C - Implementing fast push of many elements to the end of array

Question

I have a simple struct to hold an array:

struct array_of_a_type {
        size_t allocated_size;
        size_t elements; /* 1-index based */
        a_type *array;
};

I want to write a simple function, something like this:

bool simple_function(struct array_of_a_type *my_array, int a, int b, int c, int d)
{
    a_type new_chunk[] = {
        a,   b,   a+b, d,   c,
        c,   c,   c+d, b+d, a,
        a+c, b+c, c+d, c+d, c,
    };
    size_t size = sizeof(new_chunk) / sizeof(a_type);
    return push_to_array(my_array, new_chunk, size);
}

The my_array is a static, global variable. Below is an implementation of push_to_array.

static bool push_to_array(struct array_of_a_type *a, a_type *new_chunk, size_t size)
{
    const size_t new_size = a->elements + size;
    const size_t old_size = a->elements;
    if (new_size > a->allocated_size) {
        /* The allocated_size is most of the time big enough.
           I’ve stripped this part of code to minimum. */
        a_type *tmp = realloc(a->array, new_size * sizeof(a_type));
        if (!tmp) {
            return true;
        } else {
            a->array = tmp;
            a->allocated_size = new_size;
        }
    }
    a->elements = new_size;
    memcpy(a->array + old_size, new_chunk, size * sizeof(a_type));
    return false;
}

My question:
How can I rewrite ‘simple_function’ to make more compilers generate code that will write directly to the destination? I would like the code to stay quite short and flexible.

My code works. Unfortunately the gcc (and an old clang) create temporary data on the stack and then copy it to destination. Below if a fragment of generated x86_64 assembler.

movq    8(%rsp), %rdx
movq    %rdx, 8(%rax)
movq    16(%rsp), %rdx
movq    %rdx, 16(%rax)
movq    24(%rsp), %rdx
movq    %rdx, 24(%rax)
movq    32(%rsp), %rdx
movq    %rdx, 32(%rax)

For AMD the assembler have this:

rep movsq

The new clang works fine. I've compiled with -O3.

I have tried with code that added one element a time. There was a lot of conditional jumps to call realloc, unfortunately.

Answer 1

For efficiency, you need to separate the logic for growing the array, and assigning the values to the (unused) slots, to avoid the extra copy (from stack to array).

To prettify the code, you can create a set of helper macros. I shall assume that by "push" you mean "append to the array". If you really meant "prepend", then there is an additional memmove() needed.

Let's assume you have

#include <stdlib.h>
#include <stdio.h>

typedef int  array_data_type;

typedef struct {
    size_t           size;
    size_t           used;
    array_data_type *item;
} array_type;

#define ARRAY_INITIALIZER { 0, 0, NULL }

void array_free(array_type *const array)
{
    free(array->item);
    array->size = 0;
    array->used = 0;
    array->item = NULL;
}

void array_init(array_type *const array)
{
    array->size = 0;
    array->used = 0;
    array->item = NULL;
}

void array_init_size(array_type *const array, const size_t size)
{
    if (!size) {
        array->size = 0;
        array->used = 0;
        array->item = NULL;
        return;
    }

    array->item = malloc(size * sizeof array->item[0]);
    if (!array->item) {
        fprintf(stderr, "array_init_size(%p, %zu): Out of memory.\n", (void *)array, size);
        exit(EXIT_FAILURE);
    }
    array->size = size;
    array->used  = 0;
}

void array_grow_to(array_type *const array, size_t size)
{
    array_data_type *temp;

    if (size < 4)
        size = 4;
    else
    if (size < 16777216) {
        size |= size >> 1;
        size |= size >> 2;
        size |= size >> 4;
        size |= size >> 8;
        size |= size >> 16;
        size++;
    } else
        size = (size | 8388607) + 8388609;

    temp = realloc(array->item, size * sizeof array->item[0]);
    if (!temp) {
        fprintf(stderr, "array_grow_to(%p, %zu): Out of memory.\n", (void *)array, size);
        exit(EXIT_FAILURE);
    }

    array->item = temp;
    array->size = size;
}

static inline array_data_type *array_grow_by(array_type *const array, size_t const count)
{
    array_data_type *retval;

    if (array->used + count > array->size)
        array_grow_to(array, array->used + count);

    retval = array->item + array->used;
    array->used += count;
    return retval;
}

I like to use used for the number of elements in the array, and size for the number of elements the array has memory allocated for. Do a search and replace if you're used to some other names.

array_grow_to() adjusts the new size to at least 4, or the next power of two if less than 16,777,216, or to a larger multiple of 8,388,608. This limits the amount of allocated but unused memory for very large lists.

array_grow_by() ensures the array has room for count new elements, and returns the pointer to the first new unused element.

