path: root/otherlibs/bigarray/bigarray_stubs.c

/***********************************************************************/
/*                                                                     */
/*                                OCaml                                */
/*                                                                     */
/*         Manuel Serrano and Xavier Leroy, INRIA Rocquencourt         */
/*                                                                     */
/*  Copyright 2000 Institut National de Recherche en Informatique et   */
/*  en Automatique.  All rights reserved.  This file is distributed    */
/*  under the terms of the GNU Library General Public License, with    */
/*  the special exception on linking described in file ../../LICENSE.  */
/*                                                                     */
/***********************************************************************/

/* $Id$ */

#include <stddef.h>
#include <stdarg.h>
#include <string.h>
#include "alloc.h"
#include "bigarray.h"
#include "custom.h"
#include "fail.h"
#include "intext.h"
#include "hash.h"
#include "memory.h"
#include "mlvalues.h"

#define int8 caml_ba_int8
#define uint8 caml_ba_uint8
#define int16 caml_ba_int16
#define uint16 caml_ba_uint16

extern void caml_ba_unmap_file(void * addr, uintnat len);
                                          /* from mmap_xxx.c */

/* Compute the number of elements of a big array */

static uintnat caml_ba_num_elts(struct caml_ba_array * b)
{
  uintnat num_elts;
  int i;
  num_elts = 1;
  for (i = 0; i < b->num_dims; i++) num_elts = num_elts * b->dim[i];
  return num_elts;
}

/* Size in bytes of a bigarray element, indexed by bigarray kind */

int caml_ba_element_size[] =
{ 4 /*FLOAT32*/, 8 /*FLOAT64*/,
  1 /*SINT8*/, 1 /*UINT8*/,
  2 /*SINT16*/, 2 /*UINT16*/,
  4 /*INT32*/, 8 /*INT64*/,
  sizeof(value) /*CAML_INT*/, sizeof(value) /*NATIVE_INT*/,
  8 /*COMPLEX32*/, 16 /*COMPLEX64*/
};

/* Compute the number of bytes for the elements of a big array */

CAMLexport uintnat caml_ba_byte_size(struct caml_ba_array * b)
{
  return caml_ba_num_elts(b)
         * caml_ba_element_size[b->flags & CAML_BA_KIND_MASK];
}

/* Operation table for bigarrays */

static void caml_ba_finalize(value v);
static int caml_ba_compare(value v1, value v2);
static intnat caml_ba_hash(value v);
static void caml_ba_serialize(value, uintnat *, uintnat *);
uintnat caml_ba_deserialize(void * dst);
static struct custom_operations caml_ba_ops = {
  "_bigarray",
  caml_ba_finalize,
  caml_ba_compare,
  caml_ba_hash,
  caml_ba_serialize,
  caml_ba_deserialize,
  custom_compare_ext_default
};

/* Multiplication of unsigned longs with overflow detection */

static uintnat
caml_ba_multov(uintnat a, uintnat b, int * overflow)
{
#define HALF_SIZE (sizeof(uintnat) * 4)
#define HALF_MASK (((uintnat)1 << HALF_SIZE) - 1)
#define LOW_HALF(x) ((x) & HALF_MASK)
#define HIGH_HALF(x) ((x) >> HALF_SIZE)
  /* Cut in half words */
  uintnat al = LOW_HALF(a);
  uintnat ah = HIGH_HALF(a);
  uintnat bl = LOW_HALF(b);
  uintnat bh = HIGH_HALF(b);
  /* Exact product is:
              al * bl
           +  ah * bl  << HALF_SIZE
           +  al * bh  << HALF_SIZE
           +  ah * bh  << 2*HALF_SIZE
     Overflow occurs if:
        ah * bh is not 0, i.e. ah != 0 and bh != 0
     OR ah * bl has high half != 0
     OR ah * bl has high half != 0
     OR the sum al * bl + LOW_HALF(ah * bl) << HALF_SIZE
                        + LOW_HALF(al * bh) << HALF_SIZE overflows.
     This sum is equal to p = (a * b) modulo word size. */
  uintnat p1 = al * bh;
  uintnat p2 = ah * bl;
  uintnat p = a * b;
  if (ah != 0 && bh != 0) *overflow = 1;
  if (HIGH_HALF(p1) != 0 || HIGH_HALF(p2) != 0) *overflow = 1;
  p1 <<= HALF_SIZE;
  p2 <<= HALF_SIZE;
  p1 += p2;
  if (p < p1 || p1 < p2) *overflow = 1; /* overflow in sums */
  return p;
#undef HALF_SIZE
#undef LOW_HALF
#undef HIGH_HALF
}

/* Allocation of a big array */

#define CAML_BA_MAX_MEMORY 1024*1024*1024
/* 1 Gb -- after allocating that much, it's probably worth speeding
   up the major GC */

/* [caml_ba_alloc] will allocate a new bigarray object in the heap.
   If [data] is NULL, the memory for the contents is also allocated
   (with [malloc]) by [caml_ba_alloc].
   [data] cannot point into the OCaml heap.
   [dim] may point into an object in the OCaml heap.
*/
CAMLexport value
caml_ba_alloc(int flags, int num_dims, void * data, intnat * dim)
{
  uintnat num_elts, asize, size;
  int overflow, i;
  value res;
  struct caml_ba_array * b;
  intnat dimcopy[CAML_BA_MAX_NUM_DIMS];

