A Quick Guide to Go's Assembler

This document is a quick outline of the unusual form of assembly language used by the gc Go compiler. The document is not comprehensive.

The assembler is based on the input style of the Plan 9 assemblers, which is documented in detail elsewhere. If you plan to write assembly language, you should read that document although much of it is Plan 9-specific. The current document provides a summary of the syntax and the differences with what is explained in that document, and describes the peculiarities that apply when writing assembly code to interact with Go.

The most important thing to know about Go's assembler is that it is not a direct representation of the underlying machine. Some of the details map precisely to the machine, but some do not. This is because the compiler suite (see this description) needs no assembler pass in the usual pipeline. Instead, the compiler operates on a kind of semi-abstract instruction set, and instruction selection occurs partly after code generation. The assembler works on the semi-abstract form, so when you see an instruction like MOV what the tool chain actually generates for that operation might not be a move instruction at all, perhaps a clear or load. Or it might correspond exactly to the machine instruction with that name. In general, machine-specific operations tend to appear as themselves, while more general concepts like memory move and subroutine call and return are more abstract. The details vary with architecture, and we apologize for the imprecision; the situation is not well-defined.

The assembler program is a way to parse a description of that semi-abstract instruction set and turn it into instructions to be input to the linker. If you want to see what the instructions look like in assembly for a given architecture, say amd64, there are many examples in the sources of the standard library, in packages such as runtime and math/big. You can also examine what the compiler emits as assembly code (the actual output may differ from what you see here):

$ cat x.go
package main

func main() {
	println(3)
}
$ GOOS=linux GOARCH=amd64 go tool compile -S x.go        # or: go build -gcflags -S x.go

--- prog list "main" ---
0000 (x.go:3) TEXT    main+0(SB),$8-0
0001 (x.go:3) FUNCDATA $0,gcargs·0+0(SB)
0002 (x.go:3) FUNCDATA $1,gclocals·0+0(SB)
0003 (x.go:4) MOVQ    $3,(SP)
0004 (x.go:4) PCDATA  $0,$8
0005 (x.go:4) CALL    ,runtime.printint+0(SB)
0006 (x.go:4) PCDATA  $0,$-1
0007 (x.go:4) PCDATA  $0,$0
0008 (x.go:4) CALL    ,runtime.printnl+0(SB)
0009 (x.go:4) PCDATA  $0,$-1
0010 (x.go:5) RET     ,
...

The FUNCDATA and PCDATA directives contain information for use by the garbage collector; they are introduced by the compiler.

Constants

Although the assembler takes its guidance from the Plan 9 assemblers, it is a distinct program, so there are some differences. One is in constant evaluation. Constant expressions in the assembler are parsed using Go's operator precedence, not the C-like precedence of the original. Thus 3&1<<2 is 4, not 0—it parses as (3&1)<<2 not 3&(1<<2). Also, constants are always evaluated as 64-bit unsigned integers. Thus -2 is not the integer value minus two, but the unsigned 64-bit integer with the same bit pattern. The distinction rarely matters but to avoid ambiguity, division or right shift where the right operand's high bit is set is rejected.

Symbols

Some symbols, such as R1 or LR, are predefined and refer to registers. The exact set depends on the architecture.

There are four predeclared symbols that refer to pseudo-registers. These are not real registers, but rather virtual registers maintained by the tool chain, such as a frame pointer. The set of pseudo-registers is the same for all architectures:

All user-defined symbols are written as offsets to the pseudo-registers FP (arguments and locals) and SB (globals).

The SB pseudo-register can be thought of as the origin of memory, so the symbol foo(SB) is the name foo as an address in memory. This form is used to name global functions and data. Adding <> to the name, as in foo<>(SB), makes the name visible only in the current source file, like a top-level static declaration in a C file. Adding an offset to the name refers to that offset from the symbol's address, so a+4(SB) is four bytes past the start of foo.

The FP pseudo-register is a virtual frame pointer used to refer to function arguments. The compilers maintain a virtual frame pointer and refer to the arguments on the stack as offsets from that pseudo-register. Thus 0(FP) is the first argument to the function, 8(FP) is the second (on a 64-bit machine), and so on. However, when referring to a function argument this way, it is necessary to place a name at the beginning, as in first_arg+0(FP) and second_arg+8(FP). (The meaning of the offset—offset from the frame pointer—distinct from its use with SB, where it is an offset from the symbol.) The assembler enforces this convention, rejecting plain 0(FP) and 8(FP). The actual name is semantically irrelevant but should be used to document the argument's name. It is worth stressing that FP is always a pseudo-register, not a hardware register, even on architectures with a hardware frame pointer.

