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_CBRANCH_EXECZ label1 // Use our branch optimization |
45 | /// // instruction again. |
46 | /// %vgpr0 = V_SUB_F32 %vgpr0, %vgpr // Do the ELSE block |
47 | /// label1: |
48 | /// %exec = S_OR_B64 %exec, %sgpr0 // Re-enable saved exec mask bits |
49 | //===----------------------------------------------------------------------===// |
50 | |
51 | #include "SILowerControlFlow.h" |
52 | #include "AMDGPU.h" |
53 | #include "GCNSubtarget.h" |
54 | #include "MCTargetDesc/AMDGPUMCTargetDesc.h" |
55 | #include "llvm/ADT/SmallSet.h" |
56 | #include "llvm/CodeGen/LiveIntervals.h" |
57 | #include "llvm/CodeGen/LiveVariables.h" |
58 | #include "llvm/CodeGen/MachineDominators.h" |
59 | #include "llvm/CodeGen/MachineFunctionPass.h" |
60 | #include "llvm/Target/TargetMachine.h" |
61 | |
62 | using namespace llvm; |
63 | |
64 | #define DEBUG_TYPE "si-lower-control-flow" |
65 | |
66 | static cl::opt<bool> |
67 | RemoveRedundantEndcf("amdgpu-remove-redundant-endcf" , |
68 | cl::init(Val: true), cl::ReallyHidden); |
69 | |
70 | namespace { |
71 | |
72 | class SILowerControlFlow { |
73 | private: |
74 | const SIRegisterInfo *TRI = nullptr; |
75 | const SIInstrInfo *TII = nullptr; |
76 | LiveIntervals *LIS = nullptr; |
77 | LiveVariables *LV = nullptr; |
78 | MachineDominatorTree *MDT = nullptr; |
79 | MachineRegisterInfo *MRI = nullptr; |
80 | SetVector<MachineInstr*> LoweredEndCf; |
81 | DenseSet<Register> LoweredIf; |
82 | SmallSet<MachineBasicBlock *, 4> KillBlocks; |
83 | SmallSet<Register, 8> RecomputeRegs; |
84 | |
85 | const TargetRegisterClass *BoolRC = nullptr; |
86 | unsigned AndOpc; |
87 | unsigned OrOpc; |
88 | unsigned XorOpc; |
89 | unsigned MovTermOpc; |
90 | unsigned Andn2TermOpc; |
91 | unsigned XorTermrOpc; |
92 | unsigned OrTermrOpc; |
93 | unsigned OrSaveExecOpc; |
94 | unsigned Exec; |
95 | |
96 | bool EnableOptimizeEndCf = false; |
97 | |
98 | bool hasKill(const MachineBasicBlock *Begin, const MachineBasicBlock *End); |
99 | |
100 | void emitIf(MachineInstr &MI); |
101 | void emitElse(MachineInstr &MI); |
102 | void emitIfBreak(MachineInstr &MI); |
103 | void emitLoop(MachineInstr &MI); |
104 | |
105 | MachineBasicBlock *emitEndCf(MachineInstr &MI); |
106 | |
107 | void findMaskOperands(MachineInstr &MI, unsigned OpNo, |
108 | SmallVectorImpl<MachineOperand> &Src) const; |
109 | |
110 | void combineMasks(MachineInstr &MI); |
111 | |
112 | bool removeMBBifRedundant(MachineBasicBlock &MBB); |
113 | |
114 | MachineBasicBlock *process(MachineInstr &MI); |
115 | |
116 | // Skip to the next instruction, ignoring debug instructions, and trivial |
117 | // block boundaries (blocks that have one (typically fallthrough) successor, |
118 | // and the successor has one predecessor. |
119 | MachineBasicBlock::iterator |
120 | skipIgnoreExecInstsTrivialSucc(MachineBasicBlock &MBB, |
121 | MachineBasicBlock::iterator It) const; |
122 | |
123 | /// Find the insertion point for a new conditional branch. |
124 | MachineBasicBlock::iterator |
125 | skipToUncondBrOrEnd(MachineBasicBlock &MBB, |
126 | MachineBasicBlock::iterator I) const { |
127 | assert(I->isTerminator()); |
128 | |
129 | // FIXME: What if we had multiple pre-existing conditional branches? |
130 | MachineBasicBlock::iterator End = MBB.end(); |
131 | while (I != End && !I->isUnconditionalBranch()) |
132 | ++I; |
133 | return I; |
134 | } |
135 | |
136 | // Remove redundant SI_END_CF instructions. |
137 | void optimizeEndCf(); |
138 | |
139 | public: |
140 | SILowerControlFlow(LiveIntervals *LIS, LiveVariables *LV, |
141 | MachineDominatorTree *MDT) |
142 | : LIS(LIS), LV(LV), MDT(MDT) {} |
143 | bool run(MachineFunction &MF); |
144 | }; |
145 | |
146 | class SILowerControlFlowLegacy : public MachineFunctionPass { |
147 | public: |
148 | static char ID; |
149 | |
150 | SILowerControlFlowLegacy() : MachineFunctionPass(ID) {} |
151 | |
152 | bool runOnMachineFunction(MachineFunction &MF) override; |
153 | |
154 | StringRef getPassName() const override { |
