I'm trying to produce code (currently using clang++-3.8) that adds two numbers consisting of multiple machine words. To simplify things for the moment I'm only adding 128bit numbers, but I'd like to be able to generalise this.
First some typedefs:
typedef unsigned long long unsigned_word;
typedef __uint128_t unsigned_128;
And a "result" type:
struct Result
{
unsigned_word lo;
unsigned_word hi;
};
The first function, f, takes two pairs of unsigned words and returns a result, by as an intermediate step putting both of these 64 bit words into a 128 bit word before adding them, like so:
Result f (unsigned_word lo1, unsigned_word hi1, unsigned_word lo2, unsigned_word hi2)
{
Result x;
unsigned_128 n1 = lo1 + (static_cast<unsigned_128>(hi1) << 64);
unsigned_128 n2 = lo2 + (static_cast<unsigned_128>(hi2) << 64);
unsigned_128 r1 = n1 + n2;
x.lo = r1 & ((static_cast<unsigned_128>(1) << 64) - 1);
x.hi = r1 >> 64;
return x;
}
This actually gets inlined quite nicely like so:
movq 8(%rsp), %rsi
movq (%rsp), %rbx
addq 24(%rsp), %rsi
adcq 16(%rsp), %rbx
Now, instead I've written a simpler function using the clang multi-precision primatives, as below:
static Result g (unsigned_word lo1, unsigned_word hi1, unsigned_word lo2, unsigned_word hi2)
{
Result x;
unsigned_word carryout;
x.lo = __builtin_addcll(lo1, lo2, 0, &carryout);
x.hi = __builtin_addcll(hi1, hi2, carryout, &x.carry);
return x;
}
This produces the following assembly:
movq 24(%rsp), %rsi
movq (%rsp), %rbx
addq 16(%rsp), %rbx
addq 8(%rsp), %rsi
adcq $0, %rbx
In this case, there's an extra add. Instead of doing an ordinary add on the lo-words, then an adc on the hi-words, it just adds the hi-words, then adds the lo-words, then does an adc on the hi-word again with an argument of zero.
This may not look too bad, but when you try this with larger words (say 192bit, 256bit) you soon get a mess of ors and other instructions dealing with the carries up the chain, instead of a simple chain of add, adc, adc, ... adc.
The multi-precision primitives seem to be doing a terrible job at exactly what they're intended to do.
So what I'm looking for is code that I could generalise to any length (no need to do it, just enough so I can work out how to), which clang produces additions in an manner with is as efficient as what it does with it's built in 128 bit type (which unfortunately I can't easily generalise). I presume this should just a chain of adcs, but I'm welcome to arguments and code that it should be something else.
There is an intrinsic to do this: _addcarry_u64. However, only Visual Studio and ICC (at least VS 2013 and 2015 and ICC 13 and ICC 15) do this efficiently. Clang 3.7 and GCC 5.2 still don't produce efficient code with this intrinsic.
Clang in addition has a built-in which one would think does this, __builtin_addcll, but it does not produce efficient code either.
The reason Visual Studio does this is that it does not allow inline assembly in 64-bit mode so the compiler should provide a way to do this with an intrinsic (though Microsoft took their time implementing this).
Therefore, with Visual Studio use _addcarry_u64. With ICC use _addcarry_u64 or inline assembly. With Clang and GCC use inline assembly.
Note that since the Broadwell microarchitecture there are two new instructions: adcx and adox which you can access with the _addcarryx_u64 intrinsic . Intel's documentation for these intrinsics used to be different then the assembly produced by the compiler but it appears their documentation is correct now. However, Visual Studio still only appears to produce adcx with _addcarryx_u64 whereas ICC produces both adcx and adox with this intrinsic. But even though ICC produces both instructions it does not produce the most optimal code (ICC 15) and so inline assembly is still necessary.
Personally, I think the fact that a non-standard feature of C/C++, such as inline assembly or intrinsics, is required to do this is a weakness of C/C++ but others might disagree. The adc instruction has been in the x86 instruction set since 1979. I would not hold my breath on C/C++ compilers being able to optimally figure out when you want adc. Sure they can have built-in types such as __int128 but the moment you want a larger type that's not built-in you have to use some non-standard C/C++ feature such as inline assembly or intrinsics.
In terms of inline assembly code to do this I already posted a solution for 256-bit addition for eight 64-bit integers in register at multi-word addition using the carry flag.
Here is that code reposted.
#define ADD256(X1, X2, X3, X4, Y1, Y2, Y3, Y4) \
__asm__ __volatile__ ( \
"addq %[v1], %[u1] \n" \
"adcq %[v2], %[u2] \n" \
"adcq %[v3], %[u3] \n" \
"adcq %[v4], %[u4] \n" \
: [u1] "+&r" (X1), [u2] "+&r" (X2), [u3] "+&r" (X3), [u4] "+&r" (X4) \
: [v1] "r" (Y1), [v2] "r" (Y2), [v3] "r" (Y3), [v4] "r" (Y4))
If you want to explicitly load the values from memory you can do something like this
//uint64_t dst[4] = {1,1,1,1};
//uint64_t src[4] = {1,2,3,4};
asm (
"movq (%[in]), %%rax\n"
"addq %%rax, %[out]\n"
"movq 8(%[in]), %%rax\n"
"adcq %%rax, 8%[out]\n"
"movq 16(%[in]), %%rax\n"
"adcq %%rax, 16%[out]\n"
"movq 24(%[in]), %%rax\n"
"adcq %%rax, 24%[out]\n"
: [out] "=m" (dst)
: [in]"r" (src)
: "%rax"
);
That produces nearlly identical assembly as from the following function in ICC
void add256(uint256 *x, uint256 *y) {
unsigned char c = 0;
c = _addcarry_u64(c, x->x1, y->x1, &x->x1);
c = _addcarry_u64(c, x->x2, y->x2, &x->x2);
c = _addcarry_u64(c, x->x3, y->x3, &x->x3);
_addcarry_u64(c, x->x4, y->x4, &x->x4);
}
I have limited experience with GCC inline assembly (or inline assembly in general - I usually use an assembler such as NASM) so maybe there are better inline assembly solutions.
