asm.html

<!--{
	"Title": "A Quick Guide to Go's Assembler",
	"Path":  "/doc/asm"
}-->

<h2 id="introduction">A Quick Guide to Go's Assembler</h2>

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

<p>
The assembler is based on the input style of the Plan 9 assemblers, which is documented in detail
<a href="https://9p.io/sys/doc/asm.html">elsewhere</a>.
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.
</p>

<p>
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
<a href="https://9p.io/sys/doc/compiler.html">this description</a>)
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 <code>MOV</code>
what the toolchain 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.
</p>

<p>
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
<a href="/pkg/runtime/"><code>runtime</code></a> and
<a href="/pkg/math/big/"><code>math/big</code></a>.
You can also examine what the compiler emits as assembly code
(the actual output may differ from what you see here):
</p>

<pre>
$ 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
"".main STEXT size=74 args=0x0 locals=0x10
	0x0000 00000 (x.go:3)	TEXT	"".main(SB), $16-0
	0x0000 00000 (x.go:3)	MOVQ	(TLS), CX
	0x0009 00009 (x.go:3)	CMPQ	SP, 16(CX)
	0x000d 00013 (x.go:3)	JLS	67
	0x000f 00015 (x.go:3)	SUBQ	$16, SP
	0x0013 00019 (x.go:3)	MOVQ	BP, 8(SP)
	0x0018 00024 (x.go:3)	LEAQ	8(SP), BP
	0x001d 00029 (x.go:3)	FUNCDATA	$0, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
	0x001d 00029 (x.go:3)	FUNCDATA	$1, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
	0x001d 00029 (x.go:3)	FUNCDATA	$2, gclocals·33cdeccccebe80329f1fdbee7f5874cb(SB)
	0x001d 00029 (x.go:4)	PCDATA	$0, $0
	0x001d 00029 (x.go:4)	PCDATA	$1, $0
	0x001d 00029 (x.go:4)	CALL	runtime.printlock(SB)
	0x0022 00034 (x.go:4)	MOVQ	$3, (SP)
	0x002a 00042 (x.go:4)	CALL	runtime.printint(SB)
	0x002f 00047 (x.go:4)	CALL	runtime.printnl(SB)
	0x0034 00052 (x.go:4)	CALL	runtime.printunlock(SB)
	0x0039 00057 (x.go:5)	MOVQ	8(SP), BP
	0x003e 00062 (x.go:5)	ADDQ	$16, SP
	0x0042 00066 (x.go:5)	RET
	0x0043 00067 (x.go:5)	NOP
	0x0043 00067 (x.go:3)	PCDATA	$1, $-1
	0x0043 00067 (x.go:3)	PCDATA	$0, $-1
	0x0043 00067 (x.go:3)	CALL	runtime.morestack_noctxt(SB)
	0x0048 00072 (x.go:3)	JMP	0
...
</pre>

<p>
The <code>FUNCDATA</code> and <code>PCDATA</code> directives contain information
for use by the garbage collector; they are introduced by the compiler.
</p>

<p>
To see what gets put in the binary after linking, use <code>go tool objdump</code>:
</p>

<pre>
$ go build -o x.exe x.go
$ go tool objdump -s main.main x.exe
TEXT main.main(SB) /tmp/x.go
  x.go:3		0x10501c0		65488b0c2530000000	MOVQ GS:0x30, CX
  x.go:3		0x10501c9		483b6110		CMPQ 0x10(CX), SP
  x.go:3		0x10501cd		7634			JBE 0x1050203
  x.go:3		0x10501cf		4883ec10		SUBQ $0x10, SP
  x.go:3		0x10501d3		48896c2408		MOVQ BP, 0x8(SP)
  x.go:3		0x10501d8		488d6c2408		LEAQ 0x8(SP), BP
  x.go:4		0x10501dd		e86e45fdff		CALL runtime.printlock(SB)
  x.go:4		0x10501e2		48c7042403000000	MOVQ $0x3, 0(SP)
  x.go:4		0x10501ea		e8e14cfdff		CALL runtime.printint(SB)
  x.go:4		0x10501ef		e8ec47fdff		CALL runtime.printnl(SB)
  x.go:4		0x10501f4		e8d745fdff		CALL runtime.printunlock(SB)
  x.go:5		0x10501f9		488b6c2408		MOVQ 0x8(SP), BP
  x.go:5		0x10501fe		4883c410		ADDQ $0x10, SP
  x.go:5		0x1050202		c3			RET
  x.go:3		0x1050203		e83882ffff		CALL runtime.morestack_noctxt(SB)
  x.go:3		0x1050208		ebb6			JMP main.main(SB)
</pre>

<h3 id="constants">Constants</h3>

<p>
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 <code>3&amp;1&lt;&lt;2</code> is 4, not 0—it parses as <code>(3&amp;1)&lt;&lt;2</code>
not <code>3&amp;(1&lt;&lt;2)</code>.
Also, constants are always evaluated as 64-bit unsigned integers.
Thus <code>-2</code> 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.
</p>

<h3 id="symbols">Symbols</h3>

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

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

<ul>

<li>
<code>FP</code>: Frame pointer: arguments and locals.
</li>

<li>
<code>PC</code>: Program counter:
jumps and branches.
</li>

<li>
<code>SB</code>: Static base pointer: global symbols.
</li>

<li>
<code>SP</code>: Stack pointer: the highest address within the local stack frame.
</li>

</ul>

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

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

<p>
The <code>FP</code> 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 <code>0(FP)</code> is the first argument to the function,
<code>8(FP)</code> 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 <code>first_arg+0(FP)</code> and <code>second_arg+8(FP)</code>.
(The meaning of the offset—offset from the frame pointer—distinct
from its use with <code>SB</code>, where it is an offset from the symbol.)
The assembler enforces this convention, rejecting plain <code>0(FP)</code> and <code>8(FP)</code>.
The actual name is semantically irrelevant but should be used to document
the argument's name.
It is worth stressing that <code>FP</code> is always a
pseudo-register, not a hardware
register, even on architectures with a hardware frame pointer.
</p>

<p>
For assembly functions with Go prototypes, <code>go</code> <code>vet</code> 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 <code>_lo</code> or <code>_hi</code> suffix to the name, as in <code>arg_lo+0(FP)</code> or <code>arg_hi+4(FP)</code>.
If a Go prototype does not name its result, the expected assembly name is <code>ret</code>.
</p>

<p>
The <code>SP</code> 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 highest address within the local stack frame, so references should use negative offsets
in the range [−framesize, 0):
<code>x-8(SP)</code>, <code>y-4(SP)</code>, and so on.
</p>

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

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

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

<pre>
label:
	MOVW $0, R1
	JMP label
</pre>

<p>
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 <code>name(SB)</code>, but not offsets from symbols,
such as <code>name+4(SB)</code>.
</p>

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

<p>
In Go object files and binaries, the full name of a symbol is the
package path followed by a period and the symbol name:
<code>fmt.Printf</code> or <code>math/rand.Int</code>.
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
<code>fmt·Printf</code> and <code>math∕rand·Int</code>.
The assembly listings generated by the compilers when using the <code>-S</code> flag
show the period and slash directly instead of the Unicode replacements
required by the assemblers.
</p>

<p>
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 <code>·Int</code>.
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.
</p>

<h3 id="directives">Directives</h3>

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

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

<p>
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 <code>$24-8</code> 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 <code>NOSPLIT</code> is not specified for the <code>TEXT</code>,
the argument size must be provided.
For assembly functions with Go prototypes, <code>go</code> <code>vet</code> will check that the
argument size is correct.
</p>

<p>
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 <code>SB</code>.
This function would be called from Go source for package <code>runtime</code> using the
simple name <code>profileloop</code>.
</p>

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

<pre>
DATA	symbol+offset(SB)/width, value
</pre>

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

<p>
The <code>GLOBL</code> 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 <code>DATA</code> directive
has initialized it.
The <code>GLOBL</code> directive must follow any corresponding <code>DATA</code> directives.
</p>

<p>
For example,
</p>

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

GLOBL runtime·tlsoffset(SB), NOPTR, $4
</pre>

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

<p>
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 <code>#include</code>  file <code>textflag.h</code>, are:
</p>

<ul>
<li>
<code>NOPROF</code> = 1
<br>
(For <code>TEXT</code> items.)
Don't profile the marked function.  This flag is deprecated.
</li>
<li>
<code>DUPOK</code> = 2
<br>
It is legal to have multiple instances of this symbol in a single binary.
The linker will choose one of the duplicates to use.
</li>
<li>
<code>NOSPLIT</code> = 4
<br>
(For <code>TEXT</code> items.)
Don't insert the preamble to check if the stack must be split.
The frame for the routine, plus anything it calls, must fit in the
spare space remaining in the current stack segment.
Used to protect routines such as the stack splitting code itself.
</li>
<li>
<code>RODATA</code> = 8
<br>
(For <code>DATA</code> and <code>GLOBL</code> items.)
Put this data in a read-only section.
</li>
<li>
<code>NOPTR</code> = 16
<br>
(For <code>DATA</code> and <code>GLOBL</code> items.)
This data contains no pointers and therefore does not need to be
scanned by the garbage collector.
</li>
<li>
<code>WRAPPER</code> = 32
<br>
(For <code>TEXT</code> items.)
This is a wrapper function and should not count as disabling <code>recover</code>.
</li>
<li>
<code>NEEDCTXT</code> = 64
<br>
(For <code>TEXT</code> items.)
This function is a closure so it uses its incoming context register.
</li>
<li>
<code>LOCAL</code> = 128
<br>
This symbol is local to the dynamic shared object.
</li>
<li>
<code>TLSBSS</code> = 256
<br>
(For <code>DATA</code> and <code>GLOBL</code> items.)
Put this data in thread local storage.
</li>
<li>
<code>NOFRAME</code> = 512
<br>
(For <code>TEXT</code> items.)
Do not insert instructions to allocate a stack frame and save/restore the return
address, even if this is not a leaf function.
Only valid on functions that declare a frame size of 0.
</li>
<li>
<code>TOPFRAME</code> = 2048
<br>
(For <code>TEXT</code> items.)
Function is the outermost frame of the call stack. Traceback should stop at this function.
</li>
</ul>

<h3 id="special-instructions">Special instructions</h3>

<p>
The <code>PCALIGN</code> pseudo-instruction is used to indicate that the next instruction should be aligned
to a specified boundary by padding with no-op instructions.
</p>

<p>
It is currently supported on arm64, amd64, ppc64, loong64 and riscv64.

For example, the start of the <code>MOVD</code> instruction below is aligned to 32 bytes:
<pre>
PCALIGN $32
MOVD $2, R0
</pre>
</p>

<h3 id="data-offsets">Interacting with Go types and constants</h3>

<p>
If a package has any .s files, then <code>go build</code> will direct
the compiler to emit a special header called <code>go_asm.h</code>,
which the .s files can then <code>#include</code>.
The file contains symbolic <code>#define</code> constants for the
offsets of Go struct fields, the sizes of Go struct types, and most
Go <code>const</code> declarations defined in the current package.
Go assembly should avoid making assumptions about the layout of Go
types and instead use these constants.
This improves the readability of assembly code, and keeps it robust to
changes in data layout either in the Go type definitions or in the
layout rules used by the Go compiler.
</p>

<p>
Constants are of the form <code>const_<i>name</i></code>.
For example, given the Go declaration <code>const bufSize =
1024</code>, assembly code can refer to the value of this constant
as <code>const_bufSize</code>.
</p>

<p>
Field offsets are of the form <code><i>type</i>_<i>field</i></code>.
Struct sizes are of the form <code><i>type</i>__size</code>.
For example, consider the following Go definition:
</p>

<pre>
type reader struct {
	buf [bufSize]byte
	r   int
}
</pre>

<p>
Assembly can refer to the size of this struct
as <code>reader__size</code> and the offsets of the two fields
as <code>reader_buf</code> and <code>reader_r</code>.
Hence, if register <code>R1</code> contains a pointer to
a <code>reader</code>, assembly can reference the <code>r</code> field
as <code>reader_r(R1)</code>.
</p>

<p>
If any of these <code>#define</code> names are ambiguous (for example,
a struct with a <code>_size</code> field), <code>#include
"go_asm.h"</code> will fail with a "redefinition of macro" error.
</p>

<h3 id="runtime">Runtime Coordination</h3>

<p>
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.
</p>

<p>
A data symbol marked with the <code>NOPTR</code> flag (see above)
is treated as containing no pointers to runtime-allocated data.
A data symbol with the <code>RODATA</code> flag
is allocated in read-only memory and is therefore treated
as implicitly marked <code>NOPTR</code>.
A data symbol with a total size smaller than a pointer
is also treated as implicitly marked <code>NOPTR</code>.
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 <code>DATA</code> and <code>GLOBL</code> directives.
A good general rule of thumb is to define all non-<code>RODATA</code>
symbols in Go instead of in assembly.
</p>

