1 | //===-- SILowerControlFlow.cpp - Use predicates for control flow ----------===// |
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 |
10 | /// This pass lowers the pseudo control flow instructions to real |
11 | /// machine instructions. |
12 | /// |
13 | /// All control flow is handled using predicated instructions and |
14 | /// a predicate stack. Each Scalar ALU controls the operations of 64 Vector |
15 | /// ALUs. The Scalar ALU can update the predicate for any of the Vector ALUs |
16 | /// by writing to the 64-bit EXEC register (each bit corresponds to a |
17 | /// single vector ALU). Typically, for predicates, a vector ALU will write |
18 | /// to its bit of the VCC register (like EXEC VCC is 64-bits, one for each |
19 | /// Vector ALU) and then the ScalarALU will AND the VCC register with the |
20 | /// EXEC to update the predicates. |
21 | /// |
22 | /// For example: |
23 | /// %vcc = V_CMP_GT_F32 %vgpr1, %vgpr2 |
24 | /// %sgpr0 = SI_IF %vcc |
25 | /// %vgpr0 = V_ADD_F32 %vgpr0, %vgpr0 |
26 | /// %sgpr0 = SI_ELSE %sgpr0 |
27 | /// %vgpr0 = V_SUB_F32 %vgpr0, %vgpr0 |
28 | /// SI_END_CF %sgpr0 |
29 | /// |
30 | /// becomes: |
31 | /// |
32 | /// %sgpr0 = S_AND_SAVEEXEC_B64 %vcc // Save and update the exec mask |
33 | /// %sgpr0 = S_XOR_B64 %sgpr0, %exec // Clear live bits from saved exec mask |
34 | /// S_CBRANCH_EXECZ label0 // This instruction is an optional |
35 | /// // optimization which allows us to |
36 | /// // branch if all the bits of |
37 | /// // EXEC are zero. |
38 | /// %vgpr0 = V_ADD_F32 %vgpr0, %vgpr0 // Do the IF block of the branch |
39 | /// |
40 | /// label0: |
41 | /// %sgpr0 = S_OR_SAVEEXEC_B64 %sgpr0 // Restore the exec mask for the Then |
42 | /// // block |
43 | /// %exec = S_XOR_B64 %sgpr0, %exec // Update the exec mask |
44 | /// S_BRANCH_EXECZ label1 // Use our branch optimization |
45 | /// // instruction again. |
46 | /// %vgpr0 = V_SUB_F32 %vgpr0, %vgpr // Do the THEN block |
47 | /// label1: |
48 | /// %exec = S_OR_B64 %exec, %sgpr0 // Re-enable saved exec mask bits |
49 | //===----------------------------------------------------------------------===// |
50 | |
51 | #include "AMDGPU.h" |
52 | #include "GCNSubtarget.h" |
53 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
54 | #include "llvm/ADT/SmallSet.h" |
55 | #include "llvm/CodeGen/LiveIntervals.h" |
56 | #include "llvm/CodeGen/LiveVariables.h" |
57 | #include "llvm/CodeGen/MachineDominators.h" |
58 | #include "llvm/CodeGen/MachineFunctionPass.h" |
59 | #include "llvm/Target/TargetMachine.h" |
60 | |
61 | using namespace llvm; |
62 | |
63 | #define DEBUG_TYPE "si-lower-control-flow" |
64 | |
65 | static cl::opt<bool> |
66 | RemoveRedundantEndcf("amdgpu-remove-redundant-endcf" , |
67 | cl::init(Val: true), cl::ReallyHidden); |
68 | |
69 | namespace { |
70 | |
71 | class SILowerControlFlow : public MachineFunctionPass { |
72 | private: |
73 | const SIRegisterInfo *TRI = nullptr; |
74 | const SIInstrInfo *TII = nullptr; |
75 | LiveIntervals *LIS = nullptr; |
76 | LiveVariables *LV = nullptr; |
77 | MachineDominatorTree *MDT = nullptr; |
78 | MachineRegisterInfo *MRI = nullptr; |
79 | SetVector<MachineInstr*> LoweredEndCf; |
80 | DenseSet<Register> LoweredIf; |
81 | SmallSet<MachineBasicBlock *, 4> KillBlocks; |
82 | SmallSet<Register, 8> RecomputeRegs; |
83 | |
84 | const TargetRegisterClass *BoolRC = nullptr; |
85 | unsigned AndOpc; |
86 | unsigned OrOpc; |
87 | unsigned XorOpc; |
88 | unsigned MovTermOpc; |
89 | unsigned Andn2TermOpc; |
90 | unsigned XorTermrOpc; |
91 | unsigned OrTermrOpc; |
92 | unsigned OrSaveExecOpc; |
93 | unsigned Exec; |
94 | |
95 | bool EnableOptimizeEndCf = false; |
96 | |
97 | bool hasKill(const MachineBasicBlock *Begin, const MachineBasicBlock *End); |
98 | |
99 | void emitIf(MachineInstr &MI); |
100 | void emitElse(MachineInstr &MI); |
101 | void emitIfBreak(MachineInstr &MI); |
102 | void emitLoop(MachineInstr &MI); |
103 | |
104 | MachineBasicBlock *emitEndCf(MachineInstr &MI); |
105 | |
106 | void findMaskOperands(MachineInstr &MI, unsigned OpNo, |
