ALTREP implementation of growable_*

Currently broken:
- 1 errors out of 11637
 - won't work with expression vectors at all

src/assign.c

@@ -323,7 +323,7 @@ SEXP shallowwrapper(SEXP dt, SEXP cols) {
 }
 
 SEXP truelength(SEXP x) {
-  return ScalarInteger(isNull(x) ? 0 : growable_max_size(x));
+  return ScalarInteger(is_growable(x) ? growable_max_size(x) : 0);
 }
 
 SEXP selfrefokwrapper(SEXP x, SEXP verbose) {
@@ -520,7 +520,7 @@ SEXP assign(SEXP dt, SEXP rows, SEXP cols, SEXP newcolnames, SEXP values)
   // modify DT by reference. Other than if new columns are being added and the allocVec() fails with
   // out-of-memory. In that case the user will receive hard halt and know to rerun.
   if (length(newcolnames)) {
-    oldtncol = growable_max_size(dt);   // TO DO: oldtncol can be just called tl now, as we won't realloc here any more.
+    oldtncol = is_growable(dt) ? growable_max_size(dt) : 0;   // TO DO: oldtncol can be just called tl now, as we won't realloc here any more.
 
     if (oldtncol<oldncol) {
       if (oldtncol==0) error(_("This data.table has either been loaded from disk (e.g. using readRDS()/load()) or constructed manually (e.g. using structure()). Please run setDT() or setalloccol() on it first (to pre-allocate space for new columns) before assigning by reference to it."));   // #2996

src/data.table.h

@@ -17,6 +17,9 @@
 #if R_VERSION < R_Version(3, 4, 0)
 #  define SET_GROWABLE_BIT(x)  // #3292
 #endif
+#if R_VERSION >= R_Version(4, 3, 0)
+#  define USE_GROWABLE_ALTREP
+#endif
 #include <Rinternals.h>
 #define SEXPPTR_RO(x) ((const SEXP *)DATAPTR_RO(x))  // to avoid overhead of looped STRING_ELT and VECTOR_ELT
 #include <stdint.h>    // for uint64_t rather than unsigned long long
@@ -309,6 +312,9 @@ void growable_resize(SEXP x, R_xlen_t newsize);
 Rboolean is_growable(SEXP x);
 // Transform x into a growable vector. The return value must be reprotected in place of x. What happens to x is deliberately not specified, but no copying occurs.
 SEXP make_growable(SEXP x);
+#if R_VERSION >= R_Version(4, 3, 0)
+void register_altrep_classes(DllInfo*);
+#endif
 
 // functions called from R level .Call/.External and registered in init.c
 // these now live here to pass -Wstrict-prototypes, #5477

src/dogroups.c

@@ -45,11 +45,10 @@ static bool anySpecialStatic(SEXP x, hashtab * specials) {
   // (see data.table.h), and isNewList() is true for NULL
   if (n==0)
     return false;
+  if (hash_lookup(specials, x, 0)<0) return true; // test 2158
   if (isVectorAtomic(x))
-    return ALTREP(x) || hash_lookup(specials, x, 0)<0;
+    return ALTREP(x); // see test 2156: ALTREP is a source of sharing we can't trace reliably
   if (isNewList(x)) {
-    if (hash_lookup(specials, x, 0)<0)
-      return true;  // test 2158
     for (int i=0; i<n; ++i) {
       list_el = VECTOR_ELT(x,i);
       if (anySpecialStatic(list_el, specials))

src/growable.c

@@ -1,5 +1,7 @@
 #include "data.table.h"
 
+#ifndef USE_GROWABLE_ALTREP
+
 SEXP growable_allocate(SEXPTYPE type, R_xlen_t size, R_xlen_t max_size) {
   SEXP ret = PROTECT(allocVector(type, max_size));
   SET_TRUELENGTH(ret, max_size);
@@ -32,9 +34,243 @@ Rboolean is_growable(SEXP x) {
   ;
 }
 
