LLVM 17.0.0git
SIPreAllocateWWMRegs.cpp
Go to the documentation of this file.
1//===- SIPreAllocateWWMRegs.cpp - WWM Register Pre-allocation -------------===//
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/// Pass to pre-allocated WWM registers
11//
12//===----------------------------------------------------------------------===//
13
14#include "AMDGPU.h"
15#include "GCNSubtarget.h"
26
27using namespace llvm;
28
29#define DEBUG_TYPE "si-pre-allocate-wwm-regs"
30
31namespace {
32
33class SIPreAllocateWWMRegs : public MachineFunctionPass {
34private:
35 const SIInstrInfo *TII;
36 const SIRegisterInfo *TRI;
38 LiveIntervals *LIS;
40 VirtRegMap *VRM;
41 RegisterClassInfo RegClassInfo;
42
43 std::vector<unsigned> RegsToRewrite;
44#ifndef NDEBUG
45 void printWWMInfo(const MachineInstr &MI);
46#endif
47
48public:
49 static char ID;
50
51 SIPreAllocateWWMRegs() : MachineFunctionPass(ID) {
53 }
54
55 bool runOnMachineFunction(MachineFunction &MF) override;
56
57 void getAnalysisUsage(AnalysisUsage &AU) const override {
63 AU.setPreservesCFG();
65 }
66
67private:
68 bool processDef(MachineOperand &MO);
69 void rewriteRegs(MachineFunction &MF);
70};
71
72} // End anonymous namespace.
73
74INITIALIZE_PASS_BEGIN(SIPreAllocateWWMRegs, DEBUG_TYPE,
75 "SI Pre-allocate WWM Registers", false, false)
79INITIALIZE_PASS_END(SIPreAllocateWWMRegs, DEBUG_TYPE,
80 "SI Pre-allocate WWM Registers", false, false)
81
82char SIPreAllocateWWMRegs::ID = 0;
83
84char &llvm::SIPreAllocateWWMRegsID = SIPreAllocateWWMRegs::ID;
85
87 return new SIPreAllocateWWMRegs();
88}
89
90bool SIPreAllocateWWMRegs::processDef(MachineOperand &MO) {
91 Register Reg = MO.getReg();
92 if (Reg.isPhysical())
93 return false;
94
95 if (!TRI->isVGPR(*MRI, Reg))
96 return false;
97
98 if (VRM->hasPhys(Reg))
99 return false;
100
101 LiveInterval &LI = LIS->getInterval(Reg);
102
103 for (MCRegister PhysReg : RegClassInfo.getOrder(MRI->getRegClass(Reg))) {
104 if (!MRI->isPhysRegUsed(PhysReg) &&
105 Matrix->checkInterference(LI, PhysReg) == LiveRegMatrix::IK_Free) {
106 Matrix->assign(LI, PhysReg);
107 assert(PhysReg != 0);
108 RegsToRewrite.push_back(Reg);
109 return true;
110 }
111 }
112
113 llvm_unreachable("physreg not found for WWM expression");
114}
115
116void SIPreAllocateWWMRegs::rewriteRegs(MachineFunction &MF) {
117 for (MachineBasicBlock &MBB : MF) {
118 for (MachineInstr &MI : MBB) {
119 for (MachineOperand &MO : MI.operands()) {
120 if (!MO.isReg())
121 continue;
122
123 const Register VirtReg = MO.getReg();
124 if (VirtReg.isPhysical())
125 continue;
126
127 if (!VRM->hasPhys(VirtReg))
128 continue;
129
130 Register PhysReg = VRM->getPhys(VirtReg);
131 const unsigned SubReg = MO.getSubReg();
132 if (SubReg != 0) {
133 PhysReg = TRI->getSubReg(PhysReg, SubReg);
134 MO.setSubReg(0);
135 }
136
137 MO.setReg(PhysReg);
138 MO.setIsRenamable(false);
139 }
140 }
141 }
142
143 SIMachineFunctionInfo *MFI = MF.getInfo<SIMachineFunctionInfo>();
144
145 for (unsigned Reg : RegsToRewrite) {
146 LIS->removeInterval(Reg);
147
148 const Register PhysReg = VRM->getPhys(Reg);
149 assert(PhysReg != 0);
150
151 MFI->reserveWWMRegister(PhysReg);
152 }
153
154 RegsToRewrite.clear();
155
156 // Update the set of reserved registers to include WWM ones.
