Middle-end and optimization regressions: what should we do?

Middle-end and optimization regressions: what should we do?

[François-Xavier Coudert]

Hi all,

PR 22619 and PR 22509 are two examples of recent 4.1 regressions that
showed up in gfortran, due to middle-end or optimization bugs (only
happen at -O3). Since these are regressions, they should be treated
before a long time passes, but since both source codes are Fortran, I
guess people don't (and won't) want to look at them.

How can we help here? Is there a way to make gfortran output a
complete GIMPLE tree, that could be used for middle-end hackers to
determine where the problem is? Or are we doomed to a dichotomy to
know which patch caused these regressions?

FX

PS: PR 22619 appeared somewhere between 20050716 and 20050717, so
patches that could possible have messed up are:

005-07-17  Daniel Berlin  <dberlin@dberlin.org>

        Fix PR tree-optimization/22531
        * tree-ssa-pre.c (do_eustores):  Make sure LHS is a decl for the
         moment.

2005-07-17  Daniel Berlin  <dberlin@dberlin.org>

        * tree-promote-statics.c (pass_promote_statics): Change dump file
         name.

2005-07-17  Daniel Berlin  <dberlin@dberlin.org>

        * tree-optimize.c (init_tree_optimization_passes): Add
        pass_eliminate_useless_stores pass.
        * tree-pass.h (pass_eliminate_useless_stores): New pass structure.
        * tree-ssa-pre.c (is_copy_stmt): New function.
        (follow_copies_till_vuse): Ditto. 
        (do_eustores): Ditto.
        (gate_eustores): Ditto. 
        
2005-07-16  Richard Henderson  <rth@redhat.com>

        * gcc.c (MFWRAP_SPEC): Don't wrap pthread_join or pthread_exit.

2005-07-16 Danny Berlin <dberlin@dberlin.org>
           Kenneth Zadeck <zadeck@naturalbridge.com>

        * Makefile.in: Added rules for ipa-pure-const.c, ipa-reference.c,
        ipa-reference.h, ipa-utils.c, ipa-utils.h, ipa-type-escape.c,
        ipa-type-escape.h, tree-promote-statics.c
        * ipa-pure-const.c, ipa-reference.c, ipa-reference.h, ipa-utils.c,
        ipa-utils.h, ipa-type-escape.c, ipa-type-escape.h, 
        tree-promote-statics.c: new files.
        * alias.c: (nonlocal_mentioned_p_1, nonlocal_mentioned_p,
        nonlocal_referenced_p_1, nonlocal_referenced_p, nonlocal_set_p_1,
        int nonlocal_set_p, mark_constant_function): Deleted.
        (rest_of_handle_cfg): Removed call to mark_constant_function.
        (nonoverlapping_component_refs_p): Added calls to support
        type based aliasing.
        * tree-ssa-alias.c (may_alias_p,
        compute_flow_insensitive_aliasing): Ditto.
        * calls.c (flags_from_decl_or_type): Removed reference to
        cgraph_rtl_info.
        * c-typeck.c (convert_arguments): Make builtins tolerant of having
        too many arguments.  This is necessary for Spec 2000.
        * cgraph.h (const_function, pure_function): Removed.
        * common.opt: Added "fipa-pure-const", "fipa-reference",
        "fipa-type-escape", and "ftree-promote-static".
        * opts.c: Ditto.
        * passes.c: Added ipa and tree-promote-statics passes.
        * timevar.def: Added TV_IPA_PURE_CONST, TV_IPA_REFERENCE,
        TV_IPA_TYPE_ESCAPE, and TV_PROMOTE_STATICS.
        * tree-dfa.c (referenced_var_lookup_if_exists): New function.
        * tree-flow.h: Added exposed sra calls and addition of
        reference_vars_info field for FUNCTION_DECLS.
        * tree-pass.h: Added passes.
        * tree-sra.c: (sra_init_cache): New function.
        (sra_insert_before, sra_insert_after) Made public.
        (type_can_be_decomposed_p): Renamed from type_can_be_decomposed_p
        and made public.
        * tree-ssa-alias.c (dump_alias_stats): Added stats for type based
        aliasing. (may_alias_p): Added code to use type escape analysis to
        improve alias sets.
        * tree-ssa-operands.c (add_call_clobber_ops): Added parameter and
        code to prune clobbers of static variables based on information
        produced in ipa-reference pass.  Changed call clobbering so that
        statics are not marked as clobbered if the call does not clobber
        them.

2005-07-16  Daniel Berlin  <dberlin@dberlin.org>

        * tree-ssa-structalias.c (need_to_solve): Need to check for preds,
        too.

2005-07-16  Eric Botcazou <ebotcazou@libertysurf.fr>

        * doc/install.texi (*-*-solaris2*): Document recommended version
        of GNU binutils and mention GNU linker problem on Solaris 10.

2005-07-16  Joseph S. Myers  <joseph@codesourcery.com>

        PR c/22421
        * c-decl.c (c_build_bitfield_integer_type): New function.
        (finish_struct): Call it.
        * c-pretty-print.c (pp_c_type_specifier): Handle bit-field types.

2005-07-16  Kaveh R. Ghazi  <ghazi@caip.rutgers.edu>

        * c-typeck.c (digest_init): Call 'convert_for_assignment'
        before returning.

2005-07-16  Jan Hubicka  <jh@suse.cz>

        * cfg.c (update_bb_profile_for_threading): Fix profile updating.
        (scale_bbs_frequencies_int): Watch roundoff errors.
        * predict.c (return_prediction): Initialize return_stmt.

2005-07-16  Jan Hubicka  <jh@suse.cz>

        * profile.c (rest_of_handle_branch_prob): Fix handling of estimation
        after RTL profiling.


PR 22509 appeared between 20050713 and 20050714, so possible guilty patches are:

2005-07-14  David Edelsohn  <edelsohn@gnu.org>

        * config/rs6000/rs6000.md (UNSPEC_SYNC, UNSPEC_LWSYNC,
        UNSPEC_ISYNC, UNSPEC_SYNC_OP, UNSPEC_ATOMIC, UNSPEC_CMPXCHG,
        UNSPEC_XCHG, UNSPEC_AND): New.
        (UNSPECV_ATOMIC, UNSPECV_SYNC, UNSPECV_SYNC_OP, UNSPECV_CMPXCHG,
        UNSPECV_LWSYNC, UNSPECV_ISYNC): Delete.
        * config/rs6000/sync.md (FETCHOP): New code macro.
        (fetchop_name, fetchop_pred, fetchopsi_constr, fetchopdi_constr):
        New code attrs.
        (memory_barrier, sync_internal): Use unspec instead of unspec_volatile.
        (sync_compare_and_swap<mode>): Same.
        (sync_lock_test_and_set<mode>): Same.
        (sync_<fetchop><mode>, sync_nand<mode>): Only use rs6000_emit_sync
        for QImode and HImode, and not PPC405.
        (sync_old_<fetchop><mode>, sync_old_nand<mode>): Same.
        (sync_new_<fetchop><mode>, sync_new_nand<mode>): Same.
        (sync_<fetchop>{si,di}_internal): New.
        (sync_nand{si,di}_internal): New.
        (sync_old_<fetchop>{si,di}_internal): New.
        (sync_old_nand{si,di}_internal): New
        (sync_new_<fetchop>{si,di}_internal): New.
        (sync_new_nand{si,di}_internal): New.
        (atomic_and{si,di}): New.
        (sync_new_nand{si,di}_internal): New.
        (atomic_and{si,di}): New.
        (sync_add<mode>_internal): Delete.
        (sync_addshort_internal): Use unspec instead of unspec_volatile.
        (sync_sub<mode>_internal): Delte.
        (sync_subshort_internal): New.
        (sync_andsi_internal): Use unspec instead of unspec_volatile.
        (sync_anddi_internal): Delete.
        (sync_boolsi_internal): Use unspec instead of unspec_volatile.
        (sync_booldi_internal): Delete.
        (sync_boolc<mode>_internal): Delete.
        (sync_boolcshort_internal): Use unspec instead of unspec_volatile.
        (sync_boolc<mode>_internal2): Delete.
        (sync_boolcc<mode>_internal): Delete.
        (isync, lwsync): Use unspec instead of unspec_volatile.
        * config/rs6000/rs6000.c (rs6000_emit_sync): Implement MINUS.
        Revert UNSPEC_VOLATILE.
        (rs6000_split_atomic_op): New.
        * config/rs6000/rs6000-protos.h (rs6000_split_atomic_op): Declare.

2005-07-14  Eric Christopher  <echristo@redhat.com>

        * config/mips/mips.c (mips_canonicalize_comparison): Cast
        argument of trunc_int_for_mode to unsigned HOST_WIDE_INT.

2005-07-14  Eric Christopher  <echristo@redhat.com>

        * config/s390/t-tpf (SHLIB_MAPFILES): Remove.

2005-07-14  Steven Bosscher  <stevenb@suse.de>

        PR tree-optimization/22230
        * tree-vrp.c (extract_range_from_binary_expr): Fix logics thinko in
        the computation of the four cross productions for "range op range".

2005-07-14  Alexandre Oliva  <aoliva@redhat.com>
            Ulrich Weigand  <uweigand@de.ibm.com>

        PR target/20126
        * loop.c (loop_givs_rescan): Do not ICE if unable to reduce an IV
        in some insn.

2005-07-14  Ulrich Weigand  <uweigand@de.ibm.com>

        * config/s390/s390.h (TARGET_TPF_PROFILING): Add default definition.

2005-07-14  Steve Ellcey  <sje@cup.hp.com>

        * config/ia64/ia64.c (ia64_output_dwarf_dtprel): Support ILP32 mode.

2005-07-14  Richard Guenther  <rguenther@suse.de>

        PR middle-end/22347
        * config/i386/i386-protos.h (ix86_function_value): Change
        prototype to match new target hook.
        * config/i386/i386.c (ix86_value_regno): Change prototype
        to take extra type argument.
        (TARGET_FUNCTION_VALUE): Define.
        (ix86_function_ok_for_sibcall): Pass extra argument to
        ix86_value_regno, check return slot rtx for exact match.
        (ix86_function_value): Take extra parameter.  Dispatch to
        ix86_value_regno with fndecl/fntype as provided.
        (ix86_value_regno): Handle extra type argument.
        * config/i386/i386.h (FUNCTION_VALUE): No longer define.

        * testsuite/gcc.target/i386/sseregparm-3.c: New testcase.
        * testsuite/gcc.target/i386/sseregparm-4.c: New testcase.
        * testsuite/gcc.target/i386/sseregparm-5.c: New testcase.
        * testsuite/gcc.target/i386/sseregparm-6.c: New testcase.
        * testsuite/gcc.target/i386/sseregparm-7.c: New testcase.

2005-07-14  Richard Guenther  <rguenther@suse.de>

        * Makefile.in (explow.o, reg-stack.o): Depend on target.h.
        * calls.c (expand_call): Pass fntype to hard_function_value.
        (emit_library_call_value_1): Likewise.
        * explow.c: Include target.h.
        (hard_function_value): Take extra argument, the fntype.
        Use new target hook for function_value.
        * expr.h (hard_function_value): Change prototype.
        * function.c (aggregate_value_p): Pass 0 as fntype to
        hard_function_value.
        (assign_parms): Use new target hook for function_value.
        Pass 0 as fntype to hard_function_value.
        (expand_function_end): Likewise.
        * reg-stack.c: Include target.h.
        (stack_result): Use new target hook for function_value.
        * target-def.h: New target hook function_value.
        * target.h: Likewise.
        * targhooks.c (default_function_value): New function.
        * targhooks.h (default_function_value): Declare.

2005-07-13  Ian Lance Taylor  <ian@airs.com>

        * config/mips/mips.h (CLZ_DEFINED_VALUE_AT_ZERO): Define.

2005-07-14  Jan Hubicka  <jh@suse.cz>

        * tree-dfa.c (dump_variable): Use default_def function.
        * tree-ssa-alias.c (dump_points_to_info): Likewise.
        * tree-ssa.c (verify_use): Likewise.
        * tree-ssanames.c (release_ssa_name): Likewise.
        * tree-tailcall.c (eliminate_tail_call): Likewise.
        (tree_optimize_tail_calls_1): Likewise.
        * tree-vrp.c (get_value_range): Likewise.

2005-07-14  Ben Elliston  <bje@au.ibm.com>

        * gcc.c (main): Compare language[0] with '*' when iterating over
        the infiles.

2005-07-13  Adrian Strae½tling  <straetling@de.ibm.com>

        * config/s390/s390.c: (s390_cc_modes_compatible): Move before
        "s390_emit_compare".  Add handling of CCZ1mode.
        (s390_canonicalize_comparison): Simplify cascaded EQ, NE.
        (390_emit_compare): Use "s390_cc_modes_compatible" for mode
        checking.
        (s390_branch_condition_mask): Add CCZ1mode handling.
        * config/s390/s390.md: ("seq", "*seq"): New pattern.
        ("sync_compare_and_swap_cc<mode>", "*sync_compare_and_swap_cc<mode>"):
        Use CCZ1mode instead of CCZmode.
        * config/s390/s390-modes.def: Add CCZ1mode.  Comment new mode.

2006-07-13  Adrian Strae½tling  <straetling@de.ibm.com>

        * config/s390/s390.md: ("cmpstrsi", "*cmpstr<mode>"): New
        pattern.
        ("strlen<mode>", "*strlen<mode>"): Use hard reg 0 in SImode.

2005-07-13  Eric Christopher  <echristo@redhat.com>

        * config/mips/mips.c (mips_canonicalize_comparison): New.
        (mips_emit_int_relational): Use.

2005-07-13  Eric Christopher  <echristo@redhat.com>

        * config.gcc (s390x-ibm-tpf*): Add extra_options. Remove
        static extra parts.
        * config/s390/s390.md: Include tpf.md. Move tpf specific
        patterns...
        * config/s390/tpf.md: To here.
        * config/s390/s390.opt: Move tpf specific options...
        * config/s390/tpf.opt: to here. Add mmain option.
        * config/s390/tpf-unwind.h: Remove unnecessary defines.
        * config/s390/tpf.h: Rewrite.

2005-07-13  H.J. Lu  <hongjiu.lu@intel.com>

        * doc/tm.texi: Remove @xref{Cross-profiling}.

2005-07-13  Jeff Law  <law@redhat.com>

        * fold-const.c (fold_binary): When comparing two simple ADDR_EXPR
        expressions, test their _DECL operands for pointer equality rather
        than using operand_equal_p.

2005-07-13  H.J. Lu  <hongjiu.lu@intel.com>

        * config/alpha/linux.h (TARGET_HAS_F_SETLKW): Renamed to ...
        (TARGET_POSIX_IO): This.
        * config/darwin.h: Likewise.
        * config/freebsd.h: Likewise.
        * config/linux.h: Likewise.
        * config/lynx.h: Likewise.
        * config/netbsd.h: Likewise.
        * config/rs6000/linux64.h: Likewise.
        * config/rs6000/linux.h: Likewise.
        * config/s390/tpf.h: Likewise.
        * config/sh/embed-elf.h: Likewise.
        * config/sparc/linux64.h: Likewise.
        * config/sparc/linux.h: Likewise.
        * config/svr4.h: Likewise.
        * gcov-io.h: Likewise.

        * doc/tm.texi: Updated.

        * libgcov.c (create_file_directory): Defined only if
        TARGET_POSIX_IO is defined.
        (gcov_exit): Call create_file_directory only if TARGET_POSIX_IO
        is defined.

2005-07-13  Jan Hubicka  <jh@suse.cz>

        * tree-ssa-operands.c (get_expr_operands): Fix typo in previous patch.

2005-07-13  David Edelsohn  <edelsohn@gnu.org>

        * tree-ssa-dom.c (lookup_avail_expr): Do not pass member in freed
        structure as argument.

2005-07-13  Paolo Bonzini  <bonzini@gnu.org>

        PR tree-optimization/21921
        * tree-iterator.c (tsi_link_before): Support the case when
        tsi_end_p (tsi) == true.

Re: Middle-end and optimization regressions: what should we do?

[Steve Kargl]

On Thu, Jul 28, 2005 at 07:26:22PM +0200, Fran?ois-Xavier Coudert wrote:
> 
> PR 22619 and PR 22509 are two examples of recent 4.1 regressions that
> showed up in gfortran, due to middle-end or optimization bugs (only
> happen at -O3). Since these are regressions, they should be treated
> before a long time passes, but since both source codes are Fortran, I
> guess people don't (and won't) want to look at them.
> 
> How can we help here? Is there a way to make gfortran output a
> complete GIMPLE tree, that could be used for middle-end hackers to
> determine where the problem is? Or are we doomed to a dichotomy to
> know which patch caused these regressions?
> 

These types of regressions have essentially halted my testing
and development on gfortran because I usually try to identify
the exact ChangeLog entry associated with the problem.  This
typically involves a binary search for the problem with a
bootstrap in a clean directory for each "cvs update -D <date>". 

As far as providing info to the middle-end people, you can
do -fdump-tree-all and try to sift through the volumes of
data.

-- 
Steve

Re: Middle-end and optimization regressions: what should we do?

[Daniel Berlin]

On Thu, 2005-07-28 at 19:26 +0200, François-Xavier Coudert wrote:
> Hi all,
> 
> PR 22619 and PR 22509 are two examples of recent 4.1 regressions that
> showed up in gfortran, due to middle-end or optimization bugs (only
> happen at -O3). Since these are regressions, they should be treated
> before a long time passes, but since both source codes are Fortran, I
> guess people don't (and won't) want to look at them.

I was getting these failures before my patch that added eustores, and
kenny was responsible for regtesting the tree for the ipa-* patches :)

(I always bootstrap and regtest with fortran on).



> PS: PR 22619 appeared somewhere between 20050716 and 20050717, so
> patches that could possible have messed up are:

The culprit in this case is indirectly the patch that added
ipa-reference AFAICT.  -fno-ipa-reference seems to fix the bug on my
computer, see if it helps you too.

No clue why.

Given that it happens after CSE, I imagine it's enabling CSE to
eliminate something it didn't expect to be able to eliminate.

Re: Middle-end and optimization regressions: what should we do?

[Andrew Pinski]

> 
> Hi all,
> 
> PR 22619 and PR 22509 are two examples of recent 4.1 regressions that
> showed up in gfortran, due to middle-end or optimization bugs (only
> happen at -O3). Since these are regressions, they should be treated
> before a long time passes, but since both source codes are Fortran, I
> guess people don't (and won't) want to look at them.
> 
> How can we help here? Is there a way to make gfortran output a
> complete GIMPLE tree, that could be used for middle-end hackers to
> determine where the problem is? Or are we doomed to a dichotomy to
> know which patch caused these regressions?
> 
> FX
> 
> PS: PR 22619 appeared somewhere between 20050716 and 20050717, so
> patches that could possible have messed up are:

Patch which exposed it:
> 2005-07-16 Danny Berlin <dberlin@dberlin.org>
>            Kenneth Zadeck <zadeck@naturalbridge.com>

Kenny's patch exposed the latent bug.  I attached a testcase to RR
22619 showing that it is a regression from 3.2.3 and it fails in
3.3 and up.  Kenny's patch changes a static variable to a static
const variable which gets "inlined".





> PR 22509 appeared between 20050713 and 20050714, so possible guilty patches are:

this was caused exposed by:
2005-07-12  Zdenek Dvorak  <dvorakz@suse.cz>

        PR rtl-optimization/20376
        * toplev.c (process_options): Enable -fweb and -frename-registers when
        unrolling.
        * doc/invoke.texi: Update the information about when -fweb and
        -frename-registers are enabled.

Thanks,
Andrew Pinski

Re: Middle-end and optimization regressions: what should we do?

[Janis Johnson]

On Thu, Jul 28, 2005 at 10:41:48AM -0700, Steve Kargl wrote:
> On Thu, Jul 28, 2005 at 07:26:22PM +0200, Fran?ois-Xavier Coudert wrote:
> > 
> > PR 22619 and PR 22509 are two examples of recent 4.1 regressions that
> > showed up in gfortran, due to middle-end or optimization bugs (only
> > happen at -O3). Since these are regressions, they should be treated
> > before a long time passes, but since both source codes are Fortran, I
> > guess people don't (and won't) want to look at them.
> > 
> > How can we help here? Is there a way to make gfortran output a
> > complete GIMPLE tree, that could be used for middle-end hackers to
> > determine where the problem is? Or are we doomed to a dichotomy to
> > know which patch caused these regressions?
> 
> These types of regressions have essentially halted my testing
> and development on gfortran because I usually try to identify
> the exact ChangeLog entry associated with the problem.  This
> typically involves a binary search for the problem with a
> bootstrap in a clean directory for each "cvs update -D <date>". 

In case you're not already aware of them, see contrib/reghunt and
http://gcc.gnu.org/bugs/reghunt.html.

Janis