<p>
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 <code>Syscall</code> in package <code>syscall</code> should
use the name <code>·Syscall</code> instead of the equivalent name
<code>syscall·Syscall</code> in its <code>TEXT</code> directive).
For more complex situations, explicit annotation is needed.
These annotations use pseudo-instructions defined in the standard
<code>#include</code> file <code>funcdata.h</code>.
</p>

<p>
If a function has no arguments and no results,
the pointer information can be omitted.
This is indicated by an argument size annotation of <code>$<i>n</i>-0</code>
on the <code>TEXT</code> 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 <code>go</code> <code>vet</code> 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 <code>GO_RESULTS_INITIALIZED</code>.
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
<code>GO_RESULTS_INITIALIZED</code>.
</p>

<p>
If a function has no local stack frame,
the pointer information can be omitted.
This is indicated by a local frame size annotation of <code>$0-<i>n</i></code>
on the <code>TEXT</code> 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 <code>NO_LOCAL_POINTERS</code>.
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.
</p>

<p>
Assembly functions should always be given Go prototypes,
both to provide pointer information for the arguments and results
and to let <code>go</code> <code>vet</code> check that
the offsets being used to access them are correct.
</p>

<h2 id="architectures">Architecture-specific details</h2>

<p>
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 <code>obj</code> support library for
that architecture, located in the directory <code>src/cmd/internal/obj/arm</code>.
In that directory is a file <code>a.out.go</code>; it contains
a long list of constants starting with <code>A</code>, like this:
</p>

<pre>
const (
	AAND = obj.ABaseARM + obj.A_ARCHSPECIFIC + iota
	AEOR
	ASUB
	ARSB
	AADD
	...
</pre>

<p>
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 <code>A</code> in this list, so <code>AAND</code>
represents the bitwise and instruction,
<code>AND</code> (without the leading <code>A</code>),
and is written in assembly source as <code>AND</code>.
The enumeration is mostly in alphabetical order.
(The architecture-independent <code>AXXX</code>, defined in the
<code>cmd/internal/obj</code> package,
represents an invalid instruction).
The sequence of the <code>A</code> names has nothing to do with the actual
encoding of the machine instructions.
The <code>cmd/internal/obj</code> package takes care of that detail.
</p>

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

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

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

<p>
Here follow some descriptions of key Go-specific details for the supported architectures.
</p>

<h3 id="x86">32-bit Intel 386</h3>

<p>
The runtime pointer to the <code>g</code> structure is maintained
through the value of an otherwise unused (as far as Go is concerned) register in the MMU.
In the runtime package, assembly code can include <code>go_tls.h</code>, which defines
an OS- and architecture-dependent macro <code>get_tls</code> for accessing this register.
The <code>get_tls</code> macro takes one argument, which is the register to load the
<code>g</code> pointer into.
</p>

<p>
For example, the sequence to load <code>g</code> and <code>m</code>
using <code>CX</code> looks like this:
</p>

<pre>
#include "go_tls.h"
#include "go_asm.h"
...
get_tls(CX)
MOVL	g(CX), AX     // Move g into AX.
MOVL	g_m(AX), BX   // Move g.m into BX.
</pre>

<p>
The <code>get_tls</code> macro is also defined on <a href="#amd64">amd64</a>.
</p>

<p>
Addressing modes:
</p>

<ul>

<li>
<code>(DI)(BX*2)</code>: The location at address <code>DI</code> plus <code>BX*2</code>.
</li>

<li>
<code>64(DI)(BX*2)</code>: The location at address <code>DI</code> plus <code>BX*2</code> plus 64.
These modes accept only 1, 2, 4, and 8 as scale factors.
</li>

</ul>

<p>
When using the compiler and assembler's
<code>-dynlink</code> or <code>-shared</code> modes,
any load or store of a fixed memory location such as a global variable
must be assumed to overwrite <code>CX</code>.
Therefore, to be safe for use with these modes,
assembly sources should typically avoid CX except between memory references.
</p>

