//
// TODO: Handle multiple loops at a time.
//
-// TODO: Should AddrMode::BaseGV be changed to a ConstantExpr
+// TODO: Should TargetLowering::AddrMode::BaseGV be changed to a ConstantExpr
// instead of a GlobalValue?
//
// TODO: When truncation is free, truncate ICmp users' operands to make it a
#include "llvm/Transforms/Scalar.h"
#include "llvm/Transforms/Utils/BasicBlockUtils.h"
#include "llvm/Transforms/Utils/Local.h"
-#include "llvm/TargetTransformInfo.h"
#include "llvm/ADT/SmallBitVector.h"
#include "llvm/ADT/SetVector.h"
#include "llvm/ADT/DenseSet.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ValueHandle.h"
#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetLowering.h"
#include <algorithm>
using namespace llvm;
enum KindType {
Basic, ///< A normal use, with no folding.
Special, ///< A special case of basic, allowing -1 scales.
- Address, ///< An address use; folding according to ScalarTargetTransformInfo.
+ Address, ///< An address use; folding according to TargetLowering
ICmpZero ///< An equality icmp with both operands folded into one.
// TODO: Add a generic icmp too?
};
/// address-mode folding and special icmp tricks.
static bool isLegalUse(const AddrMode &AM,
LSRUse::KindType Kind, Type *AccessTy,
- const ScalarTargetTransformInfo *STTI) {
+ const TargetLowering *TLI) {
switch (Kind) {
case LSRUse::Address:
// If we have low-level target information, ask the target if it can
// completely fold this address.
- if (STTI) return STTI->isLegalAddressingMode(AM, AccessTy);
+ if (TLI) return TLI->isLegalAddressingMode(AM, AccessTy);
// Otherwise, just guess that reg+reg addressing is legal.
return !AM.BaseGV && AM.BaseOffs == 0 && AM.Scale <= 1;
// If we have low-level target information, ask the target if it can fold an
// integer immediate on an icmp.
if (AM.BaseOffs != 0) {
- if (!STTI)
+ if (!TLI)
return false;
// We have one of:
// ICmpZero BaseReg + Offset => ICmp BaseReg, -Offset
int64_t Offs = AM.BaseOffs;
if (AM.Scale == 0)
Offs = -(uint64_t)Offs; // The cast does the right thing with INT64_MIN.
- return STTI->isLegalICmpImmediate(Offs);
+ return TLI->isLegalICmpImmediate(Offs);
}
// ICmpZero BaseReg + -1*ScaleReg => ICmp BaseReg, ScaleReg
static bool isLegalUse(AddrMode AM,
int64_t MinOffset, int64_t MaxOffset,
LSRUse::KindType Kind, Type *AccessTy,
- const ScalarTargetTransformInfo *LTTI) {
+ const TargetLowering *TLI) {
// Check for overflow.
if (((int64_t)((uint64_t)AM.BaseOffs + MinOffset) > AM.BaseOffs) !=
(MinOffset > 0))
return false;
AM.BaseOffs = (uint64_t)AM.BaseOffs + MinOffset;
- if (isLegalUse(AM, Kind, AccessTy, LTTI)) {
+ if (isLegalUse(AM, Kind, AccessTy, TLI)) {
AM.BaseOffs = (uint64_t)AM.BaseOffs - MinOffset;
// Check for overflow.
if (((int64_t)((uint64_t)AM.BaseOffs + MaxOffset) > AM.BaseOffs) !=
(MaxOffset > 0))
return false;
AM.BaseOffs = (uint64_t)AM.BaseOffs + MaxOffset;
- return isLegalUse(AM, Kind, AccessTy, LTTI);
+ return isLegalUse(AM, Kind, AccessTy, TLI);
}
return false;
}
GlobalValue *BaseGV,
bool HasBaseReg,
LSRUse::KindType Kind, Type *AccessTy,
- const ScalarTargetTransformInfo *LTTI) {
+ const TargetLowering *TLI) {
// Fast-path: zero is always foldable.
if (BaseOffs == 0 && !BaseGV) return true;
AM.HasBaseReg = true;
}
- return isLegalUse(AM, Kind, AccessTy, LTTI);
+ return isLegalUse(AM, Kind, AccessTy, TLI);
}
static bool isAlwaysFoldable(const SCEV *S,
int64_t MinOffset, int64_t MaxOffset,
bool HasBaseReg,
LSRUse::KindType Kind, Type *AccessTy,
- const ScalarTargetTransformInfo *LTTI,
+ const TargetLowering *TLI,
ScalarEvolution &SE) {
// Fast-path: zero is always foldable.
