xref: /llvm-project/llvm/lib/Target/AMDGPU/SILowerControlFlow.cpp (revision 78fc9daf8d1825d32f170f8e60f9158550f93e93)
1 //===-- SILowerControlFlow.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 lowers the pseudo control flow instructions to real
12 /// machine instructions.
13 ///
14 /// All control flow is handled using predicated instructions and
15 /// a predicate stack.  Each Scalar ALU controls the operations of 64 Vector
16 /// ALUs.  The Scalar ALU can update the predicate for any of the Vector ALUs
17 /// by writting to the 64-bit EXEC register (each bit corresponds to a
18 /// single vector ALU).  Typically, for predicates, a vector ALU will write
19 /// to its bit of the VCC register (like EXEC VCC is 64-bits, one for each
20 /// Vector ALU) and then the ScalarALU will AND the VCC register with the
21 /// EXEC to update the predicates.
22 ///
23 /// For example:
24 /// %VCC = V_CMP_GT_F32 %VGPR1, %VGPR2
25 /// %SGPR0 = SI_IF %VCC
26 ///   %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0
27 /// %SGPR0 = SI_ELSE %SGPR0
28 ///   %VGPR0 = V_SUB_F32 %VGPR0, %VGPR0
29 /// SI_END_CF %SGPR0
30 ///
31 /// becomes:
32 ///
33 /// %SGPR0 = S_AND_SAVEEXEC_B64 %VCC  // Save and update the exec mask
34 /// %SGPR0 = S_XOR_B64 %SGPR0, %EXEC  // Clear live bits from saved exec mask
35 /// S_CBRANCH_EXECZ label0            // This instruction is an optional
36 ///                                   // optimization which allows us to
37 ///                                   // branch if all the bits of
38 ///                                   // EXEC are zero.
39 /// %VGPR0 = V_ADD_F32 %VGPR0, %VGPR0 // Do the IF block of the branch
40 ///
41 /// label0:
42 /// %SGPR0 = S_OR_SAVEEXEC_B64 %EXEC   // Restore the exec mask for the Then block
43 /// %EXEC = S_XOR_B64 %SGPR0, %EXEC    // Clear live bits from saved 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 "AMDGPUSubtarget.h"
53 #include "SIInstrInfo.h"
54 #include "SIMachineFunctionInfo.h"
55 #include "llvm/CodeGen/LivePhysRegs.h"
56 #include "llvm/CodeGen/MachineFrameInfo.h"
57 #include "llvm/CodeGen/MachineFunction.h"
58 #include "llvm/CodeGen/MachineFunctionPass.h"
59 #include "llvm/CodeGen/MachineInstrBuilder.h"
60 #include "llvm/CodeGen/MachineRegisterInfo.h"
61 
62 using namespace llvm;
63 
64 #define DEBUG_TYPE "si-lower-control-flow"
65 
66 namespace {
67 
68 class SILowerControlFlow : public MachineFunctionPass {
69 private:
70   const SIRegisterInfo *TRI;
71   const SIInstrInfo *TII;
72   LiveIntervals *LIS;
73 
74   void emitIf(MachineInstr &MI);
75   void emitElse(MachineInstr &MI);
76   void emitBreak(MachineInstr &MI);
77   void emitIfBreak(MachineInstr &MI);
78   void emitElseBreak(MachineInstr &MI);
79   void emitLoop(MachineInstr &MI);
80   void emitEndCf(MachineInstr &MI);
81 
82 public:
83   static char ID;
84 
85   SILowerControlFlow() :
86     MachineFunctionPass(ID),
87     TRI(nullptr),
88     TII(nullptr),
89     LIS(nullptr) {}
90 
91   bool runOnMachineFunction(MachineFunction &MF) override;
92 
93   const char *getPassName() const override {
94     return "SI Lower control flow pseudo instructions";
95   }
96 
97   void getAnalysisUsage(AnalysisUsage &AU) const override {
98     AU.addPreserved<LiveIntervals>();
99     AU.addPreserved<SlotIndexes>();
100     AU.setPreservesCFG();
101     MachineFunctionPass::getAnalysisUsage(AU);
102   }
103 };
104 
105 } // End anonymous namespace
106 
107 char SILowerControlFlow::ID = 0;
108 
109 INITIALIZE_PASS(SILowerControlFlow, DEBUG_TYPE,
110                "SI lower control flow", false, false)
111 
112 char &llvm::SILowerControlFlowID = SILowerControlFlow::ID;
113 
114 void SILowerControlFlow::emitIf(MachineInstr &MI) {
115   MachineBasicBlock &MBB = *MI.getParent();
116   const DebugLoc &DL = MI.getDebugLoc();
117   MachineBasicBlock::iterator I(&MI);
118 
119   MachineOperand &SaveExec = MI.getOperand(0);
120   MachineOperand &Cond = MI.getOperand(1);
121   assert(SaveExec.getSubReg() == AMDGPU::NoSubRegister &&
122          Cond.getSubReg() == AMDGPU::NoSubRegister);
123 
124   unsigned SaveExecReg = SaveExec.getReg();
125 
126   MachineInstr *AndSaveExec =
127     BuildMI(MBB, I, DL, TII->get(AMDGPU::S_AND_SAVEEXEC_B64), SaveExecReg)
128     .addOperand(Cond);
129 
130   MachineInstr *Xor =
131     BuildMI(MBB, I, DL, TII->get(AMDGPU::S_XOR_B64), SaveExecReg)
132     .addReg(AMDGPU::EXEC)
133     .addReg(SaveExecReg);
134 
135   // Insert a pseudo terminator to help keep the verifier happy. This will also
136   // be used later when inserting skips.
