Efficient integer compare function

Question

The compare function is a function that takes two arguments a and b and returns an integer describing their order. If a is smaller than b, the result is some negative integer. If a is bigger than b, the result is some positive integer. Otherwise, a and b are equal, and the result is zero.

This function is often used to parameterize sorting and searching algorithms from standard libraries.

Implementing the compare function for characters is quite easy; you simply subtract the arguments:

int compare_char(char a, char b)
{
    return a - b;
}

This works because the difference between two characters is generally assumed to fit into an integer. (Note that this assumption does not hold for systems where sizeof(char) == sizeof(int).)

This trick cannot work to compare integers, because the difference between two integers generally does not fit into an integer. For example, INT_MAX - (-1) = INT_MIN suggests that INT_MAX is smaller than -1 (technically, the overflow leads to undefined behavior, but let's assume modulo arithmetic).

So how can we implement the compare function efficiently for integers? Here is my first attempt:

int compare_int(int a, int b)
{
    int temp;
    int result;
    __asm__ __volatile__ (
        "cmp %3, %2 \n\t"
        "mov $0, %1 \n\t"

        "mov $1, %0 \n\t"
        "cmovg %0, %1 \n\t"

        "mov $-1, %0 \n\t"
        "cmovl %0, %1 \n\t"
    : "=r"(temp), "=r"(result)
    : "r"(a), "r"(b)
    : "cc");
    return result;
}

Can it be done in less than 6 instructions? Is there a less straightforward way that is more efficient?

Answer 1

The following has always proven to be fairly efficient for me:

return (a < b) ? -1 : (a > b);

With gcc -O2 -S, this compiles down to the following five instructions:

xorl    %edx, %edx
cmpl    %esi, %edi
movl    $-1, %eax
setg    %dl
cmovge  %edx, %eax

---

As a follow-up to Ambroz Bizjak's excellent companion answer, I was not convinced that his program tested the same assembly code what was posted above. And, when I was studying the compiler output more closely, I noticed that the compiler was not generating the same instructions as was posted in either of our answers. So, I took his test program, hand modified the assembly output to match what we posted, and compared the resulting times. It seems the two versions compare roughly identically.

./opt_cmp_branchless: 0m1.070s
./opt_cmp_branch:     0m1.037s

I am posting the assembly of each program in full so that others may attempt the same experiment, and confirm or contradict my observation.

The following is the version with the cmovge instruction ((a < b) ? -1 : (a > b)):

        .file   "cmp.c"
        .text
        .section        .rodata.str1.1,"aMS",@progbits,1
.LC0:
        .string "%d=0\n"
        .text
        .p2align 4,,15
.globl main
        .type   main, @function
main:
.LFB20:
        .cfi_startproc
        pushq   %rbp
        .cfi_def_cfa_offset 16
        .cfi_offset 6, -16
        pushq   %rbx
        .cfi_def_cfa_offset 24
        .cfi_offset 3, -24
        movl    $arr.2789, %ebx
        subq    $8, %rsp
        .cfi_def_cfa_offset 32
.L9:
        leaq    4(%rbx), %rbp
.L10:
        call    rand
        movb    %al, (%rbx)
        addq    $1, %rbx
        cmpq    %rbx, %rbp
        jne     .L10
        cmpq    $arr.2789+4096, %rbp
        jne     .L9
        xorl    %r8d, %r8d
        xorl    %esi, %esi
        orl     $-1, %edi
.L12:
        xorl    %ebp, %ebp
        .p2align 4,,10
        .p2align 3
.L18:
        movl    arr.2789(%rbp), %ecx
        xorl    %eax, %eax
        .p2align 4,,10
        .p2align 3
.L15:
        movl    arr.2789(%rax), %edx
        xorl    %ebx, %ebx
        cmpl    %ecx, %edx
        movl    $-1, %edx
        setg    %bl
        cmovge  %ebx, %edx
        addq    $4, %rax
        addl    %edx, %esi
        cmpq    $4096, %rax
        jne     .L15
        addq    $4, %rbp
        cmpq    $4096, %rbp
        jne     .L18
        addl    $1, %r8d
        cmpl    $500, %r8d
        jne     .L12
        movl    $.LC0, %edi
        xorl    %eax, %eax
        call    printf
        addq    $8, %rsp
        .cfi_def_cfa_offset 24
        xorl    %eax, %eax
        popq    %rbx
        .cfi_def_cfa_offset 16
        popq    %rbp
        .cfi_def_cfa_offset 8
        ret
        .cfi_endproc
.LFE20:
        .size   main, .-main
        .local  arr.2789
        .comm   arr.2789,4096,32
        .section        .note.GNU-stack,"",@progbits

The version below uses the branchless method ((a > b) - (a < b)):

