1 //===-- X86PreTileConfig.cpp - Tile Register Pre-configure-----------------===//
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 Pass to pre-config the shapes of AMX registers
10 /// AMX register needs to be configured before use. The shapes of AMX register
11 /// are encoded in the 1st and 2nd machine operand of AMX pseudo instructions.
12 ///
13 /// The instruction ldtilecfg is used to config the shapes. It must be reachable
14 /// for all variable shapes. ldtilecfg will be inserted more than once if we
15 /// cannot find a dominating point for all AMX instructions.
16 ///
17 /// The configure register is caller saved according to ABI. We need to insert
18 /// ldtilecfg again after the call instruction if callee clobbers any AMX
19 /// registers.
20 ///
21 /// This pass calculates all points that ldtilecfg need to be inserted to and
22 /// insert them. It reports error if the reachability conditions aren't met.
23 //
24 //===----------------------------------------------------------------------===//
25 
26 #include "X86.h"
27 #include "X86InstrBuilder.h"
28 #include "X86RegisterInfo.h"
29 #include "X86Subtarget.h"
30 #include "llvm/CodeGen/MachineFunctionPass.h"
31 #include "llvm/CodeGen/MachineInstr.h"
32 #include "llvm/CodeGen/MachineLoopInfo.h"
33 #include "llvm/CodeGen/MachineRegisterInfo.h"
34 #include "llvm/CodeGen/Passes.h"
35 #include "llvm/CodeGen/TargetInstrInfo.h"
36 #include "llvm/CodeGen/TargetRegisterInfo.h"
37 #include "llvm/InitializePasses.h"
38 
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "tile-pre-config"
42 #define REPORT_CONFIG_FAIL                                                     \
43   report_fatal_error(                                                          \
44       MF.getName() +                                                           \
45       ": Failed to config tile register, please define the shape earlier");
46 
47 namespace {
48 
49 struct MIRef {
50   MachineInstr *MI = nullptr;
51   MachineBasicBlock *MBB = nullptr;
52   // A virtual position for instruction that will be inserted after MI.
53   size_t Pos = 0;
54   MIRef() = default;
55   MIRef(MachineBasicBlock *MBB) : MBB(MBB) {
56     for (auto I = MBB->begin(), E = MBB->end(); I != E && I->isPHI();
57          ++I, ++Pos)
58       MI = &*I;
59   }
60   MIRef(MachineInstr *MI)
61       : MI(MI), MBB(MI->getParent()),
62         Pos(std::distance(MBB->instr_begin(), ++MI->getIterator())) {}
63   MIRef(MachineInstr *MI, MachineBasicBlock *MBB)
64       : MI(MI), MBB(MBB),
65         Pos(std::distance(MBB->instr_begin(), ++MI->getIterator())) {}
66   MIRef(MachineInstr *MI, MachineBasicBlock *MBB, size_t Pos)
67       : MI(MI), MBB(MBB), Pos(Pos) {}
68   operator bool() const { return MBB != nullptr; }
69   bool operator==(const MIRef &RHS) const {
70     return MI == RHS.MI && MBB == RHS.MBB;
71   }
72   bool operator!=(const MIRef &RHS) const { return !(*this == RHS); }
73   bool operator<(const MIRef &RHS) const {
74     // Comparison between different BBs happens when inserting a MIRef into set.
75     // So we compare MBB first to make the insertion happy.
76     return MBB < RHS.MBB || (MBB == RHS.MBB && Pos < RHS.Pos);
77   }
78   bool operator>(const MIRef &RHS) const {
79     // Comparison between different BBs happens when inserting a MIRef into set.
80     // So we compare MBB first to make the insertion happy.
81     return MBB > RHS.MBB || (MBB == RHS.MBB && Pos > RHS.Pos);
82   }
83 };
84 
85 struct BBInfo {
86   MIRef FirstAMX;
87   MIRef LastCall;
88   bool HasAMXRegLiveIn = false;
89   bool TileCfgForbidden = false;
90   bool NeedTileCfgLiveIn = false;
91 };
92 
93 class X86PreTileConfig : public MachineFunctionPass {
94   MachineRegisterInfo *MRI;
95   const MachineLoopInfo *MLI;
96   SmallSet<MachineInstr *, 8> DefVisited;
97   DenseMap<MachineBasicBlock *, BBInfo> BBVisitedInfo;
98   DenseMap<MachineBasicBlock *, SmallVector<MIRef, 8>> ShapeBBs;
99 
100   /// Check if the callee will clobber AMX registers.
101   bool isDestructiveCall(MachineInstr &MI, BitVector UsableRegs) {
102     auto Iter = llvm::find_if(
103         MI.operands(), [](MachineOperand &MO) { return MO.isRegMask(); });
104     if (Iter == MI.operands_end())
105       return false;
106     UsableRegs.clearBitsInMask(Iter->getRegMask());
107     return !UsableRegs.none();
108   }
109 
110   /// Check if MI is AMX pseudo instruction.
