[RISCV] Fix mgather -> riscv.masked.strided.load combine not extending indices (...
[llvm-project.git] / llvm / lib / CodeGen / RegUsageInfoCollector.cpp
blob6657cf3c1ef4abcdf162c1bc2108c847128d7a5e
1 //===-- RegUsageInfoCollector.cpp - Register Usage Information Collector --===//
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 /// This pass is required to take advantage of the interprocedural register
10 /// allocation infrastructure.
11 ///
12 /// This pass is simple MachineFunction pass which collects register usage
13 /// details by iterating through each physical registers and checking
14 /// MRI::isPhysRegUsed() then creates a RegMask based on this details.
15 /// The pass then stores this RegMask in PhysicalRegisterUsageInfo.cpp
16 ///
17 //===----------------------------------------------------------------------===//
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/CodeGen/MachineFunctionPass.h"
21 #include "llvm/CodeGen/MachineOperand.h"
22 #include "llvm/CodeGen/MachineRegisterInfo.h"
23 #include "llvm/CodeGen/Passes.h"
24 #include "llvm/CodeGen/RegisterUsageInfo.h"
25 #include "llvm/CodeGen/TargetFrameLowering.h"
26 #include "llvm/IR/Function.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/raw_ostream.h"
30 using namespace llvm;
32 #define DEBUG_TYPE "ip-regalloc"
34 STATISTIC(NumCSROpt,
35 "Number of functions optimized for callee saved registers");
37 namespace {
39 class RegUsageInfoCollector : public MachineFunctionPass {
40 public:
41 RegUsageInfoCollector() : MachineFunctionPass(ID) {
42 PassRegistry &Registry = *PassRegistry::getPassRegistry();
43 initializeRegUsageInfoCollectorPass(Registry);
46 StringRef getPassName() const override {
47 return "Register Usage Information Collector Pass";
50 void getAnalysisUsage(AnalysisUsage &AU) const override {
51 AU.addRequired<PhysicalRegisterUsageInfo>();
52 AU.setPreservesAll();
53 MachineFunctionPass::getAnalysisUsage(AU);
56 bool runOnMachineFunction(MachineFunction &MF) override;
58 // Call getCalleeSaves and then also set the bits for subregs and
59 // fully saved superregs.
60 static void computeCalleeSavedRegs(BitVector &SavedRegs, MachineFunction &MF);
62 static char ID;
65 } // end of anonymous namespace
67 char RegUsageInfoCollector::ID = 0;
69 INITIALIZE_PASS_BEGIN(RegUsageInfoCollector, "RegUsageInfoCollector",
70 "Register Usage Information Collector", false, false)
71 INITIALIZE_PASS_DEPENDENCY(PhysicalRegisterUsageInfo)
72 INITIALIZE_PASS_END(RegUsageInfoCollector, "RegUsageInfoCollector",
73 "Register Usage Information Collector", false, false)
75 FunctionPass *llvm::createRegUsageInfoCollector() {
76 return new RegUsageInfoCollector();
79 // TODO: Move to hook somwehere?
81 // Return true if it is useful to track the used registers for IPRA / no CSR
82 // optimizations. This is not useful for entry points, and computing the
83 // register usage information is expensive.
84 static bool isCallableFunction(const MachineFunction &MF) {
85 switch (MF.getFunction().getCallingConv()) {
86 case CallingConv::AMDGPU_VS:
87 case CallingConv::AMDGPU_GS:
88 case CallingConv::AMDGPU_PS:
89 case CallingConv::AMDGPU_CS:
90 case CallingConv::AMDGPU_HS:
91 case CallingConv::AMDGPU_ES:
92 case CallingConv::AMDGPU_LS:
93 case CallingConv::AMDGPU_KERNEL:
94 return false;
95 default:
96 return true;
100 bool RegUsageInfoCollector::runOnMachineFunction(MachineFunction &MF) {
101 MachineRegisterInfo *MRI = &MF.getRegInfo();
102 const TargetRegisterInfo *TRI = MF.getSubtarget().getRegisterInfo();
103 const LLVMTargetMachine &TM = MF.getTarget();
105 LLVM_DEBUG(dbgs() << " -------------------- " << getPassName()
106 << " -------------------- \nFunction Name : "
107 << MF.getName() << '\n');
109 // Analyzing the register usage may be expensive on some targets.
110 if (!isCallableFunction(MF)) {
111 LLVM_DEBUG(dbgs() << "Not analyzing non-callable function\n");
112 return false;
115 // If there are no callers, there's no point in computing more precise
116 // register usage here.
117 if (MF.getFunction().use_empty()) {
118 LLVM_DEBUG(dbgs() << "Not analyzing function with no callers\n");
119 return false;
122 std::vector<uint32_t> RegMask;
124 // Compute the size of the bit vector to represent all the registers.
125 // The bit vector is broken into 32-bit chunks, thus takes the ceil of
126 // the number of registers divided by 32 for the size.
127 unsigned RegMaskSize = MachineOperand::getRegMaskSize(TRI->getNumRegs());
128 RegMask.resize(RegMaskSize, ~((uint32_t)0));
130 const Function &F = MF.getFunction();
132 PhysicalRegisterUsageInfo &PRUI = getAnalysis<PhysicalRegisterUsageInfo>();
133 PRUI.setTargetMachine(TM);
135 LLVM_DEBUG(dbgs() << "Clobbered Registers: ");
137 BitVector SavedRegs;
138 computeCalleeSavedRegs(SavedRegs, MF);
140 const BitVector &UsedPhysRegsMask = MRI->getUsedPhysRegsMask();
141 auto SetRegAsDefined = [&RegMask] (unsigned Reg) {
142 RegMask[Reg / 32] &= ~(1u << Reg % 32);
145 // Some targets can clobber registers "inside" a call, typically in
146 // linker-generated code.
147 for (const MCPhysReg Reg : TRI->getIntraCallClobberedRegs(&MF))
148 for (MCRegAliasIterator AI(Reg, TRI, true); AI.isValid(); ++AI)
149 SetRegAsDefined(*AI);
151 // Scan all the physical registers. When a register is defined in the current
152 // function set it and all the aliasing registers as defined in the regmask.
153 // FIXME: Rewrite to use regunits.
154 for (unsigned PReg = 1, PRegE = TRI->getNumRegs(); PReg < PRegE; ++PReg) {
155 // Don't count registers that are saved and restored.
156 if (SavedRegs.test(PReg))
157 continue;
158 // If a register is defined by an instruction mark it as defined together
159 // with all it's unsaved aliases.
160 if (!MRI->def_empty(PReg)) {
161 for (MCRegAliasIterator AI(PReg, TRI, true); AI.isValid(); ++AI)
162 if (!SavedRegs.test(*AI))
163 SetRegAsDefined(*AI);
164 continue;
166 // If a register is in the UsedPhysRegsMask set then mark it as defined.
167 // All clobbered aliases will also be in the set, so we can skip setting
168 // as defined all the aliases here.
169 if (UsedPhysRegsMask.test(PReg))
170 SetRegAsDefined(PReg);
173 if (TargetFrameLowering::isSafeForNoCSROpt(F) &&
174 MF.getSubtarget().getFrameLowering()->isProfitableForNoCSROpt(F)) {
175 ++NumCSROpt;
176 LLVM_DEBUG(dbgs() << MF.getName()
177 << " function optimized for not having CSR.\n");
180 LLVM_DEBUG(
181 for (unsigned PReg = 1, PRegE = TRI->getNumRegs(); PReg < PRegE; ++PReg) {
182 if (MachineOperand::clobbersPhysReg(&(RegMask[0]), PReg))
183 dbgs() << printReg(PReg, TRI) << " ";
186 dbgs() << " \n----------------------------------------\n";
189 PRUI.storeUpdateRegUsageInfo(F, RegMask);
191 return false;
194 void RegUsageInfoCollector::
195 computeCalleeSavedRegs(BitVector &SavedRegs, MachineFunction &MF) {
196 const TargetFrameLowering &TFI = *MF.getSubtarget().getFrameLowering();
197 const TargetRegisterInfo &TRI = *MF.getSubtarget().getRegisterInfo();
199 // Target will return the set of registers that it saves/restores as needed.
200 SavedRegs.clear();
201 TFI.getCalleeSaves(MF, SavedRegs);
202 if (SavedRegs.none())
203 return;
205 // Insert subregs.
206 const MCPhysReg *CSRegs = TRI.getCalleeSavedRegs(&MF);
207 for (unsigned i = 0; CSRegs[i]; ++i) {
208 MCPhysReg Reg = CSRegs[i];
209 if (SavedRegs.test(Reg)) {
210 // Save subregisters
211 for (MCPhysReg SR : TRI.subregs(Reg))
212 SavedRegs.set(SR);