[M3commit] CVS Update: cm3

[Jay Krell]

CVSROOT:	/usr/cvs
Changes by:	jkrell at birch.	10/02/14 13:02:21

Modified files:
	cm3/m3-sys/m3back/src/: M3x86.m3 

Log message:
	fix indirection in atomic fetch_and_op
	just my ignorance using movOp instead of load_ind
	code is like:
	_Main__Test_AtomicInteger_FetchOr:
	00000250: 8D 35 38 00 00 00  lea         esi,[_MM_Main+38h] ; atomic variable
	00000256: 8B 15 98 00 00 00  mov         edx,dword ptr [_MM_Main+98h] ; value to or with
	0000025C: 8B 06              mov         eax,dword ptr [esi] ; fetch old value
	0000025E: 8B D8              mov         ebx,eax ; move old value to tmp (this is needed)
	00000260: 0B DA              or          ebx,edx ; compute new value
	00000262: F0 0F B1 1E        lock cmpxchg dword ptr [esi],ebx ; try to swap
	00000266: 75 F6              jne         0000025E
	; old value is in eax (always eax; or edx:eax pair for 64bit operations, not yet implemented)
	; new value is in, in this case, ebx
	
	still could be better, could be:
	mov         edx,dword ptr [_MM_Main+98h]
	mov         eax,[_MM_Main+38h]
	retry:
	mov         ebx,eax
	or          ebx,edx
	lock cmpxchg [_MM_Main+38h],ebx
	jne         retry
	; old value is in eax
	; new value is in, in this case, ebx
	
	six instructions instead of seven
	three registers (eax, ebx, edx) instead of four (eax, ebx, edx, esi)
	(problem is that we get load_address on the variable first,
	and we go ahead and enregister that address unnecessarily)
	
	And even so: if six instructions is considered too large, we can
	use a function instead.
	
	As well, add, sub, we can do more efficiently with xadd.
	
	(still need to check what libatomic does here)
	
	also we can probably optimize
	or(-1) => store_ordered(-1)
	or(0) => do nothing
	and(-1) => do nothing
	and(0) => store_ordered(0)
	
	probably they are rare/nonexistant