tree-optimization/102880 - improve CD-DCE

The PR shows a missed control-dependent DCE caused by CFG cleanup
merging a forwarder resulting in a partially degenerate PHI node.
With control-dependent DCE we need to mark control dependences
of incoming edges into PHIs as necessary but that is unnecessarily
conservative for the case when two edges have the same value.
There is no easy way to mark only a subset of control dependences
of both edges necessary so the fix is to produce forwarder blocks
where then the control dependence captures the requirements more
precisely.

For gcc.dg/tree-ssa/ssa-dom-thread-7.c the number of edges in the
CFG decrease as we have commonized PHI arguments which in turn
results in different threadings.  The testcase is too complex
and the dump scanning too simple to do anything meaningful here
but to adjust the number of expected threads.

The same CFG massaging could be useful at RTL expansion time to
reduce the number of copies we need to insert on edges.

FAIL: gcc.dg/tree-ssa/ssa-hoist-4.c scan-tree-dump-times optimized "MAX_EXPR" 1

2021-11-12  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102880
* tree-ssa-dce.c (sort_phi_args): New function.
(make_forwarders_with_degenerate_phis): Likewise.
(perform_tree_ssa_dce): Call
make_forwarders_with_degenerate_phis.

* gcc.dg/tree-ssa/pr102880.c: New testcase.
* gcc.dg/tree-ssa/pr69270-3.c: Robustify.
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Change the number of
expected threadings.

tree-optimization/102880 - make PHI-OPT recognize more CFGs

This allows extra edges into the middle BB for the PHI-OPT
transforms using replace_phi_edge_with_variable that do not
end up moving stmts from that middle BB.  This avoids regressing
gcc.dg/tree-ssa/ssa-hoist-4.c with the actual fix for PR102880
where CFG cleanup has the choice to remove two forwarders and
picks "the wrong" leading to

   if (a > b) /
       /\    /
      /  <BB>
     /    |
  # PHI <a, b>

rather than

   if (a > b)  |
       /\      |
    <BB> \     |
     /    \    |
  # PHI <a, b, b>

but it's relatively straight-forward to support extra edges
into the middle-BB in paths ending in replace_phi_edge_with_variable
and that do not require moving stmts.  That's because we really
only want to remove the edge from the condition to the middle BB.
Of course actually doing that means updating dominators in non-trival
ways which is why I kept the original code for the single edge
case and simply defer to CFG cleanup by adjusting the condition for
the complicated case.

The testcase needs to be a GIMPLE one since it's quite unreliable
to produce the desired CFG.

2021-11-15  Richard Biener  <rguenther@suse.de>

PR tree-optimization/102880
* tree-ssa-phiopt.c (tree_ssa_phiopt_worker): Push
single_pred (bb1) condition to places that really need it.
(match_simplify_replacement): Likewise.
(value_replacement): Likewise.
(replace_phi_edge_with_variable): Deal with extra edges
into the middle BB.

* gcc.dg/tree-ssa/phi-opt-26.c: New testcase.

arc: Update arc specific tests

Update assembly output test pattern. Take into consideration also for
which platform we do execute the test (baremetal or linux).

gcc/testsuite/ChangeLog:

* gcc.target/arc/add_n-combine.c: Update test patterns.
* gcc.target/arc/builtin_eh.c: Update test for linux platforms.
* gcc.target/arc/mul64-1.c: Disable this test while running on
linux.
* gcc.target/arc/tls-gd.c: Update matching patterns.
* gcc.target/arc/tls-ie.c: Likewise.
* gcc.target/arc/tls-ld.c: Likewise.
* gcc.target/arc/uncached-8.c: Likewise.

Signed-off-by: Claudiu Zissulescu <claziss@synopsys.com>

Replace more DEBUG_EXPR_DECL creations with build_debug_expr_decl

As discussed on the mailing list, this patch replaces all but one
remaining open coded constructions of DEBUG_EXPR_DECL with calls to
build_debug_expr_decl, even if - in order not to introduce any
functional change - the mode of the constructed decl is then
overwritten.

It is not clear if changing the mode has any effect in practice and
therefore I have added a FIXME note to code which does it, as
requested.

After this patch, DEBUG_EXPR_DECLs are created only by
build_debug_expr_decl and make_debug_expr_from_rtl which looks like
it should be left alone.

gcc/ChangeLog:

2021-11-11  Martin Jambor  <mjambor@suse.cz>

* cfgexpand.c (expand_gimple_basic_block): Use build_debug_expr_decl,
add a fixme note about the mode assignment perhaps being unnecessary.
* ipa-param-manipulation.c (ipa_param_adjustments::modify_call):
Likewise.
(ipa_param_body_adjustments::mark_dead_statements): Likewise.
(ipa_param_body_adjustments::reset_debug_stmts): Likewise.
* tree-inline.c (remap_ssa_name): Likewise.
(tree_function_versioning): Likewise.
* tree-into-ssa.c (rewrite_debug_stmt_uses): Likewise.
* tree-ssa-loop-ivopts.c (remove_unused_ivs): Likewise.
* tree-ssa.c (insert_debug_temp_for_var_def): Likewise.

ipa-sra: Testcase that removing a "returns_nonnull" retval works

Since we can now remove return values of functions with return_nonnull
type attribute, I'll feel a bit safer if we can test this does not ICE
when someone attempts to access a non-existent call LHS.  Eventually
we should probably drop the attribute when this happens.

gcc/testsuite/ChangeLog:

2021-11-15  Martin Jambor  <mjambor@suse.cz>

* gcc.dg/ipa/ipa-sra-ret-nonull.c: New test.

libgomp: Mark thread_limit clause to target construct as implemented

After the Fortran changes we can mark it as implemented...

2021-11-16  Jakub Jelinek  <jakub@redhat.com>

* libgomp.texi (OpenMP 5.1): Mark thread_limit clause to target
construct as implemented.

openmp: Regimplify operands of GIMPLE_COND in a few more places [PR103208]

As the testcase shows, the non-rectangular loop expansion code didn't
try to regimplify operands of GIMPLE_CONDs it built in some cases.
I have added a helper function which does that and used it in some places
that were regimplifying already to simplify those spots, plus added it
in a couple of other places where it was needed.

2021-11-16  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/103208
* omp-expand.c (expand_omp_build_cond): New function.
(expand_omp_for_init_counts, expand_omp_for_init_vars,
expand_omp_for_static_nochunk, expand_omp_for_static_chunk): Use it.

* c-c++-common/gomp/loop-11.c: New test.

waccess: Fix up pass_waccess::check_alloc_size_call [PR102009]

This function punts if the builtins have no arguments, but as can be seen
on the testcase, even if it has some arguments but alloc_size attribute's
arguments point to arguments that aren't passed, we get a warning earlier
from the FE but should punt rather than ICE on it.
Other users of alloc_size attribute e.g. in
tree-object-size.c (alloc_object_size) punt similarly and similarly
even in the same TU maybe_warn_nonstring_arg correctly verifies calls have
enough arguments.

2021-11-16  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/102009
* gimple-ssa-warn-access.cc (pass_waccess::check_alloc_size_call):
Punt if any of alloc_size arguments is out of bounds vs. number of
call arguments.

* gcc.dg/pr102009.c: New test.

x86_64: Avoid rorx rotation instructions with -Os.

This patch teaches the i386 backend to avoid using BMI2's rorx
instructions when optimizing for size.  The benefits are shown
with the following example:

unsigned int ror1(unsigned int x) { return (x >> 1) | (x << 31); }
unsigned int ror2(unsigned int x) { return (x >> 2) | (x << 30); }
unsigned int rol2(unsigned int x) { return (x >> 30) | (x << 2); }
unsigned int rol1(unsigned int x) { return (x >> 31) | (x << 1); }

which currently with -Os -march=cascadelake generates:

ror1: rorx    $1, %edi, %eax // 6 bytes
        ret
ror2: rorx    $2, %edi, %eax // 6 bytes
        ret
rol2: rorx    $30, %edi, %eax // 6 bytes
        ret
rol1: rorx    $31, %edi, %eax // 6 bytes
        ret

but with this patch now generates:

ror1: movl    %edi, %eax // 2 bytes
        rorl    %eax // 2 bytes
        ret
ror2: movl    %edi, %eax // 2 bytes
        rorl    $2, %eax // 3 bytes
        ret
rol2: movl    %edi, %eax // 2 bytes
        roll    $2, %eax // 3 bytes
        ret
rol1: movl    %edi, %eax // 2 bytes
        roll    %eax // 2 bytes
        ret

I've confirmed that this patch is a win on the CSiBE benchmark,
even though rotations are rare, where for example libmspack/test/md5.o
shrinks from 5824 bytes to 5632 bytes.

2021-11-16  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/i386/i386.md (*bmi2_rorx<mode3>_1): Make conditional
on !optimize_function_for_size_p.
(*<any_rotate><mode>3_1): Add preferred_for_size attribute.
(define_splits): Conditionalize on !optimize_function_for_size_p.
(*bmi2_rorxsi3_1_zext): Likewise.
(*<any_rotate>si2_1_zext): Add preferred_for_size attribute.
(define_splits): Conditionalize on !optimize_function_for_size_p.

Fix uninitialized access in merge_call_side_effects

gcc/ChangeLog:

PR ipa/103262
* ipa-modref.c (merge_call_side_effects): Fix uninitialized
access.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/modref-dse-5.c: New test.

tree-optimization: [PR103245] Improve detection of abs pattern using multiplication

So while working on PR 103228 (and a few others), I noticed the testcase for PR 94785
was failing. The problem is that the nop_convert moved from being inside the IOR to be
outside of it. I also noticed the patch for PR 103228 was not needed to reproduce the
issue either.
This patch combines the two patterns together for the abs match when using multiplication
and adds a few places where nop_convert are optional.

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

PR tree-optimization/103245

gcc/ChangeLog:

* match.pd: Combine the abs pattern matching using multiplication.
Adding optional nop_convert too.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/pr103245-1.c: New test.

Add a missing return when transforming atomic bit test and operations

When failing to transform equivalent, but slighly different cases of
atomic bit test and operations to their canonical forms, return
immediately.

gcc/

PR middle-end/103268
* tree-ssa-ccp.c (optimize_atomic_bit_test_and): Add a missing
return.

gcc/testsuite/

PR middle-end/103268
* gcc.dg/pr103268-1.c: New test.
* gcc.dg/pr103268-2.c: Likewise.

Update my email address.

* MAINTAINERS: Update my address.

Daily bump.

c++: Add -fimplicit-constexpr

With each successive C++ standard the restrictions on the use of the
constexpr keyword for functions get weaker and weaker; it recently occurred
to me that it is heading toward the same fate as the C register keyword,
which was once useful for optimization but became obsolete.  Similarly, it
seems to me that we should be able to just treat inlines as constexpr
functions and not make people add the extra keyword everywhere.