If you define the following C99 preprocessor macros,

#define MACRO_CONCATENATE(part1, ...)   part1 ## __VA_ARGS__

#define ARRAY_SET_N(array, count, ...)  MACRO_CONCATENATE(ARRAY_SET_, count)(array, count, __VA_ARGS__)
#define ARRAY_SET_0(...)
#define ARRAY_SET_1(a, n, v)        a[n-1] = v
#define ARRAY_SET_2(a, n, v, ...)   a[n-2] = v; ARRAY_SET_1(a, n, __VA_ARGS__)
#define ARRAY_SET_3(a, n, v, ...)   a[n-3] = v; ARRAY_SET_2(a, n, __VA_ARGS__)
#define ARRAY_SET_4(a, n, v, ...)   a[n-4] = v; ARRAY_SET_3(a, n, __VA_ARGS__)
#define ARRAY_SET_5(a, n, v, ...)   a[n-5] = v; ARRAY_SET_4(a, n, __VA_ARGS__)
#define ARRAY_SET_6(a, n, v, ...)   a[n-6] = v; ARRAY_SET_5(a, n, __VA_ARGS__)
#define ARRAY_SET_7(a, n, v, ...)   a[n-7] = v; ARRAY_SET_6(a, n, __VA_ARGS__)
#define ARRAY_SET_8(a, n, v, ...)   a[n-8] = v; ARRAY_SET_7(a, n, __VA_ARGS__)
#define ARRAY_SET_9(a, n, v, ...)   a[n-9] = v; ARRAY_SET_8(a, n, __VA_ARGS__)
#define ARRAY_SET_10(a, n, v, ...)  a[n-10] = v; ARRAY_SET_9(a, n, __VA_ARGS__)
#define ARRAY_SET_11(a, n, v, ...)  a[n-11] = v; ARRAY_SET_10(a, n, __VA_ARGS__)
#define ARRAY_SET_12(a, n, v, ...)  a[n-12] = v; ARRAY_SET_11(a, n, __VA_ARGS__)
#define ARRAY_SET_13(a, n, v, ...)  a[n-13] = v; ARRAY_SET_12(a, n, __VA_ARGS__)
#define ARRAY_SET_14(a, n, v, ...)  a[n-14] = v; ARRAY_SET_13(a, n, __VA_ARGS__)
#define ARRAY_SET_15(a, n, v, ...)  a[n-15] = v; ARRAY_SET_14(a, n, __VA_ARGS__)
#define ARRAY_SET_16(a, n, v, ...)  a[n-16] = v; ARRAY_SET_15(a, n, __VA_ARGS__)
#define ARRAY_SET_17(a, n, v, ...)  a[n-17] = v; ARRAY_SET_16(a, n, __VA_ARGS__)
#define ARRAY_SET_18(a, n, v, ...)  a[n-18] = v; ARRAY_SET_17(a, n, __VA_ARGS__)
#define ARRAY_SET_19(a, n, v, ...)  a[n-19] = v; ARRAY_SET_18(a, n, __VA_ARGS__)
#define ARRAY_SET_20(a, n, v, ...)  a[n-20] = v; ARRAY_SET_19(a, n, __VA_ARGS__)
#define ARRAY_SET_21(a, n, v, ...)  a[n-21] = v; ARRAY_SET_20(a, n, __VA_ARGS__)
#define ARRAY_SET_22(a, n, v, ...)  a[n-22] = v; ARRAY_SET_21(a, n, __VA_ARGS__)
#define ARRAY_SET_23(a, n, v, ...)  a[n-23] = v; ARRAY_SET_22(a, n, __VA_ARGS__)
#define ARRAY_SET_24(a, n, v, ...)  a[n-24] = v; ARRAY_SET_23(a, n, __VA_ARGS__)
#define ARRAY_SET_25(a, n, v, ...)  a[n-25] = v; ARRAY_SET_24(a, n, __VA_ARGS__)
#define ARRAY_SET_26(a, n, v, ...)  a[n-26] = v; ARRAY_SET_25(a, n, __VA_ARGS__)
#define ARRAY_SET_27(a, n, v, ...)  a[n-27] = v; ARRAY_SET_26(a, n, __VA_ARGS__)
#define ARRAY_SET_28(a, n, v, ...)  a[n-28] = v; ARRAY_SET_27(a, n, __VA_ARGS__)
#define ARRAY_SET_29(a, n, v, ...)  a[n-29] = v; ARRAY_SET_28(a, n, __VA_ARGS__)
#define ARRAY_SET_30(a, n, v, ...)  a[n-30] = v; ARRAY_SET_29(a, n, __VA_ARGS__)
#define ARRAY_SET_31(a, n, v, ...)  a[n-31] = v; ARRAY_SET_30(a, n, __VA_ARGS__)
#define ARRAY_SET_32(a, n, v, ...)  a[n-32] = v; ARRAY_SET_31(a, n, __VA_ARGS__)
#define ARRAY_SET_33(a, n, v, ...)  a[n-33] = v; ARRAY_SET_32(a, n, __VA_ARGS__)
#define ARRAY_SET_34(a, n, v, ...)  a[n-34] = v; ARRAY_SET_33(a, n, __VA_ARGS__)
#define ARRAY_SET_35(a, n, v, ...)  a[n-35] = v; ARRAY_SET_34(a, n, __VA_ARGS__)
#define ARRAY_SET_36(a, n, v, ...)  a[n-36] = v; ARRAY_SET_35(a, n, __VA_ARGS__)
#define ARRAY_SET_37(a, n, v, ...)  a[n-37] = v; ARRAY_SET_36(a, n, __VA_ARGS__)
#define ARRAY_SET_38(a, n, v, ...)  a[n-38] = v; ARRAY_SET_37(a, n, __VA_ARGS__)
#define ARRAY_SET_39(a, n, v, ...)  a[n-39] = v; ARRAY_SET_38(a, n, __VA_ARGS__)
#define ARRAY_SET_40(a, n, v, ...)  a[n-40] = v; ARRAY_SET_39(a, n, __VA_ARGS__)
#define ARRAY_SET_41(a, n, v, ...)  a[n-41] = v; ARRAY_SET_40(a, n, __VA_ARGS__)
#define ARRAY_SET_42(a, n, v, ...)  a[n-42] = v; ARRAY_SET_41(a, n, __VA_ARGS__)
#define ARRAY_SET_43(a, n, v, ...)  a[n-43] = v; ARRAY_SET_42(a, n, __VA_ARGS__)
#define ARRAY_SET_44(a, n, v, ...)  a[n-44] = v; ARRAY_SET_43(a, n, __VA_ARGS__)
#define ARRAY_SET_45(a, n, v, ...)  a[n-45] = v; ARRAY_SET_44(a, n, __VA_ARGS__)
#define ARRAY_SET_46(a, n, v, ...)  a[n-46] = v; ARRAY_SET_45(a, n, __VA_ARGS__)
#define ARRAY_SET_47(a, n, v, ...)  a[n-47] = v; ARRAY_SET_46(a, n, __VA_ARGS__)
#define ARRAY_SET_48(a, n, v, ...)  a[n-48] = v; ARRAY_SET_47(a, n, __VA_ARGS__)
#define ARRAY_SET_49(a, n, v, ...)  a[n-49] = v; ARRAY_SET_48(a, n, __VA_ARGS__)
#define ARRAY_SET_50(a, n, v, ...)  a[n-50] = v; ARRAY_SET_49(a, n, __VA_ARGS__)
#define ARRAY_SET_51(a, n, v, ...)  a[n-51] = v; ARRAY_SET_50(a, n, __VA_ARGS__)
#define ARRAY_SET_52(a, n, v, ...)  a[n-52] = v; ARRAY_SET_51(a, n, __VA_ARGS__)
#define ARRAY_SET_53(a, n, v, ...)  a[n-53] = v; ARRAY_SET_52(a, n, __VA_ARGS__)
#define ARRAY_SET_54(a, n, v, ...)  a[n-54] = v; ARRAY_SET_53(a, n, __VA_ARGS__)
#define ARRAY_SET_55(a, n, v, ...)  a[n-55] = v; ARRAY_SET_54(a, n, __VA_ARGS__)
#define ARRAY_SET_56(a, n, v, ...)  a[n-56] = v; ARRAY_SET_55(a, n, __VA_ARGS__)
#define ARRAY_SET_57(a, n, v, ...)  a[n-57] = v; ARRAY_SET_56(a, n, __VA_ARGS__)
#define ARRAY_SET_58(a, n, v, ...)  a[n-58] = v; ARRAY_SET_57(a, n, __VA_ARGS__)
#define ARRAY_SET_59(a, n, v, ...)  a[n-59] = v; ARRAY_SET_58(a, n, __VA_ARGS__)
#define ARRAY_SET_60(a, n, v, ...)  a[n-60] = v; ARRAY_SET_59(a, n, __VA_ARGS__)
#define ARRAY_SET_61(a, n, v, ...)  a[n-61] = v; ARRAY_SET_60(a, n, __VA_ARGS__)
#define ARRAY_SET_62(a, n, v, ...)  a[n-62] = v; ARRAY_SET_61(a, n, __VA_ARGS__)
#define ARRAY_SET_63(a, n, v, ...)  a[n-63] = v; ARRAY_SET_62(a, n, __VA_ARGS__)
#define ARRAY_SET_64(a, n, v, ...)  a[n-64] = v; ARRAY_SET_63(a, n, __VA_ARGS__)

