6502/llvm-6502
1
0
Fork 0
mirror of https://github.com/c64scene-ar/llvm-6502.git synced 2025-06-18 11:24:01 +00:00
Code Issues Projects Releases Wiki Activity

LoopVectorizer: Add support for loop-unrolling during vectorization for increasing the ILP. At the moment this feature is disabled by default and this commit should not cause any functional changes.

Browse Source 
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@171436 91177308-0d34-0410-b5e6-96231b3b80d8
This commit is contained in:
Nadav Rotem
2013-01-03 00:52:27 +00:00
parent 251ed7f3e5
commit e4159491a7
2 changed files with 329 additions and 169 deletions
Download Patch File Download Diff File Expand all files Collapse all files

422
lib/Transforms/Vectorize/LoopVectorize.cpp
View File

@ -42,6 +42,11 @@ static cl::opt<unsigned>
VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden, VectorizationFactor("force-vector-width", cl::init(0), cl::Hidden,
cl::desc("Sets the SIMD width. Zero is autoselect.")); cl::desc("Sets the SIMD width. Zero is autoselect."));
static cl::opt<unsigned>
VectorizationUnroll("force-vector-unroll", cl::init(1), cl::Hidden,
cl::desc("Sets the vectorization unroll count. "
"Zero is autoselect."));
static cl::opt<bool> static cl::opt<bool>
EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden, EnableIfConversion("enable-if-conversion", cl::init(true), cl::Hidden,
cl::desc("Enable if-conversion during vectorization.")); cl::desc("Enable if-conversion during vectorization."));
@ -117,7 +122,7 @@ struct LoopVectorize : public LoopPass {
F->getParent()->getModuleIdentifier()<<"\n"); F->getParent()->getModuleIdentifier()<<"\n");
// If we decided that it is *legal* to vectorizer the loop then do it. // If we decided that it is *legal* to vectorizer the loop then do it.
InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF); InnerLoopVectorizer LB(L, SE, LI, DT, DL, VF, VectorizationUnroll);
LB.vectorize(&LVL); LB.vectorize(&LVL);
DEBUG(verifyFunction(*L->getHeader()->getParent())); DEBUG(verifyFunction(*L->getHeader()->getParent()));
@ -180,7 +185,8 @@ Value *InnerLoopVectorizer::getBroadcastInstrs(Value *V) {
return Shuf; return Shuf;
} }
Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) { Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, unsigned StartIdx,
bool Negate) {
assert(Val->getType()->isVectorTy() && "Must be a vector"); assert(Val->getType()->isVectorTy() && "Must be a vector");
assert(Val->getType()->getScalarType()->isIntegerTy() && assert(Val->getType()->getScalarType()->isIntegerTy() &&
"Elem must be an integer"); "Elem must be an integer");
@ -191,8 +197,10 @@ Value *InnerLoopVectorizer::getConsecutiveVector(Value* Val, bool Negate) {
SmallVector<Constant*, 8> Indices; SmallVector<Constant*, 8> Indices;
// Create a vector of consecutive numbers from zero to VF. // Create a vector of consecutive numbers from zero to VF.
for (int i = 0; i < VLen; ++i) for (int i = 0; i < VLen; ++i) {
Indices.push_back(ConstantInt::get(ITy, Negate ? (-i): i )); int Idx = Negate ? (-i): i;
Indices.push_back(ConstantInt::get(ITy, StartIdx + Idx));
}
// Add the consecutive indices to the vector value. // Add the consecutive indices to the vector value.
Constant *Cv = ConstantVector::get(Indices); Constant *Cv = ConstantVector::get(Indices);
@ -244,18 +252,20 @@ bool LoopVectorizationLegality::isUniform(Value *V) {
return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop)); return (SE->isLoopInvariant(SE->getSCEV(V), TheLoop));
} }
Value *InnerLoopVectorizer::getVectorValue(Value *V) { InnerLoopVectorizer::VectorParts&
InnerLoopVectorizer::getVectorValue(Value *V) {
assert(V != Induction && "The new induction variable should not be used."); assert(V != Induction && "The new induction variable should not be used.");
assert(!V->getType()->isVectorTy() && "Can't widen a vector"); assert(!V->getType()->isVectorTy() && "Can't widen a vector");
// If we saved a vectorized copy of V, use it.
Value *&MapEntry = WidenMap[V];
if (MapEntry)
return MapEntry;
// Broadcast V and save the value for future uses. // If we have this scalar in the map, return it.
if (WidenMap.has(V))
return WidenMap.get(V);
// If this scalar is unknown, assume that it is a constant or that it is
// loop invariant. Broadcast V and save the value for future uses.
Value *B = getBroadcastInstrs(V); Value *B = getBroadcastInstrs(V);
MapEntry = B; WidenMap.splat(V, B);
return B; return WidenMap.get(V);
} }
Constant* Constant*
@ -277,7 +287,7 @@ Value *InnerLoopVectorizer::reverseVector(Value *Vec) {
void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) { void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
assert(!Instr->getType()->isAggregateType() && "Can't handle vectors"); assert(!Instr->getType()->isAggregateType() && "Can't handle vectors");
// Holds vector parameters or scalars, in case of uniform vals. // Holds vector parameters or scalars, in case of uniform vals.
