[binutils-gdb] [gdb/symtab] Use task size in parallel_for_each in dwarf2_build_psymtabs_hard

[Tom de Vries <vries@sourceware.org>]

https://sourceware.org/git/gitweb.cgi?p=binutils-gdb.git;h=b069b588cfe10e6bf20ed723cf796686ba4fc0dc

commit b069b588cfe10e6bf20ed723cf796686ba4fc0dc
Author: Tom de Vries <tdevries@suse.de>
Date:   Fri Aug 5 16:12:56 2022 +0200

    [gdb/symtab] Use task size in parallel_for_each in dwarf2_build_psymtabs_hard
    
    In dwarf2_build_psymtabs_hard, we use a parallel_for_each to distribute CUs
    over threads.
    
    Ensuring a fair distribution over the worker threads and main thread in terms
    of number of CUs might not be the most efficient way, given that CUs can vary
    in size.
    
    Fix this by using per_cu->get_length () as the task size.
    
    I've used this experiment to verify the performance impact:
    ...
    $ for n in $(seq 1 10); do \
        time gdb -q -batch ~/firefox/libxul.so-93.0-1.1.x86_64.debug \
        2>&1 \
        | grep "real:"; \
      done
    ...
    and without the patch got:
    ...
    real: 4.71
    real: 4.88
    real: 4.29
    real: 4.30
    real: 4.65
    real: 4.27
    real: 4.27
    real: 4.27
    real: 4.75
    real: 4.41
    ...
    and with the patch:
    ...
    real: 3.68
    real: 3.81
    real: 3.80
    real: 3.68
    real: 3.75
    real: 3.69
    real: 3.69
    real: 3.74
    real: 3.67
    real: 3.74
    ...
    so that seems a reasonable improvement.
    
    With parallel_for_each_debug set to true, we get some more detail about
    the difference in behaviour.  Without the patch we have:
    ...
    Parallel for: n_elements: 2818
    Parallel for: minimum elements per thread: 1
    Parallel for: elts_per_thread: 704
    Parallel for: elements on worker thread 0       : 705
    Parallel for: elements on worker thread 1       : 705
    Parallel for: elements on worker thread 2       : 704
    Parallel for: elements on worker thread 3       : 0
    Parallel for: elements on main thread           : 704
    ...
    and with the patch:
    ...
    Parallel for: n_elements: 2818
    Parallel for: total_size: 1483674865
    Parallel for: size_per_thread: 370918716
    Parallel for: elements on worker thread 0       : 752   (size: 371811790)
    Parallel for: elements on worker thread 1       : 360   (size: 371509370)
    Parallel for: elements on worker thread 2       : 1130  (size: 372681710)
    Parallel for: elements on worker thread 3       : 0     (size: 0)
    Parallel for: elements on main thread           : 576   (size: 367671995)
    ...
    
    Tested on x86_64-linux.

Diff:
---
 gdb/dwarf2/read.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/gdb/dwarf2/read.c b/gdb/dwarf2/read.c
index f03151983dc..348fbe12da1 100644
--- a/gdb/dwarf2/read.c
+++ b/gdb/dwarf2/read.c
@@ -7085,6 +7085,13 @@ dwarf2_build_psymtabs_hard (dwarf2_per_objfile *per_objfile)
 
     using iter_type = decltype (per_bfd->all_comp_units.begin ());
 
+    auto task_size_ = [] (iter_type iter)
+      {
+	dwarf2_per_cu_data *per_cu = iter->get ();
+	return (size_t)per_cu->length ();
+      };
+    auto task_size = gdb::make_function_view (task_size_);
+
     /* Each thread returns a pair holding a cooked index, and a vector
        of errors that should be printed.  The latter is done because
        GDB's I/O system is not thread-safe.  run_on_main_thread could be
@@ -7113,7 +7120,7 @@ dwarf2_build_psymtabs_hard (dwarf2_per_objfile *per_objfile)
 	      }
 	  }
 	return result_type (thread_storage.release (), std::move (errors));
-      });
+      }, task_size);
 
     /* Only show a given exception a single time.  */
     std::unordered_set<gdb_exception> seen_exceptions;