1 //===-- X86SelectionDAGInfo.cpp - X86 SelectionDAG Info -------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 // This file implements the X86SelectionDAGInfo class.
11 //
12 //===----------------------------------------------------------------------===//
13
14 #include "X86InstrInfo.h"
15 #include "X86ISelLowering.h"
16 #include "X86RegisterInfo.h"
17 #include "X86Subtarget.h"
18 #include "X86SelectionDAGInfo.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/IR/DerivedTypes.h"
21 #include "llvm/Target/TargetLowering.h"
22
23 using namespace llvm;
24
25 #define DEBUG_TYPE "x86-selectiondag-info"
26
isBaseRegConflictPossible(SelectionDAG & DAG,ArrayRef<unsigned> ClobberSet) const27 bool X86SelectionDAGInfo::isBaseRegConflictPossible(
28 SelectionDAG &DAG, ArrayRef<unsigned> ClobberSet) const {
29 // We cannot use TRI->hasBasePointer() until *after* we select all basic
30 // blocks. Legalization may introduce new stack temporaries with large
31 // alignment requirements. Fall back to generic code if there are any
32 // dynamic stack adjustments (hopefully rare) and the base pointer would
33 // conflict if we had to use it.
34 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
35 if (!MFI->hasVarSizedObjects() && !MFI->hasOpaqueSPAdjustment())
36 return false;
37
38 const X86RegisterInfo *TRI = static_cast<const X86RegisterInfo *>(
39 DAG.getSubtarget().getRegisterInfo());
40 unsigned BaseReg = TRI->getBaseRegister();
41 for (unsigned R : ClobberSet)
42 if (BaseReg == R)
43 return true;
44 return false;
45 }
46
EmitTargetCodeForMemset(SelectionDAG & DAG,SDLoc dl,SDValue Chain,SDValue Dst,SDValue Src,SDValue Size,unsigned Align,bool isVolatile,MachinePointerInfo DstPtrInfo) const47 SDValue X86SelectionDAGInfo::EmitTargetCodeForMemset(
48 SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src,
49 SDValue Size, unsigned Align, bool isVolatile,
50 MachinePointerInfo DstPtrInfo) const {
51 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
52 const X86Subtarget &Subtarget =
53 DAG.getMachineFunction().getSubtarget<X86Subtarget>();
54
55 #ifndef NDEBUG
56 // If the base register might conflict with our physical registers, bail out.
57 const unsigned ClobberSet[] = {X86::RCX, X86::RAX, X86::RDI,
58 X86::ECX, X86::EAX, X86::EDI};
59 assert(!isBaseRegConflictPossible(DAG, ClobberSet));
60 #endif
61
62 // If to a segment-relative address space, use the default lowering.
63 if (DstPtrInfo.getAddrSpace() >= 256)
64 return SDValue();
65
66 // If not DWORD aligned or size is more than the threshold, call the library.
67 // The libc version is likely to be faster for these cases. It can use the
68 // address value and run time information about the CPU.
69 if ((Align & 3) != 0 || !ConstantSize ||
70 ConstantSize->getZExtValue() > Subtarget.getMaxInlineSizeThreshold()) {
71 // Check to see if there is a specialized entry-point for memory zeroing.
72 ConstantSDNode *V = dyn_cast<ConstantSDNode>(Src);
73
74 if (const char *bzeroEntry = V &&
75 V->isNullValue() ? Subtarget.getBZeroEntry() : nullptr) {
76 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
77 EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
78 Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
79 TargetLowering::ArgListTy Args;
80 TargetLowering::ArgListEntry Entry;
81 Entry.Node = Dst;
82 Entry.Ty = IntPtrTy;
83 Args.push_back(Entry);
84 Entry.Node = Size;
85 Args.push_back(Entry);
86
87 TargetLowering::CallLoweringInfo CLI(DAG);
88 CLI.setDebugLoc(dl).setChain(Chain)
89 .setCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
90 DAG.getExternalSymbol(bzeroEntry, IntPtr), std::move(Args),
91 0)
92 .setDiscardResult();
93
94 std::pair<SDValue,SDValue> CallResult = TLI.LowerCallTo(CLI);
95 return CallResult.second;
96 }
97
98 // Otherwise have the target-independent code call memset.
