//===----------------------------------------------------------------------===//
#include "AMDGPU.h"
+#include "AMDGPUSubtarget.h"
+#include "SIDefines.h"
#include "SIInstrInfo.h"
#include "SIMachineFunctionInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
} Counters;
+typedef enum {
+ OTHER,
+ SMEM,
+ VMEM
+} InstType;
+
typedef Counters RegCounters[512];
typedef std::pair<unsigned, unsigned> RegInterval;
private:
static char ID;
const SIInstrInfo *TII;
- const SIRegisterInfo &TRI;
+ const SIRegisterInfo *TRI;
const MachineRegisterInfo *MRI;
/// \brief Constant hardware limits
/// \brief Different export instruction types seen since last wait.
unsigned ExpInstrTypesSeen;
+ /// \brief Type of the last opcode.
+ InstType LastOpcodeType;
+
+ bool LastInstWritesM0;
+
/// \brief Get increment/decrement amount for this instruction.
Counters getHwCounts(MachineInstr &MI);
RegInterval getRegInterval(MachineOperand &Op);
/// \brief Handle instructions async components
- void pushInstruction(MachineInstr &MI);
+ void pushInstruction(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I);
/// \brief Insert the actual wait instruction
bool insertWait(MachineBasicBlock &MBB,
/// \brief Resolve all operand dependencies to counter requirements
Counters handleOperands(MachineInstr &MI);
+ /// \brief Insert S_NOP between an instruction writing M0 and S_SENDMSG.
+ void handleSendMsg(MachineBasicBlock &MBB, MachineBasicBlock::iterator I);
+
public:
SIInsertWaits(TargetMachine &tm) :
MachineFunctionPass(ID),
- TII(static_cast<const SIInstrInfo*>(tm.getInstrInfo())),
- TRI(TII->getRegisterInfo()) { }
+ TII(nullptr),
+ TRI(nullptr),
+ ExpInstrTypesSeen(0) { }
- virtual bool runOnMachineFunction(MachineFunction &MF);
+ bool runOnMachineFunction(MachineFunction &MF) override;
- const char *getPassName() const {
+ const char *getPassName() const override {
return "SI insert wait instructions";
}
// LGKM may uses larger values
if (TSFlags & SIInstrFlags::LGKM_CNT) {
- MachineOperand &Op = MI.getOperand(0);
- assert(Op.isReg() && "First LGKM operand must be a register!");
+ if (TII->isSMRD(MI.getOpcode())) {
+
+ MachineOperand &Op = MI.getOperand(0);
+ assert(Op.isReg() && "First LGKM operand must be a register!");
+
+ unsigned Reg = Op.getReg();
+ unsigned Size = TRI->getMinimalPhysRegClass(Reg)->getSize();
+ Result.Named.LGKM = Size > 4 ? 2 : 1;
- unsigned Reg = Op.getReg();
- unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize();
- Result.Named.LGKM = Size > 4 ? 2 : 1;
+ } else {
+ // DS
+ Result.Named.LGKM = 1;
+ }
} else {
Result.Named.LGKM = 0;
if (!MI.getDesc().mayStore())
return false;
+ // Check if this operand is the value being stored.
+ // Special case for DS instructions, since the address
+ // operand comes before the value operand and it may have
+ // multiple data operands.
+
+ if (TII->isDS(MI.getOpcode())) {
+ MachineOperand *Data = TII->getNamedOperand(MI, AMDGPU::OpName::data);
+ if (Data && Op.isIdenticalTo(*Data))
+ return true;
+
+ MachineOperand *Data0 = TII->getNamedOperand(MI, AMDGPU::OpName::data0);
+ if (Data0 && Op.isIdenticalTo(*Data0))
+ return true;
+
+ MachineOperand *Data1 = TII->getNamedOperand(MI, AMDGPU::OpName::data1);
+ if (Data1 && Op.isIdenticalTo(*Data1))
+ return true;
+
+ return false;
+ }
+
+ // NOTE: This assumes that the value operand is before the
+ // address operand, and that there is only one value operand.
for (MachineInstr::mop_iterator I = MI.operands_begin(),
E = MI.operands_end(); I != E; ++I) {
RegInterval SIInsertWaits::getRegInterval(MachineOperand &Op) {
- if (!Op.isReg())
+ if (!Op.isReg() || !TRI->isInAllocatableClass(Op.getReg()))
return std::make_pair(0, 0);
unsigned Reg = Op.getReg();
- unsigned Size = TRI.getMinimalPhysRegClass(Reg)->getSize();
+ unsigned Size = TRI->getMinimalPhysRegClass(Reg)->getSize();
assert(Size >= 4);
RegInterval Result;
- Result.first = TRI.getEncodingValue(Reg);
+ Result.first = TRI->getEncodingValue(Reg);
Result.second = Result.first + Size / 4;
return Result;
}
-void SIInsertWaits::pushInstruction(MachineInstr &MI) {
+void SIInsertWaits::pushInstruction(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I) {
// Get the hardware counter increments and sum them up
- Counters Increment = getHwCounts(MI);
+ Counters Increment = getHwCounts(*I);
unsigned Sum = 0;
for (unsigned i = 0; i < 3; ++i) {
}
// If we don't increase anything then that's it
- if (Sum == 0)
+ if (Sum == 0) {
+ LastOpcodeType = OTHER;
return;
+ }
+
+ if (TRI->ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS) {
+ // Any occurence of consecutive VMEM or SMEM instructions forms a VMEM
+ // or SMEM clause, respectively.
+ //
+ // The temporary workaround is to break the clauses with S_NOP.
