[gcc(refs/users/kubaneko/heads/histogram)] some refactoring and added lto-streaming for histograms

[gcc(refs/users/kubaneko/heads/histogram)] some refactoring and added lto-streaming for histograms

[Ondrej Kubanek]

https://gcc.gnu.org/g:e755feaf21fe3b58357ac90440e67733a032db47

commit e755feaf21fe3b58357ac90440e67733a032db47
Author: kubaneko <kubanek0ondrej@gmail.com>
Date:   Mon Mar 13 21:31:03 2023 +0000

    some refactoring and added lto-streaming for histograms

Diff:
---
 gcc/cfgloop.h           | 73 -------------------------------------------------
 gcc/lto-streamer-in.cc  | 11 ++++++++
 gcc/lto-streamer-out.cc |  7 +++++
 gcc/tree-vect-loop.cc   |  9 ++----
 4 files changed, 20 insertions(+), 80 deletions(-)

diff --git a/gcc/cfgloop.h b/gcc/cfgloop.h
index 48badc22648..6a7ef44b8e1 100644
--- a/gcc/cfgloop.h
+++ b/gcc/cfgloop.h
@@ -99,79 +99,6 @@ struct GTY(()) histogram_counters{
     vec<gcov_type, va_heap, vl_embed> * GTY((skip)) hist;
 };
 
-// void histogram_counters_minus_upper_bound (histogram_counters* hist_c, gcov_type_unsigned difference){
-//     if (hist_c==NULL || difference==0)
-//         return;
-//     auto hist=*(hist_c->hist);
-//     unsigned int lin_size=param_profile_histogram_size_lin;
-//     unsigned int lin_pow2=floor_log2(lin_size-1);
-//     unsigned int tot_size=param_profile_histogram_size;
-//     gcov_type_unsigned  log_diff=floor_log2(difference);
-//     unsigned int i=1;
-//     for(; i<lin_size-1 &&  i<tot_size-1; i++){
-//         if (i<=difference){
-//             hist[0]+=hist[i];
-//         } else {
-//             hist[i-difference]+=hist[i];
-//         }
-//         hist[i]=0;
-//     }
-//     // we try to restore some of the linear histogram
-//     // for the last linear counter containing also lesser values then nearest pow2
-//     // assumption is uniform
-//     gcov_type_unsigned upper_bound_lin=(1<<(ceil_log2(lin_size)-1));
-//     if (lin_size>2 && difference<upper_bound_lin && i<tot_size-1){
-//         gcov_type_unsigned lin_size_val=hist[lin_size-1]/((upper_bound_lin-1)-(lin_size-2));
-//         for (int j=int(i)-int(difference);j<gcov_type(upper_bound_lin)-gcov_type(difference)+1 && j<int(lin_size)-1;i++) {
-//             if (j<0) {
-//                 hist[0]+=lin_size_val;
-//             } else {
-//                 hist[j]+=lin_size_val;
-//             }
-//             hist[lin_size-1]-=lin_size_val;
-//         }
-//         i++;
-//     }
-//     // Index where difference would be placed
-//     unsigned diff_index=log_diff+(lin_size-lin_pow2)-1;
-//     for (; i<diff_index && difference<lin_size && i<tot_size-1;i++){
-//         hist[0]+=hist[i];
-//         hist[i]=0;
-//     }
-//     for (; i<tot_size-1 && (1<<(i-diff_index+2))<10000;i++){
-//         // We assume uniform distribution
-//         // tot_size -1 should not be touched We know little about it
-//         gcov_type_unsigned share=hist[i]/(1<<(i-diff_index+2));
-//         hist[i-1]+=share;
-//         hist[i]-=share;
-//     }
-// }
-// 
-// void histogram_counters_div_upper_bound (histogram_counters* hist_c, unsigned int divisor){
-//     if (hist_c==NULL || divisor<2)
-//         return;
-//     auto hist=*(hist_c->hist);
-//     unsigned int lin_size=param_profile_histogram_size_lin;
-//     unsigned int tot_size=param_profile_histogram_size;
-//     gcov_type_unsigned  log_div=floor_log2(divisor);
-//     unsigned int i=1;
-//     for(; i<lin_size-1 && i<tot_size-1; i++){
-//         hist[i/divisor]+=hist[i];
-//         hist[i]=0;
-//     }
-//     
-//     for (;i<tot_size-1;i++){
-//         gcov_type_unsigned upper_diff=((1<<(ceil_log2(lin_size)+i-lin_size))-1)/divisor;
-//         if (upper_diff<lin_size-1 && lin_size>1){
-//             hist[upper_diff==0 ? 1 : upper_diff]+=hist[i];
-//             hist[i]=0;
-//         } else {
-//             hist[i-log_div>0?i-log_div : 1]+=hist[i];
-//             hist[i]=0;
-//         }
-//     }
-// }
-
 typedef class loop *loop_p;
 
 /* An integer estimation of the number of iterations.  Estimate_state
diff --git a/gcc/lto-streamer-in.cc b/gcc/lto-streamer-in.cc
index 03cb41cfa16..2cf59858a54 100644
--- a/gcc/lto-streamer-in.cc
+++ b/gcc/lto-streamer-in.cc
@@ -1126,6 +1126,17 @@ input_cfg (class lto_input_block *ib, class data_in *data_in,
       loop->any_estimate = streamer_read_hwi (ib);
       if (loop->any_estimate)
 	loop->nb_iterations_estimate = streamer_read_widest_int (ib);
+    // whether the loop has histogram counter
+    if (streamer_read_hwi (ib)){
+       loop->counters=ggc_alloc<histogram_counters>();
+       loop->counters->sum=streamer_read_gcov_count (ib);
+       loop->counters->hist=NULL;
+       vec_safe_grow_cleared(loop->counters->hist,param_profile_histogram_size);
+       for (int i=0;i<param_profile_histogram_size;++i){
+           (*loop->counters->hist)[i]=streamer_read_gcov_count (ib);
+       }
+    }
+
 
       /* Read OMP SIMD related info.  */
       loop->safelen = streamer_read_hwi (ib);
diff --git a/gcc/lto-streamer-out.cc b/gcc/lto-streamer-out.cc
index 0bca530313c..8901a82dd67 100644
--- a/gcc/lto-streamer-out.cc
+++ b/gcc/lto-streamer-out.cc
@@ -2181,6 +2181,13 @@ output_cfg (struct output_block *ob, struct function *fn)
       streamer_write_hwi (ob, loop->any_estimate);
       if (loop->any_estimate)
 	streamer_write_widest_int (ob, loop->nb_iterations_estimate);
+      streamer_write_hwi (ob, loop->counters!=NULL);
+      if (loop->counters) {
+          streamer_write_gcov_count (ob, loop->counters->sum);
+          for (unsigned int i=0; i<param_profile_histogram_size; ++i) {
+            streamer_write_gcov_count (ob, (*loop->counters->hist)[i]);
+          }
+      }
 
       /* Write OMP SIMD related info.  */
       streamer_write_hwi (ob, loop->safelen);
diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 74622aef0f9..cc2ed78b52e 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -11067,13 +11067,8 @@ vect_transform_loop (loop_vec_info loop_vinfo, gimple *loop_vectorized_call)
 			    lowest_vf) - 1
 	   : wi::udiv_floor (loop->nb_iterations_upper_bound + bias_for_lowest,
 			     lowest_vf) - 1);
-    if (final_iter_may_be_partial){
-        histogram_counters_div_upper_bound(loop->counters,lowest_vf);
-        histogram_counters_minus_upper_bound(loop->counters,1);
-    } else {
-        histogram_counters_div_upper_bound(loop->counters,lowest_vf);
-        histogram_counters_minus_upper_bound(loop->counters,1);
-    }
+    histogram_counters_div_upper_bound(loop->counters,lowest_vf);
+    histogram_counters_minus_upper_bound(loop->counters,1);
       if (main_vinfo
 	  /* Both peeling for alignment and peeling for gaps can end up
 	     with the scalar epilogue running for more than VF-1 iterations.  */