107 | SmallVectorImpl<MachineOperand> &Src) const; |
108 | |
109 | void combineMasks(MachineInstr &MI); |
110 | |
111 | bool removeMBBifRedundant(MachineBasicBlock &MBB); |
112 | |
113 | MachineBasicBlock *process(MachineInstr &MI); |
114 | |
115 | // Skip to the next instruction, ignoring debug instructions, and trivial |
116 | // block boundaries (blocks that have one (typically fallthrough) successor, |
117 | // and the successor has one predecessor. |
118 | MachineBasicBlock::iterator |
119 | skipIgnoreExecInstsTrivialSucc(MachineBasicBlock &MBB, |
120 | MachineBasicBlock::iterator It) const; |
121 | |
122 | /// Find the insertion point for a new conditional branch. |
123 | MachineBasicBlock::iterator |
124 | skipToUncondBrOrEnd(MachineBasicBlock &MBB, |
125 | MachineBasicBlock::iterator I) const { |
126 | assert(I->isTerminator()); |
127 | |
128 | // FIXME: What if we had multiple pre-existing conditional branches? |
129 | MachineBasicBlock::iterator End = MBB.end(); |
130 | while (I != End && !I->isUnconditionalBranch()) |
131 | ++I; |
132 | return I; |
133 | } |
134 | |
135 | // Remove redundant SI_END_CF instructions. |
136 | void optimizeEndCf(); |
137 | |
138 | public: |
139 | static char ID; |
140 | |
141 | SILowerControlFlow() : MachineFunctionPass(ID) {} |
142 | |
143 | bool runOnMachineFunction(MachineFunction &MF) override; |
144 | |
145 | StringRef getPassName() const override { |
146 | return "SI Lower control flow pseudo instructions" ; |
147 | } |
148 | |
149 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
150 | AU.addUsedIfAvailable<LiveIntervalsWrapperPass>(); |
151 | // Should preserve the same set that TwoAddressInstructions does. |
152 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
153 | AU.addPreserved<SlotIndexesWrapperPass>(); |
154 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
155 | AU.addPreservedID(ID&: LiveVariablesID); |
156 | MachineFunctionPass::getAnalysisUsage(AU); |
157 | } |
158 | }; |
159 | |
160 | } // end anonymous namespace |
161 | |
162 | char SILowerControlFlow::ID = 0; |
163 | |
164 | INITIALIZE_PASS(SILowerControlFlow, DEBUG_TYPE, |
165 | "SI lower control flow" , false, false) |
166 | |
167 | static void setImpSCCDefDead(MachineInstr &MI, bool IsDead) { |
168 | MachineOperand &ImpDefSCC = MI.getOperand(i: 3); |
169 | assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef()); |
170 | |
171 | ImpDefSCC.setIsDead(IsDead); |
172 | } |
173 | |
174 | char &llvm::SILowerControlFlowID = SILowerControlFlow::ID; |
175 | |
176 | bool SILowerControlFlow::hasKill(const MachineBasicBlock *Begin, |
177 | const MachineBasicBlock *End) { |
178 | DenseSet<const MachineBasicBlock*> Visited; |
179 | SmallVector<MachineBasicBlock *, 4> Worklist(Begin->successors()); |
180 | |
181 | while (!Worklist.empty()) { |
182 | MachineBasicBlock *MBB = Worklist.pop_back_val(); |
183 | |
184 | if (MBB == End || !Visited.insert(V: MBB).second) |
185 | continue; |
186 | if (KillBlocks.contains(Ptr: MBB)) |
187 | return true; |
188 | |
189 | Worklist.append(in_start: MBB->succ_begin(), in_end: MBB->succ_end()); |
190 | } |
191 | |
192 | return false; |
193 | } |
194 | |
195 | static bool isSimpleIf(const MachineInstr &MI, const MachineRegisterInfo *MRI) { |
196 | Register SaveExecReg = MI.getOperand(i: 0).getReg(); |
197 | auto U = MRI->use_instr_nodbg_begin(RegNo: SaveExecReg); |
198 | |
199 | if (U == MRI->use_instr_nodbg_end() || |
200 | std::next(x: U) != MRI->use_instr_nodbg_end() || |
201 | U->getOpcode() != AMDGPU::SI_END_CF) |
202 | return false; |
203 | |
204 | return true; |
205 | } |
206 | |
207 | void SILowerControlFlow::emitIf(MachineInstr &MI) { |
208 | MachineBasicBlock &MBB = *MI.getParent(); |
209 | const DebugLoc &DL = MI.getDebugLoc(); |
210 | MachineBasicBlock::iterator I(&MI); |
211 | Register SaveExecReg = MI.getOperand(i: 0).getReg(); |
212 | MachineOperand& Cond = MI.getOperand(i: 1); |
213 | assert(Cond.getSubReg() == AMDGPU::NoSubRegister); |
214 | |
215 | MachineOperand &ImpDefSCC = MI.getOperand(i: 4); |
216 | assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef()); |