SmallVector<Value*, 8> Params; SmallVector<VectorParts, 4> Params;
// Find all of the vectorized parameters. // Find all of the vectorized parameters.
for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
@ -295,12 +305,14 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
// If the src is an instruction that appeared earlier in the basic block // If the src is an instruction that appeared earlier in the basic block
// then it should already be vectorized. // then it should already be vectorized.
if (SrcInst && OrigLoop->contains(SrcInst)) { if (SrcInst && OrigLoop->contains(SrcInst)) {
assert(WidenMap.count(SrcInst) && "Source operand is unavailable"); assert(WidenMap.has(SrcInst) && "Source operand is unavailable");
// The parameter is a vector value from earlier. // The parameter is a vector value from earlier.
Params.push_back(WidenMap[SrcInst]); Params.push_back(WidenMap.get(SrcInst));
} else { } else {
// The parameter is a scalar from outside the loop. Maybe even a constant. // The parameter is a scalar from outside the loop. Maybe even a constant.
Params.push_back(SrcOp); VectorParts Scalars;
Scalars.append(UF, SrcOp);
Params.push_back(Scalars);
} }
} }
@ -309,39 +321,38 @@ void InnerLoopVectorizer::scalarizeInstruction(Instruction *Instr) {
// Does this instruction return a value ? // Does this instruction return a value ?
bool IsVoidRetTy = Instr->getType()->isVoidTy(); bool IsVoidRetTy = Instr->getType()->isVoidTy();
Value *VecResults = 0;
// If we have a return value, create an empty vector. We place the scalarized Value *UndefVec = IsVoidRetTy ? 0 :
// instructions in this vector. UndefValue::get(VectorType::get(Instr->getType(), VF));
if (!IsVoidRetTy) // Create a new entry in the WidenMap and initialize it to Undef or Null.
VecResults = UndefValue::get(VectorType::get(Instr->getType(), VF)); VectorParts &VecResults = WidenMap.splat(Instr, UndefVec);
// For each scalar that we create: // For each scalar that we create:
for (unsigned i = 0; i < VF; ++i) { for (unsigned Width = 0; Width < VF; ++Width) {
Instruction *Cloned = Instr->clone(); // For each vector unroll 'part':
if (!IsVoidRetTy) for (unsigned Part = 0; Part < UF; ++Part) {
Cloned->setName(Instr->getName() + ".cloned"); Instruction *Cloned = Instr->clone();
// Replace the operands of the cloned instrucions with extracted scalars. if (!IsVoidRetTy)
for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) { Cloned->setName(Instr->getName() + ".cloned");
Value *Op = Params[op]; // Replace the operands of the cloned instrucions with extracted scalars.
// Param is a vector. Need to extract the right lane. for (unsigned op = 0, e = Instr->getNumOperands(); op != e; ++op) {
if (Op->getType()->isVectorTy()) Value *Op = Params[op][Part];
Op = Builder.CreateExtractElement(Op, Builder.getInt32(i)); // Param is a vector. Need to extract the right lane.
Cloned->setOperand(op, Op); if (Op->getType()->isVectorTy())
Op = Builder.CreateExtractElement(Op, Builder.getInt32(Width));
Cloned->setOperand(op, Op);
}
// Place the cloned scalar in the new loop.
Builder.Insert(Cloned);
// If the original scalar returns a value we need to place it in a vector
// so that future users will be able to use it.
if (!IsVoidRetTy)
VecResults[Part] = Builder.CreateInsertElement(VecResults[Part], Cloned,
Builder.getInt32(Width));
} }
// Place the cloned scalar in the new loop.
Builder.Insert(Cloned);
// If the original scalar returns a value we need to place it in a vector
// so that future users will be able to use it.
if (!IsVoidRetTy)
VecResults = Builder.CreateInsertElement(VecResults, Cloned,
Builder.getInt32(i));
} }
if (!IsVoidRetTy)
WidenMap[Instr] = VecResults;
} }
Value* Value*
@ -503,7 +514,9 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
// Generate the induction variable. // Generate the induction variable.
Induction = Builder.CreatePHI(IdxTy, 2, "index"); Induction = Builder.CreatePHI(IdxTy, 2, "index");
Constant *Step = ConstantInt::get(IdxTy, VF); // The loop step is equal to the vectorization factor (num of SIMD elements)
// times the unroll factor (num of SIMD instructions).
Constant *Step = ConstantInt::get(IdxTy, VF * UF);
// We may need to extend the index in case there is a type mismatch. // We may need to extend the index in case there is a type mismatch.
// We know that the count starts at zero and does not overflow. // We know that the count starts at zero and does not overflow.
@ -521,8 +534,7 @@ InnerLoopVectorizer::createEmptyLoop(LoopVectorizationLegality *Legal) {
// Now we need to generate the expression for N - (N % VF), which is // Now we need to generate the expression for N - (N % VF), which is
// the part that the vectorized body will execute. // the part that the vectorized body will execute.
Constant *CIVF = ConstantInt::get(IdxTy, VF); Value *R = BinaryOperator::CreateURem(Count, Step, "n.mod.vf", Loc);
Value *R = BinaryOperator::CreateURem(Count, CIVF, "n.mod.vf", Loc);
Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc); Value *CountRoundDown = BinaryOperator::CreateSub(Count, R, "n.vec", Loc);
Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx, Value *IdxEndRoundDown = BinaryOperator::CreateAdd(CountRoundDown, StartIdx,
"end.idx.rnd.down", Loc); "end.idx.rnd.down", Loc);
@ -775,7 +787,6 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end(); for (PhiVector::iterator it = RdxPHIsToFix.begin(), e = RdxPHIsToFix.end();
it != e; ++it) { it != e; ++it) {
PHINode *RdxPhi = *it; PHINode *RdxPhi = *it;
PHINode *VecRdxPhi = dyn_cast<PHINode>(WidenMap[RdxPhi]);
assert(RdxPhi && "Unable to recover vectorized PHI"); assert(RdxPhi && "Unable to recover vectorized PHI");
// Find the reduction variable descriptor. // Find the reduction variable descriptor.
