This document describes techniques for writing compiler backends that convert the LLVM IR (intermediate representation) to code for a specified machine or other languages. Code intended for a specific machine can take the form of either assembly code or binary code (usable for a JIT compiler).
The backend of LLVM features a target-independent code generator that may create output for several types of target CPUs, including X86, PowerPC, Alpha, and SPARC. The backend may also be used to generate code targeted at SPUs of the Cell processor or GPUs to support the execution of compute kernels.
The document focuses on existing examples found in subdirectories of llvm/lib/Target in a downloaded LLVM release. In particular, this document focuses on the example of creating a static compiler (one that emits text assembly) for a SPARC target, because SPARC has fairly standard characteristics, such as a RISC instruction set and straightforward calling conventions.
The audience for this document is anyone who needs to write an LLVM backend to generate code for a specific hardware or software target.
To write a compiler backend for LLVM that converts the LLVM IR (intermediate representation) to code for a specified target (machine or other language), follow these steps:
In the .cpp and .h files, initially stub up these methods and then implement them later. Initially, you may not know which private members that the class will need and which components will need to be subclassed.
To actually create your compiler backend, you need to create and modify a few files. The absolute minimum is discussed here, but to actually use the LLVM target-independent code generator, you must perform the steps described in the LLVM Target-Independent Code Generator document.
First, you should create a subdirectory under lib/Target to hold all the files related to your target. If your target is called "Dummy", create the directory lib/Target/Dummy.
In this new directory, create a Makefile. It is easiest to copy a Makefile of another target and modify it. It should at least contain the LEVEL, LIBRARYNAME and TARGET variables, and then include $(LEVEL)/Makefile.common. The library can be named LLVMDummy (for example, see the MIPS target). Alternatively, you can split the library into LLVMDummyCodeGen and LLVMDummyAsmPrinter, the latter of which should be implemented in a subdirectory below lib/Target/Dummy (for example, see the PowerPC target).
Note that these two naming schemes are hardcoded into llvm-config. Using any other naming scheme will confuse llvm-config and produce lots of (seemingly unrelated) linker errors when linking llc.
To make your target actually do something, you need to implement a subclass of TargetMachine. This implementation should typically be in the file lib/Target/DummyTargetMachine.cpp, but any file in the lib/Target directory will be built and should work. To use LLVM's target independent code generator, you should do what all current machine backends do: create a subclass of LLVMTargetMachine. (To create a target from scratch, create a subclass of TargetMachine.)
To get LLVM to actually build and link your target, you need to add it to the TARGETS_TO_BUILD variable. To do this, you modify the configure script to know about your target when parsing the --enable-targets option. Search the configure script for TARGETS_TO_BUILD, add your target to the lists there (some creativity required) and then reconfigure. Alternatively, you can change autotools/configure.ac and regenerate configure by running ./autoconf/AutoRegen.sh
LLVMTargetMachine is designed as a base class for targets implemented with the LLVM target-independent code generator. The LLVMTargetMachine class should be specialized by a concrete target class that implements the various virtual methods. LLVMTargetMachine is defined as a subclass of TargetMachine in include/llvm/Target/TargetMachine.h. The TargetMachine class implementation (TargetMachine.cpp) also processes numerous command-line options.
To create a concrete target-specific subclass of LLVMTargetMachine, start by copying an existing TargetMachine class and header. You should name the files that you create to reflect your specific target. For instance, for the SPARC target, name the files SparcTargetMachine.h and SparcTargetMachine.cpp
For a target machine XXX, the implementation of XXXTargetMachine must have access methods to obtain objects that represent target components. These methods are named get*Info and are intended to obtain the instruction set (getInstrInfo), register set (getRegisterInfo), stack frame layout (getFrameInfo), and similar information. XXXTargetMachine must also implement the getTargetData method to access an object with target-specific data characteristics, such as data type size and alignment requirements.
For instance, for the SPARC target, the header file SparcTargetMachine.h declares prototypes for several get*Info and getTargetData methods that simply return a class member.
namespace llvm { class Module; class SparcTargetMachine : public LLVMTargetMachine { const TargetData DataLayout; // Calculates type size & alignment SparcSubtarget Subtarget; SparcInstrInfo InstrInfo; TargetFrameInfo FrameInfo; protected: virtual const TargetAsmInfo *createTargetAsmInfo() const; public: SparcTargetMachine(const Module &M, const std::string &FS); virtual const SparcInstrInfo *getInstrInfo() const {return &InstrInfo; } virtual const TargetFrameInfo *getFrameInfo() const {return &FrameInfo; } virtual const TargetSubtarget *getSubtargetImpl() const{return &Subtarget; } virtual const TargetRegisterInfo *getRegisterInfo() const { return &InstrInfo.getRegisterInfo(); } virtual const TargetData *getTargetData() const { return &DataLayout; } static unsigned getModuleMatchQuality(const Module &M); // Pass Pipeline Configuration virtual bool addInstSelector(PassManagerBase &PM, bool Fast); virtual bool addPreEmitPass(PassManagerBase &PM, bool Fast); virtual bool addAssemblyEmitter(PassManagerBase &PM, bool Fast, std::ostream &Out); }; } // end namespace llvm
For some targets, you also need to support the following methods:
In addition, the XXXTargetMachine constructor should specify a TargetDescription string that determines the data layout for the target machine, including characteristics such as pointer size, alignment, and endianness. For example, the constructor for SparcTargetMachine contains the following:
SparcTargetMachine::SparcTargetMachine(const Module &M, const std::string &FS) : DataLayout("E-p:32:32-f128:128:128"), Subtarget(M, FS), InstrInfo(Subtarget), FrameInfo(TargetFrameInfo::StackGrowsDown, 8, 0) { }
Hyphens separate portions of the TargetDescription string.
You must also register your target using the RegisterTarget template. (See the TargetMachineRegistry class.) For example, in SparcTargetMachine.cpp, the target is registered with:
namespace { // Register the target. RegisterTarget<SparcTargetMachine>X("sparc", "SPARC"); }
You should describe a concrete target-specific class that represents the register file of a target machine. This class is called XXXRegisterInfo (where XXX identifies the target) and represents the class register file data that is used for register allocation and also describes the interactions between registers.
You also need to define register classes to categorize related registers. A register class should be added for groups of registers that are all treated the same way for some instruction. Typical examples are register classes that include integer, floating-point, or vector registers. A register allocator allows an instruction to use any register in a specified register class to perform the instruction in a similar manner. Register classes allocate virtual registers to instructions from these sets, and register classes let the target-independent register allocator automatically choose the actual registers.
