[M3commit] CVS Update: cm3

[Jay Krell]

CVSROOT:	/usr/cvs
Changes by:	jkrell at birch.	10/02/28 01:32:53

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

Log message:
	inline 64bit shifts by a constant, or at least more instances (e.g. shifting a non-const by a const?)
	Oh, actually left_shift and right_shift already did this, this now affects "signed shift",
	which I just don't see much utility in really...
	
	including:
	
	before:
	
	< 	start_call_direct   p.71[_m3_shift64 at 12] 0 Word.64
	< 	load_stack_param   Word.32 0
	<  00004803: 68FFFFFFFF          PUSH $-1
	< 	load_stack_param   Word.64 0
	<  00004808: 8B5DF8              MOV EBX tv.71[_a]
	<  0000480B: 8B55FC              MOV EDX tv.71[_a]+4
	<  0000480E: 52                  PUSH EDX
	<  0000480F: 53                  PUSH EBX
	< 	call_direct       p.71[_m3_shift64 at 12] Word.64
	<  00004810: FF1500000000        CALL p.71[_m3_shift64 at 12]
	
	after:
	
	>  00004803: 8B5DF8              MOV EBX tv.71[_a]
	>  00004806: 8B55FC              MOV EDX tv.71[_a]+4
	>  00004809: 0FACD301            SHRD EBX EDX $1
	>  0000480D: D1EA                SHR EDX $1
	
	(though this really should use other addressing modes
	and that'd probably save an instruction)
	
	and then tricks like, shifting by more than 32:
	
	before:
	
	<  00004CA7: 6828000000          PUSH $40
	< 	load_stack_param   Word.64 0
	<  00004CAC: 8B5DF8              MOV EBX tv.71[_a]
	<  00004CAF: 8B55FC              MOV EDX tv.71[_a]+4
	<  00004CB2: 52                  PUSH EDX
	<  00004CB3: 53                  PUSH EBX
	< 	call_direct       p.71[_m3_shift64 at 12] Word.64
	<  00004CB4: FF1500000000        CALL p.71[_m3_shift64 at 12]
	
	after:
	
	>  00004B50: 8B5DF8              MOV EBX tv.71[_a]
	>  00004B53: 8B55FC              MOV EDX tv.71[_a]+4
	>  00004B56: 8BD3                MOV EDX EBX
	>  00004B58: C1E208              SHL EDX $8
	>  00004B5B: 33DB                XOR EBX EBX
	
	move low to high, zero low, shift high by count - 32
	
	though this too could be a lot better, at the very least, move edx ebx
	could be replaced just with a "virtual move" that reassigns the registers
	
	I'm strongly considering inlining all shifts and rotates, though I couldn't
	get it to work yet. This is a step.
	
	also notice this is less than one line changed, just to let the
	code run that was already there