-// Assuming no ALTREP for now
+// Assuming no ALTREP columns
 SEXP make_growable(SEXP x) {
   if (TRUELENGTH(x) < XLENGTH(x)) SET_TRUELENGTH(x, XLENGTH(x));
   SET_GROWABLE_BIT(x);
   return x;
 }
+
+#else
+
+#include <R_ext/Altrep.h>
+
+static R_altrep_class_t dta_grow_string, dta_grow_integer, dta_grow_logical, dta_grow_real, dta_grow_complex, dta_grow_raw, dta_grow_list;
+static Rcomplex NA_COMPLEX = { 0, };
+
+/*
+ALTREP class layout:
+data1 = underlying vector
+data2 = its current length stored as a length-1 REALSXP
+Unless we implement an Unserialize method, this can be changed any time.
+Classes have been released on CRAN with a Serialized_state/Unserialize pair will have to stay as they have been defined in order to keep *.rds files readable.
+*/
+
+static R_xlen_t altall_Length(SEXP x) {
+  return (R_xlen_t)REAL(R_altrep_data2(x))[0];
+}
+
+#define make_inspect_method(classname) \
+  static Rboolean alt##classname##_Inspect( \
+    SEXP x, int pre, int deep, int pvec, \
+    void (*inspect_subtree)(SEXP x, int pre, int deep, int pvec) \
+  ) { \
+    (void)pre; (void)deep; (void)pvec; (void)inspect_subtree; \
+    Rprintf("data.table::growable" #classname "_v0(truelength=%g) ", (double)XLENGTH(R_altrep_data1(x))); \
+    return FALSE; \
+  }
+make_inspect_method(string)
+make_inspect_method(integer)
+make_inspect_method(logical)
+make_inspect_method(real)
+make_inspect_method(complex)
+make_inspect_method(raw)
+make_inspect_method(list)
+#undef make_inspect_method
+
+#define make_dataptr_method(class, accessor) \
+  static void * alt##class##_Dataptr(SEXP x, Rboolean writable) { \
+    (void)writable; \
+    return (void*)accessor(R_altrep_data1(x)); \
+  }
+make_dataptr_method(string, STRING_PTR_RO)
+make_dataptr_method(integer, INTEGER)
+make_dataptr_method(logical, LOGICAL)
+make_dataptr_method(real, REAL)
+make_dataptr_method(complex, COMPLEX)
+make_dataptr_method(raw, RAW)
+make_dataptr_method(list, DATAPTR_RO) // VECTOR_PTR_RO to appear in R-4.5
+#undef make_dataptr_method
+
+static const void * altall_Dataptr_or_null(SEXP x) { return DATAPTR_RO(x); }
+
+// lots of boilerplate, but R calling *_ELT one by one would be far too slow
+#define make_extract_subset_method(class, type, accessor, NA) \
+  static SEXP alt##class##_Extract_subset(SEXP x, SEXP indx, SEXP call) { \
+    (void)call; \
+    indx = PROTECT(coerceVector(indx, REALSXP)); \
+    double * ii = REAL(indx); \
+    R_xlen_t rlen = XLENGTH(indx), mylen = XLENGTH(x); \
+    SEXP ret = PROTECT(allocVector(TYPEOF(x), rlen)); \
+    type *rdata = accessor(ret), *mydata = accessor(x); \
+    for (R_xlen_t i = 0; i < rlen; ++i) \
+      rdata[i] = (ii[i] >= 1 && ii[i] <= mylen) ? mydata[(R_xlen_t)ii[i]-1] : NA; \
+    UNPROTECT(2); \
+    return ret; \
+  }
+make_extract_subset_method(integer, int, INTEGER, NA_INTEGER)
+make_extract_subset_method(logical, int, LOGICAL, NA_LOGICAL)
+make_extract_subset_method(real, double, REAL, NA_REAL)
+make_extract_subset_method(complex, Rcomplex, COMPLEX, NA_COMPLEX)
+make_extract_subset_method(raw, Rbyte, RAW, 0)
+// not implementing the string and list methods because those do require the write barrier and are thus no better than calling *_ELT one by one
+#undef make_extract_subset_method
+
+#define make_elt_method(class, accessor) \
