1 | //===--------------------- DispatchStage.cpp --------------------*- C++ -*-===// |
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 | /// \file |
9 | /// |
10 | /// This file models the dispatch component of an instruction pipeline. |
11 | /// |
12 | /// The DispatchStage is responsible for updating instruction dependencies |
13 | /// and communicating to the simulated instruction scheduler that an instruction |
14 | /// is ready to be scheduled for execution. |
15 | /// |
16 | //===----------------------------------------------------------------------===// |
17 | |
18 | #include "llvm/MCA/Stages/DispatchStage.h" |
19 | #include "llvm/MCA/HWEventListener.h" |
20 | #include "llvm/MCA/HardwareUnits/Scheduler.h" |
21 | #include "llvm/Support/Debug.h" |
22 | |
23 | #define DEBUG_TYPE "llvm-mca" |
24 | |
25 | namespace llvm { |
26 | namespace mca { |
27 | |
28 | DispatchStage::DispatchStage(const MCSubtargetInfo &Subtarget, |
29 | const MCRegisterInfo &MRI, |
30 | unsigned MaxDispatchWidth, RetireControlUnit &R, |
31 | RegisterFile &F) |
32 | : DispatchWidth(MaxDispatchWidth), AvailableEntries(MaxDispatchWidth), |
33 | CarryOver(0U), STI(Subtarget), RCU(R), PRF(F) { |
34 | if (!DispatchWidth) |
35 | DispatchWidth = Subtarget.getSchedModel().IssueWidth; |
36 | } |
37 | |
38 | void DispatchStage::notifyInstructionDispatched(const InstRef &IR, |
39 | ArrayRef<unsigned> UsedRegs, |
40 | unsigned UOps) const { |
41 | LLVM_DEBUG(dbgs() << "[E] Instruction Dispatched: #" << IR << '\n'); |
42 | notifyEvent<HWInstructionEvent>( |
43 | Event: HWInstructionDispatchedEvent(IR, UsedRegs, UOps)); |
44 | } |
45 | |
46 | bool DispatchStage::checkPRF(const InstRef &IR) const { |
47 | SmallVector<MCPhysReg, 4> RegDefs; |
48 | for (const WriteState &RegDef : IR.getInstruction()->getDefs()) |
49 | RegDefs.emplace_back(Args: RegDef.getRegisterID()); |
50 | |
51 | const unsigned RegisterMask = PRF.isAvailable(Regs: RegDefs); |
52 | // A mask with all zeroes means: register files are available. |
53 | if (RegisterMask) { |
54 | notifyEvent<HWStallEvent>( |
55 | Event: HWStallEvent(HWStallEvent::RegisterFileStall, IR)); |
56 | return false; |
57 | } |
58 | |
59 | return true; |
60 | } |
61 | |
62 | bool DispatchStage::checkRCU(const InstRef &IR) const { |
63 | const unsigned NumMicroOps = IR.getInstruction()->getNumMicroOps(); |
64 | if (RCU.isAvailable(Quantity: NumMicroOps)) |
65 | return true; |
66 | notifyEvent<HWStallEvent>( |
67 | Event: HWStallEvent(HWStallEvent::RetireControlUnitStall, IR)); |
68 | return false; |
69 | } |
70 | |
71 | bool DispatchStage::canDispatch(const InstRef &IR) const { |
72 | bool CanDispatch = checkRCU(IR); |
73 | CanDispatch &= checkPRF(IR); |
74 | CanDispatch &= checkNextStage(IR); |
75 | return CanDispatch; |
76 | } |
77 | |
78 | Error DispatchStage::dispatch(InstRef IR) { |
79 | assert(!CarryOver && "Cannot dispatch another instruction!" ); |
80 | Instruction &IS = *IR.getInstruction(); |
81 | const unsigned NumMicroOps = IS.getNumMicroOps(); |
82 | if (NumMicroOps > DispatchWidth) { |
83 | assert(AvailableEntries == DispatchWidth); |
84 | AvailableEntries = 0; |
85 | CarryOver = NumMicroOps - DispatchWidth; |
86 | CarriedOver = IR; |
87 | } else { |
88 | assert(AvailableEntries >= NumMicroOps); |
89 | AvailableEntries -= NumMicroOps; |
90 | } |
91 | |
92 | // Check if this instructions ends the dispatch group. |
93 | if (IS.getEndGroup()) |
94 | AvailableEntries = 0; |
95 | |
96 | // Check if this is an optimizable reg-reg move or an XCHG-like instruction. |
97 | if (IS.isOptimizableMove()) |
98 | if (PRF.tryEliminateMoveOrSwap(Writes: IS.getDefs(), Reads: IS.getUses())) |
99 | IS.setEliminated(); |
100 | |
101 | // A dependency-breaking instruction doesn't have to wait on the register |
102 | // input operands, and it is often optimized at register renaming stage. |
103 | // Update RAW dependencies if this instruction is not a dependency-breaking |
104 | // instruction. A dependency-breaking instruction is a zero-latency |
105 | // instruction that doesn't consume hardware resources. |
106 | // An example of dependency-breaking instruction on X86 is a zero-idiom XOR. |
107 | // |
108 | // We also don't update data dependencies for instructions that have been |
109 | // eliminated at register renaming stage. |
110 | if (!IS.isEliminated()) { |
111 | for (ReadState &RS : IS.getUses()) |
112 | PRF.addRegisterRead(RS, STI); |
113 | } |
114 | |
115 | // By default, a dependency-breaking zero-idiom is expected to be optimized |
116 | // at register renaming stage. That means, no physical register is allocated |
117 | // to the instruction. |
118 | SmallVector<unsigned, 4> RegisterFiles(PRF.getNumRegisterFiles()); |
119 | for (WriteState &WS : IS.getDefs()) |
120 | PRF.addRegisterWrite(Write: WriteRef(IR.getSourceIndex(), &WS), UsedPhysRegs: RegisterFiles); |
121 | |
122 | // Reserve entries in the reorder buffer. |
123 | unsigned RCUTokenID = RCU.dispatch(IS: IR); |
124 | // Notify the instruction that it has been dispatched. |
125 | IS.dispatch(RCUTokenID); |
126 | |
127 | // Notify listeners of the "instruction dispatched" event, |
128 | // and move IR to the next stage. |
129 | notifyInstructionDispatched(IR, UsedRegs: RegisterFiles, |
130 | UOps: std::min(a: DispatchWidth, b: NumMicroOps)); |
131 | return moveToTheNextStage(IR); |
132 | } |
133 | |
134 | Error DispatchStage::cycleStart() { |
135 | // The retire stage is responsible for calling method `cycleStart` |
136 | // on the PRF. |
137 | if (!CarryOver) { |
138 | AvailableEntries = DispatchWidth; |
139 | return ErrorSuccess(); |
140 | } |
141 | |
142 | AvailableEntries = CarryOver >= DispatchWidth ? 0 : DispatchWidth - CarryOver; |
143 | unsigned DispatchedOpcodes = DispatchWidth - AvailableEntries; |
144 | CarryOver -= DispatchedOpcodes; |
145 | assert(CarriedOver && "Invalid dispatched instruction" ); |
146 | |
147 | SmallVector<unsigned, 8> RegisterFiles(PRF.getNumRegisterFiles(), 0U); |
148 | notifyInstructionDispatched(IR: CarriedOver, UsedRegs: RegisterFiles, UOps: DispatchedOpcodes); |
149 | if (!CarryOver) |
150 | CarriedOver = InstRef(); |
151 | return ErrorSuccess(); |
152 | } |
153 | |
154 | bool DispatchStage::isAvailable(const InstRef &IR) const { |
155 | // Conservatively bail out if there are no available dispatch entries. |
156 | if (!AvailableEntries) |
157 | return false; |
158 | |
159 | const Instruction &Inst = *IR.getInstruction(); |
160 | unsigned NumMicroOps = Inst.getNumMicroOps(); |
161 | unsigned Required = std::min(a: NumMicroOps, b: DispatchWidth); |
162 | if (Required > AvailableEntries) |
163 | return false; |
164 | |
165 | if (Inst.getBeginGroup() && AvailableEntries != DispatchWidth) |
166 | return false; |
167 | |
168 | // The dispatch logic doesn't internally buffer instructions. It only accepts |
169 | // instructions that can be successfully moved to the next stage during this |
170 | // same cycle. |
171 | return canDispatch(IR); |
172 | } |
173 | |
174 | Error DispatchStage::execute(InstRef &IR) { |
175 | assert(canDispatch(IR) && "Cannot dispatch another instruction!" ); |
176 | return dispatch(IR); |
177 | } |
178 | |
179 | #ifndef NDEBUG |
180 | void DispatchStage::dump() const { |
181 | PRF.dump(); |
182 | RCU.dump(); |
183 | } |
184 | #endif |
185 | } // namespace mca |
186 | } // namespace llvm |
187 | |