For assembly functions with Go prototypes, go vet will check that the argument names and offsets match. On 32-bit systems, the low and high 32 bits of a 64-bit value are distinguished by adding a _lo or _hi suffix to the name, as in arg_lo+0(FP) or arg_hi+4(FP). If a Go prototype does not name its result, the expected assembly name is ret.

The SP pseudo-register is a virtual stack pointer used to refer to frame-local variables and the arguments being prepared for function calls. It points to the top of the local stack frame, so references should use negative offsets in the range [−framesize, 0): x-8(SP), y-4(SP), and so on.

On architectures with a hardware register named SP, the name prefix distinguishes references to the virtual stack pointer from references to the architectural SP register. That is, x-8(SP) and -8(SP) are different memory locations: the first refers to the virtual stack pointer pseudo-register, while the second refers to the hardware's SP register.

On machines where SP and PC are traditionally aliases for a physical, numbered register, in the Go assembler the names SP and PC are still treated specially; for instance, references to SP require a symbol, much like FP. To access the actual hardware register use the true R name. For example, on the ARM architecture the hardware SP and PC are accessible as R13 and R15.

Branches and direct jumps are always written as offsets to the PC, or as jumps to labels:

label:
	MOVW $0, R1
	JMP label

Each label is visible only within the function in which it is defined. It is therefore permitted for multiple functions in a file to define and use the same label names. Direct jumps and call instructions can target text symbols, such as name(SB), but not offsets from symbols, such as name+4(SB).

Instructions, registers, and assembler directives are always in UPPER CASE to remind you that assembly programming is a fraught endeavor. (Exception: the g register renaming on ARM.)

In Go object files and binaries, the full name of a symbol is the package path followed by a period and the symbol name: fmt.Printf or math/rand.Int. Because the assembler's parser treats period and slash as punctuation, those strings cannot be used directly as identifier names. Instead, the assembler allows the middle dot character U+00B7 and the division slash U+2215 in identifiers and rewrites them to plain period and slash. Within an assembler source file, the symbols above are written as fmt·Printf and math∕rand·Int. The assembly listings generated by the compilers when using the -S flag show the period and slash directly instead of the Unicode replacements required by the assemblers.

Most hand-written assembly files do not include the full package path in symbol names, because the linker inserts the package path of the current object file at the beginning of any name starting with a period: in an assembly source file within the math/rand package implementation, the package's Int function can be referred to as ·Int. This convention avoids the need to hard-code a package's import path in its own source code, making it easier to move the code from one location to another.

Directives

The assembler uses various directives to bind text and data to symbol names. For example, here is a simple complete function definition. The TEXT directive declares the symbol runtime·profileloop and the instructions that follow form the body of the function. The last instruction in a TEXT block must be some sort of jump, usually a RET (pseudo-)instruction. (If it's not, the linker will append a jump-to-itself instruction; there is no fallthrough in TEXTs.) After the symbol, the arguments are flags (see below) and the frame size, a constant (but see below):

TEXT runtime·profileloop(SB),NOSPLIT,$8
	MOVQ	$runtime·profileloop1(SB), CX
	MOVQ	CX, 0(SP)
	CALL	runtime·externalthreadhandler(SB)
	RET

In the general case, the frame size is followed by an argument size, separated by a minus sign. (It's not a subtraction, just idiosyncratic syntax.) The frame size $24-8 states that the function has a 24-byte frame and is called with 8 bytes of argument, which live on the caller's frame. If NOSPLIT is not specified for the TEXT, the argument size must be provided. For assembly functions with Go prototypes, go vet will check that the argument size is correct.

Note that the symbol name uses a middle dot to separate the components and is specified as an offset from the static base pseudo-register SB. This function would be called from Go source for package runtime using the simple name profileloop.

Global data symbols are defined by a sequence of initializing DATA directives followed by a GLOBL directive. Each DATA directive initializes a section of the corresponding memory. The memory not explicitly initialized is zeroed. The general form of the DATA directive is

DATA	symbol+offset(SB)/width, value

which initializes the symbol memory at the given offset and width with the given value. The DATA directives for a given symbol must be written with increasing offsets.

