The LLVM Target-Independent Code Generator
  1. Introduction
  2. Target description classes
  3. Machine code description classes
  4. Target-independent code generation algorithms
  5. Target-specific Implementation Notes

Written by Chris Lattner, Bill Wendling, Fernando Magno Quintao Pereira and Jim Laskey

Warning: This is a work in progress.

Introduction

The LLVM target-independent code generator is a framework that provides a suite of reusable components for translating the LLVM internal representation to the machine code for a specified target—either in assembly form (suitable for a static compiler) or in binary machine code format (usable for a JIT compiler). The LLVM target-independent code generator consists of five main components:

  1. Abstract target description interfaces which capture important properties about various aspects of the machine, independently of how they will be used. These interfaces are defined in include/llvm/Target/.
  2. Classes used to represent the machine code being generated for a target. These classes are intended to be abstract enough to represent the machine code for any target machine. These classes are defined in include/llvm/CodeGen/.
  3. Target-independent algorithms used to implement various phases of native code generation (register allocation, scheduling, stack frame representation, etc). This code lives in lib/CodeGen/.
  4. Implementations of the abstract target description interfaces for particular targets. These machine descriptions make use of the components provided by LLVM, and can optionally provide custom target-specific passes, to build complete code generators for a specific target. Target descriptions live in lib/Target/.
  5. The target-independent JIT components. The LLVM JIT is completely target independent (it uses the TargetJITInfo structure to interface for target-specific issues. The code for the target-independent JIT lives in lib/ExecutionEngine/JIT.

Depending on which part of the code generator you are interested in working on, different pieces of this will be useful to you. In any case, you should be familiar with the target description and machine code representation classes. If you want to add a backend for a new target, you will need to implement the target description classes for your new target and understand the LLVM code representation. If you are interested in implementing a new code generation algorithm, it should only depend on the target-description and machine code representation classes, ensuring that it is portable.

Required components in the code generator

The two pieces of the LLVM code generator are the high-level interface to the code generator and the set of reusable components that can be used to build target-specific backends. The two most important interfaces (TargetMachine and TargetData) are the only ones that are required to be defined for a backend to fit into the LLVM system, but the others must be defined if the reusable code generator components are going to be used.

This design has two important implications. The first is that LLVM can support completely non-traditional code generation targets. For example, the C backend does not require register allocation, instruction selection, or any of the other standard components provided by the system. As such, it only implements these two interfaces, and does its own thing. Another example of a code generator like this is a (purely hypothetical) backend that converts LLVM to the GCC RTL form and uses GCC to emit machine code for a target.

This design also implies that it is possible to design and implement radically different code generators in the LLVM system that do not make use of any of the built-in components. Doing so is not recommended at all, but could be required for radically different targets that do not fit into the LLVM machine description model: FPGAs for example.

The high-level design of the code generator

The LLVM target-independent code generator is designed to support efficient and quality code generation for standard register-based microprocessors. Code generation in this model is divided into the following stages:

  1. Instruction Selection - This phase determines an efficient way to express the input LLVM code in the target instruction set. This stage produces the initial code for the program in the target instruction set, then makes use of virtual registers in SSA form and physical registers that represent any required register assignments due to target constraints or calling conventions. This step turns the LLVM code into a DAG of target instructions.
  2. Scheduling and Formation - This phase takes the DAG of target instructions produced by the instruction selection phase, determines an ordering of the instructions, then emits the instructions as MachineInstrs with that ordering. Note that we describe this in the instruction selection section because it operates on a SelectionDAG.
  3. SSA-based Machine Code Optimizations - This optional stage consists of a series of machine-code optimizations that operate on the SSA-form produced by the instruction selector. Optimizations like modulo-scheduling or peephole optimization work here.
  4. Register Allocation - The target code is transformed from an infinite virtual register file in SSA form to the concrete register file used by the target. This phase introduces spill code and eliminates all virtual register references from the program.
  5. Prolog/Epilog Code Insertion - Once the machine code has been generated for the function and the amount of stack space required is known (used for LLVM alloca's and spill slots), the prolog and epilog code for the function can be inserted and "abstract stack location references" can be eliminated. This stage is responsible for implementing optimizations like frame-pointer elimination and stack packing.
  6. Late Machine Code Optimizations - Optimizations that operate on "final" machine code can go here, such as spill code scheduling and peephole optimizations.
  7. Code Emission - The final stage actually puts out the code for the current function, either in the target assembler format or in machine code.

The code generator is based on the assumption that the instruction selector will use an optimal pattern matching selector to create high-quality sequences of native instructions. Alternative code generator designs based on pattern expansion and aggressive iterative peephole optimization are much slower. This design permits efficient compilation (important for JIT environments) and aggressive optimization (used when generating code offline) by allowing components of varying levels of sophistication to be used for any step of compilation.

In addition to these stages, target implementations can insert arbitrary target-specific passes into the flow. For example, the X86 target uses a special pass to handle the 80x87 floating point stack architecture. Other targets with unusual requirements can be supported with custom passes as needed.

Using TableGen for target description

The target description classes require a detailed description of the target architecture. These target descriptions often have a large amount of common information (e.g., an add instruction is almost identical to a sub instruction). In order to allow the maximum amount of commonality to be factored out, the LLVM code generator uses the TableGen tool to describe big chunks of the target machine, which allows the use of domain-specific and target-specific abstractions to reduce the amount of repetition.

As LLVM continues to be developed and refined, we plan to move more and more of the target description to the .td form. Doing so gives us a number of advantages. The most important is that it makes it easier to port LLVM because it reduces the amount of C++ code that has to be written, and the surface area of the code generator that needs to be understood before someone can get something working. Second, it makes it easier to change things. In particular, if tables and other things are all emitted by tblgen, we only need a change in one place (tblgen) to update all of the targets to a new interface.

