1 //===- SIInsertHardClauses.cpp - Insert Hard Clauses ----------------------===//
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 /// Insert s_clause instructions to form hard clauses.
11 ///
12 /// Clausing load instructions can give cache coherency benefits. Before gfx10,
13 /// the hardware automatically detected "soft clauses", which were sequences of
14 /// memory instructions of the same type. In gfx10 this detection was removed,
15 /// and the s_clause instruction was introduced to explicitly mark "hard
16 /// clauses".
17 ///
18 /// It's the scheduler's job to form the clauses by putting similar memory
19 /// instructions next to each other. Our job is just to insert an s_clause
20 /// instruction to mark the start of each clause.
21 ///
22 /// Note that hard clauses are very similar to, but logically distinct from, the
23 /// groups of instructions that have to be restartable when XNACK is enabled.
24 /// The rules are slightly different in each case. For example an s_nop
25 /// instruction breaks a restartable group, but can appear in the middle of a
26 /// hard clause. (Before gfx10 there wasn't a distinction, and both were called
27 /// "soft clauses" or just "clauses".)
28 ///
29 /// The SIFormMemoryClauses pass and GCNHazardRecognizer deal with restartable
30 /// groups, not hard clauses.
31 //
32 //===----------------------------------------------------------------------===//
33 
34 #include "AMDGPU.h"
35 #include "GCNSubtarget.h"
36 #include "MCTargetDesc/AMDGPUMCTargetDesc.h"
37 #include "llvm/ADT/SmallVector.h"
38 
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "si-insert-hard-clauses"
42 
43 namespace {
44 
45 enum HardClauseType {
46   // Texture, buffer, global or scratch memory instructions.
47   HARDCLAUSE_VMEM,
48   // Flat (not global or scratch) memory instructions.
49   HARDCLAUSE_FLAT,
50   // Instructions that access LDS.
51   HARDCLAUSE_LDS,
52   // Scalar memory instructions.
53   HARDCLAUSE_SMEM,
54   // VALU instructions.
55   HARDCLAUSE_VALU,
56   LAST_REAL_HARDCLAUSE_TYPE = HARDCLAUSE_VALU,
57 
58   // Internal instructions, which are allowed in the middle of a hard clause,
59   // except for s_waitcnt.
60   HARDCLAUSE_INTERNAL,
61   // Instructions that are not allowed in a hard clause: SALU, export, branch,
62   // message, GDS, s_waitcnt and anything else not mentioned above.
63   HARDCLAUSE_ILLEGAL,
64 };
65 
66 HardClauseType getHardClauseType(const MachineInstr &MI) {
67   // On current architectures we only get a benefit from clausing loads.
68   if (MI.mayLoad()) {
69     if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI))
70       return HARDCLAUSE_VMEM;
71     if (SIInstrInfo::isFLAT(MI))
72       return HARDCLAUSE_FLAT;
73     // TODO: LDS
74     if (SIInstrInfo::isSMRD(MI))
75       return HARDCLAUSE_SMEM;
76   }
77 
78   // Don't form VALU clauses. It's not clear what benefit they give, if any.
79 
80   // In practice s_nop is the only internal instruction we're likely to see.
81   // It's safe to treat the rest as illegal.
82   if (MI.getOpcode() == AMDGPU::S_NOP)
83     return HARDCLAUSE_INTERNAL;
84   return HARDCLAUSE_ILLEGAL;
85 }
86 
87 class SIInsertHardClauses : public MachineFunctionPass {
88 public:
89   static char ID;
90 
91   SIInsertHardClauses() : MachineFunctionPass(ID) {}
92 
93   void getAnalysisUsage(AnalysisUsage &AU) const override {
94     AU.setPreservesCFG();
95     MachineFunctionPass::getAnalysisUsage(AU);
96   }
97 
98   // Track information about a clause as we discover it.
99   struct ClauseInfo {
100     // The type of all (non-internal) instructions in the clause.
101     HardClauseType Type = HARDCLAUSE_ILLEGAL;
102     // The first (necessarily non-internal) instruction in the clause.
