AMD Athlon Processor x86 Optimization Manual page 106

AMD Athlon™ Processor x86 Code Optimization
90
$r_two_divs:
MOV ECX, EAX
MOV EAX, EDX
XOR EDX, EDX
DIV EBX
MOV EAX, ECX
DIV EBX
MOV EAX, EDX
XOR EDX, EDX
POP EBX
RET
$r_big_divisor:
PUSH EDI
MOV EDI, ECX
SHR EDX, 1
RCR EAX, 1
ROR EDI, 1
RCR EBX, 1
BSR ECX, ECX
SHRD EBX, EDI, CL
SHRD EAX, EDX, CL
SHR EDX, CL
ROL EDI, 1
DIV EBX
MOV EBX, [ESP+12] ;dividend lo-word
MOV ECX, EAX
IMUL EDI, EAX
MUL DWORD PTR [ESP+20] ;quotient * divisor lo-word
ADD EDX, EDI
SUB EBX, EAX
MOV ECX, [ESP+16] ;dividend_hi
MOV EAX, [ESP+20] ;divisor_lo
SBB ECX, EDX
SBB EDX, EDX
AND EAX, EDX
AND EDX, [ESP+24] ;(remainder < 0)? divisor_hi : 0
ADD EAX, EBX
ADC EDX, ECX
POP EDI
POP EBX
RET
_ullrem ENDP
;save dividend_lo in ECX
;get dividend_hi
;zero extend it into EDX:EAX
;EAX = quotient_hi, EDX = intermediate
; remainder
;EAX = dividend_lo
;EAX = quotient_lo
;EAX = remainder_lo
;EDX = remainder_hi = 0
;restore EBX as per calling convention
;done, return to caller
;save EDI as per calling convention
;save divisor_hi
;shift both divisor and dividend right
; by 1 bit
;ECX = number of remaining shifts
;scale down divisor and dividend such
; that divisor is less than 2^32
; (i.e. fits in EBX)
;restore original divisor (EDI:ESI)
;compute quotient
;save quotient
;quotient * divisor hi-word (low only)
;EDX:EAX = quotient * divisor
;dividend_lo – (quot.*divisor)–lo
;subtract divisor * quot. from
; dividend
;(remainder < 0)? 0xFFFFFFFF : 0
;(remainder < 0)? divisor_lo : 0
;remainder += (remainder < 0)?
; divisor : 0
;restore EDI as per calling convention
;restore EBX as per calling convention
;done, return to caller
Efficient 64-Bit Integer Arithmetic
22007E/0—November 1999