
   MI.eraseFromParent();
 
   if (LIS)
     LIS->handleMove(*NewMI);
 }
 
 // Returns replace operands for a logical operation, either single result
if (Def && LoweredIf.count(SavedExec)) {
       LLVM_DEBUG(dbgs() << "Skip redundant "; MI->dump());
       if (LIS)
         LIS->RemoveMachineInstrFromMaps(*MI);
       MI->eraseFromParent();
     }
   }
 }
 
+void SILowerControlFlow::process(MachineInstr &MI) {
+  MachineBasicBlock &MBB = *MI.getParent();
+  MachineBasicBlock::iterator I(MI);
+  MachineInstr *Prev = (I != MBB.begin()) ? &*(std::prev(I)) : nullptr;
+
+  switch (MI.getOpcode()) {
+  case AMDGPU::SI_IF:
+    emitIf(MI);
+    break;
+
+  case AMDGPU::SI_ELSE:
+    emitElse(MI);
+    break;
+
+  case AMDGPU::SI_IF_BREAK:
+    emitIfBreak(MI);
+    break;
+
+  case AMDGPU::SI_LOOP:
+    emitLoop(MI);
+    break;
+
+  case AMDGPU::SI_END_CF:
+    emitEndCf(MI);
+    break;
+
+  default:
+    assert(false && "Attempt to process unsupported instruction");
+    break;
+  }
+
+  MachineBasicBlock::iterator Next;
+  for (I = Prev ? Prev->getIterator() : MBB.begin(); I != MBB.end(); I = Next) {
+    Next = std::next(I);
+    MachineInstr &MaskMI = *I;
+    switch (MaskMI.getOpcode()) {
+    case AMDGPU::S_AND_B64:
+    case AMDGPU::S_OR_B64:
+    case AMDGPU::S_AND_B32:
+    case AMDGPU::S_OR_B32:
+      // Cleanup bit manipulations on exec mask
+      combineMasks(MaskMI);
+      break;
+    default:
+      I = MBB.end();
+      break;
+    }
+  }
+}
+
 bool SILowerControlFlow::runOnMachineFunction(MachineFunction &MF) {
   const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
   TII = ST.getInstrInfo();
   TRI = &TII->getRegisterInfo();
 
   // This doesn't actually need LiveIntervals, but we can preserve them.
   LIS = getAnalysisIfAvailable<LiveIntervals>();
   MRI = &MF.getRegInfo();
   BoolRC = TRI->getBoolRC();
+  InsertKillCleanups =
+      MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS;
 
   if (ST.isWave32()) {
     AndOpc = AMDGPU::S_AND_B32;
     OrOpc = AMDGPU::S_OR_B32;
     XorOpc = AMDGPU::S_XOR_B32;
     MovTermOpc = AMDGPU::S_MOV_B32_term;
     Andn2TermOpc = AMDGPU::S_ANDN2_B32_term;
     XorTermrOpc = AMDGPU::S_XOR_B32_term;
     OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B32;
     Exec = AMDGPU::EXEC_LO;
   } else {
     AndOpc = AMDGPU::S_AND_B64;
     OrOpc = AMDGPU::S_OR_B64;
     XorOpc = AMDGPU::S_XOR_B64;
     MovTermOpc = AMDGPU::S_MOV_B64_term;
     Andn2TermOpc = AMDGPU::S_ANDN2_B64_term;
     XorTermrOpc = AMDGPU::S_XOR_B64_term;
     OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B64;
     Exec = AMDGPU::EXEC;
   }
 
+  SmallVector<MachineInstr *, 32> Worklist;
+
   MachineFunction::iterator NextBB;
   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
        BI != BE; BI = NextBB) {
     NextBB = std::next(BI);
     MachineBasicBlock &MBB = *BI;
 
-    MachineBasicBlock::iterator I, Next, Last;
-
-    for (I = MBB.begin(), Last = MBB.end(); I != MBB.end(); I = Next) {
+    MachineBasicBlock::iterator I, Next;
+    for (I = MBB.begin(); I != MBB.end(); I = Next) {
       Next = std::next(I);
       MachineInstr &MI = *I;
 
       switch (MI.getOpcode()) {
       case AMDGPU::SI_IF:
-        emitIf(MI);
+        process(MI);
         break;
 
       case AMDGPU::SI_ELSE:
-        emitElse(MI);
-        break;
-
       case AMDGPU::SI_IF_BREAK:
-        emitIfBreak(MI);
-        break;
-
       case AMDGPU::SI_LOOP:
-        emitLoop(MI);
-        break;
-
       case AMDGPU::SI_END_CF:
-        emitEndCf(MI);
+        // Only build worklist if SI_IF instructions must be processed first.
+        if (InsertKillCleanups)
+          Worklist.push_back(&MI);
+        else
+          process(MI);
         break;
 
-      case AMDGPU::S_AND_B64:
-      case AMDGPU::S_OR_B64:
-      case AMDGPU::S_AND_B32:
-      case AMDGPU::S_OR_B32:
-        // Cleanup bit manipulations on exec mask
-        combineMasks(MI);
-        Last = I;
-        continue;
-
       default:
-        Last = I;
-        continue;
+        break;
       }
-
-      // Replay newly inserted code to combine masks
-      Next = (Last == MBB.end()) ? MBB.begin() : Last;
     }
   }
 
+  for (MachineInstr *MI : Worklist)
+    process(*MI);
+
   optimizeEndCf();
 
   LoweredEndCf.clear();
   LoweredIf.clear();
+  NeedsKillCleanup.clear();
 
   return true;
 }

llvm/test/CodeGen/AMDGPU/kill-infinite-loop.ll

 main_body:
   br label %loop
 
 loop:
   call void @llvm.amdgcn.kill(i1 false) #3
   br label %loop
 }
 
-; In case there's an epilog, we shouldn't have to do this.
+; Check that the epilog is the final block
 ; CHECK-LABEL: return_nonvoid
-; CHECK-NOT: exp null off, off, off, off done vm
+; CHECK: exp null off, off, off, off done vm
+; CHECK-NEXT: s_endpgm
+; CHECK-NEXT: BB{{[0-9]+}}_{{[0-9]+}}:
 define amdgpu_ps float @return_nonvoid(float %0) #0 {
 main_body:
   %cmp = fcmp olt float %0, 1.000000e+01
   br i1 %cmp, label %end, label %loop
 
 loop:
   call void @llvm.amdgcn.kill(i1 false) #3
   br label %loop

llvm/test/CodeGen/AMDGPU/skip-if-dead.ll

bb8:                                              ; preds = %bb9, %bb4
   store volatile i32 9, i32 addrspace(1)* undef
   ret void
 
 bb9:                                              ; preds = %bb4
   ret void
 }
 
 ; CHECK-LABEL: {{^}}cbranch_kill:
-; CHECK-NOT: exp null off, off, off, off done vm
+; CHECK: ; %bb.{{[0-9]+}}: ; %export
+; CHECK-NEXT: s_or_b64
+; CHECK-NEXT: s_cbranch_execz [[EXIT:BB[0-9]+_[0-9]+]]
+; CHECK: [[EXIT]]:
+; CHECK-NEXT: exp null off, off, off, off done vm
 define amdgpu_ps void @cbranch_kill(i32 inreg %0, <2 x float> %1) {
 .entry:
   %val0 = extractelement <2 x float> %1, i32 0
   %val1 = extractelement <2 x float> %1, i32 1
   %p0 = call float @llvm.amdgcn.interp.p1(float %val0, i32 immarg 0, i32 immarg 1, i32 %0) #2
   %sample = call float @llvm.amdgcn.image.sample.l.2darray.f32.f32(i32 1, float %p0, float %p0, float %p0, float 0.000000e+00, <8 x i32> undef, <4 x i32> undef, i1 false, i32 0, i32 0)
   %cond0 = fcmp ugt float %sample, 0.000000e+00
   br i1 %cond0, label %live, label %kill