| 1 | //===-- AMDGPURewriteAGPRCopyMFMA.cpp -------------------------------------===// |
| 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 | /// \file \brief Try to replace MFMA instructions using VGPRs with MFMA |
| 10 | /// instructions using AGPRs. We expect MFMAs to be selected using VGPRs, and |
| 11 | /// only use AGPRs if it helps avoid spilling. In this case, the MFMA will have |
| 12 | /// copies between AGPRs and VGPRs and the AGPR variant of an MFMA pseudo. This |
| 13 | /// pass will attempt to delete the cross register bank copy and replace the |
| 14 | /// MFMA opcode. |
| 15 | /// |
| 16 | /// TODO: |
| 17 | /// - Handle rewrites of phis. This must be more careful than normal about the |
| 18 | /// reassignment. We do not want to introduce an AGPR-to-AGPR copy inside of a |
| 19 | /// loop, so it depends on the exact assignment of the copy. |
| 20 | /// |
| 21 | /// - Update LiveIntervals incrementally instead of recomputing from scratch |
| 22 | /// |
| 23 | //===----------------------------------------------------------------------===// |
| 24 | |
| 25 | #include "AMDGPU.h" |
| 26 | #include "GCNSubtarget.h" |
| 27 | #include "SIMachineFunctionInfo.h" |
| 28 | #include "SIRegisterInfo.h" |
| 29 | #include "llvm/ADT/Statistic.h" |
| 30 | #include "llvm/CodeGen/LiveIntervals.h" |
| 31 | #include "llvm/CodeGen/LiveRegMatrix.h" |
| 32 | #include "llvm/CodeGen/LiveStacks.h" |
| 33 | #include "llvm/CodeGen/MachineDominators.h" |
| 34 | #include "llvm/CodeGen/MachineFrameInfo.h" |
| 35 | #include "llvm/CodeGen/MachineFunctionPass.h" |
| 36 | #include "llvm/CodeGen/SlotIndexes.h" |
| 37 | #include "llvm/CodeGen/VirtRegMap.h" |
| 38 | #include "llvm/InitializePasses.h" |
| 39 | #include "llvm/Support/DebugCounter.h" |
| 40 | |
| 41 | using namespace llvm; |
| 42 | |
| 43 | #define DEBUG_TYPE "amdgpu-rewrite-agpr-copy-mfma" |
| 44 | |
| 45 | DEBUG_COUNTER(RewriteAGPRCopyMFMACounter, DEBUG_TYPE, |
| 46 | "Controls which MFMA chains are rewritten to AGPR form" ); |
| 47 | |
| 48 | namespace { |
| 49 | |
| 50 | STATISTIC(NumMFMAsRewrittenToAGPR, |
| 51 | "Number of MFMA instructions rewritten to use AGPR form" ); |
| 52 | |
| 53 | /// Map from spill slot frame index to list of instructions which reference it. |
| 54 | using SpillReferenceMap = DenseMap<int, SmallVector<MachineInstr *, 4>>; |
| 55 | |
| 56 | class AMDGPURewriteAGPRCopyMFMAImpl { |
| 57 | MachineFunction &MF; |
| 58 | const GCNSubtarget &ST; |
| 59 | const SIInstrInfo &TII; |
| 60 | const SIRegisterInfo &TRI; |
| 61 | MachineRegisterInfo &MRI; |
| 62 | VirtRegMap &VRM; |
| 63 | LiveRegMatrix &LRM; |
| 64 | LiveIntervals &LIS; |
| 65 | LiveStacks &LSS; |
| 66 | const RegisterClassInfo &RegClassInfo; |
| 67 | MachineDominatorTree &MDT; |
| 68 | |
| 69 | bool attemptReassignmentsToAGPR(SmallSetVector<Register, 4> &InterferingRegs, |
| 70 | MCPhysReg PrefPhysReg) const; |
| 71 | |
| 72 | public: |
| 73 | AMDGPURewriteAGPRCopyMFMAImpl(MachineFunction &MF, VirtRegMap &VRM, |
| 74 | LiveRegMatrix &LRM, LiveIntervals &LIS, |
| 75 | LiveStacks &LSS, |
| 76 | const RegisterClassInfo &RegClassInfo, |
| 77 | MachineDominatorTree &MDT) |
| 78 | : MF(MF), ST(MF.getSubtarget<GCNSubtarget>()), TII(*ST.getInstrInfo()), |
| 79 | TRI(*ST.getRegisterInfo()), MRI(MF.getRegInfo()), VRM(VRM), LRM(LRM), |
| 80 | LIS(LIS), LSS(LSS), RegClassInfo(RegClassInfo), MDT(MDT) {} |
| 81 | |
| 82 | bool isRewriteCandidate(const MachineInstr &MI) const { |
| 83 | return TII.isMAI(MI) && AMDGPU::getAGPRFormOp(Opcode: MI.getOpcode()) != -1; |
| 84 | } |
| 85 | |
| 86 | /// Find AV_* registers assigned to AGPRs (or virtual registers which were |
| 87 | /// already required to be AGPR). |
| 88 | /// |
| 89 | /// \return the assigned physical register that \p VReg is assigned to if it |
| 90 | /// is an AGPR, otherwise MCRegister(). |
| 91 | MCRegister getAssignedAGPR(Register VReg) const { |
| 92 | MCRegister PhysReg = VRM.getPhys(virtReg: VReg); |
| 93 | if (!PhysReg) |
| 94 | return MCRegister(); |
| 95 | |
| 96 | // If this is an AV register, we have to check if the actual assignment is |
| 97 | // to an AGPR |
| 98 | const TargetRegisterClass *AssignedRC = TRI.getPhysRegBaseClass(Reg: PhysReg); |
| 99 | return TRI.isAGPRClass(RC: AssignedRC) ? PhysReg : MCRegister(); |
| 100 | } |
| 101 | |
| 102 | bool tryReassigningMFMAChain(MachineInstr &MFMA, Register MFMAHintReg, |