The GLOBL directive declares a symbol to be global. The arguments are optional flags and the size of the data being declared as a global, which will have initial value all zeros unless a DATA directive has initialized it. The GLOBL directive must follow any corresponding DATA directives.

For example,

DATA divtab<>+0x00(SB)/4, $0xf4f8fcff
DATA divtab<>+0x04(SB)/4, $0xe6eaedf0
...
DATA divtab<>+0x3c(SB)/4, $0x81828384
GLOBL divtab<>(SB), RODATA, $64

GLOBL runtime·tlsoffset(SB), NOPTR, $4

declares and initializes divtab<>, a read-only 64-byte table of 4-byte integer values, and declares runtime·tlsoffset, a 4-byte, implicitly zeroed variable that contains no pointers.

There may be one or two arguments to the directives. If there are two, the first is a bit mask of flags, which can be written as numeric expressions, added or or-ed together, or can be set symbolically for easier absorption by a human. Their values, defined in the standard #include file textflag.h, are:

Runtime Coordination

For garbage collection to run correctly, the runtime must know the location of pointers in all global data and in most stack frames. The Go compiler emits this information when compiling Go source files, but assembly programs must define it explicitly.

A data symbol marked with the NOPTR flag (see above) is treated as containing no pointers to runtime-allocated data. A data symbol with the RODATA flag is allocated in read-only memory and is therefore treated as implicitly marked NOPTR. A data symbol with a total size smaller than a pointer is also treated as implicitly marked NOPTR. It is not possible to define a symbol containing pointers in an assembly source file; such a symbol must be defined in a Go source file instead. Assembly source can still refer to the symbol by name even without DATA and GLOBL directives. A good general rule of thumb is to define all non-RODATA symbols in Go instead of in assembly.

Each function also needs annotations giving the location of live pointers in its arguments, results, and local stack frame. For an assembly function with no pointer results and either no local stack frame or no function calls, the only requirement is to define a Go prototype for the function in a Go source file in the same package. The name of the assembly function must not contain the package name component (for example, function Syscall in package syscall should use the name ·Syscall instead of the equivalent name syscall·Syscall in its TEXT directive). For more complex situations, explicit annotation is needed. These annotations use pseudo-instructions defined in the standard #include file funcdata.h.

If a function has no arguments and no results, the pointer information can be omitted. This is indicated by an argument size annotation of $n-0 on the TEXT instruction. Otherwise, pointer information must be provided by a Go prototype for the function in a Go source file, even for assembly functions not called directly from Go. (The prototype will also let go vet check the argument references.) At the start of the function, the arguments are assumed to be initialized but the results are assumed uninitialized. If the results will hold live pointers during a call instruction, the function should start by zeroing the results and then executing the pseudo-instruction GO_RESULTS_INITIALIZED. This instruction records that the results are now initialized and should be scanned during stack movement and garbage collection. It is typically easier to arrange that assembly functions do not return pointers or do not contain call instructions; no assembly functions in the standard library use GO_RESULTS_INITIALIZED.

If a function has no local stack frame, the pointer information can be omitted. This is indicated by a local frame size annotation of $0-n on the TEXT instruction. The pointer information can also be omitted if the function contains no call instructions. Otherwise, the local stack frame must not contain pointers, and the assembly must confirm this fact by executing the pseudo-instruction NO_LOCAL_POINTERS. Because stack resizing is implemented by moving the stack, the stack pointer may change during any function call: even pointers to stack data must not be kept in local variables.

Architecture-specific details

It is impractical to list all the instructions and other details for each machine. To see what instructions are defined for a given machine, say ARM, look in the source for the obj support library for that architecture, located in the directory src/cmd/internal/obj/arm. In that directory is a file a.out.go; it contains a long list of constants starting with A, like this:

const (
	AAND = obj.ABaseARM + obj.A_ARCHSPECIFIC + iota
	AEOR
	ASUB
	ARSB
	AADD
	...

This is the list of instructions and their spellings as known to the assembler and linker for that architecture. Each instruction begins with an initial capital A in this list, so AAND represents the bitwise and instruction, AND (without the leading A), and is written in assembly source as AND. The enumeration is mostly in alphabetical order. (The architecture-independent AXXX, defined in the cmd/internal/obj package, represents an invalid instruction). The sequence of the A names has nothing to do with the actual encoding of the machine instructions. The cmd/internal/obj package takes care of that detail.