Target description classes

The LLVM target description classes (located in the include/llvm/Target directory) provide an abstract description of the target machine independent of any particular client. These classes are designed to capture the abstract properties of the target (such as the instructions and registers it has), and do not incorporate any particular pieces of code generation algorithms.

All of the target description classes (except the TargetData class) are designed to be subclassed by the concrete target implementation, and have virtual methods implemented. To get to these implementations, the TargetMachine class provides accessors that should be implemented by the target.

The TargetMachine class

The TargetMachine class provides virtual methods that are used to access the target-specific implementations of the various target description classes via the get*Info methods (getInstrInfo, getRegisterInfo, getFrameInfo, etc.). This class is designed to be specialized by a concrete target implementation (e.g., X86TargetMachine) which implements the various virtual methods. The only required target description class is the TargetData class, but if the code generator components are to be used, the other interfaces should be implemented as well.

The TargetData class

The TargetData class is the only required target description class, and it is the only class that is not extensible (you cannot derived a new class from it). TargetData specifies information about how the target lays out memory for structures, the alignment requirements for various data types, the size of pointers in the target, and whether the target is little-endian or big-endian.

The TargetLowering class

The TargetLowering class is used by SelectionDAG based instruction selectors primarily to describe how LLVM code should be lowered to SelectionDAG operations. Among other things, this class indicates:

The TargetRegisterInfo class

The TargetRegisterInfo class is used to describe the register file of the target and any interactions between the registers.

Registers in the code generator are represented in the code generator by unsigned integers. Physical registers (those that actually exist in the target description) are unique small numbers, and virtual registers are generally large. Note that register #0 is reserved as a flag value.

Each register in the processor description has an associated TargetRegisterDesc entry, which provides a textual name for the register (used for assembly output and debugging dumps) and a set of aliases (used to indicate whether one register overlaps with another).

In addition to the per-register description, the TargetRegisterInfo class exposes a set of processor specific register classes (instances of the TargetRegisterClass class). Each register class contains sets of registers that have the same properties (for example, they are all 32-bit integer registers). Each SSA virtual register created by the instruction selector has an associated register class. When the register allocator runs, it replaces virtual registers with a physical register in the set.

The target-specific implementations of these classes is auto-generated from a TableGen description of the register file.

The TargetInstrInfo class

The TargetInstrInfo class is used to describe the machine instructions supported by the target. It is essentially an array of TargetInstrDescriptor objects, each of which describes one instruction the target supports. Descriptors define things like the mnemonic for the opcode, the number of operands, the list of implicit register uses and defs, whether the instruction has certain target-independent properties (accesses memory, is commutable, etc), and holds any target-specific flags.

The TargetFrameInfo class

The TargetFrameInfo class is used to provide information about the stack frame layout of the target. It holds the direction of stack growth, the known stack alignment on entry to each function, and the offset to the local area. The offset to the local area is the offset from the stack pointer on function entry to the first location where function data (local variables, spill locations) can be stored.

The TargetSubtarget class

The TargetSubtarget class is used to provide information about the specific chip set being targeted. A sub-target informs code generation of which instructions are supported, instruction latencies and instruction execution itinerary; i.e., which processing units are used, in what order, and for how long.

The TargetJITInfo class

The TargetJITInfo class exposes an abstract interface used by the Just-In-Time code generator to perform target-specific activities, such as emitting stubs. If a TargetMachine supports JIT code generation, it should provide one of these objects through the getJITInfo method.

Machine code description classes

At the high-level, LLVM code is translated to a machine specific representation formed out of MachineFunction, MachineBasicBlock, and MachineInstr instances (defined in include/llvm/CodeGen). This representation is completely target agnostic, representing instructions in their most abstract form: an opcode and a series of operands. This representation is designed to support both an SSA representation for machine code, as well as a register allocated, non-SSA form.

The MachineInstr class

Target machine instructions are represented as instances of the MachineInstr class. This class is an extremely abstract way of representing machine instructions. In particular, it only keeps track of an opcode number and a set of operands.

The opcode number is a simple unsigned integer that only has meaning to a specific backend. All of the instructions for a target should be defined in the *InstrInfo.td file for the target. The opcode enum values are auto-generated from this description. The MachineInstr class does not have any information about how to interpret the instruction (i.e., what the semantics of the instruction are); for that you must refer to the TargetInstrInfo class.

The operands of a machine instruction can be of several different types: a register reference, a constant integer, a basic block reference, etc. In addition, a machine operand should be marked as a def or a use of the value (though only registers are allowed to be defs).

By convention, the LLVM code generator orders instruction operands so that all register definitions come before the register uses, even on architectures that are normally printed in other orders. For example, the SPARC add instruction: "add %i1, %i2, %i3" adds the "%i1", and "%i2" registers and stores the result into the "%i3" register. In the LLVM code generator, the operands should be stored as "%i3, %i1, %i2": with the destination first.

Keeping destination (definition) operands at the beginning of the operand list has several advantages. In particular, the debugging printer will print the instruction like this:

%r3 = add %i1, %i2

Also if the first operand is a def, it is easier to create instructions whose only def is the first operand.

Using the MachineInstrBuilder.h functions

Machine instructions are created by using the BuildMI functions, located in the include/llvm/CodeGen/MachineInstrBuilder.h file. The BuildMI functions make it easy to build arbitrary machine instructions. Usage of the BuildMI functions look like this:

// Create a 'DestReg = mov 42' (rendered in X86 assembly as 'mov DestReg, 42')
// instruction.  The '1' specifies how many operands will be added.
MachineInstr *MI = BuildMI(X86::MOV32ri, 1, DestReg).addImm(42);

// Create the same instr, but insert it at the end of a basic block.
MachineBasicBlock &MBB = ...
BuildMI(MBB, X86::MOV32ri, 1, DestReg).addImm(42);

// Create the same instr, but insert it before a specified iterator point.
MachineBasicBlock::iterator MBBI = ...
BuildMI(MBB, MBBI, X86::MOV32ri, 1, DestReg).addImm(42);

// Create a 'cmp Reg, 0' instruction, no destination reg.
MI = BuildMI(X86::CMP32ri, 2).addReg(Reg).addImm(0);
// Create an 'sahf' instruction which takes no operands and stores nothing.
MI = BuildMI(X86::SAHF, 0);

// Create a self looping branch instruction.
BuildMI(MBB, X86::JNE, 1).addMBB(&MBB);

The key thing to remember with the BuildMI functions is that you have to specify the number of operands that the machine instruction will take. This allows for efficient memory allocation. You also need to specify if operands default to be uses of values, not definitions. If you need to add a definition operand (other than the optional destination register), you must explicitly mark it as such:

MI.addReg(Reg, MachineOperand::Def);
Fixed (preassigned) registers

One important issue that the code generator needs to be aware of is the presence of fixed registers. In particular, there are often places in the instruction stream where the register allocator must arrange for a particular value to be in a particular register. This can occur due to limitations of the instruction set (e.g., the X86 can only do a 32-bit divide with the EAX/EDX registers), or external factors like calling conventions. In any case, the instruction selector should emit code that copies a virtual register into or out of a physical register when needed.

For example, consider this simple LLVM example:

define i32 @test(i32 %X, i32 %Y) {
  %Z = udiv i32 %X, %Y
  ret i32 %Z
}

The X86 instruction selector produces this machine code for the div and ret (use "llc X.bc -march=x86 -print-machineinstrs" to get this):

;; Start of div
%EAX = mov %reg1024           ;; Copy X (in reg1024) into EAX
%reg1027 = sar %reg1024, 31
%EDX = mov %reg1027           ;; Sign extend X into EDX
idiv %reg1025                 ;; Divide by Y (in reg1025)
%reg1026 = mov %EAX           ;; Read the result (Z) out of EAX

;; Start of ret
%EAX = mov %reg1026           ;; 32-bit return value goes in EAX
ret

By the end of code generation, the register allocator has coalesced the registers and deleted the resultant identity moves producing the following code:

;; X is in EAX, Y is in ECX
mov %EAX, %EDX
sar %EDX, 31
idiv %ECX
ret 

This approach is extremely general (if it can handle the X86 architecture, it can handle anything!) and allows all of the target specific knowledge about the instruction stream to be isolated in the instruction selector. Note that physical registers should have a short lifetime for good code generation, and all physical registers are assumed dead on entry to and exit from basic blocks (before register allocation). Thus, if you need a value to be live across basic block boundaries, it must live in a virtual register.

Machine code in SSA form

MachineInstr's are initially selected in SSA-form, and are maintained in SSA-form until register allocation happens. For the most part, this is trivially simple since LLVM is already in SSA form; LLVM PHI nodes become machine code PHI nodes, and virtual registers are only allowed to have a single definition.

After register allocation, machine code is no longer in SSA-form because there are no virtual registers left in the code.

The MachineBasicBlock class

The MachineBasicBlock class contains a list of machine instructions (MachineInstr instances). It roughly corresponds to the LLVM code input to the instruction selector, but there can be a one-to-many mapping (i.e. one LLVM basic block can map to multiple machine basic blocks). The MachineBasicBlock class has a "getBasicBlock" method, which returns the LLVM basic block that it comes from.

The MachineFunction class

The MachineFunction class contains a list of machine basic blocks (MachineBasicBlock instances). It corresponds one-to-one with the LLVM function input to the instruction selector. In addition to a list of basic blocks, the MachineFunction contains a a MachineConstantPool, a MachineFrameInfo, a MachineFunctionInfo, and a MachineRegisterInfo. See include/llvm/CodeGen/MachineFunction.h for more information.

Target-independent code generation algorithms

This section documents the phases described in the high-level design of the code generator. It explains how they work and some of the rationale behind their design.

Instruction Selection

Instruction Selection is the process of translating LLVM code presented to the code generator into target-specific machine instructions. There are several well-known ways to do this in the literature. LLVM uses a SelectionDAG based instruction selector.

Portions of the DAG instruction selector are generated from the target description (*.td) files. Our goal is for the entire instruction selector to be generated from these .td files, though currently there are still things that require custom C++ code.

Introduction to SelectionDAGs

The SelectionDAG provides an abstraction for code representation in a way that is amenable to instruction selection using automatic techniques (e.g. dynamic-programming based optimal pattern matching selectors). It is also well-suited to other phases of code generation; in particular, instruction scheduling (SelectionDAG's are very close to scheduling DAGs post-selection). Additionally, the SelectionDAG provides a host representation where a large variety of very-low-level (but target-independent) optimizations may be performed; ones which require extensive information about the instructions efficiently supported by the target.

The SelectionDAG is a Directed-Acyclic-Graph whose nodes are instances of the SDNode class. The primary payload of the SDNode is its operation code (Opcode) that indicates what operation the node performs and the operands to the operation. The various operation node types are described at the top of the include/llvm/CodeGen/SelectionDAGNodes.h file.

Although most operations define a single value, each node in the graph may define multiple values. For example, a combined div/rem operation will define both the dividend and the remainder. Many other situations require multiple values as well. Each node also has some number of operands, which are edges to the node defining the used value. Because nodes may define multiple values, edges are represented by instances of the SDValue class, which is a <SDNode, unsigned> pair, indicating the node and result value being used, respectively. Each value produced by an SDNode has an associated MVT (Machine Value Type) indicating what the type of the value is.

