//===----------------------------------------------------------------------===//
#include "llvm/CodeGen/SelectionDAG.h"
-#include "SDNodeOrdering.h"
#include "SDNodeDbgValue.h"
+#include "SDNodeOrdering.h"
+#include "llvm/ADT/SetVector.h"
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/Assembly/Writer.h"
#include "llvm/CallingConv.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineConstantPool.h"
+#include "llvm/CodeGen/MachineFrameInfo.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/Constants.h"
+#include "llvm/DataLayout.h"
#include "llvm/DebugInfo.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/GlobalAlias.h"
#include "llvm/GlobalVariable.h"
#include "llvm/Intrinsics.h"
-#include "llvm/Analysis/ValueTracking.h"
-#include "llvm/Assembly/Writer.h"
-#include "llvm/CodeGen/MachineBasicBlock.h"
-#include "llvm/CodeGen/MachineConstantPool.h"
-#include "llvm/CodeGen/MachineFrameInfo.h"
-#include "llvm/CodeGen/MachineModuleInfo.h"
-#include "llvm/Target/TargetRegisterInfo.h"
-#include "llvm/DataLayout.h"
-#include "llvm/Target/TargetLowering.h"
-#include "llvm/Target/TargetSelectionDAGInfo.h"
-#include "llvm/Target/TargetOptions.h"
-#include "llvm/Target/TargetInstrInfo.h"
-#include "llvm/Target/TargetIntrinsicInfo.h"
-#include "llvm/Target/TargetMachine.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/ManagedStatic.h"
#include "llvm/Support/MathExtras.h"
-#include "llvm/Support/raw_ostream.h"
#include "llvm/Support/Mutex.h"
-#include "llvm/ADT/SetVector.h"
-#include "llvm/ADT/SmallPtrSet.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/SmallVector.h"
-#include "llvm/ADT/StringExtras.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/TargetTransformInfo.h"
+#include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetIntrinsicInfo.h"
+#include "llvm/Target/TargetLowering.h"
+#include "llvm/Target/TargetMachine.h"
+#include "llvm/Target/TargetOptions.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+#include "llvm/Target/TargetSelectionDAGInfo.h"
#include <algorithm>
#include <cmath>
using namespace llvm;
const APFloat& Val) {
assert(VT.isFloatingPoint() && "Can only convert between FP types");
- // PPC long double cannot be converted to any other type.
- if (VT == MVT::ppcf128 ||
- &Val.getSemantics() == &APFloat::PPCDoubleDouble)
- return false;
-
// convert modifies in place, so make a copy.
APFloat Val2 = APFloat(Val);
bool losesInfo;
}
if (ConstantFPSDNode *N1C = dyn_cast<ConstantFPSDNode>(N1.getNode())) {
if (ConstantFPSDNode *N2C = dyn_cast<ConstantFPSDNode>(N2.getNode())) {
- // No compile time operations on this type yet.
- if (N1C->getValueType(0) == MVT::ppcf128)
- return SDValue();
-
APFloat::cmpResult R = N1C->getValueAPF().compare(N2C->getValueAPF());
switch (Cond) {
default: break;
return getConstant(Val.zextOrTrunc(VT.getSizeInBits()), VT);
case ISD::UINT_TO_FP:
case ISD::SINT_TO_FP: {
- // No compile time operations on ppcf128.
- if (VT == MVT::ppcf128) break;
APFloat apf(APInt::getNullValue(VT.getSizeInBits()));
(void)apf.convertFromAPInt(Val,
Opcode==ISD::SINT_TO_FP,
// Constant fold unary operations with a floating point constant operand.
if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(Operand.getNode())) {
APFloat V = C->getValueAPF(); // make copy
- if (VT != MVT::ppcf128 && Operand.getValueType() != MVT::ppcf128) {
- switch (Opcode) {
- case ISD::FNEG:
- V.changeSign();
+ switch (Opcode) {
+ case ISD::FNEG:
+ V.changeSign();
+ return getConstantFP(V, VT);
+ case ISD::FABS:
+ V.clearSign();
+ return getConstantFP(V, VT);
+ case ISD::FCEIL: {
+ APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardPositive);
+ if (fs == APFloat::opOK || fs == APFloat::opInexact)
return getConstantFP(V, VT);
- case ISD::FABS:
- V.clearSign();
+ break;
+ }
+ case ISD::FTRUNC: {
+ APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardZero);
+ if (fs == APFloat::opOK || fs == APFloat::opInexact)
return getConstantFP(V, VT);
- case ISD::FCEIL: {
- APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardPositive);
- if (fs == APFloat::opOK || fs == APFloat::opInexact)
- return getConstantFP(V, VT);
- break;
- }
- case ISD::FTRUNC: {
- APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardZero);
- if (fs == APFloat::opOK || fs == APFloat::opInexact)
- return getConstantFP(V, VT);
- break;
- }
- case ISD::FFLOOR: {
- APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardNegative);
- if (fs == APFloat::opOK || fs == APFloat::opInexact)
- return getConstantFP(V, VT);
- break;
- }
- case ISD::FP_EXTEND: {
- bool ignored;
- // This can return overflow, underflow, or inexact; we don't care.
- // FIXME need to be more flexible about rounding mode.
- (void)V.convert(*EVTToAPFloatSemantics(VT),
- APFloat::rmNearestTiesToEven, &ignored);
+ break;
+ }
+ case ISD::FFLOOR: {
+ APFloat::opStatus fs = V.roundToIntegral(APFloat::rmTowardNegative);
+ if (fs == APFloat::opOK || fs == APFloat::opInexact)
return getConstantFP(V, VT);
- }
- case ISD::FP_TO_SINT:
- case ISD::FP_TO_UINT: {
- integerPart x[2];
- bool ignored;
- assert(integerPartWidth >= 64);
- // FIXME need to be more flexible about rounding mode.
- APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
- Opcode==ISD::FP_TO_SINT,
- APFloat::rmTowardZero, &ignored);
- if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
- break;
- APInt api(VT.getSizeInBits(), x);
- return getConstant(api, VT);
- }
- case ISD::BITCAST:
- if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
- return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
- else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
- return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
+ break;
+ }
+ case ISD::FP_EXTEND: {
+ bool ignored;
+ // This can return overflow, underflow, or inexact; we don't care.
+ // FIXME need to be more flexible about rounding mode.
+ (void)V.convert(*EVTToAPFloatSemantics(VT),
+ APFloat::rmNearestTiesToEven, &ignored);
+ return getConstantFP(V, VT);
+ }
+ case ISD::FP_TO_SINT:
+ case ISD::FP_TO_UINT: {
+ integerPart x[2];
+ bool ignored;
+ assert(integerPartWidth >= 64);
+ // FIXME need to be more flexible about rounding mode.
+ APFloat::opStatus s = V.convertToInteger(x, VT.getSizeInBits(),
+ Opcode==ISD::FP_TO_SINT,
+ APFloat::rmTowardZero, &ignored);
+ if (s==APFloat::opInvalidOp) // inexact is OK, in fact usual
break;
- }
+ APInt api(VT.getSizeInBits(), x);
+ return getConstant(api, VT);
+ }
+ case ISD::BITCAST:
+ if (VT == MVT::i32 && C->getValueType(0) == MVT::f32)
+ return getConstant((uint32_t)V.bitcastToAPInt().getZExtValue(), VT);
+ else if (VT == MVT::i64 && C->getValueType(0) == MVT::f64)
+ return getConstant(V.bitcastToAPInt().getZExtValue(), VT);
+ break;
}
}
// Cannonicalize constant to RHS if commutative
std::swap(N1CFP, N2CFP);
std::swap(N1, N2);
- } else if (N2CFP && VT != MVT::ppcf128) {
+ } else if (N2CFP) {
APFloat V1 = N1CFP->getValueAPF(), V2 = N2CFP->getValueAPF();
APFloat::opStatus s;
switch (Opcode) {
unsigned NumVTBytes = VT.getSizeInBits() / 8;
unsigned NumBytes = std::min(NumVTBytes, unsigned(Str.size()));
- uint64_t Val = 0;
+ APInt Val(NumBytes*8, 0);
if (TLI.isLittleEndian()) {
for (unsigned i = 0; i != NumBytes; ++i)
Val |= (uint64_t)(unsigned char)Str[i] << i*8;
Val |= (uint64_t)(unsigned char)Str[i] << (NumVTBytes-i-1)*8;
}
- return DAG.getConstant(Val, VT);
+ // If the "cost" of materializing the integer immediate is 1 or free, then
+ // it is cost effective to turn the load into the immediate.
