Skip to content

Instantly share code, notes, and snippets.

@orlp
Last active November 15, 2024 18:29
Show Gist options
  • Save orlp/3551590 to your computer and use it in GitHub Desktop.
Save orlp/3551590 to your computer and use it in GitHub Desktop.
int64_t ipow(int64_t base, uint8_t exp) {
static const uint8_t highest_bit_set[] = {
0, 1, 2, 2, 3, 3, 3, 3,
4, 4, 4, 4, 4, 4, 4, 4,
5, 5, 5, 5, 5, 5, 5, 5,
5, 5, 5, 5, 5, 5, 5, 5,
6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 6,
6, 6, 6, 6, 6, 6, 6, 255, // anything past 63 is a guaranteed overflow with base > 1
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255,
};
int64_t result = 1;
switch (highest_bit_set[exp]) {
case 255: // we use 255 as an overflow marker and return 0 on overflow/underflow
if (base == 1) {
return 1;
}
if (base == -1) {
return 1 - 2 * (exp & 1);
}
return 0;
case 6:
if (exp & 1) result *= base;
exp >>= 1;
base *= base;
case 5:
if (exp & 1) result *= base;
exp >>= 1;
base *= base;
case 4:
if (exp & 1) result *= base;
exp >>= 1;
base *= base;
case 3:
if (exp & 1) result *= base;
exp >>= 1;
base *= base;
case 2:
if (exp & 1) result *= base;
exp >>= 1;
base *= base;
case 1:
if (exp & 1) result *= base;
default:
return result;
}
}
@KevinTyrrell
Copy link

A lot of repetitious code here. if (exp & 1) result *= base; is written six times, only needs to be written once. Also the array will be created and destroyed every function call which has overhead. You can move it outside the function header to prevent it from being continuously re-instantiated.

@ian-abbott
Copy link

@kevintyrell The code repetition is to optimize for speed. The array is static so will not be created and destroyed at every function call. The placement of the array within the function merely limits its static scope to that function.

@brdann
Copy link

brdann commented Aug 4, 2018

That's cool! But it computes unsigned int64_t and returns signed int64_t, what I'm missing?

@shengyang998
Copy link

shengyang998 commented Nov 27, 2018

@brdann Yeah, I think it should return signed.
Whatever it is, signed or unsigned, somehow the answer is wrong on my device.

// I got it. The int64_t doesn't work. Change to long long did solve it.

@WarpspeedSCP
Copy link

I feel like I'm missing something. There aren't any loops here. Does the function have to be called recursively?

@XMB5
Copy link

XMB5 commented Feb 13, 2019

The switch cases fall through

@alejandro-colomar
Copy link

You should check for overflow, because squaring base without any checks is asking for trouble. ipow(INT64_MAX, 2) invokes Undefined Behavior. ipow(3, 63) also invokes UB.

@alejandro-colomar
Copy link

For the code repetition I would make a static inline function, which avoids repetition, but keeps performance

@Clemapfel
Copy link

this can easily be made constexpr / consteval by moving highest_bit_set into the same namespace as the function and making that constexpr too which I recommend as it would be in the spirit of optimal runtime

@Zorgatone
Copy link

@orlp

Does this work with uint64_t as well?
Do I need to add elements to highest_bit_set past exponent 63?

@unclefunctor
Copy link

The optimization here is at a whole new level. Thanks for sharing and well done sir!

@xsbee
Copy link

xsbee commented Jun 18, 2022

Use switch fallthroughs.

@chemoelectric
Copy link

chemoelectric commented Feb 5, 2023

That is some dang clever loop unrolling.

@QVRE
Copy link

QVRE commented Sep 20, 2023

Wouldn't renaming 255 to 7 make this code easier to turn into an array lookup for the compiler ?

@MOJNICK
Copy link

MOJNICK commented Jan 29, 2024

@ian-abbott Have you tested performance for static and non-static "highest_bit_set version"?

@MOJNICK
Copy link

MOJNICK commented Jan 29, 2024

@Zorgatone yes you need to add, It will be equal to 7. And you need to add switch-case 7: . + Change signed parameters/values to unsigned. Or if you are sure the results you are aiming are <= int_64_max then of course you can use this function straight away ( in that case casting rules of C applies )

@ian-abbott
Copy link

ian-abbott commented Jan 30, 2024

@MOJNICK If highest_bit_set is non-static, the code initializes the array contents every time the function is called, so it is slower than the static version.

@nschaeff
Copy link

nschaeff commented Mar 8, 2024

the lookup table for highest bit set can be replaced with __builtin_clz if it is available (gcc, clang). It will make it faster (especially if the lookup table is not in cache, which is likely in a real-world application doing other things than that) and smaller.

Simply replace highest_bit_set[exp] with (exp>0) ? 32 - __builtin_clz(exp) : 0

@mogando668
Copy link

your code should end the cases sooner - when exp is down to { 0 , 1 , 2 }

Cuz from there it's very obvious what needs to be returned :::

   return { <= 0 := result
               1 := result * base
               2 := result * base * base  
          }

or more succinctly ::

  return ( base ** exp ) * result

Because when exp == 2 , you already know ahead of time that ( exp >>= 1 ) & 1 must be true, so why bother evaluating the obvious ?

Another speed up trick you can contemplate is that if you already know up-front that the exponent is 1 short of a power of 2, then you already know every single bit is a 1.

In that case, make it into a vanilla countdown loop for # of bits, and simply doing the exact same thing every round without bothering to either right shift or check for ( exp & 1 ). In my own code, for this scenario, I usually make the starting point slightly lagged, so each round I simply do a compound statement of

    result *= base *= base       # intentionally lagged, so one must be careful regarding when you 
                                 # need to perform the extra one to sync them back up with each other

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment