1 //===-- X86SelectionDAGInfo.cpp - X86 SelectionDAG Info -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file implements the X86SelectionDAGInfo class.
12 //===----------------------------------------------------------------------===//
14 #define DEBUG_TYPE "x86-selectiondag-info"
15 #include "X86TargetMachine.h"
16 #include "llvm/DerivedTypes.h"
17 #include "llvm/CodeGen/SelectionDAG.h"
20 X86SelectionDAGInfo::X86SelectionDAGInfo(const X86TargetMachine &TM) :
21 TargetSelectionDAGInfo(TM),
22 Subtarget(&TM.getSubtarget<X86Subtarget>()),
23 TLI(*TM.getTargetLowering()) {
26 X86SelectionDAGInfo::~X86SelectionDAGInfo() {
30 X86SelectionDAGInfo::EmitTargetCodeForMemset(SelectionDAG &DAG, DebugLoc dl,
32 SDValue Dst, SDValue Src,
33 SDValue Size, unsigned Align,
35 MachinePointerInfo DstPtrInfo) const {
36 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
38 // If to a segment-relative address space, use the default lowering.
39 if (DstPtrInfo.getAddrSpace() >= 256)
42 // If not DWORD aligned or size is more than the threshold, call the library.
43 // The libc version is likely to be faster for these cases. It can use the
44 // address value and run time information about the CPU.
45 if ((Align & 3) != 0 ||
47 ConstantSize->getZExtValue() >
48 Subtarget->getMaxInlineSizeThreshold()) {
51 // Check to see if there is a specialized entry-point for memory zeroing.
52 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
54 if (const char *bzeroEntry = V &&
55 V->isNullValue() ? Subtarget->getBZeroEntry() : 0) {
56 EVT IntPtr = TLI.getPointerTy();
57 unsigned AS = DstPtrInfo.getAddrSpace();
58 Type *IntPtrTy = getDataLayout()->getIntPtrType(*DAG.getContext(), AS);
59 TargetLowering::ArgListTy Args;
60 TargetLowering::ArgListEntry Entry;
63 Args.push_back(Entry);
65 Args.push_back(Entry);
67 CallLoweringInfo CLI(Chain, Type::getVoidTy(*DAG.getContext()),
68 false, false, false, false,
69 0, CallingConv::C, /*isTailCall=*/false,
70 /*doesNotRet=*/false, /*isReturnValueUsed=*/false,
71 DAG.getExternalSymbol(bzeroEntry, IntPtr), Args,
73 std::pair<SDValue,SDValue> CallResult =
75 return CallResult.second;
78 // Otherwise have the target-independent code call memset.
82 uint64_t SizeVal = ConstantSize->getZExtValue();
86 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
87 unsigned BytesLeft = 0;
88 bool TwoRepStos = false;
91 uint64_t Val = ValC->getZExtValue() & 255;
93 // If the value is a constant, then we can potentially use larger sets.
95 case 2: // WORD aligned
98 Val = (Val << 8) | Val;
100 case 0: // DWORD aligned
103 Val = (Val << 8) | Val;
104 Val = (Val << 16) | Val;
105 if (Subtarget->is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
108 Val = (Val << 32) | Val;
111 default: // Byte aligned
114 Count = DAG.getIntPtrConstant(SizeVal);
118 if (AVT.bitsGT(MVT::i8)) {
119 unsigned UBytes = AVT.getSizeInBits() / 8;
120 Count = DAG.getIntPtrConstant(SizeVal / UBytes);
121 BytesLeft = SizeVal % UBytes;
124 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, AVT),
126 InFlag = Chain.getValue(1);
129 Count = DAG.getIntPtrConstant(SizeVal);
130 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
131 InFlag = Chain.getValue(1);
134 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
137 InFlag = Chain.getValue(1);
138 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
141 InFlag = Chain.getValue(1);
143 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
144 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
145 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
148 InFlag = Chain.getValue(1);
150 EVT CVT = Count.getValueType();
151 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
152 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, CVT));
153 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX :
156 InFlag = Chain.getValue(1);
157 Tys = DAG.getVTList(MVT::Other, MVT::Glue);
158 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
159 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops, array_lengthof(Ops));
160 } else if (BytesLeft) {
161 // Handle the last 1 - 7 bytes.
162 unsigned Offset = SizeVal - BytesLeft;
163 EVT AddrVT = Dst.getValueType();
164 EVT SizeVT = Size.getValueType();
166 Chain = DAG.getMemset(Chain, dl,
167 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
168 DAG.getConstant(Offset, AddrVT)),
170 DAG.getConstant(BytesLeft, SizeVT),
171 Align, isVolatile, DstPtrInfo.getWithOffset(Offset));
174 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
179 X86SelectionDAGInfo::EmitTargetCodeForMemcpy(SelectionDAG &DAG, DebugLoc dl,
180 SDValue Chain, SDValue Dst, SDValue Src,
181 SDValue Size, unsigned Align,
182 bool isVolatile, bool AlwaysInline,
183 MachinePointerInfo DstPtrInfo,
184 MachinePointerInfo SrcPtrInfo) const {
185 // This requires the copy size to be a constant, preferably
186 // within a subtarget-specific limit.
187 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
190 uint64_t SizeVal = ConstantSize->getZExtValue();
191 if (!AlwaysInline && SizeVal > Subtarget->getMaxInlineSizeThreshold())
194 /// If not DWORD aligned, it is more efficient to call the library. However
195 /// if calling the library is not allowed (AlwaysInline), then soldier on as
196 /// the code generated here is better than the long load-store sequence we
197 /// would otherwise get.
198 if (!AlwaysInline && (Align & 3) != 0)
201 // If to a segment-relative address space, use the default lowering.
202 if (DstPtrInfo.getAddrSpace() >= 256 ||
203 SrcPtrInfo.getAddrSpace() >= 256)
216 AVT = Subtarget->is64Bit() ? MVT::i64 : MVT::i32;
218 unsigned UBytes = AVT.getSizeInBits() / 8;
219 unsigned CountVal = SizeVal / UBytes;
220 SDValue Count = DAG.getIntPtrConstant(CountVal);
221 unsigned BytesLeft = SizeVal % UBytes;
223 SDValue InFlag(0, 0);
224 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RCX :
227 InFlag = Chain.getValue(1);
228 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RDI :
231 InFlag = Chain.getValue(1);
232 Chain = DAG.getCopyToReg(Chain, dl, Subtarget->is64Bit() ? X86::RSI :
235 InFlag = Chain.getValue(1);
237 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
238 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
239 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops,
240 array_lengthof(Ops));
242 SmallVector<SDValue, 4> Results;
243 Results.push_back(RepMovs);
245 // Handle the last 1 - 7 bytes.
246 unsigned Offset = SizeVal - BytesLeft;
247 EVT DstVT = Dst.getValueType();
248 EVT SrcVT = Src.getValueType();
249 EVT SizeVT = Size.getValueType();
250 Results.push_back(DAG.getMemcpy(Chain, dl,
251 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
252 DAG.getConstant(Offset, DstVT)),
253 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
254 DAG.getConstant(Offset, SrcVT)),
255 DAG.getConstant(BytesLeft, SizeVT),
256 Align, isVolatile, AlwaysInline,
257 DstPtrInfo.getWithOffset(Offset),
258 SrcPtrInfo.getWithOffset(Offset)));
261 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
262 &Results[0], Results.size());