From 52b899d60d8c61b8affe0206014173912de94940 Mon Sep 17 00:00:00 2001
From: Roberto Ierusalimschy <roberto@inf.puc-rio.br>
Date: Fri, 24 Nov 2023 14:41:07 -0300
Subject: [PATCH] Simpler coding for new representation for arrays

With the tags comming first in a cell, we can define the whole cell
as a C type and let C do part of the address computations.
---
 lobject.h |  2 +-
 ltable.c  | 46 +++++++++++++++++++++++++++-------------------
 ltable.h  | 38 ++++++++------------------------------
 3 files changed, 36 insertions(+), 50 deletions(-)

diff --git a/lobject.h b/lobject.h
index 8688a842..6da50215 100644
--- a/lobject.h
+++ b/lobject.h
@@ -762,7 +762,7 @@ typedef union Node {
 #define setnorealasize(t)	((t)->flags |= BITRAS)
 
 
-typedef union ArrayCell ArrayCell;
+typedef struct ArrayCell ArrayCell;
 
 
 typedef struct Table {
diff --git a/ltable.c b/ltable.c
index c9f66b3c..d3e90696 100644
--- a/ltable.c
+++ b/ltable.c
@@ -541,29 +541,28 @@ static int numusehash (const Table *t, unsigned int *nums, unsigned int *pna) {
 
 
 /*
-** Convert an "abstract size" (number of values in an array) to
-** "concrete size" (number of cell elements in the array). Cells
-** do not need to be full; we only must make sure it has the values
-** needed and its 'tag' element. So, we compute the concrete tag index
-** and the concrete value index of the last element, get their maximum
-** and adds 1.
+** Convert an "abstract size" (number of slots in an array) to
+** "concrete size" (number of bytes in the array).
+** If the abstract size is not a multiple of NM, the last cell is
+** incomplete, so we don't need to allocate memory for the whole cell.
+** 'extra' computes how many values are not needed in that last cell.
+** It will be zero when 'size' is a multiple of NM, and from there it
+** increases as 'size' decreases, up to (NM - 1).
 */
-static unsigned int concretesize (unsigned int size) {
-  if (size == 0) return 0;
-  else {
-    unsigned int ts = TagIndex(size - 1);
-    unsigned int vs = ValueIndex(size - 1);
-    return ((ts >= vs) ? ts : vs) + 1;
-  }
+static size_t concretesize (unsigned int size) {
+  unsigned int numcells = (size + NM - 1) / NM;   /* (size / NM) rounded up */
+  unsigned int extra = NM - 1 - ((size + NM - 1) % NM);
+  return numcells * sizeof(ArrayCell) - extra * sizeof(Value);
 }
 
 
 static ArrayCell *resizearray (lua_State *L , Table *t,
                                unsigned int oldasize,
                                unsigned int newasize) {
-  oldasize = concretesize(oldasize);
-  newasize = concretesize(newasize);
-  return luaM_reallocvector(L, t->array, oldasize, newasize, ArrayCell);
+  size_t oldasizeb = concretesize(oldasize);
+  size_t newasizeb = concretesize(newasize);
+  void *a = luaM_reallocvector(L, t->array, oldasizeb, newasizeb, lu_byte);
+  return cast(ArrayCell*, a);
 }
 
 
@@ -747,10 +746,19 @@ Table *luaH_new (lua_State *L) {
 }
 
 
+/*
+** Frees a table. The assert ensures the correctness of 'concretesize',
+** checking its result against the address of the last element in the
+** array part of the table, computed abstractly.
+*/
 void luaH_free (lua_State *L, Table *t) {
-  unsigned ps = concretesize(luaH_realasize(t));
+  unsigned int realsize = luaH_realasize(t);
+  size_t sizeb = concretesize(realsize);
+  lua_assert((sizeb == 0 && realsize == 0) ||
+             cast_charp(t->array) + sizeb - sizeof(Value) ==
+             cast_charp(getArrVal(t, realsize - 1)));
   freehash(L, t);
-  luaM_freearray(L, t->array, ps);
+  luaM_freemem(L, t->array, sizeb);
   luaM_free(L, t);
 }
 
@@ -944,7 +952,7 @@ TString *luaH_getstrkey (Table *t, TString *key) {
 ** main search function
 */
 int luaH_get (Table *t, const TValue *key, TValue *res) {
-  const TValue *slot; 
+  const TValue *slot;
   switch (ttypetag(key)) {
     case LUA_VSHRSTR:
       slot = luaH_Hgetshortstr(t, tsvalue(key));
diff --git a/ltable.h b/ltable.h
index d488a1f7..5581efb1 100644
--- a/ltable.h
+++ b/ltable.h
@@ -69,44 +69,22 @@
 
 /*
 ** The array part of a table is represented by an array of cells.
-** Each cell is composed of (NM + 1) elements, and each element has the
-** type 'ArrayCell'.  In each cell, only one element has the variant
-** 'tag', while the other NM elements have the variant 'value'. The
-** array in the 'tag' element holds the tags of the other elements in
-** that cell.
+** Each cell is composed of NM tags followed by NM values, so that
+** no space is wasted in padding.
 */
-#define NM      ((unsigned int)sizeof(Value))
+#define NM      cast_uint(sizeof(Value))
 
-union ArrayCell {
-  unsigned char tag[NM];
-  Value value;
+struct ArrayCell {
+  lu_byte tag[NM];
+  Value value[NM];
 };
 
 
-/*
-** 'NMTag' defines which cell element has the tags; that could be any
-** value between 0 (tags come before all values) and NM (tags come after
-** all values).
-*/
-#define NMTag     0
-
-
-/*
-** Computes the concrete index that holds the tag of abstract index 'i'
-*/
-#define TagIndex(i)     (((i)/NM * (NM + 1u)) + NMTag)
-
-/*
-** Computes the concrete index that holds the value of abstract index 'i'
-*/
-#define ValueIndex(i)   ((i) + (((i) + (NM - NMTag))/NM))
-
-
 /* Computes the address of the tag for the abstract index 'k' */
-#define getArrTag(t,k)	(&(t)->array[TagIndex(k)].tag[(k)%NM])
+#define getArrTag(t,k)	(&(t)->array[(k)/NM].tag[(k)%NM])
 
 /* Computes the address of the value for the abstract index 'k' */
-#define getArrVal(t,k)	(&(t)->array[ValueIndex(k)].value)
+#define getArrVal(t,k)	(&(t)->array[(k)/NM].value[(k)%NM])
 
 
 /*