What was the original reason for the design of AT&T assembly syntax?

UNIX was for a long time developed on the PDP-11, a 16 bit computer from DEC, which had a fairly simple instruction set. Nearly every instruction has two operands, each of which can have one of the following eight addressing modes, here shown in the MACRO 16 assembly language:

0n  Rn        register
1n  (Rn)      deferred
2n  (Rn)+     autoincrement
3n  @(Rn)+    autoincrement deferred
4n  -(Rn)     autodecrement
5n  @-(Rn)    autodecrement deferred
6n  X(Rn)     index
7n  @X(Rn)    index deferred

Immediates and direct addresses can be encoded by cleverly re-using some addressing modes on R7, the program counter:

27  #imm      immediate
37  @#imm     absolute
67  addr      relative
77  @addr     relative deferred

As the UNIX tty driver used @ and # as control characters, $ was substituted for # and * for @.

The first operand in a PDP11 instruction word refers to the source operand while the second operand refers to the destination. This is reflected in the assembly language's operand order which is source, then destination. For example, the opcode

011273

refers to the instruction

mov (R2),R3

which moves the word pointed to by R2 to R3.

This syntax was adapted to the 8086 CPU and its addressing modes:

mr0 X(bx,si)  bx + si indexed
mr1 X(bx,di)  bx + di indexed
mr2 X(bp,si)  bp + si indexed
mr3 X(bp,di)  bp + di indexed
mr4 X(si)     si indexed
mr5 X(di)     di indexed
mr6 X(bp)     bp indexed
mr7 X(bx)     bx indexed
3rR R         register
0r6 addr      direct

Where m is 0 if there is no index, m is 1 if there is a one-byte index, m is 2 if there is a two-byte index and m is 3 if instead of a memory operand, a register is used. If two operands exist, the other operand is always a register and encoded in the r digit. Otherwise, r encodes another three bits of the opcode.

Immediates aren't possible in this addressing scheme, all instructions that take immediates encode that fact in their opcode. Immediates are spelled $imm just like in the PDP-11 syntax.

While Intel always used a dst, src operand ordering for its assembler, there was no particularly compelling reason to adapt this convention and the UNIX assembler was written to use the src, dst operand ordering known from the PDP11.

They made some inconsistencies with this ordering in their implementation of the 8087 floating point instructions, possibly because Intel gave the two possible directions of non-commutative floating point instructions different mnemonics which do not match the operand ordering used by AT&T's syntax.

The PDP11 instructions jmp (jump) and jsr (jump to subroutine) jump to the address of their operand. Thus, jmp foo would jump to foo and jmp *foo would jump to the address stored in the variable foo, similar to how lea works in the 8086.

The syntax for the x86's jmp and call instructions was designed as if these instructions worked like on the PDP11, which is why jmp foo jumps to foo and jmp *foo jumps to the value at address foo, even though the 8086 doesn't actually have deferred addressing. This has the advantage and convenience of syntactically distinguishing direct jumps from indirect jumps without requiring an $ prefix for every direct jump target but doesn't make a lot of sense logically.

The syntax was expanded to specify segment prefixes using a colon:

seg:addr

When the 80386 was introduced, this scheme was adapted to its new SIB addressing modes using a four-part generic addressing mode:

disp(base,index,scale)

where disp is a displacement, base is a base register, index an index register and scale is 1, 2, 4, or 8 to scale the index register by one of these amounts. This is equal to Intel syntax:

[disp+base+index*scale]

Another remarkable feature of the PDP-11 is that most instructions are available in a byte and a word variant. Which one you use is indicated by a b or w suffix to the opcode, which directly toggles the first bit of the opcode:

 010001   movw r0,r1
 110001   movb r0,r1

this also was adapted for AT&T syntax as most 8086 instructions are indeed also available in a byte mode and a word mode. Later the 80386 and AMD K6 introduced 32 bit instructions (suffixed l for long) and 64 bit instructions (suffixed q for quad).

Last but not least, originally the convention was to prefix C language symbols with an underscore (as is still done on Windows) so you can distinguish a C function named ax from the register ax. When Unix System Laboratories developed the ELF binary format, they decided to get rid of this decoration. As there is no way to distinguish a direct address from a register otherwise, a % prefix was added to every register:

mov direct,%eax # move memory at direct to %eax

And that's how we got today's AT&T syntax.