#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+using namespace llvm;
+
namespace llvm {
+ void initializeR600EmitClauseMarkersPass(PassRegistry&);
+}
-class R600EmitClauseMarkersPass : public MachineFunctionPass {
+namespace {
+
+class R600EmitClauseMarkers : public MachineFunctionPass {
private:
- static char ID;
const R600InstrInfo *TII;
+ int Address;
unsigned OccupiedDwords(MachineInstr *MI) const {
switch (MI->getOpcode()) {
case AMDGPU::INTERP_PAIR_XY:
case AMDGPU::INTERP_PAIR_ZW:
case AMDGPU::INTERP_VEC_LOAD:
- case AMDGPU::DOT4_eg_pseudo:
- case AMDGPU::DOT4_r600_pseudo:
+ case AMDGPU::DOT_4:
return 4;
case AMDGPU::KILL:
return 0;
break;
}
+ // These will be expanded to two ALU instructions in the
+ // ExpandSpecialInstructions pass.
+ if (TII->isLDSRetInstr(MI->getOpcode()))
+ return 2;
+
if(TII->isVector(*MI) ||
TII->isCubeOp(MI->getOpcode()) ||
TII->isReductionOp(MI->getOpcode()))
case AMDGPU::INTERP_PAIR_ZW:
case AMDGPU::INTERP_VEC_LOAD:
case AMDGPU::COPY:
- case AMDGPU::DOT4_eg_pseudo:
- case AMDGPU::DOT4_r600_pseudo:
+ case AMDGPU::DOT_4:
return true;
default:
return false;
switch (MI->getOpcode()) {
case AMDGPU::KILL:
case AMDGPU::RETURN:
+ case AMDGPU::IMPLICIT_DEF:
return true;
default:
return false;
}
bool SubstituteKCacheBank(MachineInstr *MI,
- std::vector<std::pair<unsigned, unsigned> > &CachedConsts) const {
+ std::vector<std::pair<unsigned, unsigned> > &CachedConsts,
+ bool UpdateInstr = true) const {
std::vector<std::pair<unsigned, unsigned> > UsedKCache;
- const SmallVector<std::pair<MachineOperand *, int64_t>, 3> &Consts =
+
+ if (!TII->isALUInstr(MI->getOpcode()) && MI->getOpcode() != AMDGPU::DOT_4)
+ return true;
+
+ const SmallVectorImpl<std::pair<MachineOperand *, int64_t> > &Consts =
TII->getSrcs(MI);
- assert(TII->isALUInstr(MI->getOpcode()) && "Can't assign Const");
+ assert((TII->isALUInstr(MI->getOpcode()) ||
+ MI->getOpcode() == AMDGPU::DOT_4) && "Can't assign Const");
for (unsigned i = 0, n = Consts.size(); i < n; ++i) {
if (Consts[i].first->getReg() != AMDGPU::ALU_CONST)
continue;
return false;
}
+ if (!UpdateInstr)
+ return true;
+
for (unsigned i = 0, j = 0, n = Consts.size(); i < n; ++i) {
if (Consts[i].first->getReg() != AMDGPU::ALU_CONST)
continue;
return true;
}
+ bool canClauseLocalKillFitInClause(
+ unsigned AluInstCount,
+ std::vector<std::pair<unsigned, unsigned> > KCacheBanks,
+ MachineBasicBlock::iterator Def,
+ MachineBasicBlock::iterator BBEnd) {
+ const R600RegisterInfo &TRI = TII->getRegisterInfo();
+ for (MachineInstr::const_mop_iterator
+ MOI = Def->operands_begin(),
+ MOE = Def->operands_end(); MOI != MOE; ++MOI) {
+ if (!MOI->isReg() || !MOI->isDef() ||
+ TRI.isPhysRegLiveAcrossClauses(MOI->getReg()))
+ continue;
+
+ // Def defines a clause local register, so check that its use will fit
+ // in the clause.
+ unsigned LastUseCount = 0;
+ for (MachineBasicBlock::iterator UseI = Def; UseI != BBEnd; ++UseI) {
+ AluInstCount += OccupiedDwords(UseI);
+ // Make sure we won't need to end the clause due to KCache limitations.
+ if (!SubstituteKCacheBank(UseI, KCacheBanks, false))
+ return false;
+
+ // We have reached the maximum instruction limit before finding the
+ // use that kills this register, so we cannot use this def in the
+ // current clause.
+ if (AluInstCount >= TII->getMaxAlusPerClause())
+ return false;
+
+ // Register kill flags have been cleared by the time we get to this
+ // pass, but it is safe to assume that all uses of this register
+ // occur in the same basic block as its definition, because
+ // it is illegal for the scheduler to schedule them in
+ // different blocks.
