xref: /freebsd/contrib/llvm-project/llvm/lib/Target/AArch64/AArch64Subtarget.cpp (revision 8bcb0991864975618c09697b1aca10683346d9f0)
1 //===-- AArch64Subtarget.cpp - AArch64 Subtarget Information ----*- 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 //
9 // This file implements the AArch64 specific subclass of TargetSubtarget.
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "AArch64Subtarget.h"
14 
15 #include "AArch64.h"
16 #include "AArch64CallLowering.h"
17 #include "AArch64InstrInfo.h"
18 #include "AArch64LegalizerInfo.h"
19 #include "AArch64PBQPRegAlloc.h"
20 #include "AArch64RegisterBankInfo.h"
21 #include "AArch64TargetMachine.h"
22 #include "MCTargetDesc/AArch64AddressingModes.h"
23 #include "llvm/CodeGen/GlobalISel/InstructionSelect.h"
24 #include "llvm/CodeGen/MachineScheduler.h"
25 #include "llvm/IR/GlobalValue.h"
26 #include "llvm/Support/TargetParser.h"
27 
28 using namespace llvm;
29 
30 #define DEBUG_TYPE "aarch64-subtarget"
31 
32 #define GET_SUBTARGETINFO_CTOR
33 #define GET_SUBTARGETINFO_TARGET_DESC
34 #include "AArch64GenSubtargetInfo.inc"
35 
36 static cl::opt<bool>
37 EnableEarlyIfConvert("aarch64-early-ifcvt", cl::desc("Enable the early if "
38                      "converter pass"), cl::init(true), cl::Hidden);
39 
40 // If OS supports TBI, use this flag to enable it.
41 static cl::opt<bool>
42 UseAddressTopByteIgnored("aarch64-use-tbi", cl::desc("Assume that top byte of "
43                          "an address is ignored"), cl::init(false), cl::Hidden);
44 
45 static cl::opt<bool>
46     UseNonLazyBind("aarch64-enable-nonlazybind",
47                    cl::desc("Call nonlazybind functions via direct GOT load"),
48                    cl::init(false), cl::Hidden);
49 
50 AArch64Subtarget &
51 AArch64Subtarget::initializeSubtargetDependencies(StringRef FS,
52                                                   StringRef CPUString) {
53   // Determine default and user-specified characteristics
54 
55   if (CPUString.empty())
56     CPUString = "generic";
57 
58   ParseSubtargetFeatures(CPUString, FS);
59   initializeProperties();
60 
61   return *this;
62 }
63 
64 void AArch64Subtarget::initializeProperties() {
65   // Initialize CPU specific properties. We should add a tablegen feature for
66   // this in the future so we can specify it together with the subtarget
67   // features.
68   switch (ARMProcFamily) {
69   case Others:
70     break;
71   case CortexA35:
72     break;
73   case CortexA53:
74     PrefFunctionLogAlignment = 3;
75     break;
76   case CortexA55:
77     break;
78   case CortexA57:
79     MaxInterleaveFactor = 4;
80     PrefFunctionLogAlignment = 4;
81     break;
82   case CortexA65:
83     PrefFunctionLogAlignment = 3;
84     break;
85   case CortexA72:
86   case CortexA73:
87   case CortexA75:
88   case CortexA76:
89     PrefFunctionLogAlignment = 4;
90     break;
91   case Cyclone:
92     CacheLineSize = 64;
93     PrefetchDistance = 280;
94     MinPrefetchStride = 2048;
95     MaxPrefetchIterationsAhead = 3;
96     break;
97   case ExynosM1:
98     MaxInterleaveFactor = 4;
99     MaxJumpTableSize = 8;
100     PrefFunctionLogAlignment = 4;
101     PrefLoopLogAlignment = 3;
102     break;
103   case ExynosM3:
104     MaxInterleaveFactor = 4;
105     MaxJumpTableSize = 20;
106     PrefFunctionLogAlignment = 5;
107     PrefLoopLogAlignment = 4;
108     break;
109   case Falkor:
110     MaxInterleaveFactor = 4;
111     // FIXME: remove this to enable 64-bit SLP if performance looks good.
112     MinVectorRegisterBitWidth = 128;
113     CacheLineSize = 128;
114     PrefetchDistance = 820;
115     MinPrefetchStride = 2048;
116     MaxPrefetchIterationsAhead = 8;
117     break;
118   case Kryo:
119     MaxInterleaveFactor = 4;
120     VectorInsertExtractBaseCost = 2;
121     CacheLineSize = 128;
122     PrefetchDistance = 740;
123     MinPrefetchStride = 1024;
124     MaxPrefetchIterationsAhead = 11;
125     // FIXME: remove this to enable 64-bit SLP if performance looks good.
126     MinVectorRegisterBitWidth = 128;
127     break;
128   case NeoverseE1:
129     PrefFunctionLogAlignment = 3;
130     break;
131   case NeoverseN1:
132     PrefFunctionLogAlignment = 4;
133     break;
134   case Saphira:
135     MaxInterleaveFactor = 4;
136     // FIXME: remove this to enable 64-bit SLP if performance looks good.
137     MinVectorRegisterBitWidth = 128;
138     break;
139   case ThunderX2T99:
140     CacheLineSize = 64;
141     PrefFunctionLogAlignment = 3;
142     PrefLoopLogAlignment = 2;
143     MaxInterleaveFactor = 4;
144     PrefetchDistance = 128;
145     MinPrefetchStride = 1024;
146     MaxPrefetchIterationsAhead = 4;
147     // FIXME: remove this to enable 64-bit SLP if performance looks good.
148     MinVectorRegisterBitWidth = 128;
149     break;
150   case ThunderX:
151   case ThunderXT88:
152   case ThunderXT81:
153   case ThunderXT83:
154     CacheLineSize = 128;
155     PrefFunctionLogAlignment = 3;
156     PrefLoopLogAlignment = 2;
157     // FIXME: remove this to enable 64-bit SLP if performance looks good.
158     MinVectorRegisterBitWidth = 128;
159     break;
160   case TSV110:
161     CacheLineSize = 64;
162     PrefFunctionLogAlignment = 4;
163     PrefLoopLogAlignment = 2;
164     break;
165   }
166 }
167 
168 AArch64Subtarget::AArch64Subtarget(const Triple &TT, const std::string &CPU,
169                                    const std::string &FS,
170                                    const TargetMachine &TM, bool LittleEndian)
171     : AArch64GenSubtargetInfo(TT, CPU, FS),
172       ReserveXRegister(AArch64::GPR64commonRegClass.getNumRegs()),
173       CustomCallSavedXRegs(AArch64::GPR64commonRegClass.getNumRegs()),
174       IsLittle(LittleEndian),
175       TargetTriple(TT), FrameLowering(),
176       InstrInfo(initializeSubtargetDependencies(FS, CPU)), TSInfo(),
177       TLInfo(TM, *this) {
178   if (AArch64::isX18ReservedByDefault(TT))
179     ReserveXRegister.set(18);
180 
181   CallLoweringInfo.reset(new AArch64CallLowering(*getTargetLowering()));
182   Legalizer.reset(new AArch64LegalizerInfo(*this));
183 
184   auto *RBI = new AArch64RegisterBankInfo(*getRegisterInfo());
185 
186   // FIXME: At this point, we can't rely on Subtarget having RBI.
187   // It's awkward to mix passing RBI and the Subtarget; should we pass
188   // TII/TRI as well?
189   InstSelector.reset(createAArch64InstructionSelector(
190       *static_cast<const AArch64TargetMachine *>(&TM), *this, *RBI));
191 
192   RegBankInfo.reset(RBI);
193 }
194 
195 const CallLowering *AArch64Subtarget::getCallLowering() const {
196   return CallLoweringInfo.get();
197 }
198 
199 InstructionSelector *AArch64Subtarget::getInstructionSelector() const {
200   return InstSelector.get();
201 }
202 
203 const LegalizerInfo *AArch64Subtarget::getLegalizerInfo() const {
204   return Legalizer.get();
205 }
206 
207 const RegisterBankInfo *AArch64Subtarget::getRegBankInfo() const {
208   return RegBankInfo.get();
209 }
210 
211 /// Find the target operand flags that describe how a global value should be
212 /// referenced for the current subtarget.
213 unsigned
214 AArch64Subtarget::ClassifyGlobalReference(const GlobalValue *GV,
215                                           const TargetMachine &TM) const {
216   // MachO large model always goes via a GOT, simply to get a single 8-byte
217   // absolute relocation on all global addresses.
