1 //===-- X86SelectionDAGInfo.cpp - X86 SelectionDAG Info -------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This file implements the X86SelectionDAGInfo class.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "X86SelectionDAGInfo.h"
14 #include "X86ISelLowering.h"
15 #include "X86InstrInfo.h"
16 #include "X86RegisterInfo.h"
17 #include "X86Subtarget.h"
18 #include "llvm/CodeGen/MachineFrameInfo.h"
19 #include "llvm/CodeGen/SelectionDAG.h"
20 #include "llvm/CodeGen/TargetLowering.h"
21 #include "llvm/IR/DerivedTypes.h"
22 
23 using namespace llvm;
24 
25 #define DEBUG_TYPE "x86-selectiondag-info"
26 
27 static cl::opt<bool>
28     UseFSRMForMemcpy("x86-use-fsrm-for-memcpy", cl::Hidden, cl::init(false),
29                      cl::desc("Use fast short rep mov in memcpy lowering"));
30 
31 bool X86SelectionDAGInfo::isBaseRegConflictPossible(
32     SelectionDAG &DAG, ArrayRef<MCPhysReg> ClobberSet) const {
33   // We cannot use TRI->hasBasePointer() until *after* we select all basic
34   // blocks.  Legalization may introduce new stack temporaries with large
35   // alignment requirements.  Fall back to generic code if there are any
36   // dynamic stack adjustments (hopefully rare) and the base pointer would
37   // conflict if we had to use it.
38   MachineFrameInfo &MFI = DAG.getMachineFunction().getFrameInfo();
39   if (!MFI.hasVarSizedObjects() && !MFI.hasOpaqueSPAdjustment())
40     return false;
41 
42   const X86RegisterInfo *TRI = static_cast<const X86RegisterInfo *>(
43       DAG.getSubtarget().getRegisterInfo());
44   return llvm::is_contained(ClobberSet, TRI->getBaseRegister());
45 }
46 
47 SDValue X86SelectionDAGInfo::EmitTargetCodeForMemset(
48     SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Val,
49     SDValue Size, Align Alignment, 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 MCPhysReg 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 (Alignment < Align(4) || !ConstantSize ||
70       ConstantSize->getZExtValue() > Subtarget.getMaxInlineSizeThreshold()) {
71     // Check to see if there is a specialized entry-point for memory zeroing.
72     ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Val);
73 
74     if (const char *bzeroName = (ValC && ValC->isNullValue())
75         ? DAG.getTargetLoweringInfo().getLibcallName(RTLIB::BZERO)
76         : nullptr) {
77       const TargetLowering &TLI = DAG.getTargetLoweringInfo();
78       EVT IntPtr = TLI.getPointerTy(DAG.getDataLayout());
79       Type *IntPtrTy = DAG.getDataLayout().getIntPtrType(*DAG.getContext());
80       TargetLowering::ArgListTy Args;
81       TargetLowering::ArgListEntry Entry;
82       Entry.Node = Dst;
83       Entry.Ty = IntPtrTy;
84       Args.push_back(Entry);
85       Entry.Node = Size;
86       Args.push_back(Entry);
87 
88       TargetLowering::CallLoweringInfo CLI(DAG);
89       CLI.setDebugLoc(dl)
90           .setChain(Chain)
91           .setLibCallee(CallingConv::C, Type::getVoidTy(*DAG.getContext()),
92                         DAG.getExternalSymbol(bzeroName, IntPtr),
93                         std::move(Args))
94           .setDiscardResult();
95 
96       std::pair<SDValue,SDValue> CallResult = TLI.LowerCallTo(CLI);
97       return CallResult.second;
98     }
99 
100     // Otherwise have the target-independent code call memset.
101     return SDValue();
102   }
103 
104   uint64_t SizeVal = ConstantSize->getZExtValue();
105   SDValue InFlag;
106   EVT AVT;
107   SDValue Count;
108   ConstantSDNode *ValC = dyn_cast<ConstantSDNode>(Val);
109   unsigned BytesLeft = 0;
110   if (ValC) {
111     unsigned ValReg;
112     uint64_t Val = ValC->getZExtValue() & 255;
113 
114     // If the value is a constant, then we can potentially use larger sets.