217 | |
218 | // If there is only one use of save exec register and that use is SI_END_CF, |
219 | // we can optimize SI_IF by returning the full saved exec mask instead of |
220 | // just cleared bits. |
221 | bool SimpleIf = isSimpleIf(MI, MRI); |
222 | |
223 | if (SimpleIf) { |
224 | // Check for SI_KILL_*_TERMINATOR on path from if to endif. |
225 | // if there is any such terminator simplifications are not safe. |
226 | auto UseMI = MRI->use_instr_nodbg_begin(RegNo: SaveExecReg); |
227 | SimpleIf = !hasKill(Begin: MI.getParent(), End: UseMI->getParent()); |
228 | } |
229 | |
230 | // Add an implicit def of exec to discourage scheduling VALU after this which |
231 | // will interfere with trying to form s_and_saveexec_b64 later. |
232 | Register CopyReg = SimpleIf ? SaveExecReg |
233 | : MRI->createVirtualRegister(RegClass: BoolRC); |
234 | MachineInstr *CopyExec = |
235 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: CopyReg) |
236 | .addReg(RegNo: Exec) |
237 | .addReg(RegNo: Exec, flags: RegState::ImplicitDefine); |
238 | LoweredIf.insert(V: CopyReg); |
239 | |
240 | Register Tmp = MRI->createVirtualRegister(RegClass: BoolRC); |
241 | |
242 | MachineInstr *And = |
243 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: Tmp) |
244 | .addReg(RegNo: CopyReg) |
245 | .add(MO: Cond); |
246 | if (LV) |
247 | LV->replaceKillInstruction(Reg: Cond.getReg(), OldMI&: MI, NewMI&: *And); |
248 | |
249 | setImpSCCDefDead(MI&: *And, IsDead: true); |
250 | |
251 | MachineInstr *Xor = nullptr; |
252 | if (!SimpleIf) { |
253 | Xor = |
254 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: XorOpc), DestReg: SaveExecReg) |
255 | .addReg(RegNo: Tmp) |
256 | .addReg(RegNo: CopyReg); |
257 | setImpSCCDefDead(MI&: *Xor, IsDead: ImpDefSCC.isDead()); |
258 | } |
259 | |
260 | // Use a copy that is a terminator to get correct spill code placement it with |
261 | // fast regalloc. |
262 | MachineInstr *SetExec = |
263 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: MovTermOpc), DestReg: Exec) |
264 | .addReg(RegNo: Tmp, flags: RegState::Kill); |
265 | if (LV) |
266 | LV->getVarInfo(Reg: Tmp).Kills.push_back(x: SetExec); |
267 | |
268 | // Skip ahead to the unconditional branch in case there are other terminators |
269 | // present. |
270 | I = skipToUncondBrOrEnd(MBB, I); |
271 | |
272 | // Insert the S_CBRANCH_EXECZ instruction which will be optimized later |
273 | // during SIRemoveShortExecBranches. |
274 | MachineInstr *NewBr = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECZ)) |
275 | .add(MO: MI.getOperand(i: 2)); |
276 | |
277 | if (!LIS) { |
278 | MI.eraseFromParent(); |
279 | return; |
280 | } |
281 | |
282 | LIS->InsertMachineInstrInMaps(MI&: *CopyExec); |
283 | |
284 | // Replace with and so we don't need to fix the live interval for condition |
285 | // register. |
286 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *And); |
287 | |
288 | if (!SimpleIf) |
289 | LIS->InsertMachineInstrInMaps(MI&: *Xor); |
290 | LIS->InsertMachineInstrInMaps(MI&: *SetExec); |
291 | LIS->InsertMachineInstrInMaps(MI&: *NewBr); |
292 | |
293 | LIS->removeAllRegUnitsForPhysReg(Reg: AMDGPU::EXEC); |
294 | MI.eraseFromParent(); |
295 | |
296 | // FIXME: Is there a better way of adjusting the liveness? It shouldn't be |
297 | // hard to add another def here but I'm not sure how to correctly update the |
298 | // valno. |
299 | RecomputeRegs.insert(V: SaveExecReg); |
300 | LIS->createAndComputeVirtRegInterval(Reg: Tmp); |
301 | if (!SimpleIf) |
302 | LIS->createAndComputeVirtRegInterval(Reg: CopyReg); |
303 | } |
304 | |
305 | void SILowerControlFlow::emitElse(MachineInstr &MI) { |
306 | MachineBasicBlock &MBB = *MI.getParent(); |
307 | const DebugLoc &DL = MI.getDebugLoc(); |
308 | |
309 | Register DstReg = MI.getOperand(i: 0).getReg(); |
310 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
311 | |
312 | MachineBasicBlock::iterator Start = MBB.begin(); |
313 | |
314 | // This must be inserted before phis and any spill code inserted before the |
315 | // else. |
316 | Register SaveReg = MRI->createVirtualRegister(RegClass: BoolRC); |
317 | MachineInstr *OrSaveExec = |
318 | BuildMI(BB&: MBB, I: Start, MIMD: DL, MCID: TII->get(Opcode: OrSaveExecOpc), DestReg: SaveReg) |
319 | .add(MO: MI.getOperand(i: 1)); // Saved EXEC |
320 | if (LV) |
321 | LV->replaceKillInstruction(Reg: SrcReg, OldMI&: MI, NewMI&: *OrSaveExec); |
322 | |
323 | MachineBasicBlock *DestBB = MI.getOperand(i: 2).getMBB(); |
324 | |
325 | MachineBasicBlock::iterator ElsePt(MI); |
326 | |
327 | // This accounts for any modification of the EXEC mask within the block and |
328 | // can be optimized out pre-RA when not required. |
329 | MachineInstr *And = BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: DstReg) |
330 | .addReg(RegNo: Exec) |
331 | .addReg(RegNo: SaveReg); |
332 | |
333 | MachineInstr *Xor = |
334 | BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: XorTermrOpc), DestReg: Exec) |
335 | .addReg(RegNo: Exec) |
336 | .addReg(RegNo: DstReg); |
337 | |
338 | // Skip ahead to the unconditional branch in case there are other terminators |
339 | // present. |
340 | ElsePt = skipToUncondBrOrEnd(MBB, I: ElsePt); |
341 | |
342 | MachineInstr *Branch = |
343 | BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECZ)) |
344 | .addMBB(MBB: DestBB); |
345 | |
346 | if (!LIS) { |
347 | MI.eraseFromParent(); |
348 | return; |
349 | } |
350 | |
351 | LIS->RemoveMachineInstrFromMaps(MI); |
352 | MI.eraseFromParent(); |
353 | |
354 | LIS->InsertMachineInstrInMaps(MI&: *OrSaveExec); |
355 | LIS->InsertMachineInstrInMaps(MI&: *And); |
356 | |
357 | LIS->InsertMachineInstrInMaps(MI&: *Xor); |
358 | LIS->InsertMachineInstrInMaps(MI&: *Branch); |
359 | |
360 | RecomputeRegs.insert(V: SrcReg); |
361 | RecomputeRegs.insert(V: DstReg); |
362 | LIS->createAndComputeVirtRegInterval(Reg: SaveReg); |
363 | |
364 | // Let this be recomputed. |
365 | LIS->removeAllRegUnitsForPhysReg(Reg: AMDGPU::EXEC); |
366 | } |
367 | |
368 | void SILowerControlFlow::emitIfBreak(MachineInstr &MI) { |
369 | MachineBasicBlock &MBB = *MI.getParent(); |
370 | const DebugLoc &DL = MI.getDebugLoc(); |
371 | auto Dst = MI.getOperand(i: 0).getReg(); |
372 | |
373 | // Skip ANDing with exec if the break condition is already masked by exec |
374 | // because it is a V_CMP in the same basic block. (We know the break |
375 | // condition operand was an i1 in IR, so if it is a VALU instruction it must |
376 | // be one with a carry-out.) |
377 | bool SkipAnding = false; |
378 | if (MI.getOperand(i: 1).isReg()) { |
379 | if (MachineInstr *Def = MRI->getUniqueVRegDef(Reg: MI.getOperand(i: 1).getReg())) { |
380 | SkipAnding = Def->getParent() == MI.getParent() |
381 | && SIInstrInfo::isVALU(MI: *Def); |
382 | } |
383 | } |
384 | |
385 | // AND the break condition operand with exec, then OR that into the "loop |
386 | // exit" mask. |
387 | MachineInstr *And = nullptr, *Or = nullptr; |
388 | Register AndReg; |
389 | if (!SkipAnding) { |
390 | AndReg = MRI->createVirtualRegister(RegClass: BoolRC); |
391 | And = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: AndReg) |
392 | .addReg(RegNo: Exec) |
393 | .add(MO: MI.getOperand(i: 1)); |
394 | if (LV) |
395 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 1).getReg(), OldMI&: MI, NewMI&: *And); |
396 | Or = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: OrOpc), DestReg: Dst) |
397 | .addReg(RegNo: AndReg) |
398 | .add(MO: MI.getOperand(i: 2)); |
399 | } else { |
400 | Or = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: OrOpc), DestReg: Dst) |
401 | .add(MO: MI.getOperand(i: 1)) |
402 | .add(MO: MI.getOperand(i: 2)); |
403 | if (LV) |
404 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 1).getReg(), OldMI&: MI, NewMI&: *Or); |
405 | } |
406 | if (LV) |
407 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 2).getReg(), OldMI&: MI, NewMI&: *Or); |
408 | |
409 | if (LIS) { |
410 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *Or); |
411 | if (And) { |
412 | // Read of original operand 1 is on And now not Or. |
413 | RecomputeRegs.insert(V: And->getOperand(i: 2).getReg()); |
414 | LIS->InsertMachineInstrInMaps(MI&: *And); |
