[intercu] Add a cache
Daniel Jacobowitz
drow@false.org
Sun Feb 22 21:06:00 GMT 2004
Cache the compilation units loaded in the processing of the last 20
psymtabs. This actually has no substantial performance impact on my
testcases; I am going to have to make larger ones :) I think it will help
us scale to even larger binaries at some point, and in any case I'll need
the same caching information for full symbols, I expect.
So, in the news department. All my changes have no perceptable effect on
startup time using Monotone as a testcase; if I build it with
-feliminate-dwarf2-dups its .debug_info goes down by about 40% but we seem
to spend so little of our time dealing with the things that are in its
headers that not much changes. Note that the binary itself becomes huge.
This is an unfortunate side effect of the elimination of
-feliminate-dwarf2-dups, which adds a ton of symbols to the symbol table; I
need to think about that...
On the other hand, I also built GDB with -feliminate-dwarf2-dups. The total
size of .debug_info is cut in 4, and startup time is cut in 3! This is an
altogether happy result, and I may be able to squeeze it a little further.
As I said, though, I will definitely need to invest in some larger
testcases; total startup time for GDB went from 1.4 seconds to 0.4 seconds
and we're closing in fast on the resolution of my measurements.
Committed to the intercu branch.
--
Daniel Jacobowitz
MontaVista Software Debian GNU/Linux Developer
2004-02-22 Daniel Jacobowitz <drow@mvista.com>
* dwarf2read.c (MAX_CACHE_AGE): Define.
(struct dwarf2_cu): Add last_used counter.
(free_comp_units_worker): New function, based on
clear_per_cu_pointer. Support aging.
(clear_per_cu_pointer): Use it.
(free_cached_comp_units, free_one_cached_comp_unit): New functions.
(dwarf2_build_psymtabs_hard): Keep read_in_chain across iterations.
Free all cached comp units after building psymtabs. Free any cached
copy of a comp unit before building a psymtab for it.
(find_partial_die): Clear the last_used counter.
Index: dwarf2read.c
===================================================================
RCS file: /cvs/src/src/gdb/dwarf2read.c,v
retrieving revision 1.135.2.19
diff -u -p -r1.135.2.19 dwarf2read.c
--- dwarf2read.c 22 Feb 2004 19:26:43 -0000 1.135.2.19
+++ dwarf2read.c 22 Feb 2004 20:55:45 -0000
@@ -52,6 +52,10 @@
#include "gdb_assert.h"
#include <sys/types.h>
+/* Loaded secondary compilation units are kept in memory until they have not
+ been referenced for the processing of this many compilation units. */
+#define MAX_CACHE_AGE 20
+
#ifndef DWARF2_REG_TO_REGNUM
#define DWARF2_REG_TO_REGNUM(REG) (REG)
#endif
@@ -274,6 +278,9 @@ struct dwarf2_cu
/* Backchain to our per_cu entry if the tree has been built. */
struct dwarf2_per_cu_data *per_cu;
+
+ /* How many compilation units ago was this CU last referenced? */
+ int last_used;
};
static const struct objfile_data *dwarf2_cu_tree;
@@ -970,7 +977,11 @@ dwarf2_symbol_mark_computed (struct attr
static struct dwarf2_per_cu_data *dwarf2_find_containing_comp_unit
(unsigned long offset, struct dwarf2_cu *cu);
-static void clear_per_cu_pointer (void *data);
+static void clear_per_cu_pointer (void *);
+
+static void free_cached_comp_units (void *);
+
+static void free_one_cached_comp_unit (struct dwarf2_cu *, struct dwarf2_cu *);
/* Allocation function for the libiberty splay tree which uses an obstack. */
static void *
@@ -1270,6 +1281,7 @@ dwarf2_build_psymtabs_hard (struct objfi
struct cleanup *back_to;
CORE_ADDR lowpc, highpc, baseaddr;
splay_tree cu_tree = NULL;
+ struct dwarf2_cu cu;
info_ptr = dwarf_info_buffer;
@@ -1304,6 +1316,13 @@ dwarf2_build_psymtabs_hard (struct objfi
obstack_init (&dwarf2_tmp_obstack);
back_to = make_cleanup (dwarf2_free_tmp_obstack, NULL);
+ /* read_in_chain, unlike every other field in the dwarf2_cu object,
+ is live across the loop. The function clear_per_cu_pointer will
+ free any allocated compilation units that have not been used in
+ several psymtabs. Others will remain on the list. */
+ cu.read_in_chain = NULL;
+ make_cleanup (free_cached_comp_units, &cu);
+
/* Since the objects we're extracting from dwarf_info_buffer vary in
length, only the individual functions to extract them (like
read_comp_unit_head and load_partial_die) can really know whether
@@ -1320,7 +1339,6 @@ dwarf2_build_psymtabs_hard (struct objfi
while (info_ptr < dwarf_info_buffer + dwarf_info_size)
{
struct cleanup *back_to_inner;
- struct dwarf2_cu cu;
struct abbrev_info *abbrev;
unsigned int bytes_read;
struct dwarf2_per_cu_data *this_cu;
@@ -1390,6 +1408,14 @@ dwarf2_build_psymtabs_hard (struct objfi
node = splay_tree_lookup (cu_tree, cu.header.offset);
gdb_assert (node != NULL);
per_cu = (struct dwarf2_per_cu_data *) node->value;
+
+ /* If we were already read in, free ourselves to read in again.
+ Yes, this is pointless duplication. Fixing this will provide
+ a nice speed boost but require a lot of editing in this
+ function. */
+ if (per_cu->cu != NULL)
+ free_one_cached_comp_unit (&cu, per_cu->cu);
+
cu.per_cu = per_cu;
/* Note that this is a pointer to our stack frame. It will
@@ -1401,7 +1427,6 @@ dwarf2_build_psymtabs_hard (struct objfi
cu.per_cu = NULL;
make_cleanup (clear_per_cu_pointer, &cu);
- cu.read_in_chain = NULL;
/* Check if comp unit has_children.
If so, read the rest of the partial symbols from this comp unit.
@@ -5064,6 +5089,7 @@ find_partial_die (unsigned long offset,
cu->read_in_chain = per_cu;
}
+ per_cu->cu->last_used = 0;
*target_cu = per_cu->cu;
return find_partial_die_in_comp_unit (offset, per_cu->cu);
}
@@ -8765,31 +8791,66 @@ dwarf2_find_containing_comp_unit (unsign
}
static void
-clear_per_cu_pointer (void *data)
+free_comp_units_worker (struct dwarf2_cu *cu, int aging,
+ struct dwarf2_cu *target_cu)
{
- struct dwarf2_per_cu_data *this_cu, *per_cu;
- struct dwarf2_cu *cu = data;
+ struct dwarf2_per_cu_data *this_cu, *per_cu, **last_chain;
this_cu = cu->per_cu;
- if (this_cu == NULL)
+ if (this_cu == NULL && target_cu == NULL)
return;
- per_cu = this_cu->cu->read_in_chain;
+ per_cu = cu->read_in_chain;
+ last_chain = &cu->read_in_chain;
while (per_cu != NULL)
{
struct dwarf2_per_cu_data *next_cu;
- obstack_free (&per_cu->cu->partial_die_obstack, NULL);
-
next_cu = per_cu->cu->read_in_chain;
- xfree (per_cu->cu);
- per_cu->cu = NULL;
+
+ if ((aging && per_cu->cu->last_used > MAX_CACHE_AGE)
+ || (target_cu && per_cu->cu == target_cu)
+ || (!aging && target_cu == NULL))
+ {
+ obstack_free (&per_cu->cu->partial_die_obstack, NULL);
+ xfree (per_cu->cu);
+ per_cu->cu = NULL;
+ *last_chain = next_cu;
+ }
+ else
+ {
+ per_cu->cu->last_used++;
+ last_chain = &per_cu->cu->read_in_chain;
+ }
+
per_cu = next_cu;
}
/* This compilation unit is on the stack in dwarf2_build_psymtabs_hard,
- so we should not xfree it. */
- this_cu->cu = NULL;
+ so we should not xfree it. Just unlink it. */
+ if (target_cu == NULL)
+ {
+ cu->per_cu = NULL;
+ this_cu->cu = NULL;
+ }
+}
+
+static void
+clear_per_cu_pointer (void *data)
+{
+ free_comp_units_worker (data, 1, NULL);
+}
+
+static void
+free_cached_comp_units (void *data)
+{
+ free_comp_units_worker (data, 0, NULL);
+}
+
+static void
+free_one_cached_comp_unit (struct dwarf2_cu *cu, struct dwarf2_cu *target_cu)
+{
+ free_comp_units_worker (cu, 0, target_cu);
}
void _initialize_dwarf2_read (void);
More information about the Gdb-patches
mailing list