This is designed to be a short summary. For more, see also Bryan and O’Halleron (from CMU)’s 46-page chapter on writing assembly using AT&T syntax; Bloomfield (from UVA)’s 12-page chapter on assembly using Intel syntax and 6-page chapter on calling conventions; Intel’s 2,226-page decription of all instruction; and AMD’s multi-volume manual.

1 Addressing modes

Operands of most operations may be either a register, an immediate value, or the contents of memory. A memory address in general is made of an immediate, two registers, and a scale on one of the registers: imm + rA + rB*s where s is one of the four specific values 1, 2, 4, or 8.

2 Two syntaxes

For mostly historical reasons, x86-64 has two different syntaxes.

Feature Intel syntax AT&T syntax Mnemonic
Register rsp %rsp AT&T has a & in its name, and wants similar symbols elsewhere
Immediate 23 $23 AT&T has a & in its name, and wants similar symbols elsewhere
Reg+Imm Addr [rsp+23] 23(%rsp) Intel uses infix operators
R+R*4+Imm Addr [rsp+r8*4+23] 23(%rsp,%r8,4) Intel uses infix operators
a += b add rax,rbx addq %rbx, %rax Intel starts with what that answer goes into; AT&T starts with the argument.

In general, AT&T syntax is more explicit: there are prefixes for types, operations have width suffixes, etc. Intel syntax, on the other hand, is more loose, and has to add things like QWORD PTR if the instructions operands do not make the width of a command obvious.

Intel syntax AT&T syntax
mov QWORD PTR [rdx+0x227],rax mov %rax,0x227(%rdx)

The width specifiers are

bits historical name Intel name AT&T Suffix register names
8 byte BYTE b ah, al, r9b, …
16 word, as this was the native size of the 8086 processor WORD w ax, r9w, …
32 double word DWORD l eax, r9d, …
64 quad word QWORD q rax, r9, …

The most popular *nix toolchains default to AT&T syntax. The most popular Windows toolchains default to Intel syntax.

3 Registers

The general-purpose program registers in x86-64 have somewhat ideosyncratic names:

8-bit 16-bit 32-bit 64-bit calling callee-save notes
al, ah ax eax rax return special meaning for multiply and divide instructions
cl, ch cx ecx rcx arg 4
dl, dh dx edx rdx arg 3 special meaning for multiply and divide instructions
bl, bh bx ebx rbx yes
spl sp esp rsp yes stack pointer
bpl bp ebp rbp yes
sil si esi rsi arg 2
dil di edi rdi arg 1
r8b r8w r8d r8 arg 5
r9b r9w r9d r9 arg 6
r10b r10w r10d r10
r11b r11w r11d r11
r12b r12w r12d r12 yes
r13b r13w r13d r13 yes
r14b r14w r14d r14 yes
r15b r15w r15d r15 yes

The registers overlap in the low-order bits. Thus, if r15 is 0x0123456789abcedf then r15d is 0x89abcdef, r15w is 0xcdef, and r15b is 0xef. Some registers also have names for both the lowest-byte _l and next-higher-byte _h. Thus, if rax is 0x0123456789abcedf then eax is 0x89abcdef, ax is 0xcdef, al is 0xef, and ah is 0xcd.

In part because x86-64 has preserved backwards compatibility with many previous architectures all the way back to the 16-bit integer-only 8086, many newer operations have been placed in their own register bank with their own operations. As one of the larger examples, floating-point operations are not handled from the main registers. However, we’ll restrict ourselves to the registers above and the operations that work on them.

4 The most important instructions

x86-64 has thousands of instructions, but many of them are used only in fairly specialized cases. The following instructions are the most important to understand x86-64 code.

4.1 Move

The various mov instructions implement, in effect, the assignment operator =. Moves can be done between registers, memory, and immediates, with some limitations; as a rule of thumb, either the source or destination must be a register.

When moving from a smaller source to a larger destination, mov has two variants: movzx (zero-extend) fills in the extra high-order bits with zeros, and movsx (sign-extend) fills in the extra high-order bits with copies of the high-order bit of the source.

There is a special “swap” instruction xchg, though it is usually used to implement a no-op not a move.

There are special moves for moving between register banks (as, e.g., moves to and from the XMM registers, etc).

There are also conditional moves which only moves if the condition codes indicate the last compared value had a particular relationship to 0, although those are fairly uncommon in compiled code.

4.2 Jumps

Jumps move the pc to a new location. jmp does this unconditionally, and various other instructions do so conditionally. Conditions in x86-64 are based on the “condition codes”, a set of single-bit flags that store enough information to compare a value to 0. Condition codes are set by most ALU operations, as well as by the special cmp and test operations.

Because comparisons are done differently for signed and unsigned values, there are multiple versions of comparions:

je, jne
Jump if the compared values were equal (je) or not equal (jne) or the result of the last operations was equal to 0 (je) or not (jne)
ja, jae, jb, jbe
Jump if the first compared values was above/below the other, using unsigned comparisons.
jg, jge, jl, jle
Jump if the first compared values was greater/lesser that the other, using signed comparisons.

There are also conditional jumps that check just single bits of the conditions codes, one of moderate commonness being js which checks the sign bit.

4.3 Load Effective Address

One specific instruction, lea, is widely used. It looks like a memory-to-register move, but instead of loading the contents of memory at an address it loads the address itself.

Because addresses are computed by adding two registers and an immediate, with one address being multiplied by a small power-of-two constant, lea is commonly used to perform basic arithmetic. For example, code like a = 5*b + 20 can be written in AT&T syntax x86-64 as lea 20(rbx,rbx,4),%rax.

4.4 ALU operations

Most ALU operations are implemented in x86-64 as assignment operators in code.

Instruction Is like
add +=
sub -=
and &=
or |=
xor ^=
shl <<=
shr >>=, zero-extending
sar >>=, sign-extending

These instructions also set the condition codes. Additionally, cmp sets the condition codes like sub, and test like and, but both without storing the result in a register.

Multiplication and division are implemented differently. The result of addition can be one bit larger than its largest operand, which makes += a relatively safe way to handle it; but multiplication can result in twice as many bits as the largest operand, and the circuitry that does division also does modulus as the same time, meaning both effectively have multiple registers of return value(s). There are several variants, but to get a feel two are described below:

imul X
multiplies rax by register X, storing the 128-bit result with the high-order 64-bits in rdx and the low-order in rax.
idiv X
divides a 128-bit numerator (high-order bits in rdx, low-order in rax) by register X, storing the quotient in rax and the remainder in rdx.

4.5 Push and pop

The behavior of push X can be described as

rsp -= 8
memory[rsp] = X

The behavior of pop X can be described as

X = memory[rsp]
rsp += 8

Note that some programs use only 32-bit and smaller values, and use a variant of push and pop that adjust esp by 4 instead of rsp by 8.

Push and pop are widely used in common function call protocols, for argument passing and register saving, as explained in Calling Conventions.

4.6 Call and return

call X means "push the address of the next instruction, then jmp X. ret means pop PC – an instruction not otherwise writeable using pop because PC is not a program register.

4.7 No operation

Compilers generate a surprising number of operations that do nothing. Called “no-ops” or “nops,” these are used to align certain instructions with multiple-of-8 addresses and other less-than-obvious optimizations.

Some no-ops may the specific no-op instruction nop, and others are encoded as meaningless instructions like xchg %eax,%eax

5 Calling Conventions

Although not intrinsically dictated by the ISA itself, it is common for ISAs to be accompanied by a recommended calling convention. This involves three primary components:

Argument passing

Invoking a function (with call) involves jumping to its code and storing where to return to. That code needs to know where to find it’s arguments.

x86-64’s most common calling convention1 puts the arguments, in order, in rdi, rsi, rdx, rcx, r8, and r9. Remaining arguments, if any, are pushed onto the stack, last to first, before the call.

Return value passing
The code that invokes a function needs to know from where to retrieve it’s return value. x86-64’s most common calling conventions put the return value into rax.
Callee- and caller-save registers

In general, both the code that invokes a function and the code of the function itself will use all the program registers. This means that the old values of these registers must be saved and restored.

x86-64 calling conventions distinguish between callee-save and caller-save registers.

A caller-save register is one that the invoking code must assume the invoked code might have changed, thus necessitating saving it before the call if it contains meaningful data to the invoking code. It is also one that the invoked code can use without first saving and later restoring.

A callee-save register is one that the invoking code can assume the invoked code will not change, and thus the invoking code does not need to save before the call. It is also one that the invoked code cannot use unless it first saves its value and restores that saved value to the register before returning.

The most common way to save a register is to push its contents onto the stack using push (as, e.g., push %rax) or a similar rsp-based mov (as, e.g., mov %rax,-32(rsp)).

x86-64’s most common calling convention2 identifies rsp, rbp, and r12 through r15 as callee-save registers and all others (rax, rbx, rcx, rdx, rsi, rdi, and r8 through r11) as caller-save.

Note that all of the above is merely convention. A program could violate all of these rules and still work fine, but it might have some difficulty interacting with other functions if it does. In that way it is similar to conventions about what side of the street to drive on or what color of traffic signal light means “stop”: the decision is fairly arbitrary, but if you make a different arbitrary decision than others do then things are not likely to go well for you.

6 The most common x86-64 instructions

For the curious, I counted how many times different instructions occurred in the 200,723,121 instructions comprising the programs in the /usr/bin directory of my installation of Manjaro Linux. The following table lists the most frequent, omitting those that use different register sets.

Frequency instruction meaning
72,239,722 mov =
14,145,074 lea

“load effective address,” usually used for addition

lea rbp,[rip+0x20aa7e] is equivalent to rbp = rip + 0x20aa7e

12,327,021 call push the PC and jump to address
9,228,101 add +=
8,346,941 cmp set flags as if performing subtraction
7,897,873 jmp unconditionally jump to new address
7,572,220 test set flags as if performing &
7,539,235 je jump if and only if flags indicate == 0
5,651,123 pop

pops a value off of the stack

First reads from address in rsp, then increases rsp by the size of the value read.

5,555,926 push

pushes a value onto the stack

First decreases rsp by the size of the value, then writes the value into memory at the address thereafter stored in rsp.

5,534,972 xor ^=
5,272,088 jne jump if and only if flags indicate != 0
5,216,558 nop do nothing
4,902,189 int3 used to contact the operating system
2,696,014 sub -=
2,366,491 ret pop the PC from the stack
1,683,246 movzx move, zero-extending (for assigning from smaller- to larger-sized register or memory region)
1,083,966 and &=
863,217 shl <<=
859,984 movsx move, sign-extending (for assigning from smaller- to larger-sized register or memory region)
685,614 jbe jump if and only if flags indicate <= 0 using unsigned comparison (b indicates “below”)
662,053 ja jump if and only if flags indicate > 0 using unsigned comparison (a indicates “above”)
643,204 or |=
596,148 shr >>=, zero-extending
493,925 xchg swap the contents of the two arguments
444,418 jle jump if and only if flags indicate <= 0 using unsigned comparison (l indicates “less”)
403,570 jb jump if and only if flags indicate < 0 using unsigned comparison (b indicates “below”)
371,140 jae jump if and only if flags indicate >= 0 using unsigned comparison (a indicates “above”)
360,654 jg jump if and only if flags indicate > 0 using unsigned comparison (g indicates “greater”)
335,154 sar >>=, sign-extending
320,519 movabs (special move for large immediates)
307,118 js jump if sign bit set
292,134 imul integer multiply of 32-bit values. ???
259,750 ud2 unreachable code

  1. Unlike other platforms, Microsoft toolchains use just 4 registers for arguments: rdx, rcx, r8, and r9

  2. Unlike other platforms, Microsoft toolchains use rbx, rbp, rdi, rsi, rsp, r12, r13, r14, and r15 as callee-save and rax, rcx, rdx, r8, r9, r10 and r11 as caller-save