+ //
+ // The proper solution would be to allocate registers such that all source
+ // and destination registers don't overlap, e.g. this is illegal:
+ // r0 = load r2
+ // r2 = load r0
+ if ((LastOpcodeType == SMEM && TII->isSMRD(I->getOpcode())) ||
+ (LastOpcodeType == VMEM && Increment.Named.VM)) {
+ // Insert a NOP to break the clause.
+ BuildMI(MBB, I, DebugLoc(), TII->get(AMDGPU::S_NOP))
+ .addImm(0);
+ LastInstWritesM0 = false;
+ }
+
+ if (TII->isSMRD(I->getOpcode()))
+ LastOpcodeType = SMEM;
+ else if (Increment.Named.VM)
+ LastOpcodeType = VMEM;
+ }
// Remember which export instructions we have seen
if (Increment.Named.EXP) {
- ExpInstrTypesSeen |= MI.getOpcode() == AMDGPU::EXP ? 1 : 2;
+ ExpInstrTypesSeen |= I->getOpcode() == AMDGPU::EXP ? 1 : 2;
}
- for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
+ for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
- MachineOperand &Op = MI.getOperand(i);
+ MachineOperand &Op = I->getOperand(i);
if (!isOpRelevant(Op))
continue;
continue;
NeedWait = true;
-
+
if (Ordered[i]) {
unsigned Value = LastIssued.Array[i] - Required.Array[i];
- // adjust the value to the real hardware posibilities
+ // Adjust the value to the real hardware possibilities.
Counts.Array[i] = std::min(Value, WaitCounts.Array[i]);
} else
Counts.Array[i] = 0;
- // Remember on what we have waited on
+ // Remember on what we have waited on.
WaitedOn.Array[i] = LastIssued.Array[i] - Counts.Array[i];
}
((Counts.Named.EXP & 0x7) << 4) |
((Counts.Named.LGKM & 0x7) << 8));
+ LastOpcodeType = OTHER;
+ LastInstWritesM0 = false;
return true;
}
Dst.Array[i] = std::max(Dst.Array[i], Src.Array[i]);
}
-bool SIInsertWaits::unorderedDefines(MachineInstr &MI) {
-
- uint64_t TSFlags = TII->get(MI.getOpcode()).TSFlags;
- if (TSFlags & SIInstrFlags::LGKM_CNT)
- return true;
-
- if (TSFlags & SIInstrFlags::EXP_CNT)
- return ExpInstrTypesSeen == 3;
-
- return false;
-}
-
Counters SIInsertWaits::handleOperands(MachineInstr &MI) {
- bool UnorderedDefines = unorderedDefines(MI);
Counters Result = ZeroCounts;
+ // S_SENDMSG implicitly waits for all outstanding LGKM transfers to finish,
+ // but we also want to wait for any other outstanding transfers before
+ // signalling other hardware blocks
+ if (MI.getOpcode() == AMDGPU::S_SENDMSG)
+ return LastIssued;
+
// For each register affected by this
// instruction increase the result sequence
for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
if (Op.isDef()) {
increaseCounters(Result, UsedRegs[j]);
- if (UnorderedDefines)
- increaseCounters(Result, DefinedRegs[j]);
+ increaseCounters(Result, DefinedRegs[j]);
}
if (Op.isUse())
return Result;
}
-bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
+void SIInsertWaits::handleSendMsg(MachineBasicBlock &MBB,
+ MachineBasicBlock::iterator I) {
+ if (TRI->ST.getGeneration() < AMDGPUSubtarget::VOLCANIC_ISLANDS)
+ return;
+ // There must be "S_NOP 0" between an instruction writing M0 and S_SENDMSG.
+ if (LastInstWritesM0 && I->getOpcode() == AMDGPU::S_SENDMSG) {
+ BuildMI(MBB, I, DebugLoc(), TII->get(AMDGPU::S_NOP)).addImm(0);
+ LastInstWritesM0 = false;
+ return;
+ }
+
+ // Set whether this instruction sets M0
+ LastInstWritesM0 = false;
+
+ unsigned NumOperands = I->getNumOperands();
+ for (unsigned i = 0; i < NumOperands; i++) {
+ const MachineOperand &Op = I->getOperand(i);
+
+ if (Op.isReg() && Op.isDef() && Op.getReg() == AMDGPU::M0)
+ LastInstWritesM0 = true;
+ }
+}
+
+// FIXME: Insert waits listed in Table 4.2 "Required User-Inserted Wait States"
+// around other non-memory instructions.
+bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
bool Changes = false;
+ TII = static_cast<const SIInstrInfo *>(MF.getSubtarget().getInstrInfo());
+ TRI =
+ static_cast<const SIRegisterInfo *>(MF.getSubtarget().getRegisterInfo());
+
MRI = &MF.getRegInfo();
WaitedOn = ZeroCounts;
LastIssued = ZeroCounts;
+ LastOpcodeType = OTHER;
+ LastInstWritesM0 = false;
memset(&UsedRegs, 0, sizeof(UsedRegs));
memset(&DefinedRegs, 0, sizeof(DefinedRegs));
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
I != E; ++I) {
- Changes |= insertWait(MBB, I, handleOperands(*I));
- pushInstruction(*I);
+ // Wait for everything before a barrier.
+ if (I->getOpcode() == AMDGPU::S_BARRIER)
+ Changes |= insertWait(MBB, I, LastIssued);
+ else
+ Changes |= insertWait(MBB, I, handleOperands(*I));
+
+ pushInstruction(MBB, I);
+ handleSendMsg(MBB, I);
}
// Wait for everything at the end of the MBB