<h3 id="amd64">64-bit Intel 386 (a.k.a. amd64)</h3>

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

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

<p>
Register <code>BP</code> is callee-save.
The assembler automatically inserts <code>BP</code> save/restore when frame size is larger than zero.
Using <code>BP</code> as a general purpose register is allowed,
however it can interfere with sampling-based profiling.
</p>

<h3 id="arm">ARM</h3>

<p>
The registers <code>R10</code> and <code>R11</code>
are reserved by the compiler and linker.
</p>

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

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

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

<p>
The name <code>SP</code> always refers to the virtual stack pointer described earlier.
For the hardware register, use <code>R13</code>.
</p>

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

<p>
Addressing modes:
</p>

<ul>

<li>
<code>R0-&gt;16</code>
<br>
<code>R0&gt;&gt;16</code>
<br>
<code>R0&lt;&lt;16</code>
<br>
<code>R0@&gt;16</code>:
For <code>&lt;&lt;</code>, left shift <code>R0</code> by 16 bits.
The other codes are <code>-&gt;</code> (arithmetic right shift),
<code>&gt;&gt;</code> (logical right shift), and
<code>@&gt;</code> (rotate right).
</li>

<li>
<code>R0-&gt;R1</code>
<br>
<code>R0&gt;&gt;R1</code>
<br>
<code>R0&lt;&lt;R1</code>
<br>
<code>R0@&gt;R1</code>:
For <code>&lt;&lt;</code>, left shift <code>R0</code> by the count in <code>R1</code>.
The other codes are <code>-&gt;</code> (arithmetic right shift),
<code>&gt;&gt;</code> (logical right shift), and
<code>@&gt;</code> (rotate right).

</li>

<li>
<code>[R0,g,R12-R15]</code>: For multi-register instructions, the set comprising
<code>R0</code>, <code>g</code>, and <code>R12</code> through <code>R15</code> inclusive.
</li>

<li>
<code>(R5, R6)</code>: Destination register pair.
</li>

</ul>

<h3 id="arm64">ARM64</h3>

<p>
<code>R18</code> is the "platform register", reserved on the Apple platform.
To prevent accidental misuse, the register is named <code>R18_PLATFORM</code>.
<code>R27</code> and <code>R28</code> are reserved by the compiler and linker.
<code>R29</code> is the frame pointer.
<code>R30</code> is the link register.
</p>

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

<p>
Addressing modes:
</p>

<ul>

<li>
<code>R0-&gt;16</code>
<br>
<code>R0&gt;&gt;16</code>
<br>
<code>R0&lt;&lt;16</code>
<br>
<code>R0@&gt;16</code>:
These are the same as on the 32-bit ARM.
</li>

<li>
<code>$(8&lt;&lt;12)</code>:
Left shift the immediate value <code>8</code> by <code>12</code> bits.
</li>

<li>
<code>8(R0)</code>:
Add the value of <code>R0</code> and <code>8</code>.
</li>

<li>
<code>(R2)(R0)</code>:
The location at <code>R0</code> plus <code>R2</code>.
</li>

<li>
<code>R0.UXTB</code>
<br>
<code>R0.UXTB&lt;&lt;imm</code>:
<code>UXTB</code>: extract an 8-bit value from the low-order bits of <code>R0</code> and zero-extend it to the size of <code>R0</code>.
<code>R0.UXTB&lt;&lt;imm</code>: left shift the result of <code>R0.UXTB</code> by <code>imm</code> bits.
The <code>imm</code> value can be 0, 1, 2, 3, or 4.
The other extensions include <code>UXTH</code> (16-bit), <code>UXTW</code> (32-bit), and <code>UXTX</code> (64-bit).
</li>

<li>
<code>R0.SXTB</code>
<br>
<code>R0.SXTB&lt;&lt;imm</code>:
<code>SXTB</code>: extract an 8-bit value from the low-order bits of <code>R0</code> and sign-extend it to the size of <code>R0</code>.
<code>R0.SXTB&lt;&lt;imm</code>: left shift the result of <code>R0.SXTB</code> by <code>imm</code> bits.
The <code>imm</code> value can be 0, 1, 2, 3, or 4.
The other extensions include <code>SXTH</code> (16-bit), <code>SXTW</code> (32-bit), and <code>SXTX</code> (64-bit).
</li>

<li>
<code>(R5, R6)</code>: Register pair for <code>LDAXP</code>/<code>LDP</code>/<code>LDXP</code>/<code>STLXP</code>/<code>STP</code>/<code>STP</code>.
</li>