The instructions for both the 386 and AMD64 architectures are listed in cmd/internal/obj/x86/a.out.go.

The architectures share syntax for common addressing modes such as (R1) (register indirect), 4(R1) (register indirect with offset), and $foo(SB) (absolute address). The assembler also supports some (not necessarily all) addressing modes specific to each architecture. The sections below list these.

One detail evident in the examples from the previous sections is that data in the instructions flows from left to right: MOVQ $0, CX clears CX. This rule applies even on architectures where the conventional notation uses the opposite direction.

Here follow some descriptions of key Go-specific details for the supported architectures.

32-bit Intel 386

The runtime pointer to the g structure is maintained through the value of an otherwise unused (as far as Go is concerned) register in the MMU. A OS-dependent macro get_tls is defined for the assembler if the source includes a special header, go_asm.h:

#include "go_asm.h"

Within the runtime, the get_tls macro loads its argument register with a pointer to the g pointer, and the g struct contains the m pointer. The sequence to load g and m using CX looks like this:

get_tls(CX)
MOVL	g(CX), AX     // Move g into AX.
MOVL	g_m(AX), BX   // Move g.m into BX.

Addressing modes:

64-bit Intel 386 (a.k.a. amd64)

The two architectures behave largely the same at the assembler level. Assembly code to access the m and g pointers on the 64-bit version is the same as on the 32-bit 386, except it uses MOVQ rather than MOVL:

get_tls(CX)
MOVQ	g(CX), AX     // Move g into AX.
MOVQ	g_m(AX), BX   // Move g.m into BX.

ARM

The registers R10 and R11 are reserved by the compiler and linker.

R10 points to the g (goroutine) structure. Within assembler source code, this pointer must be referred to as g; the name R10 is not recognized.

To make it easier for people and compilers to write assembly, the ARM linker allows general addressing forms and pseudo-operations like DIV or MOD that may not be expressible using a single hardware instruction. It implements these forms as multiple instructions, often using the R11 register to hold temporary values. Hand-written assembly can use R11, but doing so requires being sure that the linker is not also using it to implement any of the other instructions in the function.

When defining a TEXT, specifying frame size $-4 tells the linker that this is a leaf function that does not need to save LR on entry.

The name SP always refers to the virtual stack pointer described earlier. For the hardware register, use R13.

Condition code syntax is to append a period and the one- or two-letter code to the instruction, as in MOVW.EQ. Multiple codes may be appended: MOVM.IA.W. The order of the code modifiers is irrelevant.

Addressing modes:

ARM64

The ARM64 port is in an experimental state.

Instruction modifiers are appended to the instruction following a period. The only modifiers are P (postincrement) and W (preincrement): MOVW.P, MOVW.W

Addressing modes:

Power64, a.k.a. ppc64

The Power 64 port is in an experimental state.

Addressing modes:

Unsupported opcodes

The assemblers are designed to support the compiler so not all hardware instructions are defined for all architectures: if the compiler doesn't generate it, it might not be there. If you need to use a missing instruction, there are two ways to proceed. One is to update the assembler to support that instruction, which is straightforward but only worthwhile if it's likely the instruction will be used again. Instead, for simple one-off cases, it's possible to use the BYTE and WORD directives to lay down explicit data into the instruction stream within a TEXT. Here's how the 386 runtime defines the 64-bit atomic load function.

// uint64 atomicload64(uint64 volatile* addr);
// so actually
// void atomicload64(uint64 *res, uint64 volatile *addr);
TEXT runtime·atomicload64(SB), NOSPLIT, $0-12
	MOVL	ptr+0(FP), AX
	TESTL	$7, AX
	JZ	2(PC)
	MOVL	0, AX // crash with nil ptr deref
	LEAL	ret_lo+4(FP), BX
	// MOVQ (%EAX), %MM0
	BYTE $0x0f; BYTE $0x6f; BYTE $0x00
	// MOVQ %MM0, 0(%EBX)
	BYTE $0x0f; BYTE $0x7f; BYTE $0x03
	// EMMS
	BYTE $0x0F; BYTE $0x77
	RET