There were a lot of testcase changes needed; many disabling errors about
non-constexpr functions that are now constexpr, and many disabling implicit
constexpr so that the tests can check the same thing as before, whether
that's mangling or whatever.

gcc/c-family/ChangeLog:

* c.opt: Add -fimplicit-constexpr.
* c-cppbuiltin.c: Define __cpp_implicit_constexpr.
* c-opts.c (c_common_post_options): Disable below C++14.

gcc/cp/ChangeLog:

* cp-tree.h (struct lang_decl_fn): Add implicit_constexpr.
(decl_implicit_constexpr_p): New.
* class.c (type_maybe_constexpr_destructor): Use
TYPE_HAS_TRIVIAL_DESTRUCTOR and maybe_constexpr_fn.
(finalize_literal_type_property): Simplify.
* constexpr.c (is_valid_constexpr_fn): Check for dtor.
(maybe_save_constexpr_fundef): Try to set DECL_DECLARED_CONSTEXPR_P
on inlines.
(cxx_eval_call_expression): Use maybe_constexpr_fn.
(maybe_constexpr_fn): Handle flag_implicit_constexpr.
(var_in_maybe_constexpr_fn): Use maybe_constexpr_fn.
(potential_constant_expression_1): Likewise.
(decl_implicit_constexpr_p): New.
* decl.c (validate_constexpr_redeclaration): Allow change with
-fimplicit-constexpr.
(grok_special_member_properties): Use maybe_constexpr_fn.
* error.c (dump_function_decl): Don't print 'constexpr'
if it's implicit.
* Make-lang.in (check-c++-all): Update.

libstdc++-v3/ChangeLog:

* testsuite/20_util/to_address/1_neg.cc: Adjust error.
* testsuite/26_numerics/random/concept.cc: Adjust asserts.

gcc/testsuite/ChangeLog:

* lib/g++-dg.exp: Handle "impcx".
* lib/target-supports.exp
(check_effective_target_implicit_constexpr): New.
* g++.dg/abi/abi-tag16.C:
* g++.dg/abi/abi-tag18a.C:
* g++.dg/abi/guard4.C:
* g++.dg/abi/lambda-defarg1.C:
* g++.dg/abi/mangle26.C:
* g++.dg/cpp0x/constexpr-diag3.C:
* g++.dg/cpp0x/constexpr-ex1.C:
* g++.dg/cpp0x/constexpr-ice5.C:
* g++.dg/cpp0x/constexpr-incomplete2.C:
* g++.dg/cpp0x/constexpr-memfn1.C:
* g++.dg/cpp0x/constexpr-neg3.C:
* g++.dg/cpp0x/constexpr-specialization.C:
* g++.dg/cpp0x/inh-ctor19.C:
* g++.dg/cpp0x/inh-ctor30.C:
* g++.dg/cpp0x/lambda/lambda-mangle3.C:
* g++.dg/cpp0x/lambda/lambda-mangle5.C:
* g++.dg/cpp1y/auto-fn12.C:
* g++.dg/cpp1y/constexpr-loop5.C:
* g++.dg/cpp1z/constexpr-lambda7.C:
* g++.dg/cpp2a/constexpr-dtor3.C:
* g++.dg/cpp2a/constexpr-new13.C:
* g++.dg/cpp2a/constinit11.C:
* g++.dg/cpp2a/constinit12.C:
* g++.dg/cpp2a/constinit14.C:
* g++.dg/cpp2a/constinit15.C:
* g++.dg/cpp2a/spaceship-constexpr1.C:
* g++.dg/cpp2a/spaceship-eq3.C:
* g++.dg/cpp2a/udlit-class-nttp-neg2.C:
* g++.dg/debug/dwarf2/auto1.C:
* g++.dg/debug/dwarf2/cdtor-1.C:
* g++.dg/debug/dwarf2/lambda1.C:
* g++.dg/debug/dwarf2/pr54508.C:
* g++.dg/debug/dwarf2/pubnames-2.C:
* g++.dg/debug/dwarf2/pubnames-3.C:
* g++.dg/ext/is_literal_type3.C:
* g++.dg/ext/visibility/template7.C:
* g++.dg/gcov/gcov-12.C:
* g++.dg/gcov/gcov-2.C:
* g++.dg/ipa/devirt-35.C:
* g++.dg/ipa/devirt-36.C:
* g++.dg/ipa/devirt-37.C:
* g++.dg/ipa/devirt-44.C:
* g++.dg/ipa/imm-devirt-1.C:
* g++.dg/lookup/builtin5.C:
* g++.dg/lto/inline-crossmodule-1_0.C:
* g++.dg/modules/enum-1_a.C:
* g++.dg/modules/fn-inline-1_c.C:
* g++.dg/modules/pmf-1_b.C:
* g++.dg/modules/used-1_c.C:
* g++.dg/tls/thread_local11.C:
* g++.dg/tls/thread_local11a.C:
* g++.dg/tm/pr46653.C:
* g++.dg/ubsan/pr70035.C:
* g++.old-deja/g++.other/delete6.C:
* g++.dg/modules/pmf-1_a.H:
Adjust for implicit constexpr.

c++: split_nonconstant_init and flexarrays

split_nonconstant_init was doing the wrong thing for both the initialization
and cleanup here; we know the size from the initializer, and we can pass it
along.  This doesn't make the testcase work, since the y destructor is still
broken, but it removes the wrong error for the aggregate initialization.

gcc/cp/ChangeLog:

* typeck2.c (split_nonconstant_init_1): Handle flexarrays better.

gcc/testsuite/ChangeLog:

* g++.dg/ext/flexary37.C: Remove expected error.

gimple-fold: Use ranges to simplify strncat and snprintf

Use ranges for lengths and object sizes in strncat and snprintf to
determine if they can be transformed into simpler operations.

gcc/ChangeLog:

* gimple-fold.c (gimple_fold_builtin_strncat): Use ranges to
determine if it is safe to transform to strcat.
(gimple_fold_builtin_snprintf): Likewise.

gcc/testsuite/ChangeLog:

* gcc.dg/fold-stringops-2.c: Define size_t.
(safe1): Adjust.
(safe4): New test.
* gcc.dg/fold-stringops-3.c: New test.

Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>

gimple-fold: Use ranges to simplify _chk calls

Instead of comparing LEN and SIZE only if they are constants, use their
ranges to decide if LEN will always be lower than or same as SIZE.

This change ends up putting the stringop-overflow warning line number
against the strcpy implementation, so adjust the warning check to be
line number agnostic.

gcc/ChangeLog:

* gimple-fold.c (known_lower): New function.
(gimple_fold_builtin_strncat_chk,
gimple_fold_builtin_memory_chk, gimple_fold_builtin_stxcpy_chk,
gimple_fold_builtin_stxncpy_chk,
gimple_fold_builtin_snprintf_chk,
gimple_fold_builtin_sprintf_chk): Use it.

gcc/testsuite/ChangeLog:

* gcc.dg/Wobjsize-1.c: Make warning change line agnostic.
* gcc.dg/fold-stringops-2.c: New test.

Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>

gimple-fold: Transform stp*cpy_chk to str*cpy directly

Avoid going through another folding cycle and use the ignore flag to
directly transform BUILT_IN_STPCPY_CHK to BUILT_IN_STRCPY when set,
likewise for BUILT_IN_STPNCPY_CHK to BUILT_IN_STPNCPY.

Dump the transformation in dump_file so that we can verify in tests that
the direct transformation actually happened.

gcc/ChangeLog:

* gimple-fold.c (dump_transformation): New function.
(gimple_fold_builtin_stxcpy_chk,
gimple_fold_builtin_stxncpy_chk): Use it.  Simplify to
BUILT_IN_STRNCPY if return value is not used.

gcc/testsuite/ChangeLog:

* gcc.dg/fold-stringops-1.c: New test.

Signed-off-by: Siddhesh Poyarekar <siddhesh@gotplt.org>

Check optab before transforming atomic bit test and operations

Check optab before transforming equivalent, but slighly different cases
of atomic bit test and operations to their canonical forms.

gcc/

PR middle-end/103184
* tree-ssa-ccp.c (optimize_atomic_bit_test_and): Check optab
before transforming equivalent, but slighly different cases to
their canonical forms.

gcc/testsuite/

PR middle-end/103184
* gcc.dg/pr103184-1.c: New test.
* gcc.dg/pr103184-2.c: Likewise.

IPA: Provide a mechanism to register static DTORs via cxa_atexit.

For at least one target (Darwin) the platform convention is to
register static destructors (i.e. __attribute__((destructor)))
with __cxa_atexit rather than placing them into a list that is
run by some other mechanism.

This patch provides a target hook that allows a target to opt
into this and handling for the process in ipa_cdtor_merge ().

When the mode is enabled (dtors_from_cxa_atexit is set) we:

 * Generate new CTORs to register static destructors with
   __cxa_atexit and add them to the existing list of CTORs;
   we then process the revised CTORs list.

 * We sort the DTORs into priority and then TU order, this
   means that they are registered in that order with
   __cxa_atexit () and therefore will be run in the reverse
   order.

 * Likewise, CTORs are sorted into priority and then TU order,
   which means that they will run in that order.

This matches the behavior of using init/fini (or
mod_init_func/mod_term_func) sections.

This also fixes a bug where Fortran needs a DTOR to be run to
close IO.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
PR fortran/102992

gcc/ChangeLog:

* config/darwin.h (TARGET_DTORS_FROM_CXA_ATEXIT): New.
* doc/tm.texi: Regenerated.
* doc/tm.texi.in: Add TARGET_DTORS_FROM_CXA_ATEXIT hook.
* ipa.c (cgraph_build_static_cdtor_1): Return the built
function decl.
(build_cxa_atexit_decl): New.
(build_dso_handle_decl): New.
(build_cxa_dtor_registrations): New.
(compare_cdtor_tu_order): New.
(build_cxa_atexit_fns): New.
(ipa_cdtor_merge): If dtors_from_cxa_atexit is set,
process the DTORs/CTORs accordingly.
(pass_ipa_cdtor_merge::gate): Also run if
dtors_from_cxa_atexit is set.
* target.def (dtors_from_cxa_atexit): New hook.

configure, Darwin: Check ld64 support for -platform-version.

Newer versions of ld64 allow specifiying the OS target (e.g.
macos or ios) the version and the SDK version all in a single
command.  This checks the availability of the command for the
current toolchain.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:

* config.in: Regenerate.
* configure: Regenerate.
* configure.ac: Test ld64 for -platform-version support.

testsuite, Darwin: In tsvc.h, use malloc for Darwin <= 9.

Earlier Darwin versions fdo not have posix_memalign() but the
malloc implementation is guaranteed to produce memory suitably
aligned for the largest vector type.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/testsuite/ChangeLog:

* gcc.dg/vect/tsvc/tsvc.h: Use malloc for Darwin 9 and
earlier.

Ada, Darwin : Use DSYMUTIL_FOR_TARGET in libgnat/gnarl builds.

Most of the time we get away with using the dsymutil that is
installed with the latest Xcode, however for some cross-compilation
cases that does not work.

We now have the ability to specify the correct dsymutil to use for
the toolchain (--with-dsymutil=) and we should use that specified
tool for debug link.  Fixes cross-compilers from x86-64 to powerpc.

Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ada/ChangeLog:

* gcc-interface/Makefile.in: Use DSYMUTIL_FOR_TARGET in
libgnat/libgnarl recipies.

libstdc++: Unordered containers merge re-use hash code

When merging 2 unordered containers with same hasher we can re-use the hash code from
the cache if any.

Also in the context of the merge operation on multi-container use previous insert iterator as a hint
for the next insert.

libstdc++-v3/ChangeLog:

* include/bits/hashtable_policy.h:
(_Hash_code_base<>::_M_hash_code(const _Hash&, const _Hash_node_value<_Value, true>&)): New.
(_Hash_code_base<>::_M_hash_code<_H2>(const _H2&, const _Hash_node_value<>&)): New.
* include/bits/hashtable.h (_Hashtable<>::_M_merge_unique): Use latter.
(_Hashtable<>::_M_merge_multi): Likewise.
* testsuite/23_containers/unordered_multiset/modifiers/merge.cc (test05): New test.
* testsuite/23_containers/unordered_set/modifiers/merge.cc (test04): New test.

Use 'location_hash' for 'gcc/diagnostic-spec.h:nowarn_map'

Instead of hard-coded '0'/'UINT_MAX', we now use the 'RESERVED_LOCATION_P'
values 'UNKNOWN_LOCATION'/'BUILTINS_LOCATION' as spare values for
'Empty'/'Deleted', and generally simplify the code.

gcc/
* diagnostic-spec.h (typedef xint_hash_t)
(typedef xint_hash_map_t): Replace with...
(typedef nowarn_map_t): ... this.
(nowarn_map): Adjust.
* diagnostic-spec.c (nowarn_map, suppress_warning_at): Likewise.

Use 'location_hash' for 'seen_locations' in 'gcc/profile.c:branch_prob'

Follow-up to commit 102fcf94e625a2016a65829c73a42bd6c2420376
"Fix GCOV CFG related issues": considering the current
'int_hash <location_t, 0, 2>', per 'libcpp/include/line-map.h':

      Actual     | Value                         | Meaning
      -----------+-------------------------------+-------------------------------
      0x00000000 | UNKNOWN_LOCATION (gcc/input.h)| Unknown/invalid location.
      -----------+-------------------------------+-------------------------------
      0x00000001 | BUILTINS_LOCATION             | The location for declarations
                 |   (gcc/input.h)               | in "<built-in>"
      -----------+-------------------------------+-------------------------------
      0x00000002 | RESERVED_LOCATION_COUNT       | The first location to be
                 | (also                         | handed out, and the
                 |  ordmap[0]->start_location)   | first line in ordmap 0

... this currently uses value '0' ('UNKNOWN_LOCATION') as spare values for
'Empty', and value '2' ('RESERVED_LOCATION_COUNT') as spare values for
'Deleted', which is questionable?

What actually does get put into 'seen_locations' is (mostly...)
restricted/gated by '!RESERVED_LOCATION_P' (which is true unless
'UNKNOWN_LOCATION' or 'BUILTINS_LOCATION'), thus we may simply use
'location_hash'.

gcc/
* profile.c (branch_prob): Use 'location_hash' for
'seen_locations'.

Drop tree overflow in irange setter.

Drop meaningless overflow that may creep into the IL.

gcc/ChangeLog:

PR tree-optimization/103207
* value-range.cc (irange::set): Drop overflow.

gcc/testsuite/ChangeLog:

* gcc.dg/pr103207.c: New test.

Fortran: openmp: Add support for thread_limit clause on target

gcc/fortran/ChangeLog:

* openmp.c (OMP_TARGET_CLAUSES): Add thread_limit.
* trans-openmp.c (gfc_split_omp_clauses): Add thread_limit also to
teams.

libgomp/ChangeLog:

* testsuite/libgomp.fortran/thread-limit-1.f90: New test.

testsuite: Add testcase for already fixed PR [PR100469]

This bug introduced in r11-7448-gff92ede8d269375f800e1b347a48f4698874b4a3
has been fixed already by r12-1354-g2d2ed777b23ab6503027039e0adbfe1162f52b2f
aka PR100852 fix.

2021-11-15  Jakub Jelinek  <jakub@redhat.com>

PR debug/100469
* g++.dg/opt/pr100469.C: New test.

x86: Add gcc.target/i386/pr103205-2.c

PR target/103205
* gcc.target/i386/pr103205-2.c: New test.

libffi: Update LOCAL_PATCHES

Add

commit a91f844ef449d0dd1cf2e0e47b0ade0d8a6304e1
Author: Rainer Orth <ro@CeBiTec.Uni-Bielefeld.DE>
Date:   Mon Nov 15 10:24:27 2021 +0100

    libffi: Use #define instead of .macro in  src/x86/win64.S [PR102874]

to LOCAL_PATCHES.

* LOCAL_PATCHES: Add commit a91f844ef44.

openmp: Add support for thread_limit clause on target

OpenMP 5.1 says that thread_limit clause can also appear on target,
and similarly to teams should affect the thread-limit-var ICV.
On combined target teams, the clause goes to both.

We actually passed thread_limit internally on target already before,
but only used it for gcn/ptx offloading to hint how many threads should be
created and for ptx didn't set thread_limit_var in that case.
Similarly for host fallback.
Also, I found that we weren't copying the args array that contains encoded
thread_limit and num_teams clause for target (etc.) for async target.

2021-11-15  Jakub Jelinek  <jakub@redhat.com>

gcc/
* gimplify.c (optimize_target_teams): Only add OMP_CLAUSE_THREAD_LIMIT
to OMP_TARGET_CLAUSES if it isn't there already.
gcc/c-family/
* c-omp.c (c_omp_split_clauses) <case OMP_CLAUSE_THREAD_LIMIT>:
Duplicate to both OMP_TARGET and OMP_TEAMS.
gcc/c/
* c-parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
gcc/cp/
* parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
libgomp/
* task.c (gomp_create_target_task): Copy args array as well.
* target.c (gomp_target_fallback): Add args argument.
Set gomp_icv (true)->thread_limit_var if thread_limit is present.
(GOMP_target): Adjust gomp_target_fallback caller.
(GOMP_target_ext): Likewise.
(gomp_target_task_fn): Likewise.
* config/nvptx/team.c (gomp_nvptx_main): Set
gomp_global_icv.thread_limit_var.
* testsuite/libgomp.c-c++-common/thread-limit-1.c: New test.

Fix PHI ordering problems in the path solver.

After auditing the PHI range calculations, I'm not convinced we've
caught all the corner cases.  They haven't shown up in the wild (yet),
but better safe than sorry.

We shouldn't write anything to the cache or trigger additional
lookups while calculating a PHI, as this may cause ordering problems.
We should resolve the PHI with either the cache as it stands, or by
asking for ranges on entry to the path.  I've documented this.

There was one dubious case where we called fold_range in
ssa_range_in_phi, which mostly by luck wasn't triggering lookups,
because fold_range solves a PHI by calling range_on_edge, which is set
to pick up global ranges by default in path_range_query.  This is
fragile, so I've rewritten the call to explicitly use cached or global
ranges.

Also, the cache should be avoided in ssa_range_in_phi when the arg is
defined in the PHI's block, as not doing so could create an ordering
problem.  We have a similar check when calculating relations in PHIs.

Tested on x86-64 & ppc64le Linux.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::internal_range_of_expr):
Remove useless code.
(path_range_query::ssa_defined_in_bb): New.
(path_range_query::ssa_range_in_phi): Avoid fold_range call that
could trigger additional lookups.
Do not use the cache for ARGs defined in this block.
(path_range_query::compute_ranges_in_block): Use ssa_defined_in_bb.
(path_range_query::maybe_register_phi_relation): Same.
(path_range_query::range_of_stmt): Adjust comment.
* gimple-range-path.h (ssa_defined_in_bb): New.

path solver: Default to global range if nothing found.

This has been a long time coming, but we weren't able to make the
change because of some unrelated regressions.

Tested on x86-64 & ppc64le Linux.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::internal_range_of_expr):
Default to global range if nothing found.

gcc/testsuite/ChangeLog:

* g++.dg/tree-ssa/pr31146-2.C: Add -fno-thread-jumps.

tree-optimization/103237 - avoid vectorizing unhandled double reductions

Double reductions which have multiple LC PHIs in the inner loop
are not handled correctly during transformation since those PHIs
are not properly classified as reduction.  The following disables
vectorizing them.

2021-11-15  Richard Biener  <rguenther@suse.de>

PR tree-optimization/103237
* tree-vect-loop.c (vect_is_simple_reduction): Fail for
double reductions with multiple inner loop LC PHI nodes.

* gcc.dg/torture/pr103237.c: New testcase.

PR target/103069: Relax cmpxchg loop for x86 target

From the CPU's point of view, getting a cache line for writing is more
expensive than reading.  See Appendix A.2 Spinlock in:

https://www.intel.com/content/dam/www/public/us/en/documents/white-papers/
xeon-lock-scaling-analysis-paper.pdf

The full compare and swap will grab the cache line exclusive and causes
excessive cache line bouncing.

The atomic_fetch_{or,xor,and,nand} builtins generates cmpxchg loop under
-march=x86-64 like:

movl v(%rip), %eax
.L2:
movl %eax, %ecx
movl %eax, %edx
orl $1, %ecx
lock cmpxchgl %ecx, v(%rip)
jne .L2
movl %edx, %eax
andl $1, %eax
ret

To relax above loop, GCC should first emit a normal load, check and jump to
.L2 if cmpxchgl may fail. Before jump to .L2, PAUSE should be inserted to
yield the CPU to another hyperthread and to save power, so the code is
like

.L84:
        movl    (%rdi), %ecx
        movl    %eax, %edx
        orl     %esi, %edx
        cmpl    %eax, %ecx
        jne     .L82
        lock cmpxchgl   %edx, (%rdi)
        jne     .L84
.L82:
        rep nop
        jmp     .L84

This patch adds corresponding atomic_fetch_op expanders to insert load/
compare and pause for all the atomic logic fetch builtins. Add flag
-mrelax-cmpxchg-loop to control whether to generate relaxed loop.

gcc/ChangeLog:

PR target/103069
* config/i386/i386-expand.c (ix86_expand_atomic_fetch_op_loop):
New expand function.
* config/i386/i386-options.c (ix86_target_string): Add
-mrelax-cmpxchg-loop flag.
(ix86_valid_target_attribute_inner_p): Likewise.
* config/i386/i386-protos.h (ix86_expand_atomic_fetch_op_loop):
New expand function prototype.
* config/i386/i386.opt: Add -mrelax-cmpxchg-loop.
* config/i386/sync.md (atomic_fetch_<logic><mode>): New expander
for SI,HI,QI modes.
(atomic_<logic>_fetch<mode>): Likewise.
(atomic_fetch_nand<mode>): Likewise.
(atomic_nand_fetch<mode>): Likewise.
(atomic_fetch_<logic><mode>): New expander for DI,TI modes.
(atomic_<logic>_fetch<mode>): Likewise.
(atomic_fetch_nand<mode>): Likewise.
(atomic_nand_fetch<mode>): Likewise.
* doc/invoke.texi: Document -mrelax-cmpxchg-loop.

gcc/testsuite/ChangeLog:

PR target/103069
* gcc.target/i386/pr103069-1.c: New test.
* gcc.target/i386/pr103069-2.c: Ditto.

tree-optimization/103219 - avoid ICE in unroll-and-jam

For no particularly good reason unroll-and-jam uses single_dom_exit
to determine the exit for the region it wants to run VN on.  That
happens to ICE because of the dominance restriction.  Use single_exit
instead.

