strcoll: improve performance by removing the cache (#15884)

this is a path that should solve bug 15884. It complains about the performance
of strcoll(). It was found out that the runtime of strcoll() is actually bound
to strlen which is needed for calculating the size of a cache that was
installed to improve the comparison performance.

The idea for this patch was that the cache is only useful in rare cases
(strings of same length and same first-level-chars) and that it would be
better to avoid memory allocation at all. To prove this I wrote a performance
test bench-strcoll.c with test data in benchtests-strcoll.tar.gz. Also
modifications in benchtests/Makefile and localedata/Makefile are necessary to
make it work.

After removing the cache the strcoll method showed the predicted behavior
(getting slightly faster) in all but the test case for hindi word sorting.
This was due the hindi text having much more equal words than the other ones.
For equal strings the performance was worse since all comparison levels were
run through and from the second level on the cache improved the comparison
performance of the original version.

Therefore I added a bytewise test via strcmp iff the first level comparison
found that both strings did match because in this case it is very likely that
equal strings are compared. This solved the problem with the hindi test case
and improved the performance of the others.

Performance comparison:

glibc files     -33.77%
vi_VN.UTF-8     -34.12%
en_US.UTF-8     -42.42%
ar_SA.UTF-8     -27.49%
zh_CN.UTF-8     +07.90%
cs_CZ.UTF-8     -29.67%
en_GB.UTF-8     -28.50%
da_DK.UTF-8     -36.57%
pl_PL.UTF-8     -39.31%
fr_FR.UTF-8     -28.57%
pt_PT.UTF-8     -22.82%
el_GR.UTF-8     -26.77%
ru_RU.UTF-8     -35.81%
iw_IL.UTF-8     -35.34%
es_ES.UTF-8     -34.46%
hi_IN.UTF-8     -00.38%
sv_SE.UTF-8     -36.99%
hu_HU.UTF-8     -16.35%
tr_TR.UTF-8     -27.80%
is_IS.UTF-8     -33.24%
it_IT.UTF-8     -24.39%
sr_RS.UTF-8     -37.55%
ja_JP.UTF-8     +02.84%

1 files changed, 26 insertions, 318 deletions

diff --git a/string/strcoll_l.c b/string/strcoll_l.c
index d4f42a3..6677eaf 100644
--- a/string/strcoll_l.c
+++ b/string/strcoll_l.c
@@ -22,7 +22,6 @@
 #include <locale.h>
 #include <stddef.h>
 #include <stdint.h>
-#include <stdlib.h>
 #include <string.h>
 #include <sys/param.h>
 
@@ -55,8 +54,6 @@ typedef struct
   size_t backw;			/* Current Backward sequence index.  */
   size_t backw_stop;		/* Index where the backward sequences stop.  */
   const USTRING_TYPE *us;	/* The string.  */
-  int32_t *idxarr;		/* Array to cache weight indices.  */
-  unsigned char *rulearr;	/* Array to cache rules.  */
   unsigned char rule;		/* Saved rule for the first sequence.  */
   int32_t idx;			/* Index to weight of the current sequence.  */
   int32_t save_idx;		/* Save looked up index of a forward
@@ -65,179 +62,9 @@ typedef struct
   const USTRING_TYPE *back_us;	/* Beginning of the backward sequence.  */
 } coll_seq;
 