#define ARRAY_APPEND_N(array, count, ...)                           \
    do {                                                            \
        array_data_type *const _base = array_grow_by(array, count); \
        ARRAY_SET_N(_base, count, __VA_ARGS__);                     \
    } while(0)

you can then write your simple function as

void simple_function(array_type *array,
                     const array_data_type a, const array_data_type b,
                     const array_data_type c, const array_data_type d)
{
    ARRAY_APPEND_N(array, 15, a,   b,   a+b, d,   c,
                              c,   c,   c+d, b+d, a,
                              a+c, b+c, c+d, c+d, c);
}

and have it preprocessed into (except for indentation)

void simple_function(array_type *array,
                     const array_data_type a, const array_data_type b,
                     const array_data_type c, const array_data_type d)
{
    do {
        array_data_type *const _base = array_grow_by(array, 15);
        _base[15 - 15] = a;
        _base[15 - 14] = b;
        _base[15 - 13] = a+b;
        _base[15 - 12] = d;
        _base[15 - 11] = c;
        _base[15 - 10] = c;
        _base[15 -  9] = c;
        _base[15 -  8] = c+d;
        _base[15 -  7] = b+d;
        _base[15 -  6] = a;
        _base[15 -  5] = a+c;
        _base[15 -  4] = b+c;
        _base[15 -  3] = c+d;
        _base[15 -  2] = c+d;
        _base[15 -  1] = c;
    } while(0);
}

which generally compiles to excellent machine code on Intel/AMD64 architectures (and other architectures that support relative addressing modes). On some other architectures, it is better to not make _base a constant, and instead autoincrement it (*(_base++) = v;).

If you implement a PP_NARG() macro to count the number of macro arguments, you can add macro

#define ARRAY_APPEND(array, ...) ARRAY_APPEND_N(array, PP_NARG(__VA_ARGS__), __VA_ARGS__)

in which case your function simplifies to

void simple_function(array_type *array,
                     const array_data_type a, const array_data_type b,
                     const array_data_type c, const array_data_type d)
{
    ARRAY_APPEND(array, a,   b,   a+b, d,   c,
                        c,   c,   c+d, b+d, a,
                        a+c, b+c, c+d, c+d, c);
}

The number of preprocessor macro arguments is limited to 64 in some compilers, which limits the maximum number of elements a single macro can add to 62. Depending on the compiler you use, you can extend the macros to support larger number of arguments, but other compilers may choke on those.

Answer 2

Some code refactoring must be done.

First you need a helper function that is similar to the function push_to_array, but this one only allocates new memory for elements:

static inline bool increase_size(struct array_of_a_type *a, size_t size)
{
    const size_t new_size = a->elements + size;
    if (new_size > a->allocated_size) {
        a_type *tmp = realloc(a->array, new_size * sizeof(a_type));
        if (!tmp) {
            return true;
        } else {
            a->array = tmp;
            a->allocated_size = new_size;
        }
    }
    a->elements = new_size;
    return false;
}

Coincidentally, the function push_to_array must be changed to avoid code duplication:

static bool push_to_array(struct array_of_a_type *a, a_type *new_chunk, size_t size)
{
    bool failed = increase_size( a , size );
    if( failed )
    {
        return failed;
    }
    memcpy(a->array + ( a->elements - size ), new_chunk, size * sizeof(a_type));
    return false;
}

The simple_function is now really easy to write, without using a temporary array:

bool simple_function(struct array_of_a_type *my_array, int a, int b, int c, int d)
{
    bool failed = increase_size( my_array , 15 );
    if( failed )
    {
        return failed;
    }

    size_t i = my_array->elements - 15;
    my_array->array[i] = a;
    my_array->array[i+1] = b;
    my_array->array[i+2] = a+b;
    my_array->array[i+3] = d;
    //... write the rest of the assignments
    my_array->array[i+14] = c;

    return false;
}

Answer 3

Are you mad that your simple_function's a_type array is on the stack? That's because you made it an array with the [] which creates it on the stack. You need to make the array like this:

a_type *ap = malloc(<size> * sizeof(a_type));
atype[0] = a;
...

then you can return ap at the end.

Also you'll probably want to push to the array a member at a time, so you can keep the static array, and then do this:

int i;
for (i = 0; i < <size>; i++)
    push_to_array(&my_array, new[i]);

and have your push_to_array function change up a bit.

An implementation of push for a stack can be found here, note that the grow function handles the reallocation: https://github.com/minshallj/my_clib/blob/master/stack.c#L24-L31 You should be able to adjust this to your array "class".

Also, is my_array a global that lives somewhere else in your program? I don't see it declared anywhere.

Answer 4

malloc() a fixed sized array instead of using realloc(). Every time you realloc() it's possible that a copy will happen if realloc() can't grow the existing memory block.

One possible solution is to malloc() a fixed sized array and then double the size of that array when it is full. Then copy the data to the newly doubled array. This will reduce the number of potential copies.

Answer 5

You need to avoid actually making a tmp array. That's only possible if the push routine is inlined into the caller. There's no way to pass a variable-length list through a function call other than in memory. Having that memory be the final array where you want them is only possible if the caller has all the logic needed to safely put them there: i.e. inlined.

Actually, with clang-3.5 on amd64, the temp array is optimized away completely, and simple_function writes directly to the final location at the end of the array. So the call to memcpy never happens. This isn't the case for gcc 4.9.2.

This wouldn't work well here, I think, but you could have a two-stage function, with a check for a common-case (like realloc not needed) that can get inlined, otherwise call the full function.

You could see what kind of asm you get from inlining a variadic function, like bool push_many(struct array_of_a_type *a, size_t size, ...). Nvm, I tried this, and neither gcc nor clang could inline the variadic function. Gcc did generate a custom version of the function, push_many.constprop.0:, but it still looked FAR slower than block-coping the args off the stack where the caller put them.

static bool push_many(struct array_of_a_type *a, size_t size, ...)
{
    va_list ap;
    va_start(ap, size);

    const size_t new_size = a->elements + size;
    if (new_size > a->allocated_size) {
        a_type *tmp = realloc(a->array, new_size * sizeof(a_type));
        if (!tmp) {
            return true;
        } else {
            a->array = tmp;
        a->allocated_size = new_size;
        }
    }

    a_type *p = a->array + a->elements;  // points to spot after last used
    va_start(ap, size);
    for (int i = a->elements; i < new_size; i++) {
        p[i] = va_arg(ap, a_type); /* Increments ap to the next argument. */
    }
    va_end(ap);
//    memcpy(a->array, new_chunk, size * sizeof(a_type));
    a->elements = new_size;
    return false;
}

The copy loop compiled to 14 instructions, including 3 cmovcc conditional moves. Each iteration copies one int. (typedef int a_type;) So gcc 4.9.2 for amd64 is useless at inlining variadic functions, unfortunately. clang-3.5 generates pretty nasty code, too.