if (S->isZero()) return true;
AM.HasBaseReg = HasBaseReg;
AM.Scale = Kind == LSRUse::ICmpZero ? -1 : 1;
- return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, LTTI);
+ return isLegalUse(AM, MinOffset, MaxOffset, Kind, AccessTy, TLI);
}
namespace {
ScalarEvolution &SE;
DominatorTree &DT;
LoopInfo &LI;
- const ScalarTargetTransformInfo *const STTI;
+ const TargetLowering *const TLI;
Loop *const L;
bool Changed;
Pass *P);
public:
- LSRInstance(const ScalarTargetTransformInfo *ltti, Loop *l, Pass *P);
+ LSRInstance(const TargetLowering *tli, Loop *l, Pass *P);
bool getChanged() const { return Changed; }
}
if (!DestTy) continue;
- if (STTI) {
+ if (TLI) {
// If target does not support DestTy natively then do not apply
// this transformation.
- if (!STTI->isTypeLegal(DestTy)) continue;
+ EVT DVT = TLI->getValueType(DestTy);
+ if (!TLI->isTypeLegal(DVT)) continue;
}
PHINode *PH = dyn_cast<PHINode>(ShadowUse->getOperand(0));
if (C->getValue().getMinSignedBits() >= 64 ||
C->getValue().isMinSignedValue())
goto decline_post_inc;
- // Without STTI, assume that any stride might be valid, and so any
+ // Without TLI, assume that any stride might be valid, and so any
// use might be shared.
- if (!STTI)
+ if (!TLI)
goto decline_post_inc;
// Check for possible scaled-address reuse.
Type *AccessTy = getAccessType(UI->getUser());
AddrMode AM;
AM.Scale = C->getSExtValue();
- if (STTI->isLegalAddressingMode(AM, AccessTy))
+ if (TLI->isLegalAddressingMode(AM, AccessTy))
goto decline_post_inc;
AM.Scale = -AM.Scale;
- if (STTI->isLegalAddressingMode(AM, AccessTy))
+ if (TLI->isLegalAddressingMode(AM, AccessTy))
goto decline_post_inc;
}
}
// Conservatively assume HasBaseReg is true for now.
if (NewOffset < LU.MinOffset) {
if (!isAlwaysFoldable(LU.MaxOffset - NewOffset, 0, HasBaseReg,
- Kind, AccessTy, STTI))
+ Kind, AccessTy, TLI))
return false;
NewMinOffset = NewOffset;
} else if (NewOffset > LU.MaxOffset) {
if (!isAlwaysFoldable(NewOffset - LU.MinOffset, 0, HasBaseReg,
- Kind, AccessTy, STTI))
+ Kind, AccessTy, TLI))
return false;
NewMaxOffset = NewOffset;
}
int64_t Offset = ExtractImmediate(Expr, SE);
// Basic uses can't accept any offset, for example.
- if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, STTI)) {
+ if (!isAlwaysFoldable(Offset, 0, /*HasBaseReg=*/true, Kind, AccessTy, TLI)) {
Expr = Copy;
Offset = 0;
}
/// TODO: Consider IVInc free if it's already used in another chains.
static bool
isProfitableChain(IVChain &Chain, SmallPtrSet<Instruction*, 4> &Users,
- ScalarEvolution &SE, const ScalarTargetTransformInfo *STTI) {
+ ScalarEvolution &SE, const TargetLowering *TLI) {
if (StressIVChain)
return true;
for (unsigned UsersIdx = 0, NChains = IVChainVec.size();
UsersIdx < NChains; ++UsersIdx) {
if (!isProfitableChain(IVChainVec[UsersIdx],
- ChainUsersVec[UsersIdx].FarUsers, SE, STTI))
+ ChainUsersVec[UsersIdx].FarUsers, SE, TLI))
continue;
// Preserve the chain at UsesIdx.
if (ChainIdx != UsersIdx)
/// Return true if the IVInc can be folded into an addressing mode.
static bool canFoldIVIncExpr(const SCEV *IncExpr, Instruction *UserInst,
- Value *Operand,
- const ScalarTargetTransformInfo *STTI) {
+ Value *Operand, const TargetLowering *TLI) {
const SCEVConstant *IncConst = dyn_cast<SCEVConstant>(IncExpr);
if (!IncConst || !isAddressUse(UserInst, Operand))
return false;
int64_t IncOffset = IncConst->getValue()->getSExtValue();
if (!isAlwaysFoldable(IncOffset, /*BaseGV=*/0, /*HaseBaseReg=*/false,
- LSRUse::Address, getAccessType(UserInst), STTI))
+ LSRUse::Address, getAccessType(UserInst), TLI))
return false;
return true;
// If an IV increment can't be folded, use it as the next IV value.
if (!canFoldIVIncExpr(LeftOverExpr, IncI->UserInst, IncI->IVOperand,
- STTI)) {
+ TLI)) {
assert(IVTy == IVOper->getType() && "inconsistent IV increment type");
IVSrc = IVOper;
LeftOverExpr = 0;
// into an immediate field.
