AMDGPU: Insert mem_viol check loop around GWS pre-GFX9

It is necessary to emit this loop around GWS operations in case the
wave is preempted pre-GFX9.

llvm-svn: 363979

5 files changed, 129 insertions, 19 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
index 3b1ed618677..e1be0105aea 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h
@@ -715,6 +715,15 @@ public:
     return getGeneration() < GFX9;
   }
 
+  // True if the hardware rewinds and replays GWS operations if a wave is
+  // preempted.
+  //
+  // If this is false, a GWS operation requires testing if a nack set the
+  // MEM_VIOL bit, and repeating if so.
+  bool hasGWSAutoReplay() const {
+    return getGeneration() >= GFX9;
+  }
+
   bool hasAddNoCarry() const {
     return AddNoCarryInsts;
   }
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index 248f6599f7e..0ef5d79d66a 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -467,7 +467,7 @@ defm DS_WRXCHG_RTN_B64 : DS_1A1D_RET_mc<"ds_wrxchg_rtn_b64", VReg_64>;
 defm DS_WRXCHG2_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2_rtn_b64", VReg_128, VReg_64>;
 defm DS_WRXCHG2ST64_RTN_B64 : DS_1A2D_Off8_RET_mc<"ds_wrxchg2st64_rtn_b64", VReg_128, VReg_64>;
 
-let isConvergent = 1 in {
+let isConvergent = 1, usesCustomInserter = 1 in {
 def DS_GWS_INIT       : DS_GWS_1D<"ds_gws_init"> {
   let mayLoad = 0;
 }
diff --git a/llvm/lib/Target/AMDGPU/SIDefines.h b/llvm/lib/Target/AMDGPU/SIDefines.h
index c21f88f52d1..cc96f1de43a 100644
--- a/llvm/lib/Target/AMDGPU/SIDefines.h
+++ b/llvm/lib/Target/AMDGPU/SIDefines.h
@@ -323,6 +323,8 @@ enum Offset : unsigned { // Offset, (5) [10:6]
   OFFSET_WIDTH_ = 5,
   OFFSET_MASK_ = (((1 << OFFSET_WIDTH_) - 1) << OFFSET_SHIFT_),
 
+  OFFSET_MEM_VIOL = 8,
+
   OFFSET_SRC_SHARED_BASE = 16,
   OFFSET_SRC_PRIVATE_BASE = 0
 };
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index a159f208aca..fbe88ca2246 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -2922,6 +2922,109 @@ MachineBasicBlock *SITargetLowering::splitKillBlock(MachineInstr &MI,
   return SplitBB;
 }
 
