Re: [PATCH] Add option for dumping to stderr (issue6190057)
Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. Thanks, Sharad On Thu, Sep 27, 2012 at 9:10 AM, Xinliang David Li davi...@google.com wrote: On Thu, Sep 27, 2012 at 4:35 AM, Sharad Singhai sing...@google.com wrote: Thanks for the review. A couple of comments inline: Some minor issues: * c/c-decl.c (c_write_global_declarations): Use different method to determine if the dump has ben initialized. * cp/decl2.c (cp_write_global_declarations): Ditto. * testsuite/gcc.target/i386/vect-double-1.c: Fix test. these subdirs all have their separate ChangeLog entry from where the directory name is omitted. Index: tree-dump.c === --- tree-dump.c (revision 191490) +++ tree-dump.c (working copy) @@ -24,9 +24,11 @@ along with GCC; see the file COPYING3. If not see #include coretypes.h #include tm.h #include tree.h +#include gimple-pretty-print.h #include splay-tree.h #include filenames.h #include diagnostic-core.h +#include rtl.h what do you need gimple-pretty-print.h and rtl.h for? + +extern void dump_bb (FILE *, basic_block, int, int); + that should be declared in some header +/* Dump gimple statement GS with SPC indentation spaces and + EXTRA_DUMP_FLAGS on the dump streams if DUMP_KIND is enabled. */ + +void +dump_gimple_stmt (int dump_kind, int extra_dump_flags, gimple gs, int spc) +{ the gimple stuff really belongs in to gimple-pretty-print.c This dump_gimple_stmt () is just a dispatcher, which uses internal data structure such as dump streams/flags. If I move it into gimple-pretty-print.c, then I would have to export those streams/flags. I was hoping to avoid it by keeping all dump_* () methods together in dumpfile.c (earlier in tree-dump.c). Thus, later one could just make dump_file/dump_flags static when all the passes have converted to this scheme. You can make the flags/streams global but only expose them via inline accessors in the header file. David (parts of tree-dump.c should be moved to a new file dumpfile.c) +/* Dump tree T using EXTRA_DUMP_FLAGS on dump streams if DUMP_KIND is + enabled. */ + +void +dump_generic_expr (int dump_kind, int extra_dump_flags, tree t) +{ belongs to tree-pretty-print.c (to where the routines are it calls) This is again a dispatcher for dump_generic_expr () which writes to the appropriate stream depending upon dump_kind. +int +dump_start (int phase, int *flag_ptr) +{ perfect candidate for dumpfile.c You can do this re-shuffling as followup, but please try to not include rtl.h or gimple-pretty-print.h from tree-dump.c. Thus re-shuffling required by that do now. tree-dump.c should only know about dumping 'tree'. Okay, I have moved relevant methods into dumpfile.c. Index: tree-dump.h === --- tree-dump.h (revision 191490) +++ tree-dump.h (working copy) @@ -22,6 +22,7 @@ along with GCC; see the file COPYING3. If not see #ifndef GCC_TREE_DUMP_H #define GCC_TREE_DUMP_H +#include input.h probably no longer required. Index: dumpfile.h === --- dumpfile.h (revision 191490) +++ dumpfile.h (working copy) @@ -22,6 +22,9 @@ along with GCC; see the file COPYING3. If not see #ifndef GCC_DUMPFILE_H #define GCC_DUMPFILE_H 1 +#include coretypes.h +#include input.h likewise for input.h. Index: testsuite/gcc.target/i386/vect-double-1.c === --- testsuite/gcc.target/i386/vect-double-1.c (revision 191490) +++ testsuite/gcc.target/i386/vect-double-1.c (working copy) @@ -32,5 +32,5 @@ sse2_test (void) } } -/* { dg-final { scan-tree-dump-times vectorized 1 loops 1 vect } } */ +/* { dg-final { scan-tree-dump-times Vectorized loops: 1 1 vect } } */ /* { dg-final { cleanup-tree-dump vect } } */ I am sure you need a gazillion more testsuite adjustments? Thus, did you really test the patch by a bootstrap and a
Re: [Patch contrib] check_GNU_style: remove tmp file
Ping? May I commit this small patch? http://gcc.gnu.org/ml/gcc-patches/2012-09/msg00580.html Thanks, Christophe. On 10 September 2012 14:23, Christophe Lyon christophe.l...@linaro.org wrote: On 9 September 2012 12:46, Gerald Pfeifer ger...@pfeifer.com wrote: On Mon, 3 Sep 2012, Christophe Lyon wrote: check_GNU_style.sh currently leaves a temporary file in the current directory. This patch removes it upon exit. Christophe. 2012-09-03 Christophe Lyon christophe.l...@linaro.org * check_GNU_style.sh: Remove temporay file upon exit. Shouldn't this also be removed upon abort? See contrib/warn_summary, for an example, Gerald Good point. Here is a new version, catching the same signals as warn_summary. Christophe.
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 01, 2012 at 08:47:13AM +0200, Steven Bosscher wrote: The test case compiles just fine at -O2, only VRP has trouble with it. Let's try to stick with facts, not speculation. I was talking about the other PR, PR26854, which from what I remember when trying it myself and even the latest -O3 time reports from the reduced testcase show that IRA/reload aren't there very significant (for -O3 IRA takes ~ 6% and reload ~ 1%). I've put a lot of hard work into it to fix almost all scalability problems on this PR for gcc 4.8. LRA undoes all of that work. I understand it is painful for some people to hear, but I remain of opinion that LRA cannot be considered ready if it scales so much worse than everything else in the compiler. Judging the whole implementation from just these corner cases and not how it performs on other testcases (SPEC, rebuild a distro, ...) is IMHO not the right thing, if Vlad thinks the corner cases are fixable during stage3; IMHO we should allow LRA in, worst case it can be disabled by default even for i?86/x86_64. Jakub
Re: [rtl] combine a vec_concat of 2 vec_selects from the same vector
2012-09-09 Marc Glisse marc.gli...@inria.fr gcc/ * simplify-rtx.c (simplify_binary_operation_1) VEC_SELECT: Detect the identity. VEC_CONCAT: Handle VEC_SELECTs from the same vector. gcc/testsuite/ * gcc.target/i386/vect-rebuild.c: New testcase. OK if you adjust the above date and add the missing space at the end of: /* Try to merge 2 VEC_SELECTs from the same vector into a single one. */ -- Eric Botcazou
[Ada] Validity checks on subprogram parameters and results
This patch introduces two new validity checks to the GNAT compiler: 1) -gnatVl Check non-overlapping parameters When this check is enabled, each subprogram call is preceded by a sequence of checks that ensure no overlap between actual parameters. 2) -gnatVv Check proper initialization of scalars on parameters and results When this check is enabled, each IN, IN OUT and OUT formal parameter along with a possible function result is checked on entry and exit of a subprogram for properly initialized scalars. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Hristian Kirtchev kirtc...@adacore.com * checks.ads, checks.adb (Apply_Parameter_Aliasing_Checks): New routine. (Apply_Parameter_Validity_Checks): New routines. * exp_ch6.adb (Expand_Call): Add aliasing checks to detect overlapping objects. * freeze.adb: Add with and use clauses for Checks and Validsw. (Freeze_Entity): Add checks to detect proper initialization of scalars. * sem_ch4.adb: Add with and use clauses for Checks and Validsw. (Analyze_Call): Add aliasing checks to detect overlapping objects. * sem_ch13.adb: Add with and use clauses for Validsw. (Analyze_Aspect_Specifications): Add checks to detect proper initialization of scalars. * sem_prag.adb (Chain_PPC): Correct the extraction of the subprogram name. * sem_util.adb (Is_Object_Reference): Attribute 'Result now produces an object. * usage.adb (Usage): Add usage lines for validity switches 'l', 'L', 'v' and 'V'. * validsw.adb (Reset_Validity_Check_Options): Include processing for flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params. Code reformatting. (Save_Validity_Check_Options): Include processing for flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params. (Set_Validity_Check_Options): Add processing for validity switches 'a', 'l', 'L', 'n', 'v' and 'V'. Code reformatting. * validsw.ads: Add new flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params along with comments on usage. Index: usage.adb === --- usage.adb (revision 191888) +++ usage.adb (working copy) @@ -399,6 +399,8 @@ Write_Line (Fturn off checking for floating-point); Write_Line (iturn on checking for in params); Write_Line (Iturn off checking for in params); + Write_Line (lturn on checking for non-overlapping params); + Write_Line (Lturn off checking for non-overlapping params); Write_Line (mturn on checking for in out params); Write_Line (Mturn off checking for in out params); Write_Line (oturn on checking for operators/attributes); @@ -411,6 +413,8 @@ Write_Line (Sturn off checking for subscripts); Write_Line (tturn on checking for tests); Write_Line (Tturn off checking for tests); + Write_Line (vturn on checking for 'Valid_Scalars on params); + Write_Line (Vturn off checking for 'Valid_Scalars on params); Write_Line (nturn off all validity checks (including RM)); -- Lines for -gnatw switch Index: checks.adb === --- checks.adb (revision 191888) +++ checks.adb (working copy) @@ -1774,6 +1774,353 @@ (Ck_Node, Target_Typ, Source_Typ, Do_Static = False); end Apply_Length_Check; + - + -- Apply_Parameter_Aliasing_Checks -- + - + + procedure Apply_Parameter_Aliasing_Checks (Call : Node_Id) is + Loc: constant Source_Ptr := Sloc (Call); + Actual : Node_Id; + Actual_Typ : Entity_Id; + Check : Node_Id; + Cond : Node_Id := Empty; + Param : Node_Id; + Param_Typ : Entity_Id; + + begin + -- Do not generate the checks in Ada 83, 95 or 05 mode because they + -- require an Ada 2012 construct. + + if Ada_Version_Explicit Ada_2012 then + return; + end if; + + -- Inspect all pairs of parameters + + Actual := First_Actual (Call); + while Present (Actual) loop + Actual_Typ := Base_Type (Etype (Actual)); + + if Nkind (Actual) = N_Identifier + and then Is_Object_Reference (Actual) + then +Param := Next_Actual (Actual); +while Present (Param) loop + Param_Typ := Base_Type (Etype (Param)); + + if Nkind (Param) = N_Identifier + and then Is_Object_Reference (Param) + and then Actual_Typ = Param_Typ + then + --
Re: RFC: LRA for x86/x86-64 [0/9]
[ Sorry for re-send, it seems that mobile gmail sends text/html and the sourceware mailer daemon rejects that. ] On Monday, October 1, 2012, Jakub Jelinek ja...@redhat.com wrote: On Sun, Sep 30, 2012 at 06:50:50PM -0400, Vladimir Makarov wrote: I think this testcase shouldn't be a show stopper for LRA inclusion into 4.8, but something to look at for stage3. I think a lot of GCC passes have scalability issues on that testcase, that is why it must be compiled with -O1 and not higher optimization options, The test case compiles just fine at -O2, only VRP has trouble with it. Let's try to stick with facts, not speculation. And the test case is not generated, it is the Eigen template library applied to mpfr. I've put a lot of hard work into it to fix almost all scalability problems on this PR for gcc 4.8. LRA undoes all of that work. I understand it is painful for some people to hear, but I remain of opinion that LRA cannot be considered ready if it scales so much worse than everything else in the compiler. Ciao! Steven
[Ada] Detect more cases of possible infinite loops
In cases where GNAT did not detect the possibility of an infinite loop, it now issues a warning. For example, on the following code: $ gcc -c bad.adb bad.adb:9:13: warning: variable Cur is not modified in loop body bad.adb:9:13: warning: possible infinite loop 1. package body Bad is 2.procedure P (Y : Integer) is 3.begin 4. null; 5.end P; 6.procedure Q (X : Integer) is 7. Cur : Integer := X; 8.begin 9. while Cur /= 0 loop 10. P (Cur); 11. end loop; 12.end Q; 13. end Bad; Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Yannick Moy m...@adacore.com * sem_warn.adb (Check_Infinite_Loop_Warning/Test_Ref): Improve the detection of modifications to the loop variable by noting that, if the type of variable is elementary and the condition does not contain a function call, then the condition cannot be modified by side-effects from a procedure call. Index: sem_warn.adb === --- sem_warn.adb(revision 191890) +++ sem_warn.adb(working copy) @@ -472,32 +472,41 @@ return Abandon; end if; --- If we appear in the context of a procedure call, then also --- abandon, since there may be issues of non-visible side --- effects going on in the call. +-- If the condition contains a function call, we consider it may +-- be modified by side-effects from a procedure call. Otherwise, +-- we consider the condition may not be modified, although that +-- might happen if Variable is itself a by-reference parameter, +-- and the procedure called modifies the global object referred to +-- by Variable, but we actually prefer to issue a warning in this +-- odd case. Note that the case where the procedure called has +-- visibility over Variable is treated in another case below. -declare - P : Node_Id; +if Function_Call_Found then + declare + P : Node_Id; -begin - P := N; - loop - P := Parent (P); - exit when P = Loop_Statement; + begin + P := N; + loop + P := Parent (P); + exit when P = Loop_Statement; - -- Abandon if at procedure call, or something strange is - -- going on (perhaps a node with no parent that should - -- have one but does not?) As always, for a warning we - -- prefer to just abandon the warning than get into the - -- business of complaining about the tree structure here! + -- Abandon if at procedure call, or something strange is + -- going on (perhaps a node with no parent that should + -- have one but does not?) As always, for a warning we + -- prefer to just abandon the warning than get into the + -- business of complaining about the tree structure here! - if No (P) or else Nkind (P) = N_Procedure_Call_Statement then - return Abandon; - end if; - end loop; -end; + if No (P) + or else Nkind (P) = N_Procedure_Call_Statement + then +return Abandon; + end if; + end loop; + end; +end if; --- Reference to variable renaming variable in question + -- Reference to variable renaming variable in question elsif Is_Entity_Name (N) and then Present (Entity (N)) @@ -509,7 +518,7 @@ then return Abandon; --- Call to subprogram + -- Call to subprogram elsif Nkind (N) in N_Subprogram_Call then
[Ada] Ada 2012 legality checks on uses of names of protected procedures
Ada 2012 AI05-0225 clarifies that most uses of the names of protected procedures and entries require that the target object (explicit or implicit) be a variable. This applies to calls, generic actuals, and prefixes of 'Access. It applies in particular to such uses within the body a protected function. Example is ACATS Test b950001. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Ed Schonberg schonb...@adacore.com * sem_util.ads sem_util.adb (Check_Internal_Protected_Use): reject use of protected procedure or entry within the body of a protected function of the same protected type, when usage is a call, an actual in an instantiation, a or prefix of 'Access. * sem_ch8.adb (Analyze_Subprogram_Renaming): Verify that target object in renaming of protected procedure is a variable, and apply Check_Internal_Protected_Use. * sem_res.adb (Analyze_Call, Analyze_Entry_Call): apply Check_Internal_Protected_Use rather than on-line code. * sem_attr.adb (Analyze_Access_Attribute): Verify that target object in accsss to protected procedure is a variable, and apply Check_Internal_Protected_Use. Index: sem_util.adb === --- sem_util.adb(revision 191890) +++ sem_util.adb(working copy) @@ -1191,6 +1191,50 @@ end if; end Check_Implicit_Dereference; + -- + -- Check_Internal_Protected_Use -- + -- + + procedure Check_Internal_Protected_Use (N : Node_Id; Nam : Entity_Id) is + S: Entity_Id; + Prot : Entity_Id; + + begin + S := Current_Scope; + while Present (S) loop + if S = Standard_Standard then +return; + + elsif Ekind (S) = E_Function + and then Ekind (Scope (S)) = E_Protected_Type + then +Prot := Scope (S); +exit; + end if; + + S := Scope (S); + end loop; + + if Scope (Nam) = Prot and then Ekind (Nam) /= E_Function then + if Nkind (N) = N_Subprogram_Renaming_Declaration then +Error_Msg_N + (within protected function cannot use protected +procedure in renaming or as generic actual, N); + + elsif Nkind (N) = N_Attribute_Reference then +Error_Msg_N + (within protected function cannot take access of + protected procedure, N); + + else +Error_Msg_N + (within protected function, protected object is constant, N); +Error_Msg_N + (\cannot call operation that may modify it, N); + end if; + end if; + end Check_Internal_Protected_Use; + --- -- Check_Later_Vs_Basic_Declarations -- --- Index: sem_util.ads === --- sem_util.ads(revision 191888) +++ sem_util.ads(working copy) @@ -170,6 +170,12 @@ -- checks whether T is a reference type, and if so it adds an interprettion -- to Expr whose type is the designated type of the reference_discriminant. + procedure Check_Internal_Protected_Use (N : Node_Id; Nam : Entity_Id); + -- Within a protected function, the current object is a constant, and + -- internal calls to a procedure or entry are illegal. Similarly, other + -- uses of a protected procedure in a renaming or a generic instantiation + -- in the context of a protected function are illegal (AI05-0225). + procedure Check_Later_Vs_Basic_Declarations (Decls : List_Id; During_Parsing : Boolean); Index: sem_res.adb === --- sem_res.adb (revision 191888) +++ sem_res.adb (working copy) @@ -5314,15 +5314,7 @@ -- Check that this is not a call to a protected procedure or entry from -- within a protected function. - if Ekind (Current_Scope) = E_Function -and then Ekind (Scope (Current_Scope)) = E_Protected_Type -and then Ekind (Nam) /= E_Function -and then Scope (Nam) = Scope (Current_Scope) - then - Error_Msg_N (within protected function, protected - object is constant, N); - Error_Msg_N (\cannot call operation that may modify it, N); - end if; + Check_Internal_Protected_Use (N, Nam); -- Freeze the subprogram name if not in a spec-expression. Note that we -- freeze procedure calls as well as function calls. Procedure calls are @@ -6732,6 +6724,7 @@ end if; Resolve_Actuals (N, Nam); + Check_Internal_Protected_Use (N, Nam); -- Create a call reference to the entry Index: sem_attr.adb === --- sem_attr.adb
[Ada] Tagged /= operator in GNAT tree doesn't get fully resolved with -gnatc
When an inequality operator is used for a tagged type, the tree node for the inequality prior to expansion (such as with -gnatc) reflects /= operator in Standard rather than being resolved to be a logical negation of the tagged type's equality function. This is a problem for ASIS in Corresponding_Equality_Operator. We now ensure that the rewriting of a tagged inequality happens during analysis rather than being deferred to expansion. No simple test available. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Gary Dismukes dismu...@adacore.com * sem_ch4.adb (Find_Equality_Types.Try_One_Interp): Exclude the predefined interpretation from consideration if it's for a /= operator of a tagged type. This will allow Analyze_Equality_Op to rewrite the /= as a logical negation of a call to the appropriate dispatching equality function. This needs to be done during analysis rather than expansion for the benefit of ASIS, which otherwise gets the unresolved N_Op_Ne operator from Standard. Index: sem_ch4.adb === --- sem_ch4.adb (revision 191890) +++ sem_ch4.adb (working copy) @@ -5612,8 +5612,24 @@ return; end if; + -- If the right operand has a type compatible with T1, check for an + -- acceptable interpretation, unless T1 is limited (no predefined + -- equality available), or this is use of a /= for a tagged type. + -- In the latter case, possible interpretations of equality need to + -- be considered, we don't want the default inequality declared in + -- Standard to be chosen, and the /= will be rewritten as a + -- negation of = (see the end of Analyze_Equality_Op). This ensures + -- that that rewriting happens during analysis rather than being + -- delayed until expansion (this is needed for ASIS, which only sees + -- the unexpanded tree). Note that if the node is N_Op_Ne, but Op_Id + -- is Name_Op_Eq then we still proceed with the interpretation, + -- because that indicates the potential rewriting case where the + -- interpretation to consider is actually = and the node may be + -- about to be rewritten by Analyze_Equality_Op. + if T1 /= Standard_Void_Type and then Has_Compatible_Type (R, T1) + and then ((not Is_Limited_Type (T1) and then not Is_Limited_Composite (T1)) @@ -5622,6 +5638,11 @@ (Is_Array_Type (T1) and then not Is_Limited_Type (Component_Type (T1)) and then Available_Full_View_Of_Component (T1))) + + and then + (Nkind (N) /= N_Op_Ne + or else not Is_Tagged_Type (T1) + or else Chars (Op_Id) = Name_Op_Eq) then if Found and then Base_Type (T1) /= Base_Type (T_F)
[Ada] Next step in implementing extended overflow checking
This patch defines the four modes of overflow handling (SUPPRESSED, CHECKED, MINIMIZED, ELIMINATED), and adds the Overflow_Checks pragma and extedned -gnato switch to et them. But for now Checked, Minimized, and Eliminated are all treated as Checked, so the behavior is unchanged. The following program: 1. procedure over2 is 2.function Ident (X : Integer) return Integer is 3.begin 4. return X; 5.end; 6.x : integer := Ident (Integer'Last); 7.procedure g; 8.pragma Postcondition (x + 2 = 0); 9.procedure g is begin null; end; 10. begin 11.g; 12. end; compiled with -gnata -gnato10 generates raised SYSTEM.ASSERTIONS.ASSERT_FAILURE : failed postcondition from over2.adb:8 since -gnato10 turns off overflow checking in assertions resulting in the postcondition giving a result of false. compiled with -gnata -gnato01 generates raised CONSTRAINT_ERROR : over2.adb:8 overflow check failed since -gnato01 turns on overflow checking in assertions resulting in the overflow being detected. The following test program: 1. pragma Overflow_Checks 2. (Suppressed, Assertions = Checked); 3. procedure over21 is 4.function Ident (X : Integer) return Integer is 5.begin return X; end; 6.x : integer := Ident (Integer'Last); 7.procedure g; 8.pragma Postcondition (x + 2 = 0); 9.procedure g is begin null; end; 10. begin 11.g; 12. end; compiled with -gnata result in raised CONSTRAINT_ERROR : over21.adb:8 overflow check failed The following test program: 1. pragma Overflow_Checks 2. (Checked, Assertions = Suppressed); 3. procedure over22 is 4.function Ident (X : Integer) return Integer is 5.begin return X; end; 6.x : integer := Ident (Integer'Last); 7.procedure g; 8.pragma Postcondition (x + 2 = 0); 9.procedure g is begin null; end; 10. begin 11.g; 12. end; compiled with -gnata generates raised SYSTEM.ASSERTIONS.ASSERT_FAILURE : failed postcondition from over22.adb:8 Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Robert Dewar de...@adacore.com * checks.adb: Remove reference to Enable_Overflow_Checks Use Suppress_Options rather than Scope_Suppress. * gnat1drv.adb (Adjust_Global_Switches): Handle new overflow settings (Adjust_Global_Switches): Initialize Scope_Suppress from Suppress_Options. * opt.adb: Remove Enable_Overflow_Checks (use Suppress_Options instead). * opt.ads: Remove Overflow_Checks_Unsuppressed (not used) Remove Enable_Overflow_Checks (use Suppress_Options instead) Suppress_Options is now current setting (replaces Scope_Suppress). * osint.adb (Initialize): Deal with initializing overflow checking. * par-prag.adb: Add dummy entry for pragma Overflow_Checks. * sem.adb (Semantics): Save and restore In_Assertion_Expr Use Suppress_Options instead of Scope_Suppress. * sem.ads (In_Assertion_Expr): New flag (Scope_Suppress): Removed, use Suppress_Options instead. * sem_eval.adb (Compile_Time_Compare): Return Unknown in preanalysis mode. * sem_prag.adb (Process_Suppress_Unsuppress): Setting of Overflow_Checks_Unsuppressed removed (not used anywhere!) (Analyze_Pragma, case Check): Set In_Assertion_Expression (Analyze_Pragma, case Overflow_Checks): Implement new pragma * snames.ads-tmpl: Add names needed for handling pragma Overflow_Checks * switch-c.adb (Scan_Front_End_Switches) Handle -gnato? and -gnato?? where ? is 0-3 * types.ads: Updates and fixes to comment on Suppress_Record. Index: switch-c.adb === --- switch-c.adb(revision 191888) +++ switch-c.adb(working copy) @@ -128,9 +128,8 @@ -- Handle switches that do not start with -gnat - if Ptr + 3 Max -or else Switch_Chars (Ptr .. Ptr + 3) /= gnat - then + if Ptr + 3 Max or else Switch_Chars (Ptr .. Ptr + 3) /= gnat then + -- There are two front-end switches that do not start with -gnat: -- -I, --RTS @@ -755,11 +754,78 @@ when 'o' = Ptr := Ptr + 1; - Suppress_Options.Suppress (Overflow_Check) := False; - Suppress_Options.Overflow_Checks_General := Check_All; - Suppress_Options.Overflow_Checks_Assertions := Check_All; - Opt.Enable_Overflow_Checks := True; + -- Case of no digits after the -gnato + + if Ptr Max or else Switch_Chars (Ptr) not in '0' .. '3' then + Suppress_Options.Overflow_Checks_General:= Checked; + Suppress_Options.Overflow_Checks_Assertions :=
Re: [rtl] combine a vec_concat of 2 vec_selects from the same vector
/* Try to merge 2 VEC_SELECTs from the same vector into a single one. */ I was trying to avoid splitting in 2 lines, but ok I'll split. Indeed. Then you can remove the '2' above, it doesn't add much. -- Eric Botcazou
[Ada] Checks on aliasing and initialization of scalars for parameters
This patch reimplements the checks related to aliasing and initialization of scalars for subprogram parameters and ties them to compilation flags -gnateA and -gnateV respectively. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Hristian Kirtchev kirtc...@adacore.com * checks.adb (Apply_Parameter_Aliasing_Checks): Removed. (Apply_Parameter_Aliasing_And_Validity_Checks): New routine. (Apply_Parameter_Validity_Checks): Removed. * checks.ads (Apply_Parameter_Aliasing_Checks): Removed. (Apply_Parameter_Aliasing_And_Validity_Checks): New routine. (Apply_Parameter_Validity_Checks): Removed. * exp_ch6.adb (Expand_Call): Remove the generation of parameter aliasing checks. * freeze.adb: Remove with and use clauses for Validsw. (Freeze_Entity): Update the guard and generation of aliasing and scalar initialization checks for subprogram parameters. * opt.ads: Add new flags Check_Aliasing_Of_Parameters and Check_Validity_Of_Parameters along with comments on usage. * sem_attr.adb (Analyze_Attribute): Pragma Overlaps_Storage is no longer an Ada 2012 feature. * sem_ch4.adb: Remove with and use clauses for Checks and Validsw. (Analyze_Call): Remove the generation of aliasing checks for subprogram parameters. * sem_ch13.adb: Remove with and use clauses for Validsw. (Analyze_Aspect_Specifications): Remove the generation of scalar initialization checks. * switch-c.adb (Scan_Front_End_Switches): Add processing for -gnateA and -gnateV. * usage.adb (Usage): Add information on switches -gnateA and -gnateV. Remove information on validity switches 'l', 'L', 'v' and 'V'. * validsw.adb (Reset_Validity_Check_Options): Remove the reset of flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params. (Save_Validity_Check_Options): Remove the processing for flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params. (Set_Validity_Check_Options): Remove the processing for flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params. * validsw.ads: Remove flags Validity_Check_Non_Overlapping_Params and Validity_Check_Valid_Scalars_On_Params along with their comments on usage. Index: switch-c.adb === --- switch-c.adb(revision 191895) +++ switch-c.adb(working copy) @@ -380,6 +380,12 @@ Enable_Switch_Storing; Ptr := Ptr + 1; + -- -gnateA (aliasing checks on parameters) + + when 'A' = + Ptr := Ptr + 1; + Check_Aliasing_Of_Parameters := True; + -- -gnatec (configuration pragmas) when 'c' = @@ -566,6 +572,22 @@ when 'P' = Treat_Categorization_Errors_As_Warnings := True; + -- -gnateS (generate SCO information) + + -- Include Source Coverage Obligation information in ALI + -- files for the benefit of source coverage analysis tools + -- (xcov). + + when 'S' = + Generate_SCO := True; + Ptr := Ptr + 1; + + -- -gnateV (validity checks on parameters) + + when 'V' = + Ptr := Ptr + 1; + Check_Validity_Of_Parameters := True; + -- -gnatez (final delimiter of explicit switches) -- All switches that come after -gnatez have been added by @@ -577,16 +599,6 @@ Disable_Switch_Storing; Ptr := Ptr + 1; - -- -gnateS (generate SCO information) - - -- Include Source Coverage Obligation information in ALI - -- files for the benefit of source coverage analysis tools - -- (xcov). - - when 'S' = - Generate_SCO := True; - Ptr := Ptr + 1; - -- All other -gnate? switches are unassigned when others = Index: usage.adb === --- usage.adb (revision 191890) +++ usage.adb (working copy) @@ -167,6 +167,11 @@ Write_Switch_Char (Dnn); Write_Line (Debug expanded generated code (max line length = nn)); + -- Line for -gnateA switch + + Write_Switch_Char (eA); + Write_Line (Aliasing checks on subprogram parameters); + -- Line for -gnatec switch Write_Switch_Char (ec=?); @@ -227,6 +232,11 @@ Write_Switch_Char (eS); Write_Line (Generate SCO (Source Coverage
[Ada] Ada 2012 invariant checks on subcomponents