+ if (UseI->findRegisterUseOperandIdx(MOI->getReg()))
+ LastUseCount = AluInstCount;
+
+ if (UseI != Def && UseI->findRegisterDefOperandIdx(MOI->getReg()) != -1)
+ break;
+ }
+ if (LastUseCount)
+ return LastUseCount <= TII->getMaxAlusPerClause();
+ llvm_unreachable("Clause local register live at end of clause.");
+ }
+ return true;
+ }
+
MachineBasicBlock::iterator
- MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const {
+ MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator I) {
MachineBasicBlock::iterator ClauseHead = I;
std::vector<std::pair<unsigned, unsigned> > KCacheBanks;
bool PushBeforeModifier = false;
if (AluInstCount > TII->getMaxAlusPerClause())
break;
if (I->getOpcode() == AMDGPU::PRED_X) {
+ // We put PRED_X in its own clause to ensure that ifcvt won't create
+ // clauses with more than 128 insts.
+ // IfCvt is indeed checking that "then" and "else" branches of an if
+ // statement have less than ~60 insts thus converted clauses can't be
+ // bigger than ~121 insts (predicate setter needs to be in the same
+ // clause as predicated alus).
+ if (AluInstCount > 0)
+ break;
if (TII->getFlagOp(I).getImm() & MO_FLAG_PUSH)
PushBeforeModifier = true;
AluInstCount ++;
continue;
}
- if (I->getOpcode() == AMDGPU::KILLGT) {
+ // XXX: GROUP_BARRIER instructions cannot be in the same ALU clause as:
+ //
+ // * KILL or INTERP instructions
+ // * Any instruction that sets UPDATE_EXEC_MASK or UPDATE_PRED bits
+ // * Uses waterfalling (i.e. INDEX_MODE = AR.X)
+ //
+ // XXX: These checks have not been implemented yet.
+ if (TII->mustBeLastInClause(I->getOpcode())) {
I++;
break;
}
- if (TII->isALUInstr(I->getOpcode()) &&
- !SubstituteKCacheBank(I, KCacheBanks))
+
+ // If this instruction defines a clause local register, make sure
+ // its use can fit in this clause.
+ if (!canClauseLocalKillFitInClause(AluInstCount, KCacheBanks, I, E))
+ break;
+
+ if (!SubstituteKCacheBank(I, KCacheBanks))
break;
AluInstCount += OccupiedDwords(I);
}
unsigned Opcode = PushBeforeModifier ?
AMDGPU::CF_ALU_PUSH_BEFORE : AMDGPU::CF_ALU;
BuildMI(MBB, ClauseHead, MBB.findDebugLoc(ClauseHead), TII->get(Opcode))
- .addImm(0) // ADDR
+ // We don't use the ADDR field until R600ControlFlowFinalizer pass, where
+ // it is safe to assume it is 0. However if we always put 0 here, the ifcvt
+ // pass may assume that identical ALU clause starter at the beginning of a
+ // true and false branch can be factorized which is not the case.
+ .addImm(Address++) // ADDR
.addImm(KCacheBanks.empty()?0:KCacheBanks[0].first) // KB0
.addImm((KCacheBanks.size() < 2)?0:KCacheBanks[1].first) // KB1
.addImm(KCacheBanks.empty()?0:2) // KM0
.addImm((KCacheBanks.size() < 2)?0:2) // KM1
.addImm(KCacheBanks.empty()?0:KCacheBanks[0].second) // KLINE0
.addImm((KCacheBanks.size() < 2)?0:KCacheBanks[1].second) // KLINE1
- .addImm(AluInstCount); // COUNT
+ .addImm(AluInstCount) // COUNT
+ .addImm(1); // Enabled
return I;
}
public:
- R600EmitClauseMarkersPass(TargetMachine &tm) : MachineFunctionPass(ID),
- TII (static_cast<const R600InstrInfo *>(tm.getInstrInfo())) { }
+ static char ID;
+ R600EmitClauseMarkers() : MachineFunctionPass(ID), TII(nullptr), Address(0) {
+
+ initializeR600EmitClauseMarkersPass(*PassRegistry::getPassRegistry());
+ }
virtual bool runOnMachineFunction(MachineFunction &MF) {
+ TII = static_cast<const R600InstrInfo *>(MF.getTarget().getInstrInfo());
+
for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
BB != BB_E; ++BB) {
MachineBasicBlock &MBB = *BB;
}
};
-char R600EmitClauseMarkersPass::ID = 0;
+char R600EmitClauseMarkers::ID = 0;
-}
+} // end anonymous namespace
+INITIALIZE_PASS_BEGIN(R600EmitClauseMarkers, "emitclausemarkers",
+ "R600 Emit Clause Markters", false, false)
+INITIALIZE_PASS_END(R600EmitClauseMarkers, "emitclausemarkers",
+ "R600 Emit Clause Markters", false, false)
-llvm::FunctionPass *llvm::createR600EmitClauseMarkers(TargetMachine &TM) {
- return new R600EmitClauseMarkersPass(TM);
+llvm::FunctionPass *llvm::createR600EmitClauseMarkers() {
+ return new R600EmitClauseMarkers();
}