SelectionDAGs contain two different kinds of values: those that represent data flow and those that represent control flow dependencies. Data values are simple edges with an integer or floating point value type. Control edges are represented as "chain" edges which are of type MVT::Other. These edges provide an ordering between nodes that have side effects (such as loads, stores, calls, returns, etc). All nodes that have side effects should take a token chain as input and produce a new one as output. By convention, token chain inputs are always operand #0, and chain results are always the last value produced by an operation.

A SelectionDAG has designated "Entry" and "Root" nodes. The Entry node is always a marker node with an Opcode of ISD::EntryToken. The Root node is the final side-effecting node in the token chain. For example, in a single basic block function it would be the return node.

One important concept for SelectionDAGs is the notion of a "legal" vs. "illegal" DAG. A legal DAG for a target is one that only uses supported operations and supported types. On a 32-bit PowerPC, for example, a DAG with a value of type i1, i8, i16, or i64 would be illegal, as would a DAG that uses a SREM or UREM operation. The legalize types and legalize operations phases are responsible for turning an illegal DAG into a legal DAG.

SelectionDAG Instruction Selection Process

SelectionDAG-based instruction selection consists of the following steps:

  1. Build initial DAG - This stage performs a simple translation from the input LLVM code to an illegal SelectionDAG.
  2. Optimize SelectionDAG - This stage performs simple optimizations on the SelectionDAG to simplify it, and recognize meta instructions (like rotates and div/rem pairs) for targets that support these meta operations. This makes the resultant code more efficient and the select instructions from DAG phase (below) simpler.
  3. Legalize SelectionDAG Types - This stage transforms SelectionDAG nodes to eliminate any types that are unsupported on the target.
  4. Optimize SelectionDAG - The SelectionDAG optimizer is run to clean up redundancies exposed by type legalization.
  5. Legalize SelectionDAG Types - This stage transforms SelectionDAG nodes to eliminate any types that are unsupported on the target.
  6. Optimize SelectionDAG - The SelectionDAG optimizer is run to eliminate inefficiencies introduced by operation legalization.
  7. Select instructions from DAG - Finally, the target instruction selector matches the DAG operations to target instructions. This process translates the target-independent input DAG into another DAG of target instructions.
  8. SelectionDAG Scheduling and Formation - The last phase assigns a linear order to the instructions in the target-instruction DAG and emits them into the MachineFunction being compiled. This step uses traditional prepass scheduling techniques.

After all of these steps are complete, the SelectionDAG is destroyed and the rest of the code generation passes are run.

One great way to visualize what is going on here is to take advantage of a few LLC command line options. The following options pop up a window displaying the SelectionDAG at specific times (if you only get errors printed to the console while using this, you probably need to configure your system to add support for it).

The -view-sunit-dags displays the Scheduler's dependency graph. This graph is based on the final SelectionDAG, with nodes that must be scheduled together bundled into a single scheduling-unit node, and with immediate operands and other nodes that aren't relevant for scheduling omitted.

Initial SelectionDAG Construction

The initial SelectionDAG is naïvely peephole expanded from the LLVM input by the SelectionDAGLowering class in the lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp file. The intent of this pass is to expose as much low-level, target-specific details to the SelectionDAG as possible. This pass is mostly hard-coded (e.g. an LLVM add turns into an SDNode add while a getelementptr is expanded into the obvious arithmetic). This pass requires target-specific hooks to lower calls, returns, varargs, etc. For these features, the TargetLowering interface is used.

SelectionDAG LegalizeTypes Phase

The Legalize phase is in charge of converting a DAG to only use the types that are natively supported by the target.

There are two main ways of converting values of unsupported scalar types to values of supported types: converting small types to larger types ("promoting"), and breaking up large integer types into smaller ones ("expanding"). For example, a target might require that all f32 values are promoted to f64 and that all i1/i8/i16 values are promoted to i32. The same target might require that all i64 values be expanded into pairs of i32 values. These changes can insert sign and zero extensions as needed to make sure that the final code has the same behavior as the input.

There are two main ways of converting values of unsupported vector types to value of supported types: splitting vector types, multiple times if necessary, until a legal type is found, and extending vector types by adding elements to the end to round them out to legal types ("widening"). If a vector gets split all the way down to single-element parts with no supported vector type being found, the elements are converted to scalars ("scalarizing").

A target implementation tells the legalizer which types are supported (and which register class to use for them) by calling the addRegisterClass method in its TargetLowering constructor.

SelectionDAG Legalize Phase

The Legalize phase is in charge of converting a DAG to only use the operations that are natively supported by the target.

Targets often have weird constraints, such as not supporting every operation on every supported datatype (e.g. X86 does not support byte conditional moves and PowerPC does not support sign-extending loads from a 16-bit memory location). Legalize takes care of this by open-coding another sequence of operations to emulate the operation ("expansion"), by promoting one type to a larger type that supports the operation ("promotion"), or by using a target-specific hook to implement the legalization ("custom").

A target implementation tells the legalizer which operations are not supported (and which of the above three actions to take) by calling the setOperationAction method in its TargetLowering constructor.

Prior to the existence of the Legalize passes, we required that every target selector supported and handled every operator and type even if they are not natively supported. The introduction of the Legalize phases allows all of the canonicalization patterns to be shared across targets, and makes it very easy to optimize the canonicalized code because it is still in the form of a DAG.

SelectionDAG Optimization Phase: the DAG Combiner

The SelectionDAG optimization phase is run multiple times for code generation, immediately after the DAG is built and once after each legalization. The first run of the pass allows the initial code to be cleaned up (e.g. performing optimizations that depend on knowing that the operators have restricted type inputs). Subsequent runs of the pass clean up the messy code generated by the Legalize passes, which allows Legalize to be very simple (it can focus on making code legal instead of focusing on generating good and legal code).