137   MachineInstr *NewBr =
138     BuildMI(MBB, I, DL, TII->get(AMDGPU::SI_MASK_BRANCH))
139     .addOperand(MI.getOperand(2))
140     .addReg(SaveExecReg, getKillRegState(SaveExec.isKill()));
141 
142   if (!LIS) {
143     MI.eraseFromParent();
144     return;
145   }
146 
147 
148   LIS->ReplaceMachineInstrInMaps(MI, *AndSaveExec);
149   LIS->InsertMachineInstrInMaps(*Xor);
150   LIS->InsertMachineInstrInMaps(*NewBr);
151 
152   MI.eraseFromParent();
153 
154   // FIXME: Is there a better way of adjusting the liveness? It shouldn't be
155   // hard to add another def here but I'm not sure how to correctly update the
156   // valno.
157   LIS->removeInterval(SaveExecReg);
158   LIS->createAndComputeVirtRegInterval(SaveExecReg);
159 }
160 
161 void SILowerControlFlow::emitElse(MachineInstr &MI) {
162   MachineBasicBlock &MBB = *MI.getParent();
163   const DebugLoc &DL = MI.getDebugLoc();
164 
165   unsigned DstReg = MI.getOperand(0).getReg();
166   assert(MI.getOperand(0).getSubReg() == AMDGPU::NoSubRegister);
167 
168   bool ExecModified = MI.getOperand(3).getImm() != 0;
169   MachineBasicBlock::iterator Start = MBB.begin();
170 
171   // This must be inserted before phis and any spill code inserted before the
172   // else.
173   MachineInstr *OrSaveExec =
174     BuildMI(MBB, Start, DL, TII->get(AMDGPU::S_OR_SAVEEXEC_B64), DstReg)
175     .addOperand(MI.getOperand(1)); // Saved EXEC
176   MachineBasicBlock *DestBB = MI.getOperand(2).getMBB();
177 
178   MachineBasicBlock::iterator ElsePt(MI);
179 
180   if (ExecModified) {
181     MachineInstr *And =
182       BuildMI(MBB, ElsePt, DL, TII->get(AMDGPU::S_AND_B64), DstReg)
183       .addReg(AMDGPU::EXEC)
184       .addReg(DstReg);
185 
186     if (LIS)
187       LIS->InsertMachineInstrInMaps(*And);
188   }
189 
190   MachineInstr *Xor =
191     BuildMI(MBB, ElsePt, DL, TII->get(AMDGPU::S_XOR_B64), AMDGPU::EXEC)
192     .addReg(AMDGPU::EXEC)
193     .addReg(DstReg);
194 
195   MachineBasicBlock::iterator Term = MBB.getFirstTerminator();
196   // Insert a pseudo terminator to help keep the verifier happy.
197   MachineInstr *Branch =
198     BuildMI(MBB, Term, DL, TII->get(AMDGPU::SI_MASK_BRANCH))
199     .addMBB(DestBB)
200     .addReg(DstReg);
201 
202   if (!LIS) {
203     MI.eraseFromParent();
204     return;
205   }
206 
207   LIS->RemoveMachineInstrFromMaps(MI);
208   MI.eraseFromParent();
209 
210   LIS->InsertMachineInstrInMaps(*OrSaveExec);
211 
212   LIS->InsertMachineInstrInMaps(*Xor);
213   LIS->InsertMachineInstrInMaps(*Branch);
214 
215   // src reg is tied to dst reg.
216   LIS->removeInterval(DstReg);
217   LIS->createAndComputeVirtRegInterval(DstReg);
218 
219   // Let this be recomputed.
220   LIS->removeRegUnit(*MCRegUnitIterator(AMDGPU::EXEC, TRI));
221 }
222 
223 void SILowerControlFlow::emitBreak(MachineInstr &MI) {
224   MachineBasicBlock &MBB = *MI.getParent();
225   const DebugLoc &DL = MI.getDebugLoc();
226   unsigned Dst = MI.getOperand(0).getReg();
227 
228   MachineInstr *Or =
229     BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_OR_B64), Dst)
230     .addReg(AMDGPU::EXEC)
231     .addOperand(MI.getOperand(1));
232 
233   if (LIS)
234     LIS->ReplaceMachineInstrInMaps(MI, *Or);
235   MI.eraseFromParent();
236 }
237 
238 void SILowerControlFlow::emitIfBreak(MachineInstr &MI) {
239   MI.setDesc(TII->get(AMDGPU::S_OR_B64));
240 }
241 
242 void SILowerControlFlow::emitElseBreak(MachineInstr &MI) {
243   MI.setDesc(TII->get(AMDGPU::S_OR_B64));
244 }
245 
246 void SILowerControlFlow::emitLoop(MachineInstr &MI) {
247   MachineBasicBlock &MBB = *MI.getParent();
248   const DebugLoc &DL = MI.getDebugLoc();
249 
250   MachineInstr *AndN2 =
251     BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_ANDN2_B64), AMDGPU::EXEC)
252     .addReg(AMDGPU::EXEC)
253     .addOperand(MI.getOperand(0));
254 
255   MachineInstr *Branch =
256     BuildMI(MBB, &MI, DL, TII->get(AMDGPU::S_CBRANCH_EXECNZ))
257     .addOperand(MI.getOperand(1));
258 
259   if (LIS) {
260     LIS->ReplaceMachineInstrInMaps(MI, *AndN2);
261     LIS->InsertMachineInstrInMaps(*Branch);
262   }
263 
264   MI.eraseFromParent();
265 }
266 
267 void SILowerControlFlow::emitEndCf(MachineInstr &MI) {
268   MachineBasicBlock &MBB = *MI.getParent();
269   const DebugLoc &DL = MI.getDebugLoc();
270 
271   MachineBasicBlock::iterator InsPt = MBB.begin();
272   MachineInstr *NewMI =
273     BuildMI(MBB, InsPt, DL, TII->get(AMDGPU::S_OR_B64), AMDGPU::EXEC)
274     .addReg(AMDGPU::EXEC)
275     .addOperand(MI.getOperand(0));
276 
277   if (LIS)
278     LIS->ReplaceMachineInstrInMaps(MI, *NewMI);
279 
280   MI.eraseFromParent();
281 
282   if (LIS)
283     LIS->handleMove(*NewMI);
284 }
285 
286 bool SILowerControlFlow::runOnMachineFunction(MachineFunction &MF) {
287   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
288   TII = ST.getInstrInfo();
289   TRI = &TII->getRegisterInfo();
290 
291   // This doesn't actually need LiveIntervals, but we can preserve them.
292   LIS = getAnalysisIfAvailable<LiveIntervals>();
293 
294   MachineFunction::iterator NextBB;
295   for (MachineFunction::iterator BI = MF.begin(), BE = MF.end();
296        BI != BE; BI = NextBB) {
297     NextBB = std::next(BI);
298     MachineBasicBlock &MBB = *BI;
299 
300     MachineBasicBlock::iterator I, Next;
301 
302     for (I = MBB.begin(); I != MBB.end(); I = Next) {
303       Next = std::next(I);
304       MachineInstr &MI = *I;
305 
306       switch (MI.getOpcode()) {
307       case AMDGPU::SI_IF:
308         emitIf(MI);
309         break;
310 
311       case AMDGPU::SI_ELSE:
312         emitElse(MI);
313         break;
314 
315       case AMDGPU::SI_BREAK:
316         emitBreak(MI);
317         break;
318 
319       case AMDGPU::SI_IF_BREAK:
320         emitIfBreak(MI);
321         break;
322 
323       case AMDGPU::SI_ELSE_BREAK:
324         emitElseBreak(MI);
325         break;
326 
327       case AMDGPU::SI_LOOP:
328         emitLoop(MI);
329         break;
330 
331       case AMDGPU::SI_END_CF:
332         emitEndCf(MI);
333         break;
334 
335       default:
336         break;
337       }
338     }
339   }
340 
341   return true;
342 }
343