+  static SEXP alt##class##_Elt(SEXP x, R_xlen_t i) { \
+    return accessor(R_altrep_data1(x), i); \
+  }
+make_elt_method(string, STRING_ELT)
+make_elt_method(list, VECTOR_ELT)
+#undef make_elt_method
+
+#define make_set_elt_method(class, accessor) \
+  static void alt##class##_Set_elt(SEXP x, R_xlen_t i, SEXP v) { \
+    accessor(R_altrep_data1(x), i, v); \
+  }
+make_set_elt_method(string, SET_STRING_ELT)
+make_set_elt_method(list, SET_VECTOR_ELT)
+#undef make_set_elt_method
+
+// liked the Extract_subset methods? say hello to Get_region
+#define make_get_region_method(class, type, accessor) \
+  static R_xlen_t alt##class##_Get_region( \
+    SEXP x, R_xlen_t i, R_xlen_t n, type * buf \
+  ) { \
+    R_xlen_t j = 0, mylen = XLENGTH(x); \
+    type * data = accessor(x); \
+    for (; j < n && i < mylen; ++i, ++j) buf[j] = data[i]; \
+    return j; \
+  }
+make_get_region_method(integer, int, INTEGER)
+make_get_region_method(logical, int, LOGICAL)
+make_get_region_method(real, double, REAL)
+make_get_region_method(complex, Rcomplex, COMPLEX)
+make_get_region_method(raw, Rbyte, RAW)
+#undef make_get_region_method
+
+void register_altrep_classes(DllInfo * info) {
+  // Used by the altcomplex_Extract_subset method
+  NA_COMPLEX = (Rcomplex){ .r = NA_REAL, .i = NA_REAL };
+
+  dta_grow_string = R_make_altstring_class("growable_string_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_string, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_string, altstring_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_string, altstring_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_string, altall_Dataptr_or_null);
+  R_set_altstring_Elt_method(dta_grow_string, altstring_Elt);
+  R_set_altstring_Set_elt_method(dta_grow_string, altstring_Set_elt);
+  dta_grow_integer = R_make_altinteger_class("growable_integer_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_integer, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_integer, altinteger_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_integer, altinteger_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_integer, altall_Dataptr_or_null);
+  R_set_altvec_Extract_subset_method(dta_grow_integer, altinteger_Extract_subset);
+  R_set_altinteger_Get_region_method(dta_grow_integer, altinteger_Get_region);
+  dta_grow_logical = R_make_altlogical_class("growable_logical_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_logical, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_logical, altlogical_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_logical, altlogical_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_logical, altall_Dataptr_or_null);
+  R_set_altvec_Extract_subset_method(dta_grow_logical, altlogical_Extract_subset);
+  R_set_altlogical_Get_region_method(dta_grow_logical, altlogical_Get_region);
+  dta_grow_real = R_make_altreal_class("growable_real_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_real, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_real, altreal_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_real, altreal_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_real, altall_Dataptr_or_null);
+  R_set_altvec_Extract_subset_method(dta_grow_real, altreal_Extract_subset);
+  R_set_altreal_Get_region_method(dta_grow_real, altreal_Get_region);