If a record has a subvomponent whose type has a defined invariant, then there must be a invariant check on that component whenever a value of the record type is created or modified by a visible primitive operation of the type. The command: gnatmake -q -gnat12 -gnata main main must yield: CHECK CHECK CHECK CHECK OK CHECK CHECK CHECK CHECK CHECK CHECK --- with Text_IO; use Text_IO; with Ada.Assertions; use Ada.Assertions; with P; procedure Main is V : P.R2; -- Check on the T default initialization? NOK Table : P.A2 (1..3); begin begin P.P3 (Table); exception when Assertion_Error = Put_Line (OK); end; P.Prim (V.V.V); -- Check on the T reference? OK P.P1 (V.V); -- Check on the part? NOK P.P2 (V); -- Check on the part? NOK declare Thing1 : P.R3 := P.F3 (True); Thing2 : P.R3 := P.F3 (False); begin null; end; end Main; -- package P is type T is private with Type_Invariant = Check (T); procedure Prim (V : in out T); function Check (V : in T) return Boolean; type R1 is record V : T; end record; type R2 is record V : R1; end record; type R3 (D : Boolean) is record V1 : T; case D is when True = V2 : T; when False = Zero : Integer := 0; end case; end record; type A2 is array (integer range ) of T; procedure P1 (V : in out R1); procedure P2 (V : in out R2); procedure P3 (T : in out A2); function F3 (Yes: Boolean) return R3; private type T is record Val : Integer := 17; end record; end P; --- with Ada.Text_IO; use Ada.Text_IO; package body P is --- -- Check -- --- function Check (V : in T) return Boolean is begin Put_Line (CHECK); return V.Val = 17; end Check; procedure Prim (V : in out T) is begin null; end Prim; procedure P1 (V : in out R1) is begin null; end P1; procedure P2 (V : in out R2) is begin null; end P2; procedure P3 (T : in out A2) is begin T (T'Last).Val := 18; null; end P3; function F3 (Yes : Boolean) return R3 is Result : R3 (Yes); begin return Result; end; end P; Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Ed Schonberg schonb...@adacore.com * exp_ch3.adb (Build_Record_Invariant_Proc): new procedure to generate a checking procedure for record types that may have components whose types have type invariants declared. Index: exp_ch3.adb === --- exp_ch3.adb (revision 191894) +++ exp_ch3.adb (working copy) @@ -118,6 +118,10 @@ -- Build record initialization procedure. N is the type declaration -- node, and Rec_Ent is the corresponding entity for the record type. + procedure Build_Record_Invariant_Proc (R_Type : Entity_Id; Nod : Node_Id); + -- If the record type has components whose types have invariant, build + -- an invariant procedure for the record type itself. + procedure Build_Slice_Assignment (Typ : Entity_Id); -- Build assignment procedure for one-dimensional arrays of controlled -- types. Other array and slice assignments are expanded in-line, but @@ -3611,6 +3615,174 @@ end if; end Build_Record_Init_Proc; + + -- Build_Record_Invariant_Proc -- + + + procedure Build_Record_Invariant_Proc (R_Type : Entity_Id; Nod : Node_Id) is + Loc : constant Source_Ptr := Sloc (Nod); + + Object_Name : constant Name_Id := New_Internal_Name ('I'); + -- Name for argument of invariant procedure + + Object_Entity : constant Node_Id := +Make_Defining_Identifier (Loc, Object_Name); + -- The procedure declaration entity for the argument + + Invariant_Found : Boolean; + -- Set if any component needs an invariant check. + + Proc_Id : Entity_Id; + Proc_Body : Node_Id; + Stmts : List_Id; + Type_Def : Node_Id; + + function Build_Invariant_Checks (Comp_List : Node_Id) return List_Id; + -- Recursive procedure that generates a list of checks for components + -- that need it, and recurses through variant parts when present. + + function Build_Component_Invariant_Call (Comp : Entity_Id) + return Node_Id; + -- Build call to invariant procedure for a record component. + + + -- Build_Component_Invariant_Call -- + + + function Build_Component_Invariant_Call (Comp : Entity_Id) + return Node_Id + is + Sel_Comp : Node_Id; + + begin + Invariant_Found := True; + Sel_Comp := + Make_Selected_Component (Loc, + Prefix = New_Occurrence_Of (Object_Entity, Loc), + Selector_Name = New_Occurrence_Of (Comp, Loc)); + +
[Ada] Additional invariant checks on composite types
If a composite type has a declared invariant, and some of its compoents are of types that have their own invariants, the invariant checks on those compoents must be added to the invariant checks for the enclosing type. The command; gnatmake -q -gnat12 -gnata test_bars test_bars must yield: chart invariant violation detected value invariant violation detected with Ada.Assertions; use Ada.Assertions; with Bars; use Bars; with Text_IO; use Text_IO; procedure Test_Bars is B : Bar_chart := Bare_Bar (5); D : Data (1 .. 5) := (20, 20, 20, 20, 19); begin begin Assemble (D, B); exception when Assertion_Error = Put_Line (chart invariant violation detected); end; declare D : Data (1 .. 5) := (30, 30, 30, 30, -20); begin Assemble (D, B); exception when Assertion_Error = Put_Line (value invariant violation detected); end; end; --- package Bars is type Value is private with Invariant = Legal (Value); type Bar_Chart () is private with Invariant = Complete (Bar_Chart); type Data is array (positive range ) of Integer; function Legal (It : Value) return Boolean; function Complete (It : Bar_Chart) return Boolean; function Bare_Bar (N : Positive) return Bar_Chart; procedure Assemble (From : Data; Result : out Bar_Chart); private type Value is new Integer; type Bar_Chart is array (positive range ) of Value; end; --- package body Bars is -- type Value is private --with Invariant = Legal (Value); -- type Bar_Chart is private --with Invariant = Complete (Bar_Chart); function Legal (It : Value) return Boolean is begin return It = 0 and It = 100; end; function Complete (It : Bar_Chart) return Boolean is Total : Value := 0; begin for B of It loop Total := Total + B; end loop; return Total = 100; end; function Bare_Bar (N : Positive) return Bar_Chart is Result : Bar_Chart (1 .. N) := (100, others = 0); begin return Result; end; procedure Assemble (From : Data; Result : out Bar_Chart) is begin for J in From'range loop Result (J) := Value (From (J)); end loop; end; end; Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Ed Schonberg schonb...@adacore.com * exp_ch3.ads (Build_Array_Invariant_Proc): moved to body. * exp_ch3.adb (Build_Array_Invariant_Proc, Build_Record_Invariant_Proc): transform into functions. (Insert_Component_Invariant_Checks): for composite types that have components with specified invariants, build a checking procedure, and make into the invariant procedure of the composite type, or incorporate it into the user- defined invariant procedure if one has been created. * sem_ch3.adb (Array_Type_Declaration): Checking for invariants on the component type is defered to the expander. Index: exp_ch3.adb === --- exp_ch3.adb (revision 191902) +++ exp_ch3.adb (working copy) @@ -88,6 +88,22 @@ -- used for attachment of any actions required in its construction. -- It also supplies the source location used for the procedure. + function Build_Array_Invariant_Proc + (A_Type : Entity_Id; + Nod: Node_Id) return Node_Id; + -- If the component of type of array type has invariants, build procedure + -- that checks invariant on all components of the array. Ada 2012 specifies + -- that an invariant on some type T must be applied to in-out parameters + -- and return values that include a part of type T. If the array type has + -- an otherwise specified invariant, the component check procedure is + -- called from within the user-specified invariant. Otherwise this becomes + -- the invariant procedure for the array type. + + function Build_Record_Invariant_Proc + (R_Type : Entity_Id; + Nod: Node_Id) return Node_Id; + -- Ditto for record types. + function Build_Discriminant_Formals (Rec_Id : Entity_Id; Use_Dl : Boolean) return List_Id; @@ -118,10 +134,6 @@ -- Build record initialization procedure. N is the type declaration -- node, and Rec_Ent is the corresponding entity for the record type. - procedure Build_Record_Invariant_Proc (R_Type : Entity_Id; Nod : Node_Id); - -- If the record type has components whose types have invariant, build - -- an invariant procedure for the record type itself. - procedure Build_Slice_Assignment (Typ : Entity_Id); -- Build assignment procedure for one-dimensional arrays of controlled -- types. Other array and slice assignments are expanded in-line, but @@ -184,6 +196,14 @@ -- Treat user-defined stream operations as renaming_as_body if the -- subprogram they rename is not frozen when the type is frozen. + procedure Insert_Component_Invariant_Checks + (N : Node_Id; +
Profile housekeeping 5/n (make RTL loop optimizers to use loop bounds better)
Hi, this patch commonizes the maximal iteration estimate logic in between SCEV and loop-iv. Both are now using loop-nb_iterations_upper_bound. I decided to keep same API for SCEV code as for RTL code, so I made estimated_loop_iterations and max_loop_iterations to not try to recompute bounds and ICE when invoked without SCEV fired on. The patch updates RTL optimizers to use the estimated_loop_iterations and max_loop_iterations. This has few advantages: 1) loop unroller can now take into account estimates stored into loop structure by earlier pass (I think none exist though) It is however better then using expected_loop_iterations since profile might get out of date with expansion. 2) loop peeling code now use max iterations bounds. This makes it i.e. to peel vectorizer prologues/epilogues/scalar loops so -fpeel-loops now improves my low iteration count testcase by about 10% 3) Same for loop unswithcing. I am not really friend with the new double_int API. I copied some existing examples but find it ugly. Why do we miss operators for comparsions and division? Why from_*/to_* can't be a cast at least for basic integer types? Regtested/bootstrapped x86_64-linux, seems sane? I also wonder if loop vectorizer should not update the estimates after loop iteration count is reduced by vectorizing. Honza * loop-unswitch.c (unswitch_single_loop): Use estimated_loop_iterations_int to prevent unswitching when loop is known to not roll. * tree-ssa-loop-niter.c (estimated_loop_iterations): Do not segfault when SCEV is not initialized. (max_loop_iterations): Likewise. * tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops): Use estimated_loop_iterations_int to prevent unswithcing when loop is known to not roll. * tree-scalar-evolution.c (scev_initialized_p): New function. * tree-scalar-evolution.h (scev_initialized_p): Likewise. * loop-unroll.c (decide_peel_once_rolling): Use max_loop_iterations_int. (unroll_loop_constant_iterations): Update nb_iterations_upper_bound and nb_iterations_estimate. (decide_unroll_runtime_iterations): Use estimated_loop_iterations or max_loop_iterations; (unroll_loop_runtime_iterations): fix profile updating. (decide_peel_simple): Use estimated_loop_iterations and max_loop_iterations. (decide_unroll_stupid): Use estimated_loop_iterations ad max_loop_iterations. * loop-doloop.c (doloop_modify): Use max_loop_iterations_int. (doloop_optimize): Likewise. * loop-iv.c (iv_number_of_iterations): Use record_niter_bound. (find_simple_exit): Likewise. * cfgloop.h (struct niter_desc): Remove niter_max. Index: loop-unswitch.c === *** loop-unswitch.c (revision 191867) --- loop-unswitch.c (working copy) *** unswitch_single_loop (struct loop *loop, *** 257,262 --- 257,263 rtx cond, rcond = NULL_RTX, conds, rconds, acond, cinsn; int repeat; edge e; + HOST_WIDE_INT iterations; /* Do not unswitch too much. */ if (num PARAM_VALUE (PARAM_MAX_UNSWITCH_LEVEL)) *** unswitch_single_loop (struct loop *loop, *** 299,305 } /* Nor if the loop usually does not roll. */ ! if (expected_loop_iterations (loop) 1) { if (dump_file) fprintf (dump_file, ;; Not unswitching, loop iterations 1\n); --- 300,307 } /* Nor if the loop usually does not roll. */ ! iterations = estimated_loop_iterations_int (loop); ! if (iterations = 0 iterations = 1) { if (dump_file) fprintf (dump_file, ;; Not unswitching, loop iterations 1\n); Index: tree-ssa-loop-niter.c === *** tree-ssa-loop-niter.c (revision 191867) --- tree-ssa-loop-niter.c (working copy) *** estimate_numbers_of_iterations_loop (str *** 3012,3020 bool estimated_loop_iterations (struct loop *loop, double_int *nit) { ! estimate_numbers_of_iterations_loop (loop); if (!loop-any_estimate) ! return false; *nit = loop-nb_iterations_estimate; return true; --- 3012,3034 bool estimated_loop_iterations (struct loop *loop, double_int *nit) { ! /* When SCEV information is available, try to update loop iterations ! estimate. Otherwise just return whatever we recorded earlier. */ ! if (scev_initialized_p ()) ! estimate_numbers_of_iterations_loop (loop); ! ! /* Even if the bound is not recorded, possibly we can derrive one from ! profile. */ if (!loop-any_estimate) ! { ! if (loop-header-count) ! { ! *nit = gcov_type_to_double_int ! (expected_loop_iterations_unbounded (loop) + 1); ! return true; ! } !
Re: [patch] Minor TARGET_MEM_REF cleanup
On Sat, Sep 29, 2012 at 1:17 PM, Eric Botcazou ebotca...@adacore.com wrote: Hi, for simple loops like: extern int a[]; extern int b[]; void foo (int l) { int i; for (i = 0; i l; i++) a[i] = b [i]; } you get in the .lim3 dump: Unanalyzed memory reference 0: _5 = MEM[symbol: b, index: ivtmp.3_1, step: 4, offset: 0B]; Memory reference 1: MEM[symbol: a, index: ivtmp.3_1, step: 4, offset: 0B] so the pass analyzes the store but not the load, which seems an oversight. The patch also folds copy_mem_ref_info into its only user and removes it. Tested on x86_64-suse-linux, OK for mainline? Please take the opportunity to clean up simple_mem_ref_in_stmt some more. Both loads and stores in assigns require gimple_assign_single_p, thus do if (!gimple_assign_single_p (stmt)) return NULL; before deciding on store/load. To decide that the stmt is a load then do if (TREE_CODE (*lhs) == SSA_NAME gimple_vuse (stmt)) it is a store if gimple_vdef (stmt) (TREE_CODE (*rhs) == SSA_NAME || is_gimple_min_invariant (*rhs)) else it may still be an aggregate copy but LIM doesn't handle those (though they may still be interesting for disambiguation ...) The tree-ssa-address parts are ok as-is. Thanks, Richard. 2012-09-29 Eric Botcazou ebotca...@adacore.com * tree.h (copy_mem_ref_info): Delete. * tree-ssa-address.c (copy_mem_ref_info): Likewise. (maybe_fold_tmr): Copy flags manually. * tree-ssa-loop-im.c (simple_mem_ref_in_stmt): Accept TARGET_MEM_REF on the RHS as well. -- Eric Botcazou
Re: [PATCH RFA] Implement register pressure directed hoist pass
On Sat, Sep 29, 2012 at 8:37 AM, Bin Cheng bin.ch...@arm.com wrote: This is the updated patch according to your comments. Please review. I also re-collected code size data and found it is improved by about 0.24% for mips, which is better than previous data. I believe this should be caused by recent changes in trunk, rather than by using DF caches to calculate register pressure. Hello, Thanks for the update. The first look wasn't a very thorough review, so I have more comments now. Sorry for that, I should have taken the time for this the first time round... First, as a general note: Please add a fat comment somewhere before the code of the pass to explain how the register pressure driven hoisting pass works. The existing implementation used to be an almost one-to-one copy from Muchnick's book, but lately the code has moved away from that (e.g. with Maxim's patches from last year) and it gets a bit hard to follow without some explanation. You should at least document your own algorithm, but I'd really appreciate if you could also spend some time writing down how things work in general and/or how the GCC implementation differs from Muchnick's original algorithm. +Use IRA to evaluate register pressure in hoist pass for decisions to hoist the code hoisting pass (brownie points for an xref to the flag for code hoisting). - - do rough calc of how many regs are needed in each block, and a rough - calc of how many regs are available in each class and use that to - throttle back the code in cases where RTX_COST is minimal. This comment still applies to LCM. +/* Record all regs that are set in any one insn. Communication from + mark_reg_{store,clobber} and global_conflicts. Asm can refer to + all hard-registers. */ +static rtx regs_set[(FIRST_PSEUDO_REGISTER MAX_RECOG_OPERANDS + ? FIRST_PSEUDO_REGISTER : MAX_RECOG_OPERANDS) * 2]; +/* Number of regs stored in the previous array. */ +static int n_regs_set; You can use DF_INSN_DEFS for this in calculate_bb_reg_pressure. But then again... + while (n_regs_set-- 0) + { + rtx note = find_regno_note (insn, REG_UNUSED, + REGNO (regs_set[n_regs_set])); + if (! note) + continue; + + mark_reg_death (XEXP (note, 0)); + } Why not just mark all registers mentioned in REG_UNUSED notes as death, i.e.: for (note = REG_NOTES (insn); note; note = XEXP (note, 1)) if ((REG_NOTE_KIND (note) == REG_UNUSED) mark_reg_death (XEXP (note, 0)); ? +static int hoist_expr_reaches_here_p (basic_block, struct expr*, basic_block, space between struct expr and *. + int *, bitmap_head *); bitmap_head * = bitmap. Likewise here: /* Basic blocks that have occurrences reachable from BB. */ bitmap_head _from_bbs, *from_bbs = _from_bbs; +/* Basic blocks through which expr is hoisted. */ +bitmap_head _hoisted_bbs, *hoisted_bbs = _hoisted_bbs; And using BITMAP_ALLOC here: bitmap_initialize (from_bbs, 0); + if (flag_ira_hoist_pressure == 1) + bitmap_initialize (hoisted_bbs, 0); (Some older code uses the bitmap_head * form, but that pre-dates coretypes.h. Newer code that uses the old form should really be fixed also.) Please also use an obstack for the hoisted_bbs and from_bbs bitmaps. They're currently allocated in GC space but that's completely unnecessary for objects with such a clearly identifiable life time. You can even allocate these bitmaps outside the loop, you're already calling bitmap_clear on them. EXPR_BB. Stop Two spaces after a period in comments. @@ -2863,7 +2909,8 @@ static int if (visited == NULL) { visited_allocated_locally = 1; - visited = XCNEWVEC (char, last_basic_block); + visited = sbitmap_alloc (last_basic_block); + sbitmap_zero (visited); } Can you please submit this as a separate patch? I'd go ahead and commit it as obvious, but it's always good to have one change per patch and this bit is independent of the reg-pressure dependent hoisting. +fira-hoist-pressure +Common Report Var(flag_ira_hoist_pressure) Init(-1) +Use IRA based register pressure calculation +in hoist optimizations. Please add the Optimization marker. Why initialize to -1? The default initialization is 0, and if the flag is set it takes a value 1. If you follow that common behavior, you can replace all occurrences of if (flag_ira_hoist_pressure == 1) with just if (flag_ira_hoist_pressure). + /* Enable register pressure hoist when optimizing for size on Thumb1 set. */ + if (TARGET_THUMB1 optimize_function_for_size_p (cfun) + flag_ira_hoist_pressure == -1) +flag_ira_hoist_pressure = 1; One would expect this to be a win on all targets, but you probably looked at this (at least Thumb2 and plain ARM). Did your patch not help
Re: [PATCH] Fix PR middle-end/54759
On Sun, Sep 30, 2012 at 9:03 PM, Dehao Chen de...@google.com wrote: Hi, This patch fixes the bug when comparing location to UNKNOWN_LOC. Bootstrapped and passed gcc regression test. Okay for trunk? Ok. Thanks, Richard. Thanks, Dehao 2012-09-30 Dehao Chen de...@google.com PR middle-end/54759 * gcc/tree-vect-loop-manip.c (slpeel_make_loop_iterate_ntimes): Use LOCATION_LOCUS to compare with UNKNOWN_LOCATION. (slpeel_tree_peel_loop_to_edge): Likewise. * gcc/tree-vectorizer.c (vectorize_loops): Likewise.
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 7:48 AM, Jakub Jelinek ja...@redhat.com wrote: On Sun, Sep 30, 2012 at 06:50:50PM -0400, Vladimir Makarov wrote: But I think that LRA cpu time problem for this test can be fixed. But I don't think I can fix it for 2 weeks. So if people believe that current LRA behaviour on this PR is a stopper to include it into gcc4.8 than we should postpone its inclusion until gcc4.9 when I hope to fix it. I think this testcase shouldn't be a show stopper for LRA inclusion into 4.8, but something to look at for stage3. I agree here. I think a lot of GCC passes have scalability issues on that testcase, that is why it must be compiled with -O1 and not higher optimization options, so perhaps it would be enough to choose a faster algorithm generating worse code for the huge functions and -O1. Yes, we spent quite some time in making basic optimization work for insane testcases (basically avoid quadratic or bigger complexity in any IL size variable (number of basic-blocks, edges, instructions, pseudos, etc.)). And indeed if you use -O2 we do have issues with existing passes (and even at -O1 points-to analysis can wreck things, or even profile guessing - see existing bugs for that). Basically I would tune -O1 towards being able to compile and optimize insane testcases with memory and compile-time requirements that are linear in any of the above complexity measures. Thus, falling back to the -O0 register allocating strathegy at certain thresholds for the above complexity measures is fine (existing IRA for example has really bad scaling on the number of loops in the function, but you can tweak with flags to make it not consider that). And I agree it is primarily a bug in the generator that it creates such huge functions, that can't perform very well. Well, not for -O2+, yes, but at least we should try(!) hard. Thanks, Richard. Jakub
Re: [Patch,avr]: Ad PR rtl-optimization/52543: Undo the MEM-UNSPEC hack
2012/9/30 Georg-Johann Lay a...@gjlay.de: Denis Chertykov wrote: I have tried to use secondary a few years ago (may be 5 or 7). I have definitely remember only one thing: secondary reload should be avoided as long as possible. Currently each mov has to be decorated with moving the segment to RAMPZ and (depending on target) restoring RAMPZ afterwards. GCC has no concept of a segmented layout and there is no way to describe that. One way is to hack with UNSPEC and bypass ira/reload altogether but IMO that is no good solution. Besides that is only works because the mov insns have special constraints (there will be writes to flash, flash does not change after load time, etc.) The better way to got a knowledge about it is a GDB ;-) I think reload.c:push_secondary_reload() should be the right place but it does not call targetm.secondary_reload so that no secondary is generated. It's hard to tell where the place is that is responsible for the bypassing of calling the hook. From the internals I don't see why it is skipped and the responsiveness in the gcc-help@ list on such topics is zero :-( IMHO it's a question to gcc@ not to gcc-help@ Ok, I will try my luck again. Do you have an idea for a better approach, i.e. not set RAMPZ over and over again? May be it's a question similar to addressing with infinite displacement optimization that we discuss in long long thread - [Patch, AVR]: Fix PR46779 Generally, the answer is here: http://gcc.gnu.org/ml/gcc-patches/2011-06/msg01281.html and here: The biggest problem that I see, from the 950612-1.c test case with the current handling of the infinite displacement frame pointer, is that the adjustments to the frame pointer are never exposed as separate instructions, so there's never a chance to optimize them. You can set RAMPZ over and over but you can give GCC a chance to optimize it out in cse or cse-postreload passes. Denis. PS: as you know GCC has no concept of a segmented layout may be better just drop it.
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 9:16 AM, Jakub Jelinek ja...@redhat.com wrote: On Mon, Oct 01, 2012 at 08:47:13AM +0200, Steven Bosscher wrote: The test case compiles just fine at -O2, only VRP has trouble with it. Let's try to stick with facts, not speculation. I was talking about the other PR, PR26854, which from what I remember when trying it myself and even the latest -O3 time reports from the reduced testcase show that IRA/reload aren't there very significant (for -O3 IRA takes ~ 6% and reload ~ 1%). OK, but what does LRA take? Vlad's numbers for 64-bits and looking at user time: Reload: 503.26user LRA: 598.70user So if reload is ~1% of 503s then that'd be ~5s. And the only difference between the two timings is LRA instead of reload, so LRA takes ~100s, or 20%. I've put a lot of hard work into it to fix almost all scalability problems on this PR for gcc 4.8. LRA undoes all of that work. I understand it is painful for some people to hear, but I remain of opinion that LRA cannot be considered ready if it scales so much worse than everything else in the compiler. Judging the whole implementation from just these corner cases and not how it performs on other testcases (SPEC, rebuild a distro, ...) is IMHO not the right thing, if Vlad thinks the corner cases are fixable during stage3; IMHO we should allow LRA in, worst case it can be disabled by default even for i?86/x86_64. I'd be asked to do a guest lecture on compiler construction (to be clear: I'd be highly surprised if anyone would ask me to, but for sake of argument, bear with me ;-) then I'd start by stating that algorithms should be designed for the corner cases, because the devil's always in the details. But more to the point regarding stage3: It will already be a busy stage3 if the other, probably even more significant, scalability issues have to be fixed, i.e. var-tracking and macro expansion. And there's also the symtab work that's bound to cause some interesting bugs still to be shaken out. With all due respect to Vlad, and, seriously, hats off to Vlad for tacking reload and coming up with a much easier to understand and nicely phase-split replacement, I just don't believe that these scalability issues can be addressed in stage3. It's now very late stage1, and LRA was originally scheduled for GCC 4.9. Why the sudden hurrying? Did I miss the 2 minute warning? Ciao! Steven
Re: Profile housekeeping 5/n (make RTL loop optimizers to use loop bounds better)
On Mon, Oct 1, 2012 at 11:37 AM, Jan Hubicka hubi...@ucw.cz wrote: Hi, this patch commonizes the maximal iteration estimate logic in between SCEV and loop-iv. Both are now using loop-nb_iterations_upper_bound. I decided to keep same API for SCEV code as for RTL code, so I made estimated_loop_iterations and max_loop_iterations to not try to recompute bounds and ICE when invoked without SCEV fired on. The patch updates RTL optimizers to use the estimated_loop_iterations and max_loop_iterations. This has few advantages: 1) loop unroller can now take into account estimates stored into loop structure by earlier pass (I think none exist though) It is however better then using expected_loop_iterations since profile might get out of date with expansion. 2) loop peeling code now use max iterations bounds. This makes it i.e. to peel vectorizer prologues/epilogues/scalar loops so -fpeel-loops now improves my low iteration count testcase by about 10% 3) Same for loop unswithcing. I am not really friend with the new double_int API. I copied some existing examples but find it ugly. Why do we miss operators for comparsions and division? Why from_*/to_* can't be a cast at least for basic integer types? Regtested/bootstrapped x86_64-linux, seems sane? Yes. Can you add a testcase or two? I tweaked RTL unroll/peel after preserving loops to not blindly unroll/peel everything 8 times (not sure if _I_ added testcases ...). I also wonder if loop vectorizer should not update the estimates after loop iteration count is reduced by vectorizing. Probably yes. Thanks, Richard. Honza * loop-unswitch.c (unswitch_single_loop): Use estimated_loop_iterations_int to prevent unswitching when loop is known to not roll. * tree-ssa-loop-niter.c (estimated_loop_iterations): Do not segfault when SCEV is not initialized. (max_loop_iterations): Likewise. * tree-ssa-loop-unswitch.c (tree_ssa_unswitch_loops): Use estimated_loop_iterations_int to prevent unswithcing when loop is known to not roll. * tree-scalar-evolution.c (scev_initialized_p): New function. * tree-scalar-evolution.h (scev_initialized_p): Likewise. * loop-unroll.c (decide_peel_once_rolling): Use max_loop_iterations_int. (unroll_loop_constant_iterations): Update nb_iterations_upper_bound and nb_iterations_estimate. (decide_unroll_runtime_iterations): Use estimated_loop_iterations or max_loop_iterations; (unroll_loop_runtime_iterations): fix profile updating. (decide_peel_simple): Use estimated_loop_iterations and max_loop_iterations. (decide_unroll_stupid): Use estimated_loop_iterations ad max_loop_iterations. * loop-doloop.c (doloop_modify): Use max_loop_iterations_int. (doloop_optimize): Likewise. * loop-iv.c (iv_number_of_iterations): Use record_niter_bound. (find_simple_exit): Likewise. * cfgloop.h (struct niter_desc): Remove niter_max. Index: loop-unswitch.c === *** loop-unswitch.c (revision 191867) --- loop-unswitch.c (working copy) *** unswitch_single_loop (struct loop *loop, *** 257,262 --- 257,263 rtx cond, rcond = NULL_RTX, conds, rconds, acond, cinsn; int repeat; edge e; + HOST_WIDE_INT iterations; /* Do not unswitch too much. */ if (num PARAM_VALUE (PARAM_MAX_UNSWITCH_LEVEL)) *** unswitch_single_loop (struct loop *loop, *** 299,305 } /* Nor if the loop usually does not roll. */ ! if (expected_loop_iterations (loop) 1) { if (dump_file) fprintf (dump_file, ;; Not unswitching, loop iterations 1\n); --- 300,307 } /* Nor if the loop usually does not roll. */ ! iterations = estimated_loop_iterations_int (loop); ! if (iterations = 0 iterations = 1) { if (dump_file) fprintf (dump_file, ;; Not unswitching, loop iterations 1\n); Index: tree-ssa-loop-niter.c === *** tree-ssa-loop-niter.c (revision 191867) --- tree-ssa-loop-niter.c (working copy) *** estimate_numbers_of_iterations_loop (str *** 3012,3020 bool estimated_loop_iterations (struct loop *loop, double_int *nit) { ! estimate_numbers_of_iterations_loop (loop); if (!loop-any_estimate) ! return false; *nit = loop-nb_iterations_estimate; return true; --- 3012,3034 bool estimated_loop_iterations (struct loop *loop, double_int *nit) { ! /* When SCEV information is available, try to update loop iterations ! estimate. Otherwise just return whatever we recorded earlier. */ ! if (scev_initialized_p ()) !