  Assert(num_dims >= 1 && num_dims <= CAML_BA_MAX_NUM_DIMS);
  Assert((flags & CAML_BA_KIND_MASK) <= CAML_BA_COMPLEX64);
  for (i = 0; i < num_dims; i++) dimcopy[i] = dim[i];
  size = 0;
  if (data == NULL) {
    overflow = 0;
    num_elts = 1;
    for (i = 0; i < num_dims; i++) {
      num_elts = caml_ba_multov(num_elts, dimcopy[i], &overflow);
    }
    size = caml_ba_multov(num_elts,
                          caml_ba_element_size[flags & CAML_BA_KIND_MASK],
                          &overflow);
    if (overflow) caml_raise_out_of_memory();
    data = malloc(size);
    if (data == NULL && size != 0) caml_raise_out_of_memory();
    flags |= CAML_BA_MANAGED;
  }
  /* PR#5516: use C99's flexible array types if possible */
#if (__STDC_VERSION__ >= 199901L)
  asize = sizeof(struct caml_ba_array) + num_dims * sizeof(intnat);
#else
  asize = sizeof(struct caml_ba_array) + (num_dims - 1) * sizeof(intnat);
#endif
  res = caml_alloc_custom(&caml_ba_ops, asize, size, CAML_BA_MAX_MEMORY);
  b = Caml_ba_array_val(res);
  b->data = data;
  b->num_dims = num_dims;
  b->flags = flags;
  b->proxy = NULL;
  for (i = 0; i < num_dims; i++) b->dim[i] = dimcopy[i];
  return res;
}

/* Same as caml_ba_alloc, but dimensions are passed as a list of
   arguments */

CAMLexport value caml_ba_alloc_dims(int flags, int num_dims, void * data, ...)
{
  va_list ap;
  intnat dim[CAML_BA_MAX_NUM_DIMS];
  int i;
  value res;

  Assert(num_dims <= CAML_BA_MAX_NUM_DIMS);
  va_start(ap, data);
  for (i = 0; i < num_dims; i++) dim[i] = va_arg(ap, intnat);
  va_end(ap);
  res = caml_ba_alloc(flags, num_dims, data, dim);
  return res;
}

/* Allocate a bigarray from OCaml */

CAMLprim value caml_ba_create(value vkind, value vlayout, value vdim)
{
  intnat dim[CAML_BA_MAX_NUM_DIMS];
  mlsize_t num_dims;
  int i, flags;

  num_dims = Wosize_val(vdim);
  if (num_dims < 1 || num_dims > CAML_BA_MAX_NUM_DIMS)
    caml_invalid_argument("Bigarray.create: bad number of dimensions");
  for (i = 0; i < num_dims; i++) {
    dim[i] = Long_val(Field(vdim, i));
    if (dim[i] < 0)
      caml_invalid_argument("Bigarray.create: negative dimension");
  }
  flags = Int_val(vkind) | Int_val(vlayout);
  return caml_ba_alloc(flags, num_dims, NULL, dim);
}

/* Given a big array and a vector of indices, check that the indices
   are within the bounds and return the offset of the corresponding
   array element in the data part of the array. */

static long caml_ba_offset(struct caml_ba_array * b, intnat * index)
{
  intnat offset;
  int i;

  offset = 0;
  if ((b->flags & CAML_BA_LAYOUT_MASK) == CAML_BA_C_LAYOUT) {
    /* C-style layout: row major, indices start at 0 */
    for (i = 0; i < b->num_dims; i++) {
      if ((uintnat) index[i] >= (uintnat) b->dim[i])
        caml_array_bound_error();
      offset = offset * b->dim[i] + index[i];
    }
  } else {
    /* Fortran-style layout: column major, indices start at 1 */
    for (i = b->num_dims - 1; i >= 0; i--) {
      if ((uintnat) (index[i] - 1) >= (uintnat) b->dim[i])
        caml_array_bound_error();
      offset = offset * b->dim[i] + (index[i] - 1);
    }
  }
  return offset;
}

/* Helper function to allocate a record of two double floats */

static value copy_two_doubles(double d0, double d1)
{
  value res = caml_alloc_small(2 * Double_wosize, Double_array_tag);
  Store_double_field(res, 0, d0);
  Store_double_field(res, 1, d1);
  return res;
}

/* Generic code to read from a big array */

value caml_ba_get_N(value vb, value * vind, int nind)
{
  struct caml_ba_array * b = Caml_ba_array_val(vb);
  intnat index[CAML_BA_MAX_NUM_DIMS];
  int i;
  intnat offset;

