static uint64_t umul128(uint64_t a, uint64_t b, uint64_t *productHi)
{
// The casts here help MSVC to avoid calls to the __allmul library function.
uint32_t aLo = (uint32_t)a;
uint32_t aHi = (uint32_t)(a >> 32);
uint32_t bLo = (uint32_t)b;
uint32_t bHi = (uint32_t)(b >> 32);
uint64_t b00 = (uint64_t)aLo * bLo;
uint64_t b01 = (uint64_t)aLo * bHi;
uint64_t b10 = (uint64_t)aHi * bLo;
uint64_t b11 = (uint64_t)aHi * bHi;
uint32_t b00Lo = (uint32_t)b00;
uint32_t b00Hi = (uint32_t)(b00 >> 32);
uint64_t mid1 = b10 + b00Hi;
uint32_t mid1Lo = (uint32_t)(mid1);
uint32_t mid1Hi = (uint32_t)(mid1 >> 32);
uint64_t mid2 = b01 + mid1Lo;
uint32_t mid2Lo = (uint32_t)(mid2);
uint32_t mid2Hi = (uint32_t)(mid2 >> 32);
uint64_t pHi = b11 + mid1Hi + mid2Hi;
uint64_t pLo = ((uint64_t)mid2Lo << 32) | b00Lo;
*productHi = pHi;
return(pLo);
}
public void sfmt_fill_array64(uint64_t *array, int size)
{
assert(idx == SFMT_N32);
assert(size % 2 == 0);
assert(size >= SFMT_N64);
gen_rand_array((w128_t *)array, size / 2);
idx = SFMT_N32;
}
// if needed, allocate memory so that the object is able to process JSON
// documents having up to len bytes and maxdepth "depth"
public bool AllocateCapacity(size_t len, size_t maxdepth = DEFAULTMAXDEPTH)
{
if ((maxdepth == 0) || (len == 0))
{
return(false);
}
if (len > SIMDJSON_MAXSIZE_BYTES)
{
return(false);
}
if ((len <= bytecapacity) && (depthcapacity < maxdepth))
{
return(true);
}
Deallocate();
isvalid = false;
bytecapacity = 0; // will only set it to len after allocations are a success
n_structural_indexes = 0;
uint32_t max_structures = (uint32_t)(ROUNDUP_N(len, 64) + 2 + 7);
structural_indexes = allocate <uint32_t>(max_structures);
// a pathological input like "[[[[..." would generate len tape elements, so need a capacity of len + 1
size_t localtapecapacity = ROUNDUP_N(len + 1, 64);
// a document with only zero-length strings... could have len/3 string
// and we would need len/3 * 5 bytes on the string buffer
size_t localstringcapacity = ROUNDUP_N(5 * len / 3 + 32, 64);
string_buf = allocate <uint8_t>(localstringcapacity);
tape = allocate <uint64_t>(localtapecapacity);
containing_scope_offset = allocate <uint32_t>(maxdepth);
ret_address = allocate <char1>(maxdepth);
if ((string_buf == null) || (tape == null) ||
(containing_scope_offset == null) || (ret_address == null) || (structural_indexes == null))
{
delete(ret_address);
delete(containing_scope_offset);
delete(tape);
delete(string_buf);
delete(structural_indexes);
return(false);
}
/*
* // We do not need to initialize this content for parsing, though we could
* // need to initialize it for safety.
* memset(string_buf, 0 , localstringcapacity);
* memset(structural_indexes, 0, max_structures * sizeof(uint32_t));
* memset(tape, 0, localtapecapacity * sizeof(uint64_t));
*/
bytecapacity = len;
depthcapacity = maxdepth;
tapecapacity = localtapecapacity;
stringcapacity = localstringcapacity;
return(true);
}
public uint32_t dsfmt_genrand_uint32()
{
if (idx >= DSFMT_N64)
{
dsfmt_gen_rand_all();
idx = 0;
}
fixed(w128_t *status = this.status)
{
uint64_t *psfmt64 = &status[0].u64_0;
return((uint)psfmt64[idx++]);
}
}
/**
* This function certificate the period of 2^{SFMT_MEXP}-1.
* @param dsfmt dsfmt state vector.
*/
void period_certification()
{
uint64_t *tmp = stackalloc uint64_t[2];
uint64_t inner;
uint64_t work;
fixed(w128_t *status = this.status)
{
tmp[0] = (status[DSFMT_N].u64_0 ^ DSFMT_FIX1);
tmp[1] = (status[DSFMT_N].u64_1 ^ DSFMT_FIX2);
inner = tmp[0] & pcv[0];
inner ^= tmp[1] & pcv[1];
for (int i = 32; i > 0; i >>= 1)
{
inner ^= inner >> i;
}
inner &= 1;
/* check OK */
if (inner == 1)
{
return;
}
/* check NG, and modification */
if ((DSFMT_PCV2 & 1) == 1)
{
status[DSFMT_N].u64_1 ^= 1;
}
else
{
#pragma warning disable 162
for (int i = 1; i >= 0; i--)
{
work = 1;
for (int j = 0; j < 64; j++)
{
if ((work & pcv[i]) != 0)
{
(&status[DSFMT_N].u64_0)[i] ^= work;
return;
}
work = work << 1;
}
}
#pragma warning restore
}
}
}
// We need a 64x128-bit multiplication and a subsequent 128-bit shift.
// Multiplication:
// The 64-bit factor is variable and passed in, the 128-bit factor comes
// from a lookup table. We know that the 64-bit factor only has 55
// significant bits (i.e., the 9 topmost bits are zeros). The 128-bit
// factor only has 124 significant bits (i.e., the 4 topmost bits are
// zeros).
// Shift:
// In principle, the multiplication result requires 55 + 124 = 179 bits to
// represent. However, we then shift this value to the right by j, which is
// at least j >= 115, so the result is guaranteed to fit into 179 - 115 = 64
// bits. This means that we only need the topmost 64 significant bits of
// the 64x128-bit multiplication.
//
// There are several ways to do this:
// 1. Best case: the compiler exposes a 128-bit type.
// We perform two 64x64-bit multiplications, add the higher 64 bits of the
// lower result to the higher result, and shift by j - 64 bits.
//
// We explicitly cast from 64-bit to 128-bit, so the compiler can tell
// that these are only 64-bit inputs, and can map these to the best
// possible sequence of assembly instructions.
// x64 machines happen to have matching assembly instructions for
// 64x64-bit multiplications and 128-bit shifts.
//
// 2. Second best case: the compiler exposes intrinsics for the x64 assembly
// instructions mentioned in 1.
//
// 3. We only have 64x64 bit instructions that return the lower 64 bits of
// the result, i.e., we have to use plain C.
// Our inputs are less than the full width, so we have three options:
// a. Ignore this fact and just implement the intrinsics manually.
// b. Split both into 31-bit pieces, which guarantees no internal overflow,
// but requires extra work upfront (unless we change the lookup table).
// c. Split only the first factor into 31-bit pieces, which also guarantees
// no internal overflow, but requires extra work since the intermediate
// results are not perfectly aligned.
static uint64_t mulShift64(uint64_t m, uint64_t *mul, int32_t j)
{
// m is maximum 55 bits
uint64_t high1; // 128
uint64_t low1 = umul128(m, mul[1], &high1); // 64
uint64_t high0; // 64
umul128(m, mul[0], &high0); // 0
uint64_t sum = high0 + low1;
if (sum < high0)
{
++high1; // overflow into high1
}
return(shiftright128(sum, high1, (uint)j - 64));
}
// if needed, allocate memory so that the object is able to process JSON
// documents having up to len butes and maxdepth "depth"
public bool AllocateCapacity(size_t len, size_t maxdepth = DEFAULTMAXDEPTH)
{
if ((maxdepth == 0) || (len == 0))
{
Debug.WriteLine("capacities must be non-zero ");
return(false);
}
if (len > 0)
{
if ((len <= bytecapacity) && (depthcapacity < maxdepth))
{
return(true);
}
Deallocate();
}
isvalid = false;
bytecapacity = 0; // will only set it to len after allocations are a success
n_structural_indexes = 0;
uint32_t max_structures = (uint32_t)ROUNDUP_N(len, 64) + 2 + 7;
structural_indexes = Utils.allocate <uint32_t>(max_structures);
size_t localtapecapacity = ROUNDUP_N(len, 64);
size_t localstringcapacity = ROUNDUP_N(len, 64);
string_buf = Utils.allocate <uint8_t>(localstringcapacity);
tape = Utils.allocate <uint64_t>(localtapecapacity);
containing_scope_offset = Utils.allocate <uint32_t>(maxdepth);
ret_address = Utils.allocate <bytechar>(maxdepth);
if ((string_buf == null) || (tape == null) ||
(containing_scope_offset == null) || (ret_address == null) || (structural_indexes == null))
{
Deallocate();
return(false);
}
bytecapacity = len;
depthcapacity = maxdepth;
tapecapacity = localtapecapacity;
stringcapacity = localstringcapacity;
return(true);
}
private void Deallocate()
{
isvalid = false;
bytecapacity = 0;
depthcapacity = 0;
tapecapacity = 0;
stringcapacity = 0;
if (ret_address != null)
{
delete(ret_address);
ret_address = null;
}
if (containing_scope_offset != null)
{
delete(containing_scope_offset);
containing_scope_offset = null;
}
if (tape != null)
{
delete(tape);
tape = null;
}
if (string_buf != null)
{
delete(string_buf);
string_buf = null;
}
if (structural_indexes != null)
{
delete(structural_indexes);
structural_indexes = null;
}
}
static uint32_t mulShift_mod1e9(uint64_t m, uint64_t *mul, int32_t j)
{
uint64_t high0; // 64
uint64_t low0 = umul128(m, mul[0], &high0); // 0
uint64_t high1; // 128
uint64_t low1 = umul128(m, mul[1], &high1); // 64
uint64_t high2; // 192
uint64_t low2 = umul128(m, mul[2], &high2); // 128
uint64_t s0low = low0; // 0
uint64_t s0high = low1 + high0; // 64
uint32_t c1 = s0high < low1 ? 1U : 0;
uint64_t s1low = low2 + high1 + c1; // 128
uint32_t c2 = s1low < low2 ? 1U : 0; // high1 + c1 can't overflow, so compare against low2
uint64_t s1high = high2 + c2; // 192
assert(j >= 128);
assert(j <= 180);
uint32_t dist = (uint32_t)(j - 128); // dist: [0, 52]
uint64_t shiftedhigh = s1high >> (int)dist;
uint64_t shiftedlow = shiftright128(s1low, s1high, dist);
return(uint128_mod1e9(shiftedhigh, shiftedlow));
}
// This is faster if we don't have a 64x64->128-bit multiplication.
static uint64_t mulShiftAll64(uint64_t m, uint64_t *mul, int32_t j,
uint64_t *vp, uint64_t *vm, uint32_t mmShift)
{
m <<= 1;
// m is maximum 55 bits
uint64_t tmp;
uint64_t lo = umul128(m, mul[0], &tmp);
uint64_t hi;
uint64_t mid = tmp + umul128(m, mul[1], &hi);
if (mid < tmp)
{
++hi; // overflow into hi
}
uint64_t lo2 = lo + mul[0];
uint64_t mid2 = mid + mul[1];
if (lo2 < lo)
{
++mid2;
}
uint64_t hi2 = hi;
if (mid2 < mid)
{
++hi2;
}
*vp = shiftright128(mid2, hi2, (uint32_t)(j - 64 - 1));
if (mmShift == 1)
{
uint64_t lo3 = lo - mul[0];
uint64_t mid3 = mid - mul[1];
if (lo3 > lo)
{
--mid3;
}
uint64_t hi3 = hi;
if (mid3 > mid)
{
--hi3;
}
*vm = shiftright128(mid3, hi3, (uint32_t)(j - 64 - 1));
}
else
{
uint64_t lo3 = lo + lo;
uint64_t mid3 = mid + mid;
if (lo3 < lo)
{
++mid3;
}
uint64_t hi3 = hi + hi;
if (mid3 < mid)
{
++hi3;
}
uint64_t lo4 = lo3 - mul[0];
uint64_t mid4 = mid3 - mul[1];
if (lo4 > lo3)
{
--mid4;
}
uint64_t hi4 = hi3;
if (mid4 > mid3)
{
--hi4;
}
*vm = shiftright128(mid4, hi4, (uint32_t)(j - 64));
}
return(shiftright128(mid, hi, (uint32_t)(j - 64 - 1)));
}