| 103 | MCPhysReg PhysRegHint) const; |
| 104 | |
| 105 | /// Compute the register class constraints based on the uses of \p Reg, |
| 106 | /// excluding MFMA uses from which can be rewritten to change the register |
| 107 | /// class constraint. MFMA scale operands need to be constraint checked. |
| 108 | /// This should be nearly identical to MachineRegisterInfo::recomputeRegClass. |
| 109 | |
| 110 | /// \p RewriteCandidates will collect the set of MFMA instructions that need |
| 111 | /// to have the opcode mutated to perform the replacement. |
| 112 | /// |
| 113 | /// \p RewriteRegs will accumulate the set of register used by those MFMAs |
| 114 | /// that need to have the register classes adjusted. |
| 115 | bool recomputeRegClassExceptRewritable( |
| 116 | Register Reg, SmallVectorImpl<MachineInstr *> &RewriteCandidates, |
| 117 | SmallSetVector<Register, 4> &RewriteRegs) const; |
| 118 | |
| 119 | bool tryFoldCopiesToAGPR(Register VReg, MCRegister AssignedAGPR) const; |
| 120 | bool tryFoldCopiesFromAGPR(Register VReg, MCRegister AssignedAGPR) const; |
| 121 | |
| 122 | /// Replace spill instruction \p SpillMI which loads/stores from/to \p SpillFI |
| 123 | /// with a COPY to the replacement register value \p VReg. |
| 124 | void replaceSpillWithCopyToVReg(MachineInstr &SpillMI, int SpillFI, |
| 125 | Register VReg) const; |
| 126 | |
| 127 | /// Create a map from frame index to use instructions for spills. If a use of |
| 128 | /// the frame index does not consist only of spill instructions, it will not |
| 129 | /// be included in the map. |
| 130 | void collectSpillIndexUses(ArrayRef<LiveInterval *> StackIntervals, |
| 131 | SpillReferenceMap &Map) const; |
| 132 | |
| 133 | /// Return true if the reload \p LoadMI of the stack slot with live interval |
| 134 | /// \p SlotLI is jointly dominated by the slot's spill stores, i.e. every path |
| 135 | /// from the entry block to the load passes through a store to the slot before |
| 136 | /// the load. \p StoreFreeReachable is the set of blocks reachable from the |
| 137 | /// entry block without passing through any store block for the slot. |
| 138 | bool isLoadJointlyDominatedByStores( |
| 139 | const MachineInstr &LoadMI, const LiveInterval &SlotLI, |
| 140 | const SmallPtrSetImpl<MachineBasicBlock *> &StoreFreeReachable) const; |
| 141 | |
| 142 | /// Attempt to unspill VGPRs by finding a free register and replacing the |
| 143 | /// spill instructions with copies. |
| 144 | void eliminateSpillsOfReassignedVGPRs() const; |
| 145 | |
| 146 | bool run(MachineFunction &MF) const; |
| 147 | }; |
| 148 | |
| 149 | bool AMDGPURewriteAGPRCopyMFMAImpl::recomputeRegClassExceptRewritable( |
| 150 | Register StartReg, SmallVectorImpl<MachineInstr *> &RewriteCandidates, |
| 151 | SmallSetVector<Register, 4> &RewriteRegs) const { |
| 152 | SmallVector<Register, 8> Worklist = {StartReg}; |
| 153 | |
| 154 | // Recursively visit all transitive MFMA users |
| 155 | while (!Worklist.empty()) { |
| 156 | Register Reg = Worklist.pop_back_val(); |
| 157 | const TargetRegisterClass *OldRC = MRI.getRegClass(Reg); |
| 158 | |
| 159 | // Inflate to the equivalent AV_* class. |
| 160 | const TargetRegisterClass *NewRC = TRI.getLargestLegalSuperClass(RC: OldRC, MF); |
| 161 | if (OldRC == NewRC) |
| 162 | return false; |
| 163 | |
| 164 | // Accumulate constraints from all uses. |
| 165 | for (MachineOperand &MO : MRI.reg_nodbg_operands(Reg)) { |
| 166 | // Apply the effect of the given operand to NewRC. |
| 167 | MachineInstr *MI = MO.getParent(); |
| 168 | |
| 169 | // We can swap the classes of dst + src2 as a pair to AGPR, so ignore the |
| 170 | // effects of rewrite candidates. It just so happens that we can use |
| 171 | // either AGPR or VGPR in src0/src1. We still need to check constraint |
| 172 | // effects for scale variant, which does not allow AGPR. |
| 173 | if (isRewriteCandidate(MI: *MI)) { |
| 174 | int AGPROp = AMDGPU::getAGPRFormOp(Opcode: MI->getOpcode()); |
| 175 | const MCInstrDesc &AGPRDesc = TII.get(Opcode: AGPROp); |
| 176 | const TargetRegisterClass *NewRC = |
| 177 | TII.getRegClass(MCID: AGPRDesc, OpNum: MO.getOperandNo()); |
| 178 | if (!TRI.hasAGPRs(RC: NewRC)) |
| 179 | return false; |
| 180 | |
| 181 | const MachineOperand *VDst = |
| 182 | TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::vdst); |