+  dta_grow_complex = R_make_altcomplex_class("growable_complex_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_complex, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_complex, altcomplex_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_complex, altcomplex_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_complex, altall_Dataptr_or_null);
+  R_set_altvec_Extract_subset_method(dta_grow_complex, altcomplex_Extract_subset);
+  R_set_altcomplex_Get_region_method(dta_grow_complex, altcomplex_Get_region);
+  dta_grow_raw = R_make_altraw_class("growable_raw_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_raw, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_raw, altraw_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_raw, altraw_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_raw, altall_Dataptr_or_null);
+  R_set_altvec_Extract_subset_method(dta_grow_raw, altraw_Extract_subset);
+  R_set_altraw_Get_region_method(dta_grow_raw, altraw_Get_region);
+  dta_grow_list = R_make_altlist_class("growable_list_v0", "data.table", info);
+  R_set_altrep_Length_method(dta_grow_list, altall_Length);
+  R_set_altrep_Inspect_method(dta_grow_list, altlist_Inspect);
+  R_set_altvec_Dataptr_method(dta_grow_list, altlist_Dataptr);
+  R_set_altvec_Dataptr_or_null_method(dta_grow_list, altall_Dataptr_or_null);
+  R_set_altlist_Elt_method(dta_grow_list, altlist_Elt);
+  R_set_altlist_Set_elt_method(dta_grow_list, altlist_Set_elt);
+}
+
+static R_altrep_class_t dta_grow_string, dta_grow_integer, dta_grow_logical, dta_grow_real, dta_grow_complex, dta_grow_raw, dta_grow_list;
+
+static R_altrep_class_t type2class(SEXPTYPE type) {
+  switch(type) {
+  case STRSXP:
+    return dta_grow_string;
+  case INTSXP:
+    return dta_grow_integer;
+  case LGLSXP:
+    return dta_grow_logical;
+  case REALSXP:
+    return dta_grow_real;
+  case CPLXSXP:
+    return dta_grow_complex;
+  case RAWSXP:
+    return dta_grow_raw;
+  case VECSXP:
+  case EXPRSXP:
+    return dta_grow_list;
+  default:
+    internal_error(__func__, "Can't create a growable vector of type '%s'", type2char(type));
+  }
+}
+
+SEXP growable_allocate(SEXPTYPE type, R_xlen_t size, R_xlen_t max_size) {
+  SEXP ret = PROTECT(R_new_altrep(type2class(type), R_NilValue, R_NilValue));
+  R_set_altrep_data1(ret, allocVector(type, max_size));
+  R_set_altrep_data2(ret, ScalarReal(size));
+  UNPROTECT(1);
+  return ret;
+}
+
+R_xlen_t growable_max_size(SEXP x) {
+  return XLENGTH(R_altrep_data1(x));
+}
+
+void growable_resize(SEXP x, R_xlen_t newsize) {
+  R_xlen_t max_size;
+  if (newsize > (max_size = growable_max_size(x))) internal_error(
+    __func__, "newsize=%g > max_size=%g",
+    (double)newsize, (double)max_size
+  );
+  REAL(R_altrep_data2(x))[0] = newsize;
+}
+
+Rboolean is_growable(SEXP x) {
+  switch(TYPEOF(x)) {
+  case STRSXP:
+  case INTSXP:
+  case LGLSXP:
+  case REALSXP:
+  case CPLXSXP:
+  case RAWSXP:
+  case VECSXP:
+    return R_altrep_inherits(x, type2class(TYPEOF(x)));
+  default:
+    return FALSE;
+  }
+}
+
+SEXP make_growable(SEXP x) {
+  SEXP ret = PROTECT(R_new_altrep(type2class(TYPEOF(x)), R_NilValue, R_NilValue));
+  R_set_altrep_data1(ret, x);
+  R_set_altrep_data2(ret, ScalarReal(XLENGTH(x)));
+  SHALLOW_DUPLICATE_ATTRIB(ret, x);
+  UNPROTECT(1);
+  return ret;
+}
+
+#endif