So what I'm looking for is code that I could generalize to any length
To answer this question here is another solution using template meta programming. I used this same trick for loop unrolling. This produces optimal code with ICC. If Clang or GCC ever implement _addcarry_u64 efficiently this would be a good general solution.
#include <x86intrin.h>
#include <inttypes.h>
#define LEN 4 // N = N*64-bit add e.g. 4=256-bit add, 3=192-bit add, ...
static unsigned char c = 0;
template<int START, int N>
struct Repeat {
static void add (uint64_t *x, uint64_t *y) {
c = _addcarry_u64(c, x[START], y[START], &x[START]);
Repeat<START+1, N>::add(x,y);
}
};
template<int N>
struct Repeat<LEN, N> {
static void add (uint64_t *x, uint64_t *y) {}
};
void sum_unroll(uint64_t *x, uint64_t *y) {
Repeat<0,LEN>::add(x,y);
}
Assembly from ICC
xorl %r10d, %r10d #12.13
movzbl c(%rip), %eax #12.13
cmpl %eax, %r10d #12.13
movq (%rsi), %rdx #12.13
adcq %rdx, (%rdi) #12.13
movq 8(%rsi), %rcx #12.13
adcq %rcx, 8(%rdi) #12.13
movq 16(%rsi), %r8 #12.13
adcq %r8, 16(%rdi) #12.13
movq 24(%rsi), %r9 #12.13
adcq %r9, 24(%rdi) #12.13
setb %r10b
Meta programming is a basic feature of assemblers so it's too bad C and C++ (except through template meta programming hacks) have no solution for this either (the D language does).
The inline assembly I used above which referenced memory was causing some problems in a function. Here is a new version which seems to work better
void foo(uint64_t *dst, uint64_t *src)
{
__asm (
"movq (%[in]), %%rax\n"
"addq %%rax, (%[out])\n"
"movq 8(%[in]), %%rax\n"
"adcq %%rax, 8(%[out])\n"
"movq 16(%[in]), %%rax\n"
"addq %%rax, 16(%[out])\n"
"movq 24(%[in]), %%rax\n"
"adcq %%rax, 24(%[out])\n"
:
: [in] "r" (src), [out] "r" (dst)
: "%rax"
);
}
Starting with clang 5.0 it is possible to get good results using __uint128_t-addition and getting the carry bit by shifting:
inline uint64_t add_with_carry(uint64_t &a, const uint64_t &b, const uint64_t &c)
{
__uint128_t s = __uint128_t(a) + b + c;
a = s;
return s >> 64;
}
In many situations clang still does strange operations (I assume because of possible aliasing?), but usually copying one variable into a temporary helps.
Usage examples with
template<int size> struct LongInt
{
uint64_t data[size];
};
Manual usage:
void test(LongInt<3> &a, const LongInt<3> &b_)
{
const LongInt<3> b = b_; // need to copy b_ into local temporary
uint64_t c0 = add_with_carry(a.data[0], b.data[0], 0);
uint64_t c1 = add_with_carry(a.data[1], b.data[1], c0);
uint64_t c2 = add_with_carry(a.data[2], b.data[2], c1);
}
Generic solution:
template<int size>
void addTo(LongInt<size> &a, const LongInt<size> b)
{
__uint128_t c = __uint128_t(a.data[0]) + b.data[0];
for(int i=1; i<size; ++i)
{
c = __uint128_t(a.data[i]) + b.data[i] + (c >> 64);
a.data[i] = c;
}
}
Godbolt Link: All examples above are compiled to only mov, add and adc instructions (starting with clang 5.0, and at least -O2).
The examples don't produce good code with gcc (up to 8.1, which at the moment is the highest version on godbolt).
And I did not yet manage to get anything usable with __builtin_addcll ...
On Clang 6, both __builtin_addcl and __builtin_add_overflow produce the same, optimal disassembly.
Result g(unsigned_word lo1, unsigned_word hi1, unsigned_word lo2, unsigned_word hi2)
{
Result x;
unsigned_word carryout;
x.lo = __builtin_addcll(lo1, lo2, 0, &carryout);
x.hi = __builtin_addcll(hi1, hi2, carryout, &carryout);
return x;
}
Result h(unsigned_word lo1, unsigned_word hi1, unsigned_word lo2, unsigned_word hi2)
{
Result x;
unsigned_word carryout;
carryout = __builtin_add_overflow(lo1, lo2, &x.lo);
carryout = __builtin_add_overflow(hi1, carryout, &hi1);
__builtin_add_overflow(hi1, hi2, &x.hi);
return x;
}
Assembly for both:
add rdi, rdx
adc rsi, rcx
mov rax, rdi
mov rdx, rsi
ret
来源：https://stackoverflow.com/questions/33690791/producing-good-add-with-carry-code-from-clang