@ -791,8 +802,8 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
Builder.SetInsertPoint(LoopBypassBlock->getTerminator()); Builder.SetInsertPoint(LoopBypassBlock->getTerminator());
// This is the vector-clone of the value that leaves the loop. // This is the vector-clone of the value that leaves the loop.
Value *VectorExit = getVectorValue(RdxDesc.LoopExitInstr); VectorParts &VectorExit = getVectorValue(RdxDesc.LoopExitInstr);
Type *VecTy = VectorExit->getType(); Type *VecTy = VectorExit[0]->getType();
// Find the reduction identity variable. Zero for addition, or, xor, // Find the reduction identity variable. Zero for addition, or, xor,
// one for multiplication, -1 for And. // one for multiplication, -1 for And.
@ -811,10 +822,17 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
// Reductions do not have to start at zero. They can start with // Reductions do not have to start at zero. They can start with
// any loop invariant values. // any loop invariant values.
VecRdxPhi->addIncoming(VectorStart, VecPreheader); VectorParts &VecRdxPhi = WidenMap.get(RdxPhi);
Value *Val = BasicBlock *Latch = OrigLoop->getLoopLatch();
getVectorValue(RdxPhi->getIncomingValueForBlock(OrigLoop->getLoopLatch())); Value *LoopVal = RdxPhi->getIncomingValueForBlock(Latch);
VecRdxPhi->addIncoming(Val, LoopVectorBody); VectorParts &Val = getVectorValue(LoopVal);
for (unsigned part = 0; part < UF; ++part) {
// Make sure to add the reduction stat value only to the
// first unroll part.
Value *StartVal = (part == 0) ? VectorStart : Identity;
cast<PHINode>(VecRdxPhi[part])->addIncoming(StartVal, VecPreheader);
cast<PHINode>(VecRdxPhi[part])->addIncoming(Val[part], LoopVectorBody);
}
// Before each round, move the insertion point right between // Before each round, move the insertion point right between
// the PHIs and the values we are going to write. // the PHIs and the values we are going to write.
@ -822,18 +840,54 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
// instructions. // instructions.
Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt()); Builder.SetInsertPoint(LoopMiddleBlock->getFirstInsertionPt());
// This PHINode contains the vectorized reduction variable, or VectorParts RdxParts;
// the initial value vector, if we bypass the vector loop. for (unsigned part = 0; part < UF; ++part) {
PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi"); // This PHINode contains the vectorized reduction variable, or
NewPhi->addIncoming(VectorStart, LoopBypassBlock); // the initial value vector, if we bypass the vector loop.
NewPhi->addIncoming(getVectorValue(RdxDesc.LoopExitInstr), LoopVectorBody); VectorParts &RdxExitVal = getVectorValue(RdxDesc.LoopExitInstr);
PHINode *NewPhi = Builder.CreatePHI(VecTy, 2, "rdx.vec.exit.phi");
Value *StartVal = (part == 0) ? VectorStart : Identity;
NewPhi->addIncoming(StartVal, LoopBypassBlock);
NewPhi->addIncoming(RdxExitVal[part], LoopVectorBody);
RdxParts.push_back(NewPhi);
}
// Reduce all of the unrolled parts into a single vector.
Value *ReducedPartRdx = RdxParts[0];
for (unsigned part = 1; part < UF; ++part) {
switch (RdxDesc.Kind) {
case LoopVectorizationLegality::IntegerAdd:
ReducedPartRdx =
Builder.CreateAdd(RdxParts[part], ReducedPartRdx, "add.rdx");
break;
case LoopVectorizationLegality::IntegerMult:
ReducedPartRdx =
Builder.CreateMul(RdxParts[part], ReducedPartRdx, "mul.rdx");
break;
case LoopVectorizationLegality::IntegerOr:
ReducedPartRdx =
Builder.CreateOr(RdxParts[part], ReducedPartRdx, "or.rdx");
break;
case LoopVectorizationLegality::IntegerAnd:
ReducedPartRdx =
Builder.CreateAnd(RdxParts[part], ReducedPartRdx, "and.rdx");
break;
case LoopVectorizationLegality::IntegerXor:
ReducedPartRdx =
Builder.CreateXor(RdxParts[part], ReducedPartRdx, "xor.rdx");
break;
default:
llvm_unreachable("Unknown reduction operation");
}
}
// VF is a power of 2 so we can emit the reduction using log2(VF) shuffles // VF is a power of 2 so we can emit the reduction using log2(VF) shuffles
// and vector ops, reducing the set of values being computed by half each // and vector ops, reducing the set of values being computed by half each
// round. // round.
assert(isPowerOf2_32(VF) && assert(isPowerOf2_32(VF) &&
"Reduction emission only supported for pow2 vectors!"); "Reduction emission only supported for pow2 vectors!");
Value *TmpVec = NewPhi; Value *TmpVec = ReducedPartRdx;
SmallVector<Constant*, 32> ShuffleMask(VF, 0); SmallVector<Constant*, 32> ShuffleMask(VF, 0);
for (unsigned i = VF; i != 1; i >>= 1) { for (unsigned i = VF; i != 1; i >>= 1) {
// Move the upper half of the vector to the lower half. // Move the upper half of the vector to the lower half.
