| File: | llvm/include/llvm/CodeGen/SelectionDAGNodes.h |
| Warning: | line 1150, column 10 Called C++ object pointer is null |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | //===-- TargetLowering.cpp - Implement the TargetLowering class -----------===// | ||||||
| 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 implements the TargetLowering class. | ||||||
| 10 | // | ||||||
| 11 | //===----------------------------------------------------------------------===// | ||||||
| 12 | |||||||
| 13 | #include "llvm/CodeGen/TargetLowering.h" | ||||||
| 14 | #include "llvm/ADT/STLExtras.h" | ||||||
| 15 | #include "llvm/CodeGen/CallingConvLower.h" | ||||||
| 16 | #include "llvm/CodeGen/MachineFrameInfo.h" | ||||||
| 17 | #include "llvm/CodeGen/MachineFunction.h" | ||||||
| 18 | #include "llvm/CodeGen/MachineJumpTableInfo.h" | ||||||
| 19 | #include "llvm/CodeGen/MachineRegisterInfo.h" | ||||||
| 20 | #include "llvm/CodeGen/SelectionDAG.h" | ||||||
| 21 | #include "llvm/CodeGen/TargetRegisterInfo.h" | ||||||
| 22 | #include "llvm/CodeGen/TargetSubtargetInfo.h" | ||||||
| 23 | #include "llvm/IR/DataLayout.h" | ||||||
| 24 | #include "llvm/IR/DerivedTypes.h" | ||||||
| 25 | #include "llvm/IR/GlobalVariable.h" | ||||||
| 26 | #include "llvm/IR/LLVMContext.h" | ||||||
| 27 | #include "llvm/MC/MCAsmInfo.h" | ||||||
| 28 | #include "llvm/MC/MCExpr.h" | ||||||
| 29 | #include "llvm/Support/ErrorHandling.h" | ||||||
| 30 | #include "llvm/Support/KnownBits.h" | ||||||
| 31 | #include "llvm/Support/MathExtras.h" | ||||||
| 32 | #include "llvm/Target/TargetLoweringObjectFile.h" | ||||||
| 33 | #include "llvm/Target/TargetMachine.h" | ||||||
| 34 | #include <cctype> | ||||||
| 35 | using namespace llvm; | ||||||
| 36 | |||||||
| 37 | /// NOTE: The TargetMachine owns TLOF. | ||||||
| 38 | TargetLowering::TargetLowering(const TargetMachine &tm) | ||||||
| 39 | : TargetLoweringBase(tm) {} | ||||||
| 40 | |||||||
| 41 | const char *TargetLowering::getTargetNodeName(unsigned Opcode) const { | ||||||
| 42 | return nullptr; | ||||||
| 43 | } | ||||||
| 44 | |||||||
| 45 | bool TargetLowering::isPositionIndependent() const { | ||||||
| 46 | return getTargetMachine().isPositionIndependent(); | ||||||
| 47 | } | ||||||
| 48 | |||||||
| 49 | /// Check whether a given call node is in tail position within its function. If | ||||||
| 50 | /// so, it sets Chain to the input chain of the tail call. | ||||||
| 51 | bool TargetLowering::isInTailCallPosition(SelectionDAG &DAG, SDNode *Node, | ||||||
| 52 | SDValue &Chain) const { | ||||||
| 53 | const Function &F = DAG.getMachineFunction().getFunction(); | ||||||
| 54 | |||||||
| 55 | // First, check if tail calls have been disabled in this function. | ||||||
| 56 | if (F.getFnAttribute("disable-tail-calls").getValueAsString() == "true") | ||||||
| 57 | return false; | ||||||
| 58 | |||||||
| 59 | // Conservatively require the attributes of the call to match those of | ||||||
| 60 | // the return. Ignore NoAlias and NonNull because they don't affect the | ||||||
| 61 | // call sequence. | ||||||
| 62 | AttributeList CallerAttrs = F.getAttributes(); | ||||||
| 63 | if (AttrBuilder(CallerAttrs, AttributeList::ReturnIndex) | ||||||
| 64 | .removeAttribute(Attribute::NoAlias) | ||||||
| 65 | .removeAttribute(Attribute::NonNull) | ||||||
| 66 | .hasAttributes()) | ||||||
| 67 | return false; | ||||||
| 68 | |||||||
| 69 | // It's not safe to eliminate the sign / zero extension of the return value. | ||||||
| 70 | if (CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::ZExt) || | ||||||
| 71 | CallerAttrs.hasAttribute(AttributeList::ReturnIndex, Attribute::SExt)) | ||||||
| 72 | return false; | ||||||
| 73 | |||||||
| 74 | // Check if the only use is a function return node. | ||||||
| 75 | return isUsedByReturnOnly(Node, Chain); | ||||||
| 76 | } | ||||||
| 77 | |||||||
| 78 | bool TargetLowering::parametersInCSRMatch(const MachineRegisterInfo &MRI, | ||||||
| 79 | const uint32_t *CallerPreservedMask, | ||||||
| 80 | const SmallVectorImpl<CCValAssign> &ArgLocs, | ||||||
| 81 | const SmallVectorImpl<SDValue> &OutVals) const { | ||||||
| 82 | for (unsigned I = 0, E = ArgLocs.size(); I != E; ++I) { | ||||||
| 83 | const CCValAssign &ArgLoc = ArgLocs[I]; | ||||||
| 84 | if (!ArgLoc.isRegLoc()) | ||||||
| 85 | continue; | ||||||
| 86 | Register Reg = ArgLoc.getLocReg(); | ||||||
| 87 | // Only look at callee saved registers. | ||||||
| 88 | if (MachineOperand::clobbersPhysReg(CallerPreservedMask, Reg)) | ||||||
| 89 | continue; | ||||||
| 90 | // Check that we pass the value used for the caller. | ||||||
| 91 | // (We look for a CopyFromReg reading a virtual register that is used | ||||||
| 92 | // for the function live-in value of register Reg) | ||||||
| 93 | SDValue Value = OutVals[I]; | ||||||
| 94 | if (Value->getOpcode() != ISD::CopyFromReg) | ||||||
| 95 | return false; | ||||||
| 96 | unsigned ArgReg = cast<RegisterSDNode>(Value->getOperand(1))->getReg(); | ||||||
| 97 | if (MRI.getLiveInPhysReg(ArgReg) != Reg) | ||||||
| 98 | return false; | ||||||
| 99 | } | ||||||
| 100 | return true; | ||||||
| 101 | } | ||||||
| 102 | |||||||
| 103 | /// Set CallLoweringInfo attribute flags based on a call instruction | ||||||
| 104 | /// and called function attributes. | ||||||
| 105 | void TargetLoweringBase::ArgListEntry::setAttributes(const CallBase *Call, | ||||||
| 106 | unsigned ArgIdx) { | ||||||
| 107 | IsSExt = Call->paramHasAttr(ArgIdx, Attribute::SExt); | ||||||
| 108 | IsZExt = Call->paramHasAttr(ArgIdx, Attribute::ZExt); | ||||||
| 109 | IsInReg = Call->paramHasAttr(ArgIdx, Attribute::InReg); | ||||||
| 110 | IsSRet = Call->paramHasAttr(ArgIdx, Attribute::StructRet); | ||||||
| 111 | IsNest = Call->paramHasAttr(ArgIdx, Attribute::Nest); | ||||||
| 112 | IsByVal = Call->paramHasAttr(ArgIdx, Attribute::ByVal); | ||||||
| 113 | IsInAlloca = Call->paramHasAttr(ArgIdx, Attribute::InAlloca); | ||||||
| 114 | IsReturned = Call->paramHasAttr(ArgIdx, Attribute::Returned); | ||||||
| 115 | IsSwiftSelf = Call->paramHasAttr(ArgIdx, Attribute::SwiftSelf); | ||||||
| 116 | IsSwiftError = Call->paramHasAttr(ArgIdx, Attribute::SwiftError); | ||||||
| 117 | Alignment = Call->getParamAlignment(ArgIdx); | ||||||
| 118 | ByValType = nullptr; | ||||||
| 119 | if (Call->paramHasAttr(ArgIdx, Attribute::ByVal)) | ||||||
| 120 | ByValType = Call->getParamByValType(ArgIdx); | ||||||
| 121 | } | ||||||
| 122 | |||||||
| 123 | /// Generate a libcall taking the given operands as arguments and returning a | ||||||
| 124 | /// result of type RetVT. | ||||||
| 125 | std::pair<SDValue, SDValue> | ||||||
| 126 | TargetLowering::makeLibCall(SelectionDAG &DAG, RTLIB::Libcall LC, EVT RetVT, | ||||||
| 127 | ArrayRef<SDValue> Ops, | ||||||
| 128 | MakeLibCallOptions CallOptions, | ||||||
| 129 | const SDLoc &dl, | ||||||
| 130 | SDValue InChain) const { | ||||||
| 131 | if (!InChain) | ||||||
| 132 | InChain = DAG.getEntryNode(); | ||||||
| 133 | |||||||
| 134 | TargetLowering::ArgListTy Args; | ||||||
| 135 | Args.reserve(Ops.size()); | ||||||
| 136 | |||||||
| 137 | TargetLowering::ArgListEntry Entry; | ||||||
| 138 | for (unsigned i = 0; i < Ops.size(); ++i) { | ||||||
| 139 | SDValue NewOp = Ops[i]; | ||||||
| 140 | Entry.Node = NewOp; | ||||||
| 141 | Entry.Ty = Entry.Node.getValueType().getTypeForEVT(*DAG.getContext()); | ||||||
| 142 | Entry.IsSExt = shouldSignExtendTypeInLibCall(NewOp.getValueType(), | ||||||
| 143 | CallOptions.IsSExt); | ||||||
| 144 | Entry.IsZExt = !Entry.IsSExt; | ||||||
| 145 | |||||||
| 146 | if (CallOptions.IsSoften && | ||||||
| 147 | !shouldExtendTypeInLibCall(CallOptions.OpsVTBeforeSoften[i])) { | ||||||
| 148 | Entry.IsSExt = Entry.IsZExt = false; | ||||||
| 149 | } | ||||||
| 150 | Args.push_back(Entry); | ||||||
| 151 | } | ||||||
| 152 | |||||||
| 153 | if (LC == RTLIB::UNKNOWN_LIBCALL) | ||||||
| 154 | report_fatal_error("Unsupported library call operation!"); | ||||||
| 155 | SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC), | ||||||
| 156 | getPointerTy(DAG.getDataLayout())); | ||||||
| 157 | |||||||
| 158 | Type *RetTy = RetVT.getTypeForEVT(*DAG.getContext()); | ||||||
| 159 | TargetLowering::CallLoweringInfo CLI(DAG); | ||||||
| 160 | bool signExtend = shouldSignExtendTypeInLibCall(RetVT, CallOptions.IsSExt); | ||||||
| 161 | bool zeroExtend = !signExtend; | ||||||
| 162 | |||||||
| 163 | if (CallOptions.IsSoften && | ||||||
| 164 | !shouldExtendTypeInLibCall(CallOptions.RetVTBeforeSoften)) { | ||||||
| 165 | signExtend = zeroExtend = false; | ||||||
| 166 | } | ||||||
| 167 | |||||||
| 168 | CLI.setDebugLoc(dl) | ||||||
| 169 | .setChain(InChain) | ||||||
| 170 | .setLibCallee(getLibcallCallingConv(LC), RetTy, Callee, std::move(Args)) | ||||||
| 171 | .setNoReturn(CallOptions.DoesNotReturn) | ||||||
| 172 | .setDiscardResult(!CallOptions.IsReturnValueUsed) | ||||||
| 173 | .setIsPostTypeLegalization(CallOptions.IsPostTypeLegalization) | ||||||
| 174 | .setSExtResult(signExtend) | ||||||
| 175 | .setZExtResult(zeroExtend); | ||||||
| 176 | return LowerCallTo(CLI); | ||||||
| 177 | } | ||||||
| 178 | |||||||
| 179 | bool | ||||||
| 180 | TargetLowering::findOptimalMemOpLowering(std::vector<EVT> &MemOps, | ||||||
| 181 | unsigned Limit, uint64_t Size, | ||||||
| 182 | unsigned DstAlign, unsigned SrcAlign, | ||||||
| 183 | bool IsMemset, | ||||||
| 184 | bool ZeroMemset, | ||||||
| 185 | bool MemcpyStrSrc, | ||||||
| 186 | bool AllowOverlap, | ||||||
| 187 | unsigned DstAS, unsigned SrcAS, | ||||||
| 188 | const AttributeList &FuncAttributes) const { | ||||||
| 189 | // If 'SrcAlign' is zero, that means the memory operation does not need to | ||||||
| 190 | // load the value, i.e. memset or memcpy from constant string. Otherwise, | ||||||
| 191 | // it's the inferred alignment of the source. 'DstAlign', on the other hand, | ||||||
| 192 | // is the specified alignment of the memory operation. If it is zero, that | ||||||
| 193 | // means it's possible to change the alignment of the destination. | ||||||
| 194 | // 'MemcpyStrSrc' indicates whether the memcpy source is constant so it does | ||||||
| 195 | // not need to be loaded. | ||||||
| 196 | if (!(SrcAlign == 0 || SrcAlign >= DstAlign)) | ||||||
| 197 | return false; | ||||||
| 198 | |||||||
| 199 | EVT VT = getOptimalMemOpType(Size, DstAlign, SrcAlign, | ||||||
| 200 | IsMemset, ZeroMemset, MemcpyStrSrc, | ||||||
| 201 | FuncAttributes); | ||||||
| 202 | |||||||
| 203 | if (VT == MVT::Other) { | ||||||
| 204 | // Use the largest integer type whose alignment constraints are satisfied. | ||||||
| 205 | // We only need to check DstAlign here as SrcAlign is always greater or | ||||||
| 206 | // equal to DstAlign (or zero). | ||||||
| 207 | VT = MVT::i64; | ||||||
| 208 | while (DstAlign && DstAlign < VT.getSizeInBits() / 8 && | ||||||
| 209 | !allowsMisalignedMemoryAccesses(VT, DstAS, DstAlign)) | ||||||
| 210 | VT = (MVT::SimpleValueType)(VT.getSimpleVT().SimpleTy - 1); | ||||||
| 211 | assert(VT.isInteger())((VT.isInteger()) ? static_cast<void> (0) : __assert_fail ("VT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 211, __PRETTY_FUNCTION__)); | ||||||
| 212 | |||||||
| 213 | // Find the largest legal integer type. | ||||||
| 214 | MVT LVT = MVT::i64; | ||||||
| 215 | while (!isTypeLegal(LVT)) | ||||||
| 216 | LVT = (MVT::SimpleValueType)(LVT.SimpleTy - 1); | ||||||
| 217 | assert(LVT.isInteger())((LVT.isInteger()) ? static_cast<void> (0) : __assert_fail ("LVT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 217, __PRETTY_FUNCTION__)); | ||||||
| 218 | |||||||
| 219 | // If the type we've chosen is larger than the largest legal integer type | ||||||
| 220 | // then use that instead. | ||||||
| 221 | if (VT.bitsGT(LVT)) | ||||||
| 222 | VT = LVT; | ||||||
| 223 | } | ||||||
| 224 | |||||||
| 225 | unsigned NumMemOps = 0; | ||||||
| 226 | while (Size != 0) { | ||||||
| 227 | unsigned VTSize = VT.getSizeInBits() / 8; | ||||||
| 228 | while (VTSize > Size) { | ||||||
| 229 | // For now, only use non-vector load / store's for the left-over pieces. | ||||||
| 230 | EVT NewVT = VT; | ||||||
| 231 | unsigned NewVTSize; | ||||||
| 232 | |||||||
| 233 | bool Found = false; | ||||||
| 234 | if (VT.isVector() || VT.isFloatingPoint()) { | ||||||
| 235 | NewVT = (VT.getSizeInBits() > 64) ? MVT::i64 : MVT::i32; | ||||||
| 236 | if (isOperationLegalOrCustom(ISD::STORE, NewVT) && | ||||||
| 237 | isSafeMemOpType(NewVT.getSimpleVT())) | ||||||
| 238 | Found = true; | ||||||
| 239 | else if (NewVT == MVT::i64 && | ||||||
| 240 | isOperationLegalOrCustom(ISD::STORE, MVT::f64) && | ||||||
| 241 | isSafeMemOpType(MVT::f64)) { | ||||||
| 242 | // i64 is usually not legal on 32-bit targets, but f64 may be. | ||||||
| 243 | NewVT = MVT::f64; | ||||||
| 244 | Found = true; | ||||||
| 245 | } | ||||||
| 246 | } | ||||||
| 247 | |||||||
| 248 | if (!Found) { | ||||||
| 249 | do { | ||||||
| 250 | NewVT = (MVT::SimpleValueType)(NewVT.getSimpleVT().SimpleTy - 1); | ||||||
| 251 | if (NewVT == MVT::i8) | ||||||
| 252 | break; | ||||||
| 253 | } while (!isSafeMemOpType(NewVT.getSimpleVT())); | ||||||
| 254 | } | ||||||
| 255 | NewVTSize = NewVT.getSizeInBits() / 8; | ||||||
| 256 | |||||||
| 257 | // If the new VT cannot cover all of the remaining bits, then consider | ||||||
| 258 | // issuing a (or a pair of) unaligned and overlapping load / store. | ||||||
| 259 | bool Fast; | ||||||
| 260 | if (NumMemOps && AllowOverlap && NewVTSize < Size && | ||||||
| 261 | allowsMisalignedMemoryAccesses(VT, DstAS, DstAlign, | ||||||
| 262 | MachineMemOperand::MONone, &Fast) && | ||||||
| 263 | Fast) | ||||||
| 264 | VTSize = Size; | ||||||
| 265 | else { | ||||||
| 266 | VT = NewVT; | ||||||
| 267 | VTSize = NewVTSize; | ||||||
| 268 | } | ||||||
| 269 | } | ||||||
| 270 | |||||||
| 271 | if (++NumMemOps > Limit) | ||||||
| 272 | return false; | ||||||
| 273 | |||||||
| 274 | MemOps.push_back(VT); | ||||||
| 275 | Size -= VTSize; | ||||||
| 276 | } | ||||||
| 277 | |||||||
| 278 | return true; | ||||||
| 279 | } | ||||||
| 280 | |||||||
| 281 | /// Soften the operands of a comparison. This code is shared among BR_CC, | ||||||
| 282 | /// SELECT_CC, and SETCC handlers. | ||||||
| 283 | void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT, | ||||||
| 284 | SDValue &NewLHS, SDValue &NewRHS, | ||||||
| 285 | ISD::CondCode &CCCode, | ||||||
| 286 | const SDLoc &dl, const SDValue OldLHS, | ||||||
| 287 | const SDValue OldRHS) const { | ||||||
| 288 | SDValue Chain; | ||||||
| 289 | return softenSetCCOperands(DAG, VT, NewLHS, NewRHS, CCCode, dl, OldLHS, | ||||||
| 290 | OldRHS, Chain); | ||||||
| 291 | } | ||||||
| 292 | |||||||
| 293 | void TargetLowering::softenSetCCOperands(SelectionDAG &DAG, EVT VT, | ||||||
| 294 | SDValue &NewLHS, SDValue &NewRHS, | ||||||
| 295 | ISD::CondCode &CCCode, | ||||||
| 296 | const SDLoc &dl, const SDValue OldLHS, | ||||||
| 297 | const SDValue OldRHS, | ||||||
| 298 | SDValue &Chain, | ||||||
| 299 | bool IsSignaling) const { | ||||||
| 300 | // FIXME: Currently we cannot really respect all IEEE predicates due to libgcc | ||||||
| 301 | // not supporting it. We can update this code when libgcc provides such | ||||||
| 302 | // functions. | ||||||
| 303 | |||||||
| 304 | assert((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128)(((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128) && "Unsupported setcc type!") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128) && \"Unsupported setcc type!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 305, __PRETTY_FUNCTION__)) | ||||||
| 305 | && "Unsupported setcc type!")(((VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128) && "Unsupported setcc type!") ? static_cast <void> (0) : __assert_fail ("(VT == MVT::f32 || VT == MVT::f64 || VT == MVT::f128 || VT == MVT::ppcf128) && \"Unsupported setcc type!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 305, __PRETTY_FUNCTION__)); | ||||||
| 306 | |||||||
| 307 | // Expand into one or more soft-fp libcall(s). | ||||||
| 308 | RTLIB::Libcall LC1 = RTLIB::UNKNOWN_LIBCALL, LC2 = RTLIB::UNKNOWN_LIBCALL; | ||||||
| 309 | bool ShouldInvertCC = false; | ||||||
| 310 | switch (CCCode) { | ||||||
| 311 | case ISD::SETEQ: | ||||||
| 312 | case ISD::SETOEQ: | ||||||
| 313 | LC1 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : | ||||||
| 314 | (VT == MVT::f64) ? RTLIB::OEQ_F64 : | ||||||
| 315 | (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128; | ||||||
| 316 | break; | ||||||
| 317 | case ISD::SETNE: | ||||||
| 318 | case ISD::SETUNE: | ||||||
| 319 | LC1 = (VT == MVT::f32) ? RTLIB::UNE_F32 : | ||||||
| 320 | (VT == MVT::f64) ? RTLIB::UNE_F64 : | ||||||
| 321 | (VT == MVT::f128) ? RTLIB::UNE_F128 : RTLIB::UNE_PPCF128; | ||||||
| 322 | break; | ||||||
| 323 | case ISD::SETGE: | ||||||
| 324 | case ISD::SETOGE: | ||||||
| 325 | LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : | ||||||
| 326 | (VT == MVT::f64) ? RTLIB::OGE_F64 : | ||||||
| 327 | (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128; | ||||||
| 328 | break; | ||||||
| 329 | case ISD::SETLT: | ||||||
| 330 | case ISD::SETOLT: | ||||||
| 331 | LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : | ||||||
| 332 | (VT == MVT::f64) ? RTLIB::OLT_F64 : | ||||||
| 333 | (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128; | ||||||
| 334 | break; | ||||||
| 335 | case ISD::SETLE: | ||||||
| 336 | case ISD::SETOLE: | ||||||
| 337 | LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : | ||||||
| 338 | (VT == MVT::f64) ? RTLIB::OLE_F64 : | ||||||
| 339 | (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128; | ||||||
| 340 | break; | ||||||
| 341 | case ISD::SETGT: | ||||||
| 342 | case ISD::SETOGT: | ||||||
| 343 | LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : | ||||||
| 344 | (VT == MVT::f64) ? RTLIB::OGT_F64 : | ||||||
| 345 | (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128; | ||||||
| 346 | break; | ||||||
| 347 | case ISD::SETO: | ||||||
| 348 | ShouldInvertCC = true; | ||||||
| 349 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 350 | case ISD::SETUO: | ||||||
| 351 | LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : | ||||||
| 352 | (VT == MVT::f64) ? RTLIB::UO_F64 : | ||||||
| 353 | (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128; | ||||||
| 354 | break; | ||||||
| 355 | case ISD::SETONE: | ||||||
| 356 | // SETONE = O && UNE | ||||||
| 357 | ShouldInvertCC = true; | ||||||
| 358 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 359 | case ISD::SETUEQ: | ||||||
| 360 | LC1 = (VT == MVT::f32) ? RTLIB::UO_F32 : | ||||||
| 361 | (VT == MVT::f64) ? RTLIB::UO_F64 : | ||||||
| 362 | (VT == MVT::f128) ? RTLIB::UO_F128 : RTLIB::UO_PPCF128; | ||||||
| 363 | LC2 = (VT == MVT::f32) ? RTLIB::OEQ_F32 : | ||||||
| 364 | (VT == MVT::f64) ? RTLIB::OEQ_F64 : | ||||||
| 365 | (VT == MVT::f128) ? RTLIB::OEQ_F128 : RTLIB::OEQ_PPCF128; | ||||||
| 366 | break; | ||||||
| 367 | default: | ||||||
| 368 | // Invert CC for unordered comparisons | ||||||
| 369 | ShouldInvertCC = true; | ||||||
| 370 | switch (CCCode) { | ||||||
| 371 | case ISD::SETULT: | ||||||
| 372 | LC1 = (VT == MVT::f32) ? RTLIB::OGE_F32 : | ||||||
| 373 | (VT == MVT::f64) ? RTLIB::OGE_F64 : | ||||||
| 374 | (VT == MVT::f128) ? RTLIB::OGE_F128 : RTLIB::OGE_PPCF128; | ||||||
| 375 | break; | ||||||
| 376 | case ISD::SETULE: | ||||||
| 377 | LC1 = (VT == MVT::f32) ? RTLIB::OGT_F32 : | ||||||
| 378 | (VT == MVT::f64) ? RTLIB::OGT_F64 : | ||||||
| 379 | (VT == MVT::f128) ? RTLIB::OGT_F128 : RTLIB::OGT_PPCF128; | ||||||
| 380 | break; | ||||||
| 381 | case ISD::SETUGT: | ||||||
| 382 | LC1 = (VT == MVT::f32) ? RTLIB::OLE_F32 : | ||||||
| 383 | (VT == MVT::f64) ? RTLIB::OLE_F64 : | ||||||
| 384 | (VT == MVT::f128) ? RTLIB::OLE_F128 : RTLIB::OLE_PPCF128; | ||||||
| 385 | break; | ||||||
| 386 | case ISD::SETUGE: | ||||||
| 387 | LC1 = (VT == MVT::f32) ? RTLIB::OLT_F32 : | ||||||
| 388 | (VT == MVT::f64) ? RTLIB::OLT_F64 : | ||||||
| 389 | (VT == MVT::f128) ? RTLIB::OLT_F128 : RTLIB::OLT_PPCF128; | ||||||
| 390 | break; | ||||||
| 391 | default: llvm_unreachable("Do not know how to soften this setcc!")::llvm::llvm_unreachable_internal("Do not know how to soften this setcc!" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 391); | ||||||
| 392 | } | ||||||
| 393 | } | ||||||
| 394 | |||||||
| 395 | // Use the target specific return value for comparions lib calls. | ||||||
| 396 | EVT RetVT = getCmpLibcallReturnType(); | ||||||
| 397 | SDValue Ops[2] = {NewLHS, NewRHS}; | ||||||
| 398 | TargetLowering::MakeLibCallOptions CallOptions; | ||||||
| 399 | EVT OpsVT[2] = { OldLHS.getValueType(), | ||||||
| 400 | OldRHS.getValueType() }; | ||||||
| 401 | CallOptions.setTypeListBeforeSoften(OpsVT, RetVT, true); | ||||||
| 402 | auto Call = makeLibCall(DAG, LC1, RetVT, Ops, CallOptions, dl, Chain); | ||||||
| 403 | NewLHS = Call.first; | ||||||
| 404 | NewRHS = DAG.getConstant(0, dl, RetVT); | ||||||
| 405 | |||||||
| 406 | CCCode = getCmpLibcallCC(LC1); | ||||||
| 407 | if (ShouldInvertCC) { | ||||||
| 408 | assert(RetVT.isInteger())((RetVT.isInteger()) ? static_cast<void> (0) : __assert_fail ("RetVT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 408, __PRETTY_FUNCTION__)); | ||||||
| 409 | CCCode = getSetCCInverse(CCCode, RetVT); | ||||||
| 410 | } | ||||||
| 411 | |||||||
| 412 | if (LC2 == RTLIB::UNKNOWN_LIBCALL) { | ||||||
| 413 | // Update Chain. | ||||||
| 414 | Chain = Call.second; | ||||||
| 415 | } else { | ||||||
| 416 | EVT SetCCVT = | ||||||
| 417 | getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), RetVT); | ||||||
| 418 | SDValue Tmp = DAG.getSetCC(dl, SetCCVT, NewLHS, NewRHS, CCCode); | ||||||
| 419 | auto Call2 = makeLibCall(DAG, LC2, RetVT, Ops, CallOptions, dl, Chain); | ||||||
| 420 | CCCode = getCmpLibcallCC(LC2); | ||||||
| 421 | if (ShouldInvertCC) | ||||||
| 422 | CCCode = getSetCCInverse(CCCode, RetVT); | ||||||
| 423 | NewLHS = DAG.getSetCC(dl, SetCCVT, Call2.first, NewRHS, CCCode); | ||||||
| 424 | if (Chain) | ||||||
| 425 | Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Call.second, | ||||||
| 426 | Call2.second); | ||||||
| 427 | NewLHS = DAG.getNode(ShouldInvertCC ? ISD::AND : ISD::OR, dl, | ||||||
| 428 | Tmp.getValueType(), Tmp, NewLHS); | ||||||
| 429 | NewRHS = SDValue(); | ||||||
| 430 | } | ||||||
| 431 | } | ||||||
| 432 | |||||||
| 433 | /// Return the entry encoding for a jump table in the current function. The | ||||||
| 434 | /// returned value is a member of the MachineJumpTableInfo::JTEntryKind enum. | ||||||
| 435 | unsigned TargetLowering::getJumpTableEncoding() const { | ||||||
| 436 | // In non-pic modes, just use the address of a block. | ||||||
| 437 | if (!isPositionIndependent()) | ||||||
| 438 | return MachineJumpTableInfo::EK_BlockAddress; | ||||||
| 439 | |||||||
| 440 | // In PIC mode, if the target supports a GPRel32 directive, use it. | ||||||
| 441 | if (getTargetMachine().getMCAsmInfo()->getGPRel32Directive() != nullptr) | ||||||
| 442 | return MachineJumpTableInfo::EK_GPRel32BlockAddress; | ||||||
| 443 | |||||||
| 444 | // Otherwise, use a label difference. | ||||||
| 445 | return MachineJumpTableInfo::EK_LabelDifference32; | ||||||
| 446 | } | ||||||
| 447 | |||||||
| 448 | SDValue TargetLowering::getPICJumpTableRelocBase(SDValue Table, | ||||||
| 449 | SelectionDAG &DAG) const { | ||||||
| 450 | // If our PIC model is GP relative, use the global offset table as the base. | ||||||
| 451 | unsigned JTEncoding = getJumpTableEncoding(); | ||||||
| 452 | |||||||
| 453 | if ((JTEncoding == MachineJumpTableInfo::EK_GPRel64BlockAddress) || | ||||||
| 454 | (JTEncoding == MachineJumpTableInfo::EK_GPRel32BlockAddress)) | ||||||
| 455 | return DAG.getGLOBAL_OFFSET_TABLE(getPointerTy(DAG.getDataLayout())); | ||||||
| 456 | |||||||
| 457 | return Table; | ||||||
| 458 | } | ||||||
| 459 | |||||||
| 460 | /// This returns the relocation base for the given PIC jumptable, the same as | ||||||
| 461 | /// getPICJumpTableRelocBase, but as an MCExpr. | ||||||
| 462 | const MCExpr * | ||||||
| 463 | TargetLowering::getPICJumpTableRelocBaseExpr(const MachineFunction *MF, | ||||||
| 464 | unsigned JTI,MCContext &Ctx) const{ | ||||||
| 465 | // The normal PIC reloc base is the label at the start of the jump table. | ||||||
| 466 | return MCSymbolRefExpr::create(MF->getJTISymbol(JTI, Ctx), Ctx); | ||||||
| 467 | } | ||||||
| 468 | |||||||
| 469 | bool | ||||||
| 470 | TargetLowering::isOffsetFoldingLegal(const GlobalAddressSDNode *GA) const { | ||||||
| 471 | const TargetMachine &TM = getTargetMachine(); | ||||||
| 472 | const GlobalValue *GV = GA->getGlobal(); | ||||||
| 473 | |||||||
| 474 | // If the address is not even local to this DSO we will have to load it from | ||||||
| 475 | // a got and then add the offset. | ||||||
| 476 | if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) | ||||||
| 477 | return false; | ||||||
| 478 | |||||||
| 479 | // If the code is position independent we will have to add a base register. | ||||||
| 480 | if (isPositionIndependent()) | ||||||
| 481 | return false; | ||||||
| 482 | |||||||
| 483 | // Otherwise we can do it. | ||||||
| 484 | return true; | ||||||
| 485 | } | ||||||
| 486 | |||||||
| 487 | //===----------------------------------------------------------------------===// | ||||||
| 488 | // Optimization Methods | ||||||
| 489 | //===----------------------------------------------------------------------===// | ||||||
| 490 | |||||||
| 491 | /// If the specified instruction has a constant integer operand and there are | ||||||
| 492 | /// bits set in that constant that are not demanded, then clear those bits and | ||||||
| 493 | /// return true. | ||||||
| 494 | bool TargetLowering::ShrinkDemandedConstant(SDValue Op, const APInt &Demanded, | ||||||
| 495 | TargetLoweringOpt &TLO) const { | ||||||
| 496 | SDLoc DL(Op); | ||||||
| 497 | unsigned Opcode = Op.getOpcode(); | ||||||
| 498 | |||||||
| 499 | // Do target-specific constant optimization. | ||||||
| 500 | if (targetShrinkDemandedConstant(Op, Demanded, TLO)) | ||||||
| 501 | return TLO.New.getNode(); | ||||||
| 502 | |||||||
| 503 | // FIXME: ISD::SELECT, ISD::SELECT_CC | ||||||
| 504 | switch (Opcode) { | ||||||
| 505 | default: | ||||||
| 506 | break; | ||||||
| 507 | case ISD::XOR: | ||||||
| 508 | case ISD::AND: | ||||||
| 509 | case ISD::OR: { | ||||||
| 510 | auto *Op1C = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | ||||||
| 511 | if (!Op1C) | ||||||
| 512 | return false; | ||||||
| 513 | |||||||
| 514 | // If this is a 'not' op, don't touch it because that's a canonical form. | ||||||
| 515 | const APInt &C = Op1C->getAPIntValue(); | ||||||
| 516 | if (Opcode == ISD::XOR && Demanded.isSubsetOf(C)) | ||||||
| 517 | return false; | ||||||
| 518 | |||||||
| 519 | if (!C.isSubsetOf(Demanded)) { | ||||||
| 520 | EVT VT = Op.getValueType(); | ||||||
| 521 | SDValue NewC = TLO.DAG.getConstant(Demanded & C, DL, VT); | ||||||
| 522 | SDValue NewOp = TLO.DAG.getNode(Opcode, DL, VT, Op.getOperand(0), NewC); | ||||||
| 523 | return TLO.CombineTo(Op, NewOp); | ||||||
| 524 | } | ||||||
| 525 | |||||||
| 526 | break; | ||||||
| 527 | } | ||||||
| 528 | } | ||||||
| 529 | |||||||
| 530 | return false; | ||||||
| 531 | } | ||||||
| 532 | |||||||
| 533 | /// Convert x+y to (VT)((SmallVT)x+(SmallVT)y) if the casts are free. | ||||||
| 534 | /// This uses isZExtFree and ZERO_EXTEND for the widening cast, but it could be | ||||||
| 535 | /// generalized for targets with other types of implicit widening casts. | ||||||
| 536 | bool TargetLowering::ShrinkDemandedOp(SDValue Op, unsigned BitWidth, | ||||||
| 537 | const APInt &Demanded, | ||||||
| 538 | TargetLoweringOpt &TLO) const { | ||||||
| 539 | assert(Op.getNumOperands() == 2 &&((Op.getNumOperands() == 2 && "ShrinkDemandedOp only supports binary operators!" ) ? static_cast<void> (0) : __assert_fail ("Op.getNumOperands() == 2 && \"ShrinkDemandedOp only supports binary operators!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 540, __PRETTY_FUNCTION__)) | ||||||
| 540 | "ShrinkDemandedOp only supports binary operators!")((Op.getNumOperands() == 2 && "ShrinkDemandedOp only supports binary operators!" ) ? static_cast<void> (0) : __assert_fail ("Op.getNumOperands() == 2 && \"ShrinkDemandedOp only supports binary operators!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 540, __PRETTY_FUNCTION__)); | ||||||
| 541 | assert(Op.getNode()->getNumValues() == 1 &&((Op.getNode()->getNumValues() == 1 && "ShrinkDemandedOp only supports nodes with one result!" ) ? static_cast<void> (0) : __assert_fail ("Op.getNode()->getNumValues() == 1 && \"ShrinkDemandedOp only supports nodes with one result!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 542, __PRETTY_FUNCTION__)) | ||||||
| 542 | "ShrinkDemandedOp only supports nodes with one result!")((Op.getNode()->getNumValues() == 1 && "ShrinkDemandedOp only supports nodes with one result!" ) ? static_cast<void> (0) : __assert_fail ("Op.getNode()->getNumValues() == 1 && \"ShrinkDemandedOp only supports nodes with one result!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 542, __PRETTY_FUNCTION__)); | ||||||
| 543 | |||||||
| 544 | SelectionDAG &DAG = TLO.DAG; | ||||||
| 545 | SDLoc dl(Op); | ||||||
| 546 | |||||||
| 547 | // Early return, as this function cannot handle vector types. | ||||||
| 548 | if (Op.getValueType().isVector()) | ||||||
| 549 | return false; | ||||||
| 550 | |||||||
| 551 | // Don't do this if the node has another user, which may require the | ||||||
| 552 | // full value. | ||||||
| 553 | if (!Op.getNode()->hasOneUse()) | ||||||
| 554 | return false; | ||||||
| 555 | |||||||
| 556 | // Search for the smallest integer type with free casts to and from | ||||||
| 557 | // Op's type. For expedience, just check power-of-2 integer types. | ||||||
| 558 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | ||||||
| 559 | unsigned DemandedSize = Demanded.getActiveBits(); | ||||||
| 560 | unsigned SmallVTBits = DemandedSize; | ||||||
| 561 | if (!isPowerOf2_32(SmallVTBits)) | ||||||
| 562 | SmallVTBits = NextPowerOf2(SmallVTBits); | ||||||
| 563 | for (; SmallVTBits < BitWidth; SmallVTBits = NextPowerOf2(SmallVTBits)) { | ||||||
| 564 | EVT SmallVT = EVT::getIntegerVT(*DAG.getContext(), SmallVTBits); | ||||||
| 565 | if (TLI.isTruncateFree(Op.getValueType(), SmallVT) && | ||||||
| 566 | TLI.isZExtFree(SmallVT, Op.getValueType())) { | ||||||
| 567 | // We found a type with free casts. | ||||||
| 568 | SDValue X = DAG.getNode( | ||||||
| 569 | Op.getOpcode(), dl, SmallVT, | ||||||
| 570 | DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(0)), | ||||||
| 571 | DAG.getNode(ISD::TRUNCATE, dl, SmallVT, Op.getOperand(1))); | ||||||
| 572 | assert(DemandedSize <= SmallVTBits && "Narrowed below demanded bits?")((DemandedSize <= SmallVTBits && "Narrowed below demanded bits?" ) ? static_cast<void> (0) : __assert_fail ("DemandedSize <= SmallVTBits && \"Narrowed below demanded bits?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 572, __PRETTY_FUNCTION__)); | ||||||
| 573 | SDValue Z = DAG.getNode(ISD::ANY_EXTEND, dl, Op.getValueType(), X); | ||||||
| 574 | return TLO.CombineTo(Op, Z); | ||||||
| 575 | } | ||||||
| 576 | } | ||||||
| 577 | return false; | ||||||
| 578 | } | ||||||
| 579 | |||||||
| 580 | bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, | ||||||
| 581 | DAGCombinerInfo &DCI) const { | ||||||
| 582 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 583 | TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), | ||||||
| 584 | !DCI.isBeforeLegalizeOps()); | ||||||
| 585 | KnownBits Known; | ||||||
| 586 | |||||||
| 587 | bool Simplified = SimplifyDemandedBits(Op, DemandedBits, Known, TLO); | ||||||
| 588 | if (Simplified) { | ||||||
| 589 | DCI.AddToWorklist(Op.getNode()); | ||||||
| 590 | DCI.CommitTargetLoweringOpt(TLO); | ||||||
| 591 | } | ||||||
| 592 | return Simplified; | ||||||
| 593 | } | ||||||
| 594 | |||||||
| 595 | bool TargetLowering::SimplifyDemandedBits(SDValue Op, const APInt &DemandedBits, | ||||||
| 596 | KnownBits &Known, | ||||||
| 597 | TargetLoweringOpt &TLO, | ||||||
| 598 | unsigned Depth, | ||||||
| 599 | bool AssumeSingleUse) const { | ||||||
| 600 | EVT VT = Op.getValueType(); | ||||||
| 601 | APInt DemandedElts = VT.isVector() | ||||||
| 602 | ? APInt::getAllOnesValue(VT.getVectorNumElements()) | ||||||
| 603 | : APInt(1, 1); | ||||||
| 604 | return SimplifyDemandedBits(Op, DemandedBits, DemandedElts, Known, TLO, Depth, | ||||||
| 605 | AssumeSingleUse); | ||||||
| 606 | } | ||||||
| 607 | |||||||
| 608 | // TODO: Can we merge SelectionDAG::GetDemandedBits into this? | ||||||
| 609 | // TODO: Under what circumstances can we create nodes? Constant folding? | ||||||
| 610 | SDValue TargetLowering::SimplifyMultipleUseDemandedBits( | ||||||
| 611 | SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, | ||||||
| 612 | SelectionDAG &DAG, unsigned Depth) const { | ||||||
| 613 | // Limit search depth. | ||||||
| 614 | if (Depth >= SelectionDAG::MaxRecursionDepth) | ||||||
| 615 | return SDValue(); | ||||||
| 616 | |||||||
| 617 | // Ignore UNDEFs. | ||||||
| 618 | if (Op.isUndef()) | ||||||
| 619 | return SDValue(); | ||||||
| 620 | |||||||
| 621 | // Not demanding any bits/elts from Op. | ||||||
| 622 | if (DemandedBits == 0 || DemandedElts == 0) | ||||||
| 623 | return DAG.getUNDEF(Op.getValueType()); | ||||||
| 624 | |||||||
| 625 | unsigned NumElts = DemandedElts.getBitWidth(); | ||||||
| 626 | KnownBits LHSKnown, RHSKnown; | ||||||
| 627 | switch (Op.getOpcode()) { | ||||||
| 628 | case ISD::BITCAST: { | ||||||
| 629 | SDValue Src = peekThroughBitcasts(Op.getOperand(0)); | ||||||
| 630 | EVT SrcVT = Src.getValueType(); | ||||||
| 631 | EVT DstVT = Op.getValueType(); | ||||||
| 632 | unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits(); | ||||||
| 633 | unsigned NumDstEltBits = DstVT.getScalarSizeInBits(); | ||||||
| 634 | |||||||
| 635 | if (NumSrcEltBits == NumDstEltBits) | ||||||
| 636 | if (SDValue V = SimplifyMultipleUseDemandedBits( | ||||||
| 637 | Src, DemandedBits, DemandedElts, DAG, Depth + 1)) | ||||||
| 638 | return DAG.getBitcast(DstVT, V); | ||||||
| 639 | |||||||
| 640 | // TODO - bigendian once we have test coverage. | ||||||
| 641 | if (SrcVT.isVector() && (NumDstEltBits % NumSrcEltBits) == 0 && | ||||||
| 642 | DAG.getDataLayout().isLittleEndian()) { | ||||||
| 643 | unsigned Scale = NumDstEltBits / NumSrcEltBits; | ||||||
| 644 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | ||||||
| 645 | APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits); | ||||||
| 646 | APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts); | ||||||
| 647 | for (unsigned i = 0; i != Scale; ++i) { | ||||||
| 648 | unsigned Offset = i * NumSrcEltBits; | ||||||
| 649 | APInt Sub = DemandedBits.extractBits(NumSrcEltBits, Offset); | ||||||
| 650 | if (!Sub.isNullValue()) { | ||||||
| 651 | DemandedSrcBits |= Sub; | ||||||
| 652 | for (unsigned j = 0; j != NumElts; ++j) | ||||||
| 653 | if (DemandedElts[j]) | ||||||
| 654 | DemandedSrcElts.setBit((j * Scale) + i); | ||||||
| 655 | } | ||||||
| 656 | } | ||||||
| 657 | |||||||
| 658 | if (SDValue V = SimplifyMultipleUseDemandedBits( | ||||||
| 659 | Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1)) | ||||||
| 660 | return DAG.getBitcast(DstVT, V); | ||||||
| 661 | } | ||||||
| 662 | |||||||
| 663 | // TODO - bigendian once we have test coverage. | ||||||
| 664 | if ((NumSrcEltBits % NumDstEltBits) == 0 && | ||||||
| 665 | DAG.getDataLayout().isLittleEndian()) { | ||||||
| 666 | unsigned Scale = NumSrcEltBits / NumDstEltBits; | ||||||
| 667 | unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; | ||||||
| 668 | APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits); | ||||||
| 669 | APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts); | ||||||
| 670 | for (unsigned i = 0; i != NumElts; ++i) | ||||||
| 671 | if (DemandedElts[i]) { | ||||||
| 672 | unsigned Offset = (i % Scale) * NumDstEltBits; | ||||||
| 673 | DemandedSrcBits.insertBits(DemandedBits, Offset); | ||||||
| 674 | DemandedSrcElts.setBit(i / Scale); | ||||||
| 675 | } | ||||||
| 676 | |||||||
| 677 | if (SDValue V = SimplifyMultipleUseDemandedBits( | ||||||
| 678 | Src, DemandedSrcBits, DemandedSrcElts, DAG, Depth + 1)) | ||||||
| 679 | return DAG.getBitcast(DstVT, V); | ||||||
| 680 | } | ||||||
| 681 | |||||||
| 682 | break; | ||||||
| 683 | } | ||||||
| 684 | case ISD::AND: { | ||||||
| 685 | LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); | ||||||
| 686 | RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); | ||||||
| 687 | |||||||
| 688 | // If all of the demanded bits are known 1 on one side, return the other. | ||||||
| 689 | // These bits cannot contribute to the result of the 'and' in this | ||||||
| 690 | // context. | ||||||
| 691 | if (DemandedBits.isSubsetOf(LHSKnown.Zero | RHSKnown.One)) | ||||||
| 692 | return Op.getOperand(0); | ||||||
| 693 | if (DemandedBits.isSubsetOf(RHSKnown.Zero | LHSKnown.One)) | ||||||
| 694 | return Op.getOperand(1); | ||||||
| 695 | break; | ||||||
| 696 | } | ||||||
| 697 | case ISD::OR: { | ||||||
| 698 | LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); | ||||||
| 699 | RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); | ||||||
| 700 | |||||||
| 701 | // If all of the demanded bits are known zero on one side, return the | ||||||
| 702 | // other. These bits cannot contribute to the result of the 'or' in this | ||||||
| 703 | // context. | ||||||
| 704 | if (DemandedBits.isSubsetOf(LHSKnown.One | RHSKnown.Zero)) | ||||||
| 705 | return Op.getOperand(0); | ||||||
| 706 | if (DemandedBits.isSubsetOf(RHSKnown.One | LHSKnown.Zero)) | ||||||
| 707 | return Op.getOperand(1); | ||||||
| 708 | break; | ||||||
| 709 | } | ||||||
| 710 | case ISD::XOR: { | ||||||
| 711 | LHSKnown = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1); | ||||||
| 712 | RHSKnown = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1); | ||||||
| 713 | |||||||
| 714 | // If all of the demanded bits are known zero on one side, return the | ||||||
| 715 | // other. | ||||||
| 716 | if (DemandedBits.isSubsetOf(RHSKnown.Zero)) | ||||||
| 717 | return Op.getOperand(0); | ||||||
| 718 | if (DemandedBits.isSubsetOf(LHSKnown.Zero)) | ||||||
| 719 | return Op.getOperand(1); | ||||||
| 720 | break; | ||||||
| 721 | } | ||||||
| 722 | case ISD::SETCC: { | ||||||
| 723 | SDValue Op0 = Op.getOperand(0); | ||||||
| 724 | SDValue Op1 = Op.getOperand(1); | ||||||
| 725 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); | ||||||
| 726 | // If (1) we only need the sign-bit, (2) the setcc operands are the same | ||||||
| 727 | // width as the setcc result, and (3) the result of a setcc conforms to 0 or | ||||||
| 728 | // -1, we may be able to bypass the setcc. | ||||||
| 729 | if (DemandedBits.isSignMask() && | ||||||
| 730 | Op0.getScalarValueSizeInBits() == DemandedBits.getBitWidth() && | ||||||
| 731 | getBooleanContents(Op0.getValueType()) == | ||||||
| 732 | BooleanContent::ZeroOrNegativeOneBooleanContent) { | ||||||
| 733 | // If we're testing X < 0, then this compare isn't needed - just use X! | ||||||
| 734 | // FIXME: We're limiting to integer types here, but this should also work | ||||||
| 735 | // if we don't care about FP signed-zero. The use of SETLT with FP means | ||||||
| 736 | // that we don't care about NaNs. | ||||||
| 737 | if (CC == ISD::SETLT && Op1.getValueType().isInteger() && | ||||||
| 738 | (isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode()))) | ||||||
| 739 | return Op0; | ||||||
| 740 | } | ||||||
| 741 | break; | ||||||
| 742 | } | ||||||
| 743 | case ISD::SIGN_EXTEND_INREG: { | ||||||
| 744 | // If none of the extended bits are demanded, eliminate the sextinreg. | ||||||
| 745 | EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); | ||||||
| 746 | if (DemandedBits.getActiveBits() <= ExVT.getScalarSizeInBits()) | ||||||
| 747 | return Op.getOperand(0); | ||||||
| 748 | break; | ||||||
| 749 | } | ||||||
| 750 | case ISD::INSERT_VECTOR_ELT: { | ||||||
| 751 | // If we don't demand the inserted element, return the base vector. | ||||||
| 752 | SDValue Vec = Op.getOperand(0); | ||||||
| 753 | auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | ||||||
| 754 | EVT VecVT = Vec.getValueType(); | ||||||
| 755 | if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements()) && | ||||||
| 756 | !DemandedElts[CIdx->getZExtValue()]) | ||||||
| 757 | return Vec; | ||||||
| 758 | break; | ||||||
| 759 | } | ||||||
| 760 | case ISD::VECTOR_SHUFFLE: { | ||||||
| 761 | ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask(); | ||||||
| 762 | |||||||
| 763 | // If all the demanded elts are from one operand and are inline, | ||||||
| 764 | // then we can use the operand directly. | ||||||
| 765 | bool AllUndef = true, IdentityLHS = true, IdentityRHS = true; | ||||||
| 766 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 767 | int M = ShuffleMask[i]; | ||||||
| 768 | if (M < 0 || !DemandedElts[i]) | ||||||
| 769 | continue; | ||||||
| 770 | AllUndef = false; | ||||||
| 771 | IdentityLHS &= (M == (int)i); | ||||||
| 772 | IdentityRHS &= ((M - NumElts) == i); | ||||||
| 773 | } | ||||||
| 774 | |||||||
| 775 | if (AllUndef) | ||||||
| 776 | return DAG.getUNDEF(Op.getValueType()); | ||||||
| 777 | if (IdentityLHS) | ||||||
| 778 | return Op.getOperand(0); | ||||||
| 779 | if (IdentityRHS) | ||||||
| 780 | return Op.getOperand(1); | ||||||
| 781 | break; | ||||||
| 782 | } | ||||||
| 783 | default: | ||||||
| 784 | if (Op.getOpcode() >= ISD::BUILTIN_OP_END) | ||||||
| 785 | if (SDValue V = SimplifyMultipleUseDemandedBitsForTargetNode( | ||||||
| 786 | Op, DemandedBits, DemandedElts, DAG, Depth)) | ||||||
| 787 | return V; | ||||||
| 788 | break; | ||||||
| 789 | } | ||||||
| 790 | return SDValue(); | ||||||
| 791 | } | ||||||
| 792 | |||||||
| 793 | /// Look at Op. At this point, we know that only the OriginalDemandedBits of the | ||||||
| 794 | /// result of Op are ever used downstream. If we can use this information to | ||||||
| 795 | /// simplify Op, create a new simplified DAG node and return true, returning the | ||||||
| 796 | /// original and new nodes in Old and New. Otherwise, analyze the expression and | ||||||
| 797 | /// return a mask of Known bits for the expression (used to simplify the | ||||||
| 798 | /// caller). The Known bits may only be accurate for those bits in the | ||||||
| 799 | /// OriginalDemandedBits and OriginalDemandedElts. | ||||||
| 800 | bool TargetLowering::SimplifyDemandedBits( | ||||||
| 801 | SDValue Op, const APInt &OriginalDemandedBits, | ||||||
| 802 | const APInt &OriginalDemandedElts, KnownBits &Known, TargetLoweringOpt &TLO, | ||||||
| 803 | unsigned Depth, bool AssumeSingleUse) const { | ||||||
| 804 | unsigned BitWidth = OriginalDemandedBits.getBitWidth(); | ||||||
| 805 | assert(Op.getScalarValueSizeInBits() == BitWidth &&((Op.getScalarValueSizeInBits() == BitWidth && "Mask size mismatches value type size!" ) ? static_cast<void> (0) : __assert_fail ("Op.getScalarValueSizeInBits() == BitWidth && \"Mask size mismatches value type size!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 806, __PRETTY_FUNCTION__)) | ||||||
| 806 | "Mask size mismatches value type size!")((Op.getScalarValueSizeInBits() == BitWidth && "Mask size mismatches value type size!" ) ? static_cast<void> (0) : __assert_fail ("Op.getScalarValueSizeInBits() == BitWidth && \"Mask size mismatches value type size!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 806, __PRETTY_FUNCTION__)); | ||||||
| 807 | |||||||
| 808 | unsigned NumElts = OriginalDemandedElts.getBitWidth(); | ||||||
| 809 | assert((!Op.getValueType().isVector() ||(((!Op.getValueType().isVector() || NumElts == Op.getValueType ().getVectorNumElements()) && "Unexpected vector size" ) ? static_cast<void> (0) : __assert_fail ("(!Op.getValueType().isVector() || NumElts == Op.getValueType().getVectorNumElements()) && \"Unexpected vector size\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 811, __PRETTY_FUNCTION__)) | ||||||
| 810 | NumElts == Op.getValueType().getVectorNumElements()) &&(((!Op.getValueType().isVector() || NumElts == Op.getValueType ().getVectorNumElements()) && "Unexpected vector size" ) ? static_cast<void> (0) : __assert_fail ("(!Op.getValueType().isVector() || NumElts == Op.getValueType().getVectorNumElements()) && \"Unexpected vector size\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 811, __PRETTY_FUNCTION__)) | ||||||
| 811 | "Unexpected vector size")(((!Op.getValueType().isVector() || NumElts == Op.getValueType ().getVectorNumElements()) && "Unexpected vector size" ) ? static_cast<void> (0) : __assert_fail ("(!Op.getValueType().isVector() || NumElts == Op.getValueType().getVectorNumElements()) && \"Unexpected vector size\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 811, __PRETTY_FUNCTION__)); | ||||||
| 812 | |||||||
| 813 | APInt DemandedBits = OriginalDemandedBits; | ||||||
| 814 | APInt DemandedElts = OriginalDemandedElts; | ||||||
| 815 | SDLoc dl(Op); | ||||||
| 816 | auto &DL = TLO.DAG.getDataLayout(); | ||||||
| 817 | |||||||
| 818 | // Don't know anything. | ||||||
| 819 | Known = KnownBits(BitWidth); | ||||||
| 820 | |||||||
| 821 | // Undef operand. | ||||||
| 822 | if (Op.isUndef()) | ||||||
| 823 | return false; | ||||||
| 824 | |||||||
| 825 | if (Op.getOpcode() == ISD::Constant) { | ||||||
| 826 | // We know all of the bits for a constant! | ||||||
| 827 | Known.One = cast<ConstantSDNode>(Op)->getAPIntValue(); | ||||||
| 828 | Known.Zero = ~Known.One; | ||||||
| 829 | return false; | ||||||
| 830 | } | ||||||
| 831 | |||||||
| 832 | // Other users may use these bits. | ||||||
| 833 | EVT VT = Op.getValueType(); | ||||||
| 834 | if (!Op.getNode()->hasOneUse() && !AssumeSingleUse) { | ||||||
| 835 | if (Depth != 0) { | ||||||
| 836 | // If not at the root, Just compute the Known bits to | ||||||
| 837 | // simplify things downstream. | ||||||
| 838 | Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); | ||||||
| 839 | return false; | ||||||
| 840 | } | ||||||
| 841 | // If this is the root being simplified, allow it to have multiple uses, | ||||||
| 842 | // just set the DemandedBits/Elts to all bits. | ||||||
| 843 | DemandedBits = APInt::getAllOnesValue(BitWidth); | ||||||
| 844 | DemandedElts = APInt::getAllOnesValue(NumElts); | ||||||
| 845 | } else if (OriginalDemandedBits == 0 || OriginalDemandedElts == 0) { | ||||||
| 846 | // Not demanding any bits/elts from Op. | ||||||
| 847 | return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); | ||||||
| 848 | } else if (Depth >= SelectionDAG::MaxRecursionDepth) { | ||||||
| 849 | // Limit search depth. | ||||||
| 850 | return false; | ||||||
| 851 | } | ||||||
| 852 | |||||||
| 853 | KnownBits Known2, KnownOut; | ||||||
| 854 | switch (Op.getOpcode()) { | ||||||
| 855 | case ISD::TargetConstant: | ||||||
| 856 | llvm_unreachable("Can't simplify this node")::llvm::llvm_unreachable_internal("Can't simplify this node", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 856); | ||||||
| 857 | case ISD::SCALAR_TO_VECTOR: { | ||||||
| 858 | if (!DemandedElts[0]) | ||||||
| 859 | return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); | ||||||
| 860 | |||||||
| 861 | KnownBits SrcKnown; | ||||||
| 862 | SDValue Src = Op.getOperand(0); | ||||||
| 863 | unsigned SrcBitWidth = Src.getScalarValueSizeInBits(); | ||||||
| 864 | APInt SrcDemandedBits = DemandedBits.zextOrSelf(SrcBitWidth); | ||||||
| 865 | if (SimplifyDemandedBits(Src, SrcDemandedBits, SrcKnown, TLO, Depth + 1)) | ||||||
| 866 | return true; | ||||||
| 867 | Known = SrcKnown.zextOrTrunc(BitWidth, false); | ||||||
| 868 | break; | ||||||
| 869 | } | ||||||
| 870 | case ISD::BUILD_VECTOR: | ||||||
| 871 | // Collect the known bits that are shared by every demanded element. | ||||||
| 872 | // TODO: Call SimplifyDemandedBits for non-constant demanded elements. | ||||||
| 873 | Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); | ||||||
| 874 | return false; // Don't fall through, will infinitely loop. | ||||||
| 875 | case ISD::LOAD: { | ||||||
| 876 | LoadSDNode *LD = cast<LoadSDNode>(Op); | ||||||
| 877 | if (getTargetConstantFromLoad(LD)) { | ||||||
| 878 | Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); | ||||||
| 879 | return false; // Don't fall through, will infinitely loop. | ||||||
| 880 | } | ||||||
| 881 | break; | ||||||
| 882 | } | ||||||
| 883 | case ISD::INSERT_VECTOR_ELT: { | ||||||
| 884 | SDValue Vec = Op.getOperand(0); | ||||||
| 885 | SDValue Scl = Op.getOperand(1); | ||||||
| 886 | auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | ||||||
| 887 | EVT VecVT = Vec.getValueType(); | ||||||
| 888 | |||||||
| 889 | // If index isn't constant, assume we need all vector elements AND the | ||||||
| 890 | // inserted element. | ||||||
| 891 | APInt DemandedVecElts(DemandedElts); | ||||||
| 892 | if (CIdx && CIdx->getAPIntValue().ult(VecVT.getVectorNumElements())) { | ||||||
| 893 | unsigned Idx = CIdx->getZExtValue(); | ||||||
| 894 | DemandedVecElts.clearBit(Idx); | ||||||
| 895 | |||||||
| 896 | // Inserted element is not required. | ||||||
| 897 | if (!DemandedElts[Idx]) | ||||||
| 898 | return TLO.CombineTo(Op, Vec); | ||||||
| 899 | } | ||||||
| 900 | |||||||
| 901 | KnownBits KnownScl; | ||||||
| 902 | unsigned NumSclBits = Scl.getScalarValueSizeInBits(); | ||||||
| 903 | APInt DemandedSclBits = DemandedBits.zextOrTrunc(NumSclBits); | ||||||
| 904 | if (SimplifyDemandedBits(Scl, DemandedSclBits, KnownScl, TLO, Depth + 1)) | ||||||
| 905 | return true; | ||||||
| 906 | |||||||
| 907 | Known = KnownScl.zextOrTrunc(BitWidth, false); | ||||||
| 908 | |||||||
| 909 | KnownBits KnownVec; | ||||||
| 910 | if (SimplifyDemandedBits(Vec, DemandedBits, DemandedVecElts, KnownVec, TLO, | ||||||
| 911 | Depth + 1)) | ||||||
| 912 | return true; | ||||||
| 913 | |||||||
| 914 | if (!!DemandedVecElts) { | ||||||
| 915 | Known.One &= KnownVec.One; | ||||||
| 916 | Known.Zero &= KnownVec.Zero; | ||||||
| 917 | } | ||||||
| 918 | |||||||
| 919 | return false; | ||||||
| 920 | } | ||||||
| 921 | case ISD::INSERT_SUBVECTOR: { | ||||||
| 922 | SDValue Base = Op.getOperand(0); | ||||||
| 923 | SDValue Sub = Op.getOperand(1); | ||||||
| 924 | EVT SubVT = Sub.getValueType(); | ||||||
| 925 | unsigned NumSubElts = SubVT.getVectorNumElements(); | ||||||
| 926 | |||||||
| 927 | // If index isn't constant, assume we need the original demanded base | ||||||
| 928 | // elements and ALL the inserted subvector elements. | ||||||
| 929 | APInt BaseElts = DemandedElts; | ||||||
| 930 | APInt SubElts = APInt::getAllOnesValue(NumSubElts); | ||||||
| 931 | if (isa<ConstantSDNode>(Op.getOperand(2))) { | ||||||
| 932 | const APInt &Idx = Op.getConstantOperandAPInt(2); | ||||||
| 933 | if (Idx.ule(NumElts - NumSubElts)) { | ||||||
| 934 | unsigned SubIdx = Idx.getZExtValue(); | ||||||
| 935 | SubElts = DemandedElts.extractBits(NumSubElts, SubIdx); | ||||||
| 936 | BaseElts.insertBits(APInt::getNullValue(NumSubElts), SubIdx); | ||||||
| 937 | } | ||||||
| 938 | } | ||||||
| 939 | |||||||
| 940 | KnownBits KnownSub, KnownBase; | ||||||
| 941 | if (SimplifyDemandedBits(Sub, DemandedBits, SubElts, KnownSub, TLO, | ||||||
| 942 | Depth + 1)) | ||||||
| 943 | return true; | ||||||
| 944 | if (SimplifyDemandedBits(Base, DemandedBits, BaseElts, KnownBase, TLO, | ||||||
| 945 | Depth + 1)) | ||||||
| 946 | return true; | ||||||
| 947 | |||||||
| 948 | Known.Zero.setAllBits(); | ||||||
| 949 | Known.One.setAllBits(); | ||||||
| 950 | if (!!SubElts) { | ||||||
| 951 | Known.One &= KnownSub.One; | ||||||
| 952 | Known.Zero &= KnownSub.Zero; | ||||||
| 953 | } | ||||||
| 954 | if (!!BaseElts) { | ||||||
| 955 | Known.One &= KnownBase.One; | ||||||
| 956 | Known.Zero &= KnownBase.Zero; | ||||||
| 957 | } | ||||||
| 958 | break; | ||||||
| 959 | } | ||||||
| 960 | case ISD::EXTRACT_SUBVECTOR: { | ||||||
| 961 | // If index isn't constant, assume we need all the source vector elements. | ||||||
| 962 | SDValue Src = Op.getOperand(0); | ||||||
| 963 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | ||||||
| 964 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); | ||||||
| 965 | APInt SrcElts = APInt::getAllOnesValue(NumSrcElts); | ||||||
| 966 | if (SubIdx && SubIdx->getAPIntValue().ule(NumSrcElts - NumElts)) { | ||||||
| 967 | // Offset the demanded elts by the subvector index. | ||||||
| 968 | uint64_t Idx = SubIdx->getZExtValue(); | ||||||
| 969 | SrcElts = DemandedElts.zextOrSelf(NumSrcElts).shl(Idx); | ||||||
| 970 | } | ||||||
| 971 | if (SimplifyDemandedBits(Src, DemandedBits, SrcElts, Known, TLO, Depth + 1)) | ||||||
| 972 | return true; | ||||||
| 973 | break; | ||||||
| 974 | } | ||||||
| 975 | case ISD::CONCAT_VECTORS: { | ||||||
| 976 | Known.Zero.setAllBits(); | ||||||
| 977 | Known.One.setAllBits(); | ||||||
| 978 | EVT SubVT = Op.getOperand(0).getValueType(); | ||||||
| 979 | unsigned NumSubVecs = Op.getNumOperands(); | ||||||
| 980 | unsigned NumSubElts = SubVT.getVectorNumElements(); | ||||||
| 981 | for (unsigned i = 0; i != NumSubVecs; ++i) { | ||||||
| 982 | APInt DemandedSubElts = | ||||||
| 983 | DemandedElts.extractBits(NumSubElts, i * NumSubElts); | ||||||
| 984 | if (SimplifyDemandedBits(Op.getOperand(i), DemandedBits, DemandedSubElts, | ||||||
| 985 | Known2, TLO, Depth + 1)) | ||||||
| 986 | return true; | ||||||
| 987 | // Known bits are shared by every demanded subvector element. | ||||||
| 988 | if (!!DemandedSubElts) { | ||||||
| 989 | Known.One &= Known2.One; | ||||||
| 990 | Known.Zero &= Known2.Zero; | ||||||
| 991 | } | ||||||
| 992 | } | ||||||
| 993 | break; | ||||||
| 994 | } | ||||||
| 995 | case ISD::VECTOR_SHUFFLE: { | ||||||
| 996 | ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask(); | ||||||
| 997 | |||||||
| 998 | // Collect demanded elements from shuffle operands.. | ||||||
| 999 | APInt DemandedLHS(NumElts, 0); | ||||||
| 1000 | APInt DemandedRHS(NumElts, 0); | ||||||
| 1001 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 1002 | if (!DemandedElts[i]) | ||||||
| 1003 | continue; | ||||||
| 1004 | int M = ShuffleMask[i]; | ||||||
| 1005 | if (M < 0) { | ||||||
| 1006 | // For UNDEF elements, we don't know anything about the common state of | ||||||
| 1007 | // the shuffle result. | ||||||
| 1008 | DemandedLHS.clearAllBits(); | ||||||
| 1009 | DemandedRHS.clearAllBits(); | ||||||
| 1010 | break; | ||||||
| 1011 | } | ||||||
| 1012 | assert(0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range")((0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range" ) ? static_cast<void> (0) : __assert_fail ("0 <= M && M < (int)(2 * NumElts) && \"Shuffle index out of range\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1012, __PRETTY_FUNCTION__)); | ||||||
| 1013 | if (M < (int)NumElts) | ||||||
| 1014 | DemandedLHS.setBit(M); | ||||||
| 1015 | else | ||||||
| 1016 | DemandedRHS.setBit(M - NumElts); | ||||||
| 1017 | } | ||||||
| 1018 | |||||||
| 1019 | if (!!DemandedLHS || !!DemandedRHS) { | ||||||
| 1020 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1021 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1022 | |||||||
| 1023 | Known.Zero.setAllBits(); | ||||||
| 1024 | Known.One.setAllBits(); | ||||||
| 1025 | if (!!DemandedLHS) { | ||||||
| 1026 | if (SimplifyDemandedBits(Op0, DemandedBits, DemandedLHS, Known2, TLO, | ||||||
| 1027 | Depth + 1)) | ||||||
| 1028 | return true; | ||||||
| 1029 | Known.One &= Known2.One; | ||||||
| 1030 | Known.Zero &= Known2.Zero; | ||||||
| 1031 | } | ||||||
| 1032 | if (!!DemandedRHS) { | ||||||
| 1033 | if (SimplifyDemandedBits(Op1, DemandedBits, DemandedRHS, Known2, TLO, | ||||||
| 1034 | Depth + 1)) | ||||||
| 1035 | return true; | ||||||
| 1036 | Known.One &= Known2.One; | ||||||
| 1037 | Known.Zero &= Known2.Zero; | ||||||
| 1038 | } | ||||||
| 1039 | |||||||
| 1040 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1041 | SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( | ||||||
| 1042 | Op0, DemandedBits, DemandedLHS, TLO.DAG, Depth + 1); | ||||||
| 1043 | SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( | ||||||
| 1044 | Op1, DemandedBits, DemandedRHS, TLO.DAG, Depth + 1); | ||||||
| 1045 | if (DemandedOp0 || DemandedOp1) { | ||||||
| 1046 | Op0 = DemandedOp0 ? DemandedOp0 : Op0; | ||||||
| 1047 | Op1 = DemandedOp1 ? DemandedOp1 : Op1; | ||||||
| 1048 | SDValue NewOp = TLO.DAG.getVectorShuffle(VT, dl, Op0, Op1, ShuffleMask); | ||||||
| 1049 | return TLO.CombineTo(Op, NewOp); | ||||||
| 1050 | } | ||||||
| 1051 | } | ||||||
| 1052 | break; | ||||||
| 1053 | } | ||||||
| 1054 | case ISD::AND: { | ||||||
| 1055 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1056 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1057 | |||||||
| 1058 | // If the RHS is a constant, check to see if the LHS would be zero without | ||||||
| 1059 | // using the bits from the RHS. Below, we use knowledge about the RHS to | ||||||
| 1060 | // simplify the LHS, here we're using information from the LHS to simplify | ||||||
| 1061 | // the RHS. | ||||||
| 1062 | if (ConstantSDNode *RHSC = isConstOrConstSplat(Op1)) { | ||||||
| 1063 | // Do not increment Depth here; that can cause an infinite loop. | ||||||
| 1064 | KnownBits LHSKnown = TLO.DAG.computeKnownBits(Op0, DemandedElts, Depth); | ||||||
| 1065 | // If the LHS already has zeros where RHSC does, this 'and' is dead. | ||||||
| 1066 | if ((LHSKnown.Zero & DemandedBits) == | ||||||
| 1067 | (~RHSC->getAPIntValue() & DemandedBits)) | ||||||
| 1068 | return TLO.CombineTo(Op, Op0); | ||||||
| 1069 | |||||||
| 1070 | // If any of the set bits in the RHS are known zero on the LHS, shrink | ||||||
| 1071 | // the constant. | ||||||
| 1072 | if (ShrinkDemandedConstant(Op, ~LHSKnown.Zero & DemandedBits, TLO)) | ||||||
| 1073 | return true; | ||||||
| 1074 | |||||||
| 1075 | // Bitwise-not (xor X, -1) is a special case: we don't usually shrink its | ||||||
| 1076 | // constant, but if this 'and' is only clearing bits that were just set by | ||||||
| 1077 | // the xor, then this 'and' can be eliminated by shrinking the mask of | ||||||
| 1078 | // the xor. For example, for a 32-bit X: | ||||||
| 1079 | // and (xor (srl X, 31), -1), 1 --> xor (srl X, 31), 1 | ||||||
| 1080 | if (isBitwiseNot(Op0) && Op0.hasOneUse() && | ||||||
| 1081 | LHSKnown.One == ~RHSC->getAPIntValue()) { | ||||||
| 1082 | SDValue Xor = TLO.DAG.getNode(ISD::XOR, dl, VT, Op0.getOperand(0), Op1); | ||||||
| 1083 | return TLO.CombineTo(Op, Xor); | ||||||
| 1084 | } | ||||||
| 1085 | } | ||||||
| 1086 | |||||||
| 1087 | if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, | ||||||
| 1088 | Depth + 1)) | ||||||
| 1089 | return true; | ||||||
| 1090 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1090, __PRETTY_FUNCTION__)); | ||||||
| 1091 | if (SimplifyDemandedBits(Op0, ~Known.Zero & DemandedBits, DemandedElts, | ||||||
| 1092 | Known2, TLO, Depth + 1)) | ||||||
| 1093 | return true; | ||||||
| 1094 | assert(!Known2.hasConflict() && "Bits known to be one AND zero?")((!Known2.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known2.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1094, __PRETTY_FUNCTION__)); | ||||||
| 1095 | |||||||
| 1096 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1097 | if (!DemandedBits.isAllOnesValue() || !DemandedElts.isAllOnesValue()) { | ||||||
| 1098 | SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( | ||||||
| 1099 | Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1100 | SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( | ||||||
| 1101 | Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1102 | if (DemandedOp0 || DemandedOp1) { | ||||||
| 1103 | Op0 = DemandedOp0 ? DemandedOp0 : Op0; | ||||||
| 1104 | Op1 = DemandedOp1 ? DemandedOp1 : Op1; | ||||||
| 1105 | SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); | ||||||
| 1106 | return TLO.CombineTo(Op, NewOp); | ||||||
| 1107 | } | ||||||
| 1108 | } | ||||||
| 1109 | |||||||
| 1110 | // If all of the demanded bits are known one on one side, return the other. | ||||||
| 1111 | // These bits cannot contribute to the result of the 'and'. | ||||||
| 1112 | if (DemandedBits.isSubsetOf(Known2.Zero | Known.One)) | ||||||
| 1113 | return TLO.CombineTo(Op, Op0); | ||||||
| 1114 | if (DemandedBits.isSubsetOf(Known.Zero | Known2.One)) | ||||||
| 1115 | return TLO.CombineTo(Op, Op1); | ||||||
| 1116 | // If all of the demanded bits in the inputs are known zeros, return zero. | ||||||
| 1117 | if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero)) | ||||||
| 1118 | return TLO.CombineTo(Op, TLO.DAG.getConstant(0, dl, VT)); | ||||||
| 1119 | // If the RHS is a constant, see if we can simplify it. | ||||||
| 1120 | if (ShrinkDemandedConstant(Op, ~Known2.Zero & DemandedBits, TLO)) | ||||||
| 1121 | return true; | ||||||
| 1122 | // If the operation can be done in a smaller type, do so. | ||||||
| 1123 | if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) | ||||||
| 1124 | return true; | ||||||
| 1125 | |||||||
| 1126 | // Output known-1 bits are only known if set in both the LHS & RHS. | ||||||
| 1127 | Known.One &= Known2.One; | ||||||
| 1128 | // Output known-0 are known to be clear if zero in either the LHS | RHS. | ||||||
| 1129 | Known.Zero |= Known2.Zero; | ||||||
| 1130 | break; | ||||||
| 1131 | } | ||||||
| 1132 | case ISD::OR: { | ||||||
| 1133 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1134 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1135 | |||||||
| 1136 | if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, | ||||||
| 1137 | Depth + 1)) | ||||||
| 1138 | return true; | ||||||
| 1139 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1139, __PRETTY_FUNCTION__)); | ||||||
| 1140 | if (SimplifyDemandedBits(Op0, ~Known.One & DemandedBits, DemandedElts, | ||||||
| 1141 | Known2, TLO, Depth + 1)) | ||||||
| 1142 | return true; | ||||||
| 1143 | assert(!Known2.hasConflict() && "Bits known to be one AND zero?")((!Known2.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known2.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1143, __PRETTY_FUNCTION__)); | ||||||
| 1144 | |||||||
| 1145 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1146 | if (!DemandedBits.isAllOnesValue() || !DemandedElts.isAllOnesValue()) { | ||||||
| 1147 | SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( | ||||||
| 1148 | Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1149 | SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( | ||||||
| 1150 | Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1151 | if (DemandedOp0 || DemandedOp1) { | ||||||
| 1152 | Op0 = DemandedOp0 ? DemandedOp0 : Op0; | ||||||
| 1153 | Op1 = DemandedOp1 ? DemandedOp1 : Op1; | ||||||
| 1154 | SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); | ||||||
| 1155 | return TLO.CombineTo(Op, NewOp); | ||||||
| 1156 | } | ||||||
| 1157 | } | ||||||
| 1158 | |||||||
| 1159 | // If all of the demanded bits are known zero on one side, return the other. | ||||||
| 1160 | // These bits cannot contribute to the result of the 'or'. | ||||||
| 1161 | if (DemandedBits.isSubsetOf(Known2.One | Known.Zero)) | ||||||
| 1162 | return TLO.CombineTo(Op, Op0); | ||||||
| 1163 | if (DemandedBits.isSubsetOf(Known.One | Known2.Zero)) | ||||||
| 1164 | return TLO.CombineTo(Op, Op1); | ||||||
| 1165 | // If the RHS is a constant, see if we can simplify it. | ||||||
| 1166 | if (ShrinkDemandedConstant(Op, DemandedBits, TLO)) | ||||||
| 1167 | return true; | ||||||
| 1168 | // If the operation can be done in a smaller type, do so. | ||||||
| 1169 | if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) | ||||||
| 1170 | return true; | ||||||
| 1171 | |||||||
| 1172 | // Output known-0 bits are only known if clear in both the LHS & RHS. | ||||||
| 1173 | Known.Zero &= Known2.Zero; | ||||||
| 1174 | // Output known-1 are known to be set if set in either the LHS | RHS. | ||||||
| 1175 | Known.One |= Known2.One; | ||||||
| 1176 | break; | ||||||
| 1177 | } | ||||||
| 1178 | case ISD::XOR: { | ||||||
| 1179 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1180 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1181 | |||||||
| 1182 | if (SimplifyDemandedBits(Op1, DemandedBits, DemandedElts, Known, TLO, | ||||||
| 1183 | Depth + 1)) | ||||||
| 1184 | return true; | ||||||
| 1185 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1185, __PRETTY_FUNCTION__)); | ||||||
| 1186 | if (SimplifyDemandedBits(Op0, DemandedBits, DemandedElts, Known2, TLO, | ||||||
| 1187 | Depth + 1)) | ||||||
| 1188 | return true; | ||||||
| 1189 | assert(!Known2.hasConflict() && "Bits known to be one AND zero?")((!Known2.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known2.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1189, __PRETTY_FUNCTION__)); | ||||||
| 1190 | |||||||
| 1191 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1192 | if (!DemandedBits.isAllOnesValue() || !DemandedElts.isAllOnesValue()) { | ||||||
| 1193 | SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( | ||||||
| 1194 | Op0, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1195 | SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( | ||||||
| 1196 | Op1, DemandedBits, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 1197 | if (DemandedOp0 || DemandedOp1) { | ||||||
| 1198 | Op0 = DemandedOp0 ? DemandedOp0 : Op0; | ||||||
| 1199 | Op1 = DemandedOp1 ? DemandedOp1 : Op1; | ||||||
| 1200 | SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1); | ||||||
| 1201 | return TLO.CombineTo(Op, NewOp); | ||||||
| 1202 | } | ||||||
| 1203 | } | ||||||
| 1204 | |||||||
| 1205 | // If all of the demanded bits are known zero on one side, return the other. | ||||||
| 1206 | // These bits cannot contribute to the result of the 'xor'. | ||||||
| 1207 | if (DemandedBits.isSubsetOf(Known.Zero)) | ||||||
| 1208 | return TLO.CombineTo(Op, Op0); | ||||||
| 1209 | if (DemandedBits.isSubsetOf(Known2.Zero)) | ||||||
| 1210 | return TLO.CombineTo(Op, Op1); | ||||||
| 1211 | // If the operation can be done in a smaller type, do so. | ||||||
| 1212 | if (ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) | ||||||
| 1213 | return true; | ||||||
| 1214 | |||||||
| 1215 | // If all of the unknown bits are known to be zero on one side or the other | ||||||
| 1216 | // (but not both) turn this into an *inclusive* or. | ||||||
| 1217 | // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0 | ||||||
| 1218 | if (DemandedBits.isSubsetOf(Known.Zero | Known2.Zero)) | ||||||
| 1219 | return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::OR, dl, VT, Op0, Op1)); | ||||||
| 1220 | |||||||
| 1221 | // Output known-0 bits are known if clear or set in both the LHS & RHS. | ||||||
| 1222 | KnownOut.Zero = (Known.Zero & Known2.Zero) | (Known.One & Known2.One); | ||||||
| 1223 | // Output known-1 are known to be set if set in only one of the LHS, RHS. | ||||||
| 1224 | KnownOut.One = (Known.Zero & Known2.One) | (Known.One & Known2.Zero); | ||||||
| 1225 | |||||||
| 1226 | if (ConstantSDNode *C = isConstOrConstSplat(Op1)) { | ||||||
| 1227 | // If one side is a constant, and all of the known set bits on the other | ||||||
| 1228 | // side are also set in the constant, turn this into an AND, as we know | ||||||
| 1229 | // the bits will be cleared. | ||||||
| 1230 | // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2 | ||||||
| 1231 | // NB: it is okay if more bits are known than are requested | ||||||
| 1232 | if (C->getAPIntValue() == Known2.One) { | ||||||
| 1233 | SDValue ANDC = | ||||||
| 1234 | TLO.DAG.getConstant(~C->getAPIntValue() & DemandedBits, dl, VT); | ||||||
| 1235 | return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::AND, dl, VT, Op0, ANDC)); | ||||||
| 1236 | } | ||||||
| 1237 | |||||||
| 1238 | // If the RHS is a constant, see if we can change it. Don't alter a -1 | ||||||
| 1239 | // constant because that's a 'not' op, and that is better for combining | ||||||
| 1240 | // and codegen. | ||||||
| 1241 | if (!C->isAllOnesValue()) { | ||||||
| 1242 | if (DemandedBits.isSubsetOf(C->getAPIntValue())) { | ||||||
| 1243 | // We're flipping all demanded bits. Flip the undemanded bits too. | ||||||
| 1244 | SDValue New = TLO.DAG.getNOT(dl, Op0, VT); | ||||||
| 1245 | return TLO.CombineTo(Op, New); | ||||||
| 1246 | } | ||||||
| 1247 | // If we can't turn this into a 'not', try to shrink the constant. | ||||||
| 1248 | if (ShrinkDemandedConstant(Op, DemandedBits, TLO)) | ||||||
| 1249 | return true; | ||||||
| 1250 | } | ||||||
| 1251 | } | ||||||
| 1252 | |||||||
| 1253 | Known = std::move(KnownOut); | ||||||
| 1254 | break; | ||||||
| 1255 | } | ||||||
| 1256 | case ISD::SELECT: | ||||||
| 1257 | if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, Known, TLO, | ||||||
| 1258 | Depth + 1)) | ||||||
| 1259 | return true; | ||||||
| 1260 | if (SimplifyDemandedBits(Op.getOperand(1), DemandedBits, Known2, TLO, | ||||||
| 1261 | Depth + 1)) | ||||||
| 1262 | return true; | ||||||
| 1263 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1263, __PRETTY_FUNCTION__)); | ||||||
| 1264 | assert(!Known2.hasConflict() && "Bits known to be one AND zero?")((!Known2.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known2.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1264, __PRETTY_FUNCTION__)); | ||||||
| 1265 | |||||||
| 1266 | // If the operands are constants, see if we can simplify them. | ||||||
| 1267 | if (ShrinkDemandedConstant(Op, DemandedBits, TLO)) | ||||||
| 1268 | return true; | ||||||
| 1269 | |||||||
| 1270 | // Only known if known in both the LHS and RHS. | ||||||
| 1271 | Known.One &= Known2.One; | ||||||
| 1272 | Known.Zero &= Known2.Zero; | ||||||
| 1273 | break; | ||||||
| 1274 | case ISD::SELECT_CC: | ||||||
| 1275 | if (SimplifyDemandedBits(Op.getOperand(3), DemandedBits, Known, TLO, | ||||||
| 1276 | Depth + 1)) | ||||||
| 1277 | return true; | ||||||
| 1278 | if (SimplifyDemandedBits(Op.getOperand(2), DemandedBits, Known2, TLO, | ||||||
| 1279 | Depth + 1)) | ||||||
| 1280 | return true; | ||||||
| 1281 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1281, __PRETTY_FUNCTION__)); | ||||||
| 1282 | assert(!Known2.hasConflict() && "Bits known to be one AND zero?")((!Known2.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known2.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1282, __PRETTY_FUNCTION__)); | ||||||
| 1283 | |||||||
| 1284 | // If the operands are constants, see if we can simplify them. | ||||||
| 1285 | if (ShrinkDemandedConstant(Op, DemandedBits, TLO)) | ||||||
| 1286 | return true; | ||||||
| 1287 | |||||||
| 1288 | // Only known if known in both the LHS and RHS. | ||||||
| 1289 | Known.One &= Known2.One; | ||||||
| 1290 | Known.Zero &= Known2.Zero; | ||||||
| 1291 | break; | ||||||
| 1292 | case ISD::SETCC: { | ||||||
| 1293 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1294 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1295 | ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get(); | ||||||
| 1296 | // If (1) we only need the sign-bit, (2) the setcc operands are the same | ||||||
| 1297 | // width as the setcc result, and (3) the result of a setcc conforms to 0 or | ||||||
| 1298 | // -1, we may be able to bypass the setcc. | ||||||
| 1299 | if (DemandedBits.isSignMask() && | ||||||
| 1300 | Op0.getScalarValueSizeInBits() == BitWidth && | ||||||
| 1301 | getBooleanContents(Op0.getValueType()) == | ||||||
| 1302 | BooleanContent::ZeroOrNegativeOneBooleanContent) { | ||||||
| 1303 | // If we're testing X < 0, then this compare isn't needed - just use X! | ||||||
| 1304 | // FIXME: We're limiting to integer types here, but this should also work | ||||||
| 1305 | // if we don't care about FP signed-zero. The use of SETLT with FP means | ||||||
| 1306 | // that we don't care about NaNs. | ||||||
| 1307 | if (CC == ISD::SETLT && Op1.getValueType().isInteger() && | ||||||
| 1308 | (isNullConstant(Op1) || ISD::isBuildVectorAllZeros(Op1.getNode()))) | ||||||
| 1309 | return TLO.CombineTo(Op, Op0); | ||||||
| 1310 | |||||||
| 1311 | // TODO: Should we check for other forms of sign-bit comparisons? | ||||||
| 1312 | // Examples: X <= -1, X >= 0 | ||||||
| 1313 | } | ||||||
| 1314 | if (getBooleanContents(Op0.getValueType()) == | ||||||
| 1315 | TargetLowering::ZeroOrOneBooleanContent && | ||||||
| 1316 | BitWidth > 1) | ||||||
| 1317 | Known.Zero.setBitsFrom(1); | ||||||
| 1318 | break; | ||||||
| 1319 | } | ||||||
| 1320 | case ISD::SHL: { | ||||||
| 1321 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1322 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1323 | |||||||
| 1324 | if (ConstantSDNode *SA = isConstOrConstSplat(Op1, DemandedElts)) { | ||||||
| 1325 | // If the shift count is an invalid immediate, don't do anything. | ||||||
| 1326 | if (SA->getAPIntValue().uge(BitWidth)) | ||||||
| 1327 | break; | ||||||
| 1328 | |||||||
| 1329 | unsigned ShAmt = SA->getZExtValue(); | ||||||
| 1330 | if (ShAmt == 0) | ||||||
| 1331 | return TLO.CombineTo(Op, Op0); | ||||||
| 1332 | |||||||
| 1333 | // If this is ((X >>u C1) << ShAmt), see if we can simplify this into a | ||||||
| 1334 | // single shift. We can do this if the bottom bits (which are shifted | ||||||
| 1335 | // out) are never demanded. | ||||||
| 1336 | // TODO - support non-uniform vector amounts. | ||||||
| 1337 | if (Op0.getOpcode() == ISD::SRL) { | ||||||
| 1338 | if (!DemandedBits.intersects(APInt::getLowBitsSet(BitWidth, ShAmt))) { | ||||||
| 1339 | if (ConstantSDNode *SA2 = | ||||||
| 1340 | isConstOrConstSplat(Op0.getOperand(1), DemandedElts)) { | ||||||
| 1341 | if (SA2->getAPIntValue().ult(BitWidth)) { | ||||||
| 1342 | unsigned C1 = SA2->getZExtValue(); | ||||||
| 1343 | unsigned Opc = ISD::SHL; | ||||||
| 1344 | int Diff = ShAmt - C1; | ||||||
| 1345 | if (Diff < 0) { | ||||||
| 1346 | Diff = -Diff; | ||||||
| 1347 | Opc = ISD::SRL; | ||||||
| 1348 | } | ||||||
| 1349 | |||||||
| 1350 | SDValue NewSA = TLO.DAG.getConstant(Diff, dl, Op1.getValueType()); | ||||||
| 1351 | return TLO.CombineTo( | ||||||
| 1352 | Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA)); | ||||||
| 1353 | } | ||||||
| 1354 | } | ||||||
| 1355 | } | ||||||
| 1356 | } | ||||||
| 1357 | |||||||
| 1358 | if (SimplifyDemandedBits(Op0, DemandedBits.lshr(ShAmt), DemandedElts, | ||||||
| 1359 | Known, TLO, Depth + 1)) | ||||||
| 1360 | return true; | ||||||
| 1361 | |||||||
| 1362 | // Try shrinking the operation as long as the shift amount will still be | ||||||
| 1363 | // in range. | ||||||
| 1364 | if ((ShAmt < DemandedBits.getActiveBits()) && | ||||||
| 1365 | ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) | ||||||
| 1366 | return true; | ||||||
| 1367 | |||||||
| 1368 | // Convert (shl (anyext x, c)) to (anyext (shl x, c)) if the high bits | ||||||
| 1369 | // are not demanded. This will likely allow the anyext to be folded away. | ||||||
| 1370 | if (Op0.getOpcode() == ISD::ANY_EXTEND) { | ||||||
| 1371 | SDValue InnerOp = Op0.getOperand(0); | ||||||
| 1372 | EVT InnerVT = InnerOp.getValueType(); | ||||||
| 1373 | unsigned InnerBits = InnerVT.getScalarSizeInBits(); | ||||||
| 1374 | if (ShAmt < InnerBits && DemandedBits.getActiveBits() <= InnerBits && | ||||||
| 1375 | isTypeDesirableForOp(ISD::SHL, InnerVT)) { | ||||||
| 1376 | EVT ShTy = getShiftAmountTy(InnerVT, DL); | ||||||
| 1377 | if (!APInt(BitWidth, ShAmt).isIntN(ShTy.getSizeInBits())) | ||||||
| 1378 | ShTy = InnerVT; | ||||||
| 1379 | SDValue NarrowShl = | ||||||
| 1380 | TLO.DAG.getNode(ISD::SHL, dl, InnerVT, InnerOp, | ||||||
| 1381 | TLO.DAG.getConstant(ShAmt, dl, ShTy)); | ||||||
| 1382 | return TLO.CombineTo( | ||||||
| 1383 | Op, TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, NarrowShl)); | ||||||
| 1384 | } | ||||||
| 1385 | // Repeat the SHL optimization above in cases where an extension | ||||||
| 1386 | // intervenes: (shl (anyext (shr x, c1)), c2) to | ||||||
| 1387 | // (shl (anyext x), c2-c1). This requires that the bottom c1 bits | ||||||
| 1388 | // aren't demanded (as above) and that the shifted upper c1 bits of | ||||||
| 1389 | // x aren't demanded. | ||||||
| 1390 | if (Op0.hasOneUse() && InnerOp.getOpcode() == ISD::SRL && | ||||||
| 1391 | InnerOp.hasOneUse()) { | ||||||
| 1392 | if (ConstantSDNode *SA2 = | ||||||
| 1393 | isConstOrConstSplat(InnerOp.getOperand(1))) { | ||||||
| 1394 | unsigned InnerShAmt = SA2->getLimitedValue(InnerBits); | ||||||
| 1395 | if (InnerShAmt < ShAmt && InnerShAmt < InnerBits && | ||||||
| 1396 | DemandedBits.getActiveBits() <= | ||||||
| 1397 | (InnerBits - InnerShAmt + ShAmt) && | ||||||
| 1398 | DemandedBits.countTrailingZeros() >= ShAmt) { | ||||||
| 1399 | SDValue NewSA = TLO.DAG.getConstant(ShAmt - InnerShAmt, dl, | ||||||
| 1400 | Op1.getValueType()); | ||||||
| 1401 | SDValue NewExt = TLO.DAG.getNode(ISD::ANY_EXTEND, dl, VT, | ||||||
| 1402 | InnerOp.getOperand(0)); | ||||||
| 1403 | return TLO.CombineTo( | ||||||
| 1404 | Op, TLO.DAG.getNode(ISD::SHL, dl, VT, NewExt, NewSA)); | ||||||
| 1405 | } | ||||||
| 1406 | } | ||||||
| 1407 | } | ||||||
| 1408 | } | ||||||
| 1409 | |||||||
| 1410 | Known.Zero <<= ShAmt; | ||||||
| 1411 | Known.One <<= ShAmt; | ||||||
| 1412 | // low bits known zero. | ||||||
| 1413 | Known.Zero.setLowBits(ShAmt); | ||||||
| 1414 | } | ||||||
| 1415 | break; | ||||||
| 1416 | } | ||||||
| 1417 | case ISD::SRL: { | ||||||
| 1418 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1419 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1420 | |||||||
| 1421 | if (ConstantSDNode *SA = isConstOrConstSplat(Op1, DemandedElts)) { | ||||||
| 1422 | // If the shift count is an invalid immediate, don't do anything. | ||||||
| 1423 | if (SA->getAPIntValue().uge(BitWidth)) | ||||||
| 1424 | break; | ||||||
| 1425 | |||||||
| 1426 | unsigned ShAmt = SA->getZExtValue(); | ||||||
| 1427 | if (ShAmt == 0) | ||||||
| 1428 | return TLO.CombineTo(Op, Op0); | ||||||
| 1429 | |||||||
| 1430 | EVT ShiftVT = Op1.getValueType(); | ||||||
| 1431 | APInt InDemandedMask = (DemandedBits << ShAmt); | ||||||
| 1432 | |||||||
| 1433 | // If the shift is exact, then it does demand the low bits (and knows that | ||||||
| 1434 | // they are zero). | ||||||
| 1435 | if (Op->getFlags().hasExact()) | ||||||
| 1436 | InDemandedMask.setLowBits(ShAmt); | ||||||
| 1437 | |||||||
| 1438 | // If this is ((X << C1) >>u ShAmt), see if we can simplify this into a | ||||||
| 1439 | // single shift. We can do this if the top bits (which are shifted out) | ||||||
| 1440 | // are never demanded. | ||||||
| 1441 | // TODO - support non-uniform vector amounts. | ||||||
| 1442 | if (Op0.getOpcode() == ISD::SHL) { | ||||||
| 1443 | if (ConstantSDNode *SA2 = | ||||||
| 1444 | isConstOrConstSplat(Op0.getOperand(1), DemandedElts)) { | ||||||
| 1445 | if (!DemandedBits.intersects( | ||||||
| 1446 | APInt::getHighBitsSet(BitWidth, ShAmt))) { | ||||||
| 1447 | if (SA2->getAPIntValue().ult(BitWidth)) { | ||||||
| 1448 | unsigned C1 = SA2->getZExtValue(); | ||||||
| 1449 | unsigned Opc = ISD::SRL; | ||||||
| 1450 | int Diff = ShAmt - C1; | ||||||
| 1451 | if (Diff < 0) { | ||||||
| 1452 | Diff = -Diff; | ||||||
| 1453 | Opc = ISD::SHL; | ||||||
| 1454 | } | ||||||
| 1455 | |||||||
| 1456 | SDValue NewSA = TLO.DAG.getConstant(Diff, dl, ShiftVT); | ||||||
| 1457 | return TLO.CombineTo( | ||||||
| 1458 | Op, TLO.DAG.getNode(Opc, dl, VT, Op0.getOperand(0), NewSA)); | ||||||
| 1459 | } | ||||||
| 1460 | } | ||||||
| 1461 | } | ||||||
| 1462 | } | ||||||
| 1463 | |||||||
| 1464 | // Compute the new bits that are at the top now. | ||||||
| 1465 | if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO, | ||||||
| 1466 | Depth + 1)) | ||||||
| 1467 | return true; | ||||||
| 1468 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1468, __PRETTY_FUNCTION__)); | ||||||
| 1469 | Known.Zero.lshrInPlace(ShAmt); | ||||||
| 1470 | Known.One.lshrInPlace(ShAmt); | ||||||
| 1471 | |||||||
| 1472 | Known.Zero.setHighBits(ShAmt); // High bits known zero. | ||||||
| 1473 | } | ||||||
| 1474 | break; | ||||||
| 1475 | } | ||||||
| 1476 | case ISD::SRA: { | ||||||
| 1477 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1478 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1479 | |||||||
| 1480 | // If this is an arithmetic shift right and only the low-bit is set, we can | ||||||
| 1481 | // always convert this into a logical shr, even if the shift amount is | ||||||
| 1482 | // variable. The low bit of the shift cannot be an input sign bit unless | ||||||
| 1483 | // the shift amount is >= the size of the datatype, which is undefined. | ||||||
| 1484 | if (DemandedBits.isOneValue()) | ||||||
| 1485 | return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1)); | ||||||
| 1486 | |||||||
| 1487 | if (ConstantSDNode *SA = isConstOrConstSplat(Op1, DemandedElts)) { | ||||||
| 1488 | // If the shift count is an invalid immediate, don't do anything. | ||||||
| 1489 | if (SA->getAPIntValue().uge(BitWidth)) | ||||||
| 1490 | break; | ||||||
| 1491 | |||||||
| 1492 | unsigned ShAmt = SA->getZExtValue(); | ||||||
| 1493 | if (ShAmt == 0) | ||||||
| 1494 | return TLO.CombineTo(Op, Op0); | ||||||
| 1495 | |||||||
| 1496 | APInt InDemandedMask = (DemandedBits << ShAmt); | ||||||
| 1497 | |||||||
| 1498 | // If the shift is exact, then it does demand the low bits (and knows that | ||||||
| 1499 | // they are zero). | ||||||
| 1500 | if (Op->getFlags().hasExact()) | ||||||
| 1501 | InDemandedMask.setLowBits(ShAmt); | ||||||
| 1502 | |||||||
| 1503 | // If any of the demanded bits are produced by the sign extension, we also | ||||||
| 1504 | // demand the input sign bit. | ||||||
| 1505 | if (DemandedBits.countLeadingZeros() < ShAmt) | ||||||
| 1506 | InDemandedMask.setSignBit(); | ||||||
| 1507 | |||||||
| 1508 | if (SimplifyDemandedBits(Op0, InDemandedMask, DemandedElts, Known, TLO, | ||||||
| 1509 | Depth + 1)) | ||||||
| 1510 | return true; | ||||||
| 1511 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1511, __PRETTY_FUNCTION__)); | ||||||
| 1512 | Known.Zero.lshrInPlace(ShAmt); | ||||||
| 1513 | Known.One.lshrInPlace(ShAmt); | ||||||
| 1514 | |||||||
| 1515 | // If the input sign bit is known to be zero, or if none of the top bits | ||||||
| 1516 | // are demanded, turn this into an unsigned shift right. | ||||||
| 1517 | if (Known.Zero[BitWidth - ShAmt - 1] || | ||||||
| 1518 | DemandedBits.countLeadingZeros() >= ShAmt) { | ||||||
| 1519 | SDNodeFlags Flags; | ||||||
| 1520 | Flags.setExact(Op->getFlags().hasExact()); | ||||||
| 1521 | return TLO.CombineTo( | ||||||
| 1522 | Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, Op1, Flags)); | ||||||
| 1523 | } | ||||||
| 1524 | |||||||
| 1525 | int Log2 = DemandedBits.exactLogBase2(); | ||||||
| 1526 | if (Log2 >= 0) { | ||||||
| 1527 | // The bit must come from the sign. | ||||||
| 1528 | SDValue NewSA = | ||||||
| 1529 | TLO.DAG.getConstant(BitWidth - 1 - Log2, dl, Op1.getValueType()); | ||||||
| 1530 | return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::SRL, dl, VT, Op0, NewSA)); | ||||||
| 1531 | } | ||||||
| 1532 | |||||||
| 1533 | if (Known.One[BitWidth - ShAmt - 1]) | ||||||
| 1534 | // New bits are known one. | ||||||
| 1535 | Known.One.setHighBits(ShAmt); | ||||||
| 1536 | } | ||||||
| 1537 | break; | ||||||
| 1538 | } | ||||||
| 1539 | case ISD::FSHL: | ||||||
| 1540 | case ISD::FSHR: { | ||||||
| 1541 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1542 | SDValue Op1 = Op.getOperand(1); | ||||||
| 1543 | SDValue Op2 = Op.getOperand(2); | ||||||
| 1544 | bool IsFSHL = (Op.getOpcode() == ISD::FSHL); | ||||||
| 1545 | |||||||
| 1546 | if (ConstantSDNode *SA = isConstOrConstSplat(Op2, DemandedElts)) { | ||||||
| 1547 | unsigned Amt = SA->getAPIntValue().urem(BitWidth); | ||||||
| 1548 | |||||||
| 1549 | // For fshl, 0-shift returns the 1st arg. | ||||||
| 1550 | // For fshr, 0-shift returns the 2nd arg. | ||||||
| 1551 | if (Amt == 0) { | ||||||
| 1552 | if (SimplifyDemandedBits(IsFSHL ? Op0 : Op1, DemandedBits, DemandedElts, | ||||||
| 1553 | Known, TLO, Depth + 1)) | ||||||
| 1554 | return true; | ||||||
| 1555 | break; | ||||||
| 1556 | } | ||||||
| 1557 | |||||||
| 1558 | // fshl: (Op0 << Amt) | (Op1 >> (BW - Amt)) | ||||||
| 1559 | // fshr: (Op0 << (BW - Amt)) | (Op1 >> Amt) | ||||||
| 1560 | APInt Demanded0 = DemandedBits.lshr(IsFSHL ? Amt : (BitWidth - Amt)); | ||||||
| 1561 | APInt Demanded1 = DemandedBits << (IsFSHL ? (BitWidth - Amt) : Amt); | ||||||
| 1562 | if (SimplifyDemandedBits(Op0, Demanded0, DemandedElts, Known2, TLO, | ||||||
| 1563 | Depth + 1)) | ||||||
| 1564 | return true; | ||||||
| 1565 | if (SimplifyDemandedBits(Op1, Demanded1, DemandedElts, Known, TLO, | ||||||
| 1566 | Depth + 1)) | ||||||
| 1567 | return true; | ||||||
| 1568 | |||||||
| 1569 | Known2.One <<= (IsFSHL ? Amt : (BitWidth - Amt)); | ||||||
| 1570 | Known2.Zero <<= (IsFSHL ? Amt : (BitWidth - Amt)); | ||||||
| 1571 | Known.One.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt); | ||||||
| 1572 | Known.Zero.lshrInPlace(IsFSHL ? (BitWidth - Amt) : Amt); | ||||||
| 1573 | Known.One |= Known2.One; | ||||||
| 1574 | Known.Zero |= Known2.Zero; | ||||||
| 1575 | } | ||||||
| 1576 | break; | ||||||
| 1577 | } | ||||||
| 1578 | case ISD::BITREVERSE: { | ||||||
| 1579 | SDValue Src = Op.getOperand(0); | ||||||
| 1580 | APInt DemandedSrcBits = DemandedBits.reverseBits(); | ||||||
| 1581 | if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO, | ||||||
| 1582 | Depth + 1)) | ||||||
| 1583 | return true; | ||||||
| 1584 | Known.One = Known2.One.reverseBits(); | ||||||
| 1585 | Known.Zero = Known2.Zero.reverseBits(); | ||||||
| 1586 | break; | ||||||
| 1587 | } | ||||||
| 1588 | case ISD::BSWAP: { | ||||||
| 1589 | SDValue Src = Op.getOperand(0); | ||||||
| 1590 | APInt DemandedSrcBits = DemandedBits.byteSwap(); | ||||||
| 1591 | if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedElts, Known2, TLO, | ||||||
| 1592 | Depth + 1)) | ||||||
| 1593 | return true; | ||||||
| 1594 | Known.One = Known2.One.byteSwap(); | ||||||
| 1595 | Known.Zero = Known2.Zero.byteSwap(); | ||||||
| 1596 | break; | ||||||
| 1597 | } | ||||||
| 1598 | case ISD::SIGN_EXTEND_INREG: { | ||||||
| 1599 | SDValue Op0 = Op.getOperand(0); | ||||||
| 1600 | EVT ExVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); | ||||||
| 1601 | unsigned ExVTBits = ExVT.getScalarSizeInBits(); | ||||||
| 1602 | |||||||
| 1603 | // If we only care about the highest bit, don't bother shifting right. | ||||||
| 1604 | if (DemandedBits.isSignMask()) { | ||||||
| 1605 | unsigned NumSignBits = TLO.DAG.ComputeNumSignBits(Op0); | ||||||
| 1606 | bool AlreadySignExtended = NumSignBits >= BitWidth - ExVTBits + 1; | ||||||
| 1607 | // However if the input is already sign extended we expect the sign | ||||||
| 1608 | // extension to be dropped altogether later and do not simplify. | ||||||
| 1609 | if (!AlreadySignExtended) { | ||||||
| 1610 | // Compute the correct shift amount type, which must be getShiftAmountTy | ||||||
| 1611 | // for scalar types after legalization. | ||||||
| 1612 | EVT ShiftAmtTy = VT; | ||||||
| 1613 | if (TLO.LegalTypes() && !ShiftAmtTy.isVector()) | ||||||
| 1614 | ShiftAmtTy = getShiftAmountTy(ShiftAmtTy, DL); | ||||||
| 1615 | |||||||
| 1616 | SDValue ShiftAmt = | ||||||
| 1617 | TLO.DAG.getConstant(BitWidth - ExVTBits, dl, ShiftAmtTy); | ||||||
| 1618 | return TLO.CombineTo(Op, | ||||||
| 1619 | TLO.DAG.getNode(ISD::SHL, dl, VT, Op0, ShiftAmt)); | ||||||
| 1620 | } | ||||||
| 1621 | } | ||||||
| 1622 | |||||||
| 1623 | // If none of the extended bits are demanded, eliminate the sextinreg. | ||||||
| 1624 | if (DemandedBits.getActiveBits() <= ExVTBits) | ||||||
| 1625 | return TLO.CombineTo(Op, Op0); | ||||||
| 1626 | |||||||
| 1627 | APInt InputDemandedBits = DemandedBits.getLoBits(ExVTBits); | ||||||
| 1628 | |||||||
| 1629 | // Since the sign extended bits are demanded, we know that the sign | ||||||
| 1630 | // bit is demanded. | ||||||
| 1631 | InputDemandedBits.setBit(ExVTBits - 1); | ||||||
| 1632 | |||||||
| 1633 | if (SimplifyDemandedBits(Op0, InputDemandedBits, Known, TLO, Depth + 1)) | ||||||
| 1634 | return true; | ||||||
| 1635 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1635, __PRETTY_FUNCTION__)); | ||||||
| 1636 | |||||||
| 1637 | // If the sign bit of the input is known set or clear, then we know the | ||||||
| 1638 | // top bits of the result. | ||||||
| 1639 | |||||||
| 1640 | // If the input sign bit is known zero, convert this into a zero extension. | ||||||
| 1641 | if (Known.Zero[ExVTBits - 1]) | ||||||
| 1642 | return TLO.CombineTo( | ||||||
| 1643 | Op, TLO.DAG.getZeroExtendInReg(Op0, dl, ExVT.getScalarType())); | ||||||
| 1644 | |||||||
| 1645 | APInt Mask = APInt::getLowBitsSet(BitWidth, ExVTBits); | ||||||
| 1646 | if (Known.One[ExVTBits - 1]) { // Input sign bit known set | ||||||
| 1647 | Known.One.setBitsFrom(ExVTBits); | ||||||
| 1648 | Known.Zero &= Mask; | ||||||
| 1649 | } else { // Input sign bit unknown | ||||||
| 1650 | Known.Zero &= Mask; | ||||||
| 1651 | Known.One &= Mask; | ||||||
| 1652 | } | ||||||
| 1653 | break; | ||||||
| 1654 | } | ||||||
| 1655 | case ISD::BUILD_PAIR: { | ||||||
| 1656 | EVT HalfVT = Op.getOperand(0).getValueType(); | ||||||
| 1657 | unsigned HalfBitWidth = HalfVT.getScalarSizeInBits(); | ||||||
| 1658 | |||||||
| 1659 | APInt MaskLo = DemandedBits.getLoBits(HalfBitWidth).trunc(HalfBitWidth); | ||||||
| 1660 | APInt MaskHi = DemandedBits.getHiBits(HalfBitWidth).trunc(HalfBitWidth); | ||||||
| 1661 | |||||||
| 1662 | KnownBits KnownLo, KnownHi; | ||||||
| 1663 | |||||||
| 1664 | if (SimplifyDemandedBits(Op.getOperand(0), MaskLo, KnownLo, TLO, Depth + 1)) | ||||||
| 1665 | return true; | ||||||
| 1666 | |||||||
| 1667 | if (SimplifyDemandedBits(Op.getOperand(1), MaskHi, KnownHi, TLO, Depth + 1)) | ||||||
| 1668 | return true; | ||||||
| 1669 | |||||||
| 1670 | Known.Zero = KnownLo.Zero.zext(BitWidth) | | ||||||
| 1671 | KnownHi.Zero.zext(BitWidth).shl(HalfBitWidth); | ||||||
| 1672 | |||||||
| 1673 | Known.One = KnownLo.One.zext(BitWidth) | | ||||||
| 1674 | KnownHi.One.zext(BitWidth).shl(HalfBitWidth); | ||||||
| 1675 | break; | ||||||
| 1676 | } | ||||||
| 1677 | case ISD::ZERO_EXTEND: | ||||||
| 1678 | case ISD::ZERO_EXTEND_VECTOR_INREG: { | ||||||
| 1679 | SDValue Src = Op.getOperand(0); | ||||||
| 1680 | EVT SrcVT = Src.getValueType(); | ||||||
| 1681 | unsigned InBits = SrcVT.getScalarSizeInBits(); | ||||||
| 1682 | unsigned InElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; | ||||||
| 1683 | bool IsVecInReg = Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG; | ||||||
| 1684 | |||||||
| 1685 | // If none of the top bits are demanded, convert this into an any_extend. | ||||||
| 1686 | if (DemandedBits.getActiveBits() <= InBits) { | ||||||
| 1687 | // If we only need the non-extended bits of the bottom element | ||||||
| 1688 | // then we can just bitcast to the result. | ||||||
| 1689 | if (IsVecInReg && DemandedElts == 1 && | ||||||
| 1690 | VT.getSizeInBits() == SrcVT.getSizeInBits() && | ||||||
| 1691 | TLO.DAG.getDataLayout().isLittleEndian()) | ||||||
| 1692 | return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); | ||||||
| 1693 | |||||||
| 1694 | unsigned Opc = | ||||||
| 1695 | IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND; | ||||||
| 1696 | if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) | ||||||
| 1697 | return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src)); | ||||||
| 1698 | } | ||||||
| 1699 | |||||||
| 1700 | APInt InDemandedBits = DemandedBits.trunc(InBits); | ||||||
| 1701 | APInt InDemandedElts = DemandedElts.zextOrSelf(InElts); | ||||||
| 1702 | if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, | ||||||
| 1703 | Depth + 1)) | ||||||
| 1704 | return true; | ||||||
| 1705 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1705, __PRETTY_FUNCTION__)); | ||||||
| 1706 | assert(Known.getBitWidth() == InBits && "Src width has changed?")((Known.getBitWidth() == InBits && "Src width has changed?" ) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == InBits && \"Src width has changed?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1706, __PRETTY_FUNCTION__)); | ||||||
| 1707 | Known = Known.zext(BitWidth, true /* ExtendedBitsAreKnownZero */); | ||||||
| 1708 | break; | ||||||
| 1709 | } | ||||||
| 1710 | case ISD::SIGN_EXTEND: | ||||||
| 1711 | case ISD::SIGN_EXTEND_VECTOR_INREG: { | ||||||
| 1712 | SDValue Src = Op.getOperand(0); | ||||||
| 1713 | EVT SrcVT = Src.getValueType(); | ||||||
| 1714 | unsigned InBits = SrcVT.getScalarSizeInBits(); | ||||||
| 1715 | unsigned InElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; | ||||||
| 1716 | bool IsVecInReg = Op.getOpcode() == ISD::SIGN_EXTEND_VECTOR_INREG; | ||||||
| 1717 | |||||||
| 1718 | // If none of the top bits are demanded, convert this into an any_extend. | ||||||
| 1719 | if (DemandedBits.getActiveBits() <= InBits) { | ||||||
| 1720 | // If we only need the non-extended bits of the bottom element | ||||||
| 1721 | // then we can just bitcast to the result. | ||||||
| 1722 | if (IsVecInReg && DemandedElts == 1 && | ||||||
| 1723 | VT.getSizeInBits() == SrcVT.getSizeInBits() && | ||||||
| 1724 | TLO.DAG.getDataLayout().isLittleEndian()) | ||||||
| 1725 | return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); | ||||||
| 1726 | |||||||
| 1727 | unsigned Opc = | ||||||
| 1728 | IsVecInReg ? ISD::ANY_EXTEND_VECTOR_INREG : ISD::ANY_EXTEND; | ||||||
| 1729 | if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) | ||||||
| 1730 | return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src)); | ||||||
| 1731 | } | ||||||
| 1732 | |||||||
| 1733 | APInt InDemandedBits = DemandedBits.trunc(InBits); | ||||||
| 1734 | APInt InDemandedElts = DemandedElts.zextOrSelf(InElts); | ||||||
| 1735 | |||||||
| 1736 | // Since some of the sign extended bits are demanded, we know that the sign | ||||||
| 1737 | // bit is demanded. | ||||||
| 1738 | InDemandedBits.setBit(InBits - 1); | ||||||
| 1739 | |||||||
| 1740 | if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, | ||||||
| 1741 | Depth + 1)) | ||||||
| 1742 | return true; | ||||||
| 1743 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1743, __PRETTY_FUNCTION__)); | ||||||
| 1744 | assert(Known.getBitWidth() == InBits && "Src width has changed?")((Known.getBitWidth() == InBits && "Src width has changed?" ) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == InBits && \"Src width has changed?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1744, __PRETTY_FUNCTION__)); | ||||||
| 1745 | |||||||
| 1746 | // If the sign bit is known one, the top bits match. | ||||||
| 1747 | Known = Known.sext(BitWidth); | ||||||
| 1748 | |||||||
| 1749 | // If the sign bit is known zero, convert this to a zero extend. | ||||||
| 1750 | if (Known.isNonNegative()) { | ||||||
| 1751 | unsigned Opc = | ||||||
| 1752 | IsVecInReg ? ISD::ZERO_EXTEND_VECTOR_INREG : ISD::ZERO_EXTEND; | ||||||
| 1753 | if (!TLO.LegalOperations() || isOperationLegal(Opc, VT)) | ||||||
| 1754 | return TLO.CombineTo(Op, TLO.DAG.getNode(Opc, dl, VT, Src)); | ||||||
| 1755 | } | ||||||
| 1756 | break; | ||||||
| 1757 | } | ||||||
| 1758 | case ISD::ANY_EXTEND: | ||||||
| 1759 | case ISD::ANY_EXTEND_VECTOR_INREG: { | ||||||
| 1760 | SDValue Src = Op.getOperand(0); | ||||||
| 1761 | EVT SrcVT = Src.getValueType(); | ||||||
| 1762 | unsigned InBits = SrcVT.getScalarSizeInBits(); | ||||||
| 1763 | unsigned InElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; | ||||||
| 1764 | bool IsVecInReg = Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG; | ||||||
| 1765 | |||||||
| 1766 | // If we only need the bottom element then we can just bitcast. | ||||||
| 1767 | // TODO: Handle ANY_EXTEND? | ||||||
| 1768 | if (IsVecInReg && DemandedElts == 1 && | ||||||
| 1769 | VT.getSizeInBits() == SrcVT.getSizeInBits() && | ||||||
| 1770 | TLO.DAG.getDataLayout().isLittleEndian()) | ||||||
| 1771 | return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); | ||||||
| 1772 | |||||||
| 1773 | APInt InDemandedBits = DemandedBits.trunc(InBits); | ||||||
| 1774 | APInt InDemandedElts = DemandedElts.zextOrSelf(InElts); | ||||||
| 1775 | if (SimplifyDemandedBits(Src, InDemandedBits, InDemandedElts, Known, TLO, | ||||||
| 1776 | Depth + 1)) | ||||||
| 1777 | return true; | ||||||
| 1778 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1778, __PRETTY_FUNCTION__)); | ||||||
| 1779 | assert(Known.getBitWidth() == InBits && "Src width has changed?")((Known.getBitWidth() == InBits && "Src width has changed?" ) ? static_cast<void> (0) : __assert_fail ("Known.getBitWidth() == InBits && \"Src width has changed?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1779, __PRETTY_FUNCTION__)); | ||||||
| 1780 | Known = Known.zext(BitWidth, false /* => any extend */); | ||||||
| 1781 | break; | ||||||
| 1782 | } | ||||||
| 1783 | case ISD::TRUNCATE: { | ||||||
| 1784 | SDValue Src = Op.getOperand(0); | ||||||
| 1785 | |||||||
| 1786 | // Simplify the input, using demanded bit information, and compute the known | ||||||
| 1787 | // zero/one bits live out. | ||||||
| 1788 | unsigned OperandBitWidth = Src.getScalarValueSizeInBits(); | ||||||
| 1789 | APInt TruncMask = DemandedBits.zext(OperandBitWidth); | ||||||
| 1790 | if (SimplifyDemandedBits(Src, TruncMask, Known, TLO, Depth + 1)) | ||||||
| 1791 | return true; | ||||||
| 1792 | Known = Known.trunc(BitWidth); | ||||||
| 1793 | |||||||
| 1794 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1795 | if (SDValue NewSrc = SimplifyMultipleUseDemandedBits( | ||||||
| 1796 | Src, TruncMask, DemandedElts, TLO.DAG, Depth + 1)) | ||||||
| 1797 | return TLO.CombineTo(Op, TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, NewSrc)); | ||||||
| 1798 | |||||||
| 1799 | // If the input is only used by this truncate, see if we can shrink it based | ||||||
| 1800 | // on the known demanded bits. | ||||||
| 1801 | if (Src.getNode()->hasOneUse()) { | ||||||
| 1802 | switch (Src.getOpcode()) { | ||||||
| 1803 | default: | ||||||
| 1804 | break; | ||||||
| 1805 | case ISD::SRL: | ||||||
| 1806 | // Shrink SRL by a constant if none of the high bits shifted in are | ||||||
| 1807 | // demanded. | ||||||
| 1808 | if (TLO.LegalTypes() && !isTypeDesirableForOp(ISD::SRL, VT)) | ||||||
| 1809 | // Do not turn (vt1 truncate (vt2 srl)) into (vt1 srl) if vt1 is | ||||||
| 1810 | // undesirable. | ||||||
| 1811 | break; | ||||||
| 1812 | |||||||
| 1813 | SDValue ShAmt = Src.getOperand(1); | ||||||
| 1814 | auto *ShAmtC = dyn_cast<ConstantSDNode>(ShAmt); | ||||||
| 1815 | if (!ShAmtC || ShAmtC->getAPIntValue().uge(BitWidth)) | ||||||
| 1816 | break; | ||||||
| 1817 | uint64_t ShVal = ShAmtC->getZExtValue(); | ||||||
| 1818 | |||||||
| 1819 | APInt HighBits = | ||||||
| 1820 | APInt::getHighBitsSet(OperandBitWidth, OperandBitWidth - BitWidth); | ||||||
| 1821 | HighBits.lshrInPlace(ShVal); | ||||||
| 1822 | HighBits = HighBits.trunc(BitWidth); | ||||||
| 1823 | |||||||
| 1824 | if (!(HighBits & DemandedBits)) { | ||||||
| 1825 | // None of the shifted in bits are needed. Add a truncate of the | ||||||
| 1826 | // shift input, then shift it. | ||||||
| 1827 | if (TLO.LegalTypes()) | ||||||
| 1828 | ShAmt = TLO.DAG.getConstant(ShVal, dl, getShiftAmountTy(VT, DL)); | ||||||
| 1829 | SDValue NewTrunc = | ||||||
| 1830 | TLO.DAG.getNode(ISD::TRUNCATE, dl, VT, Src.getOperand(0)); | ||||||
| 1831 | return TLO.CombineTo( | ||||||
| 1832 | Op, TLO.DAG.getNode(ISD::SRL, dl, VT, NewTrunc, ShAmt)); | ||||||
| 1833 | } | ||||||
| 1834 | break; | ||||||
| 1835 | } | ||||||
| 1836 | } | ||||||
| 1837 | |||||||
| 1838 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1838, __PRETTY_FUNCTION__)); | ||||||
| 1839 | break; | ||||||
| 1840 | } | ||||||
| 1841 | case ISD::AssertZext: { | ||||||
| 1842 | // AssertZext demands all of the high bits, plus any of the low bits | ||||||
| 1843 | // demanded by its users. | ||||||
| 1844 | EVT ZVT = cast<VTSDNode>(Op.getOperand(1))->getVT(); | ||||||
| 1845 | APInt InMask = APInt::getLowBitsSet(BitWidth, ZVT.getSizeInBits()); | ||||||
| 1846 | if (SimplifyDemandedBits(Op.getOperand(0), ~InMask | DemandedBits, Known, | ||||||
| 1847 | TLO, Depth + 1)) | ||||||
| 1848 | return true; | ||||||
| 1849 | assert(!Known.hasConflict() && "Bits known to be one AND zero?")((!Known.hasConflict() && "Bits known to be one AND zero?" ) ? static_cast<void> (0) : __assert_fail ("!Known.hasConflict() && \"Bits known to be one AND zero?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 1849, __PRETTY_FUNCTION__)); | ||||||
| 1850 | |||||||
| 1851 | Known.Zero |= ~InMask; | ||||||
| 1852 | break; | ||||||
| 1853 | } | ||||||
| 1854 | case ISD::EXTRACT_VECTOR_ELT: { | ||||||
| 1855 | SDValue Src = Op.getOperand(0); | ||||||
| 1856 | SDValue Idx = Op.getOperand(1); | ||||||
| 1857 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); | ||||||
| 1858 | unsigned EltBitWidth = Src.getScalarValueSizeInBits(); | ||||||
| 1859 | |||||||
| 1860 | // Demand the bits from every vector element without a constant index. | ||||||
| 1861 | APInt DemandedSrcElts = APInt::getAllOnesValue(NumSrcElts); | ||||||
| 1862 | if (auto *CIdx = dyn_cast<ConstantSDNode>(Idx)) | ||||||
| 1863 | if (CIdx->getAPIntValue().ult(NumSrcElts)) | ||||||
| 1864 | DemandedSrcElts = APInt::getOneBitSet(NumSrcElts, CIdx->getZExtValue()); | ||||||
| 1865 | |||||||
| 1866 | // If BitWidth > EltBitWidth the value is anyext:ed. So we do not know | ||||||
| 1867 | // anything about the extended bits. | ||||||
| 1868 | APInt DemandedSrcBits = DemandedBits; | ||||||
| 1869 | if (BitWidth > EltBitWidth) | ||||||
| 1870 | DemandedSrcBits = DemandedSrcBits.trunc(EltBitWidth); | ||||||
| 1871 | |||||||
| 1872 | if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, Known2, TLO, | ||||||
| 1873 | Depth + 1)) | ||||||
| 1874 | return true; | ||||||
| 1875 | |||||||
| 1876 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 1877 | if (!DemandedSrcBits.isAllOnesValue() || | ||||||
| 1878 | !DemandedSrcElts.isAllOnesValue()) { | ||||||
| 1879 | if (SDValue DemandedSrc = SimplifyMultipleUseDemandedBits( | ||||||
| 1880 | Src, DemandedSrcBits, DemandedSrcElts, TLO.DAG, Depth + 1)) { | ||||||
| 1881 | SDValue NewOp = | ||||||
| 1882 | TLO.DAG.getNode(Op.getOpcode(), dl, VT, DemandedSrc, Idx); | ||||||
| 1883 | return TLO.CombineTo(Op, NewOp); | ||||||
| 1884 | } | ||||||
| 1885 | } | ||||||
| 1886 | |||||||
| 1887 | Known = Known2; | ||||||
| 1888 | if (BitWidth > EltBitWidth) | ||||||
| 1889 | Known = Known.zext(BitWidth, false /* => any extend */); | ||||||
| 1890 | break; | ||||||
| 1891 | } | ||||||
| 1892 | case ISD::BITCAST: { | ||||||
| 1893 | SDValue Src = Op.getOperand(0); | ||||||
| 1894 | EVT SrcVT = Src.getValueType(); | ||||||
| 1895 | unsigned NumSrcEltBits = SrcVT.getScalarSizeInBits(); | ||||||
| 1896 | |||||||
| 1897 | // If this is an FP->Int bitcast and if the sign bit is the only | ||||||
| 1898 | // thing demanded, turn this into a FGETSIGN. | ||||||
| 1899 | if (!TLO.LegalOperations() && !VT.isVector() && !SrcVT.isVector() && | ||||||
| 1900 | DemandedBits == APInt::getSignMask(Op.getValueSizeInBits()) && | ||||||
| 1901 | SrcVT.isFloatingPoint()) { | ||||||
| 1902 | bool OpVTLegal = isOperationLegalOrCustom(ISD::FGETSIGN, VT); | ||||||
| 1903 | bool i32Legal = isOperationLegalOrCustom(ISD::FGETSIGN, MVT::i32); | ||||||
| 1904 | if ((OpVTLegal || i32Legal) && VT.isSimple() && SrcVT != MVT::f16 && | ||||||
| 1905 | SrcVT != MVT::f128) { | ||||||
| 1906 | // Cannot eliminate/lower SHL for f128 yet. | ||||||
| 1907 | EVT Ty = OpVTLegal ? VT : MVT::i32; | ||||||
| 1908 | // Make a FGETSIGN + SHL to move the sign bit into the appropriate | ||||||
| 1909 | // place. We expect the SHL to be eliminated by other optimizations. | ||||||
| 1910 | SDValue Sign = TLO.DAG.getNode(ISD::FGETSIGN, dl, Ty, Src); | ||||||
| 1911 | unsigned OpVTSizeInBits = Op.getValueSizeInBits(); | ||||||
| 1912 | if (!OpVTLegal && OpVTSizeInBits > 32) | ||||||
| 1913 | Sign = TLO.DAG.getNode(ISD::ZERO_EXTEND, dl, VT, Sign); | ||||||
| 1914 | unsigned ShVal = Op.getValueSizeInBits() - 1; | ||||||
| 1915 | SDValue ShAmt = TLO.DAG.getConstant(ShVal, dl, VT); | ||||||
| 1916 | return TLO.CombineTo(Op, | ||||||
| 1917 | TLO.DAG.getNode(ISD::SHL, dl, VT, Sign, ShAmt)); | ||||||
| 1918 | } | ||||||
| 1919 | } | ||||||
| 1920 | |||||||
| 1921 | // Bitcast from a vector using SimplifyDemanded Bits/VectorElts. | ||||||
| 1922 | // Demand the elt/bit if any of the original elts/bits are demanded. | ||||||
| 1923 | // TODO - bigendian once we have test coverage. | ||||||
| 1924 | if (SrcVT.isVector() && (BitWidth % NumSrcEltBits) == 0 && | ||||||
| 1925 | TLO.DAG.getDataLayout().isLittleEndian()) { | ||||||
| 1926 | unsigned Scale = BitWidth / NumSrcEltBits; | ||||||
| 1927 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | ||||||
| 1928 | APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits); | ||||||
| 1929 | APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts); | ||||||
| 1930 | for (unsigned i = 0; i != Scale; ++i) { | ||||||
| 1931 | unsigned Offset = i * NumSrcEltBits; | ||||||
| 1932 | APInt Sub = DemandedBits.extractBits(NumSrcEltBits, Offset); | ||||||
| 1933 | if (!Sub.isNullValue()) { | ||||||
| 1934 | DemandedSrcBits |= Sub; | ||||||
| 1935 | for (unsigned j = 0; j != NumElts; ++j) | ||||||
| 1936 | if (DemandedElts[j]) | ||||||
| 1937 | DemandedSrcElts.setBit((j * Scale) + i); | ||||||
| 1938 | } | ||||||
| 1939 | } | ||||||
| 1940 | |||||||
| 1941 | APInt KnownSrcUndef, KnownSrcZero; | ||||||
| 1942 | if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef, | ||||||
| 1943 | KnownSrcZero, TLO, Depth + 1)) | ||||||
| 1944 | return true; | ||||||
| 1945 | |||||||
| 1946 | KnownBits KnownSrcBits; | ||||||
| 1947 | if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, | ||||||
| 1948 | KnownSrcBits, TLO, Depth + 1)) | ||||||
| 1949 | return true; | ||||||
| 1950 | } else if ((NumSrcEltBits % BitWidth) == 0 && | ||||||
| 1951 | TLO.DAG.getDataLayout().isLittleEndian()) { | ||||||
| 1952 | unsigned Scale = NumSrcEltBits / BitWidth; | ||||||
| 1953 | unsigned NumSrcElts = SrcVT.isVector() ? SrcVT.getVectorNumElements() : 1; | ||||||
| 1954 | APInt DemandedSrcBits = APInt::getNullValue(NumSrcEltBits); | ||||||
| 1955 | APInt DemandedSrcElts = APInt::getNullValue(NumSrcElts); | ||||||
| 1956 | for (unsigned i = 0; i != NumElts; ++i) | ||||||
| 1957 | if (DemandedElts[i]) { | ||||||
| 1958 | unsigned Offset = (i % Scale) * BitWidth; | ||||||
| 1959 | DemandedSrcBits.insertBits(DemandedBits, Offset); | ||||||
| 1960 | DemandedSrcElts.setBit(i / Scale); | ||||||
| 1961 | } | ||||||
| 1962 | |||||||
| 1963 | if (SrcVT.isVector()) { | ||||||
| 1964 | APInt KnownSrcUndef, KnownSrcZero; | ||||||
| 1965 | if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, KnownSrcUndef, | ||||||
| 1966 | KnownSrcZero, TLO, Depth + 1)) | ||||||
| 1967 | return true; | ||||||
| 1968 | } | ||||||
| 1969 | |||||||
| 1970 | KnownBits KnownSrcBits; | ||||||
| 1971 | if (SimplifyDemandedBits(Src, DemandedSrcBits, DemandedSrcElts, | ||||||
| 1972 | KnownSrcBits, TLO, Depth + 1)) | ||||||
| 1973 | return true; | ||||||
| 1974 | } | ||||||
| 1975 | |||||||
| 1976 | // If this is a bitcast, let computeKnownBits handle it. Only do this on a | ||||||
| 1977 | // recursive call where Known may be useful to the caller. | ||||||
| 1978 | if (Depth > 0) { | ||||||
| 1979 | Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); | ||||||
| 1980 | return false; | ||||||
| 1981 | } | ||||||
| 1982 | break; | ||||||
| 1983 | } | ||||||
| 1984 | case ISD::ADD: | ||||||
| 1985 | case ISD::MUL: | ||||||
| 1986 | case ISD::SUB: { | ||||||
| 1987 | // Add, Sub, and Mul don't demand any bits in positions beyond that | ||||||
| 1988 | // of the highest bit demanded of them. | ||||||
| 1989 | SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1); | ||||||
| 1990 | SDNodeFlags Flags = Op.getNode()->getFlags(); | ||||||
| 1991 | unsigned DemandedBitsLZ = DemandedBits.countLeadingZeros(); | ||||||
| 1992 | APInt LoMask = APInt::getLowBitsSet(BitWidth, BitWidth - DemandedBitsLZ); | ||||||
| 1993 | if (SimplifyDemandedBits(Op0, LoMask, DemandedElts, Known2, TLO, | ||||||
| 1994 | Depth + 1) || | ||||||
| 1995 | SimplifyDemandedBits(Op1, LoMask, DemandedElts, Known2, TLO, | ||||||
| 1996 | Depth + 1) || | ||||||
| 1997 | // See if the operation should be performed at a smaller bit width. | ||||||
| 1998 | ShrinkDemandedOp(Op, BitWidth, DemandedBits, TLO)) { | ||||||
| 1999 | if (Flags.hasNoSignedWrap() || Flags.hasNoUnsignedWrap()) { | ||||||
| 2000 | // Disable the nsw and nuw flags. We can no longer guarantee that we | ||||||
| 2001 | // won't wrap after simplification. | ||||||
| 2002 | Flags.setNoSignedWrap(false); | ||||||
| 2003 | Flags.setNoUnsignedWrap(false); | ||||||
| 2004 | SDValue NewOp = | ||||||
| 2005 | TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1, Flags); | ||||||
| 2006 | return TLO.CombineTo(Op, NewOp); | ||||||
| 2007 | } | ||||||
| 2008 | return true; | ||||||
| 2009 | } | ||||||
| 2010 | |||||||
| 2011 | // Attempt to avoid multi-use ops if we don't need anything from them. | ||||||
| 2012 | if (!LoMask.isAllOnesValue() || !DemandedElts.isAllOnesValue()) { | ||||||
| 2013 | SDValue DemandedOp0 = SimplifyMultipleUseDemandedBits( | ||||||
| 2014 | Op0, LoMask, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 2015 | SDValue DemandedOp1 = SimplifyMultipleUseDemandedBits( | ||||||
| 2016 | Op1, LoMask, DemandedElts, TLO.DAG, Depth + 1); | ||||||
| 2017 | if (DemandedOp0 || DemandedOp1) { | ||||||
| 2018 | Flags.setNoSignedWrap(false); | ||||||
| 2019 | Flags.setNoUnsignedWrap(false); | ||||||
| 2020 | Op0 = DemandedOp0 ? DemandedOp0 : Op0; | ||||||
| 2021 | Op1 = DemandedOp1 ? DemandedOp1 : Op1; | ||||||
| 2022 | SDValue NewOp = | ||||||
| 2023 | TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Op1, Flags); | ||||||
| 2024 | return TLO.CombineTo(Op, NewOp); | ||||||
| 2025 | } | ||||||
| 2026 | } | ||||||
| 2027 | |||||||
| 2028 | // If we have a constant operand, we may be able to turn it into -1 if we | ||||||
| 2029 | // do not demand the high bits. This can make the constant smaller to | ||||||
| 2030 | // encode, allow more general folding, or match specialized instruction | ||||||
| 2031 | // patterns (eg, 'blsr' on x86). Don't bother changing 1 to -1 because that | ||||||
| 2032 | // is probably not useful (and could be detrimental). | ||||||
| 2033 | ConstantSDNode *C = isConstOrConstSplat(Op1); | ||||||
| 2034 | APInt HighMask = APInt::getHighBitsSet(BitWidth, DemandedBitsLZ); | ||||||
| 2035 | if (C && !C->isAllOnesValue() && !C->isOne() && | ||||||
| 2036 | (C->getAPIntValue() | HighMask).isAllOnesValue()) { | ||||||
| 2037 | SDValue Neg1 = TLO.DAG.getAllOnesConstant(dl, VT); | ||||||
| 2038 | // Disable the nsw and nuw flags. We can no longer guarantee that we | ||||||
| 2039 | // won't wrap after simplification. | ||||||
| 2040 | Flags.setNoSignedWrap(false); | ||||||
| 2041 | Flags.setNoUnsignedWrap(false); | ||||||
| 2042 | SDValue NewOp = TLO.DAG.getNode(Op.getOpcode(), dl, VT, Op0, Neg1, Flags); | ||||||
| 2043 | return TLO.CombineTo(Op, NewOp); | ||||||
| 2044 | } | ||||||
| 2045 | |||||||
| 2046 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 2047 | } | ||||||
| 2048 | default: | ||||||
| 2049 | if (Op.getOpcode() >= ISD::BUILTIN_OP_END) { | ||||||
| 2050 | if (SimplifyDemandedBitsForTargetNode(Op, DemandedBits, DemandedElts, | ||||||
| 2051 | Known, TLO, Depth)) | ||||||
| 2052 | return true; | ||||||
| 2053 | break; | ||||||
| 2054 | } | ||||||
| 2055 | |||||||
| 2056 | // Just use computeKnownBits to compute output bits. | ||||||
| 2057 | Known = TLO.DAG.computeKnownBits(Op, DemandedElts, Depth); | ||||||
| 2058 | break; | ||||||
| 2059 | } | ||||||
| 2060 | |||||||
| 2061 | // If we know the value of all of the demanded bits, return this as a | ||||||
| 2062 | // constant. | ||||||
| 2063 | if (DemandedBits.isSubsetOf(Known.Zero | Known.One)) { | ||||||
| 2064 | // Avoid folding to a constant if any OpaqueConstant is involved. | ||||||
| 2065 | const SDNode *N = Op.getNode(); | ||||||
| 2066 | for (SDNodeIterator I = SDNodeIterator::begin(N), | ||||||
| 2067 | E = SDNodeIterator::end(N); | ||||||
| 2068 | I != E; ++I) { | ||||||
| 2069 | SDNode *Op = *I; | ||||||
| 2070 | if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) | ||||||
| 2071 | if (C->isOpaque()) | ||||||
| 2072 | return false; | ||||||
| 2073 | } | ||||||
| 2074 | // TODO: Handle float bits as well. | ||||||
| 2075 | if (VT.isInteger()) | ||||||
| 2076 | return TLO.CombineTo(Op, TLO.DAG.getConstant(Known.One, dl, VT)); | ||||||
| 2077 | } | ||||||
| 2078 | |||||||
| 2079 | return false; | ||||||
| 2080 | } | ||||||
| 2081 | |||||||
| 2082 | bool TargetLowering::SimplifyDemandedVectorElts(SDValue Op, | ||||||
| 2083 | const APInt &DemandedElts, | ||||||
| 2084 | APInt &KnownUndef, | ||||||
| 2085 | APInt &KnownZero, | ||||||
| 2086 | DAGCombinerInfo &DCI) const { | ||||||
| 2087 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 2088 | TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(), | ||||||
| 2089 | !DCI.isBeforeLegalizeOps()); | ||||||
| 2090 | |||||||
| 2091 | bool Simplified = | ||||||
| 2092 | SimplifyDemandedVectorElts(Op, DemandedElts, KnownUndef, KnownZero, TLO); | ||||||
| 2093 | if (Simplified) { | ||||||
| 2094 | DCI.AddToWorklist(Op.getNode()); | ||||||
| 2095 | DCI.CommitTargetLoweringOpt(TLO); | ||||||
| 2096 | } | ||||||
| 2097 | |||||||
| 2098 | return Simplified; | ||||||
| 2099 | } | ||||||
| 2100 | |||||||
| 2101 | /// Given a vector binary operation and known undefined elements for each input | ||||||
| 2102 | /// operand, compute whether each element of the output is undefined. | ||||||
| 2103 | static APInt getKnownUndefForVectorBinop(SDValue BO, SelectionDAG &DAG, | ||||||
| 2104 | const APInt &UndefOp0, | ||||||
| 2105 | const APInt &UndefOp1) { | ||||||
| 2106 | EVT VT = BO.getValueType(); | ||||||
| 2107 | assert(DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() &&((DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() && "Vector binop only") ? static_cast< void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() && \"Vector binop only\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2108, __PRETTY_FUNCTION__)) | ||||||
| 2108 | "Vector binop only")((DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() && "Vector binop only") ? static_cast< void> (0) : __assert_fail ("DAG.getTargetLoweringInfo().isBinOp(BO.getOpcode()) && VT.isVector() && \"Vector binop only\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2108, __PRETTY_FUNCTION__)); | ||||||
| 2109 | |||||||
| 2110 | EVT EltVT = VT.getVectorElementType(); | ||||||
| 2111 | unsigned NumElts = VT.getVectorNumElements(); | ||||||
| 2112 | assert(UndefOp0.getBitWidth() == NumElts &&((UndefOp0.getBitWidth() == NumElts && UndefOp1.getBitWidth () == NumElts && "Bad type for undef analysis") ? static_cast <void> (0) : __assert_fail ("UndefOp0.getBitWidth() == NumElts && UndefOp1.getBitWidth() == NumElts && \"Bad type for undef analysis\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2113, __PRETTY_FUNCTION__)) | ||||||
| 2113 | UndefOp1.getBitWidth() == NumElts && "Bad type for undef analysis")((UndefOp0.getBitWidth() == NumElts && UndefOp1.getBitWidth () == NumElts && "Bad type for undef analysis") ? static_cast <void> (0) : __assert_fail ("UndefOp0.getBitWidth() == NumElts && UndefOp1.getBitWidth() == NumElts && \"Bad type for undef analysis\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2113, __PRETTY_FUNCTION__)); | ||||||
| 2114 | |||||||
| 2115 | auto getUndefOrConstantElt = [&](SDValue V, unsigned Index, | ||||||
| 2116 | const APInt &UndefVals) { | ||||||
| 2117 | if (UndefVals[Index]) | ||||||
| 2118 | return DAG.getUNDEF(EltVT); | ||||||
| 2119 | |||||||
| 2120 | if (auto *BV = dyn_cast<BuildVectorSDNode>(V)) { | ||||||
| 2121 | // Try hard to make sure that the getNode() call is not creating temporary | ||||||
| 2122 | // nodes. Ignore opaque integers because they do not constant fold. | ||||||
| 2123 | SDValue Elt = BV->getOperand(Index); | ||||||
| 2124 | auto *C = dyn_cast<ConstantSDNode>(Elt); | ||||||
| 2125 | if (isa<ConstantFPSDNode>(Elt) || Elt.isUndef() || (C && !C->isOpaque())) | ||||||
| 2126 | return Elt; | ||||||
| 2127 | } | ||||||
| 2128 | |||||||
| 2129 | return SDValue(); | ||||||
| 2130 | }; | ||||||
| 2131 | |||||||
| 2132 | APInt KnownUndef = APInt::getNullValue(NumElts); | ||||||
| 2133 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2134 | // If both inputs for this element are either constant or undef and match | ||||||