115     if (Alignment > Align(2)) {
116       // DWORD aligned
117       AVT = MVT::i32;
118       ValReg = X86::EAX;
119       Val = (Val << 8)  | Val;
120       Val = (Val << 16) | Val;
121       if (Subtarget.is64Bit() && Alignment > Align(8)) { // QWORD aligned
122         AVT = MVT::i64;
123         ValReg = X86::RAX;
124         Val = (Val << 32) | Val;
125       }
126     } else if (Alignment == Align(2)) {
127       // WORD aligned
128       AVT = MVT::i16;
129       ValReg = X86::AX;
130       Val = (Val << 8) | Val;
131     } else {
132       // Byte aligned
133       AVT = MVT::i8;
134       ValReg = X86::AL;
135       Count = DAG.getIntPtrConstant(SizeVal, dl);
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, Val, InFlag);
151     InFlag = Chain.getValue(1);
152   }
153 
154   bool Use64BitRegs = Subtarget.isTarget64BitLP64();
155   Chain = DAG.getCopyToReg(Chain, dl, Use64BitRegs ? X86::RCX : X86::ECX,
156                            Count, InFlag);
157   InFlag = Chain.getValue(1);
158   Chain = DAG.getCopyToReg(Chain, dl, Use64BitRegs ? X86::RDI : X86::EDI,
159                            Dst, InFlag);
160   InFlag = Chain.getValue(1);
161 
162   SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
163   SDValue Ops[] = { Chain, DAG.getValueType(AVT), InFlag };
164   Chain = DAG.getNode(X86ISD::REP_STOS, dl, Tys, Ops);
165 
166   if (BytesLeft) {
167     // Handle the last 1 - 7 bytes.
168     unsigned Offset = SizeVal - BytesLeft;
169     EVT AddrVT = Dst.getValueType();
170     EVT SizeVT = Size.getValueType();
171 
172     Chain =
173         DAG.getMemset(Chain, dl,
174                       DAG.getNode(ISD::ADD, dl, AddrVT, Dst,
175                                   DAG.getConstant(Offset, dl, AddrVT)),
176                       Val, DAG.getConstant(BytesLeft, dl, SizeVT), Alignment,
177                       isVolatile, false, DstPtrInfo.getWithOffset(Offset));
178   }
179 
180   // TODO: Use a Tokenfactor, as in memcpy, instead of a single chain.
181   return Chain;
182 }
183 
184 /// Emit a single REP MOVS{B,W,D,Q} instruction.
185 static SDValue emitRepmovs(const X86Subtarget &Subtarget, SelectionDAG &DAG,
186                            const SDLoc &dl, SDValue Chain, SDValue Dst,
187                            SDValue Src, SDValue Size, MVT AVT) {
188   const bool Use64BitRegs = Subtarget.isTarget64BitLP64();
189   const unsigned CX = Use64BitRegs ? X86::RCX : X86::ECX;
190   const unsigned DI = Use64BitRegs ? X86::RDI : X86::EDI;
191   const unsigned SI = Use64BitRegs ? X86::RSI : X86::ESI;
192 
193   SDValue InFlag;
194   Chain = DAG.getCopyToReg(Chain, dl, CX, Size, InFlag);
195   InFlag = Chain.getValue(1);
196   Chain = DAG.getCopyToReg(Chain, dl, DI, Dst, InFlag);
197   InFlag = Chain.getValue(1);
198   Chain = DAG.getCopyToReg(Chain, dl, SI, Src, InFlag);
199   InFlag = Chain.getValue(1);
200 
201   SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
202   SDValue Ops[] = {Chain, DAG.getValueType(AVT), InFlag};
203   return DAG.getNode(X86ISD::REP_MOVS, dl, Tys, Ops);
204 }
205 
206 /// Emit a single REP MOVSB instruction for a particular constant size.
207 static SDValue emitRepmovsB(const X86Subtarget &Subtarget, SelectionDAG &DAG,
208                             const SDLoc &dl, SDValue Chain, SDValue Dst,
209                             SDValue Src, uint64_t Size) {
210   return emitRepmovs(Subtarget, DAG, dl, Chain, Dst, Src,
211                      DAG.getIntPtrConstant(Size, dl), MVT::i8);
212 }
213 
214 /// Returns the best type to use with repmovs depending on alignment.
215 static MVT getOptimalRepmovsType(const X86Subtarget &Subtarget,
216                                  uint64_t Align) {
217   assert((Align != 0) && "Align is normalized");
218   assert(isPowerOf2_64(Align) && "Align is a power of 2");
219   switch (Align) {
220   case 1:
221     return MVT::i8;
222   case 2:
223     return MVT::i16;
224   case 4:
225     return MVT::i32;
226   default:
227     return Subtarget.is64Bit() ? MVT::i64 : MVT::i32;
228   }
229 }
230 
231 /// Returns a REP MOVS instruction, possibly with a few load/stores to implement
232 /// a constant size memory copy. In some cases where we know REP MOVS is
233 /// inefficient we return an empty SDValue so the calling code can either
234 /// generate a load/store sequence or call the runtime memcpy function.