Much of the code for registers, including register definition, register aliases, and register classes, is generated by TableGen from XXXRegisterInfo.td input files and placed in XXXGenRegisterInfo.h.inc and XXXGenRegisterInfo.inc output files. Some of the code in the implementation of XXXRegisterInfo requires hand-coding.
The XXXRegisterInfo.td file typically starts with register definitions for a target machine. The Register class (specified in Target.td) is used to define an object for each register. The specified string n becomes the Name of the register. The basic Register object does not have any subregisters and does not specify any aliases.
class Register<string n> { string Namespace = ""; string AsmName = n; string Name = n; int SpillSize = 0; int SpillAlignment = 0; list<Register> Aliases = []; list<Register> SubRegs = []; list<int> DwarfNumbers = []; }
For example, in the X86RegisterInfo.td file, there are register definitions that utilize the Register class, such as:
def AL : Register<"AL">, DwarfRegNum<[0, 0, 0]>;
This defines the register AL and assigns it values (with DwarfRegNum) that are used by gcc, gdb, or a debug information writer (such as DwarfWriter in llvm/lib/CodeGen) to identify a register. For register AL, DwarfRegNum takes an array of 3 values, representing 3 different modes: the first element is for X86-64, the second for EH (exception handling) on X86-32, and the third is generic. -1 is a special Dwarf number that indicates the gcc number is undefined, and -2 indicates the register number is invalid for this mode.
From the previously described line in the X86RegisterInfo.td file, TableGen generates this code in the X86GenRegisterInfo.inc file:
static const unsigned GR8[] = { X86::AL, ... }; const unsigned AL_AliasSet[] = { X86::AX, X86::EAX, X86::RAX, 0 }; const TargetRegisterDesc RegisterDescriptors[] = { ... { "AL", "AL", AL_AliasSet, Empty_SubRegsSet, Empty_SubRegsSet, AL_SuperRegsSet }, ...
From the register info file, TableGen generates a TargetRegisterDesc object for each register. TargetRegisterDesc is defined in include/llvm/Target/TargetRegisterInfo.h with the following fields:
struct TargetRegisterDesc { const char *AsmName; // Assembly language name for the register const char *Name; // Printable name for the reg (for debugging) const unsigned *AliasSet; // Register Alias Set const unsigned *SubRegs; // Sub-register set const unsigned *ImmSubRegs; // Immediate sub-register set const unsigned *SuperRegs; // Super-register set };
TableGen uses the entire target description file (.td) to determine text names for the register (in the AsmName and Name fields of TargetRegisterDesc) and the relationships of other registers to the defined register (in the other TargetRegisterDesc fields). In this example, other definitions establish the registers "AX", "EAX", and "RAX" as aliases for one another, so TableGen generates a null-terminated array (AL_AliasSet) for this register alias set.
The Register class is commonly used as a base class for more complex classes. In Target.td, the Register class is the base for the RegisterWithSubRegs class that is used to define registers that need to specify subregisters in the SubRegs list, as shown here:
class RegisterWithSubRegs<string n, list<Register> subregs> : Register<n> { let SubRegs = subregs; }
In SparcRegisterInfo.td, additional register classes are defined for SPARC: a Register subclass, SparcReg, and further subclasses: Ri, Rf, and Rd. SPARC registers are identified by 5-bit ID numbers, which is a feature common to these subclasses. Note the use of ‘let’ expressions to override values that are initially defined in a superclass (such as SubRegs field in the Rd class).
class SparcReg<string n> : Register<n> { field bits<5> Num; let Namespace = "SP"; } // Ri - 32-bit integer registers class Ri<bits<5> num, string n> : SparcReg<n> { let Num = num; } // Rf - 32-bit floating-point registers class Rf<bits<5> num, string n> : SparcReg<n> { let Num = num; } // Rd - Slots in the FP register file for 64-bit floating-point values. class Rd<bits<5> num, string n, list<Register> subregs> : SparcReg<n> { let Num = num; let SubRegs = subregs; }
In the SparcRegisterInfo.td file, there are register definitions that utilize these subclasses of Register, such as:
def G0 : Ri< 0, "G0">, DwarfRegNum<[0]>; def G1 : Ri< 1, "G1">, DwarfRegNum<[1]>; ... def F0 : Rf< 0, "F0">, DwarfRegNum<[32]>; def F1 : Rf< 1, "F1">, DwarfRegNum<[33]>; ... def D0 : Rd< 0, "F0", [F0, F1]>, DwarfRegNum<[32]>; def D1 : Rd< 2, "F2", [F2, F3]>, DwarfRegNum<[34]>;
The last two registers shown above (D0 and D1) are double-precision floating-point registers that are aliases for pairs of single-precision floating-point sub-registers. In addition to aliases, the sub-register and super-register relationships of the defined register are in fields of a register’s TargetRegisterDesc.
The RegisterClass class (specified in Target.td) is used to define an object that represents a group of related registers and also defines the default allocation order of the registers. A target description file XXXRegisterInfo.td that uses Target.td can construct register classes using the following class:
class RegisterClass<string namespace, list<ValueType> regTypes, int alignment, list<Register> regList> { string Namespace = namespace; list<ValueType> RegTypes = regTypes; int Size = 0; // spill size, in bits; zero lets tblgen pick the size int Alignment = alignment; // CopyCost is the cost of copying a value between two registers // default value 1 means a single instruction // A negative value means copying is extremely expensive or impossible int CopyCost = 1; list<Register> MemberList = regList; // for register classes that are subregisters of this class list<RegisterClass> SubRegClassList = []; code MethodProtos = [{}]; // to insert arbitrary code code MethodBodies = [{}]; }
To define a RegisterClass, use the following 4 arguments:
In SparcRegisterInfo.td, three RegisterClass objects are defined: FPRegs, DFPRegs, and IntRegs. For all three register classes, the first argument defines the namespace with the string “SP”. FPRegs defines a group of 32 single-precision floating-point registers (F0 to F31); DFPRegs defines a group of 16 double-precision registers (D0-D15). For IntRegs, the MethodProtos and MethodBodies methods are used by TableGen to insert the specified code into generated output.
