MIPS Assembly/MIPS Details

Registers
MIPS has 32 general-purpose registers and another 32 floating-point registers. Registers all begin with a dollar-symbol ($). The floating point registers are named $f0, $f1, ..., $f31. The general-purpose registers have both names and numbers, and are listed below. When programming in MIPS assembly, it is usually best to use the register names.

In general, there are many registers that can be used in your programs: the ten temporary registers and the eight saved registers, and the arg $a and return-value $v registers. Temporary registers are general-purpose registers that can be used for arithmetic and other instructions freely (call clobbered), while saved registers must keep their value across function calls. (If you want to use one, you have to save it on procedure entry, and restore at procedure exit). The easiest way to handle the call-preserved $s0..7 registers is to not touch them at all.

Temporary register names all start with a $t. For instance, there are $t0, $t1 ... $t9. this means there are 10 temporary registers that can be used without worrying about saving and restoring their contents. The saved registers are named $s0 to $s7.

The zero register, is named $zero ($0), and is a static register: it always contains the value zero. This register may not be used as the target of a store operation, because its value is hardwired in, and cannot be changed by the program.

There are also several registers to which the programmer does not have direct access with most instructions. Among these are the Program Counter (PC), which stores the address of the instruction executing (read by JAL to calculate a return address, written by jumps and branches), and the "hi" and "lo" registers, which are used in multiplication and division, which have results longer than 32 bits (multiplication may result in a 64-bit product and division results in a quotient and remainder). There are special instructions to move data to and from the hi and lo registers.

Instruction Formats
There are 3 instruction formats: R Instructions, I Instructions, and J Instructions.

R Instructions
R Instructions take three arguments: two source registers (rt and rs), and a destination register (rd). R instructions are written using the following format:


 * instruction rd, rs, rt

where each one stands for as follow:

For example,

add $t0, $t1, $t2

Adds the values of $t1 and $t2 and stores the result in $t0.

When assembled into machine code, an R instruction is represented as follows:

For R-format instructions, the opcode, or "operation code" is always zero. rs, rt, and rd correspond to the two source and one destination registers, respectively. shamt is used in shift instructions instead of rt to make the hardware simpler. In assembly, to shift the value in $t4 two bits to the left and place the result in $t5:

sll $t5, $t4, 2

Since the opcode is zero for all R-format instructions, func specifies to the hardware exactly which R-format instruction to execute. The add example above would be encoded as follows:

opcode rs   rt     rd   shamt funct 000000 01001 01010 01000 00000 100000

Since it is an R-format instruction, the first six bits (opcode) are 0. The next 5 bits correspond to rs, which in this example is $t1. From the table above, we find that $t1 is $9, which in binary is 01001. Likewise, the next five bits encode $t2 = $10 = 01010. The destination is $t0 = $8 = 01000. We are not performing a shift, so shamt is 00000. Finally, since the func for the add instruction is 100000.

For the shift example above, the opcode field is again 0 since this is an R format instruction. The rs field is unused in shifts, so we leave the next five bits at 0. The rt field is $t4 = $12 = 01100. The rd field is $t5 = $13 = 01101. The shift amount, shamt, is 2 = 00010. Finally, the func field for sll is 000000. Thus, the encoding for sll $t5, $t4, 2 is:

opcode rs   rt     rd   shamt  funct 000000 00000 01100 01101 00010 000000

I Instructions
I instructions take two register arguments and a 16-bit "immediate" value. An immediate is a value that is stored as part of the instruction instead of in memory. This makes accessing constants much faster than if we had to put constants in memory and then load them (hence the name). I-format instructions, like R-format instructions, specify the target register (rt) first. Next comes one source register (rs) and finally the immediate value.


 * instruction rt, rs, imm

For example, let's say that we want to add the value 5 to the register $t1, and store the result in $t0:

addi $t0, $t1, 5

Or compare-and-branch to a nearby label (range is 16-bit signed displacement, left shifted by 2). The assembler calculates `(target - branch_insn_address + 4) >> 2` as the immediate:

beq $t0, $zero, t0_equals_zero_branch_target

I-format instructions are represented in machine code as follows:

The opcode specifies which operation is requested. rs and rt are five bits each, as before, and in the same positions as the R-format instructions. The imm field holds the immediate value. Depending on the instruction, the immediate constant may either be sign-extended or zero-extended. If a 32-bit immediate is needed, a special instruction, lui ("load upper immediate") exists for loading an immediate into the upper 16 bits of a register. That register can then be logically ORed with another 16-bit immediate to store the final value in that register. That value can then be used in normal R-format instructions. The following sequence of instructions stores the bit pattern 0101 0101 0101 ... into register $t0:

lui $t0, 0x5555 ori $t0, $t0, 0x5555

Typically, the assembler will automatically split 32-bit constants in this way so the programmer doesn't have to worry about it if you write li $t0, 0x5555555. Some assemblers (like MARS with extended pseudo-instructions mode enabled) even support large immediates for addi / addiu / xori / etc. this way.

The addi example from above would be encoded as follows. The addi instruction has an opcode of 001000. The source register, $t1, is number 9, or 01001 in binary. The target register, $t0, is number 8, or 01000 in binary. Five is 101 in binary, so addi $t0, $t1, 5 in machine code is:

opcode  rs   rt           imm 001000 01001 01000 0000 0000 0000 0101

addi and addiu sign-extend their 16-bit immediate to 32 bits (so can represent signed 2's complement values from -32768 .. 32767, i.e. unsigned values of 0..0x7fff and 0xffff8000 .. 0xffffffff)

Bitwise boolean logical instructions like andi, ori, and xori zero-extend the immediate (so can use values from 0..65535)

Other instructions that use the I format include load/store (address = register + signed 16-bit displacement), compare-immediate such as slti $t1, $t0, 1234 (producing a 0 or 1 in a register), lui for loading an immediate to the upper 16 bits of a register, and branch (but not jump) instructions.

J Instructions
J instructions are used to transfer program flow to a given hardcoded absolute address within the 256MB region around the PC register. J instructions are almost always written with labels: the assembler and linker will convert the label into a numerical value. A J instruction takes only one argument: the address to jump to.

(Truly PC-relative control transfers can be done instead with 'b label' that are usable in position-independent code. MIPS branch instructions are I-format instructions with a 16-bit relative displacement, left-shifted by 2, unlike jumps.)

instruction addr

There are two J-format instructions: j and jal. The latter will be discussed later. The j ("jump") instruction tells the processor to immediately skip to the instruction addressed by addr. To example, to jump to a label1:

j label1

A J-format instruction is encoded as

On MIPS32 machines, addresses are 32-bits wide, so 26 bits may not be enough to specify which instruction to jump to. Fortunately, since all instructions are 32-bits (four bytes) wide, we can assume that all instructions start at a byte address that's divisible by 4 (we are in fact guaranteed this by the loader). In binary, a number that is divisible by 4 ends with two zeros (just like a number that's divisible by 100 in decimal always ends in two zeros). Therefore, we can allow the assembler to leave out the last two zeros and have the hardware reinsert them. This effectively makes the address field 30 bits. The final four bits will be borrowed from the address of the next instruction, so we cannot let a program straddle a 256MB boundary, because a jump across the boundary would require a change in the 4 uppermost bits. (The PC while a J instruction is processed has already been updated to point to the next instruction, the branch-delay slot.)

In the example above, if label1 specified an instruction as address 120, or 1111000 in binary, we can encode the jump example above in machine code as follows. The opcode for j is 2, or 10 in binary, and we must chop off the last two bits of the jump address, leaving it at 11110. Thus, the machine code for j 120 is:

opcode |---addr---| 000010 0000 0000 0000 0000 0000 0111 10