  /* Check number of indices = number of dimensions of array
     (maybe not necessary if ML typing guarantees this) */
  if (nind != b->num_dims)
    caml_invalid_argument("Bigarray.get: wrong number of indices");
  /* Compute offset and check bounds */
  for (i = 0; i < b->num_dims; i++) index[i] = Long_val(vind[i]);
  offset = caml_ba_offset(b, index);
  /* Perform read */
  switch ((b->flags) & CAML_BA_KIND_MASK) {
  default:
    Assert(0);
  case CAML_BA_FLOAT32:
    return caml_copy_double(((float *) b->data)[offset]);
  case CAML_BA_FLOAT64:
    return caml_copy_double(((double *) b->data)[offset]);
  case CAML_BA_SINT8:
    return Val_int(((int8 *) b->data)[offset]);
  case CAML_BA_UINT8:
    return Val_int(((uint8 *) b->data)[offset]);
  case CAML_BA_SINT16:
    return Val_int(((int16 *) b->data)[offset]);
  case CAML_BA_UINT16:
    return Val_int(((uint16 *) b->data)[offset]);
  case CAML_BA_INT32:
    return caml_copy_int32(((int32 *) b->data)[offset]);
  case CAML_BA_INT64:
    return caml_copy_int64(((int64 *) b->data)[offset]);
  case CAML_BA_NATIVE_INT:
    return caml_copy_nativeint(((intnat *) b->data)[offset]);
  case CAML_BA_CAML_INT:
    return Val_long(((intnat *) b->data)[offset]);
  case CAML_BA_COMPLEX32:
    { float * p = ((float *) b->data) + offset * 2;
      return copy_two_doubles(p[0], p[1]); }
  case CAML_BA_COMPLEX64:
    { double * p = ((double *) b->data) + offset * 2;
      return copy_two_doubles(p[0], p[1]); }
  }
}

CAMLprim value caml_ba_get_1(value vb, value vind1)
{
  return caml_ba_get_N(vb, &vind1, 1);
}

CAMLprim value caml_ba_get_2(value vb, value vind1, value vind2)
{
  value vind[2];
  vind[0] = vind1; vind[1] = vind2;
  return caml_ba_get_N(vb, vind, 2);
}

CAMLprim value caml_ba_get_3(value vb, value vind1, value vind2, value vind3)
{
  value vind[3];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  return caml_ba_get_N(vb, vind, 3);
}

#if 0
CAMLprim value caml_ba_get_4(value vb, value vind1, value vind2,
                     value vind3, value vind4)
{
  value vind[4];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3; vind[3] = vind4;
  return caml_ba_get_N(vb, vind, 4);
}

CAMLprim value caml_ba_get_5(value vb, value vind1, value vind2,
                     value vind3, value vind4, value vind5)
{
  value vind[5];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  vind[3] = vind4; vind[4] = vind5;
  return caml_ba_get_N(vb, vind, 5);
}

CAMLprim value caml_ba_get_6(value vb, value vind1, value vind2,
                     value vind3, value vind4, value vind5, value vind6)
{
  value vind[6];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  vind[3] = vind4; vind[4] = vind5; vind[5] = vind6;
  return caml_ba_get_N(vb, vind, 6);
}
#endif

CAMLprim value caml_ba_get_generic(value vb, value vind)
{
  return caml_ba_get_N(vb, &Field(vind, 0), Wosize_val(vind));
}

/* Generic write to a big array */

static value caml_ba_set_aux(value vb, value * vind, intnat nind, value newval)
{
  struct caml_ba_array * b = Caml_ba_array_val(vb);
  intnat index[CAML_BA_MAX_NUM_DIMS];
  int i;
  intnat offset;

  /* Check number of indices = number of dimensions of array
     (maybe not necessary if ML typing guarantees this) */
  if (nind != b->num_dims)
    caml_invalid_argument("Bigarray.set: wrong number of indices");
  /* Compute offset and check bounds */
  for (i = 0; i < b->num_dims; i++) index[i] = Long_val(vind[i]);
  offset = caml_ba_offset(b, index);
  /* Perform write */
  switch (b->flags & CAML_BA_KIND_MASK) {
  default:
    Assert(0);
  case CAML_BA_FLOAT32:
    ((float *) b->data)[offset] = Double_val(newval); break;
  case CAML_BA_FLOAT64:
    ((double *) b->data)[offset] = Double_val(newval); break;
  case CAML_BA_SINT8:
  case CAML_BA_UINT8:
    ((int8 *) b->data)[offset] = Int_val(newval); break;
  case CAML_BA_SINT16:
  case CAML_BA_UINT16:
    ((int16 *) b->data)[offset] = Int_val(newval); break;
  case CAML_BA_INT32:
    ((int32 *) b->data)[offset] = Int32_val(newval); break;
  case CAML_BA_INT64:
    ((int64 *) b->data)[offset] = Int64_val(newval); break;
  case CAML_BA_NATIVE_INT:
    ((intnat *) b->data)[offset] = Nativeint_val(newval); break;
  case CAML_BA_CAML_INT:
    ((intnat *) b->data)[offset] = Long_val(newval); break;
  case CAML_BA_COMPLEX32:
    { float * p = ((float *) b->data) + offset * 2;
      p[0] = Double_field(newval, 0);
      p[1] = Double_field(newval, 1);
      break; }
  case CAML_BA_COMPLEX64:
    { double * p = ((double *) b->data) + offset * 2;
      p[0] = Double_field(newval, 0);
      p[1] = Double_field(newval, 1);
      break; }
  }
  return Val_unit;
}

