Objc_msgSend(), instruction by instruction

Here's the disassembly for objc_msgSend on ARMv6, iOS4.2.1 (sorry no ARMv7 devices on my desk at the moment). I'll try to annotate it:

  0x32d98f0c <objc_msgSend+0>:	teq	r0, #0	; 0x0
  0x32d98f10 <objc_msgSend+4>:	moveq	r1, #0	; 0x0
  0x32d98f14 <objc_msgSend+8>:	bxeq	lr

Tests if r0 (the receiver) is nil, and if so, sets r1 selector to 0x0 and returns. I guess there's no nil handler on ARM. Simple function return values are in registers r0-r3, so I guess r1 is set to 0 in case the caller is expecting a long long.

If receiver is non-nil:

  0x32d98f18 <objc_msgSend+12>:	push	{r3, r4, r5, r6}
  0x32d98f1c <objc_msgSend+16>:	ldr	r4, [r0]

^ this loads the class pointer (isa) into r4

This looks similar to the bit twiddling x86 cache lookup in the class; it's somewhat harder to read than the x86 due to ARM's bit packing shortcuts[1]:

  0x32d98f20 <objc_msgSend+20>:	ldr	r5, [r4, #8]
  0x32d98f24 <objc_msgSend+24>:	ldr	r6, [r5]
  0x32d98f28 <objc_msgSend+28>:	add	r3, r5, #8	; 0x8
  0x32d98f2c <objc_msgSend+32>:	and	r5, r6, r1, lsr #2
  0x32d98f30 <objc_msgSend+36>:	ldr	r4, [r3, r5, lsl #2]

Check if the method is NULL, and if so, jump to the cache miss at the end:

  0x32d98f34 <objc_msgSend+40>:	teq	r4, #0	; 0x0
  0x32d98f38 <objc_msgSend+44>:	add	r5, r5, #1	; 0x1
  0x32d98f3c <objc_msgSend+48>:	beq	0x32d98f60 <objc_msgSend+84>

This would appear to be the part checking if this is the cache entry we're looking for:

  0x32d98f40 <objc_msgSend+52>:	ldr	r12, [r4]
  0x32d98f44 <objc_msgSend+56>:	teq	r1, r12
  0x32d98f48 <objc_msgSend+60>:	and	r5, r5, r6

If it isn't, loop:

  0x32d98f4c <objc_msgSend+64>:	bne	0x32d98f30 <objc_msgSend+36>

If it is, restore the registers and do an indirect jump into the method we found (I'm not sure what the teq r4,r4 is for):

  0x32d98f50 <objc_msgSend+68>:	ldr	r12, [r4, #8]
  0x32d98f54 <objc_msgSend+72>:	teq	r4, r4
  0x32d98f58 <objc_msgSend+76>:	pop	{r3, r4, r5, r6}
  0x32d98f5c <objc_msgSend+80>:	bx	r12

Tail call into the slow version with full lookup (after restoring the clobbered registers and the stack pointer):

  0x32d98f60 <objc_msgSend+84>:	pop	{r3, r4, r5, r6}
  0x32d98f64 <objc_msgSend+88>:	b	0x32d98f68 <objc_msgSend_uncached>

[1] A lazy excuse of course. I just haven't read or written anywhere near as much ARM assembly as x86.