Stephen Hines | 36b5688 | 2014-04-23 16:57:46 -0700 | [diff] [blame] | 1 | //===-- ARM64Subtarget.cpp - ARM64 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 ARM64 specific subclass of TargetSubtarget. |
| 11 | // |
| 12 | //===----------------------------------------------------------------------===// |
| 13 | |
| 14 | #include "ARM64InstrInfo.h" |
| 15 | #include "ARM64Subtarget.h" |
| 16 | #include "llvm/ADT/SmallVector.h" |
| 17 | #include "llvm/CodeGen/MachineScheduler.h" |
| 18 | #include "llvm/IR/GlobalValue.h" |
| 19 | #include "llvm/Support/TargetRegistry.h" |
| 20 | |
| 21 | #define GET_SUBTARGETINFO_CTOR |
| 22 | #define GET_SUBTARGETINFO_TARGET_DESC |
| 23 | #include "ARM64GenSubtargetInfo.inc" |
| 24 | |
| 25 | using namespace llvm; |
| 26 | |
| 27 | ARM64Subtarget::ARM64Subtarget(const std::string &TT, const std::string &CPU, |
| 28 | const std::string &FS) |
| 29 | : ARM64GenSubtargetInfo(TT, CPU, FS), HasZeroCycleRegMove(false), |
| 30 | HasZeroCycleZeroing(false), CPUString(CPU), TargetTriple(TT) { |
| 31 | // Determine default and user-specified characteristics |
| 32 | |
| 33 | if (CPUString.empty()) |
| 34 | // We default to Cyclone for now. |
| 35 | CPUString = "cyclone"; |
| 36 | |
| 37 | ParseSubtargetFeatures(CPUString, FS); |
| 38 | } |
| 39 | |
| 40 | /// ClassifyGlobalReference - Find the target operand flags that describe |
| 41 | /// how a global value should be referenced for the current subtarget. |
| 42 | unsigned char |
| 43 | ARM64Subtarget::ClassifyGlobalReference(const GlobalValue *GV, |
| 44 | const TargetMachine &TM) const { |
| 45 | |
| 46 | // Determine whether this is a reference to a definition or a declaration. |
| 47 | // Materializable GVs (in JIT lazy compilation mode) do not require an extra |
| 48 | // load from stub. |
| 49 | bool isDecl = GV->hasAvailableExternallyLinkage(); |
| 50 | if (GV->isDeclaration() && !GV->isMaterializable()) |
| 51 | isDecl = true; |
| 52 | |
| 53 | // MachO large model always goes via a GOT, simply to get a single 8-byte |
| 54 | // absolute relocation on all global addresses. |
| 55 | if (TM.getCodeModel() == CodeModel::Large && isTargetMachO()) |
| 56 | return ARM64II::MO_GOT; |
| 57 | |
| 58 | // The small code mode's direct accesses use ADRP, which cannot necessarily |
| 59 | // produce the value 0 (if the code is above 4GB). Therefore they must use the |
| 60 | // GOT. |
| 61 | if (TM.getCodeModel() == CodeModel::Small && GV->isWeakForLinker() && isDecl) |
| 62 | return ARM64II::MO_GOT; |
| 63 | |
| 64 | // If symbol visibility is hidden, the extra load is not needed if |
| 65 | // the symbol is definitely defined in the current translation unit. |
| 66 | |
| 67 | // The handling of non-hidden symbols in PIC mode is rather target-dependent: |
| 68 | // + On MachO, if the symbol is defined in this module the GOT can be |
| 69 | // skipped. |
| 70 | // + On ELF, the R_AARCH64_COPY relocation means that even symbols actually |
| 71 | // defined could end up in unexpected places. Use a GOT. |
| 72 | if (TM.getRelocationModel() != Reloc::Static && GV->hasDefaultVisibility()) { |
| 73 | if (isTargetMachO()) |
| 74 | return (isDecl || GV->isWeakForLinker()) ? ARM64II::MO_GOT |
| 75 | : ARM64II::MO_NO_FLAG; |
| 76 | else |
| 77 | return ARM64II::MO_GOT; |
| 78 | } |
| 79 | |
| 80 | return ARM64II::MO_NO_FLAG; |
| 81 | } |
| 82 | |
| 83 | /// This function returns the name of a function which has an interface |
| 84 | /// like the non-standard bzero function, if such a function exists on |
| 85 | /// the current subtarget and it is considered prefereable over |
| 86 | /// memset with zero passed as the second argument. Otherwise it |
| 87 | /// returns null. |
| 88 | const char *ARM64Subtarget::getBZeroEntry() const { |
| 89 | // At the moment, always prefer bzero. |
| 90 | return "bzero"; |
| 91 | } |
| 92 | |
| 93 | void ARM64Subtarget::overrideSchedPolicy(MachineSchedPolicy &Policy, |
| 94 | MachineInstr *begin, MachineInstr *end, |
| 95 | unsigned NumRegionInstrs) const { |
| 96 | // LNT run (at least on Cyclone) showed reasonably significant gains for |
| 97 | // bi-directional scheduling. 253.perlbmk. |
| 98 | Policy.OnlyTopDown = false; |
| 99 | Policy.OnlyBottomUp = false; |
| 100 | } |