415 | LIS->createAndComputeVirtRegInterval(Reg: AndReg); |
416 | } |
417 | } |
418 | |
419 | MI.eraseFromParent(); |
420 | } |
421 | |
422 | void SILowerControlFlow::emitLoop(MachineInstr &MI) { |
423 | MachineBasicBlock &MBB = *MI.getParent(); |
424 | const DebugLoc &DL = MI.getDebugLoc(); |
425 | |
426 | MachineInstr *AndN2 = |
427 | BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: Andn2TermOpc), DestReg: Exec) |
428 | .addReg(RegNo: Exec) |
429 | .add(MO: MI.getOperand(i: 0)); |
430 | if (LV) |
431 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 0).getReg(), OldMI&: MI, NewMI&: *AndN2); |
432 | |
433 | auto BranchPt = skipToUncondBrOrEnd(MBB, I: MI.getIterator()); |
434 | MachineInstr *Branch = |
435 | BuildMI(BB&: MBB, I: BranchPt, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECNZ)) |
436 | .add(MO: MI.getOperand(i: 1)); |
437 | |
438 | if (LIS) { |
439 | RecomputeRegs.insert(V: MI.getOperand(i: 0).getReg()); |
440 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *AndN2); |
441 | LIS->InsertMachineInstrInMaps(MI&: *Branch); |
442 | } |
443 | |
444 | MI.eraseFromParent(); |
445 | } |
446 | |
447 | MachineBasicBlock::iterator |
448 | SILowerControlFlow::skipIgnoreExecInstsTrivialSucc( |
449 | MachineBasicBlock &MBB, MachineBasicBlock::iterator It) const { |
450 | |
451 | SmallSet<const MachineBasicBlock *, 4> Visited; |
452 | MachineBasicBlock *B = &MBB; |
453 | do { |
454 | if (!Visited.insert(Ptr: B).second) |
455 | return MBB.end(); |
456 | |
457 | auto E = B->end(); |
458 | for ( ; It != E; ++It) { |
459 | if (TII->mayReadEXEC(MRI: *MRI, MI: *It)) |
460 | break; |
461 | } |
462 | |
463 | if (It != E) |
464 | return It; |
465 | |
466 | if (B->succ_size() != 1) |
467 | return MBB.end(); |
468 | |
469 | // If there is one trivial successor, advance to the next block. |
470 | MachineBasicBlock *Succ = *B->succ_begin(); |
471 | |
472 | It = Succ->begin(); |
473 | B = Succ; |
474 | } while (true); |
475 | } |
476 | |
477 | MachineBasicBlock *SILowerControlFlow::emitEndCf(MachineInstr &MI) { |
478 | MachineBasicBlock &MBB = *MI.getParent(); |
479 | const DebugLoc &DL = MI.getDebugLoc(); |
480 | |
481 | MachineBasicBlock::iterator InsPt = MBB.begin(); |
482 | |
483 | // If we have instructions that aren't prolog instructions, split the block |
484 | // and emit a terminator instruction. This ensures correct spill placement. |
485 | // FIXME: We should unconditionally split the block here. |
486 | bool NeedBlockSplit = false; |
487 | Register DataReg = MI.getOperand(i: 0).getReg(); |
488 | for (MachineBasicBlock::iterator I = InsPt, E = MI.getIterator(); |
489 | I != E; ++I) { |
490 | if (I->modifiesRegister(Reg: DataReg, TRI)) { |
491 | NeedBlockSplit = true; |
492 | break; |
493 | } |
494 | } |
495 | |
496 | unsigned Opcode = OrOpc; |
497 | MachineBasicBlock *SplitBB = &MBB; |
498 | if (NeedBlockSplit) { |
499 | SplitBB = MBB.splitAt(SplitInst&: MI, /*UpdateLiveIns*/true, LIS); |
500 | if (MDT && SplitBB != &MBB) { |
501 | MachineDomTreeNode *MBBNode = (*MDT)[&MBB]; |
502 | SmallVector<MachineDomTreeNode *> Children(MBBNode->begin(), |
503 | MBBNode->end()); |
504 | MachineDomTreeNode *SplitBBNode = MDT->addNewBlock(BB: SplitBB, DomBB: &MBB); |
505 | for (MachineDomTreeNode *Child : Children) |
506 | MDT->changeImmediateDominator(N: Child, NewIDom: SplitBBNode); |
507 | } |
508 | Opcode = OrTermrOpc; |
509 | InsPt = MI; |
510 | } |
511 | |
512 | MachineInstr *NewMI = |
513 | BuildMI(BB&: MBB, I: InsPt, MIMD: DL, MCID: TII->get(Opcode), DestReg: Exec) |
514 | .addReg(RegNo: Exec) |
515 | .add(MO: MI.getOperand(i: 0)); |
516 | if (LV) { |
517 | LV->replaceKillInstruction(Reg: DataReg, OldMI&: MI, NewMI&: *NewMI); |
518 | |
519 | if (SplitBB != &MBB) { |
520 | // Track the set of registers defined in the original block so we don't |
521 | // accidentally add the original block to AliveBlocks. AliveBlocks only |
522 | // includes blocks which are live through, which excludes live outs and |
523 | // local defs. |
524 | DenseSet<Register> DefInOrigBlock; |
525 | |
526 | for (MachineBasicBlock *BlockPiece : {&MBB, SplitBB}) { |
527 | for (MachineInstr &X : *BlockPiece) { |
528 | for (MachineOperand &Op : X.all_defs()) { |
529 | if (Op.getReg().isVirtual()) |
530 | DefInOrigBlock.insert(V: Op.getReg()); |
531 | } |
532 | } |
533 | } |
534 | |
535 | for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) { |
536 | Register Reg = Register::index2VirtReg(Index: i); |
537 | LiveVariables::VarInfo &VI = LV->getVarInfo(Reg); |
538 | |
539 | if (VI.AliveBlocks.test(Idx: MBB.getNumber())) |
540 | VI.AliveBlocks.set(SplitBB->getNumber()); |
541 | else { |
542 | for (MachineInstr *Kill : VI.Kills) { |
543 | if (Kill->getParent() == SplitBB && !DefInOrigBlock.contains(V: Reg)) |
544 | VI.AliveBlocks.set(MBB.getNumber()); |
545 | } |
546 | } |
547 | } |
548 | } |
549 | } |
550 | |
551 | LoweredEndCf.insert(X: NewMI); |
552 | |
553 | if (LIS) |
554 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *NewMI); |
555 | |
556 | MI.eraseFromParent(); |
557 | |
558 | if (LIS) |
559 | LIS->handleMove(MI&: *NewMI); |
560 | return SplitBB; |
561 | } |
562 | |
563 | // Returns replace operands for a logical operation, either single result |
564 | // for exec or two operands if source was another equivalent operation. |
565 | void SILowerControlFlow::findMaskOperands(MachineInstr &MI, unsigned OpNo, |
566 | SmallVectorImpl<MachineOperand> &Src) const { |
567 | MachineOperand &Op = MI.getOperand(i: OpNo); |
568 | if (!Op.isReg() || !Op.getReg().isVirtual()) { |
569 | Src.push_back(Elt: Op); |
570 | return; |
571 | } |
572 | |
573 | MachineInstr *Def = MRI->getUniqueVRegDef(Reg: Op.getReg()); |
574 | if (!Def || Def->getParent() != MI.getParent() || |
575 | !(Def->isFullCopy() || (Def->getOpcode() == MI.getOpcode()))) |
576 | return; |
577 | |
578 | // Make sure we do not modify exec between def and use. |
579 | // A copy with implicitly defined exec inserted earlier is an exclusion, it |
580 | // does not really modify exec. |
581 | for (auto I = Def->getIterator(); I != MI.getIterator(); ++I) |
582 | if (I->modifiesRegister(Reg: AMDGPU::EXEC, TRI) && |
583 | !(I->isCopy() && I->getOperand(i: 0).getReg() != Exec)) |
584 | return; |
585 | |
586 | for (const auto &SrcOp : Def->explicit_operands()) |
587 | if (SrcOp.isReg() && SrcOp.isUse() && |
588 | (SrcOp.getReg().isVirtual() || SrcOp.getReg() == Exec)) |
589 | Src.push_back(Elt: SrcOp); |
590 | } |
591 | |
592 | // Search and combine pairs of equivalent instructions, like |
593 | // S_AND_B64 x, (S_AND_B64 x, y) => S_AND_B64 x, y |
594 | // S_OR_B64 x, (S_OR_B64 x, y) => S_OR_B64 x, y |
595 | // One of the operands is exec mask. |
596 | void SILowerControlFlow::combineMasks(MachineInstr &MI) { |
597 | assert(MI.getNumExplicitOperands() == 3); |
598 | SmallVector<MachineOperand, 4> Ops; |
599 | unsigned OpToReplace = 1; |
600 | findMaskOperands(MI, OpNo: 1, Src&: Ops); |
601 | if (Ops.size() == 1) OpToReplace = 2; // First operand can be exec or its copy |
602 | findMaskOperands(MI, OpNo: 2, Src&: Ops); |
603 | if (Ops.size() != 3) return; |
604 | |
605 | unsigned UniqueOpndIdx; |
606 | if (Ops[0].isIdenticalTo(Other: Ops[1])) UniqueOpndIdx = 2; |
607 | else if (Ops[0].isIdenticalTo(Other: Ops[2])) UniqueOpndIdx = 1; |
608 | else if (Ops[1].isIdenticalTo(Other: Ops[2])) UniqueOpndIdx = 1; |
609 | else return; |
610 | |
611 | Register Reg = MI.getOperand(i: OpToReplace).getReg(); |
612 | MI.removeOperand(OpNo: OpToReplace); |
613 | MI.addOperand(Op: Ops[UniqueOpndIdx]); |
614 | if (MRI->use_empty(RegNo: Reg)) |
615 | MRI->getUniqueVRegDef(Reg)->eraseFromParent(); |
616 | } |
617 | |
618 | void SILowerControlFlow::optimizeEndCf() { |
619 | // If the only instruction immediately following this END_CF is another |
620 | // END_CF in the only successor we can avoid emitting exec mask restore here. |
621 | if (!EnableOptimizeEndCf) |
622 | return; |
623 | |
624 | for (MachineInstr *MI : reverse(C&: LoweredEndCf)) { |
625 | MachineBasicBlock &MBB = *MI->getParent(); |
626 | auto Next = |
627 | skipIgnoreExecInstsTrivialSucc(MBB, It: std::next(x: MI->getIterator())); |
628 | if (Next == MBB.end() || !LoweredEndCf.count(key: &*Next)) |
629 | continue; |
630 | // Only skip inner END_CF if outer ENDCF belongs to SI_IF. |
631 | // If that belongs to SI_ELSE then saved mask has an inverted value. |
632 | Register SavedExec |
633 | = TII->getNamedOperand(MI&: *Next, OperandName: AMDGPU::OpName::src1)->getReg(); |
634 | assert(SavedExec.isVirtual() && "Expected saved exec to be src1!" ); |
635 | |
636 | const MachineInstr *Def = MRI->getUniqueVRegDef(Reg: SavedExec); |
637 | if (Def && LoweredIf.count(V: SavedExec)) { |
638 | LLVM_DEBUG(dbgs() << "Skip redundant " ; MI->dump()); |
639 | if (LIS) |
640 | LIS->RemoveMachineInstrFromMaps(MI&: *MI); |
641 | Register Reg; |
642 | if (LV) |
643 | Reg = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src1)->getReg(); |
644 | MI->eraseFromParent(); |
645 | if (LV) |
646 | LV->recomputeForSingleDefVirtReg(Reg); |
647 | removeMBBifRedundant(MBB); |
648 | } |
649 | } |
650 | } |
651 | |
652 | MachineBasicBlock *SILowerControlFlow::process(MachineInstr &MI) { |
653 | MachineBasicBlock &MBB = *MI.getParent(); |
654 | MachineBasicBlock::iterator I(MI); |
655 | MachineInstr *Prev = (I != MBB.begin()) ? &*(std::prev(x: I)) : nullptr; |
656 | |
657 | MachineBasicBlock *SplitBB = &MBB; |
658 | |
659 | switch (MI.getOpcode()) { |
660 | case AMDGPU::SI_IF: |
661 | emitIf(MI); |
662 | break; |
663 | |
664 | case AMDGPU::SI_ELSE: |
665 | emitElse(MI); |
666 | break; |
667 | |
668 | case AMDGPU::SI_IF_BREAK: |
669 | emitIfBreak(MI); |
670 | break; |
671 | |
672 | case AMDGPU::SI_LOOP: |
673 | emitLoop(MI); |
674 | break; |
675 | |
676 | case AMDGPU::SI_WATERFALL_LOOP: |
677 | MI.setDesc(TII->get(Opcode: AMDGPU::S_CBRANCH_EXECNZ)); |
678 | break; |
679 | |
680 | case AMDGPU::SI_END_CF: |
681 | SplitBB = emitEndCf(MI); |
682 | break; |
683 | |
684 | default: |
685 | assert(false && "Attempt to process unsupported instruction" ); |
686 | break; |
687 | } |
688 | |
689 | MachineBasicBlock::iterator Next; |
690 | for (I = Prev ? Prev->getIterator() : MBB.begin(); I != MBB.end(); I = Next) { |
691 | Next = std::next(x: I); |
692 | MachineInstr &MaskMI = *I; |
693 | switch (MaskMI.getOpcode()) { |
694 | case AMDGPU::S_AND_B64: |
695 | case AMDGPU::S_OR_B64: |
696 | case AMDGPU::S_AND_B32: |
697 | case AMDGPU::S_OR_B32: |
698 | // Cleanup bit manipulations on exec mask |
699 | combineMasks(MI&: MaskMI); |
700 | break; |
701 | default: |
702 | I = MBB.end(); |
703 | break; |
704 | } |
705 | } |
706 | |
707 | return SplitBB; |
708 | } |
709 | |
710 | bool SILowerControlFlow::removeMBBifRedundant(MachineBasicBlock &MBB) { |
711 | for (auto &I : MBB.instrs()) { |
712 | if (!I.isDebugInstr() && !I.isUnconditionalBranch()) |
713 | return false; |
714 | } |
715 | |
716 | assert(MBB.succ_size() == 1 && "MBB has more than one successor" ); |
717 | |
718 | MachineBasicBlock *Succ = *MBB.succ_begin(); |
719 | MachineBasicBlock *FallThrough = nullptr; |
720 | |
721 | while (!MBB.predecessors().empty()) { |
722 | MachineBasicBlock *P = *MBB.pred_begin(); |
723 | if (P->getFallThrough(JumpToFallThrough: false) == &MBB) |
724 | FallThrough = P; |
725 | P->ReplaceUsesOfBlockWith(Old: &MBB, New: Succ); |
726 | } |
727 | MBB.removeSuccessor(Succ); |
728 | if (LIS) { |
729 | for (auto &I : MBB.instrs()) |
730 | LIS->RemoveMachineInstrFromMaps(MI&: I); |
731 | } |
732 | if (MDT) { |
733 | // If Succ, the single successor of MBB, is dominated by MBB, MDT needs |
734 | // updating by changing Succ's idom to the one of MBB; otherwise, MBB must |
735 | // be a leaf node in MDT and could be erased directly. |
736 | if (MDT->dominates(A: &MBB, B: Succ)) |
737 | MDT->changeImmediateDominator(N: MDT->getNode(BB: Succ), |
738 | NewIDom: MDT->getNode(BB: &MBB)->getIDom()); |
739 | MDT->eraseNode(BB: &MBB); |
740 | } |
741 | MBB.clear(); |
742 | MBB.eraseFromParent(); |
743 | if (FallThrough && !FallThrough->isLayoutSuccessor(MBB: Succ)) { |
744 | // Note: we cannot update block layout and preserve live intervals; |
745 | // hence we must insert a branch. |
746 | MachineInstr *BranchMI = BuildMI(BB&: *FallThrough, I: FallThrough->end(), |
747 | MIMD: FallThrough->findBranchDebugLoc(), MCID: TII->get(Opcode: AMDGPU::S_BRANCH)) |
748 | .addMBB(MBB: Succ); |
749 | if (LIS) |
750 | LIS->InsertMachineInstrInMaps(MI&: *BranchMI); |
751 | } |
752 | |
753 | return true; |
754 | } |
755 | |
756 | bool SILowerControlFlow::runOnMachineFunction(MachineFunction &MF) { |
757 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
758 | TII = ST.getInstrInfo(); |
759 | TRI = &TII->getRegisterInfo(); |
760 | EnableOptimizeEndCf = RemoveRedundantEndcf && |
761 | MF.getTarget().getOptLevel() > CodeGenOptLevel::None; |
762 | |
763 | // This doesn't actually need LiveIntervals, but we can preserve them. |
764 | auto *LISWrapper = getAnalysisIfAvailable<LiveIntervalsWrapperPass>(); |
765 | LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr; |
766 | // This doesn't actually need LiveVariables, but we can preserve them. |
767 | auto *LVWrapper = getAnalysisIfAvailable<LiveVariablesWrapperPass>(); |
768 | LV = LVWrapper ? &LVWrapper->getLV() : nullptr; |
769 | auto *MDTWrapper = getAnalysisIfAvailable<MachineDominatorTreeWrapperPass>(); |
770 | MDT = MDTWrapper ? &MDTWrapper->getDomTree() : nullptr; |
771 | MRI = &MF.getRegInfo(); |
772 | BoolRC = TRI->getBoolRC(); |
773 | |
774 | if (ST.isWave32()) { |
775 | AndOpc = AMDGPU::S_AND_B32; |
776 | OrOpc = AMDGPU::S_OR_B32; |
777 | XorOpc = AMDGPU::S_XOR_B32; |
778 | MovTermOpc = AMDGPU::S_MOV_B32_term; |
779 | Andn2TermOpc = AMDGPU::S_ANDN2_B32_term; |
780 | XorTermrOpc = AMDGPU::S_XOR_B32_term; |
781 | OrTermrOpc = AMDGPU::S_OR_B32_term; |
782 | OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B32; |
783 | Exec = AMDGPU::EXEC_LO; |
784 | } else { |
785 | AndOpc = AMDGPU::S_AND_B64; |
786 | OrOpc = AMDGPU::S_OR_B64; |
787 | XorOpc = AMDGPU::S_XOR_B64; |
788 | MovTermOpc = AMDGPU::S_MOV_B64_term; |
789 | Andn2TermOpc = AMDGPU::S_ANDN2_B64_term; |
790 | XorTermrOpc = AMDGPU::S_XOR_B64_term; |
791 | OrTermrOpc = AMDGPU::S_OR_B64_term; |
792 | OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B64; |
793 | Exec = AMDGPU::EXEC; |
794 | } |
795 | |
796 | // Compute set of blocks with kills |
797 | const bool CanDemote = |
798 | MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS; |
799 | for (auto &MBB : MF) { |
800 | bool IsKillBlock = false; |
801 | for (auto &Term : MBB.terminators()) { |
802 | if (TII->isKillTerminator(Opcode: Term.getOpcode())) { |
803 | KillBlocks.insert(Ptr: &MBB); |
804 | IsKillBlock = true; |
805 | break; |
806 | } |
807 | } |
808 | if (CanDemote && !IsKillBlock) { |
809 | for (auto &MI : MBB) { |
810 | if (MI.getOpcode() == AMDGPU::SI_DEMOTE_I1) { |
811 | KillBlocks.insert(Ptr: &MBB); |
812 | break; |
813 | } |
814 | } |
815 | } |
816 | } |
817 | |
818 | bool Changed = false; |
819 | MachineFunction::iterator NextBB; |
820 | for (MachineFunction::iterator BI = MF.begin(); |
821 | BI != MF.end(); BI = NextBB) { |
822 | NextBB = std::next(x: BI); |
823 | MachineBasicBlock *MBB = &*BI; |
824 | |
825 | MachineBasicBlock::iterator I, E, Next; |
826 | E = MBB->end(); |
827 | for (I = MBB->begin(); I != E; I = Next) { |
828 | Next = std::next(x: I); |
829 | MachineInstr &MI = *I; |
830 | MachineBasicBlock *SplitMBB = MBB; |
831 | |
832 | switch (MI.getOpcode()) { |
833 | case AMDGPU::SI_IF: |
834 | case AMDGPU::SI_ELSE: |
835 | case AMDGPU::SI_IF_BREAK: |
836 | case AMDGPU::SI_WATERFALL_LOOP: |
837 | case AMDGPU::SI_LOOP: |
838 | case AMDGPU::SI_END_CF: |
839 | SplitMBB = process(MI); |
840 | Changed = true; |
841 | break; |
842 | } |
843 | |
844 | if (SplitMBB != MBB) { |
845 | MBB = Next->getParent(); |
846 | E = MBB->end(); |
847 | } |
848 | } |
849 | } |
850 | |
851 | optimizeEndCf(); |
852 | |
853 | if (LIS) { |
854 | for (Register Reg : RecomputeRegs) { |
855 | LIS->removeInterval(Reg); |
856 | LIS->createAndComputeVirtRegInterval(Reg); |
857 | } |
858 | } |
859 | |
860 | RecomputeRegs.clear(); |
861 | LoweredEndCf.clear(); |
862 | LoweredIf.clear(); |
863 | KillBlocks.clear(); |
864 | |
865 | return Changed; |
866 | } |
867 | |