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 11:52 AM, Richard Guenther richard.guent...@gmail.com wrote: I think this testcase shouldn't be a show stopper for LRA inclusion into 4.8, but something to look at for stage3. I agree here. I would also agree if it were not for the fact that IRA is already a scalability bottle-neck and that has been known for a long time, too. I have no confidence at all that if LRA goes in now, these scalability problems will be solved in stage3 or at any next release cycle. It's always the same thing with GCC: Once a patch is in, everyone moves on to the next fancy new thing dropping the not-quite-broken but also not-quite-working things on the floor. Ciao! Steven
Re: RFC: LRA for x86/x86-64 [0/9]
On Sun, Sep 30, 2012 at 7:03 PM, Richard Guenther richard.guent...@gmail.com wrote: On Sun, Sep 30, 2012 at 6:52 PM, Steven Bosscher stevenb@gmail.com wrote: Hi, To look at it in yet another way: integrated RA : 189.34 (16%) usr LRA non-specific: 59.82 ( 5%) usr LRA virtuals eliminatenon: 56.79 ( 5%) usr LRA create live ranges : 175.30 (15%) usr LRA hard reg assignment : 130.85 (11%) usr The IRA pass is slower than the next-slowest pass (tree PRA) by almost a factor 2.5. Each of the individually-measured *phases* of LRA is slower than the complete IRA *pass*. These 5 timevars together make up for 52% of all compile time. That figure indeed makes IRA + LRA look bad. Did you by chance identify anything obvious that can be done to improve the situation? The LRA create live range time is mostly spent in merge_live_ranges walking lists. Perhaps the live ranges can be represented better with a sorted VEC, so that the start and finish points can be looked up on log-time instead of linear. Ciao! Steven
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 01, 2012 at 12:01:36PM +0200, Steven Bosscher wrote: I would also agree if it were not for the fact that IRA is already a scalability bottle-neck and that has been known for a long time, too. I have no confidence at all that if LRA goes in now, these scalability problems will be solved in stage3 or at any next release cycle. It's always the same thing with GCC: Once a patch is in, everyone moves on to the next fancy new thing dropping the not-quite-broken but also not-quite-working things on the floor. If we open a P1 bug for it for 4.8, then it will need to be resolved some way before branching. I think Vlad is committed to bugfixing LRA, after all the intent is for 4.9 to enable it on more (all?) targets, and all the bugfixing and scalability work on LRA is needed for that anyway. Jakub
[Ada] Implement extended overflow handling for comparison ops
This patch enables extended overflow handling for comparison ops so that the comparison can be done in an expanded type, or even in bignum mode if operating in ELIMINATED overflow check mode. The following test program: 1. with Text_IO; use Text_IO; 2. procedure Overflowm2 is 3.function r1 4. (a, b, c, d : Integer) return Boolean is 5.begin 6. return a + b + c + d = Integer'Last; 7.end; 8.function r2 9. (a, b, c, d : Integer) return Boolean is 10.begin 11. return a * b * c * d = Integer'First; 12.end; 13. begin 14.begin 15. Put_Line 16. (r1 returns 17.Boolean'Image 18. (r1 (Integer'Last, Integer'Last, 19. -Integer'Last, -Integer'Last))); 20.exception 21. when Constraint_Error = 22. Put_Line (r1 raises exception); 23.end; 24. 25.begin 26. Put_Line 27. (r2 returns 28.Boolean'Image 29. (r2 (Integer'Last, Integer'Last, 30. Integer'Last, 0))); 31.exception 32. when Constraint_Error = 33. Put_Line (r2 raises exception); 34.end; 35. end Overflowm2; In CHECKED mode (-gnato1) we get: r1 raises exception r2 raises exception since the first addition in r1 and the first multiplication in r2 result in values outside the bounds of Integer'Base. In MINIMIZED mode (-gnato2) we get: r1 returns TRUE r2 raises exception since we can compute the addition result in Long_Long_Integer, and do the comparison in Long_Long_Integer mode, but the second multiplication yields a value outside this range, so that causes an overflow. In ELIMINATE mode (-gnato3) we get: r1 returns TRUE r2 returns TRUE Because now we use Bignum arithmetic for the intermediate multiplication results, and the final comparison is also done in bignum mode. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Robert Dewar de...@adacore.com * exp_ch4.adb (Expand_Compare_Minimize_Eliminate_Overflow): New procedure. Index: exp_ch4.adb === --- exp_ch4.adb (revision 191912) +++ exp_ch4.adb (working copy) @@ -140,6 +140,10 @@ procedure Expand_Short_Circuit_Operator (N : Node_Id); -- Common expansion processing for short-circuit boolean operators + procedure Expand_Compare_Minimize_Eliminate_Overflow (N : Node_Id); + -- Deal with comparison in Minimize/Eliminate overflow mode. This is where + -- we allow comparison of out of range values. + function Expand_Composite_Equality (Nod: Node_Id; Typ: Entity_Id; @@ -2276,6 +2280,237 @@ end; end Expand_Boolean_Operator; + + -- Expand_Compare_Minimize_Eliminate_Overflow -- + + + procedure Expand_Compare_Minimize_Eliminate_Overflow (N : Node_Id) is + Loc : constant Source_Ptr := Sloc (N); + + Llo, Lhi : Uint; + Rlo, Rhi : Uint; + + LLIB : constant Entity_Id := Base_Type (Standard_Long_Long_Integer); + -- Entity for Long_Long_Integer'Base + + Check : constant Overflow_Check_Type := Overflow_Check_Mode (Empty); + -- Current checking mode + + procedure Set_True; + procedure Set_False; + -- These procedures rewrite N with an occurrence of Standard_True or + -- Standard_False, and then makes a call to Warn_On_Known_Condition. + + --- + -- Set_False -- + --- + + procedure Set_False is + begin + Rewrite (N, New_Occurrence_Of (Standard_False, Loc)); + Warn_On_Known_Condition (N); + end Set_False; + + -- + -- Set_True -- + -- + + procedure Set_True is + begin + Rewrite (N, New_Occurrence_Of (Standard_True, Loc)); + Warn_On_Known_Condition (N); + end Set_True; + + -- Start of processing for Expand_Compare_Minimize_Eliminate_Overflow + + begin + -- Nothing to do unless we have a comparison operator with operands + -- that are signed integer types, and we are operating in either + -- MINIMIZED or ELIMINATED overflow checking mode. + + if Nkind (N) not in N_Op_Compare +or else Check not in Minimized_Or_Eliminated +or else not Is_Signed_Integer_Type (Etype (Left_Opnd (N))) + then + return; + end if; + + -- OK, this is the case we are interested in. First step is to process + -- our operands using the Minimize_Eliminate circuitry which applies + -- this processing to the two operand subtrees. + + Minimize_Eliminate_Overflow_Checks (Left_Opnd (N), Llo, Lhi); + Minimize_Eliminate_Overflow_Checks (Right_Opnd (N), Rlo, Rhi); + + --
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 12:14 PM, Jakub Jelinek ja...@redhat.com wrote: On Mon, Oct 01, 2012 at 12:01:36PM +0200, Steven Bosscher wrote: I would also agree if it were not for the fact that IRA is already a scalability bottle-neck and that has been known for a long time, too. I have no confidence at all that if LRA goes in now, these scalability problems will be solved in stage3 or at any next release cycle. It's always the same thing with GCC: Once a patch is in, everyone moves on to the next fancy new thing dropping the not-quite-broken but also not-quite-working things on the floor. If we open a P1 bug for it for 4.8, then it will need to be resolved some way before branching. I think Vlad is committed to bugfixing LRA, after all the intent is for 4.9 to enable it on more (all?) targets, and all the bugfixing and scalability work on LRA is needed for that anyway. I don't question Vlad's commitment, but the last time I met him, he only had two hands just like everyone else. But I've made my point and it seems that I'm not voting with the majority. Ciao! Steven
Re: RFC: LRA for x86/x86-64 [0/9]
On 10/01/2012 12:14 PM, Jakub Jelinek wrote: On Mon, Oct 01, 2012 at 12:01:36PM +0200, Steven Bosscher wrote: I would also agree if it were not for the fact that IRA is already a scalability bottle-neck and that has been known for a long time, too. I have no confidence at all that if LRA goes in now, these scalability problems will be solved in stage3 or at any next release cycle. It's always the same thing with GCC: Once a patch is in, everyone moves on to the next fancy new thing dropping the not-quite-broken but also not-quite-working things on the floor. If we open a P1 bug for it for 4.8, then it will need to be resolved some way before branching. I think Vlad is committed to bugfixing LRA, after all the intent is for 4.9 to enable it on more (all?) targets, and all the bugfixing and scalability work on LRA is needed for that anyway. Why can't this be done on the branch? We've made the mistake of rushing things into mainline too early a few times before, we should have learned by now. And adding more half transitions is not something we really want either. Bernd
[Patch, Committed] Fix declared inline after being called warning
Hi, I've committed to branch 4.7 as obvious attached patch that fixes a compiler warning 'declared inline after being called'. I ran into this warning when building the 4.7 branch with a gcc 4.3 compiler: ... var-tracking.c:558: warning: 'set_dv_changed' declared inline after being called var-tracking.c:558: warning: previous declaration of 'set_dv_changed' was here ... Other instances of this problem are: - http://gcc.gnu.org/ml/gcc-patches/2011-12/msg00256.html - http://gcc.gnu.org/ml/gcc-patches/2011-04/msg01426.html I've fixed this in the 4.7 branch. The problem is not present in the 4.6 branch, and with trunk the warning doesn't trigger, I suppose because we're using g++ now. Build on i686-pc-linux-gnu. Thanks, - Tom 2012-10-01 Tom de Vries t...@codesourcery.com * var-tracking.c (set_dv_changed): Add an 'inline' function specifier to the prototype. Index: gcc/var-tracking.c === --- gcc/var-tracking.c (revision 191792) +++ gcc/var-tracking.c (working copy) @@ -570,7 +570,7 @@ static void dump_vars (htab_t); static void dump_dataflow_set (dataflow_set *); static void dump_dataflow_sets (void); -static void set_dv_changed (decl_or_value, bool); +static inline void set_dv_changed (decl_or_value, bool); static void variable_was_changed (variable, dataflow_set *); static void **set_slot_part (dataflow_set *, rtx, void **, decl_or_value, HOST_WIDE_INT,
[patch][lra] Use XNEWVEC and friends instead of xmalloc/xrealloc, and add some timevars
Hello, This patch uses the libiberty new-like operators instead of using xmalloc/xrealloc. It also adds timevars for the main LRA phases, and it fixes a warning suggesting a space before a ';' in an only-looping for loop. Bootstrapped (lra-branch, of course) on x86_64-unknown-linux-gnu. OK? Ciao! Steven lra_XALLOC.diff Description: Binary data
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 12:10 PM, Steven Bosscher stevenb@gmail.com wrote: The LRA create live range time is mostly spent in merge_live_ranges walking lists. Hmm no, that's just gcc17's ancient debugger telling me lies. lra_live_range_in_p is not even used. /me upgrades to something newer than gdb 6.8... Ciao! Steven
Re: Constant-fold vector comparisons
On Sat, Sep 29, 2012 at 3:25 PM, Marc Glisse marc.gli...@inria.fr wrote: Hello, this patch does 2 things (I should have split it in 2, but the questions go together): 1) it handles constant folding of vector comparisons, 2) it fixes another place where vectors are not expected (I'll probably wait to have front-end support and testcases to do more of those, but there is something to discuss). I wasn't sure what integer_truep should test exactly. For integer: == 1 or != 0? For vectors: == -1 or 0? I chose the one that worked best for the forwprop case where I used it. It seems that before this patch, the middle-end didn't know how comparison results were encoded (a good reason for VEC_COND_EXPR to require a comparison as its first argument). I am using the OpenCL encoding that what matters is the high bit of each vector element. I am not quite sure what happens for targets (are there any?) that use a different encoding. When expanding vcond, they can do the comparison as they like. When expanding an isolated comparison, I expect they have to expand it as vcond(ab,-1,0). So it should be ok, but I could easily have missed something. Comments below 2012-10-01 Marc Glisse marc.gli...@inria.fr gcc/ * tree.c (integer_truep): New function. * tree.h (integer_truep): Declare. * tree-ssa-forwprop.c (forward_propagate_into_cond): Call it. Don't use boolean_type_node for vectors. * fold-const.c (fold_relational_const): Handle VECTOR_CST. gcc/testsuite/ * gcc.dg/tree-ssa/foldconst-6.c: New testcase. -- Marc Glisse Index: gcc/tree.h === --- gcc/tree.h (revision 191850) +++ gcc/tree.h (working copy) @@ -5272,20 +5272,25 @@ extern int integer_zerop (const_tree); /* integer_onep (tree x) is nonzero if X is an integer constant of value 1. */ extern int integer_onep (const_tree); /* integer_all_onesp (tree x) is nonzero if X is an integer constant all of whose significant bits are 1. */ extern int integer_all_onesp (const_tree); +/* integer_truep (tree x) is nonzero if X is an integer constant of value 1, + or a vector constant of value 0. */ + +extern bool integer_truep (const_tree); + /* integer_pow2p (tree x) is nonzero is X is an integer constant with exactly one bit 1. */ extern int integer_pow2p (const_tree); /* integer_nonzerop (tree x) is nonzero if X is an integer constant with a nonzero value. */ extern int integer_nonzerop (const_tree); Index: gcc/tree-ssa-forwprop.c === --- gcc/tree-ssa-forwprop.c (revision 191850) +++ gcc/tree-ssa-forwprop.c (working copy) @@ -564,46 +564,46 @@ forward_propagate_into_cond (gimple_stmt enum tree_code code; tree name = cond; gimple def_stmt = get_prop_source_stmt (name, true, NULL); if (!def_stmt || !can_propagate_from (def_stmt)) return 0; code = gimple_assign_rhs_code (def_stmt); if (TREE_CODE_CLASS (code) == tcc_comparison) tmp = fold_build2_loc (gimple_location (def_stmt), code, - boolean_type_node, + TREE_TYPE (cond), That's obvious. gimple_assign_rhs1 (def_stmt), gimple_assign_rhs2 (def_stmt)); else if ((code == BIT_NOT_EXPR TYPE_PRECISION (TREE_TYPE (cond)) == 1) || (code == BIT_XOR_EXPR - integer_onep (gimple_assign_rhs2 (def_stmt + integer_truep (gimple_assign_rhs2 (def_stmt See below. { tmp = gimple_assign_rhs1 (def_stmt); swap = true; } } if (tmp is_gimple_condexpr (tmp)) { if (dump_file tmp) { fprintf (dump_file, Replaced '); print_generic_expr (dump_file, cond, 0); fprintf (dump_file, ' with '); print_generic_expr (dump_file, tmp, 0); fprintf (dump_file, '\n); } - if (integer_onep (tmp)) + if (integer_truep (tmp)) gimple_assign_set_rhs_from_tree (gsi_p, gimple_assign_rhs2 (stmt)); else if (integer_zerop (tmp)) gimple_assign_set_rhs_from_tree (gsi_p, gimple_assign_rhs3 (stmt)); else { gimple_assign_set_rhs1 (stmt, unshare_expr (tmp)); if (swap) { tree t = gimple_assign_rhs2 (stmt); gimple_assign_set_rhs2 (stmt, gimple_assign_rhs3 (stmt)); Index: gcc/testsuite/gcc.dg/tree-ssa/foldconst-6.c === --- gcc/testsuite/gcc.dg/tree-ssa/foldconst-6.c (revision 0) +++ gcc/testsuite/gcc.dg/tree-ssa/foldconst-6.c (revision 0) @@ -0,0 +1,14 @@ +/* {
Re: [PATCH] Add option for dumping to stderr (issue6190057)
On Sun, Sep 30, 2012 at 11:36 PM, Sharad Singhai sing...@google.com wrote: Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. It caused: FAIL: gcc.dg/tree-ssa/gen-vect-11.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11a.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11a.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11b.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11b.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11c.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11c.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-2.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-2.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-25.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-25.c scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-32.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-32.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/O3-pr36119.f90 (test for excess errors) FAIL: gfortran.dg/vect/O3-pr39595.f (test for excess errors) FAIL: gfortran.dg/vect/Ofast-pr50414.f90 (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445.f (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445a.f (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 scan-tree-dump vect vectorized 1 loops FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/no-fre-no-copy-prop-O3-pr51704.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gfortran.dg/vect/pr19049.f90 -O scan-tree-dump-times vect complicated access pattern 1 FAIL: gfortran.dg/vect/pr19049.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32377.f90 -O scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/pr32377.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32380.f -O scan-tree-dump-times vect vectorized 6 loops 1 FAIL: gfortran.dg/vect/pr32380.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr33301.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr50178.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr50412.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058-2.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51285.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-1.f90 -O scan-tree-dump-times vect vectorized 3 loops 1 FAIL: gfortran.dg/vect/vect-1.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-2.f90 -O scan-tree-dump-times vect Alignment of access forced using peeling 3 FAIL: gfortran.dg/vect/vect-2.f90 -O scan-tree-dump-times vect Vectorizing an unaligned access 2 FAIL: gfortran.dg/vect/vect-2.f90 -O scan-tree-dump-times vect vectorized 3 loops 1 FAIL: gfortran.dg/vect/vect-2.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-3.f90 -O
[patch] experimenting with renumbering of pseudos after expand
Hello, For most code, expand creates a lot of pseudos that are cleaned up in subsequent passes, if they even live long enough to make it there. On average, for cc1 preprocessed source, the number of holes in regno_reg_rtx is about half the size of that array, or in other words: regno_reg_rtx is almost a sparse array. I've been experimenting with the attached patch to renumber pseudos just before initializing DF. For the already-notorious test case of PR54146 on x86_64, the patch reduces max_reg_num from 348404 to 180502 for the largest function. This reduces the memory foot print of the test case to less than 6GB, an almost 25% reduction, and it speeds up the DF_LR and DF_LIVE problems a bit (~15% but that's not much on the total compile time). I had hopes that it would help for LRA compile times also, but unfortunately the only significant change is this one: without patch: LRA hard reg assignment : 130.85 (11%) usr 0.20 ( 1%) sys 131.17 (11%) wall 0 kB ( 0%) ggc with patch: LRA hard reg assignment : 108.92 ( 9%) usr 0.07 ( 0%) sys 109.03 ( 9%) wall 0 kB ( 0%) ggc Anyway, putting the patch out there to show that I've done more than just complaining ;-) It also would suggest that the scalability challenge for LRA is not the number of pseudos but something else (number of insns, I guess). Ciao! Steven compact_regno_reg_rtx.diff Description: Binary data
[PATCH] Fix PR47799 - debug info for early-inlining with LTO
This tries to emit proper debug information for early-inlined functions from LTO LTRANS phase (thus, emit DW_TAG_inlined_subroutine and allow gdb to set breakpoints). We need to avoid confusing LTO and dwarf2out with the full abstract block tree, so this patch flattens the abstract block tree by always using the ultimate origin for BLOCK_ABSTRACT_ORIGIN on blocks which are inlined_function_outer_scope_p. Thus, it tries to output the minimal info dwarf2out.c needs to emit the desired debug information. As with LTO all abstract inline instances get generated late for early inlined functions I had to amend the hack for extern inline functions to always output dies for decls that come their way through dwarf2out_abstract_function. And further down not crash on a NULL DECL_INITIAL (when LTO decided to output the function body in another LTRANS unit or if it does not get output at all). Currently LTO-bootstrapping and testing on x86_64-unknown-linux-gnu. Jason, are the dwarf2out bits ok with you? I've sofar toyed with examples like int x, y; static inline int foo (int i) { y = i; return y; } static inline int bar (int i) { x = i; return foo (x); } int main () { int k = 0; int res = bar (k); return res; } and debug information with/without LTO is now reasonably the same and I can set breakpoints on the inlined instances. Thanks, Richard. 2012-10-01 Richard Guenther rguent...@suse.de PR lto/47788 * tree-streamer-out.c (write_ts_block_tree_pointers): For inlined functions outer scopes write the ultimate origin as BLOCK_ABSTRACT_ORIGIN and BLOCK_SOURCE_LOCATION. Do not stream the fragment chains. (lto_input_ts_block_tree_pointers): Likewise. * dwarf2out.c (gen_subprogram_die): Handle NULL DECL_INITIAL. (dwarf2out_decl): Always output DECL_ABSTRACT function decls. Index: gcc/tree-streamer-in.c === *** gcc/tree-streamer-in.c (revision 191824) --- gcc/tree-streamer-in.c (working copy) *** static void *** 789,810 lto_input_ts_block_tree_pointers (struct lto_input_block *ib, struct data_in *data_in, tree expr) { - /* Do not stream BLOCK_SOURCE_LOCATION. We cannot handle debug information - for early inlining so drop it on the floor instead of ICEing in - dwarf2out.c. */ BLOCK_VARS (expr) = streamer_read_chain (ib, data_in); - /* Do not stream BLOCK_NONLOCALIZED_VARS. We cannot handle debug information - for early inlining so drop it on the floor instead of ICEing in - dwarf2out.c. */ - BLOCK_SUPERCONTEXT (expr) = stream_read_tree (ib, data_in); ! /* Do not stream BLOCK_ABSTRACT_ORIGIN. We cannot handle debug information ! for early inlining so drop it on the floor instead of ICEing in dwarf2out.c. */ ! BLOCK_FRAGMENT_ORIGIN (expr) = stream_read_tree (ib, data_in); ! BLOCK_FRAGMENT_CHAIN (expr) = stream_read_tree (ib, data_in); /* We re-compute BLOCK_SUBBLOCKS of our parent here instead of streaming it. For non-BLOCK BLOCK_SUPERCONTEXTs we still --- 789,810 lto_input_ts_block_tree_pointers (struct lto_input_block *ib, struct data_in *data_in, tree expr) { BLOCK_VARS (expr) = streamer_read_chain (ib, data_in); BLOCK_SUPERCONTEXT (expr) = stream_read_tree (ib, data_in); ! /* Stream BLOCK_ABSTRACT_ORIGIN and BLOCK_SOURCE_LOCATION for ! the limited cases we can handle - those that represent inlined ! function scopes. For the rest them on the floor instead of ICEing in dwarf2out.c. */ ! BLOCK_ABSTRACT_ORIGIN (expr) = stream_read_tree (ib, data_in); ! BLOCK_SOURCE_LOCATION (expr) = lto_input_location (ib, data_in); ! /* Do not stream BLOCK_NONLOCALIZED_VARS. We cannot handle debug information ! for early inlined BLOCKs so drop it on the floor instead of ICEing in ! dwarf2out.c. */ ! ! /* BLOCK_FRAGMENT_ORIGIN and BLOCK_FRAGMENT_CHAIN is not live at LTO ! streaming time. */ /* We re-compute BLOCK_SUBBLOCKS of our parent here instead of streaming it. For non-BLOCK BLOCK_SUPERCONTEXTs we still Index: gcc/tree-streamer-out.c === *** gcc/tree-streamer-out.c (revision 191824) --- gcc/tree-streamer-out.c (working copy) *** write_ts_exp_tree_pointers (struct outpu *** 682,702 static void write_ts_block_tree_pointers (struct output_block *ob, tree expr, bool ref_p) { - /* Do not stream BLOCK_SOURCE_LOCATION. We cannot handle debug information - for early inlining so drop it on the floor instead of ICEing in - dwarf2out.c. */ streamer_write_chain (ob, BLOCK_VARS (expr), ref_p); /* Do not stream BLOCK_NONLOCALIZED_VARS. We cannot handle debug information ! for early inlining so drop it on the floor
Re: [RFC] Make vectorizer to skip loops with small iteration estimate
On Sun, 30 Sep 2012, Jan Hubicka wrote: Hi, the point of the following patch is to make vectorizer to not vectorize the following testcase with profile feedback: int a[1]; int i=5; int k=2; int val; __attribute__ ((noinline,noclone)) test() { int j; for(j=0;jk;j++) a[j]=val; } main() { while (i) { test (); i--; } } Here the compiler should work out that the second loop iterates 2 times at the average and thus it is not good candidate for vectorizing. In my first attempt I added the following: @@ -1474,6 +1478,18 @@ vect_analyze_loop_operations (loop_vec_i return false; } + if ((estimated_niter = estimated_stmt_executions_int (loop)) != -1 + (unsigned HOST_WIDE_INT) estimated_niter = th) +{ + if (vect_print_dump_info (REPORT_UNVECTORIZED_LOCATIONS)) +fprintf (vect_dump, not vectorized: estimated iteration count too small.); + if (vect_print_dump_info (REPORT_DETAILS)) +fprintf (vect_dump, not vectorized: estimated iteration count smaller than + user specified loop bound parameter or minimum + profitable iterations (whichever is more conservative).); + return false; +} + But to my surprise it does not help. There are two things: 1) the value of TH is bit low. In a way the cost model works is that it finds minimal niters where vectorized loop with all the setup costs is cheaper than the vector loop with all the setup costs. I.e. /* Calculate number of iterations required to make the vector version profitable, relative to the loop bodies only. The following condition must hold true: SIC * niters + SOC VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC(A) where SIC = scalar iteration cost, VIC = vector iteration cost, VOC = vector outside cost, VF = vectorization factor, PL_ITERS = prologue iterations, EP_ITERS= epilogue iterations SOC = scalar outside cost for run time cost model check. */ This value is used for both 1) decision if number of iterations is too low (max iterations is known) 2) decision on runtime whether we want to take the vectorized path or the scalar path. The vectoried loop looks like: k.1_10 = k; if (k.1_10 0) { pretmp_2 = val; niters.8_4 = (unsigned int) k.1_10; bnd.9_13 = niters.8_4 2; ratio_mult_vf.10_1 = bnd.9_13 2; _18 = niters.8_4 = 3; _19 = ratio_mult_vf.10_1 == 0; _20 = _19 | _18; if (_20 != 0) scalar loop else vector prologue } So the unvectorized cost is SIC * niters The vectorized path is SOC + VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC The scalar path of vectorizer loop is SIC * niters + SOC Note that 'th' is used for the runtime profitability check which is done at the time the setup cost has already been taken (yes, we probably should make it more conservative but then guard the whole set of loops by the check, not only the vectorized path). See PR53355 for the general issue. It makes sense to vectorize if SIC * niters SOC + VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC (B) That is in the optimal cse where we actually vectorize the overall speed of vectorized loop including the runtime check is better. It makes sense to take the vector loop if SIC * niters VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC (C) Because the scalar loop is taken. The attached patch implements the formula (C) and uses it to deterine the decision based on number of iterations estimate (that is usually provided by the feedback) As a reality check, I tried my testcase. 9: Cost model analysis: Vector inside of loop cost: 1 Vector prologue cost: 7 Vector epilogue cost: 2 Scalar iteration cost: 1 Scalar outside cost: 6 Vector outside cost: 9 prologue iterations: 0 epilogue iterations: 2 Calculated minimum iters for profitability: 4 9: Profitability threshold = 3 9: Profitability estimated iterations threshold = 20 This is overrated. The loop starts to be benefical at about 4 iterations in reality. I guess the values are kind of wrong. Vector inside of loop cost and Scalar iteration cost seems to ignore the fact that the loops do contain some control flow that should account at least one extra cycle. Vector prologue cost seems bit overrated for one pack operation. Of course this is very simple benchmark, in reality the vectorizatoin can be a lot more harmful by complicating more complex control flows. So I guess we have two options 1) go with the new formula and try to make cost model a bit more realistic. 2) stay with original formula that is quite close to
Re: [PATCH] Do not mark pseudo-copies decomposable during first lower-subreg pass
* gcc.dg/lower-subreg-1.c: Disable on arm-*-* targets. I just noticed that the triple is incomplete; we're supposed to use arm*-*-* instead of just arm-*-*. Checked in the the following fix as obvious. Bye, Ulrich 2012-10-01 Ulrich Weigand ulrich.weig...@linaro.org * gcc.dg/lower-subreg-1.c: Disable on arm*-*-* targets. Index: gcc/testsuite/gcc.dg/lower-subreg-1.c === *** gcc/testsuite/gcc.dg/lower-subreg-1.c (revision 191805) --- gcc/testsuite/gcc.dg/lower-subreg-1.c (working copy) *** *** 1,4 ! /* { dg-do compile { target { ! { mips64 || { arm-*-* ia64-*-* spu-*-* tilegx-*-* } } } } } */ /* { dg-options -O -fdump-rtl-subreg1 } */ /* { dg-skip-if { { i?86-*-* x86_64-*-* } x32 } { * } { } } */ /* { dg-require-effective-target ilp32 } */ --- 1,4 ! /* { dg-do compile { target { ! { mips64 || { arm*-*-* ia64-*-* spu-*-* tilegx-*-* } } } } } */ /* { dg-options -O -fdump-rtl-subreg1 } */ /* { dg-skip-if { { i?86-*-* x86_64-*-* } x32 } { * } { } } */ /* { dg-require-effective-target ilp32 } */ -- Dr. Ulrich Weigand GNU Toolchain for Linux on System z and Cell BE ulrich.weig...@de.ibm.com
Re: [PATCH] Fix instability of -fschedule-insn for x86
We also plan to test these changes along with LRA On Sun, Sep 30, 2012 at 4:33 PM, Uros Bizjak ubiz...@gmail.com wrote: On Tue, Sep 18, 2012 at 1:31 PM, Uros Bizjak ubiz...@gmail.com wrote: This patch aims to fix all stability issues related to using the first scheduler in gcc for x86 target (there several reported issues related to this problem). Main idea of this activity is mostly to provide user a possibility to safely turn on first scheduler for his codes. In some cases this could positively affect performance, especially for in-order Atom. Below is short description of proposed changes. 2012-09-18 Yuri Rumyantsev ysrum...@gmail.com * config/i386/i386.c (ix86_dep_by_shift_count_body) : Add check on reload_completed since it can be invoked before register allocation phase in 1st scheduler. (ia32_multipass_dfa_lookahead) : Do not use dfa_lookahead for 1st Scheduler to save compile time. (ix86_sched_reorder) : Do not perform ready list reordering for 1st Scheduler to save compile time. (insn_is_function_arg) : New function. Returns true if lhs of insn is HW function argument register. (add_parameter_dependencies) : New function. Add output dependencies for chain of function adjacent arguments if only there is a move to likely spilled HW registers. Return first argument if at least one dependence was added or NULL otherwise. (avoid_func_arg_motion) : New function. Add output or anti dependency from insn to first_arg to restrict code motion. (add_dependee_for_func_arg) : New function. Avoid cross block motion of function argument through adding dependency from the first non-jump insn in bb. (ix86_dependencies_evaluation_hook) : New function. Hook for schedule1: avoid motion of function arguments passed in passed in likely spilled HW registers. (ix86_adjust_priority) : New function. Hook for schedule1: set priority of moves from likely spilled HW registers to maximum to schedule them as soon as possible. (ix86_sched_init_global): Do not perform multipass scheduling for 1st Scheduler to save compile time. I would kindly ask scheduler expert to review the patch from the scheduler functionality POV. I have received opinion from Vladimir from off-line discussion, quoted below: --quote-- I think, it is ok. Switching off first cycle multipass scheduling is ok. It is mostly useful when the order of insns issued on the same cycle is important (mostly VLIW or quasy-VLIW processors). Other solutions are necessary to decrease spills and avoid reload crash (can not find a register in a class) when the 1st insn scheduling is on. I don't think it fully avoids possibility of the reload crashes but it takes into account most of cases resulting in the crashes and makes the crash possibility really negligible. Register pressure sensitive insn scheduling decreased the possibility. This patch will make it negligible. And LRA will solve all the rest cases of the crashes. I don't like a bit absence in freedom of moving argument insns with likely spilled hard-regs between each other as they are chained in the original order but it is debatable because it still decreases the possibility of spills. In overall, the patch is ok for me. --/quote-- Based on this opinion, the patch is OK for mainline, if there are no objections from other x86 maintainers in the next couple of days (48h). However, please watch for possible fallout from the patch, compile-time ICEs and performance problems. x86 and scheduler didn't play well together in the past, but your patch and (in the near future) LRA seems to fix all these problems. Thanks, Uros.
