1 //===-- ARMSubtarget.cpp - ARM Subtarget Information ------------*- C++ -*-===// 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 // This file implements the ARM specific subclass of TargetSubtargetInfo. 11 // 12 //===----------------------------------------------------------------------===// 13 14 #include "ARMSubtarget.h" 15 #include "ARMBaseRegisterInfo.h" 16 #include "llvm/GlobalValue.h" 17 #include "llvm/Target/TargetSubtargetInfo.h" 18 #include "llvm/Support/CommandLine.h" 19 #include "llvm/ADT/SmallVector.h" 20 21 #define GET_SUBTARGETINFO_TARGET_DESC 22 #define GET_SUBTARGETINFO_CTOR 23 #include "ARMGenSubtargetInfo.inc" 24 25 using namespace llvm; 26 27 static cl::opt<bool> 28 ReserveR9("arm-reserve-r9", cl::Hidden, 29 cl::desc("Reserve R9, making it unavailable as GPR")); 30 31 static cl::opt<bool> 32 DarwinUseMOVT("arm-darwin-use-movt", cl::init(true), cl::Hidden); 33 34 static cl::opt<bool> 35 StrictAlign("arm-strict-align", cl::Hidden, 36 cl::desc("Disallow all unaligned memory accesses")); 37 38 ARMSubtarget::ARMSubtarget(const std::string &TT, const std::string &CPU, 39 const std::string &FS) 40 : ARMGenSubtargetInfo(TT, CPU, FS) 41 , ARMProcFamily(Others) 42 , HasV4TOps(false) 43 , HasV5TOps(false) 44 , HasV5TEOps(false) 45 , HasV6Ops(false) 46 , HasV6T2Ops(false) 47 , HasV7Ops(false) 48 , HasVFPv2(false) 49 , HasVFPv3(false) 50 , HasNEON(false) 51 , UseNEONForSinglePrecisionFP(false) 52 , SlowFPVMLx(false) 53 , HasVMLxForwarding(false) 54 , SlowFPBrcc(false) 55 , InThumbMode(false) 56 , InNaClMode(false) 57 , HasThumb2(false) 58 , IsMClass(false) 59 , NoARM(false) 60 , PostRAScheduler(false) 61 , IsR9Reserved(ReserveR9) 62 , UseMovt(false) 63 , SupportsTailCall(false) 64 , HasFP16(false) 65 , HasD16(false) 66 , HasHardwareDivide(false) 67 , HasT2ExtractPack(false) 68 , HasDataBarrier(false) 69 , Pref32BitThumb(false) 70 , AvoidCPSRPartialUpdate(false) 71 , HasMPExtension(false) 72 , FPOnlySP(false) 73 , AllowsUnalignedMem(false) 74 , Thumb2DSP(false) 75 , stackAlignment(4) 76 , CPUString(CPU) 77 , TargetTriple(TT) 78 , TargetABI(ARM_ABI_APCS) { 79 // Determine default and user specified characteristics 80 if (CPUString.empty()) 81 CPUString = "generic"; 82 83 // Insert the architecture feature derived from the target triple into the 84 // feature string. This is important for setting features that are implied 85 // based on the architecture version. 86 std::string ArchFS = ARM_MC::ParseARMTriple(TT); 87 if (!FS.empty()) { 88 if (!ArchFS.empty()) 89 ArchFS = ArchFS + "," + FS; 90 else 91 ArchFS = FS; 92 } 93 ParseSubtargetFeatures(CPUString, ArchFS); 94 95 // Thumb2 implies at least V6T2. FIXME: Fix tests to explicitly specify a 96 // ARM version or CPU and then remove this. 97 if (!HasV6T2Ops && hasThumb2()) 98 HasV4TOps = HasV5TOps = HasV5TEOps = HasV6Ops = HasV6T2Ops = true; 99 100 // Initialize scheduling itinerary for the specified CPU. 101 InstrItins = getInstrItineraryForCPU(CPUString); 102 103 // After parsing Itineraries, set ItinData.IssueWidth. 104 computeIssueWidth(); 105 106 if (TT.find("eabi") != std::string::npos) 107 TargetABI = ARM_ABI_AAPCS; 108 109 if (isAAPCS_ABI()) 110 stackAlignment = 8; 111 112 if (!isTargetDarwin()) 113 UseMovt = hasV6T2Ops(); 114 else { 115 IsR9Reserved = ReserveR9 | !HasV6Ops; 116 UseMovt = DarwinUseMOVT && hasV6T2Ops(); 117 const Triple &T = getTargetTriple(); 118 SupportsTailCall = T.getOS() == Triple::IOS && !T.isOSVersionLT(5, 0); 119 } 120 121 if (!isThumb() || hasThumb2()) 122 PostRAScheduler = true; 123 124 // v6+ may or may not support unaligned mem access depending on the system 125 // configuration. 