+// Split block \p MBB at \p MI, as to insert a loop. If \p InstInLoop is true,
+// \p MI will be the only instruction in the loop body block. Otherwise, it will
+// be the first instruction in the remainder block.
+//
+/// \returns { LoopBody, Remainder }
+static std::pair<MachineBasicBlock *, MachineBasicBlock *>
+splitBlockForLoop(MachineInstr &MI, MachineBasicBlock &MBB, bool InstInLoop) {
+  MachineFunction *MF = MBB.getParent();
+  MachineBasicBlock::iterator I(&MI);
+
+  // To insert the loop we need to split the block. Move everything after this
+  // point to a new block, and insert a new empty block between the two.
+  MachineBasicBlock *LoopBB = MF->CreateMachineBasicBlock();
+  MachineBasicBlock *RemainderBB = MF->CreateMachineBasicBlock();
+  MachineFunction::iterator MBBI(MBB);
+  ++MBBI;
+
+  MF->insert(MBBI, LoopBB);
+  MF->insert(MBBI, RemainderBB);
+
+  LoopBB->addSuccessor(LoopBB);
+  LoopBB->addSuccessor(RemainderBB);
+
+  // Move the rest of the block into a new block.
+  RemainderBB->transferSuccessorsAndUpdatePHIs(&MBB);
+
+  if (InstInLoop) {
+    auto Next = std::next(I);
+
+    // Move instruction to loop body.
+    LoopBB->splice(LoopBB->begin(), &MBB, I, Next);
+
+    // Move the rest of the block.
+    RemainderBB->splice(RemainderBB->begin(), &MBB, Next, MBB.end());
+  } else {
+    RemainderBB->splice(RemainderBB->begin(), &MBB, I, MBB.end());
+  }
+
+  MBB.addSuccessor(LoopBB);
+
+  return std::make_pair(LoopBB, RemainderBB);
+}
+
+MachineBasicBlock *
+SITargetLowering::emitGWSMemViolTestLoop(MachineInstr &MI,
+                                         MachineBasicBlock *BB) const {
+  const DebugLoc &DL = MI.getDebugLoc();
+
+  MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
+
+  MachineBasicBlock *LoopBB;
+  MachineBasicBlock *RemainderBB;
+  const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
+
+  MachineBasicBlock::iterator Prev = std::prev(MI.getIterator());
+
+  std::tie(LoopBB, RemainderBB) = splitBlockForLoop(MI, *BB, true);
+
+  MachineBasicBlock::iterator I = LoopBB->end();
+
+  MachineOperand *Src = TII->getNamedOperand(MI, AMDGPU::OpName::data0);
+  assert(Src && "missing operand from GWS instruction");
+
+  const unsigned EncodedReg = AMDGPU::Hwreg::encodeHwreg(
+    AMDGPU::Hwreg::ID_TRAPSTS, AMDGPU::Hwreg::OFFSET_MEM_VIOL, 1);
+
+  // Clear TRAP_STS.MEM_VIOL
+  BuildMI(*LoopBB, LoopBB->begin(), DL, TII->get(AMDGPU::S_SETREG_IMM32_B32))
+    .addImm(0)
+    .addImm(EncodedReg);
+
+  // This is a pain, but we're not allowed to have physical register live-ins
+  // yet. Insert a pair of copies if the VGPR0 hack is necessary.
+  if (TargetRegisterInfo::isPhysicalRegister(Src->getReg())) {
+    unsigned Data0 = MRI.createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    BuildMI(*BB, std::next(Prev), DL, TII->get(AMDGPU::COPY), Data0)
+      .add(*Src);
+
+    BuildMI(*LoopBB, LoopBB->begin(), DL, TII->get(AMDGPU::COPY), Src->getReg())
+      .addReg(Data0);
+
+    MRI.setSimpleHint(Data0, Src->getReg());
+  }
+
+  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_WAITCNT))
+    .addImm(0);
+
+  unsigned Reg = MRI.createVirtualRegister(&AMDGPU::SReg_32_XM0RegClass);
+
+  // Load and check TRAP_STS.MEM_VIOL
+  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_GETREG_B32), Reg)
+    .addImm(EncodedReg);
+
+  // FIXME: Do we need to use an isel pseudo that may clobber scc?
+  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_CMP_LG_U32))
+    .addReg(Reg, RegState::Kill)
+    .addImm(0);
+  BuildMI(*LoopBB, I, DL, TII->get(AMDGPU::S_CBRANCH_SCC1))
+    .addMBB(LoopBB);
+
+  return RemainderBB;
+}
+
 // Do a v_movrels_b32 or v_movreld_b32 for each unique value of \p IdxReg in the
 // wavefront. If the value is uniform and just happens to be in a VGPR, this
 // will only do one iteration. In the worst case, this will loop 64 times.
@@ -3061,24 +3164,9 @@ static MachineBasicBlock::iterator loadM0FromVGPR(const SIInstrInfo *TII,
   BuildMI(MBB, I, DL, TII->get(MovExecOpc), SaveExec)
     .addReg(Exec);
 
-  // To insert the loop we need to split the block. Move everything after this
-  // point to a new block, and insert a new empty block between the two.
-  MachineBasicBlock *LoopBB = MF->CreateMachineBasicBlock();
-  MachineBasicBlock *RemainderBB = MF->CreateMachineBasicBlock();
-  MachineFunction::iterator MBBI(MBB);
-  ++MBBI;
-
-  MF->insert(MBBI, LoopBB);
-  MF->insert(MBBI, RemainderBB);
-
-  LoopBB->addSuccessor(LoopBB);
-  LoopBB->addSuccessor(RemainderBB);
-
-  // Move the rest of the block into a new block.
-  RemainderBB->transferSuccessorsAndUpdatePHIs(&MBB);
-  RemainderBB->splice(RemainderBB->begin(), &MBB, I, MBB.end());
-
-  MBB.addSuccessor(LoopBB);
+  MachineBasicBlock *LoopBB;
+  MachineBasicBlock *RemainderBB;
+  std::tie(LoopBB, RemainderBB) = splitBlockForLoop(MI, MBB, false);
 
   const MachineOperand *Idx = TII->getNamedOperand(MI, AMDGPU::OpName::idx);
 
@@ -3630,6 +3718,14 @@ MachineBasicBlock *SITargetLowering::EmitInstrWithCustomInserter(
     MI.eraseFromParent();
     return BB;
   }
+  case AMDGPU::DS_GWS_INIT:
+  case AMDGPU::DS_GWS_SEMA_V:
+  case AMDGPU::DS_GWS_SEMA_BR:
+  case AMDGPU::DS_GWS_SEMA_P:
+  case AMDGPU::DS_GWS_BARRIER:
+    if (getSubtarget()->hasGWSAutoReplay())
+      return BB;
+    return emitGWSMemViolTestLoop(MI, BB);
   default:
     return AMDGPUTargetLowering::EmitInstrWithCustomInserter(MI, BB);
   }
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.h b/llvm/lib/Target/AMDGPU/SIISelLowering.h
index 7d9c05d81ad..b3762dc3483 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.h
@@ -313,6 +313,9 @@ public:
   MachineBasicBlock *splitKillBlock(MachineInstr &MI,
                                     MachineBasicBlock *BB) const;
 
+  MachineBasicBlock *emitGWSMemViolTestLoop(MachineInstr &MI,
+                                            MachineBasicBlock *BB) const;
+
   MachineBasicBlock *
   EmitInstrWithCustomInserter(MachineInstr &MI,
                               MachineBasicBlock *BB) const override;