Re: RFC: LRA for x86/x86-64 [0/9]
On Sat, Sep 29, 2012 at 10:26 PM, Steven Bosscher stevenb@gmail.com wrote: LRA create live ranges : 175.30 (15%) usr 2.14 (13%) sys 177.44 (15%) wall2761 kB ( 0%) ggc I've tried to split this up a bit more: process_bb_lives ~50% create_start_finish_chains ~25% remove_some_program_points_and_update_live_ranges ~25% The latter two have a common structure with loops that look like this: for (i = FIRST_PSEUDO_REGISTER; i max_regno; i++) { for (r = lra_reg_info[i].live_ranges; r != NULL; r = r-next) Perhaps it's possible to do some of the work of compress_live_ranges during process_bb_lives, to create shorter live_ranges chains. Also, maybe doing something else than a linked list of live_ranges will help (unfortunately that's not a trivial change, it seems, because the lra_live_range_t type is used everywhere and there are no iterators or anything abstracted out like that -- just chain walks...). Still it does seem to me that a sorted VEC of lra_live_range objects probably would speed things up. Question is of course how much... :-) Ciao! Steven
[Ada] Membership tests work with extended overflow checks
This patch implements membership tests in which the operands can be out of range in extended overflow checkig modes. The following is a test program: 1. pragma Ada_2012; 2. with Text_IO; use Text_IO; 3. procedure Overflowm3 is 4.subtype Int10 is Integer range 1 .. 5; 5.subtype IntP is Integer with Predicate = Intp = 0; 6. 7.function r1 8. (a, b, c, d : Integer) return Boolean is 9.begin 10. return a + b + c + d in Integer'First .. Integer'Last 11. and then a + b + c + d in Integer 12. and then a + b + c + d in Intp 13. and then a + b + c + d not in Int10; 14.end; 15.function r2 16. (a, b, c, d : Integer) return Boolean is 17.begin 18. return a * b * c * d in Integer'First .. Integer'Last 19. and then a * b * c * d in Integer 20. and then a * b * c * d in Intp 21. and then a * b * c * d not in Int10; 22.end; 23. 24. begin 25.begin 26. Put_Line 27. (r1 returns 28.Boolean'Image 29. (r1 (Integer'Last, Integer'Last, 30. -Integer'Last, -Integer'Last))); 31.exception 32. when Constraint_Error = 33. Put_Line (r1 raises exception); 34.end; 35. 36.begin 37. Put_Line 38. (r2 returns 39.Boolean'Image 40. (r2 (Integer'Last, Integer'Last, 41. Integer'Last, 0))); 42.exception 43. when Constraint_Error = 44. Put_Line (r2 raises exception); 45.end; 46. end Overflowm3; In CHECKED mode (-gnato1) we get: r1 raises exception r2 raises exception since the first addition in r1 and the first multiplication in r2 result in values outside the bounds of Integer'Base. In MINIMIZED mode (-gnato2) we get: r1 returns TRUE r2 raises exception since we can compute the addition result in Long_Long_Integer, but the second multiplication yields a value outside this range, so that causes an overflow. In ELIMINATE mode (-gnato3) we get: r1 returns TRUE r2 returns TRUE Because now we use Bignum arithmetic for the intermediate multiplication results, and the final result is in range. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Robert Dewar de...@adacore.com * checks.adb (Apply_Arithmetic_Overflow_Minimized_Eliminated): Handle case of appearing in range in membership test. * exp_ch4.adb (Expand_Membership_Minimize_Eliminate_Overflow): New procedure (Expand_N_In): Use Expand_Membership_Minimize_Eliminate_Overflow. * rtsfind.ads: Add RE_Bignum_In_LLI_Range. * s-bignum.ads, s-bignum.adb (Bignum_In_LLI_Range): New function. * sinfo.ads, sinfo.adb (No_Minimize_Eliminate): New flag. Index: sinfo.adb === --- sinfo.adb (revision 191888) +++ sinfo.adb (working copy) @@ -2235,6 +2235,15 @@ return Flag13 (N); end No_Initialization; + function No_Minimize_Eliminate + (N : Node_Id) return Boolean is + begin + pragma Assert (False +or else NT (N).Nkind = N_In +or else NT (N).Nkind = N_Not_In); + return Flag17 (N); + end No_Minimize_Eliminate; + function No_Truncation (N : Node_Id) return Boolean is begin @@ -5288,6 +5297,15 @@ Set_Flag13 (N, Val); end Set_No_Initialization; + procedure Set_No_Minimize_Eliminate + (N : Node_Id; Val : Boolean := True) is + begin + pragma Assert (False +or else NT (N).Nkind = N_In +or else NT (N).Nkind = N_Not_In); + Set_Flag17 (N, Val); + end Set_No_Minimize_Eliminate; + procedure Set_No_Truncation (N : Node_Id; Val : Boolean := True) is begin Index: sinfo.ads === --- sinfo.ads (revision 191913) +++ sinfo.ads (working copy) @@ -1545,6 +1545,11 @@ --should not be taken into account (needed for in place initialization --with aggregates). + -- No_Minimize_Eliminate (Flag17-Sem) + --This flag is present in membership operator nodes (N_In/N_Not_In). + --It is used to indicate that processing for extended overflow checking + --modes is not required (this is used to prevent infinite recursion). + -- No_Truncation (Flag17-Sem) --Present in N_Unchecked_Type_Conversion node. This flag has an effect --only if the RM_Size of the source is greater than the RM_Size of the @@ -3675,6 +3680,7 @@ -- Left_Opnd (Node2) -- Right_Opnd (Node3) -- Alternatives (List4) (set to No_List if only one set alternative) + -- No_Minimize_Eliminate (Flag17) -- plus fields for expression -- N_Not_In @@ -3682,6 +3688,7 @@ --
[Ada] Exponentiation works with extended overflow checks
This patch implements extended overflow checking modes with the exonentiation operator. The following is a test program: 1. with Text_IO; use Text_IO; 2. procedure Overflowm4 is 3.function r1 (a, b : Integer) return Boolean is 4.begin 5. return a ** 2 - b ** 2 = Integer'Last; 6.end; 7.function r2 (a, b : Integer) return Boolean is 8.begin 9. return a ** 10 - b ** 10 in Integer; 10.end; 11. begin 12.begin 13. Put_Line 14. (r1 returns 15. Boolean'Image (r1 (Integer'Last, Integer'Last))); 16.exception 17. when Constraint_Error = 18. Put_Line (r1 raises exception); 19.end; 20. 21.begin 22. Put_Line 23. (r2 returns 24. Boolean'Image (r2 (Integer'Last, Integer'Last))); 25.exception 26. when Constraint_Error = 27. Put_Line (r2 raises exception); 28.end; 29. end Overflowm4; In CHECKED mode (-gnato1) we get: r1 raises exception r2 raises exception since the first exponentiation in both r1 and r2 result in values outside the bounds of Integer'Base. In MINIMIZED mode (-gnato2) we get: r1 returns TRUE r2 raises exception since we can compute the exponentiation results in r1 in Long_Long_Integer mode, but that's not true for r2. In ELIMINATE mode (-gnato3) we get: r1 returns TRUE r2 returns TRUE Because now we use Bignum arithmetic for the exponentiation operations in r2. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Robert Dewar de...@adacore.com * checks.adb (Minimize_Eliminate_Overflow_Checks): Changes for exponentiation. * exp_ch4.adb (Expand_N_Op_Expon): Changes for Minimize/Eliminate overflow checks. * s-bignum.adb (Compare): Fix bad precondition. Index: checks.adb === --- checks.adb (revision 191918) +++ checks.adb (working copy) @@ -6548,7 +6548,7 @@ when N_Op_Abs = Lo := Uint_0; - Hi := UI_Max (UI_Abs (Rlo), UI_Abs (Rhi)); + Hi := UI_Max (abs Rlo, abs Rhi); -- Addition @@ -6564,8 +6564,80 @@ -- Exponentiation when N_Op_Expon = - raise Program_Error; + -- Discard negative values for the exponent, since they will + -- simply result in an exception in any case. + + if Rhi 0 then + Rhi := Uint_0; + elsif Rlo 0 then + Rlo := Uint_0; + end if; + + -- Estimate number of bits in result before we go computing + -- giant useless bounds. Basically the number of bits in the + -- result is the number of bits in the base multiplied by the + -- value of the exponent. If this is big enough that the result + -- definitely won't fit in Long_Long_Integer, switch to bignum + -- mode immediately, and avoid computing giant bounds. + + -- The comparison here is approximate, but conservative, it + -- only clicks on cases that are sure to exceed the bounds. + + if Num_Bits (UI_Max (abs Llo, abs Lhi)) * Rhi + 1 100 then + Lo := No_Uint; + Hi := No_Uint; + + -- If right operand is zero then result is 1 + + elsif Rhi = 0 then + Lo := Uint_1; + Hi := Uint_1; + + else + -- High bound comes either from exponentiation of largest + -- positive value to largest exponent value, or from the + -- exponentiation of most negative value to an odd exponent. + + declare + Hi1, Hi2 : Uint; + + begin + if Lhi = 0 then +Hi1 := Lhi ** Rhi; + else +Hi1 := Uint_0; + end if; + + if Llo 0 then +if Rhi mod 2 = 0 then + Hi2 := Llo ** (Rhi - 1); +else + Hi2 := Llo ** Rhi; +end if; + else +Hi2 := Uint_0; + end if; + + Hi := UI_Max (Hi1, Hi2); + end; + + -- Result can only be negative if base can be negative + + if Llo 0 then + if UI_Mod (Rhi, 2) = 0 then +Lo := Llo ** (Rhi - 1); + else +Lo := Llo ** Rhi; + end if; + + -- Otherwise low bound is minimium ** minimum + + else
[Ada] Division/Rem/Mod work with extended overflow checks
This patch implements extended overflow checking modes with the division, rem, and mod operators. This completes the work on extended overflow checking. The following is a test program: 1. with Text_IO; use Text_IO; 2. procedure Overflowm5 is 3.function r1 (a, b, c : Integer) 4. return Integer is 5.begin 6. return a / b - c; 7.end; 8.function r2 (a, b, c : Long_Long_Integer) 9. return Long_Long_Integer is 10.begin 11. return a / b - c; 12.end; 13. begin 14.begin 15. Put_Line 16. (r1 returns 17.Integer'Image 18.(r1 (Integer'First, - 1, Integer'Last))); 19.exception 20. when Constraint_Error = 21. Put_Line (r1 raises exception); 22.end; 23. 24.begin 25. Put_Line 26. (r2 returns 27.Long_Long_Integer'Image 28.(r2 (Long_Long_Integer'First, -1, 29. Long_Long_Integer'Last))); 30.exception 31. when Constraint_Error = 32. Put_Line (r2 raises exception); 33.end; 34. end Overflowm5; In CHECKED mode (-gnato1) we get: r1 raises exception r2 raises exception since in both cases we are dividing the largest negative integer by minus one, which generates a value one greater than the largest positive value. In MINIMIZED mode (-gnato2) we get: r1 returns 1 r2 raises exception since we can now compute the division and subtraction for r1 in Long_Long_Integer mode, but that's not true for r2. In ELIMINATE mode (-gnato3) we get: r1 returns 1 r2 returns 1 Because now we use Bignum arithmetic for the division and subtraction operations in r2. Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Robert Dewar de...@adacore.com * checks.adb (Apply_Divide_Checks): New name for Apply_Divide_Check (Minimize_Eliminate_Overflow_Checks): Add code to handle division (and rem and mod) properly. (Apply_Division_Check): New procedure (Apply_Divide_Checks): Use Apply_Division_Check (Apply_Divide_Checks): Use Apply_Arithmetic_Overflow_Minimized_Eliminated. * checks.ads (Apply_Divide_Checks): New name for Apply_Divide_Check, also add clearer documentation for this routine and put in alfa order. * exp_ch4.adb (Apply_Divide_Checks): New name for Apply_Divide_Check. * s-bignum.adb (To_Bignum): Handle largest negative integer properly. * sem.adb (Analyze): Handle overflow suppression correctly (Analyze_List): Handle overflow suppression correctly * sem_res.adb (Analyze_And_Resolve): Handle overflow suppression correctly. Index: checks.adb === --- checks.adb (revision 191919) +++ checks.adb (working copy) @@ -193,14 +193,6 @@ -- Local Subprograms -- --- - procedure Apply_Float_Conversion_Check - (Ck_Node: Node_Id; - Target_Typ : Entity_Id); - -- The checks on a conversion from a floating-point type to an integer - -- type are delicate. They have to be performed before conversion, they - -- have to raise an exception when the operand is a NaN, and rounding must - -- be taken into account to determine the safe bounds of the operand. - procedure Apply_Arithmetic_Overflow_Normal (N : Node_Id); -- Used to apply arithmetic overflow checks for all cases except operators -- on signed arithmetic types in Minimized/Eliminate case (for which we @@ -211,6 +203,24 @@ -- checking mode is Minimized or Eliminated (and the Do_Overflow_Check flag -- is known to be set) and we have an signed integer arithmetic op. + procedure Apply_Division_Check + (N : Node_Id; + Rlo : Uint; + Rhi : Uint; + ROK : Boolean); + -- N is an N_Op_Div, N_Op_Rem, or N_Op_Mod node. This routine applies + -- division checks as required if the Do_Division_Check flag is set. + -- Rlo and Rhi give the possible range of the right operand, these values + -- can be referenced and trusted only if ROK is set True. + + procedure Apply_Float_Conversion_Check + (Ck_Node: Node_Id; + Target_Typ : Entity_Id); + -- The checks on a conversion from a floating-point type to an integer + -- type are delicate. They have to be performed before conversion, they + -- have to raise an exception when the operand is a NaN, and rounding must + -- be taken into account to determine the safe bounds of the operand. + procedure Apply_Selected_Length_Checks (Ck_Node: Node_Id; Target_Typ : Entity_Id; @@ -1641,52 +1651,69 @@ Reason= CE_Discriminant_Check_Failed)); end Apply_Discriminant_Check; - - -- Apply_Divide_Check -- - +
[Ada] Static predicate checks on type conversions
In Ada 2012, if a subtype has predicates, a predicate check must be applied to the expression in a type conversion to the subtype. Furthermore, if the expression is a scalar static constant, the predicate must be evluated at compile-time, and the program must be rejected if the predicate is false. Compiling gcc -c -gnat12 -gnata main.adb must yield: main.adb:6:16: static expression fails static predicate check on T --- procedure Main is subtype T is Integer with Static_Predicate = T = 10; V : T := 10; begin V := 1000 / T (9); end Main; Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Ed Schonberg schonb...@adacore.com * checks.adb (Apply_Predicate_Check): If the predicate is a static one and the operand is static, evaluate the predicate at compile time. * sem_eval.ads, sem_eval.adb (Eval_Static_Predicate_Check): new procedure, to evaluate a static predicate check whenever possible. * sem_res.adb (Resolve_Type_Conversion): Apply predicate check on the conversion if the target type has predicates. Index: checks.adb === --- checks.adb (revision 191920) +++ checks.adb (working copy) @@ -2337,6 +2337,23 @@ (Sloc (N), Reason = SE_Infinite_Recursion)); else + +-- If the predicate is a static predicate and the operand is +-- static, the predicate must be evaluated statically. If the +-- evaluation fails this is a static constraint error. + +if Is_OK_Static_Expression (N) then + if Present (Static_Predicate (Typ)) then + if Eval_Static_Predicate_Check (N, Typ) then + return; + else + Error_Msg_NE + (static expression fails static predicate check on, + N, Typ); + end if; + end if; +end if; + Insert_Action (N, Make_Predicate_Check (Typ, Duplicate_Subexpr (N))); end if; Index: sem_res.adb === --- sem_res.adb (revision 191920) +++ sem_res.adb (working copy) @@ -9713,6 +9713,22 @@ end if; end; end if; + + -- Ada 2012: if target type has predicates, the result requires a + -- predicate check. If the context is a call to another predicate + -- check we must prevent infinite recursion. + + if Has_Predicates (Target_Typ) then + if Nkind (Parent (N)) = N_Function_Call + and then Present (Name (Parent (N))) + and then Has_Predicates (Entity (Name (Parent (N + then +null; + + else +Apply_Predicate_Check (N, Target_Typ); + end if; + end if; end Resolve_Type_Conversion; -- Index: sem_eval.adb === --- sem_eval.adb(revision 191895) +++ sem_eval.adb(working copy) @@ -3249,6 +3249,37 @@ end if; end Eval_Slice; + - + -- Eval_Static_Predicate_Check -- + - + + function Eval_Static_Predicate_Check + (N : Node_Id; + Typ : Entity_Id) return Boolean + is + Loc : constant Source_Ptr := Sloc (N); + Pred : constant List_Id := Static_Predicate (Typ); + Test : Node_Id; + begin + if No (Pred) then + return True; + end if; + + -- The static predicate is a list of alternatives in the proper format + -- for an Ada 2012 membership test. If the argument is a literal, the + -- membership test can be evaluated statically. The caller transforms + -- a result of False into a static contraint error. + + Test := Make_In (Loc, + Left_Opnd= New_Copy_Tree (N), + Right_Opnd = Empty, + Alternatives = Pred); + Analyze_And_Resolve (Test, Standard_Boolean); + + return Nkind (Test) = N_Identifier +and then Entity (Test) = Standard_True; + end Eval_Static_Predicate_Check; + - -- Eval_String_Literal -- - Index: sem_eval.ads === --- sem_eval.ads(revision 191888) +++ sem_eval.ads(working copy) @@ -317,6 +317,11 @@ procedure Eval_Unary_Op (N : Node_Id); procedure Eval_Unchecked_Conversion (N : Node_Id); + function Eval_Static_Predicate_Check + (N : Node_Id; + Typ : Entity_Id) return Boolean; + -- Evaluate a static predicate check applied to a scalar literal. + procedure Fold_Str (N : Node_Id; Val : String_Id; Static : Boolean); -- Rewrite N with a new N_String_Literal node as the result of the compile -- time
[Ada] Complain when actual Symbol is present in any dimension output call
This patch prevents the user to provide parameter Symbol (reserved for compiler use only) in any dimension output call. -- Source -- with System.Dim.Mks;use System.Dim.Mks; with System.Dim.Mks_IO; use System.Dim.Mks_IO; procedure Main is begin Put (8.0**(1 / 3) * m , 1, 2, 0, error); end Main; - -- Compilation Execution -- - $ gcc -c -gnat12 main.adb main.adb:6:37: Symbol parameter should not be provided main.adb:6:37: reserved for compiler use only Tested on x86_64-pc-linux-gnu, committed on trunk 2012-10-01 Vincent Pucci pu...@adacore.com * sem_dim.adb (Has_Symbols): Complain if parameter Symbol has been provided by the user in the dimension output call. Index: sem_dim.adb === --- sem_dim.adb (revision 191911) +++ sem_dim.adb (working copy) @@ -2703,7 +2703,8 @@ - function Has_Symbols return Boolean is - Actual : Node_Id; + Actual : Node_Id; + Actual_Str : Node_Id; begin Actual := First (Actuals); @@ -2711,16 +2712,49 @@ -- Look for a symbols parameter association in the list of actuals while Present (Actual) loop -if Nkind (Actual) = N_Parameter_Association +-- Positional parameter association case when the actual is a +-- string literal. + +if Nkind (Actual) = N_String_Literal then + Actual_Str := Actual; + +-- Named parameter association case when the selector name is +-- Symbol. + +elsif Nkind (Actual) = N_Parameter_Association and then Chars (Selector_Name (Actual)) = Name_Symbol then + Actual_Str := Explicit_Actual_Parameter (Actual); + +-- Ignore all other cases + +else + Actual_Str := Empty; +end if; + +if Present (Actual_Str) then -- Return True if the actual comes from source or if the string -- of symbols doesn't have the default value (i.e. it is ). - return Comes_From_Source (Actual) - or else - String_Length - (Strval (Explicit_Actual_Parameter (Actual))) /= 0; + if Comes_From_Source (Actual) + or else String_Length (Strval (Actual_Str)) /= 0 + then + -- Complain only if the actual comes from source or if it + -- hasn't been fully analyzed yet. + + if Comes_From_Source (Actual) +or else not Analyzed (Actual) + then + Error_Msg_N (Symbol parameter should not be provided, + Actual); + Error_Msg_N (\reserved for compiler use only, Actual); + end if; + + return True; + + else + return False; + end if; end if; Next (Actual);
[PATCH] Fix -frounding-math builtins
I noticed that we attach the no-vops attribute to -frounding-math math functions. That's bogus as can be seen from the testcase int fesetround(int); double asinh(double x); double foo (double x, int b) { double y = 0.0, z; if (b) y = asinh (x); fesetround (0x400 /*FE_DOWNWARD*/); z = asinh (x); return y + z; } where PRE rightfully so removes a seeming partial redundancy by inserting a asinh call into the else block. That's because it exactly does _not_ get to see the rounding mode clobbering fesetround call as asinh does not have a virtual operand. Fixed as follows. Bootstrapped and tested on x86_64-unknown-linux-gnu. Richard. 2012-10-01 Richard Guenther rguent...@suse.de * builtins.def (ATTR_MATHFN_FPROUNDING): Do not use no-vops with -frounding-math. * builtin-attrs.def (ATTR_PURE_NOTHROW_NOVOPS_LIST): Remove. (ATTR_PURE_NOTHROW_NOVOPS_LEAF_LIST): Likewise. Index: gcc/builtins.def === *** gcc/builtins.def(revision 191917) --- gcc/builtins.def(working copy) *** along with GCC; see the file COPYING3. *** 163,169 memory. */ #undef ATTR_MATHFN_FPROUNDING #define ATTR_MATHFN_FPROUNDING (flag_rounding_math ? \ ! ATTR_PURE_NOTHROW_NOVOPS_LEAF_LIST : ATTR_CONST_NOTHROW_LEAF_LIST) /* Define an attribute list for math functions that are normally impure because some of them may write into global memory for --- 163,169 memory. */ #undef ATTR_MATHFN_FPROUNDING #define ATTR_MATHFN_FPROUNDING (flag_rounding_math ? \ ! ATTR_PURE_NOTHROW_LEAF_LIST : ATTR_CONST_NOTHROW_LEAF_LIST) /* Define an attribute list for math functions that are normally impure because some of them may write into global memory for Index: gcc/builtin-attrs.def === *** gcc/builtin-attrs.def (revision 191917) --- gcc/builtin-attrs.def (working copy) *** DEF_ATTR_TREE_LIST (ATTR_PURE_NOTHROW_LI *** 127,136 ATTR_NULL, ATTR_NOTHROW_LIST) DEF_ATTR_TREE_LIST (ATTR_PURE_NOTHROW_LEAF_LIST, ATTR_PURE, \ ATTR_NULL, ATTR_NOTHROW_LEAF_LIST) - DEF_ATTR_TREE_LIST (ATTR_PURE_NOTHROW_NOVOPS_LIST, ATTR_NOVOPS, \ - ATTR_NULL, ATTR_PURE_NOTHROW_LIST) - DEF_ATTR_TREE_LIST (ATTR_PURE_NOTHROW_NOVOPS_LEAF_LIST, ATTR_NOVOPS,\ - ATTR_NULL, ATTR_PURE_NOTHROW_LEAF_LIST) DEF_ATTR_TREE_LIST (ATTR_NORETURN_NOTHROW_LIST, ATTR_NORETURN,\ ATTR_NULL, ATTR_NOTHROW_LIST) DEF_ATTR_TREE_LIST (ATTR_NORETURN_NOTHROW_LEAF_LIST, ATTR_NORETURN,\ --- 127,132
Re: [PATCH] Add option for dumping to stderr (issue6190057)
I am sorry, I didn't enable all the languages. Will fix the fortran test breakage shortly. Thanks, Sharad Sharad On Mon, Oct 1, 2012 at 4:50 AM, H.J. Lu hjl.to...@gmail.com wrote: On Sun, Sep 30, 2012 at 11:36 PM, Sharad Singhai sing...@google.com wrote: Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. It caused: FAIL: gcc.dg/tree-ssa/gen-vect-11.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11a.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11a.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11b.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11b.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11c.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11c.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-2.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-2.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-25.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-25.c scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-32.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-32.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/O3-pr36119.f90 (test for excess errors) FAIL: gfortran.dg/vect/O3-pr39595.f (test for excess errors) FAIL: gfortran.dg/vect/Ofast-pr50414.f90 (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445.f (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445a.f (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 scan-tree-dump vect vectorized 1 loops FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/no-fre-no-copy-prop-O3-pr51704.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gfortran.dg/vect/pr19049.f90 -O scan-tree-dump-times vect complicated access pattern 1 FAIL: gfortran.dg/vect/pr19049.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32377.f90 -O scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/pr32377.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32380.f -O scan-tree-dump-times vect vectorized 6 loops 1 FAIL: gfortran.dg/vect/pr32380.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr33301.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr50178.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr50412.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058-2.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51285.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-1.f90 -O scan-tree-dump-times vect vectorized 3 loops 1 FAIL: gfortran.dg/vect/vect-1.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-2.f90 -O scan-tree-dump-times vect Alignment of access forced using peeling 3 FAIL: gfortran.dg/vect/vect-2.f90 -O
Re: [PATCH] Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 6:49 AM, Sharad Singhai sing...@google.com wrote: I am sorry, I didn't enable all the languages. Will fix the fortran test breakage shortly. It is not just Fortran. There are some failures in C testcases. Thanks, Sharad Sharad On Mon, Oct 1, 2012 at 4:50 AM, H.J. Lu hjl.to...@gmail.com wrote: On Sun, Sep 30, 2012 at 11:36 PM, Sharad Singhai sing...@google.com wrote: Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. It caused: FAIL: gcc.dg/tree-ssa/gen-vect-11.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11a.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11a.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11b.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11b.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11c.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11c.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-2.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-2.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-25.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-25.c scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-32.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-32.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/O3-pr36119.f90 (test for excess errors) FAIL: gfortran.dg/vect/O3-pr39595.f (test for excess errors) FAIL: gfortran.dg/vect/Ofast-pr50414.f90 (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445.f (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445a.f (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 scan-tree-dump vect vectorized 1 loops FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/no-fre-no-copy-prop-O3-pr51704.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gfortran.dg/vect/pr19049.f90 -O scan-tree-dump-times vect complicated access pattern 1 FAIL: gfortran.dg/vect/pr19049.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32377.f90 -O scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/pr32377.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr32380.f -O scan-tree-dump-times vect vectorized 6 loops 1 FAIL: gfortran.dg/vect/pr32380.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr33301.f -O (test for excess errors) FAIL: gfortran.dg/vect/pr50178.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr50412.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058-2.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51058.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/pr51285.f90 -O (test for excess errors) FAIL: gfortran.dg/vect/vect-1.f90 -O scan-tree-dump-times vect vectorized 3 loops 1 FAIL: gfortran.dg/vect/vect-1.f90 -O (test for excess errors) FAIL:
Re: [PATCH] Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 3:55 PM, Sharad Singhai sing...@google.com wrote: On Mon, Oct 1, 2012 at 6:52 AM, H.J. Lu hjl.to...@gmail.com wrote: On Mon, Oct 1, 2012 at 6:49 AM, Sharad Singhai sing...@google.com wrote: I am sorry, I didn't enable all the languages. Will fix the fortran test breakage shortly. It is not just Fortran. There are some failures in C testcases. I checked and those files looked like generator files for Fortran tests and thus were not exercised in my configuration. I am really sorry about that. I am fixing it. As I said, you should not enable/disable anything special but configure with all default languages enabled (no --enable-languages) and do toplevel make -k check, preferably also excercising multilibs with RUNTESTFLAGS=--target_board=unix/\{,-m32\} Richard. UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11a.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11b.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11c.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-2.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-25.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-26.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-28.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-32.c Thanks, Sharad Thanks, Sharad Sharad On Mon, Oct 1, 2012 at 4:50 AM, H.J. Lu hjl.to...@gmail.com wrote: On Sun, Sep 30, 2012 at 11:36 PM, Sharad Singhai sing...@google.com wrote: Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. It caused: FAIL: gcc.dg/tree-ssa/gen-vect-11.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11a.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11a.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11b.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11b.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11c.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11c.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-2.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-2.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-25.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-25.c scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-32.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-32.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/O3-pr36119.f90 (test for excess errors) FAIL: gfortran.dg/vect/O3-pr39595.f (test for excess errors) FAIL: gfortran.dg/vect/Ofast-pr50414.f90 (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445.f (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445a.f (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 scan-tree-dump vect vectorized 1 loops FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/no-fre-no-copy-prop-O3-pr51704.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32457.f90 scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gfortran.dg/vect/pr19049.f90
Re: [PATCH] Add option for dumping to stderr (issue6190057)
Okay, I am retesting without any special configs and with multilibs as you suggested. Thanks, Sharad On Mon, Oct 1, 2012 at 7:00 AM, Richard Guenther richard.guent...@gmail.com wrote: On Mon, Oct 1, 2012 at 3:55 PM, Sharad Singhai sing...@google.com wrote: On Mon, Oct 1, 2012 at 6:52 AM, H.J. Lu hjl.to...@gmail.com wrote: On Mon, Oct 1, 2012 at 6:49 AM, Sharad Singhai sing...@google.com wrote: I am sorry, I didn't enable all the languages. Will fix the fortran test breakage shortly. It is not just Fortran. There are some failures in C testcases. I checked and those files looked like generator files for Fortran tests and thus were not exercised in my configuration. I am really sorry about that. I am fixing it. As I said, you should not enable/disable anything special but configure with all default languages enabled (no --enable-languages) and do toplevel make -k check, preferably also excercising multilibs with RUNTESTFLAGS=--target_board=unix/\{,-m32\} Richard. UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11a.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11b.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-11c.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-2.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-25.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-26.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-28.c UNSUPPORTED: gcc.dg/tree-ssa/gen-vect-32.c Thanks, Sharad Thanks, Sharad Sharad On Mon, Oct 1, 2012 at 4:50 AM, H.J. Lu hjl.to...@gmail.com wrote: On Sun, Sep 30, 2012 at 11:36 PM, Sharad Singhai sing...@google.com wrote: Resend to gcc-patches I have addressed the comments by fixing all the minor issues, bootstrapped and tested on x86_64. I did the recommended reshuffling by moving non-tree code from tree-dump.c into a new file dumpfile.c. I committed two successive revisions r191883 Main patch with the dump infrastructure changes. However, I accidentally left out a new file, dumpfile.c. r191884 Added dumpfile.c, and did the renaming of dump_* functions from gimple_pretty_print.[ch]. As things stand right now, r191883 is broken because of the missing file 'dumpfile.c', which the very next commit fixes. Anyone who got broken revision r191883, please svn update. I am really very sorry about that. I have a couple more minor patches which deal with renaming; I plan to address those later. It caused: FAIL: gcc.dg/tree-ssa/gen-vect-11.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11a.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11a.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11b.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11b.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-11c.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-11c.c scan-tree-dump-times vect vectorized 0 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-2.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-2.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-25.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-25.c scan-tree-dump-times vect vectorized 2 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-26.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect Alignment of access forced using peeling 1 FAIL: gcc.dg/tree-ssa/gen-vect-28.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gcc.dg/tree-ssa/gen-vect-32.c (test for excess errors) FAIL: gcc.dg/tree-ssa/gen-vect-32.c scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/O3-pr36119.f90 (test for excess errors) FAIL: gfortran.dg/vect/O3-pr39595.f (test for excess errors) FAIL: gfortran.dg/vect/Ofast-pr50414.f90 (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445.f (test for excess errors) FAIL: gfortran.dg/vect/cost-model-pr34445a.f (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-pr38968.f90 scan-tree-dump vect vectorized 1 loops FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-real8-pr40801.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 (test for excess errors) FAIL: gfortran.dg/vect/fast-math-vect-8.f90 scan-tree-dump-times vect vectorized 1 loops 1 FAIL: gfortran.dg/vect/no-fre-no-copy-prop-O3-pr51704.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 (test for excess errors) FAIL: gfortran.dg/vect/no-vfa-pr32377.f90 scan-tree-dump-times vect vectorized 2 loops 1
Second ping: Re: Add a configure option to disable system header canonicalizations (issue6495088)
Ping, again. On 21 September 2012 12:45, Simon Baldwin sim...@google.com wrote: Ping. http://gcc.gnu.org/ml/gcc-patches/2012-09/msg00459.html Full text of previous message and context at URL above. No comments or code changes since. Patch description left below for convenience. Add flags to disable system header canonicalizations. Libcpp may canonicalize system header paths with lrealpath() for diagnostics, dependency output, and similar. If gcc is held in a symlink farm the canonicalized paths may be meaningless to users, and will also conflict with build frameworks that (for example) disallow absolute paths to header files. This change adds -f[no-]canonical-system-headers to the gcc command line, and a configure option --[en/dis]able-canonical-system-headers to set default behaviour, allowing the user to select whether or not to implement r186991. Default is enabled. See also PR c++/52974. Tested for regressions with bootstrap builds of C and C++, both with and without configure --disable-canonical-system-headers. -- Google UK Limited | Registered Office: Belgrave House, 76 Buckingham Palace Road, London SW1W 9TQ | Registered in England Number: 3977902 -- Google UK Limited | Registered Office: Belgrave House, 76 Buckingham Palace Road, London SW1W 9TQ | Registered in England Number: 3977902
[Patch] Fix PR53397
Hi, The below patch fixes the FFT/Scimark regression caused by useless prefetch generation. This fix tries to make prefetch less aggressive by prefetching arrays in the inner loop, when the step is invariant in the entire loop nest. GCC currently tries to prefetch invariant steps when they are in the inner loop. But does not check if the step is variant in outer loops. In the scimark FFT case, the trip count of the inner loop varies by a non constant step, which is invariant in the inner loop. But the step variable is varying in outer loop. This makes inner loop trip count small (at run time varies sometimes as small as 1 iteration) Prefetching ahead x iteration when the inner loop trip count is smaller than x leads to useless prefetches. Flag used: -O3 -march=amdfam10 Before ** ** ** SciMark2 Numeric Benchmark, see http://math.nist.gov/scimark ** ** for details. (Results can be submitted to p...@nist.gov) ** ** ** Using 2.00 seconds min time per kenel. Composite Score: 550.50 FFT Mflops:38.66(N=1024) SOR Mflops: 617.61(100 x 100) MonteCarlo: Mflops: 173.74 Sparse matmult Mflops: 675.63(N=1000, nz=5000) LU Mflops: 1246.88(M=100, N=100) After ** ** ** SciMark2 Numeric Benchmark, see http://math.nist.gov/scimark ** ** for details. (Results can be submitted to p...@nist.gov) ** ** ** Using 2.00 seconds min time per kenel. Composite Score: 639.20 FFT Mflops: 479.19(N=1024) SOR Mflops: 617.61(100 x 100) MonteCarlo: Mflops: 173.18 Sparse matmult Mflops: 679.13(N=1000, nz=5000) LU Mflops: 1246.88(M=100, N=100) GCC regression make check -k passes with x86_64-unknown-linux-gnu New tests that PASS: gcc.dg/pr53397-1.c scan-assembler prefetcht0 gcc.dg/pr53397-1.c scan-tree-dump aprefetch Issued prefetch gcc.dg/pr53397-1.c (test for excess errors) gcc.dg/pr53397-2.c scan-tree-dump aprefetch loop variant step gcc.dg/pr53397-2.c scan-tree-dump aprefetch Not prefetching gcc.dg/pr53397-2.c (test for excess errors) Checked CPU2006 and polyhedron on latest AMD processor, no regressions noted. Ok to commit in trunk? regards, Venkat gcc/ChangeLog +2012-10-01 Venkataramanan Kumar venkataramanan.ku...@amd.com + + * tree-ssa-loop-prefetch.c (gather_memory_references_ref):$ + Perform non constant step prefetching in inner loop, only $ + when it is invariant in the entire loop nest. $ + * testsuite/gcc.dg/pr53397-1.c: New test case $ + Checks we are prefecthing for loop invariant steps$ + * testsuite/gcc.dg/pr53397-2.c: New test case$ + Checks we are not prefecthing for loop variant steps + Index: gcc/testsuite/gcc.dg/pr53397-1.c === --- gcc/testsuite/gcc.dg/pr53397-1.c(revision 0) +++ gcc/testsuite/gcc.dg/pr53397-1.c(revision 0) @@ -0,0 +1,28 @@ +/* Prefetching when the step is loop invariant. */ + +/* { dg-do compile } */ +/* { dg-options -O3 -fprefetch-loop-arrays -fdump-tree-aprefetch-details --param min-insn-to-prefetch-ratio=3 --param simultaneous-prefetches=10 -fdump-tree-aprefetch-details } */ + + +double data[16384]; +void prefetch_when_non_constant_step_is_invariant(int step, int n) +{ + int a; + int b; + for (a = 1; a step; a++) { +for (b = 0; b n; b += 2 * step) { + + int i = 2*(b + a); + int j = 2*(b + a + step); + + + data[j] = data[i]; + data[j+1] = data[i+1]; +} + } +} + +/* { dg-final { scan-tree-dump Issued prefetch aprefetch } } */ +/* { dg-final { scan-assembler prefetcht0 } } */ + +/* { dg-final { cleanup-tree-dump aprefetch } } */ Index: gcc/testsuite/gcc.dg/pr53397-2.c === --- gcc/testsuite/gcc.dg/pr53397-2.c(revision 0) +++ gcc/testsuite/gcc.dg/pr53397-2.c(revision 0) @@ -0,0 +1,29 @@ +/* Not prefetching when the step is loop variant. */ + +/* { dg-do compile } */ +/* { dg-options -O3 -fprefetch-loop-arrays -fdump-tree-aprefetch-details --param min-insn-to-prefetch-ratio=3 --param simultaneous-prefetches=10 -fdump-tree-aprefetch-details } */ + + +double data[16384]; +void donot_prefetch_when_non_constant_step_is_variant(int step, int n) +{ + int a; + int b; + for (a = 1; a step; a++,step*=2) { +for (b = 0; b n; b += 2 * step) { + + int i = 2*(b + a); + int j = 2*(b + a + step); + + + data[j] = data[i]; + data[j+1] = data[i+1]; +} + } +} + +/* { dg-final { scan-tree-dump Not prefetching aprefetch } } */ +/* { dg-final { scan-tree-dump loop variant step aprefetch } }
Re: [PATCH, i386]: Implement atomic_fetch_sub
On 08/30/2012 05:33 PM, Richard Henderson wrote: On 08/23/2012 08:59 AM, Andrew MacLeod wrote: 2012-08-23 Andrew MacLeod amacl...@redhat.com gcc PR target/54087 * optabs.c (expand_atomic_fetch_op_no_fallback): New. Factored code from expand_atomic_fetch_op. (expand_atomic_fetch_op): iTry atomic_{add|sub} operations in terms of the other one if direct opcode fails. testsuite * gcc.dg/pr54087.c: New testcase for atomic_sub - atomic_add when atomic_sub fails. Ok. Oops, approved but never checked in.Just did so after a new bootstrap/test cycle with no issues. Andrew
Re: [patch, mips] Patch for new mips triplet - mips-mti-elf
On Sun, 2012-09-30 at 19:53 +0100, Richard Sandiford wrote: Sorry for only noticing now, but this produced: ERROR: gcc.target/mips/pr37362.c -O0 : syntax error in target selector target ! mips*-sde-elf mips*-mti-elf for dg-do 2 compile { target { ! mips*-sde-elf mips*-mti-elf } } ... We need another set of braces. Tested on mipsisa64-elf and applied. Richard Thanks for fixing this, I am not sure why I didn't notice it in my testing. Steve Ellcey sell...@mips.com
Re: Tweak IRA checks for singleton register classes
On 12-09-30 2:21 PM, Richard Sandiford wrote: IRA has code to check whether there is only a single acceptable register for a given operand. This code uses conditions like: ira_class_hard_regs_num[cl] != 0 (ira_class_hard_regs_num[cl] = ira_reg_class_max_nregs[cl][mode]) i.e. the number of registers needed to store the mode is = the number of alloctable registers in the class. Then: ira_class_hard_regs[cl][0] gives the register in question. MIPS has a slightly strange situation in which HI can only be allocated alongside LO; it can't be allocated independently. At the moment, HI and LO have their own register classes (MD0_REG and MD1_REG, with the mapping depending on endianness) and MD_REGS is used when both HI and LO are required. There is also ACC_REGS, which is equivalent to MD_REGS when the DSP ASE is not being used. MD_REGS and ACC_REGS are already mapped to constraints. Having MD0_REG and MD1_REG leads to some confusing costs and makes HI and LO irregular WRT the DSP ASE accumulator registers. I've been experimenting with patches to remove these classes and just have MD_REGS. I wanted to get to a situtation where this change has no effect on cc1 .ii files for -mno-dsp; the patch below is one of those needed to get to that stage. MD_REGS has only one SImode register. As describe above, the same goes for ACC_REGS unless the DSP ASE is being used. However, both classes fail the check above because HI (which doesn't accept SImode) is also allocatable. That is, the classes have two allocatable registers, but only one of them can be used for SImode. The patch below adds a new array for tracking which class/mode combinations specify a single register, and for recording which register that is. The net effect will be the same on almost all targets. I deliberately didn't change: for (p2 = reg_class_subclasses[cl2][0]; *p2 != LIM_REG_CLASSES; p2++) if (ira_class_hard_regs_num[*p2] 0 (ira_reg_class_max_nregs[*p2][mode] = ira_class_hard_regs_num[*p2])) cost = MAX (cost, ira_register_move_cost[mode][cl1][*p2]); for (p1 = reg_class_subclasses[cl1][0]; *p1 != LIM_REG_CLASSES; p1++) if (ira_class_hard_regs_num[*p1] 0 (ira_reg_class_max_nregs[*p1][mode] = ira_class_hard_regs_num[*p1])) cost = MAX (cost, ira_register_move_cost[mode][*p1][cl2]); from ira_init_register_move_cost because that had more effect than I was expecting and wasn't needed for the MIPS patch. It could be done as a follow-up if I ever find time... I checked that this produced no difference in assembly output for a set of x86_64 gcc .ii files (tested with -O2 -march=native on gcc20). Also tested on x86_64-linux-gnu (including -m32) and mipsisa64-elf. OK to install? It is OK for me, Richard. Although it was hard to me understand correctness of the following changes. Thanks for the patch. Index: gcc/ira-lives.c === --- gcc/ira-lives.c 2012-09-30 12:56:14.344185269 +0100 +++ gcc/ira-lives.c 2012-09-30 17:45:14.962463976 +0100 @@ -849,9 +849,10 @@ single_reg_class (const char *constraint next_cl = (c == 'r' ? GENERAL_REGS : REG_CLASS_FROM_CONSTRAINT (c, constraints)); - if ((cl != NO_REGS next_cl != cl) - || (ira_class_hard_regs_num[next_cl] - ira_reg_class_max_nregs[next_cl][GET_MODE (op)])) + if (cl == NO_REGS + ? ira_class_singleton[next_cl][GET_MODE (op)] 0 + : (ira_class_singleton[cl][GET_MODE (op)] +!= ira_class_singleton[next_cl][GET_MODE (op)])) return NO_REGS; cl = next_cl; break; @@ -861,10 +862,10 @@ single_reg_class (const char *constraint next_cl = single_reg_class (recog_data.constraints[c - '0'], recog_data.operand[c - '0'], NULL_RTX); - if ((cl != NO_REGS next_cl != cl) - || next_cl == NO_REGS - || (ira_class_hard_regs_num[next_cl] - ira_reg_class_max_nregs[next_cl][GET_MODE (op)])) + if (cl == NO_REGS + ? ira_class_singleton[next_cl][GET_MODE (op)] 0 + : (ira_class_singleton[cl][GET_MODE (op)] +!= ira_class_singleton[next_cl][GET_MODE (op)])) return NO_REGS; cl = next_cl; break;
Re: [PATCH, i386]: Implement atomic_fetch_sub
Il 03/08/2012 17:08, Richard Henderson ha scritto: On 2012-08-03 08:01, Uros Bizjak wrote: On Fri, Aug 3, 2012 at 4:40 PM, Richard Henderson r...@redhat.com wrote: On 2012-08-03 01:51, Uros Bizjak wrote: The same reasoning goes for dynamic negation: for neg %eax,%eax value 0x8000 stays the same, but we have changed (x)sub to an (x)add in the code stream. So? Did you think the xadd will trap? No, but can we ignore the fact that we changed xsub -0x8000, mem to xadd -0x08000, mem? Yes, since it'll have the same effect on the bits. In fact we can even use this trick for xxor... Paolo
Re: RFC: LRA for x86/x86-64 [0/9]
On 12-10-01 6:30 AM, Bernd Schmidt wrote: On 10/01/2012 12:14 PM, Jakub Jelinek wrote: On Mon, Oct 01, 2012 at 12:01:36PM +0200, Steven Bosscher wrote: I would also agree if it were not for the fact that IRA is already a scalability bottle-neck and that has been known for a long time, too. I have no confidence at all that if LRA goes in now, these scalability problems will be solved in stage3 or at any next release cycle. It's always the same thing with GCC: Once a patch is in, everyone moves on to the next fancy new thing dropping the not-quite-broken but also not-quite-working things on the floor. If we open a P1 bug for it for 4.8, then it will need to be resolved some way before branching. I think Vlad is committed to bugfixing LRA, after all the intent is for 4.9 to enable it on more (all?) targets, and all the bugfixing and scalability work on LRA is needed for that anyway. Why can't this be done on the branch? We've made the mistake of rushing things into mainline too early a few times before, we should have learned by now. And adding more half transitions is not something we really want either. I should clearly express that the transition will be not happen for short time because of the task complexity. I believe that lra will coexist with reload for 1-2 releases. I only ported LRA for 9 major targets. The transition completion will be dependent on secondary target maintainers too because I alone can not do porting LRA for all supported targets. It was discussed with a lot of people on 2012 GNU Tools Cauldron. Maintenance of LRA on the branch is a big burden, even x86-64 is sometimes broken after merge with the trunk. When I proposed merge LRA to gcc4.8, I had in mind that: o moving most changes from LRA branch will help LRA maintenance on the branch and I'll have more time to work on other targets and problems. o the earlier we start the transition, the better it will be for LRA because LRA on the trunk will have more feedback and better testing. I've chosen x86/x86-64 for this because I am confident in this port. On majority of tests, it generates faster, smaller code (even for these two extreme tests it generates 15% smaller code) for less time. IMO, the slow compilation of the extreme tests are much less important than what I've just mentioned. But because I got clear objections from at least two people and no clear support for the LRA inclusion (there were just no objections to include it), I will not insists on LRA merge now. I believe in the importance of this work as LLVM catches GCC on RA front by implementing a new RA for LLVM3.0. I believe we should get rid off reload as outdated, hard to maintain, and preventing implementation of new RA optimizations. In any case submitting the patches was a good thing to do because I got a lot of feedback. I still appreciate any comments on the patches.
Re: [RFC] Make vectorizer to skip loops with small iteration estimate
So the unvectorized cost is SIC * niters The vectorized path is SOC + VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC The scalar path of vectorizer loop is SIC * niters + SOC Note that 'th' is used for the runtime profitability check which is done at the time the setup cost has already been taken (yes, we Yes, I understand that. probably should make it more conservative but then guard the whole set of loops by the check, not only the vectorized path). See PR53355 for the general issue. Yep, we may reduce the cost of SOC by outputting early guard for non-vectorized path better than we do now. However... Of course this is very simple benchmark, in reality the vectorizatoin can be a lot more harmful by complicating more complex control flows. So I guess we have two options 1) go with the new formula and try to make cost model a bit more realistic. 2) stay with original formula that is quite close to reality, but I think more by an accident. I think we need to improve it as whole, thus I'd prefer 2). ... I do not see why. Even if we make the check cheaper we will only distribute part of SOC to vector prologues/epilogues. Still I think the formula is wrong, I.e. accounting SOC where it should not. The cost of scalar path without vectorization is niters * SIC while with vectorization we have scalar path niters * SIC + SOC and vector path SOC + VIC * ((niters-PL_ITERS-EP_ITERS)/VF) + VOC So SOC cancels out in the runtime check. I still think we need two formulas - one determining if vectorization is profitable, other specifying the threshold for scalar path at runtime (that will generally give lower values). 2) Even when loop iterates 2 times, it is estimated to 4 iterations by estimated_stmt_executions_int with the profile feedback. The reason is loop_ch pass. Given a rolled loop with exit probability 30%, proceeds by duplicating the header with original probabilities. This makes the loop to be executed with 60% probability. Because the loop body counts remain the same (and they should), the expected number of iterations increase by the decrease of entry edge to the header. I wonder what to do about this. Obviously without path profiling loop_ch can not really do a good job. We can artifically make header to suceed more likely, that is the reality, but that requires non-trivial loop profile updating. We can also simply record the iteration bound into loop structure and ignore that the profile is not realistic But we don't preserve loop structure from header copying ... From what time we keep loop structure? In general I would like to eventualy drop value histograms to loop structure specifying number of iterations with profile feedback. Finally we can duplicate loop headers before profilng. I implemented that via early_ch pass executed only with profile generation or feedback. I guess it makes sense to do, even if it breaks the assumption that we should do strictly -Os generation on paths where Well, there are CH cases that do not increase code size and I doubt that loop header copying is generally bad for -Os ... we are not good at handling non-copied loop headers. There is comment saying /* Loop header copying usually increases size of the code. This used not to be true, since quite often it is possible to verify that the condition is satisfied in the first iteration and therefore to eliminate it. Jump threading handles these cases now. */ if (optimize_loop_for_size_p (loop)) return false; I am not sure how much backing it has. Schedule loop_ch as part of early passes just after profile pass makes optimize_loop_for_size_p to return true even for functions that are later found cold by profile feedback. I do not see that being big issue. I tested enabling loop_ch in early passes with -fprofile-feedback and it is SPEC neutral. Given that it improves loop count estimates, I would still like mainline doing that. I do not like these quite important estimates to be wrong most of time. Btw, I added a similar check in vect_analyze_loop_operations: if ((LOOP_VINFO_NITERS_KNOWN_P (loop_vinfo) (LOOP_VINFO_INT_NITERS (loop_vinfo) vectorization_factor)) || ((max_niter = max_stmt_executions_int (loop)) != -1 (unsigned HOST_WIDE_INT) max_niter vectorization_factor)) { if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, not vectorized: iteration count too small.); if (dump_kind_p (MSG_MISSED_OPTIMIZATION)) dump_printf_loc (MSG_MISSED_OPTIMIZATION, vect_location, not vectorized: iteration count smaller than vectorization factor.); return false; } maybe you simply
Re: [PATCH] Add option for dumping to stderr (issue6190057)
Your change on September 30th, breaks the powerpc port because the REPORT_DETAILS value in the enumeration is no longer there, and the rs6000_density_test function was using that. Please in the future, when you are making global changes, grep for uses of enum values in all of the machine dependent directories so we can avoid breakage like this. -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH] Add option for dumping to stderr (issue6190057)
On Mon, Oct 01, 2012 at 02:02:26PM -0400, Michael Meissner wrote: Your change on September 30th, breaks the powerpc port because the REPORT_DETAILS value in the enumeration is no longer there, and the rs6000_density_test function was using that. Please in the future, when you are making global changes, grep for uses of enum values in all of the machine dependent directories so we can avoid breakage like this. Also, in looking at the changes, given we are already up to 28 TDF_ flags, I would recommend immediately adding a new type that is the TDF flagword type. Thus it will be a lot simpler when we add 4 more TDF flags and have to change the type from int to HOST_WIDE_INT. -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [patch, libfortran] Fix PR 54736, memory corruption with GFORTRAN_CONVERT_UNIT
Hello world, the previous version of the patch has an issue that Shane pointed out in the PR. This version should work; at least it survived all the test cases I could come up with. Regression-tested (again). OK for trunk? Also for 4.6 and 4.7? Thomas 2012-10-01 Thomas König tkoe...@gcc.gnu.org PR libfortran/54736 * runtime/environ.c (search_unit): Correct logic for binary search. (mark_single): Fix index errors. Index: runtime/environ.c === --- runtime/environ.c (Revision 191857) +++ runtime/environ.c (Arbeitskopie) @@ -459,21 +459,35 @@ search_unit (int unit, int *ip) { int low, high, mid; - low = -1; - high = n_elist; - while (high - low 1) + if (n_elist == 0) { + *ip = 0; + return 0; +} + + low = 0; + high = n_elist - 1; + + do +{ mid = (low + high) / 2; - if (unit = elist[mid].unit) - high = mid; + if (unit == elist[mid].unit) + { + *ip = mid; + return 1; + } + else if (unit elist[mid].unit) + low = mid + 1; else - low = mid; -} - *ip = high; - if (elist[high].unit == unit) -return 1; + high = mid - 1; +} while (low = high); + + if (unit elist[mid].unit) +*ip = mid + 1; else -return 0; +*ip = mid; + + return 0; } /* This matches a keyword. If it is found, return the token supplied, @@ -588,13 +602,13 @@ mark_single (int unit) } if (search_unit (unit, i)) { - elist[unit].conv = endian; + elist[i].conv = endian; } else { - for (j=n_elist; j=i; j--) + for (j=n_elist-1; j=i; j--) elist[j+1] = elist[j]; - + n_elist += 1; elist[i].unit = unit; elist[i].conv = endian;
Re: [PATCH] Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 1:27 PM, Michael Meissner meiss...@linux.vnet.ibm.com wrote: On Mon, Oct 01, 2012 at 02:02:26PM -0400, Michael Meissner wrote: Your change on September 30th, breaks the powerpc port because the REPORT_DETAILS value in the enumeration is no longer there, and the rs6000_density_test function was using that. Please in the future, when you are making global changes, grep for uses of enum values in all of the machine dependent directories so we can avoid breakage like this. Also, in looking at the changes, given we are already up to 28 TDF_ flags, I would recommend immediately adding a new type that is the TDF flagword type. Thus it will be a lot simpler when we add 4 more TDF flags and have to change the type from int to HOST_WIDE_INT. Agreed that we need an abstraction here. -- Gaby
Re: RFC: LRA for x86/x86-64 [4/9]
Thanks a lot for doing this. When you finally get to the stage of rm reload.c reload1.c, please do it in a screen session and save the log for posterity. Vladimir Makarov vmaka...@redhat.com writes: +/* Return register bank of given hard regno for the current target. */ +DEFHOOK +(register_bank, + A target hook which returns the register bank number to which the\ + register @var{hard_regno} belongs to. The smaller the number, the\ + more preferable the hard register usage (when all other conditions are\ + the same). This hook can be used to prefer some hard register over\ + others in LRA. For example, some x86-64 register usage needs\ + additional prefix which makes instructions longer. The hook can\ + return bigger bank number for such registers make them less favorable\ + and as result making the generated code smaller.\ + \ + The default version of this target hook returns always zero., + int, (int), + default_register_bank) This is a horribly bikeshed-level comment, sorry, but I wonder if something like register_priority would be better. Register classes are in some ways an extension of register banks, so it wasn't obvious from the name why we needed both. +/* Return true if maximal address displacement can be different. */ +DEFHOOK +(different_addr_displacement_p, + A target hook which returns true if an address with the same structure\ + can have different maximal legitimate displacement. For example, the\ + displacement can depend on memory mode or on operand combinations in\ + the insn.\ + \ + The default version of this target hook returns always false., + bool, (void), + default_different_addr_displacement_p) If I read the patch correctly, this is only used in: + if (lra_reg_spill_p || targetm.different_addr_displacement_p ()) + lra_set_used_insn_alternative (insn, -1); and so we keep the current alternative when neither spill_class_mode nor different_addr_displacement_p is defined. How many targets on the LRA branch are like that? I would have expected most targets with limited address displacements would have to return true for the above hook, because multiword loads and stores typically have to be split into word loads and stores. Same goes for strict-alignment targets, where wider modes often have slightly lower maximal displacements. E.g. for MIPS, SImode loads and stores have a displacement range of [-32768, 32764], but DImode loads and stores only accept [-32768, 32760]. So the maximal displacement depends on mode, even though the instruction set is pretty regular. Targets with full address-size displacements can use the default false return, but it looks like the x86 port defines spill_class_mode instead, so AIUI the value isn't really tested on Core i7. What's the impact of that compared to the other x86 targets that don't set X86_TUNE_GENERAL_REGS_SSE_SPILL? Is LRA just quicker for them, or will it make different decisions (compared to Core i7) even for non-SSE insns? +/* Determine class of registers which could be used for spilled + pseudos instead of memory. */ +DEFHOOK +(spill_class, + This hook defines a class of registers which could be used for spilled pseudos\ + of given class instead of memory, + reg_class_t, (reg_class_t), + NULL) Should probably say that NO_REGS means none. +/* Determine mode for spilling pseudos into registers instead of memory. */ +DEFHOOK +(spill_class_mode, + This hook defines mode in which a pseudo of given mode and of the first\ + register class can be spilled into the second register class, + enum machine_mode, (reg_class_t, reg_class_t, enum machine_mode), + NULL) It looks like the only use is in: + || (targetm.spill_class_mode (rclass, spill_class, + PSEUDO_REGNO_MODE (regno)) + != PSEUDO_REGNO_MODE (regno)) So would it make sense to have a single hook like: /* Determine mode for spilling pseudos into registers instead of memory. */ DEFHOOK (spill_class, This hook defines a class of registers which could be used for spilling\ pseudos of the given mode and class, or @code{NO_REGS} if only memory\ should be used. Not defining this hook is equivalent to returning\ @code{NO_REGS} for all inputs. reg_class_t, (reg_class_t, enum machine_mode), NULL) ? It means that setup_reg_spill_flag needs a class-x-mode walk, but (bad excuse ahoy) we have plenty of those already. If we really wanted to avoid the extra loop, we could make VOIDmode mean any mode, although that does make the interface a bit more clunky. Richard
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 10:51 AM, Vladimir Makarov vmaka...@redhat.com wrote: When I proposed merge LRA to gcc4.8, I had in mind that: o moving most changes from LRA branch will help LRA maintenance on the branch and I'll have more time to work on other targets and problems. o the earlier we start the transition, the better it will be for LRA because LRA on the trunk will have more feedback and better testing. I've chosen x86/x86-64 for this because I am confident in this port. On majority of tests, it generates faster, smaller code (even for these two extreme tests it generates 15% smaller code) for less time. IMO, the slow compilation of the extreme tests are much less important than what I've just mentioned. But because I got clear objections from at least two people and no clear support for the LRA inclusion (there were just no objections to include it), I will not insists on LRA merge now. I believe that we should proceed with the LRA merge as Vlad has proposed, and treat the compilation time slowdowns on specific test cases as bugs to be addressed. Clearly these slowdowns are not good. However, requiring significant work like LRA to be better or equal to the current code in every single case is making the perfect the enemy of the good. We must weigh the benefits and drawbacks, not require that there be no drawbacks at all. In this case I believe that the benefits of LRA significantly outweigh the drawbacks. Steven is correct in saying that there is a tendency to move on and never address GCC bugs. However, there is also a counter-vailing tendency to fix GCC bugs. Anyhow I'm certainly not saying that in all cases it's OK to accept a merge with regressions; I'm saying that in this specific case it is OK. (I say all this based on Vlad's descriptions, I have not actually looked at the patches.) Ian
Re: RFC: LRA for x86/x86-64 [4/9]
On Oct 1, 2012, at 2:51 PM, Richard Sandiford wrote: ... E.g. for MIPS, SImode loads and stores have a displacement range of [-32768, 32764], but DImode loads and stores only accept [-32768, 32760]. So the maximal displacement depends on mode, even though the instruction set is pretty regular. It may be that the case doesn't arise in code GCC generates, but I don't think that's true. The offset field is always a 2's complement 16 bit integer, hence in the range -32768..32767. The alignment required in loading multibyte data with aligned load/store instructions applies to the final address, not the offset. For example, if R1 contains 1, then LD r2,32767(r1) will work. paul
Re: [Patch contrib] check_GNU_style: remove tmp file
On Mon, Sep 10, 2012 at 5:23 AM, Christophe Lyon christophe.l...@linaro.org wrote: Good point. Here is a new version, catching the same signals as warn_summary. This is OK. Thanks. (Minor note: it's easier if you the ChangeLog entry in the body of the message rather than the patch.) Ian
Re: RFC: LRA for x86/x86-64 [0/9]
From: Ian Lance Taylor i...@google.com Date: Mon, 1 Oct 2012 11:55:56 -0700 Steven is correct in saying that there is a tendency to move on and never address GCC bugs. However, there is also a counter-vailing tendency to fix GCC bugs. Anyhow I'm certainly not saying that in all cases it's OK to accept a merge with regressions; I'm saying that in this specific case it is OK. I think it's more important in this case to recognize Steven's real point, which is that for an identical situation (IRA), and with an identical patch author, we had similar bugs. They were promised to be worked on, and yet some of those regressions are still very much with us. The likelyhood of a repeat is therefore very real. I really don't have a lot of confidence given what has happened in the past. I also don't understand what's so evil about sorting this out on a branch. It's the perfect carrot to get the compile time regressions fixed.
Re: PR 53889: Add __gthread_recursive_mutex_destroy
On Sun, Sep 30, 2012 at 11:41 AM, Jonathan Wakely jwakely@gmail.com wrote: There is no __gthread_recursive_mutex_destroy function in the gthreads API. Trying to use __gthread_mutex_destroy fails to compile on platforms where the mutex types are different. To avoid resource leaks libstdc++ needs to hack around the missing function with overloaded functions and SFINAE tricks to detect how a recursive mutex can be destroyed. This patch extends the gthreads API to include __gthread_recursive_mutex_destroy, defining it for each gthread model, and removing the hacks from libstdc++. +return rtems_gxx_mutex_destroy( __mutex ); Space before '(', not space after. Doing anything else here is going to be painful, but this assumes that RTEMS uses the same representation for non-recursive and recursive mutexes. That is currently true, but it deserves a comment. --- a/libgcc/config/i386/gthr-win32.h +++ b/libgcc/config/i386/gthr-win32.h +static inline void +__gthread_recursive_mutex_destroy (__gthread_recursive_mutex_t *mutex) +{ + __gthread_mutex_t __mutex2; + __mutex2.sema = mutex-sema; + __gthr_win32_mutex_destroy (__mutex2); +} I think it would be better to put this in libgcc/config/i386/gthr-win32.c, like the other functions. Then you can just call CloseHandle. --- a/libgcc/config/mips/gthr-mipssde.h +++ b/libgcc/config/mips/gthr-mipssde.h +static inline int +__gthread_recursive_mutex_destroy (__gthread_recursive_mutex_t *__mutex) +{ + return __gthread_mutex_destroy(__mutex); +} Will this even compile? It doesn't look like it. Ian
[SH] PR 51244 - Handle T bit - 0x7FFFFFFF / 0x80000000
Hello, This handles the case where the T bit is stored to a reg as the value 0x7FFF or 0x8000. Tested on rev 191894 with make -k check RUNTESTFLAGS=--target_board=sh-sim \{-m2/-ml,-m2/-mb,-m2a/-mb,-m4/-ml,-m4/-mb,-m4a/-ml,-m4a/-mb} and no new failures. OK? Cheers, Oleg gcc/ChangeLog: PR target/51244 * config/sh/sh.md (*mov_t_msb_neg): New insn and two accompanying unnamed split patterns. testsuite/ChangeLog: PR target/51244 * gcc.target/sh/pr51244-12.c: New. Index: gcc/testsuite/gcc.target/sh/pr51244-12.c === --- gcc/testsuite/gcc.target/sh/pr51244-12.c (revision 0) +++ gcc/testsuite/gcc.target/sh/pr51244-12.c (revision 0) @@ -0,0 +1,68 @@ +/* Check that the negc instruction is generated as expected for the cases + below. If we see a movrt or #-1 negc sequence it means that the pattern + which handles the inverted case does not work properly. */ +/* { dg-do compile { target sh*-*-* } } */ +/* { dg-options -O1 } */ +/* { dg-skip-if { sh*-*-* } { -m5* } { } } */ +/* { dg-final { scan-assembler-times negc 10 } } */ +/* { dg-final { scan-assembler-not movrt|#-1|add|sub } } */ + +int +test00 (int a, int b, int* x) +{ + return (a == b) ? 0x7FFF : 0x8000; +} + +int +test00_inv (int a, int b) +{ + return (a != b) ? 0x8000 : 0x7FFF; +} + +int +test01 (int a, int b) +{ + return (a = b) ? 0x7FFF : 0x8000; +} + +int +test01_inv (int a, int b) +{ + return (a b) ? 0x8000 : 0x7FFF; +} + +int +test02 (int a, int b) +{ + return (a b) ? 0x7FFF : 0x8000; +} + +int +test02_inv (int a, int b) +{ + return (a = b) ? 0x8000 : 0x7FFF; +} + +int +test03 (int a, int b) +{ + return ((a b) == 0) ? 0x7FFF : 0x8000; +} + +int +test03_inv (int a, int b) +{ + return ((a b) != 0) ? 0x8000 : 0x7FFF; +} + +int +test04 (int a) +{ + return ((a 0x55) == 0) ? 0x7FFF : 0x8000; +} + +int +test04_inv (int a) +{ + return ((a 0x55) != 0) ? 0x8000 : 0x7FFF; +} Index: gcc/config/sh/sh.md === --- gcc/config/sh/sh.md (revision 191894) +++ gcc/config/sh/sh.md (working copy) @@ -10769,6 +10769,51 @@ (set (reg:SI T_REG) (const_int 1)) (use (match_dup 2))])]) +;; Use negc to store the T bit in a MSB of a reg in the following way: +;; T = 1: 0x8000 - reg +;; T = 0: 0x7FFF - reg +;; This works because 0 - 0x8000 = 0x8000. +(define_insn_and_split *mov_t_msb_neg + [(set (match_operand:SI 0 arith_reg_dest) + (minus:SI (const_int -2147483648) ;; 0x8000 + (match_operand 1 t_reg_operand))) + (clobber (reg:SI T_REG))] + TARGET_SH1 + # + can_create_pseudo_p () + [(set (match_dup 2) (const_int -2147483648)) + (parallel [(set (match_dup 0) (minus:SI (neg:SI (match_dup 2)) + (reg:SI T_REG))) + (clobber (reg:SI T_REG))])] +{ + operands[2] = gen_reg_rtx (SImode); +}) + +;; These are essentially the same as above, but with the inverted T bit. +;; Combine recognizes the split patterns, but does not take them sometimes +;; if the T_REG clobber is specified. Instead it tries to split out the +;; T bit negation. Since these splits are supposed to be taken only by +;; combine, it will see the T_REG clobber of the *mov_t_msb_neg insn, so this +;; should be fine. +(define_split + [(set (match_operand:SI 0 arith_reg_dest) + (plus:SI (match_operand 1 negt_reg_operand) + (const_int 2147483647)))] ;; 0x7fff + TARGET_SH1 can_create_pseudo_p () + [(parallel [(set (match_dup 0) + (minus:SI (const_int -2147483648) (reg:SI T_REG))) + (clobber (reg:SI T_REG))])]) + +(define_split + [(set (match_operand:SI 0 arith_reg_dest) + (if_then_else:SI (match_operand 1 t_reg_operand) + (const_int 2147483647) ;; 0x7fff + (const_int -2147483648)))] ;; 0x8000 + TARGET_SH1 can_create_pseudo_p () + [(parallel [(set (match_dup 0) + (minus:SI (const_int -2147483648) (reg:SI T_REG))) + (clobber (reg:SI T_REG))])]) + ;; The *negnegt pattern helps the combine pass to figure out how to fold ;; an explicit double T bit negation. (define_insn_and_split *negnegt
Re: RFC: LRA for x86/x86-64 [0/9]
On 10/01/2012 03:19 PM, David Miller wrote: From: Ian Lance Taylor i...@google.com Date: Mon, 1 Oct 2012 11:55:56 -0700 Steven is correct in saying that there is a tendency to move on and never address GCC bugs. However, there is also a counter-vailing tendency to fix GCC bugs. Anyhow I'm certainly not saying that in all cases it's OK to accept a merge with regressions; I'm saying that in this specific case it is OK. I think it's more important in this case to recognize Steven's real point, which is that for an identical situation (IRA), and with an identical patch author, we had similar bugs. They were promised to be worked on, and yet some of those regressions are still very much with us. That is not true. I worked on many compiler time regression bugs. I remeber one serious degradation of compilation time on all_cp2k_gfortran.f90. I solved the problem and make IRA working faster and generating much better code than the old RA. http://blog.gmane.org/gmane.comp.gcc.patches/month=20080501/page=15 About other two mentioned PRs by Steven: PR26854. I worked on this bug even when IRA was on the branch and make again GCC with IRA 5% faster on this test than GCC with the old RA. PR 54146 is 3 months old. There were a lot work on other optimizations before IRA became important. It happens only 2 months ago. I had no time to work on it but I am going to. People sometimes see that RA takes a lot of compilation time but it is in the nature of RA. I'd recommend first to check how the old RA behaves and then call it a degradation. And please, don't listen just one side. The likelyhood of a repeat is therefore very real. I really don't have a lot of confidence given what has happened in the past. I also don't understand what's so evil about sorting this out on a branch. It's the perfect carrot to get the compile time regressions fixed. Wrong assumptions result in wrong conclusions.
Re: RFC: LRA for x86/x86-64 [0/9]
On Mon, Oct 1, 2012 at 9:51 PM, Vladimir Makarov vmaka...@redhat.com wrote: I think it's more important in this case to recognize Steven's real point, which is that for an identical situation (IRA), and with an identical patch author, we had similar bugs. They were promised to be worked on, and yet some of those regressions are still very much with us. That is not true. I worked on many compiler time regression bugs. I remeber one serious degradation of compilation time on all_cp2k_gfortran.f90. I solved the problem and make IRA working faster and generating much better code than the old RA. http://blog.gmane.org/gmane.comp.gcc.patches/month=20080501/page=15 About other two mentioned PRs by Steven: PR26854. I worked on this bug even when IRA was on the branch and make again GCC with IRA 5% faster on this test than GCC with the old RA. PR 54146 is 3 months old. There were a lot work on other optimizations before IRA became important. It happens only 2 months ago. I had no time to work on it but I am going to. This is also not quite true, see PR37448, which shows the problems as the test case for PR54146. I just think scalability is a very important issue. If some pass or algorithm scales bad on some measure, then users _will_ run into that at some point and report bugs about it (if you're lucky enough to have a user patient enough to sit out the long compile time :-) ). Also, good scalability opens up opportunities. For example, historically GCC has been conservative on inlining heuristics to avoid compile time explosions. I think it's better to address the causes of that explosion and to avoid introducing new potential bottlenecks. People sometimes see that RA takes a lot of compilation time but it is in the nature of RA. I'd recommend first to check how the old RA behaves and then call it a degradation. There's no question that RA is one of the hardest problems the compiler has to solve, being NP-complete and all that. I like LRA's iterative approach, but if you know you're going to solve a hard problem with a number potentially expensive iterations, there's even more reason to make scalability a design goal! As I said earlier in this thread, I was really looking forward to IRA at the time you worked on it, because it is supposed to be a regional allocator and I had expected that to mean it could, well, allocate per-region which is usually very helpful for scalability (partition your function and insert compensation code on strategically picked region boundaries). But that's not what IRA has turned out to be. (Instead, its regional nature is one of the reasons for its scalability problems.) IRA is certainly not worse than old global.c in very many ways, and LRA looks like a well thought-through and welcome replacement of old reload. But scalability is an issue in the design of IRA and LRA looks to be the same in that regard. Ciao! Steven
Re: RFC: LRA for x86/x86-64 [4/9]
On 09/27/2012 07:44 PM, Vladimir Makarov wrote: On 09/27/2012 08:07 PM, Joseph S. Myers wrote: On Thu, 27 Sep 2012, Vladimir Makarov wrote: Hook spill_class returns a value of enum reg_class which is defined in target-depend include file. That's what reg_class_t is for: avoiding enum reg_class in hook interfaces. Ok. Thanks for pointing this out. Here is the modified patch. 2012-09-27 Vladimir Makarov vmaka...@redhat.com * targhooks.h (default_lra_p): Declare. (default_register_bank): Ditto. (default_different_addr_displacement_p): Ditto. * targhooks.c (default_lra_p): New function. (default_register_bank): Ditto. (default_different_addr_displacement_p): Ditto. * target.def (lra_p): New hook. (register_bank): Ditto. (different_addr_displacement_p): Ditto. (spill_class, spill_class_mode): New hooks. * doc/tm.texi.in: Add TARGET_LRA_P, TARGET_REGISTER_BANK, TARGET_DIFFERENT_ADDR_DISPLACEMENT_P, TARGET_SPILL_CLASS, and TARGET_SPILL_CLASS_MODE. * doc/tm.texi: Update. The change also requires some modification in the 9th patch. The ChangeLog for the patch should be the same as before. This looks fine to me. jeff
[PATCH, libbacktrace]: Compile with -fasynchronous-unwind-tables
Hello! Without -fasynchronous-unwind-tables, FDE is not generated for backtrace_full and backtrace_simple wrappers. Without FDE, unwinding terminates at these functions. Attached patch fixes this problem by adding -fasynchronous-unwind-tables, and this way forcing FDEs for all functions. With this change, btest passes OK, failing log and runtime/pprof from libgo testsuite also pass OK. BTW: It would be enough to compile only backtrace.c and simple.c with -fasynchronous-unwind-tables, since critical wrapper functions live here. 2012-10-01 Uros Bizjak ubiz...@gmail.com PR other/54761 * Makefile.am (AM_CFLAGS): Add -fasynchronous-unwind-tables. * Makefile.in: Regenerate. Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32} and alphaev68-pc-linux-gnu (where fixes all mentioned unwinding failures). OK for mainline? Uros. Index: ChangeLog === --- ChangeLog (revision 191932) +++ ChangeLog (working copy) @@ -1,3 +1,9 @@ +2012-10-01 Uros Bizjak ubiz...@gmail.com + + PR other/54761 + * Makefile.am (AM_CFLAGS): Add -fasynchronous-unwind-tables. + * Makefile.in: Regenerate. + 2012-09-29 Ian Lance Taylor i...@google.com PR other/54749 Index: Makefile.am === --- Makefile.am (revision 191932) +++ Makefile.am (working copy) @@ -34,7 +34,7 @@ AM_CPPFLAGS = -I $(top_srcdir)/../include -I $(top_srcdir)/../libgcc \ -I ../libgcc -I ../gcc/include -I $(MULTIBUILDTOP)../../gcc/include -AM_CFLAGS = $(WARN_FLAGS) $(PIC_FLAG) +AM_CFLAGS = $(WARN_FLAGS) $(PIC_FLAG) -fasynchronous-unwind-tables noinst_LTLIBRARIES = libbacktrace.la Index: Makefile.in === --- Makefile.in (revision 191932) +++ Makefile.in (working copy) @@ -253,7 +253,7 @@ AM_CPPFLAGS = -I $(top_srcdir)/../include -I $(top_srcdir)/../libgcc \ -I ../libgcc -I ../gcc/include -I $(MULTIBUILDTOP)../../gcc/include -AM_CFLAGS = $(WARN_FLAGS) $(PIC_FLAG) +AM_CFLAGS = $(WARN_FLAGS) $(PIC_FLAG) -fasynchronous-unwind-tables noinst_LTLIBRARIES = libbacktrace.la libbacktrace_la_SOURCES = \ backtrace.h \
Re: [patch][lra] Use XNEWVEC and friends instead of xmalloc/xrealloc, and add some timevars
On Oct 1, 2012, at 4:05 AM, Steven Bosscher stevenb@gmail.com wrote: This patch uses the libiberty new-like operators instead of using xmalloc/xrealloc. So, in headers that can be used by C compiles, we should use the abstraction… Should we prefer it for translation units that are C++ only?
[PATCH] Fix powerpc breakage, was: Add option for dumping to stderr (issue6190057)
I tracked down some of the other code that previously used REPORT_DETAILS, and MSG_NOTE is the new way to do the same thing. This bootstraps and no unexpected errors occur during make check. Is it ok to install? 2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (toplevel): Include dumpfile.h. (rs6000_density_test): Rework to accomidate 09-30 change by Sharad Singhai. * config/rs6000/t-rs6000 (rs6000.o): Add dumpfile.h dependency. Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191932) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -58,6 +58,7 @@ #include tm-constrs.h #include opts.h #include tree-vectorizer.h +#include dumpfile.h #if TARGET_XCOFF #include xcoffout.h /* get declarations of xcoff_*_section_name */ #endif @@ -3518,11 +3519,11 @@ rs6000_density_test (rs6000_cost_data *d vec_cost + not_vec_cost DENSITY_SIZE_THRESHOLD) { data-cost[vect_body] = vec_cost * (100 + DENSITY_PENALTY) / 100; - if (vect_print_dump_info (REPORT_DETAILS)) - fprintf (vect_dump, -density %d%%, cost %d exceeds threshold, penalizing -loop body cost by %d%%, density_pct, -vec_cost + not_vec_cost, DENSITY_PENALTY); + if (dump_kind_p (MSG_NOTE)) + dump_printf_loc (MSG_NOTE, vect_location, +density %d%%, cost %d exceeds threshold, penalizing +loop body cost by %d%%, density_pct, +vec_cost + not_vec_cost, DENSITY_PENALTY); } } Index: gcc/config/rs6000/t-rs6000 === --- gcc/config/rs6000/t-rs6000 (revision 191932) +++ gcc/config/rs6000/t-rs6000 (working copy) @@ -26,7 +26,7 @@ rs6000.o: $(CONFIG_H) $(SYSTEM_H) corety $(OBSTACK_H) $(TREE_H) $(EXPR_H) $(OPTABS_H) except.h function.h \ output.h dbxout.h $(BASIC_BLOCK_H) toplev.h $(GGC_H) $(HASHTAB_H) \ $(TM_P_H) $(TARGET_H) $(TARGET_DEF_H) langhooks.h reload.h gt-rs6000.h \ - cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) + cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) dumpfile.h rs6000-c.o: $(srcdir)/config/rs6000/rs6000-c.c \ $(srcdir)/config/rs6000/rs6000-protos.h \ -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH] Fix powerpc breakage, was: Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 4:37 PM, Michael Meissner meiss...@linux.vnet.ibm.com wrote: I tracked down some of the other code that previously used REPORT_DETAILS, and MSG_NOTE is the new way to do the same thing. This bootstraps and no unexpected errors occur during make check. Is it ok to install? yes -- qualifies as obvious. Thanks! 2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (toplevel): Include dumpfile.h. (rs6000_density_test): Rework to accomidate 09-30 change by Sharad Singhai. * config/rs6000/t-rs6000 (rs6000.o): Add dumpfile.h dependency. Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191932) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -58,6 +58,7 @@ #include tm-constrs.h #include opts.h #include tree-vectorizer.h +#include dumpfile.h #if TARGET_XCOFF #include xcoffout.h /* get declarations of xcoff_*_section_name */ #endif @@ -3518,11 +3519,11 @@ rs6000_density_test (rs6000_cost_data *d vec_cost + not_vec_cost DENSITY_SIZE_THRESHOLD) { data-cost[vect_body] = vec_cost * (100 + DENSITY_PENALTY) / 100; - if (vect_print_dump_info (REPORT_DETAILS)) - fprintf (vect_dump, -density %d%%, cost %d exceeds threshold, penalizing -loop body cost by %d%%, density_pct, -vec_cost + not_vec_cost, DENSITY_PENALTY); + if (dump_kind_p (MSG_NOTE)) + dump_printf_loc (MSG_NOTE, vect_location, +density %d%%, cost %d exceeds threshold, penalizing +loop body cost by %d%%, density_pct, +vec_cost + not_vec_cost, DENSITY_PENALTY); } } Index: gcc/config/rs6000/t-rs6000 === --- gcc/config/rs6000/t-rs6000 (revision 191932) +++ gcc/config/rs6000/t-rs6000 (working copy) @@ -26,7 +26,7 @@ rs6000.o: $(CONFIG_H) $(SYSTEM_H) corety $(OBSTACK_H) $(TREE_H) $(EXPR_H) $(OPTABS_H) except.h function.h \ output.h dbxout.h $(BASIC_BLOCK_H) toplev.h $(GGC_H) $(HASHTAB_H) \ $(TM_P_H) $(TARGET_H) $(TARGET_DEF_H) langhooks.h reload.h gt-rs6000.h \ - cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) + cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) dumpfile.h rs6000-c.o: $(srcdir)/config/rs6000/rs6000-c.c \ $(srcdir)/config/rs6000/rs6000-protos.h \ -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Convert more non-GTY htab_t to hash_table.
Change more non-GTY hash tables to use the new type-safe template hash table. Constify member function parameters that can be const. Correct a couple of expressions in formerly uninstantiated templates. The new code is 0.362% faster in bootstrap, with a 99.5% confidence of being faster. Tested on x86-64. Okay for trunk? Index: gcc/java/ChangeLog 2012-10-01 Lawrence Crowl cr...@google.com * Make-lang.in (JAVA_OBJS): Add dependence on hash-table.o. (JCFDUMP_OBJS): Add dependence on hash-table.o. (jcf-io.o): Add dependence on hash-table.h. * jcf-io.c (memoized_class_lookups): Change to use type-safe hash table. Index: gcc/c/ChangeLog 2012-10-01 Lawrence Crowl cr...@google.com * Make-lang.in (c-decl.o): Add dependence on hash-table.h. * c-decl.c (detect_field_duplicates_hash): Change to new type-safe hash table. Index: gcc/objc/ChangeLog 2012-10-01 Lawrence Crowl cr...@google.com * Make-lang.in (OBJC_OBJS): Add dependence on hash-table.o. (objc-act.o): Add dependence on hash-table.h. * objc-act.c (objc_detect_field_duplicates): Change to new type-safe hash table. Index: gcc/ChangeLog 2012-10-01 Lawrence Crowl cr...@google.com * Makefile.in (fold-const.o): Add depencence on hash-table.h. (dse.o): Likewise. (cfg.o): Likewise. * fold-const.c (fold_checksum_tree): Change to new type-safe hash table. * (print_fold_checksum): Likewise. * cfg.c (var bb_original): Likewise. * (var bb_copy): Likewise. * (var loop_copy): Likewise. * hash-table.h (template hash_table): Constify parameters for find... and remove_elt... member functions. (hash_table::empty) Correct size expression. (hash_table::clear_slot) Correct deleted entry assignment. * dse.c (var rtx_group_table): Change to new type-safe hash table. Index: gcc/cp/ChangeLog 2012-10-01 Lawrence Crowl cr...@google.com * Make-lang.in (class.o): Add dependence on hash-table.h. (tree.o): Likewise. (semantics.o): Likewise. * class.c (fixed_type_or_null): Change to new type-safe hash table. * tree.c (verify_stmt_tree): Likewise. (verify_stmt_tree_r): Likewise. * semantics.c (struct nrv_data): Likewise. Index: gcc/java/Make-lang.in === --- gcc/java/Make-lang.in (revision 191941) +++ gcc/java/Make-lang.in (working copy) @@ -83,10 +83,10 @@ JAVA_OBJS = java/class.o java/decl.o jav java/zextract.o java/jcf-io.o java/win32-host.o java/jcf-parse.o java/mangle.o \ java/mangle_name.o java/builtins.o java/resource.o \ java/jcf-depend.o \ - java/jcf-path.o java/boehm.o java/java-gimplify.o + java/jcf-path.o java/boehm.o java/java-gimplify.o hash-table.o JCFDUMP_OBJS = java/jcf-dump.o java/jcf-io.o java/jcf-depend.o java/jcf-path.o \ - java/win32-host.o java/zextract.o ggc-none.o + java/win32-host.o java/zextract.o ggc-none.o hash-table.o JVGENMAIN_OBJS = java/jvgenmain.o java/mangle_name.o @@ -326,7 +326,7 @@ java/java-gimplify.o: java/java-gimplify # jcf-io.o needs $(ZLIBINC) added to cflags. CFLAGS-java/jcf-io.o += $(ZLIBINC) java/jcf-io.o: java/jcf-io.c $(CONFIG_H) $(SYSTEM_H) coretypes.h \ - $(JAVA_TREE_H) java/zipfile.h + $(JAVA_TREE_H) java/zipfile.h $(HASH_TABLE_H) # jcf-path.o needs a -D. CFLAGS-java/jcf-path.o += \ Index: gcc/java/jcf-io.c === --- gcc/java/jcf-io.c (revision 191941) +++ gcc/java/jcf-io.c (working copy) @@ -31,7 +31,7 @@ The Free Software Foundation is independ #include jcf.h #include tree.h #include java-tree.h -#include hashtab.h +#include hash-table.h #include dirent.h #include zlib.h @@ -271,20 +271,34 @@ find_classfile (char *filename, JCF *jcf return open_class (filename, jcf, fd, dep_name); } -/* Returns 1 if the CLASSNAME (really a char *) matches the name - stored in TABLE_ENTRY (also a char *). */ -static int -memoized_class_lookup_eq (const void *table_entry, const void *classname) +/* Hash table helper. */ + +struct charstar_hash : typed_noop_remove char { - return strcmp ((const char *)classname, (const char *)table_entry) == 0; + typedef const char T; + static inline hashval_t hash (const T *candidate); + static inline bool equal (const T *existing, const T *candidate); +}; + +inline hashval_t +charstar_hash::hash (const T *candidate) +{ + return htab_hash_string (candidate); } +inline bool +charstar_hash::equal (const T *existing, const T *candidate) +{ + return strcmp (existing, candidate) == 0; +} + + /* A hash table keeping track of class names that were not found during class lookup. (There is no need to cache the values associated with names that were found; they are saved in IDENTIFIER_CLASS_VALUE.) */ -static htab_t
[patch][lra] a few bitmap obstacks for lra-assigns
Hello, This eliminates a few large loops in lra-assigns.c. They're not the most costly loops but the life times of the bitmaps is well-defined and destroying a bitmap obstack is much cheaper than looping over all bitmaps calling bitmap_clear. The saving is small but you have to start somewhere... Bootstrapped lra-branch and tested on x86_64-unknown-linux-gnu, along with the patch from earlier today. OK for the branch? Ciao! Steven lra_assign_bitobstacks.diff Description: Binary data
[PATCH rs6000 testsuite] Fix a couple tests for VSX scalar instructions
This patch fixes a couple failures that occur if the testsuite is run with -mvsx and the VSX scalar sqrt instructions are generated. Ok for trunk? -Pat testsuite/ChangeLog: 2012-10-01 Pat Haugen pthau...@us.ibm.com * gcc.target/powerpc/pr46728-1.c: Accept xssqrtdp. * gcc.target/powerpc/pr46728-2.c: Likewise. Index: gcc/testsuite/gcc.target/powerpc/pr46728-1.c === --- gcc/testsuite/gcc.target/powerpc/pr46728-1.c(revision 191713) +++ gcc/testsuite/gcc.target/powerpc/pr46728-1.c(working copy) @@ -27,5 +27,5 @@ main (int argc, char *argv[]) } -/* { dg-final { scan-assembler-times fsqrt 2 { target powerpc*-*-* } } } */ +/* { dg-final { scan-assembler-times fsqrt|xssqrtdp 2 { target powerpc*-*-* } } } */ /* { dg-final { scan-assembler-not pow { target powerpc*-*-* } } } */ Index: gcc/testsuite/gcc.target/powerpc/pr46728-2.c === --- gcc/testsuite/gcc.target/powerpc/pr46728-2.c(revision 191713) +++ gcc/testsuite/gcc.target/powerpc/pr46728-2.c(working copy) @@ -27,5 +27,5 @@ main (int argc, char *argv[]) } -/* { dg-final { scan-assembler-times fsqrt 4 { target powerpc*-*-* } } } */ +/* { dg-final { scan-assembler-times fsqrt|xssqrtdp 4 { target powerpc*-*-* } } } */ /* { dg-final { scan-assembler-not pow { target powerpc*-*-* } } } */
[patch][lra] Comment typo fix
I suppose no-one would object if I commit this as obvious at some point? Index: lra-constraints.c === --- lra-constraints.c (revision 191858) +++ lra-constraints.c (working copy) @@ -4293,7 +4293,7 @@ update_ebb_live_info (rtx head, rtx tail { if (prev_bb != NULL) { - /* Udpate DF_LR_IN (prev_bb): */ + /* Update DF_LR_IN (prev_bb): */ EXECUTE_IF_SET_IN_BITMAP (check_only_regs, 0, j, bi) if (bitmap_bit_p (live_regs, j)) bitmap_set_bit (DF_LR_IN (prev_bb), j);
Re: [patch][lra] Comment typo fix
On 10/1/2012 6:09 PM, Steven Bosscher wrote: I suppose no-one would object if I commit this as obvious at some point? Index: lra-constraints.c === --- lra-constraints.c (revision 191858) +++ lra-constraints.c (working copy) @@ -4293,7 +4293,7 @@ update_ebb_live_info (rtx head, rtx tail { if (prev_bb != NULL) { - /* Udpate DF_LR_IN (prev_bb): */ + /* Update DF_LR_IN (prev_bb): */ EXECUTE_IF_SET_IN_BITMAP (check_only_regs, 0, j, bi) if (bitmap_bit_p (live_regs, j)) bitmap_set_bit (DF_LR_IN (prev_bb), j); took me a few readings to see the change you had made, amazing how the brain reads what it expects to see :-)
MAINTAINERS: add myself as dwarf debugging code maintainer
2012-10-01 Cary Coutant ccout...@google.com * MAINTAINERS: Add myself as dwarf debugging code maintainer. Index: MAINTAINERS === --- MAINTAINERS (revision 191942) +++ MAINTAINERS (working copy) @@ -185,6 +185,7 @@ caller-save.c Jeff Law l...@redhat.com callgraph Jan Hubicka j...@suse.cz debugging code Jim Wilson wil...@tuliptree.org dwarf debugging code Jason Merrill ja...@redhat.com +dwarf debugging code Cary Coutantccout...@google.com c++ runtime libs Paolo Carlini paolo.carl...@oracle.com c++ runtime libs Gabriel Dos Reisg...@integrable-solutions.net c++ runtime libs Ulrich Drepper drep...@gmail.com
Re: [PATCH] Fix powerpc breakage, was: Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 2:37 PM, Michael Meissner meiss...@linux.vnet.ibm.com wrote: I tracked down some of the other code that previously used REPORT_DETAILS, and MSG_NOTE is the new way to do the same thing. This bootstraps and no unexpected errors occur during make check. Is it ok to install? 2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (toplevel): Include dumpfile.h. (rs6000_density_test): Rework to accomidate 09-30 change by Sharad Singhai. * config/rs6000/t-rs6000 (rs6000.o): Add dumpfile.h dependency. Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191932) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -58,6 +58,7 @@ #include tm-constrs.h #include opts.h #include tree-vectorizer.h +#include dumpfile.h #if TARGET_XCOFF #include xcoffout.h /* get declarations of xcoff_*_section_name */ #endif @@ -3518,11 +3519,11 @@ rs6000_density_test (rs6000_cost_data *d vec_cost + not_vec_cost DENSITY_SIZE_THRESHOLD) { data-cost[vect_body] = vec_cost * (100 + DENSITY_PENALTY) / 100; - if (vect_print_dump_info (REPORT_DETAILS)) - fprintf (vect_dump, -density %d%%, cost %d exceeds threshold, penalizing -loop body cost by %d%%, density_pct, -vec_cost + not_vec_cost, DENSITY_PENALTY); + if (dump_kind_p (MSG_NOTE)) Is this check needed? Seems redundant. David + dump_printf_loc (MSG_NOTE, vect_location, +density %d%%, cost %d exceeds threshold, penalizing +loop body cost by %d%%, density_pct, +vec_cost + not_vec_cost, DENSITY_PENALTY); } } Index: gcc/config/rs6000/t-rs6000 === --- gcc/config/rs6000/t-rs6000 (revision 191932) +++ gcc/config/rs6000/t-rs6000 (working copy) @@ -26,7 +26,7 @@ rs6000.o: $(CONFIG_H) $(SYSTEM_H) corety $(OBSTACK_H) $(TREE_H) $(EXPR_H) $(OPTABS_H) except.h function.h \ output.h dbxout.h $(BASIC_BLOCK_H) toplev.h $(GGC_H) $(HASHTAB_H) \ $(TM_P_H) $(TARGET_H) $(TARGET_DEF_H) langhooks.h reload.h gt-rs6000.h \ - cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) + cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) dumpfile.h rs6000-c.o: $(srcdir)/config/rs6000/rs6000-c.c \ $(srcdir)/config/rs6000/rs6000-protos.h \ -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH, libbacktrace]: Compile with -fasynchronous-unwind-tables
On Mon, Oct 1, 2012 at 2:12 PM, Uros Bizjak ubiz...@gmail.com wrote: Without -fasynchronous-unwind-tables, FDE is not generated for backtrace_full and backtrace_simple wrappers. Without FDE, unwinding terminates at these functions. I'm not opposed to -fasynchronous-unwind-tables, but now that you bring it up I'm fairly certain that it would suffice to use -funwind-tables. I've been testing mainly on x86_64, and I forgot that on x86_64 -funwind-tables is the default. Sorry about that. And -fasynchronous-unwind-tables is the default also, so I could be wrong that -funwind-tables is all that is needed. Attached patch fixes this problem by adding -fasynchronous-unwind-tables, and this way forcing FDEs for all functions. With this change, btest passes OK, failing log and runtime/pprof from libgo testsuite also pass OK. This is basically fine but libbacktrace may be compiled by the host compiler and that may not be GCC, so please add a configure test to see if the compiler accepts the -fasynchronous-unwind-tables option. Ian
Re: [PATCH] Fix powerpc breakage, was: Add option for dumping to stderr (issue6190057)
Thanks for tracking down and fixing the powerpc port. The dump_kind_p () check is redundant but canonical form here. I think blocks of dump code guarded by if dump_kind_p (...) might be easier to read/maintain. Sharad Sharad On Mon, Oct 1, 2012 at 3:45 PM, Xinliang David Li davi...@google.com wrote: On Mon, Oct 1, 2012 at 2:37 PM, Michael Meissner meiss...@linux.vnet.ibm.com wrote: I tracked down some of the other code that previously used REPORT_DETAILS, and MSG_NOTE is the new way to do the same thing. This bootstraps and no unexpected errors occur during make check. Is it ok to install? 2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (toplevel): Include dumpfile.h. (rs6000_density_test): Rework to accomidate 09-30 change by Sharad Singhai. * config/rs6000/t-rs6000 (rs6000.o): Add dumpfile.h dependency. Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191932) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -58,6 +58,7 @@ #include tm-constrs.h #include opts.h #include tree-vectorizer.h +#include dumpfile.h #if TARGET_XCOFF #include xcoffout.h /* get declarations of xcoff_*_section_name */ #endif @@ -3518,11 +3519,11 @@ rs6000_density_test (rs6000_cost_data *d vec_cost + not_vec_cost DENSITY_SIZE_THRESHOLD) { data-cost[vect_body] = vec_cost * (100 + DENSITY_PENALTY) / 100; - if (vect_print_dump_info (REPORT_DETAILS)) - fprintf (vect_dump, -density %d%%, cost %d exceeds threshold, penalizing -loop body cost by %d%%, density_pct, -vec_cost + not_vec_cost, DENSITY_PENALTY); + if (dump_kind_p (MSG_NOTE)) Is this check needed? Seems redundant. David + dump_printf_loc (MSG_NOTE, vect_location, +density %d%%, cost %d exceeds threshold, penalizing +loop body cost by %d%%, density_pct, +vec_cost + not_vec_cost, DENSITY_PENALTY); } } Index: gcc/config/rs6000/t-rs6000 === --- gcc/config/rs6000/t-rs6000 (revision 191932) +++ gcc/config/rs6000/t-rs6000 (working copy) @@ -26,7 +26,7 @@ rs6000.o: $(CONFIG_H) $(SYSTEM_H) corety $(OBSTACK_H) $(TREE_H) $(EXPR_H) $(OPTABS_H) except.h function.h \ output.h dbxout.h $(BASIC_BLOCK_H) toplev.h $(GGC_H) $(HASHTAB_H) \ $(TM_P_H) $(TARGET_H) $(TARGET_DEF_H) langhooks.h reload.h gt-rs6000.h \ - cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) + cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) dumpfile.h rs6000-c.o: $(srcdir)/config/rs6000/rs6000-c.c \ $(srcdir)/config/rs6000/rs6000-protos.h \ -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH] Fix powerpc breakage, was: Add option for dumping to stderr (issue6190057)
On Mon, Oct 1, 2012 at 4:05 PM, Sharad Singhai sing...@google.com wrote: Thanks for tracking down and fixing the powerpc port. The dump_kind_p () check is redundant but canonical form here. I think blocks of dump code guarded by if dump_kind_p (...) might be easier to read/maintain. I find it confusing to be honest. The redundant check serves no purpose. David Sharad Sharad On Mon, Oct 1, 2012 at 3:45 PM, Xinliang David Li davi...@google.com wrote: On Mon, Oct 1, 2012 at 2:37 PM, Michael Meissner meiss...@linux.vnet.ibm.com wrote: I tracked down some of the other code that previously used REPORT_DETAILS, and MSG_NOTE is the new way to do the same thing. This bootstraps and no unexpected errors occur during make check. Is it ok to install? 2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (toplevel): Include dumpfile.h. (rs6000_density_test): Rework to accomidate 09-30 change by Sharad Singhai. * config/rs6000/t-rs6000 (rs6000.o): Add dumpfile.h dependency. Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191932) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -58,6 +58,7 @@ #include tm-constrs.h #include opts.h #include tree-vectorizer.h +#include dumpfile.h #if TARGET_XCOFF #include xcoffout.h /* get declarations of xcoff_*_section_name */ #endif @@ -3518,11 +3519,11 @@ rs6000_density_test (rs6000_cost_data *d vec_cost + not_vec_cost DENSITY_SIZE_THRESHOLD) { data-cost[vect_body] = vec_cost * (100 + DENSITY_PENALTY) / 100; - if (vect_print_dump_info (REPORT_DETAILS)) - fprintf (vect_dump, -density %d%%, cost %d exceeds threshold, penalizing -loop body cost by %d%%, density_pct, -vec_cost + not_vec_cost, DENSITY_PENALTY); + if (dump_kind_p (MSG_NOTE)) Is this check needed? Seems redundant. David + dump_printf_loc (MSG_NOTE, vect_location, +density %d%%, cost %d exceeds threshold, penalizing +loop body cost by %d%%, density_pct, +vec_cost + not_vec_cost, DENSITY_PENALTY); } } Index: gcc/config/rs6000/t-rs6000 === --- gcc/config/rs6000/t-rs6000 (revision 191932) +++ gcc/config/rs6000/t-rs6000 (working copy) @@ -26,7 +26,7 @@ rs6000.o: $(CONFIG_H) $(SYSTEM_H) corety $(OBSTACK_H) $(TREE_H) $(EXPR_H) $(OPTABS_H) except.h function.h \ output.h dbxout.h $(BASIC_BLOCK_H) toplev.h $(GGC_H) $(HASHTAB_H) \ $(TM_P_H) $(TARGET_H) $(TARGET_DEF_H) langhooks.h reload.h gt-rs6000.h \ - cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) + cfgloop.h $(OPTS_H) $(COMMON_TARGET_H) dumpfile.h rs6000-c.o: $(srcdir)/config/rs6000/rs6000-c.c \ $(srcdir)/config/rs6000/rs6000-protos.h \ -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH] Rs6000 infrastructure cleanup (switches), revised patch #2d
2012-10-01 Michael Meissner meiss...@linux.vnet.ibm.com * config/rs6000/rs6000.c (rs6000_option_override_internal): If -mcpu=xxx is not specified and the compiler is not configured using --with-cpu=xxx, use the bits from the TARGET_DEFAULT to set the initial options. I reworked the patch to allow TARGET_DEFAULT bits to be set if there is no -mcpu=xxx and the compiler was not configured using --with-cpu=xxx, so that we don't first clear all of the ISA bits, set them from the cpu, and then merge back in the TARGET_DEFAULT bits. Somebody asked about what is set, when this function gets called. The target_flags variable is set with the initial settings (TARGET_DEFAULT) and then all of the switches that the user sets or resets are then applied. The target_flags_explicit variable is only set if the user explicitly used that switch. So for instance, if the user passed -mpopcntb -mno-vsx on Linux 64-bit systems, target_flags would be 0x150001 (MASK_PPC_GFXOPT | MASK_POWERPC64 | MASK_64BIT | MASK_POPCNTB) and target_flags_explicit would be 0x201 (MASK_POPCNTB | MASK_VSX). Index: gcc/config/rs6000/rs6000.c === --- gcc/config/rs6000/rs6000.c (revision 191942) +++ gcc/config/rs6000/rs6000.c (working copy) @@ -2446,21 +2446,34 @@ rs6000_option_override_internal (bool gl rs6000_cpu_index = cpu_index = main_target_opt-x_rs6000_cpu_index; have_cpu = true; } + else if (implicit_cpu) +{ + rs6000_cpu_index = cpu_index = rs6000_cpu_name_lookup (implicit_cpu); + have_cpu = true; +} else { - const char *default_cpu = -(implicit_cpu ? implicit_cpu - : (TARGET_POWERPC64 ? powerpc64 : powerpc)); - + const char *default_cpu = (TARGET_POWERPC64 ? powerpc64 : powerpc); rs6000_cpu_index = cpu_index = rs6000_cpu_name_lookup (default_cpu); - have_cpu = implicit_cpu != 0; + have_cpu = false; } gcc_assert (cpu_index = 0); - target_flags = ~set_masks; - target_flags |= (processor_target_table[cpu_index].target_enable - set_masks); + /* If we have a cpu, either through an explicit -mcpu=xxx or if the + compiler was configured with --with-cpu=xxx, replace all of the ISA bits + with those from the cpu, except for options that were explicitly set. If + we don't have a cpu, do not override the target bits set in + TARGET_DEFAULT. */ + if (have_cpu) +{ + target_flags = ~set_masks; + target_flags |= (processor_target_table[cpu_index].target_enable + set_masks); +} + else +target_flags |= (processor_target_table[cpu_index].target_enable + ~target_flags_explicit); if (rs6000_tune_index = 0) tune_index = rs6000_tune_index; -- Michael Meissner, IBM 5 Technology Place Drive, M/S 2757, Westford, MA 01886-3141, USA meiss...@linux.vnet.ibm.com fax +1 (978) 399-6899
Re: [PATCH rs6000 testsuite] Fix a couple tests for VSX scalar instructions
On Mon, Oct 1, 2012 at 6:03 PM, Pat Haugen pthau...@linux.vnet.ibm.com wrote: This patch fixes a couple failures that occur if the testsuite is run with -mvsx and the VSX scalar sqrt instructions are generated. Ok for trunk? -Pat testsuite/ChangeLog: 2012-10-01 Pat Haugen pthau...@us.ibm.com * gcc.target/powerpc/pr46728-1.c: Accept xssqrtdp. * gcc.target/powerpc/pr46728-2.c: Likewise. LGTM. Thanks, David
Re: [PATCH] Fix test breakage, was: Add option for dumping to stderr (issue6190057)
Here is a patch to fix test breakage caused by r191883. Bootstrapped on x86_64 and tested with make -k check RUNTESTFLAGS=--target_board=unix/\{,-m32\}. Okay for trunk? Thanks, Sharad 2012-10-01 Sharad Singhai sing...@google.com * tree-vect-stmts.c (vectorizable_operation): Add missing return. testsuite/Changelog * gfortran.dg/vect/vect.exp: Change verbose vectorizor dump options to fix test failures caused by r191883. * gcc.dg/tree-ssa/gen-vect-11.c: Likewise. * gcc.dg/tree-ssa/gen-vect-2.c: Likewise. * gcc.dg/tree-ssa/gen-vect-32.c: Likewise. * gcc.dg/tree-ssa/gen-vect-25.c: Likewise. * gcc.dg/tree-ssa/gen-vect-11a.c: Likewise. * gcc.dg/tree-ssa/gen-vect-26.c: Likewise. * gcc.dg/tree-ssa/gen-vect-11b.c: Likewise. * gcc.dg/tree-ssa/gen-vect-11c.c: Likewise. * gcc.dg/tree-ssa/gen-vect-28.c: Likewise. * testsuite/gcc.target/i386/vect-double-1.c: Fix test. Missing entry from r191883. Index: testsuite/gfortran.dg/vect/vect.exp === --- testsuite/gfortran.dg/vect/vect.exp (revision 191883) +++ testsuite/gfortran.dg/vect/vect.exp (working copy) @@ -26,7 +26,7 @@ set DEFAULT_VECTCFLAGS # These flags are used for all targets. lappend DEFAULT_VECTCFLAGS -O2 -ftree-vectorize -fno-vect-cost-model \ - -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats + -fdump-tree-vect-details # If the target system supports vector instructions, the default action # for a test is 'run', otherwise it's 'compile'. Save current default. Index: testsuite/gcc.dg/tree-ssa/gen-vect-11.c === --- testsuite/gcc.dg/tree-ssa/gen-vect-11.c (revision 191883) +++ testsuite/gcc.dg/tree-ssa/gen-vect-11.c (working copy) @@ -1,6 +1,6 @@ /* { dg-do run { target vect_cmdline_needed } } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=3 -fwrapv -fdump-tree-vect-stats } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=3 -fwrapv -fdump-tree-vect-stats -mno-sse { target { i?86-*-* x86_64-*-* } } } */ +/* { dg-options -O2 -ftree-vectorize -fwrapv -fdump-tree-vect-details } */ +/* { dg-options -O2 -ftree-vectorize -fwrapv -fdump-tree-vect-details -mno-sse { target { i?86-*-* x86_64-*-* } } } */ #include stdlib.h Index: testsuite/gcc.dg/tree-ssa/gen-vect-2.c === --- testsuite/gcc.dg/tree-ssa/gen-vect-2.c (revision 191883) +++ testsuite/gcc.dg/tree-ssa/gen-vect-2.c (working copy) @@ -1,6 +1,6 @@ /* { dg-do run { target vect_cmdline_needed } } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats -mno-sse { target { i?86-*-* x86_64-*-* } } } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details -mno-sse { target { i?86-*-* x86_64-*-* } } } */ #include stdlib.h Index: testsuite/gcc.dg/tree-ssa/gen-vect-32.c === --- testsuite/gcc.dg/tree-ssa/gen-vect-32.c (revision 191883) +++ testsuite/gcc.dg/tree-ssa/gen-vect-32.c (working copy) @@ -1,6 +1,6 @@ /* { dg-do run { target vect_cmdline_needed } } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats -mno-sse { target { i?86-*-* x86_64-*-* } } } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details -mno-sse { target { i?86-*-* x86_64-*-* } } } */ #include stdlib.h Index: testsuite/gcc.dg/tree-ssa/gen-vect-25.c === --- testsuite/gcc.dg/tree-ssa/gen-vect-25.c (revision 191883) +++ testsuite/gcc.dg/tree-ssa/gen-vect-25.c (working copy) @@ -1,6 +1,6 @@ /* { dg-do run { target vect_cmdline_needed } } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=4 -fdump-tree-vect-stats -mno-sse { target { i?86-*-* x86_64-*-* } } } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details } */ +/* { dg-options -O2 -ftree-vectorize -fdump-tree-vect-details -mno-sse { target { i?86-*-* x86_64-*-* } } } */ #include stdlib.h Index: testsuite/gcc.dg/tree-ssa/gen-vect-11a.c === --- testsuite/gcc.dg/tree-ssa/gen-vect-11a.c (revision 191883) +++ testsuite/gcc.dg/tree-ssa/gen-vect-11a.c (working copy) @@ -1,6 +1,6 @@ /* { dg-do run { target vect_cmdline_needed } } */ -/* { dg-options -O2 -ftree-vectorize -ftree-vectorizer-verbose=3 -fdump-tree-vect-stats
[v3] update doxygen config
to 1.8.2, yay, fixes for template aliases in! tested x86/linux -benjamindiff --git a/libstdc++-v3/ChangeLog b/libstdc++-v3/ChangeLog index 9302b71..33a46cc 100644 --- a/libstdc++-v3/ChangeLog +++ b/libstdc++-v3/ChangeLog @@ -1,3 +1,7 @@ +2012-09-30 Benjamin Kosnik b...@redhat.com + + * doc/doxygen/user.cfg.in: Update to doxygen 1.8.2. + 2012-09-30 Jonathan Wakely jwakely@gmail.com * include/ext/ropeimpl.h (__uninitialized_fill_n_a): Fix using diff --git a/libstdc++-v3/doc/doxygen/user.cfg.in b/libstdc++-v3/doc/doxygen/user.cfg.in index b9fc4bf..3bf306f 100644 --- a/libstdc++-v3/doc/doxygen/user.cfg.in +++ b/libstdc++-v3/doc/doxygen/user.cfg.in @@ -1,4 +1,4 @@ -# Doxyfile 1.8.0 +# Doxyfile 1.8.2 # This file describes the settings to be used by the documentation system # doxygen (www.doxygen.org) for a project. @@ -126,7 +126,9 @@ FULL_PATH_NAMES= NO # only done if one of the specified strings matches the left-hand part of # the path. The tag can be used to show relative paths in the file list. # If left blank the directory from which doxygen is run is used as the -# path to strip. +# path to strip. Note that you specify absolute paths here, but also +# relative paths, which will be relative from the directory where doxygen is +# started. STRIP_FROM_PATH= @@ -231,14 +233,15 @@ OPTIMIZE_FOR_FORTRAN = NO OPTIMIZE_OUTPUT_VHDL = NO # Doxygen selects the parser to use depending on the extension of the files it -# parses. With this tag you can assign which parser to use for a given extension. -# Doxygen has a built-in mapping, but you can override or extend it using this -# tag. The format is ext=language, where ext is a file extension, and language -# is one of the parsers supported by doxygen: IDL, Java, Javascript, CSharp, C, -# C++, D, PHP, Objective-C, Python, Fortran, VHDL, C, C++. For instance to make -# doxygen treat .inc files as Fortran files (default is PHP), and .f files as C -# (default is Fortran), use: inc=Fortran f=C. Note that for custom extensions -# you also need to set FILE_PATTERNS otherwise the files are not read by doxygen. +# parses. With this tag you can assign which parser to use for a given +# extension. Doxygen has a built-in mapping, but you can override or extend it +# using this tag. The format is ext=language, where ext is a file extension, +# and language is one of the parsers supported by doxygen: IDL, Java, +# Javascript, CSharp, C, C++, D, PHP, Objective-C, Python, Fortran, VHDL, C, +# C++. For instance to make doxygen treat .inc files as Fortran files (default +# is PHP), and .f files as C (default is Fortran), use: inc=Fortran f=C. Note +# that for custom extensions you also need to set FILE_PATTERNS otherwise the +# files are not read by doxygen. EXTENSION_MAPPING = @@ -251,6 +254,13 @@ EXTENSION_MAPPING = MARKDOWN_SUPPORT = YES +# When enabled doxygen tries to link words that correspond to documented classes, +# or namespaces to their corresponding documentation. Such a link can be +# prevented in individual cases by by putting a % sign in front of the word or +# globally by setting AUTOLINK_SUPPORT to NO. + +AUTOLINK_SUPPORT = YES + # If you use STL classes (i.e. std::string, std::vector, etc.) but do not want # to include (a tag file for) the STL sources as input, then you should # set this tag to YES in order to let doxygen match functions declarations and @@ -271,12 +281,7 @@ CPP_CLI_SUPPORT= NO SIP_SUPPORT= NO -# For Microsoft's IDL there are propget and propput attributes to indicate getter -# and setter methods for a property. Setting this option to YES (the default) -# will make doxygen replace the get and set methods by a property in the -# documentation. This will only work if the methods are indeed getting or -# setting a simple type. If this is not the case, or you want to show the -# methods anyway, you should set this option to NO. +# For Microsoft's IDL there are propget and propput attributes to indicate getter and setter methods for a property. Setting this option to YES (the default) will make doxygen replace the get and set methods by a property in the documentation. This will only work if the methods are indeed getting or setting a simple type. If this is not the case, or you want to show the methods anyway, you should set this option to NO. IDL_PROPERTY_SUPPORT = NO @@ -300,7 +305,8 @@ SUBGROUPING= YES # @ingroup) instead of on a separate page (for HTML and Man pages) or # section (for LaTeX and RTF). -INLINE_GROUPED_CLASSES = NO +#INLINE_GROUPED_CLASSES = NO +INLINE_GROUPED_CLASSES = YES # When the INLINE_SIMPLE_STRUCTS tag is set to YES, structs, classes, and # unions with only public data fields will be shown inline in the documentation @@ -308,7 +314,7 @@ INLINE_GROUPED_CLASSES = NO # documentation), provided this scope is documented. If set to NO (the default), # structs, classes, and unions are shown on a
Re: profitable_hard_regs vs. PR 48435
On 12-09-30 2:42 PM, Richard Sandiford wrote: This is another patch needed for the MIPS MD_REGS change described here: http://gcc.gnu.org/ml/gcc-patches/2012-09/msg01992.html The profitable_hard_regs set used during IRA colouring used to be restricted to registers that are valid for the allocno's mode. That caused problems for multi-register modes that can only start with an even register (say), because profitable_hard_regs would only include the even start registers, not the pairing odd registers. Vlad fixed it with: 2011-04-08 Vladimir Makarov vmaka...@redhat.com PR inline-asm/48435 * ira-color.c (setup_profitable_hard_regs): Add comments. Don't take prohibited hard regs into account. (setup_conflict_profitable_regs): Rename to get_conflict_profitable_regs. (check_hard_reg_p): Check prohibited hard regs. However, one effect of that change is that if register R belongs to class CL but can never be used anywhere in a register of mode M, it will still be included in profitable_hard_regs. That's the case with MD_REGS and register HI on MIPS. The patch below is a half-way house between the original behaviour and the post-48435 one. It restricts profitable_hard_regs to registers that can be used for the allocno's mode, but doesn't restrict it to starting registers. Most of the ira.c change is reindentation, so I've included a -b diff as well. As with the patch linked above, I checked that this produced no difference in assembly output for a set of x86_64 gcc .ii files (tested with -O2 -march=native on gcc20). Also tested on x86_64-linux-gnu (including -m32) and mipsisa64-elf. OK to install? gcc/ * ira-int.h (target_ira_int): Add x_ira_useful_class_mode_regs. (ira_useful_class_mode_regs): New macro. * ira.c (clarify_prohibited_class_mode_regs): Set up ira_useful_class_mode_regs. * ira-color.c (setup_profitable_hard_regs): Use it to initialise profitable_hard_regs. Ok. thanks, Richard.
Various minor C++ PATCHes
1) I wanted a lang-specific flag for my thread_local work, so I took the one away from __PRETTY_FUNCTION__, which really doesn't need it. 2) I didn't end up wanting to use a flag in DECL_LANG_SPECIFIC, but it still makes sense to check for appropriate tree codes in DECL_FRIEND_P and DECL_ANTICIPATED. 3) DECL_NONTRIVIALLY_INITIALIZED_P currently means has a user-written initializer and it would be more useful to me if it actually meant has non-trivial initialization so that a non-trivial default constructor counts even if there is nothing written in the variable declaration. The one existing reader of this flag seems to agree with me. Tested x86_64-pc-linux-gnu, applying to trunk. commit 6c286ba0a085136d7efecd6d70d0cf96a4e8386a Author: Jason Merrill ja...@redhat.com Date: Wed Sep 26 12:14:15 2012 -0400 * cp-tree.h (DECL_PRETTY_FUNCTION_P): Just look at the name. * decl.c (cp_make_fname_decl): Adjust. diff --git a/gcc/c-family/c-common.h b/gcc/c-family/c-common.h index 9298e3d..cefe92d 100644 --- a/gcc/c-family/c-common.h +++ b/gcc/c-family/c-common.h @@ -45,7 +45,6 @@ never after. /* Usage of TREE_LANG_FLAG_?: 0: IDENTIFIER_MARKED (used by search routines). - DECL_PRETTY_FUNCTION_P (in VAR_DECL) C_MAYBE_CONST_EXPR_INT_OPERANDS (in C_MAYBE_CONST_EXPR, for C) 1: C_DECLARED_LABEL_FLAG (in LABEL_DECL) STATEMENT_LIST_STMT_EXPR (in STATEMENT_LIST) diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h index f437022..cee8590 100644 --- a/gcc/cp/cp-tree.h +++ b/gcc/cp/cp-tree.h @@ -56,7 +56,6 @@ c-common.h, not after. AGGR_INIT_VIA_CTOR_P (in AGGR_INIT_EXPR) PTRMEM_OK_P (in ADDR_EXPR, OFFSET_REF, SCOPE_REF) PAREN_STRING_LITERAL (in STRING_CST) - DECL_PRETTY_FUNCTION_P (in VAR_DECL) KOENIG_LOOKUP_P (in CALL_EXPR) STATEMENT_LIST_NO_SCOPE (in STATEMENT_LIST). EXPR_STMT_STMT_EXPR_RESULT (in EXPR_STMT) @@ -2410,7 +2409,8 @@ struct GTY((variable_size)) lang_decl { /* Nonzero if this DECL is the __PRETTY_FUNCTION__ variable in a template function. */ #define DECL_PRETTY_FUNCTION_P(NODE) \ - (TREE_LANG_FLAG_0 (VAR_DECL_CHECK (NODE))) + (DECL_NAME (NODE) \ +!strcmp (IDENTIFIER_POINTER (DECL_NAME (NODE)), __PRETTY_FUNCTION__)) /* The _TYPE context in which this _DECL appears. This field holds the class where a virtual function instance is actually defined. */ diff --git a/gcc/cp/decl.c b/gcc/cp/decl.c index 078b148..d4c78e1 100644 --- a/gcc/cp/decl.c +++ b/gcc/cp/decl.c @@ -3829,7 +3829,6 @@ cp_make_fname_decl (location_t loc, tree id, int type_dep) /* As we're using pushdecl_with_scope, we must set the context. */ DECL_CONTEXT (decl) = current_function_decl; - DECL_PRETTY_FUNCTION_P (decl) = type_dep; TREE_STATIC (decl) = 1; TREE_READONLY (decl) = 1; commit 21960ac8d1debe452eef3776fe2d638fa55dbe52 Author: Jason Merrill ja...@redhat.com Date: Wed Sep 26 12:19:21 2012 -0400 * cp-tree.h (TYPE_FUNCTION_OR_TEMPLATE_DECL_CHECK): New. (DECL_FRIEND_P, DECL_ANTICIPATED): Use it. (TYPE_FUNCTION_OR_TEMPLATE_DECL_P): New. * name-lookup.c (hidden_name_p): Use it. diff --git a/gcc/cp/cp-tree.h b/gcc/cp/cp-tree.h index cee8590..e4f3761 100644 --- a/gcc/cp/cp-tree.h +++ b/gcc/cp/cp-tree.h @@ -202,6 +202,13 @@ c-common.h, not after. #define VAR_OR_FUNCTION_DECL_CHECK(NODE) \ TREE_CHECK2(NODE,VAR_DECL,FUNCTION_DECL) +#define TYPE_FUNCTION_OR_TEMPLATE_DECL_CHECK(NODE) \ + TREE_CHECK3(NODE,TYPE_DECL,TEMPLATE_DECL,FUNCTION_DECL) + +#define TYPE_FUNCTION_OR_TEMPLATE_DECL_P(NODE) \ + (TREE_CODE (NODE) == TYPE_DECL || TREE_CODE (NODE) == TEMPLATE_DECL \ + || TREE_CODE (NODE) == FUNCTION_DECL) + #define VAR_FUNCTION_OR_PARM_DECL_CHECK(NODE) \ TREE_CHECK3(NODE,VAR_DECL,FUNCTION_DECL,PARM_DECL) @@ -1875,8 +1882,8 @@ struct GTY(()) lang_decl_base { unsigned initialized_in_class : 1; /* var or fn */ unsigned repo_available_p : 1; /* var or fn */ unsigned threadprivate_or_deleted_p : 1; /* var or fn */ - unsigned anticipated_p : 1; /* fn or type */ - unsigned friend_attr : 1; /* fn or type */ + unsigned anticipated_p : 1; /* fn, type or template */ + unsigned friend_attr : 1; /* fn, type or template */ unsigned template_conv_p : 1; /* var or template */ unsigned odr_used : 1; /* var or fn */ unsigned u2sel : 1; @@ -2293,7 +2300,9 @@ struct GTY((variable_size)) lang_decl { /* Nonzero for DECL means that this decl is just a friend declaration, and should not be added to the list of members for this class. */ -#define DECL_FRIEND_P(NODE) (DECL_LANG_SPECIFIC (NODE)-u.base.friend_attr) +#define DECL_FRIEND_P(NODE) \ + (DECL_LANG_SPECIFIC (TYPE_FUNCTION_OR_TEMPLATE_DECL_CHECK (NODE)) \ + -u.base.friend_attr) /* A TREE_LIST of the types which have befriended this FUNCTION_DECL. */ #define DECL_BEFRIENDING_CLASSES(NODE) \ @@ -3101,7 +3110,8 @@ more_aggr_init_expr_args_p (const aggr_init_expr_arg_iterator *iter)
Re: [patch][lra] a few bitmap obstacks for lra-assigns
On 12-10-01 6:02 PM, Steven Bosscher wrote: Hello, This eliminates a few large loops in lra-assigns.c. They're not the most costly loops but the life times of the bitmaps is well-defined and destroying a bitmap obstack is much cheaper than looping over all bitmaps calling bitmap_clear. The saving is small but you have to start somewhere... Bootstrapped lra-branch and tested on x86_64-unknown-linux-gnu, along with the patch from earlier today. OK for the branch? Ok. Thanks, Steven.
Re: [v3] libstdc++/54314
adds in exports for more vtable bits, as per bugzilla missed some, added here and tested with: tested x86/linux --enable-symvers=gnu-versioned-namespace -benjamin2012-10-01 Benjamin Kosnik b...@redhat.com * config/abi/pre/gnu-versioned-namespace.ver: Add more typeinfo/vtable exports. diff --git a/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver b/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver index 5f25374..84210e4 100644 --- a/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver +++ b/libstdc++-v3/config/abi/pre/gnu-versioned-namespace.ver @@ -62,15 +62,18 @@ GLIBCXX_7.0 { _ZTTNSt*; # vtable +_ZTVSt*; _ZTVNSt*; # thunk _ZTv0_n24_NS*; # typeinfo structure +_ZTISt*; _ZTINSt*; # typeinfo name +_ZTSSt*; _ZTSNSt*; # locale
Re: PR 53889: Add __gthread_recursive_mutex_destroy
On 1 October 2012 20:22, Ian Lance Taylor wrote: On Sun, Sep 30, 2012 at 11:41 AM, Jonathan Wakely jwakely@gmail.com wrote: There is no __gthread_recursive_mutex_destroy function in the gthreads API. Trying to use __gthread_mutex_destroy fails to compile on platforms where the mutex types are different. To avoid resource leaks libstdc++ needs to hack around the missing function with overloaded functions and SFINAE tricks to detect how a recursive mutex can be destroyed. This patch extends the gthreads API to include __gthread_recursive_mutex_destroy, defining it for each gthread model, and removing the hacks from libstdc++. +return rtems_gxx_mutex_destroy( __mutex ); Space before '(', not space after. Oops, yes. Doing anything else here is going to be painful, but this assumes that RTEMS uses the same representation for non-recursive and recursive mutexes. That is currently true, but it deserves a comment. Good point ... I have a vague recollection that I looked into it and decided they were the same, but I wrote this patch back in July and don't remember the details. If anyone knows better please let me know. The attached revised patch adds a comment. --- a/libgcc/config/i386/gthr-win32.h +++ b/libgcc/config/i386/gthr-win32.h +static inline void +__gthread_recursive_mutex_destroy (__gthread_recursive_mutex_t *mutex) +{ + __gthread_mutex_t __mutex2; + __mutex2.sema = mutex-sema; + __gthr_win32_mutex_destroy (__mutex2); +} I think it would be better to put this in libgcc/config/i386/gthr-win32.c, like the other functions. Then you can just call CloseHandle. Done. I've also made __gthread_recursive_mutex_destroy return int, unlike the Win32 __ghtread_mutex_destroy (see PR 53888 for that) --- a/libgcc/config/mips/gthr-mipssde.h +++ b/libgcc/config/mips/gthr-mipssde.h +static inline int +__gthread_recursive_mutex_destroy (__gthread_recursive_mutex_t *__mutex) +{ + return __gthread_mutex_destroy(__mutex); +} Will this even compile? It doesn't look like it. Nope, you're right. I've replaced it with this, which is all __ghtread_mutex_destroy does anyway: static inline int __gthread_recursive_mutex_destroy (__gthread_recursive_mutex_t * UNUSED(__mutex)) { return 0; } Is that indentation right? (the asterisk is in the same column as the parameter type in a fixed-width font.) Thanks for the careful review. The ChangeLog entry is the same, I haven't rested this because the changes since the first patch are only to targets I don't have access to. libgcc: PR other/53889 * gthr.h (__gthread_recursive_mutex_destroy): Document new required function. * gthr-posix.h (__gthread_recursive_mutex_destroy): Define. * gthr-single.h (__gthread_recursive_mutex_destroy): Likewise. * config/gthr-rtems.h (__gthread_recursive_mutex_destroy): Likewise. * config/gthr-vxworks.h (__gthread_recursive_mutex_destroy): Likewise. * config/i386/gthr-win32.h (__gthread_recursive_mutex_destroy): Likewise. * config/mips/gthr-mipssde.h (__gthread_recursive_mutex_destroy): Likewise. * config/pa/gthr-dce.h (__gthread_recursive_mutex_destroy): Likewise. * config/s390/gthr-tpf.h (__gthread_recursive_mutex_destroy): Likewise. libstdc++-v3: PR other/53889 * include/std/mutex (__recursive_mutex_base::~__recursive_mutex_base): Use __gthread_recursive_mutex_destroy. (__recursive_mutex_base::_S_destroy): Remove. (__recursive_mutex_base::_S_destroy_win32): Likewise. * include/ext/concurrence.h (__recursive_mutex::~__recursive_mutex): Use __gthread_recursive_mutex_destroy. (__recursive_mutex::_S_destroy): Remove. (__recursive_mutex::_S_destroy_win32): Likewise. commit 7877201b61694279811c0276c92d85aaec2b30a2 Author: Jonathan Wakely jwakely@gmail.com Date: Tue Oct 2 01:40:46 2012 +0100 libgcc: PR other/53889 * gthr.h (__gthread_recursive_mutex_destroy): Document new required function. * gthr-posix.h (__gthread_recursive_mutex_destroy): Define. * gthr-single.h (__gthread_recursive_mutex_destroy): Likewise. * config/gthr-rtems.h (__gthread_recursive_mutex_destroy): Likewise. * config/gthr-vxworks.h (__gthread_recursive_mutex_destroy): Likewise. * config/i386/gthr-win32.h (__gthread_recursive_mutex_destroy): Likewise. * config/mips/gthr-mipssde.h (__gthread_recursive_mutex_destroy): Likewise. * config/pa/gthr-dce.h (__gthread_recursive_mutex_destroy): Likewise. * config/s390/gthr-tpf.h (__gthread_recursive_mutex_destroy): Likewise. libstdc++-v3: PR other/53889 * include/std/mutex (__recursive_mutex_base::~__recursive_mutex_base): Use __gthread_recursive_mutex_destroy. (__recursive_mutex_base::_S_destroy): Remove.
Re: RFC: LRA for x86/x86-64 [0/9]
On 12-10-01 4:24 PM, Steven Bosscher wrote: On Mon, Oct 1, 2012 at 9:51 PM, Vladimir Makarov vmaka...@redhat.com wrote: I think it's more important in this case to recognize Steven's real point, which is that for an identical situation (IRA), and with an identical patch author, we had similar bugs. They were promised to be worked on, and yet some of those regressions are still very much with us. That is not true. I worked on many compiler time regression bugs. I remeber one serious degradation of compilation time on all_cp2k_gfortran.f90. I solved the problem and make IRA working faster and generating much better code than the old RA. http://blog.gmane.org/gmane.comp.gcc.patches/month=20080501/page=15 About other two mentioned PRs by Steven: PR26854. I worked on this bug even when IRA was on the branch and make again GCC with IRA 5% faster on this test than GCC with the old RA. PR 54146 is 3 months old. There were a lot work on other optimizations before IRA became important. It happens only 2 months ago. I had no time to work on it but I am going to. This is also not quite true, see PR37448, which shows the problems as the test case for PR54146. I worked on this PR too and did some progress but it was much less than on other PRs. Sometimes I think that I should have maintained the old RA to compare with it and show that IRA is still a step ahead. Unfortunately, comparison with old releases has no sense now because more aggressive optimizations (inlining). I just think scalability is a very important issue. If some pass or algorithm scales bad on some measure, then users _will_ run into that at some point and report bugs about it (if you're lucky enough to have a user patient enough to sit out the long compile time :-) ). Also, good scalability opens up opportunities. For example, historically GCC has been conservative on inlining heuristics to avoid compile time explosions. I think it's better to address the causes of that explosion and to avoid introducing new potential bottlenecks. As I wrote, scalability sometimes misleading as in case PR (a Scheme interpreter) where 30% more compilation time with LRA is translated into 15% decrease in code size and most probably better performance. Now we can manage to achieve scalability with worse performance but that is not what user expects and even worse he did know it. It is even a bigger problem IMO. Ideally, we should have more scalable algorithms as fallback but we should warn the programmer that worse performance algorithms are used and he could achieve a better performance by dividing a huge function into several ones. People sometimes see that RA takes a lot of compilation time but it is in the nature of RA. I'd recommend first to check how the old RA behaves and then call it a degradation. There's no question that RA is one of the hardest problems the compiler has to solve, being NP-complete and all that. I like LRA's iterative approach, but if you know you're going to solve a hard problem with a number potentially expensive iterations, there's even more reason to make scalability a design goal! When I designed IRA I kept this goal in my mind too. Although my first priority was the performance. The old RA was ok when the register pressure was not high. I knew aggressive inlining and LTO were coming. And my goal was to design RA generating a good code for bigger programs with much higher register pressure where the old RA drawbacks would be obvious. As I said earlier in this thread, I was really looking forward to IRA at the time you worked on it, because it is supposed to be a regional allocator and I had expected that to mean it could, well, allocate per-region which is usually very helpful for scalability (partition your function and insert compensation code on strategically picked region boundaries). But that's not what IRA has turned out to be. (Instead, its regional nature is one of the reasons for its scalability problems.) IRA is certainly not worse than old global.c in very many ways, and LRA looks like a well thought-through and welcome replacement of old reload. But scalability is an issue in the design of IRA and LRA looks to be the same in that regard. Regional allocator means that allocation is made taking mostly a region into account to achieve a better allocation. But a good regional RA (e.g. Callahan-Koblenz or Intel fusion based RA) takes interactions with other regions too and such allocators might be even slower than non-regional. Although I should say that in my impressions CK-allocator (I tried and implemented this about 6 years ago) is probably better scalable than IRA but it generates worse code.
Re: RFC: LRA for x86/x86-64 [0/9]
On 10/01/2012 09:03 AM, Steven Bosscher wrote: On Sat, Sep 29, 2012 at 10:26 PM, Steven Bosscher stevenb@gmail.com wrote: LRA create live ranges : 175.30 (15%) usr 2.14 (13%) sys 177.44 (15%) wall2761 kB ( 0%) ggc I've tried to split this up a bit more: process_bb_lives ~50% create_start_finish_chains ~25% remove_some_program_points_and_update_live_ranges ~25% The latter two have a common structure with loops that look like this: for (i = FIRST_PSEUDO_REGISTER; i max_regno; i++) { for (r = lra_reg_info[i].live_ranges; r != NULL; r = r-next) Perhaps it's possible to do some of the work of compress_live_ranges during process_bb_lives, to create shorter live_ranges chains. It is aready done. Moreover program points are compressed to minimal to guarantee right conflict resolution. For example, if only 2 pseudos live in 1..10 and 2..14, they actually will have the same range like 1..1. Analogous live ranges are used in IRA as intermidiate step to build a conflict graph. Actually, the first approach was to use IRA code to assign hard registers to pseudos (e.g. Jeff Law tried this approach) but it was rejected as requiring much more compilation time. In some way, one can look at the assignment in LRA is a compromise between quality (which could achieved through repeated buidling conflict graphs and using graph coloring) and compilation speed. Also, maybe doing something else than a linked list of live_ranges will help (unfortunately that's not a trivial change, it seems, because the lra_live_range_t type is used everywhere and there are no iterators or anything abstracted out like that -- just chain walks...). Still it does seem to me that a sorted VEC of lra_live_range objects probably would speed things up. Question is of course how much... :-) My experience shows that these lists are usually 1-2 elements. Although in this case, there are pseudos with huge number elements (hundreeds). I tried -fweb for this tests because it can decrease the number elements but GCC (I don't know what pass) scales even worse: after 20 min of waiting and when virt memory achieved 20GB I stoped it. I guess the same speed or close to reload one on this test can be achieved through implementing other simpler algorithms generating worse code. I need some time to design them and implement this.
Re: [PATCH v2, rtl-optimization]: Fix PR54457, [x32] Fail to combine 64bit index + constant
On Thu, Sep 27, 2012 at 11:13 AM, Uros Bizjak ubiz...@gmail.com wrote: On Thu, Sep 27, 2012 at 8:08 PM, paul_kon...@dell.com wrote: I agree (subreg:M (op:N A C) 0) to (op:M (subreg:N (A 0)) C) is a good transformation, but why do we need to handle as special the case where the subreg is itself the operand of a plus or minus? I think it should happen regardless of where the subreg occurs. Don't we need to restrict this to the low part though? ... After some off-line discussion with Richard, attached is v2 of the patch. 2012-09-27 Uros Bizjak ubiz...@gmail.com PR rtl-optimization/54457 * simplify-rtx.c (simplify_subreg): Simplify (subreg:SI (op:DI ((x:DI) (y:DI)), 0) to (op:SI (subreg:SI (x:DI) 0) (subreg:SI (x:DI) 0)). ... Is it just specific to DI - SI, or is it for any large mode - smaller mode, like SI - HI? Oh, I just copied v1 ChangeLog. The patch converts all modes where size of mode M size of mode N. Updated ChangeLog reads: 2012-09-27 Uros Bizjak ubiz...@gmail.com PR rtl-optimization/54457 * simplify-rtx.c (simplify_subreg): Simplify (subreg:M (op:N ((x:N) (y:N)), 0) to (op:M (subreg:M (x:N) 0) (subreg:M (x:N) 0)), where the outer subreg is effectively a truncation to the original mode M. When I was doing something similar on our internal toolchain at Cavium. I found doing this caused a regression on MIPS64 n32 in gcc.c-torture/execute/20040709-1.c Where: (insn 15 14 16 2 (set (reg/v:DI 200 [ y ]) (reg:DI 2 $2)) t.c:16 301 {*movdi_64bit} (expr_list:REG_DEAD (reg:DI 2 $2) (nil))) (insn 16 15 17 2 (set (reg:DI 210) (zero_extract:DI (reg/v:DI 200 [ y ]) (const_int 29 [0x1d]) (const_int 0 [0]))) t.c:16 249 {extzvdi} (expr_list:REG_DEAD (reg/v:DI 200 [ y ]) (nil))) (insn 17 16 23 2 (set (reg:SI 211) (truncate:SI (reg:DI 210))) t.c:16 175 {truncdisi2} (expr_list:REG_DEAD (reg:DI 210) (nil))) Gets converted to: (insn 23 17 26 2 (set (reg/i:SI 2 $2) (and:SI (reg:SI 2 $2 [+4 ]) (const_int 536870911 [0x1fff]))) t.c:18 156 {*andsi3} (nil)) Which is considered an ext instruction And with the Octeon simulator which causes undefined arguments to 32bit word operations to come out as 0xDEADBEEF which showed the regression. I fixed it by changing it to produce TRUNCATE instead of the subreg. I did the simplification on ior/and rather than plus/minus/mult so the issue is only when expanding to this to and/ior. Thanks, Andrew Pinski testsuite/ChangeLog: 2012-09-27 Uros Bizjak ubiz...@gmail.com PR rtl-optimization/54457 * gcc.target/i386/pr54457.c: New test. Uros.
Re: RFC: LRA for x86/x86-64 [0/9]
On 10/01/2012 07:14 PM, Vladimir Makarov wrote: Analogous live ranges are used in IRA as intermidiate step to build a conflict graph. Actually, the first approach was to use IRA code to assign hard registers to pseudos (e.g. Jeff Law tried this approach) but it was rejected as requiring much more compilation time. In some way, one can look at the assignment in LRA is a compromise between quality (which could achieved through repeated buidling conflict graphs and using graph coloring) and compilation speed. Not only was it slow (iterating IRA), guaranteeing termination was a major problem. There's some algorithmic games that have to be played (they're at least discussed in literature, but not under the heading of termination) and there's some issues specific to the IRA implementation which make ensuring termination difficult. I got nearly as good of results by conservative updates of the conflicts after splitting ranges and (ab)using ira's reload hooks to give the new pseudos for the split range a chance to be allocated again. The biggest problem with that approach was getting the costing right for the new pseudos. That requires running a fair amount of IRA a second time. I'd still like to return to some of the ideas from that work as I think some of the bits are still relevant in the IRA+LRA world. My experience shows that these lists are usually 1-2 elements. That's been my experience as well. The vast majority of the time the range lists are very small. Jeff
Re: RFC: LRA for x86/x86-64 [9/9]
On 09/27/2012 04:59 PM, Vladimir Makarov wrote: This is the last patch switching on LRA for x86/x86-64. The patch also contains code deciding when to use spilling general regs into SSE instead of memory. 2012-09-27 Vladimir Makarov vmaka...@redhat.com * config/i386/i386.h (enum ix86_tune_indices): Add X86_TUNE_GENERAL_REGS_SSE_SPILL. (TARGET_GENERAL_REGS_SSE_SPILL): New macro. * config/i386/i386.c (initial_ix86_tune_features): Set up X86_TUNE_GENERAL_REGS_SSE_SPILL for m_COREI7 and m_CORE2I7. (ix86_lra_p, ix86_register_bank): New functions. (ix86_secondary_reload): Add NON_Q_REGS, SIREG, DIREG. (inline_secondary_memory_needed): Change assert. (ix86_spill_class, ix86_spill_class_mode): New function. (TARGET_LRA_P, TARGET_REGISTER_BANK, TARGET_SPILL_CLASS): New macros. (TARGET_SPILL_CLASS_MODE): New macro. So for the register_bank stuff, aren't we really just defining some kind of alternate costing model? ie, we can't really get the costs we want from register classes, but it's dependent on the precise register used. I guess I'm rethinking if register bank is the right name. Otherwise it seems reasonable. Looks like there's an extra newline at EOF. jeff
Re: RFC: LRA for x86/x86-64 [8/9]
On 09/27/2012 04:59 PM, Vladimir Makarov wrote: The following patch adds a code neccessary for correct work of LRA (function ira_setup_eliminable_regset) and for correct work of the compiler when LRA is used (see file dwarf2out.c). 2012-09-27 Vladimir Makarov vmaka...@redhat.com * loop-invariant.c (calculate_loop_reg_pressure): Pass new argument to ira_setup_eliminable_regset. * haifa-sched.c (sched_init): Pass new argument to ira_setup_eliminable_regset. * dwarf2out.c: Include lra.h. (based_loc_descr, compute_frame_pointer_to_fb_displacement): Use lra_eliminate_regs for LRA instead of eliminate_regs. * ira.c: (ira_setup_eliminable_regset): Add parameter. Remove need_fp. Call lra_init_elemination and mark HARD_FRAME_POINTER_REGNUM as living forever if frame_pointer_needed. (ira): Call ira_setup_eliminable_regset with a new argument. * ira.h (ira_setup_eliminable_regset): Add an argument. * Makefile.in (dwarf2out.o): Add dependence on ira.h and lra.h. This is OK. Obviously it's useless without 7a/7b. But I just wanted to go ahead and review the ancillary bits before going to the meat of the submission. So it's just the 7a/7b patch that needs review, right? jeff