235 static SDValue emitConstantSizeRepmov(
236     SelectionDAG &DAG, const X86Subtarget &Subtarget, const SDLoc &dl,
237     SDValue Chain, SDValue Dst, SDValue Src, uint64_t Size, EVT SizeVT,
238     unsigned Align, bool isVolatile, bool AlwaysInline,
239     MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) {
240 
241   /// TODO: Revisit next line: big copy with ERMSB on march >= haswell are very
242   /// efficient.
243   if (!AlwaysInline && Size > Subtarget.getMaxInlineSizeThreshold())
244     return SDValue();
245 
246   /// If we have enhanced repmovs we use it.
247   if (Subtarget.hasERMSB())
248     return emitRepmovsB(Subtarget, DAG, dl, Chain, Dst, Src, Size);
249 
250   assert(!Subtarget.hasERMSB() && "No efficient RepMovs");
251   /// We assume runtime memcpy will do a better job for unaligned copies when
252   /// ERMS is not present.
253   if (!AlwaysInline && (Align & 3) != 0)
254     return SDValue();
255 
256   const MVT BlockType = getOptimalRepmovsType(Subtarget, Align);
257   const uint64_t BlockBytes = BlockType.getSizeInBits() / 8;
258   const uint64_t BlockCount = Size / BlockBytes;
259   const uint64_t BytesLeft = Size % BlockBytes;
260   SDValue RepMovs =
261       emitRepmovs(Subtarget, DAG, dl, Chain, Dst, Src,
262                   DAG.getIntPtrConstant(BlockCount, dl), BlockType);
263 
264   /// RepMov can process the whole length.
265   if (BytesLeft == 0)
266     return RepMovs;
267 
268   assert(BytesLeft && "We have leftover at this point");
269 
270   /// In case we optimize for size we use repmovsb even if it's less efficient
271   /// so we can save the loads/stores of the leftover.
272   if (DAG.getMachineFunction().getFunction().hasMinSize())
273     return emitRepmovsB(Subtarget, DAG, dl, Chain, Dst, Src, Size);
274 
275   // Handle the last 1 - 7 bytes.
276   SmallVector<SDValue, 4> Results;
277   Results.push_back(RepMovs);
278   unsigned Offset = Size - BytesLeft;
279   EVT DstVT = Dst.getValueType();
280   EVT SrcVT = Src.getValueType();
281   Results.push_back(DAG.getMemcpy(
282       Chain, dl,
283       DAG.getNode(ISD::ADD, dl, DstVT, Dst, DAG.getConstant(Offset, dl, DstVT)),
284       DAG.getNode(ISD::ADD, dl, SrcVT, Src, DAG.getConstant(Offset, dl, SrcVT)),
285       DAG.getConstant(BytesLeft, dl, SizeVT), llvm::Align(Align), isVolatile,
286       /*AlwaysInline*/ true, /*isTailCall*/ false,
287       DstPtrInfo.getWithOffset(Offset), SrcPtrInfo.getWithOffset(Offset)));
288   return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Results);
289 }
290 
291 SDValue X86SelectionDAGInfo::EmitTargetCodeForMemcpy(
292     SelectionDAG &DAG, const SDLoc &dl, SDValue Chain, SDValue Dst, SDValue Src,
293     SDValue Size, Align Alignment, bool isVolatile, bool AlwaysInline,
294     MachinePointerInfo DstPtrInfo, MachinePointerInfo SrcPtrInfo) const {
295   // If to a segment-relative address space, use the default lowering.
296   if (DstPtrInfo.getAddrSpace() >= 256 || SrcPtrInfo.getAddrSpace() >= 256)
297     return SDValue();
298 
299   // If the base registers conflict with our physical registers, use the default
300   // lowering.
301   const MCPhysReg ClobberSet[] = {X86::RCX, X86::RSI, X86::RDI,
302                                   X86::ECX, X86::ESI, X86::EDI};
303   if (isBaseRegConflictPossible(DAG, ClobberSet))
304     return SDValue();
305 
306   const X86Subtarget &Subtarget =
307       DAG.getMachineFunction().getSubtarget<X86Subtarget>();
308 
309   // If enabled and available, use fast short rep mov.
310   if (UseFSRMForMemcpy && Subtarget.hasFSRM())
311     return emitRepmovs(Subtarget, DAG, dl, Chain, Dst, Src, Size, MVT::i8);
312 
313   /// Handle constant sizes,
314   if (ConstantSDNode *ConstantSize = dyn_cast<ConstantSDNode>(Size))
315     return emitConstantSizeRepmov(
316         DAG, Subtarget, dl, Chain, Dst, Src, ConstantSize->getZExtValue(),
317         Size.getValueType(), Alignment.value(), isVolatile, AlwaysInline,
318         DstPtrInfo, SrcPtrInfo);
319 
320   return SDValue();
321 }
322