-/* Get next sequence.  The weight indices are cached, so we don't need to
-   traverse the string.  */
-static void
-get_next_seq_cached (coll_seq *seq, int nrules, int pass,
-		     const unsigned char *rulesets,
-		     const USTRING_TYPE *weights)
-{
-  size_t val = seq->val = 0;
-  int len = seq->len;
-  size_t backw_stop = seq->backw_stop;
-  size_t backw = seq->backw;
-  size_t idxcnt = seq->idxcnt;
-  size_t idxmax = seq->idxmax;
-  size_t idxnow = seq->idxnow;
-  unsigned char *rulearr = seq->rulearr;
-  int32_t *idxarr = seq->idxarr;
-
-  while (len == 0)
-    {
-      ++val;
-      if (backw_stop != ~0ul)
-	{
-	  /* There is something pushed.  */
-	  if (backw == backw_stop)
-	    {
-	      /* The last pushed character was handled.  Continue
-		 with forward characters.  */
-	      if (idxcnt < idxmax)
-		{
-		  idxnow = idxcnt;
-		  backw_stop = ~0ul;
-		}
-	      else
-		{
-		  /* Nothing any more.  The backward sequence
-		     ended with the last sequence in the string.  */
-		  idxnow = ~0ul;
-		  break;
-		}
-	    }
-	  else
-	    idxnow = --backw;
-	}
-      else
-	{
-	  backw_stop = idxcnt;
-
-	  while (idxcnt < idxmax)
-	    {
-	      if ((rulesets[rulearr[idxcnt] * nrules + pass]
-		   & sort_backward) == 0)
-		/* No more backward characters to push.  */
-		break;
-	      ++idxcnt;
-	    }
-
-	  if (backw_stop == idxcnt)
-	    {
-	      /* No sequence at all or just one.  */
-	      if (idxcnt == idxmax)
-		/* Note that LEN is still zero.  */
-		break;
-
-	      backw_stop = ~0ul;
-	      idxnow = idxcnt++;
-	    }
-	  else
-	    /* We pushed backward sequences.  */
-	    idxnow = backw = idxcnt - 1;
-	}
-      len = weights[idxarr[idxnow]++];
-    }
-
-  /* Update the structure.  */
-  seq->val = val;
-  seq->len = len;
-  seq->backw_stop = backw_stop;
-  seq->backw = backw;
-  seq->idxcnt = idxcnt;
-  seq->idxnow = idxnow;
-}
-
 /* Get next sequence.  Traverse the string as required.  */
 static void
 get_next_seq (coll_seq *seq, int nrules, const unsigned char *rulesets,
-	      const USTRING_TYPE *weights, const int32_t *table,
-	      const USTRING_TYPE *extra, const int32_t *indirect)
-{
-  size_t val = seq->val = 0;
-  int len = seq->len;
-  size_t backw_stop = seq->backw_stop;
-  size_t backw = seq->backw;
-  size_t idxcnt = seq->idxcnt;
-  size_t idxmax = seq->idxmax;
-  size_t idxnow = seq->idxnow;
-  unsigned char *rulearr = seq->rulearr;
-  int32_t *idxarr = seq->idxarr;
-  const USTRING_TYPE *us = seq->us;
-
-  while (len == 0)
-    {
-      ++val;
-      if (backw_stop != ~0ul)
-	{
-	  /* There is something pushed.  */
-	  if (backw == backw_stop)
-	    {
-	      /* The last pushed character was handled.  Continue
-		 with forward characters.  */
-	      if (idxcnt < idxmax)
-		{
-		  idxnow = idxcnt;
-		  backw_stop = ~0ul;
-		}
-	      else
-		/* Nothing any more.  The backward sequence ended with
-		   the last sequence in the string.  Note that LEN
-		   is still zero.  */
-		break;
-	    }
-	  else
-	    idxnow = --backw;
-	}
-      else
-	{
-	  backw_stop = idxmax;
-
-	  while (*us != L('\0'))
-	    {
-	      int32_t tmp = findidx (table, indirect, extra, &us, -1);
-	      rulearr[idxmax] = tmp >> 24;
-	      idxarr[idxmax] = tmp & 0xffffff;
-	      idxcnt = idxmax++;
-
-	      if ((rulesets[rulearr[idxcnt] * nrules]
-		   & sort_backward) == 0)
-		/* No more backward characters to push.  */
-		break;
-	      ++idxcnt;
-	    }
-
-	  if (backw_stop >= idxcnt)
-	    {
-	      /* No sequence at all or just one.  */
-	      if (idxcnt == idxmax || backw_stop > idxcnt)
-		/* Note that LEN is still zero.  */
-		break;
-
-	      backw_stop = ~0ul;
-	      idxnow = idxcnt;
-	    }
-	  else
-	    /* We pushed backward sequences.  */
-	    idxnow = backw = idxcnt - 1;
-	}
-      len = weights[idxarr[idxnow]++];
-    }
-
-  /* Update the structure.  */
-  seq->val = val;
-  seq->len = len;
-  seq->backw_stop = backw_stop;
-  seq->backw = backw;
-  seq->idxcnt = idxcnt;
-  seq->idxmax = idxmax;
-  seq->idxnow = idxnow;
-  seq->us = us;
-}
-
-/* Get next sequence.  Traverse the string as required.  This function does not
-   set or use any index or rule cache.  */
-static void
-get_next_seq_nocache (coll_seq *seq, int nrules, const unsigned char *rulesets,
 		      const USTRING_TYPE *weights, const int32_t *table,
 		      const USTRING_TYPE *extra, const int32_t *indirect,
 		      int pass)
@@ -366,10 +193,9 @@ get_next_seq_nocache (coll_seq *seq, int nrules, const unsigned char *rulesets,
   seq->idx = idx;
 }
 
-/* Compare two sequences.  This version does not use the index and rules
-   cache.  */
+/* Compare two sequences.  */
 static int
-do_compare_nocache (coll_seq *seq1, coll_seq *seq2, int position,
+do_compare (coll_seq *seq1, coll_seq *seq2, int position,
 		    const USTRING_TYPE *weights)
 {
   int seq1len = seq1->len;
@@ -417,56 +243,6 @@ out:
   return result;
 }
 
-/* Compare two sequences using the index cache.  */
-static int
-do_compare (coll_seq *seq1, coll_seq *seq2, int position,
-	    const USTRING_TYPE *weights)
-{
-  int seq1len = seq1->len;
-  int seq2len = seq2->len;
-  size_t val1 = seq1->val;
-  size_t val2 = seq2->val;
-  int32_t *idx1arr = seq1->idxarr;
-  int32_t *idx2arr = seq2->idxarr;
-  int idx1now = seq1->idxnow;
-  int idx2now = seq2->idxnow;
-  int result = 0;
-
-  /* Test for position if necessary.  */
-  if (position && val1 != val2)
-    {
-      result = val1 > val2 ? 1 : -1;
-      goto out;
-    }
-
-  /* Compare the two sequences.  */
-  do
-    {
-      if (weights[idx1arr[idx1now]] != weights[idx2arr[idx2now]])
-	{
-	  /* The sequences differ.  */
-	  result = weights[idx1arr[idx1now]] - weights[idx2arr[idx2now]];
-	  goto out;
-	}
-
-      /* Increment the offsets.  */
-      ++idx1arr[idx1now];
-      ++idx2arr[idx2now];
-
-      --seq1len;
-      --seq2len;
-    }
-  while (seq1len > 0 && seq2len > 0);
-
-  if (position && seq1len != seq2len)
-    result = seq1len - seq2len;
-
-out:
-  seq1->len = seq1len;
-  seq2->len = seq2len;
-  return result;
-}
-
 int
 STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
 {
@@ -483,6 +259,10 @@ STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
   if (nrules == 0)
     return STRCMP (s1, s2);
 
+  /* Catch empty strings.  */
+  if (__glibc_unlikely (*s1 == '\0') || __glibc_unlikely (*s2 == '\0'))
+    return (*s1 != '\0') - (*s2 != '\0');
+
   rulesets = (const unsigned char *)
     current->values[_NL_ITEM_INDEX (_NL_COLLATE_RULESETS)].string;
   table = (const int32_t *)
@@ -499,65 +279,12 @@ STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
   assert (((uintptr_t) extra) % __alignof__ (extra[0]) == 0);
   assert (((uintptr_t) indirect) % __alignof__ (indirect[0]) == 0);
 
-  /* We need this a few times.  */
-  size_t s1len = STRLEN (s1);
-  size_t s2len = STRLEN (s2);
-
-  /* Catch empty strings.  */
-  if (__glibc_unlikely (s1len == 0) || __glibc_unlikely (s2len == 0))
-    return (s1len != 0) - (s2len != 0);
-
-  /* Perform the first pass over the string and while doing this find
-     and store the weights for each character.  Since we want this to
-     be as fast as possible we are using `alloca' to store the temporary
-     values.  But since there is no limit on the length of the string
-     we have to use `malloc' if the string is too long.  We should be
-     very conservative here.
-
-     Please note that the localedef programs makes sure that `position'
-     is not used at the first level.  */
+  int result = 0, rule = 0;
 
   coll_seq seq1, seq2;
-  bool use_malloc = false;
-  int result = 0;
-
   memset (&seq1, 0, sizeof (seq1));
   seq2 = seq1;
 