CAMLprim value caml_ba_set_1(value vb, value vind1, value newval)
{
  return caml_ba_set_aux(vb, &vind1, 1, newval);
}

CAMLprim value caml_ba_set_2(value vb, value vind1, value vind2, value newval)
{
  value vind[2];
  vind[0] = vind1; vind[1] = vind2;
  return caml_ba_set_aux(vb, vind, 2, newval);
}

CAMLprim value caml_ba_set_3(value vb, value vind1, value vind2, value vind3,
                     value newval)
{
  value vind[3];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  return caml_ba_set_aux(vb, vind, 3, newval);
}

#if 0
CAMLprim value caml_ba_set_4(value vb, value vind1, value vind2,
                     value vind3, value vind4, value newval)
{
  value vind[4];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3; vind[3] = vind4;
  return caml_ba_set_aux(vb, vind, 4, newval);
}

CAMLprim value caml_ba_set_5(value vb, value vind1, value vind2,
                     value vind3, value vind4, value vind5, value newval)
{
  value vind[5];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  vind[3] = vind4; vind[4] = vind5;
  return caml_ba_set_aux(vb, vind, 5, newval);
}

CAMLprim value caml_ba_set_6(value vb, value vind1, value vind2,
                     value vind3, value vind4, value vind5,
                     value vind6, value newval)
{
  value vind[6];
  vind[0] = vind1; vind[1] = vind2; vind[2] = vind3;
  vind[3] = vind4; vind[4] = vind5; vind[5] = vind6;
  return caml_ba_set_aux(vb, vind, 6, newval);
}

value caml_ba_set_N(value vb, value * vind, int nargs)
{
  return caml_ba_set_aux(vb, vind, nargs - 1, vind[nargs - 1]);
}
#endif

CAMLprim value caml_ba_set_generic(value vb, value vind, value newval)
{
  return caml_ba_set_aux(vb, &Field(vind, 0), Wosize_val(vind), newval);
}

/* Return the number of dimensions of a big array */

CAMLprim value caml_ba_num_dims(value vb)
{
  struct caml_ba_array * b = Caml_ba_array_val(vb);
  return Val_long(b->num_dims);
}

/* Return the n-th dimension of a big array */

CAMLprim value caml_ba_dim(value vb, value vn)
{
  struct caml_ba_array * b = Caml_ba_array_val(vb);
  intnat n = Long_val(vn);
  if (n >= b->num_dims) caml_invalid_argument("Bigarray.dim");
  return Val_long(b->dim[n]);
}

/* Return the kind of a big array */

CAMLprim value caml_ba_kind(value vb)
{
  return Val_int(Caml_ba_array_val(vb)->flags & CAML_BA_KIND_MASK);
}

/* Return the layout of a big array */

CAMLprim value caml_ba_layout(value vb)
{
  return Val_int(Caml_ba_array_val(vb)->flags & CAML_BA_LAYOUT_MASK);
}

/* Finalization of a big array */

static void caml_ba_finalize(value v)
{
  struct caml_ba_array * b = Caml_ba_array_val(v);

  switch (b->flags & CAML_BA_MANAGED_MASK) {
  case CAML_BA_EXTERNAL:
    break;
  case CAML_BA_MANAGED:
    if (b->proxy == NULL) {
      free(b->data);
    } else {
      if (-- b->proxy->refcount == 0) {
        free(b->proxy->data);
        caml_stat_free(b->proxy);
      }
    }
    break;
  case CAML_BA_MAPPED_FILE:
    if (b->proxy == NULL) {
      caml_ba_unmap_file(b->data, caml_ba_byte_size(b));
    } else {
      if (-- b->proxy->refcount == 0) {
        caml_ba_unmap_file(b->proxy->data, b->proxy->size);
        caml_stat_free(b->proxy);
      }
    }
    break;
  }
}

/* Comparison of two big arrays */

static int caml_ba_compare(value v1, value v2)
{
  struct caml_ba_array * b1 = Caml_ba_array_val(v1);
  struct caml_ba_array * b2 = Caml_ba_array_val(v2);
  uintnat n, num_elts;
  intnat flags1, flags2;
  int i;

  /* Compare kind & layout in case the arguments are of different types */
  flags1 = b1->flags & (CAML_BA_KIND_MASK | CAML_BA_LAYOUT_MASK);
  flags2 = b2->flags & (CAML_BA_KIND_MASK | CAML_BA_LAYOUT_MASK);
  if (flags1 != flags2) return flags2 - flags1;
  /* Compare number of dimensions */
  if (b1->num_dims != b2->num_dims) return b2->num_dims - b1->num_dims;
  /* Same number of dimensions: compare dimensions lexicographically */
  for (i = 0; i < b1->num_dims; i++) {
    intnat d1 = b1->dim[i];
    intnat d2 = b2->dim[i];
    if (d1 != d2) return d1 < d2 ? -1 : 1;
  }
  /* Same dimensions: compare contents lexicographically */
  num_elts = caml_ba_num_elts(b1);