| 183 | const MachineOperand *Src2 = |
| 184 | TII.getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src2); |
| 185 | for (const MachineOperand *Op : {VDst, Src2}) { |
| 186 | if (!Op->isReg()) |
| 187 | continue; |
| 188 | |
| 189 | Register OtherReg = Op->getReg(); |
| 190 | if (OtherReg.isPhysical()) |
| 191 | return false; |
| 192 | |
| 193 | if (OtherReg != Reg && RewriteRegs.insert(X: OtherReg)) |
| 194 | Worklist.push_back(Elt: OtherReg); |
| 195 | } |
| 196 | |
| 197 | if (!is_contained(Range&: RewriteCandidates, Element: MI)) { |
| 198 | LLVM_DEBUG({ |
| 199 | Register VDstPhysReg = VRM.getPhys(VDst->getReg()); |
| 200 | dbgs() << "Attempting to replace VGPR MFMA with AGPR version:" |
| 201 | << " Dst=[" << printReg(VDst->getReg()) << " => " |
| 202 | << printReg(VDstPhysReg, &TRI); |
| 203 | |
| 204 | if (Src2->isReg()) { |
| 205 | Register Src2PhysReg = VRM.getPhys(Src2->getReg()); |
| 206 | dbgs() << "], Src2=[" << printReg(Src2->getReg(), &TRI) << " => " |
| 207 | << printReg(Src2PhysReg, &TRI); |
| 208 | } |
| 209 | |
| 210 | dbgs() << "]: " << MI; |
| 211 | }); |
| 212 | |
| 213 | RewriteCandidates.push_back(Elt: MI); |
| 214 | } |
| 215 | |
| 216 | continue; |
| 217 | } |
| 218 | |
| 219 | unsigned OpNo = &MO - &MI->getOperand(i: 0); |
| 220 | NewRC = MI->getRegClassConstraintEffect(OpIdx: OpNo, CurRC: NewRC, TII: &TII, TRI: &TRI); |
| 221 | if (!NewRC || NewRC == OldRC) { |
| 222 | LLVM_DEBUG(dbgs() << "User of " << printReg(Reg, &TRI) |
| 223 | << " cannot be reassigned to " |
| 224 | << (NewRC ? TRI.getRegClassName(NewRC) : "NULL" ) |
| 225 | << ": " << *MI); |
| 226 | return false; |
| 227 | } |
| 228 | } |
| 229 | } |
| 230 | |
| 231 | return true; |
| 232 | } |
| 233 | |
| 234 | bool AMDGPURewriteAGPRCopyMFMAImpl::tryReassigningMFMAChain( |
| 235 | MachineInstr &MFMA, Register MFMAHintReg, MCPhysReg PhysRegHint) const { |
| 236 | // src2 and dst have the same physical class constraint; try to preserve |
| 237 | // the original src2 subclass if one were to exist. |
| 238 | SmallVector<MachineInstr *, 4> RewriteCandidates = {&MFMA}; |
| 239 | SmallSetVector<Register, 4> RewriteRegs; |
| 240 | |
| 241 | // Make sure we reassign the MFMA we found the copy from first. We want |
| 242 | // to ensure dst ends up in the physreg we were originally copying to. |
| 243 | RewriteRegs.insert(X: MFMAHintReg); |
| 244 | |
| 245 | // We've found av = COPY (MFMA) (or MFMA (v = COPY av)) and need to verify |
| 246 | // that we can trivially rewrite src2 to use the new AGPR. If we can't |
| 247 | // trivially replace it, we're going to induce as many copies as we would have |
| 248 | // emitted in the first place, as well as need to assign another register, and |
| 249 | // need to figure out where to put them. The live range splitting is smarter |
| 250 | // than anything we're doing here, so trust it did something reasonable. |
| 251 | // |
| 252 | // Note recomputeRegClassExceptRewritable will consider the constraints of |
| 253 | // this MFMA's src2 as well as the src2/dst of any transitive MFMA users. |
| 254 | if (!recomputeRegClassExceptRewritable(StartReg: MFMAHintReg, RewriteCandidates, |
| 255 | RewriteRegs)) { |
| 256 | LLVM_DEBUG(dbgs() << "Could not recompute the regclass of dst reg " |
| 257 | << printReg(MFMAHintReg, &TRI) << '\n'); |
| 258 | return false; |
| 259 | } |
| 260 | |
| 261 | // If src2 and dst are different registers, we need to also reassign the |
| 262 | // input to an available AGPR if it is compatible with all other uses. |
| 263 | // |
| 264 | // If we can't reassign it, we'd need to introduce a different copy |
| 265 | // which is likely worse than the copy we'd be saving. |
| 266 | // |
| 267 | // It's likely that the MFMA is used in sequence with other MFMAs; if we |
| 268 | // cannot migrate the full use/def chain of MFMAs, we would need to |
| 269 | // introduce intermediate copies somewhere. So we only make the |
| 270 | // transform if all the interfering MFMAs can also be migrated. Collect |
| 271 | // the set of rewritable MFMAs and check if we can assign an AGPR at |
| 272 | // that point. |
| 273 | // |
| 274 | // If any of the MFMAs aren't reassignable, we give up and rollback to |
| 275 | // the original register assignments. |
| 276 | |
| 277 | using RecoloringStack = |
| 278 | SmallVector<std::pair<const LiveInterval *, MCRegister>, 8>; |
| 279 | RecoloringStack TentativeReassignments; |
| 280 | |
| 281 | for (Register RewriteReg : RewriteRegs) { |
| 282 | LiveInterval &LI = LIS.getInterval(Reg: RewriteReg); |
| 283 | TentativeReassignments.push_back(Elt: {&LI, VRM.getPhys(virtReg: RewriteReg)}); |
| 284 | LRM.unassign(VirtReg: LI); |
| 285 | } |
| 286 | |
| 287 | if (!DebugCounter::shouldExecute(Counter&: RewriteAGPRCopyMFMACounter) || |
| 288 | !attemptReassignmentsToAGPR(InterferingRegs&: RewriteRegs, PrefPhysReg: PhysRegHint)) { |
| 289 | // Roll back the register assignments to the original state. |
| 290 | for (auto [LI, OldAssign] : TentativeReassignments) { |
| 291 | if (VRM.hasPhys(virtReg: LI->reg())) |
| 292 | LRM.unassign(VirtReg: *LI); |
| 293 | LRM.assign(VirtReg: *LI, PhysReg: OldAssign); |
| 294 | } |
| 295 | |
| 296 | return false; |
| 297 | } |
| 298 | |
| 299 | // Fixup the register classes of the virtual registers now that we've |
| 300 | // committed to the reassignments. |
| 301 | for (Register InterferingReg : RewriteRegs) { |
| 302 | const TargetRegisterClass *EquivalentAGPRRegClass = |
| 303 | TRI.getEquivalentAGPRClass(SRC: MRI.getRegClass(Reg: InterferingReg)); |
| 304 | MRI.setRegClass(Reg: InterferingReg, RC: EquivalentAGPRRegClass); |
| 305 | } |
| 306 | |
| 307 | for (MachineInstr *RewriteCandidate : RewriteCandidates) { |
| 308 | int NewMFMAOp = AMDGPU::getAGPRFormOp(Opcode: RewriteCandidate->getOpcode()); |
| 309 | RewriteCandidate->setDesc(TII.get(Opcode: NewMFMAOp)); |
| 310 | ++NumMFMAsRewrittenToAGPR; |
| 311 | } |
| 312 | |
| 313 | return true; |
| 314 | } |
| 315 | |
| 316 | /// Attempt to reassign the registers in \p InterferingRegs to be AGPRs, with a |
| 317 | /// preference to use \p PhysReg first. Returns false if the reassignments |
| 318 | /// cannot be trivially performed. |
| 319 | bool AMDGPURewriteAGPRCopyMFMAImpl::attemptReassignmentsToAGPR( |
| 320 | SmallSetVector<Register, 4> &InterferingRegs, MCPhysReg PrefPhysReg) const { |
| 321 | // FIXME: The ordering may matter here, but we're just taking uselistorder |
| 322 | // with the special case of ensuring to process the starting instruction |
| 323 | // first. We probably should extract the priority advisor out of greedy and |
| 324 | // use that ordering. |
| 325 | for (Register InterferingReg : InterferingRegs) { |
| 326 | LiveInterval &ReassignLI = LIS.getInterval(Reg: InterferingReg); |
| 327 | const TargetRegisterClass *EquivalentAGPRRegClass = |
| 328 | TRI.getEquivalentAGPRClass(SRC: MRI.getRegClass(Reg: InterferingReg)); |
| 329 | |
| 330 | MCPhysReg Assignable = AMDGPU::NoRegister; |
| 331 | if (EquivalentAGPRRegClass->contains(Reg: PrefPhysReg) && |
| 332 | LRM.checkInterference(VirtReg: ReassignLI, PhysReg: PrefPhysReg) == |
| 333 | LiveRegMatrix::IK_Free) { |
| 334 | // First try to assign to the AGPR we were already copying to. This |
| 335 | // should be the first assignment we attempt. We have to guard |
| 336 | // against the use being a subregister (which doesn't have an exact |
| 337 | // class match). |
| 338 | |
| 339 | // TODO: If this does happen to be a subregister use, we should |
| 340 | // still try to assign to a subregister of the original copy result. |
| 341 | Assignable = PrefPhysReg; |
| 342 | } else { |
| 343 | ArrayRef<MCPhysReg> AllocOrder = |
| 344 | RegClassInfo.getOrder(RC: EquivalentAGPRRegClass); |
| 345 | for (MCPhysReg Reg : AllocOrder) { |
| 346 | if (LRM.checkInterference(VirtReg: ReassignLI, PhysReg: Reg) == LiveRegMatrix::IK_Free) { |
| 347 | Assignable = Reg; |
| 348 | break; |
| 349 | } |
| 350 | } |
| 351 | } |
| 352 | |
| 353 | if (!Assignable) { |
| 354 | LLVM_DEBUG(dbgs() << "Unable to reassign VGPR " |
| 355 | << printReg(InterferingReg, &TRI) |
| 356 | << " to a free AGPR\n" ); |
| 357 | return false; |
| 358 | } |
| 359 | |
| 360 | LLVM_DEBUG(dbgs() << "Reassigning VGPR " << printReg(InterferingReg, &TRI) |
| 361 | << " to " << printReg(Assignable, &TRI) << '\n'); |
| 362 | LRM.assign(VirtReg: ReassignLI, PhysReg: Assignable); |
| 363 | } |
| 364 | |
| 365 | return true; |
| 366 | } |
| 367 | |
| 368 | /// Identify copies that look like: |
| 369 | /// %vdst:vgpr = V_MFMA_.. %src0:av, %src1:av, %src2:vgpr |
| 370 | /// %agpr = COPY %vgpr |
| 371 | /// |
| 372 | /// Then try to replace the transitive uses of %src2 and %vdst with the AGPR |
| 373 | /// versions of the MFMA. This should cover the common case. |