One important class of optimizations performed is optimizing inserted sign and zero extension instructions. We currently use ad-hoc techniques, but could move to more rigorous techniques in the future. Here are some good papers on the subject:

"Widening integer arithmetic"
Kevin Redwine and Norman Ramsey
International Conference on Compiler Construction (CC) 2004

"Effective sign extension elimination"
Motohiro Kawahito, Hideaki Komatsu, and Toshio Nakatani
Proceedings of the ACM SIGPLAN 2002 Conference on Programming Language Design and Implementation.

SelectionDAG Select Phase

The Select phase is the bulk of the target-specific code for instruction selection. This phase takes a legal SelectionDAG as input, pattern matches the instructions supported by the target to this DAG, and produces a new DAG of target code. For example, consider the following LLVM fragment:

%t1 = add float %W, %X
%t2 = mul float %t1, %Y
%t3 = add float %t2, %Z

This LLVM code corresponds to a SelectionDAG that looks basically like this:

(fadd:f32 (fmul:f32 (fadd:f32 W, X), Y), Z)

If a target supports floating point multiply-and-add (FMA) operations, one of the adds can be merged with the multiply. On the PowerPC, for example, the output of the instruction selector might look like this DAG:

(FMADDS (FADDS W, X), Y, Z)

The FMADDS instruction is a ternary instruction that multiplies its first two operands and adds the third (as single-precision floating-point numbers). The FADDS instruction is a simple binary single-precision add instruction. To perform this pattern match, the PowerPC backend includes the following instruction definitions:

def FMADDS : AForm_1<59, 29,
                    (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRC, F4RC:$FRB),
                    "fmadds $FRT, $FRA, $FRC, $FRB",
                    [(set F4RC:$FRT, (fadd (fmul F4RC:$FRA, F4RC:$FRC),
                                           F4RC:$FRB))]>;
def FADDS : AForm_2<59, 21,
                    (ops F4RC:$FRT, F4RC:$FRA, F4RC:$FRB),
                    "fadds $FRT, $FRA, $FRB",
                    [(set F4RC:$FRT, (fadd F4RC:$FRA, F4RC:$FRB))]>;

The portion of the instruction definition in bold indicates the pattern used to match the instruction. The DAG operators (like fmul/fadd) are defined in the lib/Target/TargetSelectionDAG.td file. "F4RC" is the register class of the input and result values.

The TableGen DAG instruction selector generator reads the instruction patterns in the .td file and automatically builds parts of the pattern matching code for your target. It has the following strengths:

While it has many strengths, the system currently has some limitations, primarily because it is a work in progress and is not yet finished:

Despite these limitations, the instruction selector generator is still quite useful for most of the binary and logical operations in typical instruction sets. If you run into any problems or can't figure out how to do something, please let Chris know!

SelectionDAG Scheduling and Formation Phase

The scheduling phase takes the DAG of target instructions from the selection phase and assigns an order. The scheduler can pick an order depending on various constraints of the machines (i.e. order for minimal register pressure or try to cover instruction latencies). Once an order is established, the DAG is converted to a list of MachineInstrs and the SelectionDAG is destroyed.

Note that this phase is logically separate from the instruction selection phase, but is tied to it closely in the code because it operates on SelectionDAGs.

Future directions for the SelectionDAG
  1. Optional function-at-a-time selection.
  2. Auto-generate entire selector from .td file.
SSA-based Machine Code Optimizations

To Be Written

Live Intervals

Live Intervals are the ranges (intervals) where a variable is live. They are used by some register allocator passes to determine if two or more virtual registers which require the same physical register are live at the same point in the program (i.e., they conflict). When this situation occurs, one virtual register must be spilled.

Live Variable Analysis

The first step in determining the live intervals of variables is to calculate the set of registers that are immediately dead after the instruction (i.e., the instruction calculates the value, but it is never used) and the set of registers that are used by the instruction, but are never used after the instruction (i.e., they are killed). Live variable information is computed for each virtual register and register allocatable physical register in the function. This is done in a very efficient manner because it uses SSA to sparsely compute lifetime information for virtual registers (which are in SSA form) and only has to track physical registers within a block. Before register allocation, LLVM can assume that physical registers are only live within a single basic block. This allows it to do a single, local analysis to resolve physical register lifetimes within each basic block. If a physical register is not register allocatable (e.g., a stack pointer or condition codes), it is not tracked.

Physical registers may be live in to or out of a function. Live in values are typically arguments in registers. Live out values are typically return values in registers. Live in values are marked as such, and are given a dummy "defining" instruction during live intervals analysis. If the last basic block of a function is a return, then it's marked as using all live out values in the function.

PHI nodes need to be handled specially, because the calculation of the live variable information from a depth first traversal of the CFG of the function won't guarantee that a virtual register used by the PHI node is defined before it's used. When a PHI node is encountered, only the definition is handled, because the uses will be handled in other basic blocks.

For each PHI node of the current basic block, we simulate an assignment at the end of the current basic block and traverse the successor basic blocks. If a successor basic block has a PHI node and one of the PHI node's operands is coming from the current basic block, then the variable is marked as alive within the current basic block and all of its predecessor basic blocks, until the basic block with the defining instruction is encountered.

Live Intervals Analysis

We now have the information available to perform the live intervals analysis and build the live intervals themselves. We start off by numbering the basic blocks and machine instructions. We then handle the "live-in" values. These are in physical registers, so the physical register is assumed to be killed by the end of the basic block. Live intervals for virtual registers are computed for some ordering of the machine instructions [1, N]. A live interval is an interval [i, j), where 1 <= i <= j < N, for which a variable is live.