155 | return "SI Lower control flow pseudo instructions" ; |
156 | } |
157 | |
158 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
159 | AU.addUsedIfAvailable<LiveIntervalsWrapperPass>(); |
160 | // Should preserve the same set that TwoAddressInstructions does. |
161 | AU.addPreserved<MachineDominatorTreeWrapperPass>(); |
162 | AU.addPreserved<SlotIndexesWrapperPass>(); |
163 | AU.addPreserved<LiveIntervalsWrapperPass>(); |
164 | AU.addPreserved<LiveVariablesWrapperPass>(); |
165 | MachineFunctionPass::getAnalysisUsage(AU); |
166 | } |
167 | }; |
168 | |
169 | } // end anonymous namespace |
170 | |
171 | char SILowerControlFlowLegacy::ID = 0; |
172 | |
173 | INITIALIZE_PASS(SILowerControlFlowLegacy, DEBUG_TYPE, "SI lower control flow" , |
174 | false, false) |
175 | |
176 | static void setImpSCCDefDead(MachineInstr &MI, bool IsDead) { |
177 | MachineOperand &ImpDefSCC = MI.getOperand(i: 3); |
178 | assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef()); |
179 | |
180 | ImpDefSCC.setIsDead(IsDead); |
181 | } |
182 | |
183 | char &llvm::SILowerControlFlowLegacyID = SILowerControlFlowLegacy::ID; |
184 | |
185 | bool SILowerControlFlow::hasKill(const MachineBasicBlock *Begin, |
186 | const MachineBasicBlock *End) { |
187 | DenseSet<const MachineBasicBlock*> Visited; |
188 | SmallVector<MachineBasicBlock *, 4> Worklist(Begin->successors()); |
189 | |
190 | while (!Worklist.empty()) { |
191 | MachineBasicBlock *MBB = Worklist.pop_back_val(); |
192 | |
193 | if (MBB == End || !Visited.insert(V: MBB).second) |
194 | continue; |
195 | if (KillBlocks.contains(Ptr: MBB)) |
196 | return true; |
197 | |
198 | Worklist.append(in_start: MBB->succ_begin(), in_end: MBB->succ_end()); |
199 | } |
200 | |
201 | return false; |
202 | } |
203 | |
204 | static bool isSimpleIf(const MachineInstr &MI, const MachineRegisterInfo *MRI) { |
205 | Register SaveExecReg = MI.getOperand(i: 0).getReg(); |
206 | auto U = MRI->use_instr_nodbg_begin(RegNo: SaveExecReg); |
207 | |
208 | if (U == MRI->use_instr_nodbg_end() || |
209 | std::next(x: U) != MRI->use_instr_nodbg_end() || |
210 | U->getOpcode() != AMDGPU::SI_END_CF) |
211 | return false; |
212 | |
213 | return true; |
214 | } |
215 | |
216 | void SILowerControlFlow::emitIf(MachineInstr &MI) { |
217 | MachineBasicBlock &MBB = *MI.getParent(); |
218 | const DebugLoc &DL = MI.getDebugLoc(); |
219 | MachineBasicBlock::iterator I(&MI); |
220 | Register SaveExecReg = MI.getOperand(i: 0).getReg(); |
221 | MachineOperand& Cond = MI.getOperand(i: 1); |
222 | assert(Cond.getSubReg() == AMDGPU::NoSubRegister); |
223 | |
224 | MachineOperand &ImpDefSCC = MI.getOperand(i: 4); |
225 | assert(ImpDefSCC.getReg() == AMDGPU::SCC && ImpDefSCC.isDef()); |
226 | |
227 | // If there is only one use of save exec register and that use is SI_END_CF, |
228 | // we can optimize SI_IF by returning the full saved exec mask instead of |
229 | // just cleared bits. |
230 | bool SimpleIf = isSimpleIf(MI, MRI); |
231 | |
232 | if (SimpleIf) { |
233 | // Check for SI_KILL_*_TERMINATOR on path from if to endif. |
234 | // if there is any such terminator simplifications are not safe. |
235 | auto UseMI = MRI->use_instr_nodbg_begin(RegNo: SaveExecReg); |
236 | SimpleIf = !hasKill(Begin: MI.getParent(), End: UseMI->getParent()); |
237 | } |
238 | |
239 | // Add an implicit def of exec to discourage scheduling VALU after this which |
240 | // will interfere with trying to form s_and_saveexec_b64 later. |
241 | Register CopyReg = SimpleIf ? SaveExecReg |
242 | : MRI->createVirtualRegister(RegClass: BoolRC); |
243 | MachineInstr *CopyExec = |
244 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::COPY), DestReg: CopyReg) |
245 | .addReg(RegNo: Exec) |
246 | .addReg(RegNo: Exec, flags: RegState::ImplicitDefine); |
247 | LoweredIf.insert(V: CopyReg); |
248 | |
249 | Register Tmp = MRI->createVirtualRegister(RegClass: BoolRC); |
250 | |
251 | MachineInstr *And = |
252 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: Tmp) |
253 | .addReg(RegNo: CopyReg) |
254 | .add(MO: Cond); |
255 | if (LV) |
256 | LV->replaceKillInstruction(Reg: Cond.getReg(), OldMI&: MI, NewMI&: *And); |
257 | |
258 | setImpSCCDefDead(MI&: *And, IsDead: true); |
259 | |
260 | MachineInstr *Xor = nullptr; |
261 | if (!SimpleIf) { |
262 | Xor = |
263 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: XorOpc), DestReg: SaveExecReg) |
264 | .addReg(RegNo: Tmp) |
265 | .addReg(RegNo: CopyReg); |
266 | setImpSCCDefDead(MI&: *Xor, IsDead: ImpDefSCC.isDead()); |
267 | } |
268 | |
269 | // Use a copy that is a terminator to get correct spill code placement it with |
270 | // fast regalloc. |
271 | MachineInstr *SetExec = |
272 | BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: MovTermOpc), DestReg: Exec) |
273 | .addReg(RegNo: Tmp, flags: RegState::Kill); |
274 | if (LV) |
275 | LV->getVarInfo(Reg: Tmp).Kills.push_back(x: SetExec); |
276 | |
277 | // Skip ahead to the unconditional branch in case there are other terminators |
278 | // present. |
279 | I = skipToUncondBrOrEnd(MBB, I); |
280 | |
281 | // Insert the S_CBRANCH_EXECZ instruction which will be optimized later |
282 | // during SIPreEmitPeephole. |
283 | MachineInstr *NewBr = BuildMI(BB&: MBB, I, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECZ)) |
284 | .add(MO: MI.getOperand(i: 2)); |
285 | |
286 | if (!LIS) { |
287 | MI.eraseFromParent(); |
288 | return; |
289 | } |
290 | |
291 | LIS->InsertMachineInstrInMaps(MI&: *CopyExec); |
292 | |
293 | // Replace with and so we don't need to fix the live interval for condition |
294 | // register. |
295 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *And); |
296 | |
297 | if (!SimpleIf) |
298 | LIS->InsertMachineInstrInMaps(MI&: *Xor); |
299 | LIS->InsertMachineInstrInMaps(MI&: *SetExec); |
300 | LIS->InsertMachineInstrInMaps(MI&: *NewBr); |
301 | |
302 | LIS->removeAllRegUnitsForPhysReg(Reg: AMDGPU::EXEC); |
303 | MI.eraseFromParent(); |
304 | |
305 | // FIXME: Is there a better way of adjusting the liveness? It shouldn't be |
306 | // hard to add another def here but I'm not sure how to correctly update the |
307 | // valno. |
308 | RecomputeRegs.insert(V: SaveExecReg); |
309 | LIS->createAndComputeVirtRegInterval(Reg: Tmp); |
310 | if (!SimpleIf) |
311 | LIS->createAndComputeVirtRegInterval(Reg: CopyReg); |
312 | } |
313 | |
314 | void SILowerControlFlow::emitElse(MachineInstr &MI) { |
315 | MachineBasicBlock &MBB = *MI.getParent(); |
316 | const DebugLoc &DL = MI.getDebugLoc(); |
317 | |
318 | Register DstReg = MI.getOperand(i: 0).getReg(); |
319 | Register SrcReg = MI.getOperand(i: 1).getReg(); |
320 | |
321 | MachineBasicBlock::iterator Start = MBB.begin(); |
322 | |
323 | // This must be inserted before phis and any spill code inserted before the |
324 | // else. |
325 | Register SaveReg = MRI->createVirtualRegister(RegClass: BoolRC); |
326 | MachineInstr *OrSaveExec = |
327 | BuildMI(BB&: MBB, I: Start, MIMD: DL, MCID: TII->get(Opcode: OrSaveExecOpc), DestReg: SaveReg) |
328 | .add(MO: MI.getOperand(i: 1)); // Saved EXEC |
329 | if (LV) |
330 | LV->replaceKillInstruction(Reg: SrcReg, OldMI&: MI, NewMI&: *OrSaveExec); |
331 | |
332 | MachineBasicBlock *DestBB = MI.getOperand(i: 2).getMBB(); |
333 | |
334 | MachineBasicBlock::iterator ElsePt(MI); |
335 | |
336 | // This accounts for any modification of the EXEC mask within the block and |
337 | // can be optimized out pre-RA when not required. |
338 | MachineInstr *And = BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: DstReg) |
339 | .addReg(RegNo: Exec) |
340 | .addReg(RegNo: SaveReg); |
341 | |
342 | MachineInstr *Xor = |
343 | BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: XorTermrOpc), DestReg: Exec) |
344 | .addReg(RegNo: Exec) |
345 | .addReg(RegNo: DstReg); |
346 | |
347 | // Skip ahead to the unconditional branch in case there are other terminators |
348 | // present. |
349 | ElsePt = skipToUncondBrOrEnd(MBB, I: ElsePt); |
350 | |
351 | MachineInstr *Branch = |
352 | BuildMI(BB&: MBB, I: ElsePt, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECZ)) |
353 | .addMBB(MBB: DestBB); |
354 | |
355 | if (!LIS) { |
356 | MI.eraseFromParent(); |
357 | return; |
358 | } |
359 | |
360 | LIS->RemoveMachineInstrFromMaps(MI); |
361 | MI.eraseFromParent(); |
362 | |
363 | LIS->InsertMachineInstrInMaps(MI&: *OrSaveExec); |
364 | LIS->InsertMachineInstrInMaps(MI&: *And); |
365 | |
366 | LIS->InsertMachineInstrInMaps(MI&: *Xor); |
367 | LIS->InsertMachineInstrInMaps(MI&: *Branch); |
368 | |
369 | RecomputeRegs.insert(V: SrcReg); |
370 | RecomputeRegs.insert(V: DstReg); |
371 | LIS->createAndComputeVirtRegInterval(Reg: SaveReg); |
372 | |
373 | // Let this be recomputed. |
374 | LIS->removeAllRegUnitsForPhysReg(Reg: AMDGPU::EXEC); |
375 | } |
376 | |
377 | void SILowerControlFlow::emitIfBreak(MachineInstr &MI) { |
378 | MachineBasicBlock &MBB = *MI.getParent(); |
379 | const DebugLoc &DL = MI.getDebugLoc(); |
380 | auto Dst = MI.getOperand(i: 0).getReg(); |
381 | |
382 | // Skip ANDing with exec if the break condition is already masked by exec |
383 | // because it is a V_CMP in the same basic block. (We know the break |
384 | // condition operand was an i1 in IR, so if it is a VALU instruction it must |
385 | // be one with a carry-out.) |
386 | bool SkipAnding = false; |
387 | if (MI.getOperand(i: 1).isReg()) { |
388 | if (MachineInstr *Def = MRI->getUniqueVRegDef(Reg: MI.getOperand(i: 1).getReg())) { |
389 | SkipAnding = Def->getParent() == MI.getParent() |
390 | && SIInstrInfo::isVALU(MI: *Def); |
391 | } |
392 | } |
393 | |
394 | // AND the break condition operand with exec, then OR that into the "loop |
395 | // exit" mask. |
396 | MachineInstr *And = nullptr, *Or = nullptr; |
397 | Register AndReg; |
398 | if (!SkipAnding) { |
399 | AndReg = MRI->createVirtualRegister(RegClass: BoolRC); |
400 | And = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: AndOpc), DestReg: AndReg) |
401 | .addReg(RegNo: Exec) |
402 | .add(MO: MI.getOperand(i: 1)); |
403 | if (LV) |
404 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 1).getReg(), OldMI&: MI, NewMI&: *And); |
405 | Or = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: OrOpc), DestReg: Dst) |
406 | .addReg(RegNo: AndReg) |
407 | .add(MO: MI.getOperand(i: 2)); |
408 | } else { |
409 | Or = BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: OrOpc), DestReg: Dst) |
410 | .add(MO: MI.getOperand(i: 1)) |
411 | .add(MO: MI.getOperand(i: 2)); |
412 | if (LV) |
413 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 1).getReg(), OldMI&: MI, NewMI&: *Or); |
414 | } |
415 | if (LV) |
416 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 2).getReg(), OldMI&: MI, NewMI&: *Or); |
417 | |
418 | if (LIS) { |
419 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *Or); |
420 | if (And) { |
421 | // Read of original operand 1 is on And now not Or. |
422 | RecomputeRegs.insert(V: And->getOperand(i: 2).getReg()); |
423 | LIS->InsertMachineInstrInMaps(MI&: *And); |
424 | LIS->createAndComputeVirtRegInterval(Reg: AndReg); |
425 | } |
426 | } |
427 | |
428 | MI.eraseFromParent(); |
429 | } |
430 | |
431 | void SILowerControlFlow::emitLoop(MachineInstr &MI) { |
432 | MachineBasicBlock &MBB = *MI.getParent(); |
433 | const DebugLoc &DL = MI.getDebugLoc(); |
434 | |
435 | MachineInstr *AndN2 = |
436 | BuildMI(BB&: MBB, I: &MI, MIMD: DL, MCID: TII->get(Opcode: Andn2TermOpc), DestReg: Exec) |
437 | .addReg(RegNo: Exec) |
438 | .add(MO: MI.getOperand(i: 0)); |
439 | if (LV) |
440 | LV->replaceKillInstruction(Reg: MI.getOperand(i: 0).getReg(), OldMI&: MI, NewMI&: *AndN2); |
441 | |
442 | auto BranchPt = skipToUncondBrOrEnd(MBB, I: MI.getIterator()); |
443 | MachineInstr *Branch = |
444 | BuildMI(BB&: MBB, I: BranchPt, MIMD: DL, MCID: TII->get(Opcode: AMDGPU::S_CBRANCH_EXECNZ)) |
445 | .add(MO: MI.getOperand(i: 1)); |
446 | |
447 | if (LIS) { |
448 | RecomputeRegs.insert(V: MI.getOperand(i: 0).getReg()); |
449 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *AndN2); |
450 | LIS->InsertMachineInstrInMaps(MI&: *Branch); |
451 | } |
452 | |
453 | MI.eraseFromParent(); |
454 | } |
455 | |
456 | MachineBasicBlock::iterator |
457 | SILowerControlFlow::skipIgnoreExecInstsTrivialSucc( |
458 | MachineBasicBlock &MBB, MachineBasicBlock::iterator It) const { |
459 | |
460 | SmallSet<const MachineBasicBlock *, 4> Visited; |
461 | MachineBasicBlock *B = &MBB; |
462 | do { |
463 | if (!Visited.insert(Ptr: B).second) |
464 | return MBB.end(); |
465 | |
466 | auto E = B->end(); |
467 | for ( ; It != E; ++It) { |
468 | if (TII->mayReadEXEC(MRI: *MRI, MI: *It)) |
469 | break; |
470 | } |
471 | |
472 | if (It != E) |
473 | return It; |
474 | |
475 | if (B->succ_size() != 1) |
476 | return MBB.end(); |
477 | |
478 | // If there is one trivial successor, advance to the next block. |
479 | MachineBasicBlock *Succ = *B->succ_begin(); |
480 | |
481 | It = Succ->begin(); |
482 | B = Succ; |
483 | } while (true); |
484 | } |
485 | |
486 | MachineBasicBlock *SILowerControlFlow::emitEndCf(MachineInstr &MI) { |
487 | MachineBasicBlock &MBB = *MI.getParent(); |
488 | const DebugLoc &DL = MI.getDebugLoc(); |
489 | |
490 | MachineBasicBlock::iterator InsPt = MBB.begin(); |
491 | |
492 | // If we have instructions that aren't prolog instructions, split the block |
493 | // and emit a terminator instruction. This ensures correct spill placement. |
494 | // FIXME: We should unconditionally split the block here. |
495 | bool NeedBlockSplit = false; |
496 | Register DataReg = MI.getOperand(i: 0).getReg(); |
497 | for (MachineBasicBlock::iterator I = InsPt, E = MI.getIterator(); |
498 | I != E; ++I) { |
499 | if (I->modifiesRegister(Reg: DataReg, TRI)) { |
500 | NeedBlockSplit = true; |
501 | break; |
502 | } |
503 | } |
504 | |
505 | unsigned Opcode = OrOpc; |
506 | MachineBasicBlock *SplitBB = &MBB; |
507 | if (NeedBlockSplit) { |
508 | SplitBB = MBB.splitAt(SplitInst&: MI, /*UpdateLiveIns*/true, LIS); |
509 | if (MDT && SplitBB != &MBB) { |
510 | MachineDomTreeNode *MBBNode = (*MDT)[&MBB]; |
511 | SmallVector<MachineDomTreeNode *> Children(MBBNode->begin(), |
512 | MBBNode->end()); |
513 | MachineDomTreeNode *SplitBBNode = MDT->addNewBlock(BB: SplitBB, DomBB: &MBB); |
514 | for (MachineDomTreeNode *Child : Children) |
515 | MDT->changeImmediateDominator(N: Child, NewIDom: SplitBBNode); |
516 | } |
517 | Opcode = OrTermrOpc; |
518 | InsPt = MI; |
519 | } |
520 | |
521 | MachineInstr *NewMI = |
522 | BuildMI(BB&: MBB, I: InsPt, MIMD: DL, MCID: TII->get(Opcode), DestReg: Exec) |
523 | .addReg(RegNo: Exec) |
524 | .add(MO: MI.getOperand(i: 0)); |
525 | if (LV) { |
526 | LV->replaceKillInstruction(Reg: DataReg, OldMI&: MI, NewMI&: *NewMI); |
527 | |
528 | if (SplitBB != &MBB) { |
529 | // Track the set of registers defined in the original block so we don't |
530 | // accidentally add the original block to AliveBlocks. AliveBlocks only |
531 | // includes blocks which are live through, which excludes live outs and |
532 | // local defs. |
533 | DenseSet<Register> DefInOrigBlock; |
534 | |
535 | for (MachineBasicBlock *BlockPiece : {&MBB, SplitBB}) { |
536 | for (MachineInstr &X : *BlockPiece) { |
537 | for (MachineOperand &Op : X.all_defs()) { |
538 | if (Op.getReg().isVirtual()) |
539 | DefInOrigBlock.insert(V: Op.getReg()); |
540 | } |
541 | } |
542 | } |
543 | |
544 | for (unsigned i = 0, e = MRI->getNumVirtRegs(); i != e; ++i) { |
545 | Register Reg = Register::index2VirtReg(Index: i); |
546 | LiveVariables::VarInfo &VI = LV->getVarInfo(Reg); |
547 | |
548 | if (VI.AliveBlocks.test(Idx: MBB.getNumber())) |
549 | VI.AliveBlocks.set(SplitBB->getNumber()); |
550 | else { |
551 | for (MachineInstr *Kill : VI.Kills) { |
552 | if (Kill->getParent() == SplitBB && !DefInOrigBlock.contains(V: Reg)) |
553 | VI.AliveBlocks.set(MBB.getNumber()); |
554 | } |
555 | } |
556 | } |
557 | } |
558 | } |
559 | |
560 | LoweredEndCf.insert(X: NewMI); |
561 | |
562 | if (LIS) |
563 | LIS->ReplaceMachineInstrInMaps(MI, NewMI&: *NewMI); |
564 | |
565 | MI.eraseFromParent(); |
566 | |
567 | if (LIS) |
568 | LIS->handleMove(MI&: *NewMI); |
569 | return SplitBB; |
570 | } |
571 | |
572 | // Returns replace operands for a logical operation, either single result |
573 | // for exec or two operands if source was another equivalent operation. |
574 | void SILowerControlFlow::findMaskOperands(MachineInstr &MI, unsigned OpNo, |
575 | SmallVectorImpl<MachineOperand> &Src) const { |
576 | MachineOperand &Op = MI.getOperand(i: OpNo); |
577 | if (!Op.isReg() || !Op.getReg().isVirtual()) { |
578 | Src.push_back(Elt: Op); |
579 | return; |
580 | } |
581 | |
582 | MachineInstr *Def = MRI->getUniqueVRegDef(Reg: Op.getReg()); |
583 | if (!Def || Def->getParent() != MI.getParent() || |
584 | !(Def->isFullCopy() || (Def->getOpcode() == MI.getOpcode()))) |
585 | return; |
586 | |
587 | // Make sure we do not modify exec between def and use. |
588 | // A copy with implicitly defined exec inserted earlier is an exclusion, it |
589 | // does not really modify exec. |
590 | for (auto I = Def->getIterator(); I != MI.getIterator(); ++I) |
591 | if (I->modifiesRegister(Reg: AMDGPU::EXEC, TRI) && |
592 | !(I->isCopy() && I->getOperand(i: 0).getReg() != Exec)) |
593 | return; |
594 | |
595 | for (const auto &SrcOp : Def->explicit_operands()) |
596 | if (SrcOp.isReg() && SrcOp.isUse() && |
597 | (SrcOp.getReg().isVirtual() || SrcOp.getReg() == Exec)) |
598 | Src.push_back(Elt: SrcOp); |
599 | } |
600 | |
601 | // Search and combine pairs of equivalent instructions, like |
602 | // S_AND_B64 x, (S_AND_B64 x, y) => S_AND_B64 x, y |
603 | // S_OR_B64 x, (S_OR_B64 x, y) => S_OR_B64 x, y |
604 | // One of the operands is exec mask. |
605 | void SILowerControlFlow::combineMasks(MachineInstr &MI) { |
606 | assert(MI.getNumExplicitOperands() == 3); |
607 | SmallVector<MachineOperand, 4> Ops; |
608 | unsigned OpToReplace = 1; |
609 | findMaskOperands(MI, OpNo: 1, Src&: Ops); |
610 | if (Ops.size() == 1) OpToReplace = 2; // First operand can be exec or its copy |
611 | findMaskOperands(MI, OpNo: 2, Src&: Ops); |
612 | if (Ops.size() != 3) return; |
613 | |
614 | unsigned UniqueOpndIdx; |
615 | if (Ops[0].isIdenticalTo(Other: Ops[1])) UniqueOpndIdx = 2; |
616 | else if (Ops[0].isIdenticalTo(Other: Ops[2])) UniqueOpndIdx = 1; |
617 | else if (Ops[1].isIdenticalTo(Other: Ops[2])) UniqueOpndIdx = 1; |
618 | else return; |
619 | |
620 | Register Reg = MI.getOperand(i: OpToReplace).getReg(); |
621 | MI.removeOperand(OpNo: OpToReplace); |
622 | MI.addOperand(Op: Ops[UniqueOpndIdx]); |
623 | if (MRI->use_empty(RegNo: Reg)) |
624 | MRI->getUniqueVRegDef(Reg)->eraseFromParent(); |
625 | } |
626 | |
627 | void SILowerControlFlow::optimizeEndCf() { |
628 | // If the only instruction immediately following this END_CF is another |
629 | // END_CF in the only successor we can avoid emitting exec mask restore here. |
630 | if (!EnableOptimizeEndCf) |
631 | return; |
632 | |
633 | for (MachineInstr *MI : reverse(C&: LoweredEndCf)) { |
634 | MachineBasicBlock &MBB = *MI->getParent(); |
635 | auto Next = |
636 | skipIgnoreExecInstsTrivialSucc(MBB, It: std::next(x: MI->getIterator())); |
637 | if (Next == MBB.end() || !LoweredEndCf.count(key: &*Next)) |
638 | continue; |
639 | // Only skip inner END_CF if outer ENDCF belongs to SI_IF. |
640 | // If that belongs to SI_ELSE then saved mask has an inverted value. |
641 | Register SavedExec |
642 | = TII->getNamedOperand(MI&: *Next, OperandName: AMDGPU::OpName::src1)->getReg(); |
643 | assert(SavedExec.isVirtual() && "Expected saved exec to be src1!" ); |
644 | |
645 | const MachineInstr *Def = MRI->getUniqueVRegDef(Reg: SavedExec); |
646 | if (Def && LoweredIf.count(V: SavedExec)) { |
647 | LLVM_DEBUG(dbgs() << "Skip redundant " ; MI->dump()); |
648 | if (LIS) |
649 | LIS->RemoveMachineInstrFromMaps(MI&: *MI); |
650 | Register Reg; |
651 | if (LV) |
652 | Reg = TII->getNamedOperand(MI&: *MI, OperandName: AMDGPU::OpName::src1)->getReg(); |
653 | MI->eraseFromParent(); |
654 | if (LV) |
655 | LV->recomputeForSingleDefVirtReg(Reg); |
656 | removeMBBifRedundant(MBB); |
657 | } |
658 | } |
659 | } |
660 | |
661 | MachineBasicBlock *SILowerControlFlow::process(MachineInstr &MI) { |
662 | MachineBasicBlock &MBB = *MI.getParent(); |
663 | MachineBasicBlock::iterator I(MI); |
664 | MachineInstr *Prev = (I != MBB.begin()) ? &*(std::prev(x: I)) : nullptr; |
665 | |
666 | MachineBasicBlock *SplitBB = &MBB; |
667 | |
668 | switch (MI.getOpcode()) { |
669 | case AMDGPU::SI_IF: |
670 | emitIf(MI); |
671 | break; |
672 | |
673 | case AMDGPU::SI_ELSE: |
674 | emitElse(MI); |
675 | break; |
676 | |
677 | case AMDGPU::SI_IF_BREAK: |
678 | emitIfBreak(MI); |
679 | break; |
680 | |
681 | case AMDGPU::SI_LOOP: |
682 | emitLoop(MI); |
683 | break; |
684 | |
685 | case AMDGPU::SI_WATERFALL_LOOP: |
686 | MI.setDesc(TII->get(Opcode: AMDGPU::S_CBRANCH_EXECNZ)); |
687 | break; |
688 | |
689 | case AMDGPU::SI_END_CF: |
690 | SplitBB = emitEndCf(MI); |
691 | break; |
692 | |
693 | default: |
694 | assert(false && "Attempt to process unsupported instruction" ); |
695 | break; |
696 | } |
697 | |
698 | MachineBasicBlock::iterator Next; |
699 | for (I = Prev ? Prev->getIterator() : MBB.begin(); I != MBB.end(); I = Next) { |
700 | Next = std::next(x: I); |
701 | MachineInstr &MaskMI = *I; |
702 | switch (MaskMI.getOpcode()) { |
703 | case AMDGPU::S_AND_B64: |
704 | case AMDGPU::S_OR_B64: |
705 | case AMDGPU::S_AND_B32: |
706 | case AMDGPU::S_OR_B32: |
707 | // Cleanup bit manipulations on exec mask |
708 | combineMasks(MI&: MaskMI); |
709 | break; |
710 | default: |
711 | I = MBB.end(); |
712 | break; |
713 | } |
714 | } |
715 | |
716 | return SplitBB; |
717 | } |
718 | |
719 | bool SILowerControlFlow::removeMBBifRedundant(MachineBasicBlock &MBB) { |
720 | for (auto &I : MBB.instrs()) { |
721 | if (!I.isDebugInstr() && !I.isUnconditionalBranch()) |
722 | return false; |
723 | } |
724 | |
725 | assert(MBB.succ_size() == 1 && "MBB has more than one successor" ); |
726 | |
727 | MachineBasicBlock *Succ = *MBB.succ_begin(); |
728 | MachineBasicBlock *FallThrough = nullptr; |
729 | |
730 | while (!MBB.predecessors().empty()) { |
731 | MachineBasicBlock *P = *MBB.pred_begin(); |
732 | if (P->getFallThrough(JumpToFallThrough: false) == &MBB) |
733 | FallThrough = P; |
734 | P->ReplaceUsesOfBlockWith(Old: &MBB, New: Succ); |
735 | } |
736 | MBB.removeSuccessor(Succ); |
737 | if (LIS) { |
738 | for (auto &I : MBB.instrs()) |
739 | LIS->RemoveMachineInstrFromMaps(MI&: I); |
740 | } |
741 | if (MDT) { |
742 | // If Succ, the single successor of MBB, is dominated by MBB, MDT needs |
743 | // updating by changing Succ's idom to the one of MBB; otherwise, MBB must |
744 | // be a leaf node in MDT and could be erased directly. |
745 | if (MDT->dominates(A: &MBB, B: Succ)) |
746 | MDT->changeImmediateDominator(N: MDT->getNode(BB: Succ), |
747 | NewIDom: MDT->getNode(BB: &MBB)->getIDom()); |
748 | MDT->eraseNode(BB: &MBB); |
749 | } |
750 | MBB.clear(); |
751 | MBB.eraseFromParent(); |
752 | if (FallThrough && !FallThrough->isLayoutSuccessor(MBB: Succ)) { |
753 | // Note: we cannot update block layout and preserve live intervals; |
754 | // hence we must insert a branch. |
755 | MachineInstr *BranchMI = BuildMI(BB&: *FallThrough, I: FallThrough->end(), |
756 | MIMD: FallThrough->findBranchDebugLoc(), MCID: TII->get(Opcode: AMDGPU::S_BRANCH)) |
757 | .addMBB(MBB: Succ); |
758 | if (LIS) |
759 | LIS->InsertMachineInstrInMaps(MI&: *BranchMI); |
760 | } |
761 | |
762 | return true; |
763 | } |
764 | |
765 | bool SILowerControlFlow::run(MachineFunction &MF) { |
766 | const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>(); |
767 | TII = ST.getInstrInfo(); |
768 | TRI = &TII->getRegisterInfo(); |
769 | EnableOptimizeEndCf = RemoveRedundantEndcf && |
770 | MF.getTarget().getOptLevel() > CodeGenOptLevel::None; |
771 | |
772 | MRI = &MF.getRegInfo(); |
773 | BoolRC = TRI->getBoolRC(); |
774 | |
775 | if (ST.isWave32()) { |
776 | AndOpc = AMDGPU::S_AND_B32; |
777 | OrOpc = AMDGPU::S_OR_B32; |
778 | XorOpc = AMDGPU::S_XOR_B32; |
779 | MovTermOpc = AMDGPU::S_MOV_B32_term; |
780 | Andn2TermOpc = AMDGPU::S_ANDN2_B32_term; |
781 | XorTermrOpc = AMDGPU::S_XOR_B32_term; |
782 | OrTermrOpc = AMDGPU::S_OR_B32_term; |
783 | OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B32; |
784 | Exec = AMDGPU::EXEC_LO; |
785 | } else { |
786 | AndOpc = AMDGPU::S_AND_B64; |
787 | OrOpc = AMDGPU::S_OR_B64; |
788 | XorOpc = AMDGPU::S_XOR_B64; |
789 | MovTermOpc = AMDGPU::S_MOV_B64_term; |
790 | Andn2TermOpc = AMDGPU::S_ANDN2_B64_term; |
791 | XorTermrOpc = AMDGPU::S_XOR_B64_term; |
792 | OrTermrOpc = AMDGPU::S_OR_B64_term; |
793 | OrSaveExecOpc = AMDGPU::S_OR_SAVEEXEC_B64; |
794 | Exec = AMDGPU::EXEC; |
795 | } |
796 | |
797 | // Compute set of blocks with kills |
798 | const bool CanDemote = |
799 | MF.getFunction().getCallingConv() == CallingConv::AMDGPU_PS; |
800 | for (auto &MBB : MF) { |
801 | bool IsKillBlock = false; |
802 | for (auto &Term : MBB.terminators()) { |
803 | if (TII->isKillTerminator(Opcode: Term.getOpcode())) { |
804 | KillBlocks.insert(Ptr: &MBB); |
805 | IsKillBlock = true; |
806 | break; |
807 | } |
808 | } |
809 | if (CanDemote && !IsKillBlock) { |
810 | for (auto &MI : MBB) { |
811 | if (MI.getOpcode() == AMDGPU::SI_DEMOTE_I1) { |
812 | KillBlocks.insert(Ptr: &MBB); |
813 | break; |
814 | } |
815 | } |
816 | } |
817 | } |
818 | |
819 | bool Changed = false; |
820 | MachineFunction::iterator NextBB; |
821 | for (MachineFunction::iterator BI = MF.begin(); |
822 | BI != MF.end(); BI = NextBB) { |
823 | NextBB = std::next(x: BI); |
824 | MachineBasicBlock *MBB = &*BI; |
825 | |
826 | MachineBasicBlock::iterator I, E, Next; |
827 | E = MBB->end(); |
828 | for (I = MBB->begin(); I != E; I = Next) { |
829 | Next = std::next(x: I); |
830 | MachineInstr &MI = *I; |
831 | MachineBasicBlock *SplitMBB = MBB; |
832 | |
833 | switch (MI.getOpcode()) { |
834 | case AMDGPU::SI_IF: |
835 | case AMDGPU::SI_ELSE: |
836 | case AMDGPU::SI_IF_BREAK: |
837 | case AMDGPU::SI_WATERFALL_LOOP: |
838 | case AMDGPU::SI_LOOP: |
839 | case AMDGPU::SI_END_CF: |
840 | SplitMBB = process(MI); |
841 | Changed = true; |
842 | break; |
843 | } |
844 | |
845 | if (SplitMBB != MBB) { |
846 | MBB = Next->getParent(); |
847 | E = MBB->end(); |
848 | } |
849 | } |
850 | } |
851 | |
852 | optimizeEndCf(); |
853 | |
854 | if (LIS) { |
855 | for (Register Reg : RecomputeRegs) { |
856 | LIS->removeInterval(Reg); |
857 | LIS->createAndComputeVirtRegInterval(Reg); |
858 | } |
859 | } |
860 | |
861 | RecomputeRegs.clear(); |
862 | LoweredEndCf.clear(); |
863 | LoweredIf.clear(); |
864 | KillBlocks.clear(); |
865 | |
866 | return Changed; |
867 | } |
868 | |
869 | bool SILowerControlFlowLegacy::runOnMachineFunction(MachineFunction &MF) { |
870 | // This doesn't actually need LiveIntervals, but we can preserve them. |
871 | auto *LISWrapper = getAnalysisIfAvailable<LiveIntervalsWrapperPass>(); |
872 | LiveIntervals *LIS = LISWrapper ? &LISWrapper->getLIS() : nullptr; |
873 | // This doesn't actually need LiveVariables, but we can preserve them. |
874 | auto *LVWrapper = getAnalysisIfAvailable<LiveVariablesWrapperPass>(); |
875 | LiveVariables *LV = LVWrapper ? &LVWrapper->getLV() : nullptr; |
876 | auto *MDTWrapper = getAnalysisIfAvailable<MachineDominatorTreeWrapperPass>(); |
877 | MachineDominatorTree *MDT = MDTWrapper ? &MDTWrapper->getDomTree() : nullptr; |
878 | return SILowerControlFlow(LIS, LV, MDT).run(MF); |
879 | } |
880 | |
881 | PreservedAnalyses |
882 | SILowerControlFlowPass::run(MachineFunction &MF, |
883 | MachineFunctionAnalysisManager &MFAM) { |
884 | LiveIntervals *LIS = MFAM.getCachedResult<LiveIntervalsAnalysis>(IR&: MF); |
885 | LiveVariables *LV = MFAM.getCachedResult<LiveVariablesAnalysis>(IR&: MF); |
886 | MachineDominatorTree *MDT = |
887 | MFAM.getCachedResult<MachineDominatorTreeAnalysis>(IR&: MF); |
888 | |
889 | bool Changed = SILowerControlFlow(LIS, LV, MDT).run(MF); |
890 | if (!Changed) |
891 | return PreservedAnalyses::all(); |
892 | |
893 | auto PA = getMachineFunctionPassPreservedAnalyses(); |
894 | PA.preserve<MachineDominatorTreeAnalysis>(); |
895 | PA.preserve<SlotIndexesAnalysis>(); |
896 | PA.preserve<LiveIntervalsAnalysis>(); |
897 | PA.preserve<LiveVariablesAnalysis>(); |
898 | return PA; |
899 | } |
900 | |