Or another approach would be to use macros to get a push inlined into the calling function. GCC or C99 variadic macros won't work, though; you can't iterate over the args in a macro, only pass them to a variadic function like printf. So you'd need the check-for-realloc in every macro invocation, and GCC would have to optimize it away. I'm pretty sure gcc would have a hard time optimizing away all the check-for-space and increment-the-counter operations, though, because the one-at-a-time as written could result in a different number of calls to realloc than batching everything into one. (And that would be a really hard optimization for a compiler to see, I think.)

You could use PUSH4, PUSH8, and PUSH16 macros, that take a large fixed number of args.

Or you could make your code really brittle and have a ALLOC_SPACE_FOR_N_MORE macro/function, and then a simple NOCHECK_PUSH macro that assumed there was enough space, and just incremented the counter. (And hopefully gcc would just do a single add at the end for a sequence of pushes.)

Answer 6

Your push function has a bug. It always overwrites the front of the array.

Using the intermediate new_chunk array makes the compiler work harder. It's better [and faster] to refactor the code to allocate more space in the array and write directly to the array.

The code goes even faster if when doing a realloc a "speculative growth length" is added. This cuts down on the number of realloc calls. That is, when realloc is called, allocated_size and/or new_size is incremented by the growth factor (e.g. 15), so that more space is allocated than currently needed, in anticipation of the next push_to_array. By doing this, the second call can avoid a realloc because it still has enough remaining space.

I've created a test program that shows all this. I've come up with four versions that each show an incremental improvement.

The "best" version is about 2.7x faster

---

The bug:

In your push_to_array function, this is the original code:

    // NOTE/BUG: this is broken -- it always writes the low part of the array
    a->elements = new_size;
    memcpy(a->array,new_chunk,size * sizeof(a_type));

Here is what that code needs to be:

    // NOTE/FIX: this writes to the correct part of the array
    memcpy(a->array + a->elements,new_chunk,size * sizeof(a_type));
    a->elements = new_size;

---

Refactoring:

I've come up with four additional versions:

1. Just [minimally] fixes the bug
2. Similar, uses new_chunk, but allows the growth parameter
3. Uses growth parameter and writes directly to array (i.e. no new_chunk)
4. Similar to 3, but inlines all code

---

Code:

In particular, note the differences between the _fix_push and _fix_array_space functions for speed.

Also, see the SIMPLE_INIT macro that mimics what you did when constructing new_chunk.

// pushary.c -- test program

// pushary.h -- push to array

#ifndef _pushary_h_
#define _pushary_h_

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <string.h>
#include <errno.h>
#include <time.h>
#include <sys/types.h>
#include <sys/wait.h>

#define bool    int
#define true    1
#define false   0

#ifndef MCHECK
#define MCHECK  0
#endif

#ifndef ABIG
#define ABIG    0
#endif

#ifndef NOMACRO
#define NOMACRO 0
#endif

#ifndef NODATA1
#define NODATA1 0
#endif

#ifndef NODATA2
#define NODATA2 0
#endif

#ifndef NODATA3
#define NODATA3 0
#endif

#ifndef NODATA4
#define NODATA4 0
#endif

#define sysfault(_fmt...) \
    do { \
        printf(_fmt); \
        exit(1); \
    } while (0)

#if MCHECK
#include <mcheck.h>
#define MCHECKALL       mcheck_check_all()
#else
#define MCHECKALL       /**/
#endif

#if ABIG
typedef long long a_type;
#else
typedef int a_type;
#endif

// macro for simple_function
// NOTE: different functions could have different macros of this form
#define SIMPLE_INIT(_cmd,_a,_b,_c,_d) \
    _cmd(_a) _cmd(_b) _cmd(_a + _b) _cmd(_d) _cmd(_c) \
    _cmd(_c) _cmd(_c) _cmd(_c + _d) _cmd(_b + _d) _cmd(_a) \
    _cmd(_a + _c) _cmd(_b + _c) _cmd(_c + _d) _cmd(_c + _d) _cmd(_c)

#define _SIZE(_val) \
    + 1

#define _SET(_val) \
    ptr[idx++] = _val;

struct array_of_a_type {
    const char *sym;
    size_t allocated_size;
    size_t grow_size;                   // amount to grow on realloc
    size_t elements;                    // 1-index based
    a_type *array;
    double elap;                        // elapsed time
    double rate;                        // rate
};
typedef struct array_of_a_type a_list;

typedef bool (*simple_p)(a_list *ary,int a,int b,int c,int d);

#if 0
#define INLINE  static inline
#else
#define INLINE  __attribute__((__always_inline__)) static inline
#endif

// test control
typedef struct tstctl tstctl_t;
struct tstctl {
    tstctl_t *tst_next;                 // tstorder linkage
    const char *tst_tag;                // test name
    simple_p tst_proc;                  // simple function
    double tst_bestrat;                 // best ratio
    int tst_bestgrow;                   // best growlen
    double tst_elap;                    // current/best elapsed time
    double tst_rate;                    // current rate
    int tst_trybest;                    // best trial
    a_list tst_lst;                     // array/list
};

// _fix_push -- original push (with bug fix)
INLINE bool
_fix_push(a_list *a,a_type *new_chunk,size_t size)
{
    const size_t new_size = a->elements + size;

    if (new_size > a->allocated_size) {
        a_type *tmp = realloc(a->array,new_size * sizeof(a_type));

        if (!tmp) {
            sysfault("_fix_push: realloc error -- %s\n",strerror(errno));
            return true;
        }
        else {
            a->array = tmp;
            a->allocated_size = new_size;
        }
    }

    // NOTE/FIX: this writes to the correct part of the array
    memcpy(a->array + a->elements,new_chunk,size * sizeof(a_type));
    a->elements = new_size;

    return false;
}

// _fix_array_space -- allocate space in array
// RETURNS: pointer to place to store (or NULL)
INLINE a_type *
_fix_array_space(a_list *a,size_t count)
{
    size_t new_size = a->elements + count;
    size_t newmax;
    a_type *tmp;

    newmax = a->allocated_size;

    if (new_size > newmax) {
        // prevent us from doing realloc on every push
        // NOTE: grow_size is arbitrary -- pick any optimal value
        newmax += new_size;
        newmax += a->grow_size;

        tmp = realloc(a->array,newmax * sizeof(a_type));
        if (tmp == NULL) {
            sysfault("_fix_array_space: realloc error -- %s\n",strerror(errno));
            return tmp;
        }

        a->array = tmp;
        a->allocated_size = newmax;
    }

    tmp = a->array + a->elements;
    a->elements = new_size;

    return tmp;
}

// /home/cae/OBJ/ovrgen/pushary/pushary.proto -- prototypes

// FILE: /home/cae/preserve/ovrbnc/pushary/com.c
// com.c -- common routines

    // fix_array_space -- allocate space in array
    // RETURNS: pointer to place to store (or NULL)
    a_type *
    fix_array_space(a_list *a,size_t count);

    // fix_push -- original push (with bug fix)
    bool
    fix_push(a_list *a,a_type *new_chunk,size_t size);

// FILE: /home/cae/preserve/ovrbnc/pushary/fix1.c
// fix1.c -- push to array
//
// fixes bug in orig

    bool
    fix1_simple(a_list *my_array,int a,int b,int c,int d);