111   bool isAMXInstruction(MachineInstr &MI) {
112     if (MI.isPHI() || MI.isDebugInstr() || MI.getNumOperands() < 3)
113       return false;
114     MachineOperand &MO = MI.getOperand(0);
115     // We can simply check if it is AMX instruction by its def.
116     // But we should exclude old API which uses physical registers.
117     if (MO.isReg() && MO.getReg().isVirtual() &&
118         MRI->getRegClass(MO.getReg())->getID() == X86::TILERegClassID) {
119       collectShapeInfo(MI);
120       return true;
121     }
122     // PTILESTOREDV is the only exception that doesn't def a AMX register.
123     return MI.getOpcode() == X86::PTILESTOREDV;
124   }
125 
126   /// Check if it is an edge from loop bottom to loop head.
127   bool isLoopBackEdge(MachineBasicBlock *Header, MachineBasicBlock *Bottom) {
128     if (!MLI->isLoopHeader(Header))
129       return false;
130     auto *ML = MLI->getLoopFor(Header);
131     if (ML->contains(Bottom) && ML->isLoopLatch(Bottom))
132       return true;
133 
134     return false;
135   }
136 
137   /// Collect the shape def information for later use.
138   void collectShapeInfo(MachineInstr &MI);
139 
140   /// Try to hoist shapes definded below AMX instructions.
141   bool hoistShapesInBB(MachineBasicBlock *MBB, SmallVectorImpl<MIRef> &Shapes) {
142     MIRef &FirstAMX = BBVisitedInfo[MBB].FirstAMX;
143     auto FirstShapeBelowAMX = llvm::lower_bound(Shapes, FirstAMX);
144     auto InsertPoint = FirstAMX.MI->getIterator();
145     for (auto I = FirstShapeBelowAMX, E = Shapes.end(); I != E; ++I) {
146       // Do not hoist instructions that access memory.
147       if (I->MI->mayLoadOrStore())
148         return false;
149       for (auto &MO : I->MI->operands()) {
150         if (MO.isDef())
151           continue;
152         // Do not hoist instructions if the sources' def under AMX instruction.
153         // TODO: We can handle isMoveImmediate MI here.
154         if (MO.isReg() && MIRef(MRI->getVRegDef(MO.getReg())) > FirstAMX)
155           return false;
156         // TODO: Maybe need more checks here.
157       }
158       MBB->insert(InsertPoint, I->MI->removeFromParent());
159     }
160     // We only need to mark the last shape in the BB now.
161     Shapes.clear();
162     Shapes.push_back(MIRef(&*--InsertPoint, MBB));
163     return true;
164   }
165 
166 public:
167   X86PreTileConfig() : MachineFunctionPass(ID) {}
168 
169   /// Return the pass name.
170   StringRef getPassName() const override {
171     return "Tile Register Pre-configure";
172   }
173 
174   /// X86PreTileConfig analysis usage.
175   void getAnalysisUsage(AnalysisUsage &AU) const override {
176     AU.setPreservesAll();
177     AU.addRequired<MachineLoopInfo>();
178     MachineFunctionPass::getAnalysisUsage(AU);
179   }
180 
181   /// Clear MF related structures.
182   void releaseMemory() override {
183     ShapeBBs.clear();
184     DefVisited.clear();
185     BBVisitedInfo.clear();
186   }
187 
188   /// Perform ldtilecfg instructions inserting.
189   bool runOnMachineFunction(MachineFunction &MF) override;
190 
191   static char ID;
192 };
193 
194 } // end anonymous namespace
195 
196 char X86PreTileConfig::ID = 0;
197 
198 INITIALIZE_PASS_BEGIN(X86PreTileConfig, "tilepreconfig",
199                       "Tile Register Pre-configure", false, false)
200 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
201 INITIALIZE_PASS_END(X86PreTileConfig, "tilepreconfig",
202                     "Tile Register Pre-configure", false, false)
203 
204 void X86PreTileConfig::collectShapeInfo(MachineInstr &MI) {
205   auto RecordShape = [&](MachineInstr *MI, MachineBasicBlock *MBB) {
206     MIRef MIR(MI, MBB);
207     auto I = llvm::lower_bound(ShapeBBs[MBB], MIR);
208     if (I == ShapeBBs[MBB].end() || *I != MIR)
209       ShapeBBs[MBB].insert(I, MIR);
210   };
211 
212   SmallVector<Register, 8> WorkList(
213       {MI.getOperand(1).getReg(), MI.getOperand(2).getReg()});
214   while (!WorkList.empty()) {
215     Register R = WorkList.pop_back_val();
216     MachineInstr *DefMI = MRI->getVRegDef(R);
217     assert(DefMI && "R must has one define instruction");
218     MachineBasicBlock *DefMBB = DefMI->getParent();
219     if (DefMI->isMoveImmediate() || !DefVisited.insert(DefMI).second)
220       continue;
221     if (DefMI->isPHI()) {
222       for (unsigned I = 1; I < DefMI->getNumOperands(); I += 2)
223         if (isLoopBackEdge(DefMBB, DefMI->getOperand(I + 1).getMBB()))
