| Message ID | 20151211085231.GC5527@kam.mff.cuni.cz |
|---|---|
| State | New |
| Headers | show |
On Fri, 11 Dec 2015, Jan Hubicka wrote: > Hi, > this patch further reduce memory use and time of WPA stage, especially without -g > phase opt and generate : 75.66 (39%) usr 1.78 (14%) sys 77.44 (37%) wall 855644 kB (21%) ggc > phase stream in : 34.62 (18%) usr 1.95 (16%) sys 36.57 (18%) wall 3245604 kB (79%) ggc > phase stream out : 81.89 (42%) usr 8.49 (69%) sys 90.37 (44%) wall 50 kB ( 0%) ggc > ipa dead code removal : 4.33 ( 2%) usr 0.06 ( 0%) sys 4.24 ( 2%) wall 0 kB ( 0%) ggc > ipa virtual call target : 25.15 (13%) usr 0.14 ( 1%) sys 25.42 (12%) wall 0 kB ( 0%) ggc > ipa cp : 3.92 ( 2%) usr 0.21 ( 2%) sys 4.18 ( 2%) wall 340698 kB ( 8%) ggc > ipa inlining heuristics : 24.12 (12%) usr 0.38 ( 3%) sys 24.37 (12%) wall 500427 kB (12%) ggc > lto stream inflate : 7.07 ( 4%) usr 0.38 ( 3%) sys 7.33 ( 4%) wall 0 kB ( 0%) ggc > ipa lto gimple in : 1.95 ( 1%) usr 0.61 ( 5%) sys 2.42 ( 1%) wall 324875 kB ( 8%) ggc > ipa lto gimple out : 9.16 ( 5%) usr 1.64 (13%) sys 10.49 ( 5%) wall 50 kB ( 0%) ggc > ipa lto decl in : 21.25 (11%) usr 1.01 ( 8%) sys 22.37 (11%) wall 2348869 kB (57%) ggc > ipa lto decl out : 67.33 (34%) usr 1.66 (13%) sys 68.96 (33%) wall 0 kB ( 0%) ggc > ipa lto constructors out: 1.39 ( 1%) usr 0.38 ( 3%) sys 2.18 ( 1%) wall 0 kB ( 0%) ggc > ipa lto decl merge : 2.12 ( 2%) usr 0.00 ( 0%) sys 2.12 ( 2%) wall 13737 kB ( 0%) ggc > ipa reference : 2.14 ( 2%) usr 0.00 ( 0%) sys 2.13 ( 2%) wall 0 kB ( 0%) ggc > ipa pure const : 2.29 ( 2%) usr 0.01 ( 0%) sys 2.35 ( 2%) wall 0 kB ( 0%) ggc > ipa icf : 9.02 ( 7%) usr 0.18 ( 2%) sys 9.72 ( 7%) wall 19203 kB ( 0%) ggc > TOTAL : 195.27 12.37 207.64 4103297 kB > > to: > > phase setup : 0.00 ( 0%) usr 0.00 ( 0%) sys 0.01 ( 0%) wall 1996 kB ( 0%) ggc > phase opt and generate : 77.17 (53%) usr 1.69 ( 9%) sys 79.45 (48%) wall 856874 kB (26%) ggc > phase stream in : 25.92 (18%) usr 1.75 (10%) sys 27.66 (17%) wall 2418654 kB (74%) ggc > phase stream out : 39.90 (27%) usr 14.74 (81%) sys 54.82 (33%) wall 50 kB ( 0%) ggc > phase finalize : 2.52 ( 2%) usr 0.11 ( 1%) sys 2.63 ( 2%) wall 0 kB ( 0%) ggc > garbage collection : 4.56 ( 3%) usr 0.01 ( 0%) sys 4.56 ( 3%) wall 0 kB ( 0%) ggc > ipa dead code removal : 4.32 ( 3%) usr 0.03 ( 0%) sys 4.59 ( 3%) wall 2 kB ( 0%) ggc > ipa virtual call target : 23.19 (16%) usr 0.18 ( 1%) sys 23.31 (14%) wall 0 kB ( 0%) ggc > ipa cp : 4.06 ( 3%) usr 0.18 ( 1%) sys 4.10 ( 2%) wall 339974 kB (10%) ggc > ipa inlining heuristics : 25.05 (17%) usr 0.32 ( 2%) sys 25.86 (16%) wall 500986 kB (15%) ggc > lto stream inflate : 5.50 ( 4%) usr 0.42 ( 2%) sys 5.73 ( 3%) wall 0 kB ( 0%) ggc > ipa lto gimple in : 1.97 ( 1%) usr 0.51 ( 3%) sys 2.70 ( 2%) wall 324937 kB (10%) ggc > ipa lto gimple out : 9.00 ( 6%) usr 1.59 ( 9%) sys 10.22 ( 6%) wall 50 kB ( 0%) ggc > ipa lto decl in : 14.29 (10%) usr 0.73 ( 4%) sys 15.18 ( 9%) wall 1522854 kB (46%) ggc > ipa lto decl out : 25.35 (17%) usr 0.59 ( 3%) sys 25.91 (16%) wall 0 kB ( 0%) ggc > ipa lto constructors out: 1.48 ( 1%) usr 0.51 ( 3%) sys 2.38 ( 1%) wall 0 kB ( 0%) ggc > ipa lto cgraph I/O : 0.74 ( 1%) usr 0.22 ( 1%) sys 0.97 ( 1%) wall 408576 kB (12%) ggc > ipa lto decl merge : 1.94 ( 1%) usr 0.00 ( 0%) sys 1.95 ( 1%) wall 13556 kB ( 0%) ggc > whopr wpa I/O : 2.95 ( 2%) usr 12.03 (66%) sys 15.17 ( 9%) wall 0 kB ( 0%) ggc > whopr partitioning : 3.99 ( 3%) usr 0.03 ( 0%) sys 4.01 ( 