Align EVRP and ranger for how ranges of ADDR_EXPR are calculated.
gcc/
* gimple-range.cc: (gimple_ranger::range_of_range_op): Check for
ADDR_EXPR and call range_of_address.
(gimple_ranger::range_of_address): Rename from
range_of_non_trivial_assignment and match vrp_stmt_computes_nonzero.
* gimple-range.h: (range_of_address): Renamed.
* range-op.cc: (pointer_table): Add INTEGER_CST handler.
gcc/testsuite/
* gcc.dg/tree-ssa/pr78655.c: New.
This patch makes the entry point to loop invariant motion public, so
that it can be called after loop interchange when that pass has
swapped loops. This avoids the non-LTO -Ofast run-time regressions of
410.bwaves and 503.bwaves_r (which are 19% and 15% faster than current
master on an AMD zen2 machine) while not introducing a full LIM pass
into the pass pipeline.
The patch also adds a parameter which allows not to perform any store
motion so that it is not done after an interchange.
gcc/ChangeLog:
2020-11-12 Martin Jambor <mjambor@suse.cz>
PR tree-optimization/94406
* tree-ssa-loop-im.c (tree_ssa_lim): Renamed to
loop_invariant_motion_in_fun, added a parameter to control store
motion.
(pass_lim::execute): Adjust call to tree_ssa_lim, now
loop_invariant_motion_in_fun.
* tree-ssa-loop-manip.h (loop_invariant_motion_in_fun): Declare.
* gimple-loop-interchange.cc (pass_linterchange::execute): Call
loop_invariant_motion_in_fun if any interchange has been done.
This reduces the number of collisions for PHIs in the VN hashtable
by always hashing the number of predecessors and separately hashing
the block number when we never merge PHIs from different blocks.
This improves collisions seen for the PR69609 testcase dramatically.
2020-11-13 Richard Biener <rguenther@suse.de>
* tree-ssa-sccvn.c (vn_phi_compute_hash): Always hash the
number of predecessors. Hash the block number also for
loop header PHIs.
(expressions_equal_p): Short-cut SSA name compares, remove
test for NULL operands.
(vn_phi_eq): Cache number of predecessors, change inlined
test from expressions_equal_p.
This is so that the library is not to conflict with gcc-10.
libphobos/ChangeLog:
* configure: Regenerate.
* configure.ac (libtool_VERSION): Update to 2:0.0.
For some targets, mathfn_built_in returns NULL as copysign is not
implicitly available, causing an ICE. Now copysign is explicitly
requested when expanding the intrinsic.
gcc/d/ChangeLog:
* intrinsics.cc (expand_intrinsic_copysign): Explicitly determine
which built-in copysign function to call.
Because this what the upstream reference compiler did, thunks for the D
front-end were associated with the class definition, so were forced
code-gen even if the target function was extern. This has now been
changed so there are now only generated if there is a function
definition, fixing the ICE that occurred in PR 97644, which was caused
by calling expand_thunk() early.
gcc/d/ChangeLog:
PR d/97644
* dmd/MERGE: Merge upstream dmd 95044d8e4.
* d-target.cc (TargetCPP::thunkMangle): New function.
* decl.cc (finish_thunk): Don't force expand thunks for external
functions.
(make_thunk): Emit thunks only if the function has a definition.
Generate correct mangling for thunks to C++ classes.
gcc/testsuite/ChangeLog:
* gdc.dg/pr92216.d: Update scan-assember.
gcc/c-family/c-attribs.c:4698:5: warning: misleading indentation; statement is not part of the previous 'if' [-Wmisleading-indentation]
gcc/c-family/ChangeLog:
* c-attribs.c (build_attr_access_from_parms): Format properly.
Some tex tools don't allow the @r{} command to be split across
lines. Fixed by making the change occupy a long line.
gcc/ChangeLog:
* doc/extend.texi: Don't try to line-wrap an @r command.
This makes sure to properly extend the input range before seeing
whether it fits the target.
2020-11-13 Richard Biener <rguenther@suse.de>
PR tree-optimization/97812
* tree-vrp.c (register_edge_assert_for_2): Extend the range
according to its sign before seeing whether it fits.
* gcc.dg/torture/pr97812.c: New testcase.
For linux targets this test doesn't need -lpthread because it only uses
atomics, but for all other targets std::call_once still needs pthreads.
Add the necessary test directives to make that work.
The timings in this test might be too fragile or too target-specific, so
it might need to be adjusted in future, or restricted to only run on
specific targets. For now I've increased the allowed ratio between
wait_for calls before and after the future is made ready, because it was
failing with -O3 -march=native sometimes.
libstdc++-v3/ChangeLog:
* testsuite/30_threads/future/members/poll.cc: Require gthreads
and add -pthread for targets that require it. Relax required
ratio of wait_for calls before/after the future is ready.
This patch implements aarch64 backend expansion for __builtin_memset. Most of
the implementation is based on the expansion of __builtin_memcpy. We change the
values of SET_RATIO and MOVE_RATIO for cases where we do not have to strictly
align and where we can benefit from NEON instructions in the backend.
gcc/ChangeLog:
* config/aarch64/aarch64-protos.h (aarch64_expand_setmem): New
declaration.
* config/aarch64/aarch64.c (aarch64_gen_store_pair): Add case for
E_V16QImode.
(aarch64_set_one_block_and_progress_pointer): New helper for
aarch64_expand_setmem.
(aarch64_expand_setmem): Define the expansion for memset.
* config/aarch64/aarch64.h (CLEAR_RATIO): Tweak to favor
aarch64_expand_setmem when allowed and profitable.
(SET_RATIO): Likewise.
* config/aarch64/aarch64.md: Define pattern for setmemdi.
gcc/testsuite/ChangeLog:
* g++.dg/tree-ssa/pr90883.C: Remove xfail for aarch64.
* gcc.dg/tree-prof/stringop-2.c: Add xfail for aarch64.
* gcc.target/aarch64/memset-corner-cases.c: New test.
* gcc.target/aarch64/memset-q-reg.c: New test.
This part of the implementation covers property nullability attributes
and includes the changes to common code. Follow-on changes will be needed
to cover Objective-C method definitions, but those are expected to be
local to the Objective-C front end.
The basis of the implementation is to translate the Objective-C-specific
keywords into an attribute (objc_nullability) which has the required
states to carry the attribute markup.
We introduce the keywords, and these are parsed and validated in the same
manner as other property attributes. The resulting value is attached to
the property as an objc_nullability attribute.
gcc/c-family/ChangeLog:
PR objc/90707
* c-common.c (c_common_reswords): null_unspecified, nullable,
nonnull, null_resettable: New keywords.
* c-common.h (enum rid): RID_NULL_UNSPECIFIED, RID_NULLABLE,
RID_NONNULL, RID_NULL_RESETTABLE: New.
(OBJC_IS_PATTR_KEYWORD): Include nullability keywords in the
ranges accepted for property attributes.
* c-attribs.c (handle_objc_nullability_attribute): New.
* c-objc.h (enum objc_property_attribute_group): Add
OBJC_PROPATTR_GROUP_NULLABLE.
(enum objc_property_attribute_kind):Add
OBJC_PROPERTY_ATTR_NULL_UNSPECIFIED, OBJC_PROPERTY_ATTR_NULLABLE,
OBJC_PROPERTY_ATTR_NONNULL, OBJC_PROPERTY_ATTR_NULL_RESETTABLE.
gcc/objc/ChangeLog:
PR objc/90707
* objc-act.c (objc_prop_attr_kind_for_rid): Handle nullability.
(objc_add_property_declaration): Handle nullability attributes.
Check that these are applicable to the property type.
* objc-act.h (enum objc_property_nullability): New.
gcc/testsuite/ChangeLog:
PR objc/90707
* obj-c++.dg/property/at-property-4.mm: Add basic nullability
tests.
* objc.dg/property/at-property-4.m: Likewise.
* obj-c++.dg/attributes/nullability-00.mm: New test.
* obj-c++.dg/property/nullability-00.mm: New test.
* objc.dg/attributes/nullability-00.m: New test.
* objc.dg/property/nullability-00.m: New test.
gcc/ChangeLog:
PR objc/90707
* doc/extend.texi: Document the objc_nullability attribute.
This warning catches the case that the user has left the
superclass specification from a class interface. Root
classes are, of course, permitted and an attribute is added
to mark these so that the diagnostic is suppressed.
The warning and attribute spellings have been kept in sync
with the language reference implementation (clang).
The diagnostic location information present in the objective-c
interface and class definitions is relatively poor. This patch
adds a location for the class name to the interface and makes use
of it in existing warnings.
Part 1 is the changes to code and added tests.
Many entries in the testsuite make use of root classes so
there are a large number of mechanical changes there adding
"-Wno-objc-root-class" to the options.
The test changes are parts 2 (objective-c) and 3 (objective-c++)
in the patch series.
gcc/c-family/ChangeLog:
PR objc/77404
* c-attribs.c (handle_objc_root_class_attribute): New
* c-objc.h (objc_start_class_interface): Add a location
value for the position of the class name.
* c.opt: Add Wobjc-root-class.
* stub-objc.c (objc_start_class_interface): Add a location
value for the position of the class name.
gcc/c/ChangeLog:
PR objc/77404
* c-parser.c (c_parser_objc_class_definition): Pass the
location of the class name to the interface declaration.
gcc/cp/ChangeLog:
PR objc/77404
* parser.c (cp_parser_objc_class_interface): Pass the
location of the class name to the interface declaration.
gcc/objc/ChangeLog:
PR objc/77404
* objc-act.c (objc_start_class_interface): Accept the location
of the class name, use it in existing diagnostic.
(start_class): Accept obj_root_class type attributes. Warn when
the interface for an implementation does not contain a super
class (unless the diagnostic is suppressed by the the command
line flag or the objc_root_class type attribute).
gcc/testsuite/ChangeLog:
PR objc/77404
* objc.dg/attributes/root-class-01.m: New test.
* objc.dg/root-class-00.m: New test.
* obj-c++.dg/attributes/root-class-01.mm: New test.
* obj-c++.dg/root-class-00.mm: New test.
gcc/ChangeLog:
PR objc/77404
* doc/extend.texi: Document the objc_root_class attribute.
* doc/invoke.texi: Document -Wobjc-root-class.
This replaces the old-school gimple_expr_code with more selective
functions throughout the compiler, in all cases making the code
shorter or more clear.
2020-11-13 Richard Biener <rguenther@suse.de>
* cfgexpand.c (gimple_assign_rhs_to_tree): Use
gimple_assign_rhs_class.
(expand_gimple_stmt_1): Likewise.
* gimplify-me.c (gimple_regimplify_operands): Use
gimple_assign_single_p.
* ipa-icf-gimple.c (func_checker::compare_gimple_assign):
Remove redundant compare.
(func_checker::compare_gimple_cond): Use gimple_cond_code.
* tree-ssa-tail-merge.c (gimple_equal_p): Likewise.
* predict.c (predict_loops): Use gimple_assign_rhs_code.
vr_values is being shared among the propagator and the folder and
passed around. I've pulled it out from the propagator so it can be
passed around to each, instead of being publicly accessible from the
propagator.
gcc/ChangeLog:
* tree-vrp.c (class vrp_prop): Rename vr_values to m_vr_values.
(vrp_prop::vrp_prop): New.
(vrp_prop::initialize): Rename vr_values to m_vr_values.
(vrp_prop::visit_stmt): Same.
(vrp_prop::visit_phi): Same.
(vrp_prop::finalize): Same.
(execute_vrp): Instantiate vrp_vr_values and pass it to folder
and propagator.
Earlier in this cycle there was some work by Giuliano Belinassi and
myself to refactor tree-vrp.c. A lot of functions and globals were
moved into independent classes, but the haphazard layout remained.
Assertion methods were indispersed with the propagation code, and with
the jump threading code, etc etc.
This series of patches moves things around so that common
functionality is geographically close. There is no change in
behavior.
I know this is all slated to go in the next release, but finding
things in the current code base, even if just to compare with the
ranger, is difficult.
Since I keep getting bit by aarch64 regressions, I've tested the whole
set of patches on aarch64, as well as individually on x86-64 Linux.
gcc/ChangeLog:
* tree-vrp.c (struct assert_locus): Move.
(class vrp_insert): Rename to vrp_asserts.
(vrp_insert::build_assert_expr_for): Move to vrp_asserts.
(fp_predicate): Same.
(vrp_insert::dump): Same.
(vrp_insert::register_new_assert_for): Same.
(extract_code_and_val_from_cond_with_ops): Move.
(vrp_insert::finish_register_edge_assert_for): Move to vrp_asserts.
(maybe_set_nonzero_bits): Move.
(vrp_insert::find_conditional_asserts): Move to vrp_asserts.
(stmt_interesting_for_vrp): Move.
(struct case_info): Move.
(compare_case_labels): Move.
(lhs_of_dominating_assert): Move.
(find_case_label_index): Move.
(find_case_label_range): Move.
(class vrp_asserts): New.
(vrp_asserts::build_assert_expr_for): Rename from vrp_insert.
(vrp_asserts::dump): Same.
(vrp_asserts::register_new_assert_for): Same.
(vrp_asserts::finish_register_edge_assert_for): Same.
(vrp_asserts::find_conditional_asserts): Same.
(vrp_asserts::compare_case_labels): Same.
(vrp_asserts::find_switch_asserts): Same.
(vrp_asserts::find_assert_locations_in_bb): Same.
(vrp_asserts::find_assert_locations): Same.
(vrp_asserts::process_assert_insertions_for): Same.
(vrp_asserts::compare_assert_loc): Same.
(vrp_asserts::process_assert_insertions): Same.
(vrp_asserts::insert_range_assertions): Same.
(vrp_asserts::all_imm_uses_in_stmt_or_feed_cond): Same.
(vrp_asserts::remove_range_assertions): Same.
(class vrp_prop): Move.
(all_imm_uses_in_stmt_or_feed_cond): Move.
(vrp_prop::vrp_initialize): Move.
(class vrp_folder): Move.
(vrp_folder::fold_predicate_in): Move.
(vrp_folder::fold_stmt): Move.
(vrp_prop::initialize): Move.
(vrp_prop::visit_stmt): Move.
(enum ssa_prop_result): Move.
(vrp_prop::visit_phi): Move.
(vrp_prop::finalize): Move.
(class vrp_dom_walker): Rename to...
(class vrp_jump_threader): ...this.
(vrp_jump_threader::before_dom_children): Rename from
vrp_dom_walker.
(simplify_stmt_for_jump_threading): Rename to...
(vrp_jump_threader::simplify_stmt): ...here.
(vrp_jump_threader::after_dom_children): Same.
(identify_jump_threads): Move.
(vrp_prop::vrp_finalize): Move array bounds setup code to...
(execute_vrp): ...here.
Adjust the range_handler to not use gimple_expr_code/type.
* gimple-range.h (gimple_range_handler): Use gimple_assign and
gimple_cond routines to get type and code.
* range-op.cc (range_op_handler): Check for integral types.
The atom_cache in normalize_atom relies on the assumption that two
equivalent (templated) trees (in the sense of cp_tree_equal) must use
the same template parameters (according to find_template_parameters).
This assumption unfortunately doesn't always hold for TARGET_EXPRs,
because cp_tree_equal ignores an artificial target of a TARGET_EXPR, but
find_template_parameters walks this target (and its DECL_CONTEXT).
Hence two TARGET_EXPRs built by force_target_expr with the same
initializer and under different settings of current_function_decl will
compare equal according to cp_tree_equal, but find_template_parameters
may return a different set of template parameters for them. This breaks
the below testcase because during normalization we build two such
TARGET_EXPRs (one under current_function_decl=f and another under =g),
and then share the same ATOMIC_CONSTR for the two corresponding atoms,
leading to a crash during satisfaction of g's associated constraints.
This patch works around this issue by removing the source of these
templated TARGET_EXPRs. The relevant call to get_target_expr_sfinae was
added in r9-6043, and it seems it's no longer necessary (according to
https://gcc.gnu.org/pipermail/gcc-patches/2019-February/517323.html, the
call was added in order to avoid regressing on initlist109.C at the time).
gcc/cp/ChangeLog:
* semantics.c (finish_compound_literal): Don't wrap the original
compound literal in a TARGET_EXPR when inside a template.
gcc/testsuite/ChangeLog:
* g++.dg/cpp2a/concepts-decltype3.C: New test.
To poll a std::future to see if it's ready you have to call one of the
timed waiting functions. The most obvious way is wait_for(0s) but this
was previously very inefficient because it would turn the relative
timeout to an absolute one by calling system_clock::now(). When the
relative timeout is zero (or less) we're obviously going to get a time
that has already passed, but the overhead of obtaining the current time
can be dozens of microseconds. The alternative is to call wait_until
with an absolute timeout that is in the past. If you know the clock's
epoch is in the past you can use a default constructed time_point.
Alternatively, using some_clock::time_point::min() gives the earliest
time point supported by the clock, which should be safe to assume is in
the past. However, using a futex wait with an absolute timeout before
the UNIX epoch fails and sets errno=EINVAL. The new code using futex
waits with absolute timeouts was not checking for this case, which could
result in hangs (or killing the process if the libray is built with
assertions enabled).
This patch checks for times before the epoch before attempting to wait
on a futex with an absolute timeout, which fixes the hangs or crashes.
It also makes it very fast to poll using an absolute timeout before the
epoch (because we skip the futex syscall).
It also makes future::wait_for avoid waiting at all when the relative
timeout is zero or less, to avoid the unnecessary overhead of getting
the current time. This makes polling with wait_for(0s) take only a few
cycles instead of dozens of milliseconds.
libstdc++-v3/ChangeLog:
* include/std/future (future::wait_for): Do not wait for
durations less than or equal to zero.
* src/c++11/futex.cc (_M_futex_wait_until)
(_M_futex_wait_until_steady): Do not wait for timeouts before
the epoch.
* testsuite/30_threads/future/members/poll.cc: New test.
This patch fixes a bug in recording::string::make_debug_string in which
'\t' and '\n' were "escaped" by simply prepending a '\', thus emitting
'\' then '\n', rather than '\' then 'n'. It also removes a hack that
determined if a string is to be escaped by checking for a leading '"',
by instead adding a flag.
gcc/jit/ChangeLog:
* jit-recording.c (recording::context::new_string): Add "escaped"
param and use it when creating the new recording::string instance.
(recording::string::string): Add "escaped" param and use it to
initialize m_escaped.
(recording::string::make_debug_string): Replace check that first
char is double-quote with use of m_escaped. Fix escaping of
'\t' and '\n'. Set "escaped" on the result.
* jit-recording.h (recording::context::new_string): Add "escaped"
param.
(recording::string::string): Add "escaped" param.
(recording::string::m_escaped): New field.
gcc/testsuite/ChangeLog:
* jit.dg/test-debug-strings.c (create_code): Add tests of
string literal escaping.
gcc/
* configure: Regenerated.
* configure.ac: If ifunc was supported in the binutils for
linux toolchain, then set enable_gnu_indirect_function to yes.
C2x adds the __has_c_attribute preprocessor operator, similar to C++
__has_cpp_attribute.
GCC implements __has_cpp_attribute as exactly equivalent to
__has_attribute. (The documentation says they differ regarding the
values returned for standard attributes, but that's actually only a
matter of the particular nonzero value returned not being specified in
the documentation for __has_attribute; the implementation makes no
distinction between the two.)
I don't think having them exactly equivalent is actually correct,
either for __has_cpp_attribute or for __has_c_attribute.
Specifically, I think it is only correct for __has_cpp_attribute or
__has_c_attribute to return nonzero if the given attribute is
supported, with the particular pp-tokens passed to __has_cpp_attribute
or __has_c_attribute, with [[]] syntax, not if it's only accepted in
__attribute__ or with gnu:: added in [[]]. For example, they should
return nonzero for gnu::packed, but zero for plain packed, because
[[gnu::packed]] is accepted but [[packed]] is ignored as not a
standard attribute.
This patch implements that for __has_c_attribute, leaving any changes
to __has_cpp_attribute for the C++ maintainers. A new
BT_HAS_STD_ATTRIBUTE is added for __has_c_attribute (which I think,
based on the above, would actually be correct to use for
__has_cpp_attribute as well). The code in c_common_has_attribute that
deals with scopes has its C++ conditional removed; instead, whether
the language is C or C++ is used only to determine the numeric values
returned for standard attributes (and which standard attributes are
handled there at all). A new argument is passed to
c_common_has_attribute to distinguish BT_HAS_STD_ATTRIBUTE from
BT_HAS_ATTRIBUTE, and that argument is used to stop attributes with no
scope specified from being accepted with __has_c_attribute unless they
are one of the known standard attributes and so handled specially.
Although the standard specify constants ending with 'L' as the values
for the standard attributes, there is no correctness issue with the
lack of code in GCC to add that 'L' to the expansion:
__has_c_attribute and __has_cpp_attribute are expanded in #if after
other macro expansion has occurred, with no semantics being specified
if they occur outside #if, so there is no way for a conforming program
to inspect the exact text of the expansion of those macros, only to
use the resulting pp-number in a #if expression, where long and int
have the same set of values.
Bootstrapped with no regressions for x86_64-pc-linux-gnu.
gcc/
2020-11-12 Joseph Myers <joseph@codesourcery.com>
* doc/cpp.texi (__has_attribute): Document when scopes are allowed
for C.
(__has_c_attribute): New.
gcc/c-family/
2020-11-12 Joseph Myers <joseph@codesourcery.com>
* c-lex.c (c_common_has_attribute): Take argument std_syntax.
Allow scope for C. Handle standard attributes for C. Do not
accept unscoped attributes if std_syntax and not handled as
standard attributes.
* c-common.h (c_common_has_attribute): Update prototype.
gcc/testsuite/
2020-11-12 Joseph Myers <joseph@codesourcery.com>
* gcc.dg/c2x-has-c-attribute-1.c, gcc.dg/c2x-has-c-attribute-2.c,
gcc.dg/c2x-has-c-attribute-3.c, gcc.dg/c2x-has-c-attribute-4.c:
New tests.
libcpp/
2020-11-12 Joseph Myers <joseph@codesourcery.com>
* include/cpplib.h (struct cpp_callbacks): Add bool argument to
has_attribute.
(enum cpp_builtin_type): Add BT_HAS_STD_ATTRIBUTE.
* init.c (builtin_array): Add __has_c_attribute.
(cpp_init_special_builtins): Handle BT_HAS_STD_ATTRIBUTE.
* macro.c (_cpp_builtin_macro_text): Handle BT_HAS_STD_ATTRIBUTE.
Update call to has_attribute for BT_HAS_ATTRIBUTE.
* traditional.c (fun_like_macro): Handle BT_HAS_STD_ATTRIBUTE.
For now, task/taskloop constructs aren't handled and C/C++ array reductions
and reductions with task or inscan modifiers need further work.
Instead of calling omp_alloc/omp_free (where the former doesn't have
alignment argument and omp_aligned_alloc is 5.1 only feature), this calls
GOMP_alloc/GOMP_free, so that the library can fail if it would fall back
into NULL (exception is zero length allocations).
2020-11-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* builtin-types.def (BT_FN_PTR_SIZE_SIZE_PTRMODE): New function type.
* omp-builtins.def (BUILT_IN_GOACC_DECLARE): Move earlier.
(BUILT_IN_GOMP_ALLOC, BUILT_IN_GOMP_FREE): New builtins.
* gimplify.c (gimplify_scan_omp_clauses): Force allocator into a
decl if it is not NULL, INTEGER_CST or decl.
(gimplify_adjust_omp_clauses): Clear GOVD_EXPLICIT on explicit clauses
which are being removed. Remove allocate clauses for variables not seen
if they are private, firstprivate or linear too. Call
omp_notice_variable on the allocator otherwise.
(gimplify_omp_for): Handle iterator vars mentioned in allocate clauses
similarly to non-is_gimple_reg iterators.
* omp-low.c (struct omp_context): Add allocate_map field.
(delete_omp_context): Delete it.
(scan_sharing_clauses): Fill it from allocate clauses. Remove it
if mentioned also in shared clause.
(lower_private_allocate): New function.
(lower_rec_input_clauses): Handle allocate clause for privatized
variables, except for task/taskloop, C/C++ array reductions for now
and task/inscan variables.
(lower_send_shared_vars): Don't consider variables in allocate_map
as shared.
* omp-expand.c (expand_omp_for_generic, expand_omp_for_static_nochunk,
expand_omp_for_static_chunk): Use expand_omp_build_assign instead of
gimple_build_assign + gsi_insert_after.
* builtins.c (builtin_fnspec): Handle BUILTIN_GOMP_ALLOC and
BUILTIN_GOMP_FREE.
* tree-ssa-ccp.c (evaluate_stmt): Handle BUILTIN_GOMP_ALLOC.
* tree-ssa-dce.c (mark_stmt_if_obviously_necessary): Handle
BUILTIN_GOMP_ALLOC.
(mark_all_reaching_defs_necessary_1): Handle BUILTIN_GOMP_ALLOC
and BUILTIN_GOMP_FREE.
(propagate_necessity): Likewise.
gcc/fortran/
* f95-lang.c (ATTR_ALLOC_WARN_UNUSED_RESULT_SIZE_2_NOTHROW_LIST):
Define.
(gfc_init_builtin_functions): Add alloc_size and warn_unused_result
attributes to __builtin_GOMP_alloc.
* types.def (BT_PTRMODE): New primitive type.
(BT_FN_VOID_PTR_PTRMODE, BT_FN_PTR_SIZE_SIZE_PTRMODE): New function
types.
libgomp/
* libgomp.map (GOMP_alloc, GOMP_free): Export at GOMP_5.0.1.
* omp.h.in (omp_alloc): Add malloc and alloc_size attributes.
* libgomp_g.h (GOMP_alloc, GOMP_free): Declare.
* allocator.c (omp_aligned_alloc): New for now static function,
add alignment argument and handle it.
(omp_alloc): Reimplement using omp_aligned_alloc.
(GOMP_alloc, GOMP_free): New functions.
(omp_free): Add ialias.
* testsuite/libgomp.c-c++-common/allocate-1.c: New test.
* testsuite/libgomp.c++/allocate-1.C: New test.
cgraph_node::materialize_clone segfaulted when I tried compiling
Tramp3D with -fdump-ipa-all because there was no clone_info - IPA-CP
created a clone only for an aggregate constant, adding a note to its
transformation summary but not creating any tree_map nor
param_adjustements.
Fixed with the following obvious extra checks which has passed
bootstrap and testing on x86_64-linux.
gcc/ChangeLog:
2020-11-12 Martin Jambor <mjambor@suse.cz>
* cgraphclones.c (cgraph_node::materialize_clone): Check that clone
info is not NULL before attempting to dump it.
This patch converts the variables that hold time benefits and
frequencies in IPA-CP from plain integers to sreals, avoiding the need
to cap them to avoid overflows and also fixing a potential underflows.
Size costs corresponding to individual constants are left as ints so
that they do not take up too much space. Care must be taken that
adding it up does not overflow, especially in the case of
prop_size_cost, because in cases of extremely long chains of lattice
dependencies it can overflow (e.g. in testsuite/gcc.dg/ipa/pr50744.c).
The overall size is already tracked in long ints.
gcc/ChangeLog:
2020-11-11 Martin Jambor <mjambor@suse.cz>
* ipa-cp.c (class ipcp_value_base): Change the type of
local_time_benefit and prop_time_benefit to sreal. Adjust the
constructor initializer.
(ipcp_lattice::print): Dump sreals.
(struct caller_statistics): Change the type of freq_sum to sreal.
(gather_caller_stats): Work with sreal freq_sum.
(incorporate_penalties): Work with sreal evaluation.
(good_cloning_opportunity_p): Adjusted for sreal sreal time_benefit
and freq_sum. Bail out if size_cost is INT_MAX.
(perform_estimation_of_a_value): Work with sreal time_benefit. Avoid
unnecessary capping.
(estimate_local_effects): Pass sreal time benefit to
good_cloning_opportunity_p without capping it. Adjust dumping.
(safe_add): If there can be overflow, return INT_MAX.
(propagate_effects): Work with sreal times.
(get_info_about_necessary_edges): Work with sreal frequencies.
(decide_about_value): Likewise and with sreal time benefits.
I'd like to have the option of marking functions with
__attribute__ ((__warn_unused_result__)), so this patch adds a macro.
And use it for maybe_wrap_with_location, it's always a bug if the
return value is not used, which happened to me and got me confused.
gcc/ChangeLog:
* system.h (WARN_UNUSED_RESULT): Define for GCC >= 3.4.
* tree.h (maybe_wrap_with_location): Add WARN_UNUSED_RESULT.
* fold-const.c (operand_compare::operand_equal_p): Compare field
offsets in operand_equal_p and OEP_ADDRESS_OF.
(operand_compare::hash_operand): