| 1 | //===- HexagonNewValueJump.cpp - Hexagon Backend New Value Jump -----------===// |
| 2 | // |
| 3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | // See https://llvm.org/LICENSE.txt for license information. |
| 5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | // |
| 7 | //===----------------------------------------------------------------------===// |
| 8 | // |
| 9 | // This implements NewValueJump pass in Hexagon. |
| 10 | // Ideally, we should merge this as a Peephole pass prior to register |
| 11 | // allocation, but because we have a spill in between the feeder and new value |
| 12 | // jump instructions, we are forced to write after register allocation. |
| 13 | // Having said that, we should re-attempt to pull this earlier at some point |
| 14 | // in future. |
| 15 | |
| 16 | // The basic approach looks for sequence of predicated jump, compare instruction |
| 17 | // that generates the predicate and, the feeder to the predicate. Once it finds |
| 18 | // all, it collapses compare and jump instruction into a new value jump |
| 19 | // instructions. |
| 20 | // |
| 21 | //===----------------------------------------------------------------------===// |
| 22 | |
| 23 | #include "llvm/InitializePasses.h" |
| 24 | #include "Hexagon.h" |
| 25 | #include "HexagonInstrInfo.h" |
| 26 | #include "HexagonRegisterInfo.h" |
| 27 | #include "HexagonSubtarget.h" |
| 28 | #include "llvm/ADT/Statistic.h" |
| 29 | #include "llvm/CodeGen/MachineBasicBlock.h" |
| 30 | #include "llvm/CodeGen/MachineBranchProbabilityInfo.h" |
| 31 | #include "llvm/CodeGen/MachineFunction.h" |
| 32 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 33 | #include "llvm/CodeGen/MachineInstr.h" |
| 34 | #include "llvm/CodeGen/MachineInstrBuilder.h" |
| 35 | #include "llvm/CodeGen/MachineOperand.h" |
| 36 | #include "llvm/CodeGen/MachineRegisterInfo.h" |
| 37 | #include "llvm/CodeGen/TargetOpcodes.h" |
| 38 | #include "llvm/CodeGen/TargetRegisterInfo.h" |
| 39 | #include "llvm/CodeGen/TargetSubtargetInfo.h" |
| 40 | #include "llvm/IR/DebugLoc.h" |
| 41 | #include "llvm/MC/MCInstrDesc.h" |
| 42 | #include "llvm/Pass.h" |
| 43 | #include "llvm/Support/BranchProbability.h" |
| 44 | #include "llvm/Support/CommandLine.h" |
| 45 | #include "llvm/Support/Debug.h" |
| 46 | #include "llvm/Support/ErrorHandling.h" |
| 47 | #include "llvm/Support/MathExtras.h" |
| 48 | #include "llvm/Support/raw_ostream.h" |
| 49 | #include <cassert> |
| 50 | #include <cstdint> |
| 51 | #include <iterator> |
| 52 | |
| 53 | using namespace llvm; |
| 54 | |
| 55 | #define DEBUG_TYPE "hexagon-nvj" |
| 56 | |
| 57 | STATISTIC(NumNVJGenerated, "Number of New Value Jump Instructions created" ); |
| 58 | |
| 59 | static cl::opt<int> DbgNVJCount("nvj-count" , cl::init(Val: -1), cl::Hidden, |
| 60 | cl::desc("Maximum number of predicated jumps to be converted to " |
| 61 | "New Value Jump" )); |
| 62 | |
| 63 | static cl::opt<bool> DisableNewValueJumps("disable-nvjump" , cl::Hidden, |
| 64 | cl::desc("Disable New Value Jumps" )); |
| 65 | |
| 66 | namespace { |
| 67 | |
| 68 | struct HexagonNewValueJump : public MachineFunctionPass { |
| 69 | static char ID; |
| 70 | |
| 71 | HexagonNewValueJump() : MachineFunctionPass(ID) {} |
| 72 | |
| 73 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 74 | AU.addRequired<MachineBranchProbabilityInfoWrapperPass>(); |
| 75 | MachineFunctionPass::getAnalysisUsage(AU); |
| 76 | } |
| 77 | |
| 78 | StringRef getPassName() const override { return "Hexagon NewValueJump" ; } |
| 79 | |
| 80 | bool runOnMachineFunction(MachineFunction &Fn) override; |
| 81 | |
| 82 | MachineFunctionProperties getRequiredProperties() const override { |
| 83 | return MachineFunctionProperties().setNoVRegs(); |
| 84 | } |
| 85 | |
| 86 | private: |
| 87 | const HexagonInstrInfo *QII; |
| 88 | const HexagonRegisterInfo *QRI; |
| 89 | |
| 90 | /// A handle to the branch probability pass. |
| 91 | const MachineBranchProbabilityInfo *MBPI; |
| 92 | |
| 93 | bool isNewValueJumpCandidate(const MachineInstr &MI) const; |
| 94 | }; |
| 95 | |
| 96 | } // end anonymous namespace |
| 97 | |
| 98 | char HexagonNewValueJump::ID = 0; |
| 99 | |
| 100 | INITIALIZE_PASS_BEGIN(HexagonNewValueJump, "hexagon-nvj" , |
| 101 | "Hexagon NewValueJump" , false, false) |
| 102 | INITIALIZE_PASS_DEPENDENCY(MachineBranchProbabilityInfoWrapperPass) |
| 103 | INITIALIZE_PASS_END(HexagonNewValueJump, "hexagon-nvj" , |
| 104 | "Hexagon NewValueJump" , false, false) |
| 105 | |
| 106 | // We have identified this II could be feeder to NVJ, |
| 107 | // verify that it can be. |
| 108 | static bool canBeFeederToNewValueJump(const HexagonInstrInfo *QII, |
| 109 | const TargetRegisterInfo *TRI, |
| 110 | MachineBasicBlock::iterator II, |
| 111 | MachineBasicBlock::iterator end, |
| 112 | MachineBasicBlock::iterator skip, |
| 113 | MachineFunction &MF) { |
| 114 | // Predicated instruction can not be feeder to NVJ. |
| 115 | if (QII->isPredicated(MI: *II)) |
| 116 | return false; |
| 117 | |
| 118 | // Bail out if feederReg is a paired register (double regs in |
| 119 | // our case). One would think that we can check to see if a given |
| 120 | // register cmpReg1 or cmpReg2 is a sub register of feederReg |
| 121 | // using -- if (QRI->isSubRegister(feederReg, cmpReg1) logic |
| 122 | // before the callsite of this function |
| 123 | // But we can not as it comes in the following fashion. |
| 124 | // %d0 = Hexagon_S2_lsr_r_p killed %d0, killed %r2 |
| 125 | // %r0 = KILL %r0, implicit killed %d0 |
| 126 | // %p0 = CMPEQri killed %r0, 0 |
| 127 | // Hence, we need to check if it's a KILL instruction. |
| 128 | if (II->getOpcode() == TargetOpcode::KILL) |
| 129 | return false; |
| 130 | |
| 131 | if (II->isImplicitDef()) |
| 132 | return false; |
| 133 | |
| 134 | if (QII->isSolo(MI: *II)) |
| 135 | return false; |
| 136 | |
| 137 | if (QII->isFloat(MI: *II)) |
| 138 | return false; |
| 139 | |
| 140 | // Make sure that the (unique) def operand is a register from IntRegs. |
| 141 | bool HadDef = false; |
| 142 | for (const MachineOperand &Op : II->operands()) { |
| 143 | if (!Op.isReg() || !Op.isDef()) |
| 144 | continue; |
| 145 | if (HadDef) |
| 146 | return false; |
| 147 | HadDef = true; |
| 148 | if (!Hexagon::IntRegsRegClass.contains(Reg: Op.getReg())) |
| 149 | return false; |
| 150 | } |
| 151 | assert(HadDef); |
| 152 | |
| 153 | // Make sure there is no 'def' or 'use' of any of the uses of |
| 154 | // feeder insn between its definition, this MI and jump, jmpInst |
| 155 | // skipping compare, cmpInst. |
| 156 | // Here's the example. |
| 157 | // r21=memub(r22+r24<<#0) |
| 158 | // p0 = cmp.eq(r21, #0) |
| 159 | // r4=memub(r3+r21<<#0) |
| 160 | // if (p0.new) jump:t .LBB29_45 |
| 161 | // Without this check, it will be converted into |
| 162 | // r4=memub(r3+r21<<#0) |
| 163 | // r21=memub(r22+r24<<#0) |
| 164 | // p0 = cmp.eq(r21, #0) |
| 165 | // if (p0.new) jump:t .LBB29_45 |
| 166 | // and result WAR hazards if converted to New Value Jump. |
| 167 | for (unsigned i = 0; i < II->getNumOperands(); ++i) { |
| 168 | if (II->getOperand(i).isReg() && |
| 169 | (II->getOperand(i).isUse() || II->getOperand(i).isDef())) { |
| 170 | MachineBasicBlock::iterator localII = II; |
| 171 | ++localII; |
| 172 | Register Reg = II->getOperand(i).getReg(); |
| 173 | for (MachineBasicBlock::iterator localBegin = localII; localBegin != end; |
| 174 | ++localBegin) { |
| 175 | if (localBegin == skip) |
| 176 | continue; |
| 177 | // Check for Subregisters too. |
| 178 | if (localBegin->modifiesRegister(Reg, TRI) || |
| 179 | localBegin->readsRegister(Reg, TRI)) |
| 180 | return false; |
| 181 | } |
| 182 | } |
| 183 | } |
| 184 | return true; |
| 185 | } |
| 186 | |
| 187 | // These are the common checks that need to performed |
| 188 | // to determine if |
| 189 | // 1. compare instruction can be moved before jump. |
| 190 | // 2. feeder to the compare instruction can be moved before jump. |
| 191 | static bool commonChecksToProhibitNewValueJump(bool afterRA, |
| 192 | MachineBasicBlock::iterator MII) { |
| 193 | // If store in path, bail out. |
| 194 | if (MII->mayStore()) |
| 195 | return false; |
| 196 | |
| 197 | // if call in path, bail out. |
| 198 | if (MII->isCall()) |
| 199 | return false; |
| 200 | |
| 201 | // if NVJ is running prior to RA, do the following checks. |
| 202 | if (!afterRA) { |
| 203 | // The following Target Opcode instructions are spurious |
| 204 | // to new value jump. If they are in the path, bail out. |
| 205 | // KILL sets kill flag on the opcode. It also sets up a |
| 206 | // single register, out of pair. |
| 207 | // %d0 = S2_lsr_r_p killed %d0, killed %r2 |
| 208 | // %r0 = KILL %r0, implicit killed %d0 |
| 209 | // %p0 = C2_cmpeqi killed %r0, 0 |
| 210 | // PHI can be anything after RA. |
| 211 | // COPY can remateriaze things in between feeder, compare and nvj. |
| 212 | if (MII->getOpcode() == TargetOpcode::KILL || |
| 213 | MII->getOpcode() == TargetOpcode::PHI || |
| 214 | MII->getOpcode() == TargetOpcode::COPY) |
| 215 | return false; |
| 216 | |
| 217 | // The following pseudo Hexagon instructions sets "use" and "def" |
| 218 | // of registers by individual passes in the backend. At this time, |
| 219 | // we don't know the scope of usage and definitions of these |
| 220 | // instructions. |
| 221 | if (MII->getOpcode() == Hexagon::LDriw_pred || |
| 222 | MII->getOpcode() == Hexagon::STriw_pred) |
| 223 | return false; |
| 224 | } |
| 225 | |
| 226 | return true; |
| 227 | } |
| 228 | |
| 229 | static bool canCompareBeNewValueJump(const HexagonInstrInfo *QII, |
| 230 | const TargetRegisterInfo *TRI, |
| 231 | MachineBasicBlock::iterator II, |
| 232 | unsigned pReg, |
| 233 | bool secondReg, |
| 234 | bool optLocation, |
| 235 | MachineBasicBlock::iterator end, |
| 236 | MachineFunction &MF) { |
| 237 | MachineInstr &MI = *II; |
| 238 | |
| 239 | // If the second operand of the compare is an imm, make sure it's in the |
| 240 | // range specified by the arch. |
| 241 | if (!secondReg) { |
| 242 | const MachineOperand &Op2 = MI.getOperand(i: 2); |
| 243 | if (!Op2.isImm()) |
| 244 | return false; |
| 245 | |
| 246 | int64_t v = Op2.getImm(); |
| 247 | bool Valid = false; |
| 248 | |
| 249 | switch (MI.getOpcode()) { |
| 250 | case Hexagon::C2_cmpeqi: |
| 251 | case Hexagon::C4_cmpneqi: |
| 252 | case Hexagon::C2_cmpgti: |
| 253 | case Hexagon::C4_cmpltei: |
| 254 | Valid = (isUInt<5>(x: v) || v == -1); |
| 255 | break; |
| 256 | case Hexagon::C2_cmpgtui: |
| 257 | case Hexagon::C4_cmplteui: |
| 258 | Valid = isUInt<5>(x: v); |
| 259 | break; |
| 260 | case Hexagon::S2_tstbit_i: |
| 261 | case Hexagon::S4_ntstbit_i: |
| 262 | Valid = (v == 0); |
| 263 | break; |
| 264 | } |
| 265 | |
| 266 | if (!Valid) |
| 267 | return false; |
| 268 | } |
| 269 | |
| 270 | unsigned cmpReg1, cmpOp2 = 0; // cmpOp2 assignment silences compiler warning. |
| 271 | cmpReg1 = MI.getOperand(i: 1).getReg(); |
| 272 | |
| 273 | if (secondReg) { |
| 274 | cmpOp2 = MI.getOperand(i: 2).getReg(); |
| 275 | |
| 276 | // If the same register appears as both operands, we cannot generate a new |
| 277 | // value compare. Only one operand may use the .new suffix. |
| 278 | if (cmpReg1 == cmpOp2) |
| 279 | return false; |
| 280 | |
| 281 | // Make sure that the second register is not from COPY |
| 282 | // at machine code level, we don't need this, but if we decide |
| 283 | // to move new value jump prior to RA, we would be needing this. |
| 284 | MachineRegisterInfo &MRI = MF.getRegInfo(); |
| 285 | if (!Register::isPhysicalRegister(Reg: cmpOp2)) { |
| 286 | MachineInstr *def = MRI.getVRegDef(Reg: cmpOp2); |
| 287 | if (def->getOpcode() == TargetOpcode::COPY) |
| 288 | return false; |
| 289 | } |
| 290 | } |
| 291 | |
| 292 | // Walk the instructions after the compare (predicate def) to the jump, |
| 293 | // and satisfy the following conditions. |
| 294 | ++II; |
| 295 | for (MachineBasicBlock::iterator localII = II; localII != end; ++localII) { |
| 296 | if (localII->isDebugInstr()) |
| 297 | continue; |
| 298 | |
| 299 | // Check 1. |
| 300 | // If "common" checks fail, bail out. |
| 301 | if (!commonChecksToProhibitNewValueJump(afterRA: optLocation, MII: localII)) |
| 302 | return false; |
| 303 | |
| 304 | // Check 2. |
| 305 | // If there is a def or use of predicate (result of compare), bail out. |
| 306 | if (localII->modifiesRegister(Reg: pReg, TRI) || |
| 307 | localII->readsRegister(Reg: pReg, TRI)) |
| 308 | return false; |
| 309 | |
| 310 | // Check 3. |
| 311 | // If there is a def of any of the use of the compare (operands of compare), |
| 312 | // bail out. |
| 313 | // Eg. |
| 314 | // p0 = cmp.eq(r2, r0) |
| 315 | // r2 = r4 |
| 316 | // if (p0.new) jump:t .LBB28_3 |
| 317 | if (localII->modifiesRegister(Reg: cmpReg1, TRI) || |
| 318 | (secondReg && localII->modifiesRegister(Reg: cmpOp2, TRI))) |
| 319 | return false; |
| 320 | } |
| 321 | return true; |
| 322 | } |
| 323 | |
| 324 | // Given a compare operator, return a matching New Value Jump compare operator. |
| 325 | // Make sure that MI here is included in isNewValueJumpCandidate. |
| 326 | static unsigned getNewValueJumpOpcode(MachineInstr *MI, int reg, |
| 327 | bool secondRegNewified, |
| 328 | MachineBasicBlock *jmpTarget, |
| 329 | const MachineBranchProbabilityInfo |
| 330 | *MBPI) { |
| 331 | bool taken = false; |
| 332 | MachineBasicBlock *Src = MI->getParent(); |
| 333 | const BranchProbability Prediction = |
| 334 | MBPI->getEdgeProbability(Src, Dst: jmpTarget); |
| 335 | |
| 336 | if (Prediction >= BranchProbability(1,2)) |
| 337 | taken = true; |
| 338 | |
| 339 | switch (MI->getOpcode()) { |
| 340 | case Hexagon::C2_cmpeq: |
| 341 | return taken ? Hexagon::J4_cmpeq_t_jumpnv_t |
| 342 | : Hexagon::J4_cmpeq_t_jumpnv_nt; |
| 343 | |
| 344 | case Hexagon::C2_cmpeqi: |
| 345 | if (reg >= 0) |
| 346 | return taken ? Hexagon::J4_cmpeqi_t_jumpnv_t |
| 347 | : Hexagon::J4_cmpeqi_t_jumpnv_nt; |
| 348 | return taken ? Hexagon::J4_cmpeqn1_t_jumpnv_t |
| 349 | : Hexagon::J4_cmpeqn1_t_jumpnv_nt; |
| 350 | |
| 351 | case Hexagon::C4_cmpneqi: |
| 352 | if (reg >= 0) |
| 353 | return taken ? Hexagon::J4_cmpeqi_f_jumpnv_t |
| 354 | : Hexagon::J4_cmpeqi_f_jumpnv_nt; |
| 355 | return taken ? Hexagon::J4_cmpeqn1_f_jumpnv_t : |
| 356 | Hexagon::J4_cmpeqn1_f_jumpnv_nt; |
| 357 | |
| 358 | case Hexagon::C2_cmpgt: |
| 359 | if (secondRegNewified) |
| 360 | return taken ? Hexagon::J4_cmplt_t_jumpnv_t |
| 361 | : Hexagon::J4_cmplt_t_jumpnv_nt; |
| 362 | return taken ? Hexagon::J4_cmpgt_t_jumpnv_t |
| 363 | : Hexagon::J4_cmpgt_t_jumpnv_nt; |
| 364 | |
| 365 | case Hexagon::C2_cmpgti: |
| 366 | if (reg >= 0) |
| 367 | return taken ? Hexagon::J4_cmpgti_t_jumpnv_t |
| 368 | : Hexagon::J4_cmpgti_t_jumpnv_nt; |
| 369 | return taken ? Hexagon::J4_cmpgtn1_t_jumpnv_t |
| 370 | : Hexagon::J4_cmpgtn1_t_jumpnv_nt; |
| 371 | |
| 372 | case Hexagon::C2_cmpgtu: |
| 373 | if (secondRegNewified) |
| 374 | return taken ? Hexagon::J4_cmpltu_t_jumpnv_t |
| 375 | : Hexagon::J4_cmpltu_t_jumpnv_nt; |
| 376 | return taken ? Hexagon::J4_cmpgtu_t_jumpnv_t |
| 377 | : Hexagon::J4_cmpgtu_t_jumpnv_nt; |
| 378 | |
| 379 | case Hexagon::C2_cmpgtui: |
| 380 | return taken ? Hexagon::J4_cmpgtui_t_jumpnv_t |
| 381 | : Hexagon::J4_cmpgtui_t_jumpnv_nt; |
| 382 | |
| 383 | case Hexagon::C4_cmpneq: |
| 384 | return taken ? Hexagon::J4_cmpeq_f_jumpnv_t |
| 385 | : Hexagon::J4_cmpeq_f_jumpnv_nt; |
| 386 | |
| 387 | case Hexagon::C4_cmplte: |
| 388 | if (secondRegNewified) |
| 389 | return taken ? Hexagon::J4_cmplt_f_jumpnv_t |
| 390 | : Hexagon::J4_cmplt_f_jumpnv_nt; |
| 391 | return taken ? Hexagon::J4_cmpgt_f_jumpnv_t |
| 392 | : Hexagon::J4_cmpgt_f_jumpnv_nt; |
| 393 | |
| 394 | case Hexagon::C4_cmplteu: |
| 395 | if (secondRegNewified) |
| 396 | return taken ? Hexagon::J4_cmpltu_f_jumpnv_t |
| 397 | : Hexagon::J4_cmpltu_f_jumpnv_nt; |
| 398 | return taken ? Hexagon::J4_cmpgtu_f_jumpnv_t |
| 399 | : Hexagon::J4_cmpgtu_f_jumpnv_nt; |
| 400 | |
| 401 | case Hexagon::C4_cmpltei: |
| 402 | if (reg >= 0) |
| 403 | return taken ? Hexagon::J4_cmpgti_f_jumpnv_t |
| 404 | : Hexagon::J4_cmpgti_f_jumpnv_nt; |
| 405 | return taken ? Hexagon::J4_cmpgtn1_f_jumpnv_t |
| 406 | : Hexagon::J4_cmpgtn1_f_jumpnv_nt; |
| 407 | |
| 408 | case Hexagon::C4_cmplteui: |
| 409 | return taken ? Hexagon::J4_cmpgtui_f_jumpnv_t |
| 410 | : Hexagon::J4_cmpgtui_f_jumpnv_nt; |
| 411 | |
| 412 | default: |
| 413 | llvm_unreachable("Could not find matching New Value Jump instruction." ); |
| 414 | } |
| 415 | // return *some value* to avoid compiler warning |
| 416 | return 0; |
| 417 | } |
| 418 | |
| 419 | bool HexagonNewValueJump::isNewValueJumpCandidate( |
| 420 | const MachineInstr &MI) const { |
| 421 | switch (MI.getOpcode()) { |
| 422 | case Hexagon::C2_cmpeq: |
| 423 | case Hexagon::C2_cmpeqi: |
| 424 | case Hexagon::C2_cmpgt: |
| 425 | case Hexagon::C2_cmpgti: |
| 426 | case Hexagon::C2_cmpgtu: |
| 427 | case Hexagon::C2_cmpgtui: |
| 428 | case Hexagon::C4_cmpneq: |
| 429 | case Hexagon::C4_cmpneqi: |
| 430 | case Hexagon::C4_cmplte: |
| 431 | case Hexagon::C4_cmplteu: |
| 432 | case Hexagon::C4_cmpltei: |
| 433 | case Hexagon::C4_cmplteui: |
| 434 | return true; |
| 435 | |
| 436 | default: |
| 437 | return false; |
| 438 | } |
| 439 | } |
| 440 | |
| 441 | bool HexagonNewValueJump::runOnMachineFunction(MachineFunction &MF) { |
| 442 | LLVM_DEBUG(dbgs() << "********** Hexagon New Value Jump **********\n" |
| 443 | << "********** Function: " << MF.getName() << "\n" ); |
| 444 | |
| 445 | if (skipFunction(F: MF.getFunction())) |
| 446 | return false; |
| 447 | |
| 448 | // If we move NewValueJump before register allocation we'll need live variable |
| 449 | // analysis here too. |
| 450 | |
| 451 | QII = static_cast<const HexagonInstrInfo *>(MF.getSubtarget().getInstrInfo()); |
| 452 | QRI = static_cast<const HexagonRegisterInfo *>( |
| 453 | MF.getSubtarget().getRegisterInfo()); |
| 454 | MBPI = &getAnalysis<MachineBranchProbabilityInfoWrapperPass>().getMBPI(); |
| 455 | |
| 456 | if (DisableNewValueJumps || |
| 457 | !MF.getSubtarget<HexagonSubtarget>().useNewValueJumps()) |
| 458 | return false; |
| 459 | |
| 460 | int nvjCount = DbgNVJCount; |
| 461 | int nvjGenerated = 0; |
| 462 | |
| 463 | // Loop through all the bb's of the function |
| 464 | for (MachineFunction::iterator MBBb = MF.begin(), MBBe = MF.end(); |
| 465 | MBBb != MBBe; ++MBBb) { |
| 466 | MachineBasicBlock *MBB = &*MBBb; |
| 467 | |
| 468 | LLVM_DEBUG(dbgs() << "** dumping bb ** " << MBB->getNumber() << "\n" ); |
| 469 | LLVM_DEBUG(MBB->dump()); |
| 470 | LLVM_DEBUG(dbgs() << "\n" |
| 471 | << "********** dumping instr bottom up **********\n" ); |
| 472 | bool foundJump = false; |
| 473 | bool foundCompare = false; |
| 474 | bool invertPredicate = false; |
| 475 | unsigned predReg = 0; // predicate reg of the jump. |
| 476 | unsigned cmpReg1 = 0; |
| 477 | int cmpOp2 = 0; |
| 478 | MachineBasicBlock::iterator jmpPos; |
| 479 | MachineBasicBlock::iterator cmpPos; |
| 480 | MachineInstr *cmpInstr = nullptr, *jmpInstr = nullptr; |
| 481 | MachineBasicBlock *jmpTarget = nullptr; |
| 482 | bool afterRA = false; |
| 483 | bool isSecondOpReg = false; |
| 484 | bool isSecondOpNewified = false; |
| 485 | // Traverse the basic block - bottom up |
| 486 | for (MachineBasicBlock::iterator MII = MBB->end(), E = MBB->begin(); |
| 487 | MII != E;) { |
| 488 | MachineInstr &MI = *--MII; |
| 489 | if (MI.isDebugInstr()) { |
| 490 | continue; |
| 491 | } |
| 492 | |
| 493 | if ((nvjCount == 0) || (nvjCount > -1 && nvjCount <= nvjGenerated)) |
| 494 | break; |
| 495 | |
| 496 | LLVM_DEBUG(dbgs() << "Instr: " ; MI.dump(); dbgs() << "\n" ); |
| 497 | |
| 498 | if (!foundJump && (MI.getOpcode() == Hexagon::J2_jumpt || |
| 499 | MI.getOpcode() == Hexagon::J2_jumptpt || |
| 500 | MI.getOpcode() == Hexagon::J2_jumpf || |
| 501 | MI.getOpcode() == Hexagon::J2_jumpfpt || |
| 502 | MI.getOpcode() == Hexagon::J2_jumptnewpt || |
| 503 | MI.getOpcode() == Hexagon::J2_jumptnew || |
| 504 | MI.getOpcode() == Hexagon::J2_jumpfnewpt || |
| 505 | MI.getOpcode() == Hexagon::J2_jumpfnew)) { |
| 506 | // This is where you would insert your compare and |
| 507 | // instr that feeds compare |
| 508 | jmpPos = MII; |
| 509 | jmpInstr = &MI; |
| 510 | predReg = MI.getOperand(i: 0).getReg(); |
| 511 | afterRA = Register::isPhysicalRegister(Reg: predReg); |
| 512 | |
| 513 | // If ifconverter had not messed up with the kill flags of the |
| 514 | // operands, the following check on the kill flag would suffice. |
| 515 | // if(!jmpInstr->getOperand(0).isKill()) break; |
| 516 | |
| 517 | // This predicate register is live out of BB |
| 518 | // this would only work if we can actually use Live |
| 519 | // variable analysis on phy regs - but LLVM does not |
| 520 | // provide LV analysis on phys regs. |
| 521 | //if(LVs.isLiveOut(predReg, *MBB)) break; |
| 522 | |
| 523 | // Get all the successors of this block - which will always |
| 524 | // be 2. Check if the predicate register is live-in in those |
| 525 | // successor. If yes, we can not delete the predicate - |
| 526 | // I am doing this only because LLVM does not provide LiveOut |
| 527 | // at the BB level. |
| 528 | bool predLive = false; |
| 529 | for (const MachineBasicBlock *SuccMBB : MBB->successors()) |
| 530 | if (SuccMBB->isLiveIn(Reg: predReg)) |
| 531 | predLive = true; |
| 532 | if (predLive) |
| 533 | break; |
| 534 | |
| 535 | if (!MI.getOperand(i: 1).isMBB()) |
| 536 | continue; |
| 537 | jmpTarget = MI.getOperand(i: 1).getMBB(); |
| 538 | foundJump = true; |
| 539 | if (MI.getOpcode() == Hexagon::J2_jumpf || |
| 540 | MI.getOpcode() == Hexagon::J2_jumpfnewpt || |
| 541 | MI.getOpcode() == Hexagon::J2_jumpfnew) { |
| 542 | invertPredicate = true; |
| 543 | } |
| 544 | continue; |
| 545 | } |
| 546 | |
| 547 | // No new value jump if there is a barrier. A barrier has to be in its |
| 548 | // own packet. A barrier has zero operands. We conservatively bail out |
| 549 | // here if we see any instruction with zero operands. |
| 550 | if (foundJump && MI.getNumOperands() == 0) |
| 551 | break; |
| 552 | |
| 553 | if (foundJump && !foundCompare && MI.getOperand(i: 0).isReg() && |
| 554 | MI.getOperand(i: 0).getReg() == predReg) { |
| 555 | // Not all compares can be new value compare. Arch Spec: 7.6.1.1 |
| 556 | if (isNewValueJumpCandidate(MI)) { |
| 557 | assert( |
| 558 | (MI.getDesc().isCompare()) && |
| 559 | "Only compare instruction can be collapsed into New Value Jump" ); |
| 560 | isSecondOpReg = MI.getOperand(i: 2).isReg(); |
| 561 | |
| 562 | if (!canCompareBeNewValueJump(QII, TRI: QRI, II: MII, pReg: predReg, secondReg: isSecondOpReg, |
| 563 | optLocation: afterRA, end: jmpPos, MF)) |
| 564 | break; |
| 565 | |
| 566 | cmpInstr = &MI; |
| 567 | cmpPos = MII; |
| 568 | foundCompare = true; |
| 569 | |
| 570 | // We need cmpReg1 and cmpOp2(imm or reg) while building |
| 571 | // new value jump instruction. |
| 572 | cmpReg1 = MI.getOperand(i: 1).getReg(); |
| 573 | |
| 574 | if (isSecondOpReg) |
| 575 | cmpOp2 = MI.getOperand(i: 2).getReg(); |
| 576 | else |
| 577 | cmpOp2 = MI.getOperand(i: 2).getImm(); |
| 578 | continue; |
| 579 | } |
| 580 | } |
| 581 | |
| 582 | if (foundCompare && foundJump) { |
| 583 | // If "common" checks fail, bail out on this BB. |
| 584 | if (!commonChecksToProhibitNewValueJump(afterRA, MII)) |
| 585 | break; |
| 586 | |
| 587 | bool foundFeeder = false; |
| 588 | MachineBasicBlock::iterator feederPos = MII; |
| 589 | if (MI.getOperand(i: 0).isReg() && MI.getOperand(i: 0).isDef() && |
| 590 | (MI.getOperand(i: 0).getReg() == cmpReg1 || |
| 591 | (isSecondOpReg && |
| 592 | MI.getOperand(i: 0).getReg() == (unsigned)cmpOp2))) { |
| 593 | |
| 594 | Register feederReg = MI.getOperand(i: 0).getReg(); |
| 595 | |
| 596 | // First try to see if we can get the feeder from the first operand |
| 597 | // of the compare. If we can not, and if secondOpReg is true |
| 598 | // (second operand of the compare is also register), try that one. |
| 599 | // TODO: Try to come up with some heuristic to figure out which |
| 600 | // feeder would benefit. |
| 601 | |
| 602 | if (feederReg == cmpReg1) { |
| 603 | if (!canBeFeederToNewValueJump(QII, TRI: QRI, II: MII, end: jmpPos, skip: cmpPos, MF)) { |
| 604 | if (!isSecondOpReg) |
| 605 | break; |
| 606 | else |
| 607 | continue; |
| 608 | } else |
| 609 | foundFeeder = true; |
| 610 | } |
| 611 | |
| 612 | if (!foundFeeder && isSecondOpReg && feederReg == (unsigned)cmpOp2) |
| 613 | if (!canBeFeederToNewValueJump(QII, TRI: QRI, II: MII, end: jmpPos, skip: cmpPos, MF)) |
| 614 | break; |
| 615 | |
| 616 | if (isSecondOpReg) { |
| 617 | // In case of CMPLT, or CMPLTU, or EQ with the second register |
| 618 | // to newify, swap the operands. |
| 619 | unsigned COp = cmpInstr->getOpcode(); |
| 620 | if ((COp == Hexagon::C2_cmpeq || COp == Hexagon::C4_cmpneq) && |