def FPRegs : RegisterClass<"SP", [f32], 32, [F0, F1, F2, F3, F4, F5, F6, F7, F8, F9, F10, F11, F12, F13, F14, F15, F16, F17, F18, F19, F20, F21, F22, F23, F24, F25, F26, F27, F28, F29, F30, F31]>; def DFPRegs : RegisterClass<"SP", [f64], 64, [D0, D1, D2, D3, D4, D5, D6, D7, D8, D9, D10, D11, D12, D13, D14, D15]>; def IntRegs : RegisterClass<"SP", [i32], 32, [L0, L1, L2, L3, L4, L5, L6, L7, I0, I1, I2, I3, I4, I5, O0, O1, O2, O3, O4, O5, O7, G1, // Non-allocatable regs: G2, G3, G4, O6, // stack ptr I6, // frame ptr I7, // return address G0, // constant zero G5, G6, G7 // reserved for kernel ]> { let MethodProtos = [{ iterator allocation_order_end(const MachineFunction &MF) const; }]; let MethodBodies = [{ IntRegsClass::iterator IntRegsClass::allocation_order_end(const MachineFunction &MF) const { return end()-10 // Don't allocate special registers -1; } }]; }
Using SparcRegisterInfo.td with TableGen generates several output files that are intended for inclusion in other source code that you write. SparcRegisterInfo.td generates SparcGenRegisterInfo.h.inc, which should be included in the header file for the implementation of the SPARC register implementation that you write (SparcRegisterInfo.h). In SparcGenRegisterInfo.h.inc a new structure is defined called SparcGenRegisterInfo that uses TargetRegisterInfo as its base. It also specifies types, based upon the defined register classes: DFPRegsClass, FPRegsClass, and IntRegsClass.
SparcRegisterInfo.td also generates SparcGenRegisterInfo.inc, which is included at the bottom of SparcRegisterInfo.cpp, the SPARC register implementation. The code below shows only the generated integer registers and associated register classes. The order of registers in IntRegs reflects the order in the definition of IntRegs in the target description file. Take special note of the use of MethodBodies in SparcRegisterInfo.td to create code in SparcGenRegisterInfo.inc. MethodProtos generates similar code in SparcGenRegisterInfo.h.inc.
// IntRegs Register Class... static const unsigned IntRegs[] = { SP::L0, SP::L1, SP::L2, SP::L3, SP::L4, SP::L5, SP::L6, SP::L7, SP::I0, SP::I1, SP::I2, SP::I3, SP::I4, SP::I5, SP::O0, SP::O1, SP::O2, SP::O3, SP::O4, SP::O5, SP::O7, SP::G1, SP::G2, SP::G3, SP::G4, SP::O6, SP::I6, SP::I7, SP::G0, SP::G5, SP::G6, SP::G7, }; // IntRegsVTs Register Class Value Types... static const MVT::ValueType IntRegsVTs[] = { MVT::i32, MVT::Other }; namespace SP { // Register class instances DFPRegsClass DFPRegsRegClass; FPRegsClass FPRegsRegClass; IntRegsClass IntRegsRegClass; ... // IntRegs Sub-register Classess... static const TargetRegisterClass* const IntRegsSubRegClasses [] = { NULL }; ... // IntRegs Super-register Classess... static const TargetRegisterClass* const IntRegsSuperRegClasses [] = { NULL }; // IntRegs Register Class sub-classes... static const TargetRegisterClass* const IntRegsSubclasses [] = { NULL }; ... // IntRegs Register Class super-classes... static const TargetRegisterClass* const IntRegsSuperclasses [] = { NULL }; ... IntRegsClass::iterator IntRegsClass::allocation_order_end(const MachineFunction &MF) const { return end()-10 // Don't allocate special registers -1; } IntRegsClass::IntRegsClass() : TargetRegisterClass(IntRegsRegClassID, IntRegsVTs, IntRegsSubclasses, IntRegsSuperclasses, IntRegsSubRegClasses, IntRegsSuperRegClasses, 4, 4, 1, IntRegs, IntRegs + 32) {} }
The final step is to hand code portions of XXXRegisterInfo, which implements the interface described in TargetRegisterInfo.h. These functions return 0, NULL, or false, unless overridden. Here’s a list of functions that are overridden for the SPARC implementation in SparcRegisterInfo.cpp:
During the early stages of code generation, the LLVM IR code is converted to a SelectionDAG with nodes that are instances of the SDNode class containing target instructions. An SDNode has an opcode, operands, type requirements, and operation properties (for example, is an operation commutative, does an operation load from memory). The various operation node types are described in the include/llvm/CodeGen/SelectionDAGNodes.h file (values of the NodeType enum in the ISD namespace).
TableGen uses the following target description (.td) input files to generate much of the code for instruction definition:
There is also a target-specific XXX.td file, where XXX is the name of the target. The XXX.td file includes the other .td input files, but its contents are only directly important for subtargets.
You should describe a concrete target-specific class XXXInstrInfo that represents machine instructions supported by a target machine. XXXInstrInfo contains an array of XXXInstrDescriptor objects, each of which describes one instruction. An instruction descriptor defines:
The Instruction class (defined in Target.td) is mostly used as a base for more complex instruction classes.
class Instruction { string Namespace = ""; dag OutOperandList; // An dag containing the MI def operand list. dag InOperandList; // An dag containing the MI use operand list. string AsmString = ""; // The .s format to print the instruction with. list<dag> Pattern; // Set to the DAG pattern for this instruction list<Register> Uses = []; list<Register> Defs = []; list<Predicate> Predicates = []; // predicates turned into isel match code ... remainder not shown for space ... }
A SelectionDAG node (SDNode) should contain an object representing a target-specific instruction that is defined in XXXInstrInfo.td. The instruction objects should represent instructions from the architecture manual of the target machine (such as the SPARC Architecture Manual for the SPARC target).
A single instruction from the architecture manual is often modeled as multiple target instructions, depending upon its operands. For example, a manual might describe an add instruction that takes a register or an immediate operand. An LLVM target could model this with two instructions named ADDri and ADDrr.
You should define a class for each instruction category and define each opcode as a subclass of the category with appropriate parameters such as the fixed binary encoding of opcodes and extended opcodes. You should map the register bits to the bits of the instruction in which they are encoded (for the JIT). Also you should specify how the instruction should be printed when the automatic assembly printer is used.
As is described in the SPARC Architecture Manual, Version 8, there are three major 32-bit formats for instructions. Format 1 is only for the CALL instruction. Format 2 is for branch on condition codes and SETHI (set high bits of a register) instructions. Format 3 is for other instructions.
Each of these formats has corresponding classes in SparcInstrFormat.td. InstSP is a base class for other instruction classes. Additional base classes are specified for more precise formats: for example in SparcInstrFormat.td, F2_1 is for SETHI, and F2_2 is for branches. There are three other base classes: F3_1 for register/register operations, F3_2 for register/immediate operations, and F3_3 for floating-point operations. SparcInstrInfo.td also adds the base class Pseudo for synthetic SPARC instructions.