99 return SDValue();
100 }
101
102 uint64_t SizeVal = ConstantSize->getZExtValue();
103 SDValue InFlag;
104 EVT AVT;
105 SDValue Count;
106 ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Src);
107 unsigned BytesLeft = 0;
108 bool TwoRepStos = false;
109 if (ValC) {
110 unsigned ValReg;
111 uint64_t Val = ValC->getZExtValue() & 255;
112
113 // If the value is a constant, then we can potentially use larger sets.
114 switch (Align & 3) {
115 case 2: // WORD aligned
116 AVT = MVT::i16;
117 ValReg = X86::AX;
118 Val = (Val << 8) | Val;
119 break;
120 case 0: // DWORD aligned
121 AVT = MVT::i32;
122 ValReg = X86::EAX;
123 Val = (Val << 8) | Val;
124 Val = (Val << 16) | Val;
125 if (Subtarget.is64Bit() && ((Align & 0x7) == 0)) { // QWORD aligned
126 AVT = MVT::i64;
127 ValReg = X86::RAX;
128 Val = (Val << 32) | Val;
129 }
130 break;
131 default: // Byte aligned
132 AVT = MVT::i8;
133 ValReg = X86::AL;
134 Count = DAG.getIntPtrConstant(SizeVal, dl);
135 break;
136 }
137
138 if (AVT.bitsGT(MVT::i8)) {
139 unsigned UBytes = AVT.getSizeInBits() / 8;
140 Count = DAG.getIntPtrConstant(SizeVal / UBytes, dl);
141 BytesLeft = SizeVal % UBytes;
142 }
143
144 Chain = DAG.getCopyToReg(Chain, dl, ValReg, DAG.getConstant(Val, dl, AVT),
145 InFlag);
146 InFlag = Chain.getValue(1);
147 } else {
148 AVT = MVT::i8;
149 Count = DAG.getIntPtrConstant(SizeVal, dl);
150 Chain = DAG.getCopyToReg(Chain, dl, X86::AL, Src, InFlag);
151 InFlag = Chain.getValue(1);
152 }
153
154 Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RCX : X86::ECX,
155 Count, InFlag);
156 InFlag = Chain.getValue(1);
157 Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RDI : X86::EDI,
158 Dst, InFlag);
159 InFlag = Chain.getValue(1);
160
161 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
162 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
163 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
164
165 if (TwoRepStos) {
166 InFlag = Chain.getValue(1);
167 Count = Size;
168 EVT CVT = Count.getValueType();
169 SDValue Left = DAG.getNode(ISD::AND, dl, CVT, Count,
170 DAG.getConstant((AVT == MVT::i64) ? 7 : 3, dl,
171 CVT));
172 Chain = DAG.getCopyToReg(Chain, dl, (CVT == MVT::i64) ? X86::RCX : X86::ECX,
173 Left, InFlag);
174 InFlag = Chain.getValue(1);
175 Tys = DAG.getVTList(MVT::Other, MVT::Glue);
176 SDValue Ops[] = { Chain, DAG.getValueType(MVT::i8), InFlag };
177 Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
178 } else if (BytesLeft) {
179 // Handle the last 1 - 7 bytes.
180 unsigned Offset = SizeVal - BytesLeft;
181 EVT AddrVT = Dst.getValueType();
182 EVT SizeVT = Size.getValueType();
183
184 Chain = DAG.getMemset(Chain, dl,
185 DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
186 DAG.getConstant(Offset, dl, AddrVT)),
187 Src,
188 DAG.getConstant(BytesLeft, dl, SizeVT),
189 Align, isVolatile, false,
190 DstPtrInfo.getWithOffset(Offset));
191 }
192
193 // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
194 return Chain;
195 }
196
EmitTargetCodeForMemcpy(SelectionDAG & DAG,SDLoc dl,SDValue Chain,SDValue Dst,SDValue Src,SDValue Size,unsigned Align,bool isVolatile,bool AlwaysInline,MachinePointerInfo DstPtrInfo,MachinePointerInfo SrcPtrInfo) const197 SDValue X86SelectionDAGInfo::EmitTargetCodeForMemcpy(
198 SelectionDAG &DAG, SDLoc dl, SDValue Chain, SDValue Dst, SDValue Src,
199 SDValue Size, unsigned Align, bool isVolatile, bool AlwaysInline,
200 MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) const {
201 // This requires the copy size to be a constant, preferably
202 // within a subtarget-specific limit.
