Before this patch, jit.dg/test-benchmark.c at optlevel 0, 100 iterations
showed "rest of compilation" as taking a significant chunk of time:
rest of compilation : 1.53 (64%) usr 0.06 (32%) sys 1.60 (27%) wall
1388 kB ( 6%) ggc
This turned out to be almost entirely spent inside initialize_rtl; adding a
TV_INITIALIZE_RTL for this indicates this cost clearly:
initialize rtl : 1.46 (61%) usr 0.01 ( 5%) sys 1.42 (24%) wall
1156 kB ( 5%) ggc
rest of compilation : 0.05 ( 2%) usr 0.00 ( 0%) sys 0.00 ( 0%) wall
232 kB ( 1%) ggc
I'm experimenting with a followup patch to amortize this away by making RTL
initialization a one-time deal, but in the meantime, this patch adds
accounting for it.
OK for trunk? (assuming bootstrap and testing)
gcc/ChangeLog:
* timevar.def (TV_INITIALIZE_RTL): New.
* toplev.c (initialize_rtl): Use an auto_timevar to account this
function's time to TV_INITIALIZE_RTL.
---
gcc/timevar.def | 1 +
gcc/toplev.c | 2 ++
2 files changed, 3 insertions(+)
diff --git a/gcc/timevar.def b/gcc/timevar.def
index 711bbed..cf8f37d 100644
--- a/gcc/timevar.def
+++ b/gcc/timevar.def
@@ -268,6 +268,7 @@ DEFTIMEVAR (TV_PLUGIN_RUN , "plugin execution")
DEFTIMEVAR (TV_GIMPLE_SLSR , "straight-line strength reduction")
DEFTIMEVAR (TV_VTABLE_VERIFICATION , "vtable verification")
DEFTIMEVAR (TV_TREE_UBSAN , "tree ubsan")
+DEFTIMEVAR (TV_INITIALIZE_RTL , "initialize rtl")
/* Everything else in rest_of_compilation not included above. */
DEFTIMEVAR (TV_EARLY_LOCAL , "early local passes")
diff --git a/gcc/toplev.c b/gcc/toplev.c
index 3c1ba38..54eefcd 100644
--- a/gcc/toplev.c
+++ b/gcc/toplev.c
@@ -1818,6 +1818,8 @@ static int rtl_initialized;
void
initialize_rtl (void)
{
+ auto_timevar tv (TV_INITIALIZE_RTL);
+
/* Initialization done just once per compilation, but delayed
till code generation. */
if (!rtl_initialized)
--
1.8.5.3