// FILE: /home/cae/preserve/ovrbnc/pushary/fix2.c
// fix2.c -- push to array
//
// uses new_chunk array
// uses push function
// uses grow length

    bool
    fix2_simple(a_list *my_array,int a,int b,int c,int d);

    bool
    fix2_push(a_list *a,a_type *new_chunk,size_t size);

// FILE: /home/cae/preserve/ovrbnc/pushary/fix3.c
// fix3.c -- push to array
//
// uses grow length
// uses non-inline space function

    bool
    fix3_simple(a_list *my_array,int a,int b,int c,int d);

// FILE: /home/cae/preserve/ovrbnc/pushary/fix4.c
// fix4.c -- push to array
//
// uses grow length
// uses inline space function

    bool
    fix4_simple(a_list *my_array,int a,int b,int c,int d);

// FILE: /home/cae/preserve/ovrbnc/pushary/orig.c
// orig.c -- push to array

    bool
    orig_simple(a_list *my_array,int a,int b,int c,int d);

    bool
    orig_push(a_list *a,a_type *new_chunk,size_t size);

// FILE: /home/cae/preserve/ovrbnc/pushary/pushary.c
// pushary.c -- test program

    // main -- main program
    int
    main(int argc,char **argv);

    // usage -- show usage
    void
    usage(void);

    // gendata -- generate data
    void
    gendata(void);

    // defall -- define all tests
    void
    defall(void);

    // defone -- define all tests
    tstctl_t *
    defone(simple_p proc,const char *tag);

    // testall -- test all
    void
    testall(void);

    // testone -- test a function
    void
    testone(tstctl_t *tst);

    // _testone -- test a function
    void
    _testone(tstctl_t *tst,int trycnt,double *elap);

    // ratshow -- show ratio
    void
    ratshow(tstctl_t *tlhs,tstctl_t *trhs,int bestflg);

    // arycmp -- compare arrays
    void
    arycmp(tstctl_t *tlhs,tstctl_t *trhs);

    // arykill -- release an array
    void
    arykill(tstctl_t *tst);

    // tvsecf -- get hi-res time
    double
    tvsecf(void);

#endif

// orig.c -- push to array

bool
orig_simple(a_list *my_array,int a,int b,int c,int d)
{
    a_type new_chunk[] = {
        a, b, a + b, d, c,
        c, c, c + d, b + d, a,
        a + c, b + c, c + d, c + d, c,
    };
    size_t size = sizeof(new_chunk) / sizeof(a_type);

    return orig_push(my_array,new_chunk,size);
}

bool
orig_push(a_list *a,a_type *new_chunk,size_t size)
{
    const size_t new_size = a->elements + size;

    if (new_size > a->allocated_size) {
        a_type *tmp = realloc(a->array,new_size * sizeof(a_type));

        if (!tmp) {
            return true;
        }
        else {
            a->array = tmp;
            a->allocated_size = new_size;
        }
    }

    // NOTE/BUG: this is broken -- it always writes the low part of the array
    a->elements = new_size;
    memcpy(a->array,new_chunk,size * sizeof(a_type));

    return false;
}
// fix1.c -- push to array
//
// fixes bug in orig

bool
fix1_simple(a_list *my_array,int a,int b,int c,int d)
{
    a_type new_chunk[] = {
        a, b, a + b, d, c,
        c, c, c + d, b + d, a,
        a + c, b + c, c + d, c + d, c,
    };
    size_t size = sizeof(new_chunk) / sizeof(a_type);

#if NODATA1 == 0
    return fix_push(my_array,new_chunk,size);
#endif
}
// fix2.c -- push to array
//
// uses new_chunk array
// uses push function
// uses grow length

bool
fix2_simple(a_list *my_array,int a,int b,int c,int d)
{
    a_type new_chunk[] = {
        a, b, a + b, d, c,
        c, c, c + d, b + d, a,
        a + c, b + c, c + d, c + d, c,
    };
    size_t size = sizeof(new_chunk) / sizeof(a_type);

    return fix2_push(my_array,new_chunk,size);
}

bool
fix2_push(a_list *a,a_type *new_chunk,size_t size)
{
    a_type *tmp;

    tmp = fix_array_space(a,size);
    if (tmp == NULL)
        return true;

    // NOTE/FIX: this writes to the correct part of the array
#if NODATA2 == 0
    memcpy(tmp,new_chunk,size * sizeof(a_type));
#endif

    return false;
}
// fix3.c -- push to array
//
// uses grow length
// uses non-inline space function

bool
fix3_simple(a_list *my_array,int a,int b,int c,int d)
{
#if NOMACRO
    size_t count = 15;
#else
    size_t count = SIMPLE_INIT(_SIZE,1,2,3,4);
#endif
    a_type *ptr;

    // use non-inline function
    ptr = fix_array_space(my_array,count);
    if (ptr == NULL)
        return true;

    // NOTE: these optimize to _exactly_ the same code
#if NODATA3 == 0
#if NOMACRO
    ptr[0] = a;
    ptr[1] = b;
    ptr[2] = a + b;
    ptr[3] = d;
    ptr[4] = c;
    ptr[5] = c;
    ptr[6] = c;
    ptr[7] = c + d;
    ptr[8] = b + d;
    ptr[9] = a;
    ptr[10] = a + c;
    ptr[11] = b + c;
    ptr[12] = c + d;
    ptr[13] = c + d;
    ptr[14] = c;
#else
    int idx = 0;
    SIMPLE_INIT(_SET,a,b,c,d)
#endif
#endif

    return false;
}
// fix4.c -- push to array
//
// uses grow length
// uses inline space function

bool
fix4_simple(a_list *my_array,int a,int b,int c,int d)
{
#if NOMACRO
    size_t count = 15;
#else
    size_t count = SIMPLE_INIT(_SIZE,1,2,3,4);
#endif
    a_type *ptr;

    // use inline function
    ptr = _fix_array_space(my_array,count);
    if (ptr == NULL)
        return true;

    // NOTE: these optimize to _exactly_ the same code
#if NODATA4 == 0
#if NOMACRO
    ptr[0] = a;
    ptr[1] = b;
    ptr[2] = a + b;
    ptr[3] = d;
    ptr[4] = c;
    ptr[5] = c;
    ptr[6] = c;
    ptr[7] = c + d;
    ptr[8] = b + d;
    ptr[9] = a;
    ptr[10] = a + c;
    ptr[11] = b + c;
    ptr[12] = c + d;
    ptr[13] = c + d;
    ptr[14] = c;
#else
    int idx = 0;
    SIMPLE_INIT(_SET,a,b,c,d)
#endif
#endif

    return false;
}
// com.c -- common routines

// fix_array_space -- allocate space in array
// RETURNS: pointer to place to store (or NULL)
a_type *
fix_array_space(a_list *a,size_t count)
{

    return _fix_array_space(a,count);
}

// fix_push -- original push (with bug fix)
bool
fix_push(a_list *a,a_type *new_chunk,size_t size)
{

    return _fix_push(a,new_chunk,size);
}

int opt_f;
int opt_m;
int opt_o;
int opt_D;
int opt_M;
int opt_G;
int opt_s;