src/init.c

@@ -203,8 +203,12 @@ void attribute_visible R_init_data_table(DllInfo *info)
 
   SEXP tmp = PROTECT(allocVector(INTSXP,2));
   if (LENGTH(tmp)!=2) error(_("Checking LENGTH(allocVector(INTSXP,2)) [%d] is 2 %s"), LENGTH(tmp), msg);
+#if R_VERSION >= R_Version(4, 3, 0)
+  register_altrep_classes(info);
+#else
   // Use (long long) to cast R_xlen_t to a fixed type to robustly avoid -Wformat compiler warnings, see #5768
   if (TRUELENGTH(tmp)!=0) error(_("Checking TRUELENGTH(allocVector(INTSXP,2)) [%lld] is 0 %s"), (long long)TRUELENGTH(tmp), msg);
+#endif
   UNPROTECT(1);
 
   // According to IEEE (http://en.wikipedia.org/wiki/IEEE_754-1985#Zero) we can rely on 0.0 being all 0 bits.

src/reorder.c

@@ -24,13 +24,17 @@ SEXP reorder(SEXP x, SEXP order)
         error(_("Column %d is length %d which differs from length of column 1 (%d). Invalid data.table."), i+1, length(v), nrow);
       if (SIZEOF(v) > maxSize)
         maxSize=SIZEOF(v);
+#ifndef USE_GROWABLE_ALTREP
       if (ALTREP(v)) SET_VECTOR_ELT(x, i, copyAsPlain(v));
+#endif
     }
     copySharedColumns(x); // otherwise two columns which point to the same vector would be reordered and then re-reordered, issues linked in PR#3768
   } else {
     if (SIZEOF(x)!=4 && SIZEOF(x)!=8 && SIZEOF(x)!=16 && SIZEOF(x)!=1)
       error(_("reorder accepts vectors but this non-VECSXP is type '%s' which isn't yet supported (SIZEOF=%zu)"), type2char(TYPEOF(x)), SIZEOF(x));
+#ifndef USE_GROWABLE_ALTREP
     if (ALTREP(x)) internal_error(__func__, "cannot reorder an ALTREP vector. Please see NEWS item 2 in v1.11.4"); // # nocov
+#endif
     maxSize = SIZEOF(x);
     nrow = length(x);
     ncol = 1;
@@ -40,7 +44,9 @@ SEXP reorder(SEXP x, SEXP order)
   if (length(order) != nrow)
     error("nrow(x)[%d]!=length(order)[%d]", nrow, length(order)); // # notranslate
   int nprotect = 0;
+#ifndef USE_GROWABLE_ALTREP
   if (ALTREP(order)) { order=PROTECT(copyAsPlain(order)); nprotect++; }  // TODO: if it's an ALTREP sequence some optimizations are possible rather than expand
+#endif
 
   const int *restrict idx = INTEGER(order);
   int i=0;

src/utils.c

@@ -202,6 +202,9 @@ inline bool INHERITS(SEXP x, SEXP char_) {
   return false;
 }
 
+#ifdef USE_GROWABLE_ALTREP
+SEXP copyAsPlain(SEXP x) { return duplicate(x); }
+#else
 SEXP copyAsPlain(SEXP x) {
   // v1.12.2 and before used standard R duplicate() to do this. But duplicate() is not guaranteed to not return an ALTREP.
   // e.g. ALTREP 'wrapper' on factor column (with materialized INTSXP) in package VIM under example(hotdeck)
@@ -256,6 +259,7 @@ SEXP copyAsPlain(SEXP x) {
   UNPROTECT(1);
   return ans;
 }
+#endif
 
 void copySharedColumns(SEXP x) {
   const int ncol = length(x);
@@ -266,7 +270,12 @@ void copySharedColumns(SEXP x) {
   int nShared=0;
   for (int i=0; i<ncol; ++i) {
     SEXP thiscol = xp[i];
-    if (ALTREP(thiscol) || hash_lookup(marks, thiscol, 0)<0) {
+    if (
+      hash_lookup(marks, thiscol, 0)<0
+#ifndef USE_GROWABLE_ALTREP
+      || ALTREP(thiscol)
+#endif
+    ) {
       shared[i] = true;  // we mark ALTREP as 'shared' too, whereas 'tocopy' would be better word to use for ALTREP
       nShared++;
     } else {