2021-11-15  Richard Biener  <rguenther@suse.de>

PR tree-optimization/103219
* gimple-loop-jam.c (tree_loop_unroll_and_jam): Use single_exit
to determine the exit for the VN region.

* gcc.dg/torture/pr103219.c: New testcase.

[tree-vectorizer.c] Merge pass_vectorize::execute with vectorize_loops and replace occurences of cfun with function param.

gcc/ChangeLog:
* tree-ssa-loop.c (pass_vectorize): Move to tree-vectorizer.c.
(pass_data_vectorize): Likewise.
(make_pass_vectorize): Likewise.
* tree-vectorizer.c (vectorize_loops): Merge with
pass_vectorize::execute and replace cfun occurences with fun param.
(adjust_simduid_builtins): Add fun param, replace cfun occurences with
fun, and adjust callers approrpiately.
(note_simd_array_uses): Likewise.
(vect_loop_dist_alias_call): Likewise.
(set_uid_loop_bbs): Likewise.
(vect_transform_loops): Likewise.
(try_vectorize_loop_1): Likewise.
(try_vectorize_loop): Likewise.

libffi: Use #define instead of .macro in  src/x86/win64.S [PR102874]

The libffi 3.4.2 import badly broke Solaris/x86 bootstrap with the native
assembler:

Assembler:
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 88 :
Illegal mnemonic
        Near line: ".macro epilogue"
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 88 : Syntax
error
        Near line: ".macro epilogue"
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 95 :
Illegal mnemonic
        Near line: ".endm"
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 95 : Syntax
error
        Near line: ".endm"
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 100 :
Illegal mnemonic
        Near line: " epilogue"
        "/vol/gcc/src/hg/master/local/libffi/src/x86/win64.S", line 100 :
Syntax error
        Near line: "epilogue"

Solaris as doesn't support .macro/.endm.

Fixed by using #define instead of the unportable .macro.

Tested on i386-pc-solaris2.11 and x86_64-pc-linux-gnu.

The bug has been reported upstream
(https://github.com/libffi/libffi/issues/665); a corresponding pull
request is also pending (https://github.com/libffi/libffi/pull/669).

2021-10-21  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

libffi:
PR libffi/102874
* src/x86/win64.S (epilogue): Use #define instead of .macro.

testsuite: i386: Require dfp in gcc.target/i386/pr101346.c

gcc.target/i386/pr101346.c currently FAILs on Solaris/x86:

FAIL: gcc.target/i386/pr101346.c (test for excess errors)

Excess errors:
/vol/gcc/src/hg/master/local/gcc/testsuite/gcc.target/i386/pr101346.c:6:1:
error: decimal floating-point not supported for this target
/vol/gcc/src/hg/master/local/gcc/testsuite/gcc.target/i386/pr101346.c:7:6:
error: decimal floating-point not supported for this target
/vol/gcc/src/hg/master/local/gcc/testsuite/gcc.target/i386/pr101346.c:9:12:
warning: implicit declaration of function '__builtin_fabsd128'; did you
mean '__builtin_fabsf128'? [-Wimplicit-function-declaration]

Fixed by requiring dfp support.  Tested on i386-pc-solaris2.11 and
x86_64-pc-linux-gnu.

2021-10-20  Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

gcc/testsuite:
* gcc.target/i386/pr101346.c: Require dfp support.

i386: Fix up x86 atomic_bit_test* expanders for !TARGET_HIMODE_MATH [PR103205]

With !TARGET_HIMODE_MATH, the OPTAB_DIRECT expand_simple_binop fail and so
we ICE.  We don't really care if they are done promoted in SImode instead.

2021-11-15  Jakub Jelinek  <jakub@redhat.com>

PR target/103205
* config/i386/sync.md (atomic_bit_test_and_set<mode>,
atomic_bit_test_and_complement<mode>,
atomic_bit_test_and_reset<mode>): Use OPTAB_WIDEN instead of
OPTAB_DIRECT.

* gcc.target/i386/pr103205.c: New test.

libgomp, nvptx: Honor OpenMP 5.1 num_teams lower bound

Here is a PTX implementation of what I was talking about, that for
num_teams_upper 0 or whenever num_teams_lower <= num_blocks, the current
implementation is fine but if the user explicitly asks for more
teams than we can provide in hardware, we need to stop assuming that
omp_get_team_num () is equal to the hw team id, but instead need to use some
team specific memory (it is .shared for PTX), or if none is
provided, array indexed by the hw team id and run some teams serially within
the same hw thread.

2021-11-15  Jakub Jelinek  <jakub@redhat.com>

* config/nvptx/team.c (__gomp_team_num): Define as
__attribute__((shared)) var.
(gomp_nvptx_main): Initialize __gomp_team_num to 0.
* config/nvptx/target.c (__gomp_team_num): Declare as
extern __attribute__((shared)) var.
(GOMP_teams4): Use __gomp_team_num as the team number instead of
%ctaid.x.  If first, initialize it to %ctaid.x.  If num_teams_lower
is bigger than num_blocks, use num_teams_lower teams and arrange for
bumping of __gomp_team_num if !first and returning false once we run
out of teams.
* config/nvptx/teams.c (__gomp_team_num): Declare as
extern __attribute__((shared)) var.
(omp_get_team_num): Return __gomp_team_num value instead of %ctaid.x.

libgomp: Add a testcase for omp_get_num_teams inside of target inside of host teams

This is https://github.com/OpenMP/spec/issues/3183
There is an agreement that we should return 1 team inside of target,
even if that target is inside of host teams.  We were doing that
when offloading and not during host fallback, r12-5151 should fix that
even for host fallback.

2021-11-15  Jakub Jelinek  <jakub@redhat.com>

* testsuite/libgomp.c/teams-5.c: New test.

c++: location of lambda object and conversion call

Two things that had poor location info: we weren't giving the TARGET_EXPR
for a lambda object any location, and the call to a conversion function was
getting whatever input_location happened to be.

gcc/cp/ChangeLog:

* call.c (perform_implicit_conversion_flags): Use the location of
the argument.
* lambda.c (build_lambda_object): Set location on the TARGET_EXPR.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/lambda/lambda-switch.C: Adjust expected location.

c++: check constexpr constructor body

The implicit constexpr patch revealed that our checks for constexpr
constructors that could possibly produce a constant value (which
otherwise are IFNDR) was failing to look at most of the function body.
Fixing that required some library tweaks.

gcc/cp/ChangeLog:

* constexpr.c (maybe_save_constexpr_fundef): Also check whether the
body of a constructor is potentially constant.

libstdc++-v3/ChangeLog:

* src/c++17/memory_resource.cc: Add missing constexpr.
* include/experimental/internet: Only mark copy constructor
as constexpr with __cpp_constexpr_dynamic_alloc.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/constexpr-89285-2.C: Expect error.
* g++.dg/cpp1y/constexpr-89285.C: Adjust error.

c++: is_this_parameter and coroutines proxies

Compiling coroutines/pr95736.C with the implicit constexpr patch broke
because is_this_parameter didn't recognize the coroutines proxy for 'this'.

gcc/cp/ChangeLog:

* semantics.c (is_this_parameter): Check DECL_HAS_VALUE_EXPR_P
instead of is_capture_proxy.

c++: c++20 constexpr default ctor and array init

The implicit constexpr patch revealed that marking the constructor in the
PR70690 testcase as constexpr made the bug reappear, because build_vec_init
assumed that a constexpr default constructor initialized the whole object,
so it was equivalent to value-initialization.  But this is no longer true in
C++20.

PR c++/70690

gcc/cp/ChangeLog:

* init.c (build_vec_init): Check default_init_uninitialized_part in
C++20.

gcc/testsuite/ChangeLog:

* g++.dg/init/array41a.C: New test.

c++: don't do constexpr folding in unevaluated context

The implicit constexpr patch revealed that we were doing constant evaluation
of arbitrary expressions in unevaluated contexts, leading to failure when we
tried to evaluate e.g. a call to declval.  This is wrong more generally;
only manifestly-constant-evaluated expressions should be evaluated within
an unevaluated operand.

Making this change revealed a case we were failing to mark as manifestly
constant-evaluated.

gcc/cp/ChangeLog:

* constexpr.c (maybe_constant_value): Don't evaluate
in an unevaluated operand unless manifestly const-evaluated.
(fold_non_dependent_expr_template): Likewise.
* decl.c (compute_array_index_type_loc): This context is
manifestly constant-evaluated.

c++: constexpr virtual and vbase thunk

C++20 allows virtual functions to be constexpr.  I don't think that calling
through a pointer to a vbase subobject is supposed to work in a constant
expression, since an object with virtual bases can't be constant, but the
call shouldn't ICE.

gcc/cp/ChangeLog:

* constexpr.c (cxx_eval_thunk_call): Error instead of ICE
on vbase thunk to constexpr function.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/constexpr-virtual20.C: New test.

gcc.dg/uninit-pred-9_b.c: Correct last adjustment for cris-elf

The change at r12-4790 should have done the same change for
CRIS as was done for powerpc64*-*-*.  (Probably MMIX too but
that may have to wait until the next weekend.)

gcc/testsuite:
* gcc.dg/uninit-pred-9_b.c: Correct last adjustment, for CRIS.

VAX: Implement the `-mlra' command-line option

Add the the `-mlra' command-line option for the VAX target, with the
usual semantics of enabling Local Register Allocation, off by default.

LRA remains unstable with the VAX target, with numerous ICEs throughout
the testsuite and worse code produced overall where successful, however
the presence of a command line option to enable it makes it easier to
experiment with it as the compiler does not have to be rebuilt to flip
between the old reload and LRA.

gcc/
* config/vax/vax.c (vax_lra_p): New prototype and function.
(TARGET_LRA_P): Wire it.
* config/vax/vax.opt (mlra): New option.
* doc/invoke.texi (Option Summary, VAX Options): Document the
new option.

Daily bump.

[Commmitted] Move some testcases to torture from tree-ssa

While writing up some testcases, I noticed some newer testcases
just had "dg-do compile/run" on them with dg-options of either -O1
or -O2. Since it is always better to run them over all optimization
levels I put them in gcc.c-torture/compile or gcc.c-torture/execute.

Committed after testing to make sure the testcases pass.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/pr100278.c: Move to ...
* gcc.c-torture/compile/pr100278.c: Here.
Remove dg-do and dg-options.
* gcc.dg/tree-ssa/pr101189.c: Move to ...
* gcc.c-torture/compile/pr101189.c: Here.
Remove dg-do and dg-options.
* gcc.dg/tree-ssa/pr100453.c: Move to ...
* gcc.c-torture/execute/pr100453.c: Here.
Remove dg-do and dg-options.
* gcc.dg/tree-ssa/pr101335.c: Move to ...
* gcc.c-torture/execute/pr101335.c: Here
Remove dg-do and dg-options.

Track nondeterminism and interposable calls in ipa-modref

Adds tracking of two new flags in ipa-modref: nondeterministic and
calls_interposable.  First is set when function does something that is not
guaranteed to be the same if run again (volatile memory access, volatile asm or
external function call).  Second is set if function calls something that
does not bind to current def.

nondeterministic enables ipa-modref to discover looping pure/const functions
and it now discovers 138 of them during cc1plus link (which about doubles
number of such functions detected late).  We however can do more

 1) We can extend FRE to eliminate redundant calls.
    I filled a PR103168 for that.
    A common case are inline functions that are not autodetected as ECF_CONST
    just becuase they do not bind to local def and can be easily handled.
    More tricky is to use modref summary to check what memory locations are
    read.
 2) DSE can eliminate redundant stores

The calls_interposable flag currently also improves tree-ssa-structalias
on functions that are not binds_to_current_def since reads_global_memory
is now not cleared by interposable functions.

gcc/ChangeLog:

* ipa-modref.h (struct modref_summary): Add nondeterministic
and calls_interposable flags.
* ipa-modref.c (modref_summary::modref_summary): Initialize new flags.
(modref_summary::useful_p): Check new flags.
(struct modref_summary_lto): Add nondeterministic and
calls_interposable flags.
(modref_summary_lto::modref_summary_lto): Initialize new flags.
(modref_summary_lto::useful_p): Check new flags.
(modref_summary::dump): Dump new flags.
(modref_summary_lto::dump): Dump new flags.
(ignore_nondeterminism_p): New function.
(merge_call_side_effects): Merge new flags.
(process_fnspec): Likewise.
(analyze_load): Volatile access is nondeterministic.
(analyze_store): Liekwise.
(analyze_stmt): Volatile ASM is nondeterministic.
(analyze_function): Clear new flags.
(modref_summaries::duplicate): Duplicate new flags.
(modref_summaries_lto::duplicate): Duplicate new flags.
(modref_write): Stream new flags.
(read_section): Stream new flags.
(propagate_unknown_call): Update new flags.
(modref_propagate_in_scc): Propagate new flags.
* tree-ssa-alias.c (ref_maybe_used_by_call_p_1): Check
calls_interposable.
* tree-ssa-structalias.c (determine_global_memory_access):
Likewise.

VAX: Add the `setmemhi' instruction

The MOVC5 machine instruction has `memset' semantics if encoded with a
zero source length[1]:

"4. MOVC5 with a zero source length operand is the preferred way
    to fill a block of memory with the fill character."

Use that instruction to implement the `setmemhi' instruction then.  Use
the AP register in the register deferred mode for the source address to
yield the shortest possible encoding of the otherwise unused operand,
observing that the address is never dereferenced if the source length is
zero.

The use of this instruction yields steadily better performance, at least
with the Mariah VAX implementation, for a variable-length `memset' call
expanded inline as a single MOVC5 operation compared to an equivalent
libcall invocation:

Length:   1, time elapsed:  0.971789 (builtin),  2.847303 (libcall)
Length:   2, time elapsed:  0.907904 (builtin),  2.728259 (libcall)
Length:   3, time elapsed:  1.038311 (builtin),  2.917245 (libcall)
Length:   4, time elapsed:  0.775305 (builtin),  2.686088 (libcall)
Length:   7, time elapsed:  1.112331 (builtin),  2.992968 (libcall)
Length:   8, time elapsed:  0.856882 (builtin),  2.764885 (libcall)
Length:  15, time elapsed:  1.256086 (builtin),  3.096660 (libcall)
Length:  16, time elapsed:  1.001962 (builtin),  2.888131 (libcall)
Length:  31, time elapsed:  1.590456 (builtin),  3.774164 (libcall)
Length:  32, time elapsed:  1.288909 (builtin),  3.629622 (libcall)
Length:  63, time elapsed:  3.430285 (builtin),  5.269789 (libcall)
Length:  64, time elapsed:  3.265147 (builtin),  5.113156 (libcall)
Length: 127, time elapsed:  6.438772 (builtin),  8.268305 (libcall)
Length: 128, time elapsed:  6.268991 (builtin),  8.114557 (libcall)
Length: 255, time elapsed: 12.417338 (builtin), 14.259678 (libcall)

(times given in seconds per 1000000 `memset' invocations for the given
length made in a loop).  It is clear from these figures that hardware
does data coalescence for consecutive bytes rather than naively copying
them one by one, as for lengths that are powers of 2 the figures are
consistently lower than ones for their respective next lower lengths.

The use of MOVC5 also requires at least 4 bytes less in terms of machine
code as it avoids encoding the address of `memset' needed for the CALLS
instruction used to make a libcall, as well as extra PUSHL instructions
needed to pass arguments to the call as those can be encoded directly as
the respective operands of the MOVC5 instruction.

It is perhaps worth noting too that for constant lengths we prefer to
emit up to 5 individual MOVx instructions rather than a single MOVC5
instruction to clear memory and for consistency we copy this behavior
here for filling memory with another value too, even though there may be
a performance advantage with a string copy in comparison to a piecemeal
copy, e.g.:

Length:  40, time elapsed:  2.183192 (string),   2.638878 (piecemeal)

But this is something for another change as it will have to be carefully
evaluated.

[1] DEC STD 032-0 "VAX Architecture Standard", Digital Equipment
    Corporation, A-DS-EL-00032-00-0 Rev J, December 15, 1989, Section
    3.10 "Character-String Instructions", p. 3-163

gcc/
* config/vax/vax.h (SET_RATIO): New macro.
* config/vax/vax.md (UNSPEC_SETMEM_FILL): New constant.
(setmemhi): New expander.
(setmemhi1): New insn and splitter.
(*setmemhi1): New insn.

gcc/testsuite/
* gcc.target/vax/setmem.c: New test.

libstdc++: [_GLIBCXX_DEBUG] Remove _Safe_container<>::_M_safe()

_GLIBCXX_DEBUG container code cleanup to get rid of _Safe_container<>::_M_safe() and just
use _Safe:: calls which use normal inheritance. Also remove several usages of _M_base()
which can be most of the time ommitted and sometimes replace with explicit _Base::
calls.

libstdc++-v3/ChangeLog:

* include/debug/safe_container.h (_Safe_container<>::_M_safe): Remove.
* include/debug/deque (deque::operator=(initializer_list<>)): Replace
_M_base() call with _Base:: call.
(deque::operator[](size_type)): Likewise.
* include/debug/forward_list (forward_list(forward_list&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(forward_list::operator=(initializer_list<>)): Remove _M_base() calls.
(forward_list::splice_after, forward_list::merge): Likewise.
* include/debug/list (list(list&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(list::operator=(initializer_list<>)): Remove _M_base() calls.
(list::splice, list::merge): Likewise.
* include/debug/map.h (map(map&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(map::operator=(initializer_list<>)): Remove _M_base() calls.
* include/debug/multimap.h (multimap(multimap&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(multimap::operator=(initializer_list<>)): Remove _M_base() calls.
* include/debug/set.h (set(set&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(set::operator=(initializer_list<>)): Remove _M_base() calls.
* include/debug/multiset.h (multiset(multiset&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(multiset::operator=(initializer_list<>)): Remove _M_base() calls.
* include/debug/string (basic_string(basic_string&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(basic_string::operator=(initializer_list<>)): Remove _M_base() call.
(basic_string::operator=(const _CharT*), basic_string::operator=(_CharT)): Likewise.
(basic_string::operator[](size_type), basic_string::operator+=(const basic_string&)):
Likewise.
(basic_string::operator+=(const _Char*), basic_string::operator+=(_CharT)): Likewise.
* include/debug/unordered_map (unordered_map(unordered_map&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(unordered_map::operator=(initializer_list<>), unordered_map::merge):
Remove _M_base() calls.
(unordered_multimap(unordered_multimap&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(unordered_multimap::operator=(initializer_list<>), unordered_multimap::merge):
Remove _M_base() calls.
* include/debug/unordered_set (unordered_set(unordered_set&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(unordered_set::operator=(initializer_list<>), unordered_set::merge):
Remove _M_base() calls.
(unordered_multiset(unordered_multiset&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(unordered_multiset::operator=(initializer_list<>), unordered_multiset::merge):
Remove _M_base() calls.
* include/debug/vector (vector(vector&&, const allocator_type&)):
Remove _M_safe() and _M_base() calls.
(vector::operator=(initializer_list<>)): Remove _M_base() calls.
(vector::operator[](size_type)): Likewise.

Extend modref to track kills

This patch adds kill tracking to ipa-modref.  This is representd by array
of accesses to memory locations that are known to be overwritten by the
function.

gcc/ChangeLog:

2021-11-14  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref-tree.c (modref_access_node::update_for_kills): New
member function.
(modref_access_node::merge_for_kills): Likewise.
(modref_access_node::insert_kill): Likewise.
* ipa-modref-tree.h (modref_access_node::update_for_kills,
modref_access_node::merge_for_kills, modref_access_node::insert_kill):
Declare.
(modref_access_node::useful_for_kill): New member function.
* ipa-modref.c (modref_summary::useful_p): Release useless kills.
(lto_modref_summary): Add kills.
(modref_summary::dump): Dump kills.
(record_access): Add mdoref_access_node parameter.
(record_access_lto): Likewise.
(merge_call_side_effects): Merge kills.
(analyze_call): Add ALWAYS_EXECUTED param and pass it around.
(struct summary_ptrs): Add always_executed filed.
(analyze_load): Update.
(analyze_store): Update; record kills.
(analyze_stmt): Add always_executed; record kills in clobbers.
(analyze_function): Track always_executed.
(modref_summaries::duplicate): Duplicate kills.
(update_signature): Release kills.
* ipa-modref.h (struct modref_summary): Add kills.
* tree-ssa-alias.c (alias_stats): Add kill stats.
(dump_alias_stats): Dump kill stats.
(store_kills_ref_p): Break out from ...
(stmt_kills_ref_p): Use it; handle modref info based kills.

gcc/testsuite/ChangeLog:

2021-11-14  Jan Hubicka  <hubicka@ucw.cz>

* gcc.dg/tree-ssa/modref-dse-3.c: New test.

Remove gcc.dg/pr103229.c

gcc/testsuite/ChangeLog:

* gcc.dg/pr103229.c: Removed.

Do not pass NULL to memset in ssa_global_cache.

The code computing ranges in PHIs in the path solver reuses the
temporary ssa_global_cache by calling its clear method.  Calling it on
an empty cache causes us to call memset with NULL.

Tested on x86-64 Linux.

gcc/ChangeLog:

PR tree-optimization/103229
* gimple-range-cache.cc (ssa_global_cache::clear): Do not pass
null value to memset.

gcc/testsuite/ChangeLog:

* gcc.dg/pr103229.c: New test.

tsan: remove not needed -ldl in options

gcc/testsuite/ChangeLog:

* c-c++-common/tsan/free_race.c: Remove unnecessary -ldl.
* c-c++-common/tsan/free_race2.c: Likewise.

Cleanup tree-ssa-alias and tree-ssa-dse use of modref summary

Move code getting tree op from access_node and stmt to a common place.  I also
commonized logic to build ao_ref. While I was on it I also replaced FOR_EACH_*
by range for since they reads better.

gcc/ChangeLog:

2021-11-14  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref-tree.c (modref_access_node::get_call_arg): New member
function.
(modref_access_node::get_ao_ref): Likewise.
* ipa-modref-tree.h (modref_access_node::get_call_arg): Declare.
(modref_access_node::get_ao_ref): Declare.
* tree-ssa-alias.c (modref_may_conflict): Use new accessors.
* tree-ssa-dse.c (dse_optimize_call): Use new accessors.

gcc/testsuite/ChangeLog:

2021-11-14  Jan Hubicka  <hubicka@ucw.cz>

* c-c++-common/asan/null-deref-1.c: Update template.
* c-c++-common/tsan/free_race.c: Update template.
* c-c++-common/tsan/free_race2.c: Update template.
* gcc.dg/ipa/ipa-sra-4.c: Update template.

Daily bump.

Fix bug in ipa-pure-const and add debug counters

gcc/ChangeLog:

PR lto/103211
* dbgcnt.def (ipa_attr): New counters.
* ipa-pure-const.c: Include dbgcnt.c
(ipa_make_function_const): Use debug counter.
(ipa_make_function_pure): Likewise.
(propagate_pure_const): Fix bug in my previous change.

More ipa-modref-tree.h cleanups

Move access dumping to member function and cleanup formating.

gcc/ChangeLog:

2021-11-13  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref-tree.c (modref_access_node::range_info_useful_p):
Offline from ipa-modref-tree.h.
(modref_access_node::dump): Move from ipa-modref.c; make member
function.
* ipa-modref-tree.h (modref_access_node::range_info_useful_p.
modref_access_node::dump): Declare.
* ipa-modref.c (dump_access): Remove.
(dump_records): Update.
(dump_lto_records): Update.
(record_access): Update.
(record_access_lto): Update.

Implement DSE of dead functions calls storing memory.

gcc/ChangeLog:

2021-11-13  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref.c (modref_summary::modref_summary): Clear new flags.
(modref_summary::dump): Dump try_dse.
(modref_summary::finalize): Add FUN attribute; compute try-dse.
(analyze_function): Update.
(read_section): Update.
(update_signature): Update.
(pass_ipa_modref::execute): Update.
* ipa-modref.h (struct modref_summary):
* tree-ssa-alias.c (ao_ref_init_from_ptr_and_range): Export.
* tree-ssa-alias.h (ao_ref_init_from_ptr_and_range): Declare.
* tree-ssa-dse.c (dse_optimize_call): New function.
(dse_optimize_stmt): Use it.

gcc/testsuite/ChangeLog:

2021-11-13  Jan Hubicka  <hubicka@ucw.cz>

* g++.dg/cpp1z/inh-ctor23.C: Fix template
* g++.dg/ipa/ipa-icf-4.C: Fix template
* gcc.dg/tree-ssa/modref-dse-1.c: New test.
* gcc.dg/tree-ssa/modref-dse-2.c: New test.

Fix checking disabled build.

gcc/ChangeLog:

2021-11-13  Jan Hubicka  <hubicka@ucw.cz>

* ipa-modref-tree.c: Move #if CHECKING_P to proper place.

fixincludes: simplify handling for access() failure [PR21283, PR80047]

POSIX says:

    On some implementations, if buf is a null pointer, getcwd() may obtain
    size bytes of memory using malloc(). In this case, the pointer returned
    by getcwd() may be used as the argument in a subsequent call to free().
    Invoking getcwd() with buf as a null pointer is not recommended in
    conforming applications.

This produces an error building GCC with --enable-werror-always:

    ../../../fixincludes/fixincl.c: In function ‘process’:
    ../../../fixincludes/fixincl.c:1356:7: error: argument 1 is null but
    the corresponding size argument 2 value is 4096 [-Werror=nonnull]

It's suggested by POSIX to call getcwd() with progressively larger
buffers until it does not give an [ERANGE] error. However, it's highly
unlikely that this error-handling route is ever used.

So we can simplify it instead of writting too much code.  We give up to
use getcwd(), because `make` will output a `Leaving directory ...` message
containing the path to cwd when we call abort().

fixincludes/ChangeLog:

PR other/21823
PR bootstrap/80047
* fixincl.c (process): Simplify the handling for highly
  unlikely access() failure, to avoid using non-standard
  extensions.

modref_access_node cleanup

move member functions of modref_access_node from ipa-modref-tree.h to
ipa-modref-tree.c since they become long and not fitting for inlines anyway.  I
also cleaned up the interface by making static insert method (which handles
inserting accesses into a vector and optimizing them) which makes it possible
to hide most of the interface handling interval merging private.

Honza

gcc/ChangeLog:

* ipa-modref-tree.h
(struct modref_access_node): Move longer member functions to
ipa-modref-tree.c
(modref_ref_node::try_merge_with): Turn into modreef_acces_node member
function.
* ipa-modref-tree.c (modref_access_node::contains): Move here
from ipa-modref-tree.h.
(modref_access_node::update): Likewise.
(modref_access_node::merge): Likewise.
(modref_access_node::closer_pair_p): Likewise.
(modref_access_node::forced_merge): Likewise.
(modref_access_node::update2): Likewise.
(modref_access_node::combined_offsets): Likewise.
(modref_access_node::try_merge_with): Likewise.
(modref_access_node::insert): Likewise.

Add finalize method to modref summary.

gcc/ChangeLog:

* ipa-modref.c (modref_summary::global_memory_read_p): Remove.
(modref_summary::global_memory_written_p): Remove.
(modref_summary::dump): Dump new flags.
(modref_summary::finalize): New member function.
(analyze_function): Call it.
(read_section): Call it.
(update_signature): Call it.
(pass_ipa_modref::execute): Call it.
* ipa-modref.h (struct modref_summary): Remove
global_memory_read_p and global_memory_written_p.
Add global_memory_read, global_memory_written.
* tree-ssa-structalias.c (determine_global_memory_access):
Update.

Whitelity type attributes for function signature change

gcc/ChangeLog:

* ipa-fnsummary.c (compute_fn_summary): Use type_attribut_allowed_p
* ipa-param-manipulation.c
(ipa_param_adjustments::type_attribute_allowed_p):
New member function.
(drop_type_attribute_if_params_changed_p): New function.
(build_adjusted_function_type): Use it.
* ipa-param-manipulation.h: Add type_attribute_allowed_p.

analyzer: add four new taint-based warnings

The initial commit of the analyzer in GCC 10 had a single warning,
  -Wanalyzer-tainted-array-index
and required manually enabling the taint checker with
-fanalyzer-checker=taint (due to scaling issues).

This patch extends the taint detection to add four new taint-based
warnings:

  -Wanalyzer-tainted-allocation-size
     for e.g. attacker-controlled malloc/alloca
  -Wanalyzer-tainted-divisor
     for detecting where an attacker can inject a divide-by-zero
  -Wanalyzer-tainted-offset
     for attacker-controlled pointer offsets
  -Wanalyzer-tainted-size
     for e.g. attacker-controlled memset

and rewords all the warnings to talk about "attacker-controlled" values
rather than "tainted" values.

Unfortunately I haven't yet addressed the scaling issues, so all of
these still require -fanalyzer-checker=taint (in addition to -fanalyzer).

gcc/analyzer/ChangeLog:
* analyzer.opt (Wanalyzer-tainted-allocation-size): New.
(Wanalyzer-tainted-divisor): New.
(Wanalyzer-tainted-offset): New.
(Wanalyzer-tainted-size): New.
* engine.cc (impl_region_model_context::get_taint_map): New.
* exploded-graph.h (impl_region_model_context::get_taint_map):
New decl.
* program-state.cc (sm_state_map::get_state): Call
alt_get_inherited_state.
(sm_state_map::impl_set_state): Modify states within
compound svalues.
(program_state::impl_call_analyzer_dump_state): Undo casts.
(selftest::test_program_state_1): Update for new context param of
create_region_for_heap_alloc.
(selftest::test_program_state_merging): Likewise.
* region-model-impl-calls.cc (region_model::impl_call_alloca):
Likewise.
(region_model::impl_call_calloc): Likewise.
(region_model::impl_call_malloc): Likewise.
(region_model::impl_call_operator_new): Likewise.
(region_model::impl_call_realloc): Likewise.
* region-model.cc (region_model::check_region_access): Call
check_region_for_taint.
(region_model::get_representative_path_var_1): Handle binops.
(region_model::create_region_for_heap_alloc): Add "ctxt" param and
pass it to set_dynamic_extents.
(region_model::create_region_for_alloca): Likewise.
(region_model::set_dynamic_extents): Add "ctxt" param and use it
to call check_dynamic_size_for_taint.
(selftest::test_state_merging): Update for new context param of
create_region_for_heap_alloc.
(selftest::test_malloc_constraints): Likewise.
(selftest::test_malloc): Likewise.
(selftest::test_alloca): Likewise for create_region_for_alloca.
* region-model.h (region_model::create_region_for_heap_alloc): Add
"ctxt" param.
(region_model::create_region_for_alloca): Likewise.
(region_model::set_dynamic_extents): Likewise.
(region_model::check_dynamic_size_for_taint): New decl.
(region_model::check_region_for_taint): New decl.
(region_model_context::get_taint_map): New vfunc.
(noop_region_model_context::get_taint_map): New.
* sm-taint.cc: Remove include of "diagnostic-event-id.h"; add
includes of "gimple-iterator.h", "tristate.h", "selftest.h",
"ordered-hash-map.h", "cgraph.h", "cfg.h", "digraph.h",
"analyzer/supergraph.h", "analyzer/call-string.h",
"analyzer/program-point.h", "analyzer/store.h",
"analyzer/region-model.h", and "analyzer/program-state.h".
(enum bounds): Move to top of file.
(class taint_diagnostic): New.
(class tainted_array_index): Convert to subclass of taint_diagnostic.
(tainted_array_index::emit): Add CWE-129.  Reword warning to use
"attacker-controlled" rather than "tainted".
(tainted_array_index::describe_state_change): Move to
taint_diagnostic::describe_state_change.
(tainted_array_index::describe_final_event): Reword to use
"attacker-controlled" rather than "tainted".
(class tainted_offset): New.
(class tainted_size): New.
(class tainted_divisor): New.
(class tainted_allocation_size): New.
(taint_state_machine::alt_get_inherited_state): New.
(taint_state_machine::on_stmt): In assignment handling, remove
ARRAY_REF handling in favor of check_region_for_taint.  Add
detection of tainted divisors.
(taint_state_machine::get_taint): New.
(taint_state_machine::combine_states): New.
(region_model::check_region_for_taint): New.
(region_model::check_dynamic_size_for_taint): New.
* sm.h (state_machine::alt_get_inherited_state): New.

gcc/ChangeLog:
* doc/invoke.texi (Static Analyzer Options): Add
-Wno-analyzer-tainted-allocation-size,
-Wno-analyzer-tainted-divisor, -Wno-analyzer-tainted-offset, and
-Wno-analyzer-tainted-size to list.  Add
-Wanalyzer-tainted-allocation-size, -Wanalyzer-tainted-divisor,
-Wanalyzer-tainted-offset, and -Wanalyzer-tainted-size to list
of options effectively enabled by -fanalyzer.
(-Wanalyzer-tainted-allocation-size): New.
(-Wanalyzer-tainted-array-index): Tweak wording; add link to CWE.
(-Wanalyzer-tainted-divisor): New.
(-Wanalyzer-tainted-offset): New.
(-Wanalyzer-tainted-size): New.

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/pr93382.c: Tweak expected wording.
* gcc.dg/analyzer/taint-alloc-1.c: New test.
* gcc.dg/analyzer/taint-alloc-2.c: New test.
* gcc.dg/analyzer/taint-divisor-1.c: New test.
* gcc.dg/analyzer/taint-1.c: Rename to...
* gcc.dg/analyzer/taint-read-index-1.c: ...this.  Tweak expected
wording.  Mark some events as xfail.
* gcc.dg/analyzer/taint-read-offset-1.c: New test.
* gcc.dg/analyzer/taint-size-1.c: New test.
* gcc.dg/analyzer/taint-write-index-1.c: New test.
* gcc.dg/analyzer/taint-write-offset-1.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

Remember fnspec based EAF flags in modref summary.

gcc/ChangeLog:

* attr-fnspec.h (attr_fnspec::arg_eaf_flags): Break out from ...
* gimple.c (gimple_call_arg_flags): ... here.
* ipa-modref.c (analyze_parms): Record flags known from fnspec.
(modref_merge_call_site_flags): Use arg_eaf_flags.

path solver: Compute all PHI ranges simultaneously.

PHIs must be resolved simulatenously, otherwise we may not pick up the
ranges incoming to the block.

For example.  If we put p3_7 in the cache before all PHIs have been
computed, we will pick up the wrong p3_7 value for p2_17:

    # p3_7 = PHI <1(2), 0(5)>
    # p2_17 = PHI <1(2), p3_7(5)>

This patch delays updating the cache until all PHIs have been
analyzed.

gcc/ChangeLog:

PR tree-optimization/103222
* gimple-range-path.cc (path_range_query::compute_ranges_in_phis):
New.
(path_range_query::compute_ranges_in_block): Call
compute_ranges_in_phis.
* gimple-range-path.h (path_range_query::compute_ranges_in_phis):
New.

gcc/testsuite/ChangeLog:

* gcc.dg/pr103222.c: New test.

libsanitizer: Update LOCAL_PATCHES

* LOCAL_PATCHES: Update to the corresponding revision.

libsanitizer: Apply local patches

libsanitizer: Merge with upstream

Merged revision: 82bc6a094e85014f1891ef9407496f44af8fe442

with the fix for PR sanitizer/102911

libstdc++: Implement std::spanstream for C++23

This implements the <spanstream> header, as proposed for C++23 by P0448R4.

libstdc++-v3/ChangeLog:

* include/Makefile.am: Add spanstream header.
* include/Makefile.in: Regenerate.
* include/precompiled/stdc++.h: Add spanstream header.
* include/std/version (__cpp_lib_spanstream): Define.
* include/std/spanstream: New file.
* testsuite/27_io/spanstream/1.cc: New test.
* testsuite/27_io/spanstream/version.cc: New test.

Enable ipa-sra with fnspec attributes

Enable some ipa-sra on fortran by allowing signature changes on functions
with "fn spec" attribute when ipa-modref is enabled.  This is possible since ipa-modref
knows how to preserve things we trace in fnspec and fnspec generated by fortran forntend
are quite simple and can be analysed automatically now.  To be sure I will also add
code that merge fnspec to parameters.

This unfortunately hits bug in ipa-param-manipulation when we remove parameter
that specifies size of variable length parameter. For this reason I added a hack
that prevent signature changes on such functions and will handle it incrementally.

I tried creating C testcase but it is blocked by another problem that we punt ipa-sra
on access attribute.  This is optimization regression we ought to fix so I filled
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103223.

As a followup I will add code classifying the type attributes (we have just few) and
get stats on access attribute.

gcc/ChangeLog:

* ipa-fnsummary.c (compute_fn_summary): Do not give up on signature
changes on "fn spec" attribute; give up on varadic types.
* ipa-param-manipulation.c: Include attribs.h.
(build_adjusted_function_type): New parameter ARG_MODIFIED; if it is
true remove "fn spec" attribute.
(ipa_param_adjustments::build_new_function_type): Update.
(ipa_param_body_adjustments::modify_formal_parameters): update.
* ipa-sra.c: Include attribs.h.
(ipa_sra_preliminary_function_checks): Do not check for TYPE_ATTRIBUTES.

path solver: Merge path_range_query constructors.

There's no need for two constructors, when we can do it all with one
that defaults to the common behavior:

path_range_query (bool resolve = true, gimple_ranger *ranger = NULL);

Tested on x86-64 Linux.

gcc/ChangeLog:

* gimple-range-path.cc (path_range_query::path_range_query): Merge
ctors.
(path_range_query::import_p): Move from header file.
(path_range_query::~path_range_query): Adjust for combined ctors.
* gimple-range-path.h: Merge ctors.
(path_range_query::import_p): Move to .cc file.

Fix wrong code with modref and some builtins.

ipa-modref gets confused by EAF flags of memcpy becuase parameter 1 is
escaping but used only directly.  In modref we do not track values saved to
memory and thus we clear all other flags on each store.  This needs to also
happen when called function escapes parameter.

gcc/ChangeLog:

PR tree-optimization/103182
* ipa-modref.c (callee_to_caller_flags): Fix merging of flags.
(modref_eaf_analysis::analyze_ssa_name): Fix merging of flags.

libstdc++: Use GCC_TRY_COMPILE_OR_LINK for getentropy, arc4random

Since r12-5056-g3439657b0286, there has been a regression in
test results; an additional 100 FAILs running the g++ and
libstdc++ testsuite on cris-elf, a newlib target.  The
failures are linker errors, not finding a definition for
getentropy.  It appears newlib has since 2017-12-03
declarations of getentropy and arc4random, and provides an
implementation of arc4random using getentropy, but provides no
definition of getentropy, not even a stub yielding ENOSYS.
This is similar to what it does for many other functions too.

While fixing newlib (like adding said stub) would likely help,
it still leaves older newlib releases hanging.  Thankfully,
the libstdc++ configury test can be improved to try linking
where possible; using the bespoke GCC_TRY_COMPILE_OR_LINK
instead of AC_TRY_COMPILE.  BTW, I see a lack of consistency;
some tests use AC_TRY_COMPILE and some GCC_TRY_COMPILE_OR_LINK
for no apparent reason, but this commit just amends
r12-5056-g3439657b0286.

libstdc++-v3:
PR libstdc++/103166
* acinclude.m4 (GLIBCXX_CHECK_GETENTROPY, GLIBCXX_CHECK_ARC4RANDOM):
Use GCC_TRY_COMPILE_OR_LINK instead of AC_TRY_COMPILE.
* configure: Regenerate.

Daily bump.

or1k: Fix clobbering of _mcount argument if fPIC is enabled

Recently we changed the PROFILE_HOOK _mcount call to pass in the link
register as an argument.  This actually does not work when the _mcount
call uses a PLT because the GOT register setup code ends up getting
inserted before the PROFILE_HOOK and clobbers the link register
argument.

These glibc tests are failing:
  gmon/tst-gmon-pie-gprof
  gmon/tst-gmon-static-gprof

This patch fixes this by saving the instruction that stores the Link
Register to the _mcount argument and then inserts the GOT register setup
instructions after that.

For example:

main.c:

    extern int e;

    int f2(int a) {
      return a + e;
    }

    int f1(int a) {
      return f2 (a + a);
    }

    int main(int argc, char ** argv) {
      return f1 (argc);
    }

Compiled:

    or1k-smh-linux-gnu-gcc -Wall -c -O2 -fPIC -pg -S main.c

Before Fix:

    main:
        l.addi  r1, r1, -16
        l.sw    8(r1), r2
        l.sw    0(r1), r16
        l.addi  r2, r1, 16   # Keeping FP, but not needed
        l.sw    4(r1), r18
        l.sw    12(r1), r9
        l.jal   8            # GOT Setup clobbers r9 (Link Register)
         l.movhi        r16, gotpchi(_GLOBAL_OFFSET_TABLE_-4)
        l.ori   r16, r16, gotpclo(_GLOBAL_OFFSET_TABLE_+0)
        l.add   r16, r16, r9
        l.or    r18, r3, r3
        l.or    r3, r9, r9    # This is not the original LR
        l.jal   plt(_mcount)
         l.nop

        l.jal   plt(f1)
         l.or    r3, r18, r18
        l.lwz   r9, 12(r1)
        l.lwz   r16, 0(r1)
        l.lwz   r18, 4(r1)
        l.lwz   r2, 8(r1)
        l.jr    r9
         l.addi  r1, r1, 16

After the fix:

    main:
        l.addi  r1, r1, -12
        l.sw    0(r1), r16
        l.sw    4(r1), r18
        l.sw    8(r1), r9
        l.or    r18, r3, r3
        l.or    r3, r9, r9    # We now have r9 (LR) set early
        l.jal   8             # Clobbers r9 (Link Register)
         l.movhi        r16, gotpchi(_GLOBAL_OFFSET_TABLE_-4)
        l.ori   r16, r16, gotpclo(_GLOBAL_OFFSET_TABLE_+0)
        l.add   r16, r16, r9
        l.jal   plt(_mcount)
         l.nop

        l.jal   plt(f1)
         l.or    r3, r18, r18
        l.lwz   r9, 8(r1)
        l.lwz   r16, 0(r1)
        l.lwz   r18, 4(r1)
        l.jr    r9
         l.addi  r1, r1, 12

Fixes: 308531d148a ("or1k: Add return address argument to _mcount call")

gcc/ChangeLog:
* config/or1k/or1k-protos.h (or1k_profile_hook): New function.
* config/or1k/or1k.h (PROFILE_HOOK): Change macro to reference
new function or1k_profile_hook.
* config/or1k/or1k.c (struct machine_function): Add new field
set_mcount_arg_insn.
(or1k_profile_hook): New function.
(or1k_init_pic_reg): Update to inject pic rtx after _mcount arg
when profiling.
(or1k_frame_pointer_required): Frame pointer no longer needed
when profiling.

Fix wrong code with pure functions

I introduced bug into find_func_aliases_for_call in handling pure functions.
Instead of reading global memory pure functions are believed to write global
memory.  This results in misoptimization of the testcase at -O1.

The change to pta-callused.c updates the template for new behaviour of the
constraint generation. We copy nonlocal memory to calluse which is correct but
also not strictly necessary because later we take care to add nonlocal_p flag
manually.

gcc/ChangeLog:

PR tree-optimization/103209
* tree-ssa-structalias.c (find_func_aliases_for_call): Fix
use of handle_rhs_call

gcc/testsuite/ChangeLog:

PR tree-optimization/103209
* gcc.dg/tree-ssa/pta-callused.c: Update template.
* gcc.c-torture/execute/pr103209.c: New test.

path solver: Solve PHI imports first for ranges.

PHIs must be resolved first while solving ranges in a block,
regardless of where they appear in the import bitmap.  We went through
a similar exercise for the relational code, but missed these.

Tested on x86-64 & ppc64le Linux.

gcc/ChangeLog:

PR tree-optimization/103202
* gimple-range-path.cc
(path_range_query::compute_ranges_in_block): Solve PHI imports first.

Fix ipa-pure-const

gcc/ChangeLog:

* ipa-pure-const.c (propagate_pure_const): Remove redundant check;
fix call of ipa_make_function_const and ipa_make_function_pure.

analyzer: "__analyzer_dump_state" has no side-effects

gcc/analyzer/ChangeLog:
* engine.cc (exploded_node::on_stmt_pre): Return when handling
"__analyzer_dump_state".

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

aarch64: Remove redundant costing code

Previous patches made some of the complex parts of the issue rate
code redundant.

gcc/
* config/aarch64/aarch64.c (aarch64_vector_op::n_advsimd_ops): Delete.
(aarch64_vector_op::m_seen_loads): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Don't push to
m_advsimd_ops.
(aarch64_vector_op::count_ops): Remove vectype and factor parameters.
Remove code that tries to predict different vec_flags from the
current loop's.
(aarch64_vector_costs::add_stmt_cost): Update accordingly.
Remove m_advsimd_ops handling.

aarch64: Use new hooks for vector comparisons

Previously we tried to account for the different issue rates of
the various vector modes by guessing what the Advanced SIMD version
of an SVE loop would look like and what its issue rate was likely to be.
We'd then increase the cost of the SVE loop if the Advanced SIMD loop
might issue more quickly.

This patch moves that logic to better_main_loop_than_p, so that we
can compare loops side-by-side rather than having to guess.  This also
means we can apply the issue rate heuristics to *any* vector loop
comparison, rather than just weighting SVE vs. Advanced SIMD.

The actual heuristics are otherwise unchanged.  We're just
applying them in a different place.

gcc/
* config/aarch64/aarch64.c (aarch64_vector_costs::m_saw_sve_only_op)
(aarch64_sve_only_stmt_p): Delete.
(aarch64_vector_costs::prefer_unrolled_loop): New function,
extracted from adjust_body_cost.
(aarch64_vector_costs::better_main_loop_than_p): New function,
using heuristics extracted from adjust_body_cost and
adjust_body_cost_sve.
(aarch64_vector_costs::adjust_body_cost_sve): Remove
advsimd_cycles_per_iter and could_use_advsimd parameters.
Update after changes above.
(aarch64_vector_costs::adjust_body_cost): Update after changes above.

aarch64: Add vf_factor to aarch64_vec_op_count

-mtune=neoverse-512tvb sets the likely SVE vector length to 128 bits,
but it also takes into account Neoverse V1, which is a 256-bit target.
This patch adds this VF (VL) factor to aarch64_vec_op_count.

gcc/
* config/aarch64/aarch64.c (aarch64_vec_op_count::m_vf_factor):
New member variable.
(aarch64_vec_op_count::aarch64_vec_op_count): Add a parameter for it.
(aarch64_vec_op_count::vf_factor): New function.
(aarch64_vector_costs::aarch64_vector_costs): When costing for
neoverse-512tvb, pass a vf_factor of 2 for the Neoverse V1 version
of an SVE loop.
(aarch64_vector_costs::adjust_body_cost): Read the vf factor
instead of hard-coding 2.

aarch64: Move cycle estimation into aarch64_vec_op_count

This patch just moves the main cycle estimation routines
into aarch64_vec_op_count.

gcc/
* config/aarch64/aarch64.c
(aarch64_vec_op_count::rename_cycles_per_iter): New function.
(aarch64_vec_op_count::min_nonpred_cycles_per_iter): Likewise.
(aarch64_vec_op_count::min_pred_cycles_per_iter): Likewise.
(aarch64_vec_op_count::min_cycles_per_iter): Likewise.
(aarch64_vec_op_count::dump): Move earlier in file.  Dump the
above properties too.
(aarch64_estimate_min_cycles_per_iter): Delete.
(adjust_body_cost): Use aarch64_vec_op_count::min_cycles_per_iter
instead of aarch64_estimate_min_cycles_per_iter.  Rely on the dump
routine to print CPI estimates.
(adjust_body_cost_sve): Likewise.  Use the other functions above
instead of doing the work inline.

aarch64: Use an array of aarch64_vec_op_counts

-mtune=neoverse-512tvb uses two issue rates, one for Neoverse V1
and one with more generic parameters.  We use both rates when
making a choice between scalar, Advanced SIMD and SVE code.

Previously we calculated the Neoverse V1 issue rates from the
more generic issue rates, but by removing m_scalar_ops and
(later) m_advsimd_ops, it becomes easier to track multiple
issue rates directly.

This patch therefore converts m_ops and (temporarily) m_advsimd_ops
into arrays.

gcc/
* config/aarch64/aarch64.c (aarch64_vec_op_count): Allow default
initialization.
(aarch64_vec_op_count::base_issue_info): Remove handling of null
issue_infos.
(aarch64_vec_op_count::simd_issue_info): Likewise.
(aarch64_vec_op_count::sve_issue_info): Likewise.
(aarch64_vector_costs::m_ops): Turn into a vector.
(aarch64_vector_costs::m_advsimd_ops): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Add entries to
the vectors based on aarch64_tune_params.
(aarch64_vector_costs::analyze_loop_vinfo): Update the pred_ops
of all entries in m_ops.
(aarch64_vector_costs::add_stmt_cost): Call count_ops for all
entries in m_ops.
(aarch64_estimate_min_cycles_per_iter): Remove issue_info
parameter and get the information from the ops instead.
(aarch64_vector_costs::adjust_body_cost_sve): Take a
aarch64_vec_issue_info instead of a aarch64_vec_op_count.
(aarch64_vector_costs::adjust_body_cost): Update call accordingly.
Exit earlier if m_ops is empty for either cost structure.

aarch64: Use real scalar op counts

Now that vector finish_costs is passed the associated scalar costs,
we can record the scalar issue information while computing the scalar
costs, rather than trying to estimate it while computing the vector
costs.

This simplifies things a little, but the main motivation is to improve
accuracy.

gcc/
* config/aarch64/aarch64.c (aarch64_vector_costs::m_scalar_ops)
(aarch64_vector_costs::m_sve_ops): Replace with...
(aarch64_vector_costs::m_ops): ...this.
(aarch64_vector_costs::analyze_loop_vinfo): Update accordingly.
(aarch64_vector_costs::adjust_body_cost_sve): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Likewise.
Initialize m_vec_flags here rather than in add_stmt_cost.
(aarch64_vector_costs::count_ops): Test for scalar reductions too.
Allow vectype to be null.
(aarch64_vector_costs::add_stmt_cost): Call count_ops for scalar
code too.  Don't require vectype to be nonnull.
(aarch64_vector_costs::adjust_body_cost): Take the loop_vec_info
and scalar costs as parameters.  Use the scalar costs to determine
the cycles per iteration of the scalar loop, then multiply it
by the estimated VF.
(aarch64_vector_costs::finish_cost): Update call accordingly.

aarch64: Get floatness from stmt_info

This patch gets the floatness of a memory access from the data
reference rather than the vectype.  This makes it more suitable
for use in scalar costing code.

gcc/
* config/aarch64/aarch64.c (aarch64_dr_type): New function.
(aarch64_vector_costs::count_ops): Use it rather than the
vectype to determine floatness.

aarch64: Remove vectype from latency tests

This patch gets the scalar mode of a reduction operation from the
gimple stmt rather than the vectype.  This makes it more suitable
for use in scalar costs.

gcc/
* config/aarch64/aarch64.c (aarch64_sve_in_loop_reduction_latency):
Remove vectype parameter and get floatness from the type of the
stmt lhs instead.
(arch64_in_loop_reduction_latency): Likewise.
(aarch64_detect_vector_stmt_subtype): Update caller.
(aarch64_vector_costs::count_ops): Likewise.

aarch64: Fold aarch64_sve_op_count into aarch64_vec_op_count

Later patches make aarch64 use the new vector hooks.  We then
only need to track one set of ops for each aarch64_vector_costs
structure.  This in turn means that it's more convenient to merge
aarch64_sve_op_count and aarch64_vec_op_count.

The patch also adds issue info and vec flags to aarch64_vec_op_count,
so that the structure is more self-descriptive.  This simplifies some
things later.

gcc/
* config/aarch64/aarch64.c (aarch64_sve_op_count): Fold into...
(aarch64_vec_op_count): ...this.  Add a constructor.
(aarch64_vec_op_count::vec_flags): New function.
(aarch64_vec_op_count::base_issue_info): Likewise.
(aarch64_vec_op_count::simd_issue_info): Likewise.
(aarch64_vec_op_count::sve_issue_info): Likewise.
(aarch64_vec_op_count::m_issue_info): New member variable.
(aarch64_vec_op_count::m_vec_flags): Likewise.
(aarch64_vector_costs): Add a constructor.
(aarch64_vector_costs::m_sve_ops): Change type to aarch64_vec_op_count.
(aarch64_vector_costs::aarch64_vector_costs): New function.
Initialize m_scalar_ops, m_advsimd_ops and m_sve_ops.
(aarch64_vector_costs::count_ops): Remove vec_flags and
issue_info parameters, using the new aarch64_vec_op_count
functions instead.
(aarch64_vector_costs::add_stmt_cost): Update call accordingly.
(aarch64_sve_op_count::dump): Fold into...
(aarch64_vec_op_count::dump): ..here.

aarch64: Detect more consecutive MEMs

For tests like:

    int res[2];
    void
    f1 (int x, int y)
    {
      res[0] = res[1] = x + y;
    }

we generated:

        add     w0, w0, w1
        adrp    x1, .LANCHOR0
        add     x2, x1, :lo12:.LANCHOR0
        str     w0, [x1, #:lo12:.LANCHOR0]
        str     w0, [x2, 4]
        ret

Using [x1, #:lo12:.LANCHOR0] for the first store prevented the
two stores being recognised as a pair.  However, the MEM_EXPR
and MEM_OFFSET information tell us that the MEMs really are
consecutive.  The peehole2 context then guarantees that the
first address is equivalent to [x2, 0].

While there: the reg_mentioned_p tests for loads were probably correct,
but seemed a bit indirect.  We're matching two consecutive loads,
so the thing we need to test is that the second MEM in the original
sequence doesn't depend on the result of the first load in the
original sequence.

gcc/
* config/aarch64/aarch64.c: Include tree-dfa.h.
(aarch64_check_consecutive_mems): New function that takes MEM_EXPR
and MEM_OFFSET into account.
(aarch64_swap_ldrstr_operands): Use it.
(aarch64_operands_ok_for_ldpstp): Likewise.  Check that the
address of the second memory doesn't depend on the result of
the first load.

gcc/testsuite/
* gcc.target/aarch64/stp_1.c: New test.

Fortran/openmp: Fix '!$omp end'

gcc/fortran/ChangeLog:

* parse.c (decode_omp_directive): Fix permitting 'nowait' for some
combined directives, add missing 'omp end ... loop'.
(gfc_ascii_statement): Fix ST_OMP_END_TEAMS_LOOP result.
* openmp.c (resolve_omp_clauses): Add missing combined loop constructs
case values to the 'if(directive-name: ...)' check.
* trans-openmp.c (gfc_split_omp_clauses): Put nowait on target if
first leaf construct accepting it.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/unexpected-end.f90: Update dg-error.
* gfortran.dg/gomp/clauses-1.f90: New test.
* gfortran.dg/gomp/nowait-2.f90: New test.
* gfortran.dg/gomp/nowait-3.f90: New test.

Fix exit condition in ipa_make_function_pure

gcc/ChangeLog:

* ipa-pure-const.c (ipa_make_function_pure): Fix exit condition.