if (isAlwaysFoldable(*J, LU.MinOffset, LU.MaxOffset,
Base.getNumRegs() > 1,
- LU.Kind, LU.AccessTy, STTI, SE))
+ LU.Kind, LU.AccessTy, TLI, SE))
continue;
// Collect all operands except *J.
if (InnerAddOps.size() == 1 &&
isAlwaysFoldable(InnerAddOps[0], LU.MinOffset, LU.MaxOffset,
Base.getNumRegs() > 1,
- LU.Kind, LU.AccessTy, STTI, SE))
+ LU.Kind, LU.AccessTy, TLI, SE))
continue;
const SCEV *InnerSum = SE.getAddExpr(InnerAddOps);
// Add the remaining pieces of the add back into the new formula.
const SCEVConstant *InnerSumSC = dyn_cast<SCEVConstant>(InnerSum);
- if (STTI && InnerSumSC &&
+ if (TLI && InnerSumSC &&
SE.getTypeSizeInBits(InnerSumSC->getType()) <= 64 &&
- STTI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
+ TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
InnerSumSC->getValue()->getZExtValue())) {
F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
InnerSumSC->getValue()->getZExtValue();
// Add J as its own register, or an unfolded immediate.
const SCEVConstant *SC = dyn_cast<SCEVConstant>(*J);
- if (STTI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
- STTI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
+ if (TLI && SC && SE.getTypeSizeInBits(SC->getType()) <= 64 &&
+ TLI->isLegalAddImmediate((uint64_t)F.UnfoldedOffset +
SC->getValue()->getZExtValue()))
F.UnfoldedOffset = (uint64_t)F.UnfoldedOffset +
SC->getValue()->getZExtValue();
Formula F = Base;
F.AM.BaseGV = GV;
if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI))
+ LU.Kind, LU.AccessTy, TLI))
continue;
F.BaseRegs[i] = G;
(void)InsertFormula(LU, LUIdx, F);
Formula F = Base;
F.AM.BaseOffs = (uint64_t)Base.AM.BaseOffs - *I;
if (isLegalUse(F.AM, LU.MinOffset - *I, LU.MaxOffset - *I,
- LU.Kind, LU.AccessTy, STTI)) {
+ LU.Kind, LU.AccessTy, TLI)) {
// Add the offset to the base register.
const SCEV *NewG = SE.getAddExpr(SE.getConstant(G->getType(), *I), G);
// If it cancelled out, drop the base register, otherwise update it.
Formula F = Base;
F.AM.BaseOffs = (uint64_t)F.AM.BaseOffs + Imm;
if (!isLegalUse(F.AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI))
+ LU.Kind, LU.AccessTy, TLI))
continue;
F.BaseRegs[i] = G;
(void)InsertFormula(LU, LUIdx, F);
F.AM.BaseOffs = NewBaseOffs;
// Check that this scale is legal.
- if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, STTI))
+ if (!isLegalUse(F.AM, Offset, Offset, LU.Kind, LU.AccessTy, TLI))
continue;
// Compensate for the use having MinOffset built into it.
Base.AM.HasBaseReg = Base.BaseRegs.size() > 1;
// Check whether this scale is going to be legal.
if (!isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI)) {
+ LU.Kind, LU.AccessTy, TLI)) {
// As a special-case, handle special out-of-loop Basic users specially.
// TODO: Reconsider this special case.
if (LU.Kind == LSRUse::Basic &&
isLegalUse(Base.AM, LU.MinOffset, LU.MaxOffset,
- LSRUse::Special, LU.AccessTy, STTI) &&
+ LSRUse::Special, LU.AccessTy, TLI) &&
LU.AllFixupsOutsideLoop)
LU.Kind = LSRUse::Special;
else
/// GenerateTruncates - Generate reuse formulae from different IV types.
void LSRInstance::GenerateTruncates(LSRUse &LU, unsigned LUIdx, Formula Base) {
- // This requires ScalarTargetTransformInfo to tell us which truncates are free.
- if (!STTI) return;
+ // This requires TargetLowering to tell us which truncates are free.