| 2135 | // the element type, compute the constant/undef result for this element of | ||||||
| 2136 | // the vector. | ||||||
| 2137 | // TODO: Ideally we would use FoldConstantArithmetic() here, but that does | ||||||
| 2138 | // not handle FP constants. The code within getNode() should be refactored | ||||||
| 2139 | // to avoid the danger of creating a bogus temporary node here. | ||||||
| 2140 | SDValue C0 = getUndefOrConstantElt(BO.getOperand(0), i, UndefOp0); | ||||||
| 2141 | SDValue C1 = getUndefOrConstantElt(BO.getOperand(1), i, UndefOp1); | ||||||
| 2142 | if (C0 && C1 && C0.getValueType() == EltVT && C1.getValueType() == EltVT) | ||||||
| 2143 | if (DAG.getNode(BO.getOpcode(), SDLoc(BO), EltVT, C0, C1).isUndef()) | ||||||
| 2144 | KnownUndef.setBit(i); | ||||||
| 2145 | } | ||||||
| 2146 | return KnownUndef; | ||||||
| 2147 | } | ||||||
| 2148 | |||||||
| 2149 | bool TargetLowering::SimplifyDemandedVectorElts( | ||||||
| 2150 | SDValue Op, const APInt &OriginalDemandedElts, APInt &KnownUndef, | ||||||
| 2151 | APInt &KnownZero, TargetLoweringOpt &TLO, unsigned Depth, | ||||||
| 2152 | bool AssumeSingleUse) const { | ||||||
| 2153 | EVT VT = Op.getValueType(); | ||||||
| 2154 | APInt DemandedElts = OriginalDemandedElts; | ||||||
| 2155 | unsigned NumElts = DemandedElts.getBitWidth(); | ||||||
| 2156 | assert(VT.isVector() && "Expected vector op")((VT.isVector() && "Expected vector op") ? static_cast <void> (0) : __assert_fail ("VT.isVector() && \"Expected vector op\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2156, __PRETTY_FUNCTION__)); | ||||||
| 2157 | assert(VT.getVectorNumElements() == NumElts &&((VT.getVectorNumElements() == NumElts && "Mask size mismatches value type element count!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == NumElts && \"Mask size mismatches value type element count!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2158, __PRETTY_FUNCTION__)) | ||||||
| 2158 | "Mask size mismatches value type element count!")((VT.getVectorNumElements() == NumElts && "Mask size mismatches value type element count!" ) ? static_cast<void> (0) : __assert_fail ("VT.getVectorNumElements() == NumElts && \"Mask size mismatches value type element count!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2158, __PRETTY_FUNCTION__)); | ||||||
| 2159 | |||||||
| 2160 | KnownUndef = KnownZero = APInt::getNullValue(NumElts); | ||||||
| 2161 | |||||||
| 2162 | // Undef operand. | ||||||
| 2163 | if (Op.isUndef()) { | ||||||
| 2164 | KnownUndef.setAllBits(); | ||||||
| 2165 | return false; | ||||||
| 2166 | } | ||||||
| 2167 | |||||||
| 2168 | // If Op has other users, assume that all elements are needed. | ||||||
| 2169 | if (!Op.getNode()->hasOneUse() && !AssumeSingleUse) | ||||||
| 2170 | DemandedElts.setAllBits(); | ||||||
| 2171 | |||||||
| 2172 | // Not demanding any elements from Op. | ||||||
| 2173 | if (DemandedElts == 0) { | ||||||
| 2174 | KnownUndef.setAllBits(); | ||||||
| 2175 | return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); | ||||||
| 2176 | } | ||||||
| 2177 | |||||||
| 2178 | // Limit search depth. | ||||||
| 2179 | if (Depth >= SelectionDAG::MaxRecursionDepth) | ||||||
| 2180 | return false; | ||||||
| 2181 | |||||||
| 2182 | SDLoc DL(Op); | ||||||
| 2183 | unsigned EltSizeInBits = VT.getScalarSizeInBits(); | ||||||
| 2184 | |||||||
| 2185 | switch (Op.getOpcode()) { | ||||||
| 2186 | case ISD::SCALAR_TO_VECTOR: { | ||||||
| 2187 | if (!DemandedElts[0]) { | ||||||
| 2188 | KnownUndef.setAllBits(); | ||||||
| 2189 | return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); | ||||||
| 2190 | } | ||||||
| 2191 | KnownUndef.setHighBits(NumElts - 1); | ||||||
| 2192 | break; | ||||||
| 2193 | } | ||||||
| 2194 | case ISD::BITCAST: { | ||||||
| 2195 | SDValue Src = Op.getOperand(0); | ||||||
| 2196 | EVT SrcVT = Src.getValueType(); | ||||||
| 2197 | |||||||
| 2198 | // We only handle vectors here. | ||||||
| 2199 | // TODO - investigate calling SimplifyDemandedBits/ComputeKnownBits? | ||||||
| 2200 | if (!SrcVT.isVector()) | ||||||
| 2201 | break; | ||||||
| 2202 | |||||||
| 2203 | // Fast handling of 'identity' bitcasts. | ||||||
| 2204 | unsigned NumSrcElts = SrcVT.getVectorNumElements(); | ||||||
| 2205 | if (NumSrcElts == NumElts) | ||||||
| 2206 | return SimplifyDemandedVectorElts(Src, DemandedElts, KnownUndef, | ||||||
| 2207 | KnownZero, TLO, Depth + 1); | ||||||
| 2208 | |||||||
| 2209 | APInt SrcZero, SrcUndef; | ||||||
| 2210 | APInt SrcDemandedElts = APInt::getNullValue(NumSrcElts); | ||||||
| 2211 | |||||||
| 2212 | // Bitcast from 'large element' src vector to 'small element' vector, we | ||||||
| 2213 | // must demand a source element if any DemandedElt maps to it. | ||||||
| 2214 | if ((NumElts % NumSrcElts) == 0) { | ||||||
| 2215 | unsigned Scale = NumElts / NumSrcElts; | ||||||
| 2216 | for (unsigned i = 0; i != NumElts; ++i) | ||||||
| 2217 | if (DemandedElts[i]) | ||||||
| 2218 | SrcDemandedElts.setBit(i / Scale); | ||||||
| 2219 | |||||||
| 2220 | if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero, | ||||||
| 2221 | TLO, Depth + 1)) | ||||||
| 2222 | return true; | ||||||
| 2223 | |||||||
| 2224 | // Try calling SimplifyDemandedBits, converting demanded elts to the bits | ||||||
| 2225 | // of the large element. | ||||||
| 2226 | // TODO - bigendian once we have test coverage. | ||||||
| 2227 | if (TLO.DAG.getDataLayout().isLittleEndian()) { | ||||||
| 2228 | unsigned SrcEltSizeInBits = SrcVT.getScalarSizeInBits(); | ||||||
| 2229 | APInt SrcDemandedBits = APInt::getNullValue(SrcEltSizeInBits); | ||||||
| 2230 | for (unsigned i = 0; i != NumElts; ++i) | ||||||
| 2231 | if (DemandedElts[i]) { | ||||||
| 2232 | unsigned Ofs = (i % Scale) * EltSizeInBits; | ||||||
| 2233 | SrcDemandedBits.setBits(Ofs, Ofs + EltSizeInBits); | ||||||
| 2234 | } | ||||||
| 2235 | |||||||
| 2236 | KnownBits Known; | ||||||
| 2237 | if (SimplifyDemandedBits(Src, SrcDemandedBits, Known, TLO, Depth + 1)) | ||||||
| 2238 | return true; | ||||||
| 2239 | } | ||||||
| 2240 | |||||||
| 2241 | // If the src element is zero/undef then all the output elements will be - | ||||||
| 2242 | // only demanded elements are guaranteed to be correct. | ||||||
| 2243 | for (unsigned i = 0; i != NumSrcElts; ++i) { | ||||||
| 2244 | if (SrcDemandedElts[i]) { | ||||||
| 2245 | if (SrcZero[i]) | ||||||
| 2246 | KnownZero.setBits(i * Scale, (i + 1) * Scale); | ||||||
| 2247 | if (SrcUndef[i]) | ||||||
| 2248 | KnownUndef.setBits(i * Scale, (i + 1) * Scale); | ||||||
| 2249 | } | ||||||
| 2250 | } | ||||||
| 2251 | } | ||||||
| 2252 | |||||||
| 2253 | // Bitcast from 'small element' src vector to 'large element' vector, we | ||||||
| 2254 | // demand all smaller source elements covered by the larger demanded element | ||||||
| 2255 | // of this vector. | ||||||
| 2256 | if ((NumSrcElts % NumElts) == 0) { | ||||||
| 2257 | unsigned Scale = NumSrcElts / NumElts; | ||||||
| 2258 | for (unsigned i = 0; i != NumElts; ++i) | ||||||
| 2259 | if (DemandedElts[i]) | ||||||
| 2260 | SrcDemandedElts.setBits(i * Scale, (i + 1) * Scale); | ||||||
| 2261 | |||||||
| 2262 | if (SimplifyDemandedVectorElts(Src, SrcDemandedElts, SrcUndef, SrcZero, | ||||||
| 2263 | TLO, Depth + 1)) | ||||||
| 2264 | return true; | ||||||
| 2265 | |||||||
| 2266 | // If all the src elements covering an output element are zero/undef, then | ||||||
| 2267 | // the output element will be as well, assuming it was demanded. | ||||||
| 2268 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2269 | if (DemandedElts[i]) { | ||||||
| 2270 | if (SrcZero.extractBits(Scale, i * Scale).isAllOnesValue()) | ||||||
| 2271 | KnownZero.setBit(i); | ||||||
| 2272 | if (SrcUndef.extractBits(Scale, i * Scale).isAllOnesValue()) | ||||||
| 2273 | KnownUndef.setBit(i); | ||||||
| 2274 | } | ||||||
| 2275 | } | ||||||
| 2276 | } | ||||||
| 2277 | break; | ||||||
| 2278 | } | ||||||
| 2279 | case ISD::BUILD_VECTOR: { | ||||||
| 2280 | // Check all elements and simplify any unused elements with UNDEF. | ||||||
| 2281 | if (!DemandedElts.isAllOnesValue()) { | ||||||
| 2282 | // Don't simplify BROADCASTS. | ||||||
| 2283 | if (llvm::any_of(Op->op_values(), | ||||||
| 2284 | [&](SDValue Elt) { return Op.getOperand(0) != Elt; })) { | ||||||
| 2285 | SmallVector<SDValue, 32> Ops(Op->op_begin(), Op->op_end()); | ||||||
| 2286 | bool Updated = false; | ||||||
| 2287 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2288 | if (!DemandedElts[i] && !Ops[i].isUndef()) { | ||||||
| 2289 | Ops[i] = TLO.DAG.getUNDEF(Ops[0].getValueType()); | ||||||
| 2290 | KnownUndef.setBit(i); | ||||||
| 2291 | Updated = true; | ||||||
| 2292 | } | ||||||
| 2293 | } | ||||||
| 2294 | if (Updated) | ||||||
| 2295 | return TLO.CombineTo(Op, TLO.DAG.getBuildVector(VT, DL, Ops)); | ||||||
| 2296 | } | ||||||
| 2297 | } | ||||||
| 2298 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2299 | SDValue SrcOp = Op.getOperand(i); | ||||||
| 2300 | if (SrcOp.isUndef()) { | ||||||
| 2301 | KnownUndef.setBit(i); | ||||||
| 2302 | } else if (EltSizeInBits == SrcOp.getScalarValueSizeInBits() && | ||||||
| 2303 | (isNullConstant(SrcOp) || isNullFPConstant(SrcOp))) { | ||||||
| 2304 | KnownZero.setBit(i); | ||||||
| 2305 | } | ||||||
| 2306 | } | ||||||
| 2307 | break; | ||||||
| 2308 | } | ||||||
| 2309 | case ISD::CONCAT_VECTORS: { | ||||||
| 2310 | EVT SubVT = Op.getOperand(0).getValueType(); | ||||||
| 2311 | unsigned NumSubVecs = Op.getNumOperands(); | ||||||
| 2312 | unsigned NumSubElts = SubVT.getVectorNumElements(); | ||||||
| 2313 | for (unsigned i = 0; i != NumSubVecs; ++i) { | ||||||
| 2314 | SDValue SubOp = Op.getOperand(i); | ||||||
| 2315 | APInt SubElts = DemandedElts.extractBits(NumSubElts, i * NumSubElts); | ||||||
| 2316 | APInt SubUndef, SubZero; | ||||||
| 2317 | if (SimplifyDemandedVectorElts(SubOp, SubElts, SubUndef, SubZero, TLO, | ||||||
| 2318 | Depth + 1)) | ||||||
| 2319 | return true; | ||||||
| 2320 | KnownUndef.insertBits(SubUndef, i * NumSubElts); | ||||||
| 2321 | KnownZero.insertBits(SubZero, i * NumSubElts); | ||||||
| 2322 | } | ||||||
| 2323 | break; | ||||||
| 2324 | } | ||||||
| 2325 | case ISD::INSERT_SUBVECTOR: { | ||||||
| 2326 | if (!isa<ConstantSDNode>(Op.getOperand(2))) | ||||||
| 2327 | break; | ||||||
| 2328 | SDValue Base = Op.getOperand(0); | ||||||
| 2329 | SDValue Sub = Op.getOperand(1); | ||||||
| 2330 | EVT SubVT = Sub.getValueType(); | ||||||
| 2331 | unsigned NumSubElts = SubVT.getVectorNumElements(); | ||||||
| 2332 | const APInt &Idx = Op.getConstantOperandAPInt(2); | ||||||
| 2333 | if (Idx.ugt(NumElts - NumSubElts)) | ||||||
| 2334 | break; | ||||||
| 2335 | unsigned SubIdx = Idx.getZExtValue(); | ||||||
| 2336 | APInt SubElts = DemandedElts.extractBits(NumSubElts, SubIdx); | ||||||
| 2337 | APInt SubUndef, SubZero; | ||||||
| 2338 | if (SimplifyDemandedVectorElts(Sub, SubElts, SubUndef, SubZero, TLO, | ||||||
| 2339 | Depth + 1)) | ||||||
| 2340 | return true; | ||||||
| 2341 | APInt BaseElts = DemandedElts; | ||||||
| 2342 | BaseElts.insertBits(APInt::getNullValue(NumSubElts), SubIdx); | ||||||
| 2343 | |||||||
| 2344 | // If none of the base operand elements are demanded, replace it with undef. | ||||||
| 2345 | if (!BaseElts && !Base.isUndef()) | ||||||
| 2346 | return TLO.CombineTo(Op, | ||||||
| 2347 | TLO.DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VT, | ||||||
| 2348 | TLO.DAG.getUNDEF(VT), | ||||||
| 2349 | Op.getOperand(1), | ||||||
| 2350 | Op.getOperand(2))); | ||||||
| 2351 | |||||||
| 2352 | if (SimplifyDemandedVectorElts(Base, BaseElts, KnownUndef, KnownZero, TLO, | ||||||
| 2353 | Depth + 1)) | ||||||
| 2354 | return true; | ||||||
| 2355 | KnownUndef.insertBits(SubUndef, SubIdx); | ||||||
| 2356 | KnownZero.insertBits(SubZero, SubIdx); | ||||||
| 2357 | break; | ||||||
| 2358 | } | ||||||
| 2359 | case ISD::EXTRACT_SUBVECTOR: { | ||||||
| 2360 | SDValue Src = Op.getOperand(0); | ||||||
| 2361 | ConstantSDNode *SubIdx = dyn_cast<ConstantSDNode>(Op.getOperand(1)); | ||||||
| 2362 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); | ||||||
| 2363 | if (SubIdx && SubIdx->getAPIntValue().ule(NumSrcElts - NumElts)) { | ||||||
| 2364 | // Offset the demanded elts by the subvector index. | ||||||
| 2365 | uint64_t Idx = SubIdx->getZExtValue(); | ||||||
| 2366 | APInt SrcElts = DemandedElts.zextOrSelf(NumSrcElts).shl(Idx); | ||||||
| 2367 | APInt SrcUndef, SrcZero; | ||||||
| 2368 | if (SimplifyDemandedVectorElts(Src, SrcElts, SrcUndef, SrcZero, TLO, | ||||||
| 2369 | Depth + 1)) | ||||||
| 2370 | return true; | ||||||
| 2371 | KnownUndef = SrcUndef.extractBits(NumElts, Idx); | ||||||
| 2372 | KnownZero = SrcZero.extractBits(NumElts, Idx); | ||||||
| 2373 | } | ||||||
| 2374 | break; | ||||||
| 2375 | } | ||||||
| 2376 | case ISD::INSERT_VECTOR_ELT: { | ||||||
| 2377 | SDValue Vec = Op.getOperand(0); | ||||||
| 2378 | SDValue Scl = Op.getOperand(1); | ||||||
| 2379 | auto *CIdx = dyn_cast<ConstantSDNode>(Op.getOperand(2)); | ||||||
| 2380 | |||||||
| 2381 | // For a legal, constant insertion index, if we don't need this insertion | ||||||
| 2382 | // then strip it, else remove it from the demanded elts. | ||||||
| 2383 | if (CIdx && CIdx->getAPIntValue().ult(NumElts)) { | ||||||
| 2384 | unsigned Idx = CIdx->getZExtValue(); | ||||||
| 2385 | if (!DemandedElts[Idx]) | ||||||
| 2386 | return TLO.CombineTo(Op, Vec); | ||||||
| 2387 | |||||||
| 2388 | APInt DemandedVecElts(DemandedElts); | ||||||
| 2389 | DemandedVecElts.clearBit(Idx); | ||||||
| 2390 | if (SimplifyDemandedVectorElts(Vec, DemandedVecElts, KnownUndef, | ||||||
| 2391 | KnownZero, TLO, Depth + 1)) | ||||||
| 2392 | return true; | ||||||
| 2393 | |||||||
| 2394 | KnownUndef.clearBit(Idx); | ||||||
| 2395 | if (Scl.isUndef()) | ||||||
| 2396 | KnownUndef.setBit(Idx); | ||||||
| 2397 | |||||||
| 2398 | KnownZero.clearBit(Idx); | ||||||
| 2399 | if (isNullConstant(Scl) || isNullFPConstant(Scl)) | ||||||
| 2400 | KnownZero.setBit(Idx); | ||||||
| 2401 | break; | ||||||
| 2402 | } | ||||||
| 2403 | |||||||
| 2404 | APInt VecUndef, VecZero; | ||||||
| 2405 | if (SimplifyDemandedVectorElts(Vec, DemandedElts, VecUndef, VecZero, TLO, | ||||||
| 2406 | Depth + 1)) | ||||||
| 2407 | return true; | ||||||
| 2408 | // Without knowing the insertion index we can't set KnownUndef/KnownZero. | ||||||
| 2409 | break; | ||||||
| 2410 | } | ||||||
| 2411 | case ISD::VSELECT: { | ||||||
| 2412 | // Try to transform the select condition based on the current demanded | ||||||
| 2413 | // elements. | ||||||
| 2414 | // TODO: If a condition element is undef, we can choose from one arm of the | ||||||
| 2415 | // select (and if one arm is undef, then we can propagate that to the | ||||||
| 2416 | // result). | ||||||
| 2417 | // TODO - add support for constant vselect masks (see IR version of this). | ||||||
| 2418 | APInt UnusedUndef, UnusedZero; | ||||||
| 2419 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, UnusedUndef, | ||||||
| 2420 | UnusedZero, TLO, Depth + 1)) | ||||||
| 2421 | return true; | ||||||
| 2422 | |||||||
| 2423 | // See if we can simplify either vselect operand. | ||||||
| 2424 | APInt DemandedLHS(DemandedElts); | ||||||
| 2425 | APInt DemandedRHS(DemandedElts); | ||||||
| 2426 | APInt UndefLHS, ZeroLHS; | ||||||
| 2427 | APInt UndefRHS, ZeroRHS; | ||||||
| 2428 | if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedLHS, UndefLHS, | ||||||
| 2429 | ZeroLHS, TLO, Depth + 1)) | ||||||
| 2430 | return true; | ||||||
| 2431 | if (SimplifyDemandedVectorElts(Op.getOperand(2), DemandedRHS, UndefRHS, | ||||||
| 2432 | ZeroRHS, TLO, Depth + 1)) | ||||||
| 2433 | return true; | ||||||
| 2434 | |||||||
| 2435 | KnownUndef = UndefLHS & UndefRHS; | ||||||
| 2436 | KnownZero = ZeroLHS & ZeroRHS; | ||||||
| 2437 | break; | ||||||
| 2438 | } | ||||||
| 2439 | case ISD::VECTOR_SHUFFLE: { | ||||||
| 2440 | ArrayRef<int> ShuffleMask = cast<ShuffleVectorSDNode>(Op)->getMask(); | ||||||
| 2441 | |||||||
| 2442 | // Collect demanded elements from shuffle operands.. | ||||||
| 2443 | APInt DemandedLHS(NumElts, 0); | ||||||
| 2444 | APInt DemandedRHS(NumElts, 0); | ||||||
| 2445 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2446 | int M = ShuffleMask[i]; | ||||||
| 2447 | if (M < 0 || !DemandedElts[i]) | ||||||
| 2448 | continue; | ||||||
| 2449 | assert(0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range")((0 <= M && M < (int)(2 * NumElts) && "Shuffle index out of range" ) ? static_cast<void> (0) : __assert_fail ("0 <= M && M < (int)(2 * NumElts) && \"Shuffle index out of range\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2449, __PRETTY_FUNCTION__)); | ||||||
| 2450 | if (M < (int)NumElts) | ||||||
| 2451 | DemandedLHS.setBit(M); | ||||||
| 2452 | else | ||||||
| 2453 | DemandedRHS.setBit(M - NumElts); | ||||||
| 2454 | } | ||||||
| 2455 | |||||||
| 2456 | // See if we can simplify either shuffle operand. | ||||||
| 2457 | APInt UndefLHS, ZeroLHS; | ||||||
| 2458 | APInt UndefRHS, ZeroRHS; | ||||||
| 2459 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedLHS, UndefLHS, | ||||||
| 2460 | ZeroLHS, TLO, Depth + 1)) | ||||||
| 2461 | return true; | ||||||
| 2462 | if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedRHS, UndefRHS, | ||||||
| 2463 | ZeroRHS, TLO, Depth + 1)) | ||||||
| 2464 | return true; | ||||||
| 2465 | |||||||
| 2466 | // Simplify mask using undef elements from LHS/RHS. | ||||||
| 2467 | bool Updated = false; | ||||||
| 2468 | bool IdentityLHS = true, IdentityRHS = true; | ||||||
| 2469 | SmallVector<int, 32> NewMask(ShuffleMask.begin(), ShuffleMask.end()); | ||||||
| 2470 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2471 | int &M = NewMask[i]; | ||||||
| 2472 | if (M < 0) | ||||||
| 2473 | continue; | ||||||
| 2474 | if (!DemandedElts[i] || (M < (int)NumElts && UndefLHS[M]) || | ||||||
| 2475 | (M >= (int)NumElts && UndefRHS[M - NumElts])) { | ||||||
| 2476 | Updated = true; | ||||||
| 2477 | M = -1; | ||||||
| 2478 | } | ||||||
| 2479 | IdentityLHS &= (M < 0) || (M == (int)i); | ||||||
| 2480 | IdentityRHS &= (M < 0) || ((M - NumElts) == i); | ||||||
| 2481 | } | ||||||
| 2482 | |||||||
| 2483 | // Update legal shuffle masks based on demanded elements if it won't reduce | ||||||
| 2484 | // to Identity which can cause premature removal of the shuffle mask. | ||||||
| 2485 | if (Updated && !IdentityLHS && !IdentityRHS && !TLO.LegalOps) { | ||||||
| 2486 | SDValue LegalShuffle = | ||||||
| 2487 | buildLegalVectorShuffle(VT, DL, Op.getOperand(0), Op.getOperand(1), | ||||||
| 2488 | NewMask, TLO.DAG); | ||||||
| 2489 | if (LegalShuffle) | ||||||
| 2490 | return TLO.CombineTo(Op, LegalShuffle); | ||||||
| 2491 | } | ||||||
| 2492 | |||||||
| 2493 | // Propagate undef/zero elements from LHS/RHS. | ||||||
| 2494 | for (unsigned i = 0; i != NumElts; ++i) { | ||||||
| 2495 | int M = ShuffleMask[i]; | ||||||
| 2496 | if (M < 0) { | ||||||
| 2497 | KnownUndef.setBit(i); | ||||||
| 2498 | } else if (M < (int)NumElts) { | ||||||
| 2499 | if (UndefLHS[M]) | ||||||
| 2500 | KnownUndef.setBit(i); | ||||||
| 2501 | if (ZeroLHS[M]) | ||||||
| 2502 | KnownZero.setBit(i); | ||||||
| 2503 | } else { | ||||||
| 2504 | if (UndefRHS[M - NumElts]) | ||||||
| 2505 | KnownUndef.setBit(i); | ||||||
| 2506 | if (ZeroRHS[M - NumElts]) | ||||||
| 2507 | KnownZero.setBit(i); | ||||||
| 2508 | } | ||||||
| 2509 | } | ||||||
| 2510 | break; | ||||||
| 2511 | } | ||||||
| 2512 | case ISD::ANY_EXTEND_VECTOR_INREG: | ||||||
| 2513 | case ISD::SIGN_EXTEND_VECTOR_INREG: | ||||||
| 2514 | case ISD::ZERO_EXTEND_VECTOR_INREG: { | ||||||
| 2515 | APInt SrcUndef, SrcZero; | ||||||
| 2516 | SDValue Src = Op.getOperand(0); | ||||||
| 2517 | unsigned NumSrcElts = Src.getValueType().getVectorNumElements(); | ||||||
| 2518 | APInt DemandedSrcElts = DemandedElts.zextOrSelf(NumSrcElts); | ||||||
| 2519 | if (SimplifyDemandedVectorElts(Src, DemandedSrcElts, SrcUndef, SrcZero, TLO, | ||||||
| 2520 | Depth + 1)) | ||||||
| 2521 | return true; | ||||||
| 2522 | KnownZero = SrcZero.zextOrTrunc(NumElts); | ||||||
| 2523 | KnownUndef = SrcUndef.zextOrTrunc(NumElts); | ||||||
| 2524 | |||||||
| 2525 | if (Op.getOpcode() == ISD::ANY_EXTEND_VECTOR_INREG && | ||||||
| 2526 | Op.getValueSizeInBits() == Src.getValueSizeInBits() && | ||||||
| 2527 | DemandedSrcElts == 1 && TLO.DAG.getDataLayout().isLittleEndian()) { | ||||||
| 2528 | // aext - if we just need the bottom element then we can bitcast. | ||||||
| 2529 | return TLO.CombineTo(Op, TLO.DAG.getBitcast(VT, Src)); | ||||||
| 2530 | } | ||||||
| 2531 | |||||||
| 2532 | if (Op.getOpcode() == ISD::ZERO_EXTEND_VECTOR_INREG) { | ||||||
| 2533 | // zext(undef) upper bits are guaranteed to be zero. | ||||||
| 2534 | if (DemandedElts.isSubsetOf(KnownUndef)) | ||||||
| 2535 | return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT)); | ||||||
| 2536 | KnownUndef.clearAllBits(); | ||||||
| 2537 | } | ||||||
| 2538 | break; | ||||||
| 2539 | } | ||||||
| 2540 | |||||||
| 2541 | // TODO: There are more binop opcodes that could be handled here - MUL, MIN, | ||||||
| 2542 | // MAX, saturated math, etc. | ||||||
| 2543 | case ISD::OR: | ||||||
| 2544 | case ISD::XOR: | ||||||
| 2545 | case ISD::ADD: | ||||||
| 2546 | case ISD::SUB: | ||||||
| 2547 | case ISD::FADD: | ||||||
| 2548 | case ISD::FSUB: | ||||||
| 2549 | case ISD::FMUL: | ||||||
| 2550 | case ISD::FDIV: | ||||||
| 2551 | case ISD::FREM: { | ||||||
| 2552 | APInt UndefRHS, ZeroRHS; | ||||||
| 2553 | if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedElts, UndefRHS, | ||||||
| 2554 | ZeroRHS, TLO, Depth + 1)) | ||||||
| 2555 | return true; | ||||||
| 2556 | APInt UndefLHS, ZeroLHS; | ||||||
| 2557 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, UndefLHS, | ||||||
| 2558 | ZeroLHS, TLO, Depth + 1)) | ||||||
| 2559 | return true; | ||||||
| 2560 | |||||||
| 2561 | KnownZero = ZeroLHS & ZeroRHS; | ||||||
| 2562 | KnownUndef = getKnownUndefForVectorBinop(Op, TLO.DAG, UndefLHS, UndefRHS); | ||||||
| 2563 | break; | ||||||
| 2564 | } | ||||||
| 2565 | case ISD::SHL: | ||||||
| 2566 | case ISD::SRL: | ||||||
| 2567 | case ISD::SRA: | ||||||
| 2568 | case ISD::ROTL: | ||||||
| 2569 | case ISD::ROTR: { | ||||||
| 2570 | APInt UndefRHS, ZeroRHS; | ||||||
| 2571 | if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedElts, UndefRHS, | ||||||
| 2572 | ZeroRHS, TLO, Depth + 1)) | ||||||
| 2573 | return true; | ||||||
| 2574 | APInt UndefLHS, ZeroLHS; | ||||||
| 2575 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, UndefLHS, | ||||||
| 2576 | ZeroLHS, TLO, Depth + 1)) | ||||||
| 2577 | return true; | ||||||
| 2578 | |||||||
| 2579 | KnownZero = ZeroLHS; | ||||||
| 2580 | KnownUndef = UndefLHS & UndefRHS; // TODO: use getKnownUndefForVectorBinop? | ||||||
| 2581 | break; | ||||||
| 2582 | } | ||||||
| 2583 | case ISD::MUL: | ||||||
| 2584 | case ISD::AND: { | ||||||
| 2585 | APInt SrcUndef, SrcZero; | ||||||
| 2586 | if (SimplifyDemandedVectorElts(Op.getOperand(1), DemandedElts, SrcUndef, | ||||||
| 2587 | SrcZero, TLO, Depth + 1)) | ||||||
| 2588 | return true; | ||||||
| 2589 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, KnownUndef, | ||||||
| 2590 | KnownZero, TLO, Depth + 1)) | ||||||
| 2591 | return true; | ||||||
| 2592 | |||||||
| 2593 | // If either side has a zero element, then the result element is zero, even | ||||||
| 2594 | // if the other is an UNDEF. | ||||||
| 2595 | // TODO: Extend getKnownUndefForVectorBinop to also deal with known zeros | ||||||
| 2596 | // and then handle 'and' nodes with the rest of the binop opcodes. | ||||||
| 2597 | KnownZero |= SrcZero; | ||||||
| 2598 | KnownUndef &= SrcUndef; | ||||||
| 2599 | KnownUndef &= ~KnownZero; | ||||||
| 2600 | break; | ||||||
| 2601 | } | ||||||
| 2602 | case ISD::TRUNCATE: | ||||||
| 2603 | case ISD::SIGN_EXTEND: | ||||||
| 2604 | case ISD::ZERO_EXTEND: | ||||||
| 2605 | if (SimplifyDemandedVectorElts(Op.getOperand(0), DemandedElts, KnownUndef, | ||||||
| 2606 | KnownZero, TLO, Depth + 1)) | ||||||
| 2607 | return true; | ||||||
| 2608 | |||||||
| 2609 | if (Op.getOpcode() == ISD::ZERO_EXTEND) { | ||||||
| 2610 | // zext(undef) upper bits are guaranteed to be zero. | ||||||
| 2611 | if (DemandedElts.isSubsetOf(KnownUndef)) | ||||||
| 2612 | return TLO.CombineTo(Op, TLO.DAG.getConstant(0, SDLoc(Op), VT)); | ||||||
| 2613 | KnownUndef.clearAllBits(); | ||||||
| 2614 | } | ||||||
| 2615 | break; | ||||||
| 2616 | default: { | ||||||
| 2617 | if (Op.getOpcode() >= ISD::BUILTIN_OP_END) { | ||||||
| 2618 | if (SimplifyDemandedVectorEltsForTargetNode(Op, DemandedElts, KnownUndef, | ||||||
| 2619 | KnownZero, TLO, Depth)) | ||||||
| 2620 | return true; | ||||||
| 2621 | } else { | ||||||
| 2622 | KnownBits Known; | ||||||
| 2623 | APInt DemandedBits = APInt::getAllOnesValue(EltSizeInBits); | ||||||
| 2624 | if (SimplifyDemandedBits(Op, DemandedBits, OriginalDemandedElts, Known, | ||||||
| 2625 | TLO, Depth, AssumeSingleUse)) | ||||||
| 2626 | return true; | ||||||
| 2627 | } | ||||||
| 2628 | break; | ||||||
| 2629 | } | ||||||
| 2630 | } | ||||||
| 2631 | assert((KnownUndef & KnownZero) == 0 && "Elements flagged as undef AND zero")(((KnownUndef & KnownZero) == 0 && "Elements flagged as undef AND zero" ) ? static_cast<void> (0) : __assert_fail ("(KnownUndef & KnownZero) == 0 && \"Elements flagged as undef AND zero\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2631, __PRETTY_FUNCTION__)); | ||||||
| 2632 | |||||||
| 2633 | // Constant fold all undef cases. | ||||||
| 2634 | // TODO: Handle zero cases as well. | ||||||
| 2635 | if (DemandedElts.isSubsetOf(KnownUndef)) | ||||||
| 2636 | return TLO.CombineTo(Op, TLO.DAG.getUNDEF(VT)); | ||||||
| 2637 | |||||||
| 2638 | return false; | ||||||
| 2639 | } | ||||||
| 2640 | |||||||
| 2641 | /// Determine which of the bits specified in Mask are known to be either zero or | ||||||
| 2642 | /// one and return them in the Known. | ||||||
| 2643 | void TargetLowering::computeKnownBitsForTargetNode(const SDValue Op, | ||||||
| 2644 | KnownBits &Known, | ||||||
| 2645 | const APInt &DemandedElts, | ||||||
| 2646 | const SelectionDAG &DAG, | ||||||
| 2647 | unsigned Depth) const { | ||||||
| 2648 | assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)) | ||||||
| 2649 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)) | ||||||
| 2650 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)) | ||||||
| 2651 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)) | ||||||
| 2652 | "Should use MaskedValueIsZero if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)) | ||||||
| 2653 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use MaskedValueIsZero if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use MaskedValueIsZero if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2653, __PRETTY_FUNCTION__)); | ||||||
| 2654 | Known.resetAll(); | ||||||
| 2655 | } | ||||||
| 2656 | |||||||
| 2657 | void TargetLowering::computeKnownBitsForTargetInstr( | ||||||
| 2658 | GISelKnownBits &Analysis, Register R, KnownBits &Known, | ||||||
| 2659 | const APInt &DemandedElts, const MachineRegisterInfo &MRI, | ||||||
| 2660 | unsigned Depth) const { | ||||||
| 2661 | Known.resetAll(); | ||||||
| 2662 | } | ||||||
| 2663 | |||||||
| 2664 | void TargetLowering::computeKnownBitsForFrameIndex(const SDValue Op, | ||||||
| 2665 | KnownBits &Known, | ||||||
| 2666 | const APInt &DemandedElts, | ||||||
| 2667 | const SelectionDAG &DAG, | ||||||
| 2668 | unsigned Depth) const { | ||||||
| 2669 | assert(isa<FrameIndexSDNode>(Op) && "expected FrameIndex")((isa<FrameIndexSDNode>(Op) && "expected FrameIndex" ) ? static_cast<void> (0) : __assert_fail ("isa<FrameIndexSDNode>(Op) && \"expected FrameIndex\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2669, __PRETTY_FUNCTION__)); | ||||||
| 2670 | |||||||
| 2671 | if (unsigned Align = DAG.InferPtrAlignment(Op)) { | ||||||
| 2672 | // The low bits are known zero if the pointer is aligned. | ||||||
| 2673 | Known.Zero.setLowBits(Log2_32(Align)); | ||||||
| 2674 | } | ||||||
| 2675 | } | ||||||
| 2676 | |||||||
| 2677 | /// This method can be implemented by targets that want to expose additional | ||||||
| 2678 | /// information about sign bits to the DAG Combiner. | ||||||
| 2679 | unsigned TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op, | ||||||
| 2680 | const APInt &, | ||||||
| 2681 | const SelectionDAG &, | ||||||
| 2682 | unsigned Depth) const { | ||||||
| 2683 | assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)) | ||||||
| 2684 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)) | ||||||
| 2685 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)) | ||||||
| 2686 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)) | ||||||
| 2687 | "Should use ComputeNumSignBits if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)) | ||||||
| 2688 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use ComputeNumSignBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use ComputeNumSignBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2688, __PRETTY_FUNCTION__)); | ||||||
| 2689 | return 1; | ||||||
| 2690 | } | ||||||
| 2691 | |||||||
| 2692 | bool TargetLowering::SimplifyDemandedVectorEltsForTargetNode( | ||||||
| 2693 | SDValue Op, const APInt &DemandedElts, APInt &KnownUndef, APInt &KnownZero, | ||||||
| 2694 | TargetLoweringOpt &TLO, unsigned Depth) const { | ||||||
| 2695 | assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)) | ||||||
| 2696 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)) | ||||||
| 2697 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)) | ||||||
| 2698 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)) | ||||||
| 2699 | "Should use SimplifyDemandedVectorElts if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)) | ||||||
| 2700 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedVectorElts if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedVectorElts if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2700, __PRETTY_FUNCTION__)); | ||||||
| 2701 | return false; | ||||||
| 2702 | } | ||||||
| 2703 | |||||||
| 2704 | bool TargetLowering::SimplifyDemandedBitsForTargetNode( | ||||||
| 2705 | SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, | ||||||
| 2706 | KnownBits &Known, TargetLoweringOpt &TLO, unsigned Depth) const { | ||||||
| 2707 | assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)) | ||||||
| 2708 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)) | ||||||
| 2709 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)) | ||||||
| 2710 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)) | ||||||
| 2711 | "Should use SimplifyDemandedBits if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)) | ||||||
| 2712 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2712, __PRETTY_FUNCTION__)); | ||||||
| 2713 | computeKnownBitsForTargetNode(Op, Known, DemandedElts, TLO.DAG, Depth); | ||||||
| 2714 | return false; | ||||||
| 2715 | } | ||||||
| 2716 | |||||||
| 2717 | SDValue TargetLowering::SimplifyMultipleUseDemandedBitsForTargetNode( | ||||||
| 2718 | SDValue Op, const APInt &DemandedBits, const APInt &DemandedElts, | ||||||
| 2719 | SelectionDAG &DAG, unsigned Depth) const { | ||||||
| 2720 | assert((((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2721 | (Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2722 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2723 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2724 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2725 | "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)) | ||||||
| 2726 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use SimplifyMultipleUseDemandedBits if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use SimplifyMultipleUseDemandedBits if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2726, __PRETTY_FUNCTION__)); | ||||||
| 2727 | return SDValue(); | ||||||
| 2728 | } | ||||||
| 2729 | |||||||
| 2730 | SDValue | ||||||
| 2731 | TargetLowering::buildLegalVectorShuffle(EVT VT, const SDLoc &DL, SDValue N0, | ||||||
| 2732 | SDValue N1, MutableArrayRef<int> Mask, | ||||||
| 2733 | SelectionDAG &DAG) const { | ||||||
| 2734 | bool LegalMask = isShuffleMaskLegal(Mask, VT); | ||||||
| 2735 | if (!LegalMask) { | ||||||
| 2736 | std::swap(N0, N1); | ||||||
| 2737 | ShuffleVectorSDNode::commuteMask(Mask); | ||||||
| 2738 | LegalMask = isShuffleMaskLegal(Mask, VT); | ||||||
| 2739 | } | ||||||
| 2740 | |||||||
| 2741 | if (!LegalMask) | ||||||
| 2742 | return SDValue(); | ||||||
| 2743 | |||||||
| 2744 | return DAG.getVectorShuffle(VT, DL, N0, N1, Mask); | ||||||
| 2745 | } | ||||||
| 2746 | |||||||
| 2747 | const Constant *TargetLowering::getTargetConstantFromLoad(LoadSDNode*) const { | ||||||
| 2748 | return nullptr; | ||||||
| 2749 | } | ||||||
| 2750 | |||||||
| 2751 | bool TargetLowering::isKnownNeverNaNForTargetNode(SDValue Op, | ||||||
| 2752 | const SelectionDAG &DAG, | ||||||
| 2753 | bool SNaN, | ||||||
| 2754 | unsigned Depth) const { | ||||||
| 2755 | assert((Op.getOpcode() >= ISD::BUILTIN_OP_END ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)) | ||||||
| 2756 | Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)) | ||||||
| 2757 | Op.getOpcode() == ISD::INTRINSIC_W_CHAIN ||(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)) | ||||||
| 2758 | Op.getOpcode() == ISD::INTRINSIC_VOID) &&(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)) | ||||||
| 2759 | "Should use isKnownNeverNaN if you don't know whether Op"(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)) | ||||||
| 2760 | " is a target node!")(((Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && "Should use isKnownNeverNaN if you don't know whether Op" " is a target node!") ? static_cast<void> (0) : __assert_fail ("(Op.getOpcode() >= ISD::BUILTIN_OP_END || Op.getOpcode() == ISD::INTRINSIC_WO_CHAIN || Op.getOpcode() == ISD::INTRINSIC_W_CHAIN || Op.getOpcode() == ISD::INTRINSIC_VOID) && \"Should use isKnownNeverNaN if you don't know whether Op\" \" is a target node!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2760, __PRETTY_FUNCTION__)); | ||||||
| 2761 | return false; | ||||||
| 2762 | } | ||||||
| 2763 | |||||||
| 2764 | // FIXME: Ideally, this would use ISD::isConstantSplatVector(), but that must | ||||||
| 2765 | // work with truncating build vectors and vectors with elements of less than | ||||||
| 2766 | // 8 bits. | ||||||
| 2767 | bool TargetLowering::isConstTrueVal(const SDNode *N) const { | ||||||
| 2768 | if (!N) | ||||||
| 2769 | return false; | ||||||
| 2770 | |||||||
| 2771 | APInt CVal; | ||||||
| 2772 | if (auto *CN = dyn_cast<ConstantSDNode>(N)) { | ||||||
| 2773 | CVal = CN->getAPIntValue(); | ||||||
| 2774 | } else if (auto *BV = dyn_cast<BuildVectorSDNode>(N)) { | ||||||
| 2775 | auto *CN = BV->getConstantSplatNode(); | ||||||
| 2776 | if (!CN) | ||||||
| 2777 | return false; | ||||||
| 2778 | |||||||
| 2779 | // If this is a truncating build vector, truncate the splat value. | ||||||
| 2780 | // Otherwise, we may fail to match the expected values below. | ||||||
| 2781 | unsigned BVEltWidth = BV->getValueType(0).getScalarSizeInBits(); | ||||||
| 2782 | CVal = CN->getAPIntValue(); | ||||||
| 2783 | if (BVEltWidth < CVal.getBitWidth()) | ||||||
| 2784 | CVal = CVal.trunc(BVEltWidth); | ||||||
| 2785 | } else { | ||||||
| 2786 | return false; | ||||||
| 2787 | } | ||||||
| 2788 | |||||||
| 2789 | switch (getBooleanContents(N->getValueType(0))) { | ||||||
| 2790 | case UndefinedBooleanContent: | ||||||
| 2791 | return CVal[0]; | ||||||
| 2792 | case ZeroOrOneBooleanContent: | ||||||
| 2793 | return CVal.isOneValue(); | ||||||
| 2794 | case ZeroOrNegativeOneBooleanContent: | ||||||
| 2795 | return CVal.isAllOnesValue(); | ||||||
| 2796 | } | ||||||
| 2797 | |||||||
| 2798 | llvm_unreachable("Invalid boolean contents")::llvm::llvm_unreachable_internal("Invalid boolean contents", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2798); | ||||||
| 2799 | } | ||||||
| 2800 | |||||||
| 2801 | bool TargetLowering::isConstFalseVal(const SDNode *N) const { | ||||||
| 2802 | if (!N) | ||||||
| 2803 | return false; | ||||||
| 2804 | |||||||
| 2805 | const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N); | ||||||
| 2806 | if (!CN) { | ||||||
| 2807 | const BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(N); | ||||||
| 2808 | if (!BV) | ||||||
| 2809 | return false; | ||||||
| 2810 | |||||||
| 2811 | // Only interested in constant splats, we don't care about undef | ||||||
| 2812 | // elements in identifying boolean constants and getConstantSplatNode | ||||||
| 2813 | // returns NULL if all ops are undef; | ||||||
| 2814 | CN = BV->getConstantSplatNode(); | ||||||
| 2815 | if (!CN) | ||||||
| 2816 | return false; | ||||||
| 2817 | } | ||||||
| 2818 | |||||||
| 2819 | if (getBooleanContents(N->getValueType(0)) == UndefinedBooleanContent) | ||||||
| 2820 | return !CN->getAPIntValue()[0]; | ||||||
| 2821 | |||||||
| 2822 | return CN->isNullValue(); | ||||||
| 2823 | } | ||||||
| 2824 | |||||||
| 2825 | bool TargetLowering::isExtendedTrueVal(const ConstantSDNode *N, EVT VT, | ||||||
| 2826 | bool SExt) const { | ||||||
| 2827 | if (VT == MVT::i1) | ||||||
| 2828 | return N->isOne(); | ||||||
| 2829 | |||||||
| 2830 | TargetLowering::BooleanContent Cnt = getBooleanContents(VT); | ||||||
| 2831 | switch (Cnt) { | ||||||
| 2832 | case TargetLowering::ZeroOrOneBooleanContent: | ||||||
| 2833 | // An extended value of 1 is always true, unless its original type is i1, | ||||||
| 2834 | // in which case it will be sign extended to -1. | ||||||
| 2835 | return (N->isOne() && !SExt) || (SExt && (N->getValueType(0) != MVT::i1)); | ||||||
| 2836 | case TargetLowering::UndefinedBooleanContent: | ||||||
| 2837 | case TargetLowering::ZeroOrNegativeOneBooleanContent: | ||||||
| 2838 | return N->isAllOnesValue() && SExt; | ||||||
| 2839 | } | ||||||
| 2840 | llvm_unreachable("Unexpected enumeration.")::llvm::llvm_unreachable_internal("Unexpected enumeration.", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2840); | ||||||
| 2841 | } | ||||||
| 2842 | |||||||
| 2843 | /// This helper function of SimplifySetCC tries to optimize the comparison when | ||||||
| 2844 | /// either operand of the SetCC node is a bitwise-and instruction. | ||||||
| 2845 | SDValue TargetLowering::foldSetCCWithAnd(EVT VT, SDValue N0, SDValue N1, | ||||||
| 2846 | ISD::CondCode Cond, const SDLoc &DL, | ||||||
| 2847 | DAGCombinerInfo &DCI) const { | ||||||
| 2848 | // Match these patterns in any of their permutations: | ||||||
| 2849 | // (X & Y) == Y | ||||||
| 2850 | // (X & Y) != Y | ||||||
| 2851 | if (N1.getOpcode() == ISD::AND && N0.getOpcode() != ISD::AND) | ||||||
| 2852 | std::swap(N0, N1); | ||||||
| 2853 | |||||||
| 2854 | EVT OpVT = N0.getValueType(); | ||||||
| 2855 | if (N0.getOpcode() != ISD::AND || !OpVT.isInteger() || | ||||||
| 2856 | (Cond != ISD::SETEQ && Cond != ISD::SETNE)) | ||||||
| 2857 | return SDValue(); | ||||||