#define MDFT 1000000

int growlen;
int growmin;
int growmax;
int growbest;

double ratbest;

int datamax;
a_type *testdata;

tstctl_t *orig1;
tstctl_t *fix1;
tstctl_t *fix2;
tstctl_t *fix3;
tstctl_t *fix4;
tstctl_t *orig2;
tstctl_t *orig3;

tstctl_t *tstref;
tstctl_t *tstorder;

// main -- main program
int
main(int argc,char **argv)
{
    char *cp;
    pid_t pid;

    --argc;
    ++argv;

    opt_G = -25;
    opt_f = 1;
    opt_M = MDFT;

    for (;  argc > 0;  --argc, ++argv) {
        cp = *argv;
        if (*cp != '-')
            break;

        switch (cp[1]) {
        case 'f':
            opt_f = ! opt_f;
            break;

        case 'o':
            cp += 2;
            opt_o = (*cp != 0) ? atoi(cp) : 2;
            break;

        case 'D':
            opt_D = ! opt_D;
            break;

        case 'm':
#if MCHECK == 0
            usage();
#endif
            opt_m = ! opt_m;
            break;

        case 'M':
            cp += 2;
            opt_M = atoi(cp);
            break;

        case 'G':
            cp += 2;
            opt_G = (*cp != 0) ? atoi(cp) : 25;
            break;

        case 's':
            cp += 2;
            opt_s = (*cp != 0) ? atoi(cp) : 3;
            break;

        default:
            usage();
            break;
        }
    }

    if (! opt_M)
        opt_M = MDFT;
    printf("M=%d\n",opt_M);
    datamax = opt_M * 4;

    printf("D=%d\n",opt_D);
    gendata();

    if (opt_G < 0) {
        growmin = 0;
        growmax = -opt_G;
    }
    else {
        growmin = opt_G;
        growmax = opt_G;
    }

    growlen = growmin;

    printf("f=%d\n",opt_f);

    if (opt_s <= 0)
        opt_s = 1;
    printf("s=%d\n",opt_s);

    defall();

    for (growlen = growmin;  growlen <= growmax;  ++growlen) {
        if (! opt_f) {
            testall();
            continue;
        }

        fflush(stdout);
        fflush(stderr);

        pid = fork();

        if (pid < 0) {
            perror("fork");
            exit(1);
        }

        if (pid == 0) {
            testall();
            exit(0);
        }

        waitpid(pid,NULL,0);
    }

    return 0;
}

// usage -- show usage
void
usage(void)
{

    printf("  -f -- invert fork mode (DEFAULT: %s)\n",opt_f ? "on" : "off");
    printf("  -D -- use real random test data (DEFAULT: off)\n");
    printf("  -G[grow_length] -- array speculative grow length (DEFAULT: %d)\n",opt_G);
    printf("    <0 -- benchmark mode range\n");
    printf("    >=0 -- single grow length with data compare\n");
    printf("  -M[push_calls] -- number of times to push to array (DEFAULT: %d)\n",
        MDFT);
    printf("  -s<subtrials> -- (DEFAULT: 1)\n");
    printf("  -o<speed reference> -- (DEFAULT: 0)\n");
    printf("    0 -- use fix1\n");
    printf("    1 -- use orig (1st invocation)\n");
    printf("    2 -- use orig (2nd invocation)\n");
    printf("  -m -- force/test mcheck failure%s\n",
        MCHECK ? "" : " (requires rebuild with -DMCHECK=1 and -lmcheck)");

    exit(1);
}

// gendata -- generate data
void
gendata(void)
{
    int *ptr;
    int idx;

    if (opt_D || opt_m) {
        MCHECKALL;
        testdata = malloc(sizeof(a_type) * datamax);

        // force an mcheck exception
        if (opt_m) {
            ptr = testdata;
            ptr -= 10;
            for (idx = 0;  idx < 20;  ++idx)
                ptr[idx] = rand();
        }
        else {
            for (idx = 0;  idx < datamax;  ++idx)
                testdata[idx] = rand();
        }

        MCHECKALL;
    }
}

// defall -- define all tests
void
defall(void)
{

    orig1 = defone(orig_simple,"org1");
    fix1 = defone(fix1_simple,"fix1");
    fix2 = defone(fix2_simple,"fix2");
    fix3 = defone(fix3_simple,"fix3");
    fix4 = defone(fix4_simple,"fix4");
    orig2 = defone(orig_simple,"org2");
    orig3 = defone(orig_simple,"org3");

    switch (opt_o) {
    case 1:
        tstref = orig1;
        break;
    case 2:
        tstref = orig2;
        break;
    default:
        opt_o = 0;
        tstref = fix1;
    }

    printf("reference test is %s\n",tstref->tst_tag);
}

// defone -- define all tests
tstctl_t *
defone(simple_p proc,const char *tag)
{
    tstctl_t *tst;

    tst = calloc(1,sizeof(tstctl_t));
    tst->tst_tag = tag;
    tst->tst_proc = proc;

    tst->tst_bestrat = 0;

    return tst;
}

// testall -- test all
void
testall(void)
{
    tstctl_t *base;
    tstctl_t *trhs;
    tstctl_t *tlhs;

    printf("\n");
    printf("G=%d\n",growlen);

    tstorder = NULL;

    // perform tests
    testone(orig1);
    testone(fix1);
    testone(orig2);
    testone(fix2);
    testone(orig3);
    testone(fix3);
    testone(fix4);

    // show benchmarks
    for (trhs = tstorder;  trhs != NULL;  trhs = trhs->tst_next)
        ratshow(tstref,trhs,1);

#if 0
    do {
        if (opt_o)
            break;

        if (base == fix1)
            break;
        base = fix1;

        ratshow(base,fix2,0);
        ratshow(base,fix3,0);
        ratshow(base,fix4,0);
    } while (0);
#endif

    // compare data
    if (opt_G >= 0) {
        base = fix1;
        for (trhs = tstorder;  trhs != NULL;  trhs = trhs->tst_next)
            arycmp(base,trhs);
    }

    // release all array memory
    for (tlhs = tstorder;  tlhs != NULL;  tlhs = trhs) {
        trhs = tlhs->tst_next;
        arykill(tlhs);
    }
}

// testone -- test a function
void
testone(tstctl_t *tst)
{
    a_list *ary;
    int trycnt;
    double elapv[opt_s];
    tstctl_t *cur;
    tstctl_t *prev;

    tst->tst_elap = 1e20;

    ary = &tst->tst_lst;
    memset(ary,0,sizeof(a_list));

    ary->sym = tst->tst_tag;
    ary->grow_size = growlen;

    for (trycnt = 0;  trycnt < opt_s;  ++trycnt)
        _testone(tst,trycnt,&elapv[trycnt]);

    prev = NULL;
    for (cur = tstorder;  cur != NULL;  cur = cur->tst_next)
        prev = cur;
    if (prev != NULL)
        prev->tst_next = tst;
    else
        tstorder = tst;
}