#define DO_INTEGER_COMPARISON(type) \
  { type * p1 = b1->data; type * p2 = b2->data; \
    for (n = 0; n < num_elts; n++) { \
      type e1 = *p1++; type e2 = *p2++; \
      if (e1 < e2) return -1; \
      if (e1 > e2) return 1; \
    } \
    return 0; \
  }
#define DO_FLOAT_COMPARISON(type) \
  { type * p1 = b1->data; type * p2 = b2->data; \
    for (n = 0; n < num_elts; n++) { \
      type e1 = *p1++; type e2 = *p2++; \
      if (e1 < e2) return -1; \
      if (e1 > e2) return 1; \
      if (e1 != e2) { \
        caml_compare_unordered = 1; \
        if (e1 == e1) return 1; \
        if (e2 == e2) return -1; \
      } \
    } \
    return 0; \
  }

  switch (b1->flags & CAML_BA_KIND_MASK) {
  case CAML_BA_COMPLEX32:
    num_elts *= 2; /*fallthrough*/
  case CAML_BA_FLOAT32:
    DO_FLOAT_COMPARISON(float);
  case CAML_BA_COMPLEX64:
    num_elts *= 2; /*fallthrough*/
  case CAML_BA_FLOAT64:
    DO_FLOAT_COMPARISON(double);
  case CAML_BA_SINT8:
    DO_INTEGER_COMPARISON(int8);
  case CAML_BA_UINT8:
    DO_INTEGER_COMPARISON(uint8);
  case CAML_BA_SINT16:
    DO_INTEGER_COMPARISON(int16);
  case CAML_BA_UINT16:
    DO_INTEGER_COMPARISON(uint16);
  case CAML_BA_INT32:
    DO_INTEGER_COMPARISON(int32);
  case CAML_BA_INT64:
#ifdef ARCH_INT64_TYPE
    DO_INTEGER_COMPARISON(int64);
#else
    { int64 * p1 = b1->data; int64 * p2 = b2->data;
      for (n = 0; n < num_elts; n++) {
        int64 e1 = *p1++; int64 e2 = *p2++;
        if ((int32)e1.h > (int32)e2.h) return 1;
        if ((int32)e1.h < (int32)e2.h) return -1;
        if (e1.l > e2.l) return 1;
        if (e1.l < e2.l) return -1;
      }
      return 0;
    }
#endif
  case CAML_BA_CAML_INT:
  case CAML_BA_NATIVE_INT:
    DO_INTEGER_COMPARISON(intnat);
  default:
    Assert(0);
    return 0;                   /* should not happen */
  }
#undef DO_INTEGER_COMPARISON
#undef DO_FLOAT_COMPARISON
}

/* Hashing of a bigarray */

static intnat caml_ba_hash(value v)
{
  struct caml_ba_array * b = Caml_ba_array_val(v);
  intnat num_elts, n;
  uint32 h, w;
  int i;

  num_elts = 1;
  for (i = 0; i < b->num_dims; i++) num_elts = num_elts * b->dim[i];
  h = 0;

  switch (b->flags & CAML_BA_KIND_MASK) {
  case CAML_BA_SINT8:
  case CAML_BA_UINT8: {
    uint8 * p = b->data;
    if (num_elts > 256) num_elts = 256;
    for (n = 0; n + 4 <= num_elts; n += 4, p += 4) {
      w = p[0] | (p[1] << 8) | (p[2] << 16) | (p[3] << 24);
      h = caml_hash_mix_uint32(h, w);
    }
    w = 0;
    switch (num_elts & 3) {
    case 3: w  = p[2] << 16;    /* fallthrough */
    case 2: w |= p[1] << 8;     /* fallthrough */
    case 1: w |= p[0];
            h = caml_hash_mix_uint32(h, w);
    }
    break;
  }
  case CAML_BA_SINT16:
  case CAML_BA_UINT16: {
    uint16 * p = b->data;
    if (num_elts > 128) num_elts = 128;
    for (n = 0; n + 2 <= num_elts; n += 2, p += 2) {
      w = p[0] | (p[1] << 16);
      h = caml_hash_mix_uint32(h, w);
    }
    if ((num_elts & 1) != 0)
      h = caml_hash_mix_uint32(h, p[0]);
    break;
  }
  case CAML_BA_INT32:
  {
    uint32 * p = b->data;
    if (num_elts > 64) num_elts = 64;
    for (n = 0; n < num_elts; n++, p++) h = caml_hash_mix_uint32(h, *p);
    break;
  }
  case CAML_BA_CAML_INT:
  case CAML_BA_NATIVE_INT:
  {
    intnat * p = b->data;
    if (num_elts > 64) num_elts = 64;
    for (n = 0; n < num_elts; n++, p++) h = caml_hash_mix_intnat(h, *p);
    break;
  }
  case CAML_BA_INT64:
  {
    int64 * p = b->data;
    if (num_elts > 32) num_elts = 32;
    for (n = 0; n < num_elts; n++, p++) h = caml_hash_mix_int64(h, *p);
    break;
  }
  case CAML_BA_COMPLEX32:
    num_elts *= 2;              /* fallthrough */
  case CAML_BA_FLOAT32:
  {
    float * p = b->data;
    if (num_elts > 64) num_elts = 64;
    for (n = 0; n < num_elts; n++, p++) h = caml_hash_mix_float(h, *p);
    break;
  }
  case CAML_BA_COMPLEX64:
    num_elts *= 2;              /* fallthrough */
  case CAML_BA_FLOAT64:
  {
    double * p = b->data;
    if (num_elts > 32) num_elts = 32;
    for (n = 0; n < num_elts; n++, p++) h = caml_hash_mix_double(h, *p);
    break;
  }
  }
  return h;
}