| 374 | bool AMDGPURewriteAGPRCopyMFMAImpl::tryFoldCopiesToAGPR( |
| 375 | Register VReg, MCRegister AssignedAGPR) const { |
| 376 | bool MadeChange = false; |
| 377 | for (MachineInstr &UseMI : MRI.def_instructions(Reg: VReg)) { |
| 378 | if (!UseMI.isCopy()) |
| 379 | continue; |
| 380 | |
| 381 | Register CopySrcReg = UseMI.getOperand(i: 1).getReg(); |
| 382 | if (!CopySrcReg.isVirtual()) |
| 383 | continue; |
| 384 | |
| 385 | // TODO: Handle loop phis copied to AGPR. e.g. |
| 386 | // |
| 387 | // loop: |
| 388 | // %phi:vgpr = COPY %mfma:vgpr |
| 389 | // %mfma:vgpr = V_MFMA_xxx_vgprcd_e64 %a, %b, %phi |
| 390 | // s_cbranch_vccnz loop |
| 391 | // |
| 392 | // endloop: |
| 393 | // %agpr = mfma |
| 394 | // |
| 395 | // We need to be sure that %phi is assigned to the same physical register as |
| 396 | // %mfma, or else we will just be moving copies into the loop. |
| 397 | |
| 398 | for (MachineInstr &CopySrcDefMI : MRI.def_instructions(Reg: CopySrcReg)) { |
| 399 | if (isRewriteCandidate(MI: CopySrcDefMI) && |
| 400 | tryReassigningMFMAChain( |
| 401 | MFMA&: CopySrcDefMI, MFMAHintReg: CopySrcDefMI.getOperand(i: 0).getReg(), PhysRegHint: AssignedAGPR)) |
| 402 | MadeChange = true; |
| 403 | } |
| 404 | } |
| 405 | |
| 406 | return MadeChange; |
| 407 | } |
| 408 | |
| 409 | /// Identify copies that look like: |
| 410 | /// %src:vgpr = COPY %src:agpr |
| 411 | /// %vdst:vgpr = V_MFMA_... %src0:av, %src1:av, %src:vgpr |
| 412 | /// |
| 413 | /// Then try to replace the transitive uses of %src2 and %vdst with the AGPR |
| 414 | /// versions of the MFMA. This should cover rarer cases, and will generally be |
| 415 | /// redundant with tryFoldCopiesToAGPR. |
| 416 | bool AMDGPURewriteAGPRCopyMFMAImpl::tryFoldCopiesFromAGPR( |
| 417 | Register VReg, MCRegister AssignedAGPR) const { |
| 418 | bool MadeChange = false; |
| 419 | for (MachineInstr &UseMI : MRI.use_instructions(Reg: VReg)) { |
| 420 | if (!UseMI.isCopy()) |
| 421 | continue; |
| 422 | |
| 423 | Register CopyDstReg = UseMI.getOperand(i: 0).getReg(); |
| 424 | if (!CopyDstReg.isVirtual()) |
| 425 | continue; |
| 426 | for (MachineOperand &CopyUseMO : MRI.reg_nodbg_operands(Reg: CopyDstReg)) { |
| 427 | if (!CopyUseMO.readsReg()) |
| 428 | continue; |
| 429 | |
| 430 | MachineInstr &CopyUseMI = *CopyUseMO.getParent(); |
| 431 | if (isRewriteCandidate(MI: CopyUseMI)) { |
| 432 | if (tryReassigningMFMAChain(MFMA&: CopyUseMI, MFMAHintReg: CopyDstReg, |
| 433 | PhysRegHint: VRM.getPhys(virtReg: CopyDstReg))) |
| 434 | MadeChange = true; |
| 435 | } |
| 436 | } |
| 437 | } |
| 438 | |
| 439 | return MadeChange; |
| 440 | } |
| 441 | |
| 442 | void AMDGPURewriteAGPRCopyMFMAImpl::replaceSpillWithCopyToVReg( |
| 443 | MachineInstr &SpillMI, int SpillFI, Register VReg) const { |
| 444 | const DebugLoc &DL = SpillMI.getDebugLoc(); |
| 445 | MachineBasicBlock &MBB = *SpillMI.getParent(); |
| 446 | MachineInstr *NewCopy; |
| 447 | if (SpillMI.mayStore()) { |
| 448 | NewCopy = BuildMI(BB&: MBB, I&: SpillMI, MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::COPY), DestReg: VReg) |
| 449 | .add(MO: SpillMI.getOperand(i: 0)); |
| 450 | } else { |
| 451 | NewCopy = BuildMI(BB&: MBB, I&: SpillMI, MIMD: DL, MCID: TII.get(Opcode: TargetOpcode::COPY)) |
| 452 | .add(MO: SpillMI.getOperand(i: 0)) |
| 453 | .addReg(RegNo: VReg); |
| 454 | } |
| 455 | |
| 456 | LIS.ReplaceMachineInstrInMaps(MI&: SpillMI, NewMI&: *NewCopy); |
| 457 | SpillMI.eraseFromParent(); |
| 458 | } |
| 459 | |
| 460 | void AMDGPURewriteAGPRCopyMFMAImpl::collectSpillIndexUses( |
| 461 | ArrayRef<LiveInterval *> StackIntervals, SpillReferenceMap &Map) const { |
| 462 | |
| 463 | SmallSet<int, 4> NeededFrameIndexes; |
| 464 | for (const LiveInterval *LI : StackIntervals) |
| 465 | NeededFrameIndexes.insert(V: LI->reg().stackSlotIndex()); |
| 466 | |
| 467 | for (MachineBasicBlock &MBB : MF) { |
| 468 | for (MachineInstr &MI : MBB) { |
| 469 | for (MachineOperand &MO : MI.operands()) { |
| 470 | if (!MO.isFI() || !NeededFrameIndexes.count(V: MO.getIndex())) |
| 471 | continue; |
| 472 | |
| 473 | if (TII.isVGPRSpill(MI)) { |
| 474 | SmallVector<MachineInstr *, 4> &References = Map[MO.getIndex()]; |
| 475 | References.push_back(Elt: &MI); |
| 476 | break; |
| 477 | } |
| 478 | |
| 479 | // Verify this was really a spill instruction, if it's not just ignore |
| 480 | // all uses. |
| 481 | |
| 482 | // TODO: This should probably be verifier enforced. |
| 483 | NeededFrameIndexes.erase(V: MO.getIndex()); |
| 484 | Map.erase(Val: MO.getIndex()); |
| 485 | } |
| 486 | } |
| 487 | } |
| 488 | } |
| 489 | |
| 490 | bool AMDGPURewriteAGPRCopyMFMAImpl::isLoadJointlyDominatedByStores( |
| 491 | const MachineInstr &LoadMI, const LiveInterval &SlotLI, |
| 492 | const SmallPtrSetImpl<MachineBasicBlock *> &StoreFreeReachable) const { |
| 493 | const MachineBasicBlock *LoadMBB = LoadMI.getParent(); |
| 494 | if (!MDT.isReachableFromEntry(A: LoadMBB)) |
| 495 | return true; |
| 496 | |
| 497 | // Check if every path passed through a store block. |
| 498 | if (!StoreFreeReachable.contains(Ptr: LoadMBB)) |
| 499 | return true; |
| 500 | |
| 501 | // Otherwise, there exists a path to this block that has not seen any store |
| 502 | // yet. We must ensure that within this block there is a store to this slot |
| 503 | // before the load. Consult the slot's LiveStacks interval: a store to the |
| 504 | // slot before the load means the slot is not live into this block but is |
| 505 | // live at the load. If the load reads an undef value, the slot is not live |
| 506 | // at the load, failing the joint-dominance check. |
| 507 | SlotIndex LoadIdx = LIS.getInstructionIndex(Instr: LoadMI); |
| 508 | return SlotLI.liveAt(index: LoadIdx) && !LIS.isLiveInToMBB(LR: SlotLI, mbb: LoadMBB); |
| 509 | } |
| 510 | |
| 511 | void AMDGPURewriteAGPRCopyMFMAImpl::eliminateSpillsOfReassignedVGPRs() const { |
| 512 | unsigned NumSlots = LSS.getNumIntervals(); |
| 513 | if (NumSlots == 0) |
| 514 | return; |
| 515 | |
| 516 | MachineFrameInfo &MFI = MF.getFrameInfo(); |
| 517 | |
| 518 | SmallVector<LiveInterval *, 32> StackIntervals; |
| 519 | StackIntervals.reserve(N: NumSlots); |
| 520 | |
| 521 | for (auto &[Slot, LI] : LSS) { |
| 522 | if (!MFI.isSpillSlotObjectIndex(ObjectIdx: Slot) || MFI.isDeadObjectIndex(ObjectIdx: Slot)) |
| 523 | continue; |
| 524 | |
| 525 | const TargetRegisterClass *RC = LSS.getIntervalRegClass(Slot); |
| 526 | if (TRI.hasVGPRs(RC)) |
| 527 | StackIntervals.push_back(Elt: &LI); |
| 528 | } |
| 529 | |
| 530 | sort(C&: StackIntervals, Comp: [](const LiveInterval *A, const LiveInterval *B) { |
| 531 | // The ordering has to be strictly weak. |
| 532 | /// Sort heaviest intervals first to prioritize their unspilling |
| 533 | if (A->weight() != B->weight()) |
| 534 | return A->weight() > B->weight(); |
| 535 | |
| 536 | if (A->getSize() != B->getSize()) |
| 537 | return A->getSize() > B->getSize(); |
| 538 | |
| 539 | // Tie breaker by number to avoid need for stable sort |
| 540 | return A->reg().stackSlotIndex() < B->reg().stackSlotIndex(); |
| 541 | }); |
| 542 | |
| 543 | // FIXME: The APIs for dealing with the LiveInterval of a frame index are |
| 544 | // cumbersome. LiveStacks owns its LiveIntervals which refer to stack |
| 545 | // slots. We cannot use the usual LiveRegMatrix::assign and unassign on these, |
| 546 | // and must create a substitute virtual register to do so. This makes |
| 547 | // incremental updating here difficult; we need to actually perform the IR |
| 548 | // mutation to get the new vreg references in place to compute the register |
| 549 | // LiveInterval to perform an assignment to track the new interference |
| 550 | // correctly, and we can't simply migrate the LiveInterval we already have. |
| 551 | // |
| 552 | // To avoid walking through the entire function for each index, pre-collect |
| 553 | // all the instructions slot referencess. |
| 554 | |
| 555 | DenseMap<int, SmallVector<MachineInstr *, 4>> SpillSlotReferences; |
| 556 | collectSpillIndexUses(StackIntervals, Map&: SpillSlotReferences); |
| 557 | |
| 558 | for (LiveInterval *LI : StackIntervals) { |
| 559 | int Slot = LI->reg().stackSlotIndex(); |
| 560 | auto SpillReferences = SpillSlotReferences.find(Val: Slot); |
| 561 | if (SpillReferences == SpillSlotReferences.end()) |
| 562 | continue; |
| 563 | |
| 564 | // For each spill reload, every path from entry to the reload must pass |
| 565 | // through at least one spill store to the same stack slot. |
| 566 | SmallPtrSet<MachineBasicBlock *, 4> StoreBlocks; |
| 567 | for (MachineInstr *MI : SpillReferences->second) { |
| 568 | if (MI->mayStore() && MDT.isReachableFromEntry(A: MI->getParent())) |
| 569 | StoreBlocks.insert(Ptr: MI->getParent()); |
