On 03/09/16 03:12, Carsten Haitzler (The Rasterman) wrote: > On Fri, 2 Sep 2016 15:15:07 +0100 Tom Hacohen <t...@osg.samsung.com> said: > >> Good job trying to tackle it, but this is wrong. >> >> This is not thread-safe and the Eo infrastructure should be. So >> essentially, this optimisation is not allowed. You could maybe do a >> cache per-thread, raster claims that __thread is actually quite alright >> nowadays with GCC, I'm not sure, but this, and the following patch >> should be reverted. > > i looked more and more. we are both right. and both wrong. > > __thread is FREE ... in a binary. it's a tls lookup in a .so. because the > compiler cant pre-allocate a memory region for this info so it has to use the > current stack frame to jump to a tls section.
I knew I wasn't imagining! :) > >> If you have anything to say to defend this approach, please do so by >> Monday, otherwise I'll revert both patches then. > > ummm i already wrapped them with spinlocks earlier this week. thats when i > started to go "there has to be a better way" because real life testing (perf > not callgrind) made eo ptr lookup go from 2% to 5% of cpu. thus my wondering > if > making this per-thread might not help. that is a whole topic on its own > > BUT i still think its a good idea: > > tls 0.33556 400000002 = 1.546 ns / get > spin 0.89498 300000000 = 7.585 ns / lock+release > > a tls lookup is almost 5 TIMES faster than a spinlock lock + unlock. given > that > and our discussion which i'll do a writeup on that adds safety to eo and > threading (ie makes it impossible to use an eo obj outside its owning thread > unless its explicitly sent/shared there) is probably totally worth it as we > can > remove basically all the locks really used outside of app init (_eoid_lock, an > eo_isa lock, klass->iterators.trash_lock). but this does bring design > considerations into play and we need to have a talk about this. i'll draft up > some mail describing this all later. > >> -- >> Tom. >> >> On 26/08/16 20:14, Cedric BAIL wrote: >>> cedric pushed a commit to branch master. >>> >>> http://git.enlightenment.org/core/efl.git/commit/?id=93a706a947cd2d6ef80f8704f4e23737fea1258f >>> >>> commit 93a706a947cd2d6ef80f8704f4e23737fea1258f >>> Author: Cedric BAIL <ced...@osg.samsung.com> >>> Date: Fri Aug 26 12:03:08 2016 -0700 >>> >>> eo: general speedup of all Eo related operation. >>> >>> This change rely on the fact that we do fetch the same >>> object id over and over again. _efl_object_call_resolve got >>> 15% faster, efl_data_scope_get 20%. >>> --- >>> src/lib/eo/eo_base_class.c | 3 +++ >>> src/lib/eo/eo_ptr_indirection.x | 20 +++++++++++++++++++- >>> 2 files changed, 22 insertions(+), 1 deletion(-) >>> >>> diff --git a/src/lib/eo/eo_base_class.c b/src/lib/eo/eo_base_class.c >>> index 732d063..35c7a9c 100644 >>> --- a/src/lib/eo/eo_base_class.c >>> +++ b/src/lib/eo/eo_base_class.c >>> @@ -11,6 +11,9 @@ >>> >>> static int event_freeze_count = 0; >>> >>> +_Eo_Object *cached_object = NULL; >>> +Eo_Id cached_id = 0; >>> + >>> typedef struct _Eo_Callback_Description Eo_Callback_Description; >>> >>> typedef struct >>> diff --git a/src/lib/eo/eo_ptr_indirection.x >>> b/src/lib/eo/eo_ptr_indirection.x index d89b1ec..cae59a4 100644 >>> --- a/src/lib/eo/eo_ptr_indirection.x >>> +++ b/src/lib/eo/eo_ptr_indirection.x >>> @@ -269,6 +269,9 @@ extern Generation_Counter _eo_generation_counter; >>> /* Macro used for readability */ >>> #define TABLE_FROM_IDS _eo_ids_tables[mid_table_id][table_id] >>> >>> +extern _Eo_Object *cached_object; >>> +extern Eo_Id cached_id; >>> + >>> static inline _Eo_Object * >>> _eo_obj_pointer_get(const Eo_Id obj_id) >>> { >>> @@ -289,6 +292,10 @@ _eo_obj_pointer_get(const Eo_Id obj_id) >>> DBG("obj_id is not a valid object id."); >>> return NULL; >>> } >>> + else if (obj_id == cached_id) >>> + { >>> + return cached_object; >>> + } >>> >>> EO_DECOMPOSE_ID(obj_id, mid_table_id, table_id, entry_id, generation); >>> >>> @@ -297,7 +304,13 @@ _eo_obj_pointer_get(const Eo_Id obj_id) >>> { >>> entry = &(TABLE_FROM_IDS->entries[entry_id]); >>> if (entry && entry->active && (entry->generation == generation)) >>> - return entry->ptr; >>> + { >>> + // Cache the result of that lookup >>> + cached_object = entry->ptr; >>> + cached_id = obj_id; >>> + >>> + return entry->ptr; >>> + } >>> } >>> >>> ERR("obj_id %p is not pointing to a valid object. Maybe it has already >>> been freed.", @@ -477,6 +490,11 @@ _eo_id_release(const Eo_Id obj_id) >>> if (_current_table == table) >>> _current_table = NULL; >>> } >>> + >>> + // In case an object is destroyed, wipe out the cache >>> + cached_id = 0; >>> + cached_object = NULL; >>> + >>> return; >>> } >>> } >>> >> >> >> ------------------------------------------------------------------------------ >> _______________________________________________ >> enlightenment-devel mailing list >> enlightenment-devel@lists.sourceforge.net >> https://lists.sourceforge.net/lists/listinfo/enlightenment-devel >> > > ------------------------------------------------------------------------------ _______________________________________________ enlightenment-devel mailing list enlightenment-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/enlightenment-devel
