1 //===-- ARMSubtarget.cpp - ARM Subtarget Information ----------------------===//
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 //===----------------------------------------------------------------------===//
10 // This file implements the ARM specific subclass of TargetSubtargetInfo.
12 //===----------------------------------------------------------------------===//
14 #include "ARMSubtarget.h"
15 #include "ARMFrameLowering.h"
16 #include "ARMISelLowering.h"
17 #include "ARMInstrInfo.h"
18 #include "ARMMachineFunctionInfo.h"
19 #include "ARMSelectionDAGInfo.h"
20 #include "ARMSubtarget.h"
21 #include "ARMTargetMachine.h"
22 #include "Thumb1FrameLowering.h"
23 #include "Thumb1InstrInfo.h"
24 #include "Thumb2InstrInfo.h"
25 #include "llvm/CodeGen/MachineRegisterInfo.h"
26 #include "llvm/IR/Attributes.h"
27 #include "llvm/IR/Function.h"
28 #include "llvm/IR/GlobalValue.h"
29 #include "llvm/Support/CommandLine.h"
30 #include "llvm/Target/TargetInstrInfo.h"
31 #include "llvm/Target/TargetOptions.h"
32 #include "llvm/Target/TargetRegisterInfo.h"
36 #define DEBUG_TYPE "arm-subtarget"
38 #define GET_SUBTARGETINFO_TARGET_DESC
39 #define GET_SUBTARGETINFO_CTOR
40 #include "ARMGenSubtargetInfo.inc"
43 UseFusedMulOps("arm-use-mulops",
44 cl::init(true), cl::Hidden);
54 static cl::opt<AlignMode>
55 Align(cl::desc("Load/store alignment support"),
56 cl::Hidden, cl::init(DefaultAlign),
58 clEnumValN(DefaultAlign, "arm-default-align",
59 "Generate unaligned accesses only on hardware/OS "
60 "combinations that are known to support them"),
61 clEnumValN(StrictAlign, "arm-strict-align",
62 "Disallow all unaligned memory accesses"),
63 clEnumValN(NoStrictAlign, "arm-no-strict-align",
64 "Allow unaligned memory accesses"),
73 static cl::opt<ITMode>
74 IT(cl::desc("IT block support"), cl::Hidden, cl::init(DefaultIT),
76 cl::values(clEnumValN(DefaultIT, "arm-default-it",
77 "Generate IT block based on arch"),
78 clEnumValN(RestrictedIT, "arm-restrict-it",
79 "Disallow deprecated IT based on ARMv8"),
80 clEnumValN(NoRestrictedIT, "arm-no-restrict-it",
81 "Allow IT blocks based on ARMv7"),
84 /// initializeSubtargetDependencies - Initializes using a CPU and feature string
85 /// so that we can use initializer lists for subtarget initialization.
86 ARMSubtarget &ARMSubtarget::initializeSubtargetDependencies(StringRef CPU,
88 initializeEnvironment();
89 initSubtargetFeatures(CPU, FS);
93 ARMFrameLowering *ARMSubtarget::initializeFrameLowering(StringRef CPU,
95 ARMSubtarget &STI = initializeSubtargetDependencies(CPU, FS);
96 if (STI.isThumb1Only())
97 return (ARMFrameLowering *)new Thumb1FrameLowering(STI);
99 return new ARMFrameLowering(STI);
102 ARMSubtarget::ARMSubtarget(const Triple &TT, const std::string &CPU,
103 const std::string &FS,
104 const ARMBaseTargetMachine &TM, bool IsLittle)
105 : ARMGenSubtargetInfo(TT, CPU, FS), ARMProcFamily(Others),
106 ARMProcClass(None), stackAlignment(4), CPUString(CPU), IsLittle(IsLittle),
107 TargetTriple(TT), Options(TM.Options), TM(TM),
108 FrameLowering(initializeFrameLowering(CPU, FS)),
109 // At this point initializeSubtargetDependencies has been called so
110 // we can query directly.
111 InstrInfo(isThumb1Only()
112 ? (ARMBaseInstrInfo *)new Thumb1InstrInfo(*this)
114 ? (ARMBaseInstrInfo *)new ARMInstrInfo(*this)
115 : (ARMBaseInstrInfo *)new Thumb2InstrInfo(*this)),
118 void ARMSubtarget::initializeEnvironment() {
134 UseNEONForSinglePrecisionFP = false;
135 UseMulOps = UseFusedMulOps;
137 HasVMLxForwarding = false;
140 UseSoftFloat = false;
145 SupportsTailCall = false;
148 HasHardwareDivide = false;
149 HasHardwareDivideInARM = false;
150 HasT2ExtractPack = false;
151 HasDataBarrier = false;
152 Pref32BitThumb = false;
153 AvoidCPSRPartialUpdate = false;
154 AvoidMOVsShifterOperand = false;
156 HasMPExtension = false;
157 HasVirtualization = false;
160 HasTrustZone = false;
163 HasZeroCycleZeroing = false;
164 AllowsUnalignedMem = false;
167 GenLongCalls = false;
168 UnsafeFPMath = false;
171 void ARMSubtarget::initSubtargetFeatures(StringRef CPU, StringRef FS) {
172 if (CPUString.empty()) {
173 if (isTargetDarwin() && TargetTriple.getArchName().endswith("v7s"))
174 // Default to the Swift CPU when targeting armv7s/thumbv7s.
177 CPUString = "generic";
180 // Insert the architecture feature derived from the target triple into the
181 // feature string. This is important for setting features that are implied
182 // based on the architecture version.