| 570 | } |
| 571 | |
| 572 | if (StoreBlocks.empty()) { |
| 573 | LLVM_DEBUG(dbgs() << "Skipping " << printReg(Slot, &TRI) |
| 574 | << ": no reachable stores\n" ); |
| 575 | continue; |
| 576 | } |
| 577 | |
| 578 | // Compute blocks reachable from entry without passing through a store |
| 579 | // block. |
| 580 | MachineBasicBlock &EntryMBB = MF.front(); |
| 581 | SmallPtrSet<MachineBasicBlock *, 16> StoreFreeReachable = {&EntryMBB}; |
| 582 | SmallVector<MachineBasicBlock *, 16> Worklist = {&EntryMBB}; |
| 583 | |
| 584 | while (!Worklist.empty()) { |
| 585 | MachineBasicBlock *MBB = Worklist.pop_back_val(); |
| 586 | if (StoreBlocks.contains(Ptr: MBB)) |
| 587 | continue; |
| 588 | |
| 589 | for (MachineBasicBlock *Succ : MBB->successors()) { |
| 590 | if (StoreFreeReachable.insert(Ptr: Succ).second) |
| 591 | Worklist.push_back(Elt: Succ); |
| 592 | } |
| 593 | } |
| 594 | |
| 595 | // Every reachable reload must be jointly dominated by the slot's stores. |
| 596 | if (!llvm::all_of(Range&: SpillReferences->second, P: [&](const MachineInstr *MI) { |
| 597 | return !MI->mayLoad() || |
| 598 | isLoadJointlyDominatedByStores(LoadMI: *MI, SlotLI: *LI, StoreFreeReachable); |
| 599 | })) { |
| 600 | LLVM_DEBUG( |
| 601 | dbgs() << "Skipping " << printReg(Slot, &TRI) |
| 602 | << ": some reachable load not jointly dominated by stores\n" ); |
| 603 | continue; |
| 604 | } |
| 605 | |
| 606 | const TargetRegisterClass *RC = LSS.getIntervalRegClass(Slot); |
| 607 | |
| 608 | LLVM_DEBUG(dbgs() << "Trying to eliminate " << printReg(Slot, &TRI) |
| 609 | << " by reassigning\n" ); |
| 610 | |
| 611 | ArrayRef<MCPhysReg> AllocOrder = RegClassInfo.getOrder(RC); |
| 612 | |
| 613 | for (MCPhysReg PhysReg : AllocOrder) { |
| 614 | if (LRM.checkInterference(VirtReg: *LI, PhysReg) != LiveRegMatrix::IK_Free) |
| 615 | continue; |
| 616 | |
| 617 | LLVM_DEBUG(dbgs() << "Reassigning " << *LI << " to " |
| 618 | << printReg(PhysReg, &TRI) << '\n'); |
| 619 | |
| 620 | const TargetRegisterClass *RC = LSS.getIntervalRegClass(Slot); |
| 621 | Register NewVReg = MRI.createVirtualRegister(RegClass: RC); |
| 622 | |
| 623 | for (MachineInstr *SpillMI : SpillReferences->second) |
| 624 | replaceSpillWithCopyToVReg(SpillMI&: *SpillMI, SpillFI: Slot, VReg: NewVReg); |
| 625 | |
| 626 | // TODO: We should be able to transfer the information from the stack |
| 627 | // slot's LiveInterval without recomputing from scratch with the |
| 628 | // replacement vreg uses. |
| 629 | LiveInterval &NewLI = LIS.createAndComputeVirtRegInterval(Reg: NewVReg); |
| 630 | VRM.grow(); |
| 631 | |
| 632 | // A spill slot can be stored to multiple times, so the replacement |
| 633 | // vreg may have multiple disconnected live range components. Split |
| 634 | // them into separate vregs to maintain the single-component invariant. |
| 635 | SmallVector<LiveInterval *, 4> SplitLIs; |
| 636 | LIS.splitSeparateComponents(LI&: NewLI, SplitLIs); |
| 637 | |
| 638 | LLVM_DEBUG({ |
| 639 | if (!SplitLIs.empty()) { |
| 640 | dbgs() << "Split unspilled interval into " << (SplitLIs.size() + 1) |
| 641 | << " components\n" ; |
| 642 | } |
| 643 | }); |
| 644 | |
| 645 | LRM.assign(VirtReg: NewLI, PhysReg); |
| 646 | for (LiveInterval *SplitLI : SplitLIs) { |
| 647 | VRM.grow(); |
| 648 | LRM.assign(VirtReg: *SplitLI, PhysReg); |
| 649 | } |
| 650 | |
| 651 | MFI.RemoveStackObject(ObjectIdx: Slot); |
| 652 | break; |
| 653 | } |
| 654 | } |
| 655 | } |
| 656 | |
| 657 | bool AMDGPURewriteAGPRCopyMFMAImpl::run(MachineFunction &MF) const { |
| 658 | // This only applies on subtargets that have a configurable AGPR vs. VGPR |
| 659 | // allocation. |
| 660 | if (!ST.hasGFX90AInsts()) |
| 661 | return false; |
| 662 | |
| 663 | // Early exit if no AGPRs were assigned. |
| 664 | if (!LRM.isPhysRegUsed(PhysReg: AMDGPU::AGPR0)) { |
| 665 | LLVM_DEBUG(dbgs() << "skipping function that did not allocate AGPRs\n" ); |
| 666 | return false; |
| 667 | } |
| 668 | |
| 669 | bool MadeChange = false; |
| 670 | |
| 671 | for (unsigned I = 0, E = MRI.getNumVirtRegs(); I != E; ++I) { |
| 672 | Register VReg = Register::index2VirtReg(Index: I); |
| 673 | MCRegister AssignedAGPR = getAssignedAGPR(VReg); |
| 674 | if (!AssignedAGPR) |
| 675 | continue; |
| 676 | |
| 677 | if (tryFoldCopiesToAGPR(VReg, AssignedAGPR)) |
| 678 | MadeChange = true; |
| 679 | if (tryFoldCopiesFromAGPR(VReg, AssignedAGPR)) |
