Is float assignment counted as using fpu? - linux

Question

Since fpu operation is very costly, I would like to less use fpu operations as far as I can. In the meanwhile, I'm wondering what kind of operations within the float variable would counted as the operation that fpu involves?
Such as the following code would involve fpu unit?
struct my_float_struct {
int f;
} g;
void func(float a)
{
g.f = a;
}
would calling func will cause lazy FPU context switch?

Answer 1

In kernel you shouldn't use float at all (use lookup tables and multiplying by coefficients to save precision on divide operations instead). In user space all operations involving float type on RHS would lead to FPU instruction usage. If you have 2 integers on RHS but assigning result to float (on LHS) -- you will end up with integer operations, but also with casting to float instruction (like
cvtsi2ss).
It's actually very easy to answer your question just looking into disassemble for corresponding C code.
C code (build with gcc -Wall -O0 -g main.c):
int main(void)
{
volatile float f1, f2;
volatile int i;
f1 = i * 356;
f2 = f1 * f2;
return 0;
}
Disassemble (using objdump -DS a.out):
int main(void)
{
4004b6: 55 push %rbp
4004b7: 48 89 e5 mov %rsp,%rbp
volatile float f1, f2;
volatile int i;
f1 = i * 356;
4004ba: 8b 45 f4 mov -0xc(%rbp),%eax
4004bd: 69 c0 64 01 00 00 imul $0x164,%eax,%eax
4004c3: 66 0f ef d2 pxor %xmm2,%xmm2
4004c7: f3 0f 2a d0 cvtsi2ss %eax,%xmm2
4004cb: 66 0f 7e d0 movd %xmm2,%eax
4004cf: 89 45 fc mov %eax,-0x4(%rbp)
f2 = f1 * f2;
4004d2: f3 0f 10 4d fc movss -0x4(%rbp),%xmm1
4004d7: f3 0f 10 45 f8 movss -0x8(%rbp),%xmm0
4004dc: f3 0f 59 c8 mulss %xmm0,%xmm1
4004e0: 66 0f 7e c8 movd %xmm1,%eax
4004e4: 89 45 f8 mov %eax,-0x8(%rbp)
return 0;
4004e7: b8 00 00 00 00 mov $0x0,%eax
}
From here you can see that:
when multiplying integer by integer, but assigning result to float, it leads to imul instruction (integer multiplying), but also to cvtsi2ss (converting to float) instruction.
when multiplying 2 floats, it leads to mulss instruction, which float multiplying
Basically you can see immediately if FPU involved by instruction operates on FPU registers, like %xmm0, %xmm1, etc.