| 621 | (feederReg == (unsigned)cmpOp2)) { |
| 622 | unsigned tmp = cmpReg1; |
| 623 | cmpReg1 = cmpOp2; |
| 624 | cmpOp2 = tmp; |
| 625 | } |
| 626 | |
| 627 | // Now we have swapped the operands, all we need to check is, |
| 628 | // if the second operand (after swap) is the feeder. |
| 629 | // And if it is, make a note. |
| 630 | if (feederReg == (unsigned)cmpOp2) |
| 631 | isSecondOpNewified = true; |
| 632 | } |
| 633 | |
| 634 | // Now that we are moving feeder close the jump, |
| 635 | // make sure we are respecting the kill values of |
| 636 | // the operands of the feeder. |
| 637 | |
| 638 | auto TransferKills = [jmpPos,cmpPos] (MachineInstr &MI) { |
| 639 | for (MachineOperand &MO : MI.operands()) { |
| 640 | if (!MO.isReg() || !MO.isUse()) |
| 641 | continue; |
| 642 | Register UseR = MO.getReg(); |
| 643 | for (auto I = std::next(x: MI.getIterator()); I != jmpPos; ++I) { |
| 644 | if (I == cmpPos) |
| 645 | continue; |
| 646 | for (MachineOperand &Op : I->operands()) { |
| 647 | if (!Op.isReg() || !Op.isUse() || !Op.isKill()) |
| 648 | continue; |
| 649 | if (Op.getReg() != UseR) |
| 650 | continue; |
| 651 | // We found that there is kill of a use register |
| 652 | // Set up a kill flag on the register |
| 653 | Op.setIsKill(false); |
| 654 | MO.setIsKill(true); |
| 655 | return; |
| 656 | } |
| 657 | } |
| 658 | } |
| 659 | }; |
| 660 | |
| 661 | TransferKills(*feederPos); |
| 662 | TransferKills(*cmpPos); |
| 663 | bool MO1IsKill = cmpPos->killsRegister(Reg: cmpReg1, TRI: QRI); |
| 664 | bool MO2IsKill = isSecondOpReg && cmpPos->killsRegister(Reg: cmpOp2, TRI: QRI); |
| 665 | |
| 666 | MBB->splice(Where: jmpPos, Other: MI.getParent(), From: MI); |
| 667 | MBB->splice(Where: jmpPos, Other: MI.getParent(), From: cmpInstr); |
| 668 | DebugLoc dl = MI.getDebugLoc(); |
| 669 | MachineInstr *NewMI; |
| 670 | |
| 671 | assert((isNewValueJumpCandidate(*cmpInstr)) && |
| 672 | "This compare is not a New Value Jump candidate." ); |
| 673 | unsigned opc = getNewValueJumpOpcode(MI: cmpInstr, reg: cmpOp2, |
| 674 | secondRegNewified: isSecondOpNewified, |
| 675 | jmpTarget, MBPI); |
| 676 | if (invertPredicate) |
| 677 | opc = QII->getInvertedPredicatedOpcode(Opc: opc); |
| 678 | |
| 679 | if (isSecondOpReg) |
| 680 | NewMI = BuildMI(BB&: *MBB, I: jmpPos, MIMD: dl, MCID: QII->get(Opcode: opc)) |
| 681 | .addReg(RegNo: cmpReg1, flags: getKillRegState(B: MO1IsKill)) |
| 682 | .addReg(RegNo: cmpOp2, flags: getKillRegState(B: MO2IsKill)) |
| 683 | .addMBB(MBB: jmpTarget); |
| 684 | |
| 685 | else |
| 686 | NewMI = BuildMI(BB&: *MBB, I: jmpPos, MIMD: dl, MCID: QII->get(Opcode: opc)) |
| 687 | .addReg(RegNo: cmpReg1, flags: getKillRegState(B: MO1IsKill)) |
| 688 | .addImm(Val: cmpOp2) |
| 689 | .addMBB(MBB: jmpTarget); |
| 690 | |
| 691 | assert(NewMI && "New Value Jump Instruction Not created!" ); |
| 692 | (void)NewMI; |
| 693 | if (cmpInstr->getOperand(i: 0).isReg() && |
| 694 | cmpInstr->getOperand(i: 0).isKill()) |
| 695 | cmpInstr->getOperand(i: 0).setIsKill(false); |
| 696 | if (cmpInstr->getOperand(i: 1).isReg() && |
| 697 | cmpInstr->getOperand(i: 1).isKill()) |
| 698 | cmpInstr->getOperand(i: 1).setIsKill(false); |
| 699 | cmpInstr->eraseFromParent(); |
| 700 | jmpInstr->eraseFromParent(); |
| 701 | ++nvjGenerated; |
| 702 | ++NumNVJGenerated; |
| 703 | break; |
| 704 | } |
| 705 | } |
| 706 | } |
| 707 | } |
| 708 | |
| 709 | return true; |
| 710 | } |
| 711 | |
| 712 | FunctionPass *llvm::createHexagonNewValueJump() { |
| 713 | return new HexagonNewValueJump(); |
| 714 | } |
| 715 | |