2%) wall 13619 kB ( 0%) ggc > ipa reference : 2.45 ( 2%) usr 0.01 ( 0%) sys 2.46 ( 1%) wall 0 kB ( 0%) ggc > ipa pure const : 2.30 ( 2%) usr 0.03 ( 0%) sys 2.33 ( 1%) wall 0 kB ( 0%) ggc > ipa icf : 8.30 ( 6%) usr 0.26 ( 1%) sys 8.37 ( 5%) wall 19276 kB ( 1%) ggc > TOTAL : 145.51 18.29 164.57 3277576 kB > > With debug output the numbers are not that impressive, but sitll about 17% down from decl in. > It also leads to about 63% code size reduction for global decl streams. > > I built WPA with -flto-partition=max and looked into one of partitions that seemed most absurd. > We used about 180k type delcs to produce about 700 lines of assembler that mostly contained > a calls to various methods. THe thing is that each method borught in a lot of declarations > so I looked into why and noticed that TYPE_FIELDS contains TYPE_DECLS that are mostly ignored > by the back-end expect for dwaf2out and dwarf2out actually ignores good portion of them, too. > > I thus made a predicate to tell waht decls are going to be useful for dwarf2out and removed > rest in free_lang_data. Clearly with early debug, we will be able to remove them all. > > Honza > > > * tree.c (free_lang_data_in_type): Skip irrelevant typedecls. > (find_decls_types_r): Likewise. > * tree.h (type_decl_relevant_for_debug_p): Declare. > * dwarf2out.c (type_decl_relevant_for_debug_p): New function. > Index: tree.c > =================================================================== > --- tree.c (revision 231546) > +++ tree.c (working copy) > @@ -5191,7 +5191,8 @@ free_lang_data_in_type (tree type) > while (member) > { > if (TREE_CODE (member) == FIELD_DECL > - || TREE_CODE (member) == TYPE_DECL) > + || (TREE_CODE (member) == TYPE_DECL > + && type_decl_relevant_for_debug_p (member))) > { > if (prev) > TREE_CHAIN (prev) = member; > @@ -5666,7 +5667,8 @@ find_decls_types_r (tree *tp, int *ws, v > while (tem) > { > if (TREE_CODE (tem) == FIELD_DECL > - || TREE_CODE (tem) == TYPE_DECL) > + || (TREE_CODE (tem) == TYPE_DECL > + && type_decl_relevant_for_debug_p (tem))) > fld_worklist_push (tem, fld); > tem = TREE_CHAIN (tem); > } > Index: tree.h > =================================================================== > --- tree.h (revision 231546) > +++ tree.h (working copy) > @@ -5417,4 +5417,6 @@ desired_pro_or_demotion_p (const_tree to > return to_type_precision <= TYPE_PRECISION (from_type); > } > > +extern bool type_decl_relevant_for_debug_p (const_tree); > + > #endif /* GCC_TREE_H */ > Index: dwarf2out.c > =================================================================== > --- dwarf2out.c (revision 231546) > +++ dwarf2out.c (working copy) > @@ -21134,6 +21134,15 @@ is_redundant_typedef (const_tree decl) > return 0; > } > > +/* Return true if DECL is going to be useful for debug output. */ > +bool > +type_decl_relevant_for_debug_p (const_tree decl) > +{ > + if (debug_info_level <= DINFO_LEVEL_TERSE) > + return false; We explicitely do not use debug-info-level tests in free-lang-data to allow mixing -g and -g0 objects. Are you sure doing the above doesn't mess up tree merging enough to effectively enlarge WPA memory use and the merged decl sections? [I'm quite sure firefox build system manages to mess up -g vs. -g0 in some places ;)] > + return (!DECL_IGNORED_P (decl) && !is_redundant_typedef (decl)); > +} > + The patch would be ok if you simply export is_redundant_typedef and inline the DECL_IGNORED_P check into free-lang-data. Thanks, Richard.
> > We explicitely do not use debug-info-level tests in free-lang-data > to allow mixing -g and -g0 objects. Are you sure doing the above > doesn't mess up tree merging enough to effectively enlarge WPA > memory use and the merged decl sections? > > [I'm quite sure firefox build system manages to mess up -g vs. -g0 > in some places ;)] Hmm, I will try the debug build with firefox on this. -fdump-ipa-devirt now dumps all main variants that are duplicates of one ODR type. We definitely have types with hundreds of duplicates, so there are quite common cases where tree merging does not fire. > > > + return (!DECL_IGNORED_P (decl) && !is_redundant_typedef (decl)); > > +} > > + > > The patch would be ok if you simply export is_redundant_typedef > and inline the DECL_IGNORED_P check into free-lang-data. OK, I had that originally, will return that back. is_redundant_typedef is declared inline. Putting it to tree.h drags bit too many dwarf2out internals, but I suppose it is OK to just turn it non-inline. It is a type of function where inliner should be able to decide. Honza > > Thanks, > Richard.
On Fri, 11 Dec 2015, Jan Hubicka wrote: > > > > We explicitely do not use debug-info-level tests in free-lang-data > > to allow mixing -g and -g0 objects. Are you sure doing the above > > doesn't mess up tree merging enough to effectively enlarge WPA > > memory use and the merged decl sections? > > > > [I'm quite sure firefox build system manages to mess up -g vs. -g0 > > in some places ;)] > > Hmm, I will try the debug build with firefox on this. -fdump-ipa-devirt > now dumps all main variants that are duplicates of one ODR type. > We definitely have types with hundreds of duplicates, so there are > quite common cases where tree merging does not fire. > > > > > + return (!DECL_IGNORED_P (decl) && !is_redundant_typedef (decl)); > > > +} > > > + > > > > The patch would be ok if you simply export is_redundant_typedef > > and inline the DECL_IGNORED_P check into free-lang-data. > > OK, I had that originally, will return that back. > is_redundant_typedef is declared inline. Putting it to tree.h drags > bit too many dwarf2out internals, but I suppose it is OK to just > turn it non-inline. It is a type of function where inliner should be > able to decide. Yeah. Richard. > Honza > > > > Thanks, > > Richard.
Index: tree.c =================================================================== --- tree.c (revision 231546) +++ tree.c (working copy) @@ -5191,7 +5191,8 @@ free_lang_data_in_type (tree type) while (member) { if (TREE_CODE (member) == FIELD_DECL - || TREE_CODE (member) == TYPE_DECL) + || (TREE_CODE (member) == TYPE_DECL + && type_decl_relevant_for_debug_p (member))) { if (prev) TREE_CHAIN (prev) = member; @@ -5666,7 +5667,8 @@ find_decls_types_r (tree *tp, int *ws, v while (tem) { if (TREE_CODE (tem) == FIELD_DECL - || TREE_CODE (tem) == TYPE_DECL) + || (TREE_CODE (tem) == TYPE_DECL + && type_decl_relevant_for_debug_p (tem))) fld_worklist_push (tem, fld); tem = TREE_CHAIN (tem); } Index: tree.h =================================================================== --- tree.h (revision 231546) +++ tree.h (working copy) @@ -5417,4 +5417,6 @@ desired_pro_or_demotion_p (const_tree to return to_type_precision <= TYPE_PRECISION (from_type); } +extern bool type_decl_relevant_for_debug_p (const_tree); + #endif /* GCC_TREE_H */ Index: dwarf2out.c =================================================================== --- dwarf2out.c (revision 231546) +++ dwarf2out.c (working copy) @@ -21134,6 +21134,15 @@ is_redundant_typedef (const_tree decl) return 0; } +/* Return true if DECL is going to be useful for debug output. */ +bool +type_decl_relevant_for_debug_p (const_tree decl) +{ + if (debug_info_level <= DINFO_LEVEL_TERSE) + return false; + return (!DECL_IGNORED_P (decl) && !is_redundant_typedef (decl)); +} + /* Return TRUE if TYPE is a typedef that names a type for linkage purposes. This kind of typedefs is produced by the C++ FE for constructs like: