1 //===-- SIInsertSkips.cpp - Use predicates for control flow ---------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// \brief This pass inserts branches on the 0 exec mask over divergent branches
12 /// branches when it's expected that jumping over the untaken control flow will
13 /// be cheaper than having every workitem no-op through it.
14 //
15 //===----------------------------------------------------------------------===//
16 
17 #include "AMDGPU.h"
18 #include "AMDGPUSubtarget.h"
19 #include "SIInstrInfo.h"
20 #include "SIMachineFunctionInfo.h"
21 #include "llvm/ADT/SmallVector.h"
22 #include "llvm/ADT/StringRef.h"
23 #include "llvm/CodeGen/MachineBasicBlock.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFunctionPass.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/CodeGen/MachineOperand.h"
29 #include "llvm/IR/CallingConv.h"
30 #include "llvm/IR/DebugLoc.h"
31 #include "llvm/MC/MCAsmInfo.h"
32 #include "llvm/Pass.h"
33 #include "llvm/Support/CommandLine.h"
34 #include "llvm/Target/TargetMachine.h"
35 #include <cassert>
36 #include <cstdint>
37 #include <iterator>
38 
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "si-insert-skips"
42 
43 static cl::opt<unsigned> SkipThresholdFlag(
44   "amdgpu-skip-threshold",
45   cl::desc("Number of instructions before jumping over divergent control flow"),
46   cl::init(12), cl::Hidden);
47 
48 namespace {
49 
50 class SIInsertSkips : public MachineFunctionPass {
51 private:
52   const SIRegisterInfo *TRI = nullptr;
53   const SIInstrInfo *TII = nullptr;
54   unsigned SkipThreshold = 0;
55 
56   bool shouldSkip(const MachineBasicBlock &From,
57                   const MachineBasicBlock &To) const;
58 
59   bool skipIfDead(MachineInstr &MI, MachineBasicBlock &NextBB);
60 
61   void kill(MachineInstr &MI);
62 
63   MachineBasicBlock *insertSkipBlock(MachineBasicBlock &MBB,
64                                      MachineBasicBlock::iterator I) const;
65 
66   bool skipMaskBranch(MachineInstr &MI, MachineBasicBlock &MBB);
67 
68 public:
69   static char ID;
70 
71   SIInsertSkips() : MachineFunctionPass(ID) {}
72 
73   bool runOnMachineFunction(MachineFunction &MF) override;
74 
75   StringRef getPassName() const override {
76     return "SI insert s_cbranch_execz instructions";
77   }
78 
79   void getAnalysisUsage(AnalysisUsage &AU) const override {
80     MachineFunctionPass::getAnalysisUsage(AU);
81   }
82 };
83 
84 } // end anonymous namespace
85 
86 char SIInsertSkips::ID = 0;
87 
88 INITIALIZE_PASS(SIInsertSkips, DEBUG_TYPE,
89                 "SI insert s_cbranch_execz instructions", false, false)
90 
91 char &llvm::SIInsertSkipsPassID = SIInsertSkips::ID;
92 
93 static bool opcodeEmitsNoInsts(unsigned Opc) {
94   switch (Opc) {
95   case TargetOpcode::IMPLICIT_DEF:
96   case TargetOpcode::KILL:
97   case TargetOpcode::BUNDLE:
98   case TargetOpcode::CFI_INSTRUCTION:
99   case TargetOpcode::EH_LABEL:
100   case TargetOpcode::GC_LABEL:
101   case TargetOpcode::DBG_VALUE:
102     return true;
103   default:
104     return false;
105   }
106 }
107 
108 bool SIInsertSkips::shouldSkip(const MachineBasicBlock &From,
109                                const MachineBasicBlock &To) const {
110   if (From.succ_empty())
111     return false;
112 
113   unsigned NumInstr = 0;
114   const MachineFunction *MF = From.getParent();
115 
116   for (MachineFunction::const_iterator MBBI(&From), ToI(&To), End = MF->end();
117        MBBI != End && MBBI != ToI; ++MBBI) {
118     const MachineBasicBlock &MBB = *MBBI;
119 
120     for (MachineBasicBlock::const_iterator I = MBB.begin(), E = MBB.end();
121          NumInstr < SkipThreshold && I != E; ++I) {
122       if (opcodeEmitsNoInsts(I->getOpcode()))
123         continue;
124 
125       // FIXME: Since this is required for correctness, this should be inserted
126       // during SILowerControlFlow.
127 
128       // When a uniform loop is inside non-uniform control flow, the branch
129       // leaving the loop might be an S_CBRANCH_VCCNZ, which is never taken
130       // when EXEC = 0. We should skip the loop lest it becomes infinite.
131       if (I->getOpcode() == AMDGPU::S_CBRANCH_VCCNZ ||
132           I->getOpcode() == AMDGPU::S_CBRANCH_VCCZ)
133         return true;
134 
135       // V_READFIRSTLANE/V_READLANE destination register may be used as operand
136       // by some SALU instruction. If exec mask is zero vector instruction
137       // defining the register that is used by the scalar one is not executed
138       // and scalar instruction will operate on undefined data. For
139       // V_READFIRSTLANE/V_READLANE we should avoid predicated execution.
140       if ((I->getOpcode() == AMDGPU::V_READFIRSTLANE_B32) ||
141           (I->getOpcode() == AMDGPU::V_READLANE_B32)) {
142         return true;
143       }
144 
145       if (I->isInlineAsm()) {
146         const MCAsmInfo *MAI = MF->getTarget().getMCAsmInfo();
147         const char *AsmStr = I->getOperand(0).getSymbolName();
148 
149         // inlineasm length estimate is number of bytes assuming the longest
150         // instruction.
151         uint64_t MaxAsmSize = TII->getInlineAsmLength(AsmStr, *MAI);
152         NumInstr += MaxAsmSize / MAI->getMaxInstLength();
153       } else {
154         ++NumInstr;
155       }
156 
157       if (NumInstr >= SkipThreshold)
158         return true;
159     }
160   }
161 
162   return false;
163 }
164 
165 bool SIInsertSkips::skipIfDead(MachineInstr &MI, MachineBasicBlock &NextBB) {
166   MachineBasicBlock &MBB = *MI.getParent();
167   MachineFunction *MF = MBB.getParent();
168 
169   if (MF->getFunction().getCallingConv() != CallingConv::AMDGPU_PS ||
170       !shouldSkip(MBB, MBB.getParent()->back()))
171     return false;
172 
173   MachineBasicBlock *SkipBB = insertSkipBlock(MBB, MI.getIterator());
174 
175   const DebugLoc &DL = MI.getDebugLoc();
176 
177   // If the exec mask is non-zero, skip the next two instructions
178   BuildMI(&MBB, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
179     .addMBB(&NextBB);
180 
181   MachineBasicBlock::iterator Insert = SkipBB->begin();
182 
183   // Exec mask is zero: Export to NULL target...
184   BuildMI(*SkipBB, Insert, DL, TII->get(AMDGPU::EXP_DONE))
185     .addImm(0x09) // V_008DFC_SQ_EXP_NULL
186     .addReg(AMDGPU::VGPR0, RegState::Undef)
187     .addReg(AMDGPU::VGPR0, RegState::Undef)
188     .addReg(AMDGPU::VGPR0, RegState::Undef)
189     .addReg(AMDGPU::VGPR0, RegState::Undef)
190     .addImm(1)  // vm
191     .addImm(0)  // compr
192     .addImm(0); // en
193 
194   // ... and terminate wavefront.
195   BuildMI(*SkipBB, Insert, DL, TII->get(AMDGPU::S_ENDPGM));
196 
197   return true;
198 }
199 
200 void SIInsertSkips::kill(MachineInstr &MI) {
201   MachineBasicBlock &MBB = *MI.getParent();
202   DebugLoc DL = MI.getDebugLoc();
203 
204   switch (MI.getOpcode()) {
205   case AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR: {
206     unsigned Opcode = 0;
207 
208     // The opcodes are inverted because the inline immediate has to be
209     // the first operand, e.g. from "x < imm" to "imm > x"
210     switch (MI.getOperand(2).getImm()) {
211     case ISD::SETOEQ:
212     case ISD::SETEQ:
213       Opcode = AMDGPU::V_CMPX_EQ_F32_e32;
214       break;
215     case ISD::SETOGT:
216     case ISD::SETGT:
217       Opcode = AMDGPU::V_CMPX_LT_F32_e32;
218       break;
219     case ISD::SETOGE:
220     case ISD::SETGE:
221       Opcode = AMDGPU::V_CMPX_LE_F32_e32;
222       break;
223     case ISD::SETOLT:
224     case ISD::SETLT:
225       Opcode = AMDGPU::V_CMPX_GT_F32_e32;
226       break;
227     case ISD::SETOLE:
228     case ISD::SETLE:
229       Opcode = AMDGPU::V_CMPX_GE_F32_e32;
230       break;
231     case ISD::SETONE:
232     case ISD::SETNE:
233       Opcode = AMDGPU::V_CMPX_LG_F32_e32;
234       break;
235     case ISD::SETO:
236       Opcode = AMDGPU::V_CMPX_O_F32_e32;
237       break;
238     case ISD::SETUO:
239       Opcode = AMDGPU::V_CMPX_U_F32_e32;
240       break;
241     case ISD::SETUEQ:
242       Opcode = AMDGPU::V_CMPX_NLG_F32_e32;
243       break;
244     case ISD::SETUGT:
245       Opcode = AMDGPU::V_CMPX_NGE_F32_e32;
246       break;
247     case ISD::SETUGE:
248       Opcode = AMDGPU::V_CMPX_NGT_F32_e32;
249       break;
250     case ISD::SETULT:
251       Opcode = AMDGPU::V_CMPX_NLE_F32_e32;
252       break;
253     case ISD::SETULE:
254       Opcode = AMDGPU::V_CMPX_NLT_F32_e32;
255       break;
256     case ISD::SETUNE:
257       Opcode = AMDGPU::V_CMPX_NEQ_F32_e32;
258       break;
259     default:
260       llvm_unreachable("invalid ISD:SET cond code");
261     }
262 
263     // TODO: Allow this:
264     if (!MI.getOperand(0).isReg() ||
265         !TRI->isVGPR(MBB.getParent()->getRegInfo(),
266                      MI.getOperand(0).getReg()))
267       llvm_unreachable("SI_KILL operand should be a VGPR");
268 
269     BuildMI(MBB, &MI, DL, TII->get(Opcode))
270         .add(MI.getOperand(1))
271         .add(MI.getOperand(0));
272     break;
273   }
274   case AMDGPU::SI_KILL_I1_TERMINATOR: {
275     const MachineOperand &Op = MI.getOperand(0);
276     int64_t KillVal = MI.getOperand(1).getImm();
277     assert(KillVal == 0 || KillVal == -1);
278 
279     // Kill all threads if Op0 is an immediate and equal to the Kill value.
280     if (Op.isImm()) {
281       int64_t Imm = Op.getImm();
282       assert(Imm == 0 || Imm == -1);
283 
284       if (Imm == KillVal)
285         BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_MOV_B64), AMDGPU::EXEC)
286           .addImm(0);
287       break;
288     }
289 
290     unsigned Opcode = KillVal ? AMDGPU::S_ANDN2_B64 : AMDGPU::S_AND_B64;
291     BuildMI(MBB, &MI, DL, TII->get(Opcode), AMDGPU::EXEC)
292         .addReg(AMDGPU::EXEC)
293         .add(Op);
294     break;
295   }
296   default:
297     llvm_unreachable("invalid opcode, expected SI_KILL_*_TERMINATOR");
298   }
299 }
300 
301 MachineBasicBlock *SIInsertSkips::insertSkipBlock(
302   MachineBasicBlock &MBB, MachineBasicBlock::iterator I) const {
303   MachineFunction *MF = MBB.getParent();
304 
305   MachineBasicBlock *SkipBB = MF->CreateMachineBasicBlock();
306   MachineFunction::iterator MBBI(MBB);
307   ++MBBI;
308 
309   MF->insert(MBBI, SkipBB);
310   MBB.addSuccessor(SkipBB);
311 
312   return SkipBB;
313 }
314 
315 // Returns true if a branch over the block was inserted.
316 bool SIInsertSkips::skipMaskBranch(MachineInstr &MI,
317                                    MachineBasicBlock &SrcMBB) {
318   MachineBasicBlock *DestBB = MI.getOperand(0).getMBB();
319 
320   if (!shouldSkip(**SrcMBB.succ_begin(), *DestBB))
321     return false;
322 
323   const DebugLoc &DL = MI.getDebugLoc();
324   MachineBasicBlock::iterator InsPt = std::next(MI.getIterator());
325 
326   BuildMI(SrcMBB, InsPt, DL, TII->get(AMDGPU::S_CBRANCH_EXECZ))
327     .addMBB(DestBB);
328 
329   return true;
330 }
331 
332 bool SIInsertSkips::runOnMachineFunction(MachineFunction &MF) {
333   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
334   TII = ST.getInstrInfo();
335   TRI = &TII->getRegisterInfo();
336   SkipThreshold = SkipThresholdFlag;
337 
338   bool HaveKill = false;
339   bool MadeChange = false;
340 
341   // Track depth of exec mask, divergent branches.
342   SmallVector<MachineBasicBlock *, 16> ExecBranchStack;
343 
344   MachineFunction::iterator NextBB;
345 
346   MachineBasicBlock *EmptyMBBAtEnd = nullptr;
347 
348   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
349        BI != BE; BI = NextBB) {
350     NextBB = std::next(BI);
351     MachineBasicBlock &MBB = *BI;
352     bool HaveSkipBlock = false;
353 
354     if (!ExecBranchStack.empty() && ExecBranchStack.back() == &MBB) {
355       // Reached convergence point for last divergent branch.
356       ExecBranchStack.pop_back();
357     }
358 
359     if (HaveKill && ExecBranchStack.empty()) {
360       HaveKill = false;
361 
362       // TODO: Insert skip if exec is 0?
363     }
364 
365     MachineBasicBlock::iterator I, Next;
366     for (I = MBB.begin(); I != MBB.end(); I = Next) {
367       Next = std::next(I);
368 
369       MachineInstr &MI = *I;
370 
371       switch (MI.getOpcode()) {
372       case AMDGPU::SI_MASK_BRANCH:
373         ExecBranchStack.push_back(MI.getOperand(0).getMBB());
374         MadeChange |= skipMaskBranch(MI, MBB);
375         break;
376 
377       case AMDGPU::S_BRANCH:
378         // Optimize out branches to the next block.
379         // FIXME: Shouldn't this be handled by BranchFolding?
380         if (MBB.isLayoutSuccessor(MI.getOperand(0).getMBB())) {
381           MI.eraseFromParent();
382         } else if (HaveSkipBlock) {
383           // Remove the given unconditional branch when a skip block has been
384           // inserted after the current one and let skip the two instructions
385           // performing the kill if the exec mask is non-zero.
386           MI.eraseFromParent();
387         }
388         break;
389 
390       case AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR:
391       case AMDGPU::SI_KILL_I1_TERMINATOR:
392         MadeChange = true;
393         kill(MI);
394 
395         if (ExecBranchStack.empty()) {
396           if (skipIfDead(MI, *NextBB)) {
397             HaveSkipBlock = true;
398             NextBB = std::next(BI);
399             BE = MF.end();
400           }
401         } else {
402           HaveKill = true;
403         }
404 
405         MI.eraseFromParent();
406         break;
407 
408       case AMDGPU::SI_RETURN_TO_EPILOG:
409         // FIXME: Should move somewhere else
410         assert(!MF.getInfo<SIMachineFunctionInfo>()->returnsVoid());
411 
412         // Graphics shaders returning non-void shouldn't contain S_ENDPGM,
413         // because external bytecode will be appended at the end.
414         if (BI != --MF.end() || I != MBB.getFirstTerminator()) {
415           // SI_RETURN_TO_EPILOG is not the last instruction. Add an empty block at
416           // the end and jump there.
417           if (!EmptyMBBAtEnd) {
418             EmptyMBBAtEnd = MF.CreateMachineBasicBlock();
419             MF.insert(MF.end(), EmptyMBBAtEnd);
420           }
421 
422           MBB.addSuccessor(EmptyMBBAtEnd);
423           BuildMI(*BI, I, MI.getDebugLoc(), TII->get(AMDGPU::S_BRANCH))
424             .addMBB(EmptyMBBAtEnd);
425           I->eraseFromParent();
426         }
427         break;
428 
429       default:
430         break;
431       }
432     }
433   }
434 
435   return MadeChange;
436 }
437