Category:ARM Assembly: Difference between revisions

ARM Assembly
This programming language may be used to instruct a computer to perform a task.

See Also:	ARM Assembly on the HOPL

Listed below are all of the tasks on Rosetta Code which have been solved using ARM Assembly.

The ARM architecture is widely used on mobile phones and tablets. It falls under the category of RISC (Reduced Instruction Set Computer) processors, which means it has fewer opcodes than a CPU such as those in the x86 family. However, it makes up for this with its speed. The ARM and its variants are used in many well-known systems such as the Raspberry Pi, Nintendo DS, iPad, and more.

Barrel Shifter

The ARM can add a bit shift or rotate to one of its operands at no additional cost to execution time or bytecode. If the operand being shifted is a register, the value of that register is not actually changed. The shift or rotate only applies during that instruction.

<lang ARM Assembly>add r0,r0,r1 lsl 2 ;shift r1 left 2 bits, add r0 to r1, store the result in r0. r1 is unchanged after this instruction</lang>

Separate Destination for Math

With the x86, 68000, and other similar processors, arithmetic functions take two operands: the source and the destination. Anytime you add two numbers, one of them gets changed. This is not the case with the ARM. The destination can be a third register that isn't involved in the arithmetic whatsoever!

<lang ARM Assembly> add r3,r2,r1 ;add r2 to r1 and store the result in r3. r1 and r2 are unchanged.</lang>

Conditional Opcodes

Checking for condition codes isn't just limited to branching on the ARM. Almost every instruction can be made conditional. If the condition is not met, the opcode will have no effect. This saves a lot of cycles that would be spent branching just to execute a single instruction.

Compare the following snippets of code. The first is written in 8086 Assembly, the second in ARM. Both do the same thing, but ARM can do it without branching.

<lang asm>mov ax, word ptr [ds:TestData] ;dereference the pointer to TestData and store the value contained within that address into ax add ax,1 ;add 1 to ax jo OverflowSet ;the addition caused an overflow, jump to this label. ret ;return from subroutine

OverflowSet: sub ax,1 ;rollback the previous addition. ret ;return from subroutine.</lang>

The same code translated to ARM doesn't need to branch: <lang ARM Assembly>mov r1,#TestData ;get the address of TestData ldr r0,[r1] ;load the 32-bit value stored at TestData into r0 adds r0,r0,#1 ;add 1 to r0 and store the result in r0, updating the flags accordingly. subvs r0,r0,#1 ;subtract 1 from r0 and store the result in r0, only if the overflow flag was set.</lang>

If your code does one thing when a flag is set and another when that same flag is clear, the ARM can do both without having to branch at all.

<lang ARM Assembly>;ARM ASSEMBLY mov r1,#TestData ;get the address of TestData ldrs r0,[r1] ;load the 32-bit value stored at TestData into r0, updating the flags accordingly. addeq r0,r0,r2 ;if r0 equals zero, add r2 to r0 and store the result in r0. subne r0,r0,r2 ;if r0 doesn't equal zero, subtract r2 from r0 and store the result in r0.</lang>

The equivalent in x86 would take at least one branch, maybe 2 depending on the outcome: <lang asm>;x86 ASSEMBLY

   mov ax, word ptr [ds:TestData]
   cmp ax,0
   bne subtract_bx
   add ax,bx
   jmp done

subtract_bx:

   sub ax,bx

done:</lang>

Setting Flags

The flags, or condition codes, are only set by instructions that end in an "s," or by compare commands such as CMP. This lets you "preserve" the processor's state after an important calculation, but do some other things before execution branches depending on the result of that calculation. On any other processor, the calculation that determines whether a branch occurs MUST happen JUST before that branch statement or the branch will be taken based on the wrong data.

<lang ARM Assembly>cmp r0,r1 ;compare r0 to r1 ldr r2,[r3] ;load r2 from the address stored in r3 ldr r3,[r4] ;load r3 from the address stored in r4 bne myLabel ;branch to myLabel if the result of "cmp r0,r1" was not equal to zero.</lang>

Most processors would have to push and pop the condition code register between the compare and the branch. Otherwise, the act of loading r2 and r3 would affect the outcome of the branch. Not so on the ARM!

Limitations of the ARM

While the ARM has a rich amount of features that other processors only dream of having, there are a few limitations. The biggest one (which was more of an issue on earlier versions such as the ARM7TDMI CPU in the Game Boy Advance) is the limitation of the MOV command. Arguably the most important command any processor has (apart from JMP), the MOV command on the ARM is often limited in what can be loaded into a register in a single command. Depending on the pattern of bits, some immediate values cannot be loaded into a register directly. The key features of the ARM instructions (barrel shifter, conditional commands, etc) all take up bytes in each command, whether they are used in a given instance of a command or not. So in order to store 32 bit numbers in a MOV command, the value has to be "8-bit rotatable," meaning that it can be expressed as an 8 bit number if you shift it enough times. Basically if there are too many 1s in the binary equivalent of the number you're trying to load, it can't be done in one go.

Most of the time, this isn't a huge deal, as the easiest way around this is to define a data block nearby containing the value you wish to load. Since each command on the ARM takes 4 bytes of storage, this is guaranteed to take equal or fewer bytes than loading the number into a register piece-by-piece.

<lang ARM Assembly> mov r0,#0x04000000 add r0,r0,#0x130

compare to

TestDataAddr: .word 0x04000130 ldr r0,TestDataAddr</lang>

Thankfully, there's an even easier solution than this. The GNU Assembler saves the day with the following special notation. <lang ARM Assembly>mov r0, =#value</lang>

This isn't actually valid ARM code, it's more of a built-in macro. Essentially, the value will be loaded in one go as an immediate if it can. If not, it will get placed nearby as a data block and the MOV will be changed to an LDR command. Basically you can take everything in the above paragraph and forget about it, since equals notation does the work for you.