| 680 | MadeChange = true; |
| 681 | } |
| 682 | |
| 683 | // If we've successfully rewritten some MFMAs, we've alleviated some VGPR |
| 684 | // pressure. See if we can eliminate some spills now that those registers are |
| 685 | // more available. |
| 686 | if (MadeChange) |
| 687 | eliminateSpillsOfReassignedVGPRs(); |
| 688 | |
| 689 | return MadeChange; |
| 690 | } |
| 691 | |
| 692 | class AMDGPURewriteAGPRCopyMFMALegacy : public MachineFunctionPass { |
| 693 | public: |
| 694 | static char ID; |
| 695 | RegisterClassInfo RegClassInfo; |
| 696 | |
| 697 | AMDGPURewriteAGPRCopyMFMALegacy() : MachineFunctionPass(ID) {} |
| 698 | |
| 699 | bool runOnMachineFunction(MachineFunction &MF) override; |
| 700 | |
| 701 | StringRef getPassName() const override { |
| 702 | return "AMDGPU Rewrite AGPR-Copy-MFMA" ; |
| 703 | } |
| 704 | |
| 705 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 706 | AU.addRequired<LiveIntervalsWrapperPass>(); |
| 707 | AU.addRequired<VirtRegMapWrapperLegacy>(); |
| 708 | AU.addRequired<LiveRegMatrixWrapperLegacy>(); |
| 709 | AU.addRequired<LiveStacksWrapperLegacy>(); |
| 710 | AU.addRequired<MachineDominatorTreeWrapperPass>(); |
| 711 | |
| 712 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
| 713 | AU.addPreserved<VirtRegMapWrapperLegacy>(); |
| 714 | AU.addPreserved<LiveRegMatrixWrapperLegacy>(); |
| 715 | AU.addPreserved<LiveStacksWrapperLegacy>(); |
| 716 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
| 717 | |
| 718 | AU.setPreservesAll(); |
| 719 | MachineFunctionPass::getAnalysisUsage(AU); |
| 720 | } |
| 721 | }; |
| 722 | |
| 723 | } // End anonymous namespace. |
| 724 | |
| 725 | INITIALIZE_PASS_BEGIN(AMDGPURewriteAGPRCopyMFMALegacy, DEBUG_TYPE, |
| 726 | "AMDGPU Rewrite AGPR-Copy-MFMA" , false, false) |
| 727 | INITIALIZE_PASS_DEPENDENCY(LiveIntervalsWrapperPass) |
| 728 | INITIALIZE_PASS_DEPENDENCY(VirtRegMapWrapperLegacy) |
| 729 | INITIALIZE_PASS_DEPENDENCY(LiveRegMatrixWrapperLegacy) |
| 730 | INITIALIZE_PASS_DEPENDENCY(LiveStacksWrapperLegacy) |
| 731 | INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass) |
| 732 | INITIALIZE_PASS_END(AMDGPURewriteAGPRCopyMFMALegacy, DEBUG_TYPE, |
| 733 | "AMDGPU Rewrite AGPR-Copy-MFMA" , false, false) |
| 734 | |
| 735 | char AMDGPURewriteAGPRCopyMFMALegacy::ID = 0; |
| 736 | |
| 737 | char &llvm::AMDGPURewriteAGPRCopyMFMALegacyID = |
| 738 | AMDGPURewriteAGPRCopyMFMALegacy::ID; |
| 739 | |
| 740 | bool AMDGPURewriteAGPRCopyMFMALegacy::runOnMachineFunction( |
| 741 | MachineFunction &MF) { |
| 742 | if (skipFunction(F: MF.getFunction())) |
| 743 | return false; |
| 744 | |
| 745 | RegClassInfo.runOnMachineFunction(MF); |
| 746 | |
| 747 | auto &VRM = getAnalysis<VirtRegMapWrapperLegacy>().getVRM(); |
| 748 | auto &LRM = getAnalysis<LiveRegMatrixWrapperLegacy>().getLRM(); |
| 749 | auto &LIS = getAnalysis<LiveIntervalsWrapperPass>().getLIS(); |
| 750 | auto &LSS = getAnalysis<LiveStacksWrapperLegacy>().getLS(); |
| 751 | auto &MDT = getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree(); |
| 752 | AMDGPURewriteAGPRCopyMFMAImpl Impl(MF, VRM, LRM, LIS, LSS, RegClassInfo, MDT); |
| 753 | return Impl.run(MF); |
| 754 | } |
| 755 | |
| 756 | PreservedAnalyses |
| 757 | AMDGPURewriteAGPRCopyMFMAPass::run(MachineFunction &MF, |
| 758 | MachineFunctionAnalysisManager &MFAM) { |
| 759 | VirtRegMap &VRM = MFAM.getResult<VirtRegMapAnalysis>(IR&: MF); |
| 760 | LiveRegMatrix &LRM = MFAM.getResult<LiveRegMatrixAnalysis>(IR&: MF); |
| 761 | LiveIntervals &LIS = MFAM.getResult<LiveIntervalsAnalysis>(IR&: MF); |
| 762 | LiveStacks &LSS = MFAM.getResult<LiveStacksAnalysis>(IR&: MF); |
| 763 | MachineDominatorTree &MDT = MFAM.getResult<MachineDominatorTreeAnalysis>(IR&: MF); |
| 764 | RegisterClassInfo RegClassInfo; |
| 765 | RegClassInfo.runOnMachineFunction(MF); |
| 766 | |
| 767 | AMDGPURewriteAGPRCopyMFMAImpl Impl(MF, VRM, LRM, LIS, LSS, RegClassInfo, MDT); |
| 768 | if (!Impl.run(MF)) |
| 769 | return PreservedAnalyses::all(); |
| 770 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
| 771 | PA.preserveSet<CFGAnalyses>() |
| 772 | .preserve<LiveStacksAnalysis>() |
| 773 | .preserve<VirtRegMapAnalysis>() |
| 774 | .preserve<SlotIndexesAnalysis>() |
| 775 | .preserve<LiveIntervalsAnalysis>() |
| 776 | .preserve<LiveRegMatrixAnalysis>(); |
| 777 | return PA; |
| 778 | } |
| 779 | |