What's the easiest way to determine if a register's value is equal to zero or not?

问题

I'm using x86 assembly with the Irvine library.

What's the easiest way to check if a register value is equal to zero or not?

I used cmp instruction but i'm searching for alternative way. This is my code using cmp instruction and the register is ebx

    cmp ebx,0
    je equ1
    mov ebx,0
    jmp cont
equ1:
    mov ebx,1
    jmp cont
cont:
    exit

This "booleanizes" a value, producing a 0 or 1 like int ebx = !!ebx would in C.

回答1:

Probably the "easiest", or simplest, "not-caring about details" answer how to determine is:

    ; here ebx is some value, flags are set to anything
    test   ebx,ebx   ; CF=0, ZF=0/1 according to ebx
    jz     whereToJumpWhenZero
    ; "non-zero ebx" will go here

    ; Or you can use the inverted "jnz" jump to take
    ; a branch when value was not zero instead of "jz".

There's a detailed answer from Peter Cordes to "testl eax against eax?" question reasoning about flags being set, etc. Also has a link to yet another similar answer, but reasoning about why it is best way from performance point of view. :)

How to set some other register (I will pick eax) to 1 when ebx is zero, and to 0 when ebx is non-zero (non destructive way for ebx itself):

    xor   eax,eax  ; eax = 0 (upper 24 bits needed to complete "al" later)
    test  ebx,ebx  ; test ebx, if it is zero (ZF=0/1)
    setz  al       ; al = 1/0 when ZF=1/0 (eax = 1/0 too)

Or how to convert ebx itself into 1/0 when ebx is zero/non-zero:

    neg   ebx      ; ZF=1/0 for zero/non-zero, CF=not(ZF)
    sbb   ebx,ebx  ; ebx = 0/-1 for CF=0/1
    inc   ebx      ; 1 when ebx was 0 at start, 0 otherwise

Or how to convert ebx itself into 1/0 when ebx is zero/non-zero, other variant (faster on "P6" to "Haswell" cores):

    test  ebx,ebx  ; ZF=1/0 for zero/non-zero ebx
    setz  bl       ; bl = 1/0 by ZF (SETcc can target only 8b r/m)
    movzx ebx,bl   ; ebx = bl extended to 32 bits by zeroes

etc, etc... It depends what happens before your testing, and also what you really want as output of test, there're many possible ways (optimal for different situations, and optimal for different target CPU).

I will add few more extremely common situations... A counter-loop down-counting from N to zero, to loop N times:

    mov   ebx,5   ; loop 5 times
exampleLoop:
    ; ... doing something, preserving ebx
    dec   ebx
    jnz   exampleLoop ; loop 5 times till ebx is zero

How to process 5 elements of word (16b) array (accessing them in array[0], array[1], ... order):

    mov   ebx,-5
    lea   esi,[array+5*2]
exampleLoop:
    mov   ax,[esi+ebx*2]  ; load value from array[i]
    ; process it ... and preserve esi and ebx
    inc   ebx
    jnz   exampleLoop  ; loop 5 times till ebx is zero

One more example, I somehow like this one a lot:

How to set target register (eax in example) to ~0 (-1)/0 when ebx is zero/non-zero and you already have value 1 in some register (ecx in example):

    ; ecx = 1, ebx = some value
    cmp   ebx,ecx  ; cmp ebx,1 => CF=1/0 for ebx zero/non-zero
    sbb   eax,eax  ; eax = -1 (~0) / 0 for CF=1/0 ; ebx/ecx intact

The -1 may look as practical as 1 (for indexing purposes at least), but -1 works also as full bitmask for further and/xor/or operations, so sometimes it is more handy.

回答2:

Use the flags, Luke
You test if a register is zero by checking the zero flag.
If the register obtained its value by some operation that affected the flags (or, more specifically, the zero flag) then you don't have to do anything, because the zero flag will already reflect the value stored in that register.

Test only if needed
If you cannot guarantee that the flags have been set, you'll have to use a test operation.
These operations come in two flavors: destructive and non-destructive.

You can see a list of instructions and the flags it alters at: http://ref.x86asm.net—more specifically, at: http://ref.x86asm.net/coder32-abc.html

The mov and lea instructions never alter flags and thus need assistance. Most other instructions set at least one flag.

Don't create false dependencies
If you need to test a register for zero, but don't want to alter its value, you use the test instruction.
You should not use a or or and instruction to check a register, because the CPU may not know that or/and can be used non-destructively and may not be able to apply certain optimizations. The technical term for this is a 'false dependency'. The register needs ebx and 'thinks' it was changed recently so it waits for the result to finalize.

test ebx, ebx  ; <-- CPU knows ebx was not altered, no stalls on subsequent reads.
or   ebx, ebx  ; <-- CPU 'thinks' ebx was changed, stall on subsequent read.

If you want the zero state to reflect in another register, you can simply mov ebx to another register.

Reduce a value to a boolean
If you want to reduce the register to a boolean (True if non-zero, False otherwise), you use one of the following sequences:

; ebx holds the value to reduce to a boolean.
; eax is an unused register.

xor eax, eax   ; eax = 0
sub eax, ebx   ; eax = 0 - ebx; CF (carry flag) = 1 if ebx <> 0
sbb ebx, ebx   ; ebx = ebx - ebx - CF
               ; <<-- ebx = -1 if non zero, 0 if zero

xor eax, eax   ; eax = 0
sub eax, ebx   ; eax = - ebx; CF = 1 if ebx <> 0
adc ebx, eax   ; ebx = (ebx + -ebx) aka 0 + CF
               ; <<== ebx = 1 if non zero, 0 if zero

test ebx, ebx  ; force ZF to be correct
setnz al       ; Store 1 if non-zero, 0 otherwise in byte register AL.

Note that the use of byte registers can be problematic due to stalls related to "partial register writes".

回答3:

You could use:

or ebx, 0 ;This does nothing, just triggers the zero flag if ebx is zero
jnz notZero 
or ebx, 1 ;ebx was zero, then ebx is 1
notZero:

来源：https://stackoverflow.com/questions/41174867/whats-the-easiest-way-to-determine-if-a-registers-value-is-equal-to-zero-or-no

标签

assembly

x86

micro-optimization