@ -922,27 +976,34 @@ InnerLoopVectorizer::vectorizeLoop(LoopVectorizationLegality *Legal) {
} }
} }
Value *InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) { InnerLoopVectorizer::VectorParts
InnerLoopVectorizer::createEdgeMask(BasicBlock *Src, BasicBlock *Dst) {
assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) && assert(std::find(pred_begin(Dst), pred_end(Dst), Src) != pred_end(Dst) &&
"Invalid edge"); "Invalid edge");
Value *SrcMask = createBlockInMask(Src); VectorParts SrcMask = createBlockInMask(Src);
// The terminator has to be a branch inst! // The terminator has to be a branch inst!
BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator()); BranchInst *BI = dyn_cast<BranchInst>(Src->getTerminator());
assert(BI && "Unexpected terminator found"); assert(BI && "Unexpected terminator found");
Value *EdgeMask = SrcMask;
if (BI->isConditional()) { if (BI->isConditional()) {
EdgeMask = getVectorValue(BI->getCondition()); VectorParts EdgeMask = getVectorValue(BI->getCondition());
if (BI->getSuccessor(0) != Dst) if (BI->getSuccessor(0) != Dst)
EdgeMask = Builder.CreateNot(EdgeMask); for (unsigned part = 0; part < UF; ++part)
EdgeMask[part] = Builder.CreateNot(EdgeMask[part]);
for (unsigned part = 0; part < UF; ++part)
EdgeMask[part] = Builder.CreateAnd(EdgeMask[part], SrcMask[part]);
return EdgeMask;
} }
return Builder.CreateAnd(EdgeMask, SrcMask); return SrcMask;
} }
Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) { InnerLoopVectorizer::VectorParts
InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
assert(OrigLoop->contains(BB) && "Block is not a part of a loop"); assert(OrigLoop->contains(BB) && "Block is not a part of a loop");
// Loop incoming mask is all-one. // Loop incoming mask is all-one.
@ -953,11 +1014,14 @@ Value *InnerLoopVectorizer::createBlockInMask(BasicBlock *BB) {
// This is the block mask. We OR all incoming edges, and with zero. // This is the block mask. We OR all incoming edges, and with zero.
Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0); Value *Zero = ConstantInt::get(IntegerType::getInt1Ty(BB->getContext()), 0);
Value *BlockMask = getVectorValue(Zero); VectorParts BlockMask = getVectorValue(Zero);
// For each pred: // For each pred:
for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) for (pred_iterator it = pred_begin(BB), e = pred_end(BB); it != e; ++it) {
BlockMask = Builder.CreateOr(BlockMask, createEdgeMask(*it, BB)); VectorParts EM = createEdgeMask(*it, BB);
for (unsigned part = 0; part < UF; ++part)
BlockMask[part] = Builder.CreateOr(BlockMask[part], EM[part]);
}
return BlockMask; return BlockMask;
} }
@ -969,6 +1033,7 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
// For each instruction in the old loop. // For each instruction in the old loop.
for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) { for (BasicBlock::iterator it = BB->begin(), e = BB->end(); it != e; ++it) {
VectorParts &Entry = WidenMap.get(it);
switch (it->getOpcode()) { switch (it->getOpcode()) {
case Instruction::Br: case Instruction::Br:
// Nothing to do for PHIs and BR, since we already took care of the // Nothing to do for PHIs and BR, since we already took care of the
@ -978,11 +1043,12 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
PHINode* P = cast<PHINode>(it); PHINode* P = cast<PHINode>(it);
// Handle reduction variables: // Handle reduction variables:
if (Legal->getReductionVars()->count(P)) { if (Legal->getReductionVars()->count(P)) {
// This is phase one of vectorizing PHIs. for (unsigned part = 0; part < UF; ++part) {
Type *VecTy = VectorType::get(it->getType(), VF); // This is phase one of vectorizing PHIs.
WidenMap[it] = Type *VecTy = VectorType::get(it->getType(), VF);
PHINode::Create(VecTy, 2, "vec.phi", Entry[part] = PHINode::Create(VecTy, 2, "vec.phi",
LoopVectorBody->getFirstInsertionPt()); LoopVectorBody-> getFirstInsertionPt());
}
PV->push_back(P); PV->push_back(P);
continue; continue;
} }
@ -996,12 +1062,15 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
// At this point we generate the predication tree. There may be // At this point we generate the predication tree. There may be
// duplications since this is a simple recursive scan, but future // duplications since this is a simple recursive scan, but future
// optimizations will clean it up. // optimizations will clean it up.
Value *Cond = createEdgeMask(P->getIncomingBlock(0), P->getParent()); VectorParts Cond = createEdgeMask(P->getIncomingBlock(0),
WidenMap[P] = P->getParent());
Builder.CreateSelect(Cond,
getVectorValue(P->getIncomingValue(0)), for (unsigned part = 0; part < UF; ++part) {
getVectorValue(P->getIncomingValue(1)), VectorParts &In0 = getVectorValue(P->getIncomingValue(0));
"predphi"); VectorParts &In1 = getVectorValue(P->getIncomingValue(1));
Entry[part] = Builder.CreateSelect(Cond[part], In0[part], In1[part],
"predphi");
}
continue; continue;
} }
@ -1021,8 +1090,8 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
Value *Broadcasted = getBroadcastInstrs(Induction); Value *Broadcasted = getBroadcastInstrs(Induction);
// After broadcasting the induction variable we need to make the // After broadcasting the induction variable we need to make the
// vector consecutive by adding 0, 1, 2 ... // vector consecutive by adding 0, 1, 2 ...
Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted); for (unsigned part = 0; part < UF; ++part)
WidenMap[OldInduction] = ConsecutiveInduction; Entry[part] = getConsecutiveVector(Broadcasted, VF * part, false);
continue; continue;
} }
case LoopVectorizationLegality::ReverseIntInduction: case LoopVectorizationLegality::ReverseIntInduction:
@ -1054,9 +1123,8 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
Value *Broadcasted = getBroadcastInstrs(ReverseInd); Value *Broadcasted = getBroadcastInstrs(ReverseInd);
// After broadcasting the induction variable we need to make the // After broadcasting the induction variable we need to make the
// vector consecutive by adding ... -3, -2, -1, 0. // vector consecutive by adding ... -3, -2, -1, 0.
Value *ConsecutiveInduction = getConsecutiveVector(Broadcasted, for (unsigned part = 0; part < UF; ++part)
true); Entry[part] = getConsecutiveVector(Broadcasted, -VF * part, true);
WidenMap[it] = ConsecutiveInduction;
continue; continue;
} }
@ -1065,19 +1133,21 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
// This is the vector of results. Notice that we don't generate // This is the vector of results. Notice that we don't generate
// vector geps because scalar geps result in better code. // vector geps because scalar geps result in better code.
Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF)); for (unsigned part = 0; part < UF; ++part) {
for (unsigned int i = 0; i < VF; ++i) { Value *VecVal = UndefValue::get(VectorType::get(P->getType(), VF));
Constant *Idx = ConstantInt::get(Induction->getType(), i); for (unsigned int i = 0; i < VF; ++i) {
Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx, Constant *Idx = ConstantInt::get(Induction->getType(),
"gep.idx"); i + part * VF);
Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx, Value *GlobalIdx = Builder.CreateAdd(NormalizedIdx, Idx,
"next.gep"); "gep.idx");
VecVal = Builder.CreateInsertElement(VecVal, SclrGep, Value *SclrGep = Builder.CreateGEP(II.StartValue, GlobalIdx,
Builder.getInt32(i), "next.gep");
"insert.gep"); VecVal = Builder.CreateInsertElement(VecVal, SclrGep,
Builder.getInt32(i),
"insert.gep");
}
Entry[part] = VecVal;
} }
WidenMap[it] = VecVal;
continue; continue;
} }
@ -1103,41 +1173,48 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
case Instruction::Xor: { case Instruction::Xor: {
// Just widen binops. // Just widen binops.
BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it); BinaryOperator *BinOp = dyn_cast<BinaryOperator>(it);
Value *A = getVectorValue(it->getOperand(0)); VectorParts &A = getVectorValue(it->getOperand(0));
Value *B = getVectorValue(it->getOperand(1)); VectorParts &B = getVectorValue(it->getOperand(1));
// Use this vector value for all users of the original instruction. // Use this vector value for all users of the original instruction.
Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A, B); for (unsigned Part = 0; Part < UF; ++Part) {
WidenMap[it] = V; Value *V = Builder.CreateBinOp(BinOp->getOpcode(), A[Part], B[Part]);
// Update the NSW, NUW and Exact flags. // Update the NSW, NUW and Exact flags.
BinaryOperator *VecOp = cast<BinaryOperator>(V); BinaryOperator *VecOp = cast<BinaryOperator>(V);
if (isa<OverflowingBinaryOperator>(BinOp)) { if (isa<OverflowingBinaryOperator>(BinOp)) {
VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap()); VecOp->setHasNoSignedWrap(BinOp->hasNoSignedWrap());
VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap()); VecOp->setHasNoUnsignedWrap(BinOp->hasNoUnsignedWrap());
}
if (isa<PossiblyExactOperator>(VecOp))
VecOp->setIsExact(BinOp->isExact());
Entry[Part] = V;
} }
if (isa<PossiblyExactOperator>(VecOp))
VecOp->setIsExact(BinOp->isExact());
break; break;
} }
case Instruction::Select: { case Instruction::Select: {
// Widen selects. // Widen selects.
// If the selector is loop invariant we can create a select // If the selector is loop invariant we can create a select
// instruction with a scalar condition. Otherwise, use vector-select. // instruction with a scalar condition. Otherwise, use vector-select.
Value *Cond = it->getOperand(0); bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(it->getOperand(0)),
bool InvariantCond = SE->isLoopInvariant(SE->getSCEV(Cond), OrigLoop); OrigLoop);
// The condition can be loop invariant but still defined inside the // The condition can be loop invariant but still defined inside the
// loop. This means that we can't just use the original 'cond' value. // loop. This means that we can't just use the original 'cond' value.
// We have to take the 'vectorized' value and pick the first lane. // We have to take the 'vectorized' value and pick the first lane.
// Instcombine will make this a no-op. // Instcombine will make this a no-op.
Cond = getVectorValue(Cond); VectorParts &Cond = getVectorValue(it->getOperand(0));
if (InvariantCond) VectorParts &Op0 = getVectorValue(it->getOperand(1));
Cond = Builder.CreateExtractElement(Cond, Builder.getInt32(0)); VectorParts &Op1 = getVectorValue(it->getOperand(2));
Value *ScalarCond = Builder.CreateExtractElement(Cond[0],
Value *Op0 = getVectorValue(it->getOperand(1)); Builder.getInt32(0));
Value *Op1 = getVectorValue(it->getOperand(2)); for (unsigned Part = 0; Part < UF; ++Part) {
WidenMap[it] = Builder.CreateSelect(Cond, Op0, Op1); Entry[Part] = Builder.CreateSelect(
InvariantCond ? ScalarCond : Cond[Part],
Op0[Part],
Op1[Part]);
}
break; break;
} }
@ -1146,12 +1223,16 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
// Widen compares. Generate vector compares. // Widen compares. Generate vector compares.
bool FCmp = (it->getOpcode() == Instruction::FCmp); bool FCmp = (it->getOpcode() == Instruction::FCmp);
CmpInst *Cmp = dyn_cast<CmpInst>(it); CmpInst *Cmp = dyn_cast<CmpInst>(it);
Value *A = getVectorValue(it->getOperand(0)); VectorParts &A = getVectorValue(it->getOperand(0));
Value *B = getVectorValue(it->getOperand(1)); VectorParts &B = getVectorValue(it->getOperand(1));
if (FCmp) for (unsigned Part = 0; Part < UF; ++Part) {
WidenMap[it] = Builder.CreateFCmp(Cmp->getPredicate(), A, B); Value *C = 0;
else if (FCmp)
WidenMap[it] = Builder.CreateICmp(Cmp->getPredicate(), A, B); C = Builder.CreateFCmp(Cmp->getPredicate(), A[Part], B[Part]);
else
C = Builder.CreateICmp(Cmp->getPredicate(), A[Part], B[Part]);
Entry[Part] = C;
}
break; break;
} }
@ -1173,12 +1254,17 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
break; break;
} }
// Handle consecutive stores.
GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr); GetElementPtrInst *Gep = dyn_cast<GetElementPtrInst>(Ptr);
if (Gep) { if (Gep) {
// The last index does not have to be the induction. It can be // The last index does not have to be the induction. It can be
// consecutive and be a function of the index. For example A[I+1]; // consecutive and be a function of the index. For example A[I+1];
unsigned NumOperands = Gep->getNumOperands(); unsigned NumOperands = Gep->getNumOperands();
Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands - 1));
Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
VectorParts &GEPParts = getVectorValue(LastGepOperand);
Value *LastIndex = GEPParts[0];
LastIndex = Builder.CreateExtractElement(LastIndex, Zero); LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
// Create the new GEP with the new induction variable. // Create the new GEP with the new induction variable.
@ -1188,19 +1274,28 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
} else { } else {
// Use the induction element ptr. // Use the induction element ptr.
assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero); VectorParts &PtrVal = getVectorValue(Ptr);
Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
} }
// If the address is consecutive but reversed, then the VectorParts &StoredVal = getVectorValue(SI->getValueOperand());
// wide load needs to start at the last vector element. for (unsigned Part = 0; Part < UF; ++Part) {
if (Reverse) // Calculate the pointer for the specific unroll-part.
Ptr = Builder.CreateGEP(Ptr, Builder.getInt32(1 - VF)); Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
Ptr = Builder.CreateBitCast(Ptr, StTy->getPointerTo()); if (Reverse) {
Value *Val = getVectorValue(SI->getValueOperand()); // If we store to reverse consecutive memory locations then we need
if (Reverse) // to reverse the order of elements in the stored value.
Val = reverseVector(Val); StoredVal[Part] = reverseVector(StoredVal[Part]);
Builder.CreateStore(Val, Ptr)->setAlignment(Alignment); // If the address is consecutive but reversed, then the
// wide store needs to start at the last vector element.
PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
}
Value *VecPtr = Builder.CreateBitCast(PartPtr, StTy->getPointerTo());
Builder.CreateStore(StoredVal[Part], VecPtr)->setAlignment(Alignment);
}
break; break;
} }
case Instruction::Load: { case Instruction::Load: {
@ -1224,7 +1319,10 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
// The last index does not have to be the induction. It can be // The last index does not have to be the induction. It can be
// consecutive and be a function of the index. For example A[I+1]; // consecutive and be a function of the index. For example A[I+1];
unsigned NumOperands = Gep->getNumOperands(); unsigned NumOperands = Gep->getNumOperands();
Value *LastIndex = getVectorValue(Gep->getOperand(NumOperands -1));
Value *LastGepOperand = Gep->getOperand(NumOperands - 1);
VectorParts &GEPParts = getVectorValue(LastGepOperand);
Value *LastIndex = GEPParts[0];
LastIndex = Builder.CreateExtractElement(LastIndex, Zero); LastIndex = Builder.CreateExtractElement(LastIndex, Zero);
// Create the new GEP with the new induction variable. // Create the new GEP with the new induction variable.
@ -1234,19 +1332,26 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
} else { } else {
// Use the induction element ptr. // Use the induction element ptr.
assert(isa<PHINode>(Ptr) && "Invalid induction ptr"); assert(isa<PHINode>(Ptr) && "Invalid induction ptr");
Ptr = Builder.CreateExtractElement(getVectorValue(Ptr), Zero); VectorParts &PtrVal = getVectorValue(Ptr);
Ptr = Builder.CreateExtractElement(PtrVal[0], Zero);
} }
// If the address is consecutive but reversed, then the
// wide load needs to start at the last vector element.
if (Reverse)
Ptr = Builder.CreateGEP(Ptr, Builder.getInt32(1 - VF));
Ptr = Builder.CreateBitCast(Ptr, RetTy->getPointerTo()); for (unsigned Part = 0; Part < UF; ++Part) {
LI = Builder.CreateLoad(Ptr); // Calculate the pointer for the specific unroll-part.
