How to update all the elements in a double array in X86?

Question

I am a newbie of X86 and I am stuck on updating a double array using the values of another double array. The following code is my function and I want to use inline assembly to replace the piece of code inside the loop. I have attached the error message below. Can anyone helps me to point out my errors? I am confused about the error messages and don't know how to revise it.

static inline void update(double * x,double * y,double * z,double * vx,
        double * vy,double * vz,uint32_t size){
        for (uint32_t i=0;i<size;++i){
            x[i] = x[i] + vx[i];
            y[i] = y[i] + vy[i];
            z[i] = z[i] + vz[i];
        }
}
uint32_t counter = 0;
__asm__ __volatile__ (  
    "loop: \n\t" 
    "faddq (%4), (%1)\n\t"
    "faddq (%5), (%2)\n\t"
    "faddq (%6), (%3)\n\t"
    "addq $8, %1\n\t"
    "addq $8, %2\n\t"
    "addq $8, %3\n\t"
    "addq $8, %4\n\t"
    "addq $8, %5\n\t"
    "addq $8, %6\n\t"
    "incq %0\n\t"
    "cmp %0, %7\n\t"
    "jne loopb"
    : "+r"(counter)
    : "r" (x),"r" (y),"r"(z),"r"(vx),"r"(vy),"r"(vz),"r"(size) 
    : "memory", "cc");

Error Messages:

update_locations_ass.c:150:15: error: invalid instruction mnemonic 'faddq'
        "loop: \n\t" 
                 ^
<inline asm>:2:2: note: instantiated into assembly here
        faddq (%rdi), (%rcx)
        ^~~~~
update_locations_ass.c:151:25: error: invalid instruction mnemonic 'faddq'
        "faddq (%4), (%1)\n\t"
                           ^
<inline asm>:3:2: note: instantiated into assembly here
        faddq (%r8), (%rdx)
        ^~~~~
update_locations_ass.c:152:28: error: invalid instruction mnemonic 'faddq'
        "faddq (%5), (%2)\n\t"
                           ^
<inline asm>:4:2: note: instantiated into assembly here
        faddq (%r9), (%rsi)
        ^~~~~
update_locations_ass.c:159:23: error: invalid operand for instruction
        "addq $8, %6\n\t"
                      ^
<inline asm>:11:7: note: instantiated into assembly here
        incq %eax

Compiler version: Configured with: --prefix=/Applications/Xcode.app/Contents/Developer/usr --with-gxx-include-dir=/usr/include/c++/4.2.1 Apple LLVM version 6.1.0 (clang-602.0.53) (based on LLVM 3.6.0svn) Target: x86_64-apple-darwin14.0.0 Thread model: posix

Answer 1

I'm equally confused here. What is faddq , and where did you get it from? Is it supposed to be fadd ? You can't use two memory operands with fadd anyway, so the code looks completely incorrect. If you're curious about the correct way to do it, try compiling with -S and -O2 so you can look at optimized compiler output.

If you want to get a faster version of the function, easiest to just do it in C anyway. Assuming that the arrays don't overlap, here is a much faster version:

// Assuming x and vx do not overlap
void update1(double *restrict x, const double *restrict vx, unsigned count) {
    for (unsigned i = 0; i < count; i++) {
        x[i] += vx[i];
    }
}

void update(/* ... */) {
    update1(x, vx, count);
    update1(y, vy, count);
    update1(z, vz, count);
}

If you compile with -O3 , the compiler will generate code that uses addpd , depending on your compilation target. This is going to be miles better than anything you could write yourself using the x87 FPU instructions.

These simple functions--just adding arrays to other arrays--are very easy for the compiler to optimize, so unless you are teaching yourself assembly language, just let the compiler do it for you.