203 ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size);
204 const X86Subtarget &Subtarget =
205 DAG.getMachineFunction().getSubtarget<X86Subtarget>();
206 if (!ConstantSize)
207 return SDValue();
208 uint64_t SizeVal = ConstantSize->getZExtValue();
209 if (!AlwaysInline && SizeVal > Subtarget.getMaxInlineSizeThreshold())
210 return SDValue();
211
212 /// If not DWORD aligned, it is more efficient to call the library. However
213 /// if calling the library is not allowed (AlwaysInline), then soldier on as
214 /// the code generated here is better than the long load-store sequence we
215 /// would otherwise get.
216 if (!AlwaysInline && (Align & 3) != 0)
217 return SDValue();
218
219 // If to a segment-relative address space, use the default lowering.
220 if (DstPtrInfo.getAddrSpace() >= 256 ||
221 SrcPtrInfo.getAddrSpace() >= 256)
222 return SDValue();
223
224 // If the base register might conflict with our physical registers, bail out.
225 const unsigned ClobberSet[] = {X86::RCX, X86::RSI, X86::RDI,
226 X86::ECX, X86::ESI, X86::EDI};
227 if (isBaseRegConflictPossible(DAG, ClobberSet))
228 return SDValue();
229
230 MVT AVT;
231 if (Align & 1)
232 AVT = MVT::i8;
233 else if (Align & 2)
234 AVT = MVT::i16;
235 else if (Align & 4)
236 // DWORD aligned
237 AVT = MVT::i32;
238 else
239 // QWORD aligned
240 AVT = Subtarget.is64Bit() ? MVT::i64 : MVT::i32;
241
242 unsigned UBytes = AVT.getSizeInBits() / 8;
243 unsigned CountVal = SizeVal / UBytes;
244 SDValue Count = DAG.getIntPtrConstant(CountVal, dl);
245 unsigned BytesLeft = SizeVal % UBytes;
246
247 SDValue InFlag;
248 Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RCX : X86::ECX,
249 Count, InFlag);
250 InFlag = Chain.getValue(1);
251 Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RDI : X86::EDI,
252 Dst, InFlag);
253 InFlag = Chain.getValue(1);
254 Chain = DAG.getCopyToReg(Chain, dl, Subtarget.is64Bit() ? X86::RSI : X86::ESI,
255 Src, InFlag);
256 InFlag = Chain.getValue(1);
257
258 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
259 SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
260 SDValue RepMovs = DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops);
261
262 SmallVector<SDValue, 4> Results;
263 Results.push_back(RepMovs);
264 if (BytesLeft) {
265 // Handle the last 1 - 7 bytes.
266 unsigned Offset = SizeVal - BytesLeft;
267 EVT DstVT = Dst.getValueType();
268 EVT SrcVT = Src.getValueType();
269 EVT SizeVT = Size.getValueType();
270 Results.push_back(DAG.getMemcpy(Chain, dl,
271 DAG.getNode(ISD::ADD, dl, DstVT, Dst,
272 DAG.getConstant(Offset, dl,
273 DstVT)),
274 DAG.getNode(ISD::ADD, dl, SrcVT, Src,
275 DAG.getConstant(Offset, dl,
276 SrcVT)),
277 DAG.getConstant(BytesLeft, dl, SizeVT),
278 Align, isVolatile, AlwaysInline, false,
279 DstPtrInfo.getWithOffset(Offset),
280 SrcPtrInfo.getWithOffset(Offset)));
281 }
282
283 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Results);
284 }
285