Stack growth defeats optimization

[brson]

After adding a call to fib(42) in shootout-fibo's main function.

Without -O:

real    0m3.098s
user    0m3.080s
sys     0m0.010s

With -O:

real    0m5.799s
user    0m5.770s
sys     0m0.020s

Edit: When tasks do a lot of calculations without using a lot of stack, the effects of growing and shrinking the stack appear to have terrible results.

[brson]

I'm having a hard time understanding what I'm seeing. The above numbers are with the following main function:

fn main() { 
    assert (fib(8) == 21); 
    assert (fib(15) == 610);                        
    fib(42);                     
    log(debug, fib(8));                               
    log(debug, fib(15));                         
}

If I instead use this main function:

fn main() {                                                  
    fib(42);                                                 
}

Thin I get the following results with -O:

real    0m8.432s
user    0m8.390s
sys     0m0.030s

It should be doing less work but it's taking longer.

[brson]

With the smaller main function for some reason a lot of time ends up being spent in check_stack_canary.

[brson]

check_stack_canary does take a lot of time, but removing it doesn't change the general behavior here that -O is slower and the smaller main function is slower.

[brson]

Setting RUST_MIN_STACK=1000000 results in the optimized code being much faster.

[brson]

My guess is that the optimized code needs less stack, so it spends more time at the bottom of the stack where segments are smaller, and ends up making more requests to grow/shrink the stack

[brson]

I checked in some optimizations to stack growth that makes it perform much better. The optimized version of fib is still slightly worse than the unoptimized one, but I don't what else to do about it.