1 //===-- ARMTargetMachine.cpp - Define TargetMachine for ARM ---------------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 //
11 //===----------------------------------------------------------------------===//
12 
13 #include "ARM.h"
14 #include "ARMFrameLowering.h"
15 #include "ARMTargetMachine.h"
16 #include "ARMTargetObjectFile.h"
17 #include "ARMTargetTransformInfo.h"
18 #include "llvm/CodeGen/Passes.h"
19 #include "llvm/CodeGen/TargetPassConfig.h"
20 #include "llvm/IR/Function.h"
21 #include "llvm/IR/LegacyPassManager.h"
22 #include "llvm/MC/MCAsmInfo.h"
23 #include "llvm/Support/CommandLine.h"
24 #include "llvm/Support/FormattedStream.h"
25 #include "llvm/Support/TargetRegistry.h"
26 #include "llvm/Target/TargetOptions.h"
27 #include "llvm/Transforms/Scalar.h"
28 using namespace llvm;
29 
30 static cl::opt<bool>
31 DisableA15SDOptimization("disable-a15-sd-optimization", cl::Hidden,
32                    cl::desc("Inhibit optimization of S->D register accesses on A15"),
33                    cl::init(false));
34 
35 static cl::opt<bool>
36 EnableAtomicTidy("arm-atomic-cfg-tidy", cl::Hidden,
37                  cl::desc("Run SimplifyCFG after expanding atomic operations"
38                           " to make use of cmpxchg flow-based information"),
39                  cl::init(true));
40 
41 static cl::opt<bool>
42 EnableARMLoadStoreOpt("arm-load-store-opt", cl::Hidden,
43                       cl::desc("Enable ARM load/store optimization pass"),
44                       cl::init(true));
45 
46 // FIXME: Unify control over GlobalMerge.
47 static cl::opt<cl::boolOrDefault>
48 EnableGlobalMerge("arm-global-merge", cl::Hidden,
49                   cl::desc("Enable the global merge pass"));
50 
51 extern "C" void LLVMInitializeARMTarget() {
52   // Register the target.
53   RegisterTargetMachine<ARMLETargetMachine> X(TheARMLETarget);
54   RegisterTargetMachine<ARMBETargetMachine> Y(TheARMBETarget);
55   RegisterTargetMachine<ThumbLETargetMachine> A(TheThumbLETarget);
56   RegisterTargetMachine<ThumbBETargetMachine> B(TheThumbBETarget);
57 }
58 
59 static std::unique_ptr<TargetLoweringObjectFile> createTLOF(const Triple &TT) {
60   if (TT.isOSBinFormatMachO())
61     return make_unique<TargetLoweringObjectFileMachO>();
62   if (TT.isOSWindows())
63     return make_unique<TargetLoweringObjectFileCOFF>();
64   return make_unique<ARMElfTargetObjectFile>();
65 }
66 
67 static ARMBaseTargetMachine::ARMABI
68 computeTargetABI(const Triple &TT, StringRef CPU,
69                  const TargetOptions &Options) {
70   if (Options.MCOptions.getABIName() == "aapcs16")
71     return ARMBaseTargetMachine::ARM_ABI_AAPCS16;
72   else if (Options.MCOptions.getABIName().startswith("aapcs"))
73     return ARMBaseTargetMachine::ARM_ABI_AAPCS;
74   else if (Options.MCOptions.getABIName().startswith("apcs"))
75     return ARMBaseTargetMachine::ARM_ABI_APCS;
76 
77   assert(Options.MCOptions.getABIName().empty() &&
78          "Unknown target-abi option!");
79 
80   ARMBaseTargetMachine::ARMABI TargetABI =
81       ARMBaseTargetMachine::ARM_ABI_UNKNOWN;
82 
83   // FIXME: This is duplicated code from the front end and should be unified.
84   if (TT.isOSBinFormatMachO()) {
85     if (TT.getEnvironment() == llvm::Triple::EABI ||
86         (TT.getOS() == llvm::Triple::UnknownOS && TT.isOSBinFormatMachO()) ||
87         CPU.startswith("cortex-m")) {
88       TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
89     } else if (TT.isWatchABI()) {
90       TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS16;
91     } else {
92       TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
93     }
94   } else if (TT.isOSWindows()) {
95     // FIXME: this is invalid for WindowsCE
96     TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
97   } else {
98     // Select the default based on the platform.
99     switch (TT.getEnvironment()) {
100     case llvm::Triple::Android:
101     case llvm::Triple::GNUEABI:
102     case llvm::Triple::GNUEABIHF:
103     case llvm::Triple::EABIHF:
104     case llvm::Triple::EABI:
105       TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
106       break;
107     case llvm::Triple::GNU:
108       TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
109       break;
110     default:
111       if (TT.isOSNetBSD())
112         TargetABI = ARMBaseTargetMachine::ARM_ABI_APCS;
113       else
114         TargetABI = ARMBaseTargetMachine::ARM_ABI_AAPCS;
115       break;
116     }
117   }
118 
119   return TargetABI;
120 }
121 
122 static std::string computeDataLayout(const Triple &TT, StringRef CPU,
123                                      const TargetOptions &Options,
124                                      bool isLittle) {
125   auto ABI = computeTargetABI(TT, CPU, Options);
126   std::string Ret = "";
127 
128   if (isLittle)
129     // Little endian.
130     Ret += "e";
131   else
132     // Big endian.
133     Ret += "E";
134 
135   Ret += DataLayout::getManglingComponent(TT);
136 
137   // Pointers are 32 bits and aligned to 32 bits.
138   Ret += "-p:32:32";
139 
140   // ABIs other than APCS have 64 bit integers with natural alignment.
141   if (ABI != ARMBaseTargetMachine::ARM_ABI_APCS)
142     Ret += "-i64:64";
143 
144   // We have 64 bits floats. The APCS ABI requires them to be aligned to 32
145   // bits, others to 64 bits. We always try to align to 64 bits.
146   if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS)
147     Ret += "-f64:32:64";
148 
149   // We have 128 and 64 bit vectors. The APCS ABI aligns them to 32 bits, others
150   // to 64. We always ty to give them natural alignment.
151   if (ABI == ARMBaseTargetMachine::ARM_ABI_APCS)
152     Ret += "-v64:32:64-v128:32:128";
153   else if (ABI != ARMBaseTargetMachine::ARM_ABI_AAPCS16)
154     Ret += "-v128:64:128";
155 
156   // Try to align aggregates to 32 bits (the default is 64 bits, which has no
157   // particular hardware support on 32-bit ARM).
158   Ret += "-a:0:32";
159 
160   // Integer registers are 32 bits.
161   Ret += "-n32";
162 
163   // The stack is 128 bit aligned on NaCl, 64 bit aligned on AAPCS and 32 bit
164   // aligned everywhere else.
165   if (TT.isOSNaCl() || ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS16)
166     Ret += "-S128";
167   else if (ABI == ARMBaseTargetMachine::ARM_ABI_AAPCS)
168     Ret += "-S64";
169   else
170     Ret += "-S32";
171 
172   return Ret;
173 }
174 
175 /// TargetMachine ctor - Create an ARM architecture model.
176 ///
177 ARMBaseTargetMachine::ARMBaseTargetMachine(const Target &T, const Triple &TT,
178                                            StringRef CPU, StringRef FS,
179                                            const TargetOptions &Options,
180                                            Reloc::Model RM, CodeModel::Model CM,
181                                            CodeGenOpt::Level OL, bool isLittle)
182     : LLVMTargetMachine(T, computeDataLayout(TT, CPU, Options, isLittle), TT,
183                         CPU, FS, Options, RM, CM, OL),
184       TargetABI(computeTargetABI(TT, CPU, Options)),
185       TLOF(createTLOF(getTargetTriple())),
186       Subtarget(TT, CPU, FS, *this, isLittle), isLittle(isLittle) {
187 
188   // Default to triple-appropriate float ABI
189   if (Options.FloatABIType == FloatABI::Default)
190     this->Options.FloatABIType =
191         Subtarget.isTargetHardFloat() ? FloatABI::Hard : FloatABI::Soft;
192 
193   // Default to triple-appropriate EABI
194   if (Options.EABIVersion == EABI::Default ||
195       Options.EABIVersion == EABI::Unknown) {
196     if (Subtarget.isTargetGNUAEABI())
197       this->Options.EABIVersion = EABI::GNU;
198     else
199       this->Options.EABIVersion = EABI::EABI5;
200   }
201 }
202 
203 ARMBaseTargetMachine::~ARMBaseTargetMachine() {}
204 
205 const ARMSubtarget *
206 ARMBaseTargetMachine::getSubtargetImpl(const Function &F) const {
207   Attribute CPUAttr = F.getFnAttribute("target-cpu");
208   Attribute FSAttr = F.getFnAttribute("target-features");
209 
210   std::string CPU = !CPUAttr.hasAttribute(Attribute::None)
211                         ? CPUAttr.getValueAsString().str()
212                         : TargetCPU;
213   std::string FS = !FSAttr.hasAttribute(Attribute::None)
214                        ? FSAttr.getValueAsString().str()
215                        : TargetFS;
216 
217   // FIXME: This is related to the code below to reset the target options,
218   // we need to know whether or not the soft float flag is set on the
219   // function before we can generate a subtarget. We also need to use
220   // it as a key for the subtarget since that can be the only difference
221   // between two functions.