103     MachineInstr *First = nullptr;
104     // The last non-internal instruction in the clause.
105     MachineInstr *Last = nullptr;
106     // The length of the clause including any internal instructions in the
107     // middle or after the end of the clause.
108     unsigned Length = 0;
109     // The base operands of *Last.
110     SmallVector<const MachineOperand *, 4> BaseOps;
111   };
112 
113   bool emitClause(const ClauseInfo &CI, const SIInstrInfo *SII) {
114     // Get the size of the clause excluding any internal instructions at the
115     // end.
116     unsigned Size =
117         std::distance(CI.First->getIterator(), CI.Last->getIterator()) + 1;
118     if (Size < 2)
119       return false;
120     assert(Size <= 64 && "Hard clause is too long!");
121 
122     auto &MBB = *CI.First->getParent();
123     auto ClauseMI =
124         BuildMI(MBB, *CI.First, DebugLoc(), SII->get(AMDGPU::S_CLAUSE))
125             .addImm(Size - 1);
126     finalizeBundle(MBB, ClauseMI->getIterator(),
127                    std::next(CI.Last->getIterator()));
128     return true;
129   }
130 
131   bool runOnMachineFunction(MachineFunction &MF) override {
132     if (skipFunction(MF.getFunction()))
133       return false;
134 
135     const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
136     if (!ST.hasHardClauses())
137       return false;
138 
139     const SIInstrInfo *SII = ST.getInstrInfo();
140     const TargetRegisterInfo *TRI = ST.getRegisterInfo();
141 
142     bool Changed = false;
143     for (auto &MBB : MF) {
144       ClauseInfo CI;
145       for (auto &MI : MBB) {
146         HardClauseType Type = getHardClauseType(MI);
147 
148         int64_t Dummy1;
149         bool Dummy2;
150         unsigned Dummy3;
151         SmallVector<const MachineOperand *, 4> BaseOps;
152         if (Type <= LAST_REAL_HARDCLAUSE_TYPE) {
153           if (!SII->getMemOperandsWithOffsetWidth(MI, BaseOps, Dummy1, Dummy2,
154                                                   Dummy3, TRI)) {
155             // We failed to get the base operands, so we'll never clause this
156             // instruction with any other, so pretend it's illegal.
157             Type = HARDCLAUSE_ILLEGAL;
158           }
159         }
160 
161         if (CI.Length == 64 ||
162             (CI.Length && Type != HARDCLAUSE_INTERNAL &&
163              (Type != CI.Type ||
164               // Note that we lie to shouldClusterMemOps about the size of the
165               // cluster. When shouldClusterMemOps is called from the machine
166               // scheduler it limits the size of the cluster to avoid increasing
167               // register pressure too much, but this pass runs after register
168               // allocation so there is no need for that kind of limit.
169               !SII->shouldClusterMemOps(CI.BaseOps, BaseOps, 2, 2)))) {
170           // Finish the current clause.
171           Changed |= emitClause(CI, SII);
172           CI = ClauseInfo();
173         }
174 
175         if (CI.Length) {
176           // Extend the current clause.
177           ++CI.Length;
178           if (Type != HARDCLAUSE_INTERNAL) {
179             CI.Last = &MI;
180             CI.BaseOps = std::move(BaseOps);
181           }
182         } else if (Type <= LAST_REAL_HARDCLAUSE_TYPE) {
183           // Start a new clause.
184           CI = ClauseInfo{Type, &MI, &MI, 1, std::move(BaseOps)};
185         }
186       }
187 
188       // Finish the last clause in the basic block if any.
189       if (CI.Length)
190         Changed |= emitClause(CI, SII);
191     }
192 
193     return Changed;
194   }
195 };
196 
197 } // namespace
198 
199 char SIInsertHardClauses::ID = 0;
200 
201 char &llvm::SIInsertHardClausesID = SIInsertHardClauses::ID;
202 
203 INITIALIZE_PASS(SIInsertHardClauses, DEBUG_TYPE, "SI Insert Hard Clauses",
204                 false, false)
205