183 std::string ArchFS = ARM_MC::ParseARMTriple(TargetTriple, CPUString);
186 ArchFS = (Twine(ArchFS) + "," + FS).str();
190 ParseSubtargetFeatures(CPUString, ArchFS);
192 // FIXME: This used enable V6T2 support implicitly for Thumb2 mode.
193 // Assert this for now to make the change obvious.
194 assert(hasV6T2Ops() || !hasThumb2());
196 // Keep a pointer to static instruction cost data for the specified CPU.
197 SchedModel = getSchedModelForCPU(CPUString);
199 // Initialize scheduling itinerary for the specified CPU.
200 InstrItins = getInstrItineraryForCPU(CPUString);
202 // FIXME: this is invalid for WindowsCE
203 if (isTargetWindows())
212 SupportsTailCall = !isTargetIOS() || !getTargetTriple().isOSVersionLT(5, 0);
214 SupportsTailCall = !isThumb1Only();
216 if (Align == DefaultAlign) {
217 // Assume pre-ARMv6 doesn't support unaligned accesses.
219 // ARMv6 may or may not support unaligned accesses depending on the
220 // SCTLR.U bit, which is architecture-specific. We assume ARMv6
221 // Darwin and NetBSD targets support unaligned accesses, and others don't.
223 // ARMv7 always has SCTLR.U set to 1, but it has a new SCTLR.A bit
224 // which raises an alignment fault on unaligned accesses. Linux
225 // defaults this bit to 0 and handles it as a system-wide (not
226 // per-process) setting. It is therefore safe to assume that ARMv7+
227 // Linux targets support unaligned accesses. The same goes for NaCl.
229 // The above behavior is consistent with GCC.
231 (hasV7Ops() && (isTargetLinux() || isTargetNaCl() ||
232 isTargetNetBSD())) ||
233 (hasV6Ops() && (isTargetMachO() || isTargetNetBSD()));
235 AllowsUnalignedMem = !(Align == StrictAlign);
238 // No v6M core supports unaligned memory access (v6M ARM ARM A3.2)
240 AllowsUnalignedMem = false;
244 RestrictIT = hasV8Ops();
254 // NEON f32 ops are non-IEEE 754 compliant. Darwin is ok with it by default.
255 const FeatureBitset &Bits = getFeatureBits();
256 if ((Bits[ARM::ProcA5] || Bits[ARM::ProcA8]) && // Where this matters
257 (Options.UnsafeFPMath || isTargetDarwin()))
258 UseNEONForSinglePrecisionFP = true;
261 bool ARMSubtarget::isAPCS_ABI() const {
262 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN);
263 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_APCS;
265 bool ARMSubtarget::isAAPCS_ABI() const {
266 assert(TM.TargetABI != ARMBaseTargetMachine::ARM_ABI_UNKNOWN);
267 return TM.TargetABI == ARMBaseTargetMachine::ARM_ABI_AAPCS;
270 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol.
272 ARMSubtarget::GVIsIndirectSymbol(const GlobalValue *GV,
273 Reloc::Model RelocM) const {
274 if (RelocM == Reloc::Static)
277 bool isDef = GV->isStrongDefinitionForLinker();
279 if (!isTargetMachO()) {
280 // Extra load is needed for all externally visible.
281 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility())
285 // If this is a strong reference to a definition, it is definitely not
290 // Unless we have a symbol with hidden visibility, we have to go through a
291 // normal $non_lazy_ptr stub because this symbol might be resolved late.
292 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference.
295 if (RelocM == Reloc::PIC_) {
296 // If symbol visibility is hidden, we have a stub for common symbol
297 // references and external declarations.
298 if (GV->isDeclarationForLinker() || GV->hasCommonLinkage())
299 // Hidden $non_lazy_ptr reference.
307 unsigned ARMSubtarget::getMispredictionPenalty() const {
308 return SchedModel.MispredictPenalty;
311 bool ARMSubtarget::hasSinCos() const {
312 return getTargetTriple().isiOS() && !getTargetTriple().isOSVersionLT(7, 0);
315 bool ARMSubtarget::enableMachineScheduler() const {
316 // Enable the MachineScheduler before register allocation for out-of-order
317 // architectures where we do not use the PostRA scheduler anymore (for now
318 // restricted to swift).
319 return getSchedModel().isOutOfOrder() && isSwift();
322 // This overrides the PostRAScheduler bit in the SchedModel for any CPU.
323 bool ARMSubtarget::enablePostRAScheduler() const {
324 // No need for PostRA scheduling on out of order CPUs (for now restricted to
326 if (getSchedModel().isOutOfOrder() && isSwift())
328 return (!isThumb() || hasThumb2());
331 bool ARMSubtarget::enableAtomicExpand() const {
332 return hasAnyDataBarrier() && !isThumb1Only();
335 bool ARMSubtarget::useMovt(const MachineFunction &MF) const {
336 // NOTE Windows on ARM needs to use mov.w/mov.t pairs to materialise 32-bit
337 // immediates as it is inherently position independent, and may be out of
339 return !NoMovt && hasV6T2Ops() &&
340 (isTargetWindows() ||
341 !MF.getFunction()->hasFnAttribute(Attribute::MinSize));
344 bool ARMSubtarget::useFastISel() const {
345 // Thumb2 support on iOS; ARM support on iOS, Linux and NaCl.
346 return TM.Options.EnableFastISel &&
347 ((isTargetMachO() && !isThumb1Only()) ||
348 (isTargetLinux() && !isThumb()) || (isTargetNaCl() && !isThumb()));