-  size_t size_max = SIZE_MAX / (sizeof (int32_t) + 1);
-
-  if (MIN (s1len, s2len) > size_max
-      || MAX (s1len, s2len) > size_max - MIN (s1len, s2len))
-    {
-      /* If the strings are long enough to cause overflow in the size request,
-         then skip the allocation and proceed with the non-cached routines.  */
-    }
-  else if (! __libc_use_alloca ((s1len + s2len) * (sizeof (int32_t) + 1)))
-    {
-      seq1.idxarr = (int32_t *) malloc ((s1len + s2len) * (sizeof (int32_t) + 1));
-
-      /* If we failed to allocate memory, we leave everything as NULL so that
-	 we use the nocache version of traversal and comparison functions.  */
-      if (seq1.idxarr != NULL)
-	{
-	  seq2.idxarr = &seq1.idxarr[s1len];
-	  seq1.rulearr = (unsigned char *) &seq2.idxarr[s2len];
-	  seq2.rulearr = &seq1.rulearr[s1len];
-	  use_malloc = true;
-	}
-    }
-  else
-    {
-      seq1.idxarr = (int32_t *) alloca (s1len * sizeof (int32_t));
-      seq2.idxarr = (int32_t *) alloca (s2len * sizeof (int32_t));
-      seq1.rulearr = (unsigned char *) alloca (s1len);
-      seq2.rulearr = (unsigned char *) alloca (s2len);
-    }
-
-  int rule = 0;
-
-  /* Cache values in the first pass and if needed, use them in subsequent
-     passes.  */
   for (int pass = 0; pass < nrules; ++pass)
     {
       seq1.idxcnt = 0;
@@ -575,64 +302,45 @@ STRCOLL (const STRING_TYPE *s1, const STRING_TYPE *s2, __locale_t l)
       seq2.us = (const USTRING_TYPE *) s2;
 
       /* We assume that if a rule has defined `position' in one section
-	 this is true for all of them.  */
+	 this is true for all of them.  Please note that the localedef programs
+	 makes sure that `position' is not used at the first level.  */
+
       int position = rulesets[rule * nrules + pass] & sort_position;
 
       while (1)
 	{
-	  if (__glibc_unlikely (seq1.idxarr == NULL))
-	    {
-	      get_next_seq_nocache (&seq1, nrules, rulesets, weights, table,
+	  get_next_seq (&seq1, nrules, rulesets, weights, table,
 				    extra, indirect, pass);
-	      get_next_seq_nocache (&seq2, nrules, rulesets, weights, table,
+	  get_next_seq (&seq2, nrules, rulesets, weights, table,
 				    extra, indirect, pass);
-	    }
-	  else if (pass == 0)
-	    {
-	      get_next_seq (&seq1, nrules, rulesets, weights, table, extra,
-			    indirect);
-	      get_next_seq (&seq2, nrules, rulesets, weights, table, extra,
-			    indirect);
-	    }
-	  else
-	    {
-	      get_next_seq_cached (&seq1, nrules, pass, rulesets, weights);
-	      get_next_seq_cached (&seq2, nrules, pass, rulesets, weights);
-	    }
-
 	  /* See whether any or both strings are empty.  */
 	  if (seq1.len == 0 || seq2.len == 0)
 	    {
 	      if (seq1.len == seq2.len)
-		/* Both ended.  So far so good, both strings are equal
-		   at this level.  */
-		break;
+		{
+		  /* Both strings ended and are equal at this level.  Do a
+		     byte-level comparison to ensure that we don't waste time
+		     going through multiple passes for totally equal strings
+		     before proceeding to subsequent passes.  */
+		  if (pass == 0 && STRCMP (s1, s2) == 0)
+		    return result;
+		  else
+		    break;
+	        }
 
 	      /* This means one string is shorter than the other.  Find out
 		 which one and return an appropriate value.  */
-	      result = seq1.len == 0 ? -1 : 1;
-	      goto free_and_return;
+	      return seq1.len == 0 ? -1 : 1;
 	    }
 
-	  if (__glibc_unlikely (seq1.idxarr == NULL))
-	    result = do_compare_nocache (&seq1, &seq2, position, weights);
-	  else
-	    result = do_compare (&seq1, &seq2, position, weights);
+	  result = do_compare (&seq1, &seq2, position, weights);
 	  if (result != 0)
-	    goto free_and_return;
+	    return result;
 	}
 
-      if (__glibc_likely (seq1.rulearr != NULL))
-	rule = seq1.rulearr[0];
-      else
-	rule = seq1.rule;
+      rule = seq1.rule;
     }
 
-  /* Free the memory if needed.  */
- free_and_return:
-  if (use_malloc)
-    free (seq1.idxarr);
-
   return result;
 }
 libc_hidden_def (STRCOLL)