224           RecordShape(DefMI, DefMBB); // In this case, PHI is also a shape def.
225         else
226           WorkList.push_back(DefMI->getOperand(I).getReg());
227     } else {
228       RecordShape(DefMI, DefMBB);
229     }
230   }
231 }
232 
233 bool X86PreTileConfig::runOnMachineFunction(MachineFunction &MF) {
234   const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>();
235   const TargetInstrInfo *TII = ST.getInstrInfo();
236   const TargetRegisterInfo *TRI = ST.getRegisterInfo();
237   const TargetRegisterClass *RC = TRI->getRegClass(X86::TILERegClassID);
238 
239   BitVector AMXRegs(TRI->getNumRegs());
240   for (unsigned I = 0; I < RC->getNumRegs(); I++)
241     AMXRegs.set(X86::TMM0 + I);
242 
243   // Iterate MF to collect information.
244   MRI = &MF.getRegInfo();
245   MLI = &getAnalysis<MachineLoopInfo>();
246   SmallSet<MIRef, 8> CfgNeedInsert;
247   SmallVector<MachineBasicBlock *, 8> CfgLiveInBBs;
248   for (auto &MBB : MF) {
249     size_t Pos = 0;
250     for (auto &MI : MBB) {
251       ++Pos;
252       if (isAMXInstruction(MI)) {
253         // If there's call before the AMX, we need to reload tile config.
254         if (BBVisitedInfo[&MBB].LastCall)
255           CfgNeedInsert.insert(BBVisitedInfo[&MBB].LastCall);
256         else // Otherwise, we need tile config to live in this BB.
257           BBVisitedInfo[&MBB].NeedTileCfgLiveIn = true;
258         // Always record the first AMX in case there's shape def after it.
259         if (!BBVisitedInfo[&MBB].FirstAMX)
260           BBVisitedInfo[&MBB].FirstAMX = MIRef(&MI, &MBB, Pos);
261       } else if (MI.isCall() && isDestructiveCall(MI, AMXRegs)) {
262         // Record the call only if the callee clobbers all AMX registers.
263         BBVisitedInfo[&MBB].LastCall = MIRef(&MI, &MBB, Pos);
264       }
265     }
266     if (BBVisitedInfo[&MBB].NeedTileCfgLiveIn) {
267       if (&MBB == &MF.front())
268         CfgNeedInsert.insert(MIRef(&MBB));
269       else
270         CfgLiveInBBs.push_back(&MBB);
271     }
272     if (BBVisitedInfo[&MBB].FirstAMX || BBVisitedInfo[&MBB].HasAMXRegLiveIn)
273       for (auto *Succ : MBB.successors())
274         if (!isLoopBackEdge(Succ, &MBB))
275           BBVisitedInfo[Succ].HasAMXRegLiveIn = true;
276   }
277 
278   // Update NeedTileCfgLiveIn for predecessors.
279   while (!CfgLiveInBBs.empty()) {
280     MachineBasicBlock *MBB = CfgLiveInBBs.pop_back_val();
281     for (auto *Pred : MBB->predecessors()) {
282       if (BBVisitedInfo[Pred].LastCall) {
283         CfgNeedInsert.insert(BBVisitedInfo[Pred].LastCall);
284       } else if (!BBVisitedInfo[Pred].NeedTileCfgLiveIn) {
285         BBVisitedInfo[Pred].NeedTileCfgLiveIn = true;
286         if (Pred == &MF.front())
287           CfgNeedInsert.insert(MIRef(Pred));
288         else
289           CfgLiveInBBs.push_back(Pred);
290       }
291     }
292   }
293 
294   // There's no AMX instruction if we didn't find a tile config live in point.
295   if (CfgNeedInsert.empty())
296     return false;
297 
298   // Avoid to insert ldtilecfg before any shape defs.
299   SmallVector<MachineBasicBlock *, 8> WorkList;
300   for (auto &I : ShapeBBs) {
301     // TODO: We can hoist shapes across BBs here.
