PPCSubtarget.cpp revision 36b56886974eae4f9c5ebc96befd3e7bfe5de338
1//===-- PowerPCSubtarget.cpp - PPC Subtarget Information ------------------===// 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 PPC specific subclass of TargetSubtargetInfo. 11// 12//===----------------------------------------------------------------------===// 13 14#include "PPCSubtarget.h" 15#include "PPC.h" 16#include "PPCRegisterInfo.h" 17#include "llvm/CodeGen/MachineFunction.h" 18#include "llvm/CodeGen/MachineScheduler.h" 19#include "llvm/IR/Attributes.h" 20#include "llvm/IR/Function.h" 21#include "llvm/IR/GlobalValue.h" 22#include "llvm/Support/Host.h" 23#include "llvm/Support/TargetRegistry.h" 24#include "llvm/Target/TargetMachine.h" 25#include <cstdlib> 26 27#define GET_SUBTARGETINFO_TARGET_DESC 28#define GET_SUBTARGETINFO_CTOR 29#include "PPCGenSubtargetInfo.inc" 30 31using namespace llvm; 32 33PPCSubtarget::PPCSubtarget(const std::string &TT, const std::string &CPU, 34 const std::string &FS, bool is64Bit, 35 CodeGenOpt::Level OptLevel) 36 : PPCGenSubtargetInfo(TT, CPU, FS) 37 , IsPPC64(is64Bit) 38 , TargetTriple(TT) { 39 initializeEnvironment(); 40 41 std::string FullFS = FS; 42 43 // At -O2 and above, track CR bits as individual registers. 44 if (OptLevel >= CodeGenOpt::Default) { 45 if (!FullFS.empty()) 46 FullFS = "+crbits," + FullFS; 47 else 48 FullFS = "+crbits"; 49 } 50 51 resetSubtargetFeatures(CPU, FullFS); 52} 53 54/// SetJITMode - This is called to inform the subtarget info that we are 55/// producing code for the JIT. 56void PPCSubtarget::SetJITMode() { 57 // JIT mode doesn't want lazy resolver stubs, it knows exactly where 58 // everything is. This matters for PPC64, which codegens in PIC mode without 59 // stubs. 60 HasLazyResolverStubs = false; 61 62 // Calls to external functions need to use indirect calls 63 IsJITCodeModel = true; 64} 65 66void PPCSubtarget::resetSubtargetFeatures(const MachineFunction *MF) { 67 AttributeSet FnAttrs = MF->getFunction()->getAttributes(); 68 Attribute CPUAttr = FnAttrs.getAttribute(AttributeSet::FunctionIndex, 69 "target-cpu"); 70 Attribute FSAttr = FnAttrs.getAttribute(AttributeSet::FunctionIndex, 71 "target-features"); 72 std::string CPU = 73 !CPUAttr.hasAttribute(Attribute::None) ? CPUAttr.getValueAsString() : ""; 74 std::string FS = 75 !FSAttr.hasAttribute(Attribute::None) ? FSAttr.getValueAsString() : ""; 76 if (!FS.empty()) { 77 initializeEnvironment(); 78 resetSubtargetFeatures(CPU, FS); 79 } 80} 81 82void PPCSubtarget::initializeEnvironment() { 83 StackAlignment = 16; 84 DarwinDirective = PPC::DIR_NONE; 85 HasMFOCRF = false; 86 Has64BitSupport = false; 87 Use64BitRegs = false; 88 UseCRBits = false; 89 HasAltivec = false; 90 HasQPX = false; 91 HasVSX = false; 92 HasFCPSGN = false; 93 HasFSQRT = false; 94 HasFRE = false; 95 HasFRES = false; 96 HasFRSQRTE = false; 97 HasFRSQRTES = false; 98 HasRecipPrec = false; 99 HasSTFIWX = false; 100 HasLFIWAX = false; 101 HasFPRND = false; 102 HasFPCVT = false; 103 HasISEL = false; 104 HasPOPCNTD = false; 105 HasLDBRX = false; 106 IsBookE = false; 107 DeprecatedMFTB = false; 108 DeprecatedDST = false; 109 HasLazyResolverStubs = false; 110 IsJITCodeModel = false; 111} 112 113void PPCSubtarget::resetSubtargetFeatures(StringRef CPU, StringRef FS) { 114 // Determine default and user specified characteristics 115 std::string CPUName = CPU; 116 if (CPUName.empty()) 117 CPUName = "generic"; 118#if (defined(__APPLE__) || defined(__linux__)) && \ 119 (defined(__ppc__) || defined(__powerpc__)) 120 if (CPUName == "generic") 121 CPUName = sys::getHostCPUName(); 122#endif 123 124 // Initialize scheduling itinerary for the specified CPU. 125 InstrItins = getInstrItineraryForCPU(CPUName); 126 127 // Make sure 64-bit features are available when CPUname is generic 128 std::string FullFS = FS; 129 130 // If we are generating code for ppc64, verify that options make sense. 131 if (IsPPC64) { 132 Has64BitSupport = true; 133 // Silently force 64-bit register use on ppc64. 134 Use64BitRegs = true; 135 if (!FullFS.empty()) 136 FullFS = "+64bit," + FullFS; 137 else 138 FullFS = "+64bit"; 139 } 140 141 // Parse features string. 142 ParseSubtargetFeatures(CPUName, FullFS); 143 144 // If the user requested use of 64-bit regs, but the cpu selected doesn't 145 // support it, ignore. 146 if (use64BitRegs() && !has64BitSupport()) 147 Use64BitRegs = false; 148 149 // Set up darwin-specific properties. 150 if (isDarwin()) 151 HasLazyResolverStubs = true; 152 153 // QPX requires a 32-byte aligned stack. Note that we need to do this if 154 // we're compiling for a BG/Q system regardless of whether or not QPX 155 // is enabled because external functions will assume this alignment. 156 if (hasQPX() || isBGQ()) 157 StackAlignment = 32; 158 159 // Determine endianness. 160 IsLittleEndian = (TargetTriple.getArch() == Triple::ppc64le); 161} 162 163/// hasLazyResolverStub - Return true if accesses to the specified global have 164/// to go through a dyld lazy resolution stub. This means that an extra load 165/// is required to get the address of the global. 166bool PPCSubtarget::hasLazyResolverStub(const GlobalValue *GV, 167 const TargetMachine &TM) const { 168 // We never have stubs if HasLazyResolverStubs=false or if in static mode. 169 if (!HasLazyResolverStubs || TM.getRelocationModel() == Reloc::Static) 170 return false; 171 // If symbol visibility is hidden, the extra load is not needed if 172 // the symbol is definitely defined in the current translation unit. 173 bool isDecl = GV->isDeclaration() && !GV->isMaterializable(); 174 if (GV->hasHiddenVisibility() && !isDecl && !GV->hasCommonLinkage()) 175 return false; 176 return GV->hasWeakLinkage() || GV->hasLinkOnceLinkage() || 177 GV->hasCommonLinkage() || isDecl; 178} 179 180bool PPCSubtarget::enablePostRAScheduler( 181 CodeGenOpt::Level OptLevel, 182 TargetSubtargetInfo::AntiDepBreakMode& Mode, 183 RegClassVector& CriticalPathRCs) const { 184 Mode = TargetSubtargetInfo::ANTIDEP_ALL; 185 186 CriticalPathRCs.clear(); 187 188 if (isPPC64()) 189 CriticalPathRCs.push_back(&PPC::G8RCRegClass); 190 else 191 CriticalPathRCs.push_back(&PPC::GPRCRegClass); 192 193 return OptLevel >= CodeGenOpt::Default; 194} 195 196// Embedded cores need aggressive scheduling (and some others also benefit). 197static bool needsAggressiveScheduling(unsigned Directive) { 198 switch (Directive) { 199 default: return false; 200 case PPC::DIR_440: 201 case PPC::DIR_A2: 202 case PPC::DIR_E500mc: 203 case PPC::DIR_E5500: 204 case PPC::DIR_PWR7: 205 return true; 206 } 207} 208 209bool PPCSubtarget::enableMachineScheduler() const { 210 // Enable MI scheduling for the embedded cores. 211 // FIXME: Enable this for all cores (some additional modeling 212 // may be necessary). 213 return needsAggressiveScheduling(DarwinDirective); 214} 215 216void PPCSubtarget::overrideSchedPolicy(MachineSchedPolicy &Policy, 217 MachineInstr *begin, 218 MachineInstr *end, 219 unsigned NumRegionInstrs) const { 220 if (needsAggressiveScheduling(DarwinDirective)) { 221 Policy.OnlyTopDown = false; 222 Policy.OnlyBottomUp = false; 223 } 224 225 // Spilling is generally expensive on all PPC cores, so always enable 226 // register-pressure tracking. 227 Policy.ShouldTrackPressure = true; 228} 229 230bool PPCSubtarget::useAA() const { 231 // Use AA during code generation for the embedded cores. 232 return needsAggressiveScheduling(DarwinDirective); 233} 234 235