| 2858 | |||||||
| 2859 | SDValue X, Y; | ||||||
| 2860 | if (N0.getOperand(0) == N1) { | ||||||
| 2861 | X = N0.getOperand(1); | ||||||
| 2862 | Y = N0.getOperand(0); | ||||||
| 2863 | } else if (N0.getOperand(1) == N1) { | ||||||
| 2864 | X = N0.getOperand(0); | ||||||
| 2865 | Y = N0.getOperand(1); | ||||||
| 2866 | } else { | ||||||
| 2867 | return SDValue(); | ||||||
| 2868 | } | ||||||
| 2869 | |||||||
| 2870 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 2871 | SDValue Zero = DAG.getConstant(0, DL, OpVT); | ||||||
| 2872 | if (DAG.isKnownToBeAPowerOfTwo(Y)) { | ||||||
| 2873 | // Simplify X & Y == Y to X & Y != 0 if Y has exactly one bit set. | ||||||
| 2874 | // Note that where Y is variable and is known to have at most one bit set | ||||||
| 2875 | // (for example, if it is Z & 1) we cannot do this; the expressions are not | ||||||
| 2876 | // equivalent when Y == 0. | ||||||
| 2877 | assert(OpVT.isInteger())((OpVT.isInteger()) ? static_cast<void> (0) : __assert_fail ("OpVT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2877, __PRETTY_FUNCTION__)); | ||||||
| 2878 | Cond = ISD::getSetCCInverse(Cond, OpVT); | ||||||
| 2879 | if (DCI.isBeforeLegalizeOps() || | ||||||
| 2880 | isCondCodeLegal(Cond, N0.getSimpleValueType())) | ||||||
| 2881 | return DAG.getSetCC(DL, VT, N0, Zero, Cond); | ||||||
| 2882 | } else if (N0.hasOneUse() && hasAndNotCompare(Y)) { | ||||||
| 2883 | // If the target supports an 'and-not' or 'and-complement' logic operation, | ||||||
| 2884 | // try to use that to make a comparison operation more efficient. | ||||||
| 2885 | // But don't do this transform if the mask is a single bit because there are | ||||||
| 2886 | // more efficient ways to deal with that case (for example, 'bt' on x86 or | ||||||
| 2887 | // 'rlwinm' on PPC). | ||||||
| 2888 | |||||||
| 2889 | // Bail out if the compare operand that we want to turn into a zero is | ||||||
| 2890 | // already a zero (otherwise, infinite loop). | ||||||
| 2891 | auto *YConst = dyn_cast<ConstantSDNode>(Y); | ||||||
| 2892 | if (YConst && YConst->isNullValue()) | ||||||
| 2893 | return SDValue(); | ||||||
| 2894 | |||||||
| 2895 | // Transform this into: ~X & Y == 0. | ||||||
| 2896 | SDValue NotX = DAG.getNOT(SDLoc(X), X, OpVT); | ||||||
| 2897 | SDValue NewAnd = DAG.getNode(ISD::AND, SDLoc(N0), OpVT, NotX, Y); | ||||||
| 2898 | return DAG.getSetCC(DL, VT, NewAnd, Zero, Cond); | ||||||
| 2899 | } | ||||||
| 2900 | |||||||
| 2901 | return SDValue(); | ||||||
| 2902 | } | ||||||
| 2903 | |||||||
| 2904 | /// There are multiple IR patterns that could be checking whether certain | ||||||
| 2905 | /// truncation of a signed number would be lossy or not. The pattern which is | ||||||
| 2906 | /// best at IR level, may not lower optimally. Thus, we want to unfold it. | ||||||
| 2907 | /// We are looking for the following pattern: (KeptBits is a constant) | ||||||
| 2908 | /// (add %x, (1 << (KeptBits-1))) srccond (1 << KeptBits) | ||||||
| 2909 | /// KeptBits won't be bitwidth(x), that will be constant-folded to true/false. | ||||||
| 2910 | /// KeptBits also can't be 1, that would have been folded to %x dstcond 0 | ||||||
| 2911 | /// We will unfold it into the natural trunc+sext pattern: | ||||||
| 2912 | /// ((%x << C) a>> C) dstcond %x | ||||||
| 2913 | /// Where C = bitwidth(x) - KeptBits and C u< bitwidth(x) | ||||||
| 2914 | SDValue TargetLowering::optimizeSetCCOfSignedTruncationCheck( | ||||||
| 2915 | EVT SCCVT, SDValue N0, SDValue N1, ISD::CondCode Cond, DAGCombinerInfo &DCI, | ||||||
| 2916 | const SDLoc &DL) const { | ||||||
| 2917 | // We must be comparing with a constant. | ||||||
| 2918 | ConstantSDNode *C1; | ||||||
| 2919 | if (!(C1 = dyn_cast<ConstantSDNode>(N1))) | ||||||
| 2920 | return SDValue(); | ||||||
| 2921 | |||||||
| 2922 | // N0 should be: add %x, (1 << (KeptBits-1)) | ||||||
| 2923 | if (N0->getOpcode() != ISD::ADD) | ||||||
| 2924 | return SDValue(); | ||||||
| 2925 | |||||||
| 2926 | // And we must be 'add'ing a constant. | ||||||
| 2927 | ConstantSDNode *C01; | ||||||
| 2928 | if (!(C01 = dyn_cast<ConstantSDNode>(N0->getOperand(1)))) | ||||||
| 2929 | return SDValue(); | ||||||
| 2930 | |||||||
| 2931 | SDValue X = N0->getOperand(0); | ||||||
| 2932 | EVT XVT = X.getValueType(); | ||||||
| 2933 | |||||||
| 2934 | // Validate constants ... | ||||||
| 2935 | |||||||
| 2936 | APInt I1 = C1->getAPIntValue(); | ||||||
| 2937 | |||||||
| 2938 | ISD::CondCode NewCond; | ||||||
| 2939 | if (Cond == ISD::CondCode::SETULT) { | ||||||
| 2940 | NewCond = ISD::CondCode::SETEQ; | ||||||
| 2941 | } else if (Cond == ISD::CondCode::SETULE) { | ||||||
| 2942 | NewCond = ISD::CondCode::SETEQ; | ||||||
| 2943 | // But need to 'canonicalize' the constant. | ||||||
| 2944 | I1 += 1; | ||||||
| 2945 | } else if (Cond == ISD::CondCode::SETUGT) { | ||||||
| 2946 | NewCond = ISD::CondCode::SETNE; | ||||||
| 2947 | // But need to 'canonicalize' the constant. | ||||||
| 2948 | I1 += 1; | ||||||
| 2949 | } else if (Cond == ISD::CondCode::SETUGE) { | ||||||
| 2950 | NewCond = ISD::CondCode::SETNE; | ||||||
| 2951 | } else | ||||||
| 2952 | return SDValue(); | ||||||
| 2953 | |||||||
| 2954 | APInt I01 = C01->getAPIntValue(); | ||||||
| 2955 | |||||||
| 2956 | auto checkConstants = [&I1, &I01]() -> bool { | ||||||
| 2957 | // Both of them must be power-of-two, and the constant from setcc is bigger. | ||||||
| 2958 | return I1.ugt(I01) && I1.isPowerOf2() && I01.isPowerOf2(); | ||||||
| 2959 | }; | ||||||
| 2960 | |||||||
| 2961 | if (checkConstants()) { | ||||||
| 2962 | // Great, e.g. got icmp ult i16 (add i16 %x, 128), 256 | ||||||
| 2963 | } else { | ||||||
| 2964 | // What if we invert constants? (and the target predicate) | ||||||
| 2965 | I1.negate(); | ||||||
| 2966 | I01.negate(); | ||||||
| 2967 | assert(XVT.isInteger())((XVT.isInteger()) ? static_cast<void> (0) : __assert_fail ("XVT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2967, __PRETTY_FUNCTION__)); | ||||||
| 2968 | NewCond = getSetCCInverse(NewCond, XVT); | ||||||
| 2969 | if (!checkConstants()) | ||||||
| 2970 | return SDValue(); | ||||||
| 2971 | // Great, e.g. got icmp uge i16 (add i16 %x, -128), -256 | ||||||
| 2972 | } | ||||||
| 2973 | |||||||
| 2974 | // They are power-of-two, so which bit is set? | ||||||
| 2975 | const unsigned KeptBits = I1.logBase2(); | ||||||
| 2976 | const unsigned KeptBitsMinusOne = I01.logBase2(); | ||||||
| 2977 | |||||||
| 2978 | // Magic! | ||||||
| 2979 | if (KeptBits != (KeptBitsMinusOne + 1)) | ||||||
| 2980 | return SDValue(); | ||||||
| 2981 | assert(KeptBits > 0 && KeptBits < XVT.getSizeInBits() && "unreachable")((KeptBits > 0 && KeptBits < XVT.getSizeInBits( ) && "unreachable") ? static_cast<void> (0) : __assert_fail ("KeptBits > 0 && KeptBits < XVT.getSizeInBits() && \"unreachable\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2981, __PRETTY_FUNCTION__)); | ||||||
| 2982 | |||||||
| 2983 | // We don't want to do this in every single case. | ||||||
| 2984 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 2985 | if (!DAG.getTargetLoweringInfo().shouldTransformSignedTruncationCheck( | ||||||
| 2986 | XVT, KeptBits)) | ||||||
| 2987 | return SDValue(); | ||||||
| 2988 | |||||||
| 2989 | const unsigned MaskedBits = XVT.getSizeInBits() - KeptBits; | ||||||
| 2990 | assert(MaskedBits > 0 && MaskedBits < XVT.getSizeInBits() && "unreachable")((MaskedBits > 0 && MaskedBits < XVT.getSizeInBits () && "unreachable") ? static_cast<void> (0) : __assert_fail ("MaskedBits > 0 && MaskedBits < XVT.getSizeInBits() && \"unreachable\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 2990, __PRETTY_FUNCTION__)); | ||||||
| 2991 | |||||||
| 2992 | // Unfold into: ((%x << C) a>> C) cond %x | ||||||
| 2993 | // Where 'cond' will be either 'eq' or 'ne'. | ||||||
| 2994 | SDValue ShiftAmt = DAG.getConstant(MaskedBits, DL, XVT); | ||||||
| 2995 | SDValue T0 = DAG.getNode(ISD::SHL, DL, XVT, X, ShiftAmt); | ||||||
| 2996 | SDValue T1 = DAG.getNode(ISD::SRA, DL, XVT, T0, ShiftAmt); | ||||||
| 2997 | SDValue T2 = DAG.getSetCC(DL, SCCVT, T1, X, NewCond); | ||||||
| 2998 | |||||||
| 2999 | return T2; | ||||||
| 3000 | } | ||||||
| 3001 | |||||||
| 3002 | // (X & (C l>>/<< Y)) ==/!= 0 --> ((X <</l>> Y) & C) ==/!= 0 | ||||||
| 3003 | SDValue TargetLowering::optimizeSetCCByHoistingAndByConstFromLogicalShift( | ||||||
| 3004 | EVT SCCVT, SDValue N0, SDValue N1C, ISD::CondCode Cond, | ||||||
| 3005 | DAGCombinerInfo &DCI, const SDLoc &DL) const { | ||||||
| 3006 | assert(isConstOrConstSplat(N1C) &&((isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C )->getAPIntValue().isNullValue() && "Should be a comparison with 0." ) ? static_cast<void> (0) : __assert_fail ("isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C)->getAPIntValue().isNullValue() && \"Should be a comparison with 0.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3008, __PRETTY_FUNCTION__)) | ||||||
| 3007 | isConstOrConstSplat(N1C)->getAPIntValue().isNullValue() &&((isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C )->getAPIntValue().isNullValue() && "Should be a comparison with 0." ) ? static_cast<void> (0) : __assert_fail ("isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C)->getAPIntValue().isNullValue() && \"Should be a comparison with 0.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3008, __PRETTY_FUNCTION__)) | ||||||
| 3008 | "Should be a comparison with 0.")((isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C )->getAPIntValue().isNullValue() && "Should be a comparison with 0." ) ? static_cast<void> (0) : __assert_fail ("isConstOrConstSplat(N1C) && isConstOrConstSplat(N1C)->getAPIntValue().isNullValue() && \"Should be a comparison with 0.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3008, __PRETTY_FUNCTION__)); | ||||||
| 3009 | assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Valid only for [in]equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Valid only for [in]equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3010, __PRETTY_FUNCTION__)) | ||||||
| 3010 | "Valid only for [in]equality comparisons.")(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Valid only for [in]equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Valid only for [in]equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3010, __PRETTY_FUNCTION__)); | ||||||
| 3011 | |||||||
| 3012 | unsigned NewShiftOpcode; | ||||||
| 3013 | SDValue X, C, Y; | ||||||
| 3014 | |||||||
| 3015 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 3016 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | ||||||
| 3017 | |||||||
| 3018 | // Look for '(C l>>/<< Y)'. | ||||||
| 3019 | auto Match = [&NewShiftOpcode, &X, &C, &Y, &TLI, &DAG](SDValue V) { | ||||||
| 3020 | // The shift should be one-use. | ||||||
| 3021 | if (!V.hasOneUse()) | ||||||
| 3022 | return false; | ||||||
| 3023 | unsigned OldShiftOpcode = V.getOpcode(); | ||||||
| 3024 | switch (OldShiftOpcode) { | ||||||
| 3025 | case ISD::SHL: | ||||||
| 3026 | NewShiftOpcode = ISD::SRL; | ||||||
| 3027 | break; | ||||||
| 3028 | case ISD::SRL: | ||||||
| 3029 | NewShiftOpcode = ISD::SHL; | ||||||
| 3030 | break; | ||||||
| 3031 | default: | ||||||
| 3032 | return false; // must be a logical shift. | ||||||
| 3033 | } | ||||||
| 3034 | // We should be shifting a constant. | ||||||
| 3035 | // FIXME: best to use isConstantOrConstantVector(). | ||||||
| 3036 | C = V.getOperand(0); | ||||||
| 3037 | ConstantSDNode *CC = | ||||||
| 3038 | isConstOrConstSplat(C, /*AllowUndefs=*/true, /*AllowTruncation=*/true); | ||||||
| 3039 | if (!CC) | ||||||
| 3040 | return false; | ||||||
| 3041 | Y = V.getOperand(1); | ||||||
| 3042 | |||||||
| 3043 | ConstantSDNode *XC = | ||||||
| 3044 | isConstOrConstSplat(X, /*AllowUndefs=*/true, /*AllowTruncation=*/true); | ||||||
| 3045 | return TLI.shouldProduceAndByConstByHoistingConstFromShiftsLHSOfAnd( | ||||||
| 3046 | X, XC, CC, Y, OldShiftOpcode, NewShiftOpcode, DAG); | ||||||
| 3047 | }; | ||||||
| 3048 | |||||||
| 3049 | // LHS of comparison should be an one-use 'and'. | ||||||
| 3050 | if (N0.getOpcode() != ISD::AND || !N0.hasOneUse()) | ||||||
| 3051 | return SDValue(); | ||||||
| 3052 | |||||||
| 3053 | X = N0.getOperand(0); | ||||||
| 3054 | SDValue Mask = N0.getOperand(1); | ||||||
| 3055 | |||||||
| 3056 | // 'and' is commutative! | ||||||
| 3057 | if (!Match(Mask)) { | ||||||
| 3058 | std::swap(X, Mask); | ||||||
| 3059 | if (!Match(Mask)) | ||||||
| 3060 | return SDValue(); | ||||||
| 3061 | } | ||||||
| 3062 | |||||||
| 3063 | EVT VT = X.getValueType(); | ||||||
| 3064 | |||||||
| 3065 | // Produce: | ||||||
| 3066 | // ((X 'OppositeShiftOpcode' Y) & C) Cond 0 | ||||||
| 3067 | SDValue T0 = DAG.getNode(NewShiftOpcode, DL, VT, X, Y); | ||||||
| 3068 | SDValue T1 = DAG.getNode(ISD::AND, DL, VT, T0, C); | ||||||
| 3069 | SDValue T2 = DAG.getSetCC(DL, SCCVT, T1, N1C, Cond); | ||||||
| 3070 | return T2; | ||||||
| 3071 | } | ||||||
| 3072 | |||||||
| 3073 | /// Try to fold an equality comparison with a {add/sub/xor} binary operation as | ||||||
| 3074 | /// the 1st operand (N0). Callers are expected to swap the N0/N1 parameters to | ||||||
| 3075 | /// handle the commuted versions of these patterns. | ||||||
| 3076 | SDValue TargetLowering::foldSetCCWithBinOp(EVT VT, SDValue N0, SDValue N1, | ||||||
| 3077 | ISD::CondCode Cond, const SDLoc &DL, | ||||||
| 3078 | DAGCombinerInfo &DCI) const { | ||||||
| 3079 | unsigned BOpcode = N0.getOpcode(); | ||||||
| 3080 | assert((BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD::XOR) &&(((BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD ::XOR) && "Unexpected binop") ? static_cast<void> (0) : __assert_fail ("(BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD::XOR) && \"Unexpected binop\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3081, __PRETTY_FUNCTION__)) | ||||||
| 3081 | "Unexpected binop")(((BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD ::XOR) && "Unexpected binop") ? static_cast<void> (0) : __assert_fail ("(BOpcode == ISD::ADD || BOpcode == ISD::SUB || BOpcode == ISD::XOR) && \"Unexpected binop\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3081, __PRETTY_FUNCTION__)); | ||||||
| 3082 | assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Unexpected condcode")(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Unexpected condcode" ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Unexpected condcode\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3082, __PRETTY_FUNCTION__)); | ||||||
| 3083 | |||||||
| 3084 | // (X + Y) == X --> Y == 0 | ||||||
| 3085 | // (X - Y) == X --> Y == 0 | ||||||
| 3086 | // (X ^ Y) == X --> Y == 0 | ||||||
| 3087 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 3088 | EVT OpVT = N0.getValueType(); | ||||||
| 3089 | SDValue X = N0.getOperand(0); | ||||||
| 3090 | SDValue Y = N0.getOperand(1); | ||||||
| 3091 | if (X == N1) | ||||||
| 3092 | return DAG.getSetCC(DL, VT, Y, DAG.getConstant(0, DL, OpVT), Cond); | ||||||
| 3093 | |||||||
| 3094 | if (Y != N1) | ||||||
| 3095 | return SDValue(); | ||||||
| 3096 | |||||||
| 3097 | // (X + Y) == Y --> X == 0 | ||||||
| 3098 | // (X ^ Y) == Y --> X == 0 | ||||||
| 3099 | if (BOpcode == ISD::ADD || BOpcode == ISD::XOR) | ||||||
| 3100 | return DAG.getSetCC(DL, VT, X, DAG.getConstant(0, DL, OpVT), Cond); | ||||||
| 3101 | |||||||
| 3102 | // The shift would not be valid if the operands are boolean (i1). | ||||||
| 3103 | if (!N0.hasOneUse() || OpVT.getScalarSizeInBits() == 1) | ||||||
| 3104 | return SDValue(); | ||||||
| 3105 | |||||||
| 3106 | // (X - Y) == Y --> X == Y << 1 | ||||||
| 3107 | EVT ShiftVT = getShiftAmountTy(OpVT, DAG.getDataLayout(), | ||||||
| 3108 | !DCI.isBeforeLegalize()); | ||||||
| 3109 | SDValue One = DAG.getConstant(1, DL, ShiftVT); | ||||||
| 3110 | SDValue YShl1 = DAG.getNode(ISD::SHL, DL, N1.getValueType(), Y, One); | ||||||
| 3111 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3112 | DCI.AddToWorklist(YShl1.getNode()); | ||||||
| 3113 | return DAG.getSetCC(DL, VT, X, YShl1, Cond); | ||||||
| 3114 | } | ||||||
| 3115 | |||||||
| 3116 | /// Try to simplify a setcc built with the specified operands and cc. If it is | ||||||
| 3117 | /// unable to simplify it, return a null SDValue. | ||||||
| 3118 | SDValue TargetLowering::SimplifySetCC(EVT VT, SDValue N0, SDValue N1, | ||||||
| 3119 | ISD::CondCode Cond, bool foldBooleans, | ||||||
| 3120 | DAGCombinerInfo &DCI, | ||||||
| 3121 | const SDLoc &dl) const { | ||||||
| 3122 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 3123 | const DataLayout &Layout = DAG.getDataLayout(); | ||||||
| 3124 | EVT OpVT = N0.getValueType(); | ||||||
| 3125 | |||||||
| 3126 | // Constant fold or commute setcc. | ||||||
| 3127 | if (SDValue Fold = DAG.FoldSetCC(VT, N0, N1, Cond, dl)) | ||||||
| 3128 | return Fold; | ||||||
| 3129 | |||||||
| 3130 | // Ensure that the constant occurs on the RHS and fold constant comparisons. | ||||||
| 3131 | // TODO: Handle non-splat vector constants. All undef causes trouble. | ||||||
| 3132 | ISD::CondCode SwappedCC = ISD::getSetCCSwappedOperands(Cond); | ||||||
| 3133 | if (isConstOrConstSplat(N0) && | ||||||
| 3134 | (DCI.isBeforeLegalizeOps() || | ||||||
| 3135 | isCondCodeLegal(SwappedCC, N0.getSimpleValueType()))) | ||||||
| 3136 | return DAG.getSetCC(dl, VT, N1, N0, SwappedCC); | ||||||
| 3137 | |||||||
| 3138 | // If we have a subtract with the same 2 non-constant operands as this setcc | ||||||
| 3139 | // -- but in reverse order -- then try to commute the operands of this setcc | ||||||
| 3140 | // to match. A matching pair of setcc (cmp) and sub may be combined into 1 | ||||||
| 3141 | // instruction on some targets. | ||||||
| 3142 | if (!isConstOrConstSplat(N0) && !isConstOrConstSplat(N1) && | ||||||
| 3143 | (DCI.isBeforeLegalizeOps() || | ||||||
| 3144 | isCondCodeLegal(SwappedCC, N0.getSimpleValueType())) && | ||||||
| 3145 | DAG.getNodeIfExists(ISD::SUB, DAG.getVTList(OpVT), { N1, N0 } ) && | ||||||
| 3146 | !DAG.getNodeIfExists(ISD::SUB, DAG.getVTList(OpVT), { N0, N1 } )) | ||||||
| 3147 | return DAG.getSetCC(dl, VT, N1, N0, SwappedCC); | ||||||
| 3148 | |||||||
| 3149 | if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { | ||||||
| 3150 | const APInt &C1 = N1C->getAPIntValue(); | ||||||
| 3151 | |||||||
| 3152 | // If the LHS is '(srl (ctlz x), 5)', the RHS is 0/1, and this is an | ||||||
| 3153 | // equality comparison, then we're just comparing whether X itself is | ||||||
| 3154 | // zero. | ||||||
| 3155 | if (N0.getOpcode() == ISD::SRL && (C1.isNullValue() || C1.isOneValue()) && | ||||||
| 3156 | N0.getOperand(0).getOpcode() == ISD::CTLZ && | ||||||
| 3157 | N0.getOperand(1).getOpcode() == ISD::Constant) { | ||||||
| 3158 | const APInt &ShAmt = N0.getConstantOperandAPInt(1); | ||||||
| 3159 | if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && | ||||||
| 3160 | ShAmt == Log2_32(N0.getValueSizeInBits())) { | ||||||
| 3161 | if ((C1 == 0) == (Cond == ISD::SETEQ)) { | ||||||
| 3162 | // (srl (ctlz x), 5) == 0 -> X != 0 | ||||||
| 3163 | // (srl (ctlz x), 5) != 1 -> X != 0 | ||||||
| 3164 | Cond = ISD::SETNE; | ||||||
| 3165 | } else { | ||||||
| 3166 | // (srl (ctlz x), 5) != 0 -> X == 0 | ||||||
| 3167 | // (srl (ctlz x), 5) == 1 -> X == 0 | ||||||
| 3168 | Cond = ISD::SETEQ; | ||||||
| 3169 | } | ||||||
| 3170 | SDValue Zero = DAG.getConstant(0, dl, N0.getValueType()); | ||||||
| 3171 | return DAG.getSetCC(dl, VT, N0.getOperand(0).getOperand(0), | ||||||
| 3172 | Zero, Cond); | ||||||
| 3173 | } | ||||||
| 3174 | } | ||||||
| 3175 | |||||||
| 3176 | SDValue CTPOP = N0; | ||||||
| 3177 | // Look through truncs that don't change the value of a ctpop. | ||||||
| 3178 | if (N0.hasOneUse() && N0.getOpcode() == ISD::TRUNCATE) | ||||||
| 3179 | CTPOP = N0.getOperand(0); | ||||||
| 3180 | |||||||
| 3181 | if (CTPOP.hasOneUse() && CTPOP.getOpcode() == ISD::CTPOP && | ||||||
| 3182 | (N0 == CTPOP || | ||||||
| 3183 | N0.getValueSizeInBits() > Log2_32_Ceil(CTPOP.getValueSizeInBits()))) { | ||||||
| 3184 | EVT CTVT = CTPOP.getValueType(); | ||||||
| 3185 | SDValue CTOp = CTPOP.getOperand(0); | ||||||
| 3186 | |||||||
| 3187 | // (ctpop x) u< 2 -> (x & x-1) == 0 | ||||||
| 3188 | // (ctpop x) u> 1 -> (x & x-1) != 0 | ||||||
| 3189 | if ((Cond == ISD::SETULT && C1 == 2) || (Cond == ISD::SETUGT && C1 == 1)){ | ||||||
| 3190 | SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT); | ||||||
| 3191 | SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne); | ||||||
| 3192 | SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add); | ||||||
| 3193 | ISD::CondCode CC = Cond == ISD::SETULT ? ISD::SETEQ : ISD::SETNE; | ||||||
| 3194 | return DAG.getSetCC(dl, VT, And, DAG.getConstant(0, dl, CTVT), CC); | ||||||
| 3195 | } | ||||||
| 3196 | |||||||
| 3197 | // If ctpop is not supported, expand a power-of-2 comparison based on it. | ||||||
| 3198 | if (C1 == 1 && !isOperationLegalOrCustom(ISD::CTPOP, CTVT) && | ||||||
| 3199 | (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { | ||||||
| 3200 | // (ctpop x) == 1 --> (x != 0) && ((x & x-1) == 0) | ||||||
| 3201 | // (ctpop x) != 1 --> (x == 0) || ((x & x-1) != 0) | ||||||
| 3202 | SDValue Zero = DAG.getConstant(0, dl, CTVT); | ||||||
| 3203 | SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT); | ||||||
| 3204 | assert(CTVT.isInteger())((CTVT.isInteger()) ? static_cast<void> (0) : __assert_fail ("CTVT.isInteger()", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3204, __PRETTY_FUNCTION__)); | ||||||
| 3205 | ISD::CondCode InvCond = ISD::getSetCCInverse(Cond, CTVT); | ||||||
| 3206 | SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne); | ||||||
| 3207 | SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add); | ||||||
| 3208 | SDValue LHS = DAG.getSetCC(dl, VT, CTOp, Zero, InvCond); | ||||||
| 3209 | SDValue RHS = DAG.getSetCC(dl, VT, And, Zero, Cond); | ||||||
| 3210 | unsigned LogicOpcode = Cond == ISD::SETEQ ? ISD::AND : ISD::OR; | ||||||
| 3211 | return DAG.getNode(LogicOpcode, dl, VT, LHS, RHS); | ||||||
| 3212 | } | ||||||
| 3213 | } | ||||||
| 3214 | |||||||
| 3215 | // (zext x) == C --> x == (trunc C) | ||||||
| 3216 | // (sext x) == C --> x == (trunc C) | ||||||
| 3217 | if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && | ||||||
| 3218 | DCI.isBeforeLegalize() && N0->hasOneUse()) { | ||||||
| 3219 | unsigned MinBits = N0.getValueSizeInBits(); | ||||||
| 3220 | SDValue PreExt; | ||||||
| 3221 | bool Signed = false; | ||||||
| 3222 | if (N0->getOpcode() == ISD::ZERO_EXTEND) { | ||||||
| 3223 | // ZExt | ||||||
| 3224 | MinBits = N0->getOperand(0).getValueSizeInBits(); | ||||||
| 3225 | PreExt = N0->getOperand(0); | ||||||
| 3226 | } else if (N0->getOpcode() == ISD::AND) { | ||||||
| 3227 | // DAGCombine turns costly ZExts into ANDs | ||||||
| 3228 | if (auto *C = dyn_cast<ConstantSDNode>(N0->getOperand(1))) | ||||||
| 3229 | if ((C->getAPIntValue()+1).isPowerOf2()) { | ||||||
| 3230 | MinBits = C->getAPIntValue().countTrailingOnes(); | ||||||
| 3231 | PreExt = N0->getOperand(0); | ||||||
| 3232 | } | ||||||
| 3233 | } else if (N0->getOpcode() == ISD::SIGN_EXTEND) { | ||||||
| 3234 | // SExt | ||||||
| 3235 | MinBits = N0->getOperand(0).getValueSizeInBits(); | ||||||
| 3236 | PreExt = N0->getOperand(0); | ||||||
| 3237 | Signed = true; | ||||||
| 3238 | } else if (auto *LN0 = dyn_cast<LoadSDNode>(N0)) { | ||||||
| 3239 | // ZEXTLOAD / SEXTLOAD | ||||||
| 3240 | if (LN0->getExtensionType() == ISD::ZEXTLOAD) { | ||||||
| 3241 | MinBits = LN0->getMemoryVT().getSizeInBits(); | ||||||
| 3242 | PreExt = N0; | ||||||
| 3243 | } else if (LN0->getExtensionType() == ISD::SEXTLOAD) { | ||||||
| 3244 | Signed = true; | ||||||
| 3245 | MinBits = LN0->getMemoryVT().getSizeInBits(); | ||||||
| 3246 | PreExt = N0; | ||||||
| 3247 | } | ||||||
| 3248 | } | ||||||
| 3249 | |||||||
| 3250 | // Figure out how many bits we need to preserve this constant. | ||||||
| 3251 | unsigned ReqdBits = Signed ? | ||||||
| 3252 | C1.getBitWidth() - C1.getNumSignBits() + 1 : | ||||||
| 3253 | C1.getActiveBits(); | ||||||
| 3254 | |||||||
| 3255 | // Make sure we're not losing bits from the constant. | ||||||
| 3256 | if (MinBits > 0 && | ||||||
| 3257 | MinBits < C1.getBitWidth() && | ||||||
| 3258 | MinBits >= ReqdBits) { | ||||||
| 3259 | EVT MinVT = EVT::getIntegerVT(*DAG.getContext(), MinBits); | ||||||
| 3260 | if (isTypeDesirableForOp(ISD::SETCC, MinVT)) { | ||||||
| 3261 | // Will get folded away. | ||||||
| 3262 | SDValue Trunc = DAG.getNode(ISD::TRUNCATE, dl, MinVT, PreExt); | ||||||
| 3263 | if (MinBits == 1 && C1 == 1) | ||||||
| 3264 | // Invert the condition. | ||||||
| 3265 | return DAG.getSetCC(dl, VT, Trunc, DAG.getConstant(0, dl, MVT::i1), | ||||||
| 3266 | Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); | ||||||
| 3267 | SDValue C = DAG.getConstant(C1.trunc(MinBits), dl, MinVT); | ||||||
| 3268 | return DAG.getSetCC(dl, VT, Trunc, C, Cond); | ||||||
| 3269 | } | ||||||
| 3270 | |||||||
| 3271 | // If truncating the setcc operands is not desirable, we can still | ||||||
| 3272 | // simplify the expression in some cases: | ||||||
| 3273 | // setcc ([sz]ext (setcc x, y, cc)), 0, setne) -> setcc (x, y, cc) | ||||||
| 3274 | // setcc ([sz]ext (setcc x, y, cc)), 0, seteq) -> setcc (x, y, inv(cc)) | ||||||
| 3275 | // setcc (zext (setcc x, y, cc)), 1, setne) -> setcc (x, y, inv(cc)) | ||||||
| 3276 | // setcc (zext (setcc x, y, cc)), 1, seteq) -> setcc (x, y, cc) | ||||||
| 3277 | // setcc (sext (setcc x, y, cc)), -1, setne) -> setcc (x, y, inv(cc)) | ||||||
| 3278 | // setcc (sext (setcc x, y, cc)), -1, seteq) -> setcc (x, y, cc) | ||||||
| 3279 | SDValue TopSetCC = N0->getOperand(0); | ||||||
| 3280 | unsigned N0Opc = N0->getOpcode(); | ||||||
| 3281 | bool SExt = (N0Opc == ISD::SIGN_EXTEND); | ||||||
| 3282 | if (TopSetCC.getValueType() == MVT::i1 && VT == MVT::i1 && | ||||||
| 3283 | TopSetCC.getOpcode() == ISD::SETCC && | ||||||
| 3284 | (N0Opc == ISD::ZERO_EXTEND || N0Opc == ISD::SIGN_EXTEND) && | ||||||
| 3285 | (isConstFalseVal(N1C) || | ||||||
| 3286 | isExtendedTrueVal(N1C, N0->getValueType(0), SExt))) { | ||||||
| 3287 | |||||||
| 3288 | bool Inverse = (N1C->isNullValue() && Cond == ISD::SETEQ) || | ||||||
| 3289 | (!N1C->isNullValue() && Cond == ISD::SETNE); | ||||||
| 3290 | |||||||
| 3291 | if (!Inverse) | ||||||
| 3292 | return TopSetCC; | ||||||
| 3293 | |||||||
| 3294 | ISD::CondCode InvCond = ISD::getSetCCInverse( | ||||||
| 3295 | cast<CondCodeSDNode>(TopSetCC.getOperand(2))->get(), | ||||||
| 3296 | TopSetCC.getOperand(0).getValueType()); | ||||||
| 3297 | return DAG.getSetCC(dl, VT, TopSetCC.getOperand(0), | ||||||
| 3298 | TopSetCC.getOperand(1), | ||||||
| 3299 | InvCond); | ||||||
| 3300 | } | ||||||
| 3301 | } | ||||||
| 3302 | } | ||||||
| 3303 | |||||||
| 3304 | // If the LHS is '(and load, const)', the RHS is 0, the test is for | ||||||
| 3305 | // equality or unsigned, and all 1 bits of the const are in the same | ||||||
| 3306 | // partial word, see if we can shorten the load. | ||||||
| 3307 | if (DCI.isBeforeLegalize() && | ||||||
| 3308 | !ISD::isSignedIntSetCC(Cond) && | ||||||
| 3309 | N0.getOpcode() == ISD::AND && C1 == 0 && | ||||||
| 3310 | N0.getNode()->hasOneUse() && | ||||||
| 3311 | isa<LoadSDNode>(N0.getOperand(0)) && | ||||||
| 3312 | N0.getOperand(0).getNode()->hasOneUse() && | ||||||
| 3313 | isa<ConstantSDNode>(N0.getOperand(1))) { | ||||||
| 3314 | LoadSDNode *Lod = cast<LoadSDNode>(N0.getOperand(0)); | ||||||
| 3315 | APInt bestMask; | ||||||
| 3316 | unsigned bestWidth = 0, bestOffset = 0; | ||||||
| 3317 | if (Lod->isSimple() && Lod->isUnindexed()) { | ||||||
| 3318 | unsigned origWidth = N0.getValueSizeInBits(); | ||||||
| 3319 | unsigned maskWidth = origWidth; | ||||||
| 3320 | // We can narrow (e.g.) 16-bit extending loads on 32-bit target to | ||||||
| 3321 | // 8 bits, but have to be careful... | ||||||
| 3322 | if (Lod->getExtensionType() != ISD::NON_EXTLOAD) | ||||||
| 3323 | origWidth = Lod->getMemoryVT().getSizeInBits(); | ||||||
| 3324 | const APInt &Mask = N0.getConstantOperandAPInt(1); | ||||||
| 3325 | for (unsigned width = origWidth / 2; width>=8; width /= 2) { | ||||||
| 3326 | APInt newMask = APInt::getLowBitsSet(maskWidth, width); | ||||||
| 3327 | for (unsigned offset=0; offset<origWidth/width; offset++) { | ||||||
| 3328 | if (Mask.isSubsetOf(newMask)) { | ||||||
| 3329 | if (Layout.isLittleEndian()) | ||||||
| 3330 | bestOffset = (uint64_t)offset * (width/8); | ||||||
| 3331 | else | ||||||
| 3332 | bestOffset = (origWidth/width - offset - 1) * (width/8); | ||||||
| 3333 | bestMask = Mask.lshr(offset * (width/8) * 8); | ||||||
| 3334 | bestWidth = width; | ||||||
| 3335 | break; | ||||||
| 3336 | } | ||||||
| 3337 | newMask <<= width; | ||||||
| 3338 | } | ||||||
| 3339 | } | ||||||
| 3340 | } | ||||||
| 3341 | if (bestWidth) { | ||||||
| 3342 | EVT newVT = EVT::getIntegerVT(*DAG.getContext(), bestWidth); | ||||||
| 3343 | if (newVT.isRound() && | ||||||
| 3344 | shouldReduceLoadWidth(Lod, ISD::NON_EXTLOAD, newVT)) { | ||||||
| 3345 | SDValue Ptr = Lod->getBasePtr(); | ||||||
| 3346 | if (bestOffset != 0) | ||||||
| 3347 | Ptr = DAG.getMemBasePlusOffset(Ptr, bestOffset, dl); | ||||||
| 3348 | unsigned NewAlign = MinAlign(Lod->getAlignment(), bestOffset); | ||||||
| 3349 | SDValue NewLoad = DAG.getLoad( | ||||||
| 3350 | newVT, dl, Lod->getChain(), Ptr, | ||||||
| 3351 | Lod->getPointerInfo().getWithOffset(bestOffset), NewAlign); | ||||||
| 3352 | return DAG.getSetCC(dl, VT, | ||||||
| 3353 | DAG.getNode(ISD::AND, dl, newVT, NewLoad, | ||||||
| 3354 | DAG.getConstant(bestMask.trunc(bestWidth), | ||||||
| 3355 | dl, newVT)), | ||||||
| 3356 | DAG.getConstant(0LL, dl, newVT), Cond); | ||||||
| 3357 | } | ||||||
| 3358 | } | ||||||
| 3359 | } | ||||||
| 3360 | |||||||
| 3361 | // If the LHS is a ZERO_EXTEND, perform the comparison on the input. | ||||||
| 3362 | if (N0.getOpcode() == ISD::ZERO_EXTEND) { | ||||||
| 3363 | unsigned InSize = N0.getOperand(0).getValueSizeInBits(); | ||||||
| 3364 | |||||||
| 3365 | // If the comparison constant has bits in the upper part, the | ||||||
| 3366 | // zero-extended value could never match. | ||||||
| 3367 | if (C1.intersects(APInt::getHighBitsSet(C1.getBitWidth(), | ||||||
| 3368 | C1.getBitWidth() - InSize))) { | ||||||
| 3369 | switch (Cond) { | ||||||
| 3370 | case ISD::SETUGT: | ||||||
| 3371 | case ISD::SETUGE: | ||||||
| 3372 | case ISD::SETEQ: | ||||||
| 3373 | return DAG.getConstant(0, dl, VT); | ||||||
| 3374 | case ISD::SETULT: | ||||||
| 3375 | case ISD::SETULE: | ||||||
| 3376 | case ISD::SETNE: | ||||||
| 3377 | return DAG.getConstant(1, dl, VT); | ||||||
| 3378 | case ISD::SETGT: | ||||||
| 3379 | case ISD::SETGE: | ||||||
| 3380 | // True if the sign bit of C1 is set. | ||||||
| 3381 | return DAG.getConstant(C1.isNegative(), dl, VT); | ||||||
| 3382 | case ISD::SETLT: | ||||||
| 3383 | case ISD::SETLE: | ||||||
| 3384 | // True if the sign bit of C1 isn't set. | ||||||
| 3385 | return DAG.getConstant(C1.isNonNegative(), dl, VT); | ||||||
| 3386 | default: | ||||||
| 3387 | break; | ||||||
| 3388 | } | ||||||
| 3389 | } | ||||||
| 3390 | |||||||
| 3391 | // Otherwise, we can perform the comparison with the low bits. | ||||||
| 3392 | switch (Cond) { | ||||||
| 3393 | case ISD::SETEQ: | ||||||
| 3394 | case ISD::SETNE: | ||||||
| 3395 | case ISD::SETUGT: | ||||||
| 3396 | case ISD::SETUGE: | ||||||
| 3397 | case ISD::SETULT: | ||||||
| 3398 | case ISD::SETULE: { | ||||||
| 3399 | EVT newVT = N0.getOperand(0).getValueType(); | ||||||
| 3400 | if (DCI.isBeforeLegalizeOps() || | ||||||
| 3401 | (isOperationLegal(ISD::SETCC, newVT) && | ||||||
| 3402 | isCondCodeLegal(Cond, newVT.getSimpleVT()))) { | ||||||
| 3403 | EVT NewSetCCVT = getSetCCResultType(Layout, *DAG.getContext(), newVT); | ||||||
| 3404 | SDValue NewConst = DAG.getConstant(C1.trunc(InSize), dl, newVT); | ||||||
| 3405 | |||||||
| 3406 | SDValue NewSetCC = DAG.getSetCC(dl, NewSetCCVT, N0.getOperand(0), | ||||||
| 3407 | NewConst, Cond); | ||||||
| 3408 | return DAG.getBoolExtOrTrunc(NewSetCC, dl, VT, N0.getValueType()); | ||||||
| 3409 | } | ||||||
| 3410 | break; | ||||||
| 3411 | } | ||||||
| 3412 | default: | ||||||
| 3413 | break; // todo, be more careful with signed comparisons | ||||||
| 3414 | } | ||||||
| 3415 | } else if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG && | ||||||
| 3416 | (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { | ||||||
| 3417 | EVT ExtSrcTy = cast<VTSDNode>(N0.getOperand(1))->getVT(); | ||||||
| 3418 | unsigned ExtSrcTyBits = ExtSrcTy.getSizeInBits(); | ||||||
| 3419 | EVT ExtDstTy = N0.getValueType(); | ||||||
| 3420 | unsigned ExtDstTyBits = ExtDstTy.getSizeInBits(); | ||||||
| 3421 | |||||||
| 3422 | // If the constant doesn't fit into the number of bits for the source of | ||||||
| 3423 | // the sign extension, it is impossible for both sides to be equal. | ||||||
| 3424 | if (C1.getMinSignedBits() > ExtSrcTyBits) | ||||||
| 3425 | return DAG.getConstant(Cond == ISD::SETNE, dl, VT); | ||||||
| 3426 | |||||||
| 3427 | SDValue ZextOp; | ||||||
| 3428 | EVT Op0Ty = N0.getOperand(0).getValueType(); | ||||||
| 3429 | if (Op0Ty == ExtSrcTy) { | ||||||
| 3430 | ZextOp = N0.getOperand(0); | ||||||
| 3431 | } else { | ||||||
| 3432 | APInt Imm = APInt::getLowBitsSet(ExtDstTyBits, ExtSrcTyBits); | ||||||
| 3433 | ZextOp = DAG.getNode(ISD::AND, dl, Op0Ty, N0.getOperand(0), | ||||||
| 3434 | DAG.getConstant(Imm, dl, Op0Ty)); | ||||||
| 3435 | } | ||||||
| 3436 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3437 | DCI.AddToWorklist(ZextOp.getNode()); | ||||||
| 3438 | // Otherwise, make this a use of a zext. | ||||||
| 3439 | return DAG.getSetCC(dl, VT, ZextOp, | ||||||
| 3440 | DAG.getConstant(C1 & APInt::getLowBitsSet( | ||||||
| 3441 | ExtDstTyBits, | ||||||
| 3442 | ExtSrcTyBits), | ||||||
| 3443 | dl, ExtDstTy), | ||||||
| 3444 | Cond); | ||||||
| 3445 | } else if ((N1C->isNullValue() || N1C->isOne()) && | ||||||
| 3446 | (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { | ||||||
| 3447 | // SETCC (SETCC), [0|1], [EQ|NE] -> SETCC | ||||||
| 3448 | if (N0.getOpcode() == ISD::SETCC && | ||||||
| 3449 | isTypeLegal(VT) && VT.bitsLE(N0.getValueType()) && | ||||||
| 3450 | (N0.getValueType() == MVT::i1 || | ||||||
| 3451 | getBooleanContents(N0.getOperand(0).getValueType()) == | ||||||
| 3452 | ZeroOrOneBooleanContent)) { | ||||||
| 3453 | bool TrueWhenTrue = (Cond == ISD::SETEQ) ^ (!N1C->isOne()); | ||||||
| 3454 | if (TrueWhenTrue) | ||||||
| 3455 | return DAG.getNode(ISD::TRUNCATE, dl, VT, N0); | ||||||
| 3456 | // Invert the condition. | ||||||
| 3457 | ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get(); | ||||||
| 3458 | CC = ISD::getSetCCInverse(CC, N0.getOperand(0).getValueType()); | ||||||
| 3459 | if (DCI.isBeforeLegalizeOps() || | ||||||
| 3460 | isCondCodeLegal(CC, N0.getOperand(0).getSimpleValueType())) | ||||||
| 3461 | return DAG.getSetCC(dl, VT, N0.getOperand(0), N0.getOperand(1), CC); | ||||||
| 3462 | } | ||||||
| 3463 | |||||||
| 3464 | if ((N0.getOpcode() == ISD::XOR || | ||||||
| 3465 | (N0.getOpcode() == ISD::AND && | ||||||
| 3466 | N0.getOperand(0).getOpcode() == ISD::XOR && | ||||||
| 3467 | N0.getOperand(1) == N0.getOperand(0).getOperand(1))) && | ||||||
| 3468 | isa<ConstantSDNode>(N0.getOperand(1)) && | ||||||
| 3469 | cast<ConstantSDNode>(N0.getOperand(1))->isOne()) { | ||||||
| 3470 | // If this is (X^1) == 0/1, swap the RHS and eliminate the xor. We | ||||||
| 3471 | // can only do this if the top bits are known zero. | ||||||
| 3472 | unsigned BitWidth = N0.getValueSizeInBits(); | ||||||
| 3473 | if (DAG.MaskedValueIsZero(N0, | ||||||
| 3474 | APInt::getHighBitsSet(BitWidth, | ||||||
| 3475 | BitWidth-1))) { | ||||||
| 3476 | // Okay, get the un-inverted input value. | ||||||
| 3477 | SDValue Val; | ||||||
| 3478 | if (N0.getOpcode() == ISD::XOR) { | ||||||
| 3479 | Val = N0.getOperand(0); | ||||||
| 3480 | } else { | ||||||
| 3481 | assert(N0.getOpcode() == ISD::AND &&((N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode () == ISD::XOR) ? static_cast<void> (0) : __assert_fail ("N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::XOR" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3482, __PRETTY_FUNCTION__)) | ||||||
| 3482 | N0.getOperand(0).getOpcode() == ISD::XOR)((N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode () == ISD::XOR) ? static_cast<void> (0) : __assert_fail ("N0.getOpcode() == ISD::AND && N0.getOperand(0).getOpcode() == ISD::XOR" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3482, __PRETTY_FUNCTION__)); | ||||||
| 3483 | // ((X^1)&1)^1 -> X & 1 | ||||||
| 3484 | Val = DAG.getNode(ISD::AND, dl, N0.getValueType(), | ||||||
| 3485 | N0.getOperand(0).getOperand(0), | ||||||
| 3486 | N0.getOperand(1)); | ||||||
| 3487 | } | ||||||
| 3488 | |||||||
| 3489 | return DAG.getSetCC(dl, VT, Val, N1, | ||||||
| 3490 | Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); | ||||||
| 3491 | } | ||||||
| 3492 | } else if (N1C->isOne()) { | ||||||
| 3493 | SDValue Op0 = N0; | ||||||
| 3494 | if (Op0.getOpcode() == ISD::TRUNCATE) | ||||||
| 3495 | Op0 = Op0.getOperand(0); | ||||||
| 3496 | |||||||
| 3497 | if ((Op0.getOpcode() == ISD::XOR) && | ||||||
| 3498 | Op0.getOperand(0).getOpcode() == ISD::SETCC && | ||||||
| 3499 | Op0.getOperand(1).getOpcode() == ISD::SETCC) { | ||||||
| 3500 | SDValue XorLHS = Op0.getOperand(0); | ||||||
| 3501 | SDValue XorRHS = Op0.getOperand(1); | ||||||
| 3502 | // Ensure that the input setccs return an i1 type or 0/1 value. | ||||||
| 3503 | if (Op0.getValueType() == MVT::i1 || | ||||||
| 3504 | (getBooleanContents(XorLHS.getOperand(0).getValueType()) == | ||||||
| 3505 | ZeroOrOneBooleanContent && | ||||||
| 3506 | getBooleanContents(XorRHS.getOperand(0).getValueType()) == | ||||||
| 3507 | ZeroOrOneBooleanContent)) { | ||||||
| 3508 | // (xor (setcc), (setcc)) == / != 1 -> (setcc) != / == (setcc) | ||||||
| 3509 | Cond = (Cond == ISD::SETEQ) ? ISD::SETNE : ISD::SETEQ; | ||||||
| 3510 | return DAG.getSetCC(dl, VT, XorLHS, XorRHS, Cond); | ||||||
| 3511 | } | ||||||
| 3512 | } | ||||||
| 3513 | if (Op0.getOpcode() == ISD::AND && | ||||||
| 3514 | isa<ConstantSDNode>(Op0.getOperand(1)) && | ||||||
| 3515 | cast<ConstantSDNode>(Op0.getOperand(1))->isOne()) { | ||||||
| 3516 | // If this is (X&1) == / != 1, normalize it to (X&1) != / == 0. | ||||||
| 3517 | if (Op0.getValueType().bitsGT(VT)) | ||||||
| 3518 | Op0 = DAG.getNode(ISD::AND, dl, VT, | ||||||
| 3519 | DAG.getNode(ISD::TRUNCATE, dl, VT, Op0.getOperand(0)), | ||||||
| 3520 | DAG.getConstant(1, dl, VT)); | ||||||
| 3521 | else if (Op0.getValueType().bitsLT(VT)) | ||||||
| 3522 | Op0 = DAG.getNode(ISD::AND, dl, VT, | ||||||
| 3523 | DAG.getNode(ISD::ANY_EXTEND, dl, VT, Op0.getOperand(0)), | ||||||
| 3524 | DAG.getConstant(1, dl, VT)); | ||||||
| 3525 | |||||||
| 3526 | return DAG.getSetCC(dl, VT, Op0, | ||||||
| 3527 | DAG.getConstant(0, dl, Op0.getValueType()), | ||||||
| 3528 | Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); | ||||||
| 3529 | } | ||||||
| 3530 | if (Op0.getOpcode() == ISD::AssertZext && | ||||||
| 3531 | cast<VTSDNode>(Op0.getOperand(1))->getVT() == MVT::i1) | ||||||
| 3532 | return DAG.getSetCC(dl, VT, Op0, | ||||||
| 3533 | DAG.getConstant(0, dl, Op0.getValueType()), | ||||||
| 3534 | Cond == ISD::SETEQ ? ISD::SETNE : ISD::SETEQ); | ||||||
| 3535 | } | ||||||
| 3536 | } | ||||||
| 3537 | |||||||
| 3538 | // Given: | ||||||
| 3539 | // icmp eq/ne (urem %x, %y), 0 | ||||||
| 3540 | // Iff %x has 0 or 1 bits set, and %y has at least 2 bits set, omit 'urem': | ||||||
| 3541 | // icmp eq/ne %x, 0 | ||||||
| 3542 | if (N0.getOpcode() == ISD::UREM && N1C->isNullValue() && | ||||||
| 3543 | (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { | ||||||
| 3544 | KnownBits XKnown = DAG.computeKnownBits(N0.getOperand(0)); | ||||||
| 3545 | KnownBits YKnown = DAG.computeKnownBits(N0.getOperand(1)); | ||||||
| 3546 | if (XKnown.countMaxPopulation() == 1 && YKnown.countMinPopulation() >= 2) | ||||||
| 3547 | return DAG.getSetCC(dl, VT, N0.getOperand(0), N1, Cond); | ||||||
| 3548 | } | ||||||
| 3549 | |||||||
| 3550 | if (SDValue V = | ||||||
| 3551 | optimizeSetCCOfSignedTruncationCheck(VT, N0, N1, Cond, DCI, dl)) | ||||||
| 3552 | return V; | ||||||
| 3553 | } | ||||||
| 3554 | |||||||
| 3555 | // These simplifications apply to splat vectors as well. | ||||||
| 3556 | // TODO: Handle more splat vector cases. | ||||||
| 3557 | if (auto *N1C = isConstOrConstSplat(N1)) { | ||||||
| 3558 | const APInt &C1 = N1C->getAPIntValue(); | ||||||
| 3559 | |||||||
| 3560 | APInt MinVal, MaxVal; | ||||||
| 3561 | unsigned OperandBitSize = N1C->getValueType(0).getScalarSizeInBits(); | ||||||
| 3562 | if (ISD::isSignedIntSetCC(Cond)) { | ||||||
| 3563 | MinVal = APInt::getSignedMinValue(OperandBitSize); | ||||||
| 3564 | MaxVal = APInt::getSignedMaxValue(OperandBitSize); | ||||||
| 3565 | } else { | ||||||
| 3566 | MinVal = APInt::getMinValue(OperandBitSize); | ||||||
| 3567 | MaxVal = APInt::getMaxValue(OperandBitSize); | ||||||
| 3568 | } | ||||||
| 3569 | |||||||
| 3570 | // Canonicalize GE/LE comparisons to use GT/LT comparisons. | ||||||
| 3571 | if (Cond == ISD::SETGE || Cond == ISD::SETUGE) { | ||||||
| 3572 | // X >= MIN --> true | ||||||
| 3573 | if (C1 == MinVal) | ||||||
| 3574 | return DAG.getBoolConstant(true, dl, VT, OpVT); | ||||||
| 3575 | |||||||
| 3576 | if (!VT.isVector()) { // TODO: Support this for vectors. | ||||||
| 3577 | // X >= C0 --> X > (C0 - 1) | ||||||
| 3578 | APInt C = C1 - 1; | ||||||
| 3579 | ISD::CondCode NewCC = (Cond == ISD::SETGE) ? ISD::SETGT : ISD::SETUGT; | ||||||
| 3580 | if ((DCI.isBeforeLegalizeOps() || | ||||||
| 3581 | isCondCodeLegal(NewCC, VT.getSimpleVT())) && | ||||||
| 3582 | (!N1C->isOpaque() || (C.getBitWidth() <= 64 && | ||||||
| 3583 | isLegalICmpImmediate(C.getSExtValue())))) { | ||||||
| 3584 | return DAG.getSetCC(dl, VT, N0, | ||||||
| 3585 | DAG.getConstant(C, dl, N1.getValueType()), | ||||||
| 3586 | NewCC); | ||||||
| 3587 | } | ||||||
| 3588 | } | ||||||
| 3589 | } | ||||||
| 3590 | |||||||
| 3591 | if (Cond == ISD::SETLE || Cond == ISD::SETULE) { | ||||||
| 3592 | // X <= MAX --> true | ||||||
| 3593 | if (C1 == MaxVal) | ||||||
| 3594 | return DAG.getBoolConstant(true, dl, VT, OpVT); | ||||||
| 3595 | |||||||
| 3596 | // X <= C0 --> X < (C0 + 1) | ||||||
| 3597 | if (!VT.isVector()) { // TODO: Support this for vectors. | ||||||
| 3598 | APInt C = C1 + 1; | ||||||
| 3599 | ISD::CondCode NewCC = (Cond == ISD::SETLE) ? ISD::SETLT : ISD::SETULT; | ||||||
| 3600 | if ((DCI.isBeforeLegalizeOps() || | ||||||
| 3601 | isCondCodeLegal(NewCC, VT.getSimpleVT())) && | ||||||
| 3602 | (!N1C->isOpaque() || (C.getBitWidth() <= 64 && | ||||||
| 3603 | isLegalICmpImmediate(C.getSExtValue())))) { | ||||||
| 3604 | return DAG.getSetCC(dl, VT, N0, | ||||||
| 3605 | DAG.getConstant(C, dl, N1.getValueType()), | ||||||
| 3606 | NewCC); | ||||||
| 3607 | } | ||||||
| 3608 | } | ||||||
| 3609 | } | ||||||
| 3610 | |||||||
| 3611 | if (Cond == ISD::SETLT || Cond == ISD::SETULT) { | ||||||
| 3612 | if (C1 == MinVal) | ||||||
| 3613 | return DAG.getBoolConstant(false, dl, VT, OpVT); // X < MIN --> false | ||||||
| 3614 | |||||||
| 3615 | // TODO: Support this for vectors after legalize ops. | ||||||
| 3616 | if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { | ||||||
| 3617 | // Canonicalize setlt X, Max --> setne X, Max | ||||||
| 3618 | if (C1 == MaxVal) | ||||||
| 3619 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); | ||||||
| 3620 | |||||||
| 3621 | // If we have setult X, 1, turn it into seteq X, 0 | ||||||
| 3622 | if (C1 == MinVal+1) | ||||||
| 3623 | return DAG.getSetCC(dl, VT, N0, | ||||||
| 3624 | DAG.getConstant(MinVal, dl, N0.getValueType()), | ||||||
| 3625 | ISD::SETEQ); | ||||||
| 3626 | } | ||||||
| 3627 | } | ||||||
| 3628 | |||||||
| 3629 | if (Cond == ISD::SETGT || Cond == ISD::SETUGT) { | ||||||
| 3630 | if (C1 == MaxVal) | ||||||
| 3631 | return DAG.getBoolConstant(false, dl, VT, OpVT); // X > MAX --> false | ||||||
| 3632 | |||||||
| 3633 | // TODO: Support this for vectors after legalize ops. | ||||||
| 3634 | if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { | ||||||
| 3635 | // Canonicalize setgt X, Min --> setne X, Min | ||||||
| 3636 | if (C1 == MinVal) | ||||||
| 3637 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETNE); | ||||||
| 3638 | |||||||
| 3639 | // If we have setugt X, Max-1, turn it into seteq X, Max | ||||||
| 3640 | if (C1 == MaxVal-1) | ||||||
| 3641 | return DAG.getSetCC(dl, VT, N0, | ||||||
| 3642 | DAG.getConstant(MaxVal, dl, N0.getValueType()), | ||||||
| 3643 | ISD::SETEQ); | ||||||
| 3644 | } | ||||||
| 3645 | } | ||||||
| 3646 | |||||||
| 3647 | if (Cond == ISD::SETEQ || Cond == ISD::SETNE) { | ||||||
| 3648 | // (X & (C l>>/<< Y)) ==/!= 0 --> ((X <</l>> Y) & C) ==/!= 0 | ||||||
| 3649 | if (C1.isNullValue()) | ||||||
| 3650 | if (SDValue CC = optimizeSetCCByHoistingAndByConstFromLogicalShift( | ||||||
| 3651 | VT, N0, N1, Cond, DCI, dl)) | ||||||
| 3652 | return CC; | ||||||
| 3653 | } | ||||||
| 3654 | |||||||
| 3655 | // If we have "setcc X, C0", check to see if we can shrink the immediate | ||||||
| 3656 | // by changing cc. | ||||||
| 3657 | // TODO: Support this for vectors after legalize ops. | ||||||
| 3658 | if (!VT.isVector() || DCI.isBeforeLegalizeOps()) { | ||||||
| 3659 | // SETUGT X, SINTMAX -> SETLT X, 0 | ||||||
| 3660 | if (Cond == ISD::SETUGT && | ||||||
| 3661 | C1 == APInt::getSignedMaxValue(OperandBitSize)) | ||||||
| 3662 | return DAG.getSetCC(dl, VT, N0, | ||||||
| 3663 | DAG.getConstant(0, dl, N1.getValueType()), | ||||||
| 3664 | ISD::SETLT); | ||||||
| 3665 | |||||||
| 3666 | // SETULT X, SINTMIN -> SETGT X, -1 | ||||||
| 3667 | if (Cond == ISD::SETULT && | ||||||
| 3668 | C1 == APInt::getSignedMinValue(OperandBitSize)) { | ||||||
| 3669 | SDValue ConstMinusOne = | ||||||
| 3670 | DAG.getConstant(APInt::getAllOnesValue(OperandBitSize), dl, | ||||||
| 3671 | N1.getValueType()); | ||||||
| 3672 | return DAG.getSetCC(dl, VT, N0, ConstMinusOne, ISD::SETGT); | ||||||
| 3673 | } | ||||||
| 3674 | } | ||||||
| 3675 | } | ||||||
| 3676 | |||||||
| 3677 | // Back to non-vector simplifications. | ||||||
| 3678 | // TODO: Can we do these for vector splats? | ||||||
| 3679 | if (auto *N1C = dyn_cast<ConstantSDNode>(N1.getNode())) { | ||||||
| 3680 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | ||||||
| 3681 | const APInt &C1 = N1C->getAPIntValue(); | ||||||
| 3682 | EVT ShValTy = N0.getValueType(); | ||||||
| 3683 | |||||||
| 3684 | // Fold bit comparisons when we can. | ||||||
| 3685 | if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && | ||||||
| 3686 | (VT == ShValTy || (isTypeLegal(VT) && VT.bitsLE(ShValTy))) && | ||||||
| 3687 | N0.getOpcode() == ISD::AND) { | ||||||
| 3688 | if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | ||||||
| 3689 | EVT ShiftTy = | ||||||
| 3690 | getShiftAmountTy(ShValTy, Layout, !DCI.isBeforeLegalize()); | ||||||
| 3691 | if (Cond == ISD::SETNE && C1 == 0) {// (X & 8) != 0 --> (X & 8) >> 3 | ||||||
| 3692 | // Perform the xform if the AND RHS is a single bit. | ||||||
| 3693 | unsigned ShCt = AndRHS->getAPIntValue().logBase2(); | ||||||
| 3694 | if (AndRHS->getAPIntValue().isPowerOf2() && | ||||||
| 3695 | !TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) { | ||||||
| 3696 | return DAG.getNode(ISD::TRUNCATE, dl, VT, | ||||||
| 3697 | DAG.getNode(ISD::SRL, dl, ShValTy, N0, | ||||||
| 3698 | DAG.getConstant(ShCt, dl, ShiftTy))); | ||||||
| 3699 | } | ||||||
| 3700 | } else if (Cond == ISD::SETEQ && C1 == AndRHS->getAPIntValue()) { | ||||||
| 3701 | // (X & 8) == 8 --> (X & 8) >> 3 | ||||||
| 3702 | // Perform the xform if C1 is a single bit. | ||||||
| 3703 | unsigned ShCt = C1.logBase2(); | ||||||
| 3704 | if (C1.isPowerOf2() && | ||||||
| 3705 | !TLI.shouldAvoidTransformToShift(ShValTy, ShCt)) { | ||||||
| 3706 | return DAG.getNode(ISD::TRUNCATE, dl, VT, | ||||||
| 3707 | DAG.getNode(ISD::SRL, dl, ShValTy, N0, | ||||||
| 3708 | DAG.getConstant(ShCt, dl, ShiftTy))); | ||||||
| 3709 | } | ||||||
| 3710 | } | ||||||
| 3711 | } | ||||||
| 3712 | } | ||||||
| 3713 | |||||||
| 3714 | if (C1.getMinSignedBits() <= 64 && | ||||||
| 3715 | !isLegalICmpImmediate(C1.getSExtValue())) { | ||||||
| 3716 | EVT ShiftTy = getShiftAmountTy(ShValTy, Layout, !DCI.isBeforeLegalize()); | ||||||
| 3717 | // (X & -256) == 256 -> (X >> 8) == 1 | ||||||
| 3718 | if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && | ||||||
| 3719 | N0.getOpcode() == ISD::AND && N0.hasOneUse()) { | ||||||
| 3720 | if (auto *AndRHS = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | ||||||
| 3721 | const APInt &AndRHSC = AndRHS->getAPIntValue(); | ||||||
| 3722 | if ((-AndRHSC).isPowerOf2() && (AndRHSC & C1) == C1) { | ||||||
| 3723 | unsigned ShiftBits = AndRHSC.countTrailingZeros(); | ||||||
| 3724 | if (!TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) { | ||||||
| 3725 | SDValue Shift = | ||||||
| 3726 | DAG.getNode(ISD::SRL, dl, ShValTy, N0.getOperand(0), | ||||||
| 3727 | DAG.getConstant(ShiftBits, dl, ShiftTy)); | ||||||
| 3728 | SDValue CmpRHS = DAG.getConstant(C1.lshr(ShiftBits), dl, ShValTy); | ||||||
| 3729 | return DAG.getSetCC(dl, VT, Shift, CmpRHS, Cond); | ||||||
| 3730 | } | ||||||
| 3731 | } | ||||||
| 3732 | } | ||||||
| 3733 | } else if (Cond == ISD::SETULT || Cond == ISD::SETUGE || | ||||||
| 3734 | Cond == ISD::SETULE || Cond == ISD::SETUGT) { | ||||||
| 3735 | bool AdjOne = (Cond == ISD::SETULE || Cond == ISD::SETUGT); | ||||||
| 3736 | // X < 0x100000000 -> (X >> 32) < 1 | ||||||
| 3737 | // X >= 0x100000000 -> (X >> 32) >= 1 | ||||||
| 3738 | // X <= 0x0ffffffff -> (X >> 32) < 1 | ||||||
| 3739 | // X > 0x0ffffffff -> (X >> 32) >= 1 | ||||||
| 3740 | unsigned ShiftBits; | ||||||
| 3741 | APInt NewC = C1; | ||||||
| 3742 | ISD::CondCode NewCond = Cond; | ||||||
| 3743 | if (AdjOne) { | ||||||
| 3744 | ShiftBits = C1.countTrailingOnes(); | ||||||
| 3745 | NewC = NewC + 1; | ||||||
| 3746 | NewCond = (Cond == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE; | ||||||
| 3747 | } else { | ||||||
| 3748 | ShiftBits = C1.countTrailingZeros(); | ||||||
| 3749 | } | ||||||
| 3750 | NewC.lshrInPlace(ShiftBits); | ||||||
| 3751 | if (ShiftBits && NewC.getMinSignedBits() <= 64 && | ||||||
| 3752 | isLegalICmpImmediate(NewC.getSExtValue()) && | ||||||
| 3753 | !TLI.shouldAvoidTransformToShift(ShValTy, ShiftBits)) { | ||||||
| 3754 | SDValue Shift = DAG.getNode(ISD::SRL, dl, ShValTy, N0, | ||||||
| 3755 | DAG.getConstant(ShiftBits, dl, ShiftTy)); | ||||||
| 3756 | SDValue CmpRHS = DAG.getConstant(NewC, dl, ShValTy); | ||||||
| 3757 | return DAG.getSetCC(dl, VT, Shift, CmpRHS, NewCond); | ||||||
| 3758 | } | ||||||
| 3759 | } | ||||||
| 3760 | } | ||||||
| 3761 | } | ||||||
| 3762 | |||||||
| 3763 | if (!isa<ConstantFPSDNode>(N0) && isa<ConstantFPSDNode>(N1)) { | ||||||
| 3764 | auto *CFP = cast<ConstantFPSDNode>(N1); | ||||||
| 3765 | assert(!CFP->getValueAPF().isNaN() && "Unexpected NaN value")((!CFP->getValueAPF().isNaN() && "Unexpected NaN value" ) ? static_cast<void> (0) : __assert_fail ("!CFP->getValueAPF().isNaN() && \"Unexpected NaN value\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3765, __PRETTY_FUNCTION__)); | ||||||
| 3766 | |||||||
| 3767 | // Otherwise, we know the RHS is not a NaN. Simplify the node to drop the | ||||||
| 3768 | // constant if knowing that the operand is non-nan is enough. We prefer to | ||||||
| 3769 | // have SETO(x,x) instead of SETO(x, 0.0) because this avoids having to | ||||||
| 3770 | // materialize 0.0. | ||||||
| 3771 | if (Cond == ISD::SETO || Cond == ISD::SETUO) | ||||||
| 3772 | return DAG.getSetCC(dl, VT, N0, N0, Cond); | ||||||
| 3773 | |||||||
| 3774 | // setcc (fneg x), C -> setcc swap(pred) x, -C | ||||||
| 3775 | if (N0.getOpcode() == ISD::FNEG) { | ||||||
| 3776 | ISD::CondCode SwapCond = ISD::getSetCCSwappedOperands(Cond); | ||||||
| 3777 | if (DCI.isBeforeLegalizeOps() || | ||||||
| 3778 | isCondCodeLegal(SwapCond, N0.getSimpleValueType())) { | ||||||
| 3779 | SDValue NegN1 = DAG.getNode(ISD::FNEG, dl, N0.getValueType(), N1); | ||||||
| 3780 | return DAG.getSetCC(dl, VT, N0.getOperand(0), NegN1, SwapCond); | ||||||
| 3781 | } | ||||||
| 3782 | } | ||||||
| 3783 | |||||||
| 3784 | // If the condition is not legal, see if we can find an equivalent one | ||||||
| 3785 | // which is legal. | ||||||
| 3786 | if (!isCondCodeLegal(Cond, N0.getSimpleValueType())) { | ||||||
| 3787 | // If the comparison was an awkward floating-point == or != and one of | ||||||
| 3788 | // the comparison operands is infinity or negative infinity, convert the | ||||||
| 3789 | // condition to a less-awkward <= or >=. | ||||||
| 3790 | if (CFP->getValueAPF().isInfinity()) { | ||||||
| 3791 | if (CFP->getValueAPF().isNegative()) { | ||||||
| 3792 | if (Cond == ISD::SETOEQ && | ||||||
| 3793 | isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) | ||||||
| 3794 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLE); | ||||||
| 3795 | if (Cond == ISD::SETUEQ && | ||||||
| 3796 | isCondCodeLegal(ISD::SETOLE, N0.getSimpleValueType())) | ||||||
| 3797 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULE); | ||||||
| 3798 | if (Cond == ISD::SETUNE && | ||||||
| 3799 | isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) | ||||||
| 3800 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGT); | ||||||
| 3801 | if (Cond == ISD::SETONE && | ||||||
| 3802 | isCondCodeLegal(ISD::SETUGT, N0.getSimpleValueType())) | ||||||
| 3803 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGT); | ||||||
| 3804 | } else { | ||||||
| 3805 | if (Cond == ISD::SETOEQ && | ||||||
| 3806 | isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) | ||||||
| 3807 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOGE); | ||||||
| 3808 | if (Cond == ISD::SETUEQ && | ||||||
| 3809 | isCondCodeLegal(ISD::SETOGE, N0.getSimpleValueType())) | ||||||
| 3810 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETUGE); | ||||||
| 3811 | if (Cond == ISD::SETUNE && | ||||||
| 3812 | isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) | ||||||
| 3813 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETULT); | ||||||
| 3814 | if (Cond == ISD::SETONE && | ||||||
| 3815 | isCondCodeLegal(ISD::SETULT, N0.getSimpleValueType())) | ||||||
| 3816 | return DAG.getSetCC(dl, VT, N0, N1, ISD::SETOLT); | ||||||
| 3817 | } | ||||||
| 3818 | } | ||||||
| 3819 | } | ||||||
| 3820 | } | ||||||
| 3821 | |||||||
| 3822 | if (N0 == N1) { | ||||||
| 3823 | // The sext(setcc()) => setcc() optimization relies on the appropriate | ||||||
| 3824 | // constant being emitted. | ||||||
| 3825 | assert(!N0.getValueType().isInteger() &&((!N0.getValueType().isInteger() && "Integer types should be handled by FoldSetCC" ) ? static_cast<void> (0) : __assert_fail ("!N0.getValueType().isInteger() && \"Integer types should be handled by FoldSetCC\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3826, __PRETTY_FUNCTION__)) | ||||||
| 3826 | "Integer types should be handled by FoldSetCC")((!N0.getValueType().isInteger() && "Integer types should be handled by FoldSetCC" ) ? static_cast<void> (0) : __assert_fail ("!N0.getValueType().isInteger() && \"Integer types should be handled by FoldSetCC\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3826, __PRETTY_FUNCTION__)); | ||||||
| 3827 | |||||||
| 3828 | bool EqTrue = ISD::isTrueWhenEqual(Cond); | ||||||
| 3829 | unsigned UOF = ISD::getUnorderedFlavor(Cond); | ||||||
| 3830 | if (UOF == 2) // FP operators that are undefined on NaNs. | ||||||
| 3831 | return DAG.getBoolConstant(EqTrue, dl, VT, OpVT); | ||||||
| 3832 | if (UOF == unsigned(EqTrue)) | ||||||
| 3833 | return DAG.getBoolConstant(EqTrue, dl, VT, OpVT); | ||||||
| 3834 | // Otherwise, we can't fold it. However, we can simplify it to SETUO/SETO | ||||||
| 3835 | // if it is not already. | ||||||
| 3836 | ISD::CondCode NewCond = UOF == 0 ? ISD::SETO : ISD::SETUO; | ||||||
| 3837 | if (NewCond != Cond && | ||||||
| 3838 | (DCI.isBeforeLegalizeOps() || | ||||||
| 3839 | isCondCodeLegal(NewCond, N0.getSimpleValueType()))) | ||||||
| 3840 | return DAG.getSetCC(dl, VT, N0, N1, NewCond); | ||||||
| 3841 | } | ||||||
| 3842 | |||||||
| 3843 | if ((Cond == ISD::SETEQ || Cond == ISD::SETNE) && | ||||||
| 3844 | N0.getValueType().isInteger()) { | ||||||
| 3845 | if (N0.getOpcode() == ISD::ADD || N0.getOpcode() == ISD::SUB || | ||||||
| 3846 | N0.getOpcode() == ISD::XOR) { | ||||||
| 3847 | // Simplify (X+Y) == (X+Z) --> Y == Z | ||||||
| 3848 | if (N0.getOpcode() == N1.getOpcode()) { | ||||||
| 3849 | if (N0.getOperand(0) == N1.getOperand(0)) | ||||||
| 3850 | return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(1), Cond); | ||||||
| 3851 | if (N0.getOperand(1) == N1.getOperand(1)) | ||||||
| 3852 | return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(0), Cond); | ||||||
| 3853 | if (isCommutativeBinOp(N0.getOpcode())) { | ||||||
| 3854 | // If X op Y == Y op X, try other combinations. | ||||||
| 3855 | if (N0.getOperand(0) == N1.getOperand(1)) | ||||||
| 3856 | return DAG.getSetCC(dl, VT, N0.getOperand(1), N1.getOperand(0), | ||||||
| 3857 | Cond); | ||||||
| 3858 | if (N0.getOperand(1) == N1.getOperand(0)) | ||||||
| 3859 | return DAG.getSetCC(dl, VT, N0.getOperand(0), N1.getOperand(1), | ||||||
| 3860 | Cond); | ||||||
| 3861 | } | ||||||
| 3862 | } | ||||||
| 3863 | |||||||
| 3864 | // If RHS is a legal immediate value for a compare instruction, we need | ||||||
| 3865 | // to be careful about increasing register pressure needlessly. | ||||||
| 3866 | bool LegalRHSImm = false; | ||||||
| 3867 | |||||||
| 3868 | if (auto *RHSC = dyn_cast<ConstantSDNode>(N1)) { | ||||||
| 3869 | if (auto *LHSR = dyn_cast<ConstantSDNode>(N0.getOperand(1))) { | ||||||
| 3870 | // Turn (X+C1) == C2 --> X == C2-C1 | ||||||
| 3871 | if (N0.getOpcode() == ISD::ADD && N0.getNode()->hasOneUse()) { | ||||||
| 3872 | return DAG.getSetCC(dl, VT, N0.getOperand(0), | ||||||
| 3873 | DAG.getConstant(RHSC->getAPIntValue()- | ||||||
| 3874 | LHSR->getAPIntValue(), | ||||||
| 3875 | dl, N0.getValueType()), Cond); | ||||||
| 3876 | } | ||||||
| 3877 | |||||||
| 3878 | // Turn (X^C1) == C2 into X == C1^C2 iff X&~C1 = 0. | ||||||
| 3879 | if (N0.getOpcode() == ISD::XOR) | ||||||
| 3880 | // If we know that all of the inverted bits are zero, don't bother | ||||||
| 3881 | // performing the inversion. | ||||||
| 3882 | if (DAG.MaskedValueIsZero(N0.getOperand(0), ~LHSR->getAPIntValue())) | ||||||
| 3883 | return | ||||||
| 3884 | DAG.getSetCC(dl, VT, N0.getOperand(0), | ||||||
| 3885 | DAG.getConstant(LHSR->getAPIntValue() ^ | ||||||
| 3886 | RHSC->getAPIntValue(), | ||||||
| 3887 | dl, N0.getValueType()), | ||||||
| 3888 | Cond); | ||||||
| 3889 | } | ||||||
| 3890 | |||||||
| 3891 | // Turn (C1-X) == C2 --> X == C1-C2 | ||||||
| 3892 | if (auto *SUBC = dyn_cast<ConstantSDNode>(N0.getOperand(0))) { | ||||||
| 3893 | if (N0.getOpcode() == ISD::SUB && N0.getNode()->hasOneUse()) { | ||||||
| 3894 | return | ||||||
| 3895 | DAG.getSetCC(dl, VT, N0.getOperand(1), | ||||||
| 3896 | DAG.getConstant(SUBC->getAPIntValue() - | ||||||
| 3897 | RHSC->getAPIntValue(), | ||||||
| 3898 | dl, N0.getValueType()), | ||||||
| 3899 | Cond); | ||||||
| 3900 | } | ||||||
| 3901 | } | ||||||
| 3902 | |||||||
| 3903 | // Could RHSC fold directly into a compare? | ||||||
| 3904 | if (RHSC->getValueType(0).getSizeInBits() <= 64) | ||||||
| 3905 | LegalRHSImm = isLegalICmpImmediate(RHSC->getSExtValue()); | ||||||
| 3906 | } | ||||||
| 3907 | |||||||
| 3908 | // (X+Y) == X --> Y == 0 and similar folds. | ||||||
| 3909 | // Don't do this if X is an immediate that can fold into a cmp | ||||||
| 3910 | // instruction and X+Y has other uses. It could be an induction variable | ||||||
| 3911 | // chain, and the transform would increase register pressure. | ||||||
| 3912 | if (!LegalRHSImm || N0.hasOneUse()) | ||||||
| 3913 | if (SDValue V = foldSetCCWithBinOp(VT, N0, N1, Cond, dl, DCI)) | ||||||
| 3914 | return V; | ||||||
| 3915 | } | ||||||
| 3916 | |||||||
| 3917 | if (N1.getOpcode() == ISD::ADD || N1.getOpcode() == ISD::SUB || | ||||||
| 3918 | N1.getOpcode() == ISD::XOR) | ||||||
| 3919 | if (SDValue V = foldSetCCWithBinOp(VT, N1, N0, Cond, dl, DCI)) | ||||||
| 3920 | return V; | ||||||
| 3921 | |||||||
| 3922 | if (SDValue V = foldSetCCWithAnd(VT, N0, N1, Cond, dl, DCI)) | ||||||
| 3923 | return V; | ||||||
| 3924 | } | ||||||
| 3925 | |||||||
| 3926 | // Fold remainder of division by a constant. | ||||||
| 3927 | if ((N0.getOpcode() == ISD::UREM || N0.getOpcode() == ISD::SREM) && | ||||||
| 3928 | N0.hasOneUse() && (Cond == ISD::SETEQ || Cond == ISD::SETNE)) { | ||||||
| 3929 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | ||||||
| 3930 | |||||||
| 3931 | // When division is cheap or optimizing for minimum size, | ||||||
| 3932 | // fall through to DIVREM creation by skipping this fold. | ||||||
| 3933 | if (!isIntDivCheap(VT, Attr) && !Attr.hasFnAttribute(Attribute::MinSize)) { | ||||||
| 3934 | if (N0.getOpcode() == ISD::UREM) { | ||||||
| 3935 | if (SDValue Folded = buildUREMEqFold(VT, N0, N1, Cond, DCI, dl)) | ||||||
| 3936 | return Folded; | ||||||
| 3937 | } else if (N0.getOpcode() == ISD::SREM) { | ||||||
| 3938 | if (SDValue Folded = buildSREMEqFold(VT, N0, N1, Cond, DCI, dl)) | ||||||
| 3939 | return Folded; | ||||||
| 3940 | } | ||||||
| 3941 | } | ||||||
| 3942 | } | ||||||
| 3943 | |||||||
| 3944 | // Fold away ALL boolean setcc's. | ||||||
| 3945 | if (N0.getValueType().getScalarType() == MVT::i1 && foldBooleans) { | ||||||
| 3946 | SDValue Temp; | ||||||
| 3947 | switch (Cond) { | ||||||
| 3948 | default: llvm_unreachable("Unknown integer setcc!")::llvm::llvm_unreachable_internal("Unknown integer setcc!", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 3948); | ||||||
| 3949 | case ISD::SETEQ: // X == Y -> ~(X^Y) | ||||||
| 3950 | Temp = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1); | ||||||
| 3951 | N0 = DAG.getNOT(dl, Temp, OpVT); | ||||||
| 3952 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3953 | DCI.AddToWorklist(Temp.getNode()); | ||||||
| 3954 | break; | ||||||
| 3955 | case ISD::SETNE: // X != Y --> (X^Y) | ||||||
| 3956 | N0 = DAG.getNode(ISD::XOR, dl, OpVT, N0, N1); | ||||||
| 3957 | break; | ||||||
| 3958 | case ISD::SETGT: // X >s Y --> X == 0 & Y == 1 --> ~X & Y | ||||||
| 3959 | case ISD::SETULT: // X <u Y --> X == 0 & Y == 1 --> ~X & Y | ||||||
| 3960 | Temp = DAG.getNOT(dl, N0, OpVT); | ||||||
| 3961 | N0 = DAG.getNode(ISD::AND, dl, OpVT, N1, Temp); | ||||||
| 3962 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3963 | DCI.AddToWorklist(Temp.getNode()); | ||||||
| 3964 | break; | ||||||
| 3965 | case ISD::SETLT: // X <s Y --> X == 1 & Y == 0 --> ~Y & X | ||||||
| 3966 | case ISD::SETUGT: // X >u Y --> X == 1 & Y == 0 --> ~Y & X | ||||||
| 3967 | Temp = DAG.getNOT(dl, N1, OpVT); | ||||||
| 3968 | N0 = DAG.getNode(ISD::AND, dl, OpVT, N0, Temp); | ||||||
| 3969 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3970 | DCI.AddToWorklist(Temp.getNode()); | ||||||
| 3971 | break; | ||||||
| 3972 | case ISD::SETULE: // X <=u Y --> X == 0 | Y == 1 --> ~X | Y | ||||||
| 3973 | case ISD::SETGE: // X >=s Y --> X == 0 | Y == 1 --> ~X | Y | ||||||
| 3974 | Temp = DAG.getNOT(dl, N0, OpVT); | ||||||
| 3975 | N0 = DAG.getNode(ISD::OR, dl, OpVT, N1, Temp); | ||||||
| 3976 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3977 | DCI.AddToWorklist(Temp.getNode()); | ||||||
| 3978 | break; | ||||||
| 3979 | case ISD::SETUGE: // X >=u Y --> X == 1 | Y == 0 --> ~Y | X | ||||||
| 3980 | case ISD::SETLE: // X <=s Y --> X == 1 | Y == 0 --> ~Y | X | ||||||
| 3981 | Temp = DAG.getNOT(dl, N1, OpVT); | ||||||
| 3982 | N0 = DAG.getNode(ISD::OR, dl, OpVT, N0, Temp); | ||||||
| 3983 | break; | ||||||
| 3984 | } | ||||||
| 3985 | if (VT.getScalarType() != MVT::i1) { | ||||||
| 3986 | if (!DCI.isCalledByLegalizer()) | ||||||
| 3987 | DCI.AddToWorklist(N0.getNode()); | ||||||
| 3988 | // FIXME: If running after legalize, we probably can't do this. | ||||||
| 3989 | ISD::NodeType ExtendCode = getExtendForContent(getBooleanContents(OpVT)); | ||||||
| 3990 | N0 = DAG.getNode(ExtendCode, dl, VT, N0); | ||||||
| 3991 | } | ||||||
| 3992 | return N0; | ||||||
| 3993 | } | ||||||
| 3994 | |||||||
| 3995 | // Could not fold it. | ||||||
| 3996 | return SDValue(); | ||||||
| 3997 | } | ||||||
| 3998 | |||||||
| 3999 | /// Returns true (and the GlobalValue and the offset) if the node is a | ||||||
| 4000 | /// GlobalAddress + offset. | ||||||
| 4001 | bool TargetLowering::isGAPlusOffset(SDNode *WN, const GlobalValue *&GA, | ||||||
| 4002 | int64_t &Offset) const { | ||||||
| 4003 | |||||||
| 4004 | SDNode *N = unwrapAddress(SDValue(WN, 0)).getNode(); | ||||||
| 4005 | |||||||
| 4006 | if (auto *GASD = dyn_cast<GlobalAddressSDNode>(N)) { | ||||||
| 4007 | GA = GASD->getGlobal(); | ||||||
| 4008 | Offset += GASD->getOffset(); | ||||||
| 4009 | return true; | ||||||
| 4010 | } | ||||||
| 4011 | |||||||
| 4012 | if (N->getOpcode() == ISD::ADD) { | ||||||
| 4013 | SDValue N1 = N->getOperand(0); | ||||||
| 4014 | SDValue N2 = N->getOperand(1); | ||||||
| 4015 | if (isGAPlusOffset(N1.getNode(), GA, Offset)) { | ||||||
| 4016 | if (auto *V = dyn_cast<ConstantSDNode>(N2)) { | ||||||
| 4017 | Offset += V->getSExtValue(); | ||||||
| 4018 | return true; | ||||||
| 4019 | } | ||||||
| 4020 | } else if (isGAPlusOffset(N2.getNode(), GA, Offset)) { | ||||||
| 4021 | if (auto *V = dyn_cast<ConstantSDNode>(N1)) { | ||||||
| 4022 | Offset += V->getSExtValue(); | ||||||
| 4023 | return true; | ||||||
| 4024 | } | ||||||
| 4025 | } | ||||||
| 4026 | } | ||||||
| 4027 | |||||||
| 4028 | return false; | ||||||
| 4029 | } | ||||||
| 4030 | |||||||
| 4031 | SDValue TargetLowering::PerformDAGCombine(SDNode *N, | ||||||
| 4032 | DAGCombinerInfo &DCI) const { | ||||||
| 4033 | // Default implementation: no optimization. | ||||||
| 4034 | return SDValue(); | ||||||
| 4035 | } | ||||||
| 4036 | |||||||
| 4037 | //===----------------------------------------------------------------------===// | ||||||
| 4038 | // Inline Assembler Implementation Methods | ||||||
| 4039 | //===----------------------------------------------------------------------===// | ||||||
| 4040 | |||||||
| 4041 | TargetLowering::ConstraintType | ||||||
| 4042 | TargetLowering::getConstraintType(StringRef Constraint) const { | ||||||
| 4043 | unsigned S = Constraint.size(); | ||||||
| 4044 | |||||||
| 4045 | if (S == 1) { | ||||||
| 4046 | switch (Constraint[0]) { | ||||||
| 4047 | default: break; | ||||||
| 4048 | case 'r': | ||||||
| 4049 | return C_RegisterClass; | ||||||
| 4050 | case 'm': // memory | ||||||
| 4051 | case 'o': // offsetable | ||||||
| 4052 | case 'V': // not offsetable | ||||||
| 4053 | return C_Memory; | ||||||
| 4054 | case 'n': // Simple Integer | ||||||
| 4055 | case 'E': // Floating Point Constant | ||||||
| 4056 | case 'F': // Floating Point Constant | ||||||
| 4057 | return C_Immediate; | ||||||
| 4058 | case 'i': // Simple Integer or Relocatable Constant | ||||||
| 4059 | case 's': // Relocatable Constant | ||||||
| 4060 | case 'p': // Address. | ||||||
| 4061 | case 'X': // Allow ANY value. | ||||||
| 4062 | case 'I': // Target registers. | ||||||
| 4063 | case 'J': | ||||||
| 4064 | case 'K': | ||||||
| 4065 | case 'L': | ||||||
| 4066 | case 'M': | ||||||
| 4067 | case 'N': | ||||||
| 4068 | case 'O': | ||||||
| 4069 | case 'P': | ||||||
| 4070 | case '<': | ||||||
| 4071 | case '>': | ||||||
| 4072 | return C_Other; | ||||||
| 4073 | } | ||||||
| 4074 | } | ||||||
| 4075 | |||||||
| 4076 | if (S > 1 && Constraint[0] == '{' && Constraint[S - 1] == '}') { | ||||||
| 4077 | if (S == 8 && Constraint.substr(1, 6) == "memory") // "{memory}" | ||||||
| 4078 | return C_Memory; | ||||||
| 4079 | return C_Register; | ||||||
| 4080 | } | ||||||
| 4081 | return C_Unknown; | ||||||
| 4082 | } | ||||||
| 4083 | |||||||
| 4084 | /// Try to replace an X constraint, which matches anything, with another that | ||||||
| 4085 | /// has more specific requirements based on the type of the corresponding | ||||||
| 4086 | /// operand. | ||||||
| 4087 | const char *TargetLowering::LowerXConstraint(EVT ConstraintVT) const { | ||||||
| 4088 | if (ConstraintVT.isInteger()) | ||||||
| 4089 | return "r"; | ||||||
| 4090 | if (ConstraintVT.isFloatingPoint()) | ||||||
| 4091 | return "f"; // works for many targets | ||||||
| 4092 | return nullptr; | ||||||
| 4093 | } | ||||||
| 4094 | |||||||
| 4095 | SDValue TargetLowering::LowerAsmOutputForConstraint( | ||||||
| 4096 | SDValue &Chain, SDValue &Flag, SDLoc DL, const AsmOperandInfo &OpInfo, | ||||||
| 4097 | SelectionDAG &DAG) const { | ||||||
| 4098 | return SDValue(); | ||||||
| 4099 | } | ||||||
| 4100 | |||||||
| 4101 | /// Lower the specified operand into the Ops vector. | ||||||
| 4102 | /// If it is invalid, don't add anything to Ops. | ||||||
| 4103 | void TargetLowering::LowerAsmOperandForConstraint(SDValue Op, | ||||||
| 4104 | std::string &Constraint, | ||||||
| 4105 | std::vector<SDValue> &Ops, | ||||||
| 4106 | SelectionDAG &DAG) const { | ||||||
| 4107 | |||||||
| 4108 | if (Constraint.length() > 1) return; | ||||||
| 4109 | |||||||
| 4110 | char ConstraintLetter = Constraint[0]; | ||||||
| 4111 | switch (ConstraintLetter) { | ||||||
| 4112 | default: break; | ||||||
| 4113 | case 'X': // Allows any operand; labels (basic block) use this. | ||||||
| 4114 | if (Op.getOpcode() == ISD::BasicBlock || | ||||||
| 4115 | Op.getOpcode() == ISD::TargetBlockAddress) { | ||||||
| 4116 | Ops.push_back(Op); | ||||||
| 4117 | return; | ||||||
| 4118 | } | ||||||
| 4119 | LLVM_FALLTHROUGH[[gnu::fallthrough]]; | ||||||
| 4120 | case 'i': // Simple Integer or Relocatable Constant | ||||||
| 4121 | case 'n': // Simple Integer | ||||||
| 4122 | case 's': { // Relocatable Constant | ||||||
| 4123 | |||||||
| 4124 | GlobalAddressSDNode *GA; | ||||||
| 4125 | ConstantSDNode *C; | ||||||
| 4126 | BlockAddressSDNode *BA; | ||||||
| 4127 | uint64_t Offset = 0; | ||||||
| 4128 | |||||||
| 4129 | // Match (GA) or (C) or (GA+C) or (GA-C) or ((GA+C)+C) or (((GA+C)+C)+C), | ||||||
| 4130 | // etc., since getelementpointer is variadic. We can't use | ||||||
| 4131 | // SelectionDAG::FoldSymbolOffset because it expects the GA to be accessible | ||||||
| 4132 | // while in this case the GA may be furthest from the root node which is | ||||||
| 4133 | // likely an ISD::ADD. | ||||||
| 4134 | while (1) { | ||||||
| 4135 | if ((GA = dyn_cast<GlobalAddressSDNode>(Op)) && ConstraintLetter != 'n') { | ||||||
| 4136 | Ops.push_back(DAG.getTargetGlobalAddress(GA->getGlobal(), SDLoc(Op), | ||||||
| 4137 | GA->getValueType(0), | ||||||
| 4138 | Offset + GA->getOffset())); | ||||||
| 4139 | return; | ||||||
| 4140 | } else if ((C = dyn_cast<ConstantSDNode>(Op)) && | ||||||
| 4141 | ConstraintLetter != 's') { | ||||||
| 4142 | // gcc prints these as sign extended. Sign extend value to 64 bits | ||||||
| 4143 | // now; without this it would get ZExt'd later in | ||||||
| 4144 | // ScheduleDAGSDNodes::EmitNode, which is very generic. | ||||||
| 4145 | bool IsBool = C->getConstantIntValue()->getBitWidth() == 1; | ||||||
| 4146 | BooleanContent BCont = getBooleanContents(MVT::i64); | ||||||
| 4147 | ISD::NodeType ExtOpc = IsBool ? getExtendForContent(BCont) | ||||||
| 4148 | : ISD::SIGN_EXTEND; | ||||||
| 4149 | int64_t ExtVal = ExtOpc == ISD::ZERO_EXTEND ? C->getZExtValue() | ||||||
| 4150 | : C->getSExtValue(); | ||||||
| 4151 | Ops.push_back(DAG.getTargetConstant(Offset + ExtVal, | ||||||
| 4152 | SDLoc(C), MVT::i64)); | ||||||
| 4153 | return; | ||||||
| 4154 | } else if ((BA = dyn_cast<BlockAddressSDNode>(Op)) && | ||||||
| 4155 | ConstraintLetter != 'n') { | ||||||
| 4156 | Ops.push_back(DAG.getTargetBlockAddress( | ||||||
| 4157 | BA->getBlockAddress(), BA->getValueType(0), | ||||||
| 4158 | Offset + BA->getOffset(), BA->getTargetFlags())); | ||||||
| 4159 | return; | ||||||
| 4160 | } else { | ||||||
| 4161 | const unsigned OpCode = Op.getOpcode(); | ||||||
| 4162 | if (OpCode == ISD::ADD || OpCode == ISD::SUB) { | ||||||
| 4163 | if ((C = dyn_cast<ConstantSDNode>(Op.getOperand(0)))) | ||||||
| 4164 | Op = Op.getOperand(1); | ||||||
| 4165 | // Subtraction is not commutative. | ||||||
| 4166 | else if (OpCode
| ||||||
| 4167 | (C = dyn_cast<ConstantSDNode>(Op.getOperand(1)))) | ||||||
| 4168 | Op = Op.getOperand(0); | ||||||
| 4169 | else | ||||||
| 4170 | return; | ||||||
| 4171 | Offset += (OpCode
| ||||||
| 4172 | continue; | ||||||
| 4173 | } | ||||||
| 4174 | } | ||||||
| 4175 | return; | ||||||
| 4176 | } | ||||||
| 4177 | break; | ||||||
| 4178 | } | ||||||
| 4179 | } | ||||||
| 4180 | } | ||||||
| 4181 | |||||||
| 4182 | std::pair<unsigned, const TargetRegisterClass *> | ||||||
| 4183 | TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *RI, | ||||||
| 4184 | StringRef Constraint, | ||||||
| 4185 | MVT VT) const { | ||||||
| 4186 | if (Constraint.empty() || Constraint[0] != '{') | ||||||
| 4187 | return std::make_pair(0u, static_cast<TargetRegisterClass *>(nullptr)); | ||||||
| 4188 | assert(*(Constraint.end() - 1) == '}' && "Not a brace enclosed constraint?")((*(Constraint.end() - 1) == '}' && "Not a brace enclosed constraint?" ) ? static_cast<void> (0) : __assert_fail ("*(Constraint.end() - 1) == '}' && \"Not a brace enclosed constraint?\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4188, __PRETTY_FUNCTION__)); | ||||||
| 4189 | |||||||
| 4190 | // Remove the braces from around the name. | ||||||
| 4191 | StringRef RegName(Constraint.data() + 1, Constraint.size() - 2); | ||||||
| 4192 | |||||||
| 4193 | std::pair<unsigned, const TargetRegisterClass *> R = | ||||||
| 4194 | std::make_pair(0u, static_cast<const TargetRegisterClass *>(nullptr)); | ||||||
| 4195 | |||||||
| 4196 | // Figure out which register class contains this reg. | ||||||
| 4197 | for (const TargetRegisterClass *RC : RI->regclasses()) { | ||||||
| 4198 | // If none of the value types for this register class are valid, we | ||||||
| 4199 | // can't use it. For example, 64-bit reg classes on 32-bit targets. | ||||||
| 4200 | if (!isLegalRC(*RI, *RC)) | ||||||
| 4201 | continue; | ||||||
| 4202 | |||||||
| 4203 | for (TargetRegisterClass::iterator I = RC->begin(), E = RC->end(); | ||||||
| 4204 | I != E; ++I) { | ||||||
| 4205 | if (RegName.equals_lower(RI->getRegAsmName(*I))) { | ||||||
| 4206 | std::pair<unsigned, const TargetRegisterClass *> S = | ||||||
| 4207 | std::make_pair(*I, RC); | ||||||
| 4208 | |||||||
| 4209 | // If this register class has the requested value type, return it, | ||||||
| 4210 | // otherwise keep searching and return the first class found | ||||||
| 4211 | // if no other is found which explicitly has the requested type. | ||||||
| 4212 | if (RI->isTypeLegalForClass(*RC, VT)) | ||||||
| 4213 | return S; | ||||||
| 4214 | if (!R.second) | ||||||
| 4215 | R = S; | ||||||
| 4216 | } | ||||||
| 4217 | } | ||||||
| 4218 | } | ||||||
| 4219 | |||||||
| 4220 | return R; | ||||||
| 4221 | } | ||||||
| 4222 | |||||||
| 4223 | //===----------------------------------------------------------------------===// | ||||||
| 4224 | // Constraint Selection. | ||||||
| 4225 | |||||||
| 4226 | /// Return true of this is an input operand that is a matching constraint like | ||||||
| 4227 | /// "4". | ||||||
| 4228 | bool TargetLowering::AsmOperandInfo::isMatchingInputConstraint() const { | ||||||
| 4229 | assert(!ConstraintCode.empty() && "No known constraint!")((!ConstraintCode.empty() && "No known constraint!") ? static_cast<void> (0) : __assert_fail ("!ConstraintCode.empty() && \"No known constraint!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4229, __PRETTY_FUNCTION__)); | ||||||
| 4230 | return isdigit(static_cast<unsigned char>(ConstraintCode[0])); | ||||||
| 4231 | } | ||||||
| 4232 | |||||||
| 4233 | /// If this is an input matching constraint, this method returns the output | ||||||
| 4234 | /// operand it matches. | ||||||
| 4235 | unsigned TargetLowering::AsmOperandInfo::getMatchedOperand() const { | ||||||
| 4236 | assert(!ConstraintCode.empty() && "No known constraint!")((!ConstraintCode.empty() && "No known constraint!") ? static_cast<void> (0) : __assert_fail ("!ConstraintCode.empty() && \"No known constraint!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4236, __PRETTY_FUNCTION__)); | ||||||
| 4237 | return atoi(ConstraintCode.c_str()); | ||||||
| 4238 | } | ||||||
| 4239 | |||||||
| 4240 | /// Split up the constraint string from the inline assembly value into the | ||||||
| 4241 | /// specific constraints and their prefixes, and also tie in the associated | ||||||
| 4242 | /// operand values. | ||||||
| 4243 | /// If this returns an empty vector, and if the constraint string itself | ||||||
| 4244 | /// isn't empty, there was an error parsing. | ||||||
| 4245 | TargetLowering::AsmOperandInfoVector | ||||||
| 4246 | TargetLowering::ParseConstraints(const DataLayout &DL, | ||||||
| 4247 | const TargetRegisterInfo *TRI, | ||||||
| 4248 | ImmutableCallSite CS) const { | ||||||
| 4249 | /// Information about all of the constraints. | ||||||
| 4250 | AsmOperandInfoVector ConstraintOperands; | ||||||
| 4251 | const InlineAsm *IA = cast<InlineAsm>(CS.getCalledValue()); | ||||||
| 4252 | unsigned maCount = 0; // Largest number of multiple alternative constraints. | ||||||
| 4253 | |||||||
| 4254 | // Do a prepass over the constraints, canonicalizing them, and building up the | ||||||
| 4255 | // ConstraintOperands list. | ||||||
| 4256 | unsigned ArgNo = 0; // ArgNo - The argument of the CallInst. | ||||||
| 4257 | unsigned ResNo = 0; // ResNo - The result number of the next output. | ||||||
| 4258 | |||||||
| 4259 | for (InlineAsm::ConstraintInfo &CI : IA->ParseConstraints()) { | ||||||
| 4260 | ConstraintOperands.emplace_back(std::move(CI)); | ||||||
| 4261 | AsmOperandInfo &OpInfo = ConstraintOperands.back(); | ||||||
| 4262 | |||||||
| 4263 | // Update multiple alternative constraint count. | ||||||
| 4264 | if (OpInfo.multipleAlternatives.size() > maCount) | ||||||
| 4265 | maCount = OpInfo.multipleAlternatives.size(); | ||||||
| 4266 | |||||||
| 4267 | OpInfo.ConstraintVT = MVT::Other; | ||||||
| 4268 | |||||||
| 4269 | // Compute the value type for each operand. | ||||||
| 4270 | switch (OpInfo.Type) { | ||||||
| 4271 | case InlineAsm::isOutput: | ||||||
| 4272 | // Indirect outputs just consume an argument. | ||||||
| 4273 | if (OpInfo.isIndirect) { | ||||||
| 4274 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); | ||||||
| 4275 | break; | ||||||
| 4276 | } | ||||||
| 4277 | |||||||
| 4278 | // The return value of the call is this value. As such, there is no | ||||||
| 4279 | // corresponding argument. | ||||||
| 4280 | assert(!CS.getType()->isVoidTy() &&((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4281, __PRETTY_FUNCTION__)) | ||||||
| 4281 | "Bad inline asm!")((!CS.getType()->isVoidTy() && "Bad inline asm!") ? static_cast<void> (0) : __assert_fail ("!CS.getType()->isVoidTy() && \"Bad inline asm!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4281, __PRETTY_FUNCTION__)); | ||||||
| 4282 | if (StructType *STy = dyn_cast<StructType>(CS.getType())) { | ||||||
| 4283 | OpInfo.ConstraintVT = | ||||||
| 4284 | getSimpleValueType(DL, STy->getElementType(ResNo)); | ||||||
| 4285 | } else { | ||||||
| 4286 | assert(ResNo == 0 && "Asm only has one result!")((ResNo == 0 && "Asm only has one result!") ? static_cast <void> (0) : __assert_fail ("ResNo == 0 && \"Asm only has one result!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4286, __PRETTY_FUNCTION__)); | ||||||
| 4287 | OpInfo.ConstraintVT = getSimpleValueType(DL, CS.getType()); | ||||||
| 4288 | } | ||||||
| 4289 | ++ResNo; | ||||||
| 4290 | break; | ||||||
| 4291 | case InlineAsm::isInput: | ||||||
| 4292 | OpInfo.CallOperandVal = const_cast<Value *>(CS.getArgument(ArgNo++)); | ||||||
| 4293 | break; | ||||||
| 4294 | case InlineAsm::isClobber: | ||||||
| 4295 | // Nothing to do. | ||||||
| 4296 | break; | ||||||
| 4297 | } | ||||||
| 4298 | |||||||
| 4299 | if (OpInfo.CallOperandVal) { | ||||||
| 4300 | llvm::Type *OpTy = OpInfo.CallOperandVal->getType(); | ||||||
| 4301 | if (OpInfo.isIndirect) { | ||||||
| 4302 | llvm::PointerType *PtrTy = dyn_cast<PointerType>(OpTy); | ||||||
| 4303 | if (!PtrTy) | ||||||
| 4304 | report_fatal_error("Indirect operand for inline asm not a pointer!"); | ||||||
| 4305 | OpTy = PtrTy->getElementType(); | ||||||
| 4306 | } | ||||||
| 4307 | |||||||
| 4308 | // Look for vector wrapped in a struct. e.g. { <16 x i8> }. | ||||||
| 4309 | if (StructType *STy = dyn_cast<StructType>(OpTy)) | ||||||
| 4310 | if (STy->getNumElements() == 1) | ||||||
| 4311 | OpTy = STy->getElementType(0); | ||||||
| 4312 | |||||||
| 4313 | // If OpTy is not a single value, it may be a struct/union that we | ||||||
| 4314 | // can tile with integers. | ||||||
| 4315 | if (!OpTy->isSingleValueType() && OpTy->isSized()) { | ||||||
| 4316 | unsigned BitSize = DL.getTypeSizeInBits(OpTy); | ||||||
| 4317 | switch (BitSize) { | ||||||
| 4318 | default: break; | ||||||
| 4319 | case 1: | ||||||
| 4320 | case 8: | ||||||
| 4321 | case 16: | ||||||
| 4322 | case 32: | ||||||
| 4323 | case 64: | ||||||
| 4324 | case 128: | ||||||
| 4325 | OpInfo.ConstraintVT = | ||||||
| 4326 | MVT::getVT(IntegerType::get(OpTy->getContext(), BitSize), true); | ||||||
| 4327 | break; | ||||||
| 4328 | } | ||||||
| 4329 | } else if (PointerType *PT = dyn_cast<PointerType>(OpTy)) { | ||||||
| 4330 | unsigned PtrSize = DL.getPointerSizeInBits(PT->getAddressSpace()); | ||||||
| 4331 | OpInfo.ConstraintVT = MVT::getIntegerVT(PtrSize); | ||||||
| 4332 | } else { | ||||||
| 4333 | OpInfo.ConstraintVT = MVT::getVT(OpTy, true); | ||||||
| 4334 | } | ||||||
| 4335 | } | ||||||
| 4336 | } | ||||||
| 4337 | |||||||
| 4338 | // If we have multiple alternative constraints, select the best alternative. | ||||||
| 4339 | if (!ConstraintOperands.empty()) { | ||||||
| 4340 | if (maCount) { | ||||||
| 4341 | unsigned bestMAIndex = 0; | ||||||
| 4342 | int bestWeight = -1; | ||||||
| 4343 | // weight: -1 = invalid match, and 0 = so-so match to 5 = good match. | ||||||
| 4344 | int weight = -1; | ||||||
| 4345 | unsigned maIndex; | ||||||
| 4346 | // Compute the sums of the weights for each alternative, keeping track | ||||||
| 4347 | // of the best (highest weight) one so far. | ||||||
| 4348 | for (maIndex = 0; maIndex < maCount; ++maIndex) { | ||||||
| 4349 | int weightSum = 0; | ||||||
| 4350 | for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); | ||||||
| 4351 | cIndex != eIndex; ++cIndex) { | ||||||
| 4352 | AsmOperandInfo &OpInfo = ConstraintOperands[cIndex]; | ||||||
| 4353 | if (OpInfo.Type == InlineAsm::isClobber) | ||||||
| 4354 | continue; | ||||||
| 4355 | |||||||
| 4356 | // If this is an output operand with a matching input operand, | ||||||
| 4357 | // look up the matching input. If their types mismatch, e.g. one | ||||||
| 4358 | // is an integer, the other is floating point, or their sizes are | ||||||
| 4359 | // different, flag it as an maCantMatch. | ||||||
| 4360 | if (OpInfo.hasMatchingInput()) { | ||||||
| 4361 | AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; | ||||||
| 4362 | if (OpInfo.ConstraintVT != Input.ConstraintVT) { | ||||||
| 4363 | if ((OpInfo.ConstraintVT.isInteger() != | ||||||
| 4364 | Input.ConstraintVT.isInteger()) || | ||||||
| 4365 | (OpInfo.ConstraintVT.getSizeInBits() != | ||||||
| 4366 | Input.ConstraintVT.getSizeInBits())) { | ||||||
| 4367 | weightSum = -1; // Can't match. | ||||||
| 4368 | break; | ||||||
| 4369 | } | ||||||
| 4370 | } | ||||||
| 4371 | } | ||||||
| 4372 | weight = getMultipleConstraintMatchWeight(OpInfo, maIndex); | ||||||
| 4373 | if (weight == -1) { | ||||||
| 4374 | weightSum = -1; | ||||||
| 4375 | break; | ||||||
| 4376 | } | ||||||
| 4377 | weightSum += weight; | ||||||
| 4378 | } | ||||||
| 4379 | // Update best. | ||||||
| 4380 | if (weightSum > bestWeight) { | ||||||
| 4381 | bestWeight = weightSum; | ||||||
| 4382 | bestMAIndex = maIndex; | ||||||
| 4383 | } | ||||||
| 4384 | } | ||||||
| 4385 | |||||||
| 4386 | // Now select chosen alternative in each constraint. | ||||||
| 4387 | for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); | ||||||
| 4388 | cIndex != eIndex; ++cIndex) { | ||||||
| 4389 | AsmOperandInfo &cInfo = ConstraintOperands[cIndex]; | ||||||
| 4390 | if (cInfo.Type == InlineAsm::isClobber) | ||||||
| 4391 | continue; | ||||||
| 4392 | cInfo.selectAlternative(bestMAIndex); | ||||||
| 4393 | } | ||||||
| 4394 | } | ||||||
| 4395 | } | ||||||
| 4396 | |||||||
| 4397 | // Check and hook up tied operands, choose constraint code to use. | ||||||
| 4398 | for (unsigned cIndex = 0, eIndex = ConstraintOperands.size(); | ||||||
| 4399 | cIndex != eIndex; ++cIndex) { | ||||||
| 4400 | AsmOperandInfo &OpInfo = ConstraintOperands[cIndex]; | ||||||
| 4401 | |||||||
| 4402 | // If this is an output operand with a matching input operand, look up the | ||||||
| 4403 | // matching input. If their types mismatch, e.g. one is an integer, the | ||||||
| 4404 | // other is floating point, or their sizes are different, flag it as an | ||||||
| 4405 | // error. | ||||||
| 4406 | if (OpInfo.hasMatchingInput()) { | ||||||
| 4407 | AsmOperandInfo &Input = ConstraintOperands[OpInfo.MatchingInput]; | ||||||
| 4408 | |||||||
| 4409 | if (OpInfo.ConstraintVT != Input.ConstraintVT) { | ||||||
| 4410 | std::pair<unsigned, const TargetRegisterClass *> MatchRC = | ||||||
| 4411 | getRegForInlineAsmConstraint(TRI, OpInfo.ConstraintCode, | ||||||
| 4412 | OpInfo.ConstraintVT); | ||||||
| 4413 | std::pair<unsigned, const TargetRegisterClass *> InputRC = | ||||||
| 4414 | getRegForInlineAsmConstraint(TRI, Input.ConstraintCode, | ||||||
| 4415 | Input.ConstraintVT); | ||||||
| 4416 | if ((OpInfo.ConstraintVT.isInteger() != | ||||||
| 4417 | Input.ConstraintVT.isInteger()) || | ||||||
| 4418 | (MatchRC.second != InputRC.second)) { | ||||||
| 4419 | report_fatal_error("Unsupported asm: input constraint" | ||||||
| 4420 | " with a matching output constraint of" | ||||||
| 4421 | " incompatible type!"); | ||||||
| 4422 | } | ||||||
| 4423 | } | ||||||
| 4424 | } | ||||||
| 4425 | } | ||||||
| 4426 | |||||||
| 4427 | return ConstraintOperands; | ||||||
| 4428 | } | ||||||
| 4429 | |||||||
| 4430 | /// Return an integer indicating how general CT is. | ||||||
| 4431 | static unsigned getConstraintGenerality(TargetLowering::ConstraintType CT) { | ||||||
| 4432 | switch (CT) { | ||||||
| 4433 | case TargetLowering::C_Immediate: | ||||||
| 4434 | case TargetLowering::C_Other: | ||||||
| 4435 | case TargetLowering::C_Unknown: | ||||||
| 4436 | return 0; | ||||||
| 4437 | case TargetLowering::C_Register: | ||||||
| 4438 | return 1; | ||||||
| 4439 | case TargetLowering::C_RegisterClass: | ||||||
| 4440 | return 2; | ||||||
| 4441 | case TargetLowering::C_Memory: | ||||||
| 4442 | return 3; | ||||||
| 4443 | } | ||||||
| 4444 | llvm_unreachable("Invalid constraint type")::llvm::llvm_unreachable_internal("Invalid constraint type", "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4444); | ||||||
| 4445 | } | ||||||
| 4446 | |||||||
| 4447 | /// Examine constraint type and operand type and determine a weight value. | ||||||
| 4448 | /// This object must already have been set up with the operand type | ||||||
| 4449 | /// and the current alternative constraint selected. | ||||||
| 4450 | TargetLowering::ConstraintWeight | ||||||
| 4451 | TargetLowering::getMultipleConstraintMatchWeight( | ||||||
| 4452 | AsmOperandInfo &info, int maIndex) const { | ||||||
| 4453 | InlineAsm::ConstraintCodeVector *rCodes; | ||||||
| 4454 | if (maIndex >= (int)info.multipleAlternatives.size()) | ||||||
| 4455 | rCodes = &info.Codes; | ||||||
| 4456 | else | ||||||
| 4457 | rCodes = &info.multipleAlternatives[maIndex].Codes; | ||||||
| 4458 | ConstraintWeight BestWeight = CW_Invalid; | ||||||
| 4459 | |||||||
| 4460 | // Loop over the options, keeping track of the most general one. | ||||||
| 4461 | for (unsigned i = 0, e = rCodes->size(); i != e; ++i) { | ||||||
| 4462 | ConstraintWeight weight = | ||||||
| 4463 | getSingleConstraintMatchWeight(info, (*rCodes)[i].c_str()); | ||||||
| 4464 | if (weight > BestWeight) | ||||||
| 4465 | BestWeight = weight; | ||||||
| 4466 | } | ||||||
| 4467 | |||||||
| 4468 | return BestWeight; | ||||||
| 4469 | } | ||||||
| 4470 | |||||||
| 4471 | /// Examine constraint type and operand type and determine a weight value. | ||||||
| 4472 | /// This object must already have been set up with the operand type | ||||||
| 4473 | /// and the current alternative constraint selected. | ||||||
| 4474 | TargetLowering::ConstraintWeight | ||||||
| 4475 | TargetLowering::getSingleConstraintMatchWeight( | ||||||
| 4476 | AsmOperandInfo &info, const char *constraint) const { | ||||||
| 4477 | ConstraintWeight weight = CW_Invalid; | ||||||
| 4478 | Value *CallOperandVal = info.CallOperandVal; | ||||||
| 4479 | // If we don't have a value, we can't do a match, | ||||||
| 4480 | // but allow it at the lowest weight. | ||||||
| 4481 | if (!CallOperandVal) | ||||||
| 4482 | return CW_Default; | ||||||
| 4483 | // Look at the constraint type. | ||||||
| 4484 | switch (*constraint) { | ||||||
| 4485 | case 'i': // immediate integer. | ||||||
| 4486 | case 'n': // immediate integer with a known value. | ||||||
| 4487 | if (isa<ConstantInt>(CallOperandVal)) | ||||||
| 4488 | weight = CW_Constant; | ||||||
| 4489 | break; | ||||||
| 4490 | case 's': // non-explicit intregal immediate. | ||||||
| 4491 | if (isa<GlobalValue>(CallOperandVal)) | ||||||
| 4492 | weight = CW_Constant; | ||||||
| 4493 | break; | ||||||
| 4494 | case 'E': // immediate float if host format. | ||||||
| 4495 | case 'F': // immediate float. | ||||||
| 4496 | if (isa<ConstantFP>(CallOperandVal)) | ||||||
| 4497 | weight = CW_Constant; | ||||||
| 4498 | break; | ||||||
| 4499 | case '<': // memory operand with autodecrement. | ||||||
| 4500 | case '>': // memory operand with autoincrement. | ||||||
| 4501 | case 'm': // memory operand. | ||||||
| 4502 | case 'o': // offsettable memory operand | ||||||
| 4503 | case 'V': // non-offsettable memory operand | ||||||
| 4504 | weight = CW_Memory; | ||||||
| 4505 | break; | ||||||
| 4506 | case 'r': // general register. | ||||||
| 4507 | case 'g': // general register, memory operand or immediate integer. | ||||||
| 4508 | // note: Clang converts "g" to "imr". | ||||||
| 4509 | if (CallOperandVal->getType()->isIntegerTy()) | ||||||
| 4510 | weight = CW_Register; | ||||||
| 4511 | break; | ||||||
| 4512 | case 'X': // any operand. | ||||||
| 4513 | default: | ||||||
| 4514 | weight = CW_Default; | ||||||
| 4515 | break; | ||||||
| 4516 | } | ||||||
| 4517 | return weight; | ||||||
| 4518 | } | ||||||
| 4519 | |||||||
| 4520 | /// If there are multiple different constraints that we could pick for this | ||||||
| 4521 | /// operand (e.g. "imr") try to pick the 'best' one. | ||||||
| 4522 | /// This is somewhat tricky: constraints fall into four classes: | ||||||
| 4523 | /// Other -> immediates and magic values | ||||||
| 4524 | /// Register -> one specific register | ||||||
| 4525 | /// RegisterClass -> a group of regs | ||||||
| 4526 | /// Memory -> memory | ||||||
| 4527 | /// Ideally, we would pick the most specific constraint possible: if we have | ||||||
| 4528 | /// something that fits into a register, we would pick it. The problem here | ||||||
| 4529 | /// is that if we have something that could either be in a register or in | ||||||
| 4530 | /// memory that use of the register could cause selection of *other* | ||||||
| 4531 | /// operands to fail: they might only succeed if we pick memory. Because of | ||||||
| 4532 | /// this the heuristic we use is: | ||||||
| 4533 | /// | ||||||
| 4534 | /// 1) If there is an 'other' constraint, and if the operand is valid for | ||||||
| 4535 | /// that constraint, use it. This makes us take advantage of 'i' | ||||||
| 4536 | /// constraints when available. | ||||||
| 4537 | /// 2) Otherwise, pick the most general constraint present. This prefers | ||||||
| 4538 | /// 'm' over 'r', for example. | ||||||
| 4539 | /// | ||||||
| 4540 | static void ChooseConstraint(TargetLowering::AsmOperandInfo &OpInfo, | ||||||
| 4541 | const TargetLowering &TLI, | ||||||
| 4542 | SDValue Op, SelectionDAG *DAG) { | ||||||
| 4543 | assert(OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options")((OpInfo.Codes.size() > 1 && "Doesn't have multiple constraint options" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.Codes.size() > 1 && \"Doesn't have multiple constraint options\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4543, __PRETTY_FUNCTION__)); | ||||||
| 4544 | unsigned BestIdx = 0; | ||||||
| 4545 | TargetLowering::ConstraintType BestType = TargetLowering::C_Unknown; | ||||||
| 4546 | int BestGenerality = -1; | ||||||
| 4547 | |||||||
| 4548 | // Loop over the options, keeping track of the most general one. | ||||||
| 4549 | for (unsigned i = 0, e = OpInfo.Codes.size(); i != e; ++i) { | ||||||
| 4550 | TargetLowering::ConstraintType CType = | ||||||
| 4551 | TLI.getConstraintType(OpInfo.Codes[i]); | ||||||
| 4552 | |||||||
| 4553 | // Indirect 'other' or 'immediate' constraints are not allowed. | ||||||
| 4554 | if (OpInfo.isIndirect && !(CType == TargetLowering::C_Memory || | ||||||
| 4555 | CType == TargetLowering::C_Register || | ||||||
| 4556 | CType == TargetLowering::C_RegisterClass)) | ||||||
| 4557 | continue; | ||||||
| 4558 | |||||||
| 4559 | // If this is an 'other' or 'immediate' constraint, see if the operand is | ||||||
| 4560 | // valid for it. For example, on X86 we might have an 'rI' constraint. If | ||||||
| 4561 | // the operand is an integer in the range [0..31] we want to use I (saving a | ||||||
| 4562 | // load of a register), otherwise we must use 'r'. | ||||||
| 4563 | if ((CType
| ||||||
| 4564 | CType == TargetLowering::C_Immediate) && Op.getNode()) { | ||||||
| 4565 | assert(OpInfo.Codes[i].size() == 1 &&((OpInfo.Codes[i].size() == 1 && "Unhandled multi-letter 'other' constraint" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.Codes[i].size() == 1 && \"Unhandled multi-letter 'other' constraint\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4566, __PRETTY_FUNCTION__)) | ||||||
| 4566 | "Unhandled multi-letter 'other' constraint")((OpInfo.Codes[i].size() == 1 && "Unhandled multi-letter 'other' constraint" ) ? static_cast<void> (0) : __assert_fail ("OpInfo.Codes[i].size() == 1 && \"Unhandled multi-letter 'other' constraint\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4566, __PRETTY_FUNCTION__)); | ||||||
| 4567 | std::vector<SDValue> ResultOps; | ||||||
| 4568 | TLI.LowerAsmOperandForConstraint(Op, OpInfo.Codes[i], | ||||||
| 4569 | ResultOps, *DAG); | ||||||
| 4570 | if (!ResultOps.empty()) { | ||||||
| 4571 | BestType = CType; | ||||||
| 4572 | BestIdx = i; | ||||||
| 4573 | break; | ||||||
| 4574 | } | ||||||
| 4575 | } | ||||||
| 4576 | |||||||
| 4577 | // Things with matching constraints can only be registers, per gcc | ||||||
| 4578 | // documentation. This mainly affects "g" constraints. | ||||||
| 4579 | if (CType == TargetLowering::C_Memory && OpInfo.hasMatchingInput()) | ||||||
| 4580 | continue; | ||||||
| 4581 | |||||||
| 4582 | // This constraint letter is more general than the previous one, use it. | ||||||
| 4583 | int Generality = getConstraintGenerality(CType); | ||||||
| 4584 | if (Generality > BestGenerality) { | ||||||
| 4585 | BestType = CType; | ||||||
| 4586 | BestIdx = i; | ||||||
| 4587 | BestGenerality = Generality; | ||||||
| 4588 | } | ||||||
| 4589 | } | ||||||
| 4590 | |||||||
| 4591 | OpInfo.ConstraintCode = OpInfo.Codes[BestIdx]; | ||||||
| 4592 | OpInfo.ConstraintType = BestType; | ||||||
| 4593 | } | ||||||
| 4594 | |||||||
| 4595 | /// Determines the constraint code and constraint type to use for the specific | ||||||
| 4596 | /// AsmOperandInfo, setting OpInfo.ConstraintCode and OpInfo.ConstraintType. | ||||||
| 4597 | void TargetLowering::ComputeConstraintToUse(AsmOperandInfo &OpInfo, | ||||||
| 4598 | SDValue Op, | ||||||
| 4599 | SelectionDAG *DAG) const { | ||||||
| 4600 | assert(!OpInfo.Codes.empty() && "Must have at least one constraint")((!OpInfo.Codes.empty() && "Must have at least one constraint" ) ? static_cast<void> (0) : __assert_fail ("!OpInfo.Codes.empty() && \"Must have at least one constraint\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4600, __PRETTY_FUNCTION__)); | ||||||
| |||||||
| 4601 | |||||||
| 4602 | // Single-letter constraints ('r') are very common. | ||||||
| 4603 | if (OpInfo.Codes.size() == 1) { | ||||||
| 4604 | OpInfo.ConstraintCode = OpInfo.Codes[0]; | ||||||
| 4605 | OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); | ||||||
| 4606 | } else { | ||||||
| 4607 | ChooseConstraint(OpInfo, *this, Op, DAG); | ||||||
| 4608 | } | ||||||
| 4609 | |||||||
| 4610 | // 'X' matches anything. | ||||||
| 4611 | if (OpInfo.ConstraintCode == "X" && OpInfo.CallOperandVal) { | ||||||
| 4612 | // Labels and constants are handled elsewhere ('X' is the only thing | ||||||
| 4613 | // that matches labels). For Functions, the type here is the type of | ||||||
| 4614 | // the result, which is not what we want to look at; leave them alone. | ||||||
| 4615 | Value *v = OpInfo.CallOperandVal; | ||||||
| 4616 | if (isa<BasicBlock>(v) || isa<ConstantInt>(v) || isa<Function>(v)) { | ||||||
| 4617 | OpInfo.CallOperandVal = v; | ||||||
| 4618 | return; | ||||||
| 4619 | } | ||||||
| 4620 | |||||||
| 4621 | if (Op.getNode() && Op.getOpcode() == ISD::TargetBlockAddress) | ||||||
| 4622 | return; | ||||||
| 4623 | |||||||
| 4624 | // Otherwise, try to resolve it to something we know about by looking at | ||||||
| 4625 | // the actual operand type. | ||||||
| 4626 | if (const char *Repl = LowerXConstraint(OpInfo.ConstraintVT)) { | ||||||
| 4627 | OpInfo.ConstraintCode = Repl; | ||||||
| 4628 | OpInfo.ConstraintType = getConstraintType(OpInfo.ConstraintCode); | ||||||
| 4629 | } | ||||||
| 4630 | } | ||||||
| 4631 | } | ||||||
| 4632 | |||||||
| 4633 | /// Given an exact SDIV by a constant, create a multiplication | ||||||
| 4634 | /// with the multiplicative inverse of the constant. | ||||||
| 4635 | static SDValue BuildExactSDIV(const TargetLowering &TLI, SDNode *N, | ||||||
| 4636 | const SDLoc &dl, SelectionDAG &DAG, | ||||||
| 4637 | SmallVectorImpl<SDNode *> &Created) { | ||||||
| 4638 | SDValue Op0 = N->getOperand(0); | ||||||
| 4639 | SDValue Op1 = N->getOperand(1); | ||||||
| 4640 | EVT VT = N->getValueType(0); | ||||||
| 4641 | EVT SVT = VT.getScalarType(); | ||||||
| 4642 | EVT ShVT = TLI.getShiftAmountTy(VT, DAG.getDataLayout()); | ||||||
| 4643 | EVT ShSVT = ShVT.getScalarType(); | ||||||
| 4644 | |||||||
| 4645 | bool UseSRA = false; | ||||||
| 4646 | SmallVector<SDValue, 16> Shifts, Factors; | ||||||
| 4647 | |||||||
| 4648 | auto BuildSDIVPattern = [&](ConstantSDNode *C) { | ||||||
| 4649 | if (C->isNullValue()) | ||||||
| 4650 | return false; | ||||||
| 4651 | APInt Divisor = C->getAPIntValue(); | ||||||
| 4652 | unsigned Shift = Divisor.countTrailingZeros(); | ||||||
| 4653 | if (Shift) { | ||||||
| 4654 | Divisor.ashrInPlace(Shift); | ||||||
| 4655 | UseSRA = true; | ||||||
| 4656 | } | ||||||
| 4657 | // Calculate the multiplicative inverse, using Newton's method. | ||||||
| 4658 | APInt t; | ||||||
| 4659 | APInt Factor = Divisor; | ||||||
| 4660 | while ((t = Divisor * Factor) != 1) | ||||||
| 4661 | Factor *= APInt(Divisor.getBitWidth(), 2) - t; | ||||||
| 4662 | Shifts.push_back(DAG.getConstant(Shift, dl, ShSVT)); | ||||||
| 4663 | Factors.push_back(DAG.getConstant(Factor, dl, SVT)); | ||||||
| 4664 | return true; | ||||||
| 4665 | }; | ||||||
| 4666 | |||||||
| 4667 | // Collect all magic values from the build vector. | ||||||
| 4668 | if (!ISD::matchUnaryPredicate(Op1, BuildSDIVPattern)) | ||||||
| 4669 | return SDValue(); | ||||||
| 4670 | |||||||
| 4671 | SDValue Shift, Factor; | ||||||
| 4672 | if (VT.isVector()) { | ||||||
| 4673 | Shift = DAG.getBuildVector(ShVT, dl, Shifts); | ||||||
| 4674 | Factor = DAG.getBuildVector(VT, dl, Factors); | ||||||
| 4675 | } else { | ||||||
| 4676 | Shift = Shifts[0]; | ||||||
| 4677 | Factor = Factors[0]; | ||||||
| 4678 | } | ||||||
| 4679 | |||||||
| 4680 | SDValue Res = Op0; | ||||||
| 4681 | |||||||
| 4682 | // Shift the value upfront if it is even, so the LSB is one. | ||||||
| 4683 | if (UseSRA) { | ||||||
| 4684 | // TODO: For UDIV use SRL instead of SRA. | ||||||
| 4685 | SDNodeFlags Flags; | ||||||
| 4686 | Flags.setExact(true); | ||||||
| 4687 | Res = DAG.getNode(ISD::SRA, dl, VT, Res, Shift, Flags); | ||||||
| 4688 | Created.push_back(Res.getNode()); | ||||||
| 4689 | } | ||||||
| 4690 | |||||||
| 4691 | return DAG.getNode(ISD::MUL, dl, VT, Res, Factor); | ||||||
| 4692 | } | ||||||
| 4693 | |||||||
| 4694 | SDValue TargetLowering::BuildSDIVPow2(SDNode *N, const APInt &Divisor, | ||||||
| 4695 | SelectionDAG &DAG, | ||||||
| 4696 | SmallVectorImpl<SDNode *> &Created) const { | ||||||
| 4697 | AttributeList Attr = DAG.getMachineFunction().getFunction().getAttributes(); | ||||||
| 4698 | const TargetLowering &TLI = DAG.getTargetLoweringInfo(); | ||||||
| 4699 | if (TLI.isIntDivCheap(N->getValueType(0), Attr)) | ||||||
| 4700 | return SDValue(N, 0); // Lower SDIV as SDIV | ||||||
| 4701 | return SDValue(); | ||||||
| 4702 | } | ||||||
| 4703 | |||||||
| 4704 | /// Given an ISD::SDIV node expressing a divide by constant, | ||||||
| 4705 | /// return a DAG expression to select that will generate the same value by | ||||||
| 4706 | /// multiplying by a magic number. | ||||||
| 4707 | /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". | ||||||
| 4708 | SDValue TargetLowering::BuildSDIV(SDNode *N, SelectionDAG &DAG, | ||||||
| 4709 | bool IsAfterLegalization, | ||||||
| 4710 | SmallVectorImpl<SDNode *> &Created) const { | ||||||
| 4711 | SDLoc dl(N); | ||||||
| 4712 | EVT VT = N->getValueType(0); | ||||||
| 4713 | EVT SVT = VT.getScalarType(); | ||||||
| 4714 | EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); | ||||||
| 4715 | EVT ShSVT = ShVT.getScalarType(); | ||||||
| 4716 | unsigned EltBits = VT.getScalarSizeInBits(); | ||||||
| 4717 | |||||||
| 4718 | // Check to see if we can do this. | ||||||
| 4719 | // FIXME: We should be more aggressive here. | ||||||
| 4720 | if (!isTypeLegal(VT)) | ||||||
| 4721 | return SDValue(); | ||||||
| 4722 | |||||||
| 4723 | // If the sdiv has an 'exact' bit we can use a simpler lowering. | ||||||
| 4724 | if (N->getFlags().hasExact()) | ||||||
| 4725 | return BuildExactSDIV(*this, N, dl, DAG, Created); | ||||||
| 4726 | |||||||
| 4727 | SmallVector<SDValue, 16> MagicFactors, Factors, Shifts, ShiftMasks; | ||||||
| 4728 | |||||||
| 4729 | auto BuildSDIVPattern = [&](ConstantSDNode *C) { | ||||||
| 4730 | if (C->isNullValue()) | ||||||
| 4731 | return false; | ||||||
| 4732 | |||||||
| 4733 | const APInt &Divisor = C->getAPIntValue(); | ||||||
| 4734 | APInt::ms magics = Divisor.magic(); | ||||||
| 4735 | int NumeratorFactor = 0; | ||||||
| 4736 | int ShiftMask = -1; | ||||||
| 4737 | |||||||
| 4738 | if (Divisor.isOneValue() || Divisor.isAllOnesValue()) { | ||||||
| 4739 | // If d is +1/-1, we just multiply the numerator by +1/-1. | ||||||
| 4740 | NumeratorFactor = Divisor.getSExtValue(); | ||||||
| 4741 | magics.m = 0; | ||||||
| 4742 | magics.s = 0; | ||||||
| 4743 | ShiftMask = 0; | ||||||
| 4744 | } else if (Divisor.isStrictlyPositive() && magics.m.isNegative()) { | ||||||
| 4745 | // If d > 0 and m < 0, add the numerator. | ||||||
| 4746 | NumeratorFactor = 1; | ||||||
| 4747 | } else if (Divisor.isNegative() && magics.m.isStrictlyPositive()) { | ||||||
| 4748 | // If d < 0 and m > 0, subtract the numerator. | ||||||
| 4749 | NumeratorFactor = -1; | ||||||
| 4750 | } | ||||||
| 4751 | |||||||
| 4752 | MagicFactors.push_back(DAG.getConstant(magics.m, dl, SVT)); | ||||||
| 4753 | Factors.push_back(DAG.getConstant(NumeratorFactor, dl, SVT)); | ||||||
| 4754 | Shifts.push_back(DAG.getConstant(magics.s, dl, ShSVT)); | ||||||
| 4755 | ShiftMasks.push_back(DAG.getConstant(ShiftMask, dl, SVT)); | ||||||
| 4756 | return true; | ||||||
| 4757 | }; | ||||||
| 4758 | |||||||
| 4759 | SDValue N0 = N->getOperand(0); | ||||||
| 4760 | SDValue N1 = N->getOperand(1); | ||||||
| 4761 | |||||||
| 4762 | // Collect the shifts / magic values from each element. | ||||||
| 4763 | if (!ISD::matchUnaryPredicate(N1, BuildSDIVPattern)) | ||||||
| 4764 | return SDValue(); | ||||||
| 4765 | |||||||
| 4766 | SDValue MagicFactor, Factor, Shift, ShiftMask; | ||||||
| 4767 | if (VT.isVector()) { | ||||||
| 4768 | MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors); | ||||||
| 4769 | Factor = DAG.getBuildVector(VT, dl, Factors); | ||||||
| 4770 | Shift = DAG.getBuildVector(ShVT, dl, Shifts); | ||||||
| 4771 | ShiftMask = DAG.getBuildVector(VT, dl, ShiftMasks); | ||||||
| 4772 | } else { | ||||||
| 4773 | MagicFactor = MagicFactors[0]; | ||||||
| 4774 | Factor = Factors[0]; | ||||||
| 4775 | Shift = Shifts[0]; | ||||||
| 4776 | ShiftMask = ShiftMasks[0]; | ||||||
| 4777 | } | ||||||
| 4778 | |||||||
| 4779 | // Multiply the numerator (operand 0) by the magic value. | ||||||
| 4780 | // FIXME: We should support doing a MUL in a wider type. | ||||||
| 4781 | SDValue Q; | ||||||
| 4782 | if (IsAfterLegalization ? isOperationLegal(ISD::MULHS, VT) | ||||||
| 4783 | : isOperationLegalOrCustom(ISD::MULHS, VT)) | ||||||
| 4784 | Q = DAG.getNode(ISD::MULHS, dl, VT, N0, MagicFactor); | ||||||
| 4785 | else if (IsAfterLegalization ? isOperationLegal(ISD::SMUL_LOHI, VT) | ||||||
| 4786 | : isOperationLegalOrCustom(ISD::SMUL_LOHI, VT)) { | ||||||
| 4787 | SDValue LoHi = | ||||||
| 4788 | DAG.getNode(ISD::SMUL_LOHI, dl, DAG.getVTList(VT, VT), N0, MagicFactor); | ||||||
| 4789 | Q = SDValue(LoHi.getNode(), 1); | ||||||
| 4790 | } else | ||||||
| 4791 | return SDValue(); // No mulhs or equivalent. | ||||||
| 4792 | Created.push_back(Q.getNode()); | ||||||
| 4793 | |||||||
| 4794 | // (Optionally) Add/subtract the numerator using Factor. | ||||||
| 4795 | Factor = DAG.getNode(ISD::MUL, dl, VT, N0, Factor); | ||||||
| 4796 | Created.push_back(Factor.getNode()); | ||||||
| 4797 | Q = DAG.getNode(ISD::ADD, dl, VT, Q, Factor); | ||||||
| 4798 | Created.push_back(Q.getNode()); | ||||||
| 4799 | |||||||
| 4800 | // Shift right algebraic by shift value. | ||||||
| 4801 | Q = DAG.getNode(ISD::SRA, dl, VT, Q, Shift); | ||||||
| 4802 | Created.push_back(Q.getNode()); | ||||||
| 4803 | |||||||
| 4804 | // Extract the sign bit, mask it and add it to the quotient. | ||||||
| 4805 | SDValue SignShift = DAG.getConstant(EltBits - 1, dl, ShVT); | ||||||
| 4806 | SDValue T = DAG.getNode(ISD::SRL, dl, VT, Q, SignShift); | ||||||
| 4807 | Created.push_back(T.getNode()); | ||||||
| 4808 | T = DAG.getNode(ISD::AND, dl, VT, T, ShiftMask); | ||||||
| 4809 | Created.push_back(T.getNode()); | ||||||
| 4810 | return DAG.getNode(ISD::ADD, dl, VT, Q, T); | ||||||
| 4811 | } | ||||||
| 4812 | |||||||
| 4813 | /// Given an ISD::UDIV node expressing a divide by constant, | ||||||
| 4814 | /// return a DAG expression to select that will generate the same value by | ||||||
| 4815 | /// multiplying by a magic number. | ||||||
| 4816 | /// Ref: "Hacker's Delight" or "The PowerPC Compiler Writer's Guide". | ||||||
| 4817 | SDValue TargetLowering::BuildUDIV(SDNode *N, SelectionDAG &DAG, | ||||||
| 4818 | bool IsAfterLegalization, | ||||||
| 4819 | SmallVectorImpl<SDNode *> &Created) const { | ||||||
| 4820 | SDLoc dl(N); | ||||||
| 4821 | EVT VT = N->getValueType(0); | ||||||
| 4822 | EVT SVT = VT.getScalarType(); | ||||||
| 4823 | EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); | ||||||
| 4824 | EVT ShSVT = ShVT.getScalarType(); | ||||||
| 4825 | unsigned EltBits = VT.getScalarSizeInBits(); | ||||||
| 4826 | |||||||
| 4827 | // Check to see if we can do this. | ||||||
| 4828 | // FIXME: We should be more aggressive here. | ||||||
| 4829 | if (!isTypeLegal(VT)) | ||||||
| 4830 | return SDValue(); | ||||||
| 4831 | |||||||
| 4832 | bool UseNPQ = false; | ||||||
| 4833 | SmallVector<SDValue, 16> PreShifts, PostShifts, MagicFactors, NPQFactors; | ||||||
| 4834 | |||||||
| 4835 | auto BuildUDIVPattern = [&](ConstantSDNode *C) { | ||||||
| 4836 | if (C->isNullValue()) | ||||||
| 4837 | return false; | ||||||
| 4838 | // FIXME: We should use a narrower constant when the upper | ||||||
| 4839 | // bits are known to be zero. | ||||||
| 4840 | APInt Divisor = C->getAPIntValue(); | ||||||
| 4841 | APInt::mu magics = Divisor.magicu(); | ||||||
| 4842 | unsigned PreShift = 0, PostShift = 0; | ||||||
| 4843 | |||||||
| 4844 | // If the divisor is even, we can avoid using the expensive fixup by | ||||||
| 4845 | // shifting the divided value upfront. | ||||||
| 4846 | if (magics.a != 0 && !Divisor[0]) { | ||||||
| 4847 | PreShift = Divisor.countTrailingZeros(); | ||||||
| 4848 | // Get magic number for the shifted divisor. | ||||||
| 4849 | magics = Divisor.lshr(PreShift).magicu(PreShift); | ||||||
| 4850 | assert(magics.a == 0 && "Should use cheap fixup now")((magics.a == 0 && "Should use cheap fixup now") ? static_cast <void> (0) : __assert_fail ("magics.a == 0 && \"Should use cheap fixup now\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4850, __PRETTY_FUNCTION__)); | ||||||
| 4851 | } | ||||||
| 4852 | |||||||
| 4853 | APInt Magic = magics.m; | ||||||
| 4854 | |||||||
| 4855 | unsigned SelNPQ; | ||||||
| 4856 | if (magics.a == 0 || Divisor.isOneValue()) { | ||||||
| 4857 | assert(magics.s < Divisor.getBitWidth() &&((magics.s < Divisor.getBitWidth() && "We shouldn't generate an undefined shift!" ) ? static_cast<void> (0) : __assert_fail ("magics.s < Divisor.getBitWidth() && \"We shouldn't generate an undefined shift!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4858, __PRETTY_FUNCTION__)) | ||||||
| 4858 | "We shouldn't generate an undefined shift!")((magics.s < Divisor.getBitWidth() && "We shouldn't generate an undefined shift!" ) ? static_cast<void> (0) : __assert_fail ("magics.s < Divisor.getBitWidth() && \"We shouldn't generate an undefined shift!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 4858, __PRETTY_FUNCTION__)); | ||||||
| 4859 | PostShift = magics.s; | ||||||
| 4860 | SelNPQ = false; | ||||||
| 4861 | } else { | ||||||
| 4862 | PostShift = magics.s - 1; | ||||||
| 4863 | SelNPQ = true; | ||||||
| 4864 | } | ||||||
| 4865 | |||||||
| 4866 | PreShifts.push_back(DAG.getConstant(PreShift, dl, ShSVT)); | ||||||
| 4867 | MagicFactors.push_back(DAG.getConstant(Magic, dl, SVT)); | ||||||
| 4868 | NPQFactors.push_back( | ||||||
| 4869 | DAG.getConstant(SelNPQ ? APInt::getOneBitSet(EltBits, EltBits - 1) | ||||||
| 4870 | : APInt::getNullValue(EltBits), | ||||||
| 4871 | dl, SVT)); | ||||||
| 4872 | PostShifts.push_back(DAG.getConstant(PostShift, dl, ShSVT)); | ||||||
| 4873 | UseNPQ |= SelNPQ; | ||||||
| 4874 | return true; | ||||||
| 4875 | }; | ||||||
| 4876 | |||||||
| 4877 | SDValue N0 = N->getOperand(0); | ||||||
| 4878 | SDValue N1 = N->getOperand(1); | ||||||
| 4879 | |||||||
| 4880 | // Collect the shifts/magic values from each element. | ||||||
| 4881 | if (!ISD::matchUnaryPredicate(N1, BuildUDIVPattern)) | ||||||
| 4882 | return SDValue(); | ||||||
| 4883 | |||||||
| 4884 | SDValue PreShift, PostShift, MagicFactor, NPQFactor; | ||||||
| 4885 | if (VT.isVector()) { | ||||||
| 4886 | PreShift = DAG.getBuildVector(ShVT, dl, PreShifts); | ||||||
| 4887 | MagicFactor = DAG.getBuildVector(VT, dl, MagicFactors); | ||||||
| 4888 | NPQFactor = DAG.getBuildVector(VT, dl, NPQFactors); | ||||||
| 4889 | PostShift = DAG.getBuildVector(ShVT, dl, PostShifts); | ||||||
| 4890 | } else { | ||||||
| 4891 | PreShift = PreShifts[0]; | ||||||
| 4892 | MagicFactor = MagicFactors[0]; | ||||||
| 4893 | PostShift = PostShifts[0]; | ||||||
| 4894 | } | ||||||
| 4895 | |||||||
| 4896 | SDValue Q = N0; | ||||||
| 4897 | Q = DAG.getNode(ISD::SRL, dl, VT, Q, PreShift); | ||||||
| 4898 | Created.push_back(Q.getNode()); | ||||||
| 4899 | |||||||
| 4900 | // FIXME: We should support doing a MUL in a wider type. | ||||||
| 4901 | auto GetMULHU = [&](SDValue X, SDValue Y) { | ||||||
| 4902 | if (IsAfterLegalization ? isOperationLegal(ISD::MULHU, VT) | ||||||
| 4903 | : isOperationLegalOrCustom(ISD::MULHU, VT)) | ||||||
| 4904 | return DAG.getNode(ISD::MULHU, dl, VT, X, Y); | ||||||
| 4905 | if (IsAfterLegalization ? isOperationLegal(ISD::UMUL_LOHI, VT) | ||||||
| 4906 | : isOperationLegalOrCustom(ISD::UMUL_LOHI, VT)) { | ||||||
| 4907 | SDValue LoHi = | ||||||
| 4908 | DAG.getNode(ISD::UMUL_LOHI, dl, DAG.getVTList(VT, VT), X, Y); | ||||||
| 4909 | return SDValue(LoHi.getNode(), 1); | ||||||
| 4910 | } | ||||||
| 4911 | return SDValue(); // No mulhu or equivalent | ||||||
| 4912 | }; | ||||||
| 4913 | |||||||
| 4914 | // Multiply the numerator (operand 0) by the magic value. | ||||||
| 4915 | Q = GetMULHU(Q, MagicFactor); | ||||||
| 4916 | if (!Q) | ||||||
| 4917 | return SDValue(); | ||||||
| 4918 | |||||||
| 4919 | Created.push_back(Q.getNode()); | ||||||
| 4920 | |||||||
| 4921 | if (UseNPQ) { | ||||||
| 4922 | SDValue NPQ = DAG.getNode(ISD::SUB, dl, VT, N0, Q); | ||||||
| 4923 | Created.push_back(NPQ.getNode()); | ||||||
| 4924 | |||||||
| 4925 | // For vectors we might have a mix of non-NPQ/NPQ paths, so use | ||||||
| 4926 | // MULHU to act as a SRL-by-1 for NPQ, else multiply by zero. | ||||||
| 4927 | if (VT.isVector()) | ||||||
| 4928 | NPQ = GetMULHU(NPQ, NPQFactor); | ||||||
| 4929 | else | ||||||
| 4930 | NPQ = DAG.getNode(ISD::SRL, dl, VT, NPQ, DAG.getConstant(1, dl, ShVT)); | ||||||
| 4931 | |||||||
| 4932 | Created.push_back(NPQ.getNode()); | ||||||
| 4933 | |||||||
| 4934 | Q = DAG.getNode(ISD::ADD, dl, VT, NPQ, Q); | ||||||
| 4935 | Created.push_back(Q.getNode()); | ||||||
| 4936 | } | ||||||
| 4937 | |||||||
| 4938 | Q = DAG.getNode(ISD::SRL, dl, VT, Q, PostShift); | ||||||
| 4939 | Created.push_back(Q.getNode()); | ||||||
| 4940 | |||||||
| 4941 | SDValue One = DAG.getConstant(1, dl, VT); | ||||||
| 4942 | SDValue IsOne = DAG.getSetCC(dl, VT, N1, One, ISD::SETEQ); | ||||||
| 4943 | return DAG.getSelect(dl, VT, IsOne, N0, Q); | ||||||
| 4944 | } | ||||||
| 4945 | |||||||
| 4946 | /// If all values in Values that *don't* match the predicate are same 'splat' | ||||||
| 4947 | /// value, then replace all values with that splat value. | ||||||
| 4948 | /// Else, if AlternativeReplacement was provided, then replace all values that | ||||||
| 4949 | /// do match predicate with AlternativeReplacement value. | ||||||
| 4950 | static void | ||||||
| 4951 | turnVectorIntoSplatVector(MutableArrayRef<SDValue> Values, | ||||||
| 4952 | std::function<bool(SDValue)> Predicate, | ||||||
| 4953 | SDValue AlternativeReplacement = SDValue()) { | ||||||
| 4954 | SDValue Replacement; | ||||||
| 4955 | // Is there a value for which the Predicate does *NOT* match? What is it? | ||||||
| 4956 | auto SplatValue = llvm::find_if_not(Values, Predicate); | ||||||
| 4957 | if (SplatValue != Values.end()) { | ||||||
| 4958 | // Does Values consist only of SplatValue's and values matching Predicate? | ||||||
| 4959 | if (llvm::all_of(Values, [Predicate, SplatValue](SDValue Value) { | ||||||
| 4960 | return Value == *SplatValue || Predicate(Value); | ||||||
| 4961 | })) // Then we shall replace values matching predicate with SplatValue. | ||||||
| 4962 | Replacement = *SplatValue; | ||||||
| 4963 | } | ||||||
| 4964 | if (!Replacement) { | ||||||
| 4965 | // Oops, we did not find the "baseline" splat value. | ||||||
| 4966 | if (!AlternativeReplacement) | ||||||
| 4967 | return; // Nothing to do. | ||||||
| 4968 | // Let's replace with provided value then. | ||||||
| 4969 | Replacement = AlternativeReplacement; | ||||||
| 4970 | } | ||||||
| 4971 | std::replace_if(Values.begin(), Values.end(), Predicate, Replacement); | ||||||
| 4972 | } | ||||||
| 4973 | |||||||
| 4974 | /// Given an ISD::UREM used only by an ISD::SETEQ or ISD::SETNE | ||||||
| 4975 | /// where the divisor is constant and the comparison target is zero, | ||||||
| 4976 | /// return a DAG expression that will generate the same comparison result | ||||||
| 4977 | /// using only multiplications, additions and shifts/rotations. | ||||||
| 4978 | /// Ref: "Hacker's Delight" 10-17. | ||||||
| 4979 | SDValue TargetLowering::buildUREMEqFold(EVT SETCCVT, SDValue REMNode, | ||||||
| 4980 | SDValue CompTargetNode, | ||||||
| 4981 | ISD::CondCode Cond, | ||||||
| 4982 | DAGCombinerInfo &DCI, | ||||||
| 4983 | const SDLoc &DL) const { | ||||||
| 4984 | SmallVector<SDNode *, 5> Built; | ||||||
| 4985 | if (SDValue Folded = prepareUREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond, | ||||||
| 4986 | DCI, DL, Built)) { | ||||||
| 4987 | for (SDNode *N : Built) | ||||||
| 4988 | DCI.AddToWorklist(N); | ||||||
| 4989 | return Folded; | ||||||
| 4990 | } | ||||||
| 4991 | |||||||
| 4992 | return SDValue(); | ||||||
| 4993 | } | ||||||
| 4994 | |||||||
| 4995 | SDValue | ||||||
| 4996 | TargetLowering::prepareUREMEqFold(EVT SETCCVT, SDValue REMNode, | ||||||
| 4997 | SDValue CompTargetNode, ISD::CondCode Cond, | ||||||
| 4998 | DAGCombinerInfo &DCI, const SDLoc &DL, | ||||||
| 4999 | SmallVectorImpl<SDNode *> &Created) const { | ||||||
| 5000 | // fold (seteq/ne (urem N, D), 0) -> (setule/ugt (rotr (mul N, P), K), Q) | ||||||
| 5001 | // - D must be constant, with D = D0 * 2^K where D0 is odd | ||||||
| 5002 | // - P is the multiplicative inverse of D0 modulo 2^W | ||||||
| 5003 | // - Q = floor(((2^W) - 1) / D) | ||||||
| 5004 | // where W is the width of the common type of N and D. | ||||||
| 5005 | assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Only applicable for (in)equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5006, __PRETTY_FUNCTION__)) | ||||||
| 5006 | "Only applicable for (in)equality comparisons.")(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Only applicable for (in)equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5006, __PRETTY_FUNCTION__)); | ||||||
| 5007 | |||||||
| 5008 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 5009 | |||||||
| 5010 | EVT VT = REMNode.getValueType(); | ||||||
| 5011 | EVT SVT = VT.getScalarType(); | ||||||
| 5012 | EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); | ||||||
| 5013 | EVT ShSVT = ShVT.getScalarType(); | ||||||
| 5014 | |||||||
| 5015 | // If MUL is unavailable, we cannot proceed in any case. | ||||||
| 5016 | if (!isOperationLegalOrCustom(ISD::MUL, VT)) | ||||||
| 5017 | return SDValue(); | ||||||
| 5018 | |||||||
| 5019 | bool ComparingWithAllZeros = true; | ||||||
| 5020 | bool AllComparisonsWithNonZerosAreTautological = true; | ||||||
| 5021 | bool HadTautologicalLanes = false; | ||||||
| 5022 | bool AllLanesAreTautological = true; | ||||||
| 5023 | bool HadEvenDivisor = false; | ||||||
| 5024 | bool AllDivisorsArePowerOfTwo = true; | ||||||
| 5025 | bool HadTautologicalInvertedLanes = false; | ||||||
| 5026 | SmallVector<SDValue, 16> PAmts, KAmts, QAmts, IAmts; | ||||||
| 5027 | |||||||
| 5028 | auto BuildUREMPattern = [&](ConstantSDNode *CDiv, ConstantSDNode *CCmp) { | ||||||
| 5029 | // Division by 0 is UB. Leave it to be constant-folded elsewhere. | ||||||
| 5030 | if (CDiv->isNullValue()) | ||||||
| 5031 | return false; | ||||||
| 5032 | |||||||
| 5033 | const APInt &D = CDiv->getAPIntValue(); | ||||||
| 5034 | const APInt &Cmp = CCmp->getAPIntValue(); | ||||||
| 5035 | |||||||
| 5036 | ComparingWithAllZeros &= Cmp.isNullValue(); | ||||||
| 5037 | |||||||
| 5038 | // x u% C1` is *always* less than C1. So given `x u% C1 == C2`, | ||||||
| 5039 | // if C2 is not less than C1, the comparison is always false. | ||||||
| 5040 | // But we will only be able to produce the comparison that will give the | ||||||
| 5041 | // opposive tautological answer. So this lane would need to be fixed up. | ||||||
| 5042 | bool TautologicalInvertedLane = D.ule(Cmp); | ||||||
| 5043 | HadTautologicalInvertedLanes |= TautologicalInvertedLane; | ||||||
| 5044 | |||||||
| 5045 | // If all lanes are tautological (either all divisors are ones, or divisor | ||||||
| 5046 | // is not greater than the constant we are comparing with), | ||||||
| 5047 | // we will prefer to avoid the fold. | ||||||
| 5048 | bool TautologicalLane = D.isOneValue() || TautologicalInvertedLane; | ||||||
| 5049 | HadTautologicalLanes |= TautologicalLane; | ||||||
| 5050 | AllLanesAreTautological &= TautologicalLane; | ||||||
| 5051 | |||||||
| 5052 | // If we are comparing with non-zero, we need'll need to subtract said | ||||||
| 5053 | // comparison value from the LHS. But there is no point in doing that if | ||||||
| 5054 | // every lane where we are comparing with non-zero is tautological.. | ||||||
| 5055 | if (!Cmp.isNullValue()) | ||||||
| 5056 | AllComparisonsWithNonZerosAreTautological &= TautologicalLane; | ||||||
| 5057 | |||||||
| 5058 | // Decompose D into D0 * 2^K | ||||||
| 5059 | unsigned K = D.countTrailingZeros(); | ||||||
| 5060 | assert((!D.isOneValue() || (K == 0)) && "For divisor '1' we won't rotate.")(((!D.isOneValue() || (K == 0)) && "For divisor '1' we won't rotate." ) ? static_cast<void> (0) : __assert_fail ("(!D.isOneValue() || (K == 0)) && \"For divisor '1' we won't rotate.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5060, __PRETTY_FUNCTION__)); | ||||||
| 5061 | APInt D0 = D.lshr(K); | ||||||
| 5062 | |||||||
| 5063 | // D is even if it has trailing zeros. | ||||||
| 5064 | HadEvenDivisor |= (K != 0); | ||||||
| 5065 | // D is a power-of-two if D0 is one. | ||||||
| 5066 | // If all divisors are power-of-two, we will prefer to avoid the fold. | ||||||
| 5067 | AllDivisorsArePowerOfTwo &= D0.isOneValue(); | ||||||
| 5068 | |||||||
| 5069 | // P = inv(D0, 2^W) | ||||||
| 5070 | // 2^W requires W + 1 bits, so we have to extend and then truncate. | ||||||
| 5071 | unsigned W = D.getBitWidth(); | ||||||
| 5072 | APInt P = D0.zext(W + 1) | ||||||
| 5073 | .multiplicativeInverse(APInt::getSignedMinValue(W + 1)) | ||||||
| 5074 | .trunc(W); | ||||||
| 5075 | assert(!P.isNullValue() && "No multiplicative inverse!")((!P.isNullValue() && "No multiplicative inverse!") ? static_cast<void> (0) : __assert_fail ("!P.isNullValue() && \"No multiplicative inverse!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5075, __PRETTY_FUNCTION__)); // unreachable | ||||||
| 5076 | assert((D0 * P).isOneValue() && "Multiplicative inverse sanity check.")(((D0 * P).isOneValue() && "Multiplicative inverse sanity check." ) ? static_cast<void> (0) : __assert_fail ("(D0 * P).isOneValue() && \"Multiplicative inverse sanity check.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5076, __PRETTY_FUNCTION__)); | ||||||
| 5077 | |||||||
| 5078 | // Q = floor((2^W - 1) u/ D) | ||||||
| 5079 | // R = ((2^W - 1) u% D) | ||||||
| 5080 | APInt Q, R; | ||||||
| 5081 | APInt::udivrem(APInt::getAllOnesValue(W), D, Q, R); | ||||||
| 5082 | |||||||
| 5083 | // If we are comparing with zero, then that comparison constant is okay, | ||||||
| 5084 | // else it may need to be one less than that. | ||||||
| 5085 | if (Cmp.ugt(R)) | ||||||
| 5086 | Q -= 1; | ||||||
| 5087 | |||||||
| 5088 | assert(APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) &&((APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && \"We are expecting that K is always less than all-ones for ShSVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5089, __PRETTY_FUNCTION__)) | ||||||
| 5089 | "We are expecting that K is always less than all-ones for ShSVT")((APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && \"We are expecting that K is always less than all-ones for ShSVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5089, __PRETTY_FUNCTION__)); | ||||||
| 5090 | |||||||
| 5091 | // If the lane is tautological the result can be constant-folded. | ||||||
| 5092 | if (TautologicalLane) { | ||||||
| 5093 | // Set P and K amount to a bogus values so we can try to splat them. | ||||||
| 5094 | P = 0; | ||||||
| 5095 | K = -1; | ||||||
| 5096 | // And ensure that comparison constant is tautological, | ||||||
| 5097 | // it will always compare true/false. | ||||||
| 5098 | Q = -1; | ||||||
| 5099 | } | ||||||
| 5100 | |||||||
| 5101 | PAmts.push_back(DAG.getConstant(P, DL, SVT)); | ||||||
| 5102 | KAmts.push_back( | ||||||
| 5103 | DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT)); | ||||||
| 5104 | QAmts.push_back(DAG.getConstant(Q, DL, SVT)); | ||||||
| 5105 | return true; | ||||||
| 5106 | }; | ||||||
| 5107 | |||||||
| 5108 | SDValue N = REMNode.getOperand(0); | ||||||
| 5109 | SDValue D = REMNode.getOperand(1); | ||||||
| 5110 | |||||||
| 5111 | // Collect the values from each element. | ||||||
| 5112 | if (!ISD::matchBinaryPredicate(D, CompTargetNode, BuildUREMPattern)) | ||||||
| 5113 | return SDValue(); | ||||||
| 5114 | |||||||
| 5115 | // If all lanes are tautological, the result can be constant-folded. | ||||||
| 5116 | if (AllLanesAreTautological) | ||||||
| 5117 | return SDValue(); | ||||||
| 5118 | |||||||
| 5119 | // If this is a urem by a powers-of-two, avoid the fold since it can be | ||||||
| 5120 | // best implemented as a bit test. | ||||||
| 5121 | if (AllDivisorsArePowerOfTwo) | ||||||
| 5122 | return SDValue(); | ||||||
| 5123 | |||||||
| 5124 | SDValue PVal, KVal, QVal; | ||||||
| 5125 | if (VT.isVector()) { | ||||||
| 5126 | if (HadTautologicalLanes) { | ||||||
| 5127 | // Try to turn PAmts into a splat, since we don't care about the values | ||||||
| 5128 | // that are currently '0'. If we can't, just keep '0'`s. | ||||||
| 5129 | turnVectorIntoSplatVector(PAmts, isNullConstant); | ||||||
| 5130 | // Try to turn KAmts into a splat, since we don't care about the values | ||||||
| 5131 | // that are currently '-1'. If we can't, change them to '0'`s. | ||||||
| 5132 | turnVectorIntoSplatVector(KAmts, isAllOnesConstant, | ||||||
| 5133 | DAG.getConstant(0, DL, ShSVT)); | ||||||
| 5134 | } | ||||||
| 5135 | |||||||
| 5136 | PVal = DAG.getBuildVector(VT, DL, PAmts); | ||||||
| 5137 | KVal = DAG.getBuildVector(ShVT, DL, KAmts); | ||||||
| 5138 | QVal = DAG.getBuildVector(VT, DL, QAmts); | ||||||
| 5139 | } else { | ||||||
| 5140 | PVal = PAmts[0]; | ||||||
| 5141 | KVal = KAmts[0]; | ||||||
| 5142 | QVal = QAmts[0]; | ||||||
| 5143 | } | ||||||
| 5144 | |||||||
| 5145 | if (!ComparingWithAllZeros && !AllComparisonsWithNonZerosAreTautological) { | ||||||
| 5146 | if (!isOperationLegalOrCustom(ISD::SUB, VT)) | ||||||
| 5147 | return SDValue(); // FIXME: Could/should use `ISD::ADD`? | ||||||
| 5148 | assert(CompTargetNode.getValueType() == N.getValueType() &&((CompTargetNode.getValueType() == N.getValueType() && "Expecting that the types on LHS and RHS of comparisons match." ) ? static_cast<void> (0) : __assert_fail ("CompTargetNode.getValueType() == N.getValueType() && \"Expecting that the types on LHS and RHS of comparisons match.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5149, __PRETTY_FUNCTION__)) | ||||||
| 5149 | "Expecting that the types on LHS and RHS of comparisons match.")((CompTargetNode.getValueType() == N.getValueType() && "Expecting that the types on LHS and RHS of comparisons match." ) ? static_cast<void> (0) : __assert_fail ("CompTargetNode.getValueType() == N.getValueType() && \"Expecting that the types on LHS and RHS of comparisons match.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5149, __PRETTY_FUNCTION__)); | ||||||
| 5150 | N = DAG.getNode(ISD::SUB, DL, VT, N, CompTargetNode); | ||||||
| 5151 | } | ||||||
| 5152 | |||||||
| 5153 | // (mul N, P) | ||||||
| 5154 | SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal); | ||||||
| 5155 | Created.push_back(Op0.getNode()); | ||||||
| 5156 | |||||||
| 5157 | // Rotate right only if any divisor was even. We avoid rotates for all-odd | ||||||
| 5158 | // divisors as a performance improvement, since rotating by 0 is a no-op. | ||||||
| 5159 | if (HadEvenDivisor) { | ||||||
| 5160 | // We need ROTR to do this. | ||||||
| 5161 | if (!isOperationLegalOrCustom(ISD::ROTR, VT)) | ||||||
| 5162 | return SDValue(); | ||||||
| 5163 | SDNodeFlags Flags; | ||||||
| 5164 | Flags.setExact(true); | ||||||
| 5165 | // UREM: (rotr (mul N, P), K) | ||||||
| 5166 | Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal, Flags); | ||||||
| 5167 | Created.push_back(Op0.getNode()); | ||||||
| 5168 | } | ||||||
| 5169 | |||||||
| 5170 | // UREM: (setule/setugt (rotr (mul N, P), K), Q) | ||||||
| 5171 | SDValue NewCC = | ||||||
| 5172 | DAG.getSetCC(DL, SETCCVT, Op0, QVal, | ||||||
| 5173 | ((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT)); | ||||||
| 5174 | if (!HadTautologicalInvertedLanes) | ||||||
| 5175 | return NewCC; | ||||||
| 5176 | |||||||
| 5177 | // If any lanes previously compared always-false, the NewCC will give | ||||||
| 5178 | // always-true result for them, so we need to fixup those lanes. | ||||||
| 5179 | // Or the other way around for inequality predicate. | ||||||
| 5180 | assert(VT.isVector() && "Can/should only get here for vectors.")((VT.isVector() && "Can/should only get here for vectors." ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Can/should only get here for vectors.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5180, __PRETTY_FUNCTION__)); | ||||||
| 5181 | Created.push_back(NewCC.getNode()); | ||||||
| 5182 | |||||||
| 5183 | // x u% C1` is *always* less than C1. So given `x u% C1 == C2`, | ||||||
| 5184 | // if C2 is not less than C1, the comparison is always false. | ||||||
| 5185 | // But we have produced the comparison that will give the | ||||||
| 5186 | // opposive tautological answer. So these lanes would need to be fixed up. | ||||||
| 5187 | SDValue TautologicalInvertedChannels = | ||||||
| 5188 | DAG.getSetCC(DL, SETCCVT, D, CompTargetNode, ISD::SETULE); | ||||||
| 5189 | Created.push_back(TautologicalInvertedChannels.getNode()); | ||||||
| 5190 | |||||||
| 5191 | if (isOperationLegalOrCustom(ISD::VSELECT, SETCCVT)) { | ||||||
| 5192 | // If we have a vector select, let's replace the comparison results in the | ||||||
| 5193 | // affected lanes with the correct tautological result. | ||||||
| 5194 | SDValue Replacement = DAG.getBoolConstant(Cond == ISD::SETEQ ? false : true, | ||||||
| 5195 | DL, SETCCVT, SETCCVT); | ||||||
| 5196 | return DAG.getNode(ISD::VSELECT, DL, SETCCVT, TautologicalInvertedChannels, | ||||||
| 5197 | Replacement, NewCC); | ||||||
| 5198 | } | ||||||
| 5199 | |||||||
| 5200 | // Else, we can just invert the comparison result in the appropriate lanes. | ||||||
| 5201 | if (isOperationLegalOrCustom(ISD::XOR, SETCCVT)) | ||||||
| 5202 | return DAG.getNode(ISD::XOR, DL, SETCCVT, NewCC, | ||||||
| 5203 | TautologicalInvertedChannels); | ||||||
| 5204 | |||||||
| 5205 | return SDValue(); // Don't know how to lower. | ||||||
| 5206 | } | ||||||
| 5207 | |||||||
| 5208 | /// Given an ISD::SREM used only by an ISD::SETEQ or ISD::SETNE | ||||||
| 5209 | /// where the divisor is constant and the comparison target is zero, | ||||||
| 5210 | /// return a DAG expression that will generate the same comparison result | ||||||
| 5211 | /// using only multiplications, additions and shifts/rotations. | ||||||
| 5212 | /// Ref: "Hacker's Delight" 10-17. | ||||||
| 5213 | SDValue TargetLowering::buildSREMEqFold(EVT SETCCVT, SDValue REMNode, | ||||||
| 5214 | SDValue CompTargetNode, | ||||||
| 5215 | ISD::CondCode Cond, | ||||||
| 5216 | DAGCombinerInfo &DCI, | ||||||
| 5217 | const SDLoc &DL) const { | ||||||
| 5218 | SmallVector<SDNode *, 7> Built; | ||||||
| 5219 | if (SDValue Folded = prepareSREMEqFold(SETCCVT, REMNode, CompTargetNode, Cond, | ||||||
| 5220 | DCI, DL, Built)) { | ||||||
| 5221 | assert(Built.size() <= 7 && "Max size prediction failed.")((Built.size() <= 7 && "Max size prediction failed." ) ? static_cast<void> (0) : __assert_fail ("Built.size() <= 7 && \"Max size prediction failed.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5221, __PRETTY_FUNCTION__)); | ||||||
| 5222 | for (SDNode *N : Built) | ||||||
| 5223 | DCI.AddToWorklist(N); | ||||||
| 5224 | return Folded; | ||||||
| 5225 | } | ||||||
| 5226 | |||||||
| 5227 | return SDValue(); | ||||||
| 5228 | } | ||||||
| 5229 | |||||||
| 5230 | SDValue | ||||||
| 5231 | TargetLowering::prepareSREMEqFold(EVT SETCCVT, SDValue REMNode, | ||||||
| 5232 | SDValue CompTargetNode, ISD::CondCode Cond, | ||||||
| 5233 | DAGCombinerInfo &DCI, const SDLoc &DL, | ||||||
| 5234 | SmallVectorImpl<SDNode *> &Created) const { | ||||||
| 5235 | // Fold: | ||||||
| 5236 | // (seteq/ne (srem N, D), 0) | ||||||
| 5237 | // To: | ||||||
| 5238 | // (setule/ugt (rotr (add (mul N, P), A), K), Q) | ||||||
| 5239 | // | ||||||
| 5240 | // - D must be constant, with D = D0 * 2^K where D0 is odd | ||||||
| 5241 | // - P is the multiplicative inverse of D0 modulo 2^W | ||||||
| 5242 | // - A = bitwiseand(floor((2^(W - 1) - 1) / D0), (-(2^k))) | ||||||
| 5243 | // - Q = floor((2 * A) / (2^K)) | ||||||
| 5244 | // where W is the width of the common type of N and D. | ||||||
| 5245 | assert((Cond == ISD::SETEQ || Cond == ISD::SETNE) &&(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Only applicable for (in)equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5246, __PRETTY_FUNCTION__)) | ||||||
| 5246 | "Only applicable for (in)equality comparisons.")(((Cond == ISD::SETEQ || Cond == ISD::SETNE) && "Only applicable for (in)equality comparisons." ) ? static_cast<void> (0) : __assert_fail ("(Cond == ISD::SETEQ || Cond == ISD::SETNE) && \"Only applicable for (in)equality comparisons.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5246, __PRETTY_FUNCTION__)); | ||||||
| 5247 | |||||||
| 5248 | SelectionDAG &DAG = DCI.DAG; | ||||||
| 5249 | |||||||
| 5250 | EVT VT = REMNode.getValueType(); | ||||||
| 5251 | EVT SVT = VT.getScalarType(); | ||||||
| 5252 | EVT ShVT = getShiftAmountTy(VT, DAG.getDataLayout()); | ||||||
| 5253 | EVT ShSVT = ShVT.getScalarType(); | ||||||
| 5254 | |||||||
| 5255 | // If MUL is unavailable, we cannot proceed in any case. | ||||||
| 5256 | if (!isOperationLegalOrCustom(ISD::MUL, VT)) | ||||||
| 5257 | return SDValue(); | ||||||
| 5258 | |||||||
| 5259 | // TODO: Could support comparing with non-zero too. | ||||||
| 5260 | ConstantSDNode *CompTarget = isConstOrConstSplat(CompTargetNode); | ||||||
| 5261 | if (!CompTarget || !CompTarget->isNullValue()) | ||||||
| 5262 | return SDValue(); | ||||||
| 5263 | |||||||
| 5264 | bool HadIntMinDivisor = false; | ||||||
| 5265 | bool HadOneDivisor = false; | ||||||
| 5266 | bool AllDivisorsAreOnes = true; | ||||||
| 5267 | bool HadEvenDivisor = false; | ||||||
| 5268 | bool NeedToApplyOffset = false; | ||||||
| 5269 | bool AllDivisorsArePowerOfTwo = true; | ||||||
| 5270 | SmallVector<SDValue, 16> PAmts, AAmts, KAmts, QAmts; | ||||||
| 5271 | |||||||
| 5272 | auto BuildSREMPattern = [&](ConstantSDNode *C) { | ||||||
| 5273 | // Division by 0 is UB. Leave it to be constant-folded elsewhere. | ||||||
| 5274 | if (C->isNullValue()) | ||||||
| 5275 | return false; | ||||||
| 5276 | |||||||
| 5277 | // FIXME: we don't fold `rem %X, -C` to `rem %X, C` in DAGCombine. | ||||||
| 5278 | |||||||
| 5279 | // WARNING: this fold is only valid for positive divisors! | ||||||
| 5280 | APInt D = C->getAPIntValue(); | ||||||
| 5281 | if (D.isNegative()) | ||||||
| 5282 | D.negate(); // `rem %X, -C` is equivalent to `rem %X, C` | ||||||
| 5283 | |||||||
| 5284 | HadIntMinDivisor |= D.isMinSignedValue(); | ||||||
| 5285 | |||||||
| 5286 | // If all divisors are ones, we will prefer to avoid the fold. | ||||||
| 5287 | HadOneDivisor |= D.isOneValue(); | ||||||
| 5288 | AllDivisorsAreOnes &= D.isOneValue(); | ||||||
| 5289 | |||||||
| 5290 | // Decompose D into D0 * 2^K | ||||||
| 5291 | unsigned K = D.countTrailingZeros(); | ||||||
| 5292 | assert((!D.isOneValue() || (K == 0)) && "For divisor '1' we won't rotate.")(((!D.isOneValue() || (K == 0)) && "For divisor '1' we won't rotate." ) ? static_cast<void> (0) : __assert_fail ("(!D.isOneValue() || (K == 0)) && \"For divisor '1' we won't rotate.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5292, __PRETTY_FUNCTION__)); | ||||||
| 5293 | APInt D0 = D.lshr(K); | ||||||
| 5294 | |||||||
| 5295 | if (!D.isMinSignedValue()) { | ||||||
| 5296 | // D is even if it has trailing zeros; unless it's INT_MIN, in which case | ||||||
| 5297 | // we don't care about this lane in this fold, we'll special-handle it. | ||||||
| 5298 | HadEvenDivisor |= (K != 0); | ||||||
| 5299 | } | ||||||
| 5300 | |||||||
| 5301 | // D is a power-of-two if D0 is one. This includes INT_MIN. | ||||||
| 5302 | // If all divisors are power-of-two, we will prefer to avoid the fold. | ||||||
| 5303 | AllDivisorsArePowerOfTwo &= D0.isOneValue(); | ||||||
| 5304 | |||||||
| 5305 | // P = inv(D0, 2^W) | ||||||
| 5306 | // 2^W requires W + 1 bits, so we have to extend and then truncate. | ||||||
| 5307 | unsigned W = D.getBitWidth(); | ||||||
| 5308 | APInt P = D0.zext(W + 1) | ||||||
| 5309 | .multiplicativeInverse(APInt::getSignedMinValue(W + 1)) | ||||||
| 5310 | .trunc(W); | ||||||
| 5311 | assert(!P.isNullValue() && "No multiplicative inverse!")((!P.isNullValue() && "No multiplicative inverse!") ? static_cast<void> (0) : __assert_fail ("!P.isNullValue() && \"No multiplicative inverse!\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5311, __PRETTY_FUNCTION__)); // unreachable | ||||||
| 5312 | assert((D0 * P).isOneValue() && "Multiplicative inverse sanity check.")(((D0 * P).isOneValue() && "Multiplicative inverse sanity check." ) ? static_cast<void> (0) : __assert_fail ("(D0 * P).isOneValue() && \"Multiplicative inverse sanity check.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5312, __PRETTY_FUNCTION__)); | ||||||
| 5313 | |||||||
| 5314 | // A = floor((2^(W - 1) - 1) / D0) & -2^K | ||||||
| 5315 | APInt A = APInt::getSignedMaxValue(W).udiv(D0); | ||||||
| 5316 | A.clearLowBits(K); | ||||||
| 5317 | |||||||
| 5318 | if (!D.isMinSignedValue()) { | ||||||
| 5319 | // If divisor INT_MIN, then we don't care about this lane in this fold, | ||||||
| 5320 | // we'll special-handle it. | ||||||
| 5321 | NeedToApplyOffset |= A != 0; | ||||||
| 5322 | } | ||||||
| 5323 | |||||||
| 5324 | // Q = floor((2 * A) / (2^K)) | ||||||
| 5325 | APInt Q = (2 * A).udiv(APInt::getOneBitSet(W, K)); | ||||||
| 5326 | |||||||
| 5327 | assert(APInt::getAllOnesValue(SVT.getSizeInBits()).ugt(A) &&((APInt::getAllOnesValue(SVT.getSizeInBits()).ugt(A) && "We are expecting that A is always less than all-ones for SVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(SVT.getSizeInBits()).ugt(A) && \"We are expecting that A is always less than all-ones for SVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5328, __PRETTY_FUNCTION__)) | ||||||
| 5328 | "We are expecting that A is always less than all-ones for SVT")((APInt::getAllOnesValue(SVT.getSizeInBits()).ugt(A) && "We are expecting that A is always less than all-ones for SVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(SVT.getSizeInBits()).ugt(A) && \"We are expecting that A is always less than all-ones for SVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5328, __PRETTY_FUNCTION__)); | ||||||
| 5329 | assert(APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) &&((APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && \"We are expecting that K is always less than all-ones for ShSVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5330, __PRETTY_FUNCTION__)) | ||||||
| 5330 | "We are expecting that K is always less than all-ones for ShSVT")((APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && "We are expecting that K is always less than all-ones for ShSVT" ) ? static_cast<void> (0) : __assert_fail ("APInt::getAllOnesValue(ShSVT.getSizeInBits()).ugt(K) && \"We are expecting that K is always less than all-ones for ShSVT\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5330, __PRETTY_FUNCTION__)); | ||||||
| 5331 | |||||||
| 5332 | // If the divisor is 1 the result can be constant-folded. Likewise, we | ||||||
| 5333 | // don't care about INT_MIN lanes, those can be set to undef if appropriate. | ||||||
| 5334 | if (D.isOneValue()) { | ||||||
| 5335 | // Set P, A and K to a bogus values so we can try to splat them. | ||||||
| 5336 | P = 0; | ||||||
| 5337 | A = -1; | ||||||
| 5338 | K = -1; | ||||||
| 5339 | |||||||
| 5340 | // x ?% 1 == 0 <--> true <--> x u<= -1 | ||||||
| 5341 | Q = -1; | ||||||
| 5342 | } | ||||||
| 5343 | |||||||
| 5344 | PAmts.push_back(DAG.getConstant(P, DL, SVT)); | ||||||
| 5345 | AAmts.push_back(DAG.getConstant(A, DL, SVT)); | ||||||
| 5346 | KAmts.push_back( | ||||||
| 5347 | DAG.getConstant(APInt(ShSVT.getSizeInBits(), K), DL, ShSVT)); | ||||||
| 5348 | QAmts.push_back(DAG.getConstant(Q, DL, SVT)); | ||||||
| 5349 | return true; | ||||||
| 5350 | }; | ||||||
| 5351 | |||||||
| 5352 | SDValue N = REMNode.getOperand(0); | ||||||
| 5353 | SDValue D = REMNode.getOperand(1); | ||||||
| 5354 | |||||||
| 5355 | // Collect the values from each element. | ||||||
| 5356 | if (!ISD::matchUnaryPredicate(D, BuildSREMPattern)) | ||||||
| 5357 | return SDValue(); | ||||||
| 5358 | |||||||
| 5359 | // If this is a srem by a one, avoid the fold since it can be constant-folded. | ||||||
| 5360 | if (AllDivisorsAreOnes) | ||||||
| 5361 | return SDValue(); | ||||||
| 5362 | |||||||
| 5363 | // If this is a srem by a powers-of-two (including INT_MIN), avoid the fold | ||||||
| 5364 | // since it can be best implemented as a bit test. | ||||||
| 5365 | if (AllDivisorsArePowerOfTwo) | ||||||
| 5366 | return SDValue(); | ||||||
| 5367 | |||||||
| 5368 | SDValue PVal, AVal, KVal, QVal; | ||||||
| 5369 | if (VT.isVector()) { | ||||||
| 5370 | if (HadOneDivisor) { | ||||||
| 5371 | // Try to turn PAmts into a splat, since we don't care about the values | ||||||
| 5372 | // that are currently '0'. If we can't, just keep '0'`s. | ||||||
| 5373 | turnVectorIntoSplatVector(PAmts, isNullConstant); | ||||||
| 5374 | // Try to turn AAmts into a splat, since we don't care about the | ||||||
| 5375 | // values that are currently '-1'. If we can't, change them to '0'`s. | ||||||
| 5376 | turnVectorIntoSplatVector(AAmts, isAllOnesConstant, | ||||||
| 5377 | DAG.getConstant(0, DL, SVT)); | ||||||
| 5378 | // Try to turn KAmts into a splat, since we don't care about the values | ||||||
| 5379 | // that are currently '-1'. If we can't, change them to '0'`s. | ||||||
| 5380 | turnVectorIntoSplatVector(KAmts, isAllOnesConstant, | ||||||
| 5381 | DAG.getConstant(0, DL, ShSVT)); | ||||||
| 5382 | } | ||||||
| 5383 | |||||||
| 5384 | PVal = DAG.getBuildVector(VT, DL, PAmts); | ||||||
| 5385 | AVal = DAG.getBuildVector(VT, DL, AAmts); | ||||||
| 5386 | KVal = DAG.getBuildVector(ShVT, DL, KAmts); | ||||||
| 5387 | QVal = DAG.getBuildVector(VT, DL, QAmts); | ||||||
| 5388 | } else { | ||||||
| 5389 | PVal = PAmts[0]; | ||||||
| 5390 | AVal = AAmts[0]; | ||||||
| 5391 | KVal = KAmts[0]; | ||||||
| 5392 | QVal = QAmts[0]; | ||||||
| 5393 | } | ||||||
| 5394 | |||||||
| 5395 | // (mul N, P) | ||||||
| 5396 | SDValue Op0 = DAG.getNode(ISD::MUL, DL, VT, N, PVal); | ||||||
| 5397 | Created.push_back(Op0.getNode()); | ||||||
| 5398 | |||||||
| 5399 | if (NeedToApplyOffset) { | ||||||
| 5400 | // We need ADD to do this. | ||||||
| 5401 | if (!isOperationLegalOrCustom(ISD::ADD, VT)) | ||||||
| 5402 | return SDValue(); | ||||||
| 5403 | |||||||
| 5404 | // (add (mul N, P), A) | ||||||
| 5405 | Op0 = DAG.getNode(ISD::ADD, DL, VT, Op0, AVal); | ||||||
| 5406 | Created.push_back(Op0.getNode()); | ||||||
| 5407 | } | ||||||
| 5408 | |||||||
| 5409 | // Rotate right only if any divisor was even. We avoid rotates for all-odd | ||||||
| 5410 | // divisors as a performance improvement, since rotating by 0 is a no-op. | ||||||
| 5411 | if (HadEvenDivisor) { | ||||||
| 5412 | // We need ROTR to do this. | ||||||
| 5413 | if (!isOperationLegalOrCustom(ISD::ROTR, VT)) | ||||||
| 5414 | return SDValue(); | ||||||
| 5415 | SDNodeFlags Flags; | ||||||
| 5416 | Flags.setExact(true); | ||||||
| 5417 | // SREM: (rotr (add (mul N, P), A), K) | ||||||
| 5418 | Op0 = DAG.getNode(ISD::ROTR, DL, VT, Op0, KVal, Flags); | ||||||
| 5419 | Created.push_back(Op0.getNode()); | ||||||
| 5420 | } | ||||||
| 5421 | |||||||
| 5422 | // SREM: (setule/setugt (rotr (add (mul N, P), A), K), Q) | ||||||
| 5423 | SDValue Fold = | ||||||
| 5424 | DAG.getSetCC(DL, SETCCVT, Op0, QVal, | ||||||
| 5425 | ((Cond == ISD::SETEQ) ? ISD::SETULE : ISD::SETUGT)); | ||||||
| 5426 | |||||||
| 5427 | // If we didn't have lanes with INT_MIN divisor, then we're done. | ||||||
| 5428 | if (!HadIntMinDivisor) | ||||||
| 5429 | return Fold; | ||||||
| 5430 | |||||||
| 5431 | // That fold is only valid for positive divisors. Which effectively means, | ||||||
| 5432 | // it is invalid for INT_MIN divisors. So if we have such a lane, | ||||||
| 5433 | // we must fix-up results for said lanes. | ||||||
| 5434 | assert(VT.isVector() && "Can/should only get here for vectors.")((VT.isVector() && "Can/should only get here for vectors." ) ? static_cast<void> (0) : __assert_fail ("VT.isVector() && \"Can/should only get here for vectors.\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5434, __PRETTY_FUNCTION__)); | ||||||
| 5435 | |||||||
| 5436 | if (!isOperationLegalOrCustom(ISD::SETEQ, VT) || | ||||||
| 5437 | !isOperationLegalOrCustom(ISD::AND, VT) || | ||||||
| 5438 | !isOperationLegalOrCustom(Cond, VT) || | ||||||
| 5439 | !isOperationLegalOrCustom(ISD::VSELECT, VT)) | ||||||
| 5440 | return SDValue(); | ||||||
| 5441 | |||||||
| 5442 | Created.push_back(Fold.getNode()); | ||||||
| 5443 | |||||||
| 5444 | SDValue IntMin = DAG.getConstant( | ||||||
| 5445 | APInt::getSignedMinValue(SVT.getScalarSizeInBits()), DL, VT); | ||||||
| 5446 | SDValue IntMax = DAG.getConstant( | ||||||
| 5447 | APInt::getSignedMaxValue(SVT.getScalarSizeInBits()), DL, VT); | ||||||
| 5448 | SDValue Zero = | ||||||
| 5449 | DAG.getConstant(APInt::getNullValue(SVT.getScalarSizeInBits()), DL, VT); | ||||||
| 5450 | |||||||
| 5451 | // Which lanes had INT_MIN divisors? Divisor is constant, so const-folded. | ||||||
| 5452 | SDValue DivisorIsIntMin = DAG.getSetCC(DL, SETCCVT, D, IntMin, ISD::SETEQ); | ||||||
| 5453 | Created.push_back(DivisorIsIntMin.getNode()); | ||||||
| 5454 | |||||||
| 5455 | // (N s% INT_MIN) ==/!= 0 <--> (N & INT_MAX) ==/!= 0 | ||||||
| 5456 | SDValue Masked = DAG.getNode(ISD::AND, DL, VT, N, IntMax); | ||||||
| 5457 | Created.push_back(Masked.getNode()); | ||||||
| 5458 | SDValue MaskedIsZero = DAG.getSetCC(DL, SETCCVT, Masked, Zero, Cond); | ||||||
| 5459 | Created.push_back(MaskedIsZero.getNode()); | ||||||
| 5460 | |||||||
| 5461 | // To produce final result we need to blend 2 vectors: 'SetCC' and | ||||||
| 5462 | // 'MaskedIsZero'. If the divisor for channel was *NOT* INT_MIN, we pick | ||||||
| 5463 | // from 'Fold', else pick from 'MaskedIsZero'. Since 'DivisorIsIntMin' is | ||||||
| 5464 | // constant-folded, select can get lowered to a shuffle with constant mask. | ||||||
| 5465 | SDValue Blended = | ||||||
| 5466 | DAG.getNode(ISD::VSELECT, DL, VT, DivisorIsIntMin, MaskedIsZero, Fold); | ||||||
| 5467 | |||||||
| 5468 | return Blended; | ||||||
| 5469 | } | ||||||
| 5470 | |||||||
| 5471 | bool TargetLowering:: | ||||||
| 5472 | verifyReturnAddressArgumentIsConstant(SDValue Op, SelectionDAG &DAG) const { | ||||||
| 5473 | if (!isa<ConstantSDNode>(Op.getOperand(0))) { | ||||||
| 5474 | DAG.getContext()->emitError("argument to '__builtin_return_address' must " | ||||||
| 5475 | "be a constant integer"); | ||||||
| 5476 | return true; | ||||||
| 5477 | } | ||||||
| 5478 | |||||||
| 5479 | return false; | ||||||
| 5480 | } | ||||||
| 5481 | |||||||
| 5482 | char TargetLowering::isNegatibleForFree(SDValue Op, SelectionDAG &DAG, | ||||||
| 5483 | bool LegalOperations, bool ForCodeSize, | ||||||
| 5484 | unsigned Depth) const { | ||||||
| 5485 | // fneg is removable even if it has multiple uses. | ||||||
| 5486 | if (Op.getOpcode() == ISD::FNEG) | ||||||
| 5487 | return 2; | ||||||
| 5488 | |||||||
| 5489 | // Don't allow anything with multiple uses unless we know it is free. | ||||||
| 5490 | EVT VT = Op.getValueType(); | ||||||
| 5491 | const SDNodeFlags Flags = Op->getFlags(); | ||||||
| 5492 | const TargetOptions &Options = DAG.getTarget().Options; | ||||||
| 5493 | if (!Op.hasOneUse() && !(Op.getOpcode() == ISD::FP_EXTEND && | ||||||
| 5494 | isFPExtFree(VT, Op.getOperand(0).getValueType()))) | ||||||
| 5495 | return 0; | ||||||
| 5496 | |||||||
| 5497 | // Don't recurse exponentially. | ||||||
| 5498 | if (Depth > SelectionDAG::MaxRecursionDepth) | ||||||
| 5499 | return 0; | ||||||
| 5500 | |||||||
| 5501 | switch (Op.getOpcode()) { | ||||||
| 5502 | case ISD::ConstantFP: { | ||||||
| 5503 | if (!LegalOperations) | ||||||
| 5504 | return 1; | ||||||
| 5505 | |||||||
| 5506 | // Don't invert constant FP values after legalization unless the target says | ||||||
| 5507 | // the negated constant is legal. | ||||||
| 5508 | return isOperationLegal(ISD::ConstantFP, VT) || | ||||||
| 5509 | isFPImmLegal(neg(cast<ConstantFPSDNode>(Op)->getValueAPF()), VT, | ||||||
| 5510 | ForCodeSize); | ||||||
| 5511 | } | ||||||
| 5512 | case ISD::BUILD_VECTOR: { | ||||||
| 5513 | // Only permit BUILD_VECTOR of constants. | ||||||
| 5514 | if (llvm::any_of(Op->op_values(), [&](SDValue N) { | ||||||
| 5515 | return !N.isUndef() && !isa<ConstantFPSDNode>(N); | ||||||
| 5516 | })) | ||||||
| 5517 | return 0; | ||||||
| 5518 | if (!LegalOperations) | ||||||
| 5519 | return 1; | ||||||
| 5520 | if (isOperationLegal(ISD::ConstantFP, VT) && | ||||||
| 5521 | isOperationLegal(ISD::BUILD_VECTOR, VT)) | ||||||
| 5522 | return 1; | ||||||
| 5523 | return llvm::all_of(Op->op_values(), [&](SDValue N) { | ||||||
| 5524 | return N.isUndef() || | ||||||
| 5525 | isFPImmLegal(neg(cast<ConstantFPSDNode>(N)->getValueAPF()), VT, | ||||||
| 5526 | ForCodeSize); | ||||||
| 5527 | }); | ||||||
| 5528 | } | ||||||
| 5529 | case ISD::FADD: | ||||||
| 5530 | if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) | ||||||
| 5531 | return 0; | ||||||
| 5532 | |||||||
| 5533 | // After operation legalization, it might not be legal to create new FSUBs. | ||||||
| 5534 | if (LegalOperations && !isOperationLegalOrCustom(ISD::FSUB, VT)) | ||||||
| 5535 | return 0; | ||||||
| 5536 | |||||||
| 5537 | // fold (fneg (fadd A, B)) -> (fsub (fneg A), B) | ||||||
| 5538 | if (char V = isNegatibleForFree(Op.getOperand(0), DAG, LegalOperations, | ||||||
| 5539 | ForCodeSize, Depth + 1)) | ||||||
| 5540 | return V; | ||||||
| 5541 | // fold (fneg (fadd A, B)) -> (fsub (fneg B), A) | ||||||
| 5542 | return isNegatibleForFree(Op.getOperand(1), DAG, LegalOperations, | ||||||
| 5543 | ForCodeSize, Depth + 1); | ||||||
| 5544 | case ISD::FSUB: | ||||||
| 5545 | // We can't turn -(A-B) into B-A when we honor signed zeros. | ||||||
| 5546 | if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) | ||||||
| 5547 | return 0; | ||||||
| 5548 | |||||||
| 5549 | // fold (fneg (fsub A, B)) -> (fsub B, A) | ||||||
| 5550 | return 1; | ||||||
| 5551 | |||||||
| 5552 | case ISD::FMUL: | ||||||
| 5553 | case ISD::FDIV: | ||||||
| 5554 | // fold (fneg (fmul X, Y)) -> (fmul (fneg X), Y) or (fmul X, (fneg Y)) | ||||||
| 5555 | if (char V = isNegatibleForFree(Op.getOperand(0), DAG, LegalOperations, | ||||||
| 5556 | ForCodeSize, Depth + 1)) | ||||||
| 5557 | return V; | ||||||
| 5558 | |||||||
| 5559 | // Ignore X * 2.0 because that is expected to be canonicalized to X + X. | ||||||
| 5560 | if (auto *C = isConstOrConstSplatFP(Op.getOperand(1))) | ||||||
| 5561 | if (C->isExactlyValue(2.0) && Op.getOpcode() == ISD::FMUL) | ||||||
| 5562 | return 0; | ||||||
| 5563 | |||||||
| 5564 | return isNegatibleForFree(Op.getOperand(1), DAG, LegalOperations, | ||||||
| 5565 | ForCodeSize, Depth + 1); | ||||||
| 5566 | |||||||
| 5567 | case ISD::FMA: | ||||||
| 5568 | case ISD::FMAD: { | ||||||
| 5569 | if (!Options.NoSignedZerosFPMath && !Flags.hasNoSignedZeros()) | ||||||
| 5570 | return 0; | ||||||
| 5571 | |||||||
| 5572 | // fold (fneg (fma X, Y, Z)) -> (fma (fneg X), Y, (fneg Z)) | ||||||
| 5573 | // fold (fneg (fma X, Y, Z)) -> (fma X, (fneg Y), (fneg Z)) | ||||||
| 5574 | char V2 = isNegatibleForFree(Op.getOperand(2), DAG, LegalOperations, | ||||||
| 5575 | ForCodeSize, Depth + 1); | ||||||
| 5576 | if (!V2) | ||||||
| 5577 | return 0; | ||||||
| 5578 | |||||||
| 5579 | // One of Op0/Op1 must be cheaply negatible, then select the cheapest. | ||||||
| 5580 | char V0 = isNegatibleForFree(Op.getOperand(0), DAG, LegalOperations, | ||||||
| 5581 | ForCodeSize, Depth + 1); | ||||||
| 5582 | char V1 = isNegatibleForFree(Op.getOperand(1), DAG, LegalOperations, | ||||||
| 5583 | ForCodeSize, Depth + 1); | ||||||
| 5584 | char V01 = std::max(V0, V1); | ||||||
| 5585 | return V01 ? std::max(V01, V2) : 0; | ||||||
| 5586 | } | ||||||
| 5587 | |||||||
| 5588 | case ISD::FP_EXTEND: | ||||||
| 5589 | case ISD::FP_ROUND: | ||||||
| 5590 | case ISD::FSIN: | ||||||
| 5591 | return isNegatibleForFree(Op.getOperand(0), DAG, LegalOperations, | ||||||
| 5592 | ForCodeSize, Depth + 1); | ||||||
| 5593 | } | ||||||
| 5594 | |||||||
| 5595 | return 0; | ||||||
| 5596 | } | ||||||
| 5597 | |||||||
| 5598 | SDValue TargetLowering::getNegatedExpression(SDValue Op, SelectionDAG &DAG, | ||||||
| 5599 | bool LegalOperations, | ||||||
| 5600 | bool ForCodeSize, | ||||||
| 5601 | unsigned Depth) const { | ||||||
| 5602 | // fneg is removable even if it has multiple uses. | ||||||
| 5603 | if (Op.getOpcode() == ISD::FNEG) | ||||||
| 5604 | return Op.getOperand(0); | ||||||
| 5605 | |||||||
| 5606 | assert(Depth <= SelectionDAG::MaxRecursionDepth &&((Depth <= SelectionDAG::MaxRecursionDepth && "getNegatedExpression doesn't match isNegatibleForFree" ) ? static_cast<void> (0) : __assert_fail ("Depth <= SelectionDAG::MaxRecursionDepth && \"getNegatedExpression doesn't match isNegatibleForFree\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5607, __PRETTY_FUNCTION__)) | ||||||
| 5607 | "getNegatedExpression doesn't match isNegatibleForFree")((Depth <= SelectionDAG::MaxRecursionDepth && "getNegatedExpression doesn't match isNegatibleForFree" ) ? static_cast<void> (0) : __assert_fail ("Depth <= SelectionDAG::MaxRecursionDepth && \"getNegatedExpression doesn't match isNegatibleForFree\"" , "/build/llvm-toolchain-snapshot-10~++20200112100611+7fa5290d5bd/llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp" , 5607, __PRETTY_FUNCTION__)); | ||||||
| 5608 | const SDNodeFlags Flags = Op->getFlags(); | ||||||
| 5609 | |||||||
| 5610 | switch (Op.getOpcode()) { | ||||||
| 5611 | case ISD::ConstantFP: { | ||||||
| 5612 | APFloat V = cast<ConstantFPSDNode>(Op)->getValueAPF(); | ||||||
| 5613 | V.changeSign(); | ||||||
| 5614 | return DAG.getConstantFP(V, SDLoc(Op), Op.getValueType()); | ||||||
| 5615 | } | ||||||
| 5616 | case ISD::BUILD_VECTOR: { | ||||||
| 5617 | SmallVector<SDValue, 4> Ops; | ||||||
| 5618 | for (SDValue C : Op->op_values()) { | ||||||
| 5619 | if (C.isUndef()) { | ||||||
| 5620 | Ops.push_back(C); | ||||||
| 5621 | continue; | ||||||
| 5622 | } | ||||||
| 5623 | APFloat V = cast<ConstantFPSDNode>(C)->getValueAPF(); | ||||||
| 5624 | V.changeSign(); | ||||||
| 5625 | Ops.push_back(DAG.getConstantFP(V, SDLoc(Op), C.getValueType())); | ||||||
| 5626 | } | ||||||
| 5627 | return DAG.getBuildVector(Op.getValueType(), SDLoc(Op), Ops); | ||||||
| 5628 | } | ||||||
| 5629 | case ISD::FADD: | ||||||
| 5630 | assert((DAG.getTarget() |