diff options
Diffstat (limited to 'llvm/lib')
-rw-r--r-- | llvm/lib/Target/X86/X86ISelDAGToDAG.cpp | 8 | ||||
-rw-r--r-- | llvm/lib/Target/X86/X86InstrCompiler.td | 4 | ||||
-rw-r--r-- | llvm/lib/Target/X86/X86InstrInfo.cpp | 56 | ||||
-rw-r--r-- | llvm/lib/Target/X86/X86MCInstLower.cpp | 35 | ||||
-rw-r--r-- | llvm/lib/Target/X86/X86Subtarget.cpp | 32 | ||||
-rw-r--r-- | llvm/lib/Target/X86/X86TargetMachine.cpp | 5 |
6 files changed, 113 insertions, 27 deletions
diff --git a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp index f7fd221b517..026e6fcc9a7 100644 --- a/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp +++ b/llvm/lib/Target/X86/X86ISelDAGToDAG.cpp @@ -940,11 +940,11 @@ bool X86DAGToDAGISel::matchWrapper(SDValue N, X86ISelAddressMode &AM) { bool IsRIPRel = N.getOpcode() == X86ISD::WrapperRIP; - // Only do this address mode folding for 64-bit if we're in the small code - // model. - // FIXME: But we can do GOTPCREL addressing in the medium code model. + // In every code model except the 64-bit large code model, we can use an + // address mode to get the GOT entry or the global itself. CodeModel::Model M = TM.getCodeModel(); - if (Subtarget->is64Bit() && M != CodeModel::Small && M != CodeModel::Kernel) + if (Subtarget->is64Bit() && M != CodeModel::Small && M != CodeModel::Kernel && + !(M == CodeModel::Medium && IsRIPRel)) return true; // Base and index reg must be 0 in order to use %rip as base. diff --git a/llvm/lib/Target/X86/X86InstrCompiler.td b/llvm/lib/Target/X86/X86InstrCompiler.td index c863bac9722..8d2403367a3 100644 --- a/llvm/lib/Target/X86/X86InstrCompiler.td +++ b/llvm/lib/Target/X86/X86InstrCompiler.td @@ -37,6 +37,10 @@ let hasSideEffects = 0, isNotDuplicable = 1, Uses = [ESP, SSP], def MOVPC32r : Ii32<0xE8, Pseudo, (outs GR32:$reg), (ins i32imm:$label), "", []>; +// 64-bit large code model PIC base construction. +let hasSideEffects = 0, mayLoad = 1, isNotDuplicable = 1, SchedRW = [WriteJump] in + def MOVGOT64r : PseudoI<(outs GR64:$reg), + (ins GR64:$scratch, i64i32imm_pcrel:$got), []>; // ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into // a stack adjustment and the codegen must know that they may modify the stack diff --git a/llvm/lib/Target/X86/X86InstrInfo.cpp b/llvm/lib/Target/X86/X86InstrInfo.cpp index cfa25731c40..604cad3abc6 100644 --- a/llvm/lib/Target/X86/X86InstrInfo.cpp +++ b/llvm/lib/Target/X86/X86InstrInfo.cpp @@ -11333,7 +11333,9 @@ isSafeToMoveRegClassDefs(const TargetRegisterClass *RC) const { /// TODO: Eliminate this and move the code to X86MachineFunctionInfo. /// unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const { - assert(!Subtarget.is64Bit() && + assert((!Subtarget.is64Bit() || + MF->getTarget().getCodeModel() == CodeModel::Medium || + MF->getTarget().getCodeModel() == CodeModel::Large) && "X86-64 PIC uses RIP relative addressing"); X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>(); @@ -11344,7 +11346,8 @@ unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const { // Create the register. The code to initialize it is inserted // later, by the CGBR pass (below). MachineRegisterInfo &RegInfo = MF->getRegInfo(); - GlobalBaseReg = RegInfo.createVirtualRegister(&X86::GR32_NOSPRegClass); + GlobalBaseReg = RegInfo.createVirtualRegister( + Subtarget.is64Bit() ? &X86::GR64_NOSPRegClass : &X86::GR32_NOSPRegClass); X86FI->setGlobalBaseReg(GlobalBaseReg); return GlobalBaseReg; } @@ -12720,7 +12723,8 @@ namespace { // Don't do anything if this is 64-bit as 64-bit PIC // uses RIP relative addressing. - if (STI.is64Bit()) + if (STI.is64Bit() && (TM->getCodeModel() == CodeModel::Small || + TM->getCodeModel() == CodeModel::Kernel)) return false; // Only emit a global base reg in PIC mode. @@ -12747,17 +12751,41 @@ namespace { else PC = GlobalBaseReg; - // Operand of MovePCtoStack is completely ignored by asm printer. It's - // only used in JIT code emission as displacement to pc. - BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0); - - // If we're using vanilla 'GOT' PIC style, we should use relative addressing - // not to pc, but to _GLOBAL_OFFSET_TABLE_ external. - if (STI.isPICStyleGOT()) { - // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register - BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg) - .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_", - X86II::MO_GOT_ABSOLUTE_ADDRESS); + if (STI.is64Bit()) { + if (TM->getCodeModel() == CodeModel::Medium) { + BuildMI(FirstMBB, MBBI, DL, TII->get(X86::LEA64r), PC) + .addReg(X86::RIP) + .addImm(0) + .addReg(0) + .addExternalSymbol("_GLOBAL_OFFSET_TABLE_") + .addReg(0); + } else if (TM->getCodeModel() == CodeModel::Large) { + // Loading the GOT in the large code model requires math with labels, + // so we use a pseudo instruction and expand it during MC emission. + // FIXME: Are we using the right register state for a scratch register + // that doesn't conflict with the destination? + unsigned Scratch = RegInfo.createVirtualRegister(&X86::GR64RegClass); + BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVGOT64r), PC) + .addReg(Scratch, RegState::Undef | RegState::Define) + .addExternalSymbol("_GLOBAL_OFFSET_TABLE_"); + } else { + llvm_unreachable("unexpected code model"); + } + } else { + // Operand of MovePCtoStack is completely ignored by asm printer. It's + // only used in JIT code emission as displacement to pc. + BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0); + + // If we're using vanilla 'GOT' PIC style, we should use relative + // addressing not to pc, but to _GLOBAL_OFFSET_TABLE_ external. + if (STI.isPICStyleGOT()) { + // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], + // %some_register + BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg) + .addReg(PC) + .addExternalSymbol("_GLOBAL_OFFSET_TABLE_", + X86II::MO_GOT_ABSOLUTE_ADDRESS); + } } return true; diff --git a/llvm/lib/Target/X86/X86MCInstLower.cpp b/llvm/lib/Target/X86/X86MCInstLower.cpp index ae4b1e79c54..ee6161ad0a2 100644 --- a/llvm/lib/Target/X86/X86MCInstLower.cpp +++ b/llvm/lib/Target/X86/X86MCInstLower.cpp @@ -1982,6 +1982,41 @@ void X86AsmPrinter::EmitInstruction(const MachineInstr *MI) { return; } + case X86::MOVGOT64r: { + // Materializes the GOT for the 64-bit large code model. + MCSymbol *DotSym = OutContext.createTempSymbol(); + OutStreamer->EmitLabel(DotSym); + + unsigned DstReg = MI->getOperand(0).getReg(); + unsigned ScratchReg = MI->getOperand(1).getReg(); + MCSymbol *GOTSym = MCInstLowering.GetSymbolFromOperand(MI->getOperand(2)); + + // .LtmpN: leaq .LtmpN(%rip), %dst + const MCExpr *DotExpr = MCSymbolRefExpr::create(DotSym, OutContext); + EmitAndCountInstruction(MCInstBuilder(X86::LEA64r) + .addReg(DstReg) // dest + .addReg(X86::RIP) // base + .addImm(1) // scale + .addReg(0) // index + .addExpr(DotExpr) // disp + .addReg(0)); // seg + + // movq $_GLOBAL_OFFSET_TABLE_ - .LtmpN, %scratch + const MCExpr *GOTSymExpr = MCSymbolRefExpr::create(GOTSym, OutContext); + const MCExpr *GOTDiffExpr = + MCBinaryExpr::createSub(GOTSymExpr, DotExpr, OutContext); + EmitAndCountInstruction(MCInstBuilder(X86::MOV64ri) + .addReg(ScratchReg) // dest + .addExpr(GOTDiffExpr)); // disp + + // addq %scratch, %dst + EmitAndCountInstruction(MCInstBuilder(X86::ADD64rr) + .addReg(DstReg) // dest + .addReg(DstReg) // dest + .addReg(ScratchReg)); // src + return; + } + case X86::ADD32ri: { // Lower the MO_GOT_ABSOLUTE_ADDRESS form of ADD32ri. if (MI->getOperand(2).getTargetFlags() != X86II::MO_GOT_ABSOLUTE_ADDRESS) diff --git a/llvm/lib/Target/X86/X86Subtarget.cpp b/llvm/lib/Target/X86/X86Subtarget.cpp index 8e1e35eebba..b7cab75f5a5 100644 --- a/llvm/lib/Target/X86/X86Subtarget.cpp +++ b/llvm/lib/Target/X86/X86Subtarget.cpp @@ -68,15 +68,31 @@ X86Subtarget::classifyGlobalReference(const GlobalValue *GV) const { unsigned char X86Subtarget::classifyLocalReference(const GlobalValue *GV) const { - // 64 bits can use %rip addressing for anything local. - if (is64Bit()) - return X86II::MO_NO_FLAG; - - // If this is for a position dependent executable, the static linker can - // figure it out. + // If we're not PIC, it's not very interesting. if (!isPositionIndependent()) return X86II::MO_NO_FLAG; + // For 64-bit, we need to consider the code model. + if (is64Bit()) { + switch (TM.getCodeModel()) { + // 64-bit small code model is simple: All rip-relative. + case CodeModel::Small: + case CodeModel::Kernel: + return X86II::MO_NO_FLAG; + + // The large PIC code model uses GOTOFF. + case CodeModel::Large: + return X86II::MO_GOTOFF; + + // Medium is a hybrid: RIP-rel for code, GOTOFF for DSO local data. + case CodeModel::Medium: + if (isa<Function>(GV)) + return X86II::MO_NO_FLAG; // All code is RIP-relative + return X86II::MO_GOTOFF; // Local symbols use GOTOFF. + } + llvm_unreachable("invalid code model"); + } + // The COFF dynamic linker just patches the executable sections. if (isTargetCOFF()) return X86II::MO_NO_FLAG; @@ -98,7 +114,7 @@ X86Subtarget::classifyLocalReference(const GlobalValue *GV) const { unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV, const Module &M) const { // Large model never uses stubs. - if (TM.getCodeModel() == CodeModel::Large) + if (TM.getCodeModel() == CodeModel::Large && !isPositionIndependent()) return X86II::MO_NO_FLAG; // Absolute symbols can be referenced directly. @@ -120,7 +136,7 @@ unsigned char X86Subtarget::classifyGlobalReference(const GlobalValue *GV, if (isTargetCOFF()) return X86II::MO_DLLIMPORT; - if (is64Bit()) + if (is64Bit() && TM.getCodeModel() != CodeModel::Large) return X86II::MO_GOTPCREL; if (isTargetDarwin()) { diff --git a/llvm/lib/Target/X86/X86TargetMachine.cpp b/llvm/lib/Target/X86/X86TargetMachine.cpp index a713bddf7bf..92f5e795d7d 100644 --- a/llvm/lib/Target/X86/X86TargetMachine.cpp +++ b/llvm/lib/Target/X86/X86TargetMachine.cpp @@ -156,6 +156,7 @@ static std::string computeDataLayout(const Triple &TT) { } static Reloc::Model getEffectiveRelocModel(const Triple &TT, + bool JIT, Optional<Reloc::Model> RM) { bool is64Bit = TT.getArch() == Triple::x86_64; if (!RM.hasValue()) { @@ -167,6 +168,8 @@ static Reloc::Model getEffectiveRelocModel(const Triple &TT, return Reloc::PIC_; return Reloc::DynamicNoPIC; } + if (JIT) + return Reloc::Static; if (TT.isOSWindows() && is64Bit) return Reloc::PIC_; return Reloc::Static; @@ -210,7 +213,7 @@ X86TargetMachine::X86TargetMachine(const Target &T, const Triple &TT, CodeGenOpt::Level OL, bool JIT) : LLVMTargetMachine( T, computeDataLayout(TT), TT, CPU, FS, Options, - getEffectiveRelocModel(TT, RM), + getEffectiveRelocModel(TT, JIT, RM), getEffectiveCodeModel(CM, JIT, TT.getArch() == Triple::x86_64), OL), TLOF(createTLOF(getTargetTriple())) { // Windows stack unwinder gets confused when execution flow "falls through" |