157 MRI->freezeReservedRegs(MF);
158}
159
160#ifndef NDEBUG
162SIPreAllocateWWMRegs::printWWMInfo(const MachineInstr &MI) {
163
164 unsigned Opc = MI.getOpcode();
165
166 if (Opc == AMDGPU::ENTER_STRICT_WWM || Opc == AMDGPU::ENTER_STRICT_WQM ||
167 Opc == AMDGPU::ENTER_PSEUDO_WM) {
168 dbgs() << "Entering ";
169 } else {
170 assert(Opc == AMDGPU::EXIT_STRICT_WWM || Opc == AMDGPU::EXIT_STRICT_WQM ||
171 Opc == AMDGPU::EXIT_PSEUDO_WM);
172 dbgs() << "Exiting ";
173 }
174
175 if (Opc == AMDGPU::ENTER_STRICT_WWM || Opc == AMDGPU::EXIT_STRICT_WWM) {
176 dbgs() << "Strict WWM ";
177 } else if (Opc == AMDGPU::ENTER_PSEUDO_WM || Opc == AMDGPU::EXIT_PSEUDO_WM) {
178 dbgs() << "Pseudo WWM/WQM ";
179 } else {
180 assert(Opc == AMDGPU::ENTER_STRICT_WQM || Opc == AMDGPU::EXIT_STRICT_WQM);
181 dbgs() << "Strict WQM ";
182 }
183
184 dbgs() << "region: " << MI;
185}
186
187#endif
188
189bool SIPreAllocateWWMRegs::runOnMachineFunction(MachineFunction &MF) {
190 LLVM_DEBUG(dbgs() << "SIPreAllocateWWMRegs: function " << MF.getName() << "\n");
191
193
194 TII = ST.getInstrInfo();
195 TRI = &TII->getRegisterInfo();
196 MRI = &MF.getRegInfo();
197
198 LIS = &getAnalysis<LiveIntervals>();
199 Matrix = &getAnalysis<LiveRegMatrix>();
200 VRM = &getAnalysis<VirtRegMap>();
201
202 RegClassInfo.runOnMachineFunction(MF);
203
204 bool RegsAssigned = false;
205
206 // We use a reverse post-order traversal of the control-flow graph to
207 // guarantee that we visit definitions in dominance order. Since WWM
208 // expressions are guaranteed to never involve phi nodes, and we can only
209 // escape WWM through the special WWM instruction, this means that this is a
210 // perfect elimination order, so we can never do any better.
212
213 for (MachineBasicBlock *MBB : RPOT) {
214 bool InWWM = false;
215 for (MachineInstr &MI : *MBB) {
216 if (MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B32 ||
217 MI.getOpcode() == AMDGPU::V_SET_INACTIVE_B64)
218 RegsAssigned |= processDef(MI.getOperand(0));
219
220 if (MI.getOpcode() == AMDGPU::ENTER_STRICT_WWM ||
221 MI.getOpcode() == AMDGPU::ENTER_STRICT_WQM ||
222 MI.getOpcode() == AMDGPU::ENTER_PSEUDO_WM) {
223 LLVM_DEBUG(printWWMInfo(MI));
224 InWWM = true;
225 continue;
226 }
227
228 if (MI.getOpcode() == AMDGPU::EXIT_STRICT_WWM ||
229 MI.getOpcode() == AMDGPU::EXIT_STRICT_WQM ||
230 MI.getOpcode() == AMDGPU::EXIT_PSEUDO_WM) {
231 LLVM_DEBUG(printWWMInfo(MI));
232 InWWM = false;
233 }
234
235 if (!InWWM)
236 continue;
237
238 LLVM_DEBUG(dbgs() << "Processing " << MI);
239
240 for (MachineOperand &DefOpnd : MI.defs()) {
241 RegsAssigned |= processDef(DefOpnd);
242 }
243 }
244 }
245
246 if (!RegsAssigned)
247 return false;
248
249 rewriteRegs(MF);
250 return true;
251}
unsigned SubReg
unsigned const MachineRegisterInfo * MRI
MachineBasicBlock & MBB
Provides AMDGPU specific target descriptions.
