1 //===-- ARMTargetMachine.cpp - Define TargetMachine for ARM ---------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
11 //===----------------------------------------------------------------------===//
13 #include "ARMTargetMachine.h"
15 #include "ARMFrameLowering.h"
16 #include "llvm/CodeGen/Passes.h"
17 #include "llvm/MC/MCAsmInfo.h"
18 #include "llvm/PassManager.h"
19 #include "llvm/Support/CommandLine.h"
20 #include "llvm/Support/FormattedStream.h"
21 #include "llvm/Support/TargetRegistry.h"
22 #include "llvm/Target/TargetOptions.h"
23 #include "llvm/Transforms/Scalar.h"
27 EnableGlobalMerge("global-merge", cl::Hidden,
28 cl::desc("Enable global merge pass"),
32 DisableA15SDOptimization("disable-a15-sd-optimization", cl::Hidden,
33 cl::desc("Inhibit optimization of S->D register accesses on A15"),
36 extern "C" void LLVMInitializeARMTarget() {
37 // Register the target.
38 RegisterTargetMachine<ARMTargetMachine> X(TheARMTarget);
39 RegisterTargetMachine<ThumbTargetMachine> Y(TheThumbTarget);
43 /// TargetMachine ctor - Create an ARM architecture model.
45 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, StringRef TT,
46 StringRef CPU, StringRef FS,
47 const TargetOptions &Options,
48 Reloc::Model RM, CodeModel::Model CM,
50 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
51 Subtarget(TT, CPU, FS, Options),
53 InstrItins(Subtarget.getInstrItineraryData()) {
55 // Default to triple-appropriate float ABI
56 if (Options.FloatABIType == FloatABI::Default)
57 this->Options.FloatABIType =
58 Subtarget.isTargetHardFloat() ? FloatABI::Hard : FloatABI::Soft;
61 void ARMBaseTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
62 // Add first the target-independent BasicTTI pass, then our ARM pass. This
63 // allows the ARM pass to delegate to the target independent layer when
65 PM.add(createBasicTargetTransformInfoPass(this));
66 PM.add(createARMTargetTransformInfoPass(this));
70 void ARMTargetMachine::anchor() { }
72 static std::string computeDataLayout(ARMSubtarget &ST) {
73 // Little endian. Pointers are 32 bits and aligned to 32 bits.
74 std::string Ret = "e-p:32:32";
76 // On thumb, i16,i18 and i1 have natural aligment requirements, but we try to
79 Ret += "-i1:8:32-i8:8:32-i16:16:32";
81 // ABIs other than APC have 64 bit integers with natural alignment.
85 // We have 64 bits floats. The APCS ABI requires them to be aligned to 32
86 // bits, others to 64 bits. We always try to align to 64 bits.
90 // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others
91 // to 64. We always ty to give them natural alignment.
93 Ret += "-v64:32:64-v128:32:128";
95 Ret += "-v128:64:128";
97 // On thumb and APCS, only try to align aggregates to 32 bits (the default is
99 if (ST.isThumb() || ST.isAPCS_ABI())
102 // Integer registers are 32 bits.
105 // The stack is 64 bit aligned on AAPCS and 32 bit aligned everywhere else.
106 if (ST.isAAPCS_ABI())
114 ARMTargetMachine::ARMTargetMachine(const Target &T, StringRef TT,
115 StringRef CPU, StringRef FS,
116 const TargetOptions &Options,
117 Reloc::Model RM, CodeModel::Model CM,
118 CodeGenOpt::Level OL)
119 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
120 InstrInfo(Subtarget),
121 DL(computeDataLayout(Subtarget)),
124 FrameLowering(Subtarget) {
126 if (!Subtarget.hasARMOps())
127 report_fatal_error("CPU: '" + Subtarget.getCPUString() + "' does not "
128 "support ARM mode execution!");
131 void ThumbTargetMachine::anchor() { }
133 ThumbTargetMachine::ThumbTargetMachine(const Target &T, StringRef TT,
134 StringRef CPU, StringRef FS,
135 const TargetOptions &Options,
136 Reloc::Model RM, CodeModel::Model CM,
137 CodeGenOpt::Level OL)
138 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
139 InstrInfo(Subtarget.hasThumb2()
140 ? ((ARMBaseInstrInfo*)new Thumb2InstrInfo(Subtarget))
141 : ((ARMBaseInstrInfo*)new Thumb1InstrInfo(Subtarget))),
142 DL(computeDataLayout(Subtarget)),
145 FrameLowering(Subtarget.hasThumb2()
146 ? new ARMFrameLowering(Subtarget)
147 : (ARMFrameLowering*)new Thumb1FrameLowering(Subtarget)) {
152 /// ARM Code Generator Pass Configuration Options.
153 class ARMPassConfig : public TargetPassConfig {
155 ARMPassConfig(ARMBaseTargetMachine *TM, PassManagerBase &PM)
156 : TargetPassConfig(TM, PM) {}
158 ARMBaseTargetMachine &getARMTargetMachine() const {
159 return getTM<ARMBaseTargetMachine>();
162 const ARMSubtarget &getARMSubtarget() const {
163 return *getARMTargetMachine().getSubtargetImpl();
166 virtual bool addPreISel();
167 virtual bool addInstSelector();
168 virtual bool addPreRegAlloc();
169 virtual bool addPreSched2();
170 virtual bool addPreEmitPass();
174 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) {
175 return new ARMPassConfig(this, PM);
178 bool ARMPassConfig::addPreISel() {
179 if (TM->getOptLevel() != CodeGenOpt::None && EnableGlobalMerge)
180 addPass(createGlobalMergePass(TM));
185 bool ARMPassConfig::addInstSelector() {
186 addPass(createARMISelDag(getARMTargetMachine(), getOptLevel()));
188 const ARMSubtarget *Subtarget = &getARMSubtarget();
189 if (Subtarget->isTargetELF() && !Subtarget->isThumb1Only() &&
190 TM->Options.EnableFastISel)
191 addPass(createARMGlobalBaseRegPass());
195 bool ARMPassConfig::addPreRegAlloc() {
196 // FIXME: temporarily disabling load / store optimization pass for Thumb1.
197 if (getOptLevel() != CodeGenOpt::None && !getARMSubtarget().isThumb1Only())
198 addPass(createARMLoadStoreOptimizationPass(true));
199 if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA9())
200 addPass(createMLxExpansionPass());
201 // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be
202 // enabled when NEON is available.
203 if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA15() &&
204 getARMSubtarget().hasNEON() && !DisableA15SDOptimization) {
205 addPass(createA15SDOptimizerPass());
210 bool ARMPassConfig::addPreSched2() {
211 // FIXME: temporarily disabling load / store optimization pass for Thumb1.
212 if (getOptLevel() != CodeGenOpt::None) {
213 if (!getARMSubtarget().isThumb1Only()) {
214 addPass(createARMLoadStoreOptimizationPass());
215 printAndVerify("After ARM load / store optimizer");
217 if (getARMSubtarget().hasNEON())
218 addPass(createExecutionDependencyFixPass(&ARM::DPRRegClass));
221 // Expand some pseudo instructions into multiple instructions to allow
222 // proper scheduling.
223 addPass(createARMExpandPseudoPass());
225 if (getOptLevel() != CodeGenOpt::None) {
226 if (!getARMSubtarget().isThumb1Only()) {
227 // in v8, IfConversion depends on Thumb instruction widths
228 if (getARMSubtarget().restrictIT() &&
229 !getARMSubtarget().prefers32BitThumb())
230 addPass(createThumb2SizeReductionPass());
231 addPass(&IfConverterID);
234 if (getARMSubtarget().isThumb2())
235 addPass(createThumb2ITBlockPass());
240 bool ARMPassConfig::addPreEmitPass() {
241 if (getARMSubtarget().isThumb2()) {
242 if (!getARMSubtarget().prefers32BitThumb())
243 addPass(createThumb2SizeReductionPass());
245 // Constant island pass work on unbundled instructions.
246 addPass(&UnpackMachineBundlesID);
249 addPass(createARMConstantIslandPass());
254 bool ARMBaseTargetMachine::addCodeEmitter(PassManagerBase &PM,
255 JITCodeEmitter &JCE) {
256 // Machine code emitter pass for ARM.
257 PM.add(createARMJITCodeEmitterPass(*this, JCE));