This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH v3] Generalize get_most_common_single_value to return k_th value & count


On 7/16/19 8:53 AM, luoxhu wrote:
> Currently get_most_common_single_value could only return the max hist
> <value, count>, add qsort to enable this function return nth value.
> Rename it to get_nth_most_common_value.
> 
> v3 Changes:
>  1. Move sort to profile.c after loading values from disk.  Simplify
>     get_nth_most_common_value.
>  2. Make qsort stable with value check if count is same.
>  3. Other comments from v2.

Hi.

Thank you for the updated patch. But you are probably using a line wrapping
in your email client and I can't apply the patch. 

> 
> gcc/ChangeLog:
> 
>     2019-07-15  Xiong Hu Luo  <luoxhu@linux.ibm.com>
> 
>     * ipa-profile.c (get_most_common_single_value): Use
>     get_nth_most_common_value.
>     * profile.c (struct value_count_t): New struct.
>     (cmp_counts): New function.
>     (sort_hist_value): New function.
>     (compute_value_histograms): Call sort_hist_value to sort the
>     values after loading from disk.
>     * value-prof.c (get_most_common_single_value): Rename to ...
>     get_nth_most_common_value.  Add input params n, return
>     the n_th value and count.
>     (gimple_divmod_fixed_value_transform): Use
>     get_nth_most_common_value.
>     (gimple_ic_transform): Likewise.
>     (gimple_stringops_transform): Likewise.
>     * value-prof.h (get_most_common_single_value): Add input params
>     n, default to 0.
> ---
>  gcc/ipa-profile.c |  4 +--
>  gcc/profile.c     | 74 +++++++++++++++++++++++++++++++++++++++++++++++
>  gcc/value-prof.c  | 53 +++++++++++++++------------------
>  gcc/value-prof.h  |  9 +++---
>  4 files changed, 103 insertions(+), 37 deletions(-)
> 
> diff --git a/gcc/ipa-profile.c b/gcc/ipa-profile.c
> index 1fb939b73d0..970dba39c80 100644
> --- a/gcc/ipa-profile.c
> +++ b/gcc/ipa-profile.c
> @@ -192,8 +192,8 @@ ipa_profile_generate_summary (void)
>            if (h)
>              {
>                gcov_type val, count, all;
> -              if (get_most_common_single_value (NULL, "indirect call",
> -                            h, &val, &count, &all))
> +              if (get_nth_most_common_value (NULL, "indirect call", h,
> +                             &val, &count, &all))
>              {
>                struct cgraph_edge * e = node->get_edge (stmt);
>                if (e && !e->indirect_unknown_callee)
> diff --git a/gcc/profile.c b/gcc/profile.c
> index 441cb8eb183..54780b44859 100644
> --- a/gcc/profile.c
> +++ b/gcc/profile.c
> @@ -743,6 +743,74 @@ compute_branch_probabilities (unsigned cfg_checksum, unsigned lineno_checksum)
>    free_aux_for_blocks ();
>  }
> 
> +struct value_count_t
> +{
> +  gcov_type value;
> +  gcov_type count;
> +};
> +

Please make a brief comment here.

> +static int
> +cmp_counts (const void *v1, const void *v2)
> +{
> +  const value_count_t *h1 = (const value_count_t *) v1;
> +  const value_count_t *h2 = (const value_count_t *) v2;
> +  if (h1->count < h2->count)
> +    return 1;
> +  if (h1->count > h2->count)
> +    return -1;
> +  if (h1->count == h2->count)
> +    {
> +      if (h1->value < h2->value)
> +    return 1;
> +      if (h1->value > h2->value)
> +    return -1;
> +    }

What about something easier like:
return (h1->count != h2->count ? h2->count - h1->count : h2->value - h1->value) ?

> +  /* There may be two entries with same count as well as value very unlikely
> +     in a multi-threaded instrumentation.  But the memory layout of the {value,
> +     count} tuple can be different.  The function will return K-th most
> +     common value.  */
> +  return 0;
> +}
> +
> +/* Sort the histogram value and count for TOPN and INDIR_CALL type.  */
> +
> +static bool
> +sort_hist_value (histogram_value hist)
> +{
> +  auto_vec<struct value_count_t, GCOV_TOPN_VALUES> value_vec;
> +  struct value_count_t temp;
> +  unsigned i;
> +
> +  if (hist->hvalue.counters[2] == -1)
> +    return false;
> +
> +  gcc_assert (hist->type == HIST_TYPE_TOPN_VALUES
> +          || hist->type == HIST_TYPE_INDIR_CALL);
> +
> +  gcc_assert (hist->n_counters == GCOV_TOPN_VALUES_COUNTERS);
> +
> +  for (i = 0; i < GCOV_TOPN_VALUES; i++)
> +    {
> +      gcov_type v = hist->hvalue.counters[2 * i + 1];
> +      gcov_type c = hist->hvalue.counters[2 * i + 2];
> +
> +      temp.value = v;
> +      temp.count = c;
> +
> +      value_vec.safe_push (temp);
> +    }
> +
> +  value_vec.qsort (cmp_counts);
> +
> +  gcc_assert (value_vec.length () == GCOV_TOPN_VALUES);
> +
> +  for (i = 0; i < GCOV_TOPN_VALUES; i++)
> +    {
> +      hist->hvalue.counters[2 * i + 1] = value_vec[i].value;
> +      hist->hvalue.counters[2 * i + 2] = value_vec[i].count;
> +    }
> +  return true;
> +}

Well, to be honest I don't like so much manipulation for sorting of 4 values.
Please implement a simple bubble sort which will do sorting in place and
will finish as soon as the sequence is sorted. Using qsort is overkill
in my opinion.

Thanks,
Martin

>  /* Load value histograms values whose description is stored in VALUES array
>     from .gcda file.
> 
> @@ -808,6 +876,12 @@ compute_value_histograms (histogram_values values, unsigned cfg_checksum,
>          else
>            hist->hvalue.counters[j] = 0;
> 
> +      if (hist->type == HIST_TYPE_TOPN_VALUES
> +      || hist->type == HIST_TYPE_INDIR_CALL)
> +    {
> +      sort_hist_value (hist);
> +    }
> +
>        /* Time profiler counter is not related to any statement,
>           so that we have to read the counter and set the value to
>           the corresponding call graph node.  */
> diff --git a/gcc/value-prof.c b/gcc/value-prof.c
> index 32e6ddd8165..97e4ae18ba3 100644
> --- a/gcc/value-prof.c
> +++ b/gcc/value-prof.c
> @@ -713,45 +713,38 @@ gimple_divmod_fixed_value (gassign *stmt, tree value, profile_probability prob,
>    return tmp2;
>  }
> 
> -/* Return most common value of TOPN_VALUE histogram.  If
> -   there's a unique value, return true and set VALUE and COUNT
> +/* Return the n-th value count of TOPN_VALUE histogram.  If
> +   there's a value, return true and set VALUE and COUNT
>     arguments.  */
> 
>  bool
> -get_most_common_single_value (gimple *stmt, const char *counter_type,
> -                  histogram_value hist,
> -                  gcov_type *value, gcov_type *count,
> -                  gcov_type *all)
> +get_nth_most_common_value (gimple *stmt, const char *counter_type,
> +               histogram_value hist, gcov_type *value,
> +               gcov_type *count, gcov_type *all, unsigned n)
>  {
>    if (hist->hvalue.counters[2] == -1)
>      return false;
> 
> +  gcc_assert (n < GCOV_TOPN_VALUES);
> +
>    *count = 0;
>    *value = 0;
> 
>    gcov_type read_all = hist->hvalue.counters[0];
> 
> -  for (unsigned i = 0; i < GCOV_TOPN_VALUES; i++)
> -    {
> -      gcov_type v = hist->hvalue.counters[2 * i + 1];
> -      gcov_type c = hist->hvalue.counters[2 * i + 2];
> -
> -      /* Indirect calls can't be vereified.  */
> -      if (stmt && check_counter (stmt, counter_type, &c, &read_all,
> -                 gimple_bb (stmt)->count))
> -    return false;
> +  gcov_type v = hist->hvalue.counters[2 * n + 1];
> +  gcov_type c = hist->hvalue.counters[2 * n + 2];
> 
> -      *all = read_all;
> +  /* Indirect calls can't be vereified.  */
> +  if (stmt
> +      && check_counter (stmt, counter_type, &c, &read_all,
> +            gimple_bb (stmt)->count))
> +    return false;
> 
> -      if (c > *count)
> -    {
> -      *value = v;
> -      *count = c;
> -    }
> -      else if (c == *count && v > *value)
> -    *value = v;
> -    }
> +  *all = read_all;
> 
> +  *value = v;
> +  *count = c;
>    return true;
>  }
> 
> @@ -784,8 +777,8 @@ gimple_divmod_fixed_value_transform (gimple_stmt_iterator *si)
>    if (!histogram)
>      return false;
> 
> -  if (!get_most_common_single_value (stmt, "divmod", histogram, &val, &count,
> -                     &all))
> +  if (!get_nth_most_common_value (stmt, "divmod", histogram, &val, &count,
> +                  &all))
>      return false;
> 
>    value = histogram->hvalue.value;
> @@ -1439,8 +1432,8 @@ gimple_ic_transform (gimple_stmt_iterator *gsi)
>    if (!histogram)
>      return false;
> 
> -  if (!get_most_common_single_value (NULL, "indirect call", histogram, &val,
> -                     &count, &all))
> +  if (!get_nth_most_common_value (NULL, "indirect call", histogram, &val,
> +                  &count, &all))
>      return false;
> 
>    if (4 * count <= 3 * all)
> @@ -1658,8 +1651,8 @@ gimple_stringops_transform (gimple_stmt_iterator *gsi)
>    if (!histogram)
>      return false;
> 
> -  if (!get_most_common_single_value (stmt, "stringops", histogram, &val,
> -                     &count, &all))
> +  if (!get_nth_most_common_value (stmt, "stringops", histogram, &val, &count,
> +                  &all))
>      return false;
> 
>    gimple_remove_histogram_value (cfun, stmt, histogram);
> diff --git a/gcc/value-prof.h b/gcc/value-prof.h
> index ca846d08cbd..1078722163b 100644
> --- a/gcc/value-prof.h
> +++ b/gcc/value-prof.h
> @@ -89,11 +89,10 @@ void free_histograms (function *);
>  void stringop_block_profile (gimple *, unsigned int *, HOST_WIDE_INT *);
>  gcall *gimple_ic (gcall *, struct cgraph_node *, profile_probability);
>  bool check_ic_target (gcall *, struct cgraph_node *);
> -bool get_most_common_single_value (gimple *stmt, const char *counter_type,
> -                   histogram_value hist,
> -                   gcov_type *value, gcov_type *count,
> -                   gcov_type *all);
> -
> +bool get_nth_most_common_value (gimple *stmt, const char *counter_type,
> +                histogram_value hist, gcov_type *value,
> +                gcov_type *count, gcov_type *all,
> +                unsigned n = 0);
> 
>  /* In tree-profile.c.  */
>  extern void gimple_init_gcov_profiler (void);


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]