Efficient 128-bit addition using carry flag

I'm using a 128 bit开发者_Python百科 integer counter in the very inner loops of my C++ code. (Irrelevant background: The actual application is evaluating finite difference equations on a regular grid, which involves repetitively incrementing large integers, and even 64 bits isn't enough precision because small rounding accumulates enough to affect the answers.)

I've represented the integer as two 64 bit unsigned longs. I now need to increment those values by a 128 bit constant. This isn't hard, but you have to manually catch the carry from the low word to the high word.

I have working code something like this:

inline void increment128(unsigned long &hiWord, unsigned long &loWord)
  {
    const unsigned long hiAdd=0x0000062DE49B5241;
    const unsigned long loAdd=0x85DC198BCDD714BA;

    loWord += loAdd;
    if (loWord < loAdd) ++hiWord; // test_and_add_carry
    hiWord += hiAdd;
  }

This is tight and simple code. It works.

Unfortunately this is about 20% of my runtime. The killer line is that loWord test. If I remove it, I obviously get the wrong answers but the runtime overhead drops from 20% to 4%! So that carry test is especially expensive!

My question: Does C++ expose the hardware carry flags, even as an extension to GCC? It seems like the additions could be done without the test-and-add-carry line above if the actual compiled instructions used an add using last carry instruction for the hiWord addition. Is there a way to rewrite the test-and-add-carry line to get the compiler to use the intrinsic opcode?

Actually gcc will use the carry automatically if you write your code carefully...

Current GCC can optimize hiWord += (loWord < loAdd); into add/adc (x86's add-with-carry). This optimization was introduced in GCC5.3.

• With separate uint64_t chunks in 64-bit mode: https://godbolt.org/z/S2kGRz.
• And the same thing in 32-bit mode with uint32_t chunks: https://godbolt.org/z/9FC9vc

(editor's note: Of course the hard part is writing a correct full-adder with carry in and carry out; that's hard in C and GCC doesn't know how to optimize any that I've seen.)

Also related: https://gcc.gnu.org/onlinedocs/gcc/Integer-Overflow-Builtins.html can give you carry-out from unsigned, or signed-overflow detection.

Older GCC, like GCC4.5, will branch or setc on the carry-out from an add, instead of using adc, and only used adc (add-with-carry) on the flag-result from an add if you used __int128. (Or uint64_t on a 32-bit target). See Is there a 128 bit integer in gcc? - only on 64-bit targets, supported since GCC4.1.

I compiled this code with gcc -O2 -Wall -Werror -S:

void increment128_1(unsigned long &hiWord, unsigned long &loWord)
{
    const unsigned long hiAdd=0x0000062DE49B5241;
    const unsigned long loAdd=0x85DC198BCDD714BA;

    loWord += loAdd;
    if (loWord < loAdd) ++hiWord; // test_and_add_carry                                                                                                             
    hiWord += hiAdd;
}

void increment128_2(unsigned long &hiWord, unsigned long &loWord)
{
    const unsigned long hiAdd=0x0000062DE49B5241;
    const unsigned long loAdd=0x85DC198BCDD714BA;

    loWord += loAdd;
    hiWord += hiAdd;
    hiWord += (loWord < loAdd); // test_and_add_carry                                                                                                               
}

This is the assembly for increment128_1:

.cfi_startproc
        movabsq     $-8801131483544218438, %rax
        addq        (%rsi), %rax
        movabsq     $-8801131483544218439, %rdx
        cmpq        %rdx, %rax
        movq        %rax, (%rsi)
        ja  .L5
        movq        (%rdi), %rax
        addq        $1, %rax
.L3:
        movabsq     $6794178679361, %rdx
        addq        %rdx, %rax
        movq        %rax, (%rdi)
        ret

...and this is the assembly for increment128_2:

        movabsq     $-8801131483544218438, %rax
        addq        %rax, (%rsi)
        movabsq     $6794178679361, %rax
        addq        (%rdi), %rax
        movabsq     $-8801131483544218439, %rdx
        movq        %rax, (%rdi)
        cmpq        %rdx, (%rsi)
        setbe       %dl
        movzbl      %dl, %edx
        leaq        (%rdx,%rax), %rax
        movq        %rax, (%rdi)
        ret

Note the lack of conditional branches in the second version.

[edit]

Also, references are often bad for performance, because GCC has to worry about aliasing... It is often better to just pass things by value. Consider:

struct my_uint128_t {
    unsigned long hi;
    unsigned long lo;
};

my_uint128_t increment128_3(my_uint128_t x)
{
    const unsigned long hiAdd=0x0000062DE49B5241;
    const unsigned long loAdd=0x85DC198BCDD714BA;

    x.lo += loAdd;
    x.hi += hiAdd + (x.lo < loAdd);
    return x;
}

Assembly:

        .cfi_startproc
        movabsq     $-8801131483544218438, %rdx
        movabsq     $-8801131483544218439, %rax
        movabsq     $6794178679362, %rcx
        addq        %rsi, %rdx
        cmpq        %rdx, %rax
        sbbq        %rax, %rax
        addq        %rcx, %rax
        addq        %rdi, %rax
        ret

This is actually the tightest code of the three.

...OK so none of them actually used the carry automatically :-). But they do avoid the conditional branch, which I bet is the slow part (since the branch prediction logic will get it wrong half the time).

[edit 2]

And one more, which I stumbled across doing a little searching. Did you know GCC has built-in support for 128-bit integers?

typedef unsigned long my_uint128_t __attribute__ ((mode(TI)));

my_uint128_t increment128_4(my_uint128_t x)
{
    const my_uint128_t hiAdd=0x0000062DE49B5241;
    const unsigned long loAdd=0x85DC198BCDD714BA;

    return x + (hiAdd << 64) + loAdd;
}

The assembly for this one is about as good as it gets:

        .cfi_startproc
        movabsq     $-8801131483544218438, %rax
        movabsq     $6794178679361, %rdx
        pushq       %rbx
        .cfi_def_cfa_offset 16
        addq        %rdi, %rax
        adcq        %rsi, %rdx
        popq        %rbx
        .cfi_offset 3, -16
        .cfi_def_cfa_offset 8
        ret

(Not sure where the push/pop of ebx came from, but this is still not bad.)

All of these are with GCC 4.5.2, by the way.

The best answer, of course, is to use the built-in __int128_t support.

Alternatively, use an inline asm. I prefer to use the named-argument form:

__asm("add %[src_lo], %[dst_lo]\n"
      "adc %[src_hi], %[dst_hi]"
      : [dst_lo] "+&r" (loWord), [dst_hi] "+r" (hiWord)
      : [src_lo] "erm" (loAdd), [src_hi] "erm" (hiAdd)
      : );

loWord is flagged as an early clobber operand, because it's written before some of the other operands are read. This avoids wrong code for hiAdd = loWord, because it will stop gcc from using the same register to hold both. It does stop the compiler from using the same register for the loAdd = loWord case, though, where it is safe.

As that early-clobber question points out, inline asm is really easy to get wrong (in hard-to-debug ways which only cause trouble up after some change to the code it's inlined into).

x86 and x86-64 inline asm is assumed to clobber the flags, so an explicit "cc" clobber isn't needed.