+ if (DAG.getTarget().getScalarTargetTransformInfo()->
+ getIntImmCost(Val, VT.getTypeForEVT(*DAG.getContext())) < 2)
+ return DAG.getConstant(Val, VT);
+ return SDValue(0, 0);
}
/// getMemBasePlusOffset - Returns base and offset node for the
static bool FindOptimalMemOpLowering(std::vector<EVT> &MemOps,
unsigned Limit, uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
- bool IsZeroVal,
+ bool IsMemset,
+ bool ZeroMemset,
bool MemcpyStrSrc,
+ bool AllowOverlap,
SelectionDAG &DAG,
const TargetLowering &TLI) {
assert((SrcAlign == 0 || SrcAlign >= DstAlign) &&
// 'MemcpyStrSrc' indicates whether the memcpy source is constant so it does
// not need to be loaded.
EVT VT = TLI.getOptimalMemOpType(Size, DstAlign, SrcAlign,
- IsZeroVal, MemcpyStrSrc,
+ IsMemset, ZeroMemset, MemcpyStrSrc,
DAG.getMachineFunction());
- Type *vtType = VT.isExtended() ? VT.getTypeForEVT(*DAG.getContext()) : NULL;
- unsigned AS = (vtType && vtType->isPointerTy()) ?
- cast<PointerType>(vtType)->getAddressSpace() : 0;
if (VT == MVT::Other) {
- if (DstAlign >= TLI.getDataLayout()->getPointerPrefAlignment(AS) ||
+ if (DstAlign >= TLI.getDataLayout()->getPointerPrefAlignment() ||
TLI.allowsUnalignedMemoryAccesses(VT)) {
VT = TLI.getPointerTy();
} else {
unsigned VTSize = VT.getSizeInBits() / 8;
while (VTSize > Size) {
// For now, only use non-vector load / store's for the left-over pieces.
+ EVT NewVT = VT;
+ unsigned NewVTSize;
+
+ bool Found = false;
if (VT.isVector() || VT.isFloatingPoint()) {
- VT = MVT::i64;
- while (!TLI.isTypeLegal(VT))
- VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
- VTSize = VT.getSizeInBits() / 8;
- } else {
- // This can result in a type that is not legal on the target, e.g.
- // 1 or 2 bytes on PPC.
- VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1);
- VTSize >>= 1;
+ NewVT = (VT.getSizeInBits() > 64) ? MVT::i64 : MVT::i32;
+ if (TLI.isOperationLegalOrCustom(ISD::STORE, NewVT) &&
+ TLI.isSafeMemOpType(NewVT.getSimpleVT()))
+ Found = true;
+ else if (NewVT == MVT::i64 &&
+ TLI.isOperationLegalOrCustom(ISD::STORE, MVT::f64) &&
+ TLI.isSafeMemOpType(MVT::f64)) {
+ // i64 is usually not legal on 32-bit targets, but f64 may be.
+ NewVT = MVT::f64;
+ Found = true;
+ }
+ }
+
+ if (!Found) {
+ do {
+ NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1);
+ if (NewVT == MVT::i8)
+ break;
+ } while (!TLI.isSafeMemOpType(NewVT.getSimpleVT()));
+ }
+ NewVTSize = NewVT.getSizeInBits() / 8;
+
+ // If the new VT cannot cover all of the remaining bits, then consider
+ // issuing a (or a pair of) unaligned and overlapping load / store.
+ // FIXME: Only does this for 64-bit or more since we don't have proper
+ // cost model for unaligned load / store.
+ bool Fast;
+ if (NumMemOps && AllowOverlap &&
+ VTSize >= 8 && NewVTSize < Size &&
+ TLI.allowsUnalignedMemoryAccesses(VT, &Fast) && Fast)
+ VTSize = Size;
+ else {
+ VT = NewVT;
+ VTSize = NewVTSize;
}
}
if (++NumMemOps > Limit)
return false;
+
MemOps.push_back(VT);
Size -= VTSize;
}
if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
(DstAlignCanChange ? 0 : Align),
(isZeroStr ? 0 : SrcAlign),
- true, CopyFromStr, DAG, TLI))
+ false, false, CopyFromStr, true, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
unsigned VTSize = VT.getSizeInBits() / 8;
SDValue Value, Store;
+ if (VTSize > Size) {
+ // Issuing an unaligned load / store pair that overlaps with the previous
+ // pair. Adjust the offset accordingly.
+ assert(i == NumMemOps-1 && i != 0);
+ SrcOff -= VTSize - Size;
+ DstOff -= VTSize - Size;
+ }
+
if (CopyFromStr &&
(isZeroStr || (VT.isInteger() && !VT.isVector()))) {
// It's unlikely a store of a vector immediate can be done in a single
// FIXME: Handle other cases where store of vector immediate is done in
// a single instruction.
Value = getMemsetStringVal(VT, dl, DAG, TLI, Str.substr(SrcOff));
- Store = DAG.getStore(Chain, dl, Value,
- getMemBasePlusOffset(Dst, DstOff, DAG),
- DstPtrInfo.getWithOffset(DstOff), isVol,
- false, Align);
- } else {
+ if (Value.getNode())
+ Store = DAG.getStore(Chain, dl, Value,
+ getMemBasePlusOffset(Dst, DstOff, DAG),
+ DstPtrInfo.getWithOffset(DstOff), isVol,
+ false, Align);
+ }
+
+ if (!Store.getNode()) {
// The type might not be legal for the target. This should only happen
// if the type is smaller than a legal type, as on PPC, so the right
// thing to do is generate a LoadExt/StoreTrunc pair. These simplify
OutChains.push_back(Store);
SrcOff += VTSize;
DstOff += VTSize;
+ Size -= VTSize;
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
unsigned Limit = AlwaysInline ? ~0U : TLI.getMaxStoresPerMemmove(OptSize);
if (!FindOptimalMemOpLowering(MemOps, Limit, Size,
- (DstAlignCanChange ? 0 : Align),
- SrcAlign, true, false, DAG, TLI))
+ (DstAlignCanChange ? 0 : Align), SrcAlign,
+ false, false, false, false, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
isa<ConstantSDNode>(Src) && cast<ConstantSDNode>(Src)->isNullValue();
if (!FindOptimalMemOpLowering(MemOps, TLI.getMaxStoresPerMemset(OptSize),
Size, (DstAlignCanChange ? 0 : Align), 0,
- IsZeroVal, false, DAG, TLI))
+ true, IsZeroVal, false, true, DAG, TLI))
return SDValue();
if (DstAlignCanChange) {
for (unsigned i = 0; i < NumMemOps; i++) {
EVT VT = MemOps[i];
+ unsigned VTSize = VT.getSizeInBits() / 8;
+ if (VTSize > Size) {
+ // Issuing an unaligned load / store pair that overlaps with the previous
+ // pair. Adjust the offset accordingly.
+ assert(i == NumMemOps-1 && i != 0);
+ DstOff -= VTSize - Size;
+ }
// If this store is smaller than the largest store see whether we can get
// the smaller value for free with a truncate.
isVol, false, Align);
OutChains.push_back(Store);
DstOff += VT.getSizeInBits() / 8;
+ Size -= VTSize;
}
return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,