Prevent optimization of arithmetic computation

I need to implement this paper

Page 5, there is a function

void Split(double a, double *x, double *y) {   double z = a * ((1 << r) + 1);   *x = z - (z - a);   *y = a - *x;}

With -O3 (which I need for the rest of the program, gcc sees that z - (z - a) can be simplified to a (but only when function is inlined) :

Split(double, double*, double*):  movsd .LC0(%rip), %xmm1  mulsd %xmm0, %xmm1  movapd %xmm1, %xmm2  subsd %xmm0, %xmm2  subsd %xmm2, %xmm1  subsd %xmm1, %xmm0  movsd %xmm1, (%rdi)  movsd %xmm0, (%rsi)  ret.LC3:  .string "%lf -- %lf\n"main:  subq $8, %rsp  movl $.LC3, %edi  movl $2, %eax  movsd .LC1(%rip), %xmm1  movsd .LC2(%rip), %xmm0  call printf  xorl %eax, %eax  addq $8, %rsp  ret.LC0:  .long 67108864  .long 1099956224.LC1:  .long 0  .long -1104234414.LC2:  .long -335544320  .long 1073920081

assembly for the function itself is not optimized but in the caller (main), all intermediate operations vanish.

How can I have this function inlined but no optimization for it

Do I really have to write inline assembly ? Or is there any compiler hint ?

[EDIT] : i can add a volatile to z and prevent optimization. But then, compiler is not able to vectorize properly