222   bool SoftFloat =
223       F.getFnAttribute("use-soft-float").getValueAsString() == "true";
224   // If the soft float attribute is set on the function turn on the soft float
225   // subtarget feature.
226   if (SoftFloat)
227     FS += FS.empty() ? "+soft-float" : ",+soft-float";
228 
229   auto &I = SubtargetMap[CPU + FS];
230   if (!I) {
231     // This needs to be done before we create a new subtarget since any
232     // creation will depend on the TM and the code generation flags on the
233     // function that reside in TargetOptions.
234     resetTargetOptions(F);
235     I = llvm::make_unique<ARMSubtarget>(TargetTriple, CPU, FS, *this, isLittle);
236   }
237   return I.get();
238 }
239 
240 TargetIRAnalysis ARMBaseTargetMachine::getTargetIRAnalysis() {
241   return TargetIRAnalysis([this](const Function &F) {
242     return TargetTransformInfo(ARMTTIImpl(this, F));
243   });
244 }
245 
246 void ARMTargetMachine::anchor() {}
247 
248 ARMTargetMachine::ARMTargetMachine(const Target &T, const Triple &TT,
249                                    StringRef CPU, StringRef FS,
250                                    const TargetOptions &Options,
251                                    Reloc::Model RM, CodeModel::Model CM,
252                                    CodeGenOpt::Level OL, bool isLittle)
253     : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, isLittle) {
254   initAsmInfo();
255   if (!Subtarget.hasARMOps())
256     report_fatal_error("CPU: '" + Subtarget.getCPUString() + "' does not "
257                        "support ARM mode execution!");
258 }
259 
260 void ARMLETargetMachine::anchor() {}
261 
262 ARMLETargetMachine::ARMLETargetMachine(const Target &T, const Triple &TT,
263                                        StringRef CPU, StringRef FS,
264                                        const TargetOptions &Options,
265                                        Reloc::Model RM, CodeModel::Model CM,
266                                        CodeGenOpt::Level OL)
267     : ARMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
268 
269 void ARMBETargetMachine::anchor() {}
270 
271 ARMBETargetMachine::ARMBETargetMachine(const Target &T, const Triple &TT,
272                                        StringRef CPU, StringRef FS,
273                                        const TargetOptions &Options,
274                                        Reloc::Model RM, CodeModel::Model CM,
275                                        CodeGenOpt::Level OL)
276     : ARMTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
277 
278 void ThumbTargetMachine::anchor() {}
279 
280 ThumbTargetMachine::ThumbTargetMachine(const Target &T, const Triple &TT,
281                                        StringRef CPU, StringRef FS,
282                                        const TargetOptions &Options,
283                                        Reloc::Model RM, CodeModel::Model CM,
284                                        CodeGenOpt::Level OL, bool isLittle)
285     : ARMBaseTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, isLittle) {
286   initAsmInfo();
287 }
288 
289 void ThumbLETargetMachine::anchor() {}
290 
291 ThumbLETargetMachine::ThumbLETargetMachine(const Target &T, const Triple &TT,
292                                            StringRef CPU, StringRef FS,
293                                            const TargetOptions &Options,
294                                            Reloc::Model RM, CodeModel::Model CM,
295                                            CodeGenOpt::Level OL)
296     : ThumbTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, true) {}
297 
298 void ThumbBETargetMachine::anchor() {}
299 
300 ThumbBETargetMachine::ThumbBETargetMachine(const Target &T, const Triple &TT,
301                                            StringRef CPU, StringRef FS,
302                                            const TargetOptions &Options,
303                                            Reloc::Model RM, CodeModel::Model CM,
304                                            CodeGenOpt::Level OL)
305     : ThumbTargetMachine(T, TT, CPU, FS, Options, RM, CM, OL, false) {}
306 
307 namespace {
308 /// ARM Code Generator Pass Configuration Options.
309 class ARMPassConfig : public TargetPassConfig {
310 public:
311   ARMPassConfig(ARMBaseTargetMachine *TM, PassManagerBase &PM)
312     : TargetPassConfig(TM, PM) {}
313 
314   ARMBaseTargetMachine &getARMTargetMachine() const {
315     return getTM<ARMBaseTargetMachine>();
316   }
317 
318   void addIRPasses() override;
319   bool addPreISel() override;
320   bool addInstSelector() override;
321   void addPreRegAlloc() override;
322   void addPreSched2() override;
323   void addPreEmitPass() override;
324 };
325 } // namespace
326 
327 TargetPassConfig *ARMBaseTargetMachine::createPassConfig(PassManagerBase &PM) {
328   return new ARMPassConfig(this, PM);
329 }
330 
331 void ARMPassConfig::addIRPasses() {
332   if (TM->Options.ThreadModel == ThreadModel::Single)
333     addPass(createLowerAtomicPass());
334   else
335     addPass(createAtomicExpandPass(TM));
336 
337   // Cmpxchg instructions are often used with a subsequent comparison to
338   // determine whether it succeeded. We can exploit existing control-flow in
339   // ldrex/strex loops to simplify this, but it needs tidying up.
340   if (TM->getOptLevel() != CodeGenOpt::None && EnableAtomicTidy)
341     addPass(createCFGSimplificationPass(-1, [this](const Function &F) {
342       const auto &ST = this->TM->getSubtarget<ARMSubtarget>(F);
343       return ST.hasAnyDataBarrier() && !ST.isThumb1Only();
344     }));
345 
346   TargetPassConfig::addIRPasses();
347 
348   // Match interleaved memory accesses to ldN/stN intrinsics.
349   if (TM->getOptLevel() != CodeGenOpt::None)
350     addPass(createInterleavedAccessPass(TM));
351 }
352 
353 bool ARMPassConfig::addPreISel() {
354   if ((TM->getOptLevel() != CodeGenOpt::None &&
355        EnableGlobalMerge == cl::BOU_UNSET) ||
356       EnableGlobalMerge == cl::BOU_TRUE) {
357     // FIXME: This is using the thumb1 only constant value for
358     // maximal global offset for merging globals. We may want
359     // to look into using the old value for non-thumb1 code of
360     // 4095 based on the TargetMachine, but this starts to become
361     // tricky when doing code gen per function.
362     bool OnlyOptimizeForSize = (TM->getOptLevel() < CodeGenOpt::Aggressive) &&
363                                (EnableGlobalMerge == cl::BOU_UNSET);
364     // Merging of extern globals is enabled by default on non-Mach-O as we
365     // expect it to be generally either beneficial or harmless. On Mach-O it
366     // is disabled as we emit the .subsections_via_symbols directive which
367     // means that merging extern globals is not safe.
368     bool MergeExternalByDefault = !TM->getTargetTriple().isOSBinFormatMachO();
369     addPass(createGlobalMergePass(TM, 127, OnlyOptimizeForSize,
370                                   MergeExternalByDefault));
371   }
372 
373   return false;
374 }
375 
376 bool ARMPassConfig::addInstSelector() {
377   addPass(createARMISelDag(getARMTargetMachine(), getOptLevel()));
378   return false;
379 }
380 
381 void ARMPassConfig::addPreRegAlloc() {
382   if (getOptLevel() != CodeGenOpt::None) {
383     addPass(createMLxExpansionPass());
384 
385     if (EnableARMLoadStoreOpt)
386       addPass(createARMLoadStoreOptimizationPass(/* pre-register alloc */ true));
387 
388     if (!DisableA15SDOptimization)
389       addPass(createA15SDOptimizerPass());
390   }
391 }
392 
393 void ARMPassConfig::addPreSched2() {
394   if (getOptLevel() != CodeGenOpt::None) {
395     if (EnableARMLoadStoreOpt)
396       addPass(createARMLoadStoreOptimizationPass());
397 
398     addPass(createExecutionDependencyFixPass(&ARM::DPRRegClass));
399   }
400 
401   // Expand some pseudo instructions into multiple instructions to allow
402   // proper scheduling.
403   addPass(createARMExpandPseudoPass());
404 
405   if (getOptLevel() != CodeGenOpt::None) {
406     // in v8, IfConversion depends on Thumb instruction widths
407     addPass(createThumb2SizeReductionPass([this](const Function &F) {
408       return this->TM->getSubtarget<ARMSubtarget>(F).restrictIT();
409     }));
410 
411     addPass(createIfConverter([this](const Function &F) {
412       return !this->TM->getSubtarget<ARMSubtarget>(F).isThumb1Only();
413     }));
414   }
415   addPass(createThumb2ITBlockPass());
416 }
417 
418 void ARMPassConfig::addPreEmitPass() {
419   addPass(createThumb2SizeReductionPass());
420 
421   // Constant island pass work on unbundled instructions.
422   addPass(createUnpackMachineBundles([this](const Function &F) {
423     return this->TM->getSubtarget<ARMSubtarget>(F).isThumb2();
424   }));
425 
426   // Don't optimize barriers at -O0.
427   if (getOptLevel() != CodeGenOpt::None)
428     addPass(createARMOptimizeBarriersPass());
429 
430   addPass(createARMConstantIslandPass());
431 }
432