SparcInstrInfo.td largely consists of operand and instruction definitions for the SPARC target. In SparcInstrInfo.td, the following target description file entry, LDrr, defines the Load Integer instruction for a Word (the LD SPARC opcode) from a memory address to a register. The first parameter, the value 3 (112), is the operation value for this category of operation. The second parameter (0000002) is the specific operation value for LD/Load Word. The third parameter is the output destination, which is a register operand and defined in the Register target description file (IntRegs).
def LDrr : F3_1 <3, 0b000000, (outs IntRegs:$dst), (ins MEMrr:$addr), "ld [$addr], $dst", [(set IntRegs:$dst, (load ADDRrr:$addr))]>;
The fourth parameter is the input source, which uses the address operand MEMrr that is defined earlier in SparcInstrInfo.td:
def MEMrr : Operand<i32> { let PrintMethod = "printMemOperand"; let MIOperandInfo = (ops IntRegs, IntRegs); }
The fifth parameter is a string that is used by the assembly printer and can be left as an empty string until the assembly printer interface is implemented. The sixth and final parameter is the pattern used to match the instruction during the SelectionDAG Select Phase described in (The LLVM Target-Independent Code Generator). This parameter is detailed in the next section, Instruction Selector.
Instruction class definitions are not overloaded for different operand types, so separate versions of instructions are needed for register, memory, or immediate value operands. For example, to perform a Load Integer instruction for a Word from an immediate operand to a register, the following instruction class is defined:
def LDri : F3_2 <3, 0b000000, (outs IntRegs:$dst), (ins MEMri:$addr), "ld [$addr], $dst", [(set IntRegs:$dst, (load ADDRri:$addr))]>;
Writing these definitions for so many similar instructions can involve a lot of cut and paste. In td files, the multiclass directive enables the creation of templates to define several instruction classes at once (using the defm directive). For example in SparcInstrInfo.td, the multiclass pattern F3_12 is defined to create 2 instruction classes each time F3_12 is invoked:
multiclass F3_12 <string OpcStr, bits<6> Op3Val, SDNode OpNode> { def rr : F3_1 <2, Op3Val, (outs IntRegs:$dst), (ins IntRegs:$b, IntRegs:$c), !strconcat(OpcStr, " $b, $c, $dst"), [(set IntRegs:$dst, (OpNode IntRegs:$b, IntRegs:$c))]>; def ri : F3_2 <2, Op3Val, (outs IntRegs:$dst), (ins IntRegs:$b, i32imm:$c), !strconcat(OpcStr, " $b, $c, $dst"), [(set IntRegs:$dst, (OpNode IntRegs:$b, simm13:$c))]>; }
So when the defm directive is used for the XOR and ADD instructions, as seen below, it creates four instruction objects: XORrr, XORri, ADDrr, and ADDri.
defm XOR : F3_12<"xor", 0b000011, xor>; defm ADD : F3_12<"add", 0b000000, add>;
SparcInstrInfo.td also includes definitions for condition codes that are referenced by branch instructions. The following definitions in SparcInstrInfo.td indicate the bit location of the SPARC condition code; for example, the 10th bit represents the ‘greater than’ condition for integers, and the 22nd bit represents the ‘greater than’ condition for floats.
def ICC_NE : ICC_VAL< 9>; // Not Equal def ICC_E : ICC_VAL< 1>; // Equal def ICC_G : ICC_VAL<10>; // Greater ... def FCC_U : FCC_VAL<23>; // Unordered def FCC_G : FCC_VAL<22>; // Greater def FCC_UG : FCC_VAL<21>; // Unordered or Greater ...
(Note that Sparc.h also defines enums that correspond to the same SPARC condition codes. Care must be taken to ensure the values in Sparc.h correspond to the values in SparcInstrInfo.td; that is, SPCC::ICC_NE = 9, SPCC::FCC_U = 23 and so on.)
The final step is to hand code portions of XXXInstrInfo, which implements the interface described in TargetInstrInfo.h. These functions return 0 or a Boolean or they assert, unless overridden. Here's a list of functions that are overridden for the SPARC implementation in SparcInstrInfo.cpp:
Performance can be improved by combining instructions or by eliminating instructions that are never reached. The AnalyzeBranch method in XXXInstrInfo may be implemented to examine conditional instructions and remove unnecessary instructions. AnalyzeBranch looks at the end of a machine basic block (MBB) for opportunities for improvement, such as branch folding and if conversion. The BranchFolder and IfConverter machine function passes (see the source files BranchFolding.cpp and IfConversion.cpp in the lib/CodeGen directory) call AnalyzeBranch to improve the control flow graph that represents the instructions.
Several implementations of AnalyzeBranch (for ARM, Alpha, and X86) can be examined as models for your own AnalyzeBranch implementation. Since SPARC does not implement a useful AnalyzeBranch, the ARM target implementation is shown below.
AnalyzeBranch returns a Boolean value and takes four parameters:
In the simplest case, if a block ends without a branch, then it falls through to the successor block. No destination blocks are specified for either TBB or FBB, so both parameters return NULL. The start of the AnalyzeBranch (see code below for the ARM target) shows the function parameters and the code for the simplest case.
bool ARMInstrInfo::AnalyzeBranch(MachineBasicBlock &MBB, MachineBasicBlock *&TBB, MachineBasicBlock *&FBB, std::vector<MachineOperand> &Cond) const { MachineBasicBlock::iterator I = MBB.end(); if (I == MBB.begin() || !isUnpredicatedTerminator(--I)) return false;
If a block ends with a single unconditional branch instruction, then AnalyzeBranch (shown below) should return the destination of that branch in the TBB parameter.
if (LastOpc == ARM::B || LastOpc == ARM::tB) { TBB = LastInst->getOperand(0).getMBB(); return false; }
If a block ends with two unconditional branches, then the second branch is never reached. In that situation, as shown below, remove the last branch instruction and return the penultimate branch in the TBB parameter.
if ((SecondLastOpc == ARM::B || SecondLastOpc==ARM::tB) && (LastOpc == ARM::B || LastOpc == ARM::tB)) { TBB = SecondLastInst->getOperand(0).getMBB(); I = LastInst; I->eraseFromParent(); return false; }
A block may end with a single conditional branch instruction that falls through to successor block if the condition evaluates to false. In that case, AnalyzeBranch (shown below) should return the destination of that conditional branch in the TBB parameter and a list of operands in the Cond parameter to evaluate the condition.