#define LLVM_DUMP_METHOD
Mark debug helper function definitions like dump() that should not be stripped from debug builds.
Definition: Compiler.h:492
#define LLVM_DEBUG(X)
Definition: Debug.h:101
AMD GCN specific subclass of TargetSubtarget.
const HexagonInstrInfo * TII
IRTranslator LLVM IR MI
Live Register Matrix
unsigned const TargetRegisterInfo * TRI
#define INITIALIZE_PASS_DEPENDENCY(depName)
Definition: PassSupport.h:55
#define INITIALIZE_PASS_END(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:59
#define INITIALIZE_PASS_BEGIN(passName, arg, name, cfg, analysis)
Definition: PassSupport.h:52
This file builds on the ADT/GraphTraits.h file to build a generic graph post order iterator.
assert(ImpDefSCC.getReg()==AMDGPU::SCC &&ImpDefSCC.isDef())
SI Pre allocate WWM Registers
#define DEBUG_TYPE
Represent the analysis usage information of a pass.
AnalysisUsage & addRequired()
AnalysisUsage & addPreserved()
Add the specified Pass class to the set of analyses preserved by this pass.
void setPreservesCFG()
This function should be called by the pass, iff they do not:
Definition: Pass.cpp:265
FunctionPass class - This class is used to implement most global optimizations.
Definition: Pass.h:311
LiveInterval - This class represents the liveness of a register, or stack slot.
Definition: LiveInterval.h:686
@ IK_Free
No interference, go ahead and assign.
Definition: LiveRegMatrix.h:85
Wrapper class representing physical registers. Should be passed by value.
Definition: MCRegister.h:24
MachineFunctionPass - This class adapts the FunctionPass interface to allow convenient creation of pa...
void getAnalysisUsage(AnalysisUsage &AU) const override
getAnalysisUsage - Subclasses that override getAnalysisUsage must call this.
virtual bool runOnMachineFunction(MachineFunction &MF)=0
runOnMachineFunction - This method must be overloaded to perform the desired machine code transformat...
const TargetSubtargetInfo & getSubtarget() const
getSubtarget - Return the subtarget for which this machine code is being compiled.
StringRef getName() const
getName - Return the name of the corresponding LLVM function.
MachineRegisterInfo & getRegInfo()
getRegInfo - Return information about the registers currently in use.
Representation of each machine instruction.
Definition: MachineInstr.h:68
MachineOperand class - Representation of each machine instruction operand.
void setSubReg(unsigned subReg)
unsigned getSubReg() const
void setIsRenamable(bool Val=true)
bool isReg() const
isReg - Tests if this is a MO_Register operand.
void setReg(Register Reg)
Change the register this operand corresponds to.
Register getReg() const
getReg - Returns the register number.
MachineRegisterInfo - Keep track of information for virtual and physical registers,...
static PassRegistry * getPassRegistry()
getPassRegistry - Access the global registry object, which is automatically initialized at applicatio...
Wrapper class representing virtual and physical registers.
Definition: Register.h:19
constexpr bool isPhysical() const
Return true if the specified register number is in the physical register namespace.
Definition: Register.h:95
This class keeps track of the SPI_SP_INPUT_ADDR config register, which tells the hardware which inter...
SlotIndexes pass.
Definition: SlotIndexes.h:319
#define llvm_unreachable(msg)
Marks that the current location is not supposed to be reachable.
unsigned ID
LLVM IR allows to use arbitrary numbers as calling convention identifiers.
Definition: CallingConv.h:24
This is an optimization pass for GlobalISel generic memory operations.
Definition: AddressRanges.h:18
char & SIPreAllocateWWMRegsID
FunctionPass * createSIPreAllocateWWMRegsPass()
raw_ostream & dbgs()
dbgs() - This returns a reference to a raw_ostream for debugging messages.
Definition: Debug.cpp:163
void initializeSIPreAllocateWWMRegsPass(PassRegistry &)