218   if (TM.getCodeModel() == CodeModel::Large && isTargetMachO())
219     return AArch64II::MO_GOT;
220 
221   if (!TM.shouldAssumeDSOLocal(*GV->getParent(), GV)) {
222     if (GV->hasDLLImportStorageClass())
223       return AArch64II::MO_GOT | AArch64II::MO_DLLIMPORT;
224     if (getTargetTriple().isOSWindows())
225       return AArch64II::MO_GOT | AArch64II::MO_COFFSTUB;
226     return AArch64II::MO_GOT;
227   }
228 
229   // The small code model's direct accesses use ADRP, which cannot
230   // necessarily produce the value 0 (if the code is above 4GB).
231   // Same for the tiny code model, where we have a pc relative LDR.
232   if ((useSmallAddressing() || TM.getCodeModel() == CodeModel::Tiny) &&
233       GV->hasExternalWeakLinkage())
234     return AArch64II::MO_GOT;
235 
236   // References to tagged globals are marked with MO_NC | MO_TAGGED to indicate
237   // that their nominal addresses are tagged and outside of the code model. In
238   // AArch64ExpandPseudo::expandMI we emit an additional instruction to set the
239   // tag if necessary based on MO_TAGGED.
240   if (AllowTaggedGlobals && !isa<FunctionType>(GV->getValueType()))
241     return AArch64II::MO_NC | AArch64II::MO_TAGGED;
242 
243   return AArch64II::MO_NO_FLAG;
244 }
245 
246 unsigned AArch64Subtarget::classifyGlobalFunctionReference(
247     const GlobalValue *GV, const TargetMachine &TM) const {
248   // MachO large model always goes via a GOT, because we don't have the
249   // relocations available to do anything else..
250   if (TM.getCodeModel() == CodeModel::Large && isTargetMachO() &&
251       !GV->hasInternalLinkage())
252     return AArch64II::MO_GOT;
253 
254   // NonLazyBind goes via GOT unless we know it's available locally.
255   auto *F = dyn_cast<Function>(GV);
256   if (UseNonLazyBind && F && F->hasFnAttribute(Attribute::NonLazyBind) &&
257       !TM.shouldAssumeDSOLocal(*GV->getParent(), GV))
258     return AArch64II::MO_GOT;
259 
260   return AArch64II::MO_NO_FLAG;
261 }
262 
263 void AArch64Subtarget::overrideSchedPolicy(MachineSchedPolicy &Policy,
264                                            unsigned NumRegionInstrs) const {
265   // LNT run (at least on Cyclone) showed reasonably significant gains for
266   // bi-directional scheduling. 253.perlbmk.
267   Policy.OnlyTopDown = false;
268   Policy.OnlyBottomUp = false;
269   // Enabling or Disabling the latency heuristic is a close call: It seems to
270   // help nearly no benchmark on out-of-order architectures, on the other hand
271   // it regresses register pressure on a few benchmarking.
272   Policy.DisableLatencyHeuristic = DisableLatencySchedHeuristic;
273 }
274 
275 bool AArch64Subtarget::enableEarlyIfConversion() const {
276   return EnableEarlyIfConvert;
277 }
278 
279 bool AArch64Subtarget::supportsAddressTopByteIgnored() const {
280   if (!UseAddressTopByteIgnored)
281     return false;
282 
283   if (TargetTriple.isiOS()) {
284     unsigned Major, Minor, Micro;
285     TargetTriple.getiOSVersion(Major, Minor, Micro);
286     return Major >= 8;
287   }
288 
289   return false;
290 }
291 
292 std::unique_ptr<PBQPRAConstraint>
293 AArch64Subtarget::getCustomPBQPConstraints() const {
294   return balanceFPOps() ? std::make_unique<A57ChainingConstraint>() : nullptr;
295 }
296 
297 void AArch64Subtarget::mirFileLoaded(MachineFunction &MF) const {
298   // We usually compute max call frame size after ISel. Do the computation now
299   // if the .mir file didn't specify it. Note that this will probably give you
300   // bogus values after PEI has eliminated the callframe setup/destroy pseudo
301   // instructions, specify explicitly if you need it to be correct.
302   MachineFrameInfo &MFI = MF.getFrameInfo();
303   if (!MFI.isMaxCallFrameSizeComputed())
304     MFI.computeMaxCallFrameSize(MF);
305 }
306