302     if (BBVisitedInfo[I.first].HasAMXRegLiveIn)
303       REPORT_CONFIG_FAIL
304     if (BBVisitedInfo[I.first].FirstAMX &&
305         BBVisitedInfo[I.first].FirstAMX < I.second.back() &&
306         !hoistShapesInBB(I.first, I.second))
307       REPORT_CONFIG_FAIL
308     WorkList.push_back(I.first);
309   }
310   while (!WorkList.empty()) {
311     MachineBasicBlock *MBB = WorkList.pop_back_val();
312     for (auto *Pred : MBB->predecessors()) {
313       if (!BBVisitedInfo[Pred].TileCfgForbidden && !isLoopBackEdge(MBB, Pred)) {
314         BBVisitedInfo[Pred].TileCfgForbidden = true;
315         WorkList.push_back(Pred);
316       }
317     }
318   }
319 
320   DebugLoc DL;
321   SmallSet<MIRef, 8> VisitedOrInserted;
322   int SS = MF.getFrameInfo().CreateStackObject(
323       ST.getTileConfigSize(), ST.getTileConfigAlignment(), false);
324 
325   // Try to insert for the tile config live in points.
326   for (auto I : CfgNeedInsert) {
327     SmallSet<MIRef, 8> InsertPoints;
328     SmallVector<MIRef, 8> WorkList({I});
329     while (!WorkList.empty()) {
330       MIRef I = WorkList.pop_back_val();
331       if (!VisitedOrInserted.count(I)) {
332         if (!BBVisitedInfo[I.MBB].TileCfgForbidden) {
333           // If the BB is all shapes reachable, stop sink and try to insert.
334           InsertPoints.insert(I);
335         } else {
336           // Avoid the BB to be multi visited.
337           VisitedOrInserted.insert(I);
338           // Sink the inserting point along the chain with NeedTileCfgLiveIn =
339           // true when MBB isn't all shapes reachable.
340           for (auto *Succ : I.MBB->successors())
341             if (BBVisitedInfo[Succ].NeedTileCfgLiveIn)
342               WorkList.push_back(MIRef(Succ));
343         }
344       }
345     }
346 
347     // A given point might be forked due to shape conditions are not met.
348     for (MIRef I : InsertPoints) {
349       // Make sure we insert ldtilecfg after the last shape def in MBB.
350       if (ShapeBBs.count(I.MBB) && I < ShapeBBs[I.MBB].back())
351         I = ShapeBBs[I.MBB].back();
352       // There're chances the MBB is sunk more than once. Record it to avoid
353       // multi insert.
354       if (VisitedOrInserted.insert(I).second) {
355         auto II = I.MI ? I.MI->getIterator() : I.MBB->instr_begin();
356         addFrameReference(BuildMI(*I.MBB, ++II, DL, TII->get(X86::LDTILECFG)),
357                           SS);
358       }
359     }
360   }
361 
362   // Zero stack slot.
363   MachineBasicBlock &MBB = MF.front();
364   MachineInstr *MI = &*MBB.begin();
365   if (ST.hasAVX512()) {
366     Register Zmm = MRI->createVirtualRegister(&X86::VR512RegClass);
367     BuildMI(MBB, MI, DL, TII->get(X86::VPXORDZrr), Zmm)
368         .addReg(Zmm, RegState::Undef)
369         .addReg(Zmm, RegState::Undef);
370     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::VMOVUPSZmr)), SS)
371         .addReg(Zmm);
372   } else if (ST.hasAVX2()) {
373     Register Ymm = MRI->createVirtualRegister(&X86::VR256RegClass);
374     BuildMI(MBB, MI, DL, TII->get(X86::VPXORYrr), Ymm)
375         .addReg(Ymm, RegState::Undef)
376         .addReg(Ymm, RegState::Undef);
377     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::VMOVUPSYmr)), SS)
378         .addReg(Ymm);
379     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::VMOVUPSYmr)), SS, 32)
380         .addReg(Ymm);
381   } else {
382     assert(ST.hasSSE2() && "AMX should assume SSE2 enabled");
383     Register Xmm = MRI->createVirtualRegister(&X86::VR128RegClass);
384     BuildMI(MBB, MI, DL, TII->get(X86::PXORrr), Xmm)
385         .addReg(Xmm, RegState::Undef)
386         .addReg(Xmm, RegState::Undef);
387     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::MOVUPSmr)), SS)
388         .addReg(Xmm);
389     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::MOVUPSmr)), SS, 16)
390         .addReg(Xmm);
391     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::MOVUPSmr)), SS, 32)
392         .addReg(Xmm);
393     addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::MOVUPSmr)), SS, 48)
394         .addReg(Xmm);
395   }
396   // Fill in the palette first.
397   addFrameReference(BuildMI(MBB, MI, DL, TII->get(X86::MOV8mi)), SS).addImm(1);
398 
399   return true;
400 }
401 
402 FunctionPass *llvm::createX86PreTileConfigPass() {
403   return new X86PreTileConfig();
404 }
405