        .file   "cmp.c"
        .text
        .section        .rodata.str1.1,"aMS",@progbits,1
.LC0:
        .string "%d=0\n"
        .text
        .p2align 4,,15
.globl main
        .type   main, @function
main:
.LFB20:
        .cfi_startproc
        pushq   %rbp
        .cfi_def_cfa_offset 16
        .cfi_offset 6, -16
        pushq   %rbx
        .cfi_def_cfa_offset 24
        .cfi_offset 3, -24
        movl    $arr.2789, %ebx
        subq    $8, %rsp
        .cfi_def_cfa_offset 32
.L9:
        leaq    4(%rbx), %rbp
.L10:
        call    rand
        movb    %al, (%rbx)
        addq    $1, %rbx
        cmpq    %rbx, %rbp
        jne     .L10
        cmpq    $arr.2789+4096, %rbp
        jne     .L9
        xorl    %r8d, %r8d
        xorl    %esi, %esi
.L19:
        movl    %ebp, %ebx
        xorl    %edi, %edi
        .p2align 4,,10
        .p2align 3
.L24:
        movl    %ebp, %ecx
        xorl    %eax, %eax
        jmp     .L22
        .p2align 4,,10
        .p2align 3
.L20:
        movl    arr.2789(%rax), %ecx
.L22:
        xorl    %edx, %edx
        cmpl    %ebx, %ecx
        setg    %cl
        setl    %dl
        movzbl  %cl, %ecx
        subl    %ecx, %edx
        addl    %edx, %esi
        addq    $4, %rax
        cmpq    $4096, %rax
        jne     .L20
        addq    $4, %rdi
        cmpq    $4096, %rdi
        je      .L21
        movl    arr.2789(%rdi), %ebx
        jmp     .L24
.L21:
        addl    $1, %r8d
        cmpl    $500, %r8d
        jne     .L19
        movl    $.LC0, %edi
        xorl    %eax, %eax
        call    printf
        addq    $8, %rsp
        .cfi_def_cfa_offset 24
        xorl    %eax, %eax
        popq    %rbx
        .cfi_def_cfa_offset 16
        popq    %rbp
        .cfi_def_cfa_offset 8
        ret
        .cfi_endproc
.LFE20:
        .size   main, .-main
        .local  arr.2789
        .comm   arr.2789,4096,32
        .section        .note.GNU-stack,"",@progbits

Answer 2

This code has no branches and uses 5 instructions. It may outperform other branch-less alternatives on recent Intel processors, where cmov* instructions are quite expensive. Disadvantage is non-symmetrical return value (INT_MIN+1, 0, 1).

int compare_int (int a, int b)
{
    int res;

    __asm__ __volatile__ (
        "xor %0, %0 \n\t"
        "cmpl %2, %1 \n\t"
        "setl %b0 \n\t"
        "rorl $1, %0 \n\t"
        "setnz %b0 \n\t"
    : "=q"(res)
    : "r"(a)
    , "r"(b)
    : "cc"
    );

    return res;
}

This variant does not need initialization, so it uses only 4 instructions:

int compare_int (int a, int b)
{
    __asm__ __volatile__ (
        "subl %1, %0 \n\t"
        "setl %b0 \n\t"
        "rorl $1, %0 \n\t"
        "setnz %b0 \n\t"
    : "+q"(a)
    : "r"(b)
    : "cc"
    );

    return a;
}

Answer 3

For what it's worth I put together an SSE2 implementation. vec_compare1 uses the same approach as compare2 but requires just three SSE2 arithmetic instructions:

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

#define COUNT 1024
#define LOOPS 500
#define COMPARE vec_compare1
#define USE_RAND 1

int arr[COUNT] __attribute__ ((aligned(16)));

typedef __m128i vSInt32;

vSInt32 vec_compare1 (vSInt32 va, vSInt32 vb)
{
    vSInt32 vcmp1 = _mm_cmpgt_epi32(va, vb);
    vSInt32 vcmp2 = _mm_cmpgt_epi32(vb, va);
    return _mm_sub_epi32(vcmp2, vcmp1);
}

int main ()
{
    for (int i = 0; i < COUNT; i++) {
#if USE_RAND
        arr[i] = rand();
#else
        for (int b = 0; b < sizeof(arr[i]); b++) {
            *((unsigned char *)&arr[i] + b) = rand();
        }
#endif
    }

    vSInt32 vsum = _mm_set1_epi32(0);

    for (int l = 0; l < LOOPS; l++) {
        for (int i = 0; i < COUNT; i++) {
            for (int j = 0; j < COUNT; j+=4) {
                vSInt32 v1 = _mm_loadu_si128(&arr[i]);
                vSInt32 v2 = _mm_load_si128(&arr[j]);
                vSInt32 v = COMPARE(v1, v2);
                vsum = _mm_add_epi32(vsum, v);
            }
        }
    }

    printf("vsum = %vd\n", vsum);

    return 0;
}

Time for this is 0.137s.

Time for compare2 with the same CPU and compiler is 0.674s.

So the SSE2 implementation is around 4x faster, as might be expected (since it's 4-wide SIMD).

Answer 4

This one has no branches, and doesn't suffer from overflow or underflow:

return (a > b) - (a < b);

With gcc -O2 -S, this compiles down to the following six instructions:

xorl    %eax, %eax
cmpl    %esi, %edi
setl    %dl
setg    %al
movzbl  %dl, %edx
subl    %edx, %eax

Here's some code to benchmark various compare implementations:

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

#define COUNT 1024
#define LOOPS 500
#define COMPARE compare2
#define USE_RAND 1

int arr[COUNT];

int compare1 (int a, int b)
{
    if (a < b) return -1;
    if (a > b) return 1;
    return 0;
}

int compare2 (int a, int b)
{
    return (a > b) - (a < b);
}

int compare3 (int a, int b)
{
    return (a < b) ? -1 : (a > b);
}

int compare4 (int a, int b)
{
    __asm__ __volatile__ (
        "sub %1, %0 \n\t"
        "jno 1f \n\t"
        "cmc \n\t"
        "rcr %0 \n\t"
        "1: "
    : "+r"(a)
    : "r"(b)
    : "cc");
    return a;
}

int main ()
{
    for (int i = 0; i < COUNT; i++) {
#if USE_RAND
        arr[i] = rand();
#else
        for (int b = 0; b < sizeof(arr[i]); b++) {
            *((unsigned char *)&arr[i] + b) = rand();
        }
#endif
    }

    int sum = 0;

    for (int l = 0; l < LOOPS; l++) {
        for (int i = 0; i < COUNT; i++) {
            for (int j = 0; j < COUNT; j++) {
                sum += COMPARE(arr[i], arr[j]);
            }
        }
    }

    printf("%d=0\n", sum);

    return 0;
}

The results on my 64-bit system, compiled with gcc -std=c99 -O2, for positive integers (USE_RAND=1):

compare1: 0m1.118s
compare2: 0m0.756s
compare3: 0m1.101s
compare4: 0m0.561s

Out of C-only solutions, the one I suggested was the fastest. user315052's solution was slower despite compiling to only 5 instructions. The slowdown is likely because, despite having one less instruction, there is a conditional instruction (cmovge).

Overall, FredOverflow's 4-instruction assembly implementation was the fastest when used with positive integers. However, this code only benchmarked the integer range RAND_MAX, so the 4-instuction test is biased, because it handles overflows separately, and these don't occur in the test; the speed may be due to successful branch prediction.

With a full range of integers (USE_RAND=0), the 4-instruction solution is in fact very slow (others are the same):

compare4: 0m1.897s

Answer 5

Okay, I managed to get it down to four instructions :) The basic idea is as follows:

Half the time, the difference is small enough to fit into an integer. In that case, just return the difference. Otherwise, shift the number one to the right. The crucial question is what bit to shift into the MSB then.

Let's look at two extreme examples, using 8 bits instead of 32 bits for the sake of simplicity:

 10000000 INT_MIN
 01111111 INT_MAX
---------
000000001 difference
 00000000 shifted

 01111111 INT_MAX
 10000000 INT_MIN
---------
111111111 difference
 11111111 shifted

Shifting the carry bit in would yield 0 for the first case (although INT_MIN is not equal to INT_MAX) and some negative number for the second case (although INT_MAX is not smaller than INT_MIN).

But if we flip the carry bit before doing the shift, we get sensible numbers:

 10000000 INT_MIN
 01111111 INT_MAX
---------
000000001 difference
100000001 carry flipped
 10000000 shifted

 01111111 INT_MAX
 10000000 INT_MIN
---------
111111111 difference
011111111 carry flipped
 01111111 shifted

I'm sure there's a deep mathematical reason why it makes sense to flip the carry bit, but I don't see it yet.

int compare_int(int a, int b)
{
    __asm__ __volatile__ (
        "sub %1, %0 \n\t"
        "jno 1f \n\t"
        "cmc \n\t"
        "rcr %0 \n\t"
        "1: "
    : "+r"(a)
    : "r"(b)
    : "cc");
    return a;
}

I have tested the code with one million random inputs plus every combination of INT_MIN, -INT_MAX, INT_MIN/2, -1, 0, 1, INT_MAX/2, INT_MAX/2+1, INT_MAX. All tests passed. Can you proove me wrong?

Answer 6

Maybe you can use the following idea (in pseudo-code; didn't write asm-code because i am not comfortable with syntax):

1. Subtract the numbers (result = a - b)
2. If no overflow, done (jo instruction and branch prediction should work very well here)
3. If there was overflow, use any robust method (return (a < b) ? -1 : (a > b))

Edit: for additional simplicity: if there was overflow, flip the sign of the result, instead of step 3.

Answer 7

You could consider promoting the integers to 64bit values.