LI->setAlignment(Alignment); Value *PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(Part * VF));
// Use this vector value for all users of the load. if (Reverse) {
WidenMap[it] = Reverse ? reverseVector(LI) : LI; // If the address is consecutive but reversed, then the
// wide store needs to start at the last vector element.
PartPtr = Builder.CreateGEP(Ptr, Builder.getInt32(-Part * VF));
PartPtr = Builder.CreateGEP(PartPtr, Builder.getInt32(1 - VF));
}
Value *VecPtr = Builder.CreateBitCast(PartPtr, RetTy->getPointerTo());
Value *LI = Builder.CreateLoad(VecPtr, "wide.load");
cast<LoadInst>(LI)->setAlignment(Alignment);
Entry[Part] = Reverse ? reverseVector(LI) : LI;
}
break; break;
} }
case Instruction::ZExt: case Instruction::ZExt:
@ -1271,13 +1376,16 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction, Value *ScalarCast = Builder.CreateCast(CI->getOpcode(), Induction,
CI->getType()); CI->getType());
Value *Broadcasted = getBroadcastInstrs(ScalarCast); Value *Broadcasted = getBroadcastInstrs(ScalarCast);
WidenMap[it] = getConsecutiveVector(Broadcasted); for (unsigned Part = 0; Part < UF; ++Part)
Entry[Part] = getConsecutiveVector(Broadcasted, VF * Part, false);
break; break;
} }
/// Vectorize casts. /// Vectorize casts.
Value *A = getVectorValue(it->getOperand(0));
Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF); Type *DestTy = VectorType::get(CI->getType()->getScalarType(), VF);
WidenMap[it] = Builder.CreateCast(CI->getOpcode(), A, DestTy);
VectorParts &A = getVectorValue(it->getOperand(0));
for (unsigned Part = 0; Part < UF; ++Part)
Entry[Part] = Builder.CreateCast(CI->getOpcode(), A[Part], DestTy);
break; break;
} }
@ -1286,12 +1394,16 @@ InnerLoopVectorizer::vectorizeBlockInLoop(LoopVectorizationLegality *Legal,
Module *M = BB->getParent()->getParent(); Module *M = BB->getParent()->getParent();
IntrinsicInst *II = cast<IntrinsicInst>(it); IntrinsicInst *II = cast<IntrinsicInst>(it);
Intrinsic::ID ID = II->getIntrinsicID(); Intrinsic::ID ID = II->getIntrinsicID();
SmallVector<Value*, 4> Args; for (unsigned Part = 0; Part < UF; ++Part) {
for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) SmallVector<Value*, 4> Args;
Args.push_back(getVectorValue(II->getArgOperand(i))); for (unsigned i = 0, ie = II->getNumArgOperands(); i != ie; ++i) {
Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) }; VectorParts &Arg = getVectorValue(II->getArgOperand(i));
Function *F = Intrinsic::getDeclaration(M, ID, Tys); Args.push_back(Arg[Part]);
WidenMap[it] = Builder.CreateCall(F, Args); }
Type *Tys[] = { VectorType::get(II->getType()->getScalarType(), VF) };
Function *F = Intrinsic::getDeclaration(M, ID, Tys);
Entry[Part] = Builder.CreateCall(F, Args);
}
break; break;
} }

76
lib/Transforms/Vectorize/LoopVectorize.h
View File

@ -54,6 +54,8 @@
#include "llvm/Analysis/ScalarEvolution.h" #include "llvm/Analysis/ScalarEvolution.h"
#include "llvm/IR/IRBuilder.h" #include "llvm/IR/IRBuilder.h"
#include <algorithm> #include <algorithm>
#include <map>
using namespace llvm; using namespace llvm;
/// We don't vectorize loops with a known constant trip count below this number. /// We don't vectorize loops with a known constant trip count below this number.
@ -91,9 +93,11 @@ class InnerLoopVectorizer {
public: public:
/// Ctor. /// Ctor.
InnerLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li, InnerLoopVectorizer(Loop *Orig, ScalarEvolution *Se, LoopInfo *Li,
DominatorTree *Dt, DataLayout *Dl, unsigned VecWidth): DominatorTree *Dt, DataLayout *Dl,
unsigned VecWidth, unsigned UnrollFactor):
OrigLoop(Orig), SE(Se), LI(Li), DT(Dt), DL(Dl), VF(VecWidth), OrigLoop(Orig), SE(Se), LI(Li), DT(Dt), DL(Dl), VF(VecWidth),
Builder(Se->getContext()), Induction(0), OldInduction(0) { } UF(UnrollFactor), Builder(Se->getContext()), Induction(0), OldInduction(0),
WidenMap(UnrollFactor) { }
// Perform the actual loop widening (vectorization). // Perform the actual loop widening (vectorization).
void vectorize(LoopVectorizationLegality *Legal) { void vectorize(LoopVectorizationLegality *Legal) {
@ -109,6 +113,10 @@ public:
private: private:
/// A small list of PHINodes. /// A small list of PHINodes.
typedef SmallVector<PHINode*, 4> PhiVector; typedef SmallVector<PHINode*, 4> PhiVector;
/// When we unroll loops we have multiple vector values for each scalar.