</ul>

<p>
Reference: <a href="/pkg/cmd/internal/obj/arm64">Go ARM64 Assembly Instructions Reference Manual</a>
</p>

<h3 id="ppc64">PPC64</h3>

<p>
This assembler is used by GOARCH values ppc64 and ppc64le.
</p>

<p>
Reference: <a href="/pkg/cmd/internal/obj/ppc64">Go PPC64 Assembly Instructions Reference Manual</a>
</p>

<h3 id="s390x">IBM z/Architecture, a.k.a. s390x</h3>

<p>
The registers <code>R10</code> and <code>R11</code> are reserved.
The assembler uses them to hold temporary values when assembling some instructions.
</p>

<p>
<code>R13</code> points to the <code>g</code> (goroutine) structure.
This register must be referred to as <code>g</code>; the name <code>R13</code> is not recognized.
</p>

<p>
<code>R15</code> points to the stack frame and should typically only be accessed using the
virtual registers <code>SP</code> and <code>FP</code>.
</p>

<p>
Load- and store-multiple instructions operate on a range of registers.
The range of registers is specified by a start register and an end register.
For example, <code>LMG</code> <code>(R9),</code> <code>R5,</code> <code>R7</code> would load
<code>R5</code>, <code>R6</code> and <code>R7</code> with the 64-bit values at
<code>0(R9)</code>, <code>8(R9)</code> and <code>16(R9)</code> respectively.
</p>

<p>
Storage-and-storage instructions such as <code>MVC</code> and <code>XC</code> are written
with the length as the first argument.
For example, <code>XC</code> <code>$8,</code> <code>(R9),</code> <code>(R9)</code> would clear
eight bytes at the address specified in <code>R9</code>.
</p>

<p>
If a vector instruction takes a length or an index as an argument then it will be the
first argument.
For example, <code>VLEIF</code> <code>$1,</code> <code>$16,</code> <code>V2</code> will load
the value sixteen into index one of <code>V2</code>.
Care should be taken when using vector instructions to ensure that they are available at
runtime.
To use vector instructions a machine must have both the vector facility (bit 129 in the
facility list) and kernel support.
Without kernel support a vector instruction will have no effect (it will be equivalent
to a <code>NOP</code> instruction).
</p>

<p>
Addressing modes:
</p>

<ul>

<li>
<code>(R5)(R6*1)</code>: The location at <code>R5</code> plus <code>R6</code>.
It is a scaled mode as on the x86, but the only scale allowed is <code>1</code>.
</li>

</ul>

<h3 id="mips">MIPS, MIPS64</h3>

<p>
General purpose registers are named <code>R0</code> through <code>R31</code>,
floating point registers are <code>F0</code> through <code>F31</code>.
</p>

<p>
<code>R30</code> is reserved to point to <code>g</code>.
<code>R23</code> is used as a temporary register.
</p>

<p>
In a <code>TEXT</code> directive, the frame size <code>$-4</code> for MIPS or
<code>$-8</code> for MIPS64 instructs the linker not to save <code>LR</code>.
</p>

<p>
<code>SP</code> refers to the virtual stack pointer.
For the hardware register, use <code>R29</code>.
</p>

<p>
Addressing modes:
</p>

<ul>

<li>
<code>16(R1)</code>: The location at <code>R1</code> plus 16.
</li>

<li>
<code>(R1)</code>: Alias for <code>0(R1)</code>.
</li>

</ul>

<p>
The value of <code>GOMIPS</code> environment variable (<code>hardfloat</code> or
<code>softfloat</code>) is made available to assembly code by predefining either
<code>GOMIPS_hardfloat</code> or <code>GOMIPS_softfloat</code>.
</p>

<p>
The value of <code>GOMIPS64</code> environment variable (<code>hardfloat</code> or
<code>softfloat</code>) is made available to assembly code by predefining either
<code>GOMIPS64_hardfloat</code> or <code>GOMIPS64_softfloat</code>.
</p>

<h3 id="unsupported_opcodes">Unsupported opcodes</h3>

<p>
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 <code>BYTE</code>
and <code>WORD</code> directives
to lay down explicit data into the instruction stream within a <code>TEXT</code>.
Here's how the 386 runtime defines the 64-bit atomic load function.
</p>

<pre>
// 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
</pre>