[KOS] optimization oddity

[Jim U. <ji...@3e...>]

While experimenting with the tsunami Vector class, I ran into a C++
feature that will be old hat to C++ gurus but was not immediately
obvious to me.

Basically, instead of this:                  you should do this:

void compute(Vector &v) {                    void compute(Vector &v) {         
    Vector vc;                                   for (int i = 0; i < 10; i++) {
    for (int i = 0; i < 10; i++) {                   Vector vc = v.normalize();
        vc = v.normalize();                          printf("%f\n", vc.x);     
        printf("%f\n", vc.x);                    }                             
    }                                        }                                 
}

The reason: v.normalize() constructs and returns a temporary object (the
normalized vector).  This temporary can be optimized away, but only in
an initializer, not in an assignment.  For the nitty-gritty, see
http://blogs.msdn.com/slippman/archive/2004/02/03/66739.aspx.

For the same reason, if you have a vector you need to normalize in
place, you should write "v.normalize_self()" instead of "v = v.normalize()",
if such a function is available.

Unfortunately, even with the temporary optimized out, gcc still writes
the normalized vector to vc (on the stack) every single iteration, even
though only vc.x is used, and only once.  I don't know why.  Built-in
types like floats could be faster, even though less convenient; the
stack may not be touched at all.

Following is the assembly output for the end of the two loops (after the
vector length has been put in fr1), showing the divide and the printf.
The extra code in the left column is where the temporary is copied from
the stack to vc (also on the stack).


.L16:                             .L16:
    ...                               ...
                                                        
    fdiv    fr1,fr2                   fdiv    fr1,fr2      
    fdiv    fr1,fr4                   fdiv    fr1,fr5      
    fmov.s  fr2,@r14                  fmov.s  fr2,@r14     
    fdiv    fr1,fr5                   fdiv    fr1,fr4      
    fmov.s  fr4,@(r0,r15)             fmov.s  fr5,@(r0,r15)
    mov     #20,r0                    mov     #8,r0            
    fdiv    fr1,fr3                   fdiv    fr1,fr3      
    fmov.s  fr5,@(r0,r15)             fmov.s  fr4,@r15     
    mov     #24,r0                    jsr     @r13               // printf
    fmov.s  fr3,@(r0,r15)             fmov.s  fr3,@(r0,r15)
    mov.l   @(16,r15),r1
    mov.l   r1,@r15
    mov.l   @(20,r15),r1
    fmov.s  @r15,fr4
    mov.l   r1,@(4,r15)
    mov.l   @(24,r15),r1
    mov.l   r1,@(8,r15)
    mov.l   @(28,r15),r1
    jsr     @r13                // printf
    mov.l   r1,@(12,r15)



-- 
"The FBI said al-Rabeei's aliases may include: Fawaz Yahia Hassan Aribii, 
Fawaz al-Rubai, Fawaz Yehia Hassan al-Rabie, Fawaz Yahya al-Rabi'i, Fawaz Yahya
al-Ribi (al-Ruba'i, al-Rabia'i, al-Rabi'i), Forqan al-Tajiki, Furgan al-Tajiki,
Furqan the Chechen, Faris al-Baraq, Sa'id, Musharraf, and Salem al-Frahan."
ji...@3e... / 0x43340710 / 517B C658 D2CB 260D 3E1F  5ED1 6DB3 FBB9 4334 0710