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 //===----------------------------------------------------------------------===//
14 #include "ARMFrameLowering.h"
15 #include "ARMTargetMachine.h"
16 #include "ARMTargetObjectFile.h"
17 #include "llvm/CodeGen/Passes.h"
18 #include "llvm/IR/Function.h"
19 #include "llvm/MC/MCAsmInfo.h"
20 #include "llvm/PassManager.h"
21 #include "llvm/Support/CommandLine.h"
22 #include "llvm/Support/FormattedStream.h"
23 #include "llvm/Support/TargetRegistry.h"
24 #include "llvm/Target/TargetOptions.h"
25 #include "llvm/Transforms/Scalar.h"
29 DisableA15SDOptimization("disable-a15-sd-optimization", cl::Hidden,
30 cl::desc("Inhibit optimization of S->D register accesses on A15"),
34 EnableAtomicTidy("arm-atomic-cfg-tidy", cl::Hidden,
35 cl::desc("Run SimplifyCFG after expanding atomic operations"
36 " to make use of cmpxchg flow-based information"),
39 extern "C" void LLVMInitializeARMTarget() {
40 // Register the target.
41 RegisterTargetMachine<ARMLETargetMachine> X(TheARMLETarget);
42 RegisterTargetMachine<ARMBETargetMachine> Y(TheARMBETarget);
43 RegisterTargetMachine<ThumbLETargetMachine> A(TheThumbLETarget);
44 RegisterTargetMachine<ThumbBETargetMachine> B(TheThumbBETarget);
47 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
48 if (TT.isOSBinFormatMachO())
49 return make_unique<TargetLoweringObjectFileMachO>();
51 return make_unique<TargetLoweringObjectFileCOFF>();
52 return make_unique<ARMElfTargetObjectFile>();
55 static ARMBaseTargetMachine::ARMABI
56 computeTargetABI(const Triple &TT, StringRef CPU,
57 const TargetOptions &Options) {
58 if (Options.MCOptions.getABIName().startswith("aapcs"))
59 return ARMBaseTargetMachine::ARM_ABI_AAPCS;
60 else if (Options.MCOptions.getABIName().startswith("apcs"))
61 return ARMBaseTargetMachine::ARM_ABI_APCS;
63 assert(Options.MCOptions.getABIName().empty() &&
64 "Unknown target-abi option!");
66 ARMBaseTargetMachine::ARMABI TargetABI =
67 ARMBaseTargetMachine::ARM_ABI_UNKNOWN;
69 // FIXME: This is duplicated code from the front end and should be unified.
70 if (TT.isOSBinFormatMachO()) {
71 if (TT.getEnvironment() == llvm::Triple::EABI ||
72 (TT.getOS() == llvm::Triple::UnknownOS &&
73 TT.getObjectFormat() == llvm::Triple::MachO) ||
74 CPU.startswith("cortex-m")) {
75 TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
77 TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
79 } else if (TT.isOSWindows()) {
80 // FIXME: this is invalid for WindowsCE
81 TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
83 // Select the default based on the platform.
84 switch (TT.getEnvironment()) {
85 case llvm::Triple::Android:
86 case llvm::Triple::GNUEABI:
87 case llvm::Triple::GNUEABIHF:
88 case llvm::Triple::EABIHF:
89 case llvm::Triple::EABI:
90 TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
92 case llvm::Triple::GNU:
93 TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
96 if (TT.getOS() == llvm::Triple::NetBSD)
97 TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
99 TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
107 static std::string computeDataLayout(const Triple &TT,
108 ARMBaseTargetMachine::ARMABI ABI,
110 std::string Ret = "";
119 Ret += DataLayout::getManglingComponent(TT);
121 // Pointers are 32 bits and aligned to 32 bits.
124 // ABIs other than APCS have 64 bit integers with natural alignment.
125 if (ABI != ARMBaseTargetMachine::ARM_ABI_APCS)
128 // We have 64 bits floats. The APCS ABI requires them to be aligned to 32
129 // bits, others to 64 bits. We always try to align to 64 bits.
130 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS)
133 // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others
134 // to 64. We always ty to give them natural alignment.
135 if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS)
136 Ret += "-v64:32:64-v128:32:128";
138 Ret += "-v128:64:128";
140 // Try to align aggregates to 32 bits (the default is 64 bits, which has no
141 // particular hardware support on 32-bit ARM).
144 // Integer registers are 32 bits.
147 // The stack is 128 bit aligned on NaCl, 64 bit aligned on AAPCS and 32 bit
148 // aligned everywhere else.
151 else if (ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS)
159 /// TargetMachine ctor - Create an ARM architecture model.
161 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, StringRef TT,
162 StringRef CPU, StringRef FS,
163 const TargetOptions &Options,
164 Reloc::Model RM, CodeModel::Model CM,
165 CodeGenOpt::Level OL, bool isLittle)
166 : LLVMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL),
167 TargetABI(computeTargetABI(Triple(TT), CPU, Options)),
168 DL(computeDataLayout(Triple(TT), TargetABI, isLittle)),
169 TLOF(createTLOF(Triple(getTargetTriple()))),
170 Subtarget(TT, CPU, FS, *this, isLittle), isLittle(isLittle) {
172 // Default to triple-appropriate float ABI
173 if (Options.FloatABIType == FloatABI::Default)
174 this->Options.FloatABIType =
175 Subtarget.isTargetHardFloat() ? FloatABI::Hard : FloatABI::Soft;
178 ARMBaseTargetMachine::~ARMBaseTargetMachine() {}
181 ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const {
182 AttributeSet FnAttrs = F.getAttributes();
184 FnAttrs.getAttribute(AttributeSet::FunctionIndex, "target-cpu");
186 FnAttrs.getAttribute(AttributeSet::FunctionIndex, "target-features");
188 std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
189 ? CPUAttr.getValueAsString().str()
191 std::string FS = !FSAttr.hasAttribute(Attribute::None)
192 ? FSAttr.getValueAsString().str()
195 // FIXME: This is related to the code below to reset the target options,
196 // we need to know whether or not the soft float flag is set on the
197 // function before we can generate a subtarget. We also need to use
198 // it as a key for the subtarget since that can be the only difference
199 // between two functions.
201 FnAttrs.getAttribute(AttributeSet::FunctionIndex, "use-soft-float");
202 bool SoftFloat = !SFAttr.hasAttribute(Attribute::None)
203 ? SFAttr.getValueAsString() == "true"
204 : Options.UseSoftFloat;
206 auto &I = SubtargetMap[CPU + FS + (SoftFloat ? "use-soft-float=true"
207 : "use-soft-float=false")];
209 // This needs to be done before we create a new subtarget since any
210 // creation will depend on the TM and the code generation flags on the
211 // function that reside in TargetOptions.
212 resetTargetOptions(F);
213 I = llvm::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle);
218 void ARMBaseTargetMachine::addAnalysisPasses(PassManagerBase &PM) {
219 // Add first the target-independent BasicTTI pass, then our ARM pass. This
220 // allows the ARM pass to delegate to the target independent layer when
222 PM.add(createBasicTargetTransformInfoPass(this));
223 PM.add(createARMTargetTransformInfoPass(this));
227 void ARMTargetMachine::anchor() { }
229 ARMTargetMachine::ARMTargetMachine(const Target &T, StringRef TT, StringRef CPU,
230 StringRef FS, const TargetOptions &Options,
231 Reloc::Model RM, CodeModel::Model CM,
232 CodeGenOpt::Level OL, bool isLittle)
233 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, isLittle) {
235 if (!Subtarget.hasARMOps())
236 report_fatal_error("CPU: '" + Subtarget.getCPUString() + "' does not "
237 "support ARM mode execution!");
240 void ARMLETargetMachine::anchor() { }
242 ARMLETargetMachine::ARMLETargetMachine(const Target &T, StringRef TT,
243 StringRef CPU, StringRef FS,
244 const TargetOptions &Options,
245 Reloc::Model RM, CodeModel::Model CM,
246 CodeGenOpt::Level OL)
247 : ARMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
249 void ARMBETargetMachine::anchor() { }
251 ARMBETargetMachine::ARMBETargetMachine(const Target &T, StringRef TT,
252 StringRef CPU, StringRef FS,
253 const TargetOptions &Options,
254 Reloc::Model RM, CodeModel::Model CM,
255 CodeGenOpt::Level OL)
256 : ARMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
258 void ThumbTargetMachine::anchor() { }
260 ThumbTargetMachine::ThumbTargetMachine(const Target &T, StringRef TT,
261 StringRef CPU, StringRef FS,
262 const TargetOptions &Options,
263 Reloc::Model RM, CodeModel::Model CM,
264 CodeGenOpt::Level OL, bool isLittle)
265 : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL,
270 void ThumbLETargetMachine::anchor() { }
272 ThumbLETargetMachine::ThumbLETargetMachine(const Target &T, StringRef TT,
273 StringRef CPU, StringRef FS,
274 const TargetOptions &Options,
275 Reloc::Model RM, CodeModel::Model CM,
276 CodeGenOpt::Level OL)
277 : ThumbTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
279 void ThumbBETargetMachine::anchor() { }
281 ThumbBETargetMachine::ThumbBETargetMachine(const Target &T, StringRef TT,
282 StringRef CPU, StringRef FS,
283 const TargetOptions &Options,
284 Reloc::Model RM, CodeModel::Model CM,
285 CodeGenOpt::Level OL)
286 : ThumbTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
289 /// ARM Code Generator Pass Configuration Options.
290 class ARMPassConfig : public TargetPassConfig {
292 ARMPassConfig(ARMBaseTargetMachine *TM, PassManagerBase &PM)
293 : TargetPassConfig(TM, PM) {}
295 ARMBaseTargetMachine &getARMTargetMachine() const {
296 return getTM<ARMBaseTargetMachine>();
299 const ARMSubtarget &getARMSubtarget() const {
300 return *getARMTargetMachine().getSubtargetImpl();
303 void addIRPasses() override;
304 bool addPreISel() override;
305 bool addInstSelector() override;
306 void addPreRegAlloc() override;
307 void addPreSched2() override;
308 void addPreEmitPass() override;
312 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) {
313 return new ARMPassConfig(this, PM);
316 void ARMPassConfig::addIRPasses() {
317 if (TM->Options.ThreadModel == ThreadModel::Single)
318 addPass(createLowerAtomicPass());
320 addPass(createAtomicExpandPass(TM));
322 // Cmpxchg instructions are often used with a subsequent comparison to
323 // determine whether it succeeded. We can exploit existing control-flow in
324 // ldrex/strex loops to simplify this, but it needs tidying up.
325 const ARMSubtarget *Subtarget = &getARMSubtarget();
326 if (Subtarget->hasAnyDataBarrier() && !Subtarget->isThumb1Only())
327 if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
328 addPass(createCFGSimplificationPass());
330 TargetPassConfig::addIRPasses();
333 bool ARMPassConfig::addPreISel() {
334 if (TM->getOptLevel() != CodeGenOpt::None)
335 addPass(createGlobalMergePass(TM));
340 bool ARMPassConfig::addInstSelector() {
341 addPass(createARMISelDag(getARMTargetMachine(), getOptLevel()));
343 const ARMSubtarget *Subtarget = &getARMSubtarget();
344 if (Subtarget->isTargetELF() && !Subtarget->isThumb1Only() &&
345 TM->Options.EnableFastISel)
346 addPass(createARMGlobalBaseRegPass());
350 void ARMPassConfig::addPreRegAlloc() {
351 if (getOptLevel() != CodeGenOpt::None)
352 addPass(createARMLoadStoreOptimizationPass(true));
353 if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA9())
354 addPass(createMLxExpansionPass());
355 // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be
356 // enabled when NEON is available.
357 if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA15() &&
358 getARMSubtarget().hasNEON() && !DisableA15SDOptimization) {
359 addPass(createA15SDOptimizerPass());
363 void ARMPassConfig::addPreSched2() {
364 if (getOptLevel() != CodeGenOpt::None) {
365 addPass(createARMLoadStoreOptimizationPass());
367 if (getARMSubtarget().hasNEON())
368 addPass(createExecutionDependencyFixPass(&ARM::DPRRegClass));
371 // Expand some pseudo instructions into multiple instructions to allow
372 // proper scheduling.
373 addPass(createARMExpandPseudoPass());
375 if (getOptLevel() != CodeGenOpt::None) {
376 if (!getARMSubtarget().isThumb1Only()) {
377 // in v8, IfConversion depends on Thumb instruction widths
378 if (getARMSubtarget().restrictIT() &&
379 !getARMSubtarget().prefers32BitThumb())
380 addPass(createThumb2SizeReductionPass());
381 addPass(&IfConverterID);
384 if (getARMSubtarget().isThumb2())
385 addPass(createThumb2ITBlockPass());
388 void ARMPassConfig::addPreEmitPass() {
389 if (getARMSubtarget().isThumb2()) {
390 if (!getARMSubtarget().prefers32BitThumb())
391 addPass(createThumb2SizeReductionPass());
393 // Constant island pass work on unbundled instructions.
394 addPass(&UnpackMachineBundlesID);
397 addPass(createARMOptimizeBarriersPass());
398 addPass(createARMConstantIslandPass());