static void caml_ba_serialize_longarray(void * data,
                                        intnat num_elts,
                                        intnat min_val, intnat max_val)
{
#ifdef ARCH_SIXTYFOUR
  int overflow_32 = 0;
  intnat * p, n;
  for (n = 0, p = data; n < num_elts; n++, p++) {
    if (*p < min_val || *p > max_val) { overflow_32 = 1; break; }
  }
  if (overflow_32) {
    caml_serialize_int_1(1);
    caml_serialize_block_8(data, num_elts);
  } else {
    caml_serialize_int_1(0);
    for (n = 0, p = data; n < num_elts; n++, p++)
      caml_serialize_int_4((int32) *p);
  }
#else
  caml_serialize_int_1(0);
  caml_serialize_block_4(data, num_elts);
#endif
}

static void caml_ba_serialize(value v,
                              uintnat * wsize_32,
                              uintnat * wsize_64)
{
  struct caml_ba_array * b = Caml_ba_array_val(v);
  intnat num_elts;
  int i;

  /* Serialize header information */
  caml_serialize_int_4(b->num_dims);
  caml_serialize_int_4(b->flags & (CAML_BA_KIND_MASK | CAML_BA_LAYOUT_MASK));
  /* On a 64-bit machine, if any of the dimensions is >= 2^32,
     the size of the marshaled data will be >= 2^32 and
     extern_value() will fail.  So, it is safe to write the dimensions
     as 32-bit unsigned integers. */
  for (i = 0; i < b->num_dims; i++) caml_serialize_int_4(b->dim[i]);
  /* Compute total number of elements */
  num_elts = 1;
  for (i = 0; i < b->num_dims; i++) num_elts = num_elts * b->dim[i];
  /* Serialize elements */
  switch (b->flags & CAML_BA_KIND_MASK) {
  case CAML_BA_SINT8:
  case CAML_BA_UINT8:
    caml_serialize_block_1(b->data, num_elts); break;
  case CAML_BA_SINT16:
  case CAML_BA_UINT16:
    caml_serialize_block_2(b->data, num_elts); break;
  case CAML_BA_FLOAT32:
  case CAML_BA_INT32:
    caml_serialize_block_4(b->data, num_elts); break;
  case CAML_BA_COMPLEX32:
    caml_serialize_block_4(b->data, num_elts * 2); break;
  case CAML_BA_FLOAT64:
  case CAML_BA_INT64:
    caml_serialize_block_8(b->data, num_elts); break;
  case CAML_BA_COMPLEX64:
    caml_serialize_block_8(b->data, num_elts * 2); break;
  case CAML_BA_CAML_INT:
    caml_ba_serialize_longarray(b->data, num_elts, -0x40000000, 0x3FFFFFFF);
    break;
  case CAML_BA_NATIVE_INT:
    caml_ba_serialize_longarray(b->data, num_elts, -0x80000000, 0x7FFFFFFF);
    break;
  }
  /* Compute required size in OCaml heap.  Assumes struct caml_ba_array
     is exactly 4 + num_dims words */
  Assert(sizeof(struct caml_ba_array) == 5 * sizeof(value));
  *wsize_32 = (4 + b->num_dims) * 4;
  *wsize_64 = (4 + b->num_dims) * 8;
}

static void caml_ba_deserialize_longarray(void * dest, intnat num_elts)
{
  int sixty = caml_deserialize_uint_1();
#ifdef ARCH_SIXTYFOUR
  if (sixty) {
    caml_deserialize_block_8(dest, num_elts);
  } else {
    intnat * p, n;
    for (n = 0, p = dest; n < num_elts; n++, p++)
      *p = caml_deserialize_sint_4();
  }
#else
  if (sixty)
    caml_deserialize_error("input_value: cannot read bigarray "
                      "with 64-bit OCaml ints");
  caml_deserialize_block_4(dest, num_elts);
#endif
}

uintnat caml_ba_deserialize(void * dst)
{
  struct caml_ba_array * b = dst;
  int i, elt_size;
  uintnat num_elts;

