| 1 | //===-- PPCISelDAGToDAG.cpp - PPC --pattern matching inst selector --------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This file defines a pattern matching instruction selector for PowerPC, |
| 10 | // converting from a legalized dag to a PPC dag. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "MCTargetDesc/PPCMCTargetDesc.h" |
| 15 | #include "MCTargetDesc/PPCPredicates.h" |
| 16 | #include "PPC.h" |
| 17 | #include "PPCISelLowering.h" |
| 18 | #include "PPCMachineFunctionInfo.h" |
| 19 | #include "PPCSelectionDAGInfo.h" |
| 20 | #include "PPCSubtarget.h" |
| 21 | #include "PPCTargetMachine.h" |
| 22 | #include "llvm/ADT/APInt.h" |
| 23 | #include "llvm/ADT/APSInt.h" |
| 24 | #include "llvm/ADT/DenseMap.h" |
| 25 | #include "llvm/ADT/STLExtras.h" |
| 26 | #include "llvm/ADT/SmallPtrSet.h" |
| 27 | #include "llvm/ADT/SmallVector.h" |
| 28 | #include "llvm/ADT/Statistic.h" |
| 29 | #include "llvm/Analysis/BranchProbabilityInfo.h" |
| 30 | #include "llvm/CodeGen/FunctionLoweringInfo.h" |
| 31 | #include "llvm/CodeGen/ISDOpcodes.h" |
| 32 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 33 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 34 | #include "llvm/CodeGen/MachineFunction.h" |
| 35 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 36 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 37 | #include "llvm/CodeGen/SelectionDAG.h" |
| 38 | #include "llvm/CodeGen/SelectionDAGISel.h" |
| 39 | #include "llvm/CodeGen/SelectionDAGNodes.h" |
| 40 | #include "llvm/CodeGen/TargetInstrInfo.h" |
| 41 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 42 | #include "llvm/CodeGen/ValueTypes.h" |
| 43 | #include "llvm/CodeGenTypes/MachineValueType.h" |
| 44 | #include "llvm/IR/BasicBlock.h" |
| 45 | #include "llvm/IR/DebugLoc.h" |
| 46 | #include "llvm/IR/Function.h" |
| 47 | #include "llvm/IR/GlobalValue.h" |
| 48 | #include "llvm/IR/InlineAsm.h" |
| 49 | #include "llvm/IR/InstrTypes.h" |
| 50 | #include "llvm/IR/IntrinsicsPowerPC.h" |
| 51 | #include "llvm/IR/Module.h" |
| 52 | #include "llvm/Support/Casting.h" |
| 53 | #include "llvm/Support/CodeGen.h" |
| 54 | #include "llvm/Support/CommandLine.h" |
| 55 | #include "llvm/Support/Compiler.h" |
| 56 | #include "llvm/Support/Debug.h" |
| 57 | #include "llvm/Support/ErrorHandling.h" |
| 58 | #include "llvm/Support/KnownBits.h" |
| 59 | #include "llvm/Support/MathExtras.h" |
| 60 | #include "llvm/Support/raw_ostream.h" |
| 61 | #include <algorithm> |
| 62 | #include <cassert> |
| 63 | #include <cstdint> |
| 64 | #include <iterator> |
| 65 | #include <limits> |
| 66 | #include <memory> |
| 67 | #include <new> |
| 68 | #include <tuple> |
| 69 | #include <utility> |
| 70 | |
| 71 | using namespace llvm; |
| 72 | |
| 73 | #define DEBUG_TYPE "ppc-isel" |
| 74 | #define PASS_NAME "PowerPC DAG->DAG Pattern Instruction Selection" |
| 75 | |
| 76 | STATISTIC(NumSextSetcc, |
| 77 | "Number of (sext(setcc)) nodes expanded into GPR sequence." ); |
| 78 | STATISTIC(NumZextSetcc, |
| 79 | "Number of (zext(setcc)) nodes expanded into GPR sequence." ); |
| 80 | STATISTIC(SignExtensionsAdded, |
| 81 | "Number of sign extensions for compare inputs added." ); |
| 82 | STATISTIC(ZeroExtensionsAdded, |
| 83 | "Number of zero extensions for compare inputs added." ); |
| 84 | STATISTIC(NumLogicOpsOnComparison, |
| 85 | "Number of logical ops on i1 values calculated in GPR." ); |
| 86 | STATISTIC(OmittedForNonExtendUses, |
| 87 | "Number of compares not eliminated as they have non-extending uses." ); |
| 88 | STATISTIC(NumP9Setb, |
| 89 | "Number of compares lowered to setb." ); |
| 90 | |
| 91 | // FIXME: Remove this once the bug has been fixed! |
| 92 | cl::opt<bool> ANDIGlueBug("expose-ppc-andi-glue-bug" , |
| 93 | cl::desc("expose the ANDI glue bug on PPC" ), cl::Hidden); |
| 94 | |
| 95 | static cl::opt<bool> |
| 96 | UseBitPermRewriter("ppc-use-bit-perm-rewriter" , cl::init(Val: true), |
| 97 | cl::desc("use aggressive ppc isel for bit permutations" ), |
| 98 | cl::Hidden); |
| 99 | static cl::opt<bool> BPermRewriterNoMasking( |
| 100 | "ppc-bit-perm-rewriter-stress-rotates" , |
| 101 | cl::desc("stress rotate selection in aggressive ppc isel for " |
| 102 | "bit permutations" ), |
| 103 | cl::Hidden); |
| 104 | |
| 105 | static cl::opt<bool> EnableBranchHint( |
| 106 | "ppc-use-branch-hint" , cl::init(Val: true), |
| 107 | cl::desc("Enable static hinting of branches on ppc" ), |
| 108 | cl::Hidden); |
| 109 | |
| 110 | static cl::opt<bool> EnableTLSOpt( |
| 111 | "ppc-tls-opt" , cl::init(Val: true), |
| 112 | cl::desc("Enable tls optimization peephole" ), |
| 113 | cl::Hidden); |
| 114 | |
| 115 | enum ICmpInGPRType { ICGPR_All, ICGPR_None, ICGPR_I32, ICGPR_I64, |
| 116 | ICGPR_NonExtIn, ICGPR_Zext, ICGPR_Sext, ICGPR_ZextI32, |
| 117 | ICGPR_SextI32, ICGPR_ZextI64, ICGPR_SextI64 }; |
| 118 | |
| 119 | static cl::opt<ICmpInGPRType> CmpInGPR( |
| 120 | "ppc-gpr-icmps" , cl::Hidden, cl::init(Val: ICGPR_All), |
| 121 | cl::desc("Specify the types of comparisons to emit GPR-only code for." ), |
| 122 | cl::values(clEnumValN(ICGPR_None, "none" , "Do not modify integer comparisons." ), |
| 123 | clEnumValN(ICGPR_All, "all" , "All possible int comparisons in GPRs." ), |
| 124 | clEnumValN(ICGPR_I32, "i32" , "Only i32 comparisons in GPRs." ), |
| 125 | clEnumValN(ICGPR_I64, "i64" , "Only i64 comparisons in GPRs." ), |
| 126 | clEnumValN(ICGPR_NonExtIn, "nonextin" , |
| 127 | "Only comparisons where inputs don't need [sz]ext." ), |
| 128 | clEnumValN(ICGPR_Zext, "zext" , "Only comparisons with zext result." ), |
| 129 | clEnumValN(ICGPR_ZextI32, "zexti32" , |
| 130 | "Only i32 comparisons with zext result." ), |
| 131 | clEnumValN(ICGPR_ZextI64, "zexti64" , |
| 132 | "Only i64 comparisons with zext result." ), |
| 133 | clEnumValN(ICGPR_Sext, "sext" , "Only comparisons with sext result." ), |
| 134 | clEnumValN(ICGPR_SextI32, "sexti32" , |
| 135 | "Only i32 comparisons with sext result." ), |
| 136 | clEnumValN(ICGPR_SextI64, "sexti64" , |
| 137 | "Only i64 comparisons with sext result." ))); |
| 138 | namespace { |
| 139 | |
| 140 | //===--------------------------------------------------------------------===// |
| 141 | /// PPCDAGToDAGISel - PPC specific code to select PPC machine |
| 142 | /// instructions for SelectionDAG operations. |
| 143 | /// |
| 144 | class PPCDAGToDAGISel : public SelectionDAGISel { |
| 145 | const PPCTargetMachine &TM; |
| 146 | const PPCSubtarget *Subtarget = nullptr; |
| 147 | const PPCTargetLowering *PPCLowering = nullptr; |
| 148 | unsigned GlobalBaseReg = 0; |
| 149 | |
| 150 | public: |
| 151 | PPCDAGToDAGISel() = delete; |
| 152 | |
| 153 | explicit PPCDAGToDAGISel(PPCTargetMachine &tm, CodeGenOptLevel OptLevel) |
| 154 | : SelectionDAGISel(tm, OptLevel), TM(tm) {} |
| 155 | |
| 156 | bool runOnMachineFunction(MachineFunction &MF) override { |
| 157 | // Make sure we re-emit a set of the global base reg if necessary |
| 158 | GlobalBaseReg = 0; |
| 159 | Subtarget = &MF.getSubtarget<PPCSubtarget>(); |
| 160 | PPCLowering = Subtarget->getTargetLowering(); |
| 161 | if (Subtarget->hasROPProtect()) { |
| 162 | // Create a place on the stack for the ROP Protection Hash. |
| 163 | // The ROP Protection Hash will always be 8 bytes and aligned to 8 |
| 164 | // bytes. |
| 165 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 166 | PPCFunctionInfo *FI = MF.getInfo<PPCFunctionInfo>(); |
| 167 | const int Result = MFI.CreateStackObject(Size: 8, Alignment: Align(8), isSpillSlot: false); |
| 168 | FI->setROPProtectionHashSaveIndex(Result); |
| 169 | } |
| 170 | SelectionDAGISel::runOnMachineFunction(mf&: MF); |
| 171 | |
| 172 | return true; |
| 173 | } |
| 174 | |
| 175 | void PreprocessISelDAG() override; |
| 176 | void PostprocessISelDAG() override; |
| 177 | |
| 178 | /// getI16Imm - Return a target constant with the specified value, of type |
| 179 | /// i16. |
| 180 | inline SDValue getI16Imm(unsigned Imm, const SDLoc &dl) { |
| 181 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i16); |
| 182 | } |
| 183 | |
| 184 | /// getI32Imm - Return a target constant with the specified value, of type |
| 185 | /// i32. |
| 186 | inline SDValue getI32Imm(unsigned Imm, const SDLoc &dl) { |
| 187 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
| 188 | } |
| 189 | |
| 190 | /// getI64Imm - Return a target constant with the specified value, of type |
| 191 | /// i64. |
| 192 | inline SDValue getI64Imm(uint64_t Imm, const SDLoc &dl) { |
| 193 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i64); |
| 194 | } |
| 195 | |
| 196 | /// getSmallIPtrImm - Return a target constant of pointer type. |
| 197 | inline SDValue getSmallIPtrImm(int64_t Imm, const SDLoc &dl) { |
| 198 | return CurDAG->getSignedTargetConstant( |
| 199 | Val: Imm, DL: dl, VT: PPCLowering->getPointerTy(DL: CurDAG->getDataLayout())); |
| 200 | } |
| 201 | |
| 202 | /// isRotateAndMask - Returns true if Mask and Shift can be folded into a |
| 203 | /// rotate and mask opcode and mask operation. |
| 204 | static bool isRotateAndMask(SDNode *N, unsigned Mask, bool isShiftMask, |
| 205 | unsigned &SH, unsigned &MB, unsigned &ME); |
| 206 | |
| 207 | /// getGlobalBaseReg - insert code into the entry mbb to materialize the PIC |
| 208 | /// base register. Return the virtual register that holds this value. |
| 209 | SDNode *getGlobalBaseReg(); |
| 210 | |
| 211 | void selectFrameIndex(SDNode *SN, SDNode *N, int64_t Offset = 0); |
| 212 | |
| 213 | // Select - Convert the specified operand from a target-independent to a |
| 214 | // target-specific node if it hasn't already been changed. |
| 215 | void Select(SDNode *N) override; |
| 216 | |
| 217 | bool tryBitfieldInsert(SDNode *N); |
| 218 | bool tryBitPermutation(SDNode *N); |
| 219 | bool tryIntCompareInGPR(SDNode *N); |
| 220 | |
| 221 | // tryTLSXFormLoad - Convert an ISD::LOAD fed by a PPCISD::ADD_TLS into |
| 222 | // an X-Form load instruction with the offset being a relocation coming from |
| 223 | // the PPCISD::ADD_TLS. |
| 224 | bool tryTLSXFormLoad(LoadSDNode *N); |
| 225 | // tryTLSXFormStore - Convert an ISD::STORE fed by a PPCISD::ADD_TLS into |
| 226 | // an X-Form store instruction with the offset being a relocation coming from |
| 227 | // the PPCISD::ADD_TLS. |
| 228 | bool tryTLSXFormStore(StoreSDNode *N); |
| 229 | /// SelectCC - Select a comparison of the specified values with the |
| 230 | /// specified condition code, returning the CR# of the expression. |
| 231 | SDValue SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 232 | const SDLoc &dl, SDValue Chain = SDValue()); |
| 233 | |
| 234 | /// SelectAddrImmOffs - Return true if the operand is valid for a preinc |
| 235 | /// immediate field. Note that the operand at this point is already the |
| 236 | /// result of a prior SelectAddressRegImm call. |
| 237 | bool SelectAddrImmOffs(SDValue N, SDValue &Out) const { |
| 238 | if (N.getOpcode() == ISD::TargetConstant || |
| 239 | N.getOpcode() == ISD::TargetGlobalAddress) { |
| 240 | Out = N; |
| 241 | return true; |
| 242 | } |
| 243 | |
| 244 | return false; |
| 245 | } |
| 246 | |
| 247 | /// SelectDSForm - Returns true if address N can be represented by the |
| 248 | /// addressing mode of DSForm instructions (a base register, plus a signed |
| 249 | /// 16-bit displacement that is a multiple of 4. |
| 250 | bool SelectDSForm(SDNode *Parent, SDValue N, SDValue &Disp, SDValue &Base) { |
| 251 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 252 | Align: Align(4)) == PPC::AM_DSForm; |
| 253 | } |
| 254 | |
| 255 | /// SelectDQForm - Returns true if address N can be represented by the |
| 256 | /// addressing mode of DQForm instructions (a base register, plus a signed |
| 257 | /// 16-bit displacement that is a multiple of 16. |
| 258 | bool SelectDQForm(SDNode *Parent, SDValue N, SDValue &Disp, SDValue &Base) { |
| 259 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 260 | Align: Align(16)) == PPC::AM_DQForm; |
| 261 | } |
| 262 | |
| 263 | /// SelectDForm - Returns true if address N can be represented by |
| 264 | /// the addressing mode of DForm instructions (a base register, plus a |
| 265 | /// signed 16-bit immediate. |
| 266 | bool SelectDForm(SDNode *Parent, SDValue N, SDValue &Disp, SDValue &Base) { |
| 267 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 268 | Align: std::nullopt) == PPC::AM_DForm; |
| 269 | } |
| 270 | |
| 271 | /// SelectPCRelForm - Returns true if address N can be represented by |
| 272 | /// PC-Relative addressing mode. |
| 273 | bool SelectPCRelForm(SDNode *Parent, SDValue N, SDValue &Disp, |
| 274 | SDValue &Base) { |
| 275 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 276 | Align: std::nullopt) == PPC::AM_PCRel; |
| 277 | } |
| 278 | |
| 279 | /// SelectPDForm - Returns true if address N can be represented by Prefixed |
| 280 | /// DForm addressing mode (a base register, plus a signed 34-bit immediate. |
| 281 | bool SelectPDForm(SDNode *Parent, SDValue N, SDValue &Disp, SDValue &Base) { |
| 282 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 283 | Align: std::nullopt) == |
| 284 | PPC::AM_PrefixDForm; |
| 285 | } |
| 286 | |
| 287 | /// SelectXForm - Returns true if address N can be represented by the |
| 288 | /// addressing mode of XForm instructions (an indexed [r+r] operation). |
| 289 | bool SelectXForm(SDNode *Parent, SDValue N, SDValue &Disp, SDValue &Base) { |
| 290 | return PPCLowering->SelectOptimalAddrMode(Parent, N, Disp, Base, DAG&: *CurDAG, |
| 291 | Align: std::nullopt) == PPC::AM_XForm; |
| 292 | } |
| 293 | |
| 294 | /// SelectForceXForm - Given the specified address, force it to be |
| 295 | /// represented as an indexed [r+r] operation (an XForm instruction). |
| 296 | bool SelectForceXForm(SDNode *Parent, SDValue N, SDValue &Disp, |
| 297 | SDValue &Base) { |
| 298 | return PPCLowering->SelectForceXFormMode(N, Disp, Base, DAG&: *CurDAG) == |
| 299 | PPC::AM_XForm; |
| 300 | } |
| 301 | |
| 302 | /// SelectAddrIdx - Given the specified address, check to see if it can be |
| 303 | /// represented as an indexed [r+r] operation. |
| 304 | /// This is for xform instructions whose associated displacement form is D. |
| 305 | /// The last parameter \p 0 means associated D form has no requirment for 16 |
| 306 | /// bit signed displacement. |
| 307 | /// Returns false if it can be represented by [r+imm], which are preferred. |
| 308 | bool SelectAddrIdx(SDValue N, SDValue &Base, SDValue &Index) { |
| 309 | return PPCLowering->SelectAddressRegReg(N, Base, Index, DAG&: *CurDAG, |
| 310 | EncodingAlignment: std::nullopt); |
| 311 | } |
| 312 | |
| 313 | /// SelectAddrIdx4 - Given the specified address, check to see if it can be |
| 314 | /// represented as an indexed [r+r] operation. |
| 315 | /// This is for xform instructions whose associated displacement form is DS. |
| 316 | /// The last parameter \p 4 means associated DS form 16 bit signed |
| 317 | /// displacement must be a multiple of 4. |
| 318 | /// Returns false if it can be represented by [r+imm], which are preferred. |
| 319 | bool SelectAddrIdxX4(SDValue N, SDValue &Base, SDValue &Index) { |
| 320 | return PPCLowering->SelectAddressRegReg(N, Base, Index, DAG&: *CurDAG, |
| 321 | EncodingAlignment: Align(4)); |
| 322 | } |
| 323 | |
| 324 | /// SelectAddrIdx16 - Given the specified address, check to see if it can be |
| 325 | /// represented as an indexed [r+r] operation. |
| 326 | /// This is for xform instructions whose associated displacement form is DQ. |
| 327 | /// The last parameter \p 16 means associated DQ form 16 bit signed |
| 328 | /// displacement must be a multiple of 16. |
| 329 | /// Returns false if it can be represented by [r+imm], which are preferred. |
| 330 | bool SelectAddrIdxX16(SDValue N, SDValue &Base, SDValue &Index) { |
| 331 | return PPCLowering->SelectAddressRegReg(N, Base, Index, DAG&: *CurDAG, |
| 332 | EncodingAlignment: Align(16)); |
| 333 | } |
| 334 | |
| 335 | /// SelectAddrIdxOnly - Given the specified address, force it to be |
| 336 | /// represented as an indexed [r+r] operation. |
| 337 | bool SelectAddrIdxOnly(SDValue N, SDValue &Base, SDValue &Index) { |
| 338 | return PPCLowering->SelectAddressRegRegOnly(N, Base, Index, DAG&: *CurDAG); |
| 339 | } |
| 340 | |
| 341 | /// SelectAddrImm - Returns true if the address N can be represented by |
| 342 | /// a base register plus a signed 16-bit displacement [r+imm]. |
| 343 | /// The last parameter \p 0 means D form has no requirment for 16 bit signed |
| 344 | /// displacement. |
| 345 | bool SelectAddrImm(SDValue N, SDValue &Disp, |
| 346 | SDValue &Base) { |
| 347 | return PPCLowering->SelectAddressRegImm(N, Disp, Base, DAG&: *CurDAG, |
| 348 | EncodingAlignment: std::nullopt); |
| 349 | } |
| 350 | |
| 351 | /// SelectAddrImmX4 - Returns true if the address N can be represented by |
| 352 | /// a base register plus a signed 16-bit displacement that is a multiple of |
| 353 | /// 4 (last parameter). Suitable for use by STD and friends. |
| 354 | bool SelectAddrImmX4(SDValue N, SDValue &Disp, SDValue &Base) { |
| 355 | return PPCLowering->SelectAddressRegImm(N, Disp, Base, DAG&: *CurDAG, EncodingAlignment: Align(4)); |
| 356 | } |
| 357 | |
| 358 | /// SelectAddrImmX16 - Returns true if the address N can be represented by |
| 359 | /// a base register plus a signed 16-bit displacement that is a multiple of |
| 360 | /// 16(last parameter). Suitable for use by STXV and friends. |
| 361 | bool SelectAddrImmX16(SDValue N, SDValue &Disp, SDValue &Base) { |
| 362 | return PPCLowering->SelectAddressRegImm(N, Disp, Base, DAG&: *CurDAG, |
| 363 | EncodingAlignment: Align(16)); |
| 364 | } |
| 365 | |
| 366 | /// SelectAddrImmX34 - Returns true if the address N can be represented by |
| 367 | /// a base register plus a signed 34-bit displacement. Suitable for use by |
| 368 | /// PSTXVP and friends. |
| 369 | bool SelectAddrImmX34(SDValue N, SDValue &Disp, SDValue &Base) { |
| 370 | return PPCLowering->SelectAddressRegImm34(N, Disp, Base, DAG&: *CurDAG); |
| 371 | } |
| 372 | |
| 373 | // Select an address into a single register. |
| 374 | bool SelectAddr(SDValue N, SDValue &Base) { |
| 375 | Base = N; |
| 376 | return true; |
| 377 | } |
| 378 | |
| 379 | bool SelectAddrPCRel(SDValue N, SDValue &Base) { |
| 380 | return PPCLowering->SelectAddressPCRel(N, Base); |
| 381 | } |
| 382 | |
| 383 | /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for |
| 384 | /// inline asm expressions. It is always correct to compute the value into |
| 385 | /// a register. The case of adding a (possibly relocatable) constant to a |
| 386 | /// register can be improved, but it is wrong to substitute Reg+Reg for |
| 387 | /// Reg in an asm, because the load or store opcode would have to change. |
| 388 | bool SelectInlineAsmMemoryOperand(const SDValue &Op, |
| 389 | InlineAsm::ConstraintCode ConstraintID, |
| 390 | std::vector<SDValue> &OutOps) override { |
| 391 | switch(ConstraintID) { |
| 392 | default: |
| 393 | errs() << "ConstraintID: " |
| 394 | << InlineAsm::getMemConstraintName(C: ConstraintID) << "\n" ; |
| 395 | llvm_unreachable("Unexpected asm memory constraint" ); |
| 396 | case InlineAsm::ConstraintCode::es: |
| 397 | case InlineAsm::ConstraintCode::m: |
| 398 | case InlineAsm::ConstraintCode::o: |
| 399 | case InlineAsm::ConstraintCode::Q: |
| 400 | case InlineAsm::ConstraintCode::Z: |
| 401 | case InlineAsm::ConstraintCode::Zy: |
| 402 | // We need to make sure that this one operand does not end up in r0 |
| 403 | // (because we might end up lowering this as 0(%op)). |
| 404 | const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo(); |
| 405 | const TargetRegisterClass *TRC = TRI->getPointerRegClass(/*Kind=*/1); |
| 406 | SDLoc dl(Op); |
| 407 | SDValue RC = CurDAG->getTargetConstant(Val: TRC->getID(), DL: dl, VT: MVT::i32); |
| 408 | SDValue NewOp = |
| 409 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::COPY_TO_REGCLASS, |
| 410 | dl, VT: Op.getValueType(), |
| 411 | Op1: Op, Op2: RC), 0); |
| 412 | |
| 413 | OutOps.push_back(x: NewOp); |
| 414 | return false; |
| 415 | } |
| 416 | return true; |
| 417 | } |
| 418 | |
| 419 | // Include the pieces autogenerated from the target description. |
| 420 | #include "PPCGenDAGISel.inc" |
| 421 | |
| 422 | private: |
| 423 | bool trySETCC(SDNode *N); |
| 424 | bool tryFoldSWTestBRCC(SDNode *N); |
| 425 | bool trySelectLoopCountIntrinsic(SDNode *N); |
| 426 | bool tryAsSingleRLDICL(SDNode *N); |
| 427 | bool tryAsSingleRLDCL(SDNode *N); |
| 428 | bool tryAsSingleRLDICR(SDNode *N); |
| 429 | bool tryAsSingleRLWINM(SDNode *N); |
| 430 | bool tryAsSingleRLWINM8(SDNode *N); |
| 431 | bool tryAsSingleRLWIMI(SDNode *N); |
| 432 | bool tryAsPairOfRLDICL(SDNode *N); |
| 433 | bool tryAsSingleRLDIMI(SDNode *N); |
| 434 | |
| 435 | void PeepholePPC64(); |
| 436 | void PeepholePPC64ZExt(); |
| 437 | void PeepholeCROps(); |
| 438 | |
| 439 | SDValue combineToCMPB(SDNode *N); |
| 440 | void foldBoolExts(SDValue &Res, SDNode *&N); |
| 441 | |
| 442 | bool AllUsersSelectZero(SDNode *N); |
| 443 | void SwapAllSelectUsers(SDNode *N); |
| 444 | |
| 445 | bool isOffsetMultipleOf(SDNode *N, unsigned Val) const; |
| 446 | void transferMemOperands(SDNode *N, SDNode *Result); |
| 447 | }; |
| 448 | |
| 449 | class PPCDAGToDAGISelLegacy : public SelectionDAGISelLegacy { |
| 450 | public: |
| 451 | static char ID; |
| 452 | explicit PPCDAGToDAGISelLegacy(PPCTargetMachine &tm, |
| 453 | CodeGenOptLevel OptLevel) |
| 454 | : SelectionDAGISelLegacy( |
| 455 | ID, std::make_unique<PPCDAGToDAGISel>(args&: tm, args&: OptLevel)) {} |
| 456 | }; |
| 457 | } // end anonymous namespace |
| 458 | |
| 459 | char PPCDAGToDAGISelLegacy::ID = 0; |
| 460 | |
| 461 | INITIALIZE_PASS(PPCDAGToDAGISelLegacy, DEBUG_TYPE, PASS_NAME, false, false) |
| 462 | |
| 463 | /// getGlobalBaseReg - Output the instructions required to put the |
| 464 | /// base address to use for accessing globals into a register. |
| 465 | /// |
| 466 | SDNode *PPCDAGToDAGISel::getGlobalBaseReg() { |
| 467 | if (!GlobalBaseReg) { |
| 468 | const TargetInstrInfo &TII = *Subtarget->getInstrInfo(); |
| 469 | // Insert the set of GlobalBaseReg into the first MBB of the function |
| 470 | MachineBasicBlock &FirstMBB = MF->front(); |
| 471 | MachineBasicBlock::iterator MBBI = FirstMBB.begin(); |
| 472 | const Module *M = MF->getFunction().getParent(); |
| 473 | DebugLoc dl; |
| 474 | |
| 475 | if (PPCLowering->getPointerTy(DL: CurDAG->getDataLayout()) == MVT::i32) { |
| 476 | if (Subtarget->isTargetELF()) { |
| 477 | GlobalBaseReg = PPC::R30; |
| 478 | if (!Subtarget->isSecurePlt() && |
| 479 | M->getPICLevel() == PICLevel::SmallPIC) { |
| 480 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MoveGOTtoLR)); |
| 481 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MFLR), DestReg: GlobalBaseReg); |
| 482 | MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); |
| 483 | } else { |
| 484 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MovePCtoLR)); |
| 485 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MFLR), DestReg: GlobalBaseReg); |
| 486 | Register TempReg = RegInfo->createVirtualRegister(RegClass: &PPC::GPRCRegClass); |
| 487 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, |
| 488 | MCID: TII.get(Opcode: PPC::UpdateGBR), DestReg: GlobalBaseReg) |
| 489 | .addReg(RegNo: TempReg, Flags: RegState::Define).addReg(RegNo: GlobalBaseReg); |
| 490 | MF->getInfo<PPCFunctionInfo>()->setUsesPICBase(true); |
| 491 | } |
| 492 | } else { |
| 493 | GlobalBaseReg = |
| 494 | RegInfo->createVirtualRegister(RegClass: &PPC::GPRC_and_GPRC_NOR0RegClass); |
| 495 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MovePCtoLR)); |
| 496 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MFLR), DestReg: GlobalBaseReg); |
| 497 | } |
| 498 | } else { |
| 499 | // We must ensure that this sequence is dominated by the prologue. |
| 500 | // FIXME: This is a bit of a big hammer since we don't get the benefits |
| 501 | // of shrink-wrapping whenever we emit this instruction. Considering |
| 502 | // this is used in any function where we emit a jump table, this may be |
| 503 | // a significant limitation. We should consider inserting this in the |
| 504 | // block where it is used and then commoning this sequence up if it |
| 505 | // appears in multiple places. |
| 506 | // Note: on ISA 3.0 cores, we can use lnia (addpcis) instead of |
| 507 | // MovePCtoLR8. |
| 508 | MF->getInfo<PPCFunctionInfo>()->setShrinkWrapDisabled(true); |
| 509 | GlobalBaseReg = RegInfo->createVirtualRegister(RegClass: &PPC::G8RC_and_G8RC_NOX0RegClass); |
| 510 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MovePCtoLR8)); |
| 511 | BuildMI(BB&: FirstMBB, I: MBBI, MIMD: dl, MCID: TII.get(Opcode: PPC::MFLR8), DestReg: GlobalBaseReg); |
| 512 | } |
| 513 | } |
| 514 | return CurDAG->getRegister(Reg: GlobalBaseReg, |
| 515 | VT: PPCLowering->getPointerTy(DL: CurDAG->getDataLayout())) |
| 516 | .getNode(); |
| 517 | } |
| 518 | |
| 519 | // Check if a SDValue has the toc-data attribute. |
| 520 | static bool hasTocDataAttr(SDValue Val) { |
| 521 | GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val); |
| 522 | if (!GA) |
| 523 | return false; |
| 524 | |
| 525 | const GlobalVariable *GV = dyn_cast_or_null<GlobalVariable>(Val: GA->getGlobal()); |
| 526 | if (!GV) |
| 527 | return false; |
| 528 | |
| 529 | if (!GV->hasAttribute(Kind: "toc-data" )) |
| 530 | return false; |
| 531 | return true; |
| 532 | } |
| 533 | |
| 534 | static CodeModel::Model getCodeModel(const PPCSubtarget &Subtarget, |
| 535 | const TargetMachine &TM, |
| 536 | const SDNode *Node) { |
| 537 | // If there isn't an attribute to override the module code model |
| 538 | // this will be the effective code model. |
| 539 | CodeModel::Model ModuleModel = TM.getCodeModel(); |
| 540 | |
| 541 | GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val: Node->getOperand(Num: 0)); |
| 542 | if (!GA) |
| 543 | return ModuleModel; |
| 544 | |
| 545 | const GlobalValue *GV = GA->getGlobal(); |
| 546 | if (!GV) |
| 547 | return ModuleModel; |
| 548 | |
| 549 | return Subtarget.getCodeModel(TM, GV); |
| 550 | } |
| 551 | |
| 552 | /// isInt32Immediate - This method tests to see if the node is a 32-bit constant |
| 553 | /// operand. If so Imm will receive the 32-bit value. |
| 554 | static bool isInt32Immediate(SDNode *N, unsigned &Imm) { |
| 555 | if (N->getOpcode() == ISD::Constant && N->getValueType(ResNo: 0) == MVT::i32) { |
| 556 | Imm = N->getAsZExtVal(); |
| 557 | return true; |
| 558 | } |
| 559 | return false; |
| 560 | } |
| 561 | |
| 562 | /// isInt64Immediate - This method tests to see if the node is a 64-bit constant |
| 563 | /// operand. If so Imm will receive the 64-bit value. |
| 564 | static bool isInt64Immediate(SDNode *N, uint64_t &Imm) { |
| 565 | if (N->getOpcode() == ISD::Constant && N->getValueType(ResNo: 0) == MVT::i64) { |
| 566 | Imm = N->getAsZExtVal(); |
| 567 | return true; |
| 568 | } |
| 569 | return false; |
| 570 | } |
| 571 | |
| 572 | // isInt32Immediate - This method tests to see if a constant operand. |
| 573 | // If so Imm will receive the 32 bit value. |
| 574 | static bool isInt32Immediate(SDValue N, unsigned &Imm) { |
| 575 | return isInt32Immediate(N: N.getNode(), Imm); |
| 576 | } |
| 577 | |
| 578 | /// isInt64Immediate - This method tests to see if the value is a 64-bit |
| 579 | /// constant operand. If so Imm will receive the 64-bit value. |
| 580 | static bool isInt64Immediate(SDValue N, uint64_t &Imm) { |
| 581 | return isInt64Immediate(N: N.getNode(), Imm); |
| 582 | } |
| 583 | |
| 584 | static unsigned getBranchHint(unsigned PCC, |
| 585 | const FunctionLoweringInfo &FuncInfo, |
| 586 | const SDValue &DestMBB) { |
| 587 | assert(isa<BasicBlockSDNode>(DestMBB)); |
| 588 | |
| 589 | if (!FuncInfo.BPI) return PPC::BR_NO_HINT; |
| 590 | |
| 591 | const BasicBlock *BB = FuncInfo.MBB->getBasicBlock(); |
| 592 | const Instruction *BBTerm = BB->getTerminator(); |
| 593 | |
| 594 | if (BBTerm->getNumSuccessors() != 2) return PPC::BR_NO_HINT; |
| 595 | |
| 596 | const BasicBlock *TBB = BBTerm->getSuccessor(Idx: 0); |
| 597 | const BasicBlock *FBB = BBTerm->getSuccessor(Idx: 1); |
| 598 | |
| 599 | auto TProb = FuncInfo.BPI->getEdgeProbability(Src: BB, Dst: TBB); |
| 600 | auto FProb = FuncInfo.BPI->getEdgeProbability(Src: BB, Dst: FBB); |
| 601 | |
| 602 | // We only want to handle cases which are easy to predict at static time, e.g. |
| 603 | // C++ throw statement, that is very likely not taken, or calling never |
| 604 | // returned function, e.g. stdlib exit(). So we set Threshold to filter |
| 605 | // unwanted cases. |
| 606 | // |
| 607 | // Below is LLVM branch weight table, we only want to handle case 1, 2 |
| 608 | // |
| 609 | // Case Taken:Nontaken Example |
| 610 | // 1. Unreachable 1048575:1 C++ throw, stdlib exit(), |
| 611 | // 2. Invoke-terminating 1:1048575 |
| 612 | // 3. Coldblock 4:64 __builtin_expect |
| 613 | // 4. Loop Branch 124:4 For loop |
| 614 | // 5. PH/ZH/FPH 20:12 |
| 615 | const uint32_t Threshold = 10000; |
| 616 | |
| 617 | if (std::max(a: TProb, b: FProb) / Threshold < std::min(a: TProb, b: FProb)) |
| 618 | return PPC::BR_NO_HINT; |
| 619 | |
| 620 | LLVM_DEBUG(dbgs() << "Use branch hint for '" << FuncInfo.Fn->getName() |
| 621 | << "::" << BB->getName() << "'\n" |
| 622 | << " -> " << TBB->getName() << ": " << TProb << "\n" |
| 623 | << " -> " << FBB->getName() << ": " << FProb << "\n" ); |
| 624 | |
| 625 | const BasicBlockSDNode *BBDN = cast<BasicBlockSDNode>(Val: DestMBB); |
| 626 | |
| 627 | // If Dest BasicBlock is False-BasicBlock (FBB), swap branch probabilities, |
| 628 | // because we want 'TProb' stands for 'branch probability' to Dest BasicBlock |
| 629 | if (BBDN->getBasicBlock()->getBasicBlock() != TBB) |
| 630 | std::swap(a&: TProb, b&: FProb); |
| 631 | |
| 632 | return (TProb > FProb) ? PPC::BR_TAKEN_HINT : PPC::BR_NONTAKEN_HINT; |
| 633 | } |
| 634 | |
| 635 | // isOpcWithIntImmediate - This method tests to see if the node is a specific |
| 636 | // opcode and that it has a immediate integer right operand. |
| 637 | // If so Imm will receive the 32 bit value. |
| 638 | static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) { |
| 639 | return N->getOpcode() == Opc |
| 640 | && isInt32Immediate(N: N->getOperand(Num: 1).getNode(), Imm); |
| 641 | } |
| 642 | |
| 643 | void PPCDAGToDAGISel::selectFrameIndex(SDNode *SN, SDNode *N, int64_t Offset) { |
| 644 | SDLoc dl(SN); |
| 645 | int FI = cast<FrameIndexSDNode>(Val: N)->getIndex(); |
| 646 | SDValue TFI = CurDAG->getTargetFrameIndex(FI, VT: N->getValueType(ResNo: 0)); |
| 647 | unsigned Opc = N->getValueType(ResNo: 0) == MVT::i32 ? PPC::ADDI : PPC::ADDI8; |
| 648 | if (SN->hasOneUse()) |
| 649 | CurDAG->SelectNodeTo(N: SN, MachineOpc: Opc, VT: N->getValueType(ResNo: 0), Op1: TFI, |
| 650 | Op2: getSmallIPtrImm(Imm: Offset, dl)); |
| 651 | else |
| 652 | ReplaceNode(F: SN, T: CurDAG->getMachineNode(Opcode: Opc, dl, VT: N->getValueType(ResNo: 0), Op1: TFI, |
| 653 | Op2: getSmallIPtrImm(Imm: Offset, dl))); |
| 654 | } |
| 655 | |
| 656 | bool PPCDAGToDAGISel::isRotateAndMask(SDNode *N, unsigned Mask, |
| 657 | bool isShiftMask, unsigned &SH, |
| 658 | unsigned &MB, unsigned &ME) { |
| 659 | // Don't even go down this path for i64, since different logic will be |
| 660 | // necessary for rldicl/rldicr/rldimi. |
| 661 | if (N->getValueType(ResNo: 0) != MVT::i32) |
| 662 | return false; |
| 663 | |
| 664 | unsigned Shift = 32; |
| 665 | unsigned Indeterminant = ~0; // bit mask marking indeterminant results |
| 666 | unsigned Opcode = N->getOpcode(); |
| 667 | if (N->getNumOperands() != 2 || |
| 668 | !isInt32Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Shift) || (Shift > 31)) |
| 669 | return false; |
| 670 | |
| 671 | if (Opcode == ISD::SHL) { |
| 672 | // apply shift left to mask if it comes first |
| 673 | if (isShiftMask) Mask = Mask << Shift; |
| 674 | // determine which bits are made indeterminant by shift |
| 675 | Indeterminant = ~(0xFFFFFFFFu << Shift); |
| 676 | } else if (Opcode == ISD::SRL) { |
| 677 | // apply shift right to mask if it comes first |
| 678 | if (isShiftMask) Mask = Mask >> Shift; |
| 679 | // determine which bits are made indeterminant by shift |
| 680 | Indeterminant = ~(0xFFFFFFFFu >> Shift); |
| 681 | // adjust for the left rotate |
| 682 | Shift = 32 - Shift; |
| 683 | } else if (Opcode == ISD::ROTL) { |
| 684 | Indeterminant = 0; |
| 685 | } else { |
| 686 | return false; |
| 687 | } |
| 688 | |
| 689 | // if the mask doesn't intersect any Indeterminant bits |
| 690 | if (Mask && !(Mask & Indeterminant)) { |
| 691 | SH = Shift & 31; |
| 692 | // make sure the mask is still a mask (wrap arounds may not be) |
| 693 | return isRunOfOnes(Val: Mask, MB, ME); |
| 694 | } |
| 695 | return false; |
| 696 | } |
| 697 | |
| 698 | // isThreadPointerAcquisitionNode - Check if the operands of an ADD_TLS |
| 699 | // instruction use the thread pointer. |
| 700 | static bool isThreadPointerAcquisitionNode(SDValue Base, SelectionDAG *CurDAG) { |
| 701 | assert( |
| 702 | Base.getOpcode() == PPCISD::ADD_TLS && |
| 703 | "Only expecting the ADD_TLS instruction to acquire the thread pointer!" ); |
| 704 | const PPCSubtarget &Subtarget = |
| 705 | CurDAG->getMachineFunction().getSubtarget<PPCSubtarget>(); |
| 706 | SDValue ADDTLSOp1 = Base.getOperand(i: 0); |
| 707 | unsigned ADDTLSOp1Opcode = ADDTLSOp1.getOpcode(); |
| 708 | |
| 709 | // Account for when ADD_TLS is used for the initial-exec TLS model on Linux. |
| 710 | // |
| 711 | // Although ADD_TLS does not explicitly use the thread pointer |
| 712 | // register when LD_GOT_TPREL_L is one of it's operands, the LD_GOT_TPREL_L |
| 713 | // instruction will have a relocation specifier, @got@tprel, that is used to |
| 714 | // generate a GOT entry. The linker replaces this entry with an offset for a |
| 715 | // for a thread local variable, which will be relative to the thread pointer. |
| 716 | if (ADDTLSOp1Opcode == PPCISD::LD_GOT_TPREL_L) |
| 717 | return true; |
| 718 | // When using PC-Relative instructions for initial-exec, a MAT_PCREL_ADDR |
| 719 | // node is produced instead to represent the aforementioned situation. |
| 720 | LoadSDNode *LD = dyn_cast<LoadSDNode>(Val&: ADDTLSOp1); |
| 721 | if (LD && LD->getBasePtr().getOpcode() == PPCISD::MAT_PCREL_ADDR) |
| 722 | return true; |
| 723 | |
| 724 | // A GET_TPOINTER PPCISD node (only produced on AIX 32-bit mode) as an operand |
| 725 | // to ADD_TLS represents a call to .__get_tpointer to get the thread pointer, |
| 726 | // later returning it into R3. |
| 727 | if (ADDTLSOp1Opcode == PPCISD::GET_TPOINTER) |
| 728 | return true; |
| 729 | |
| 730 | // The ADD_TLS note is explicitly acquiring the thread pointer (X13/R13). |
| 731 | RegisterSDNode *AddFirstOpReg = |
| 732 | dyn_cast_or_null<RegisterSDNode>(Val: ADDTLSOp1.getNode()); |
| 733 | if (AddFirstOpReg && |
| 734 | AddFirstOpReg->getReg() == Subtarget.getThreadPointerRegister()) |
| 735 | return true; |
| 736 | |
| 737 | return false; |
| 738 | } |
| 739 | |
| 740 | // canOptimizeTLSDFormToXForm - Optimize TLS accesses when an ADD_TLS |
| 741 | // instruction is present. An ADD_TLS instruction, followed by a D-Form memory |
| 742 | // operation, can be optimized to use an X-Form load or store, allowing the |
| 743 | // ADD_TLS node to be removed completely. |
| 744 | static bool canOptimizeTLSDFormToXForm(SelectionDAG *CurDAG, SDValue Base) { |
| 745 | |
| 746 | // Do not do this transformation at -O0. |
| 747 | if (CurDAG->getTarget().getOptLevel() == CodeGenOptLevel::None) |
| 748 | return false; |
| 749 | |
| 750 | // In order to perform this optimization inside tryTLSXForm[Load|Store], |
| 751 | // Base is expected to be an ADD_TLS node. |
| 752 | if (Base.getOpcode() != PPCISD::ADD_TLS) |
| 753 | return false; |
| 754 | for (auto *ADDTLSUse : Base.getNode()->users()) { |
| 755 | // The optimization to convert the D-Form load/store into its X-Form |
| 756 | // counterpart should only occur if the source value offset of the load/ |
| 757 | // store is 0. This also means that The offset should always be undefined. |
| 758 | if (LoadSDNode *LD = dyn_cast<LoadSDNode>(Val: ADDTLSUse)) { |
| 759 | if (LD->getSrcValueOffset() != 0 || !LD->getOffset().isUndef()) |
| 760 | return false; |
| 761 | } else if (StoreSDNode *ST = dyn_cast<StoreSDNode>(Val: ADDTLSUse)) { |
| 762 | if (ST->getSrcValueOffset() != 0 || !ST->getOffset().isUndef()) |
| 763 | return false; |
| 764 | } else // Don't optimize if there are ADD_TLS users that aren't load/stores. |
| 765 | return false; |
| 766 | } |
| 767 | |
| 768 | if (Base.getOperand(i: 1).getOpcode() == PPCISD::TLS_LOCAL_EXEC_MAT_ADDR) |
| 769 | return false; |
| 770 | |
| 771 | // Does the ADD_TLS node of the load/store use the thread pointer? |
| 772 | // If the thread pointer is not used as one of the operands of ADD_TLS, |
| 773 | // then this optimization is not valid. |
| 774 | return isThreadPointerAcquisitionNode(Base, CurDAG); |
| 775 | } |
| 776 | |
| 777 | bool PPCDAGToDAGISel::tryTLSXFormStore(StoreSDNode *ST) { |
| 778 | SDValue Base = ST->getBasePtr(); |
| 779 | if (!canOptimizeTLSDFormToXForm(CurDAG, Base)) |
| 780 | return false; |
| 781 | |
| 782 | SDLoc dl(ST); |
| 783 | EVT MemVT = ST->getMemoryVT(); |
| 784 | EVT RegVT = ST->getValue().getValueType(); |
| 785 | |
| 786 | unsigned Opcode; |
| 787 | switch (MemVT.getSimpleVT().SimpleTy) { |
| 788 | default: |
| 789 | return false; |
| 790 | case MVT::i8: { |
| 791 | Opcode = (RegVT == MVT::i32) ? PPC::STBXTLS_32 : PPC::STBXTLS; |
| 792 | break; |
| 793 | } |
| 794 | case MVT::i16: { |
| 795 | Opcode = (RegVT == MVT::i32) ? PPC::STHXTLS_32 : PPC::STHXTLS; |
| 796 | break; |
| 797 | } |
| 798 | case MVT::i32: { |
| 799 | Opcode = (RegVT == MVT::i32) ? PPC::STWXTLS_32 : PPC::STWXTLS; |
| 800 | break; |
| 801 | } |
| 802 | case MVT::i64: { |
| 803 | Opcode = PPC::STDXTLS; |
| 804 | break; |
| 805 | } |
| 806 | case MVT::f32: { |
| 807 | Opcode = PPC::STFSXTLS; |
| 808 | break; |
| 809 | } |
| 810 | case MVT::f64: { |
| 811 | Opcode = PPC::STFDXTLS; |
| 812 | break; |
| 813 | } |
| 814 | } |
| 815 | SDValue Chain = ST->getChain(); |
| 816 | SDVTList VTs = ST->getVTList(); |
| 817 | SDValue Ops[] = {ST->getValue(), Base.getOperand(i: 0), Base.getOperand(i: 1), |
| 818 | Chain}; |
| 819 | SDNode *MN = CurDAG->getMachineNode(Opcode, dl, VTs, Ops); |
| 820 | transferMemOperands(N: ST, Result: MN); |
| 821 | ReplaceNode(F: ST, T: MN); |
| 822 | return true; |
| 823 | } |
| 824 | |
| 825 | bool PPCDAGToDAGISel::tryTLSXFormLoad(LoadSDNode *LD) { |
| 826 | SDValue Base = LD->getBasePtr(); |
| 827 | if (!canOptimizeTLSDFormToXForm(CurDAG, Base)) |
| 828 | return false; |
| 829 | |
| 830 | SDLoc dl(LD); |
| 831 | EVT MemVT = LD->getMemoryVT(); |
| 832 | EVT RegVT = LD->getValueType(ResNo: 0); |
| 833 | bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; |
| 834 | unsigned Opcode; |
| 835 | switch (MemVT.getSimpleVT().SimpleTy) { |
| 836 | default: |
| 837 | return false; |
| 838 | case MVT::i8: { |
| 839 | Opcode = (RegVT == MVT::i32) ? PPC::LBZXTLS_32 : PPC::LBZXTLS; |
| 840 | break; |
| 841 | } |
| 842 | case MVT::i16: { |
| 843 | if (RegVT == MVT::i32) |
| 844 | Opcode = isSExt ? PPC::LHAXTLS_32 : PPC::LHZXTLS_32; |
| 845 | else |
| 846 | Opcode = isSExt ? PPC::LHAXTLS : PPC::LHZXTLS; |
| 847 | break; |
| 848 | } |
| 849 | case MVT::i32: { |
| 850 | if (RegVT == MVT::i32) |
| 851 | Opcode = isSExt ? PPC::LWAXTLS_32 : PPC::LWZXTLS_32; |
| 852 | else |
| 853 | Opcode = isSExt ? PPC::LWAXTLS : PPC::LWZXTLS; |
| 854 | break; |
| 855 | } |
| 856 | case MVT::i64: { |
| 857 | Opcode = PPC::LDXTLS; |
| 858 | break; |
| 859 | } |
| 860 | case MVT::f32: { |
| 861 | Opcode = PPC::LFSXTLS; |
| 862 | break; |
| 863 | } |
| 864 | case MVT::f64: { |
| 865 | Opcode = PPC::LFDXTLS; |
| 866 | break; |
| 867 | } |
| 868 | } |
| 869 | SDValue Chain = LD->getChain(); |
| 870 | SDVTList VTs = LD->getVTList(); |
| 871 | SDValue Ops[] = {Base.getOperand(i: 0), Base.getOperand(i: 1), Chain}; |
| 872 | SDNode *MN = CurDAG->getMachineNode(Opcode, dl, VTs, Ops); |
| 873 | transferMemOperands(N: LD, Result: MN); |
| 874 | ReplaceNode(F: LD, T: MN); |
| 875 | return true; |
| 876 | } |
| 877 | |
| 878 | /// Turn an or of two masked values into the rotate left word immediate then |
| 879 | /// mask insert (rlwimi) instruction. |
| 880 | bool PPCDAGToDAGISel::tryBitfieldInsert(SDNode *N) { |
| 881 | SDValue Op0 = N->getOperand(Num: 0); |
| 882 | SDValue Op1 = N->getOperand(Num: 1); |
| 883 | SDLoc dl(N); |
| 884 | |
| 885 | KnownBits LKnown = CurDAG->computeKnownBits(Op: Op0); |
| 886 | KnownBits RKnown = CurDAG->computeKnownBits(Op: Op1); |
| 887 | |
| 888 | unsigned TargetMask = LKnown.Zero.getZExtValue(); |
| 889 | unsigned InsertMask = RKnown.Zero.getZExtValue(); |
| 890 | |
| 891 | if ((TargetMask | InsertMask) == 0xFFFFFFFF) { |
| 892 | unsigned Op0Opc = Op0.getOpcode(); |
| 893 | unsigned Op1Opc = Op1.getOpcode(); |
| 894 | unsigned Value, SH = 0; |
| 895 | TargetMask = ~TargetMask; |
| 896 | InsertMask = ~InsertMask; |
| 897 | |
| 898 | // If the LHS has a foldable shift and the RHS does not, then swap it to the |
| 899 | // RHS so that we can fold the shift into the insert. |
| 900 | if (Op0Opc == ISD::AND && Op1Opc == ISD::AND) { |
| 901 | if (Op0.getOperand(i: 0).getOpcode() == ISD::SHL || |
| 902 | Op0.getOperand(i: 0).getOpcode() == ISD::SRL) { |
| 903 | if (Op1.getOperand(i: 0).getOpcode() != ISD::SHL && |
| 904 | Op1.getOperand(i: 0).getOpcode() != ISD::SRL) { |
| 905 | std::swap(a&: Op0, b&: Op1); |
| 906 | std::swap(a&: Op0Opc, b&: Op1Opc); |
| 907 | std::swap(a&: TargetMask, b&: InsertMask); |
| 908 | } |
| 909 | } |
| 910 | } else if (Op0Opc == ISD::SHL || Op0Opc == ISD::SRL) { |
| 911 | if (Op1Opc == ISD::AND && Op1.getOperand(i: 0).getOpcode() != ISD::SHL && |
| 912 | Op1.getOperand(i: 0).getOpcode() != ISD::SRL) { |
| 913 | std::swap(a&: Op0, b&: Op1); |
| 914 | std::swap(a&: Op0Opc, b&: Op1Opc); |
| 915 | std::swap(a&: TargetMask, b&: InsertMask); |
| 916 | } |
| 917 | } |
| 918 | |
| 919 | unsigned MB, ME; |
| 920 | if (isRunOfOnes(Val: InsertMask, MB, ME)) { |
| 921 | if ((Op1Opc == ISD::SHL || Op1Opc == ISD::SRL) && |
| 922 | isInt32Immediate(N: Op1.getOperand(i: 1), Imm&: Value)) { |
| 923 | Op1 = Op1.getOperand(i: 0); |
| 924 | SH = (Op1Opc == ISD::SHL) ? Value : 32 - Value; |
| 925 | } |
| 926 | if (Op1Opc == ISD::AND) { |
| 927 | // The AND mask might not be a constant, and we need to make sure that |
| 928 | // if we're going to fold the masking with the insert, all bits not |
| 929 | // know to be zero in the mask are known to be one. |
| 930 | KnownBits MKnown = CurDAG->computeKnownBits(Op: Op1.getOperand(i: 1)); |
| 931 | bool CanFoldMask = InsertMask == MKnown.One.getZExtValue(); |
| 932 | |
| 933 | unsigned SHOpc = Op1.getOperand(i: 0).getOpcode(); |
| 934 | if ((SHOpc == ISD::SHL || SHOpc == ISD::SRL) && CanFoldMask && |
| 935 | isInt32Immediate(N: Op1.getOperand(i: 0).getOperand(i: 1), Imm&: Value)) { |
| 936 | // Note that Value must be in range here (less than 32) because |
| 937 | // otherwise there would not be any bits set in InsertMask. |
| 938 | Op1 = Op1.getOperand(i: 0).getOperand(i: 0); |
| 939 | SH = (SHOpc == ISD::SHL) ? Value : 32 - Value; |
| 940 | } |
| 941 | } |
| 942 | |
| 943 | SH &= 31; |
| 944 | SDValue Ops[] = { Op0, Op1, getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl), |
| 945 | getI32Imm(Imm: ME, dl) }; |
| 946 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: PPC::RLWIMI, dl, VT: MVT::i32, Ops)); |
| 947 | return true; |
| 948 | } |
| 949 | } |
| 950 | return false; |
| 951 | } |
| 952 | |
| 953 | static unsigned allUsesTruncate(SelectionDAG *CurDAG, SDNode *N) { |
| 954 | unsigned MaxTruncation = 0; |
| 955 | // Cannot use range-based for loop here as we need the actual use (i.e. we |
| 956 | // need the operand number corresponding to the use). A range-based for |
| 957 | // will unbox the use and provide an SDNode*. |
| 958 | for (SDUse &Use : N->uses()) { |
| 959 | SDNode *User = Use.getUser(); |
| 960 | unsigned Opc = |
| 961 | User->isMachineOpcode() ? User->getMachineOpcode() : User->getOpcode(); |
| 962 | switch (Opc) { |
| 963 | default: return 0; |
| 964 | case ISD::TRUNCATE: |
| 965 | if (User->isMachineOpcode()) |
| 966 | return 0; |
| 967 | MaxTruncation = std::max(a: MaxTruncation, |
| 968 | b: (unsigned)User->getValueType(ResNo: 0).getSizeInBits()); |
| 969 | continue; |
| 970 | case ISD::STORE: { |
| 971 | if (User->isMachineOpcode()) |
| 972 | return 0; |
| 973 | StoreSDNode *STN = cast<StoreSDNode>(Val: User); |
| 974 | unsigned MemVTSize = STN->getMemoryVT().getSizeInBits(); |
| 975 | if (MemVTSize == 64 || Use.getOperandNo() != 0) |
| 976 | return 0; |
| 977 | MaxTruncation = std::max(a: MaxTruncation, b: MemVTSize); |
| 978 | continue; |
| 979 | } |
| 980 | case PPC::STW8: |
| 981 | case PPC::STWX8: |
| 982 | case PPC::STWU8: |
| 983 | case PPC::STWUX8: |
| 984 | if (Use.getOperandNo() != 0) |
| 985 | return 0; |
| 986 | MaxTruncation = std::max(a: MaxTruncation, b: 32u); |
| 987 | continue; |
| 988 | case PPC::STH8: |
| 989 | case PPC::STHX8: |
| 990 | case PPC::STHU8: |
| 991 | case PPC::STHUX8: |
| 992 | if (Use.getOperandNo() != 0) |
| 993 | return 0; |
| 994 | MaxTruncation = std::max(a: MaxTruncation, b: 16u); |
| 995 | continue; |
| 996 | case PPC::STB8: |
| 997 | case PPC::STBX8: |
| 998 | case PPC::STBU8: |
| 999 | case PPC::STBUX8: |
| 1000 | if (Use.getOperandNo() != 0) |
| 1001 | return 0; |
| 1002 | MaxTruncation = std::max(a: MaxTruncation, b: 8u); |
| 1003 | continue; |
| 1004 | } |
| 1005 | } |
| 1006 | return MaxTruncation; |
| 1007 | } |
| 1008 | |
| 1009 | // For any 32 < Num < 64, check if the Imm contains at least Num consecutive |
| 1010 | // zeros and return the number of bits by the left of these consecutive zeros. |
| 1011 | static int findContiguousZerosAtLeast(uint64_t Imm, unsigned Num) { |
| 1012 | unsigned HiTZ = llvm::countr_zero<uint32_t>(Val: Hi_32(Value: Imm)); |
| 1013 | unsigned LoLZ = llvm::countl_zero<uint32_t>(Val: Lo_32(Value: Imm)); |
| 1014 | if ((HiTZ + LoLZ) >= Num) |
| 1015 | return (32 + HiTZ); |
| 1016 | return 0; |
| 1017 | } |
| 1018 | |
| 1019 | // Direct materialization of 64-bit constants by enumerated patterns. |
| 1020 | static SDNode *selectI64ImmDirect(SelectionDAG *CurDAG, const SDLoc &dl, |
| 1021 | uint64_t Imm, unsigned &InstCnt) { |
| 1022 | unsigned TZ = llvm::countr_zero<uint64_t>(Val: Imm); |
| 1023 | unsigned LZ = llvm::countl_zero<uint64_t>(Val: Imm); |
| 1024 | unsigned TO = llvm::countr_one<uint64_t>(Value: Imm); |
| 1025 | unsigned LO = llvm::countl_one<uint64_t>(Value: Imm); |
| 1026 | unsigned Hi32 = Hi_32(Value: Imm); |
| 1027 | unsigned Lo32 = Lo_32(Value: Imm); |
| 1028 | SDNode *Result = nullptr; |
| 1029 | unsigned Shift = 0; |
| 1030 | |
| 1031 | auto getI32Imm = [CurDAG, dl](unsigned Imm) { |
| 1032 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
| 1033 | }; |
| 1034 | |
| 1035 | // Following patterns use 1 instructions to materialize the Imm. |
| 1036 | InstCnt = 1; |
| 1037 | // 1-1) Patterns : {zeros}{15-bit valve} |
| 1038 | // {ones}{15-bit valve} |
| 1039 | if (isInt<16>(x: Imm)) { |
| 1040 | SDValue SDImm = CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i64); |
| 1041 | return CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, Op1: SDImm); |
| 1042 | } |
| 1043 | // 1-2) Patterns : {zeros}{15-bit valve}{16 zeros} |
| 1044 | // {ones}{15-bit valve}{16 zeros} |
| 1045 | if (TZ > 15 && (LZ > 32 || LO > 32)) |
| 1046 | return CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, |
| 1047 | Op1: getI32Imm((Imm >> 16) & 0xffff)); |
| 1048 | |
| 1049 | // Following patterns use 2 instructions to materialize the Imm. |
| 1050 | InstCnt = 2; |
| 1051 | assert(LZ < 64 && "Unexpected leading zeros here." ); |
| 1052 | // Count of ones follwing the leading zeros. |
| 1053 | unsigned FO = llvm::countl_one<uint64_t>(Value: Imm << LZ); |
| 1054 | // 2-1) Patterns : {zeros}{31-bit value} |
| 1055 | // {ones}{31-bit value} |
| 1056 | if (isInt<32>(x: Imm)) { |
| 1057 | uint64_t ImmHi16 = (Imm >> 16) & 0xffff; |
| 1058 | unsigned Opcode = ImmHi16 ? PPC::LIS8 : PPC::LI8; |
| 1059 | Result = CurDAG->getMachineNode(Opcode, dl, VT: MVT::i64, Op1: getI32Imm(ImmHi16)); |
| 1060 | return CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1061 | Op2: getI32Imm(Imm & 0xffff)); |
| 1062 | } |
| 1063 | // 2-2) Patterns : {zeros}{ones}{15-bit value}{zeros} |
| 1064 | // {zeros}{15-bit value}{zeros} |
| 1065 | // {zeros}{ones}{15-bit value} |
| 1066 | // {ones}{15-bit value}{zeros} |
| 1067 | // We can take advantage of LI's sign-extension semantics to generate leading |
| 1068 | // ones, and then use RLDIC to mask off the ones in both sides after rotation. |
| 1069 | if ((LZ + FO + TZ) > 48) { |
| 1070 | Result = CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, |
| 1071 | Op1: getI32Imm((Imm >> TZ) & 0xffff)); |
| 1072 | return CurDAG->getMachineNode(Opcode: PPC::RLDIC, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1073 | Op2: getI32Imm(TZ), Op3: getI32Imm(LZ)); |
| 1074 | } |
| 1075 | // 2-3) Pattern : {zeros}{15-bit value}{ones} |
| 1076 | // Shift right the Imm by (48 - LZ) bits to construct a negtive 16 bits value, |
| 1077 | // therefore we can take advantage of LI's sign-extension semantics, and then |
| 1078 | // mask them off after rotation. |
| 1079 | // |
| 1080 | // +--LZ--||-15-bit-||--TO--+ +-------------|--16-bit--+ |
| 1081 | // |00000001bbbbbbbbb1111111| -> |00000000000001bbbbbbbbb1| |
| 1082 | // +------------------------+ +------------------------+ |
| 1083 | // 63 0 63 0 |
| 1084 | // Imm (Imm >> (48 - LZ) & 0xffff) |
| 1085 | // +----sext-----|--16-bit--+ +clear-|-----------------+ |
| 1086 | // |11111111111111bbbbbbbbb1| -> |00000001bbbbbbbbb1111111| |
| 1087 | // +------------------------+ +------------------------+ |
| 1088 | // 63 0 63 0 |
| 1089 | // LI8: sext many leading zeros RLDICL: rotate left (48 - LZ), clear left LZ |
| 1090 | if ((LZ + TO) > 48) { |
| 1091 | // Since the immediates with (LZ > 32) have been handled by previous |
| 1092 | // patterns, here we have (LZ <= 32) to make sure we will not shift right |
| 1093 | // the Imm by a negative value. |
| 1094 | assert(LZ <= 32 && "Unexpected shift value." ); |
| 1095 | Result = CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, |
| 1096 | Op1: getI32Imm((Imm >> (48 - LZ) & 0xffff))); |
| 1097 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1098 | Op2: getI32Imm(48 - LZ), Op3: getI32Imm(LZ)); |
| 1099 | } |
| 1100 | // 2-4) Patterns : {zeros}{ones}{15-bit value}{ones} |
| 1101 | // {ones}{15-bit value}{ones} |
| 1102 | // We can take advantage of LI's sign-extension semantics to generate leading |
| 1103 | // ones, and then use RLDICL to mask off the ones in left sides (if required) |
| 1104 | // after rotation. |
| 1105 | // |
| 1106 | // +-LZ-FO||-15-bit-||--TO--+ +-------------|--16-bit--+ |
| 1107 | // |00011110bbbbbbbbb1111111| -> |000000000011110bbbbbbbbb| |
| 1108 | // +------------------------+ +------------------------+ |
| 1109 | // 63 0 63 0 |
| 1110 | // Imm (Imm >> TO) & 0xffff |
| 1111 | // +----sext-----|--16-bit--+ +LZ|---------------------+ |
| 1112 | // |111111111111110bbbbbbbbb| -> |00011110bbbbbbbbb1111111| |
| 1113 | // +------------------------+ +------------------------+ |
| 1114 | // 63 0 63 0 |
| 1115 | // LI8: sext many leading zeros RLDICL: rotate left TO, clear left LZ |
| 1116 | if ((LZ + FO + TO) > 48) { |
| 1117 | Result = CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, |
| 1118 | Op1: getI32Imm((Imm >> TO) & 0xffff)); |
| 1119 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1120 | Op2: getI32Imm(TO), Op3: getI32Imm(LZ)); |
| 1121 | } |
| 1122 | // 2-5) Pattern : {32 zeros}{****}{0}{15-bit value} |
| 1123 | // If Hi32 is zero and the Lo16(in Lo32) can be presented as a positive 16 bit |
| 1124 | // value, we can use LI for Lo16 without generating leading ones then add the |
| 1125 | // Hi16(in Lo32). |
| 1126 | if (LZ == 32 && ((Lo32 & 0x8000) == 0)) { |
| 1127 | Result = CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, |
| 1128 | Op1: getI32Imm(Lo32 & 0xffff)); |
| 1129 | return CurDAG->getMachineNode(Opcode: PPC::ORIS8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1130 | Op2: getI32Imm(Lo32 >> 16)); |
| 1131 | } |
| 1132 | // 2-6) Patterns : {******}{49 zeros}{******} |
| 1133 | // {******}{49 ones}{******} |
| 1134 | // If the Imm contains 49 consecutive zeros/ones, it means that a total of 15 |
| 1135 | // bits remain on both sides. Rotate right the Imm to construct an int<16> |
| 1136 | // value, use LI for int<16> value and then use RLDICL without mask to rotate |
| 1137 | // it back. |
| 1138 | // |
| 1139 | // 1) findContiguousZerosAtLeast(Imm, 49) |
| 1140 | // +------|--zeros-|------+ +---ones--||---15 bit--+ |
| 1141 | // |bbbbbb0000000000aaaaaa| -> |0000000000aaaaaabbbbbb| |
| 1142 | // +----------------------+ +----------------------+ |
| 1143 | // 63 0 63 0 |
| 1144 | // |
| 1145 | // 2) findContiguousZerosAtLeast(~Imm, 49) |
| 1146 | // +------|--ones--|------+ +---ones--||---15 bit--+ |
| 1147 | // |bbbbbb1111111111aaaaaa| -> |1111111111aaaaaabbbbbb| |
| 1148 | // +----------------------+ +----------------------+ |
| 1149 | // 63 0 63 0 |
| 1150 | if ((Shift = findContiguousZerosAtLeast(Imm, Num: 49)) || |
| 1151 | (Shift = findContiguousZerosAtLeast(Imm: ~Imm, Num: 49))) { |
| 1152 | uint64_t RotImm = APInt(64, Imm).rotr(rotateAmt: Shift).getZExtValue(); |
| 1153 | Result = CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, |
| 1154 | Op1: getI32Imm(RotImm & 0xffff)); |
| 1155 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1156 | Op2: getI32Imm(Shift), Op3: getI32Imm(0)); |
| 1157 | } |
| 1158 | // 2-7) Patterns : High word == Low word |
| 1159 | // This may require 2 to 3 instructions, depending on whether Lo32 can be |
| 1160 | // materialized in 1 instruction. |
| 1161 | if (Hi32 == Lo32) { |
| 1162 | // Handle the first 32 bits. |
| 1163 | uint64_t ImmHi16 = (Lo32 >> 16) & 0xffff; |
| 1164 | uint64_t ImmLo16 = Lo32 & 0xffff; |
| 1165 | if (isInt<16>(x: Lo32)) |
| 1166 | Result = |
| 1167 | CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, Op1: getI32Imm(ImmLo16)); |
| 1168 | else if (!ImmLo16) |
| 1169 | Result = |
| 1170 | CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, Op1: getI32Imm(ImmHi16)); |
| 1171 | else { |
| 1172 | InstCnt = 3; |
| 1173 | Result = |
| 1174 | CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, Op1: getI32Imm(ImmHi16)); |
| 1175 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, |
| 1176 | Op1: SDValue(Result, 0), Op2: getI32Imm(ImmLo16)); |
| 1177 | } |
| 1178 | // Use rldimi to insert the Low word into High word. |
| 1179 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(32), |
| 1180 | getI32Imm(0)}; |
| 1181 | return CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops); |
| 1182 | } |
| 1183 | |
| 1184 | // Following patterns use 3 instructions to materialize the Imm. |
| 1185 | InstCnt = 3; |
| 1186 | // 3-1) Patterns : {zeros}{ones}{31-bit value}{zeros} |
| 1187 | // {zeros}{31-bit value}{zeros} |
| 1188 | // {zeros}{ones}{31-bit value} |
| 1189 | // {ones}{31-bit value}{zeros} |
| 1190 | // We can take advantage of LIS's sign-extension semantics to generate leading |
| 1191 | // ones, add the remaining bits with ORI, and then use RLDIC to mask off the |
| 1192 | // ones in both sides after rotation. |
| 1193 | if ((LZ + FO + TZ) > 32) { |
| 1194 | uint64_t ImmHi16 = (Imm >> (TZ + 16)) & 0xffff; |
| 1195 | unsigned Opcode = ImmHi16 ? PPC::LIS8 : PPC::LI8; |
| 1196 | Result = CurDAG->getMachineNode(Opcode, dl, VT: MVT::i64, Op1: getI32Imm(ImmHi16)); |
| 1197 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1198 | Op2: getI32Imm((Imm >> TZ) & 0xffff)); |
| 1199 | return CurDAG->getMachineNode(Opcode: PPC::RLDIC, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1200 | Op2: getI32Imm(TZ), Op3: getI32Imm(LZ)); |
| 1201 | } |
| 1202 | // 3-2) Pattern : {zeros}{31-bit value}{ones} |
| 1203 | // Shift right the Imm by (32 - LZ) bits to construct a negative 32 bits |
| 1204 | // value, therefore we can take advantage of LIS's sign-extension semantics, |
| 1205 | // add the remaining bits with ORI, and then mask them off after rotation. |
| 1206 | // This is similar to Pattern 2-3, please refer to the diagram there. |
| 1207 | if ((LZ + TO) > 32) { |
| 1208 | // Since the immediates with (LZ > 32) have been handled by previous |
| 1209 | // patterns, here we have (LZ <= 32) to make sure we will not shift right |
| 1210 | // the Imm by a negative value. |
| 1211 | assert(LZ <= 32 && "Unexpected shift value." ); |
| 1212 | Result = CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, |
| 1213 | Op1: getI32Imm((Imm >> (48 - LZ)) & 0xffff)); |
| 1214 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1215 | Op2: getI32Imm((Imm >> (32 - LZ)) & 0xffff)); |
| 1216 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1217 | Op2: getI32Imm(32 - LZ), Op3: getI32Imm(LZ)); |
| 1218 | } |
| 1219 | // 3-3) Patterns : {zeros}{ones}{31-bit value}{ones} |
| 1220 | // {ones}{31-bit value}{ones} |
| 1221 | // We can take advantage of LIS's sign-extension semantics to generate leading |
| 1222 | // ones, add the remaining bits with ORI, and then use RLDICL to mask off the |
| 1223 | // ones in left sides (if required) after rotation. |
| 1224 | // This is similar to Pattern 2-4, please refer to the diagram there. |
| 1225 | if ((LZ + FO + TO) > 32) { |
| 1226 | Result = CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, |
| 1227 | Op1: getI32Imm((Imm >> (TO + 16)) & 0xffff)); |
| 1228 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1229 | Op2: getI32Imm((Imm >> TO) & 0xffff)); |
| 1230 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1231 | Op2: getI32Imm(TO), Op3: getI32Imm(LZ)); |
| 1232 | } |
| 1233 | // 3-4) Patterns : {******}{33 zeros}{******} |
| 1234 | // {******}{33 ones}{******} |
| 1235 | // If the Imm contains 33 consecutive zeros/ones, it means that a total of 31 |
| 1236 | // bits remain on both sides. Rotate right the Imm to construct an int<32> |
| 1237 | // value, use LIS + ORI for int<32> value and then use RLDICL without mask to |
| 1238 | // rotate it back. |
| 1239 | // This is similar to Pattern 2-6, please refer to the diagram there. |
| 1240 | if ((Shift = findContiguousZerosAtLeast(Imm, Num: 33)) || |
| 1241 | (Shift = findContiguousZerosAtLeast(Imm: ~Imm, Num: 33))) { |
| 1242 | uint64_t RotImm = APInt(64, Imm).rotr(rotateAmt: Shift).getZExtValue(); |
| 1243 | uint64_t ImmHi16 = (RotImm >> 16) & 0xffff; |
| 1244 | unsigned Opcode = ImmHi16 ? PPC::LIS8 : PPC::LI8; |
| 1245 | Result = CurDAG->getMachineNode(Opcode, dl, VT: MVT::i64, Op1: getI32Imm(ImmHi16)); |
| 1246 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1247 | Op2: getI32Imm(RotImm & 0xffff)); |
| 1248 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1249 | Op2: getI32Imm(Shift), Op3: getI32Imm(0)); |
| 1250 | } |
| 1251 | |
| 1252 | InstCnt = 0; |
| 1253 | return nullptr; |
| 1254 | } |
| 1255 | |
| 1256 | // Try to select instructions to generate a 64 bit immediate using prefix as |
| 1257 | // well as non prefix instructions. The function will return the SDNode |
| 1258 | // to materialize that constant or it will return nullptr if it does not |
| 1259 | // find one. The variable InstCnt is set to the number of instructions that |
| 1260 | // were selected. |
| 1261 | static SDNode *selectI64ImmDirectPrefix(SelectionDAG *CurDAG, const SDLoc &dl, |
| 1262 | uint64_t Imm, unsigned &InstCnt) { |
| 1263 | unsigned TZ = llvm::countr_zero<uint64_t>(Val: Imm); |
| 1264 | unsigned LZ = llvm::countl_zero<uint64_t>(Val: Imm); |
| 1265 | unsigned TO = llvm::countr_one<uint64_t>(Value: Imm); |
| 1266 | unsigned FO = llvm::countl_one<uint64_t>(Value: LZ == 64 ? 0 : (Imm << LZ)); |
| 1267 | unsigned Hi32 = Hi_32(Value: Imm); |
| 1268 | unsigned Lo32 = Lo_32(Value: Imm); |
| 1269 | |
| 1270 | auto getI32Imm = [CurDAG, dl](unsigned Imm) { |
| 1271 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
| 1272 | }; |
| 1273 | |
| 1274 | auto getI64Imm = [CurDAG, dl](uint64_t Imm) { |
| 1275 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i64); |
| 1276 | }; |
| 1277 | |
| 1278 | // Following patterns use 1 instruction to materialize Imm. |
| 1279 | InstCnt = 1; |
| 1280 | |
| 1281 | // The pli instruction can materialize up to 34 bits directly. |
| 1282 | // If a constant fits within 34-bits, emit the pli instruction here directly. |
| 1283 | if (isInt<34>(x: Imm)) |
| 1284 | return CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, |
| 1285 | Op1: CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i64)); |
| 1286 | |
| 1287 | // Require at least two instructions. |
| 1288 | InstCnt = 2; |
| 1289 | SDNode *Result = nullptr; |
| 1290 | // Patterns : {zeros}{ones}{33-bit value}{zeros} |
| 1291 | // {zeros}{33-bit value}{zeros} |
| 1292 | // {zeros}{ones}{33-bit value} |
| 1293 | // {ones}{33-bit value}{zeros} |
| 1294 | // We can take advantage of PLI's sign-extension semantics to generate leading |
| 1295 | // ones, and then use RLDIC to mask off the ones on both sides after rotation. |
| 1296 | if ((LZ + FO + TZ) > 30) { |
| 1297 | APInt SignedInt34 = APInt(34, (Imm >> TZ) & 0x3ffffffff); |
| 1298 | APInt Extended = SignedInt34.sext(width: 64); |
| 1299 | Result = CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, |
| 1300 | Op1: getI64Imm(Extended.getZExtValue())); |
| 1301 | return CurDAG->getMachineNode(Opcode: PPC::RLDIC, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1302 | Op2: getI32Imm(TZ), Op3: getI32Imm(LZ)); |
| 1303 | } |
| 1304 | // Pattern : {zeros}{33-bit value}{ones} |
| 1305 | // Shift right the Imm by (30 - LZ) bits to construct a negative 34 bit value, |
| 1306 | // therefore we can take advantage of PLI's sign-extension semantics, and then |
| 1307 | // mask them off after rotation. |
| 1308 | // |
| 1309 | // +--LZ--||-33-bit-||--TO--+ +-------------|--34-bit--+ |
| 1310 | // |00000001bbbbbbbbb1111111| -> |00000000000001bbbbbbbbb1| |
| 1311 | // +------------------------+ +------------------------+ |
| 1312 | // 63 0 63 0 |
| 1313 | // |
| 1314 | // +----sext-----|--34-bit--+ +clear-|-----------------+ |
| 1315 | // |11111111111111bbbbbbbbb1| -> |00000001bbbbbbbbb1111111| |
| 1316 | // +------------------------+ +------------------------+ |
| 1317 | // 63 0 63 0 |
| 1318 | if ((LZ + TO) > 30) { |
| 1319 | APInt SignedInt34 = APInt(34, (Imm >> (30 - LZ)) & 0x3ffffffff); |
| 1320 | APInt Extended = SignedInt34.sext(width: 64); |
| 1321 | Result = CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, |
| 1322 | Op1: getI64Imm(Extended.getZExtValue())); |
| 1323 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1324 | Op2: getI32Imm(30 - LZ), Op3: getI32Imm(LZ)); |
| 1325 | } |
| 1326 | // Patterns : {zeros}{ones}{33-bit value}{ones} |
| 1327 | // {ones}{33-bit value}{ones} |
| 1328 | // Similar to LI we can take advantage of PLI's sign-extension semantics to |
| 1329 | // generate leading ones, and then use RLDICL to mask off the ones in left |
| 1330 | // sides (if required) after rotation. |
| 1331 | if ((LZ + FO + TO) > 30) { |
| 1332 | APInt SignedInt34 = APInt(34, (Imm >> TO) & 0x3ffffffff); |
| 1333 | APInt Extended = SignedInt34.sext(width: 64); |
| 1334 | Result = CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, |
| 1335 | Op1: getI64Imm(Extended.getZExtValue())); |
| 1336 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1337 | Op2: getI32Imm(TO), Op3: getI32Imm(LZ)); |
| 1338 | } |
| 1339 | // Patterns : {******}{31 zeros}{******} |
| 1340 | // : {******}{31 ones}{******} |
| 1341 | // If Imm contains 31 consecutive zeros/ones then the remaining bit count |
| 1342 | // is 33. Rotate right the Imm to construct a int<33> value, we can use PLI |
| 1343 | // for the int<33> value and then use RLDICL without a mask to rotate it back. |
| 1344 | // |
| 1345 | // +------|--ones--|------+ +---ones--||---33 bit--+ |
| 1346 | // |bbbbbb1111111111aaaaaa| -> |1111111111aaaaaabbbbbb| |
| 1347 | // +----------------------+ +----------------------+ |
| 1348 | // 63 0 63 0 |
| 1349 | for (unsigned Shift = 0; Shift < 63; ++Shift) { |
| 1350 | uint64_t RotImm = APInt(64, Imm).rotr(rotateAmt: Shift).getZExtValue(); |
| 1351 | if (isInt<34>(x: RotImm)) { |
| 1352 | Result = |
| 1353 | CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, Op1: getI64Imm(RotImm)); |
| 1354 | return CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 1355 | Op1: SDValue(Result, 0), Op2: getI32Imm(Shift), |
| 1356 | Op3: getI32Imm(0)); |
| 1357 | } |
| 1358 | } |
| 1359 | |
| 1360 | // Patterns : High word == Low word |
| 1361 | // This is basically a splat of a 32 bit immediate. |
| 1362 | if (Hi32 == Lo32) { |
| 1363 | Result = CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, Op1: getI64Imm(Hi32)); |
| 1364 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(32), |
| 1365 | getI32Imm(0)}; |
| 1366 | return CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops); |
| 1367 | } |
| 1368 | |
| 1369 | InstCnt = 3; |
| 1370 | // Catch-all |
| 1371 | // This pattern can form any 64 bit immediate in 3 instructions. |
| 1372 | SDNode *ResultHi = |
| 1373 | CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, Op1: getI64Imm(Hi32)); |
| 1374 | SDNode *ResultLo = |
| 1375 | CurDAG->getMachineNode(Opcode: PPC::PLI8, dl, VT: MVT::i64, Op1: getI64Imm(Lo32)); |
| 1376 | SDValue Ops[] = {SDValue(ResultLo, 0), SDValue(ResultHi, 0), getI32Imm(32), |
| 1377 | getI32Imm(0)}; |
| 1378 | return CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops); |
| 1379 | } |
| 1380 | |
| 1381 | static SDNode *selectI64Imm(SelectionDAG *CurDAG, const SDLoc &dl, uint64_t Imm, |
| 1382 | unsigned *InstCnt = nullptr) { |
| 1383 | unsigned InstCntDirect = 0; |
| 1384 | // No more than 3 instructions are used if we can select the i64 immediate |
| 1385 | // directly. |
| 1386 | SDNode *Result = selectI64ImmDirect(CurDAG, dl, Imm, InstCnt&: InstCntDirect); |
| 1387 | |
| 1388 | const PPCSubtarget &Subtarget = |
| 1389 | CurDAG->getMachineFunction().getSubtarget<PPCSubtarget>(); |
| 1390 | |
| 1391 | // If we have prefixed instructions and there is a chance we can |
| 1392 | // materialize the constant with fewer prefixed instructions than |
| 1393 | // non-prefixed, try that. |
| 1394 | if (Subtarget.hasPrefixInstrs() && InstCntDirect != 1) { |
| 1395 | unsigned InstCntDirectP = 0; |
| 1396 | SDNode *ResultP = selectI64ImmDirectPrefix(CurDAG, dl, Imm, InstCnt&: InstCntDirectP); |
| 1397 | // Use the prefix case in either of two cases: |
| 1398 | // 1) We have no result from the non-prefix case to use. |
| 1399 | // 2) The non-prefix case uses more instructions than the prefix case. |
| 1400 | // If the prefix and non-prefix cases use the same number of instructions |
| 1401 | // we will prefer the non-prefix case. |
| 1402 | if (ResultP && (!Result || InstCntDirectP < InstCntDirect)) { |
| 1403 | if (InstCnt) |
| 1404 | *InstCnt = InstCntDirectP; |
| 1405 | return ResultP; |
| 1406 | } |
| 1407 | } |
| 1408 | |
| 1409 | if (Result) { |
| 1410 | if (InstCnt) |
| 1411 | *InstCnt = InstCntDirect; |
| 1412 | return Result; |
| 1413 | } |
| 1414 | auto getI32Imm = [CurDAG, dl](unsigned Imm) { |
| 1415 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
| 1416 | }; |
| 1417 | |
| 1418 | uint32_t Hi16OfLo32 = (Lo_32(Value: Imm) >> 16) & 0xffff; |
| 1419 | uint32_t Lo16OfLo32 = Lo_32(Value: Imm) & 0xffff; |
| 1420 | |
| 1421 | // Try to use 4 instructions to materialize the immediate which is "almost" a |
| 1422 | // splat of a 32 bit immediate. |
| 1423 | if (Hi16OfLo32 && Lo16OfLo32) { |
| 1424 | uint32_t Hi16OfHi32 = (Hi_32(Value: Imm) >> 16) & 0xffff; |
| 1425 | uint32_t Lo16OfHi32 = Hi_32(Value: Imm) & 0xffff; |
| 1426 | bool IsSelected = false; |
| 1427 | |
| 1428 | auto getSplat = [CurDAG, dl, getI32Imm](uint32_t Hi16, uint32_t Lo16) { |
| 1429 | SDNode *Result = |
| 1430 | CurDAG->getMachineNode(Opcode: PPC::LIS8, dl, VT: MVT::i64, Op1: getI32Imm(Hi16)); |
| 1431 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, |
| 1432 | Op1: SDValue(Result, 0), Op2: getI32Imm(Lo16)); |
| 1433 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(32), |
| 1434 | getI32Imm(0)}; |
| 1435 | return CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops); |
| 1436 | }; |
| 1437 | |
| 1438 | if (Hi16OfHi32 == Lo16OfHi32 && Lo16OfHi32 == Lo16OfLo32) { |
| 1439 | IsSelected = true; |
| 1440 | Result = getSplat(Hi16OfLo32, Lo16OfLo32); |
| 1441 | // Modify Hi16OfHi32. |
| 1442 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(48), |
| 1443 | getI32Imm(0)}; |
| 1444 | Result = CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops); |
| 1445 | } else if (Hi16OfHi32 == Hi16OfLo32 && Hi16OfLo32 == Lo16OfLo32) { |
| 1446 | IsSelected = true; |
| 1447 | Result = getSplat(Hi16OfHi32, Lo16OfHi32); |
| 1448 | // Modify Lo16OfLo32. |
| 1449 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(16), |
| 1450 | getI32Imm(16), getI32Imm(31)}; |
| 1451 | Result = CurDAG->getMachineNode(Opcode: PPC::RLWIMI8, dl, VT: MVT::i64, Ops); |
| 1452 | } else if (Lo16OfHi32 == Lo16OfLo32 && Hi16OfLo32 == Lo16OfLo32) { |
| 1453 | IsSelected = true; |
| 1454 | Result = getSplat(Hi16OfHi32, Lo16OfHi32); |
| 1455 | // Modify Hi16OfLo32. |
| 1456 | SDValue Ops[] = {SDValue(Result, 0), SDValue(Result, 0), getI32Imm(16), |
| 1457 | getI32Imm(0), getI32Imm(15)}; |
| 1458 | Result = CurDAG->getMachineNode(Opcode: PPC::RLWIMI8, dl, VT: MVT::i64, Ops); |
| 1459 | } |
| 1460 | if (IsSelected == true) { |
| 1461 | if (InstCnt) |
| 1462 | *InstCnt = 4; |
| 1463 | return Result; |
| 1464 | } |
| 1465 | } |
| 1466 | |
| 1467 | // Handle the upper 32 bit value. |
| 1468 | Result = |
| 1469 | selectI64ImmDirect(CurDAG, dl, Imm: Imm & 0xffffffff00000000, InstCnt&: InstCntDirect); |
| 1470 | // Add in the last bits as required. |
| 1471 | if (Hi16OfLo32) { |
| 1472 | Result = CurDAG->getMachineNode(Opcode: PPC::ORIS8, dl, VT: MVT::i64, |
| 1473 | Op1: SDValue(Result, 0), Op2: getI32Imm(Hi16OfLo32)); |
| 1474 | ++InstCntDirect; |
| 1475 | } |
| 1476 | if (Lo16OfLo32) { |
| 1477 | Result = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, Op1: SDValue(Result, 0), |
| 1478 | Op2: getI32Imm(Lo16OfLo32)); |
| 1479 | ++InstCntDirect; |
| 1480 | } |
| 1481 | if (InstCnt) |
| 1482 | *InstCnt = InstCntDirect; |
| 1483 | return Result; |
| 1484 | } |
| 1485 | |
| 1486 | // Select a 64-bit constant. |
| 1487 | static SDNode *selectI64Imm(SelectionDAG *CurDAG, SDNode *N) { |
| 1488 | SDLoc dl(N); |
| 1489 | |
| 1490 | // Get 64 bit value. |
| 1491 | int64_t Imm = N->getAsZExtVal(); |
| 1492 | if (unsigned MinSize = allUsesTruncate(CurDAG, N)) { |
| 1493 | uint64_t SextImm = SignExtend64(X: Imm, B: MinSize); |
| 1494 | SDValue SDImm = CurDAG->getTargetConstant(Val: SextImm, DL: dl, VT: MVT::i64); |
| 1495 | if (isInt<16>(x: SextImm)) |
| 1496 | return CurDAG->getMachineNode(Opcode: PPC::LI8, dl, VT: MVT::i64, Op1: SDImm); |
| 1497 | } |
| 1498 | return selectI64Imm(CurDAG, dl, Imm); |
| 1499 | } |
| 1500 | |
| 1501 | namespace { |
| 1502 | |
| 1503 | class BitPermutationSelector { |
| 1504 | struct ValueBit { |
| 1505 | SDValue V; |
| 1506 | |
| 1507 | // The bit number in the value, using a convention where bit 0 is the |
| 1508 | // lowest-order bit. |
| 1509 | unsigned Idx; |
| 1510 | |
| 1511 | // ConstZero means a bit we need to mask off. |
| 1512 | // Variable is a bit comes from an input variable. |
| 1513 | // VariableKnownToBeZero is also a bit comes from an input variable, |
| 1514 | // but it is known to be already zero. So we do not need to mask them. |
| 1515 | enum Kind { |
| 1516 | ConstZero, |
| 1517 | Variable, |
| 1518 | VariableKnownToBeZero |
| 1519 | } K; |
| 1520 | |
| 1521 | ValueBit(SDValue V, unsigned I, Kind K = Variable) |
| 1522 | : V(V), Idx(I), K(K) {} |
| 1523 | ValueBit(Kind K = Variable) : Idx(UINT32_MAX), K(K) {} |
| 1524 | |
| 1525 | bool isZero() const { |
| 1526 | return K == ConstZero || K == VariableKnownToBeZero; |
| 1527 | } |
| 1528 | |
| 1529 | bool hasValue() const { |
| 1530 | return K == Variable || K == VariableKnownToBeZero; |
| 1531 | } |
| 1532 | |
| 1533 | SDValue getValue() const { |
| 1534 | assert(hasValue() && "Cannot get the value of a constant bit" ); |
| 1535 | return V; |
| 1536 | } |
| 1537 | |
| 1538 | unsigned getValueBitIndex() const { |
| 1539 | assert(hasValue() && "Cannot get the value bit index of a constant bit" ); |
| 1540 | return Idx; |
| 1541 | } |
| 1542 | }; |
| 1543 | |
| 1544 | // A bit group has the same underlying value and the same rotate factor. |
| 1545 | struct BitGroup { |
| 1546 | SDValue V; |
| 1547 | unsigned RLAmt; |
| 1548 | unsigned StartIdx, EndIdx; |
| 1549 | |
| 1550 | // This rotation amount assumes that the lower 32 bits of the quantity are |
| 1551 | // replicated in the high 32 bits by the rotation operator (which is done |
| 1552 | // by rlwinm and friends in 64-bit mode). |
| 1553 | bool Repl32; |
| 1554 | // Did converting to Repl32 == true change the rotation factor? If it did, |
| 1555 | // it decreased it by 32. |
| 1556 | bool Repl32CR; |
| 1557 | // Was this group coalesced after setting Repl32 to true? |
| 1558 | bool Repl32Coalesced; |
| 1559 | |
| 1560 | BitGroup(SDValue V, unsigned R, unsigned S, unsigned E) |
| 1561 | : V(V), RLAmt(R), StartIdx(S), EndIdx(E), Repl32(false), Repl32CR(false), |
| 1562 | Repl32Coalesced(false) { |
| 1563 | LLVM_DEBUG(dbgs() << "\tbit group for " << V.getNode() << " RLAmt = " << R |
| 1564 | << " [" << S << ", " << E << "]\n" ); |
| 1565 | } |
| 1566 | }; |
| 1567 | |
| 1568 | // Information on each (Value, RLAmt) pair (like the number of groups |
| 1569 | // associated with each) used to choose the lowering method. |
| 1570 | struct ValueRotInfo { |
| 1571 | SDValue V; |
| 1572 | unsigned RLAmt = std::numeric_limits<unsigned>::max(); |
| 1573 | unsigned NumGroups = 0; |
| 1574 | unsigned FirstGroupStartIdx = std::numeric_limits<unsigned>::max(); |
| 1575 | bool Repl32 = false; |
| 1576 | |
| 1577 | ValueRotInfo() = default; |
| 1578 | |
| 1579 | // For sorting (in reverse order) by NumGroups, and then by |
| 1580 | // FirstGroupStartIdx. |
| 1581 | bool operator < (const ValueRotInfo &Other) const { |
| 1582 | // We need to sort so that the non-Repl32 come first because, when we're |
| 1583 | // doing masking, the Repl32 bit groups might be subsumed into the 64-bit |
| 1584 | // masking operation. |
| 1585 | if (Repl32 < Other.Repl32) |
| 1586 | return true; |
| 1587 | else if (Repl32 > Other.Repl32) |
| 1588 | return false; |
| 1589 | else if (NumGroups > Other.NumGroups) |
| 1590 | return true; |
| 1591 | else if (NumGroups < Other.NumGroups) |
| 1592 | return false; |
| 1593 | else if (RLAmt == 0 && Other.RLAmt != 0) |
| 1594 | return true; |
| 1595 | else if (RLAmt != 0 && Other.RLAmt == 0) |
| 1596 | return false; |
| 1597 | else if (FirstGroupStartIdx < Other.FirstGroupStartIdx) |
| 1598 | return true; |
| 1599 | return false; |
| 1600 | } |
| 1601 | }; |
| 1602 | |
| 1603 | using ValueBitsMemoizedValue = std::pair<bool, SmallVector<ValueBit, 64>>; |
| 1604 | using ValueBitsMemoizer = |
| 1605 | DenseMap<SDValue, std::unique_ptr<ValueBitsMemoizedValue>>; |
| 1606 | ValueBitsMemoizer Memoizer; |
| 1607 | |
| 1608 | // Return a pair of bool and a SmallVector pointer to a memoization entry. |
| 1609 | // The bool is true if something interesting was deduced, otherwise if we're |
| 1610 | // providing only a generic representation of V (or something else likewise |
| 1611 | // uninteresting for instruction selection) through the SmallVector. |
| 1612 | std::pair<bool, SmallVector<ValueBit, 64> *> getValueBits(SDValue V, |
| 1613 | unsigned NumBits) { |
| 1614 | auto &ValueEntry = Memoizer[V]; |
| 1615 | if (ValueEntry) |
| 1616 | return std::make_pair(x&: ValueEntry->first, y: &ValueEntry->second); |
| 1617 | ValueEntry.reset(p: new ValueBitsMemoizedValue()); |
| 1618 | bool &Interesting = ValueEntry->first; |
| 1619 | SmallVector<ValueBit, 64> &Bits = ValueEntry->second; |
| 1620 | Bits.resize(N: NumBits); |
| 1621 | |
| 1622 | switch (V.getOpcode()) { |
| 1623 | default: break; |
| 1624 | case ISD::ROTL: |
| 1625 | if (isa<ConstantSDNode>(Val: V.getOperand(i: 1))) { |
| 1626 | assert(isPowerOf2_32(NumBits) && "rotl bits should be power of 2!" ); |
| 1627 | unsigned RotAmt = V.getConstantOperandVal(i: 1) & (NumBits - 1); |
| 1628 | |
| 1629 | const auto &LHSBits = *getValueBits(V: V.getOperand(i: 0), NumBits).second; |
| 1630 | |
| 1631 | for (unsigned i = 0; i < NumBits; ++i) |
| 1632 | Bits[i] = LHSBits[i < RotAmt ? i + (NumBits - RotAmt) : i - RotAmt]; |
| 1633 | |
| 1634 | return std::make_pair(x&: Interesting = true, y: &Bits); |
| 1635 | } |
| 1636 | break; |
| 1637 | case ISD::SHL: |
| 1638 | case PPCISD::SHL: |
| 1639 | if (isa<ConstantSDNode>(Val: V.getOperand(i: 1))) { |
| 1640 | // sld takes 7 bits, slw takes 6. |
| 1641 | unsigned ShiftAmt = V.getConstantOperandVal(i: 1) & ((NumBits << 1) - 1); |
| 1642 | |
| 1643 | const auto &LHSBits = *getValueBits(V: V.getOperand(i: 0), NumBits).second; |
| 1644 | |
| 1645 | if (ShiftAmt >= NumBits) { |
| 1646 | for (unsigned i = 0; i < NumBits; ++i) |
| 1647 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1648 | } else { |
| 1649 | for (unsigned i = ShiftAmt; i < NumBits; ++i) |
| 1650 | Bits[i] = LHSBits[i - ShiftAmt]; |
| 1651 | for (unsigned i = 0; i < ShiftAmt; ++i) |
| 1652 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1653 | } |
| 1654 | |
| 1655 | return std::make_pair(x&: Interesting = true, y: &Bits); |
| 1656 | } |
| 1657 | break; |
| 1658 | case ISD::SRL: |
| 1659 | case PPCISD::SRL: |
| 1660 | if (isa<ConstantSDNode>(Val: V.getOperand(i: 1))) { |
| 1661 | // srd takes lowest 7 bits, srw takes 6. |
| 1662 | unsigned ShiftAmt = V.getConstantOperandVal(i: 1) & ((NumBits << 1) - 1); |
| 1663 | |
| 1664 | const auto &LHSBits = *getValueBits(V: V.getOperand(i: 0), NumBits).second; |
| 1665 | |
| 1666 | if (ShiftAmt >= NumBits) { |
| 1667 | for (unsigned i = 0; i < NumBits; ++i) |
| 1668 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1669 | } else { |
| 1670 | for (unsigned i = 0; i < NumBits - ShiftAmt; ++i) |
| 1671 | Bits[i] = LHSBits[i + ShiftAmt]; |
| 1672 | for (unsigned i = NumBits - ShiftAmt; i < NumBits; ++i) |
| 1673 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1674 | } |
| 1675 | |
| 1676 | return std::make_pair(x&: Interesting = true, y: &Bits); |
| 1677 | } |
| 1678 | break; |
| 1679 | case ISD::AND: |
| 1680 | if (isa<ConstantSDNode>(Val: V.getOperand(i: 1))) { |
| 1681 | uint64_t Mask = V.getConstantOperandVal(i: 1); |
| 1682 | |
| 1683 | const SmallVector<ValueBit, 64> *LHSBits; |
| 1684 | // Mark this as interesting, only if the LHS was also interesting. This |
| 1685 | // prevents the overall procedure from matching a single immediate 'and' |
| 1686 | // (which is non-optimal because such an and might be folded with other |
| 1687 | // things if we don't select it here). |
| 1688 | std::tie(args&: Interesting, args&: LHSBits) = getValueBits(V: V.getOperand(i: 0), NumBits); |
| 1689 | |
| 1690 | for (unsigned i = 0; i < NumBits; ++i) |
| 1691 | if (((Mask >> i) & 1) == 1) |
| 1692 | Bits[i] = (*LHSBits)[i]; |
| 1693 | else { |
| 1694 | // AND instruction masks this bit. If the input is already zero, |
| 1695 | // we have nothing to do here. Otherwise, make the bit ConstZero. |
| 1696 | if ((*LHSBits)[i].isZero()) |
| 1697 | Bits[i] = (*LHSBits)[i]; |
| 1698 | else |
| 1699 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1700 | } |
| 1701 | |
| 1702 | return std::make_pair(x&: Interesting, y: &Bits); |
| 1703 | } |
| 1704 | break; |
| 1705 | case ISD::OR: { |
| 1706 | const auto &LHSBits = *getValueBits(V: V.getOperand(i: 0), NumBits).second; |
| 1707 | const auto &RHSBits = *getValueBits(V: V.getOperand(i: 1), NumBits).second; |
| 1708 | |
| 1709 | bool AllDisjoint = true; |
| 1710 | SDValue LastVal = SDValue(); |
| 1711 | unsigned LastIdx = 0; |
| 1712 | for (unsigned i = 0; i < NumBits; ++i) { |
| 1713 | if (LHSBits[i].isZero() && RHSBits[i].isZero()) { |
| 1714 | // If both inputs are known to be zero and one is ConstZero and |
| 1715 | // another is VariableKnownToBeZero, we can select whichever |
| 1716 | // we like. To minimize the number of bit groups, we select |
| 1717 | // VariableKnownToBeZero if this bit is the next bit of the same |
| 1718 | // input variable from the previous bit. Otherwise, we select |
| 1719 | // ConstZero. |
| 1720 | if (LHSBits[i].hasValue() && LHSBits[i].getValue() == LastVal && |
| 1721 | LHSBits[i].getValueBitIndex() == LastIdx + 1) |
| 1722 | Bits[i] = LHSBits[i]; |
| 1723 | else if (RHSBits[i].hasValue() && RHSBits[i].getValue() == LastVal && |
| 1724 | RHSBits[i].getValueBitIndex() == LastIdx + 1) |
| 1725 | Bits[i] = RHSBits[i]; |
| 1726 | else |
| 1727 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1728 | } |
| 1729 | else if (LHSBits[i].isZero()) |
| 1730 | Bits[i] = RHSBits[i]; |
| 1731 | else if (RHSBits[i].isZero()) |
| 1732 | Bits[i] = LHSBits[i]; |
| 1733 | else { |
| 1734 | AllDisjoint = false; |
| 1735 | break; |
| 1736 | } |
| 1737 | // We remember the value and bit index of this bit. |
| 1738 | if (Bits[i].hasValue()) { |
| 1739 | LastVal = Bits[i].getValue(); |
| 1740 | LastIdx = Bits[i].getValueBitIndex(); |
| 1741 | } |
| 1742 | else { |
| 1743 | if (LastVal) LastVal = SDValue(); |
| 1744 | LastIdx = 0; |
| 1745 | } |
| 1746 | } |
| 1747 | |
| 1748 | if (!AllDisjoint) |
| 1749 | break; |
| 1750 | |
| 1751 | return std::make_pair(x&: Interesting = true, y: &Bits); |
| 1752 | } |
| 1753 | case ISD::ZERO_EXTEND: { |
| 1754 | // We support only the case with zero extension from i32 to i64 so far. |
| 1755 | if (V.getValueType() != MVT::i64 || |
| 1756 | V.getOperand(i: 0).getValueType() != MVT::i32) |
| 1757 | break; |
| 1758 | |
| 1759 | const SmallVector<ValueBit, 64> *LHSBits; |
| 1760 | const unsigned NumOperandBits = 32; |
| 1761 | std::tie(args&: Interesting, args&: LHSBits) = getValueBits(V: V.getOperand(i: 0), |
| 1762 | NumBits: NumOperandBits); |
| 1763 | |
| 1764 | for (unsigned i = 0; i < NumOperandBits; ++i) |
| 1765 | Bits[i] = (*LHSBits)[i]; |
| 1766 | |
| 1767 | for (unsigned i = NumOperandBits; i < NumBits; ++i) |
| 1768 | Bits[i] = ValueBit(ValueBit::ConstZero); |
| 1769 | |
| 1770 | return std::make_pair(x&: Interesting, y: &Bits); |
| 1771 | } |
| 1772 | case ISD::TRUNCATE: { |
| 1773 | EVT FromType = V.getOperand(i: 0).getValueType(); |
| 1774 | EVT ToType = V.getValueType(); |
| 1775 | // We support only the case with truncate from i64 to i32. |
| 1776 | if (FromType != MVT::i64 || ToType != MVT::i32) |
| 1777 | break; |
| 1778 | const unsigned NumAllBits = FromType.getSizeInBits(); |
| 1779 | SmallVector<ValueBit, 64> *InBits; |
| 1780 | std::tie(args&: Interesting, args&: InBits) = getValueBits(V: V.getOperand(i: 0), |
| 1781 | NumBits: NumAllBits); |
| 1782 | const unsigned NumValidBits = ToType.getSizeInBits(); |
| 1783 | |
| 1784 | // A 32-bit instruction cannot touch upper 32-bit part of 64-bit value. |
| 1785 | // So, we cannot include this truncate. |
| 1786 | bool UseUpper32bit = false; |
| 1787 | for (unsigned i = 0; i < NumValidBits; ++i) |
| 1788 | if ((*InBits)[i].hasValue() && (*InBits)[i].getValueBitIndex() >= 32) { |
| 1789 | UseUpper32bit = true; |
| 1790 | break; |
| 1791 | } |
| 1792 | if (UseUpper32bit) |
| 1793 | break; |
| 1794 | |
| 1795 | for (unsigned i = 0; i < NumValidBits; ++i) |
| 1796 | Bits[i] = (*InBits)[i]; |
| 1797 | |
| 1798 | return std::make_pair(x&: Interesting, y: &Bits); |
| 1799 | } |
| 1800 | case ISD::AssertZext: { |
| 1801 | // For AssertZext, we look through the operand and |
| 1802 | // mark the bits known to be zero. |
| 1803 | const SmallVector<ValueBit, 64> *LHSBits; |
| 1804 | std::tie(args&: Interesting, args&: LHSBits) = getValueBits(V: V.getOperand(i: 0), |
| 1805 | NumBits); |
| 1806 | |
| 1807 | EVT FromType = cast<VTSDNode>(Val: V.getOperand(i: 1))->getVT(); |
| 1808 | const unsigned NumValidBits = FromType.getSizeInBits(); |
| 1809 | for (unsigned i = 0; i < NumValidBits; ++i) |
| 1810 | Bits[i] = (*LHSBits)[i]; |
| 1811 | |
| 1812 | // These bits are known to be zero but the AssertZext may be from a value |
| 1813 | // that already has some constant zero bits (i.e. from a masking and). |
| 1814 | for (unsigned i = NumValidBits; i < NumBits; ++i) |
| 1815 | Bits[i] = (*LHSBits)[i].hasValue() |
| 1816 | ? ValueBit((*LHSBits)[i].getValue(), |
| 1817 | (*LHSBits)[i].getValueBitIndex(), |
| 1818 | ValueBit::VariableKnownToBeZero) |
| 1819 | : ValueBit(ValueBit::ConstZero); |
| 1820 | |
| 1821 | return std::make_pair(x&: Interesting, y: &Bits); |
| 1822 | } |
| 1823 | case ISD::LOAD: |
| 1824 | LoadSDNode *LD = cast<LoadSDNode>(Val&: V); |
| 1825 | if (ISD::isZEXTLoad(N: V.getNode()) && V.getResNo() == 0) { |
| 1826 | EVT VT = LD->getMemoryVT(); |
| 1827 | const unsigned NumValidBits = VT.getSizeInBits(); |
| 1828 | |
| 1829 | for (unsigned i = 0; i < NumValidBits; ++i) |
| 1830 | Bits[i] = ValueBit(V, i); |
| 1831 | |
| 1832 | // These bits are known to be zero. |
| 1833 | for (unsigned i = NumValidBits; i < NumBits; ++i) |
| 1834 | Bits[i] = ValueBit(V, i, ValueBit::VariableKnownToBeZero); |
| 1835 | |
| 1836 | // Zero-extending load itself cannot be optimized. So, it is not |
| 1837 | // interesting by itself though it gives useful information. |
| 1838 | return std::make_pair(x&: Interesting = false, y: &Bits); |
| 1839 | } |
| 1840 | break; |
| 1841 | } |
| 1842 | |
| 1843 | for (unsigned i = 0; i < NumBits; ++i) |
| 1844 | Bits[i] = ValueBit(V, i); |
| 1845 | |
| 1846 | return std::make_pair(x&: Interesting = false, y: &Bits); |
| 1847 | } |
| 1848 | |
| 1849 | // For each value (except the constant ones), compute the left-rotate amount |
| 1850 | // to get it from its original to final position. |
| 1851 | void computeRotationAmounts() { |
| 1852 | NeedMask = false; |
| 1853 | RLAmt.resize(N: Bits.size()); |
| 1854 | for (unsigned i = 0; i < Bits.size(); ++i) |
| 1855 | if (Bits[i].hasValue()) { |
| 1856 | unsigned VBI = Bits[i].getValueBitIndex(); |
| 1857 | if (i >= VBI) |
| 1858 | RLAmt[i] = i - VBI; |
| 1859 | else |
| 1860 | RLAmt[i] = Bits.size() - (VBI - i); |
| 1861 | } else if (Bits[i].isZero()) { |
| 1862 | NeedMask = true; |
| 1863 | RLAmt[i] = UINT32_MAX; |
| 1864 | } else { |
| 1865 | llvm_unreachable("Unknown value bit type" ); |
| 1866 | } |
| 1867 | } |
| 1868 | |
| 1869 | // Collect groups of consecutive bits with the same underlying value and |
| 1870 | // rotation factor. If we're doing late masking, we ignore zeros, otherwise |
| 1871 | // they break up groups. |
| 1872 | void collectBitGroups(bool LateMask) { |
| 1873 | BitGroups.clear(); |
| 1874 | |
| 1875 | unsigned LastRLAmt = RLAmt[0]; |
| 1876 | SDValue LastValue = Bits[0].hasValue() ? Bits[0].getValue() : SDValue(); |
| 1877 | unsigned LastGroupStartIdx = 0; |
| 1878 | bool IsGroupOfZeros = !Bits[LastGroupStartIdx].hasValue(); |
| 1879 | for (unsigned i = 1; i < Bits.size(); ++i) { |
| 1880 | unsigned ThisRLAmt = RLAmt[i]; |
| 1881 | SDValue ThisValue = Bits[i].hasValue() ? Bits[i].getValue() : SDValue(); |
| 1882 | if (LateMask && !ThisValue) { |
| 1883 | ThisValue = LastValue; |
| 1884 | ThisRLAmt = LastRLAmt; |
| 1885 | // If we're doing late masking, then the first bit group always starts |
| 1886 | // at zero (even if the first bits were zero). |
| 1887 | if (BitGroups.empty()) |
| 1888 | LastGroupStartIdx = 0; |
| 1889 | } |
| 1890 | |
| 1891 | // If this bit is known to be zero and the current group is a bit group |
| 1892 | // of zeros, we do not need to terminate the current bit group even the |
| 1893 | // Value or RLAmt does not match here. Instead, we terminate this group |
| 1894 | // when the first non-zero bit appears later. |
| 1895 | if (IsGroupOfZeros && Bits[i].isZero()) |
| 1896 | continue; |
| 1897 | |
| 1898 | // If this bit has the same underlying value and the same rotate factor as |
| 1899 | // the last one, then they're part of the same group. |
| 1900 | if (ThisRLAmt == LastRLAmt && ThisValue == LastValue) |
| 1901 | // We cannot continue the current group if this bits is not known to |
| 1902 | // be zero in a bit group of zeros. |
| 1903 | if (!(IsGroupOfZeros && ThisValue && !Bits[i].isZero())) |
| 1904 | continue; |
| 1905 | |
| 1906 | if (LastValue.getNode()) |
| 1907 | BitGroups.push_back(Elt: BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, |
| 1908 | i-1)); |
| 1909 | LastRLAmt = ThisRLAmt; |
| 1910 | LastValue = ThisValue; |
| 1911 | LastGroupStartIdx = i; |
| 1912 | IsGroupOfZeros = !Bits[LastGroupStartIdx].hasValue(); |
| 1913 | } |
| 1914 | if (LastValue.getNode()) |
| 1915 | BitGroups.push_back(Elt: BitGroup(LastValue, LastRLAmt, LastGroupStartIdx, |
| 1916 | Bits.size()-1)); |
| 1917 | |
| 1918 | if (BitGroups.empty()) |
| 1919 | return; |
| 1920 | |
| 1921 | // We might be able to combine the first and last groups. |
| 1922 | if (BitGroups.size() > 1) { |
| 1923 | // If the first and last groups are the same, then remove the first group |
| 1924 | // in favor of the last group, making the ending index of the last group |
| 1925 | // equal to the ending index of the to-be-removed first group. |
| 1926 | if (BitGroups[0].StartIdx == 0 && |
| 1927 | BitGroups[BitGroups.size()-1].EndIdx == Bits.size()-1 && |
| 1928 | BitGroups[0].V == BitGroups[BitGroups.size()-1].V && |
| 1929 | BitGroups[0].RLAmt == BitGroups[BitGroups.size()-1].RLAmt) { |
| 1930 | LLVM_DEBUG(dbgs() << "\tcombining final bit group with initial one\n" ); |
| 1931 | BitGroups[BitGroups.size()-1].EndIdx = BitGroups[0].EndIdx; |
| 1932 | BitGroups.erase(CI: BitGroups.begin()); |
| 1933 | } |
| 1934 | } |
| 1935 | } |
| 1936 | |
| 1937 | // Take all (SDValue, RLAmt) pairs and sort them by the number of groups |
| 1938 | // associated with each. If the number of groups are same, we prefer a group |
| 1939 | // which does not require rotate, i.e. RLAmt is 0, to avoid the first rotate |
| 1940 | // instruction. If there is a degeneracy, pick the one that occurs |
| 1941 | // first (in the final value). |
| 1942 | void collectValueRotInfo() { |
| 1943 | ValueRots.clear(); |
| 1944 | |
| 1945 | for (auto &BG : BitGroups) { |
| 1946 | unsigned RLAmtKey = BG.RLAmt + (BG.Repl32 ? 64 : 0); |
| 1947 | ValueRotInfo &VRI = ValueRots[std::make_pair(x&: BG.V, y&: RLAmtKey)]; |
| 1948 | VRI.V = BG.V; |
| 1949 | VRI.RLAmt = BG.RLAmt; |
| 1950 | VRI.Repl32 = BG.Repl32; |
| 1951 | VRI.NumGroups += 1; |
| 1952 | VRI.FirstGroupStartIdx = std::min(a: VRI.FirstGroupStartIdx, b: BG.StartIdx); |
| 1953 | } |
| 1954 | |
| 1955 | // Now that we've collected the various ValueRotInfo instances, we need to |
| 1956 | // sort them. |
| 1957 | ValueRotsVec.clear(); |
| 1958 | for (auto &I : ValueRots) { |
| 1959 | ValueRotsVec.push_back(Elt: I.second); |
| 1960 | } |
| 1961 | llvm::sort(C&: ValueRotsVec); |
| 1962 | } |
| 1963 | |
| 1964 | // In 64-bit mode, rlwinm and friends have a rotation operator that |
| 1965 | // replicates the low-order 32 bits into the high-order 32-bits. The mask |
| 1966 | // indices of these instructions can only be in the lower 32 bits, so they |
| 1967 | // can only represent some 64-bit bit groups. However, when they can be used, |
| 1968 | // the 32-bit replication can be used to represent, as a single bit group, |
| 1969 | // otherwise separate bit groups. We'll convert to replicated-32-bit bit |
| 1970 | // groups when possible. Returns true if any of the bit groups were |
| 1971 | // converted. |
| 1972 | void assignRepl32BitGroups() { |
| 1973 | // If we have bits like this: |
| 1974 | // |
| 1975 | // Indices: 15 14 13 12 11 10 9 8 7 6 5 4 3 2 1 0 |
| 1976 | // V bits: ... 7 6 5 4 3 2 1 0 31 30 29 28 27 26 25 24 |
| 1977 | // Groups: | RLAmt = 8 | RLAmt = 40 | |
| 1978 | // |
| 1979 | // But, making use of a 32-bit operation that replicates the low-order 32 |
| 1980 | // bits into the high-order 32 bits, this can be one bit group with a RLAmt |
| 1981 | // of 8. |
| 1982 | |
| 1983 | auto IsAllLow32 = [this](BitGroup & BG) { |
| 1984 | if (BG.StartIdx <= BG.EndIdx) { |
| 1985 | for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) { |
| 1986 | if (!Bits[i].hasValue()) |
| 1987 | continue; |
| 1988 | if (Bits[i].getValueBitIndex() >= 32) |
| 1989 | return false; |
| 1990 | } |
| 1991 | } else { |
| 1992 | for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) { |
| 1993 | if (!Bits[i].hasValue()) |
| 1994 | continue; |
| 1995 | if (Bits[i].getValueBitIndex() >= 32) |
| 1996 | return false; |
| 1997 | } |
| 1998 | for (unsigned i = 0; i <= BG.EndIdx; ++i) { |
| 1999 | if (!Bits[i].hasValue()) |
| 2000 | continue; |
| 2001 | if (Bits[i].getValueBitIndex() >= 32) |
| 2002 | return false; |
| 2003 | } |
| 2004 | } |
| 2005 | |
| 2006 | return true; |
| 2007 | }; |
| 2008 | |
| 2009 | for (auto &BG : BitGroups) { |
| 2010 | // If this bit group has RLAmt of 0 and will not be merged with |
| 2011 | // another bit group, we don't benefit from Repl32. We don't mark |
| 2012 | // such group to give more freedom for later instruction selection. |
| 2013 | if (BG.RLAmt == 0) { |
| 2014 | auto PotentiallyMerged = [this](BitGroup & BG) { |
| 2015 | for (auto &BG2 : BitGroups) |
| 2016 | if (&BG != &BG2 && BG.V == BG2.V && |
| 2017 | (BG2.RLAmt == 0 || BG2.RLAmt == 32)) |
| 2018 | return true; |
| 2019 | return false; |
| 2020 | }; |
| 2021 | if (!PotentiallyMerged(BG)) |
| 2022 | continue; |
| 2023 | } |
| 2024 | if (BG.StartIdx < 32 && BG.EndIdx < 32) { |
| 2025 | if (IsAllLow32(BG)) { |
| 2026 | if (BG.RLAmt >= 32) { |
| 2027 | BG.RLAmt -= 32; |
| 2028 | BG.Repl32CR = true; |
| 2029 | } |
| 2030 | |
| 2031 | BG.Repl32 = true; |
| 2032 | |
| 2033 | LLVM_DEBUG(dbgs() << "\t32-bit replicated bit group for " |
| 2034 | << BG.V.getNode() << " RLAmt = " << BG.RLAmt << " [" |
| 2035 | << BG.StartIdx << ", " << BG.EndIdx << "]\n" ); |
| 2036 | } |
| 2037 | } |
| 2038 | } |
| 2039 | |
| 2040 | // Now walk through the bit groups, consolidating where possible. |
| 2041 | for (auto I = BitGroups.begin(); I != BitGroups.end();) { |
| 2042 | // We might want to remove this bit group by merging it with the previous |
| 2043 | // group (which might be the ending group). |
| 2044 | auto IP = (I == BitGroups.begin()) ? |
| 2045 | std::prev(x: BitGroups.end()) : std::prev(x: I); |
| 2046 | if (I->Repl32 && IP->Repl32 && I->V == IP->V && I->RLAmt == IP->RLAmt && |
| 2047 | I->StartIdx == (IP->EndIdx + 1) % 64 && I != IP) { |
| 2048 | |
| 2049 | LLVM_DEBUG(dbgs() << "\tcombining 32-bit replicated bit group for " |
| 2050 | << I->V.getNode() << " RLAmt = " << I->RLAmt << " [" |
| 2051 | << I->StartIdx << ", " << I->EndIdx |
| 2052 | << "] with group with range [" << IP->StartIdx << ", " |
| 2053 | << IP->EndIdx << "]\n" ); |
| 2054 | |
| 2055 | IP->EndIdx = I->EndIdx; |
| 2056 | IP->Repl32CR = IP->Repl32CR || I->Repl32CR; |
| 2057 | IP->Repl32Coalesced = true; |
| 2058 | I = BitGroups.erase(CI: I); |
| 2059 | continue; |
| 2060 | } else { |
| 2061 | // There is a special case worth handling: If there is a single group |
| 2062 | // covering the entire upper 32 bits, and it can be merged with both |
| 2063 | // the next and previous groups (which might be the same group), then |
| 2064 | // do so. If it is the same group (so there will be only one group in |
| 2065 | // total), then we need to reverse the order of the range so that it |
| 2066 | // covers the entire 64 bits. |
| 2067 | if (I->StartIdx == 32 && I->EndIdx == 63) { |
| 2068 | assert(std::next(I) == BitGroups.end() && |
| 2069 | "bit group ends at index 63 but there is another?" ); |
| 2070 | auto IN = BitGroups.begin(); |
| 2071 | |
| 2072 | if (IP->Repl32 && IN->Repl32 && I->V == IP->V && I->V == IN->V && |
| 2073 | (I->RLAmt % 32) == IP->RLAmt && (I->RLAmt % 32) == IN->RLAmt && |
| 2074 | IP->EndIdx == 31 && IN->StartIdx == 0 && I != IP && |
| 2075 | IsAllLow32(*I)) { |
| 2076 | |
| 2077 | LLVM_DEBUG(dbgs() << "\tcombining bit group for " << I->V.getNode() |
| 2078 | << " RLAmt = " << I->RLAmt << " [" << I->StartIdx |
| 2079 | << ", " << I->EndIdx |
| 2080 | << "] with 32-bit replicated groups with ranges [" |
| 2081 | << IP->StartIdx << ", " << IP->EndIdx << "] and [" |
| 2082 | << IN->StartIdx << ", " << IN->EndIdx << "]\n" ); |
| 2083 | |
| 2084 | if (IP == IN) { |
| 2085 | // There is only one other group; change it to cover the whole |
| 2086 | // range (backward, so that it can still be Repl32 but cover the |
| 2087 | // whole 64-bit range). |
| 2088 | IP->StartIdx = 31; |
| 2089 | IP->EndIdx = 30; |
| 2090 | IP->Repl32CR = IP->Repl32CR || I->RLAmt >= 32; |
| 2091 | IP->Repl32Coalesced = true; |
| 2092 | I = BitGroups.erase(CI: I); |
| 2093 | } else { |
| 2094 | // There are two separate groups, one before this group and one |
| 2095 | // after us (at the beginning). We're going to remove this group, |
| 2096 | // but also the group at the very beginning. |
| 2097 | IP->EndIdx = IN->EndIdx; |
| 2098 | IP->Repl32CR = IP->Repl32CR || IN->Repl32CR || I->RLAmt >= 32; |
| 2099 | IP->Repl32Coalesced = true; |
| 2100 | I = BitGroups.erase(CI: I); |
| 2101 | BitGroups.erase(CI: BitGroups.begin()); |
| 2102 | } |
| 2103 | |
| 2104 | // This must be the last group in the vector (and we might have |
| 2105 | // just invalidated the iterator above), so break here. |
| 2106 | break; |
| 2107 | } |
| 2108 | } |
| 2109 | } |
| 2110 | |
| 2111 | ++I; |
| 2112 | } |
| 2113 | } |
| 2114 | |
| 2115 | SDValue getI32Imm(unsigned Imm, const SDLoc &dl) { |
| 2116 | return CurDAG->getTargetConstant(Val: Imm, DL: dl, VT: MVT::i32); |
| 2117 | } |
| 2118 | |
| 2119 | uint64_t getZerosMask() { |
| 2120 | uint64_t Mask = 0; |
| 2121 | for (unsigned i = 0; i < Bits.size(); ++i) { |
| 2122 | if (Bits[i].hasValue()) |
| 2123 | continue; |
| 2124 | Mask |= (UINT64_C(1) << i); |
| 2125 | } |
| 2126 | |
| 2127 | return ~Mask; |
| 2128 | } |
| 2129 | |
| 2130 | // This method extends an input value to 64 bit if input is 32-bit integer. |
| 2131 | // While selecting instructions in BitPermutationSelector in 64-bit mode, |
| 2132 | // an input value can be a 32-bit integer if a ZERO_EXTEND node is included. |
| 2133 | // In such case, we extend it to 64 bit to be consistent with other values. |
| 2134 | SDValue ExtendToInt64(SDValue V, const SDLoc &dl) { |
| 2135 | if (V.getValueSizeInBits() == 64) |
| 2136 | return V; |
| 2137 | |
| 2138 | assert(V.getValueSizeInBits() == 32); |
| 2139 | SDValue SubRegIdx = CurDAG->getTargetConstant(Val: PPC::sub_32, DL: dl, VT: MVT::i32); |
| 2140 | SDValue ImDef = SDValue(CurDAG->getMachineNode(Opcode: PPC::IMPLICIT_DEF, dl, |
| 2141 | VT: MVT::i64), 0); |
| 2142 | SDValue ExtVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::INSERT_SUBREG, dl, |
| 2143 | VT: MVT::i64, Op1: ImDef, Op2: V, |
| 2144 | Op3: SubRegIdx), 0); |
| 2145 | return ExtVal; |
| 2146 | } |
| 2147 | |
| 2148 | SDValue TruncateToInt32(SDValue V, const SDLoc &dl) { |
| 2149 | if (V.getValueSizeInBits() == 32) |
| 2150 | return V; |
| 2151 | |
| 2152 | assert(V.getValueSizeInBits() == 64); |
| 2153 | SDValue SubRegIdx = CurDAG->getTargetConstant(Val: PPC::sub_32, DL: dl, VT: MVT::i32); |
| 2154 | SDValue SubVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::EXTRACT_SUBREG, dl, |
| 2155 | VT: MVT::i32, Op1: V, Op2: SubRegIdx), 0); |
| 2156 | return SubVal; |
| 2157 | } |
| 2158 | |
| 2159 | // Depending on the number of groups for a particular value, it might be |
| 2160 | // better to rotate, mask explicitly (using andi/andis), and then or the |
| 2161 | // result. Select this part of the result first. |
| 2162 | void SelectAndParts32(const SDLoc &dl, SDValue &Res, unsigned *InstCnt) { |
| 2163 | if (BPermRewriterNoMasking) |
| 2164 | return; |
| 2165 | |
| 2166 | for (ValueRotInfo &VRI : ValueRotsVec) { |
| 2167 | unsigned Mask = 0; |
| 2168 | for (unsigned i = 0; i < Bits.size(); ++i) { |
| 2169 | if (!Bits[i].hasValue() || Bits[i].getValue() != VRI.V) |
| 2170 | continue; |
| 2171 | if (RLAmt[i] != VRI.RLAmt) |
| 2172 | continue; |
| 2173 | Mask |= (1u << i); |
| 2174 | } |
| 2175 | |
| 2176 | // Compute the masks for andi/andis that would be necessary. |
| 2177 | unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16; |
| 2178 | assert((ANDIMask != 0 || ANDISMask != 0) && |
| 2179 | "No set bits in mask for value bit groups" ); |
| 2180 | bool NeedsRotate = VRI.RLAmt != 0; |
| 2181 | |
| 2182 | // We're trying to minimize the number of instructions. If we have one |
| 2183 | // group, using one of andi/andis can break even. If we have three |
| 2184 | // groups, we can use both andi and andis and break even (to use both |
| 2185 | // andi and andis we also need to or the results together). We need four |
| 2186 | // groups if we also need to rotate. To use andi/andis we need to do more |
| 2187 | // than break even because rotate-and-mask instructions tend to be easier |
| 2188 | // to schedule. |
| 2189 | |
| 2190 | // FIXME: We've biased here against using andi/andis, which is right for |
| 2191 | // POWER cores, but not optimal everywhere. For example, on the A2, |
| 2192 | // andi/andis have single-cycle latency whereas the rotate-and-mask |
| 2193 | // instructions take two cycles, and it would be better to bias toward |
| 2194 | // andi/andis in break-even cases. |
| 2195 | |
| 2196 | unsigned NumAndInsts = (unsigned) NeedsRotate + |
| 2197 | (unsigned) (ANDIMask != 0) + |
| 2198 | (unsigned) (ANDISMask != 0) + |
| 2199 | (unsigned) (ANDIMask != 0 && ANDISMask != 0) + |
| 2200 | (unsigned) (bool) Res; |
| 2201 | |
| 2202 | LLVM_DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() |
| 2203 | << " RL: " << VRI.RLAmt << ":" |
| 2204 | << "\n\t\t\tisel using masking: " << NumAndInsts |
| 2205 | << " using rotates: " << VRI.NumGroups << "\n" ); |
| 2206 | |
| 2207 | if (NumAndInsts >= VRI.NumGroups) |
| 2208 | continue; |
| 2209 | |
| 2210 | LLVM_DEBUG(dbgs() << "\t\t\t\tusing masking\n" ); |
| 2211 | |
| 2212 | if (InstCnt) *InstCnt += NumAndInsts; |
| 2213 | |
| 2214 | SDValue VRot; |
| 2215 | if (VRI.RLAmt) { |
| 2216 | SDValue Ops[] = |
| 2217 | { TruncateToInt32(V: VRI.V, dl), getI32Imm(Imm: VRI.RLAmt, dl), |
| 2218 | getI32Imm(Imm: 0, dl), getI32Imm(Imm: 31, dl) }; |
| 2219 | VRot = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, |
| 2220 | Ops), 0); |
| 2221 | } else { |
| 2222 | VRot = TruncateToInt32(V: VRI.V, dl); |
| 2223 | } |
| 2224 | |
| 2225 | SDValue ANDIVal, ANDISVal; |
| 2226 | if (ANDIMask != 0) |
| 2227 | ANDIVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDI_rec, dl, VT: MVT::i32, |
| 2228 | Op1: VRot, Op2: getI32Imm(Imm: ANDIMask, dl)), |
| 2229 | 0); |
| 2230 | if (ANDISMask != 0) |
| 2231 | ANDISVal = |
| 2232 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDIS_rec, dl, VT: MVT::i32, Op1: VRot, |
| 2233 | Op2: getI32Imm(Imm: ANDISMask, dl)), |
| 2234 | 0); |
| 2235 | |
| 2236 | SDValue TotalVal; |
| 2237 | if (!ANDIVal) |
| 2238 | TotalVal = ANDISVal; |
| 2239 | else if (!ANDISVal) |
| 2240 | TotalVal = ANDIVal; |
| 2241 | else |
| 2242 | TotalVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR, dl, VT: MVT::i32, |
| 2243 | Op1: ANDIVal, Op2: ANDISVal), 0); |
| 2244 | |
| 2245 | if (!Res) |
| 2246 | Res = TotalVal; |
| 2247 | else |
| 2248 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR, dl, VT: MVT::i32, |
| 2249 | Op1: Res, Op2: TotalVal), 0); |
| 2250 | |
| 2251 | // Now, remove all groups with this underlying value and rotation |
| 2252 | // factor. |
| 2253 | eraseMatchingBitGroups(F: [VRI](const BitGroup &BG) { |
| 2254 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; |
| 2255 | }); |
| 2256 | } |
| 2257 | } |
| 2258 | |
| 2259 | // Instruction selection for the 32-bit case. |
| 2260 | SDNode *Select32(SDNode *N, bool LateMask, unsigned *InstCnt) { |
| 2261 | SDLoc dl(N); |
| 2262 | SDValue Res; |
| 2263 | |
| 2264 | if (InstCnt) *InstCnt = 0; |
| 2265 | |
| 2266 | // Take care of cases that should use andi/andis first. |
| 2267 | SelectAndParts32(dl, Res, InstCnt); |
| 2268 | |
| 2269 | // If we've not yet selected a 'starting' instruction, and we have no zeros |
| 2270 | // to fill in, select the (Value, RLAmt) with the highest priority (largest |
| 2271 | // number of groups), and start with this rotated value. |
| 2272 | if ((!NeedMask || LateMask) && !Res) { |
| 2273 | ValueRotInfo &VRI = ValueRotsVec[0]; |
| 2274 | if (VRI.RLAmt) { |
| 2275 | if (InstCnt) *InstCnt += 1; |
| 2276 | SDValue Ops[] = |
| 2277 | { TruncateToInt32(V: VRI.V, dl), getI32Imm(Imm: VRI.RLAmt, dl), |
| 2278 | getI32Imm(Imm: 0, dl), getI32Imm(Imm: 31, dl) }; |
| 2279 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops), |
| 2280 | 0); |
| 2281 | } else { |
| 2282 | Res = TruncateToInt32(V: VRI.V, dl); |
| 2283 | } |
| 2284 | |
| 2285 | // Now, remove all groups with this underlying value and rotation factor. |
| 2286 | eraseMatchingBitGroups(F: [VRI](const BitGroup &BG) { |
| 2287 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt; |
| 2288 | }); |
| 2289 | } |
| 2290 | |
| 2291 | if (InstCnt) *InstCnt += BitGroups.size(); |
| 2292 | |
| 2293 | // Insert the other groups (one at a time). |
| 2294 | for (auto &BG : BitGroups) { |
| 2295 | if (!Res) { |
| 2296 | SDValue Ops[] = |
| 2297 | { TruncateToInt32(V: BG.V, dl), getI32Imm(Imm: BG.RLAmt, dl), |
| 2298 | getI32Imm(Imm: Bits.size() - BG.EndIdx - 1, dl), |
| 2299 | getI32Imm(Imm: Bits.size() - BG.StartIdx - 1, dl) }; |
| 2300 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops), 0); |
| 2301 | } else { |
| 2302 | SDValue Ops[] = |
| 2303 | { Res, TruncateToInt32(V: BG.V, dl), getI32Imm(Imm: BG.RLAmt, dl), |
| 2304 | getI32Imm(Imm: Bits.size() - BG.EndIdx - 1, dl), |
| 2305 | getI32Imm(Imm: Bits.size() - BG.StartIdx - 1, dl) }; |
| 2306 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWIMI, dl, VT: MVT::i32, Ops), 0); |
| 2307 | } |
| 2308 | } |
| 2309 | |
| 2310 | if (LateMask) { |
| 2311 | unsigned Mask = (unsigned) getZerosMask(); |
| 2312 | |
| 2313 | unsigned ANDIMask = (Mask & UINT16_MAX), ANDISMask = Mask >> 16; |
| 2314 | assert((ANDIMask != 0 || ANDISMask != 0) && |
| 2315 | "No set bits in zeros mask?" ); |
| 2316 | |
| 2317 | if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + |
| 2318 | (unsigned) (ANDISMask != 0) + |
| 2319 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
| 2320 | |
| 2321 | SDValue ANDIVal, ANDISVal; |
| 2322 | if (ANDIMask != 0) |
| 2323 | ANDIVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDI_rec, dl, VT: MVT::i32, |
| 2324 | Op1: Res, Op2: getI32Imm(Imm: ANDIMask, dl)), |
| 2325 | 0); |
| 2326 | if (ANDISMask != 0) |
| 2327 | ANDISVal = |
| 2328 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDIS_rec, dl, VT: MVT::i32, Op1: Res, |
| 2329 | Op2: getI32Imm(Imm: ANDISMask, dl)), |
| 2330 | 0); |
| 2331 | |
| 2332 | if (!ANDIVal) |
| 2333 | Res = ANDISVal; |
| 2334 | else if (!ANDISVal) |
| 2335 | Res = ANDIVal; |
| 2336 | else |
| 2337 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR, dl, VT: MVT::i32, |
| 2338 | Op1: ANDIVal, Op2: ANDISVal), 0); |
| 2339 | } |
| 2340 | |
| 2341 | return Res.getNode(); |
| 2342 | } |
| 2343 | |
| 2344 | unsigned SelectRotMask64Count(unsigned RLAmt, bool Repl32, |
| 2345 | unsigned MaskStart, unsigned MaskEnd, |
| 2346 | bool IsIns) { |
| 2347 | // In the notation used by the instructions, 'start' and 'end' are reversed |
| 2348 | // because bits are counted from high to low order. |
| 2349 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
| 2350 | InstMaskEnd = 64 - MaskStart - 1; |
| 2351 | |
| 2352 | if (Repl32) |
| 2353 | return 1; |
| 2354 | |
| 2355 | if ((!IsIns && (InstMaskEnd == 63 || InstMaskStart == 0)) || |
| 2356 | InstMaskEnd == 63 - RLAmt) |
| 2357 | return 1; |
| 2358 | |
| 2359 | return 2; |
| 2360 | } |
| 2361 | |
| 2362 | // For 64-bit values, not all combinations of rotates and masks are |
| 2363 | // available. Produce one if it is available. |
| 2364 | SDValue SelectRotMask64(SDValue V, const SDLoc &dl, unsigned RLAmt, |
| 2365 | bool Repl32, unsigned MaskStart, unsigned MaskEnd, |
| 2366 | unsigned *InstCnt = nullptr) { |
| 2367 | // In the notation used by the instructions, 'start' and 'end' are reversed |
| 2368 | // because bits are counted from high to low order. |
| 2369 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
| 2370 | InstMaskEnd = 64 - MaskStart - 1; |
| 2371 | |
| 2372 | if (InstCnt) *InstCnt += 1; |
| 2373 | |
| 2374 | if (Repl32) { |
| 2375 | // This rotation amount assumes that the lower 32 bits of the quantity |
| 2376 | // are replicated in the high 32 bits by the rotation operator (which is |
| 2377 | // done by rlwinm and friends). |
| 2378 | assert(InstMaskStart >= 32 && "Mask cannot start out of range" ); |
| 2379 | assert(InstMaskEnd >= 32 && "Mask cannot end out of range" ); |
| 2380 | SDValue Ops[] = |
| 2381 | { ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2382 | getI32Imm(Imm: InstMaskStart - 32, dl), getI32Imm(Imm: InstMaskEnd - 32, dl) }; |
| 2383 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM8, dl, VT: MVT::i64, |
| 2384 | Ops), 0); |
| 2385 | } |
| 2386 | |
| 2387 | if (InstMaskEnd == 63) { |
| 2388 | SDValue Ops[] = |
| 2389 | { ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2390 | getI32Imm(Imm: InstMaskStart, dl) }; |
| 2391 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Ops), 0); |
| 2392 | } |
| 2393 | |
| 2394 | if (InstMaskStart == 0) { |
| 2395 | SDValue Ops[] = |
| 2396 | { ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2397 | getI32Imm(Imm: InstMaskEnd, dl) }; |
| 2398 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICR, dl, VT: MVT::i64, Ops), 0); |
| 2399 | } |
| 2400 | |
| 2401 | if (InstMaskEnd == 63 - RLAmt) { |
| 2402 | SDValue Ops[] = |
| 2403 | { ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2404 | getI32Imm(Imm: InstMaskStart, dl) }; |
| 2405 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDIC, dl, VT: MVT::i64, Ops), 0); |
| 2406 | } |
| 2407 | |
| 2408 | // We cannot do this with a single instruction, so we'll use two. The |
| 2409 | // problem is that we're not free to choose both a rotation amount and mask |
| 2410 | // start and end independently. We can choose an arbitrary mask start and |
| 2411 | // end, but then the rotation amount is fixed. Rotation, however, can be |
| 2412 | // inverted, and so by applying an "inverse" rotation first, we can get the |
| 2413 | // desired result. |
| 2414 | if (InstCnt) *InstCnt += 1; |
| 2415 | |
| 2416 | // The rotation mask for the second instruction must be MaskStart. |
| 2417 | unsigned RLAmt2 = MaskStart; |
| 2418 | // The first instruction must rotate V so that the overall rotation amount |
| 2419 | // is RLAmt. |
| 2420 | unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; |
| 2421 | if (RLAmt1) |
| 2422 | V = SelectRotMask64(V, dl, RLAmt: RLAmt1, Repl32: false, MaskStart: 0, MaskEnd: 63); |
| 2423 | return SelectRotMask64(V, dl, RLAmt: RLAmt2, Repl32: false, MaskStart, MaskEnd); |
| 2424 | } |
| 2425 | |
| 2426 | // For 64-bit values, not all combinations of rotates and masks are |
| 2427 | // available. Produce a rotate-mask-and-insert if one is available. |
| 2428 | SDValue SelectRotMaskIns64(SDValue Base, SDValue V, const SDLoc &dl, |
| 2429 | unsigned RLAmt, bool Repl32, unsigned MaskStart, |
| 2430 | unsigned MaskEnd, unsigned *InstCnt = nullptr) { |
| 2431 | // In the notation used by the instructions, 'start' and 'end' are reversed |
| 2432 | // because bits are counted from high to low order. |
| 2433 | unsigned InstMaskStart = 64 - MaskEnd - 1, |
| 2434 | InstMaskEnd = 64 - MaskStart - 1; |
| 2435 | |
| 2436 | if (InstCnt) *InstCnt += 1; |
| 2437 | |
| 2438 | if (Repl32) { |
| 2439 | // This rotation amount assumes that the lower 32 bits of the quantity |
| 2440 | // are replicated in the high 32 bits by the rotation operator (which is |
| 2441 | // done by rlwinm and friends). |
| 2442 | assert(InstMaskStart >= 32 && "Mask cannot start out of range" ); |
| 2443 | assert(InstMaskEnd >= 32 && "Mask cannot end out of range" ); |
| 2444 | SDValue Ops[] = |
| 2445 | { ExtendToInt64(V: Base, dl), ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2446 | getI32Imm(Imm: InstMaskStart - 32, dl), getI32Imm(Imm: InstMaskEnd - 32, dl) }; |
| 2447 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWIMI8, dl, VT: MVT::i64, |
| 2448 | Ops), 0); |
| 2449 | } |
| 2450 | |
| 2451 | if (InstMaskEnd == 63 - RLAmt) { |
| 2452 | SDValue Ops[] = |
| 2453 | { ExtendToInt64(V: Base, dl), ExtendToInt64(V, dl), getI32Imm(Imm: RLAmt, dl), |
| 2454 | getI32Imm(Imm: InstMaskStart, dl) }; |
| 2455 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDIMI, dl, VT: MVT::i64, Ops), 0); |
| 2456 | } |
| 2457 | |
| 2458 | // We cannot do this with a single instruction, so we'll use two. The |
| 2459 | // problem is that we're not free to choose both a rotation amount and mask |
| 2460 | // start and end independently. We can choose an arbitrary mask start and |
| 2461 | // end, but then the rotation amount is fixed. Rotation, however, can be |
| 2462 | // inverted, and so by applying an "inverse" rotation first, we can get the |
| 2463 | // desired result. |
| 2464 | if (InstCnt) *InstCnt += 1; |
| 2465 | |
| 2466 | // The rotation mask for the second instruction must be MaskStart. |
| 2467 | unsigned RLAmt2 = MaskStart; |
| 2468 | // The first instruction must rotate V so that the overall rotation amount |
| 2469 | // is RLAmt. |
| 2470 | unsigned RLAmt1 = (64 + RLAmt - RLAmt2) % 64; |
| 2471 | if (RLAmt1) |
| 2472 | V = SelectRotMask64(V, dl, RLAmt: RLAmt1, Repl32: false, MaskStart: 0, MaskEnd: 63); |
| 2473 | return SelectRotMaskIns64(Base, V, dl, RLAmt: RLAmt2, Repl32: false, MaskStart, MaskEnd); |
| 2474 | } |
| 2475 | |
| 2476 | void SelectAndParts64(const SDLoc &dl, SDValue &Res, unsigned *InstCnt) { |
| 2477 | if (BPermRewriterNoMasking) |
| 2478 | return; |
| 2479 | |
| 2480 | // The idea here is the same as in the 32-bit version, but with additional |
| 2481 | // complications from the fact that Repl32 might be true. Because we |
| 2482 | // aggressively convert bit groups to Repl32 form (which, for small |
| 2483 | // rotation factors, involves no other change), and then coalesce, it might |
| 2484 | // be the case that a single 64-bit masking operation could handle both |
| 2485 | // some Repl32 groups and some non-Repl32 groups. If converting to Repl32 |
| 2486 | // form allowed coalescing, then we must use a 32-bit rotaton in order to |
| 2487 | // completely capture the new combined bit group. |
| 2488 | |
| 2489 | for (ValueRotInfo &VRI : ValueRotsVec) { |
| 2490 | uint64_t Mask = 0; |
| 2491 | |
| 2492 | // We need to add to the mask all bits from the associated bit groups. |
| 2493 | // If Repl32 is false, we need to add bits from bit groups that have |
| 2494 | // Repl32 true, but are trivially convertable to Repl32 false. Such a |
| 2495 | // group is trivially convertable if it overlaps only with the lower 32 |
| 2496 | // bits, and the group has not been coalesced. |
| 2497 | auto MatchingBG = [VRI](const BitGroup &BG) { |
| 2498 | if (VRI.V != BG.V) |
| 2499 | return false; |
| 2500 | |
| 2501 | unsigned EffRLAmt = BG.RLAmt; |
| 2502 | if (!VRI.Repl32 && BG.Repl32) { |
| 2503 | if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx <= BG.EndIdx && |
| 2504 | !BG.Repl32Coalesced) { |
| 2505 | if (BG.Repl32CR) |
| 2506 | EffRLAmt += 32; |
| 2507 | } else { |
| 2508 | return false; |
| 2509 | } |
| 2510 | } else if (VRI.Repl32 != BG.Repl32) { |
| 2511 | return false; |
| 2512 | } |
| 2513 | |
| 2514 | return VRI.RLAmt == EffRLAmt; |
| 2515 | }; |
| 2516 | |
| 2517 | for (auto &BG : BitGroups) { |
| 2518 | if (!MatchingBG(BG)) |
| 2519 | continue; |
| 2520 | |
| 2521 | if (BG.StartIdx <= BG.EndIdx) { |
| 2522 | for (unsigned i = BG.StartIdx; i <= BG.EndIdx; ++i) |
| 2523 | Mask |= (UINT64_C(1) << i); |
| 2524 | } else { |
| 2525 | for (unsigned i = BG.StartIdx; i < Bits.size(); ++i) |
| 2526 | Mask |= (UINT64_C(1) << i); |
| 2527 | for (unsigned i = 0; i <= BG.EndIdx; ++i) |
| 2528 | Mask |= (UINT64_C(1) << i); |
| 2529 | } |
| 2530 | } |
| 2531 | |
| 2532 | // We can use the 32-bit andi/andis technique if the mask does not |
| 2533 | // require any higher-order bits. This can save an instruction compared |
| 2534 | // to always using the general 64-bit technique. |
| 2535 | bool Use32BitInsts = isUInt<32>(x: Mask); |
| 2536 | // Compute the masks for andi/andis that would be necessary. |
| 2537 | unsigned ANDIMask = (Mask & UINT16_MAX), |
| 2538 | ANDISMask = (Mask >> 16) & UINT16_MAX; |
| 2539 | |
| 2540 | bool NeedsRotate = VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(x: Mask)); |
| 2541 | |
| 2542 | unsigned NumAndInsts = (unsigned) NeedsRotate + |
| 2543 | (unsigned) (bool) Res; |
| 2544 | unsigned NumOfSelectInsts = 0; |
| 2545 | selectI64Imm(CurDAG, dl, Imm: Mask, InstCnt: &NumOfSelectInsts); |
| 2546 | assert(NumOfSelectInsts > 0 && "Failed to select an i64 constant." ); |
| 2547 | if (Use32BitInsts) |
| 2548 | NumAndInsts += (unsigned) (ANDIMask != 0) + (unsigned) (ANDISMask != 0) + |
| 2549 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
| 2550 | else |
| 2551 | NumAndInsts += NumOfSelectInsts + /* and */ 1; |
| 2552 | |
| 2553 | unsigned NumRLInsts = 0; |
| 2554 | bool FirstBG = true; |
| 2555 | bool MoreBG = false; |
| 2556 | for (auto &BG : BitGroups) { |
| 2557 | if (!MatchingBG(BG)) { |
| 2558 | MoreBG = true; |
| 2559 | continue; |
| 2560 | } |
| 2561 | NumRLInsts += |
| 2562 | SelectRotMask64Count(RLAmt: BG.RLAmt, Repl32: BG.Repl32, MaskStart: BG.StartIdx, MaskEnd: BG.EndIdx, |
| 2563 | IsIns: !FirstBG); |
| 2564 | FirstBG = false; |
| 2565 | } |
| 2566 | |
| 2567 | LLVM_DEBUG(dbgs() << "\t\trotation groups for " << VRI.V.getNode() |
| 2568 | << " RL: " << VRI.RLAmt << (VRI.Repl32 ? " (32):" : ":" ) |
| 2569 | << "\n\t\t\tisel using masking: " << NumAndInsts |
| 2570 | << " using rotates: " << NumRLInsts << "\n" ); |
| 2571 | |
| 2572 | // When we'd use andi/andis, we bias toward using the rotates (andi only |
| 2573 | // has a record form, and is cracked on POWER cores). However, when using |
| 2574 | // general 64-bit constant formation, bias toward the constant form, |
| 2575 | // because that exposes more opportunities for CSE. |
| 2576 | if (NumAndInsts > NumRLInsts) |
| 2577 | continue; |
| 2578 | // When merging multiple bit groups, instruction or is used. |
| 2579 | // But when rotate is used, rldimi can inert the rotated value into any |
| 2580 | // register, so instruction or can be avoided. |
| 2581 | if ((Use32BitInsts || MoreBG) && NumAndInsts == NumRLInsts) |
| 2582 | continue; |
| 2583 | |
| 2584 | LLVM_DEBUG(dbgs() << "\t\t\t\tusing masking\n" ); |
| 2585 | |
| 2586 | if (InstCnt) *InstCnt += NumAndInsts; |
| 2587 | |
| 2588 | SDValue VRot; |
| 2589 | // We actually need to generate a rotation if we have a non-zero rotation |
| 2590 | // factor or, in the Repl32 case, if we care about any of the |
| 2591 | // higher-order replicated bits. In the latter case, we generate a mask |
| 2592 | // backward so that it actually includes the entire 64 bits. |
| 2593 | if (VRI.RLAmt || (VRI.Repl32 && !isUInt<32>(x: Mask))) |
| 2594 | VRot = SelectRotMask64(V: VRI.V, dl, RLAmt: VRI.RLAmt, Repl32: VRI.Repl32, |
| 2595 | MaskStart: VRI.Repl32 ? 31 : 0, MaskEnd: VRI.Repl32 ? 30 : 63); |
| 2596 | else |
| 2597 | VRot = VRI.V; |
| 2598 | |
| 2599 | SDValue TotalVal; |
| 2600 | if (Use32BitInsts) { |
| 2601 | assert((ANDIMask != 0 || ANDISMask != 0) && |
| 2602 | "No set bits in mask when using 32-bit ands for 64-bit value" ); |
| 2603 | |
| 2604 | SDValue ANDIVal, ANDISVal; |
| 2605 | if (ANDIMask != 0) |
| 2606 | ANDIVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDI8_rec, dl, VT: MVT::i64, |
| 2607 | Op1: ExtendToInt64(V: VRot, dl), |
| 2608 | Op2: getI32Imm(Imm: ANDIMask, dl)), |
| 2609 | 0); |
| 2610 | if (ANDISMask != 0) |
| 2611 | ANDISVal = |
| 2612 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDIS8_rec, dl, VT: MVT::i64, |
| 2613 | Op1: ExtendToInt64(V: VRot, dl), |
| 2614 | Op2: getI32Imm(Imm: ANDISMask, dl)), |
| 2615 | 0); |
| 2616 | |
| 2617 | if (!ANDIVal) |
| 2618 | TotalVal = ANDISVal; |
| 2619 | else if (!ANDISVal) |
| 2620 | TotalVal = ANDIVal; |
| 2621 | else |
| 2622 | TotalVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR8, dl, VT: MVT::i64, |
| 2623 | Op1: ExtendToInt64(V: ANDIVal, dl), Op2: ANDISVal), 0); |
| 2624 | } else { |
| 2625 | TotalVal = SDValue(selectI64Imm(CurDAG, dl, Imm: Mask), 0); |
| 2626 | TotalVal = |
| 2627 | SDValue(CurDAG->getMachineNode(Opcode: PPC::AND8, dl, VT: MVT::i64, |
| 2628 | Op1: ExtendToInt64(V: VRot, dl), Op2: TotalVal), |
| 2629 | 0); |
| 2630 | } |
| 2631 | |
| 2632 | if (!Res) |
| 2633 | Res = TotalVal; |
| 2634 | else |
| 2635 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR8, dl, VT: MVT::i64, |
| 2636 | Op1: ExtendToInt64(V: Res, dl), Op2: TotalVal), |
| 2637 | 0); |
| 2638 | |
| 2639 | // Now, remove all groups with this underlying value and rotation |
| 2640 | // factor. |
| 2641 | eraseMatchingBitGroups(F: MatchingBG); |
| 2642 | } |
| 2643 | } |
| 2644 | |
| 2645 | // Instruction selection for the 64-bit case. |
| 2646 | SDNode *Select64(SDNode *N, bool LateMask, unsigned *InstCnt) { |
| 2647 | SDLoc dl(N); |
| 2648 | SDValue Res; |
| 2649 | |
| 2650 | if (InstCnt) *InstCnt = 0; |
| 2651 | |
| 2652 | // Take care of cases that should use andi/andis first. |
| 2653 | SelectAndParts64(dl, Res, InstCnt); |
| 2654 | |
| 2655 | // If we've not yet selected a 'starting' instruction, and we have no zeros |
| 2656 | // to fill in, select the (Value, RLAmt) with the highest priority (largest |
| 2657 | // number of groups), and start with this rotated value. |
| 2658 | if ((!NeedMask || LateMask) && !Res) { |
| 2659 | // If we have both Repl32 groups and non-Repl32 groups, the non-Repl32 |
| 2660 | // groups will come first, and so the VRI representing the largest number |
| 2661 | // of groups might not be first (it might be the first Repl32 groups). |
| 2662 | unsigned MaxGroupsIdx = 0; |
| 2663 | if (!ValueRotsVec[0].Repl32) { |
| 2664 | for (unsigned i = 0, ie = ValueRotsVec.size(); i < ie; ++i) |
| 2665 | if (ValueRotsVec[i].Repl32) { |
| 2666 | if (ValueRotsVec[i].NumGroups > ValueRotsVec[0].NumGroups) |
| 2667 | MaxGroupsIdx = i; |
| 2668 | break; |
| 2669 | } |
| 2670 | } |
| 2671 | |
| 2672 | ValueRotInfo &VRI = ValueRotsVec[MaxGroupsIdx]; |
| 2673 | bool NeedsRotate = false; |
| 2674 | if (VRI.RLAmt) { |
| 2675 | NeedsRotate = true; |
| 2676 | } else if (VRI.Repl32) { |
| 2677 | for (auto &BG : BitGroups) { |
| 2678 | if (BG.V != VRI.V || BG.RLAmt != VRI.RLAmt || |
| 2679 | BG.Repl32 != VRI.Repl32) |
| 2680 | continue; |
| 2681 | |
| 2682 | // We don't need a rotate if the bit group is confined to the lower |
| 2683 | // 32 bits. |
| 2684 | if (BG.StartIdx < 32 && BG.EndIdx < 32 && BG.StartIdx < BG.EndIdx) |
| 2685 | continue; |
| 2686 | |
| 2687 | NeedsRotate = true; |
| 2688 | break; |
| 2689 | } |
| 2690 | } |
| 2691 | |
| 2692 | if (NeedsRotate) |
| 2693 | Res = SelectRotMask64(V: VRI.V, dl, RLAmt: VRI.RLAmt, Repl32: VRI.Repl32, |
| 2694 | MaskStart: VRI.Repl32 ? 31 : 0, MaskEnd: VRI.Repl32 ? 30 : 63, |
| 2695 | InstCnt); |
| 2696 | else |
| 2697 | Res = VRI.V; |
| 2698 | |
| 2699 | // Now, remove all groups with this underlying value and rotation factor. |
| 2700 | if (Res) |
| 2701 | eraseMatchingBitGroups(F: [VRI](const BitGroup &BG) { |
| 2702 | return BG.V == VRI.V && BG.RLAmt == VRI.RLAmt && |
| 2703 | BG.Repl32 == VRI.Repl32; |
| 2704 | }); |
| 2705 | } |
| 2706 | |
| 2707 | // Because 64-bit rotates are more flexible than inserts, we might have a |
| 2708 | // preference regarding which one we do first (to save one instruction). |
| 2709 | if (!Res) |
| 2710 | for (auto I = BitGroups.begin(), IE = BitGroups.end(); I != IE; ++I) { |
| 2711 | if (SelectRotMask64Count(RLAmt: I->RLAmt, Repl32: I->Repl32, MaskStart: I->StartIdx, MaskEnd: I->EndIdx, |
| 2712 | IsIns: false) < |
| 2713 | SelectRotMask64Count(RLAmt: I->RLAmt, Repl32: I->Repl32, MaskStart: I->StartIdx, MaskEnd: I->EndIdx, |
| 2714 | IsIns: true)) { |
| 2715 | if (I != BitGroups.begin()) { |
| 2716 | BitGroup BG = *I; |
| 2717 | BitGroups.erase(CI: I); |
| 2718 | BitGroups.insert(I: BitGroups.begin(), Elt: BG); |
| 2719 | } |
| 2720 | |
| 2721 | break; |
| 2722 | } |
| 2723 | } |
| 2724 | |
| 2725 | // Insert the other groups (one at a time). |
| 2726 | for (auto &BG : BitGroups) { |
| 2727 | if (!Res) |
| 2728 | Res = SelectRotMask64(V: BG.V, dl, RLAmt: BG.RLAmt, Repl32: BG.Repl32, MaskStart: BG.StartIdx, |
| 2729 | MaskEnd: BG.EndIdx, InstCnt); |
| 2730 | else |
| 2731 | Res = SelectRotMaskIns64(Base: Res, V: BG.V, dl, RLAmt: BG.RLAmt, Repl32: BG.Repl32, |
| 2732 | MaskStart: BG.StartIdx, MaskEnd: BG.EndIdx, InstCnt); |
| 2733 | } |
| 2734 | |
| 2735 | if (LateMask) { |
| 2736 | uint64_t Mask = getZerosMask(); |
| 2737 | |
| 2738 | // We can use the 32-bit andi/andis technique if the mask does not |
| 2739 | // require any higher-order bits. This can save an instruction compared |
| 2740 | // to always using the general 64-bit technique. |
| 2741 | bool Use32BitInsts = isUInt<32>(x: Mask); |
| 2742 | // Compute the masks for andi/andis that would be necessary. |
| 2743 | unsigned ANDIMask = (Mask & UINT16_MAX), |
| 2744 | ANDISMask = (Mask >> 16) & UINT16_MAX; |
| 2745 | |
| 2746 | if (Use32BitInsts) { |
| 2747 | assert((ANDIMask != 0 || ANDISMask != 0) && |
| 2748 | "No set bits in mask when using 32-bit ands for 64-bit value" ); |
| 2749 | |
| 2750 | if (InstCnt) *InstCnt += (unsigned) (ANDIMask != 0) + |
| 2751 | (unsigned) (ANDISMask != 0) + |
| 2752 | (unsigned) (ANDIMask != 0 && ANDISMask != 0); |
| 2753 | |
| 2754 | SDValue ANDIVal, ANDISVal; |
| 2755 | if (ANDIMask != 0) |
| 2756 | ANDIVal = SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDI8_rec, dl, VT: MVT::i64, |
| 2757 | Op1: ExtendToInt64(V: Res, dl), |
| 2758 | Op2: getI32Imm(Imm: ANDIMask, dl)), |
| 2759 | 0); |
| 2760 | if (ANDISMask != 0) |
| 2761 | ANDISVal = |
| 2762 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDIS8_rec, dl, VT: MVT::i64, |
| 2763 | Op1: ExtendToInt64(V: Res, dl), |
| 2764 | Op2: getI32Imm(Imm: ANDISMask, dl)), |
| 2765 | 0); |
| 2766 | |
| 2767 | if (!ANDIVal) |
| 2768 | Res = ANDISVal; |
| 2769 | else if (!ANDISVal) |
| 2770 | Res = ANDIVal; |
| 2771 | else |
| 2772 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR8, dl, VT: MVT::i64, |
| 2773 | Op1: ExtendToInt64(V: ANDIVal, dl), Op2: ANDISVal), 0); |
| 2774 | } else { |
| 2775 | unsigned NumOfSelectInsts = 0; |
| 2776 | SDValue MaskVal = |
| 2777 | SDValue(selectI64Imm(CurDAG, dl, Imm: Mask, InstCnt: &NumOfSelectInsts), 0); |
| 2778 | Res = SDValue(CurDAG->getMachineNode(Opcode: PPC::AND8, dl, VT: MVT::i64, |
| 2779 | Op1: ExtendToInt64(V: Res, dl), Op2: MaskVal), |
| 2780 | 0); |
| 2781 | if (InstCnt) |
| 2782 | *InstCnt += NumOfSelectInsts + /* and */ 1; |
| 2783 | } |
| 2784 | } |
| 2785 | |
| 2786 | return Res.getNode(); |
| 2787 | } |
| 2788 | |
| 2789 | SDNode *Select(SDNode *N, bool LateMask, unsigned *InstCnt = nullptr) { |
| 2790 | // Fill in BitGroups. |
| 2791 | collectBitGroups(LateMask); |
| 2792 | if (BitGroups.empty()) |
| 2793 | return nullptr; |
| 2794 | |
| 2795 | // For 64-bit values, figure out when we can use 32-bit instructions. |
| 2796 | if (Bits.size() == 64) |
| 2797 | assignRepl32BitGroups(); |
| 2798 | |
| 2799 | // Fill in ValueRotsVec. |
| 2800 | collectValueRotInfo(); |
| 2801 | |
| 2802 | if (Bits.size() == 32) { |
| 2803 | return Select32(N, LateMask, InstCnt); |
| 2804 | } else { |
| 2805 | assert(Bits.size() == 64 && "Not 64 bits here?" ); |
| 2806 | return Select64(N, LateMask, InstCnt); |
| 2807 | } |
| 2808 | |
| 2809 | return nullptr; |
| 2810 | } |
| 2811 | |
| 2812 | void eraseMatchingBitGroups(function_ref<bool(const BitGroup &)> F) { |
| 2813 | erase_if(C&: BitGroups, P: F); |
| 2814 | } |
| 2815 | |
| 2816 | SmallVector<ValueBit, 64> Bits; |
| 2817 | |
| 2818 | bool NeedMask = false; |
| 2819 | SmallVector<unsigned, 64> RLAmt; |
| 2820 | |
| 2821 | SmallVector<BitGroup, 16> BitGroups; |
| 2822 | |
| 2823 | DenseMap<std::pair<SDValue, unsigned>, ValueRotInfo> ValueRots; |
| 2824 | SmallVector<ValueRotInfo, 16> ValueRotsVec; |
| 2825 | |
| 2826 | SelectionDAG *CurDAG = nullptr; |
| 2827 | |
| 2828 | public: |
| 2829 | BitPermutationSelector(SelectionDAG *DAG) |
| 2830 | : CurDAG(DAG) {} |
| 2831 | |
| 2832 | // Here we try to match complex bit permutations into a set of |
| 2833 | // rotate-and-shift/shift/and/or instructions, using a set of heuristics |
| 2834 | // known to produce optimal code for common cases (like i32 byte swapping). |
| 2835 | SDNode *Select(SDNode *N) { |
| 2836 | Memoizer.clear(); |
| 2837 | auto Result = |
| 2838 | getValueBits(V: SDValue(N, 0), NumBits: N->getValueType(ResNo: 0).getSizeInBits()); |
| 2839 | if (!Result.first) |
| 2840 | return nullptr; |
| 2841 | Bits = std::move(*Result.second); |
| 2842 | |
| 2843 | LLVM_DEBUG(dbgs() << "Considering bit-permutation-based instruction" |
| 2844 | " selection for: " ); |
| 2845 | LLVM_DEBUG(N->dump(CurDAG)); |
| 2846 | |
| 2847 | // Fill it RLAmt and set NeedMask. |
| 2848 | computeRotationAmounts(); |
| 2849 | |
| 2850 | if (!NeedMask) |
| 2851 | return Select(N, LateMask: false); |
| 2852 | |
| 2853 | // We currently have two techniques for handling results with zeros: early |
| 2854 | // masking (the default) and late masking. Late masking is sometimes more |
| 2855 | // efficient, but because the structure of the bit groups is different, it |
| 2856 | // is hard to tell without generating both and comparing the results. With |
| 2857 | // late masking, we ignore zeros in the resulting value when inserting each |
| 2858 | // set of bit groups, and then mask in the zeros at the end. With early |
| 2859 | // masking, we only insert the non-zero parts of the result at every step. |
| 2860 | |
| 2861 | unsigned InstCnt = 0, InstCntLateMask = 0; |
| 2862 | LLVM_DEBUG(dbgs() << "\tEarly masking:\n" ); |
| 2863 | SDNode *RN = Select(N, LateMask: false, InstCnt: &InstCnt); |
| 2864 | LLVM_DEBUG(dbgs() << "\t\tisel would use " << InstCnt << " instructions\n" ); |
| 2865 | |
| 2866 | LLVM_DEBUG(dbgs() << "\tLate masking:\n" ); |
| 2867 | SDNode *RNLM = Select(N, LateMask: true, InstCnt: &InstCntLateMask); |
| 2868 | LLVM_DEBUG(dbgs() << "\t\tisel would use " << InstCntLateMask |
| 2869 | << " instructions\n" ); |
| 2870 | |
| 2871 | if (InstCnt <= InstCntLateMask) { |
| 2872 | LLVM_DEBUG(dbgs() << "\tUsing early-masking for isel\n" ); |
| 2873 | return RN; |
| 2874 | } |
| 2875 | |
| 2876 | LLVM_DEBUG(dbgs() << "\tUsing late-masking for isel\n" ); |
| 2877 | return RNLM; |
| 2878 | } |
| 2879 | }; |
| 2880 | |
| 2881 | class IntegerCompareEliminator { |
| 2882 | SelectionDAG *CurDAG; |
| 2883 | PPCDAGToDAGISel *S; |
| 2884 | // Conversion type for interpreting results of a 32-bit instruction as |
| 2885 | // a 64-bit value or vice versa. |
| 2886 | enum ExtOrTruncConversion { Ext, Trunc }; |
| 2887 | |
| 2888 | // Modifiers to guide how an ISD::SETCC node's result is to be computed |
| 2889 | // in a GPR. |
| 2890 | // ZExtOrig - use the original condition code, zero-extend value |
| 2891 | // ZExtInvert - invert the condition code, zero-extend value |
| 2892 | // SExtOrig - use the original condition code, sign-extend value |
| 2893 | // SExtInvert - invert the condition code, sign-extend value |
| 2894 | enum SetccInGPROpts { ZExtOrig, ZExtInvert, SExtOrig, SExtInvert }; |
| 2895 | |
| 2896 | // Comparisons against zero to emit GPR code sequences for. Each of these |
| 2897 | // sequences may need to be emitted for two or more equivalent patterns. |
| 2898 | // For example (a >= 0) == (a > -1). The direction of the comparison (</>) |
| 2899 | // matters as well as the extension type: sext (-1/0), zext (1/0). |
| 2900 | // GEZExt - (zext (LHS >= 0)) |
| 2901 | // GESExt - (sext (LHS >= 0)) |
| 2902 | // LEZExt - (zext (LHS <= 0)) |
| 2903 | // LESExt - (sext (LHS <= 0)) |
| 2904 | enum ZeroCompare { GEZExt, GESExt, LEZExt, LESExt }; |
| 2905 | |
| 2906 | SDNode *tryEXTEND(SDNode *N); |
| 2907 | SDNode *tryLogicOpOfCompares(SDNode *N); |
| 2908 | SDValue computeLogicOpInGPR(SDValue LogicOp); |
| 2909 | SDValue signExtendInputIfNeeded(SDValue Input); |
| 2910 | SDValue zeroExtendInputIfNeeded(SDValue Input); |
| 2911 | SDValue addExtOrTrunc(SDValue NatWidthRes, ExtOrTruncConversion Conv); |
| 2912 | SDValue getCompoundZeroComparisonInGPR(SDValue LHS, SDLoc dl, |
| 2913 | ZeroCompare CmpTy); |
| 2914 | SDValue get32BitZExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 2915 | int64_t RHSValue, SDLoc dl); |
| 2916 | SDValue get32BitSExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 2917 | int64_t RHSValue, SDLoc dl); |
| 2918 | SDValue get64BitZExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 2919 | int64_t RHSValue, SDLoc dl); |
| 2920 | SDValue get64BitSExtCompare(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 2921 | int64_t RHSValue, SDLoc dl); |
| 2922 | SDValue getSETCCInGPR(SDValue Compare, SetccInGPROpts ConvOpts); |
| 2923 | |
| 2924 | public: |
| 2925 | IntegerCompareEliminator(SelectionDAG *DAG, |
| 2926 | PPCDAGToDAGISel *Sel) : CurDAG(DAG), S(Sel) { |
| 2927 | assert(CurDAG->getTargetLoweringInfo() |
| 2928 | .getPointerTy(CurDAG->getDataLayout()).getSizeInBits() == 64 && |
| 2929 | "Only expecting to use this on 64 bit targets." ); |
| 2930 | } |
| 2931 | SDNode *Select(SDNode *N) { |
| 2932 | if (CmpInGPR == ICGPR_None) |
| 2933 | return nullptr; |
| 2934 | switch (N->getOpcode()) { |
| 2935 | default: break; |
| 2936 | case ISD::ZERO_EXTEND: |
| 2937 | if (CmpInGPR == ICGPR_Sext || CmpInGPR == ICGPR_SextI32 || |
| 2938 | CmpInGPR == ICGPR_SextI64) |
| 2939 | return nullptr; |
| 2940 | [[fallthrough]]; |
| 2941 | case ISD::SIGN_EXTEND: |
| 2942 | if (CmpInGPR == ICGPR_Zext || CmpInGPR == ICGPR_ZextI32 || |
| 2943 | CmpInGPR == ICGPR_ZextI64) |
| 2944 | return nullptr; |
| 2945 | return tryEXTEND(N); |
| 2946 | case ISD::AND: |
| 2947 | case ISD::OR: |
| 2948 | case ISD::XOR: |
| 2949 | return tryLogicOpOfCompares(N); |
| 2950 | } |
| 2951 | return nullptr; |
| 2952 | } |
| 2953 | }; |
| 2954 | |
| 2955 | // The obvious case for wanting to keep the value in a GPR. Namely, the |
| 2956 | // result of the comparison is actually needed in a GPR. |
| 2957 | SDNode *IntegerCompareEliminator::tryEXTEND(SDNode *N) { |
| 2958 | assert((N->getOpcode() == ISD::ZERO_EXTEND || |
| 2959 | N->getOpcode() == ISD::SIGN_EXTEND) && |
| 2960 | "Expecting a zero/sign extend node!" ); |
| 2961 | SDValue WideRes; |
| 2962 | // If we are zero-extending the result of a logical operation on i1 |
| 2963 | // values, we can keep the values in GPRs. |
| 2964 | if (ISD::isBitwiseLogicOp(Opcode: N->getOperand(Num: 0).getOpcode()) && |
| 2965 | N->getOperand(Num: 0).getValueType() == MVT::i1 && |
| 2966 | N->getOpcode() == ISD::ZERO_EXTEND) |
| 2967 | WideRes = computeLogicOpInGPR(LogicOp: N->getOperand(Num: 0)); |
| 2968 | else if (N->getOperand(Num: 0).getOpcode() != ISD::SETCC) |
| 2969 | return nullptr; |
| 2970 | else |
| 2971 | WideRes = |
| 2972 | getSETCCInGPR(Compare: N->getOperand(Num: 0), |
| 2973 | ConvOpts: N->getOpcode() == ISD::SIGN_EXTEND ? |
| 2974 | SetccInGPROpts::SExtOrig : SetccInGPROpts::ZExtOrig); |
| 2975 | |
| 2976 | if (!WideRes) |
| 2977 | return nullptr; |
| 2978 | |
| 2979 | bool Input32Bit = WideRes.getValueType() == MVT::i32; |
| 2980 | bool Output32Bit = N->getValueType(ResNo: 0) == MVT::i32; |
| 2981 | |
| 2982 | NumSextSetcc += N->getOpcode() == ISD::SIGN_EXTEND ? 1 : 0; |
| 2983 | NumZextSetcc += N->getOpcode() == ISD::SIGN_EXTEND ? 0 : 1; |
| 2984 | |
| 2985 | SDValue ConvOp = WideRes; |
| 2986 | if (Input32Bit != Output32Bit) |
| 2987 | ConvOp = addExtOrTrunc(NatWidthRes: WideRes, Conv: Input32Bit ? ExtOrTruncConversion::Ext : |
| 2988 | ExtOrTruncConversion::Trunc); |
| 2989 | return ConvOp.getNode(); |
| 2990 | } |
| 2991 | |
| 2992 | // Attempt to perform logical operations on the results of comparisons while |
| 2993 | // keeping the values in GPRs. Without doing so, these would end up being |
| 2994 | // lowered to CR-logical operations which suffer from significant latency and |
| 2995 | // low ILP. |
| 2996 | SDNode *IntegerCompareEliminator::tryLogicOpOfCompares(SDNode *N) { |
| 2997 | if (N->getValueType(ResNo: 0) != MVT::i1) |
| 2998 | return nullptr; |
| 2999 | assert(ISD::isBitwiseLogicOp(N->getOpcode()) && |
| 3000 | "Expected a logic operation on setcc results." ); |
| 3001 | SDValue LoweredLogical = computeLogicOpInGPR(LogicOp: SDValue(N, 0)); |
| 3002 | if (!LoweredLogical) |
| 3003 | return nullptr; |
| 3004 | |
| 3005 | SDLoc dl(N); |
| 3006 | bool IsBitwiseNegate = LoweredLogical.getMachineOpcode() == PPC::XORI8; |
| 3007 | unsigned = IsBitwiseNegate ? PPC::sub_eq : PPC::sub_gt; |
| 3008 | SDValue CR0Reg = CurDAG->getRegister(Reg: PPC::CR0, VT: MVT::i32); |
| 3009 | SDValue LHS = LoweredLogical.getOperand(i: 0); |
| 3010 | SDValue RHS = LoweredLogical.getOperand(i: 1); |
| 3011 | SDValue WideOp; |
| 3012 | SDValue OpToConvToRecForm; |
| 3013 | |
| 3014 | // Look through any 32-bit to 64-bit implicit extend nodes to find the |
| 3015 | // opcode that is input to the XORI. |
| 3016 | if (IsBitwiseNegate && |
| 3017 | LoweredLogical.getOperand(i: 0).getMachineOpcode() == PPC::INSERT_SUBREG) |
| 3018 | OpToConvToRecForm = LoweredLogical.getOperand(i: 0).getOperand(i: 1); |
| 3019 | else if (IsBitwiseNegate) |
| 3020 | // If the input to the XORI isn't an extension, that's what we're after. |
| 3021 | OpToConvToRecForm = LoweredLogical.getOperand(i: 0); |
| 3022 | else |
| 3023 | // If this is not an XORI, it is a reg-reg logical op and we can convert |
| 3024 | // it to record-form. |
| 3025 | OpToConvToRecForm = LoweredLogical; |
| 3026 | |
| 3027 | // Get the record-form version of the node we're looking to use to get the |
| 3028 | // CR result from. |
| 3029 | uint16_t NonRecOpc = OpToConvToRecForm.getMachineOpcode(); |
| 3030 | int NewOpc = PPCInstrInfo::getRecordFormOpcode(Opcode: NonRecOpc); |
| 3031 | |
| 3032 | // Convert the right node to record-form. This is either the logical we're |
| 3033 | // looking at or it is the input node to the negation (if we're looking at |
| 3034 | // a bitwise negation). |
| 3035 | if (NewOpc != -1 && IsBitwiseNegate) { |
| 3036 | // The input to the XORI has a record-form. Use it. |
| 3037 | assert(LoweredLogical.getConstantOperandVal(1) == 1 && |
| 3038 | "Expected a PPC::XORI8 only for bitwise negation." ); |
| 3039 | // Emit the record-form instruction. |
| 3040 | std::vector<SDValue> Ops; |
| 3041 | for (int i = 0, e = OpToConvToRecForm.getNumOperands(); i < e; i++) |
| 3042 | Ops.push_back(x: OpToConvToRecForm.getOperand(i)); |
| 3043 | |
| 3044 | WideOp = |
| 3045 | SDValue(CurDAG->getMachineNode(Opcode: NewOpc, dl, |
| 3046 | VT1: OpToConvToRecForm.getValueType(), |
| 3047 | VT2: MVT::Glue, Ops), 0); |
| 3048 | } else { |
| 3049 | assert((NewOpc != -1 || !IsBitwiseNegate) && |
| 3050 | "No record form available for AND8/OR8/XOR8?" ); |
| 3051 | WideOp = |
| 3052 | SDValue(CurDAG->getMachineNode(Opcode: NewOpc == -1 ? PPC::ANDI8_rec : NewOpc, |
| 3053 | dl, VT1: MVT::i64, VT2: MVT::Glue, Op1: LHS, Op2: RHS), |
| 3054 | 0); |
| 3055 | } |
| 3056 | |
| 3057 | // Select this node to a single bit from CR0 set by the record-form node |
| 3058 | // just created. For bitwise negation, use the EQ bit which is the equivalent |
| 3059 | // of negating the result (i.e. it is a bit set when the result of the |
| 3060 | // operation is zero). |
| 3061 | SDValue SRIdxVal = |
| 3062 | CurDAG->getTargetConstant(Val: SubRegToExtract, DL: dl, VT: MVT::i32); |
| 3063 | SDValue CRBit = |
| 3064 | SDValue(CurDAG->getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, dl, |
| 3065 | VT: MVT::i1, Op1: CR0Reg, Op2: SRIdxVal, |
| 3066 | Op3: WideOp.getValue(R: 1)), 0); |
| 3067 | return CRBit.getNode(); |
| 3068 | } |
| 3069 | |
| 3070 | // Lower a logical operation on i1 values into a GPR sequence if possible. |
| 3071 | // The result can be kept in a GPR if requested. |
| 3072 | // Three types of inputs can be handled: |
| 3073 | // - SETCC |
| 3074 | // - TRUNCATE |
| 3075 | // - Logical operation (AND/OR/XOR) |
| 3076 | // There is also a special case that is handled (namely a complement operation |
| 3077 | // achieved with xor %a, -1). |
| 3078 | SDValue IntegerCompareEliminator::computeLogicOpInGPR(SDValue LogicOp) { |
| 3079 | assert(ISD::isBitwiseLogicOp(LogicOp.getOpcode()) && |
| 3080 | "Can only handle logic operations here." ); |
| 3081 | assert(LogicOp.getValueType() == MVT::i1 && |
| 3082 | "Can only handle logic operations on i1 values here." ); |
| 3083 | SDLoc dl(LogicOp); |
| 3084 | SDValue LHS, RHS; |
| 3085 | |
| 3086 | // Special case: xor %a, -1 |
| 3087 | bool IsBitwiseNegation = isBitwiseNot(V: LogicOp); |
| 3088 | |
| 3089 | // Produces a GPR sequence for each operand of the binary logic operation. |
| 3090 | // For SETCC, it produces the respective comparison, for TRUNCATE it truncates |
| 3091 | // the value in a GPR and for logic operations, it will recursively produce |
| 3092 | // a GPR sequence for the operation. |
| 3093 | auto getLogicOperand = [&] (SDValue Operand) -> SDValue { |
| 3094 | unsigned OperandOpcode = Operand.getOpcode(); |
| 3095 | if (OperandOpcode == ISD::SETCC) |
| 3096 | return getSETCCInGPR(Compare: Operand, ConvOpts: SetccInGPROpts::ZExtOrig); |
| 3097 | else if (OperandOpcode == ISD::TRUNCATE) { |
| 3098 | SDValue InputOp = Operand.getOperand(i: 0); |
| 3099 | EVT InVT = InputOp.getValueType(); |
| 3100 | return SDValue(CurDAG->getMachineNode(Opcode: InVT == MVT::i32 ? PPC::RLDICL_32 : |
| 3101 | PPC::RLDICL, dl, VT: InVT, Op1: InputOp, |
| 3102 | Op2: S->getI64Imm(Imm: 0, dl), |
| 3103 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3104 | } else if (ISD::isBitwiseLogicOp(Opcode: OperandOpcode)) |
| 3105 | return computeLogicOpInGPR(LogicOp: Operand); |
| 3106 | return SDValue(); |
| 3107 | }; |
| 3108 | LHS = getLogicOperand(LogicOp.getOperand(i: 0)); |
| 3109 | RHS = getLogicOperand(LogicOp.getOperand(i: 1)); |
| 3110 | |
| 3111 | // If a GPR sequence can't be produced for the LHS we can't proceed. |
| 3112 | // Not producing a GPR sequence for the RHS is only a problem if this isn't |
| 3113 | // a bitwise negation operation. |
| 3114 | if (!LHS || (!RHS && !IsBitwiseNegation)) |
| 3115 | return SDValue(); |
| 3116 | |
| 3117 | NumLogicOpsOnComparison++; |
| 3118 | |
| 3119 | // We will use the inputs as 64-bit values. |
| 3120 | if (LHS.getValueType() == MVT::i32) |
| 3121 | LHS = addExtOrTrunc(NatWidthRes: LHS, Conv: ExtOrTruncConversion::Ext); |
| 3122 | if (!IsBitwiseNegation && RHS.getValueType() == MVT::i32) |
| 3123 | RHS = addExtOrTrunc(NatWidthRes: RHS, Conv: ExtOrTruncConversion::Ext); |
| 3124 | |
| 3125 | unsigned NewOpc; |
| 3126 | switch (LogicOp.getOpcode()) { |
| 3127 | default: llvm_unreachable("Unknown logic operation." ); |
| 3128 | case ISD::AND: NewOpc = PPC::AND8; break; |
| 3129 | case ISD::OR: NewOpc = PPC::OR8; break; |
| 3130 | case ISD::XOR: NewOpc = PPC::XOR8; break; |
| 3131 | } |
| 3132 | |
| 3133 | if (IsBitwiseNegation) { |
| 3134 | RHS = S->getI64Imm(Imm: 1, dl); |
| 3135 | NewOpc = PPC::XORI8; |
| 3136 | } |
| 3137 | |
| 3138 | return SDValue(CurDAG->getMachineNode(Opcode: NewOpc, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3139 | |
| 3140 | } |
| 3141 | |
| 3142 | /// If the value isn't guaranteed to be sign-extended to 64-bits, extend it. |
| 3143 | /// Otherwise just reinterpret it as a 64-bit value. |
| 3144 | /// Useful when emitting comparison code for 32-bit values without using |
| 3145 | /// the compare instruction (which only considers the lower 32-bits). |
| 3146 | SDValue IntegerCompareEliminator::signExtendInputIfNeeded(SDValue Input) { |
| 3147 | assert(Input.getValueType() == MVT::i32 && |
| 3148 | "Can only sign-extend 32-bit values here." ); |
| 3149 | unsigned Opc = Input.getOpcode(); |
| 3150 | |
| 3151 | // The value was sign extended and then truncated to 32-bits. No need to |
| 3152 | // sign extend it again. |
| 3153 | if (Opc == ISD::TRUNCATE && |
| 3154 | (Input.getOperand(i: 0).getOpcode() == ISD::AssertSext || |
| 3155 | Input.getOperand(i: 0).getOpcode() == ISD::SIGN_EXTEND)) |
| 3156 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3157 | |
| 3158 | LoadSDNode *InputLoad = dyn_cast<LoadSDNode>(Val&: Input); |
| 3159 | // The input is a sign-extending load. All ppc sign-extending loads |
| 3160 | // sign-extend to the full 64-bits. |
| 3161 | if (InputLoad && InputLoad->getExtensionType() == ISD::SEXTLOAD) |
| 3162 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3163 | |
| 3164 | ConstantSDNode *InputConst = dyn_cast<ConstantSDNode>(Val&: Input); |
| 3165 | // We don't sign-extend constants. |
| 3166 | if (InputConst) |
| 3167 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3168 | |
| 3169 | SDLoc dl(Input); |
| 3170 | SignExtensionsAdded++; |
| 3171 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::EXTSW_32_64, dl, |
| 3172 | VT: MVT::i64, Op1: Input), 0); |
| 3173 | } |
| 3174 | |
| 3175 | /// If the value isn't guaranteed to be zero-extended to 64-bits, extend it. |
| 3176 | /// Otherwise just reinterpret it as a 64-bit value. |
| 3177 | /// Useful when emitting comparison code for 32-bit values without using |
| 3178 | /// the compare instruction (which only considers the lower 32-bits). |
| 3179 | SDValue IntegerCompareEliminator::zeroExtendInputIfNeeded(SDValue Input) { |
| 3180 | assert(Input.getValueType() == MVT::i32 && |
| 3181 | "Can only zero-extend 32-bit values here." ); |
| 3182 | unsigned Opc = Input.getOpcode(); |
| 3183 | |
| 3184 | // The only condition under which we can omit the actual extend instruction: |
| 3185 | // - The value is a positive constant |
| 3186 | // - The value comes from a load that isn't a sign-extending load |
| 3187 | // An ISD::TRUNCATE needs to be zero-extended unless it is fed by a zext. |
| 3188 | bool IsTruncateOfZExt = Opc == ISD::TRUNCATE && |
| 3189 | (Input.getOperand(i: 0).getOpcode() == ISD::AssertZext || |
| 3190 | Input.getOperand(i: 0).getOpcode() == ISD::ZERO_EXTEND); |
| 3191 | if (IsTruncateOfZExt) |
| 3192 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3193 | |
| 3194 | ConstantSDNode *InputConst = dyn_cast<ConstantSDNode>(Val&: Input); |
| 3195 | if (InputConst && InputConst->getSExtValue() >= 0) |
| 3196 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3197 | |
| 3198 | LoadSDNode *InputLoad = dyn_cast<LoadSDNode>(Val&: Input); |
| 3199 | // The input is a load that doesn't sign-extend (it will be zero-extended). |
| 3200 | if (InputLoad && InputLoad->getExtensionType() != ISD::SEXTLOAD) |
| 3201 | return addExtOrTrunc(NatWidthRes: Input, Conv: ExtOrTruncConversion::Ext); |
| 3202 | |
| 3203 | // None of the above, need to zero-extend. |
| 3204 | SDLoc dl(Input); |
| 3205 | ZeroExtensionsAdded++; |
| 3206 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL_32_64, dl, VT: MVT::i64, Op1: Input, |
| 3207 | Op2: S->getI64Imm(Imm: 0, dl), |
| 3208 | Op3: S->getI64Imm(Imm: 32, dl)), 0); |
| 3209 | } |
| 3210 | |
| 3211 | // Handle a 32-bit value in a 64-bit register and vice-versa. These are of |
| 3212 | // course not actual zero/sign extensions that will generate machine code, |
| 3213 | // they're just a way to reinterpret a 32 bit value in a register as a |
| 3214 | // 64 bit value and vice-versa. |
| 3215 | SDValue IntegerCompareEliminator::addExtOrTrunc(SDValue NatWidthRes, |
| 3216 | ExtOrTruncConversion Conv) { |
| 3217 | SDLoc dl(NatWidthRes); |
| 3218 | |
| 3219 | // For reinterpreting 32-bit values as 64 bit values, we generate |
| 3220 | // INSERT_SUBREG IMPLICIT_DEF:i64, <input>, TargetConstant:i32<1> |
| 3221 | if (Conv == ExtOrTruncConversion::Ext) { |
| 3222 | SDValue ImDef(CurDAG->getMachineNode(Opcode: PPC::IMPLICIT_DEF, dl, VT: MVT::i64), 0); |
| 3223 | SDValue SubRegIdx = |
| 3224 | CurDAG->getTargetConstant(Val: PPC::sub_32, DL: dl, VT: MVT::i32); |
| 3225 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::INSERT_SUBREG, dl, VT: MVT::i64, |
| 3226 | Op1: ImDef, Op2: NatWidthRes, Op3: SubRegIdx), 0); |
| 3227 | } |
| 3228 | |
| 3229 | assert(Conv == ExtOrTruncConversion::Trunc && |
| 3230 | "Unknown convertion between 32 and 64 bit values." ); |
| 3231 | // For reinterpreting 64-bit values as 32-bit values, we just need to |
| 3232 | // EXTRACT_SUBREG (i.e. extract the low word). |
| 3233 | SDValue SubRegIdx = |
| 3234 | CurDAG->getTargetConstant(Val: PPC::sub_32, DL: dl, VT: MVT::i32); |
| 3235 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::EXTRACT_SUBREG, dl, VT: MVT::i32, |
| 3236 | Op1: NatWidthRes, Op2: SubRegIdx), 0); |
| 3237 | } |
| 3238 | |
| 3239 | // Produce a GPR sequence for compound comparisons (<=, >=) against zero. |
| 3240 | // Handle both zero-extensions and sign-extensions. |
| 3241 | SDValue |
| 3242 | IntegerCompareEliminator::getCompoundZeroComparisonInGPR(SDValue LHS, SDLoc dl, |
| 3243 | ZeroCompare CmpTy) { |
| 3244 | EVT InVT = LHS.getValueType(); |
| 3245 | bool Is32Bit = InVT == MVT::i32; |
| 3246 | SDValue ToExtend; |
| 3247 | |
| 3248 | // Produce the value that needs to be either zero or sign extended. |
| 3249 | switch (CmpTy) { |
| 3250 | case ZeroCompare::GEZExt: |
| 3251 | case ZeroCompare::GESExt: |
| 3252 | ToExtend = SDValue(CurDAG->getMachineNode(Opcode: Is32Bit ? PPC::NOR : PPC::NOR8, |
| 3253 | dl, VT: InVT, Op1: LHS, Op2: LHS), 0); |
| 3254 | break; |
| 3255 | case ZeroCompare::LEZExt: |
| 3256 | case ZeroCompare::LESExt: { |
| 3257 | if (Is32Bit) { |
| 3258 | // Upper 32 bits cannot be undefined for this sequence. |
| 3259 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3260 | SDValue Neg = |
| 3261 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, Op1: LHS), 0); |
| 3262 | ToExtend = |
| 3263 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3264 | Op1: Neg, Op2: S->getI64Imm(Imm: 1, dl), |
| 3265 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3266 | } else { |
| 3267 | SDValue Addi = |
| 3268 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: LHS, |
| 3269 | Op2: S->getI64Imm(Imm: ~0ULL, dl)), 0); |
| 3270 | ToExtend = SDValue(CurDAG->getMachineNode(Opcode: PPC::OR8, dl, VT: MVT::i64, |
| 3271 | Op1: Addi, Op2: LHS), 0); |
| 3272 | } |
| 3273 | break; |
| 3274 | } |
| 3275 | } |
| 3276 | |
| 3277 | // For 64-bit sequences, the extensions are the same for the GE/LE cases. |
| 3278 | if (!Is32Bit && |
| 3279 | (CmpTy == ZeroCompare::GEZExt || CmpTy == ZeroCompare::LEZExt)) |
| 3280 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3281 | Op1: ToExtend, Op2: S->getI64Imm(Imm: 1, dl), |
| 3282 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3283 | if (!Is32Bit && |
| 3284 | (CmpTy == ZeroCompare::GESExt || CmpTy == ZeroCompare::LESExt)) |
| 3285 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: ToExtend, |
| 3286 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3287 | |
| 3288 | assert(Is32Bit && "Should have handled the 32-bit sequences above." ); |
| 3289 | // For 32-bit sequences, the extensions differ between GE/LE cases. |
| 3290 | switch (CmpTy) { |
| 3291 | case ZeroCompare::GEZExt: { |
| 3292 | SDValue ShiftOps[] = { ToExtend, S->getI32Imm(Imm: 1, dl), S->getI32Imm(Imm: 31, dl), |
| 3293 | S->getI32Imm(Imm: 31, dl) }; |
| 3294 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, |
| 3295 | Ops: ShiftOps), 0); |
| 3296 | } |
| 3297 | case ZeroCompare::GESExt: |
| 3298 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRAWI, dl, VT: MVT::i32, Op1: ToExtend, |
| 3299 | Op2: S->getI32Imm(Imm: 31, dl)), 0); |
| 3300 | case ZeroCompare::LEZExt: |
| 3301 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, VT: MVT::i64, Op1: ToExtend, |
| 3302 | Op2: S->getI32Imm(Imm: 1, dl)), 0); |
| 3303 | case ZeroCompare::LESExt: |
| 3304 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: ToExtend, |
| 3305 | Op2: S->getI32Imm(Imm: -1, dl)), 0); |
| 3306 | } |
| 3307 | |
| 3308 | // The above case covers all the enumerators so it can't have a default clause |
| 3309 | // to avoid compiler warnings. |
| 3310 | llvm_unreachable("Unknown zero-comparison type." ); |
| 3311 | } |
| 3312 | |
| 3313 | /// Produces a zero-extended result of comparing two 32-bit values according to |
| 3314 | /// the passed condition code. |
| 3315 | SDValue |
| 3316 | IntegerCompareEliminator::get32BitZExtCompare(SDValue LHS, SDValue RHS, |
| 3317 | ISD::CondCode CC, |
| 3318 | int64_t RHSValue, SDLoc dl) { |
| 3319 | if (CmpInGPR == ICGPR_I64 || CmpInGPR == ICGPR_SextI64 || |
| 3320 | CmpInGPR == ICGPR_ZextI64 || CmpInGPR == ICGPR_Sext) |
| 3321 | return SDValue(); |
| 3322 | bool IsRHSZero = RHSValue == 0; |
| 3323 | bool IsRHSOne = RHSValue == 1; |
| 3324 | bool IsRHSNegOne = RHSValue == -1LL; |
| 3325 | switch (CC) { |
| 3326 | default: return SDValue(); |
| 3327 | case ISD::SETEQ: { |
| 3328 | // (zext (setcc %a, %b, seteq)) -> (lshr (cntlzw (xor %a, %b)), 5) |
| 3329 | // (zext (setcc %a, 0, seteq)) -> (lshr (cntlzw %a), 5) |
| 3330 | SDValue Xor = IsRHSZero ? LHS : |
| 3331 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR, dl, VT: MVT::i32, Op1: LHS, Op2: RHS), 0); |
| 3332 | SDValue Clz = |
| 3333 | SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZW, dl, VT: MVT::i32, Op1: Xor), 0); |
| 3334 | SDValue ShiftOps[] = { Clz, S->getI32Imm(Imm: 27, dl), S->getI32Imm(Imm: 5, dl), |
| 3335 | S->getI32Imm(Imm: 31, dl) }; |
| 3336 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, |
| 3337 | Ops: ShiftOps), 0); |
| 3338 | } |
| 3339 | case ISD::SETNE: { |
| 3340 | // (zext (setcc %a, %b, setne)) -> (xor (lshr (cntlzw (xor %a, %b)), 5), 1) |
| 3341 | // (zext (setcc %a, 0, setne)) -> (xor (lshr (cntlzw %a), 5), 1) |
| 3342 | SDValue Xor = IsRHSZero ? LHS : |
| 3343 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR, dl, VT: MVT::i32, Op1: LHS, Op2: RHS), 0); |
| 3344 | SDValue Clz = |
| 3345 | SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZW, dl, VT: MVT::i32, Op1: Xor), 0); |
| 3346 | SDValue ShiftOps[] = { Clz, S->getI32Imm(Imm: 27, dl), S->getI32Imm(Imm: 5, dl), |
| 3347 | S->getI32Imm(Imm: 31, dl) }; |
| 3348 | SDValue Shift = |
| 3349 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops: ShiftOps), 0); |
| 3350 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI, dl, VT: MVT::i32, Op1: Shift, |
| 3351 | Op2: S->getI32Imm(Imm: 1, dl)), 0); |
| 3352 | } |
| 3353 | case ISD::SETGE: { |
| 3354 | // (zext (setcc %a, %b, setge)) -> (xor (lshr (sub %a, %b), 63), 1) |
| 3355 | // (zext (setcc %a, 0, setge)) -> (lshr (~ %a), 31) |
| 3356 | if(IsRHSZero) |
| 3357 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GEZExt); |
| 3358 | |
| 3359 | // Not a special case (i.e. RHS == 0). Handle (%a >= %b) as (%b <= %a) |
| 3360 | // by swapping inputs and falling through. |
| 3361 | std::swap(a&: LHS, b&: RHS); |
| 3362 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3363 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3364 | [[fallthrough]]; |
| 3365 | } |
| 3366 | case ISD::SETLE: { |
| 3367 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3368 | return SDValue(); |
| 3369 | // (zext (setcc %a, %b, setle)) -> (xor (lshr (sub %b, %a), 63), 1) |
| 3370 | // (zext (setcc %a, 0, setle)) -> (xor (lshr (- %a), 63), 1) |
| 3371 | if(IsRHSZero) { |
| 3372 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3373 | return SDValue(); |
| 3374 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LEZExt); |
| 3375 | } |
| 3376 | |
| 3377 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3378 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3379 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3380 | SDValue Sub = |
| 3381 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3382 | SDValue Shift = |
| 3383 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: Sub, |
| 3384 | Op2: S->getI64Imm(Imm: 1, dl), Op3: S->getI64Imm(Imm: 63, dl)), |
| 3385 | 0); |
| 3386 | return |
| 3387 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, |
| 3388 | VT: MVT::i64, Op1: Shift, Op2: S->getI32Imm(Imm: 1, dl)), 0); |
| 3389 | } |
| 3390 | case ISD::SETGT: { |
| 3391 | // (zext (setcc %a, %b, setgt)) -> (lshr (sub %b, %a), 63) |
| 3392 | // (zext (setcc %a, -1, setgt)) -> (lshr (~ %a), 31) |
| 3393 | // (zext (setcc %a, 0, setgt)) -> (lshr (- %a), 63) |
| 3394 | // Handle SETLT -1 (which is equivalent to SETGE 0). |
| 3395 | if (IsRHSNegOne) |
| 3396 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GEZExt); |
| 3397 | |
| 3398 | if (IsRHSZero) { |
| 3399 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3400 | return SDValue(); |
| 3401 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3402 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3403 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3404 | SDValue Neg = |
| 3405 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, Op1: LHS), 0); |
| 3406 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3407 | Op1: Neg, Op2: S->getI32Imm(Imm: 1, dl), Op3: S->getI32Imm(Imm: 63, dl)), 0); |
| 3408 | } |
| 3409 | // Not a special case (i.e. RHS == 0 or RHS == -1). Handle (%a > %b) as |
| 3410 | // (%b < %a) by swapping inputs and falling through. |
| 3411 | std::swap(a&: LHS, b&: RHS); |
| 3412 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3413 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3414 | IsRHSOne = RHSConst && RHSConst->getSExtValue() == 1; |
| 3415 | [[fallthrough]]; |
| 3416 | } |
| 3417 | case ISD::SETLT: { |
| 3418 | // (zext (setcc %a, %b, setlt)) -> (lshr (sub %a, %b), 63) |
| 3419 | // (zext (setcc %a, 1, setlt)) -> (xor (lshr (- %a), 63), 1) |
| 3420 | // (zext (setcc %a, 0, setlt)) -> (lshr %a, 31) |
| 3421 | // Handle SETLT 1 (which is equivalent to SETLE 0). |
| 3422 | if (IsRHSOne) { |
| 3423 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3424 | return SDValue(); |
| 3425 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LEZExt); |
| 3426 | } |
| 3427 | |
| 3428 | if (IsRHSZero) { |
| 3429 | SDValue ShiftOps[] = { LHS, S->getI32Imm(Imm: 1, dl), S->getI32Imm(Imm: 31, dl), |
| 3430 | S->getI32Imm(Imm: 31, dl) }; |
| 3431 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, |
| 3432 | Ops: ShiftOps), 0); |
| 3433 | } |
| 3434 | |
| 3435 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3436 | return SDValue(); |
| 3437 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3438 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3439 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3440 | SDValue SUBFNode = |
| 3441 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: RHS, Op2: LHS), 0); |
| 3442 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3443 | Op1: SUBFNode, Op2: S->getI64Imm(Imm: 1, dl), |
| 3444 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3445 | } |
| 3446 | case ISD::SETUGE: |
| 3447 | // (zext (setcc %a, %b, setuge)) -> (xor (lshr (sub %b, %a), 63), 1) |
| 3448 | // (zext (setcc %a, %b, setule)) -> (xor (lshr (sub %a, %b), 63), 1) |
| 3449 | std::swap(a&: LHS, b&: RHS); |
| 3450 | [[fallthrough]]; |
| 3451 | case ISD::SETULE: { |
| 3452 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3453 | return SDValue(); |
| 3454 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3455 | LHS = zeroExtendInputIfNeeded(Input: LHS); |
| 3456 | RHS = zeroExtendInputIfNeeded(Input: RHS); |
| 3457 | SDValue Subtract = |
| 3458 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3459 | SDValue SrdiNode = |
| 3460 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3461 | Op1: Subtract, Op2: S->getI64Imm(Imm: 1, dl), |
| 3462 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3463 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, VT: MVT::i64, Op1: SrdiNode, |
| 3464 | Op2: S->getI32Imm(Imm: 1, dl)), 0); |
| 3465 | } |
| 3466 | case ISD::SETUGT: |
| 3467 | // (zext (setcc %a, %b, setugt)) -> (lshr (sub %b, %a), 63) |
| 3468 | // (zext (setcc %a, %b, setult)) -> (lshr (sub %a, %b), 63) |
| 3469 | std::swap(a&: LHS, b&: RHS); |
| 3470 | [[fallthrough]]; |
| 3471 | case ISD::SETULT: { |
| 3472 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3473 | return SDValue(); |
| 3474 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3475 | LHS = zeroExtendInputIfNeeded(Input: LHS); |
| 3476 | RHS = zeroExtendInputIfNeeded(Input: RHS); |
| 3477 | SDValue Subtract = |
| 3478 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: RHS, Op2: LHS), 0); |
| 3479 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3480 | Op1: Subtract, Op2: S->getI64Imm(Imm: 1, dl), |
| 3481 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3482 | } |
| 3483 | } |
| 3484 | } |
| 3485 | |
| 3486 | /// Produces a sign-extended result of comparing two 32-bit values according to |
| 3487 | /// the passed condition code. |
| 3488 | SDValue |
| 3489 | IntegerCompareEliminator::get32BitSExtCompare(SDValue LHS, SDValue RHS, |
| 3490 | ISD::CondCode CC, |
| 3491 | int64_t RHSValue, SDLoc dl) { |
| 3492 | if (CmpInGPR == ICGPR_I64 || CmpInGPR == ICGPR_SextI64 || |
| 3493 | CmpInGPR == ICGPR_ZextI64 || CmpInGPR == ICGPR_Zext) |
| 3494 | return SDValue(); |
| 3495 | bool IsRHSZero = RHSValue == 0; |
| 3496 | bool IsRHSOne = RHSValue == 1; |
| 3497 | bool IsRHSNegOne = RHSValue == -1LL; |
| 3498 | |
| 3499 | switch (CC) { |
| 3500 | default: return SDValue(); |
| 3501 | case ISD::SETEQ: { |
| 3502 | // (sext (setcc %a, %b, seteq)) -> |
| 3503 | // (ashr (shl (ctlz (xor %a, %b)), 58), 63) |
| 3504 | // (sext (setcc %a, 0, seteq)) -> |
| 3505 | // (ashr (shl (ctlz %a), 58), 63) |
| 3506 | SDValue CountInput = IsRHSZero ? LHS : |
| 3507 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR, dl, VT: MVT::i32, Op1: LHS, Op2: RHS), 0); |
| 3508 | SDValue Cntlzw = |
| 3509 | SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZW, dl, VT: MVT::i32, Op1: CountInput), 0); |
| 3510 | SDValue SHLOps[] = { Cntlzw, S->getI32Imm(Imm: 27, dl), |
| 3511 | S->getI32Imm(Imm: 5, dl), S->getI32Imm(Imm: 31, dl) }; |
| 3512 | SDValue Slwi = |
| 3513 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops: SHLOps), 0); |
| 3514 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG, dl, VT: MVT::i32, Op1: Slwi), 0); |
| 3515 | } |
| 3516 | case ISD::SETNE: { |
| 3517 | // Bitwise xor the operands, count leading zeros, shift right by 5 bits and |
| 3518 | // flip the bit, finally take 2's complement. |
| 3519 | // (sext (setcc %a, %b, setne)) -> |
| 3520 | // (neg (xor (lshr (ctlz (xor %a, %b)), 5), 1)) |
| 3521 | // Same as above, but the first xor is not needed. |
| 3522 | // (sext (setcc %a, 0, setne)) -> |
| 3523 | // (neg (xor (lshr (ctlz %a), 5), 1)) |
| 3524 | SDValue Xor = IsRHSZero ? LHS : |
| 3525 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR, dl, VT: MVT::i32, Op1: LHS, Op2: RHS), 0); |
| 3526 | SDValue Clz = |
| 3527 | SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZW, dl, VT: MVT::i32, Op1: Xor), 0); |
| 3528 | SDValue ShiftOps[] = |
| 3529 | { Clz, S->getI32Imm(Imm: 27, dl), S->getI32Imm(Imm: 5, dl), S->getI32Imm(Imm: 31, dl) }; |
| 3530 | SDValue Shift = |
| 3531 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops: ShiftOps), 0); |
| 3532 | SDValue Xori = |
| 3533 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI, dl, VT: MVT::i32, Op1: Shift, |
| 3534 | Op2: S->getI32Imm(Imm: 1, dl)), 0); |
| 3535 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG, dl, VT: MVT::i32, Op1: Xori), 0); |
| 3536 | } |
| 3537 | case ISD::SETGE: { |
| 3538 | // (sext (setcc %a, %b, setge)) -> (add (lshr (sub %a, %b), 63), -1) |
| 3539 | // (sext (setcc %a, 0, setge)) -> (ashr (~ %a), 31) |
| 3540 | if (IsRHSZero) |
| 3541 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GESExt); |
| 3542 | |
| 3543 | // Not a special case (i.e. RHS == 0). Handle (%a >= %b) as (%b <= %a) |
| 3544 | // by swapping inputs and falling through. |
| 3545 | std::swap(a&: LHS, b&: RHS); |
| 3546 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3547 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3548 | [[fallthrough]]; |
| 3549 | } |
| 3550 | case ISD::SETLE: { |
| 3551 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3552 | return SDValue(); |
| 3553 | // (sext (setcc %a, %b, setge)) -> (add (lshr (sub %b, %a), 63), -1) |
| 3554 | // (sext (setcc %a, 0, setle)) -> (add (lshr (- %a), 63), -1) |
| 3555 | if (IsRHSZero) |
| 3556 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LESExt); |
| 3557 | |
| 3558 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3559 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3560 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3561 | SDValue SUBFNode = |
| 3562 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3563 | Op1: LHS, Op2: RHS), 0); |
| 3564 | SDValue Srdi = |
| 3565 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3566 | Op1: SUBFNode, Op2: S->getI64Imm(Imm: 1, dl), |
| 3567 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3568 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: Srdi, |
| 3569 | Op2: S->getI32Imm(Imm: -1, dl)), 0); |
| 3570 | } |
| 3571 | case ISD::SETGT: { |
| 3572 | // (sext (setcc %a, %b, setgt)) -> (ashr (sub %b, %a), 63) |
| 3573 | // (sext (setcc %a, -1, setgt)) -> (ashr (~ %a), 31) |
| 3574 | // (sext (setcc %a, 0, setgt)) -> (ashr (- %a), 63) |
| 3575 | if (IsRHSNegOne) |
| 3576 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GESExt); |
| 3577 | if (IsRHSZero) { |
| 3578 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3579 | return SDValue(); |
| 3580 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3581 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3582 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3583 | SDValue Neg = |
| 3584 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, Op1: LHS), 0); |
| 3585 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: Neg, |
| 3586 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3587 | } |
| 3588 | // Not a special case (i.e. RHS == 0 or RHS == -1). Handle (%a > %b) as |
| 3589 | // (%b < %a) by swapping inputs and falling through. |
| 3590 | std::swap(a&: LHS, b&: RHS); |
| 3591 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3592 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3593 | IsRHSOne = RHSConst && RHSConst->getSExtValue() == 1; |
| 3594 | [[fallthrough]]; |
| 3595 | } |
| 3596 | case ISD::SETLT: { |
| 3597 | // (sext (setcc %a, %b, setgt)) -> (ashr (sub %a, %b), 63) |
| 3598 | // (sext (setcc %a, 1, setgt)) -> (add (lshr (- %a), 63), -1) |
| 3599 | // (sext (setcc %a, 0, setgt)) -> (ashr %a, 31) |
| 3600 | if (IsRHSOne) { |
| 3601 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3602 | return SDValue(); |
| 3603 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LESExt); |
| 3604 | } |
| 3605 | if (IsRHSZero) |
| 3606 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRAWI, dl, VT: MVT::i32, Op1: LHS, |
| 3607 | Op2: S->getI32Imm(Imm: 31, dl)), 0); |
| 3608 | |
| 3609 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3610 | return SDValue(); |
| 3611 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3612 | LHS = signExtendInputIfNeeded(Input: LHS); |
| 3613 | RHS = signExtendInputIfNeeded(Input: RHS); |
| 3614 | SDValue SUBFNode = |
| 3615 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: RHS, Op2: LHS), 0); |
| 3616 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, |
| 3617 | Op1: SUBFNode, Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3618 | } |
| 3619 | case ISD::SETUGE: |
| 3620 | // (sext (setcc %a, %b, setuge)) -> (add (lshr (sub %a, %b), 63), -1) |
| 3621 | // (sext (setcc %a, %b, setule)) -> (add (lshr (sub %b, %a), 63), -1) |
| 3622 | std::swap(a&: LHS, b&: RHS); |
| 3623 | [[fallthrough]]; |
| 3624 | case ISD::SETULE: { |
| 3625 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3626 | return SDValue(); |
| 3627 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3628 | LHS = zeroExtendInputIfNeeded(Input: LHS); |
| 3629 | RHS = zeroExtendInputIfNeeded(Input: RHS); |
| 3630 | SDValue Subtract = |
| 3631 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3632 | SDValue Shift = |
| 3633 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: Subtract, |
| 3634 | Op2: S->getI32Imm(Imm: 1, dl), Op3: S->getI32Imm(Imm: 63,dl)), |
| 3635 | 0); |
| 3636 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: Shift, |
| 3637 | Op2: S->getI32Imm(Imm: -1, dl)), 0); |
| 3638 | } |
| 3639 | case ISD::SETUGT: |
| 3640 | // (sext (setcc %a, %b, setugt)) -> (ashr (sub %b, %a), 63) |
| 3641 | // (sext (setcc %a, %b, setugt)) -> (ashr (sub %a, %b), 63) |
| 3642 | std::swap(a&: LHS, b&: RHS); |
| 3643 | [[fallthrough]]; |
| 3644 | case ISD::SETULT: { |
| 3645 | if (CmpInGPR == ICGPR_NonExtIn) |
| 3646 | return SDValue(); |
| 3647 | // The upper 32-bits of the register can't be undefined for this sequence. |
| 3648 | LHS = zeroExtendInputIfNeeded(Input: LHS); |
| 3649 | RHS = zeroExtendInputIfNeeded(Input: RHS); |
| 3650 | SDValue Subtract = |
| 3651 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBF8, dl, VT: MVT::i64, Op1: RHS, Op2: LHS), 0); |
| 3652 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, |
| 3653 | Op1: Subtract, Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3654 | } |
| 3655 | } |
| 3656 | } |
| 3657 | |
| 3658 | /// Produces a zero-extended result of comparing two 64-bit values according to |
| 3659 | /// the passed condition code. |
| 3660 | SDValue |
| 3661 | IntegerCompareEliminator::get64BitZExtCompare(SDValue LHS, SDValue RHS, |
| 3662 | ISD::CondCode CC, |
| 3663 | int64_t RHSValue, SDLoc dl) { |
| 3664 | if (CmpInGPR == ICGPR_I32 || CmpInGPR == ICGPR_SextI32 || |
| 3665 | CmpInGPR == ICGPR_ZextI32 || CmpInGPR == ICGPR_Sext) |
| 3666 | return SDValue(); |
| 3667 | bool IsRHSZero = RHSValue == 0; |
| 3668 | bool IsRHSOne = RHSValue == 1; |
| 3669 | bool IsRHSNegOne = RHSValue == -1LL; |
| 3670 | switch (CC) { |
| 3671 | default: return SDValue(); |
| 3672 | case ISD::SETEQ: { |
| 3673 | // (zext (setcc %a, %b, seteq)) -> (lshr (ctlz (xor %a, %b)), 6) |
| 3674 | // (zext (setcc %a, 0, seteq)) -> (lshr (ctlz %a), 6) |
| 3675 | SDValue Xor = IsRHSZero ? LHS : |
| 3676 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3677 | SDValue Clz = |
| 3678 | SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZD, dl, VT: MVT::i64, Op1: Xor), 0); |
| 3679 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: Clz, |
| 3680 | Op2: S->getI64Imm(Imm: 58, dl), |
| 3681 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3682 | } |
| 3683 | case ISD::SETNE: { |
| 3684 | // {addc.reg, addc.CA} = (addcarry (xor %a, %b), -1) |
| 3685 | // (zext (setcc %a, %b, setne)) -> (sube addc.reg, addc.reg, addc.CA) |
| 3686 | // {addcz.reg, addcz.CA} = (addcarry %a, -1) |
| 3687 | // (zext (setcc %a, 0, setne)) -> (sube addcz.reg, addcz.reg, addcz.CA) |
| 3688 | SDValue Xor = IsRHSZero ? LHS : |
| 3689 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3690 | SDValue AC = |
| 3691 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDIC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3692 | Op1: Xor, Op2: S->getI32Imm(Imm: ~0U, dl)), 0); |
| 3693 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT: MVT::i64, Op1: AC, |
| 3694 | Op2: Xor, Op3: AC.getValue(R: 1)), 0); |
| 3695 | } |
| 3696 | case ISD::SETGE: { |
| 3697 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3698 | // (zext (setcc %a, %b, setge)) -> |
| 3699 | // (adde (lshr %b, 63), (ashr %a, 63), subc.CA) |
| 3700 | // (zext (setcc %a, 0, setge)) -> (lshr (~ %a), 63) |
| 3701 | if (IsRHSZero) |
| 3702 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GEZExt); |
| 3703 | std::swap(a&: LHS, b&: RHS); |
| 3704 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3705 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3706 | [[fallthrough]]; |
| 3707 | } |
| 3708 | case ISD::SETLE: { |
| 3709 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3710 | // (zext (setcc %a, %b, setge)) -> |
| 3711 | // (adde (lshr %a, 63), (ashr %b, 63), subc.CA) |
| 3712 | // (zext (setcc %a, 0, setge)) -> (lshr (or %a, (add %a, -1)), 63) |
| 3713 | if (IsRHSZero) |
| 3714 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LEZExt); |
| 3715 | SDValue ShiftL = |
| 3716 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: LHS, |
| 3717 | Op2: S->getI64Imm(Imm: 1, dl), |
| 3718 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3719 | SDValue ShiftR = |
| 3720 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: RHS, |
| 3721 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3722 | SDValue SubtractCarry = |
| 3723 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3724 | Op1: LHS, Op2: RHS), 1); |
| 3725 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDE8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3726 | Op1: ShiftR, Op2: ShiftL, Op3: SubtractCarry), 0); |
| 3727 | } |
| 3728 | case ISD::SETGT: { |
| 3729 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3730 | // (zext (setcc %a, %b, setgt)) -> |
| 3731 | // (xor (adde (lshr %a, 63), (ashr %b, 63), subc.CA), 1) |
| 3732 | // (zext (setcc %a, 0, setgt)) -> (lshr (nor (add %a, -1), %a), 63) |
| 3733 | if (IsRHSNegOne) |
| 3734 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GEZExt); |
| 3735 | if (IsRHSZero) { |
| 3736 | SDValue Addi = |
| 3737 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: LHS, |
| 3738 | Op2: S->getI64Imm(Imm: ~0ULL, dl)), 0); |
| 3739 | SDValue Nor = |
| 3740 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NOR8, dl, VT: MVT::i64, Op1: Addi, Op2: LHS), 0); |
| 3741 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: Nor, |
| 3742 | Op2: S->getI64Imm(Imm: 1, dl), |
| 3743 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3744 | } |
| 3745 | std::swap(a&: LHS, b&: RHS); |
| 3746 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3747 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3748 | IsRHSOne = RHSConst && RHSConst->getSExtValue() == 1; |
| 3749 | [[fallthrough]]; |
| 3750 | } |
| 3751 | case ISD::SETLT: { |
| 3752 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3753 | // (zext (setcc %a, %b, setlt)) -> |
| 3754 | // (xor (adde (lshr %b, 63), (ashr %a, 63), subc.CA), 1) |
| 3755 | // (zext (setcc %a, 0, setlt)) -> (lshr %a, 63) |
| 3756 | if (IsRHSOne) |
| 3757 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LEZExt); |
| 3758 | if (IsRHSZero) |
| 3759 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: LHS, |
| 3760 | Op2: S->getI64Imm(Imm: 1, dl), |
| 3761 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3762 | SDValue SRADINode = |
| 3763 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, |
| 3764 | Op1: LHS, Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3765 | SDValue SRDINode = |
| 3766 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3767 | Op1: RHS, Op2: S->getI64Imm(Imm: 1, dl), |
| 3768 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3769 | SDValue SUBFC8Carry = |
| 3770 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3771 | Op1: RHS, Op2: LHS), 1); |
| 3772 | SDValue ADDE8Node = |
| 3773 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDE8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3774 | Op1: SRDINode, Op2: SRADINode, Op3: SUBFC8Carry), 0); |
| 3775 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, VT: MVT::i64, |
| 3776 | Op1: ADDE8Node, Op2: S->getI64Imm(Imm: 1, dl)), 0); |
| 3777 | } |
| 3778 | case ISD::SETUGE: |
| 3779 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3780 | // (zext (setcc %a, %b, setuge)) -> (add (sube %b, %b, subc.CA), 1) |
| 3781 | std::swap(a&: LHS, b&: RHS); |
| 3782 | [[fallthrough]]; |
| 3783 | case ISD::SETULE: { |
| 3784 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3785 | // (zext (setcc %a, %b, setule)) -> (add (sube %a, %a, subc.CA), 1) |
| 3786 | SDValue SUBFC8Carry = |
| 3787 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3788 | Op1: LHS, Op2: RHS), 1); |
| 3789 | SDValue SUBFE8Node = |
| 3790 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3791 | Op1: LHS, Op2: LHS, Op3: SUBFC8Carry), 0); |
| 3792 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, |
| 3793 | Op1: SUBFE8Node, Op2: S->getI64Imm(Imm: 1, dl)), 0); |
| 3794 | } |
| 3795 | case ISD::SETUGT: |
| 3796 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3797 | // (zext (setcc %a, %b, setugt)) -> -(sube %b, %b, subc.CA) |
| 3798 | std::swap(a&: LHS, b&: RHS); |
| 3799 | [[fallthrough]]; |
| 3800 | case ISD::SETULT: { |
| 3801 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3802 | // (zext (setcc %a, %b, setult)) -> -(sube %a, %a, subc.CA) |
| 3803 | SDValue SubtractCarry = |
| 3804 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3805 | Op1: RHS, Op2: LHS), 1); |
| 3806 | SDValue ExtSub = |
| 3807 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT: MVT::i64, |
| 3808 | Op1: LHS, Op2: LHS, Op3: SubtractCarry), 0); |
| 3809 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, |
| 3810 | Op1: ExtSub), 0); |
| 3811 | } |
| 3812 | } |
| 3813 | } |
| 3814 | |
| 3815 | /// Produces a sign-extended result of comparing two 64-bit values according to |
| 3816 | /// the passed condition code. |
| 3817 | SDValue |
| 3818 | IntegerCompareEliminator::get64BitSExtCompare(SDValue LHS, SDValue RHS, |
| 3819 | ISD::CondCode CC, |
| 3820 | int64_t RHSValue, SDLoc dl) { |
| 3821 | if (CmpInGPR == ICGPR_I32 || CmpInGPR == ICGPR_SextI32 || |
| 3822 | CmpInGPR == ICGPR_ZextI32 || CmpInGPR == ICGPR_Zext) |
| 3823 | return SDValue(); |
| 3824 | bool IsRHSZero = RHSValue == 0; |
| 3825 | bool IsRHSOne = RHSValue == 1; |
| 3826 | bool IsRHSNegOne = RHSValue == -1LL; |
| 3827 | switch (CC) { |
| 3828 | default: return SDValue(); |
| 3829 | case ISD::SETEQ: { |
| 3830 | // {addc.reg, addc.CA} = (addcarry (xor %a, %b), -1) |
| 3831 | // (sext (setcc %a, %b, seteq)) -> (sube addc.reg, addc.reg, addc.CA) |
| 3832 | // {addcz.reg, addcz.CA} = (addcarry %a, -1) |
| 3833 | // (sext (setcc %a, 0, seteq)) -> (sube addcz.reg, addcz.reg, addcz.CA) |
| 3834 | SDValue AddInput = IsRHSZero ? LHS : |
| 3835 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3836 | SDValue Addic = |
| 3837 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDIC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3838 | Op1: AddInput, Op2: S->getI32Imm(Imm: ~0U, dl)), 0); |
| 3839 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT: MVT::i64, Op1: Addic, |
| 3840 | Op2: Addic, Op3: Addic.getValue(R: 1)), 0); |
| 3841 | } |
| 3842 | case ISD::SETNE: { |
| 3843 | // {subfc.reg, subfc.CA} = (subcarry 0, (xor %a, %b)) |
| 3844 | // (sext (setcc %a, %b, setne)) -> (sube subfc.reg, subfc.reg, subfc.CA) |
| 3845 | // {subfcz.reg, subfcz.CA} = (subcarry 0, %a) |
| 3846 | // (sext (setcc %a, 0, setne)) -> (sube subfcz.reg, subfcz.reg, subfcz.CA) |
| 3847 | SDValue Xor = IsRHSZero ? LHS : |
| 3848 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XOR8, dl, VT: MVT::i64, Op1: LHS, Op2: RHS), 0); |
| 3849 | SDValue SC = |
| 3850 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFIC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3851 | Op1: Xor, Op2: S->getI32Imm(Imm: 0, dl)), 0); |
| 3852 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT: MVT::i64, Op1: SC, |
| 3853 | Op2: SC, Op3: SC.getValue(R: 1)), 0); |
| 3854 | } |
| 3855 | case ISD::SETGE: { |
| 3856 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3857 | // (zext (setcc %a, %b, setge)) -> |
| 3858 | // (- (adde (lshr %b, 63), (ashr %a, 63), subc.CA)) |
| 3859 | // (zext (setcc %a, 0, setge)) -> (~ (ashr %a, 63)) |
| 3860 | if (IsRHSZero) |
| 3861 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GESExt); |
| 3862 | std::swap(a&: LHS, b&: RHS); |
| 3863 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3864 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3865 | [[fallthrough]]; |
| 3866 | } |
| 3867 | case ISD::SETLE: { |
| 3868 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3869 | // (zext (setcc %a, %b, setge)) -> |
| 3870 | // (- (adde (lshr %a, 63), (ashr %b, 63), subc.CA)) |
| 3871 | // (zext (setcc %a, 0, setge)) -> (ashr (or %a, (add %a, -1)), 63) |
| 3872 | if (IsRHSZero) |
| 3873 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LESExt); |
| 3874 | SDValue ShiftR = |
| 3875 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: RHS, |
| 3876 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3877 | SDValue ShiftL = |
| 3878 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, Op1: LHS, |
| 3879 | Op2: S->getI64Imm(Imm: 1, dl), |
| 3880 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3881 | SDValue SubtractCarry = |
| 3882 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3883 | Op1: LHS, Op2: RHS), 1); |
| 3884 | SDValue Adde = |
| 3885 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDE8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3886 | Op1: ShiftR, Op2: ShiftL, Op3: SubtractCarry), 0); |
| 3887 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, Op1: Adde), 0); |
| 3888 | } |
| 3889 | case ISD::SETGT: { |
| 3890 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3891 | // (zext (setcc %a, %b, setgt)) -> |
| 3892 | // -(xor (adde (lshr %a, 63), (ashr %b, 63), subc.CA), 1) |
| 3893 | // (zext (setcc %a, 0, setgt)) -> (ashr (nor (add %a, -1), %a), 63) |
| 3894 | if (IsRHSNegOne) |
| 3895 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::GESExt); |
| 3896 | if (IsRHSZero) { |
| 3897 | SDValue Add = |
| 3898 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI8, dl, VT: MVT::i64, Op1: LHS, |
| 3899 | Op2: S->getI64Imm(Imm: -1, dl)), 0); |
| 3900 | SDValue Nor = |
| 3901 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NOR8, dl, VT: MVT::i64, Op1: Add, Op2: LHS), 0); |
| 3902 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: Nor, |
| 3903 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3904 | } |
| 3905 | std::swap(a&: LHS, b&: RHS); |
| 3906 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 3907 | IsRHSZero = RHSConst && RHSConst->isZero(); |
| 3908 | IsRHSOne = RHSConst && RHSConst->getSExtValue() == 1; |
| 3909 | [[fallthrough]]; |
| 3910 | } |
| 3911 | case ISD::SETLT: { |
| 3912 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3913 | // (zext (setcc %a, %b, setlt)) -> |
| 3914 | // -(xor (adde (lshr %b, 63), (ashr %a, 63), subc.CA), 1) |
| 3915 | // (zext (setcc %a, 0, setlt)) -> (ashr %a, 63) |
| 3916 | if (IsRHSOne) |
| 3917 | return getCompoundZeroComparisonInGPR(LHS, dl, CmpTy: ZeroCompare::LESExt); |
| 3918 | if (IsRHSZero) { |
| 3919 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, Op1: LHS, |
| 3920 | Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3921 | } |
| 3922 | SDValue SRADINode = |
| 3923 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT: MVT::i64, |
| 3924 | Op1: LHS, Op2: S->getI64Imm(Imm: 63, dl)), 0); |
| 3925 | SDValue SRDINode = |
| 3926 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl, VT: MVT::i64, |
| 3927 | Op1: RHS, Op2: S->getI64Imm(Imm: 1, dl), |
| 3928 | Op3: S->getI64Imm(Imm: 63, dl)), 0); |
| 3929 | SDValue SUBFC8Carry = |
| 3930 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3931 | Op1: RHS, Op2: LHS), 1); |
| 3932 | SDValue ADDE8Node = |
| 3933 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDE8, dl, VT: MVT::i64, |
| 3934 | Op1: SRDINode, Op2: SRADINode, Op3: SUBFC8Carry), 0); |
| 3935 | SDValue XORI8Node = |
| 3936 | SDValue(CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, VT: MVT::i64, |
| 3937 | Op1: ADDE8Node, Op2: S->getI64Imm(Imm: 1, dl)), 0); |
| 3938 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG8, dl, VT: MVT::i64, |
| 3939 | Op1: XORI8Node), 0); |
| 3940 | } |
| 3941 | case ISD::SETUGE: |
| 3942 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3943 | // (sext (setcc %a, %b, setuge)) -> ~(sube %b, %b, subc.CA) |
| 3944 | std::swap(a&: LHS, b&: RHS); |
| 3945 | [[fallthrough]]; |
| 3946 | case ISD::SETULE: { |
| 3947 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3948 | // (sext (setcc %a, %b, setule)) -> ~(sube %a, %a, subc.CA) |
| 3949 | SDValue SubtractCarry = |
| 3950 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3951 | Op1: LHS, Op2: RHS), 1); |
| 3952 | SDValue ExtSub = |
| 3953 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT1: MVT::i64, VT2: MVT::Glue, Op1: LHS, |
| 3954 | Op2: LHS, Op3: SubtractCarry), 0); |
| 3955 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::NOR8, dl, VT: MVT::i64, |
| 3956 | Op1: ExtSub, Op2: ExtSub), 0); |
| 3957 | } |
| 3958 | case ISD::SETUGT: |
| 3959 | // {subc.reg, subc.CA} = (subcarry %b, %a) |
| 3960 | // (sext (setcc %a, %b, setugt)) -> (sube %b, %b, subc.CA) |
| 3961 | std::swap(a&: LHS, b&: RHS); |
| 3962 | [[fallthrough]]; |
| 3963 | case ISD::SETULT: { |
| 3964 | // {subc.reg, subc.CA} = (subcarry %a, %b) |
| 3965 | // (sext (setcc %a, %b, setult)) -> (sube %a, %a, subc.CA) |
| 3966 | SDValue SubCarry = |
| 3967 | SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFC8, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 3968 | Op1: RHS, Op2: LHS), 1); |
| 3969 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::SUBFE8, dl, VT: MVT::i64, |
| 3970 | Op1: LHS, Op2: LHS, Op3: SubCarry), 0); |
| 3971 | } |
| 3972 | } |
| 3973 | } |
| 3974 | |
| 3975 | /// Do all uses of this SDValue need the result in a GPR? |
| 3976 | /// This is meant to be used on values that have type i1 since |
| 3977 | /// it is somewhat meaningless to ask if values of other types |
| 3978 | /// should be kept in GPR's. |
| 3979 | static bool allUsesExtend(SDValue Compare, SelectionDAG *CurDAG) { |
| 3980 | assert(Compare.getOpcode() == ISD::SETCC && |
| 3981 | "An ISD::SETCC node required here." ); |
| 3982 | |
| 3983 | // For values that have a single use, the caller should obviously already have |
| 3984 | // checked if that use is an extending use. We check the other uses here. |
| 3985 | if (Compare.hasOneUse()) |
| 3986 | return true; |
| 3987 | // We want the value in a GPR if it is being extended, used for a select, or |
| 3988 | // used in logical operations. |
| 3989 | for (auto *CompareUse : Compare.getNode()->users()) |
| 3990 | if (CompareUse->getOpcode() != ISD::SIGN_EXTEND && |
| 3991 | CompareUse->getOpcode() != ISD::ZERO_EXTEND && |
| 3992 | CompareUse->getOpcode() != ISD::SELECT && |
| 3993 | !ISD::isBitwiseLogicOp(Opcode: CompareUse->getOpcode())) { |
| 3994 | OmittedForNonExtendUses++; |
| 3995 | return false; |
| 3996 | } |
| 3997 | return true; |
| 3998 | } |
| 3999 | |
| 4000 | /// Returns an equivalent of a SETCC node but with the result the same width as |
| 4001 | /// the inputs. This can also be used for SELECT_CC if either the true or false |
| 4002 | /// values is a power of two while the other is zero. |
| 4003 | SDValue IntegerCompareEliminator::getSETCCInGPR(SDValue Compare, |
| 4004 | SetccInGPROpts ConvOpts) { |
| 4005 | assert((Compare.getOpcode() == ISD::SETCC || |
| 4006 | Compare.getOpcode() == ISD::SELECT_CC) && |
| 4007 | "An ISD::SETCC node required here." ); |
| 4008 | |
| 4009 | // Don't convert this comparison to a GPR sequence because there are uses |
| 4010 | // of the i1 result (i.e. uses that require the result in the CR). |
| 4011 | if ((Compare.getOpcode() == ISD::SETCC) && !allUsesExtend(Compare, CurDAG)) |
| 4012 | return SDValue(); |
| 4013 | |
| 4014 | SDValue LHS = Compare.getOperand(i: 0); |
| 4015 | SDValue RHS = Compare.getOperand(i: 1); |
| 4016 | |
| 4017 | // The condition code is operand 2 for SETCC and operand 4 for SELECT_CC. |
| 4018 | int CCOpNum = Compare.getOpcode() == ISD::SELECT_CC ? 4 : 2; |
| 4019 | ISD::CondCode CC = |
| 4020 | cast<CondCodeSDNode>(Val: Compare.getOperand(i: CCOpNum))->get(); |
| 4021 | EVT InputVT = LHS.getValueType(); |
| 4022 | if (InputVT != MVT::i32 && InputVT != MVT::i64) |
| 4023 | return SDValue(); |
| 4024 | |
| 4025 | if (ConvOpts == SetccInGPROpts::ZExtInvert || |
| 4026 | ConvOpts == SetccInGPROpts::SExtInvert) |
| 4027 | CC = ISD::getSetCCInverse(Operation: CC, Type: InputVT); |
| 4028 | |
| 4029 | bool Inputs32Bit = InputVT == MVT::i32; |
| 4030 | |
| 4031 | SDLoc dl(Compare); |
| 4032 | ConstantSDNode *RHSConst = dyn_cast<ConstantSDNode>(Val&: RHS); |
| 4033 | int64_t RHSValue = RHSConst ? RHSConst->getSExtValue() : INT64_MAX; |
| 4034 | bool IsSext = ConvOpts == SetccInGPROpts::SExtOrig || |
| 4035 | ConvOpts == SetccInGPROpts::SExtInvert; |
| 4036 | |
| 4037 | if (IsSext && Inputs32Bit) |
| 4038 | return get32BitSExtCompare(LHS, RHS, CC, RHSValue, dl); |
| 4039 | else if (Inputs32Bit) |
| 4040 | return get32BitZExtCompare(LHS, RHS, CC, RHSValue, dl); |
| 4041 | else if (IsSext) |
| 4042 | return get64BitSExtCompare(LHS, RHS, CC, RHSValue, dl); |
| 4043 | return get64BitZExtCompare(LHS, RHS, CC, RHSValue, dl); |
| 4044 | } |
| 4045 | |
| 4046 | } // end anonymous namespace |
| 4047 | |
| 4048 | bool PPCDAGToDAGISel::tryIntCompareInGPR(SDNode *N) { |
| 4049 | if (N->getValueType(ResNo: 0) != MVT::i32 && |
| 4050 | N->getValueType(ResNo: 0) != MVT::i64) |
| 4051 | return false; |
| 4052 | |
| 4053 | // This optimization will emit code that assumes 64-bit registers |
| 4054 | // so we don't want to run it in 32-bit mode. Also don't run it |
| 4055 | // on functions that are not to be optimized. |
| 4056 | if (TM.getOptLevel() == CodeGenOptLevel::None || !TM.isPPC64()) |
| 4057 | return false; |
| 4058 | |
| 4059 | // For POWER10, it is more profitable to use the set boolean extension |
| 4060 | // instructions rather than the integer compare elimination codegen. |
| 4061 | // Users can override this via the command line option, `--ppc-gpr-icmps`. |
| 4062 | if (!(CmpInGPR.getNumOccurrences() > 0) && Subtarget->isISA3_1()) |
| 4063 | return false; |
| 4064 | |
| 4065 | switch (N->getOpcode()) { |
| 4066 | default: break; |
| 4067 | case ISD::ZERO_EXTEND: |
| 4068 | case ISD::SIGN_EXTEND: |
| 4069 | case ISD::AND: |
| 4070 | case ISD::OR: |
| 4071 | case ISD::XOR: { |
| 4072 | IntegerCompareEliminator ICmpElim(CurDAG, this); |
| 4073 | if (SDNode *New = ICmpElim.Select(N)) { |
| 4074 | ReplaceNode(F: N, T: New); |
| 4075 | return true; |
| 4076 | } |
| 4077 | } |
| 4078 | } |
| 4079 | return false; |
| 4080 | } |
| 4081 | |
| 4082 | bool PPCDAGToDAGISel::tryBitPermutation(SDNode *N) { |
| 4083 | if (N->getValueType(ResNo: 0) != MVT::i32 && |
| 4084 | N->getValueType(ResNo: 0) != MVT::i64) |
| 4085 | return false; |
| 4086 | |
| 4087 | if (!UseBitPermRewriter) |
| 4088 | return false; |
| 4089 | |
| 4090 | switch (N->getOpcode()) { |
| 4091 | default: break; |
| 4092 | case ISD::SRL: |
| 4093 | // If we are on P10, we have a pattern for 32-bit (srl (bswap r), 16) that |
| 4094 | // uses the BRH instruction. |
| 4095 | if (Subtarget->isISA3_1() && N->getValueType(ResNo: 0) == MVT::i32 && |
| 4096 | N->getOperand(Num: 0).getOpcode() == ISD::BSWAP) { |
| 4097 | auto &OpRight = N->getOperand(Num: 1); |
| 4098 | ConstantSDNode *SRLConst = dyn_cast<ConstantSDNode>(Val: OpRight); |
| 4099 | if (SRLConst && SRLConst->getSExtValue() == 16) |
| 4100 | return false; |
| 4101 | } |
| 4102 | [[fallthrough]]; |
| 4103 | case ISD::ROTL: |
| 4104 | case ISD::SHL: |
| 4105 | case ISD::AND: |
| 4106 | case ISD::OR: { |
| 4107 | BitPermutationSelector BPS(CurDAG); |
| 4108 | if (SDNode *New = BPS.Select(N)) { |
| 4109 | ReplaceNode(F: N, T: New); |
| 4110 | return true; |
| 4111 | } |
| 4112 | return false; |
| 4113 | } |
| 4114 | } |
| 4115 | |
| 4116 | return false; |
| 4117 | } |
| 4118 | |
| 4119 | /// SelectCC - Select a comparison of the specified values with the specified |
| 4120 | /// condition code, returning the CR# of the expression. |
| 4121 | SDValue PPCDAGToDAGISel::SelectCC(SDValue LHS, SDValue RHS, ISD::CondCode CC, |
| 4122 | const SDLoc &dl, SDValue Chain) { |
| 4123 | // Always select the LHS. |
| 4124 | unsigned Opc; |
| 4125 | |
| 4126 | if (LHS.getValueType() == MVT::i32) { |
| 4127 | unsigned Imm; |
| 4128 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
| 4129 | if (isInt32Immediate(N: RHS, Imm)) { |
| 4130 | // SETEQ/SETNE comparison with 16-bit immediate, fold it. |
| 4131 | if (isUInt<16>(x: Imm)) |
| 4132 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLWI, dl, VT: MVT::i32, Op1: LHS, |
| 4133 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), |
| 4134 | 0); |
| 4135 | // If this is a 16-bit signed immediate, fold it. |
| 4136 | if (isInt<16>(x: (int)Imm)) |
| 4137 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPWI, dl, VT: MVT::i32, Op1: LHS, |
| 4138 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), |
| 4139 | 0); |
| 4140 | |
| 4141 | // For non-equality comparisons, the default code would materialize the |
| 4142 | // constant, then compare against it, like this: |
| 4143 | // lis r2, 4660 |
| 4144 | // ori r2, r2, 22136 |
| 4145 | // cmpw cr0, r3, r2 |
| 4146 | // Since we are just comparing for equality, we can emit this instead: |
| 4147 | // xoris r0,r3,0x1234 |
| 4148 | // cmplwi cr0,r0,0x5678 |
| 4149 | // beq cr0,L6 |
| 4150 | SDValue Xor(CurDAG->getMachineNode(Opcode: PPC::XORIS, dl, VT: MVT::i32, Op1: LHS, |
| 4151 | Op2: getI32Imm(Imm: Imm >> 16, dl)), 0); |
| 4152 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLWI, dl, VT: MVT::i32, Op1: Xor, |
| 4153 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), 0); |
| 4154 | } |
| 4155 | Opc = PPC::CMPLW; |
| 4156 | } else if (ISD::isUnsignedIntSetCC(Code: CC)) { |
| 4157 | if (isInt32Immediate(N: RHS, Imm) && isUInt<16>(x: Imm)) |
| 4158 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLWI, dl, VT: MVT::i32, Op1: LHS, |
| 4159 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), 0); |
| 4160 | Opc = PPC::CMPLW; |
| 4161 | } else { |
| 4162 | int16_t SImm; |
| 4163 | if (isIntS16Immediate(Op: RHS, Imm&: SImm)) |
| 4164 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPWI, dl, VT: MVT::i32, Op1: LHS, |
| 4165 | Op2: getI32Imm(Imm: (int)SImm & 0xFFFF, |
| 4166 | dl)), |
| 4167 | 0); |
| 4168 | Opc = PPC::CMPW; |
| 4169 | } |
| 4170 | } else if (LHS.getValueType() == MVT::i64) { |
| 4171 | uint64_t Imm; |
| 4172 | if (CC == ISD::SETEQ || CC == ISD::SETNE) { |
| 4173 | if (isInt64Immediate(N: RHS.getNode(), Imm)) { |
| 4174 | // SETEQ/SETNE comparison with 16-bit immediate, fold it. |
| 4175 | if (isUInt<16>(x: Imm)) |
| 4176 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLDI, dl, VT: MVT::i64, Op1: LHS, |
| 4177 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), |
| 4178 | 0); |
| 4179 | // If this is a 16-bit signed immediate, fold it. |
| 4180 | if (isInt<16>(x: Imm)) |
| 4181 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPDI, dl, VT: MVT::i64, Op1: LHS, |
| 4182 | Op2: getI32Imm(Imm: Imm & 0xFFFF, dl)), |
| 4183 | 0); |
| 4184 | |
| 4185 | // For non-equality comparisons, the default code would materialize the |
| 4186 | // constant, then compare against it, like this: |
| 4187 | // lis r2, 4660 |
| 4188 | // ori r2, r2, 22136 |
| 4189 | // cmpd cr0, r3, r2 |
| 4190 | // Since we are just comparing for equality, we can emit this instead: |
| 4191 | // xoris r0,r3,0x1234 |
| 4192 | // cmpldi cr0,r0,0x5678 |
| 4193 | // beq cr0,L6 |
| 4194 | if (isUInt<32>(x: Imm)) { |
| 4195 | SDValue Xor(CurDAG->getMachineNode(Opcode: PPC::XORIS8, dl, VT: MVT::i64, Op1: LHS, |
| 4196 | Op2: getI64Imm(Imm: Imm >> 16, dl)), 0); |
| 4197 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLDI, dl, VT: MVT::i64, Op1: Xor, |
| 4198 | Op2: getI64Imm(Imm: Imm & 0xFFFF, dl)), |
| 4199 | 0); |
| 4200 | } |
| 4201 | } |
| 4202 | Opc = PPC::CMPLD; |
| 4203 | } else if (ISD::isUnsignedIntSetCC(Code: CC)) { |
| 4204 | if (isInt64Immediate(N: RHS.getNode(), Imm) && isUInt<16>(x: Imm)) |
| 4205 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPLDI, dl, VT: MVT::i64, Op1: LHS, |
| 4206 | Op2: getI64Imm(Imm: Imm & 0xFFFF, dl)), 0); |
| 4207 | Opc = PPC::CMPLD; |
| 4208 | } else { |
| 4209 | int16_t SImm; |
| 4210 | if (isIntS16Immediate(Op: RHS, Imm&: SImm)) |
| 4211 | return SDValue(CurDAG->getMachineNode(Opcode: PPC::CMPDI, dl, VT: MVT::i64, Op1: LHS, |
| 4212 | Op2: getI64Imm(Imm: SImm & 0xFFFF, dl)), |
| 4213 | 0); |
| 4214 | Opc = PPC::CMPD; |
| 4215 | } |
| 4216 | } else if (LHS.getValueType() == MVT::f32) { |
| 4217 | if (Subtarget->hasSPE()) { |
| 4218 | switch (CC) { |
| 4219 | default: |
| 4220 | case ISD::SETEQ: |
| 4221 | case ISD::SETNE: |
| 4222 | Opc = PPC::EFSCMPEQ; |
| 4223 | break; |
| 4224 | case ISD::SETLT: |
| 4225 | case ISD::SETGE: |
| 4226 | case ISD::SETOLT: |
| 4227 | case ISD::SETOGE: |
| 4228 | case ISD::SETULT: |
| 4229 | case ISD::SETUGE: |
| 4230 | Opc = PPC::EFSCMPLT; |
| 4231 | break; |
| 4232 | case ISD::SETGT: |
| 4233 | case ISD::SETLE: |
| 4234 | case ISD::SETOGT: |
| 4235 | case ISD::SETOLE: |
| 4236 | case ISD::SETUGT: |
| 4237 | case ISD::SETULE: |
| 4238 | Opc = PPC::EFSCMPGT; |
| 4239 | break; |
| 4240 | } |
| 4241 | } else |
| 4242 | Opc = PPC::FCMPUS; |
| 4243 | } else if (LHS.getValueType() == MVT::f64) { |
| 4244 | if (Subtarget->hasSPE()) { |
| 4245 | switch (CC) { |
| 4246 | default: |
| 4247 | case ISD::SETEQ: |
| 4248 | case ISD::SETNE: |
| 4249 | Opc = PPC::EFDCMPEQ; |
| 4250 | break; |
| 4251 | case ISD::SETLT: |
| 4252 | case ISD::SETGE: |
| 4253 | case ISD::SETOLT: |
| 4254 | case ISD::SETOGE: |
| 4255 | case ISD::SETULT: |
| 4256 | case ISD::SETUGE: |
| 4257 | Opc = PPC::EFDCMPLT; |
| 4258 | break; |
| 4259 | case ISD::SETGT: |
| 4260 | case ISD::SETLE: |
| 4261 | case ISD::SETOGT: |
| 4262 | case ISD::SETOLE: |
| 4263 | case ISD::SETUGT: |
| 4264 | case ISD::SETULE: |
| 4265 | Opc = PPC::EFDCMPGT; |
| 4266 | break; |
| 4267 | } |
| 4268 | } else |
| 4269 | Opc = Subtarget->hasVSX() ? PPC::XSCMPUDP : PPC::FCMPUD; |
| 4270 | } else { |
| 4271 | assert(LHS.getValueType() == MVT::f128 && "Unknown vt!" ); |
| 4272 | assert(Subtarget->hasP9Vector() && "XSCMPUQP requires Power9 Vector" ); |
| 4273 | Opc = PPC::XSCMPUQP; |
| 4274 | } |
| 4275 | if (Chain) |
| 4276 | return SDValue( |
| 4277 | CurDAG->getMachineNode(Opcode: Opc, dl, VT1: MVT::i32, VT2: MVT::Other, Op1: LHS, Op2: RHS, Op3: Chain), |
| 4278 | 0); |
| 4279 | else |
| 4280 | return SDValue(CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i32, Op1: LHS, Op2: RHS), 0); |
| 4281 | } |
| 4282 | |
| 4283 | static PPC::Predicate getPredicateForSetCC(ISD::CondCode CC, const EVT &VT, |
| 4284 | const PPCSubtarget *Subtarget) { |
| 4285 | // For SPE instructions, the result is in GT bit of the CR |
| 4286 | bool UseSPE = Subtarget->hasSPE() && VT.isFloatingPoint(); |
| 4287 | |
| 4288 | switch (CC) { |
| 4289 | case ISD::SETUEQ: |
| 4290 | case ISD::SETONE: |
| 4291 | case ISD::SETOLE: |
| 4292 | case ISD::SETOGE: |
| 4293 | llvm_unreachable("Should be lowered by legalize!" ); |
| 4294 | default: llvm_unreachable("Unknown condition!" ); |
| 4295 | case ISD::SETOEQ: |
| 4296 | case ISD::SETEQ: |
| 4297 | return UseSPE ? PPC::PRED_GT : PPC::PRED_EQ; |
| 4298 | case ISD::SETUNE: |
| 4299 | case ISD::SETNE: |
| 4300 | return UseSPE ? PPC::PRED_LE : PPC::PRED_NE; |
| 4301 | case ISD::SETOLT: |
| 4302 | case ISD::SETLT: |
| 4303 | return UseSPE ? PPC::PRED_GT : PPC::PRED_LT; |
| 4304 | case ISD::SETULE: |
| 4305 | case ISD::SETLE: |
| 4306 | return PPC::PRED_LE; |
| 4307 | case ISD::SETOGT: |
| 4308 | case ISD::SETGT: |
| 4309 | return PPC::PRED_GT; |
| 4310 | case ISD::SETUGE: |
| 4311 | case ISD::SETGE: |
| 4312 | return UseSPE ? PPC::PRED_LE : PPC::PRED_GE; |
| 4313 | case ISD::SETO: return PPC::PRED_NU; |
| 4314 | case ISD::SETUO: return PPC::PRED_UN; |
| 4315 | // These two are invalid for floating point. Assume we have int. |
| 4316 | case ISD::SETULT: return PPC::PRED_LT; |
| 4317 | case ISD::SETUGT: return PPC::PRED_GT; |
| 4318 | } |
| 4319 | } |
| 4320 | |
| 4321 | /// getCRIdxForSetCC - Return the index of the condition register field |
| 4322 | /// associated with the SetCC condition, and whether or not the field is |
| 4323 | /// treated as inverted. That is, lt = 0; ge = 0 inverted. |
| 4324 | static unsigned getCRIdxForSetCC(ISD::CondCode CC, bool &Invert) { |
| 4325 | Invert = false; |
| 4326 | switch (CC) { |
| 4327 | default: llvm_unreachable("Unknown condition!" ); |
| 4328 | case ISD::SETOLT: |
| 4329 | case ISD::SETLT: return 0; // Bit #0 = SETOLT |
| 4330 | case ISD::SETOGT: |
| 4331 | case ISD::SETGT: return 1; // Bit #1 = SETOGT |
| 4332 | case ISD::SETOEQ: |
| 4333 | case ISD::SETEQ: return 2; // Bit #2 = SETOEQ |
| 4334 | case ISD::SETUO: return 3; // Bit #3 = SETUO |
| 4335 | case ISD::SETUGE: |
| 4336 | case ISD::SETGE: Invert = true; return 0; // !Bit #0 = SETUGE |
| 4337 | case ISD::SETULE: |
| 4338 | case ISD::SETLE: Invert = true; return 1; // !Bit #1 = SETULE |
| 4339 | case ISD::SETUNE: |
| 4340 | case ISD::SETNE: Invert = true; return 2; // !Bit #2 = SETUNE |
| 4341 | case ISD::SETO: Invert = true; return 3; // !Bit #3 = SETO |
| 4342 | case ISD::SETUEQ: |
| 4343 | case ISD::SETOGE: |
| 4344 | case ISD::SETOLE: |
| 4345 | case ISD::SETONE: |
| 4346 | llvm_unreachable("Invalid branch code: should be expanded by legalize" ); |
| 4347 | // These are invalid for floating point. Assume integer. |
| 4348 | case ISD::SETULT: return 0; |
| 4349 | case ISD::SETUGT: return 1; |
| 4350 | } |
| 4351 | } |
| 4352 | |
| 4353 | // getVCmpInst: return the vector compare instruction for the specified |
| 4354 | // vector type and condition code. Since this is for altivec specific code, |
| 4355 | // only support the altivec types (v16i8, v8i16, v4i32, v2i64, v1i128, |
| 4356 | // and v4f32). |
| 4357 | static unsigned int getVCmpInst(MVT VecVT, ISD::CondCode CC, |
| 4358 | bool HasVSX, bool &Swap, bool &Negate) { |
| 4359 | Swap = false; |
| 4360 | Negate = false; |
| 4361 | |
| 4362 | if (VecVT.isFloatingPoint()) { |
| 4363 | /* Handle some cases by swapping input operands. */ |
| 4364 | switch (CC) { |
| 4365 | case ISD::SETLE: CC = ISD::SETGE; Swap = true; break; |
| 4366 | case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; |
| 4367 | case ISD::SETOLE: CC = ISD::SETOGE; Swap = true; break; |
| 4368 | case ISD::SETOLT: CC = ISD::SETOGT; Swap = true; break; |
| 4369 | case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; |
| 4370 | case ISD::SETUGT: CC = ISD::SETULT; Swap = true; break; |
| 4371 | default: break; |
| 4372 | } |
| 4373 | /* Handle some cases by negating the result. */ |
| 4374 | switch (CC) { |
| 4375 | case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; |
| 4376 | case ISD::SETUNE: CC = ISD::SETOEQ; Negate = true; break; |
| 4377 | case ISD::SETULE: CC = ISD::SETOGT; Negate = true; break; |
| 4378 | case ISD::SETULT: CC = ISD::SETOGE; Negate = true; break; |
| 4379 | default: break; |
| 4380 | } |
| 4381 | /* We have instructions implementing the remaining cases. */ |
| 4382 | switch (CC) { |
| 4383 | case ISD::SETEQ: |
| 4384 | case ISD::SETOEQ: |
| 4385 | if (VecVT == MVT::v4f32) |
| 4386 | return HasVSX ? PPC::XVCMPEQSP : PPC::VCMPEQFP; |
| 4387 | else if (VecVT == MVT::v2f64) |
| 4388 | return PPC::XVCMPEQDP; |
| 4389 | break; |
| 4390 | case ISD::SETGT: |
| 4391 | case ISD::SETOGT: |
| 4392 | if (VecVT == MVT::v4f32) |
| 4393 | return HasVSX ? PPC::XVCMPGTSP : PPC::VCMPGTFP; |
| 4394 | else if (VecVT == MVT::v2f64) |
| 4395 | return PPC::XVCMPGTDP; |
| 4396 | break; |
| 4397 | case ISD::SETGE: |
| 4398 | case ISD::SETOGE: |
| 4399 | if (VecVT == MVT::v4f32) |
| 4400 | return HasVSX ? PPC::XVCMPGESP : PPC::VCMPGEFP; |
| 4401 | else if (VecVT == MVT::v2f64) |
| 4402 | return PPC::XVCMPGEDP; |
| 4403 | break; |
| 4404 | default: |
| 4405 | break; |
| 4406 | } |
| 4407 | llvm_unreachable("Invalid floating-point vector compare condition" ); |
| 4408 | } else { |
| 4409 | /* Handle some cases by swapping input operands. */ |
| 4410 | switch (CC) { |
| 4411 | case ISD::SETGE: CC = ISD::SETLE; Swap = true; break; |
| 4412 | case ISD::SETLT: CC = ISD::SETGT; Swap = true; break; |
| 4413 | case ISD::SETUGE: CC = ISD::SETULE; Swap = true; break; |
| 4414 | case ISD::SETULT: CC = ISD::SETUGT; Swap = true; break; |
| 4415 | default: break; |
| 4416 | } |
| 4417 | /* Handle some cases by negating the result. */ |
| 4418 | switch (CC) { |
| 4419 | case ISD::SETNE: CC = ISD::SETEQ; Negate = true; break; |
| 4420 | case ISD::SETUNE: CC = ISD::SETUEQ; Negate = true; break; |
| 4421 | case ISD::SETLE: CC = ISD::SETGT; Negate = true; break; |
| 4422 | case ISD::SETULE: CC = ISD::SETUGT; Negate = true; break; |
| 4423 | default: break; |
| 4424 | } |
| 4425 | /* We have instructions implementing the remaining cases. */ |
| 4426 | switch (CC) { |
| 4427 | case ISD::SETEQ: |
| 4428 | case ISD::SETUEQ: |
| 4429 | if (VecVT == MVT::v16i8) |
| 4430 | return PPC::VCMPEQUB; |
| 4431 | else if (VecVT == MVT::v8i16) |
| 4432 | return PPC::VCMPEQUH; |
| 4433 | else if (VecVT == MVT::v4i32) |
| 4434 | return PPC::VCMPEQUW; |
| 4435 | else if (VecVT == MVT::v2i64) |
| 4436 | return PPC::VCMPEQUD; |
| 4437 | else if (VecVT == MVT::v1i128) |
| 4438 | return PPC::VCMPEQUQ; |
| 4439 | break; |
| 4440 | case ISD::SETGT: |
| 4441 | if (VecVT == MVT::v16i8) |
| 4442 | return PPC::VCMPGTSB; |
| 4443 | else if (VecVT == MVT::v8i16) |
| 4444 | return PPC::VCMPGTSH; |
| 4445 | else if (VecVT == MVT::v4i32) |
| 4446 | return PPC::VCMPGTSW; |
| 4447 | else if (VecVT == MVT::v2i64) |
| 4448 | return PPC::VCMPGTSD; |
| 4449 | else if (VecVT == MVT::v1i128) |
| 4450 | return PPC::VCMPGTSQ; |
| 4451 | break; |
| 4452 | case ISD::SETUGT: |
| 4453 | if (VecVT == MVT::v16i8) |
| 4454 | return PPC::VCMPGTUB; |
| 4455 | else if (VecVT == MVT::v8i16) |
| 4456 | return PPC::VCMPGTUH; |
| 4457 | else if (VecVT == MVT::v4i32) |
| 4458 | return PPC::VCMPGTUW; |
| 4459 | else if (VecVT == MVT::v2i64) |
| 4460 | return PPC::VCMPGTUD; |
| 4461 | else if (VecVT == MVT::v1i128) |
| 4462 | return PPC::VCMPGTUQ; |
| 4463 | break; |
| 4464 | default: |
| 4465 | break; |
| 4466 | } |
| 4467 | llvm_unreachable("Invalid integer vector compare condition" ); |
| 4468 | } |
| 4469 | } |
| 4470 | |
| 4471 | bool PPCDAGToDAGISel::trySETCC(SDNode *N) { |
| 4472 | SDLoc dl(N); |
| 4473 | unsigned Imm; |
| 4474 | bool IsStrict = N->isStrictFPOpcode(); |
| 4475 | ISD::CondCode CC = |
| 4476 | cast<CondCodeSDNode>(Val: N->getOperand(Num: IsStrict ? 3 : 2))->get(); |
| 4477 | EVT PtrVT = |
| 4478 | CurDAG->getTargetLoweringInfo().getPointerTy(DL: CurDAG->getDataLayout()); |
| 4479 | bool isPPC64 = (PtrVT == MVT::i64); |
| 4480 | SDValue Chain = IsStrict ? N->getOperand(Num: 0) : SDValue(); |
| 4481 | |
| 4482 | SDValue LHS = N->getOperand(Num: IsStrict ? 1 : 0); |
| 4483 | SDValue RHS = N->getOperand(Num: IsStrict ? 2 : 1); |
| 4484 | |
| 4485 | if (!IsStrict && !Subtarget->useCRBits() && isInt32Immediate(N: RHS, Imm)) { |
| 4486 | // We can codegen setcc op, imm very efficiently compared to a brcond. |
| 4487 | // Check for those cases here. |
| 4488 | // setcc op, 0 |
| 4489 | if (Imm == 0) { |
| 4490 | SDValue Op = LHS; |
| 4491 | switch (CC) { |
| 4492 | default: break; |
| 4493 | case ISD::SETEQ: { |
| 4494 | Op = SDValue(CurDAG->getMachineNode(Opcode: PPC::CNTLZW, dl, VT: MVT::i32, Op1: Op), 0); |
| 4495 | SDValue Ops[] = { Op, getI32Imm(Imm: 27, dl), getI32Imm(Imm: 5, dl), |
| 4496 | getI32Imm(Imm: 31, dl) }; |
| 4497 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4498 | return true; |
| 4499 | } |
| 4500 | case ISD::SETNE: { |
| 4501 | if (isPPC64) break; |
| 4502 | SDValue AD = |
| 4503 | SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDIC, dl, VT1: MVT::i32, VT2: MVT::Glue, |
| 4504 | Op1: Op, Op2: getI32Imm(Imm: ~0U, dl)), 0); |
| 4505 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::SUBFE, VT: MVT::i32, Op1: AD, Op2: Op, Op3: AD.getValue(R: 1)); |
| 4506 | return true; |
| 4507 | } |
| 4508 | case ISD::SETLT: { |
| 4509 | SDValue Ops[] = { Op, getI32Imm(Imm: 1, dl), getI32Imm(Imm: 31, dl), |
| 4510 | getI32Imm(Imm: 31, dl) }; |
| 4511 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4512 | return true; |
| 4513 | } |
| 4514 | case ISD::SETGT: { |
| 4515 | SDValue T = |
| 4516 | SDValue(CurDAG->getMachineNode(Opcode: PPC::NEG, dl, VT: MVT::i32, Op1: Op), 0); |
| 4517 | T = SDValue(CurDAG->getMachineNode(Opcode: PPC::ANDC, dl, VT: MVT::i32, Op1: T, Op2: Op), 0); |
| 4518 | SDValue Ops[] = { T, getI32Imm(Imm: 1, dl), getI32Imm(Imm: 31, dl), |
| 4519 | getI32Imm(Imm: 31, dl) }; |
| 4520 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4521 | return true; |
| 4522 | } |
| 4523 | } |
| 4524 | } else if (Imm == ~0U) { // setcc op, -1 |
| 4525 | SDValue Op = LHS; |
| 4526 | switch (CC) { |
| 4527 | default: break; |
| 4528 | case ISD::SETEQ: |
| 4529 | if (isPPC64) break; |
| 4530 | Op = SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDIC, dl, VT1: MVT::i32, VT2: MVT::Glue, |
| 4531 | Op1: Op, Op2: getI32Imm(Imm: 1, dl)), 0); |
| 4532 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::ADDZE, VT: MVT::i32, |
| 4533 | Op1: SDValue(CurDAG->getMachineNode(Opcode: PPC::LI, dl, |
| 4534 | VT: MVT::i32, |
| 4535 | Op1: getI32Imm(Imm: 0, dl)), |
| 4536 | 0), Op2: Op.getValue(R: 1)); |
| 4537 | return true; |
| 4538 | case ISD::SETNE: { |
| 4539 | if (isPPC64) break; |
| 4540 | Op = SDValue(CurDAG->getMachineNode(Opcode: PPC::NOR, dl, VT: MVT::i32, Op1: Op, Op2: Op), 0); |
| 4541 | SDNode *AD = CurDAG->getMachineNode(Opcode: PPC::ADDIC, dl, VT1: MVT::i32, VT2: MVT::Glue, |
| 4542 | Op1: Op, Op2: getI32Imm(Imm: ~0U, dl)); |
| 4543 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::SUBFE, VT: MVT::i32, Op1: SDValue(AD, 0), Op2: Op, |
| 4544 | Op3: SDValue(AD, 1)); |
| 4545 | return true; |
| 4546 | } |
| 4547 | case ISD::SETLT: { |
| 4548 | SDValue AD = SDValue(CurDAG->getMachineNode(Opcode: PPC::ADDI, dl, VT: MVT::i32, Op1: Op, |
| 4549 | Op2: getI32Imm(Imm: 1, dl)), 0); |
| 4550 | SDValue AN = SDValue(CurDAG->getMachineNode(Opcode: PPC::AND, dl, VT: MVT::i32, Op1: AD, |
| 4551 | Op2: Op), 0); |
| 4552 | SDValue Ops[] = { AN, getI32Imm(Imm: 1, dl), getI32Imm(Imm: 31, dl), |
| 4553 | getI32Imm(Imm: 31, dl) }; |
| 4554 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4555 | return true; |
| 4556 | } |
| 4557 | case ISD::SETGT: { |
| 4558 | SDValue Ops[] = { Op, getI32Imm(Imm: 1, dl), getI32Imm(Imm: 31, dl), |
| 4559 | getI32Imm(Imm: 31, dl) }; |
| 4560 | Op = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops), 0); |
| 4561 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::XORI, VT: MVT::i32, Op1: Op, Op2: getI32Imm(Imm: 1, dl)); |
| 4562 | return true; |
| 4563 | } |
| 4564 | } |
| 4565 | } |
| 4566 | } |
| 4567 | |
| 4568 | // Altivec Vector compare instructions do not set any CR register by default and |
| 4569 | // vector compare operations return the same type as the operands. |
| 4570 | if (!IsStrict && LHS.getValueType().isVector()) { |
| 4571 | if (Subtarget->hasSPE()) |
| 4572 | return false; |
| 4573 | |
| 4574 | EVT VecVT = LHS.getValueType(); |
| 4575 | // Optimize 'Not equal to zero-vector' comparisons to 'Greater than or |
| 4576 | // less than' operators. |
| 4577 | // Example: Consider k to be any non-zero positive value. |
| 4578 | // * for k != 0, change SETNE to SETUGT (k > 0) |
| 4579 | // * for 0 != k, change SETNE to SETULT (0 < k) |
| 4580 | if (CC == ISD::SETNE) { |
| 4581 | // Only optimize for integer types (avoid FP completely) |
| 4582 | if (VecVT.getVectorElementType().isInteger()) { |
| 4583 | if (ISD::isBuildVectorAllZeros(N: RHS.getNode())) |
| 4584 | CC = ISD::SETUGT; |
| 4585 | else if (ISD::isBuildVectorAllZeros(N: LHS.getNode())) |
| 4586 | CC = ISD::SETULT; |
| 4587 | } |
| 4588 | } |
| 4589 | bool Swap, Negate; |
| 4590 | unsigned int VCmpInst = |
| 4591 | getVCmpInst(VecVT: VecVT.getSimpleVT(), CC, HasVSX: Subtarget->hasVSX(), Swap, Negate); |
| 4592 | if (Swap) |
| 4593 | std::swap(a&: LHS, b&: RHS); |
| 4594 | |
| 4595 | EVT ResVT = VecVT.changeVectorElementTypeToInteger(); |
| 4596 | if (Negate) { |
| 4597 | SDValue VCmp(CurDAG->getMachineNode(Opcode: VCmpInst, dl, VT: ResVT, Op1: LHS, Op2: RHS), 0); |
| 4598 | CurDAG->SelectNodeTo(N, MachineOpc: Subtarget->hasVSX() ? PPC::XXLNOR : PPC::VNOR, |
| 4599 | VT: ResVT, Op1: VCmp, Op2: VCmp); |
| 4600 | return true; |
| 4601 | } |
| 4602 | |
| 4603 | CurDAG->SelectNodeTo(N, MachineOpc: VCmpInst, VT: ResVT, Op1: LHS, Op2: RHS); |
| 4604 | return true; |
| 4605 | } |
| 4606 | |
| 4607 | if (Subtarget->useCRBits()) |
| 4608 | return false; |
| 4609 | |
| 4610 | bool Inv; |
| 4611 | unsigned Idx = getCRIdxForSetCC(CC, Invert&: Inv); |
| 4612 | SDValue CCReg = SelectCC(LHS, RHS, CC, dl, Chain); |
| 4613 | if (IsStrict) |
| 4614 | CurDAG->ReplaceAllUsesOfValueWith(From: SDValue(N, 1), To: CCReg.getValue(R: 1)); |
| 4615 | SDValue IntCR; |
| 4616 | |
| 4617 | // SPE e*cmp* instructions only set the 'gt' bit, so hard-code that |
| 4618 | // The correct compare instruction is already set by SelectCC() |
| 4619 | if (Subtarget->hasSPE() && LHS.getValueType().isFloatingPoint()) { |
| 4620 | Idx = 1; |
| 4621 | } |
| 4622 | |
| 4623 | // Force the ccreg into CR7. |
| 4624 | SDValue CR7Reg = CurDAG->getRegister(Reg: PPC::CR7, VT: MVT::i32); |
| 4625 | |
| 4626 | SDValue InGlue; // Null incoming flag value. |
| 4627 | CCReg = CurDAG->getCopyToReg(Chain: CurDAG->getEntryNode(), dl, Reg: CR7Reg, N: CCReg, |
| 4628 | Glue: InGlue).getValue(R: 1); |
| 4629 | |
| 4630 | IntCR = SDValue(CurDAG->getMachineNode(Opcode: PPC::MFOCRF, dl, VT: MVT::i32, Op1: CR7Reg, |
| 4631 | Op2: CCReg), 0); |
| 4632 | |
| 4633 | SDValue Ops[] = { IntCR, getI32Imm(Imm: (32 - (3 - Idx)) & 31, dl), |
| 4634 | getI32Imm(Imm: 31, dl), getI32Imm(Imm: 31, dl) }; |
| 4635 | if (!Inv) { |
| 4636 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4637 | return true; |
| 4638 | } |
| 4639 | |
| 4640 | // Get the specified bit. |
| 4641 | SDValue Tmp = |
| 4642 | SDValue(CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops), 0); |
| 4643 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::XORI, VT: MVT::i32, Op1: Tmp, Op2: getI32Imm(Imm: 1, dl)); |
| 4644 | return true; |
| 4645 | } |
| 4646 | |
| 4647 | /// Does this node represent a load/store node whose address can be represented |
| 4648 | /// with a register plus an immediate that's a multiple of \p Val: |
| 4649 | bool PPCDAGToDAGISel::isOffsetMultipleOf(SDNode *N, unsigned Val) const { |
| 4650 | LoadSDNode *LDN = dyn_cast<LoadSDNode>(Val: N); |
| 4651 | StoreSDNode *STN = dyn_cast<StoreSDNode>(Val: N); |
| 4652 | MemIntrinsicSDNode *MIN = dyn_cast<MemIntrinsicSDNode>(Val: N); |
| 4653 | SDValue AddrOp; |
| 4654 | if (LDN || (MIN && MIN->getOpcode() == PPCISD::LD_SPLAT)) |
| 4655 | AddrOp = N->getOperand(Num: 1); |
| 4656 | else if (STN) |
| 4657 | AddrOp = STN->getOperand(Num: 2); |
| 4658 | |
| 4659 | // If the address points a frame object or a frame object with an offset, |
| 4660 | // we need to check the object alignment. |
| 4661 | short Imm = 0; |
| 4662 | if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>( |
| 4663 | Val: AddrOp.getOpcode() == ISD::ADD ? AddrOp.getOperand(i: 0) : |
| 4664 | AddrOp)) { |
| 4665 | // If op0 is a frame index that is under aligned, we can't do it either, |
| 4666 | // because it is translated to r31 or r1 + slot + offset. We won't know the |
| 4667 | // slot number until the stack frame is finalized. |
| 4668 | const MachineFrameInfo &MFI = CurDAG->getMachineFunction().getFrameInfo(); |
| 4669 | unsigned SlotAlign = MFI.getObjectAlign(ObjectIdx: FI->getIndex()).value(); |
| 4670 | if ((SlotAlign % Val) != 0) |
| 4671 | return false; |
| 4672 | |
| 4673 | // If we have an offset, we need further check on the offset. |
| 4674 | if (AddrOp.getOpcode() != ISD::ADD) |
| 4675 | return true; |
| 4676 | } |
| 4677 | |
| 4678 | if (AddrOp.getOpcode() == ISD::ADD) |
| 4679 | return isIntS16Immediate(Op: AddrOp.getOperand(i: 1), Imm) && !(Imm % Val); |
| 4680 | |
| 4681 | // If the address comes from the outside, the offset will be zero. |
| 4682 | return AddrOp.getOpcode() == ISD::CopyFromReg; |
| 4683 | } |
| 4684 | |
| 4685 | void PPCDAGToDAGISel::transferMemOperands(SDNode *N, SDNode *Result) { |
| 4686 | // Transfer memoperands. |
| 4687 | MachineMemOperand *MemOp = cast<MemSDNode>(Val: N)->getMemOperand(); |
| 4688 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: Result), NewMemRefs: {MemOp}); |
| 4689 | } |
| 4690 | |
| 4691 | static bool mayUseP9Setb(SDNode *N, const ISD::CondCode &CC, SelectionDAG *DAG, |
| 4692 | bool &NeedSwapOps, bool &IsUnCmp) { |
| 4693 | |
| 4694 | assert(N->getOpcode() == ISD::SELECT_CC && "Expecting a SELECT_CC here." ); |
| 4695 | |
| 4696 | SDValue LHS = N->getOperand(Num: 0); |
| 4697 | SDValue RHS = N->getOperand(Num: 1); |
| 4698 | SDValue TrueRes = N->getOperand(Num: 2); |
| 4699 | SDValue FalseRes = N->getOperand(Num: 3); |
| 4700 | ConstantSDNode *TrueConst = dyn_cast<ConstantSDNode>(Val&: TrueRes); |
| 4701 | if (!TrueConst || (N->getSimpleValueType(ResNo: 0) != MVT::i64 && |
| 4702 | N->getSimpleValueType(ResNo: 0) != MVT::i32)) |
| 4703 | return false; |
| 4704 | |
| 4705 | // We are looking for any of: |
| 4706 | // (select_cc lhs, rhs, 1, (sext (setcc [lr]hs, [lr]hs, cc2)), cc1) |
| 4707 | // (select_cc lhs, rhs, -1, (zext (setcc [lr]hs, [lr]hs, cc2)), cc1) |
| 4708 | // (select_cc lhs, rhs, 0, (select_cc [lr]hs, [lr]hs, 1, -1, cc2), seteq) |
| 4709 | // (select_cc lhs, rhs, 0, (select_cc [lr]hs, [lr]hs, -1, 1, cc2), seteq) |
| 4710 | int64_t TrueResVal = TrueConst->getSExtValue(); |
| 4711 | if ((TrueResVal < -1 || TrueResVal > 1) || |
| 4712 | (TrueResVal == -1 && FalseRes.getOpcode() != ISD::ZERO_EXTEND) || |
| 4713 | (TrueResVal == 1 && FalseRes.getOpcode() != ISD::SIGN_EXTEND) || |
| 4714 | (TrueResVal == 0 && |
| 4715 | (FalseRes.getOpcode() != ISD::SELECT_CC || CC != ISD::SETEQ))) |
| 4716 | return false; |
| 4717 | |
| 4718 | SDValue SetOrSelCC = FalseRes.getOpcode() == ISD::SELECT_CC |
| 4719 | ? FalseRes |
| 4720 | : FalseRes.getOperand(i: 0); |
| 4721 | bool InnerIsSel = SetOrSelCC.getOpcode() == ISD::SELECT_CC; |
| 4722 | if (SetOrSelCC.getOpcode() != ISD::SETCC && |
| 4723 | SetOrSelCC.getOpcode() != ISD::SELECT_CC) |
| 4724 | return false; |
| 4725 | |
| 4726 | // Without this setb optimization, the outer SELECT_CC will be manually |
| 4727 | // selected to SELECT_CC_I4/SELECT_CC_I8 Pseudo, then expand-isel-pseudos pass |
| 4728 | // transforms pseudo instruction to isel instruction. When there are more than |
| 4729 | // one use for result like zext/sext, with current optimization we only see |
| 4730 | // isel is replaced by setb but can't see any significant gain. Since |
| 4731 | // setb has longer latency than original isel, we should avoid this. Another |
| 4732 | // point is that setb requires comparison always kept, it can break the |
| 4733 | // opportunity to get the comparison away if we have in future. |
| 4734 | if (!SetOrSelCC.hasOneUse() || (!InnerIsSel && !FalseRes.hasOneUse())) |
| 4735 | return false; |
| 4736 | |
| 4737 | SDValue InnerLHS = SetOrSelCC.getOperand(i: 0); |
| 4738 | SDValue InnerRHS = SetOrSelCC.getOperand(i: 1); |
| 4739 | ISD::CondCode InnerCC = |
| 4740 | cast<CondCodeSDNode>(Val: SetOrSelCC.getOperand(i: InnerIsSel ? 4 : 2))->get(); |
| 4741 | // If the inner comparison is a select_cc, make sure the true/false values are |
| 4742 | // 1/-1 and canonicalize it if needed. |
| 4743 | if (InnerIsSel) { |
| 4744 | ConstantSDNode *SelCCTrueConst = |
| 4745 | dyn_cast<ConstantSDNode>(Val: SetOrSelCC.getOperand(i: 2)); |
| 4746 | ConstantSDNode *SelCCFalseConst = |
| 4747 | dyn_cast<ConstantSDNode>(Val: SetOrSelCC.getOperand(i: 3)); |
| 4748 | if (!SelCCTrueConst || !SelCCFalseConst) |
| 4749 | return false; |
| 4750 | int64_t SelCCTVal = SelCCTrueConst->getSExtValue(); |
| 4751 | int64_t SelCCFVal = SelCCFalseConst->getSExtValue(); |
| 4752 | // The values must be -1/1 (requiring a swap) or 1/-1. |
| 4753 | if (SelCCTVal == -1 && SelCCFVal == 1) { |
| 4754 | std::swap(a&: InnerLHS, b&: InnerRHS); |
| 4755 | } else if (SelCCTVal != 1 || SelCCFVal != -1) |
| 4756 | return false; |
| 4757 | } |
| 4758 | |
| 4759 | // Canonicalize unsigned case |
| 4760 | if (InnerCC == ISD::SETULT || InnerCC == ISD::SETUGT) { |
| 4761 | IsUnCmp = true; |
| 4762 | InnerCC = (InnerCC == ISD::SETULT) ? ISD::SETLT : ISD::SETGT; |
| 4763 | } |
| 4764 | |
| 4765 | bool InnerSwapped = false; |
| 4766 | if (LHS == InnerRHS && RHS == InnerLHS) |
| 4767 | InnerSwapped = true; |
| 4768 | else if (LHS != InnerLHS || RHS != InnerRHS) |
| 4769 | return false; |
| 4770 | |
| 4771 | switch (CC) { |
| 4772 | // (select_cc lhs, rhs, 0, \ |
| 4773 | // (select_cc [lr]hs, [lr]hs, 1, -1, setlt/setgt), seteq) |
| 4774 | case ISD::SETEQ: |
| 4775 | if (!InnerIsSel) |
| 4776 | return false; |
| 4777 | if (InnerCC != ISD::SETLT && InnerCC != ISD::SETGT) |
| 4778 | return false; |
| 4779 | NeedSwapOps = (InnerCC == ISD::SETGT) ? InnerSwapped : !InnerSwapped; |
| 4780 | break; |
| 4781 | |
| 4782 | // (select_cc lhs, rhs, -1, (zext (setcc [lr]hs, [lr]hs, setne)), setu?lt) |
| 4783 | // (select_cc lhs, rhs, -1, (zext (setcc lhs, rhs, setgt)), setu?lt) |
| 4784 | // (select_cc lhs, rhs, -1, (zext (setcc rhs, lhs, setlt)), setu?lt) |
| 4785 | // (select_cc lhs, rhs, 1, (sext (setcc [lr]hs, [lr]hs, setne)), setu?lt) |
| 4786 | // (select_cc lhs, rhs, 1, (sext (setcc lhs, rhs, setgt)), setu?lt) |
| 4787 | // (select_cc lhs, rhs, 1, (sext (setcc rhs, lhs, setlt)), setu?lt) |
| 4788 | case ISD::SETULT: |
| 4789 | if (!IsUnCmp && InnerCC != ISD::SETNE) |
| 4790 | return false; |
| 4791 | IsUnCmp = true; |
| 4792 | [[fallthrough]]; |
| 4793 | case ISD::SETLT: |
| 4794 | if (InnerCC == ISD::SETNE || (InnerCC == ISD::SETGT && !InnerSwapped) || |
| 4795 | (InnerCC == ISD::SETLT && InnerSwapped)) |
| 4796 | NeedSwapOps = (TrueResVal == 1); |
| 4797 | else |
| 4798 | return false; |
| 4799 | break; |
| 4800 | |
| 4801 | // (select_cc lhs, rhs, 1, (sext (setcc [lr]hs, [lr]hs, setne)), setu?gt) |
| 4802 | // (select_cc lhs, rhs, 1, (sext (setcc lhs, rhs, setlt)), setu?gt) |
| 4803 | // (select_cc lhs, rhs, 1, (sext (setcc rhs, lhs, setgt)), setu?gt) |
| 4804 | // (select_cc lhs, rhs, -1, (zext (setcc [lr]hs, [lr]hs, setne)), setu?gt) |
| 4805 | // (select_cc lhs, rhs, -1, (zext (setcc lhs, rhs, setlt)), setu?gt) |
| 4806 | // (select_cc lhs, rhs, -1, (zext (setcc rhs, lhs, setgt)), setu?gt) |
| 4807 | case ISD::SETUGT: |
| 4808 | if (!IsUnCmp && InnerCC != ISD::SETNE) |
| 4809 | return false; |
| 4810 | IsUnCmp = true; |
| 4811 | [[fallthrough]]; |
| 4812 | case ISD::SETGT: |
| 4813 | if (InnerCC == ISD::SETNE || (InnerCC == ISD::SETLT && !InnerSwapped) || |
| 4814 | (InnerCC == ISD::SETGT && InnerSwapped)) |
| 4815 | NeedSwapOps = (TrueResVal == -1); |
| 4816 | else |
| 4817 | return false; |
| 4818 | break; |
| 4819 | |
| 4820 | default: |
| 4821 | return false; |
| 4822 | } |
| 4823 | |
| 4824 | LLVM_DEBUG(dbgs() << "Found a node that can be lowered to a SETB: " ); |
| 4825 | LLVM_DEBUG(N->dump()); |
| 4826 | |
| 4827 | return true; |
| 4828 | } |
| 4829 | |
| 4830 | // Return true if it's a software square-root/divide operand. |
| 4831 | static bool isSWTestOp(SDValue N) { |
| 4832 | if (N.getOpcode() == PPCISD::FTSQRT) |
| 4833 | return true; |
| 4834 | if (N.getNumOperands() < 1 || !isa<ConstantSDNode>(Val: N.getOperand(i: 0)) || |
| 4835 | N.getOpcode() != ISD::INTRINSIC_WO_CHAIN) |
| 4836 | return false; |
| 4837 | switch (N.getConstantOperandVal(i: 0)) { |
| 4838 | case Intrinsic::ppc_vsx_xvtdivdp: |
| 4839 | case Intrinsic::ppc_vsx_xvtdivsp: |
| 4840 | case Intrinsic::ppc_vsx_xvtsqrtdp: |
| 4841 | case Intrinsic::ppc_vsx_xvtsqrtsp: |
| 4842 | return true; |
| 4843 | } |
| 4844 | return false; |
| 4845 | } |
| 4846 | |
| 4847 | bool PPCDAGToDAGISel::tryFoldSWTestBRCC(SDNode *N) { |
| 4848 | assert(N->getOpcode() == ISD::BR_CC && "ISD::BR_CC is expected." ); |
| 4849 | // We are looking for following patterns, where `truncate to i1` actually has |
| 4850 | // the same semantic with `and 1`. |
| 4851 | // (br_cc seteq, (truncateToi1 SWTestOp), 0) -> (BCC PRED_NU, SWTestOp) |
| 4852 | // (br_cc seteq, (and SWTestOp, 2), 0) -> (BCC PRED_NE, SWTestOp) |
| 4853 | // (br_cc seteq, (and SWTestOp, 4), 0) -> (BCC PRED_LE, SWTestOp) |
| 4854 | // (br_cc seteq, (and SWTestOp, 8), 0) -> (BCC PRED_GE, SWTestOp) |
| 4855 | // (br_cc setne, (truncateToi1 SWTestOp), 0) -> (BCC PRED_UN, SWTestOp) |
| 4856 | // (br_cc setne, (and SWTestOp, 2), 0) -> (BCC PRED_EQ, SWTestOp) |
| 4857 | // (br_cc setne, (and SWTestOp, 4), 0) -> (BCC PRED_GT, SWTestOp) |
| 4858 | // (br_cc setne, (and SWTestOp, 8), 0) -> (BCC PRED_LT, SWTestOp) |
| 4859 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 1))->get(); |
| 4860 | if (CC != ISD::SETEQ && CC != ISD::SETNE) |
| 4861 | return false; |
| 4862 | |
| 4863 | SDValue CmpRHS = N->getOperand(Num: 3); |
| 4864 | if (!isNullConstant(V: CmpRHS)) |
| 4865 | return false; |
| 4866 | |
| 4867 | SDValue CmpLHS = N->getOperand(Num: 2); |
| 4868 | if (CmpLHS.getNumOperands() < 1 || !isSWTestOp(N: CmpLHS.getOperand(i: 0))) |
| 4869 | return false; |
| 4870 | |
| 4871 | unsigned PCC = 0; |
| 4872 | bool IsCCNE = CC == ISD::SETNE; |
| 4873 | if (CmpLHS.getOpcode() == ISD::AND && |
| 4874 | isa<ConstantSDNode>(Val: CmpLHS.getOperand(i: 1))) |
| 4875 | switch (CmpLHS.getConstantOperandVal(i: 1)) { |
| 4876 | case 1: |
| 4877 | PCC = IsCCNE ? PPC::PRED_UN : PPC::PRED_NU; |
| 4878 | break; |
| 4879 | case 2: |
| 4880 | PCC = IsCCNE ? PPC::PRED_EQ : PPC::PRED_NE; |
| 4881 | break; |
| 4882 | case 4: |
| 4883 | PCC = IsCCNE ? PPC::PRED_GT : PPC::PRED_LE; |
| 4884 | break; |
| 4885 | case 8: |
| 4886 | PCC = IsCCNE ? PPC::PRED_LT : PPC::PRED_GE; |
| 4887 | break; |
| 4888 | default: |
| 4889 | return false; |
| 4890 | } |
| 4891 | else if (CmpLHS.getOpcode() == ISD::TRUNCATE && |
| 4892 | CmpLHS.getValueType() == MVT::i1) |
| 4893 | PCC = IsCCNE ? PPC::PRED_UN : PPC::PRED_NU; |
| 4894 | |
| 4895 | if (PCC) { |
| 4896 | SDLoc dl(N); |
| 4897 | SDValue Ops[] = {getI32Imm(Imm: PCC, dl), CmpLHS.getOperand(i: 0), N->getOperand(Num: 4), |
| 4898 | N->getOperand(Num: 0)}; |
| 4899 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::BCC, VT: MVT::Other, Ops); |
| 4900 | return true; |
| 4901 | } |
| 4902 | return false; |
| 4903 | } |
| 4904 | |
| 4905 | bool PPCDAGToDAGISel::trySelectLoopCountIntrinsic(SDNode *N) { |
| 4906 | // Sometimes the promoted value of the intrinsic is ANDed by some non-zero |
| 4907 | // value, for example when crbits is disabled. If so, select the |
| 4908 | // loop_decrement intrinsics now. |
| 4909 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 1))->get(); |
| 4910 | SDValue LHS = N->getOperand(Num: 2), RHS = N->getOperand(Num: 3); |
| 4911 | |
| 4912 | if (LHS.getOpcode() != ISD::AND || !isa<ConstantSDNode>(Val: LHS.getOperand(i: 1)) || |
| 4913 | isNullConstant(V: LHS.getOperand(i: 1))) |
| 4914 | return false; |
| 4915 | |
| 4916 | if (LHS.getOperand(i: 0).getOpcode() != ISD::INTRINSIC_W_CHAIN || |
| 4917 | LHS.getOperand(i: 0).getConstantOperandVal(i: 1) != Intrinsic::loop_decrement) |
| 4918 | return false; |
| 4919 | |
| 4920 | if (!isa<ConstantSDNode>(Val: RHS)) |
| 4921 | return false; |
| 4922 | |
| 4923 | assert((CC == ISD::SETEQ || CC == ISD::SETNE) && |
| 4924 | "Counter decrement comparison is not EQ or NE" ); |
| 4925 | |
| 4926 | SDValue OldDecrement = LHS.getOperand(i: 0); |
| 4927 | assert(OldDecrement.hasOneUse() && "loop decrement has more than one use!" ); |
| 4928 | |
| 4929 | SDLoc DecrementLoc(OldDecrement); |
| 4930 | SDValue ChainInput = OldDecrement.getOperand(i: 0); |
| 4931 | SDValue DecrementOps[] = {Subtarget->isPPC64() ? getI64Imm(Imm: 1, dl: DecrementLoc) |
| 4932 | : getI32Imm(Imm: 1, dl: DecrementLoc)}; |
| 4933 | unsigned DecrementOpcode = |
| 4934 | Subtarget->isPPC64() ? PPC::DecreaseCTR8loop : PPC::DecreaseCTRloop; |
| 4935 | SDNode *NewDecrement = CurDAG->getMachineNode(Opcode: DecrementOpcode, dl: DecrementLoc, |
| 4936 | VT: MVT::i1, Ops: DecrementOps); |
| 4937 | |
| 4938 | unsigned Val = RHS->getAsZExtVal(); |
| 4939 | bool IsBranchOnTrue = (CC == ISD::SETEQ && Val) || (CC == ISD::SETNE && !Val); |
| 4940 | unsigned Opcode = IsBranchOnTrue ? PPC::BC : PPC::BCn; |
| 4941 | |
| 4942 | ReplaceUses(F: LHS.getValue(R: 0), T: LHS.getOperand(i: 1)); |
| 4943 | CurDAG->RemoveDeadNode(N: LHS.getNode()); |
| 4944 | |
| 4945 | // Mark the old loop_decrement intrinsic as dead. |
| 4946 | ReplaceUses(F: OldDecrement.getValue(R: 1), T: ChainInput); |
| 4947 | CurDAG->RemoveDeadNode(N: OldDecrement.getNode()); |
| 4948 | |
| 4949 | SDValue Chain = CurDAG->getNode(Opcode: ISD::TokenFactor, DL: SDLoc(N), VT: MVT::Other, |
| 4950 | N1: ChainInput, N2: N->getOperand(Num: 0)); |
| 4951 | |
| 4952 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VT: MVT::Other, Op1: SDValue(NewDecrement, 0), |
| 4953 | Op2: N->getOperand(Num: 4), Op3: Chain); |
| 4954 | return true; |
| 4955 | } |
| 4956 | |
| 4957 | bool PPCDAGToDAGISel::tryAsSingleRLWINM(SDNode *N) { |
| 4958 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 4959 | unsigned Imm; |
| 4960 | if (!isInt32Immediate(N: N->getOperand(Num: 1), Imm)) |
| 4961 | return false; |
| 4962 | |
| 4963 | SDLoc dl(N); |
| 4964 | SDValue Val = N->getOperand(Num: 0); |
| 4965 | unsigned SH, MB, ME; |
| 4966 | // If this is an and of a value rotated between 0 and 31 bits and then and'd |
| 4967 | // with a mask, emit rlwinm |
| 4968 | if (isRotateAndMask(N: Val.getNode(), Mask: Imm, isShiftMask: false, SH, MB, ME)) { |
| 4969 | Val = Val.getOperand(i: 0); |
| 4970 | SDValue Ops[] = {Val, getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl), |
| 4971 | getI32Imm(Imm: ME, dl)}; |
| 4972 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4973 | return true; |
| 4974 | } |
| 4975 | |
| 4976 | // If this is just a masked value where the input is not handled, and |
| 4977 | // is not a rotate-left (handled by a pattern in the .td file), emit rlwinm |
| 4978 | if (isRunOfOnes(Val: Imm, MB, ME) && Val.getOpcode() != ISD::ROTL) { |
| 4979 | // The result of LBARX/LHARX do not need to be cleared as the instructions |
| 4980 | // implicitly clear the upper bits. |
| 4981 | unsigned AlreadyCleared = 0; |
| 4982 | if (Val.getOpcode() == ISD::INTRINSIC_W_CHAIN) { |
| 4983 | auto IntrinsicID = Val.getConstantOperandVal(i: 1); |
| 4984 | if (IntrinsicID == Intrinsic::ppc_lbarx) |
| 4985 | AlreadyCleared = 24; |
| 4986 | else if (IntrinsicID == Intrinsic::ppc_lharx) |
| 4987 | AlreadyCleared = 16; |
| 4988 | if (AlreadyCleared != 0 && AlreadyCleared == MB && ME == 31) { |
| 4989 | ReplaceUses(F: SDValue(N, 0), T: N->getOperand(Num: 0)); |
| 4990 | return true; |
| 4991 | } |
| 4992 | } |
| 4993 | |
| 4994 | SDValue Ops[] = {Val, getI32Imm(Imm: 0, dl), getI32Imm(Imm: MB, dl), |
| 4995 | getI32Imm(Imm: ME, dl)}; |
| 4996 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 4997 | return true; |
| 4998 | } |
| 4999 | |
| 5000 | // AND X, 0 -> 0, not "rlwinm 32". |
| 5001 | if (Imm == 0) { |
| 5002 | ReplaceUses(F: SDValue(N, 0), T: N->getOperand(Num: 1)); |
| 5003 | return true; |
| 5004 | } |
| 5005 | |
| 5006 | return false; |
| 5007 | } |
| 5008 | |
| 5009 | bool PPCDAGToDAGISel::tryAsSingleRLWINM8(SDNode *N) { |
| 5010 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5011 | uint64_t Imm64; |
| 5012 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64)) |
| 5013 | return false; |
| 5014 | |
| 5015 | unsigned MB, ME; |
| 5016 | if (isRunOfOnes64(Val: Imm64, MB, ME) && MB >= 32 && MB <= ME) { |
| 5017 | // MB ME |
| 5018 | // +----------------------+ |
| 5019 | // |xxxxxxxxxxx00011111000| |
| 5020 | // +----------------------+ |
| 5021 | // 0 32 64 |
| 5022 | // We can only do it if the MB is larger than 32 and MB <= ME |
| 5023 | // as RLWINM will replace the contents of [0 - 32) with [32 - 64) even |
| 5024 | // we didn't rotate it. |
| 5025 | SDLoc dl(N); |
| 5026 | SDValue Ops[] = {N->getOperand(Num: 0), getI64Imm(Imm: 0, dl), getI64Imm(Imm: MB - 32, dl), |
| 5027 | getI64Imm(Imm: ME - 32, dl)}; |
| 5028 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM8, VT: MVT::i64, Ops); |
| 5029 | return true; |
| 5030 | } |
| 5031 | |
| 5032 | return false; |
| 5033 | } |
| 5034 | |
| 5035 | bool PPCDAGToDAGISel::tryAsPairOfRLDICL(SDNode *N) { |
| 5036 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5037 | uint64_t Imm64; |
| 5038 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64)) |
| 5039 | return false; |
| 5040 | |
| 5041 | // Do nothing if it is 16-bit imm as the pattern in the .td file handle |
| 5042 | // it well with "andi.". |
| 5043 | if (isUInt<16>(x: Imm64)) |
| 5044 | return false; |
| 5045 | |
| 5046 | SDLoc Loc(N); |
| 5047 | SDValue Val = N->getOperand(Num: 0); |
| 5048 | |
| 5049 | // Optimized with two rldicl's as follows: |
| 5050 | // Add missing bits on left to the mask and check that the mask is a |
| 5051 | // wrapped run of ones, i.e. |
| 5052 | // Change pattern |0001111100000011111111| |
| 5053 | // to |1111111100000011111111|. |
| 5054 | unsigned NumOfLeadingZeros = llvm::countl_zero(Val: Imm64); |
| 5055 | if (NumOfLeadingZeros != 0) |
| 5056 | Imm64 |= maskLeadingOnes<uint64_t>(N: NumOfLeadingZeros); |
| 5057 | |
| 5058 | unsigned MB, ME; |
| 5059 | if (!isRunOfOnes64(Val: Imm64, MB, ME)) |
| 5060 | return false; |
| 5061 | |
| 5062 | // ME MB MB-ME+63 |
| 5063 | // +----------------------+ +----------------------+ |
| 5064 | // |1111111100000011111111| -> |0000001111111111111111| |
| 5065 | // +----------------------+ +----------------------+ |
| 5066 | // 0 63 0 63 |
| 5067 | // There are ME + 1 ones on the left and (MB - ME + 63) & 63 zeros in between. |
| 5068 | unsigned OnesOnLeft = ME + 1; |
| 5069 | unsigned ZerosInBetween = (MB - ME + 63) & 63; |
| 5070 | // Rotate left by OnesOnLeft (so leading ones are now trailing ones) and clear |
| 5071 | // on the left the bits that are already zeros in the mask. |
| 5072 | Val = SDValue(CurDAG->getMachineNode(Opcode: PPC::RLDICL, dl: Loc, VT: MVT::i64, Op1: Val, |
| 5073 | Op2: getI64Imm(Imm: OnesOnLeft, dl: Loc), |
| 5074 | Op3: getI64Imm(Imm: ZerosInBetween, dl: Loc)), |
| 5075 | 0); |
| 5076 | // MB-ME+63 ME MB |
| 5077 | // +----------------------+ +----------------------+ |
| 5078 | // |0000001111111111111111| -> |0001111100000011111111| |
| 5079 | // +----------------------+ +----------------------+ |
| 5080 | // 0 63 0 63 |
| 5081 | // Rotate back by 64 - OnesOnLeft to undo previous rotate. Then clear on the |
| 5082 | // left the number of ones we previously added. |
| 5083 | SDValue Ops[] = {Val, getI64Imm(Imm: 64 - OnesOnLeft, dl: Loc), |
| 5084 | getI64Imm(Imm: NumOfLeadingZeros, dl: Loc)}; |
| 5085 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDICL, VT: MVT::i64, Ops); |
| 5086 | return true; |
| 5087 | } |
| 5088 | |
| 5089 | bool PPCDAGToDAGISel::tryAsSingleRLWIMI(SDNode *N) { |
| 5090 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5091 | unsigned Imm; |
| 5092 | if (!isInt32Immediate(N: N->getOperand(Num: 1), Imm)) |
| 5093 | return false; |
| 5094 | |
| 5095 | SDValue Val = N->getOperand(Num: 0); |
| 5096 | unsigned Imm2; |
| 5097 | // ISD::OR doesn't get all the bitfield insertion fun. |
| 5098 | // (and (or x, c1), c2) where isRunOfOnes(~(c1^c2)) might be a |
| 5099 | // bitfield insert. |
| 5100 | if (Val.getOpcode() != ISD::OR || !isInt32Immediate(N: Val.getOperand(i: 1), Imm&: Imm2)) |
| 5101 | return false; |
| 5102 | |
| 5103 | // The idea here is to check whether this is equivalent to: |
| 5104 | // (c1 & m) | (x & ~m) |
| 5105 | // where m is a run-of-ones mask. The logic here is that, for each bit in |
| 5106 | // c1 and c2: |
| 5107 | // - if both are 1, then the output will be 1. |
| 5108 | // - if both are 0, then the output will be 0. |
| 5109 | // - if the bit in c1 is 0, and the bit in c2 is 1, then the output will |
| 5110 | // come from x. |
| 5111 | // - if the bit in c1 is 1, and the bit in c2 is 0, then the output will |
| 5112 | // be 0. |
| 5113 | // If that last condition is never the case, then we can form m from the |
| 5114 | // bits that are the same between c1 and c2. |
| 5115 | unsigned MB, ME; |
| 5116 | if (isRunOfOnes(Val: ~(Imm ^ Imm2), MB, ME) && !(~Imm & Imm2)) { |
| 5117 | SDLoc dl(N); |
| 5118 | SDValue Ops[] = {Val.getOperand(i: 0), Val.getOperand(i: 1), getI32Imm(Imm: 0, dl), |
| 5119 | getI32Imm(Imm: MB, dl), getI32Imm(Imm: ME, dl)}; |
| 5120 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: PPC::RLWIMI, dl, VT: MVT::i32, Ops)); |
| 5121 | return true; |
| 5122 | } |
| 5123 | |
| 5124 | return false; |
| 5125 | } |
| 5126 | |
| 5127 | bool PPCDAGToDAGISel::tryAsSingleRLDCL(SDNode *N) { |
| 5128 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5129 | |
| 5130 | uint64_t Imm64; |
| 5131 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64) || !isMask_64(Value: Imm64)) |
| 5132 | return false; |
| 5133 | |
| 5134 | SDValue Val = N->getOperand(Num: 0); |
| 5135 | |
| 5136 | if (Val.getOpcode() != ISD::ROTL) |
| 5137 | return false; |
| 5138 | |
| 5139 | // Looking to try to avoid a situation like this one: |
| 5140 | // %2 = tail call i64 @llvm.fshl.i64(i64 %word, i64 %word, i64 23) |
| 5141 | // %and1 = and i64 %2, 9223372036854775807 |
| 5142 | // In this function we are looking to try to match RLDCL. However, the above |
| 5143 | // DAG would better match RLDICL instead which is not what we are looking |
| 5144 | // for here. |
| 5145 | SDValue RotateAmt = Val.getOperand(i: 1); |
| 5146 | if (RotateAmt.getOpcode() == ISD::Constant) |
| 5147 | return false; |
| 5148 | |
| 5149 | unsigned MB = 64 - llvm::countr_one(Value: Imm64); |
| 5150 | SDLoc dl(N); |
| 5151 | SDValue Ops[] = {Val.getOperand(i: 0), RotateAmt, getI32Imm(Imm: MB, dl)}; |
| 5152 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDCL, VT: MVT::i64, Ops); |
| 5153 | return true; |
| 5154 | } |
| 5155 | |
| 5156 | bool PPCDAGToDAGISel::tryAsSingleRLDICL(SDNode *N) { |
| 5157 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5158 | uint64_t Imm64; |
| 5159 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64) || !isMask_64(Value: Imm64)) |
| 5160 | return false; |
| 5161 | |
| 5162 | // If this is a 64-bit zero-extension mask, emit rldicl. |
| 5163 | unsigned MB = 64 - llvm::countr_one(Value: Imm64); |
| 5164 | unsigned SH = 0; |
| 5165 | unsigned Imm; |
| 5166 | SDValue Val = N->getOperand(Num: 0); |
| 5167 | SDLoc dl(N); |
| 5168 | |
| 5169 | if (Val.getOpcode() == ISD::ANY_EXTEND) { |
| 5170 | auto Op0 = Val.getOperand(i: 0); |
| 5171 | if (Op0.getOpcode() == ISD::SRL && |
| 5172 | isInt32Immediate(N: Op0.getOperand(i: 1).getNode(), Imm) && Imm <= MB) { |
| 5173 | |
| 5174 | auto ResultType = Val.getNode()->getValueType(ResNo: 0); |
| 5175 | auto ImDef = CurDAG->getMachineNode(Opcode: PPC::IMPLICIT_DEF, dl, VT: ResultType); |
| 5176 | SDValue IDVal(ImDef, 0); |
| 5177 | |
| 5178 | Val = SDValue(CurDAG->getMachineNode(Opcode: PPC::INSERT_SUBREG, dl, VT: ResultType, |
| 5179 | Op1: IDVal, Op2: Op0.getOperand(i: 0), |
| 5180 | Op3: getI32Imm(Imm: 1, dl)), |
| 5181 | 0); |
| 5182 | SH = 64 - Imm; |
| 5183 | } |
| 5184 | } |
| 5185 | |
| 5186 | // If the operand is a logical right shift, we can fold it into this |
| 5187 | // instruction: rldicl(rldicl(x, 64-n, n), 0, mb) -> rldicl(x, 64-n, mb) |
| 5188 | // for n <= mb. The right shift is really a left rotate followed by a |
| 5189 | // mask, and this mask is a more-restrictive sub-mask of the mask implied |
| 5190 | // by the shift. |
| 5191 | if (Val.getOpcode() == ISD::SRL && |
| 5192 | isInt32Immediate(N: Val.getOperand(i: 1).getNode(), Imm) && Imm <= MB) { |
| 5193 | assert(Imm < 64 && "Illegal shift amount" ); |
| 5194 | Val = Val.getOperand(i: 0); |
| 5195 | SH = 64 - Imm; |
| 5196 | } |
| 5197 | |
| 5198 | SDValue Ops[] = {Val, getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl)}; |
| 5199 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDICL, VT: MVT::i64, Ops); |
| 5200 | return true; |
| 5201 | } |
| 5202 | |
| 5203 | bool PPCDAGToDAGISel::tryAsSingleRLDICR(SDNode *N) { |
| 5204 | assert(N->getOpcode() == ISD::AND && "ISD::AND SDNode expected" ); |
| 5205 | uint64_t Imm64; |
| 5206 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64) || |
| 5207 | !isMask_64(Value: ~Imm64)) |
| 5208 | return false; |
| 5209 | |
| 5210 | // If this is a negated 64-bit zero-extension mask, |
| 5211 | // i.e. the immediate is a sequence of ones from most significant side |
| 5212 | // and all zero for reminder, we should use rldicr. |
| 5213 | unsigned MB = 63 - llvm::countr_one(Value: ~Imm64); |
| 5214 | unsigned SH = 0; |
| 5215 | SDLoc dl(N); |
| 5216 | SDValue Ops[] = {N->getOperand(Num: 0), getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl)}; |
| 5217 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDICR, VT: MVT::i64, Ops); |
| 5218 | return true; |
| 5219 | } |
| 5220 | |
| 5221 | bool PPCDAGToDAGISel::tryAsSingleRLDIMI(SDNode *N) { |
| 5222 | assert(N->getOpcode() == ISD::OR && "ISD::OR SDNode expected" ); |
| 5223 | uint64_t Imm64; |
| 5224 | unsigned MB, ME; |
| 5225 | SDValue N0 = N->getOperand(Num: 0); |
| 5226 | |
| 5227 | // We won't get fewer instructions if the imm is 32-bit integer. |
| 5228 | // rldimi requires the imm to have consecutive ones with both sides zero. |
| 5229 | // Also, make sure the first Op has only one use, otherwise this may increase |
| 5230 | // register pressure since rldimi is destructive. |
| 5231 | if (!isInt64Immediate(N: N->getOperand(Num: 1).getNode(), Imm&: Imm64) || |
| 5232 | isUInt<32>(x: Imm64) || !isRunOfOnes64(Val: Imm64, MB, ME) || !N0.hasOneUse()) |
| 5233 | return false; |
| 5234 | |
| 5235 | unsigned SH = 63 - ME; |
| 5236 | SDLoc Dl(N); |
| 5237 | // Use select64Imm for making LI instr instead of directly putting Imm64 |
| 5238 | SDValue Ops[] = { |
| 5239 | N->getOperand(Num: 0), |
| 5240 | SDValue(selectI64Imm(CurDAG, N: getI64Imm(Imm: -1, dl: Dl).getNode()), 0), |
| 5241 | getI32Imm(Imm: SH, dl: Dl), getI32Imm(Imm: MB, dl: Dl)}; |
| 5242 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDIMI, VT: MVT::i64, Ops); |
| 5243 | return true; |
| 5244 | } |
| 5245 | |
| 5246 | // Select - Convert the specified operand from a target-independent to a |
| 5247 | // target-specific node if it hasn't already been changed. |
| 5248 | void PPCDAGToDAGISel::Select(SDNode *N) { |
| 5249 | SDLoc dl(N); |
| 5250 | if (N->isMachineOpcode()) { |
| 5251 | N->setNodeId(-1); |
| 5252 | return; // Already selected. |
| 5253 | } |
| 5254 | |
| 5255 | // In case any misguided DAG-level optimizations form an ADD with a |
| 5256 | // TargetConstant operand, crash here instead of miscompiling (by selecting |
| 5257 | // an r+r add instead of some kind of r+i add). |
| 5258 | if (N->getOpcode() == ISD::ADD && |
| 5259 | N->getOperand(Num: 1).getOpcode() == ISD::TargetConstant) |
| 5260 | llvm_unreachable("Invalid ADD with TargetConstant operand" ); |
| 5261 | |
| 5262 | // Try matching complex bit permutations before doing anything else. |
| 5263 | if (tryBitPermutation(N)) |
| 5264 | return; |
| 5265 | |
| 5266 | // Try to emit integer compares as GPR-only sequences (i.e. no use of CR). |
| 5267 | if (tryIntCompareInGPR(N)) |
| 5268 | return; |
| 5269 | |
| 5270 | switch (N->getOpcode()) { |
| 5271 | default: break; |
| 5272 | |
| 5273 | case ISD::Constant: |
| 5274 | if (N->getValueType(ResNo: 0) == MVT::i64) { |
| 5275 | ReplaceNode(F: N, T: selectI64Imm(CurDAG, N)); |
| 5276 | return; |
| 5277 | } |
| 5278 | break; |
| 5279 | |
| 5280 | case ISD::INTRINSIC_VOID: { |
| 5281 | auto IntrinsicID = N->getConstantOperandVal(Num: 1); |
| 5282 | if (IntrinsicID != Intrinsic::ppc_tdw && IntrinsicID != Intrinsic::ppc_tw && |
| 5283 | IntrinsicID != Intrinsic::ppc_trapd && |
| 5284 | IntrinsicID != Intrinsic::ppc_trap) |
| 5285 | break; |
| 5286 | unsigned Opcode = (IntrinsicID == Intrinsic::ppc_tdw || |
| 5287 | IntrinsicID == Intrinsic::ppc_trapd) |
| 5288 | ? PPC::TDI |
| 5289 | : PPC::TWI; |
| 5290 | SmallVector<SDValue, 4> OpsWithMD; |
| 5291 | unsigned MDIndex; |
| 5292 | if (IntrinsicID == Intrinsic::ppc_tdw || |
| 5293 | IntrinsicID == Intrinsic::ppc_tw) { |
| 5294 | SDValue Ops[] = {N->getOperand(Num: 4), N->getOperand(Num: 2), N->getOperand(Num: 3)}; |
| 5295 | int16_t SImmOperand2; |
| 5296 | int16_t SImmOperand3; |
| 5297 | int16_t SImmOperand4; |
| 5298 | bool isOperand2IntS16Immediate = |
| 5299 | isIntS16Immediate(Op: N->getOperand(Num: 2), Imm&: SImmOperand2); |
| 5300 | bool isOperand3IntS16Immediate = |
| 5301 | isIntS16Immediate(Op: N->getOperand(Num: 3), Imm&: SImmOperand3); |
| 5302 | // We will emit PPC::TD or PPC::TW if the 2nd and 3rd operands are reg + |
| 5303 | // reg or imm + imm. The imm + imm form will be optimized to either an |
| 5304 | // unconditional trap or a nop in a later pass. |
| 5305 | if (isOperand2IntS16Immediate == isOperand3IntS16Immediate) |
| 5306 | Opcode = IntrinsicID == Intrinsic::ppc_tdw ? PPC::TD : PPC::TW; |
| 5307 | else if (isOperand3IntS16Immediate) |
| 5308 | // The 2nd and 3rd operands are reg + imm. |
| 5309 | Ops[2] = getI32Imm(Imm: int(SImmOperand3) & 0xFFFF, dl); |
| 5310 | else { |
| 5311 | // The 2nd and 3rd operands are imm + reg. |
| 5312 | bool isOperand4IntS16Immediate = |
| 5313 | isIntS16Immediate(Op: N->getOperand(Num: 4), Imm&: SImmOperand4); |
| 5314 | (void)isOperand4IntS16Immediate; |
| 5315 | assert(isOperand4IntS16Immediate && |
| 5316 | "The 4th operand is not an Immediate" ); |
| 5317 | // We need to flip the condition immediate TO. |
| 5318 | int16_t TO = int(SImmOperand4) & 0x1F; |
| 5319 | // We swap the first and second bit of TO if they are not same. |
| 5320 | if ((TO & 0x1) != ((TO & 0x2) >> 1)) |
| 5321 | TO = (TO & 0x1) ? TO + 1 : TO - 1; |
| 5322 | // We swap the fourth and fifth bit of TO if they are not same. |
| 5323 | if ((TO & 0x8) != ((TO & 0x10) >> 1)) |
| 5324 | TO = (TO & 0x8) ? TO + 8 : TO - 8; |
| 5325 | Ops[0] = getI32Imm(Imm: TO, dl); |
| 5326 | Ops[1] = N->getOperand(Num: 3); |
| 5327 | Ops[2] = getI32Imm(Imm: int(SImmOperand2) & 0xFFFF, dl); |
| 5328 | } |
| 5329 | OpsWithMD = {Ops[0], Ops[1], Ops[2]}; |
| 5330 | MDIndex = 5; |
| 5331 | } else { |
| 5332 | OpsWithMD = {getI32Imm(Imm: 24, dl), N->getOperand(Num: 2), getI32Imm(Imm: 0, dl)}; |
| 5333 | MDIndex = 3; |
| 5334 | } |
| 5335 | |
| 5336 | if (N->getNumOperands() > MDIndex) { |
| 5337 | SDValue MDV = N->getOperand(Num: MDIndex); |
| 5338 | const MDNode *MD = cast<MDNodeSDNode>(Val&: MDV)->getMD(); |
| 5339 | assert(MD->getNumOperands() != 0 && "Empty MDNode in operands!" ); |
| 5340 | assert((isa<MDString>(MD->getOperand(0)) && |
| 5341 | cast<MDString>(MD->getOperand(0))->getString() == |
| 5342 | "ppc-trap-reason" ) && |
| 5343 | "Unsupported annotation data type!" ); |
| 5344 | for (unsigned i = 1; i < MD->getNumOperands(); i++) { |
| 5345 | assert(isa<MDString>(MD->getOperand(i)) && |
| 5346 | "Invalid data type for annotation ppc-trap-reason!" ); |
| 5347 | OpsWithMD.push_back( |
| 5348 | Elt: getI32Imm(Imm: std::stoi(str: cast<MDString>( |
| 5349 | Val: MD->getOperand(I: i))->getString().str()), dl)); |
| 5350 | } |
| 5351 | } |
| 5352 | OpsWithMD.push_back(Elt: N->getOperand(Num: 0)); // chain |
| 5353 | CurDAG->SelectNodeTo(N, MachineOpc: Opcode, VT: MVT::Other, Ops: OpsWithMD); |
| 5354 | return; |
| 5355 | } |
| 5356 | |
| 5357 | case ISD::INTRINSIC_WO_CHAIN: { |
| 5358 | // We emit the PPC::FSELS instruction here because of type conflicts with |
| 5359 | // the comparison operand. The FSELS instruction is defined to use an 8-byte |
| 5360 | // comparison like the FSELD version. The fsels intrinsic takes a 4-byte |
| 5361 | // value for the comparison. When selecting through a .td file, a type |
| 5362 | // error is raised. Must check this first so we never break on the |
| 5363 | // !Subtarget->isISA3_1() check. |
| 5364 | auto IntID = N->getConstantOperandVal(Num: 0); |
| 5365 | if (IntID == Intrinsic::ppc_fsels) { |
| 5366 | SDValue Ops[] = {N->getOperand(Num: 1), N->getOperand(Num: 2), N->getOperand(Num: 3)}; |
| 5367 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::FSELS, VT: MVT::f32, Ops); |
| 5368 | return; |
| 5369 | } |
| 5370 | |
| 5371 | if (IntID == Intrinsic::ppc_bcdadd_p || IntID == Intrinsic::ppc_bcdsub_p) { |
| 5372 | auto Pred = N->getConstantOperandVal(Num: 1); |
| 5373 | unsigned Opcode = |
| 5374 | IntID == Intrinsic::ppc_bcdadd_p ? PPC::BCDADD_rec : PPC::BCDSUB_rec; |
| 5375 | unsigned SubReg = 0; |
| 5376 | unsigned ShiftVal = 0; |
| 5377 | bool Reverse = false; |
| 5378 | switch (Pred) { |
| 5379 | case 0: |
| 5380 | SubReg = PPC::sub_eq; |
| 5381 | ShiftVal = 1; |
| 5382 | break; |
| 5383 | case 1: |
| 5384 | SubReg = PPC::sub_eq; |
| 5385 | ShiftVal = 1; |
| 5386 | Reverse = true; |
| 5387 | break; |
| 5388 | case 2: |
| 5389 | SubReg = PPC::sub_lt; |
| 5390 | ShiftVal = 3; |
| 5391 | break; |
| 5392 | case 3: |
| 5393 | SubReg = PPC::sub_lt; |
| 5394 | ShiftVal = 3; |
| 5395 | Reverse = true; |
| 5396 | break; |
| 5397 | case 4: |
| 5398 | SubReg = PPC::sub_gt; |
| 5399 | ShiftVal = 2; |
| 5400 | break; |
| 5401 | case 5: |
| 5402 | SubReg = PPC::sub_gt; |
| 5403 | ShiftVal = 2; |
| 5404 | Reverse = true; |
| 5405 | break; |
| 5406 | case 6: |
| 5407 | SubReg = PPC::sub_un; |
| 5408 | break; |
| 5409 | case 7: |
| 5410 | SubReg = PPC::sub_un; |
| 5411 | Reverse = true; |
| 5412 | break; |
| 5413 | } |
| 5414 | |
| 5415 | EVT VTs[] = {MVT::v16i8, MVT::Glue}; |
| 5416 | SDValue Ops[] = {N->getOperand(Num: 2), N->getOperand(Num: 3), |
| 5417 | CurDAG->getTargetConstant(Val: 0, DL: dl, VT: MVT::i32)}; |
| 5418 | SDValue BCDOp = SDValue(CurDAG->getMachineNode(Opcode, dl, ResultTys: VTs, Ops), 0); |
| 5419 | SDValue CR6Reg = CurDAG->getRegister(Reg: PPC::CR6, VT: MVT::i32); |
| 5420 | // On Power10, we can use SETBC[R]. On prior architectures, we have to use |
| 5421 | // MFOCRF and shift/negate the value. |
| 5422 | if (Subtarget->isISA3_1()) { |
| 5423 | SDValue SubRegIdx = CurDAG->getTargetConstant(Val: SubReg, DL: dl, VT: MVT::i32); |
| 5424 | SDValue CRBit = SDValue( |
| 5425 | CurDAG->getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, dl, VT: MVT::i1, |
| 5426 | Op1: CR6Reg, Op2: SubRegIdx, Op3: BCDOp.getValue(R: 1)), |
| 5427 | 0); |
| 5428 | CurDAG->SelectNodeTo(N, MachineOpc: Reverse ? PPC::SETBCR : PPC::SETBC, VT: MVT::i32, |
| 5429 | Op1: CRBit); |
| 5430 | } else { |
| 5431 | SDValue Move = |
| 5432 | SDValue(CurDAG->getMachineNode(Opcode: PPC::MFOCRF, dl, VT: MVT::i32, Op1: CR6Reg, |
| 5433 | Op2: BCDOp.getValue(R: 1)), |
| 5434 | 0); |
| 5435 | SDValue Ops[] = {Move, getI32Imm(Imm: (32 - (4 + ShiftVal)) & 31, dl), |
| 5436 | getI32Imm(Imm: 31, dl), getI32Imm(Imm: 31, dl)}; |
| 5437 | if (!Reverse) |
| 5438 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 5439 | else { |
| 5440 | SDValue Shift = SDValue( |
| 5441 | CurDAG->getMachineNode(Opcode: PPC::RLWINM, dl, VT: MVT::i32, Ops), 0); |
| 5442 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::XORI, VT: MVT::i32, Op1: Shift, Op2: getI32Imm(Imm: 1, dl)); |
| 5443 | } |
| 5444 | } |
| 5445 | return; |
| 5446 | } |
| 5447 | |
| 5448 | if (!Subtarget->isISA3_1()) |
| 5449 | break; |
| 5450 | unsigned Opcode = 0; |
| 5451 | switch (IntID) { |
| 5452 | default: |
| 5453 | break; |
| 5454 | case Intrinsic::ppc_altivec_vstribr_p: |
| 5455 | Opcode = PPC::VSTRIBR_rec; |
| 5456 | break; |
| 5457 | case Intrinsic::ppc_altivec_vstribl_p: |
| 5458 | Opcode = PPC::VSTRIBL_rec; |
| 5459 | break; |
| 5460 | case Intrinsic::ppc_altivec_vstrihr_p: |
| 5461 | Opcode = PPC::VSTRIHR_rec; |
| 5462 | break; |
| 5463 | case Intrinsic::ppc_altivec_vstrihl_p: |
| 5464 | Opcode = PPC::VSTRIHL_rec; |
| 5465 | break; |
| 5466 | } |
| 5467 | if (!Opcode) |
| 5468 | break; |
| 5469 | |
| 5470 | // Generate the appropriate vector string isolate intrinsic to match. |
| 5471 | EVT VTs[] = {MVT::v16i8, MVT::Glue}; |
| 5472 | SDValue VecStrOp = |
| 5473 | SDValue(CurDAG->getMachineNode(Opcode, dl, ResultTys: VTs, Ops: N->getOperand(Num: 2)), 0); |
| 5474 | // Vector string isolate instructions update the EQ bit of CR6. |
| 5475 | // Generate a SETBC instruction to extract the bit and place it in a GPR. |
| 5476 | SDValue SubRegIdx = CurDAG->getTargetConstant(Val: PPC::sub_eq, DL: dl, VT: MVT::i32); |
| 5477 | SDValue CR6Reg = CurDAG->getRegister(Reg: PPC::CR6, VT: MVT::i32); |
| 5478 | SDValue CRBit = SDValue( |
| 5479 | CurDAG->getMachineNode(Opcode: TargetOpcode::EXTRACT_SUBREG, dl, VT: MVT::i1, |
| 5480 | Op1: CR6Reg, Op2: SubRegIdx, Op3: VecStrOp.getValue(R: 1)), |
| 5481 | 0); |
| 5482 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::SETBC, VT: MVT::i32, Op1: CRBit); |
| 5483 | return; |
| 5484 | } |
| 5485 | |
| 5486 | case ISD::SETCC: |
| 5487 | case ISD::STRICT_FSETCC: |
| 5488 | case ISD::STRICT_FSETCCS: |
| 5489 | if (trySETCC(N)) |
| 5490 | return; |
| 5491 | break; |
| 5492 | // These nodes will be transformed into GETtlsADDR32 node, which |
| 5493 | // later becomes BL_TLS __tls_get_addr(sym at tlsgd)@PLT |
| 5494 | case PPCISD::ADDI_TLSLD_L_ADDR: |
| 5495 | case PPCISD::ADDI_TLSGD_L_ADDR: { |
| 5496 | const Module *Mod = MF->getFunction().getParent(); |
| 5497 | if (PPCLowering->getPointerTy(DL: CurDAG->getDataLayout()) != MVT::i32 || |
| 5498 | !Subtarget->isSecurePlt() || !Subtarget->isTargetELF() || |
| 5499 | Mod->getPICLevel() == PICLevel::SmallPIC) |
| 5500 | break; |
| 5501 | // Attach global base pointer on GETtlsADDR32 node in order to |
| 5502 | // generate secure plt code for TLS symbols. |
| 5503 | getGlobalBaseReg(); |
| 5504 | } break; |
| 5505 | case PPCISD::CALL: |
| 5506 | case PPCISD::CALL_RM: { |
| 5507 | if (Subtarget->isPPC64() || !TM.isPositionIndependent() || |
| 5508 | !Subtarget->isSecurePlt() || !Subtarget->isTargetELF()) |
| 5509 | break; |
| 5510 | |
| 5511 | SDValue Op = N->getOperand(Num: 1); |
| 5512 | |
| 5513 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: Op)) { |
| 5514 | if (GA->getTargetFlags() == PPCII::MO_PLT) |
| 5515 | getGlobalBaseReg(); |
| 5516 | } |
| 5517 | else if (ExternalSymbolSDNode *ES = dyn_cast<ExternalSymbolSDNode>(Val&: Op)) { |
| 5518 | if (ES->getTargetFlags() == PPCII::MO_PLT) |
| 5519 | getGlobalBaseReg(); |
| 5520 | } |
| 5521 | } break; |
| 5522 | |
| 5523 | case PPCISD::GlobalBaseReg: |
| 5524 | ReplaceNode(F: N, T: getGlobalBaseReg()); |
| 5525 | return; |
| 5526 | |
| 5527 | case ISD::FrameIndex: |
| 5528 | selectFrameIndex(SN: N, N); |
| 5529 | return; |
| 5530 | |
| 5531 | case PPCISD::MFOCRF: { |
| 5532 | SDValue InGlue = N->getOperand(Num: 1); |
| 5533 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: PPC::MFOCRF, dl, VT: MVT::i32, |
| 5534 | Op1: N->getOperand(Num: 0), Op2: InGlue)); |
| 5535 | return; |
| 5536 | } |
| 5537 | |
| 5538 | case PPCISD::READ_TIME_BASE: |
| 5539 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: PPC::ReadTB, dl, VT1: MVT::i32, VT2: MVT::i32, |
| 5540 | VT3: MVT::Other, Ops: N->getOperand(Num: 0))); |
| 5541 | return; |
| 5542 | |
| 5543 | case PPCISD::SRA_ADDZE: { |
| 5544 | SDValue N0 = N->getOperand(Num: 0); |
| 5545 | SDValue ShiftAmt = |
| 5546 | CurDAG->getTargetConstant(Val: *cast<ConstantSDNode>(Val: N->getOperand(Num: 1))-> |
| 5547 | getConstantIntValue(), DL: dl, |
| 5548 | VT: N->getValueType(ResNo: 0)); |
| 5549 | if (N->getValueType(ResNo: 0) == MVT::i64) { |
| 5550 | SDNode *Op = |
| 5551 | CurDAG->getMachineNode(Opcode: PPC::SRADI, dl, VT1: MVT::i64, VT2: MVT::Glue, |
| 5552 | Op1: N0, Op2: ShiftAmt); |
| 5553 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::ADDZE8, VT: MVT::i64, Op1: SDValue(Op, 0), |
| 5554 | Op2: SDValue(Op, 1)); |
| 5555 | return; |
| 5556 | } else { |
| 5557 | assert(N->getValueType(0) == MVT::i32 && |
| 5558 | "Expecting i64 or i32 in PPCISD::SRA_ADDZE" ); |
| 5559 | SDNode *Op = |
| 5560 | CurDAG->getMachineNode(Opcode: PPC::SRAWI, dl, VT1: MVT::i32, VT2: MVT::Glue, |
| 5561 | Op1: N0, Op2: ShiftAmt); |
| 5562 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::ADDZE, VT: MVT::i32, Op1: SDValue(Op, 0), |
| 5563 | Op2: SDValue(Op, 1)); |
| 5564 | return; |
| 5565 | } |
| 5566 | } |
| 5567 | |
| 5568 | case ISD::STORE: { |
| 5569 | // Change TLS initial-exec (or TLS local-exec on AIX) D-form stores to |
| 5570 | // X-form stores. |
| 5571 | StoreSDNode *ST = cast<StoreSDNode>(Val: N); |
| 5572 | if (EnableTLSOpt && (Subtarget->isELFv2ABI() || Subtarget->isAIXABI()) && |
| 5573 | ST->getAddressingMode() != ISD::PRE_INC) |
| 5574 | if (tryTLSXFormStore(ST)) |
| 5575 | return; |
| 5576 | break; |
| 5577 | } |
| 5578 | case ISD::LOAD: { |
| 5579 | // Handle preincrement loads. |
| 5580 | LoadSDNode *LD = cast<LoadSDNode>(Val: N); |
| 5581 | EVT LoadedVT = LD->getMemoryVT(); |
| 5582 | |
| 5583 | // Normal loads are handled by code generated from the .td file. |
| 5584 | if (LD->getAddressingMode() != ISD::PRE_INC) { |
| 5585 | // Change TLS initial-exec (or TLS local-exec on AIX) D-form loads to |
| 5586 | // X-form loads. |
| 5587 | if (EnableTLSOpt && (Subtarget->isELFv2ABI() || Subtarget->isAIXABI())) |
| 5588 | if (tryTLSXFormLoad(LD)) |
| 5589 | return; |
| 5590 | break; |
| 5591 | } |
| 5592 | |
| 5593 | SDValue Offset = LD->getOffset(); |
| 5594 | if (Offset.getOpcode() == ISD::TargetConstant || |
| 5595 | Offset.getOpcode() == ISD::TargetGlobalAddress) { |
| 5596 | |
| 5597 | unsigned Opcode; |
| 5598 | bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; |
| 5599 | if (LD->getValueType(ResNo: 0) != MVT::i64) { |
| 5600 | // Handle PPC32 integer and normal FP loads. |
| 5601 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ); |
| 5602 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
| 5603 | default: llvm_unreachable("Invalid PPC load type!" ); |
| 5604 | case MVT::f64: Opcode = PPC::LFDU; break; |
| 5605 | case MVT::f32: Opcode = PPC::LFSU; break; |
| 5606 | case MVT::i32: Opcode = PPC::LWZU; break; |
| 5607 | case MVT::i16: Opcode = isSExt ? PPC::LHAU : PPC::LHZU; break; |
| 5608 | case MVT::i1: |
| 5609 | case MVT::i8: Opcode = PPC::LBZU; break; |
| 5610 | } |
| 5611 | } else { |
| 5612 | assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!" ); |
| 5613 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ); |
| 5614 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
| 5615 | default: llvm_unreachable("Invalid PPC load type!" ); |
| 5616 | case MVT::i64: Opcode = PPC::LDU; break; |
| 5617 | case MVT::i32: Opcode = PPC::LWZU8; break; |
| 5618 | case MVT::i16: Opcode = isSExt ? PPC::LHAU8 : PPC::LHZU8; break; |
| 5619 | case MVT::i1: |
| 5620 | case MVT::i8: Opcode = PPC::LBZU8; break; |
| 5621 | } |
| 5622 | } |
| 5623 | |
| 5624 | SDValue Chain = LD->getChain(); |
| 5625 | SDValue Base = LD->getBasePtr(); |
| 5626 | SDValue Ops[] = { Offset, Base, Chain }; |
| 5627 | SDNode *MN = CurDAG->getMachineNode( |
| 5628 | Opcode, dl, VT1: LD->getValueType(ResNo: 0), |
| 5629 | VT2: PPCLowering->getPointerTy(DL: CurDAG->getDataLayout()), VT3: MVT::Other, Ops); |
| 5630 | transferMemOperands(N, Result: MN); |
| 5631 | ReplaceNode(F: N, T: MN); |
| 5632 | return; |
| 5633 | } else { |
| 5634 | unsigned Opcode; |
| 5635 | bool isSExt = LD->getExtensionType() == ISD::SEXTLOAD; |
| 5636 | if (LD->getValueType(ResNo: 0) != MVT::i64) { |
| 5637 | // Handle PPC32 integer and normal FP loads. |
| 5638 | assert((!isSExt || LoadedVT == MVT::i16) && "Invalid sext update load" ); |
| 5639 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
| 5640 | default: llvm_unreachable("Invalid PPC load type!" ); |
| 5641 | case MVT::f64: Opcode = PPC::LFDUX; break; |
| 5642 | case MVT::f32: Opcode = PPC::LFSUX; break; |
| 5643 | case MVT::i32: Opcode = PPC::LWZUX; break; |
| 5644 | case MVT::i16: Opcode = isSExt ? PPC::LHAUX : PPC::LHZUX; break; |
| 5645 | case MVT::i1: |
| 5646 | case MVT::i8: Opcode = PPC::LBZUX; break; |
| 5647 | } |
| 5648 | } else { |
| 5649 | assert(LD->getValueType(0) == MVT::i64 && "Unknown load result type!" ); |
| 5650 | assert((!isSExt || LoadedVT == MVT::i16 || LoadedVT == MVT::i32) && |
| 5651 | "Invalid sext update load" ); |
| 5652 | switch (LoadedVT.getSimpleVT().SimpleTy) { |
| 5653 | default: llvm_unreachable("Invalid PPC load type!" ); |
| 5654 | case MVT::i64: Opcode = PPC::LDUX; break; |
| 5655 | case MVT::i32: Opcode = isSExt ? PPC::LWAUX : PPC::LWZUX8; break; |
| 5656 | case MVT::i16: Opcode = isSExt ? PPC::LHAUX8 : PPC::LHZUX8; break; |
| 5657 | case MVT::i1: |
| 5658 | case MVT::i8: Opcode = PPC::LBZUX8; break; |
| 5659 | } |
| 5660 | } |
| 5661 | |
| 5662 | SDValue Chain = LD->getChain(); |
| 5663 | SDValue Base = LD->getBasePtr(); |
| 5664 | SDValue Ops[] = { Base, Offset, Chain }; |
| 5665 | SDNode *MN = CurDAG->getMachineNode( |
| 5666 | Opcode, dl, VT1: LD->getValueType(ResNo: 0), |
| 5667 | VT2: PPCLowering->getPointerTy(DL: CurDAG->getDataLayout()), VT3: MVT::Other, Ops); |
| 5668 | transferMemOperands(N, Result: MN); |
| 5669 | ReplaceNode(F: N, T: MN); |
| 5670 | return; |
| 5671 | } |
| 5672 | } |
| 5673 | |
| 5674 | case ISD::AND: |
| 5675 | // If this is an 'and' with a mask, try to emit rlwinm/rldicl/rldicr |
| 5676 | if (tryAsSingleRLWINM(N) || tryAsSingleRLWIMI(N) || tryAsSingleRLDCL(N) || |
| 5677 | tryAsSingleRLDICL(N) || tryAsSingleRLDICR(N) || tryAsSingleRLWINM8(N) || |
| 5678 | tryAsPairOfRLDICL(N)) |
| 5679 | return; |
| 5680 | |
| 5681 | // Other cases are autogenerated. |
| 5682 | break; |
| 5683 | case ISD::OR: { |
| 5684 | if (N->getValueType(ResNo: 0) == MVT::i32) |
| 5685 | if (tryBitfieldInsert(N)) |
| 5686 | return; |
| 5687 | |
| 5688 | int16_t Imm; |
| 5689 | if (N->getOperand(Num: 0)->getOpcode() == ISD::FrameIndex && |
| 5690 | isIntS16Immediate(Op: N->getOperand(Num: 1), Imm)) { |
| 5691 | KnownBits LHSKnown = CurDAG->computeKnownBits(Op: N->getOperand(Num: 0)); |
| 5692 | |
| 5693 | // If this is equivalent to an add, then we can fold it with the |
| 5694 | // FrameIndex calculation. |
| 5695 | if ((LHSKnown.Zero.getZExtValue()|~(uint64_t)Imm) == ~0ULL) { |
| 5696 | selectFrameIndex(SN: N, N: N->getOperand(Num: 0).getNode(), Offset: (int64_t)Imm); |
| 5697 | return; |
| 5698 | } |
| 5699 | } |
| 5700 | |
| 5701 | // If this is 'or' against an imm with consecutive ones and both sides zero, |
| 5702 | // try to emit rldimi |
| 5703 | if (tryAsSingleRLDIMI(N)) |
| 5704 | return; |
| 5705 | |
| 5706 | // OR with a 32-bit immediate can be handled by ori + oris |
| 5707 | // without creating an immediate in a GPR. |
| 5708 | uint64_t Imm64 = 0; |
| 5709 | bool IsPPC64 = Subtarget->isPPC64(); |
| 5710 | if (IsPPC64 && isInt64Immediate(N: N->getOperand(Num: 1), Imm&: Imm64) && |
| 5711 | (Imm64 & ~0xFFFFFFFFuLL) == 0) { |
| 5712 | // If ImmHi (ImmHi) is zero, only one ori (oris) is generated later. |
| 5713 | uint64_t ImmHi = Imm64 >> 16; |
| 5714 | uint64_t ImmLo = Imm64 & 0xFFFF; |
| 5715 | if (ImmHi != 0 && ImmLo != 0) { |
| 5716 | SDNode *Lo = CurDAG->getMachineNode(Opcode: PPC::ORI8, dl, VT: MVT::i64, |
| 5717 | Op1: N->getOperand(Num: 0), |
| 5718 | Op2: getI16Imm(Imm: ImmLo, dl)); |
| 5719 | SDValue Ops1[] = { SDValue(Lo, 0), getI16Imm(Imm: ImmHi, dl)}; |
| 5720 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::ORIS8, VT: MVT::i64, Ops: Ops1); |
| 5721 | return; |
| 5722 | } |
| 5723 | } |
| 5724 | |
| 5725 | // Other cases are autogenerated. |
| 5726 | break; |
| 5727 | } |
| 5728 | case ISD::XOR: { |
| 5729 | // XOR with a 32-bit immediate can be handled by xori + xoris |
| 5730 | // without creating an immediate in a GPR. |
| 5731 | uint64_t Imm64 = 0; |
| 5732 | bool IsPPC64 = Subtarget->isPPC64(); |
| 5733 | if (IsPPC64 && isInt64Immediate(N: N->getOperand(Num: 1), Imm&: Imm64) && |
| 5734 | (Imm64 & ~0xFFFFFFFFuLL) == 0) { |
| 5735 | // If ImmHi (ImmHi) is zero, only one xori (xoris) is generated later. |
| 5736 | uint64_t ImmHi = Imm64 >> 16; |
| 5737 | uint64_t ImmLo = Imm64 & 0xFFFF; |
| 5738 | if (ImmHi != 0 && ImmLo != 0) { |
| 5739 | SDNode *Lo = CurDAG->getMachineNode(Opcode: PPC::XORI8, dl, VT: MVT::i64, |
| 5740 | Op1: N->getOperand(Num: 0), |
| 5741 | Op2: getI16Imm(Imm: ImmLo, dl)); |
| 5742 | SDValue Ops1[] = { SDValue(Lo, 0), getI16Imm(Imm: ImmHi, dl)}; |
| 5743 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::XORIS8, VT: MVT::i64, Ops: Ops1); |
| 5744 | return; |
| 5745 | } |
| 5746 | } |
| 5747 | |
| 5748 | break; |
| 5749 | } |
| 5750 | case ISD::ADD: { |
| 5751 | int16_t Imm; |
| 5752 | if (N->getOperand(Num: 0)->getOpcode() == ISD::FrameIndex && |
| 5753 | isIntS16Immediate(Op: N->getOperand(Num: 1), Imm)) { |
| 5754 | selectFrameIndex(SN: N, N: N->getOperand(Num: 0).getNode(), Offset: (int64_t)Imm); |
| 5755 | return; |
| 5756 | } |
| 5757 | |
| 5758 | break; |
| 5759 | } |
| 5760 | case ISD::SHL: { |
| 5761 | unsigned Imm, SH, MB, ME; |
| 5762 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::AND, Imm) && |
| 5763 | isRotateAndMask(N, Mask: Imm, isShiftMask: true, SH, MB, ME)) { |
| 5764 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
| 5765 | getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl), |
| 5766 | getI32Imm(Imm: ME, dl) }; |
| 5767 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 5768 | return; |
| 5769 | } |
| 5770 | |
| 5771 | // Other cases are autogenerated. |
| 5772 | break; |
| 5773 | } |
| 5774 | case ISD::SRL: { |
| 5775 | unsigned Imm, SH, MB, ME; |
| 5776 | if (isOpcWithIntImmediate(N: N->getOperand(Num: 0).getNode(), Opc: ISD::AND, Imm) && |
| 5777 | isRotateAndMask(N, Mask: Imm, isShiftMask: true, SH, MB, ME)) { |
| 5778 | SDValue Ops[] = { N->getOperand(Num: 0).getOperand(i: 0), |
| 5779 | getI32Imm(Imm: SH, dl), getI32Imm(Imm: MB, dl), |
| 5780 | getI32Imm(Imm: ME, dl) }; |
| 5781 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 5782 | return; |
| 5783 | } |
| 5784 | |
| 5785 | // Other cases are autogenerated. |
| 5786 | break; |
| 5787 | } |
| 5788 | case ISD::MUL: { |
| 5789 | SDValue Op1 = N->getOperand(Num: 1); |
| 5790 | if (Op1.getOpcode() != ISD::Constant || |
| 5791 | (Op1.getValueType() != MVT::i64 && Op1.getValueType() != MVT::i32)) |
| 5792 | break; |
| 5793 | |
| 5794 | // If the multiplier fits int16, we can handle it with mulli. |
| 5795 | int64_t Imm = Op1->getAsZExtVal(); |
| 5796 | unsigned Shift = llvm::countr_zero<uint64_t>(Val: Imm); |
| 5797 | if (isInt<16>(x: Imm) || !Shift) |
| 5798 | break; |
| 5799 | |
| 5800 | // If the shifted value fits int16, we can do this transformation: |
| 5801 | // (mul X, c1 << c2) -> (rldicr (mulli X, c1) c2). We do this in ISEL due to |
| 5802 | // DAGCombiner prefers (shl (mul X, c1), c2) -> (mul X, c1 << c2). |
| 5803 | uint64_t ImmSh = Imm >> Shift; |
| 5804 | if (!isInt<16>(x: ImmSh)) |
| 5805 | break; |
| 5806 | |
| 5807 | uint64_t SextImm = SignExtend64(X: ImmSh & 0xFFFF, B: 16); |
| 5808 | if (Op1.getValueType() == MVT::i64) { |
| 5809 | SDValue SDImm = CurDAG->getTargetConstant(Val: SextImm, DL: dl, VT: MVT::i64); |
| 5810 | SDNode *MulNode = CurDAG->getMachineNode(Opcode: PPC::MULLI8, dl, VT: MVT::i64, |
| 5811 | Op1: N->getOperand(Num: 0), Op2: SDImm); |
| 5812 | |
| 5813 | SDValue Ops[] = {SDValue(MulNode, 0), getI32Imm(Imm: Shift, dl), |
| 5814 | getI32Imm(Imm: 63 - Shift, dl)}; |
| 5815 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLDICR, VT: MVT::i64, Ops); |
| 5816 | return; |
| 5817 | } else { |
| 5818 | SDValue SDImm = CurDAG->getTargetConstant(Val: SextImm, DL: dl, VT: MVT::i32); |
| 5819 | SDNode *MulNode = CurDAG->getMachineNode(Opcode: PPC::MULLI, dl, VT: MVT::i32, |
| 5820 | Op1: N->getOperand(Num: 0), Op2: SDImm); |
| 5821 | |
| 5822 | SDValue Ops[] = {SDValue(MulNode, 0), getI32Imm(Imm: Shift, dl), |
| 5823 | getI32Imm(Imm: 0, dl), getI32Imm(Imm: 31 - Shift, dl)}; |
| 5824 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::RLWINM, VT: MVT::i32, Ops); |
| 5825 | return; |
| 5826 | } |
| 5827 | break; |
| 5828 | } |
| 5829 | // FIXME: Remove this once the ANDI glue bug is fixed: |
| 5830 | case PPCISD::ANDI_rec_1_EQ_BIT: |
| 5831 | case PPCISD::ANDI_rec_1_GT_BIT: { |
| 5832 | if (!ANDIGlueBug) |
| 5833 | break; |
| 5834 | |
| 5835 | EVT InVT = N->getOperand(Num: 0).getValueType(); |
| 5836 | assert((InVT == MVT::i64 || InVT == MVT::i32) && |
| 5837 | "Invalid input type for ANDI_rec_1_EQ_BIT" ); |
| 5838 | |
| 5839 | unsigned Opcode = (InVT == MVT::i64) ? PPC::ANDI8_rec : PPC::ANDI_rec; |
| 5840 | SDValue AndI(CurDAG->getMachineNode(Opcode, dl, VT1: InVT, VT2: MVT::Glue, |
| 5841 | Op1: N->getOperand(Num: 0), |
| 5842 | Op2: CurDAG->getTargetConstant(Val: 1, DL: dl, VT: InVT)), |
| 5843 | 0); |
| 5844 | SDValue CR0Reg = CurDAG->getRegister(Reg: PPC::CR0, VT: MVT::i32); |
| 5845 | SDValue SRIdxVal = CurDAG->getTargetConstant( |
| 5846 | Val: N->getOpcode() == PPCISD::ANDI_rec_1_EQ_BIT ? PPC::sub_eq : PPC::sub_gt, |
| 5847 | DL: dl, VT: MVT::i32); |
| 5848 | |
| 5849 | CurDAG->SelectNodeTo(N, MachineOpc: TargetOpcode::EXTRACT_SUBREG, VT: MVT::i1, Op1: CR0Reg, |
| 5850 | Op2: SRIdxVal, Op3: SDValue(AndI.getNode(), 1) /* glue */); |
| 5851 | return; |
| 5852 | } |
| 5853 | case ISD::SELECT_CC: { |
| 5854 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 4))->get(); |
| 5855 | EVT PtrVT = |
| 5856 | CurDAG->getTargetLoweringInfo().getPointerTy(DL: CurDAG->getDataLayout()); |
| 5857 | bool isPPC64 = (PtrVT == MVT::i64); |
| 5858 | |
| 5859 | // If this is a select of i1 operands, we'll pattern match it. |
| 5860 | if (Subtarget->useCRBits() && N->getOperand(Num: 0).getValueType() == MVT::i1) |
| 5861 | break; |
| 5862 | |
| 5863 | if (Subtarget->isISA3_0() && Subtarget->isPPC64()) { |
| 5864 | bool NeedSwapOps = false; |
| 5865 | bool IsUnCmp = false; |
| 5866 | if (mayUseP9Setb(N, CC, DAG: CurDAG, NeedSwapOps, IsUnCmp)) { |
| 5867 | SDValue LHS = N->getOperand(Num: 0); |
| 5868 | SDValue RHS = N->getOperand(Num: 1); |
| 5869 | if (NeedSwapOps) |
| 5870 | std::swap(a&: LHS, b&: RHS); |
| 5871 | |
| 5872 | // Make use of SelectCC to generate the comparison to set CR bits, for |
| 5873 | // equality comparisons having one literal operand, SelectCC probably |
| 5874 | // doesn't need to materialize the whole literal and just use xoris to |
| 5875 | // check it first, it leads the following comparison result can't |
| 5876 | // exactly represent GT/LT relationship. So to avoid this we specify |
| 5877 | // SETGT/SETUGT here instead of SETEQ. |
| 5878 | SDValue GenCC = |
| 5879 | SelectCC(LHS, RHS, CC: IsUnCmp ? ISD::SETUGT : ISD::SETGT, dl); |
| 5880 | CurDAG->SelectNodeTo( |
| 5881 | N, MachineOpc: N->getSimpleValueType(ResNo: 0) == MVT::i64 ? PPC::SETB8 : PPC::SETB, |
| 5882 | VT: N->getValueType(ResNo: 0), Op1: GenCC); |
| 5883 | NumP9Setb++; |
| 5884 | return; |
| 5885 | } |
| 5886 | } |
| 5887 | |
| 5888 | // Handle the setcc cases here. select_cc lhs, 0, 1, 0, cc |
| 5889 | if (!isPPC64 && isNullConstant(V: N->getOperand(Num: 1)) && |
| 5890 | isOneConstant(V: N->getOperand(Num: 2)) && isNullConstant(V: N->getOperand(Num: 3)) && |
| 5891 | CC == ISD::SETNE && |
| 5892 | // FIXME: Implement this optzn for PPC64. |
| 5893 | N->getValueType(ResNo: 0) == MVT::i32) { |
| 5894 | SDNode *Tmp = |
| 5895 | CurDAG->getMachineNode(Opcode: PPC::ADDIC, dl, VT1: MVT::i32, VT2: MVT::Glue, |
| 5896 | Op1: N->getOperand(Num: 0), Op2: getI32Imm(Imm: ~0U, dl)); |
| 5897 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::SUBFE, VT: MVT::i32, Op1: SDValue(Tmp, 0), |
| 5898 | Op2: N->getOperand(Num: 0), Op3: SDValue(Tmp, 1)); |
| 5899 | return; |
| 5900 | } |
| 5901 | |
| 5902 | SDValue CCReg = SelectCC(LHS: N->getOperand(Num: 0), RHS: N->getOperand(Num: 1), CC, dl); |
| 5903 | |
| 5904 | if (N->getValueType(ResNo: 0) == MVT::i1) { |
| 5905 | // An i1 select is: (c & t) | (!c & f). |
| 5906 | bool Inv; |
| 5907 | unsigned Idx = getCRIdxForSetCC(CC, Invert&: Inv); |
| 5908 | |
| 5909 | unsigned SRI; |
| 5910 | switch (Idx) { |
| 5911 | default: llvm_unreachable("Invalid CC index" ); |
| 5912 | case 0: SRI = PPC::sub_lt; break; |
| 5913 | case 1: SRI = PPC::sub_gt; break; |
| 5914 | case 2: SRI = PPC::sub_eq; break; |
| 5915 | case 3: SRI = PPC::sub_un; break; |
| 5916 | } |
| 5917 | |
| 5918 | SDValue CCBit = CurDAG->getTargetExtractSubreg(SRIdx: SRI, DL: dl, VT: MVT::i1, Operand: CCReg); |
| 5919 | |
| 5920 | SDValue NotCCBit(CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl, VT: MVT::i1, |
| 5921 | Op1: CCBit, Op2: CCBit), 0); |
| 5922 | SDValue C = Inv ? NotCCBit : CCBit, |
| 5923 | NotC = Inv ? CCBit : NotCCBit; |
| 5924 | |
| 5925 | SDValue CAndT(CurDAG->getMachineNode(Opcode: PPC::CRAND, dl, VT: MVT::i1, |
| 5926 | Op1: C, Op2: N->getOperand(Num: 2)), 0); |
| 5927 | SDValue NotCAndF(CurDAG->getMachineNode(Opcode: PPC::CRAND, dl, VT: MVT::i1, |
| 5928 | Op1: NotC, Op2: N->getOperand(Num: 3)), 0); |
| 5929 | |
| 5930 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::CROR, VT: MVT::i1, Op1: CAndT, Op2: NotCAndF); |
| 5931 | return; |
| 5932 | } |
| 5933 | |
| 5934 | unsigned BROpc = |
| 5935 | getPredicateForSetCC(CC, VT: N->getOperand(Num: 0).getValueType(), Subtarget); |
| 5936 | |
| 5937 | unsigned SelectCCOp; |
| 5938 | if (N->getValueType(ResNo: 0) == MVT::i32) |
| 5939 | SelectCCOp = PPC::SELECT_CC_I4; |
| 5940 | else if (N->getValueType(ResNo: 0) == MVT::i64) |
| 5941 | SelectCCOp = PPC::SELECT_CC_I8; |
| 5942 | else if (N->getValueType(ResNo: 0) == MVT::f32) { |
| 5943 | if (Subtarget->hasP8Vector()) |
| 5944 | SelectCCOp = PPC::SELECT_CC_VSSRC; |
| 5945 | else if (Subtarget->hasSPE()) |
| 5946 | SelectCCOp = PPC::SELECT_CC_SPE4; |
| 5947 | else |
| 5948 | SelectCCOp = PPC::SELECT_CC_F4; |
| 5949 | } else if (N->getValueType(ResNo: 0) == MVT::f64) { |
| 5950 | if (Subtarget->hasVSX()) |
| 5951 | SelectCCOp = PPC::SELECT_CC_VSFRC; |
| 5952 | else if (Subtarget->hasSPE()) |
| 5953 | SelectCCOp = PPC::SELECT_CC_SPE; |
| 5954 | else |
| 5955 | SelectCCOp = PPC::SELECT_CC_F8; |
| 5956 | } else if (N->getValueType(ResNo: 0) == MVT::f128) |
| 5957 | SelectCCOp = PPC::SELECT_CC_F16; |
| 5958 | else if (Subtarget->hasSPE()) |
| 5959 | SelectCCOp = PPC::SELECT_CC_SPE; |
| 5960 | else if (N->getValueType(ResNo: 0) == MVT::v2f64 || |
| 5961 | N->getValueType(ResNo: 0) == MVT::v2i64) |
| 5962 | SelectCCOp = PPC::SELECT_CC_VSRC; |
| 5963 | else |
| 5964 | SelectCCOp = PPC::SELECT_CC_VRRC; |
| 5965 | |
| 5966 | SDValue Ops[] = { CCReg, N->getOperand(Num: 2), N->getOperand(Num: 3), |
| 5967 | getI32Imm(Imm: BROpc, dl) }; |
| 5968 | CurDAG->SelectNodeTo(N, MachineOpc: SelectCCOp, VT: N->getValueType(ResNo: 0), Ops); |
| 5969 | return; |
| 5970 | } |
| 5971 | case ISD::VECTOR_SHUFFLE: |
| 5972 | if (Subtarget->hasVSX() && (N->getValueType(ResNo: 0) == MVT::v2f64 || |
| 5973 | N->getValueType(ResNo: 0) == MVT::v2i64)) { |
| 5974 | ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Val: N); |
| 5975 | |
| 5976 | SDValue Op1 = N->getOperand(Num: SVN->getMaskElt(Idx: 0) < 2 ? 0 : 1), |
| 5977 | Op2 = N->getOperand(Num: SVN->getMaskElt(Idx: 1) < 2 ? 0 : 1); |
| 5978 | unsigned DM[2]; |
| 5979 | |
| 5980 | for (int i = 0; i < 2; ++i) |
| 5981 | if (SVN->getMaskElt(Idx: i) <= 0 || SVN->getMaskElt(Idx: i) == 2) |
| 5982 | DM[i] = 0; |
| 5983 | else |
| 5984 | DM[i] = 1; |
| 5985 | |
| 5986 | if (Op1 == Op2 && DM[0] == 0 && DM[1] == 0 && |
| 5987 | Op1.getOpcode() == ISD::SCALAR_TO_VECTOR && |
| 5988 | isa<LoadSDNode>(Val: Op1.getOperand(i: 0))) { |
| 5989 | LoadSDNode *LD = cast<LoadSDNode>(Val: Op1.getOperand(i: 0)); |
| 5990 | SDValue Base, Offset; |
| 5991 | |
| 5992 | if (LD->isUnindexed() && LD->hasOneUse() && Op1.hasOneUse() && |
| 5993 | (LD->getMemoryVT() == MVT::f64 || |
| 5994 | LD->getMemoryVT() == MVT::i64) && |
| 5995 | SelectAddrIdxOnly(N: LD->getBasePtr(), Base, Index&: Offset)) { |
| 5996 | SDValue Chain = LD->getChain(); |
| 5997 | SDValue Ops[] = { Base, Offset, Chain }; |
| 5998 | MachineMemOperand *MemOp = LD->getMemOperand(); |
| 5999 | SDNode *NewN = CurDAG->SelectNodeTo(N, MachineOpc: PPC::LXVDSX, |
| 6000 | VT: N->getValueType(ResNo: 0), Ops); |
| 6001 | CurDAG->setNodeMemRefs(N: cast<MachineSDNode>(Val: NewN), NewMemRefs: {MemOp}); |
| 6002 | return; |
| 6003 | } |
| 6004 | } |
| 6005 | |
| 6006 | // For little endian, we must swap the input operands and adjust |
| 6007 | // the mask elements (reverse and invert them). |
| 6008 | if (Subtarget->isLittleEndian()) { |
| 6009 | std::swap(a&: Op1, b&: Op2); |
| 6010 | unsigned tmp = DM[0]; |
| 6011 | DM[0] = 1 - DM[1]; |
| 6012 | DM[1] = 1 - tmp; |
| 6013 | } |
| 6014 | |
| 6015 | SDValue DMV = CurDAG->getTargetConstant(Val: DM[1] | (DM[0] << 1), DL: dl, |
| 6016 | VT: MVT::i32); |
| 6017 | SDValue Ops[] = { Op1, Op2, DMV }; |
| 6018 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::XXPERMDI, VT: N->getValueType(ResNo: 0), Ops); |
| 6019 | return; |
| 6020 | } |
| 6021 | |
| 6022 | break; |
| 6023 | case PPCISD::BDNZ: |
| 6024 | case PPCISD::BDZ: { |
| 6025 | bool IsPPC64 = Subtarget->isPPC64(); |
| 6026 | SDValue Ops[] = { N->getOperand(Num: 1), N->getOperand(Num: 0) }; |
| 6027 | CurDAG->SelectNodeTo(N, MachineOpc: N->getOpcode() == PPCISD::BDNZ |
| 6028 | ? (IsPPC64 ? PPC::BDNZ8 : PPC::BDNZ) |
| 6029 | : (IsPPC64 ? PPC::BDZ8 : PPC::BDZ), |
| 6030 | VT: MVT::Other, Ops); |
| 6031 | return; |
| 6032 | } |
| 6033 | case PPCISD::COND_BRANCH: { |
| 6034 | // Op #0 is the Chain. |
| 6035 | // Op #1 is the PPC::PRED_* number. |
| 6036 | // Op #2 is the CR# |
| 6037 | // Op #3 is the Dest MBB |
| 6038 | // Op #4 is the Flag. |
| 6039 | // Prevent PPC::PRED_* from being selected into LI. |
| 6040 | unsigned PCC = N->getConstantOperandVal(Num: 1); |
| 6041 | if (EnableBranchHint) |
| 6042 | PCC |= getBranchHint(PCC, FuncInfo: *FuncInfo, DestMBB: N->getOperand(Num: 3)); |
| 6043 | |
| 6044 | SDValue Pred = getI32Imm(Imm: PCC, dl); |
| 6045 | SDValue Ops[] = { Pred, N->getOperand(Num: 2), N->getOperand(Num: 3), |
| 6046 | N->getOperand(Num: 0), N->getOperand(Num: 4) }; |
| 6047 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::BCC, VT: MVT::Other, Ops); |
| 6048 | return; |
| 6049 | } |
| 6050 | case ISD::BR_CC: { |
| 6051 | if (tryFoldSWTestBRCC(N)) |
| 6052 | return; |
| 6053 | if (trySelectLoopCountIntrinsic(N)) |
| 6054 | return; |
| 6055 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: N->getOperand(Num: 1))->get(); |
| 6056 | unsigned PCC = |
| 6057 | getPredicateForSetCC(CC, VT: N->getOperand(Num: 2).getValueType(), Subtarget); |
| 6058 | |
| 6059 | if (N->getOperand(Num: 2).getValueType() == MVT::i1) { |
| 6060 | unsigned Opc; |
| 6061 | bool Swap; |
| 6062 | switch (PCC) { |
| 6063 | default: llvm_unreachable("Unexpected Boolean-operand predicate" ); |
| 6064 | case PPC::PRED_LT: Opc = PPC::CRANDC; Swap = true; break; |
| 6065 | case PPC::PRED_LE: Opc = PPC::CRORC; Swap = true; break; |
| 6066 | case PPC::PRED_EQ: Opc = PPC::CREQV; Swap = false; break; |
| 6067 | case PPC::PRED_GE: Opc = PPC::CRORC; Swap = false; break; |
| 6068 | case PPC::PRED_GT: Opc = PPC::CRANDC; Swap = false; break; |
| 6069 | case PPC::PRED_NE: Opc = PPC::CRXOR; Swap = false; break; |
| 6070 | } |
| 6071 | |
| 6072 | // A signed comparison of i1 values produces the opposite result to an |
| 6073 | // unsigned one if the condition code includes less-than or greater-than. |
| 6074 | // This is because 1 is the most negative signed i1 number and the most |
| 6075 | // positive unsigned i1 number. The CR-logical operations used for such |
| 6076 | // comparisons are non-commutative so for signed comparisons vs. unsigned |
| 6077 | // ones, the input operands just need to be swapped. |
| 6078 | if (ISD::isSignedIntSetCC(Code: CC)) |
| 6079 | Swap = !Swap; |
| 6080 | |
| 6081 | SDValue BitComp(CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::i1, |
| 6082 | Op1: N->getOperand(Num: Swap ? 3 : 2), |
| 6083 | Op2: N->getOperand(Num: Swap ? 2 : 3)), 0); |
| 6084 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::BC, VT: MVT::Other, Op1: BitComp, Op2: N->getOperand(Num: 4), |
| 6085 | Op3: N->getOperand(Num: 0)); |
| 6086 | return; |
| 6087 | } |
| 6088 | |
| 6089 | if (EnableBranchHint) |
| 6090 | PCC |= getBranchHint(PCC, FuncInfo: *FuncInfo, DestMBB: N->getOperand(Num: 4)); |
| 6091 | |
| 6092 | SDValue CondCode = SelectCC(LHS: N->getOperand(Num: 2), RHS: N->getOperand(Num: 3), CC, dl); |
| 6093 | SDValue Ops[] = { getI32Imm(Imm: PCC, dl), CondCode, |
| 6094 | N->getOperand(Num: 4), N->getOperand(Num: 0) }; |
| 6095 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::BCC, VT: MVT::Other, Ops); |
| 6096 | return; |
| 6097 | } |
| 6098 | case ISD::BRIND: { |
| 6099 | // FIXME: Should custom lower this. |
| 6100 | SDValue Chain = N->getOperand(Num: 0); |
| 6101 | SDValue Target = N->getOperand(Num: 1); |
| 6102 | unsigned Opc = Target.getValueType() == MVT::i32 ? PPC::MTCTR : PPC::MTCTR8; |
| 6103 | unsigned Reg = Target.getValueType() == MVT::i32 ? PPC::BCTR : PPC::BCTR8; |
| 6104 | Chain = SDValue(CurDAG->getMachineNode(Opcode: Opc, dl, VT: MVT::Glue, Op1: Target, |
| 6105 | Op2: Chain), 0); |
| 6106 | CurDAG->SelectNodeTo(N, MachineOpc: Reg, VT: MVT::Other, Op1: Chain); |
| 6107 | return; |
| 6108 | } |
| 6109 | case PPCISD::TOC_ENTRY: { |
| 6110 | const bool isPPC64 = Subtarget->isPPC64(); |
| 6111 | const bool isELFABI = Subtarget->isSVR4ABI(); |
| 6112 | const bool isAIXABI = Subtarget->isAIXABI(); |
| 6113 | |
| 6114 | // PowerPC only support small, medium and large code model. |
| 6115 | const CodeModel::Model CModel = getCodeModel(Subtarget: *Subtarget, TM, Node: N); |
| 6116 | |
| 6117 | assert(!(CModel == CodeModel::Tiny || CModel == CodeModel::Kernel) && |
| 6118 | "PowerPC doesn't support tiny or kernel code models." ); |
| 6119 | |
| 6120 | if (isAIXABI && CModel == CodeModel::Medium) |
| 6121 | report_fatal_error(reason: "Medium code model is not supported on AIX." ); |
| 6122 | |
| 6123 | // For 64-bit ELF small code model, we allow SelectCodeCommon to handle |
| 6124 | // this, selecting one of LDtoc, LDtocJTI, LDtocCPT, and LDtocBA. For AIX |
| 6125 | // small code model, we need to check for a toc-data attribute. |
| 6126 | if (isPPC64 && !isAIXABI && CModel == CodeModel::Small) |
| 6127 | break; |
| 6128 | |
| 6129 | auto replaceWith = [this, &dl](unsigned OpCode, SDNode *TocEntry, |
| 6130 | EVT OperandTy) { |
| 6131 | SDValue GA = TocEntry->getOperand(Num: 0); |
| 6132 | SDValue TocBase = TocEntry->getOperand(Num: 1); |
| 6133 | SDNode *MN = nullptr; |
| 6134 | if (OpCode == PPC::ADDItoc || OpCode == PPC::ADDItoc8) |
| 6135 | // toc-data access doesn't involve in loading from got, no need to |
| 6136 | // keep memory operands. |
| 6137 | MN = CurDAG->getMachineNode(Opcode: OpCode, dl, VT: OperandTy, Op1: TocBase, Op2: GA); |
| 6138 | else { |
| 6139 | MN = CurDAG->getMachineNode(Opcode: OpCode, dl, VT: OperandTy, Op1: GA, Op2: TocBase); |
| 6140 | transferMemOperands(N: TocEntry, Result: MN); |
| 6141 | } |
| 6142 | ReplaceNode(F: TocEntry, T: MN); |
| 6143 | }; |
| 6144 | |
| 6145 | // Handle 32-bit small code model. |
| 6146 | if (!isPPC64 && CModel == CodeModel::Small) { |
| 6147 | // Transforms the ISD::TOC_ENTRY node to passed in Opcode, either |
| 6148 | // PPC::ADDItoc, or PPC::LWZtoc |
| 6149 | if (isELFABI) { |
| 6150 | assert(TM.isPositionIndependent() && |
| 6151 | "32-bit ELF can only have TOC entries in position independent" |
| 6152 | " code." ); |
| 6153 | // 32-bit ELF always uses a small code model toc access. |
| 6154 | replaceWith(PPC::LWZtoc, N, MVT::i32); |
| 6155 | return; |
| 6156 | } |
| 6157 | |
| 6158 | assert(isAIXABI && "ELF ABI already handled" ); |
| 6159 | |
| 6160 | if (hasTocDataAttr(Val: N->getOperand(Num: 0))) { |
| 6161 | replaceWith(PPC::ADDItoc, N, MVT::i32); |
| 6162 | return; |
| 6163 | } |
| 6164 | |
| 6165 | replaceWith(PPC::LWZtoc, N, MVT::i32); |
| 6166 | return; |
| 6167 | } |
| 6168 | |
| 6169 | if (isPPC64 && CModel == CodeModel::Small) { |
| 6170 | assert(isAIXABI && "ELF ABI handled in common SelectCode" ); |
| 6171 | |
| 6172 | if (hasTocDataAttr(Val: N->getOperand(Num: 0))) { |
| 6173 | replaceWith(PPC::ADDItoc8, N, MVT::i64); |
| 6174 | return; |
| 6175 | } |
| 6176 | // Break if it doesn't have toc data attribute. Proceed with common |
| 6177 | // SelectCode. |
| 6178 | break; |
| 6179 | } |
| 6180 | |
| 6181 | assert(CModel != CodeModel::Small && "All small code models handled." ); |
| 6182 | |
| 6183 | assert((isPPC64 || (isAIXABI && !isPPC64)) && "We are dealing with 64-bit" |
| 6184 | " ELF/AIX or 32-bit AIX in the following." ); |
| 6185 | |
| 6186 | // Transforms the ISD::TOC_ENTRY node for 32-bit AIX large code model mode, |
| 6187 | // 64-bit medium (ELF-only), or 64-bit large (ELF and AIX) code model code |
| 6188 | // that does not contain TOC data symbols. We generate two instructions as |
| 6189 | // described below. The first source operand is a symbol reference. If it |
| 6190 | // must be referenced via the TOC according to Subtarget, we generate: |
| 6191 | // [32-bit AIX] |
| 6192 | // LWZtocL(@sym, ADDIStocHA(%r2, @sym)) |
| 6193 | // [64-bit ELF/AIX] |
| 6194 | // LDtocL(@sym, ADDIStocHA8(%x2, @sym)) |
| 6195 | // Otherwise for medium code model ELF we generate: |
| 6196 | // ADDItocL8(ADDIStocHA8(%x2, @sym), @sym) |
| 6197 | |
| 6198 | // And finally for AIX with toc-data we generate: |
| 6199 | // [32-bit AIX] |
| 6200 | // ADDItocL(ADDIStocHA(%x2, @sym), @sym) |
| 6201 | // [64-bit AIX] |
| 6202 | // ADDItocL8(ADDIStocHA8(%x2, @sym), @sym) |
| 6203 | |
| 6204 | SDValue GA = N->getOperand(Num: 0); |
| 6205 | SDValue TOCbase = N->getOperand(Num: 1); |
| 6206 | |
| 6207 | EVT VT = Subtarget->getScalarIntVT(); |
| 6208 | SDNode *Tmp = CurDAG->getMachineNode( |
| 6209 | Opcode: isPPC64 ? PPC::ADDIStocHA8 : PPC::ADDIStocHA, dl, VT, Op1: TOCbase, Op2: GA); |
| 6210 | |
| 6211 | // On AIX, if the symbol has the toc-data attribute it will be defined |
| 6212 | // in the TOC entry, so we use an ADDItocL/ADDItocL8. |
| 6213 | if (isAIXABI && hasTocDataAttr(Val: GA)) { |
| 6214 | ReplaceNode( |
| 6215 | F: N, T: CurDAG->getMachineNode(Opcode: isPPC64 ? PPC::ADDItocL8 : PPC::ADDItocL, |
| 6216 | dl, VT, Op1: SDValue(Tmp, 0), Op2: GA)); |
| 6217 | return; |
| 6218 | } |
| 6219 | |
| 6220 | if (PPCLowering->isAccessedAsGotIndirect(N: GA)) { |
| 6221 | // If it is accessed as got-indirect, we need an extra LWZ/LD to load |
| 6222 | // the address. |
| 6223 | SDNode *MN = CurDAG->getMachineNode( |
| 6224 | Opcode: isPPC64 ? PPC::LDtocL : PPC::LWZtocL, dl, VT, Op1: GA, Op2: SDValue(Tmp, 0)); |
| 6225 | |
| 6226 | transferMemOperands(N, Result: MN); |
| 6227 | ReplaceNode(F: N, T: MN); |
| 6228 | return; |
| 6229 | } |
| 6230 | |
| 6231 | assert(isPPC64 && "TOC_ENTRY already handled for 32-bit." ); |
| 6232 | // Build the address relative to the TOC-pointer. |
| 6233 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: PPC::ADDItocL8, dl, VT: MVT::i64, |
| 6234 | Op1: SDValue(Tmp, 0), Op2: GA)); |
| 6235 | return; |
| 6236 | } |
| 6237 | case PPCISD::PPC32_PICGOT: |
| 6238 | // Generate a PIC-safe GOT reference. |
| 6239 | assert(Subtarget->is32BitELFABI() && |
| 6240 | "PPCISD::PPC32_PICGOT is only supported for 32-bit SVR4" ); |
| 6241 | CurDAG->SelectNodeTo(N, MachineOpc: PPC::PPC32PICGOT, |
| 6242 | VT1: PPCLowering->getPointerTy(DL: CurDAG->getDataLayout()), |
| 6243 | VT2: MVT::i32); |
| 6244 | return; |
| 6245 | |
| 6246 | case PPCISD::VADD_SPLAT: { |
| 6247 | // This expands into one of three sequences, depending on whether |
| 6248 | // the first operand is odd or even, positive or negative. |
| 6249 | assert(isa<ConstantSDNode>(N->getOperand(0)) && |
| 6250 | isa<ConstantSDNode>(N->getOperand(1)) && |
| 6251 | "Invalid operand on VADD_SPLAT!" ); |
| 6252 | |
| 6253 | int Elt = N->getConstantOperandVal(Num: 0); |
| 6254 | int EltSize = N->getConstantOperandVal(Num: 1); |
| 6255 | unsigned Opc1, Opc2, Opc3; |
| 6256 | EVT VT; |
| 6257 | |
| 6258 | if (EltSize == 1) { |
| 6259 | Opc1 = PPC::VSPLTISB; |
| 6260 | Opc2 = PPC::VADDUBM; |
| 6261 | Opc3 = PPC::VSUBUBM; |
| 6262 | VT = MVT::v16i8; |
| 6263 | } else if (EltSize == 2) { |
| 6264 | Opc1 = PPC::VSPLTISH; |
| 6265 | Opc2 = PPC::VADDUHM; |
| 6266 | Opc3 = PPC::VSUBUHM; |
| 6267 | VT = MVT::v8i16; |
| 6268 | } else { |
| 6269 | assert(EltSize == 4 && "Invalid element size on VADD_SPLAT!" ); |
| 6270 | Opc1 = PPC::VSPLTISW; |
| 6271 | Opc2 = PPC::VADDUWM; |
| 6272 | Opc3 = PPC::VSUBUWM; |
| 6273 | VT = MVT::v4i32; |
| 6274 | } |
| 6275 | |
| 6276 | if ((Elt & 1) == 0) { |
| 6277 | // Elt is even, in the range [-32,-18] + [16,30]. |
| 6278 | // |
| 6279 | // Convert: VADD_SPLAT elt, size |
| 6280 | // Into: tmp = VSPLTIS[BHW] elt |
| 6281 | // VADDU[BHW]M tmp, tmp |
| 6282 | // Where: [BHW] = B for size = 1, H for size = 2, W for size = 4 |
| 6283 | SDValue EltVal = getI32Imm(Imm: Elt >> 1, dl); |
| 6284 | SDNode *Tmp = CurDAG->getMachineNode(Opcode: Opc1, dl, VT, Op1: EltVal); |
| 6285 | SDValue TmpVal = SDValue(Tmp, 0); |
| 6286 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc2, dl, VT, Op1: TmpVal, Op2: TmpVal)); |
| 6287 | return; |
| 6288 | } else if (Elt > 0) { |
| 6289 | // Elt is odd and positive, in the range [17,31]. |
| 6290 | // |
| 6291 | // Convert: VADD_SPLAT elt, size |
| 6292 | // Into: tmp1 = VSPLTIS[BHW] elt-16 |
| 6293 | // tmp2 = VSPLTIS[BHW] -16 |
| 6294 | // VSUBU[BHW]M tmp1, tmp2 |
| 6295 | SDValue EltVal = getI32Imm(Imm: Elt - 16, dl); |
| 6296 | SDNode *Tmp1 = CurDAG->getMachineNode(Opcode: Opc1, dl, VT, Op1: EltVal); |
| 6297 | EltVal = getI32Imm(Imm: -16, dl); |
| 6298 | SDNode *Tmp2 = CurDAG->getMachineNode(Opcode: Opc1, dl, VT, Op1: EltVal); |
| 6299 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc3, dl, VT, Op1: SDValue(Tmp1, 0), |
| 6300 | Op2: SDValue(Tmp2, 0))); |
| 6301 | return; |
| 6302 | } else { |
| 6303 | // Elt is odd and negative, in the range [-31,-17]. |
| 6304 | // |
| 6305 | // Convert: VADD_SPLAT elt, size |
| 6306 | // Into: tmp1 = VSPLTIS[BHW] elt+16 |
| 6307 | // tmp2 = VSPLTIS[BHW] -16 |
| 6308 | // VADDU[BHW]M tmp1, tmp2 |
| 6309 | SDValue EltVal = getI32Imm(Imm: Elt + 16, dl); |
| 6310 | SDNode *Tmp1 = CurDAG->getMachineNode(Opcode: Opc1, dl, VT, Op1: EltVal); |
| 6311 | EltVal = getI32Imm(Imm: -16, dl); |
| 6312 | SDNode *Tmp2 = CurDAG->getMachineNode(Opcode: Opc1, dl, VT, Op1: EltVal); |
| 6313 | ReplaceNode(F: N, T: CurDAG->getMachineNode(Opcode: Opc2, dl, VT, Op1: SDValue(Tmp1, 0), |
| 6314 | Op2: SDValue(Tmp2, 0))); |
| 6315 | return; |
| 6316 | } |
| 6317 | } |
| 6318 | case PPCISD::LD_SPLAT: { |
| 6319 | // Here we want to handle splat load for type v16i8 and v8i16 when there is |
| 6320 | // no direct move, we don't need to use stack for this case. If target has |
| 6321 | // direct move, we should be able to get the best selection in the .td file. |
| 6322 | if (!Subtarget->hasAltivec() || Subtarget->hasDirectMove()) |
| 6323 | break; |
| 6324 | |
| 6325 | EVT Type = N->getValueType(ResNo: 0); |
| 6326 | if (Type != MVT::v16i8 && Type != MVT::v8i16) |
| 6327 | break; |
| 6328 | |
| 6329 | // If the alignment for the load is 16 or bigger, we don't need the |
| 6330 | // permutated mask to get the required value. The value must be the 0 |
| 6331 | // element in big endian target or 7/15 in little endian target in the |
| 6332 | // result vsx register of lvx instruction. |
| 6333 | // Select the instruction in the .td file. |
| 6334 | if (cast<MemIntrinsicSDNode>(Val: N)->getAlign() >= Align(16) && |
| 6335 | isOffsetMultipleOf(N, Val: 16)) |
| 6336 | break; |
| 6337 | |
| 6338 | SDValue ZeroReg = |
| 6339 | CurDAG->getRegister(Reg: Subtarget->isPPC64() ? PPC::ZERO8 : PPC::ZERO, |
| 6340 | VT: Subtarget->getScalarIntVT()); |
| 6341 | unsigned LIOpcode = Subtarget->isPPC64() ? PPC::LI8 : PPC::LI; |
| 6342 | // v16i8 LD_SPLAT addr |
| 6343 | // ======> |
| 6344 | // Mask = LVSR/LVSL 0, addr |
| 6345 | // LoadLow = LVX 0, addr |
| 6346 | // Perm = VPERM LoadLow, LoadLow, Mask |
| 6347 | // Splat = VSPLTB 15/0, Perm |
| 6348 | // |
| 6349 | // v8i16 LD_SPLAT addr |
| 6350 | // ======> |
| 6351 | // Mask = LVSR/LVSL 0, addr |
| 6352 | // LoadLow = LVX 0, addr |
| 6353 | // LoadHigh = LVX (LI, 1), addr |
| 6354 | // Perm = VPERM LoadLow, LoadHigh, Mask |
| 6355 | // Splat = VSPLTH 7/0, Perm |
| 6356 | unsigned SplatOp = (Type == MVT::v16i8) ? PPC::VSPLTB : PPC::VSPLTH; |
| 6357 | unsigned SplatElemIndex = |
| 6358 | Subtarget->isLittleEndian() ? ((Type == MVT::v16i8) ? 15 : 7) : 0; |
| 6359 | |
| 6360 | SDNode *Mask = CurDAG->getMachineNode( |
| 6361 | Opcode: Subtarget->isLittleEndian() ? PPC::LVSR : PPC::LVSL, dl, VT: Type, Op1: ZeroReg, |
| 6362 | Op2: N->getOperand(Num: 1)); |
| 6363 | |
| 6364 | SDNode *LoadLow = |
| 6365 | CurDAG->getMachineNode(Opcode: PPC::LVX, dl, VT1: MVT::v16i8, VT2: MVT::Other, |
| 6366 | Ops: {ZeroReg, N->getOperand(Num: 1), N->getOperand(Num: 0)}); |
| 6367 | |
| 6368 | SDNode *LoadHigh = LoadLow; |
| 6369 | if (Type == MVT::v8i16) { |
| 6370 | LoadHigh = CurDAG->getMachineNode( |
| 6371 | Opcode: PPC::LVX, dl, VT1: MVT::v16i8, VT2: MVT::Other, |
| 6372 | Ops: {SDValue(CurDAG->getMachineNode( |
| 6373 | Opcode: LIOpcode, dl, VT: MVT::i32, |
| 6374 | Op1: CurDAG->getTargetConstant(Val: 1, DL: dl, VT: MVT::i8)), |
| 6375 | 0), |
| 6376 | N->getOperand(Num: 1), SDValue(LoadLow, 1)}); |
| 6377 | } |
| 6378 | |
| 6379 | CurDAG->ReplaceAllUsesOfValueWith(From: SDValue(N, 1), To: SDValue(LoadHigh, 1)); |
| 6380 | transferMemOperands(N, Result: LoadHigh); |
| 6381 | |
| 6382 | SDNode *Perm = |
| 6383 | CurDAG->getMachineNode(Opcode: PPC::VPERM, dl, VT: Type, Op1: SDValue(LoadLow, 0), |
| 6384 | Op2: SDValue(LoadHigh, 0), Op3: SDValue(Mask, 0)); |
| 6385 | CurDAG->SelectNodeTo(N, MachineOpc: SplatOp, VT: Type, |
| 6386 | Op1: CurDAG->getTargetConstant(Val: SplatElemIndex, DL: dl, VT: MVT::i8), |
| 6387 | Op2: SDValue(Perm, 0)); |
| 6388 | return; |
| 6389 | } |
| 6390 | } |
| 6391 | |
| 6392 | SelectCode(N); |
| 6393 | } |
| 6394 | |
| 6395 | // If the target supports the cmpb instruction, do the idiom recognition here. |
| 6396 | // We don't do this as a DAG combine because we don't want to do it as nodes |
| 6397 | // are being combined (because we might miss part of the eventual idiom). We |
| 6398 | // don't want to do it during instruction selection because we want to reuse |
| 6399 | // the logic for lowering the masking operations already part of the |
| 6400 | // instruction selector. |
| 6401 | SDValue PPCDAGToDAGISel::combineToCMPB(SDNode *N) { |
| 6402 | SDLoc dl(N); |
| 6403 | |
| 6404 | assert(N->getOpcode() == ISD::OR && |
| 6405 | "Only OR nodes are supported for CMPB" ); |
| 6406 | |
| 6407 | SDValue Res; |
| 6408 | if (!Subtarget->hasCMPB()) |
| 6409 | return Res; |
| 6410 | |
| 6411 | if (N->getValueType(ResNo: 0) != MVT::i32 && |
| 6412 | N->getValueType(ResNo: 0) != MVT::i64) |
| 6413 | return Res; |
| 6414 | |
| 6415 | EVT VT = N->getValueType(ResNo: 0); |
| 6416 | |
| 6417 | SDValue RHS, LHS; |
| 6418 | bool BytesFound[8] = {false, false, false, false, false, false, false, false}; |
| 6419 | uint64_t Mask = 0, Alt = 0; |
| 6420 | |
| 6421 | auto IsByteSelectCC = [this](SDValue O, unsigned &b, |
| 6422 | uint64_t &Mask, uint64_t &Alt, |
| 6423 | SDValue &LHS, SDValue &RHS) { |
| 6424 | if (O.getOpcode() != ISD::SELECT_CC) |
| 6425 | return false; |
| 6426 | ISD::CondCode CC = cast<CondCodeSDNode>(Val: O.getOperand(i: 4))->get(); |
| 6427 | |
| 6428 | if (!isa<ConstantSDNode>(Val: O.getOperand(i: 2)) || |
| 6429 | !isa<ConstantSDNode>(Val: O.getOperand(i: 3))) |
| 6430 | return false; |
| 6431 | |
| 6432 | uint64_t PM = O.getConstantOperandVal(i: 2); |
| 6433 | uint64_t PAlt = O.getConstantOperandVal(i: 3); |
| 6434 | for (b = 0; b < 8; ++b) { |
| 6435 | uint64_t Mask = UINT64_C(0xFF) << (8*b); |
| 6436 | if (PM && (PM & Mask) == PM && (PAlt & Mask) == PAlt) |
| 6437 | break; |
| 6438 | } |
| 6439 | |
| 6440 | if (b == 8) |
| 6441 | return false; |
| 6442 | Mask |= PM; |
| 6443 | Alt |= PAlt; |
| 6444 | |
| 6445 | if (!isa<ConstantSDNode>(Val: O.getOperand(i: 1)) || |
| 6446 | O.getConstantOperandVal(i: 1) != 0) { |
| 6447 | SDValue Op0 = O.getOperand(i: 0), Op1 = O.getOperand(i: 1); |
| 6448 | if (Op0.getOpcode() == ISD::TRUNCATE) |
| 6449 | Op0 = Op0.getOperand(i: 0); |
| 6450 | if (Op1.getOpcode() == ISD::TRUNCATE) |
| 6451 | Op1 = Op1.getOperand(i: 0); |
| 6452 | |
| 6453 | if (Op0.getOpcode() == ISD::SRL && Op1.getOpcode() == ISD::SRL && |
| 6454 | Op0.getOperand(i: 1) == Op1.getOperand(i: 1) && CC == ISD::SETEQ && |
| 6455 | isa<ConstantSDNode>(Val: Op0.getOperand(i: 1))) { |
| 6456 | |
| 6457 | unsigned Bits = Op0.getValueSizeInBits(); |
| 6458 | if (b != Bits/8-1) |
| 6459 | return false; |
| 6460 | if (Op0.getConstantOperandVal(i: 1) != Bits-8) |
| 6461 | return false; |
| 6462 | |
| 6463 | LHS = Op0.getOperand(i: 0); |
| 6464 | RHS = Op1.getOperand(i: 0); |
| 6465 | return true; |
| 6466 | } |
| 6467 | |
| 6468 | // When we have small integers (i16 to be specific), the form present |
| 6469 | // post-legalization uses SETULT in the SELECT_CC for the |
| 6470 | // higher-order byte, depending on the fact that the |
| 6471 | // even-higher-order bytes are known to all be zero, for example: |
| 6472 | // select_cc (xor $lhs, $rhs), 256, 65280, 0, setult |
| 6473 | // (so when the second byte is the same, because all higher-order |
| 6474 | // bits from bytes 3 and 4 are known to be zero, the result of the |
| 6475 | // xor can be at most 255) |
| 6476 | if (Op0.getOpcode() == ISD::XOR && CC == ISD::SETULT && |
| 6477 | isa<ConstantSDNode>(Val: O.getOperand(i: 1))) { |
| 6478 | |
| 6479 | uint64_t ULim = O.getConstantOperandVal(i: 1); |
| 6480 | if (ULim != (UINT64_C(1) << b*8)) |
| 6481 | return false; |
| 6482 | |
| 6483 | // Now we need to make sure that the upper bytes are known to be |
| 6484 | // zero. |
| 6485 | unsigned Bits = Op0.getValueSizeInBits(); |
| 6486 | if (!CurDAG->MaskedValueIsZero( |
| 6487 | Op: Op0, Mask: APInt::getHighBitsSet(numBits: Bits, hiBitsSet: Bits - (b + 1) * 8))) |
| 6488 | return false; |
| 6489 | |
| 6490 | LHS = Op0.getOperand(i: 0); |
| 6491 | RHS = Op0.getOperand(i: 1); |
| 6492 | return true; |
| 6493 | } |
| 6494 | |
| 6495 | return false; |
| 6496 | } |
| 6497 | |
| 6498 | if (CC != ISD::SETEQ) |
| 6499 | return false; |
| 6500 | |
| 6501 | SDValue Op = O.getOperand(i: 0); |
| 6502 | if (Op.getOpcode() == ISD::AND) { |
| 6503 | if (!isa<ConstantSDNode>(Val: Op.getOperand(i: 1))) |
| 6504 | return false; |
| 6505 | if (Op.getConstantOperandVal(i: 1) != (UINT64_C(0xFF) << (8*b))) |
| 6506 | return false; |
| 6507 | |
| 6508 | SDValue XOR = Op.getOperand(i: 0); |
| 6509 | if (XOR.getOpcode() == ISD::TRUNCATE) |
| 6510 | XOR = XOR.getOperand(i: 0); |
| 6511 | if (XOR.getOpcode() != ISD::XOR) |
| 6512 | return false; |
| 6513 | |
| 6514 | LHS = XOR.getOperand(i: 0); |
| 6515 | RHS = XOR.getOperand(i: 1); |
| 6516 | return true; |
| 6517 | } else if (Op.getOpcode() == ISD::SRL) { |
| 6518 | if (!isa<ConstantSDNode>(Val: Op.getOperand(i: 1))) |
| 6519 | return false; |
| 6520 | unsigned Bits = Op.getValueSizeInBits(); |
| 6521 | if (b != Bits/8-1) |
| 6522 | return false; |
| 6523 | if (Op.getConstantOperandVal(i: 1) != Bits-8) |
| 6524 | return false; |
| 6525 | |
| 6526 | SDValue XOR = Op.getOperand(i: 0); |
| 6527 | if (XOR.getOpcode() == ISD::TRUNCATE) |
| 6528 | XOR = XOR.getOperand(i: 0); |
| 6529 | if (XOR.getOpcode() != ISD::XOR) |
| 6530 | return false; |
| 6531 | |
| 6532 | LHS = XOR.getOperand(i: 0); |
| 6533 | RHS = XOR.getOperand(i: 1); |
| 6534 | return true; |
| 6535 | } |
| 6536 | |
| 6537 | return false; |
| 6538 | }; |
| 6539 | |
| 6540 | SmallVector<SDValue, 8> Queue(1, SDValue(N, 0)); |
| 6541 | while (!Queue.empty()) { |
| 6542 | SDValue V = Queue.pop_back_val(); |
| 6543 | |
| 6544 | for (const SDValue &O : V.getNode()->ops()) { |
| 6545 | unsigned b = 0; |
| 6546 | uint64_t M = 0, A = 0; |
| 6547 | SDValue OLHS, ORHS; |
| 6548 | if (O.getOpcode() == ISD::OR) { |
| 6549 | Queue.push_back(Elt: O); |
| 6550 | } else if (IsByteSelectCC(O, b, M, A, OLHS, ORHS)) { |
| 6551 | if (!LHS) { |
| 6552 | LHS = OLHS; |
| 6553 | RHS = ORHS; |
| 6554 | BytesFound[b] = true; |
| 6555 | Mask |= M; |
| 6556 | Alt |= A; |
| 6557 | } else if ((LHS == ORHS && RHS == OLHS) || |
| 6558 | (RHS == ORHS && LHS == OLHS)) { |
| 6559 | BytesFound[b] = true; |
| 6560 | Mask |= M; |
| 6561 | Alt |= A; |
| 6562 | } else { |
| 6563 | return Res; |
| 6564 | } |
| 6565 | } else { |
| 6566 | return Res; |
| 6567 | } |
| 6568 | } |
| 6569 | } |
| 6570 | |
| 6571 | unsigned LastB = 0, BCnt = 0; |
| 6572 | for (unsigned i = 0; i < 8; ++i) |
| 6573 | if (BytesFound[LastB]) { |
| 6574 | ++BCnt; |
| 6575 | LastB = i; |
| 6576 | } |
| 6577 | |
| 6578 | if (!LastB || BCnt < 2) |
| 6579 | return Res; |
| 6580 | |
| 6581 | // Because we'll be zero-extending the output anyway if don't have a specific |
| 6582 | // value for each input byte (via the Mask), we can 'anyext' the inputs. |
| 6583 | if (LHS.getValueType() != VT) { |
| 6584 | LHS = CurDAG->getAnyExtOrTrunc(Op: LHS, DL: dl, VT); |
| 6585 | RHS = CurDAG->getAnyExtOrTrunc(Op: RHS, DL: dl, VT); |
| 6586 | } |
| 6587 | |
| 6588 | Res = CurDAG->getNode(Opcode: PPCISD::CMPB, DL: dl, VT, N1: LHS, N2: RHS); |
| 6589 | |
| 6590 | bool NonTrivialMask = ((int64_t) Mask) != INT64_C(-1); |
| 6591 | if (NonTrivialMask && !Alt) { |
| 6592 | // Res = Mask & CMPB |
| 6593 | Res = CurDAG->getNode(Opcode: ISD::AND, DL: dl, VT, N1: Res, |
| 6594 | N2: CurDAG->getConstant(Val: Mask, DL: dl, VT)); |
| 6595 | } else if (Alt) { |
| 6596 | // Res = (CMPB & Mask) | (~CMPB & Alt) |
| 6597 | // Which, as suggested here: |
| 6598 | // https://graphics.stanford.edu/~seander/bithacks.html#MaskedMerge |
| 6599 | // can be written as: |
| 6600 | // Res = Alt ^ ((Alt ^ Mask) & CMPB) |
| 6601 | // useful because the (Alt ^ Mask) can be pre-computed. |
| 6602 | Res = CurDAG->getNode(Opcode: ISD::AND, DL: dl, VT, N1: Res, |
| 6603 | N2: CurDAG->getConstant(Val: Mask ^ Alt, DL: dl, VT)); |
| 6604 | Res = CurDAG->getNode(Opcode: ISD::XOR, DL: dl, VT, N1: Res, |
| 6605 | N2: CurDAG->getConstant(Val: Alt, DL: dl, VT)); |
| 6606 | } |
| 6607 | |
| 6608 | return Res; |
| 6609 | } |
| 6610 | |
| 6611 | // When CR bit registers are enabled, an extension of an i1 variable to a i32 |
| 6612 | // or i64 value is lowered in terms of a SELECT_I[48] operation, and thus |
| 6613 | // involves constant materialization of a 0 or a 1 or both. If the result of |
| 6614 | // the extension is then operated upon by some operator that can be constant |
| 6615 | // folded with a constant 0 or 1, and that constant can be materialized using |
| 6616 | // only one instruction (like a zero or one), then we should fold in those |
| 6617 | // operations with the select. |
| 6618 | void PPCDAGToDAGISel::foldBoolExts(SDValue &Res, SDNode *&N) { |
| 6619 | if (!Subtarget->useCRBits()) |
| 6620 | return; |
| 6621 | |
| 6622 | if (N->getOpcode() != ISD::ZERO_EXTEND && |
| 6623 | N->getOpcode() != ISD::SIGN_EXTEND && |
| 6624 | N->getOpcode() != ISD::ANY_EXTEND) |
| 6625 | return; |
| 6626 | |
| 6627 | if (N->getOperand(Num: 0).getValueType() != MVT::i1) |
| 6628 | return; |
| 6629 | |
| 6630 | if (!N->hasOneUse()) |
| 6631 | return; |
| 6632 | |
| 6633 | SDLoc dl(N); |
| 6634 | EVT VT = N->getValueType(ResNo: 0); |
| 6635 | SDValue Cond = N->getOperand(Num: 0); |
| 6636 | SDValue ConstTrue = CurDAG->getSignedConstant( |
| 6637 | Val: N->getOpcode() == ISD::SIGN_EXTEND ? -1 : 1, DL: dl, VT); |
| 6638 | SDValue ConstFalse = CurDAG->getConstant(Val: 0, DL: dl, VT); |
| 6639 | |
| 6640 | do { |
| 6641 | SDNode *User = *N->user_begin(); |
| 6642 | if (User->getNumOperands() != 2) |
| 6643 | break; |
| 6644 | |
| 6645 | auto TryFold = [this, N, User, dl](SDValue Val) { |
| 6646 | SDValue UserO0 = User->getOperand(Num: 0), UserO1 = User->getOperand(Num: 1); |
| 6647 | SDValue O0 = UserO0.getNode() == N ? Val : UserO0; |
| 6648 | SDValue O1 = UserO1.getNode() == N ? Val : UserO1; |
| 6649 | |
| 6650 | return CurDAG->FoldConstantArithmetic(Opcode: User->getOpcode(), DL: dl, |
| 6651 | VT: User->getValueType(ResNo: 0), Ops: {O0, O1}); |
| 6652 | }; |
| 6653 | |
| 6654 | // FIXME: When the semantics of the interaction between select and undef |
| 6655 | // are clearly defined, it may turn out to be unnecessary to break here. |
| 6656 | SDValue TrueRes = TryFold(ConstTrue); |
| 6657 | if (!TrueRes || TrueRes.isUndef()) |
| 6658 | break; |
| 6659 | SDValue FalseRes = TryFold(ConstFalse); |
| 6660 | if (!FalseRes || FalseRes.isUndef()) |
| 6661 | break; |
| 6662 | |
| 6663 | // For us to materialize these using one instruction, we must be able to |
| 6664 | // represent them as signed 16-bit integers. |
| 6665 | uint64_t True = TrueRes->getAsZExtVal(), False = FalseRes->getAsZExtVal(); |
| 6666 | if (!isInt<16>(x: True) || !isInt<16>(x: False)) |
| 6667 | break; |
| 6668 | |
| 6669 | // We can replace User with a new SELECT node, and try again to see if we |
| 6670 | // can fold the select with its user. |
| 6671 | Res = CurDAG->getSelect(DL: dl, VT: User->getValueType(ResNo: 0), Cond, LHS: TrueRes, RHS: FalseRes); |
| 6672 | N = User; |
| 6673 | ConstTrue = TrueRes; |
| 6674 | ConstFalse = FalseRes; |
| 6675 | } while (N->hasOneUse()); |
| 6676 | } |
| 6677 | |
| 6678 | void PPCDAGToDAGISel::PreprocessISelDAG() { |
| 6679 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
| 6680 | |
| 6681 | bool MadeChange = false; |
| 6682 | while (Position != CurDAG->allnodes_begin()) { |
| 6683 | SDNode *N = &*--Position; |
| 6684 | if (N->use_empty()) |
| 6685 | continue; |
| 6686 | |
| 6687 | SDValue Res; |
| 6688 | switch (N->getOpcode()) { |
| 6689 | default: break; |
| 6690 | case ISD::OR: |
| 6691 | Res = combineToCMPB(N); |
| 6692 | break; |
| 6693 | } |
| 6694 | |
| 6695 | if (!Res) |
| 6696 | foldBoolExts(Res, N); |
| 6697 | |
| 6698 | if (Res) { |
| 6699 | LLVM_DEBUG(dbgs() << "PPC DAG preprocessing replacing:\nOld: " ); |
| 6700 | LLVM_DEBUG(N->dump(CurDAG)); |
| 6701 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
| 6702 | LLVM_DEBUG(Res.getNode()->dump(CurDAG)); |
| 6703 | LLVM_DEBUG(dbgs() << "\n" ); |
| 6704 | |
| 6705 | CurDAG->ReplaceAllUsesOfValueWith(From: SDValue(N, 0), To: Res); |
| 6706 | MadeChange = true; |
| 6707 | } |
| 6708 | } |
| 6709 | |
| 6710 | if (MadeChange) |
| 6711 | CurDAG->RemoveDeadNodes(); |
| 6712 | } |
| 6713 | |
| 6714 | /// PostprocessISelDAG - Perform some late peephole optimizations |
| 6715 | /// on the DAG representation. |
| 6716 | void PPCDAGToDAGISel::PostprocessISelDAG() { |
| 6717 | // Skip peepholes at -O0. |
| 6718 | if (TM.getOptLevel() == CodeGenOptLevel::None) |
| 6719 | return; |
| 6720 | |
| 6721 | PeepholePPC64(); |
| 6722 | PeepholeCROps(); |
| 6723 | PeepholePPC64ZExt(); |
| 6724 | } |
| 6725 | |
| 6726 | // Check if all users of this node will become isel where the second operand |
| 6727 | // is the constant zero. If this is so, and if we can negate the condition, |
| 6728 | // then we can flip the true and false operands. This will allow the zero to |
| 6729 | // be folded with the isel so that we don't need to materialize a register |
| 6730 | // containing zero. |
| 6731 | bool PPCDAGToDAGISel::(SDNode *N) { |
| 6732 | for (const SDNode *User : N->users()) { |
| 6733 | if (!User->isMachineOpcode()) |
| 6734 | return false; |
| 6735 | if (User->getMachineOpcode() != PPC::SELECT_I4 && |
| 6736 | User->getMachineOpcode() != PPC::SELECT_I8) |
| 6737 | return false; |
| 6738 | |
| 6739 | SDNode *Op1 = User->getOperand(Num: 1).getNode(); |
| 6740 | SDNode *Op2 = User->getOperand(Num: 2).getNode(); |
| 6741 | // If we have a degenerate select with two equal operands, swapping will |
| 6742 | // not do anything, and we may run into an infinite loop. |
| 6743 | if (Op1 == Op2) |
| 6744 | return false; |
| 6745 | |
| 6746 | if (!Op2->isMachineOpcode()) |
| 6747 | return false; |
| 6748 | |
| 6749 | if (Op2->getMachineOpcode() != PPC::LI && |
| 6750 | Op2->getMachineOpcode() != PPC::LI8) |
| 6751 | return false; |
| 6752 | |
| 6753 | if (!isNullConstant(V: Op2->getOperand(Num: 0))) |
| 6754 | return false; |
| 6755 | } |
| 6756 | |
| 6757 | return true; |
| 6758 | } |
| 6759 | |
| 6760 | void PPCDAGToDAGISel::SwapAllSelectUsers(SDNode *N) { |
| 6761 | SmallVector<SDNode *, 4> ToReplace; |
| 6762 | for (SDNode *User : N->users()) { |
| 6763 | assert((User->getMachineOpcode() == PPC::SELECT_I4 || |
| 6764 | User->getMachineOpcode() == PPC::SELECT_I8) && |
| 6765 | "Must have all select users" ); |
| 6766 | ToReplace.push_back(Elt: User); |
| 6767 | } |
| 6768 | |
| 6769 | for (SDNode *User : ToReplace) { |
| 6770 | SDNode *ResNode = |
| 6771 | CurDAG->getMachineNode(Opcode: User->getMachineOpcode(), dl: SDLoc(User), |
| 6772 | VT: User->getValueType(ResNo: 0), Op1: User->getOperand(Num: 0), |
| 6773 | Op2: User->getOperand(Num: 2), |
| 6774 | Op3: User->getOperand(Num: 1)); |
| 6775 | |
| 6776 | LLVM_DEBUG(dbgs() << "CR Peephole replacing:\nOld: " ); |
| 6777 | LLVM_DEBUG(User->dump(CurDAG)); |
| 6778 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
| 6779 | LLVM_DEBUG(ResNode->dump(CurDAG)); |
| 6780 | LLVM_DEBUG(dbgs() << "\n" ); |
| 6781 | |
| 6782 | ReplaceUses(F: User, T: ResNode); |
| 6783 | } |
| 6784 | } |
| 6785 | |
| 6786 | void PPCDAGToDAGISel::PeepholeCROps() { |
| 6787 | bool IsModified; |
| 6788 | do { |
| 6789 | IsModified = false; |
| 6790 | for (SDNode &Node : CurDAG->allnodes()) { |
| 6791 | MachineSDNode *MachineNode = dyn_cast<MachineSDNode>(Val: &Node); |
| 6792 | if (!MachineNode || MachineNode->use_empty()) |
| 6793 | continue; |
| 6794 | SDNode *ResNode = MachineNode; |
| 6795 | |
| 6796 | bool Op1Set = false, Op1Unset = false, |
| 6797 | Op1Not = false, |
| 6798 | Op2Set = false, Op2Unset = false, |
| 6799 | Op2Not = false; |
| 6800 | |
| 6801 | unsigned Opcode = MachineNode->getMachineOpcode(); |
| 6802 | switch (Opcode) { |
| 6803 | default: break; |
| 6804 | case PPC::CRAND: |
| 6805 | case PPC::CRNAND: |
| 6806 | case PPC::CROR: |
| 6807 | case PPC::CRXOR: |
| 6808 | case PPC::CRNOR: |
| 6809 | case PPC::CREQV: |
| 6810 | case PPC::CRANDC: |
| 6811 | case PPC::CRORC: { |
| 6812 | SDValue Op = MachineNode->getOperand(Num: 1); |
| 6813 | if (Op.isMachineOpcode()) { |
| 6814 | if (Op.getMachineOpcode() == PPC::CRSET) |
| 6815 | Op2Set = true; |
| 6816 | else if (Op.getMachineOpcode() == PPC::CRUNSET) |
| 6817 | Op2Unset = true; |
| 6818 | else if ((Op.getMachineOpcode() == PPC::CRNOR && |
| 6819 | Op.getOperand(i: 0) == Op.getOperand(i: 1)) || |
| 6820 | Op.getMachineOpcode() == PPC::CRNOT) |
| 6821 | Op2Not = true; |
| 6822 | } |
| 6823 | [[fallthrough]]; |
| 6824 | } |
| 6825 | case PPC::BC: |
| 6826 | case PPC::BCn: |
| 6827 | case PPC::SELECT_I4: |
| 6828 | case PPC::SELECT_I8: |
| 6829 | case PPC::SELECT_F4: |
| 6830 | case PPC::SELECT_F8: |
| 6831 | case PPC::SELECT_SPE: |
| 6832 | case PPC::SELECT_SPE4: |
| 6833 | case PPC::SELECT_VRRC: |
| 6834 | case PPC::SELECT_VSFRC: |
| 6835 | case PPC::SELECT_VSSRC: |
| 6836 | case PPC::SELECT_VSRC: { |
| 6837 | SDValue Op = MachineNode->getOperand(Num: 0); |
| 6838 | if (Op.isMachineOpcode()) { |
| 6839 | if (Op.getMachineOpcode() == PPC::CRSET) |
| 6840 | Op1Set = true; |
| 6841 | else if (Op.getMachineOpcode() == PPC::CRUNSET) |
| 6842 | Op1Unset = true; |
| 6843 | else if ((Op.getMachineOpcode() == PPC::CRNOR && |
| 6844 | Op.getOperand(i: 0) == Op.getOperand(i: 1)) || |
| 6845 | Op.getMachineOpcode() == PPC::CRNOT) |
| 6846 | Op1Not = true; |
| 6847 | } |
| 6848 | } |
| 6849 | break; |
| 6850 | } |
| 6851 | |
| 6852 | bool SelectSwap = false; |
| 6853 | switch (Opcode) { |
| 6854 | default: break; |
| 6855 | case PPC::CRAND: |
| 6856 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 6857 | // x & x = x |
| 6858 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 6859 | else if (Op1Set) |
| 6860 | // 1 & y = y |
| 6861 | ResNode = MachineNode->getOperand(Num: 1).getNode(); |
| 6862 | else if (Op2Set) |
| 6863 | // x & 1 = x |
| 6864 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 6865 | else if (Op1Unset || Op2Unset) |
| 6866 | // x & 0 = 0 & y = 0 |
| 6867 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRUNSET, dl: SDLoc(MachineNode), |
| 6868 | VT: MVT::i1); |
| 6869 | else if (Op1Not) |
| 6870 | // ~x & y = andc(y, x) |
| 6871 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRANDC, dl: SDLoc(MachineNode), |
| 6872 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 6873 | Op2: MachineNode->getOperand(Num: 0). |
| 6874 | getOperand(i: 0)); |
| 6875 | else if (Op2Not) |
| 6876 | // x & ~y = andc(x, y) |
| 6877 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRANDC, dl: SDLoc(MachineNode), |
| 6878 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6879 | Op2: MachineNode->getOperand(Num: 1). |
| 6880 | getOperand(i: 0)); |
| 6881 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 6882 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNAND, dl: SDLoc(MachineNode), |
| 6883 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6884 | Op2: MachineNode->getOperand(Num: 1)); |
| 6885 | SelectSwap = true; |
| 6886 | } |
| 6887 | break; |
| 6888 | case PPC::CRNAND: |
| 6889 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 6890 | // nand(x, x) -> nor(x, x) |
| 6891 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6892 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6893 | Op2: MachineNode->getOperand(Num: 0)); |
| 6894 | else if (Op1Set) |
| 6895 | // nand(1, y) -> nor(y, y) |
| 6896 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6897 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 6898 | Op2: MachineNode->getOperand(Num: 1)); |
| 6899 | else if (Op2Set) |
| 6900 | // nand(x, 1) -> nor(x, x) |
| 6901 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6902 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6903 | Op2: MachineNode->getOperand(Num: 0)); |
| 6904 | else if (Op1Unset || Op2Unset) |
| 6905 | // nand(x, 0) = nand(0, y) = 1 |
| 6906 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRSET, dl: SDLoc(MachineNode), |
| 6907 | VT: MVT::i1); |
| 6908 | else if (Op1Not) |
| 6909 | // nand(~x, y) = ~(~x & y) = x | ~y = orc(x, y) |
| 6910 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRORC, dl: SDLoc(MachineNode), |
| 6911 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 6912 | getOperand(i: 0), |
| 6913 | Op2: MachineNode->getOperand(Num: 1)); |
| 6914 | else if (Op2Not) |
| 6915 | // nand(x, ~y) = ~x | y = orc(y, x) |
| 6916 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRORC, dl: SDLoc(MachineNode), |
| 6917 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1). |
| 6918 | getOperand(i: 0), |
| 6919 | Op2: MachineNode->getOperand(Num: 0)); |
| 6920 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 6921 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRAND, dl: SDLoc(MachineNode), |
| 6922 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6923 | Op2: MachineNode->getOperand(Num: 1)); |
| 6924 | SelectSwap = true; |
| 6925 | } |
| 6926 | break; |
| 6927 | case PPC::CROR: |
| 6928 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 6929 | // x | x = x |
| 6930 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 6931 | else if (Op1Set || Op2Set) |
| 6932 | // x | 1 = 1 | y = 1 |
| 6933 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRSET, dl: SDLoc(MachineNode), |
| 6934 | VT: MVT::i1); |
| 6935 | else if (Op1Unset) |
| 6936 | // 0 | y = y |
| 6937 | ResNode = MachineNode->getOperand(Num: 1).getNode(); |
| 6938 | else if (Op2Unset) |
| 6939 | // x | 0 = x |
| 6940 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 6941 | else if (Op1Not) |
| 6942 | // ~x | y = orc(y, x) |
| 6943 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRORC, dl: SDLoc(MachineNode), |
| 6944 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 6945 | Op2: MachineNode->getOperand(Num: 0). |
| 6946 | getOperand(i: 0)); |
| 6947 | else if (Op2Not) |
| 6948 | // x | ~y = orc(x, y) |
| 6949 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRORC, dl: SDLoc(MachineNode), |
| 6950 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6951 | Op2: MachineNode->getOperand(Num: 1). |
| 6952 | getOperand(i: 0)); |
| 6953 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 6954 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6955 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6956 | Op2: MachineNode->getOperand(Num: 1)); |
| 6957 | SelectSwap = true; |
| 6958 | } |
| 6959 | break; |
| 6960 | case PPC::CRXOR: |
| 6961 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 6962 | // xor(x, x) = 0 |
| 6963 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRUNSET, dl: SDLoc(MachineNode), |
| 6964 | VT: MVT::i1); |
| 6965 | else if (Op1Set) |
| 6966 | // xor(1, y) -> nor(y, y) |
| 6967 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6968 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 6969 | Op2: MachineNode->getOperand(Num: 1)); |
| 6970 | else if (Op2Set) |
| 6971 | // xor(x, 1) -> nor(x, x) |
| 6972 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 6973 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6974 | Op2: MachineNode->getOperand(Num: 0)); |
| 6975 | else if (Op1Unset) |
| 6976 | // xor(0, y) = y |
| 6977 | ResNode = MachineNode->getOperand(Num: 1).getNode(); |
| 6978 | else if (Op2Unset) |
| 6979 | // xor(x, 0) = x |
| 6980 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 6981 | else if (Op1Not) |
| 6982 | // xor(~x, y) = eqv(x, y) |
| 6983 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CREQV, dl: SDLoc(MachineNode), |
| 6984 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 6985 | getOperand(i: 0), |
| 6986 | Op2: MachineNode->getOperand(Num: 1)); |
| 6987 | else if (Op2Not) |
| 6988 | // xor(x, ~y) = eqv(x, y) |
| 6989 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CREQV, dl: SDLoc(MachineNode), |
| 6990 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6991 | Op2: MachineNode->getOperand(Num: 1). |
| 6992 | getOperand(i: 0)); |
| 6993 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 6994 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CREQV, dl: SDLoc(MachineNode), |
| 6995 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 6996 | Op2: MachineNode->getOperand(Num: 1)); |
| 6997 | SelectSwap = true; |
| 6998 | } |
| 6999 | break; |
| 7000 | case PPC::CRNOR: |
| 7001 | if (Op1Set || Op2Set) |
| 7002 | // nor(1, y) -> 0 |
| 7003 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRUNSET, dl: SDLoc(MachineNode), |
| 7004 | VT: MVT::i1); |
| 7005 | else if (Op1Unset) |
| 7006 | // nor(0, y) = ~y -> nor(y, y) |
| 7007 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7008 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7009 | Op2: MachineNode->getOperand(Num: 1)); |
| 7010 | else if (Op2Unset) |
| 7011 | // nor(x, 0) = ~x |
| 7012 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7013 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7014 | Op2: MachineNode->getOperand(Num: 0)); |
| 7015 | else if (Op1Not) |
| 7016 | // nor(~x, y) = andc(x, y) |
| 7017 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRANDC, dl: SDLoc(MachineNode), |
| 7018 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 7019 | getOperand(i: 0), |
| 7020 | Op2: MachineNode->getOperand(Num: 1)); |
| 7021 | else if (Op2Not) |
| 7022 | // nor(x, ~y) = andc(y, x) |
| 7023 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRANDC, dl: SDLoc(MachineNode), |
| 7024 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1). |
| 7025 | getOperand(i: 0), |
| 7026 | Op2: MachineNode->getOperand(Num: 0)); |
| 7027 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 7028 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CROR, dl: SDLoc(MachineNode), |
| 7029 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7030 | Op2: MachineNode->getOperand(Num: 1)); |
| 7031 | SelectSwap = true; |
| 7032 | } |
| 7033 | break; |
| 7034 | case PPC::CREQV: |
| 7035 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 7036 | // eqv(x, x) = 1 |
| 7037 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRSET, dl: SDLoc(MachineNode), |
| 7038 | VT: MVT::i1); |
| 7039 | else if (Op1Set) |
| 7040 | // eqv(1, y) = y |
| 7041 | ResNode = MachineNode->getOperand(Num: 1).getNode(); |
| 7042 | else if (Op2Set) |
| 7043 | // eqv(x, 1) = x |
| 7044 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 7045 | else if (Op1Unset) |
| 7046 | // eqv(0, y) = ~y -> nor(y, y) |
| 7047 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7048 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7049 | Op2: MachineNode->getOperand(Num: 1)); |
| 7050 | else if (Op2Unset) |
| 7051 | // eqv(x, 0) = ~x |
| 7052 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7053 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7054 | Op2: MachineNode->getOperand(Num: 0)); |
| 7055 | else if (Op1Not) |
| 7056 | // eqv(~x, y) = xor(x, y) |
| 7057 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRXOR, dl: SDLoc(MachineNode), |
| 7058 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 7059 | getOperand(i: 0), |
| 7060 | Op2: MachineNode->getOperand(Num: 1)); |
| 7061 | else if (Op2Not) |
| 7062 | // eqv(x, ~y) = xor(x, y) |
| 7063 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRXOR, dl: SDLoc(MachineNode), |
| 7064 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7065 | Op2: MachineNode->getOperand(Num: 1). |
| 7066 | getOperand(i: 0)); |
| 7067 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 7068 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRXOR, dl: SDLoc(MachineNode), |
| 7069 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7070 | Op2: MachineNode->getOperand(Num: 1)); |
| 7071 | SelectSwap = true; |
| 7072 | } |
| 7073 | break; |
| 7074 | case PPC::CRANDC: |
| 7075 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 7076 | // andc(x, x) = 0 |
| 7077 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRUNSET, dl: SDLoc(MachineNode), |
| 7078 | VT: MVT::i1); |
| 7079 | else if (Op1Set) |
| 7080 | // andc(1, y) = ~y |
| 7081 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7082 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7083 | Op2: MachineNode->getOperand(Num: 1)); |
| 7084 | else if (Op1Unset || Op2Set) |
| 7085 | // andc(0, y) = andc(x, 1) = 0 |
| 7086 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRUNSET, dl: SDLoc(MachineNode), |
| 7087 | VT: MVT::i1); |
| 7088 | else if (Op2Unset) |
| 7089 | // andc(x, 0) = x |
| 7090 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 7091 | else if (Op1Not) |
| 7092 | // andc(~x, y) = ~(x | y) = nor(x, y) |
| 7093 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7094 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 7095 | getOperand(i: 0), |
| 7096 | Op2: MachineNode->getOperand(Num: 1)); |
| 7097 | else if (Op2Not) |
| 7098 | // andc(x, ~y) = x & y |
| 7099 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRAND, dl: SDLoc(MachineNode), |
| 7100 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7101 | Op2: MachineNode->getOperand(Num: 1). |
| 7102 | getOperand(i: 0)); |
| 7103 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 7104 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRORC, dl: SDLoc(MachineNode), |
| 7105 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7106 | Op2: MachineNode->getOperand(Num: 0)); |
| 7107 | SelectSwap = true; |
| 7108 | } |
| 7109 | break; |
| 7110 | case PPC::CRORC: |
| 7111 | if (MachineNode->getOperand(Num: 0) == MachineNode->getOperand(Num: 1)) |
| 7112 | // orc(x, x) = 1 |
| 7113 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRSET, dl: SDLoc(MachineNode), |
| 7114 | VT: MVT::i1); |
| 7115 | else if (Op1Set || Op2Unset) |
| 7116 | // orc(1, y) = orc(x, 0) = 1 |
| 7117 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRSET, dl: SDLoc(MachineNode), |
| 7118 | VT: MVT::i1); |
| 7119 | else if (Op2Set) |
| 7120 | // orc(x, 1) = x |
| 7121 | ResNode = MachineNode->getOperand(Num: 0).getNode(); |
| 7122 | else if (Op1Unset) |
| 7123 | // orc(0, y) = ~y |
| 7124 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNOR, dl: SDLoc(MachineNode), |
| 7125 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7126 | Op2: MachineNode->getOperand(Num: 1)); |
| 7127 | else if (Op1Not) |
| 7128 | // orc(~x, y) = ~(x & y) = nand(x, y) |
| 7129 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRNAND, dl: SDLoc(MachineNode), |
| 7130 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0). |
| 7131 | getOperand(i: 0), |
| 7132 | Op2: MachineNode->getOperand(Num: 1)); |
| 7133 | else if (Op2Not) |
| 7134 | // orc(x, ~y) = x | y |
| 7135 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CROR, dl: SDLoc(MachineNode), |
| 7136 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 0), |
| 7137 | Op2: MachineNode->getOperand(Num: 1). |
| 7138 | getOperand(i: 0)); |
| 7139 | else if (AllUsersSelectZero(N: MachineNode)) { |
| 7140 | ResNode = CurDAG->getMachineNode(Opcode: PPC::CRANDC, dl: SDLoc(MachineNode), |
| 7141 | VT: MVT::i1, Op1: MachineNode->getOperand(Num: 1), |
| 7142 | Op2: MachineNode->getOperand(Num: 0)); |
| 7143 | SelectSwap = true; |
| 7144 | } |
| 7145 | break; |
| 7146 | case PPC::SELECT_I4: |
| 7147 | case PPC::SELECT_I8: |
| 7148 | case PPC::SELECT_F4: |
| 7149 | case PPC::SELECT_F8: |
| 7150 | case PPC::SELECT_SPE: |
| 7151 | case PPC::SELECT_SPE4: |
| 7152 | case PPC::SELECT_VRRC: |
| 7153 | case PPC::SELECT_VSFRC: |
| 7154 | case PPC::SELECT_VSSRC: |
| 7155 | case PPC::SELECT_VSRC: |
| 7156 | if (Op1Set) |
| 7157 | ResNode = MachineNode->getOperand(Num: 1).getNode(); |
| 7158 | else if (Op1Unset) |
| 7159 | ResNode = MachineNode->getOperand(Num: 2).getNode(); |
| 7160 | else if (Op1Not) |
| 7161 | ResNode = CurDAG->getMachineNode(Opcode: MachineNode->getMachineOpcode(), |
| 7162 | dl: SDLoc(MachineNode), |
| 7163 | VT: MachineNode->getValueType(ResNo: 0), |
| 7164 | Op1: MachineNode->getOperand(Num: 0). |
| 7165 | getOperand(i: 0), |
| 7166 | Op2: MachineNode->getOperand(Num: 2), |
| 7167 | Op3: MachineNode->getOperand(Num: 1)); |
| 7168 | break; |
| 7169 | case PPC::BC: |
| 7170 | case PPC::BCn: |
| 7171 | if (Op1Not) |
| 7172 | ResNode = CurDAG->getMachineNode(Opcode: Opcode == PPC::BC ? PPC::BCn : |
| 7173 | PPC::BC, |
| 7174 | dl: SDLoc(MachineNode), |
| 7175 | VT: MVT::Other, |
| 7176 | Op1: MachineNode->getOperand(Num: 0). |
| 7177 | getOperand(i: 0), |
| 7178 | Op2: MachineNode->getOperand(Num: 1), |
| 7179 | Op3: MachineNode->getOperand(Num: 2)); |
| 7180 | // FIXME: Handle Op1Set, Op1Unset here too. |
| 7181 | break; |
| 7182 | } |
| 7183 | |
| 7184 | // If we're inverting this node because it is used only by selects that |
| 7185 | // we'd like to swap, then swap the selects before the node replacement. |
| 7186 | if (SelectSwap) |
| 7187 | SwapAllSelectUsers(N: MachineNode); |
| 7188 | |
| 7189 | if (ResNode != MachineNode) { |
| 7190 | LLVM_DEBUG(dbgs() << "CR Peephole replacing:\nOld: " ); |
| 7191 | LLVM_DEBUG(MachineNode->dump(CurDAG)); |
| 7192 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
| 7193 | LLVM_DEBUG(ResNode->dump(CurDAG)); |
| 7194 | LLVM_DEBUG(dbgs() << "\n" ); |
| 7195 | |
| 7196 | ReplaceUses(F: MachineNode, T: ResNode); |
| 7197 | IsModified = true; |
| 7198 | } |
| 7199 | } |
| 7200 | if (IsModified) |
| 7201 | CurDAG->RemoveDeadNodes(); |
| 7202 | } while (IsModified); |
| 7203 | } |
| 7204 | |
| 7205 | // Gather the set of 32-bit operations that are known to have their |
| 7206 | // higher-order 32 bits zero, where ToPromote contains all such operations. |
| 7207 | static bool PeepholePPC64ZExtGather(SDValue Op32, |
| 7208 | SmallPtrSetImpl<SDNode *> &ToPromote) { |
| 7209 | if (!Op32.isMachineOpcode()) |
| 7210 | return false; |
| 7211 | |
| 7212 | // First, check for the "frontier" instructions (those that will clear the |
| 7213 | // higher-order 32 bits. |
| 7214 | |
| 7215 | // For RLWINM and RLWNM, we need to make sure that the mask does not wrap |
| 7216 | // around. If it does not, then these instructions will clear the |
| 7217 | // higher-order bits. |
| 7218 | if ((Op32.getMachineOpcode() == PPC::RLWINM || |
| 7219 | Op32.getMachineOpcode() == PPC::RLWNM) && |
| 7220 | Op32.getConstantOperandVal(i: 2) <= Op32.getConstantOperandVal(i: 3)) { |
| 7221 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7222 | return true; |
| 7223 | } |
| 7224 | |
| 7225 | // SLW and SRW always clear the higher-order bits. |
| 7226 | if (Op32.getMachineOpcode() == PPC::SLW || |
| 7227 | Op32.getMachineOpcode() == PPC::SRW) { |
| 7228 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7229 | return true; |
| 7230 | } |
| 7231 | |
| 7232 | // For LI and LIS, we need the immediate to be positive (so that it is not |
| 7233 | // sign extended). |
| 7234 | if (Op32.getMachineOpcode() == PPC::LI || |
| 7235 | Op32.getMachineOpcode() == PPC::LIS) { |
| 7236 | if (!isUInt<15>(x: Op32.getConstantOperandVal(i: 0))) |
| 7237 | return false; |
| 7238 | |
| 7239 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7240 | return true; |
| 7241 | } |
| 7242 | |
| 7243 | // LHBRX and LWBRX always clear the higher-order bits. |
| 7244 | if (Op32.getMachineOpcode() == PPC::LHBRX || |
| 7245 | Op32.getMachineOpcode() == PPC::LWBRX) { |
| 7246 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7247 | return true; |
| 7248 | } |
| 7249 | |
| 7250 | // CNT[LT]ZW always produce a 64-bit value in [0,32], and so is zero extended. |
| 7251 | if (Op32.getMachineOpcode() == PPC::CNTLZW || |
| 7252 | Op32.getMachineOpcode() == PPC::CNTTZW) { |
| 7253 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7254 | return true; |
| 7255 | } |
| 7256 | |
| 7257 | // Next, check for those instructions we can look through. |
| 7258 | |
| 7259 | // Assuming the mask does not wrap around, then the higher-order bits are |
| 7260 | // taken directly from the first operand. |
| 7261 | if (Op32.getMachineOpcode() == PPC::RLWIMI && |
| 7262 | Op32.getConstantOperandVal(i: 3) <= Op32.getConstantOperandVal(i: 4)) { |
| 7263 | SmallPtrSet<SDNode *, 16> ToPromote1; |
| 7264 | if (!PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: 0), ToPromote&: ToPromote1)) |
| 7265 | return false; |
| 7266 | |
| 7267 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7268 | ToPromote.insert_range(R&: ToPromote1); |
| 7269 | return true; |
| 7270 | } |
| 7271 | |
| 7272 | // For OR, the higher-order bits are zero if that is true for both operands. |
| 7273 | // For SELECT_I4, the same is true (but the relevant operand numbers are |
| 7274 | // shifted by 1). |
| 7275 | if (Op32.getMachineOpcode() == PPC::OR || |
| 7276 | Op32.getMachineOpcode() == PPC::SELECT_I4) { |
| 7277 | unsigned B = Op32.getMachineOpcode() == PPC::SELECT_I4 ? 1 : 0; |
| 7278 | SmallPtrSet<SDNode *, 16> ToPromote1; |
| 7279 | if (!PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: B+0), ToPromote&: ToPromote1)) |
| 7280 | return false; |
| 7281 | if (!PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: B+1), ToPromote&: ToPromote1)) |
| 7282 | return false; |
| 7283 | |
| 7284 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7285 | ToPromote.insert_range(R&: ToPromote1); |
| 7286 | return true; |
| 7287 | } |
| 7288 | |
| 7289 | // For ORI and ORIS, we need the higher-order bits of the first operand to be |
| 7290 | // zero, and also for the constant to be positive (so that it is not sign |
| 7291 | // extended). |
| 7292 | if (Op32.getMachineOpcode() == PPC::ORI || |
| 7293 | Op32.getMachineOpcode() == PPC::ORIS) { |
| 7294 | SmallPtrSet<SDNode *, 16> ToPromote1; |
| 7295 | if (!PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: 0), ToPromote&: ToPromote1)) |
| 7296 | return false; |
| 7297 | if (!isUInt<15>(x: Op32.getConstantOperandVal(i: 1))) |
| 7298 | return false; |
| 7299 | |
| 7300 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7301 | ToPromote.insert_range(R&: ToPromote1); |
| 7302 | return true; |
| 7303 | } |
| 7304 | |
| 7305 | // The higher-order bits of AND are zero if that is true for at least one of |
| 7306 | // the operands. |
| 7307 | if (Op32.getMachineOpcode() == PPC::AND) { |
| 7308 | SmallPtrSet<SDNode *, 16> ToPromote1, ToPromote2; |
| 7309 | bool Op0OK = |
| 7310 | PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: 0), ToPromote&: ToPromote1); |
| 7311 | bool Op1OK = |
| 7312 | PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: 1), ToPromote&: ToPromote2); |
| 7313 | if (!Op0OK && !Op1OK) |
| 7314 | return false; |
| 7315 | |
| 7316 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7317 | |
| 7318 | if (Op0OK) |
| 7319 | ToPromote.insert_range(R&: ToPromote1); |
| 7320 | |
| 7321 | if (Op1OK) |
| 7322 | ToPromote.insert_range(R&: ToPromote2); |
| 7323 | |
| 7324 | return true; |
| 7325 | } |
| 7326 | |
| 7327 | // For ANDI and ANDIS, the higher-order bits are zero if either that is true |
| 7328 | // of the first operand, or if the second operand is positive (so that it is |
| 7329 | // not sign extended). |
| 7330 | if (Op32.getMachineOpcode() == PPC::ANDI_rec || |
| 7331 | Op32.getMachineOpcode() == PPC::ANDIS_rec) { |
| 7332 | SmallPtrSet<SDNode *, 16> ToPromote1; |
| 7333 | bool Op0OK = |
| 7334 | PeepholePPC64ZExtGather(Op32: Op32.getOperand(i: 0), ToPromote&: ToPromote1); |
| 7335 | bool Op1OK = isUInt<15>(x: Op32.getConstantOperandVal(i: 1)); |
| 7336 | if (!Op0OK && !Op1OK) |
| 7337 | return false; |
| 7338 | |
| 7339 | ToPromote.insert(Ptr: Op32.getNode()); |
| 7340 | |
| 7341 | if (Op0OK) |
| 7342 | ToPromote.insert_range(R&: ToPromote1); |
| 7343 | |
| 7344 | return true; |
| 7345 | } |
| 7346 | |
| 7347 | return false; |
| 7348 | } |
| 7349 | |
| 7350 | void PPCDAGToDAGISel::PeepholePPC64ZExt() { |
| 7351 | if (!Subtarget->isPPC64()) |
| 7352 | return; |
| 7353 | |
| 7354 | // When we zero-extend from i32 to i64, we use a pattern like this: |
| 7355 | // def : Pat<(i64 (zext i32:$in)), |
| 7356 | // (RLDICL (INSERT_SUBREG (i64 (IMPLICIT_DEF)), $in, sub_32), |
| 7357 | // 0, 32)>; |
| 7358 | // There are several 32-bit shift/rotate instructions, however, that will |
| 7359 | // clear the higher-order bits of their output, rendering the RLDICL |
| 7360 | // unnecessary. When that happens, we remove it here, and redefine the |
| 7361 | // relevant 32-bit operation to be a 64-bit operation. |
| 7362 | |
| 7363 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
| 7364 | |
| 7365 | bool MadeChange = false; |
| 7366 | while (Position != CurDAG->allnodes_begin()) { |
| 7367 | SDNode *N = &*--Position; |
| 7368 | // Skip dead nodes and any non-machine opcodes. |
| 7369 | if (N->use_empty() || !N->isMachineOpcode()) |
| 7370 | continue; |
| 7371 | |
| 7372 | if (N->getMachineOpcode() != PPC::RLDICL) |
| 7373 | continue; |
| 7374 | |
| 7375 | if (N->getConstantOperandVal(Num: 1) != 0 || |
| 7376 | N->getConstantOperandVal(Num: 2) != 32) |
| 7377 | continue; |
| 7378 | |
| 7379 | SDValue ISR = N->getOperand(Num: 0); |
| 7380 | if (!ISR.isMachineOpcode() || |
| 7381 | ISR.getMachineOpcode() != TargetOpcode::INSERT_SUBREG) |
| 7382 | continue; |
| 7383 | |
| 7384 | if (!ISR.hasOneUse()) |
| 7385 | continue; |
| 7386 | |
| 7387 | if (ISR.getConstantOperandVal(i: 2) != PPC::sub_32) |
| 7388 | continue; |
| 7389 | |
| 7390 | SDValue IDef = ISR.getOperand(i: 0); |
| 7391 | if (!IDef.isMachineOpcode() || |
| 7392 | IDef.getMachineOpcode() != TargetOpcode::IMPLICIT_DEF) |
| 7393 | continue; |
| 7394 | |
| 7395 | // We now know that we're looking at a canonical i32 -> i64 zext. See if we |
| 7396 | // can get rid of it. |
| 7397 | |
| 7398 | SDValue Op32 = ISR->getOperand(Num: 1); |
| 7399 | if (!Op32.isMachineOpcode()) |
| 7400 | continue; |
| 7401 | |
| 7402 | // There are some 32-bit instructions that always clear the high-order 32 |
| 7403 | // bits, there are also some instructions (like AND) that we can look |
| 7404 | // through. |
| 7405 | SmallPtrSet<SDNode *, 16> ToPromote; |
| 7406 | if (!PeepholePPC64ZExtGather(Op32, ToPromote)) |
| 7407 | continue; |
| 7408 | |
| 7409 | // If the ToPromote set contains nodes that have uses outside of the set |
| 7410 | // (except for the original INSERT_SUBREG), then abort the transformation. |
| 7411 | bool OutsideUse = false; |
| 7412 | for (SDNode *PN : ToPromote) { |
| 7413 | for (SDNode *UN : PN->users()) { |
| 7414 | if (!ToPromote.count(Ptr: UN) && UN != ISR.getNode()) { |
| 7415 | OutsideUse = true; |
| 7416 | break; |
| 7417 | } |
| 7418 | } |
| 7419 | |
| 7420 | if (OutsideUse) |
| 7421 | break; |
| 7422 | } |
| 7423 | if (OutsideUse) |
| 7424 | continue; |
| 7425 | |
| 7426 | MadeChange = true; |
| 7427 | |
| 7428 | // We now know that this zero extension can be removed by promoting to |
| 7429 | // nodes in ToPromote to 64-bit operations, where for operations in the |
| 7430 | // frontier of the set, we need to insert INSERT_SUBREGs for their |
| 7431 | // operands. |
| 7432 | for (SDNode *PN : ToPromote) { |
| 7433 | unsigned NewOpcode; |
| 7434 | switch (PN->getMachineOpcode()) { |
| 7435 | default: |
| 7436 | llvm_unreachable("Don't know the 64-bit variant of this instruction" ); |
| 7437 | case PPC::RLWINM: NewOpcode = PPC::RLWINM8; break; |
| 7438 | case PPC::RLWNM: NewOpcode = PPC::RLWNM8; break; |
| 7439 | case PPC::SLW: NewOpcode = PPC::SLW8; break; |
| 7440 | case PPC::SRW: NewOpcode = PPC::SRW8; break; |
| 7441 | case PPC::LI: NewOpcode = PPC::LI8; break; |
| 7442 | case PPC::LIS: NewOpcode = PPC::LIS8; break; |
| 7443 | case PPC::LHBRX: NewOpcode = PPC::LHBRX8; break; |
| 7444 | case PPC::LWBRX: NewOpcode = PPC::LWBRX8; break; |
| 7445 | case PPC::CNTLZW: NewOpcode = PPC::CNTLZW8; break; |
| 7446 | case PPC::CNTTZW: NewOpcode = PPC::CNTTZW8; break; |
| 7447 | case PPC::RLWIMI: NewOpcode = PPC::RLWIMI8; break; |
| 7448 | case PPC::OR: NewOpcode = PPC::OR8; break; |
| 7449 | case PPC::SELECT_I4: NewOpcode = PPC::SELECT_I8; break; |
| 7450 | case PPC::ORI: NewOpcode = PPC::ORI8; break; |
| 7451 | case PPC::ORIS: NewOpcode = PPC::ORIS8; break; |
| 7452 | case PPC::AND: NewOpcode = PPC::AND8; break; |
| 7453 | case PPC::ANDI_rec: |
| 7454 | NewOpcode = PPC::ANDI8_rec; |
| 7455 | break; |
| 7456 | case PPC::ANDIS_rec: |
| 7457 | NewOpcode = PPC::ANDIS8_rec; |
| 7458 | break; |
| 7459 | } |
| 7460 | |
| 7461 | // Note: During the replacement process, the nodes will be in an |
| 7462 | // inconsistent state (some instructions will have operands with values |
| 7463 | // of the wrong type). Once done, however, everything should be right |
| 7464 | // again. |
| 7465 | |
| 7466 | SmallVector<SDValue, 4> Ops; |
| 7467 | for (const SDValue &V : PN->ops()) { |
| 7468 | if (!ToPromote.count(Ptr: V.getNode()) && V.getValueType() == MVT::i32 && |
| 7469 | !isa<ConstantSDNode>(Val: V)) { |
| 7470 | SDValue ReplOpOps[] = { ISR.getOperand(i: 0), V, ISR.getOperand(i: 2) }; |
| 7471 | SDNode *ReplOp = |
| 7472 | CurDAG->getMachineNode(Opcode: TargetOpcode::INSERT_SUBREG, dl: SDLoc(V), |
| 7473 | VTs: ISR.getNode()->getVTList(), Ops: ReplOpOps); |
| 7474 | Ops.push_back(Elt: SDValue(ReplOp, 0)); |
| 7475 | } else { |
| 7476 | Ops.push_back(Elt: V); |
| 7477 | } |
| 7478 | } |
| 7479 | |
| 7480 | // Because all to-be-promoted nodes only have users that are other |
| 7481 | // promoted nodes (or the original INSERT_SUBREG), we can safely replace |
| 7482 | // the i32 result value type with i64. |
| 7483 | |
| 7484 | SmallVector<EVT, 2> NewVTs; |
| 7485 | SDVTList VTs = PN->getVTList(); |
| 7486 | for (unsigned i = 0, ie = VTs.NumVTs; i != ie; ++i) |
| 7487 | if (VTs.VTs[i] == MVT::i32) |
| 7488 | NewVTs.push_back(Elt: MVT::i64); |
| 7489 | else |
| 7490 | NewVTs.push_back(Elt: VTs.VTs[i]); |
| 7491 | |
| 7492 | LLVM_DEBUG(dbgs() << "PPC64 ZExt Peephole morphing:\nOld: " ); |
| 7493 | LLVM_DEBUG(PN->dump(CurDAG)); |
| 7494 | |
| 7495 | CurDAG->SelectNodeTo(N: PN, MachineOpc: NewOpcode, VTs: CurDAG->getVTList(VTs: NewVTs), Ops); |
| 7496 | |
| 7497 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
| 7498 | LLVM_DEBUG(PN->dump(CurDAG)); |
| 7499 | LLVM_DEBUG(dbgs() << "\n" ); |
| 7500 | } |
| 7501 | |
| 7502 | // Now we replace the original zero extend and its associated INSERT_SUBREG |
| 7503 | // with the value feeding the INSERT_SUBREG (which has now been promoted to |
| 7504 | // return an i64). |
| 7505 | |
| 7506 | LLVM_DEBUG(dbgs() << "PPC64 ZExt Peephole replacing:\nOld: " ); |
| 7507 | LLVM_DEBUG(N->dump(CurDAG)); |
| 7508 | LLVM_DEBUG(dbgs() << "\nNew: " ); |
| 7509 | LLVM_DEBUG(Op32.getNode()->dump(CurDAG)); |
| 7510 | LLVM_DEBUG(dbgs() << "\n" ); |
| 7511 | |
| 7512 | ReplaceUses(F: N, T: Op32.getNode()); |
| 7513 | } |
| 7514 | |
| 7515 | if (MadeChange) |
| 7516 | CurDAG->RemoveDeadNodes(); |
| 7517 | } |
| 7518 | |
| 7519 | static bool isVSXSwap(SDValue N) { |
| 7520 | if (!N->isMachineOpcode()) |
| 7521 | return false; |
| 7522 | unsigned Opc = N->getMachineOpcode(); |
| 7523 | |
| 7524 | // Single-operand XXPERMDI or the regular XXPERMDI/XXSLDWI where the immediate |
| 7525 | // operand is 2. |
| 7526 | if (Opc == PPC::XXPERMDIs) { |
| 7527 | return isa<ConstantSDNode>(Val: N->getOperand(Num: 1)) && |
| 7528 | N->getConstantOperandVal(Num: 1) == 2; |
| 7529 | } else if (Opc == PPC::XXPERMDI || Opc == PPC::XXSLDWI) { |
| 7530 | return N->getOperand(Num: 0) == N->getOperand(Num: 1) && |
| 7531 | isa<ConstantSDNode>(Val: N->getOperand(Num: 2)) && |
| 7532 | N->getConstantOperandVal(Num: 2) == 2; |
| 7533 | } |
| 7534 | |
| 7535 | return false; |
| 7536 | } |
| 7537 | |
| 7538 | // TODO: Make this complete and replace with a table-gen bit. |
| 7539 | static bool isLaneInsensitive(SDValue N) { |
| 7540 | if (!N->isMachineOpcode()) |
| 7541 | return false; |
| 7542 | unsigned Opc = N->getMachineOpcode(); |
| 7543 | |
| 7544 | switch (Opc) { |
| 7545 | default: |
| 7546 | return false; |
| 7547 | case PPC::VAVGSB: |
| 7548 | case PPC::VAVGUB: |
| 7549 | case PPC::VAVGSH: |
| 7550 | case PPC::VAVGUH: |
| 7551 | case PPC::VAVGSW: |
| 7552 | case PPC::VAVGUW: |
| 7553 | case PPC::VMAXFP: |
| 7554 | case PPC::VMAXSB: |
| 7555 | case PPC::VMAXUB: |
| 7556 | case PPC::VMAXSH: |
| 7557 | case PPC::VMAXUH: |
| 7558 | case PPC::VMAXSW: |
| 7559 | case PPC::VMAXUW: |
| 7560 | case PPC::VMINFP: |
| 7561 | case PPC::VMINSB: |
| 7562 | case PPC::VMINUB: |
| 7563 | case PPC::VMINSH: |
| 7564 | case PPC::VMINUH: |
| 7565 | case PPC::VMINSW: |
| 7566 | case PPC::VMINUW: |
| 7567 | case PPC::VADDFP: |
| 7568 | case PPC::VADDUBM: |
| 7569 | case PPC::VADDUHM: |
| 7570 | case PPC::VADDUWM: |
| 7571 | case PPC::VSUBFP: |
| 7572 | case PPC::VSUBUBM: |
| 7573 | case PPC::VSUBUHM: |
| 7574 | case PPC::VSUBUWM: |
| 7575 | case PPC::VAND: |
| 7576 | case PPC::VANDC: |
| 7577 | case PPC::VOR: |
| 7578 | case PPC::VORC: |
| 7579 | case PPC::VXOR: |
| 7580 | case PPC::VNOR: |
| 7581 | case PPC::VMULUWM: |
| 7582 | return true; |
| 7583 | } |
| 7584 | } |
| 7585 | |
| 7586 | // Try to simplify (xxswap (vec-op (xxswap) (xxswap))) where vec-op is |
| 7587 | // lane-insensitive. |
| 7588 | static void reduceVSXSwap(SDNode *N, SelectionDAG *DAG) { |
| 7589 | // Our desired xxswap might be source of COPY_TO_REGCLASS. |
| 7590 | // TODO: Can we put this a common method for DAG? |
| 7591 | auto SkipRCCopy = [](SDValue V) { |
| 7592 | while (V->isMachineOpcode() && |
| 7593 | V->getMachineOpcode() == TargetOpcode::COPY_TO_REGCLASS) { |
| 7594 | // All values in the chain should have single use. |
| 7595 | if (V->use_empty() || !V->user_begin()->isOnlyUserOf(N: V.getNode())) |
| 7596 | return SDValue(); |
| 7597 | V = V->getOperand(Num: 0); |
| 7598 | } |
| 7599 | return V.hasOneUse() ? V : SDValue(); |
| 7600 | }; |
| 7601 | |
| 7602 | SDValue VecOp = SkipRCCopy(N->getOperand(Num: 0)); |
| 7603 | if (!VecOp || !isLaneInsensitive(N: VecOp)) |
| 7604 | return; |
| 7605 | |
| 7606 | SDValue LHS = SkipRCCopy(VecOp.getOperand(i: 0)), |
| 7607 | RHS = SkipRCCopy(VecOp.getOperand(i: 1)); |
| 7608 | if (!LHS || !RHS || !isVSXSwap(N: LHS) || !isVSXSwap(N: RHS)) |
| 7609 | return; |
| 7610 | |
| 7611 | // These swaps may still have chain-uses here, count on dead code elimination |
| 7612 | // in following passes to remove them. |
| 7613 | DAG->ReplaceAllUsesOfValueWith(From: LHS, To: LHS.getOperand(i: 0)); |
| 7614 | DAG->ReplaceAllUsesOfValueWith(From: RHS, To: RHS.getOperand(i: 0)); |
| 7615 | DAG->ReplaceAllUsesOfValueWith(From: SDValue(N, 0), To: N->getOperand(Num: 0)); |
| 7616 | } |
| 7617 | |
| 7618 | // Check if an SDValue has the 'aix-small-tls' global variable attribute. |
| 7619 | static bool hasAIXSmallTLSAttr(SDValue Val) { |
| 7620 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val)) |
| 7621 | if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(Val: GA->getGlobal())) |
| 7622 | if (GV->hasAttribute(Kind: "aix-small-tls" )) |
| 7623 | return true; |
| 7624 | |
| 7625 | return false; |
| 7626 | } |
| 7627 | |
| 7628 | // Is an ADDI eligible for folding for non-TOC-based local-[exec|dynamic] |
| 7629 | // accesses? |
| 7630 | static bool isEligibleToFoldADDIForFasterLocalAccesses(SelectionDAG *DAG, |
| 7631 | SDValue ADDIToFold) { |
| 7632 | // Check if ADDIToFold (the ADDI that we want to fold into local-exec |
| 7633 | // accesses), is truly an ADDI. |
| 7634 | if (!ADDIToFold.isMachineOpcode() || |
| 7635 | (ADDIToFold.getMachineOpcode() != PPC::ADDI8)) |
| 7636 | return false; |
| 7637 | |
| 7638 | // Folding is only allowed for the AIX small-local-[exec|dynamic] TLS target |
| 7639 | // attribute or when the 'aix-small-tls' global variable attribute is present. |
| 7640 | const PPCSubtarget &Subtarget = |
| 7641 | DAG->getMachineFunction().getSubtarget<PPCSubtarget>(); |
| 7642 | SDValue TLSVarNode = ADDIToFold.getOperand(i: 1); |
| 7643 | if (!(Subtarget.hasAIXSmallLocalDynamicTLS() || |
| 7644 | Subtarget.hasAIXSmallLocalExecTLS() || hasAIXSmallTLSAttr(Val: TLSVarNode))) |
| 7645 | return false; |
| 7646 | |
| 7647 | // The second operand of the ADDIToFold should be the global TLS address |
| 7648 | // (the local-exec TLS variable). We only perform the folding if the TLS |
| 7649 | // variable is the second operand. |
| 7650 | GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: TLSVarNode); |
| 7651 | if (!GA) |
| 7652 | return false; |
| 7653 | |
| 7654 | if (DAG->getTarget().getTLSModel(GV: GA->getGlobal()) == TLSModel::LocalExec) { |
| 7655 | // The first operand of the ADDIToFold should be the thread pointer. |
| 7656 | // This transformation is only performed if the first operand of the |
| 7657 | // addi is the thread pointer. |
| 7658 | SDValue TPRegNode = ADDIToFold.getOperand(i: 0); |
| 7659 | RegisterSDNode *TPReg = dyn_cast<RegisterSDNode>(Val: TPRegNode.getNode()); |
| 7660 | if (!TPReg || (TPReg->getReg() != Subtarget.getThreadPointerRegister())) |
| 7661 | return false; |
| 7662 | } |
| 7663 | |
| 7664 | // The local-[exec|dynamic] TLS variable should only have the |
| 7665 | // [MO_TPREL_FLAG|MO_TLSLD_FLAG] target flags, so this optimization is not |
| 7666 | // performed otherwise if the flag is not set. |
| 7667 | unsigned TargetFlags = GA->getTargetFlags(); |
| 7668 | if (!(TargetFlags == PPCII::MO_TPREL_FLAG || |
| 7669 | TargetFlags == PPCII::MO_TLSLD_FLAG)) |
| 7670 | return false; |
| 7671 | |
| 7672 | // If all conditions are satisfied, the ADDI is valid for folding. |
| 7673 | return true; |
| 7674 | } |
| 7675 | |
| 7676 | // For non-TOC-based local-[exec|dynamic] access where an addi is feeding into |
| 7677 | // another addi, fold this sequence into a single addi if possible. Before this |
| 7678 | // optimization, the sequence appears as: |
| 7679 | // addi rN, r13, sym@[le|ld] |
| 7680 | // addi rM, rN, imm |
| 7681 | // After this optimization, we can fold the two addi into a single one: |
| 7682 | // addi rM, r13, sym@[le|ld] + imm |
| 7683 | static void foldADDIForFasterLocalAccesses(SDNode *N, SelectionDAG *DAG) { |
| 7684 | if (N->getMachineOpcode() != PPC::ADDI8) |
| 7685 | return; |
| 7686 | |
| 7687 | // InitialADDI is the addi feeding into N (also an addi), and the addi that |
| 7688 | // we want optimized out. |
| 7689 | SDValue InitialADDI = N->getOperand(Num: 0); |
| 7690 | |
| 7691 | if (!isEligibleToFoldADDIForFasterLocalAccesses(DAG, ADDIToFold: InitialADDI)) |
| 7692 | return; |
| 7693 | |
| 7694 | // The second operand of the InitialADDI should be the global TLS address |
| 7695 | // (the local-[exec|dynamic] TLS variable), with the |
| 7696 | // [MO_TPREL_FLAG|MO_TLSLD_FLAG] target flag. This has been checked in |
| 7697 | // isEligibleToFoldADDIForFasterLocalAccesses(). |
| 7698 | SDValue TLSVarNode = InitialADDI.getOperand(i: 1); |
| 7699 | GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: TLSVarNode); |
| 7700 | assert(GA && "Expecting a valid GlobalAddressSDNode when folding addi into " |
| 7701 | "local-[exec|dynamic] accesses!" ); |
| 7702 | unsigned TargetFlags = GA->getTargetFlags(); |
| 7703 | |
| 7704 | // The second operand of the addi that we want to preserve will be an |
| 7705 | // immediate. We add this immediate, together with the address of the TLS |
| 7706 | // variable found in InitialADDI, in order to preserve the correct TLS address |
| 7707 | // information during assembly printing. The offset is likely to be non-zero |
| 7708 | // when we end up in this case. |
| 7709 | int Offset = N->getConstantOperandVal(Num: 1); |
| 7710 | TLSVarNode = DAG->getTargetGlobalAddress(GV: GA->getGlobal(), DL: SDLoc(GA), VT: MVT::i64, |
| 7711 | offset: Offset, TargetFlags); |
| 7712 | |
| 7713 | (void)DAG->UpdateNodeOperands(N, Op1: InitialADDI.getOperand(i: 0), Op2: TLSVarNode); |
| 7714 | if (InitialADDI.getNode()->use_empty()) |
| 7715 | DAG->RemoveDeadNode(N: InitialADDI.getNode()); |
| 7716 | } |
| 7717 | |
| 7718 | void PPCDAGToDAGISel::PeepholePPC64() { |
| 7719 | SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end(); |
| 7720 | |
| 7721 | while (Position != CurDAG->allnodes_begin()) { |
| 7722 | SDNode *N = &*--Position; |
| 7723 | // Skip dead nodes and any non-machine opcodes. |
| 7724 | if (N->use_empty() || !N->isMachineOpcode()) |
| 7725 | continue; |
| 7726 | |
| 7727 | if (isVSXSwap(N: SDValue(N, 0))) |
| 7728 | reduceVSXSwap(N, DAG: CurDAG); |
| 7729 | |
| 7730 | // This optimization is performed for non-TOC-based local-[exec|dynamic] |
| 7731 | // accesses. |
| 7732 | foldADDIForFasterLocalAccesses(N, DAG: CurDAG); |
| 7733 | |
| 7734 | unsigned FirstOp; |
| 7735 | unsigned StorageOpcode = N->getMachineOpcode(); |
| 7736 | bool RequiresMod4Offset = false; |
| 7737 | |
| 7738 | switch (StorageOpcode) { |
| 7739 | default: continue; |
| 7740 | |
| 7741 | case PPC::LWA: |
| 7742 | case PPC::LD: |
| 7743 | case PPC::DFLOADf64: |
| 7744 | case PPC::DFLOADf32: |
| 7745 | RequiresMod4Offset = true; |
| 7746 | [[fallthrough]]; |
| 7747 | case PPC::LBZ: |
| 7748 | case PPC::LBZ8: |
| 7749 | case PPC::LFD: |
| 7750 | case PPC::LFS: |
| 7751 | case PPC::LHA: |
| 7752 | case PPC::LHA8: |
| 7753 | case PPC::LHZ: |
| 7754 | case PPC::LHZ8: |
| 7755 | case PPC::LWZ: |
| 7756 | case PPC::LWZ8: |
| 7757 | FirstOp = 0; |
| 7758 | break; |
| 7759 | |
| 7760 | case PPC::STD: |
| 7761 | case PPC::DFSTOREf64: |
| 7762 | case PPC::DFSTOREf32: |
| 7763 | RequiresMod4Offset = true; |
| 7764 | [[fallthrough]]; |
| 7765 | case PPC::STB: |
| 7766 | case PPC::STB8: |
| 7767 | case PPC::STFD: |
| 7768 | case PPC::STFS: |
| 7769 | case PPC::STH: |
| 7770 | case PPC::STH8: |
| 7771 | case PPC::STW: |
| 7772 | case PPC::STW8: |
| 7773 | FirstOp = 1; |
| 7774 | break; |
| 7775 | } |
| 7776 | |
| 7777 | // If this is a load or store with a zero offset, or within the alignment, |
| 7778 | // we may be able to fold an add-immediate into the memory operation. |
| 7779 | // The check against alignment is below, as it can't occur until we check |
| 7780 | // the arguments to N |
| 7781 | if (!isa<ConstantSDNode>(Val: N->getOperand(Num: FirstOp))) |
| 7782 | continue; |
| 7783 | |
| 7784 | SDValue Base = N->getOperand(Num: FirstOp + 1); |
| 7785 | if (!Base.isMachineOpcode()) |
| 7786 | continue; |
| 7787 | |
| 7788 | unsigned Flags = 0; |
| 7789 | bool ReplaceFlags = true; |
| 7790 | |
| 7791 | // When the feeding operation is an add-immediate of some sort, |
| 7792 | // determine whether we need to add relocation information to the |
| 7793 | // target flags on the immediate operand when we fold it into the |
| 7794 | // load instruction. |
| 7795 | // |
| 7796 | // For something like ADDItocL8, the relocation information is |
| 7797 | // inferred from the opcode; when we process it in the AsmPrinter, |
| 7798 | // we add the necessary relocation there. A load, though, can receive |
| 7799 | // relocation from various flavors of ADDIxxx, so we need to carry |
| 7800 | // the relocation information in the target flags. |
| 7801 | switch (Base.getMachineOpcode()) { |
| 7802 | default: continue; |
| 7803 | |
| 7804 | case PPC::ADDI8: |
| 7805 | case PPC::ADDI: |
| 7806 | // In some cases (such as TLS) the relocation information |
| 7807 | // is already in place on the operand, so copying the operand |
| 7808 | // is sufficient. |
| 7809 | ReplaceFlags = false; |
| 7810 | break; |
| 7811 | case PPC::ADDIdtprelL: |
| 7812 | Flags = PPCII::MO_DTPREL_LO; |
| 7813 | break; |
| 7814 | case PPC::ADDItlsldL: |
| 7815 | Flags = PPCII::MO_TLSLD_LO; |
| 7816 | break; |
| 7817 | case PPC::ADDItocL8: |
| 7818 | // Skip the following peephole optimizations for ADDItocL8 on AIX which |
| 7819 | // is used for toc-data access. |
| 7820 | if (Subtarget->isAIXABI()) |
| 7821 | continue; |
| 7822 | Flags = PPCII::MO_TOC_LO; |
| 7823 | break; |
| 7824 | } |
| 7825 | |
| 7826 | SDValue ImmOpnd = Base.getOperand(i: 1); |
| 7827 | |
| 7828 | // On PPC64, the TOC base pointer is guaranteed by the ABI only to have |
| 7829 | // 8-byte alignment, and so we can only use offsets less than 8 (otherwise, |
| 7830 | // we might have needed different @ha relocation values for the offset |
| 7831 | // pointers). |
| 7832 | int MaxDisplacement = 7; |
| 7833 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: ImmOpnd)) { |
| 7834 | const GlobalValue *GV = GA->getGlobal(); |
| 7835 | Align Alignment = GV->getPointerAlignment(DL: CurDAG->getDataLayout()); |
| 7836 | MaxDisplacement = std::min(a: (int)Alignment.value() - 1, b: MaxDisplacement); |
| 7837 | } |
| 7838 | |
| 7839 | bool UpdateHBase = false; |
| 7840 | SDValue HBase = Base.getOperand(i: 0); |
| 7841 | |
| 7842 | int Offset = N->getConstantOperandVal(Num: FirstOp); |
| 7843 | if (ReplaceFlags) { |
| 7844 | if (Offset < 0 || Offset > MaxDisplacement) { |
| 7845 | // If we have a addi(toc@l)/addis(toc@ha) pair, and the addis has only |
| 7846 | // one use, then we can do this for any offset, we just need to also |
| 7847 | // update the offset (i.e. the symbol addend) on the addis also. |
| 7848 | if (Base.getMachineOpcode() != PPC::ADDItocL8) |
| 7849 | continue; |
| 7850 | |
| 7851 | if (!HBase.isMachineOpcode() || |
| 7852 | HBase.getMachineOpcode() != PPC::ADDIStocHA8) |
| 7853 | continue; |
| 7854 | |
| 7855 | if (!Base.hasOneUse() || !HBase.hasOneUse()) |
| 7856 | continue; |
| 7857 | |
| 7858 | SDValue HImmOpnd = HBase.getOperand(i: 1); |
| 7859 | if (HImmOpnd != ImmOpnd) |
| 7860 | continue; |
| 7861 | |
| 7862 | UpdateHBase = true; |
| 7863 | } |
| 7864 | } else { |
| 7865 | // Global addresses can be folded, but only if they are sufficiently |
| 7866 | // aligned. |
| 7867 | if (RequiresMod4Offset) { |
| 7868 | if (GlobalAddressSDNode *GA = |
| 7869 | dyn_cast<GlobalAddressSDNode>(Val&: ImmOpnd)) { |
| 7870 | const GlobalValue *GV = GA->getGlobal(); |
| 7871 | Align Alignment = GV->getPointerAlignment(DL: CurDAG->getDataLayout()); |
| 7872 | if (Alignment < 4) |
| 7873 | continue; |
| 7874 | } |
| 7875 | } |
| 7876 | |
| 7877 | // If we're directly folding the addend from an addi instruction, then: |
| 7878 | // 1. In general, the offset on the memory access must be zero. |
| 7879 | // 2. If the addend is a constant, then it can be combined with a |
| 7880 | // non-zero offset, but only if the result meets the encoding |
| 7881 | // requirements. |
| 7882 | if (auto *C = dyn_cast<ConstantSDNode>(Val&: ImmOpnd)) { |
| 7883 | Offset += C->getSExtValue(); |
| 7884 | |
| 7885 | if (RequiresMod4Offset && (Offset % 4) != 0) |
| 7886 | continue; |
| 7887 | |
| 7888 | if (!isInt<16>(x: Offset)) |
| 7889 | continue; |
| 7890 | |
| 7891 | ImmOpnd = CurDAG->getSignedTargetConstant(Val: Offset, DL: SDLoc(ImmOpnd), |
| 7892 | VT: ImmOpnd.getValueType()); |
| 7893 | } else if (Offset != 0) { |
| 7894 | // This optimization is performed for non-TOC-based local-[exec|dynamic] |
| 7895 | // accesses. |
| 7896 | if (isEligibleToFoldADDIForFasterLocalAccesses(DAG: CurDAG, ADDIToFold: Base)) { |
| 7897 | // Add the non-zero offset information into the load or store |
| 7898 | // instruction to be used for non-TOC-based local-[exec|dynamic] |
| 7899 | // accesses. |
| 7900 | GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: ImmOpnd); |
| 7901 | assert(GA && "Expecting a valid GlobalAddressSDNode when folding " |
| 7902 | "addi into local-[exec|dynamic] accesses!" ); |
| 7903 | ImmOpnd = CurDAG->getTargetGlobalAddress(GV: GA->getGlobal(), DL: SDLoc(GA), |
| 7904 | VT: MVT::i64, offset: Offset, |
| 7905 | TargetFlags: GA->getTargetFlags()); |
| 7906 | } else |
| 7907 | continue; |
| 7908 | } |
| 7909 | } |
| 7910 | |
| 7911 | // We found an opportunity. Reverse the operands from the add |
| 7912 | // immediate and substitute them into the load or store. If |
| 7913 | // needed, update the target flags for the immediate operand to |
| 7914 | // reflect the necessary relocation information. |
| 7915 | LLVM_DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase: " ); |
| 7916 | LLVM_DEBUG(Base->dump(CurDAG)); |
| 7917 | LLVM_DEBUG(dbgs() << "\nN: " ); |
| 7918 | LLVM_DEBUG(N->dump(CurDAG)); |
| 7919 | LLVM_DEBUG(dbgs() << "\n" ); |
| 7920 | |
| 7921 | // If the relocation information isn't already present on the |
| 7922 | // immediate operand, add it now. |
| 7923 | if (ReplaceFlags) { |
| 7924 | if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(Val&: ImmOpnd)) { |
| 7925 | SDLoc dl(GA); |
| 7926 | const GlobalValue *GV = GA->getGlobal(); |
| 7927 | Align Alignment = GV->getPointerAlignment(DL: CurDAG->getDataLayout()); |
| 7928 | // We can't perform this optimization for data whose alignment |
| 7929 | // is insufficient for the instruction encoding. |
| 7930 | if (Alignment < 4 && (RequiresMod4Offset || (Offset % 4) != 0)) { |
| 7931 | LLVM_DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n" ); |
| 7932 | continue; |
| 7933 | } |
| 7934 | ImmOpnd = CurDAG->getTargetGlobalAddress(GV, DL: dl, VT: MVT::i64, offset: Offset, TargetFlags: Flags); |
| 7935 | } else if (ConstantPoolSDNode *CP = |
| 7936 | dyn_cast<ConstantPoolSDNode>(Val&: ImmOpnd)) { |
| 7937 | const Constant *C = CP->getConstVal(); |
| 7938 | ImmOpnd = CurDAG->getTargetConstantPool(C, VT: MVT::i64, Align: CP->getAlign(), |
| 7939 | Offset, TargetFlags: Flags); |
| 7940 | } |
| 7941 | } |
| 7942 | |
| 7943 | if (FirstOp == 1) // Store |
| 7944 | (void)CurDAG->UpdateNodeOperands(N, Op1: N->getOperand(Num: 0), Op2: ImmOpnd, |
| 7945 | Op3: Base.getOperand(i: 0), Op4: N->getOperand(Num: 3)); |
| 7946 | else // Load |
| 7947 | (void)CurDAG->UpdateNodeOperands(N, Op1: ImmOpnd, Op2: Base.getOperand(i: 0), |
| 7948 | Op3: N->getOperand(Num: 2)); |
| 7949 | |
| 7950 | if (UpdateHBase) |
| 7951 | (void)CurDAG->UpdateNodeOperands(N: HBase.getNode(), Op1: HBase.getOperand(i: 0), |
| 7952 | Op2: ImmOpnd); |
| 7953 | |
| 7954 | // The add-immediate may now be dead, in which case remove it. |
| 7955 | if (Base.getNode()->use_empty()) |
| 7956 | CurDAG->RemoveDeadNode(N: Base.getNode()); |
| 7957 | } |
| 7958 | } |
| 7959 | |
| 7960 | /// createPPCISelDag - This pass converts a legalized DAG into a |
| 7961 | /// PowerPC-specific DAG, ready for instruction scheduling. |
| 7962 | /// |
| 7963 | FunctionPass *llvm::createPPCISelDag(PPCTargetMachine &TM, |
| 7964 | CodeGenOptLevel OptLevel) { |
| 7965 | return new PPCDAGToDAGISelLegacy(TM, OptLevel); |
| 7966 | } |
| 7967 | |