// _testone -- test a function
void
_testone(tstctl_t *tst,int trycnt,double *elap)
{
    simple_p func;
    double tvbeg;
    double tvdif;
    a_list *ary;

    ary = &tst->tst_lst;
    arykill(tst);

    func = tst->tst_proc;

    MCHECKALL;

    tvbeg = tvsecf();

    // use real test data -- good for comparisons
    if (opt_D) {
        a_type *ptr = testdata;
        a_type *ptre = ptr + datamax;
        for (;  ptr < ptre;  ptr += 4)
            func(ary,ptr[0],ptr[1],ptr[2],ptr[3]);
    }

    // use the same test data -- faster and gives truer benchmark for function
    // being tested
    else {
        for (int loopcnt = datamax;  loopcnt > 0;  loopcnt -= 4)
            func(ary,1,2,3,4);
    }

    tvdif = tvsecf();
    tvdif -= tvbeg;

    MCHECKALL;

    ary->elap = tvdif;
    ary->rate = ary->elements;
    ary->rate /= tvdif;

    if (ary->elap < tst->tst_elap) {
        tst->tst_elap = ary->elap;
        tst->tst_rate = ary->rate;
        tst->tst_trybest = trycnt;
    }

    *elap = tvdif;
}

// ratshow -- show ratio
void
ratshow(tstctl_t *tlhs,tstctl_t *trhs,int bestflg)
{
    double ratio;
    double rhsrate;
    double lhsrate;
    int faster;

    printf("%s %.9f",trhs->tst_tag,trhs->tst_elap);

    lhsrate = tlhs->tst_rate;
    rhsrate = trhs->tst_rate;

    faster = (rhsrate > lhsrate);

    if (faster)
        ratio = rhsrate / lhsrate;
    else
        ratio = lhsrate / rhsrate;

    if (tlhs != trhs)
        printf(" is %.3fx %s",
            ratio,faster ? "faster" : "slower");

    do {
        if (! bestflg)
            break;

        if (! faster)
            ratio = -ratio;

        if (ratio <= trhs->tst_bestrat)
            break;

        trhs->tst_bestrat = ratio;
        trhs->tst_bestgrow = growlen;

        //printf(" BETTER (G=%d)",growlen);
    } while (0);

    printf("\n");
}

// arycmp -- compare arrays
void
arycmp(tstctl_t *tlhs,tstctl_t *trhs)
{
    a_list *alhs = &tlhs->tst_lst;
    a_list *arhs = &trhs->tst_lst;
    a_type lhs;
    a_type rhs;
    int matchflg;

    do {
        if (alhs->array == NULL)
            break;
        if (arhs->array == NULL)
            break;

        if (alhs->elements != arhs->elements) {
            printf("arycmp: count mismatch -- %s=%lu %s=%lu\n",
                alhs->sym,alhs->elements,arhs->sym,arhs->elements);
            break;
        }

        matchflg = 1;
        for (size_t idx = 0;  idx < alhs->elements;  ++idx) {
            lhs = alhs->array[idx];
            rhs = arhs->array[idx];
            if (lhs != rhs) {
                printf("arycmp: data mismatch -- idx=%lu %s=%d %s=%d\n",
                    idx,alhs->sym,lhs,arhs->sym,rhs);
                matchflg = 0;
                break;
            }
        }

        if (matchflg)
            printf("%s: MATCH\n",arhs->sym);
    } while (0);
}

// arykill -- release an array
void
arykill(tstctl_t *tst)
{
    a_list *ary;

    ary = &tst->tst_lst;

    if (ary->array != NULL) {
        MCHECKALL;
        free(ary->array);
        MCHECKALL;
    }

    ary->array = NULL;

    ary->allocated_size = 0;
    ary->elements = 0;
}

// tvsecf -- get hi-res time
double
tvsecf(void)
{
    struct timespec ts;
    double sec;

    clock_gettime(CLOCK_REALTIME,&ts);
    sec = ts.tv_nsec;
    sec /= 1e9;
    sec += ts.tv_sec;

    return sec;
}

---

Benchmarks:

Benchmark output for the various methods. It's too big to fit in this answer, so I'll be posting it in a second one below

Answer 7

NOTE: This answer is a continuation of my original answer because of space limitations (i.e. if upvoting, pls use original above: https://stackoverflow.com/a/39562827/5382650)

---

Benchmarks:

Because of the bug in the original code, it has skewed benchmark results because it's artificially getting better cache performance because it's not traversing the entire array as it should. That is, it appears to perform better than it actually would if the bug were fixed.

So, using fix1 would be a better indicator for baseline performance and that is what the data below uses. The original [without fix] produces 0.02 but fix1 gives 0.06, which I believe is the more correct number.

The growth factor is a tuning parameter and with the "best" value, the fix4 version is a 2.7x improvement. That is what I believe is the most trustworthy result.

However, there is an anomaly in the data that I can't explain despite considerable unit tests, longer tests, application of mcheck(3), etc. I retained the original algorithm [with the bug] as part of the test. If the original is the first test run, or is run just after fix1, it produces the "skewed" results.

However, if the original is run after fix2, fix3, or fix4, sometimes it produces 10x worse performance relative to itself. The growth value is not used by the original. But, the original's behavior seems to depend upon the speculative growth factor used by the earlier algorithm.

Sometimes, the original in the "dicey" slot gives the skewed/artificially low value (approx. 0.02). When it goes "haywire", it runs 10x slower (approx. 0.2).

There seems to be a "run of bad luck" and "good luck" with this. If the -G option were given a different value (e.g. -G-300 which will test all growth values between 0 and 300), there are multiple runs present.

I believe that the erratic results aren't relevant, but I've kept them in anyway. It could just be noise i.e. the values are okay, and fluctuate due to some internals within the memory reallocator that causes it to do more internal free block split/merge, etc.

AFAICT, it is not due to overruns of the realloc area because the program has a mode for doing mcheck and that mode gives everything a clean bill of health.

M=1000000
D=0
f=1
s=1
reference test is fix1

G=0
org1 0.028413773 is 2.462x faster
fix1 0.069955111
org2 0.035362244 is 1.978x faster
fix2 0.032926321 is 2.125x faster
org3 0.268535376 is 3.839x slower
fix3 0.026652813 is 2.625x faster
fix4 0.025245905 is 2.771x faster

G=1
org1 0.027498960 is 2.517x faster
fix1 0.069201946
org2 0.033916712 is 2.040x faster
fix2 0.031118631 is 2.224x faster
org3 0.264514446 is 3.822x slower
fix3 0.026646614 is 2.597x faster
fix4 0.025324345 is 2.733x faster

G=2
org1 0.026978731 is 2.496x faster
fix1 0.067343950
org2 0.034334421 is 1.961x faster
fix2 0.031268835 is 2.154x faster
org3 0.266630888 is 3.959x slower
fix3 0.026658535 is 2.526x faster
fix4 0.025254488 is 2.667x faster