  /* Read back header information */
  b->num_dims = caml_deserialize_uint_4();
  b->flags = caml_deserialize_uint_4() | CAML_BA_MANAGED;
  b->proxy = NULL;
  for (i = 0; i < b->num_dims; i++) b->dim[i] = caml_deserialize_uint_4();
  /* Compute total number of elements */
  num_elts = caml_ba_num_elts(b);
  /* Determine element size in bytes */
  if ((b->flags & CAML_BA_KIND_MASK) > CAML_BA_COMPLEX64)
    caml_deserialize_error("input_value: bad bigarray kind");
  elt_size = caml_ba_element_size[b->flags & CAML_BA_KIND_MASK];
  /* Allocate room for data */
  b->data = malloc(elt_size * num_elts);
  if (b->data == NULL)
    caml_deserialize_error("input_value: out of memory for bigarray");
  /* Read data */
  switch (b->flags & CAML_BA_KIND_MASK) {
  case CAML_BA_SINT8:
  case CAML_BA_UINT8:
    caml_deserialize_block_1(b->data, num_elts); break;
  case CAML_BA_SINT16:
  case CAML_BA_UINT16:
    caml_deserialize_block_2(b->data, num_elts); break;
  case CAML_BA_FLOAT32:
  case CAML_BA_INT32:
    caml_deserialize_block_4(b->data, num_elts); break;
  case CAML_BA_COMPLEX32:
    caml_deserialize_block_4(b->data, num_elts * 2); break;
  case CAML_BA_FLOAT64:
  case CAML_BA_INT64:
    caml_deserialize_block_8(b->data, num_elts); break;
  case CAML_BA_COMPLEX64:
    caml_deserialize_block_8(b->data, num_elts * 2); break;
  case CAML_BA_CAML_INT:
  case CAML_BA_NATIVE_INT:
    caml_ba_deserialize_longarray(b->data, num_elts); break;
  }
  return sizeof(struct caml_ba_array) + (b->num_dims - 1) * sizeof(intnat);
}

/* Create / update proxy to indicate that b2 is a sub-array of b1 */

static void caml_ba_update_proxy(struct caml_ba_array * b1,
                                 struct caml_ba_array * b2)
{
  struct caml_ba_proxy * proxy;
  /* Nothing to do for un-managed arrays */
  if ((b1->flags & CAML_BA_MANAGED_MASK) == CAML_BA_EXTERNAL) return;
  if (b1->proxy != NULL) {
    /* If b1 is already a proxy for a larger array, increment refcount of
       proxy */
    b2->proxy = b1->proxy;
    ++ b1->proxy->refcount;
  } else {
    /* Otherwise, create proxy and attach it to both b1 and b2 */
    proxy = caml_stat_alloc(sizeof(struct caml_ba_proxy));
    proxy->refcount = 2;      /* original array + sub array */
    proxy->data = b1->data;
    proxy->size =
      b1->flags & CAML_BA_MAPPED_FILE ? caml_ba_byte_size(b1) : 0;
    b1->proxy = proxy;
    b2->proxy = proxy;
  }
}

/* Slicing */

CAMLprim value caml_ba_slice(value vb, value vind)
{
  CAMLparam2 (vb, vind);
  #define b ((struct caml_ba_array *) Caml_ba_array_val(vb))
  CAMLlocal1 (res);
  intnat index[CAML_BA_MAX_NUM_DIMS];
  int num_inds, i;
  intnat offset;
  intnat * sub_dims;
  char * sub_data;

  /* Check number of indices < number of dimensions of array */
  num_inds = Wosize_val(vind);
  if (num_inds >= b->num_dims)
    caml_invalid_argument("Bigarray.slice: too many indices");
  /* Compute offset and check bounds */
  if ((b->flags & CAML_BA_LAYOUT_MASK) == CAML_BA_C_LAYOUT) {
    /* We slice from the left */
    for (i = 0; i < num_inds; i++) index[i] = Long_val(Field(vind, i));
    for (/*nothing*/; i < b->num_dims; i++) index[i] = 0;
    offset = caml_ba_offset(b, index);
    sub_dims = b->dim + num_inds;
  } else {
    /* We slice from the right */
    for (i = 0; i < num_inds; i++)
      index[b->num_dims - num_inds + i] = Long_val(Field(vind, i));
    for (i = 0; i < b->num_dims - num_inds; i++) index[i] = 1;
    offset = caml_ba_offset(b, index);
    sub_dims = b->dim;
  }
  sub_data =
    (char *) b->data +
    offset * caml_ba_element_size[b->flags & CAML_BA_KIND_MASK];
  /* Allocate an OCaml bigarray to hold the result */
  res = caml_ba_alloc(b->flags, b->num_dims - num_inds, sub_data, sub_dims);
  /* Create or update proxy in case of managed bigarray */
  caml_ba_update_proxy(b, Caml_ba_array_val(res));
  /* Return result */
  CAMLreturn (res);

  #undef b
}

/* Extracting a sub-array of same number of dimensions */

CAMLprim value caml_ba_sub(value vb, value vofs, value vlen)
{
  CAMLparam3 (vb, vofs, vlen);
  CAMLlocal1 (res);
  #define b ((struct caml_ba_array *) Caml_ba_array_val(vb))
  intnat ofs = Long_val(vofs);
  intnat len = Long_val(vlen);
  int i, changed_dim;
  intnat mul;
  char * sub_data;

  /* Compute offset and check bounds */
  if ((b->flags & CAML_BA_LAYOUT_MASK) == CAML_BA_C_LAYOUT) {
    /* We reduce the first dimension */
    mul = 1;
    for (i = 1; i < b->num_dims; i++) mul *= b->dim[i];
    changed_dim = 0;
  } else {
    /* We reduce the last dimension */
    mul = 1;
    for (i = 0; i < b->num_dims - 1; i++) mul *= b->dim[i];
    changed_dim = b->num_dims - 1;
    ofs--;                      /* Fortran arrays start at 1 */
  }
  if (ofs < 0 || len < 0 || ofs + len > b->dim[changed_dim])
    caml_invalid_argument("Bigarray.sub: bad sub-array");
  sub_data =
    (char *) b->data +
    ofs * mul * caml_ba_element_size[b->flags & CAML_BA_KIND_MASK];
  /* Allocate an OCaml bigarray to hold the result */
  res = caml_ba_alloc(b->flags, b->num_dims, sub_data, b->dim);
  /* Doctor the changed dimension */
  Caml_ba_array_val(res)->dim[changed_dim] = len;
  /* Create or update proxy in case of managed bigarray */
  caml_ba_update_proxy(b, Caml_ba_array_val(res));
  /* Return result */
  CAMLreturn (res);

  #undef b
}

/* Copying a big array into another one */

CAMLprim value caml_ba_blit(value vsrc, value vdst)
{
  struct caml_ba_array * src = Caml_ba_array_val(vsrc);
  struct caml_ba_array * dst = Caml_ba_array_val(vdst);
  int i;
  intnat num_bytes;

  /* Check same numbers of dimensions and same dimensions */
  if (src->num_dims != dst->num_dims) goto blit_error;
  for (i = 0; i < src->num_dims; i++)
    if (src->dim[i] != dst->dim[i]) goto blit_error;
  /* Compute number of bytes in array data */
  num_bytes =
    caml_ba_num_elts(src)
    * caml_ba_element_size[src->flags & CAML_BA_KIND_MASK];
  /* Do the copying */
  memmove (dst->data, src->data, num_bytes);
  return Val_unit;
 blit_error:
  caml_invalid_argument("Bigarray.blit: dimension mismatch");
  return Val_unit;              /* not reached */
}

/* Filling a big array with a given value */

CAMLprim value caml_ba_fill(value vb, value vinit)
{
  struct caml_ba_array * b = Caml_ba_array_val(vb);
  intnat num_elts = caml_ba_num_elts(b);

  switch (b->flags & CAML_BA_KIND_MASK) {
  default:
    Assert(0);
  case CAML_BA_FLOAT32: {
    float init = Double_val(vinit);
    float * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_FLOAT64: {
    double init = Double_val(vinit);
    double * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_SINT8:
  case CAML_BA_UINT8: {
    int init = Int_val(vinit);
    char * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_SINT16:
  case CAML_BA_UINT16: {
    int init = Int_val(vinit);
    int16 * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_INT32: {
    int32 init = Int32_val(vinit);
    int32 * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_INT64: {
    int64 init = Int64_val(vinit);
    int64 * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_NATIVE_INT: {
    intnat init = Nativeint_val(vinit);
    intnat * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_CAML_INT: {
    intnat init = Long_val(vinit);
    intnat * p;
    for (p = b->data; num_elts > 0; p++, num_elts--) *p = init;
    break;
  }
  case CAML_BA_COMPLEX32: {
    float init0 = Double_field(vinit, 0);
    float init1 = Double_field(vinit, 1);
    float * p;
    for (p = b->data; num_elts > 0; num_elts--) { *p++ = init0; *p++ = init1; }
    break;
  }
  case CAML_BA_COMPLEX64: {
    double init0 = Double_field(vinit, 0);
    double init1 = Double_field(vinit, 1);
    double * p;
    for (p = b->data; num_elts > 0; num_elts--) { *p++ = init0; *p++ = init1; }
    break;
  }
  }
  return Val_unit;
}

/* Reshape an array: change dimensions and number of dimensions, preserving
   array contents */

CAMLprim value caml_ba_reshape(value vb, value vdim)
{
  CAMLparam2 (vb, vdim);
  CAMLlocal1 (res);
#define b ((struct caml_ba_array *) Caml_ba_array_val(vb))
  intnat dim[CAML_BA_MAX_NUM_DIMS];
  mlsize_t num_dims;
  uintnat num_elts;
  int i;

  num_dims = Wosize_val(vdim);
  if (num_dims < 1 || num_dims > CAML_BA_MAX_NUM_DIMS)
    caml_invalid_argument("Bigarray.reshape: bad number of dimensions");
  num_elts = 1;
  for (i = 0; i < num_dims; i++) {
    dim[i] = Long_val(Field(vdim, i));
    if (dim[i] < 0)
      caml_invalid_argument("Bigarray.reshape: negative dimension");
    num_elts *= dim[i];
  }
  /* Check that sizes agree */
  if (num_elts != caml_ba_num_elts(b))
    caml_invalid_argument("Bigarray.reshape: size mismatch");
  /* Create bigarray with same data and new dimensions */
  res = caml_ba_alloc(b->flags, num_dims, b->data, dim);
  /* Create or update proxy in case of managed bigarray */
  caml_ba_update_proxy(b, Caml_ba_array_val(res));
  /* Return result */
  CAMLreturn (res);

#undef b
}

/* Initialization */

CAMLprim value caml_ba_init(value unit)
{
  caml_register_custom_operations(&caml_ba_ops);
  return Val_unit;
}