More to come...

Register Allocation

The Register Allocation problem consists in mapping a program Pv, that can use an unbounded number of virtual registers, to a program Pp that contains a finite (possibly small) number of physical registers. Each target architecture has a different number of physical registers. If the number of physical registers is not enough to accommodate all the virtual registers, some of them will have to be mapped into memory. These virtuals are called spilled virtuals.

How registers are represented in LLVM

In LLVM, physical registers are denoted by integer numbers that normally range from 1 to 1023. To see how this numbering is defined for a particular architecture, you can read the GenRegisterNames.inc file for that architecture. For instance, by inspecting lib/Target/X86/X86GenRegisterNames.inc we see that the 32-bit register EAX is denoted by 15, and the MMX register MM0 is mapped to 48.

Some architectures contain registers that share the same physical location. A notable example is the X86 platform. For instance, in the X86 architecture, the registers EAX, AX and AL share the first eight bits. These physical registers are marked as aliased in LLVM. Given a particular architecture, you can check which registers are aliased by inspecting its RegisterInfo.td file. Moreover, the method TargetRegisterInfo::getAliasSet(p_reg) returns an array containing all the physical registers aliased to the register p_reg.

Physical registers, in LLVM, are grouped in Register Classes. Elements in the same register class are functionally equivalent, and can be interchangeably used. Each virtual register can only be mapped to physical registers of a particular class. For instance, in the X86 architecture, some virtuals can only be allocated to 8 bit registers. A register class is described by TargetRegisterClass objects. To discover if a virtual register is compatible with a given physical, this code can be used:

bool RegMapping_Fer::compatible_class(MachineFunction &mf,
                                      unsigned v_reg,
                                      unsigned p_reg) {
  assert(TargetRegisterInfo::isPhysicalRegister(p_reg) &&
         "Target register must be physical");
  const TargetRegisterClass *trc = mf.getRegInfo().getRegClass(v_reg);
  return trc->contains(p_reg);
}

Sometimes, mostly for debugging purposes, it is useful to change the number of physical registers available in the target architecture. This must be done statically, inside the TargetRegsterInfo.td file. Just grep for RegisterClass, the last parameter of which is a list of registers. Just commenting some out is one simple way to avoid them being used. A more polite way is to explicitly exclude some registers from the allocation order. See the definition of the GR register class in lib/Target/IA64/IA64RegisterInfo.td for an example of this (e.g., numReservedRegs registers are hidden.)

Virtual registers are also denoted by integer numbers. Contrary to physical registers, different virtual registers never share the same number. The smallest virtual register is normally assigned the number 1024. This may change, so, in order to know which is the first virtual register, you should access TargetRegisterInfo::FirstVirtualRegister. Any register whose number is greater than or equal to TargetRegisterInfo::FirstVirtualRegister is considered a virtual register. Whereas physical registers are statically defined in a TargetRegisterInfo.td file and cannot be created by the application developer, that is not the case with virtual registers. In order to create new virtual registers, use the method MachineRegisterInfo::createVirtualRegister(). This method will return a virtual register with the highest code.

Before register allocation, the operands of an instruction are mostly virtual registers, although physical registers may also be used. In order to check if a given machine operand is a register, use the boolean function MachineOperand::isRegister(). To obtain the integer code of a register, use MachineOperand::getReg(). An instruction may define or use a register. For instance, ADD reg:1026 := reg:1025 reg:1024 defines the registers 1024, and uses registers 1025 and 1026. Given a register operand, the method MachineOperand::isUse() informs if that register is being used by the instruction. The method MachineOperand::isDef() informs if that registers is being defined.

We will call physical registers present in the LLVM bitcode before register allocation pre-colored registers. Pre-colored registers are used in many different situations, for instance, to pass parameters of functions calls, and to store results of particular instructions. There are two types of pre-colored registers: the ones implicitly defined, and those explicitly defined. Explicitly defined registers are normal operands, and can be accessed with MachineInstr::getOperand(int)::getReg(). In order to check which registers are implicitly defined by an instruction, use the TargetInstrInfo::get(opcode)::ImplicitDefs, where opcode is the opcode of the target instruction. One important difference between explicit and implicit physical registers is that the latter are defined statically for each instruction, whereas the former may vary depending on the program being compiled. For example, an instruction that represents a function call will always implicitly define or use the same set of physical registers. To read the registers implicitly used by an instruction, use TargetInstrInfo::get(opcode)::ImplicitUses. Pre-colored registers impose constraints on any register allocation algorithm. The register allocator must make sure that none of them is been overwritten by the values of virtual registers while still alive.

Mapping virtual registers to physical registers

There are two ways to map virtual registers to physical registers (or to memory slots). The first way, that we will call direct mapping, is based on the use of methods of the classes TargetRegisterInfo, and MachineOperand. The second way, that we will call indirect mapping, relies on the VirtRegMap class in order to insert loads and stores sending and getting values to and from memory.

The direct mapping provides more flexibility to the developer of the register allocator; however, it is more error prone, and demands more implementation work. Basically, the programmer will have to specify where load and store instructions should be inserted in the target function being compiled in order to get and store values in memory. To assign a physical register to a virtual register present in a given operand, use MachineOperand::setReg(p_reg). To insert a store instruction, use TargetRegisterInfo::storeRegToStackSlot(...), and to insert a load instruction, use TargetRegisterInfo::loadRegFromStackSlot.

The indirect mapping shields the application developer from the complexities of inserting load and store instructions. In order to map a virtual register to a physical one, use VirtRegMap::assignVirt2Phys(vreg, preg). In order to map a certain virtual register to memory, use VirtRegMap::assignVirt2StackSlot(vreg). This method will return the stack slot where vreg's value will be located. If it is necessary to map another virtual register to the same stack slot, use VirtRegMap::assignVirt2StackSlot(vreg, stack_location). One important point to consider when using the indirect mapping, is that even if a virtual register is mapped to memory, it still needs to be mapped to a physical register. This physical register is the location where the virtual register is supposed to be found before being stored or after being reloaded.

If the indirect strategy is used, after all the virtual registers have been mapped to physical registers or stack slots, it is necessary to use a spiller object to place load and store instructions in the code. Every virtual that has been mapped to a stack slot will be stored to memory after been defined and will be loaded before being used. The implementation of the spiller tries to recycle load/store instructions, avoiding unnecessary instructions. For an example of how to invoke the spiller, see RegAllocLinearScan::runOnMachineFunction in lib/CodeGen/RegAllocLinearScan.cpp.

Handling two address instructions

With very rare exceptions (e.g., function calls), the LLVM machine code instructions are three address instructions. That is, each instruction is expected to define at most one register, and to use at most two registers. However, some architectures use two address instructions. In this case, the defined register is also one of the used register. For instance, an instruction such as ADD %EAX, %EBX, in X86 is actually equivalent to %EAX = %EAX + %EBX.

In order to produce correct code, LLVM must convert three address instructions that represent two address instructions into true two address instructions. LLVM provides the pass TwoAddressInstructionPass for this specific purpose. It must be run before register allocation takes place. After its execution, the resulting code may no longer be in SSA form. This happens, for instance, in situations where an instruction such as %a = ADD %b %c is converted to two instructions such as:

%a = MOVE %b
%a = ADD %a %c

Notice that, internally, the second instruction is represented as ADD %a[def/use] %c. I.e., the register operand %a is both used and defined by the instruction.

The SSA deconstruction phase

An important transformation that happens during register allocation is called the SSA Deconstruction Phase. The SSA form simplifies many analyses that are performed on the control flow graph of programs. However, traditional instruction sets do not implement PHI instructions. Thus, in order to generate executable code, compilers must replace PHI instructions with other instructions that preserve their semantics.

There are many ways in which PHI instructions can safely be removed from the target code. The most traditional PHI deconstruction algorithm replaces PHI instructions with copy instructions. That is the strategy adopted by LLVM. The SSA deconstruction algorithm is implemented in lib/CodeGen/PHIElimination.cpp. In order to invoke this pass, the identifier PHIEliminationID must be marked as required in the code of the register allocator.

Instruction folding

Instruction folding is an optimization performed during register allocation that removes unnecessary copy instructions. For instance, a sequence of instructions such as:

%EBX = LOAD %mem_address
%EAX = COPY %EBX

can be safely substituted by the single instruction:

%EAX = LOAD %mem_address

Instructions can be folded with the TargetRegisterInfo::foldMemoryOperand(...) method. Care must be taken when folding instructions; a folded instruction can be quite different from the original instruction. See LiveIntervals::addIntervalsForSpills in lib/CodeGen/LiveIntervalAnalysis.cpp for an example of its use.

Built in register allocators

The LLVM infrastructure provides the application developer with three different register allocators:

The type of register allocator used in llc can be chosen with the command line option -regalloc=...:

$ llc -f -regalloc=simple file.bc -o sp.s;
$ llc -f -regalloc=local file.bc -o lc.s;
$ llc -f -regalloc=linearscan file.bc -o ln.s;
Prolog/Epilog Code Insertion

To Be Written

Late Machine Code Optimizations

To Be Written

Code Emission

To Be Written

Generating Assembly Code

To Be Written

Generating Binary Machine Code

For the JIT or .o file writer

Target-specific Implementation Notes

This section of the document explains features or design decisions that are specific to the code generator for a particular target.

Tail call optimization

Tail call optimization, callee reusing the stack of the caller, is currently supported on x86/x86-64 and PowerPC. It is performed if:

x86/x86-64 constraints:

PowerPC constraints:

Example:

Call as llc -tailcallopt test.ll.

declare fastcc i32 @tailcallee(i32 inreg %a1, i32 inreg %a2, i32 %a3, i32 %a4)

define fastcc i32 @tailcaller(i32 %in1, i32 %in2) {
  %l1 = add i32 %in1, %in2
  %tmp = tail call fastcc i32 @tailcallee(i32 %in1 inreg, i32 %in2 inreg, i32 %in1, i32 %l1)
  ret i32 %tmp
}

Implications of -tailcallopt:

To support tail call optimization in situations where the callee has more arguments than the caller a 'callee pops arguments' convention is used. This currently causes each fastcc call that is not tail call optimized (because one or more of above constraints are not met) to be followed by a readjustment of the stack. So performance might be worse in such cases.

On x86 and x86-64 one register is reserved for indirect tail calls (e.g via a function pointer). So there is one less register for integer argument passing. For x86 this means 2 registers (if inreg parameter attribute is used) and for x86-64 this means 5 register are used.

The X86 backend

The X86 code generator lives in the lib/Target/X86 directory. This code generator is capable of targeting a variety of x86-32 and x86-64 processors, and includes support for ISA extensions such as MMX and SSE.

X86 Target Triples supported

The following are the known target triples that are supported by the X86 backend. This is not an exhaustive list, and it would be useful to add those that people test.

X86 Calling Conventions supported

The following target-specific calling conventions are known to backend:

Representing X86 addressing modes in MachineInstrs

The x86 has a very flexible way of accessing memory. It is capable of forming memory addresses of the following expression directly in integer instructions (which use ModR/M addressing):

Base + [1,2,4,8] * IndexReg + Disp32

In order to represent this, LLVM tracks no less than 4 operands for each memory operand of this form. This means that the "load" form of 'mov' has the following MachineOperands in this order:

Index:        0     |    1        2       3           4
Meaning:   DestReg, | BaseReg,  Scale, IndexReg, Displacement
OperandTy: VirtReg, | VirtReg, UnsImm, VirtReg,   SignExtImm

Stores, and all other instructions, treat the four memory operands in the same way and in the same order.

X86 address spaces supported

x86 has the ability to perform loads and stores to different address spaces via the x86 segment registers. A segment override prefix byte on an instruction causes the instruction's memory access to go to the specified segment. LLVM address space 0 is the default address space, which includes the stack, and any unqualified memory accesses in a program. Address spaces 1-255 are currently reserved for user-defined code. The GS-segment is represented by address space 256. Other x86 segments have yet to be allocated address space numbers.

Some operating systems use the GS-segment to implement TLS, so care should be taken when reading and writing to address space 256 on these platforms.

Instruction naming

An instruction name consists of the base name, a default operand size, and a a character per operand with an optional special size. For example:

ADD8rr -> add, 8-bit register, 8-bit register
IMUL16rmi -> imul, 16-bit register, 16-bit memory, 16-bit immediate
IMUL16rmi8 -> imul, 16-bit register, 16-bit memory, 8-bit immediate
MOVSX32rm16 -> movsx, 32-bit register, 16-bit memory

The PowerPC backend

The PowerPC code generator lives in the lib/Target/PowerPC directory. The code generation is retargetable to several variations or subtargets of the PowerPC ISA; including ppc32, ppc64 and altivec.

LLVM PowerPC ABI

LLVM follows the AIX PowerPC ABI, with two deviations. LLVM uses a PC relative (PIC) or static addressing for accessing global values, so no TOC (r2) is used. Second, r31 is used as a frame pointer to allow dynamic growth of a stack frame. LLVM takes advantage of having no TOC to provide space to save the frame pointer in the PowerPC linkage area of the caller frame. Other details of PowerPC ABI can be found at PowerPC ABI. Note: This link describes the 32 bit ABI. The 64 bit ABI is similar except space for GPRs are 8 bytes wide (not 4) and r13 is reserved for system use.

Frame Layout

The size of a PowerPC frame is usually fixed for the duration of a function’s invocation. Since the frame is fixed size, all references into the frame can be accessed via fixed offsets from the stack pointer. The exception to this is when dynamic alloca or variable sized arrays are present, then a base pointer (r31) is used as a proxy for the stack pointer and stack pointer is free to grow or shrink. A base pointer is also used if llvm-gcc is not passed the -fomit-frame-pointer flag. The stack pointer is always aligned to 16 bytes, so that space allocated for altivec vectors will be properly aligned.

An invocation frame is laid out as follows (low memory at top);

Linkage

Parameter area

Dynamic area

Locals area

Saved registers area


Previous Frame

The linkage area is used by a callee to save special registers prior to allocating its own frame. Only three entries are relevant to LLVM. The first entry is the previous stack pointer (sp), aka link. This allows probing tools like gdb or exception handlers to quickly scan the frames in the stack. A function epilog can also use the link to pop the frame from the stack. The third entry in the linkage area is used to save the return address from the lr register. Finally, as mentioned above, the last entry is used to save the previous frame pointer (r31.) The entries in the linkage area are the size of a GPR, thus the linkage area is 24 bytes long in 32 bit mode and 48 bytes in 64 bit mode.

32 bit linkage area

0 Saved SP (r1)
4 Saved CR
8 Saved LR
12 Reserved
16 Reserved
20 Saved FP (r31)

64 bit linkage area

0 Saved SP (r1)
8 Saved CR
16 Saved LR
24 Reserved
32 Reserved
40 Saved FP (r31)

The parameter area is used to store arguments being passed to a callee function. Following the PowerPC ABI, the first few arguments are actually passed in registers, with the space in the parameter area unused. However, if there are not enough registers or the callee is a thunk or vararg function, these register arguments can be spilled into the parameter area. Thus, the parameter area must be large enough to store all the parameters for the largest call sequence made by the caller. The size must also be minimally large enough to spill registers r3-r10. This allows callees blind to the call signature, such as thunks and vararg functions, enough space to cache the argument registers. Therefore, the parameter area is minimally 32 bytes (64 bytes in 64 bit mode.) Also note that since the parameter area is a fixed offset from the top of the frame, that a callee can access its spilt arguments using fixed offsets from the stack pointer (or base pointer.)

Combining the information about the linkage, parameter areas and alignment. A stack frame is minimally 64 bytes in 32 bit mode and 128 bytes in 64 bit mode.

The dynamic area starts out as size zero. If a function uses dynamic alloca then space is added to the stack, the linkage and parameter areas are shifted to top of stack, and the new space is available immediately below the linkage and parameter areas. The cost of shifting the linkage and parameter areas is minor since only the link value needs to be copied. The link value can be easily fetched by adding the original frame size to the base pointer. Note that allocations in the dynamic space need to observe 16 byte alignment.

The locals area is where the llvm compiler reserves space for local variables.

The saved registers area is where the llvm compiler spills callee saved registers on entry to the callee.

Prolog/Epilog

The llvm prolog and epilog are the same as described in the PowerPC ABI, with the following exceptions. Callee saved registers are spilled after the frame is created. This allows the llvm epilog/prolog support to be common with other targets. The base pointer callee saved register r31 is saved in the TOC slot of linkage area. This simplifies allocation of space for the base pointer and makes it convenient to locate programatically and during debugging.

Dynamic Allocation

TODO - More to come.


Valid CSS Valid HTML 4.01 Chris Lattner
The LLVM Compiler Infrastructure
Last modified: $Date: 2009/03/03 05:17:36 $