G=3
org1 0.027746677 is 2.495x faster
fix1 0.069227457
org2 0.033862829 is 2.044x faster
fix2 0.031069279 is 2.228x faster
org3 0.287544250 is 4.154x slower
fix3 0.026713371 is 2.591x faster
fix4 0.025189638 is 2.748x faster

G=4
org1 0.027034283 is 2.527x faster
fix1 0.068307161
org2 0.033991575 is 2.010x faster
fix2 0.031272411 is 2.184x faster
org3 0.311707735 is 4.563x slower
fix3 0.026990414 is 2.531x faster
fix4 0.025078297 is 2.724x faster

G=5
org1 0.027446985 is 2.429x faster
fix1 0.066675663
org2 0.033823967 is 1.971x faster
fix2 0.031498909 is 2.117x faster
org3 0.331423283 is 4.971x slower
fix3 0.026667356 is 2.500x faster
fix4 0.026413918 is 2.524x faster

G=6
org1 0.027255535 is 2.428x faster
fix1 0.066179037
org2 0.033841848 is 1.956x faster
fix2 0.031159401 is 2.124x faster
org3 0.335711241 is 5.073x slower
fix3 0.026690722 is 2.479x faster
fix4 0.025039911 is 2.643x faster

G=7
org1 0.027280807 is 2.440x faster
fix1 0.066556692
org2 0.034326553 is 1.939x faster
fix2 0.031259060 is 2.129x faster
org3 0.331621408 is 4.983x slower
fix3 0.026686430 is 2.494x faster
fix4 0.025387526 is 2.622x faster

G=8
org1 0.027087212 is 2.453x faster
fix1 0.066447973
org2 0.033598185 is 1.978x faster
fix2 0.031176090 is 2.131x faster
org3 0.034165382 is 1.945x faster
fix3 0.026757479 is 2.483x faster
fix4 0.025131702 is 2.644x faster

G=9
org1 0.027328253 is 2.451x faster
fix1 0.066978931
org2 0.034043789 is 1.967x faster
fix2 0.031486034 is 2.127x faster
org3 0.033723354 is 1.986x faster
fix3 0.027368069 is 2.447x faster
fix4 0.025647879 is 2.611x faster

G=10
org1 0.027052402 is 2.458x faster
fix1 0.066498756
org2 0.033848524 is 1.965x faster
fix2 0.031741381 is 2.095x faster
org3 0.033836603 is 1.965x faster
fix3 0.027002096 is 2.463x faster
fix4 0.025351524 is 2.623x faster

G=11
org1 0.027157784 is 2.471x faster
fix1 0.067117691
org2 0.033848047 is 1.983x faster
fix2 0.031594038 is 2.124x faster
org3 0.034133911 is 1.966x faster
fix3 0.027194977 is 2.468x faster
fix4 0.025204659 is 2.663x faster

G=12
org1 0.027328730 is 2.432x faster
fix1 0.066454649
org2 0.033915043 is 1.959x faster
fix2 0.031331778 is 2.121x faster
org3 0.033701420 is 1.972x faster
fix3 0.026796579 is 2.480x faster
fix4 0.025482893 is 2.608x faster

G=13
org1 0.027091503 is 2.520x faster
fix1 0.068269968
org2 0.033600807 is 2.032x faster
fix2 0.031302691 is 2.181x faster
org3 0.034220219 is 1.995x faster
fix3 0.026732683 is 2.554x faster
fix4 0.025168657 is 2.712x faster

G=14
org1 0.027466774 is 2.403x faster
fix1 0.065990925
org2 0.034015417 is 1.940x faster
fix2 0.031306028 is 2.108x faster
org3 0.033681631 is 1.959x faster
fix3 0.026975870 is 2.446x faster
fix4 0.025142908 is 2.625x faster

G=15
org1 0.030098915 is 2.202x faster
fix1 0.066287756
org2 0.033817768 is 1.960x faster
fix2 0.031510592 is 2.104x faster
org3 0.264448166 is 3.989x slower
fix3 0.026585102 is 2.493x faster
fix4 0.025573254 is 2.592x faster

G=16
org1 0.029087305 is 2.289x faster
fix1 0.066566944
org2 0.034010649 is 1.957x faster
fix2 0.032317400 is 2.060x faster
org3 0.269736767 is 4.052x slower
fix3 0.026986122 is 2.467x faster
fix4 0.025726795 is 2.587x faster

G=17
org1 0.027568817 is 2.418x faster
fix1 0.066652775
org2 0.033725500 is 1.976x faster
fix2 0.031077385 is 2.145x faster
org3 0.270752668 is 4.062x slower
fix3 0.028372288 is 2.349x faster
fix4 0.026800632 is 2.487x faster

G=18
org1 0.028200626 is 2.466x faster
fix1 0.069550514
org2 0.035360813 is 1.967x faster
fix2 0.033010244 is 2.107x faster
org3 0.308327198 is 4.433x slower
fix3 0.028569698 is 2.434x faster
fix4 0.028189659 is 2.467x faster

G=19
org1 0.028352022 is 2.457x faster
fix1 0.069663048
org2 0.035186291 is 1.980x faster
fix2 0.033131599 is 2.103x faster
org3 0.302445412 is 4.342x slower
fix3 0.028528690 is 2.442x faster
fix4 0.026380062 is 2.641x faster

G=20
org1 0.028351307 is 2.449x faster
fix1 0.069445372
org2 0.035343409 is 1.965x faster
fix2 0.032827139 is 2.115x faster
org3 0.333808899 is 4.807x slower
fix3 0.028279066 is 2.456x faster
fix4 0.026592016 is 2.612x faster

G=21
org1 0.028333902 is 2.457x faster
fix1 0.069613457
org2 0.035215616 is 1.977x faster
fix2 0.033250570 is 2.094x faster
org3 0.326132298 is 4.685x slower
fix3 0.026517391 is 2.625x faster
fix4 0.025246382 is 2.757x faster

G=22
org1 0.027449369 is 2.421x faster
fix1 0.066462278
org2 0.033666849 is 1.974x faster
fix2 0.031057119 is 2.140x faster
org3 0.332618952 is 5.005x slower
fix3 0.028064966 is 2.368x faster
fix4 0.026383400 is 2.519x faster

G=23
org1 0.028641462 is 2.444x faster
fix1 0.070001602
org2 0.035483837 is 1.973x faster
fix2 0.033087969 is 2.116x faster
org3 0.342431068 is 4.892x slower
fix3 0.028344154 is 2.470x faster
fix4 0.026709557 is 2.621x faster

G=24
org1 0.028158426 is 2.468x faster
fix1 0.069482327
org2 0.035173178 is 1.975x faster
fix2 0.033740997 is 2.059x faster
org3 0.346288681 is 4.984x slower
fix3 0.028279781 is 2.457x faster
fix4 0.027346849 is 2.541x faster

G=25
org1 0.028361082 is 2.469x faster
fix1 0.070035458
org2 0.035205841 is 1.989x faster
fix2 0.032957315 is 2.125x faster
org3 0.035385132 is 1.979x faster
fix3 0.028091431 is 2.493x faster
fix4 0.026364803 is 2.656x faster