126 if (!StrictAlign && hasV6Ops() && isTargetDarwin()) 127 AllowsUnalignedMem = true; 128 } 129 130 /// GVIsIndirectSymbol - true if the GV will be accessed via an indirect symbol. 131 bool 132 ARMSubtarget::GVIsIndirectSymbol(const GlobalValue *GV, 133 Reloc::Model RelocM) const { 134 if (RelocM == Reloc::Static) 135 return false; 136 137 // Materializable GVs (in JIT lazy compilation mode) do not require an extra 138 // load from stub. 139 bool isDecl = GV->hasAvailableExternallyLinkage(); 140 if (GV->isDeclaration() && !GV->isMaterializable()) 141 isDecl = true; 142 143 if (!isTargetDarwin()) { 144 // Extra load is needed for all externally visible. 145 if (GV->hasLocalLinkage() || GV->hasHiddenVisibility()) 146 return false; 147 return true; 148 } else { 149 if (RelocM == Reloc::PIC_) { 150 // If this is a strong reference to a definition, it is definitely not 151 // through a stub. 152 if (!isDecl && !GV->isWeakForLinker()) 153 return false; 154 155 // Unless we have a symbol with hidden visibility, we have to go through a 156 // normal $non_lazy_ptr stub because this symbol might be resolved late. 157 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference. 158 return true; 159 160 // If symbol visibility is hidden, we have a stub for common symbol 161 // references and external declarations. 162 if (isDecl || GV->hasCommonLinkage()) 163 // Hidden $non_lazy_ptr reference. 164 return true; 165 166 return false; 167 } else { 168 // If this is a strong reference to a definition, it is definitely not 169 // through a stub. 170 if (!isDecl && !GV->isWeakForLinker()) 171 return false; 172 173 // Unless we have a symbol with hidden visibility, we have to go through a 174 // normal $non_lazy_ptr stub because this symbol might be resolved late. 175 if (!GV->hasHiddenVisibility()) // Non-hidden $non_lazy_ptr reference. 176 return true; 177 } 178 } 179 180 return false; 181 } 182 183 unsigned ARMSubtarget::getMispredictionPenalty() const { 184 // If we have a reasonable estimate of the pipeline depth, then we can 185 // estimate the penalty of a misprediction based on that. 186 if (isCortexA8()) 187 return 13; 188 else if (isCortexA9()) 189 return 8; 190 191 // Otherwise, just return a sensible default. 192 return 10; 193 } 194 195 void ARMSubtarget::computeIssueWidth() { 196 unsigned allStage1Units = 0; 197 for (const InstrItinerary *itin = InstrItins.Itineraries; 198 itin->FirstStage != ~0U; ++itin) { 199 const InstrStage *IS = InstrItins.Stages + itin->FirstStage; 200 allStage1Units |= IS->getUnits(); 201 } 202 InstrItins.IssueWidth = 0; 203 while (allStage1Units) { 204 ++InstrItins.IssueWidth; 205 // clear the lowest bit 206 allStage1Units ^= allStage1Units & ~(allStage1Units - 1); 207 } 208 assert(InstrItins.IssueWidth <= 2 && "itinerary bug, too many stage 1 units"); 209 } 210 211 bool ARMSubtarget::enablePostRAScheduler( 212 CodeGenOpt::Level OptLevel, 213 TargetSubtargetInfo::AntiDepBreakMode& Mode, 214 RegClassVector& CriticalPathRCs) const { 215 Mode = TargetSubtargetInfo::ANTIDEP_CRITICAL; 216 CriticalPathRCs.clear(); 217 CriticalPathRCs.push_back(&ARM::GPRRegClass); 218 return PostRAScheduler && OptLevel >= CodeGenOpt::Default; 219 } 220