+ if (!TLI) return;
// Don't bother truncating symbolic values.
if (Base.AM.BaseGV) return;
for (SmallSetVector<Type *, 4>::const_iterator
I = Types.begin(), E = Types.end(); I != E; ++I) {
Type *SrcTy = *I;
- if (SrcTy != DstTy && STTI->isTruncateFree(SrcTy, DstTy)) {
+ if (SrcTy != DstTy && TLI->isTruncateFree(SrcTy, DstTy)) {
Formula F = Base;
if (F.ScaledReg) F.ScaledReg = SE.getAnyExtendExpr(F.ScaledReg, *I);
Formula NewF = F;
NewF.AM.BaseOffs = Offs;
if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI))
+ LU.Kind, LU.AccessTy, TLI))
continue;
NewF.ScaledReg = SE.getAddExpr(NegImmS, NewF.ScaledReg);
Formula NewF = F;
NewF.AM.BaseOffs = (uint64_t)NewF.AM.BaseOffs + Imm;
if (!isLegalUse(NewF.AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI)) {
- if (!STTI ||
- !STTI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
+ LU.Kind, LU.AccessTy, TLI)) {
+ if (!TLI ||
+ !TLI->isLegalAddImmediate((uint64_t)NewF.UnfoldedOffset + Imm))
continue;
NewF = F;
NewF.UnfoldedOffset = (uint64_t)NewF.UnfoldedOffset + Imm;
Formula &F = LUThatHas->Formulae[i];
if (!isLegalUse(F.AM,
LUThatHas->MinOffset, LUThatHas->MaxOffset,
- LUThatHas->Kind, LUThatHas->AccessTy, STTI)) {
+ LUThatHas->Kind, LUThatHas->AccessTy, TLI)) {
DEBUG(dbgs() << " Deleting "; F.print(dbgs());
dbgs() << '\n');
LUThatHas->DeleteFormula(F);
Changed |= DeleteTriviallyDeadInstructions(DeadInsts);
}
-LSRInstance::LSRInstance(const ScalarTargetTransformInfo *stti, Loop *l, Pass *P)
+LSRInstance::LSRInstance(const TargetLowering *tli, Loop *l, Pass *P)
: IU(P->getAnalysis<IVUsers>()),
SE(P->getAnalysis<ScalarEvolution>()),
DT(P->getAnalysis<DominatorTree>()),
LI(P->getAnalysis<LoopInfo>()),
- STTI(stti), L(l), Changed(false), IVIncInsertPos(0) {
+ TLI(tli), L(l), Changed(false), IVIncInsertPos(0) {
// If LoopSimplify form is not available, stay out of trouble.
if (!L->isLoopSimplifyForm())
for (SmallVectorImpl<Formula>::const_iterator J = LU.Formulae.begin(),
JE = LU.Formulae.end(); J != JE; ++J)
assert(isLegalUse(J->AM, LU.MinOffset, LU.MaxOffset,
- LU.Kind, LU.AccessTy, STTI) &&
+ LU.Kind, LU.AccessTy, TLI) &&
"Illegal formula generated!");
};
#endif
namespace {
class LoopStrengthReduce : public LoopPass {
- /// ScalarTargetTransformInfo provides target information that is needed
- /// for strength reducing loops.
- const ScalarTargetTransformInfo *STTI;
+ /// TLI - Keep a pointer of a TargetLowering to consult for determining
+ /// transformation profitability.
+ const TargetLowering *const TLI;
public:
static char ID; // Pass ID, replacement for typeid
- LoopStrengthReduce();
+ explicit LoopStrengthReduce(const TargetLowering *tli = 0);
private:
bool runOnLoop(Loop *L, LPPassManager &LPM);
INITIALIZE_PASS_END(LoopStrengthReduce, "loop-reduce",
"Loop Strength Reduction", false, false)
-Pass *llvm::createLoopStrengthReducePass() {
- return new LoopStrengthReduce();
+
+Pass *llvm::createLoopStrengthReducePass(const TargetLowering *TLI) {
+ return new LoopStrengthReduce(TLI);
}
-LoopStrengthReduce::LoopStrengthReduce()
- : LoopPass(ID), STTI(0) {
+LoopStrengthReduce::LoopStrengthReduce(const TargetLowering *tli)
+ : LoopPass(ID), TLI(tli) {
initializeLoopStrengthReducePass(*PassRegistry::getPassRegistry());
}
bool LoopStrengthReduce::runOnLoop(Loop *L, LPPassManager & /*LPM*/) {
bool Changed = false;
- TargetTransformInfo *TTI = getAnalysisIfAvailable<TargetTransformInfo>();
-
- if (TTI)
- STTI = TTI->getScalarTargetTransformInfo();
-
// Run the main LSR transformation.
- Changed |= LSRInstance(STTI, L, this).getChanged();
+ Changed |= LSRInstance(TLI, L, this).getChanged();
// Remove any extra phis created by processing inner loops.
Changed |= DeleteDeadPHIs(L->getHeader());
Rewriter.setDebugType(DEBUG_TYPE);
#endif
unsigned numFolded = Rewriter.
- replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, STTI);
+ replaceCongruentIVs(L, &getAnalysis<DominatorTree>(), DeadInsts, TLI);
if (numFolded) {
Changed = true;
DeleteTriviallyDeadInstructions(DeadInsts);
projects / oota-llvm.git / blobdiff