-O3 optimizes away loop/switch for compile time data

o3-loop-switch.c

// https://godbolt.org/z/7zPrK8567 -O3 only

typedef struct op_t {
    const enum { ADD, SUB, DIV, MUL, SQUARE, READ, WRITE } tag;
    const float b;
    float *c;
} op_t;

typedef struct seq_t {
    const int n_ops;
    const op_t *ops;
} seq_t;

static void add(float *a, float b) { *a += b; }
static void sub(float *a, float b) { *a -= b; }
static void div(float *a, float b) { *a /= b; }
static void mul(float *a, float b) { *a *= b; }
static void square(float *a, float b) { *a = (*a)*(*a); }

static float run(const seq_t *seq, const float a0) {
    float a = a0;
    for (int i=0; i<seq->n_ops; i++) {
        op_t o = seq->ops[i];
        float val = o.c ? *o.c : o.b;
        switch(o.tag) {
            case ADD: add(&a, val); break;
            case SUB: sub(&a, val); break;
            case DIV: div(&a, val); break;
            case MUL: mul(&a, val); break;
            case SQUARE: square(&a, val); break;
            case READ: a = *o.c;
            case WRITE: *o.c = a;
        }
    }
    return a;
}

float bistable(float x) {
    float x2;
    const op_t ops[] = {
        {.tag=SQUARE},
        {.tag=WRITE, .c=&x2},
        {.tag=MUL, .b=x},
        {.tag=DIV, .b=-3.0f},
        {.tag=ADD, .b=x},
        {.tag=SQUARE},
        {.tag=ADD, .c=&x2},
    };
    const seq_t seq = {.n_ops = 7, .ops = ops};
    return run(&seq, x);
}

float bistable1(float x) {
    float x2 = x*x;
    float a = x + x2*x/-3.0f;
    return a*a + x2;
}

o3-loop-switch.S

bistable:
        movaps  xmm3, xmm0
        movaps  xmm1, xmm0
        mulss   xmm3, xmm0
        mulss   xmm1, xmm3
        divss   xmm1, DWORD PTR .LC0[rip]
        addss   xmm1, xmm0
        movaps  xmm0, xmm1
        mulss   xmm0, xmm1
        addss   xmm0, xmm3
        ret
bistable1:
        movaps  xmm3, xmm0
        movaps  xmm1, xmm0
        mulss   xmm3, xmm0
        mulss   xmm1, xmm3
        divss   xmm1, DWORD PTR .LC0[rip]
        addss   xmm1, xmm0
        movaps  xmm0, xmm1
        mulss   xmm0, xmm1
        addss   xmm0, xmm3
        ret
.LC0:
        .long   -1069547520