if (LastOpc == ARM::Bcc || LastOpc == ARM::tBcc) { // Block ends with fall-through condbranch. TBB = LastInst->getOperand(0).getMBB(); Cond.push_back(LastInst->getOperand(1)); Cond.push_back(LastInst->getOperand(2)); return false; }
If a block ends with both a conditional branch and an ensuing unconditional branch, then AnalyzeBranch (shown below) should return the conditional branch destination (assuming it corresponds to a conditional evaluation of ‘true’) in the TBB parameter and the unconditional branch destination in the FBB (corresponding to a conditional evaluation of ‘false’). A list of operands to evaluate the condition should be returned in the Cond parameter.
unsigned SecondLastOpc = SecondLastInst->getOpcode(); if ((SecondLastOpc == ARM::Bcc && LastOpc == ARM::B) || (SecondLastOpc == ARM::tBcc && LastOpc == ARM::tB)) { TBB = SecondLastInst->getOperand(0).getMBB(); Cond.push_back(SecondLastInst->getOperand(1)); Cond.push_back(SecondLastInst->getOperand(2)); FBB = LastInst->getOperand(0).getMBB(); return false; }
For the last two cases (ending with a single conditional branch or ending with one conditional and one unconditional branch), the operands returned in the Cond parameter can be passed to methods of other instructions to create new branches or perform other operations. An implementation of AnalyzeBranch requires the helper methods RemoveBranch and InsertBranch to manage subsequent operations.
AnalyzeBranch should return false indicating success in most circumstances. AnalyzeBranch should only return true when the method is stumped about what to do, for example, if a block has three terminating branches. AnalyzeBranch may return true if it encounters a terminator it cannot handle, such as an indirect branch.
LLVM uses a SelectionDAG to represent LLVM IR instructions, and nodes of the SelectionDAG ideally represent native target instructions. During code generation, instruction selection passes are performed to convert non-native DAG instructions into native target-specific instructions. The pass described in XXXISelDAGToDAG.cpp is used to match patterns and perform DAG-to-DAG instruction selection. Optionally, a pass may be defined (in XXXBranchSelector.cpp) to perform similar DAG-to-DAG operations for branch instructions. Later, the code in XXXISelLowering.cpp replaces or removes operations and data types not supported natively (legalizes) in a Selection DAG.
TableGen generates code for instruction selection using the following target description input files:
The implementation of an instruction selection pass must include a header that declares the FunctionPass class or a subclass of FunctionPass. In XXXTargetMachine.cpp, a Pass Manager (PM) should add each instruction selection pass into the queue of passes to run.
The LLVM static compiler (llc) is an excellent tool for visualizing the contents of DAGs. To display the SelectionDAG before or after specific processing phases, use the command line options for llc, described at SelectionDAG Instruction Selection Process.
To describe instruction selector behavior, you should add patterns for lowering LLVM code into a SelectionDAG as the last parameter of the instruction definitions in XXXInstrInfo.td. For example, in SparcInstrInfo.td, this entry defines a register store operation, and the last parameter describes a pattern with the store DAG operator.
def STrr : F3_1< 3, 0b000100, (outs), (ins MEMrr:$addr, IntRegs:$src), "st $src, [$addr]", [(store IntRegs:$src, ADDRrr:$addr)]>;
ADDRrr is a memory mode that is also defined in SparcInstrInfo.td:
def ADDRrr : ComplexPattern<i32, 2, "SelectADDRrr", [], []>;
The definition of ADDRrr refers to SelectADDRrr, which is a function defined in an implementation of the Instructor Selector (such as SparcISelDAGToDAG.cpp).
In lib/Target/TargetSelectionDAG.td, the DAG operator for store is defined below:
def store : PatFrag<(ops node:$val, node:$ptr), (st node:$val, node:$ptr), [{ if (StoreSDNode *ST = dyn_cast<StoreSDNode>(N)) return !ST->isTruncatingStore() && ST->getAddressingMode() == ISD::UNINDEXED; return false; }]>;
XXXInstrInfo.td also generates (in XXXGenDAGISel.inc) the SelectCode method that is used to call the appropriate processing method for an instruction. In this example, SelectCode calls Select_ISD_STORE for the ISD::STORE opcode.
SDNode *SelectCode(SDOperand N) { ... MVT::ValueType NVT = N.Val->getValueType(0); switch (N.getOpcode()) { case ISD::STORE: { switch (NVT) { default: return Select_ISD_STORE(N); break; } break; } ...
The pattern for STrr is matched, so elsewhere in XXXGenDAGISel.inc, code for STrr is created for Select_ISD_STORE. The Emit_22 method is also generated in XXXGenDAGISel.inc to complete the processing of this instruction.
SDNode *Select_ISD_STORE(const SDOperand &N) { SDOperand Chain = N.getOperand(0); if (Predicate_store(N.Val)) { SDOperand N1 = N.getOperand(1); SDOperand N2 = N.getOperand(2); SDOperand CPTmp0; SDOperand CPTmp1; // Pattern: (st:void IntRegs:i32:$src, // ADDRrr:i32:$addr)<<P:Predicate_store>> // Emits: (STrr:void ADDRrr:i32:$addr, IntRegs:i32:$src) // Pattern complexity = 13 cost = 1 size = 0 if (SelectADDRrr(N, N2, CPTmp0, CPTmp1) && N1.Val->getValueType(0) == MVT::i32 && N2.Val->getValueType(0) == MVT::i32) { return Emit_22(N, SP::STrr, CPTmp0, CPTmp1); } ...
The Legalize phase converts a DAG to use types and operations that are natively supported by the target. For natively unsupported types and operations, you need to add code to the target-specific XXXTargetLowering implementation to convert unsupported types and operations to supported ones.
In the constructor for the XXXTargetLowering class, first use the addRegisterClass method to specify which types are supports and which register classes are associated with them. The code for the register classes are generated by TableGen from XXXRegisterInfo.td and placed in XXXGenRegisterInfo.h.inc. For example, the implementation of the constructor for the SparcTargetLowering class (in SparcISelLowering.cpp) starts with the following code:
addRegisterClass(MVT::i32, SP::IntRegsRegisterClass); addRegisterClass(MVT::f32, SP::FPRegsRegisterClass); addRegisterClass(MVT::f64, SP::DFPRegsRegisterClass);
You should examine the node types in the ISD namespace (include/llvm/CodeGen/SelectionDAGNodes.h) and determine which operations the target natively supports. For operations that do not have native support, add a callback to the constructor for the XXXTargetLowering class, so the instruction selection process knows what to do. The TargetLowering class callback methods (declared in llvm/Target/TargetLowering.h) are:
Note: on older releases, setLoadXAction is used instead of setLoadExtAction. Also, on older releases, setCondCodeAction may not be supported. Examine your release to see what methods are specifically supported.
These callbacks are used to determine that an operation does or does not work with a specified type (or types). And in all cases, the third parameter is a LegalAction type enum value: Promote, Expand, Custom, or Legal. SparcISelLowering.cpp contains examples of all four LegalAction values.
For an operation without native support for a given type, the specified type may be promoted to a larger type that is supported. For example, SPARC does not support a sign-extending load for Boolean values (i1 type), so in SparcISelLowering.cpp the third parameter below, Promote, changes i1 type values to a large type before loading.
setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);
For a type without native support, a value may need to be broken down further, rather than promoted. For an operation without native support, a combination of other operations may be used to similar effect. In SPARC, the floating-point sine and cosine trig operations are supported by expansion to other operations, as indicated by the third parameter, Expand, to setOperationAction:
setOperationAction(ISD::FSIN, MVT::f32, Expand); setOperationAction(ISD::FCOS, MVT::f32, Expand);
For some operations, simple type promotion or operation expansion may be insufficient. In some cases, a special intrinsic function must be implemented.
For example, a constant value may require special treatment, or an operation may require spilling and restoring registers in the stack and working with register allocators.
As seen in SparcISelLowering.cpp code below, to perform a type conversion from a floating point value to a signed integer, first the setOperationAction should be called with Custom as the third parameter:
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
In the LowerOperation method, for each Custom operation, a case statement should be added to indicate what function to call. In the following code, an FP_TO_SINT opcode will call the LowerFP_TO_SINT method:
SDOperand SparcTargetLowering::LowerOperation( SDOperand Op, SelectionDAG &DAG) { switch (Op.getOpcode()) { case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG); ... } }
Finally, the LowerFP_TO_SINT method is implemented, using an FP register to convert the floating-point value to an integer.
static SDOperand LowerFP_TO_SINT(SDOperand Op, SelectionDAG &DAG) { assert(Op.getValueType() == MVT::i32); Op = DAG.getNode(SPISD::FTOI, MVT::f32, Op.getOperand(0)); return DAG.getNode(ISD::BIT_CONVERT, MVT::i32, Op); }
The Legal LegalizeAction enum value simply indicates that an operation is natively supported. Legal represents the default condition, so it is rarely used. In SparcISelLowering.cpp, the action for CTPOP (an operation to count the bits set in an integer) is natively supported only for SPARC v9. The following code enables the Expand conversion technique for non-v9 SPARC implementations.
setOperationAction(ISD::CTPOP, MVT::i32, Expand); ... if (TM.getSubtarget<SparcSubtarget>().isV9()) setOperationAction(ISD::CTPOP, MVT::i32, Legal); case ISD::SETULT: return SPCC::ICC_CS; case ISD::SETULE: return SPCC::ICC_LEU; case ISD::SETUGT: return SPCC::ICC_GU; case ISD::SETUGE: return SPCC::ICC_CC; } }
To support target-specific calling conventions, XXXGenCallingConv.td uses interfaces (such as CCIfType and CCAssignToReg) that are defined in lib/Target/TargetCallingConv.td. TableGen can take the target descriptor file XXXGenCallingConv.td and generate the header file XXXGenCallingConv.inc, which is typically included in XXXISelLowering.cpp. You can use the interfaces in TargetCallingConv.td to specify:
The following example demonstrates the use of the CCIfType and CCAssignToReg interfaces. If the CCIfType predicate is true (that is, if the current argument is of type f32 or f64), then the action is performed. In this case, the CCAssignToReg action assigns the argument value to the first available register: either R0 or R1.
CCIfType<[f32,f64], CCAssignToReg<[R0, R1]>>
SparcCallingConv.td contains definitions for a target-specific return-value calling convention (RetCC_Sparc32) and a basic 32-bit C calling convention (CC_Sparc32). The definition of RetCC_Sparc32 (shown below) indicates which registers are used for specified scalar return types. A single-precision float is returned to register F0, and a double-precision float goes to register D0. A 32-bit integer is returned in register I0 or I1.
def RetCC_Sparc32 : CallingConv<[ CCIfType<[i32], CCAssignToReg<[I0, I1]>>, CCIfType<[f32], CCAssignToReg<[F0]>>, CCIfType<[f64], CCAssignToReg<[D0]>> ]>;
The definition of CC_Sparc32 in SparcCallingConv.td introduces CCAssignToStack, which assigns the value to a stack slot with the specified size and alignment. In the example below, the first parameter, 4, indicates the size of the slot, and the second parameter, also 4, indicates the stack alignment along 4-byte units. (Special cases: if size is zero, then the ABI size is used; if alignment is zero, then the ABI alignment is used.)
def CC_Sparc32 : CallingConv<[ // All arguments get passed in integer registers if there is space. CCIfType<[i32, f32, f64], CCAssignToReg<[I0, I1, I2, I3, I4, I5]>>, CCAssignToStack<4, 4> ]>;
CCDelegateTo is another commonly used interface, which tries to find a specified sub-calling convention and, if a match is found, it is invoked. In the following example (in X86CallingConv.td), the definition of RetCC_X86_32_C ends with CCDelegateTo. After the current value is assigned to the register ST0 or ST1, the RetCC_X86Common is invoked.
def RetCC_X86_32_C : CallingConv<[ CCIfType<[f32], CCAssignToReg<[ST0, ST1]>>, CCIfType<[f64], CCAssignToReg<[ST0, ST1]>>, CCDelegateTo<RetCC_X86Common> ]>;
CCIfCC is an interface that attempts to match the given name to the current calling convention. If the name identifies the current calling convention, then a specified action is invoked. In the following example (in X86CallingConv.td), if the Fast calling convention is in use, then RetCC_X86_32_Fast is invoked. If the SSECall calling convention is in use, then RetCC_X86_32_SSE is invoked.
def RetCC_X86_32 : CallingConv<[ CCIfCC<"CallingConv::Fast", CCDelegateTo<RetCC_X86_32_Fast>>, CCIfCC<"CallingConv::X86_SSECall", CCDelegateTo<RetCC_X86_32_SSE>>, CCDelegateTo<RetCC_X86_32_C> ]>;
Other calling convention interfaces include:
During the code emission stage, the code generator may utilize an LLVM pass to produce assembly output. To do this, you want to implement the code for a printer that converts LLVM IR to a GAS-format assembly language for your target machine, using the following steps:
The code in XXXTargetAsmInfo.h is usually a trivial declaration of the XXXTargetAsmInfo class for use in XXXTargetAsmInfo.cpp. Similarly, XXXTargetAsmInfo.cpp usually has a few declarations of XXXTargetAsmInfo replacement values that override the default values in TargetAsmInfo.cpp. For example in SparcTargetAsmInfo.cpp,
SparcTargetAsmInfo::SparcTargetAsmInfo(const SparcTargetMachine &TM) { Data16bitsDirective = "\t.half\t"; Data32bitsDirective = "\t.word\t"; Data64bitsDirective = 0; // .xword is only supported by V9. ZeroDirective = "\t.skip\t"; CommentString = "!"; ConstantPoolSection = "\t.section \".rodata\",#alloc\n"; }
The X86 assembly printer implementation (X86TargetAsmInfo) is an example where the target specific TargetAsmInfo class uses overridden methods: ExpandInlineAsm and PreferredEHDataFormat.
A target-specific implementation of AsmPrinter is written in XXXAsmPrinter.cpp, which implements the AsmPrinter class that converts the LLVM to printable assembly. The implementation must include the following headers that have declarations for the AsmPrinter and MachineFunctionPass classes. The MachineFunctionPass is a subclass of FunctionPass.
#include "llvm/CodeGen/AsmPrinter.h" #include "llvm/CodeGen/MachineFunctionPass.h"
As a FunctionPass, AsmPrinter first calls doInitialization to set up the AsmPrinter. In SparcAsmPrinter, a Mangler object is instantiated to process variable names.
In XXXAsmPrinter.cpp, the runOnMachineFunction method (declared in MachineFunctionPass) must be implemented for XXXAsmPrinter. In MachineFunctionPass, the runOnFunction method invokes runOnMachineFunction. Target-specific implementations of runOnMachineFunction differ, but generally do the following to process each machine function:
The XXXAsmPrinter implementation must also include the code generated by TableGen that is output in the XXXGenAsmWriter.inc file. The code in XXXGenAsmWriter.inc contains an implementation of the printInstruction method that may call these methods:
The implementations of printDeclare, printImplicitDef, printInlineAsm, and printLabel in AsmPrinter.cpp are generally adequate for printing assembly and do not need to be overridden. (printBasicBlockLabel is another method that is implemented in AsmPrinter.cpp that may be directly used in an implementation of XXXAsmPrinter.)
The printOperand method is implemented with a long switch/case statement for the type of operand: register, immediate, basic block, external symbol, global address, constant pool index, or jump table index. For an instruction with a memory address operand, the printMemOperand method should be implemented to generate the proper output. Similarly, printCCOperand should be used to print a conditional operand.
doFinalization should be overridden in XXXAsmPrinter, and it should be called to shut down the assembly printer. During doFinalization, global variables and constants are printed to output.
Subtarget support is used to inform the code generation process of instruction set variations for a given chip set. For example, the LLVM SPARC implementation provided covers three major versions of the SPARC microprocessor architecture: Version 8 (V8, which is a 32-bit architecture), Version 9 (V9, a 64-bit architecture), and the UltraSPARC architecture. V8 has 16 double-precision floating-point registers that are also usable as either 32 single-precision or 8 quad-precision registers. V8 is also purely big-endian. V9 has 32 double-precision floating-point registers that are also usable as 16 quad-precision registers, but cannot be used as single-precision registers. The UltraSPARC architecture combines V9 with UltraSPARC Visual Instruction Set extensions.
If subtarget support is needed, you should implement a target-specific XXXSubtarget class for your architecture. This class should process the command-line options –mcpu= and –mattr=
TableGen uses definitions in the Target.td and Sparc.td files to generate code in SparcGenSubtarget.inc. In Target.td, shown below, the SubtargetFeature interface is defined. The first 4 string parameters of the SubtargetFeature interface are a feature name, an attribute set by the feature, the value of the attribute, and a description of the feature. (The fifth parameter is a list of features whose presence is implied, and its default value is an empty array.)
class SubtargetFeature<string n, string a, string v, string d, list<SubtargetFeature> i = []> { string Name = n; string Attribute = a; string Value = v; string Desc = d; list<SubtargetFeature> Implies = i; }
In the Sparc.td file, the SubtargetFeature is used to define the following features.
def FeatureV9 : SubtargetFeature<"v9", "IsV9", "true", "Enable SPARC-V9 instructions">; def FeatureV8Deprecated : SubtargetFeature<"deprecated-v8", "V8DeprecatedInsts", "true", "Enable deprecated V8 instructions in V9 mode">; def FeatureVIS : SubtargetFeature<"vis", "IsVIS", "true", "Enable UltraSPARC Visual Instruction Set extensions">;
Elsewhere in Sparc.td, the Proc class is defined and then is used to define particular SPARC processor subtypes that may have the previously described features.
class Proc<string Name, list<SubtargetFeature> Features> : Processor<Name, NoItineraries, Features>; def : Proc<"generic", []>; def : Proc<"v8", []>; def : Proc<"supersparc", []>; def : Proc<"sparclite", []>; def : Proc<"f934", []>; def : Proc<"hypersparc", []>; def : Proc<"sparclite86x", []>; def : Proc<"sparclet", []>; def : Proc<"tsc701", []>; def : Proc<"v9", [FeatureV9]>; def : Proc<"ultrasparc", [FeatureV9, FeatureV8Deprecated]>; def : Proc<"ultrasparc3", [FeatureV9, FeatureV8Deprecated]>; def : Proc<"ultrasparc3-vis", [FeatureV9, FeatureV8Deprecated, FeatureVIS]>;
From Target.td and Sparc.td files, the resulting SparcGenSubtarget.inc specifies enum values to identify the features, arrays of constants to represent the CPU features and CPU subtypes, and the ParseSubtargetFeatures method that parses the features string that sets specified subtarget options. The generated SparcGenSubtarget.inc file should be included in the SparcSubtarget.cpp. The target-specific implementation of the XXXSubtarget method should follow this pseudocode:
XXXSubtarget::XXXSubtarget(const Module &M, const std::string &FS) { // Set the default features // Determine default and user specified characteristics of the CPU // Call ParseSubtargetFeatures(FS, CPU) to parse the features string // Perform any additional operations }
The implementation of a target machine optionally includes a Just-In-Time (JIT) code generator that emits machine code and auxiliary structures as binary output that can be written directly to memory. To do this, implement JIT code generation by performing the following steps:
There are several different approaches to writing the JIT support code. For instance, TableGen and target descriptor files may be used for creating a JIT code generator, but are not mandatory. For the Alpha and PowerPC target machines, TableGen is used to generate XXXGenCodeEmitter.inc, which contains the binary coding of machine instructions and the getBinaryCodeForInstr method to access those codes. Other JIT implementations do not.
Both XXXJITInfo.cpp and XXXCodeEmitter.cpp must include the llvm/CodeGen/MachineCodeEmitter.h header file that defines the MachineCodeEmitter class containing code for several callback functions that write data (in bytes, words, strings, etc.) to the output stream.
In XXXCodeEmitter.cpp, a target-specific of the Emitter class is implemented as a function pass (subclass of MachineFunctionPass). The target-specific implementation of runOnMachineFunction (invoked by runOnFunction in MachineFunctionPass) iterates through the MachineBasicBlock calls emitInstruction to process each instruction and emit binary code. emitInstruction is largely implemented with case statements on the instruction types defined in XXXInstrInfo.h. For example, in X86CodeEmitter.cpp, the emitInstruction method is built around the following switch/case statements:
switch (Desc->TSFlags & X86::FormMask) { case X86II::Pseudo: // for not yet implemented instructions ... // or pseudo-instructions break; case X86II::RawFrm: // for instructions with a fixed opcode value ... break; case X86II::AddRegFrm: // for instructions that have one register operand ... // added to their opcode break; case X86II::MRMDestReg:// for instructions that use the Mod/RM byte ... // to specify a destination (register) break; case X86II::MRMDestMem:// for instructions that use the Mod/RM byte ... // to specify a destination (memory) break; case X86II::MRMSrcReg: // for instructions that use the Mod/RM byte ... // to specify a source (register) break; case X86II::MRMSrcMem: // for instructions that use the Mod/RM byte ... // to specify a source (memory) break; case X86II::MRM0r: case X86II::MRM1r: // for instructions that operate on case X86II::MRM2r: case X86II::MRM3r: // a REGISTER r/m operand and case X86II::MRM4r: case X86II::MRM5r: // use the Mod/RM byte and a field case X86II::MRM6r: case X86II::MRM7r: // to hold extended opcode data ... break; case X86II::MRM0m: case X86II::MRM1m: // for instructions that operate on case X86II::MRM2m: case X86II::MRM3m: // a MEMORY r/m operand and case X86II::MRM4m: case X86II::MRM5m: // use the Mod/RM byte and a field case X86II::MRM6m: case X86II::MRM7m: // to hold extended opcode data ... break; case X86II::MRMInitReg: // for instructions whose source and ... // destination are the same register break; }
The implementations of these case statements often first emit the opcode and then get the operand(s). Then depending upon the operand, helper methods may be called to process the operand(s). For example, in X86CodeEmitter.cpp, for the X86II::AddRegFrm case, the first data emitted (by emitByte) is the opcode added to the register operand. Then an object representing the machine operand, MO1, is extracted. The helper methods such as isImmediate, isGlobalAddress, isExternalSymbol, isConstantPoolIndex, and isJumpTableIndex determine the operand type. (X86CodeEmitter.cpp also has private methods such as emitConstant, emitGlobalAddress, emitExternalSymbolAddress, emitConstPoolAddress, and emitJumpTableAddress that emit the data into the output stream.)
case X86II::AddRegFrm: MCE.emitByte(BaseOpcode + getX86RegNum(MI.getOperand(CurOp++).getReg())); if (CurOp != NumOps) { const MachineOperand &MO1 = MI.getOperand(CurOp++); unsigned Size = X86InstrInfo::sizeOfImm(Desc); if (MO1.isImmediate()) emitConstant(MO1.getImm(), Size); else { unsigned rt = Is64BitMode ? X86::reloc_pcrel_word : (IsPIC ? X86::reloc_picrel_word : X86::reloc_absolute_word); if (Opcode == X86::MOV64ri) rt = X86::reloc_absolute_dword; // FIXME: add X86II flag? if (MO1.isGlobalAddress()) { bool NeedStub = isa<Function>(MO1.getGlobal()); bool isLazy = gvNeedsLazyPtr(MO1.getGlobal()); emitGlobalAddress(MO1.getGlobal(), rt, MO1.getOffset(), 0, NeedStub, isLazy); } else if (MO1.isExternalSymbol()) emitExternalSymbolAddress(MO1.getSymbolName(), rt); else if (MO1.isConstantPoolIndex()) emitConstPoolAddress(MO1.getIndex(), rt); else if (MO1.isJumpTableIndex()) emitJumpTableAddress(MO1.getIndex(), rt); } } break;
In the previous example, XXXCodeEmitter.cpp uses the variable rt, which is a RelocationType enum that may be used to relocate addresses (for example, a global address with a PIC base offset). The RelocationType enum for that target is defined in the short target-specific XXXRelocations.h file. The RelocationType is used by the relocate method defined in XXXJITInfo.cpp to rewrite addresses for referenced global symbols.
For example, X86Relocations.h specifies the following relocation types for the X86 addresses. In all four cases, the relocated value is added to the value already in memory. For reloc_pcrel_word and reloc_picrel_word, there is an additional initial adjustment.
enum RelocationType { reloc_pcrel_word = 0, // add reloc value after adjusting for the PC loc reloc_picrel_word = 1, // add reloc value after adjusting for the PIC base reloc_absolute_word = 2, // absolute relocation; no additional adjustment reloc_absolute_dword = 3 // absolute relocation; no additional adjustment };
XXXJITInfo.cpp implements the JIT interfaces for target-specific code-generation activities, such as emitting machine code and stubs. At minimum, a target-specific version of XXXJITInfo implements the following:
getLazyResolverFunction is generally trivial to implement. It makes the incoming parameter as the global JITCompilerFunction and returns the callback function that will be used a function wrapper. For the Alpha target (in AlphaJITInfo.cpp), the getLazyResolverFunction implementation is simply:
TargetJITInfo::LazyResolverFn AlphaJITInfo::getLazyResolverFunction( JITCompilerFn F) { JITCompilerFunction = F; return AlphaCompilationCallback; }
For the X86 target, the getLazyResolverFunction implementation is a little more complication, because it returns a different callback function for processors with SSE instructions and XMM registers.
The callback function initially saves and later restores the callee register values, incoming arguments, and frame and return address. The callback function needs low-level access to the registers or stack, so it is typically implemented with assembler.