/// This data structure holds the unrolled and vectorized values that
/// originated from one scalar instruction.
typedef SmallVector<Value*, 2> VectorParts;
/// Add code that checks at runtime if the accessed arrays overlap. /// Add code that checks at runtime if the accessed arrays overlap.
/// Returns the comparator value or NULL if no check is needed. /// Returns the comparator value or NULL if no check is needed.
@ -122,10 +130,10 @@ private:
/// A helper function that computes the predicate of the block BB, assuming /// A helper function that computes the predicate of the block BB, assuming
/// that the header block of the loop is set to True. It returns the *entry* /// that the header block of the loop is set to True. It returns the *entry*
/// mask for the block BB. /// mask for the block BB.
Value *createBlockInMask(BasicBlock *BB); VectorParts createBlockInMask(BasicBlock *BB);
/// A helper function that computes the predicate of the edge between SRC /// A helper function that computes the predicate of the edge between SRC
/// and DST. /// and DST.
Value *createEdgeMask(BasicBlock *Src, BasicBlock *Dst); VectorParts createEdgeMask(BasicBlock *Src, BasicBlock *Dst);
/// A helper function to vectorize a single BB within the innermost loop. /// A helper function to vectorize a single BB within the innermost loop.
void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB, void vectorizeBlockInLoop(LoopVectorizationLegality *Legal, BasicBlock *BB,
@ -148,14 +156,15 @@ private:
/// This function adds 0, 1, 2 ... to each vector element, starting at zero. /// This function adds 0, 1, 2 ... to each vector element, starting at zero.
/// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...). /// If Negate is set then negative numbers are added e.g. (0, -1, -2, ...).
Value *getConsecutiveVector(Value* Val, bool Negate = false); /// The sequence starts at StartIndex.
Value *getConsecutiveVector(Value* Val, unsigned StartIdx, bool Negate);
/// When we go over instructions in the basic block we rely on previous /// When we go over instructions in the basic block we rely on previous
/// values within the current basic block or on loop invariant values. /// values within the current basic block or on loop invariant values.
/// When we widen (vectorize) values we place them in the map. If the values /// When we widen (vectorize) values we place them in the map. If the values
/// are not within the map, they have to be loop invariant, so we simply /// are not within the map, they have to be loop invariant, so we simply
/// broadcast them into a vector. /// broadcast them into a vector.
Value *getVectorValue(Value *V); VectorParts &getVectorValue(Value *V);
/// Get a uniform vector of constant integers. We use this to get /// Get a uniform vector of constant integers. We use this to get
/// vectors of ones and zeros for the reduction code. /// vectors of ones and zeros for the reduction code.
@ -164,22 +173,61 @@ private:
/// Generate a shuffle sequence that will reverse the vector Vec. /// Generate a shuffle sequence that will reverse the vector Vec.
Value *reverseVector(Value *Vec); Value *reverseVector(Value *Vec);
typedef DenseMap<Value*, Value*> ValueMap; /// This is a helper class that holds the vectorizer state. It maps scalar
/// instructions to vector instructions. When the code is 'unrolled' then
/// then a single scalar value is mapped to multiple vector parts. The parts
/// are stored in the VectorPart type.
struct ValueMap {
/// C'tor. UnrollFactor controls the number of vectors ('parts') that
/// are mapped.
ValueMap(unsigned UnrollFactor) : UF(UnrollFactor) {}
/// \return True if 'Key' is saved in the Value Map.
bool has(Value *Key) { return MapStoreage.count(Key); }
/// Initializes a new entry in the map. Sets all of the vector parts to the
/// save value in 'Val'.
/// \return A reference to a vector with splat values.
VectorParts &splat(Value *Key, Value *Val) {
MapStoreage[Key].clear();
MapStoreage[Key].append(UF, Val);
return MapStoreage[Key];
}
///\return A reference to the value that is stored at 'Key'.
VectorParts &get(Value *Key) {
if (!has(Key))
MapStoreage[Key].resize(UF);
return MapStoreage[Key];
}
/// The unroll factor. Each entry in the map stores this number of vector
/// elements.
unsigned UF;
/// Map storage. We use std::map and not DenseMap because insertions to a
/// dense map invalidates its iterators.
std::map<Value*, VectorParts> MapStoreage;
};
/// The original loop. /// The original loop.
Loop *OrigLoop; Loop *OrigLoop;
// Scev analysis to use. /// Scev analysis to use.
ScalarEvolution *SE; ScalarEvolution *SE;
// Loop Info. /// Loop Info.
LoopInfo *LI; LoopInfo *LI;
// Dominator Tree. /// Dominator Tree.
DominatorTree *DT; DominatorTree *DT;
// Data Layout. /// Data Layout.
DataLayout *DL; DataLayout *DL;
// The vectorization factor to use. /// The vectorization SIMD factor to use. Each vector will have this many
/// vector elements.
unsigned VF; unsigned VF;
/// The vectorization unroll factor to use. Each scalar is vectorized to this
/// many different vector instructions.
unsigned UF;
// The builder that we use /// The builder that we use
IRBuilder<> Builder; IRBuilder<> Builder;
// --- Vectorization state --- // --- Vectorization state ---
@ -203,7 +251,7 @@ private:
PHINode *Induction; PHINode *Induction;
/// The induction variable of the old basic block. /// The induction variable of the old basic block.
PHINode *OldInduction; PHINode *OldInduction;
// Maps scalars to widened vectors. /// Maps scalars to widened vectors.
ValueMap WidenMap; ValueMap WidenMap;
}; };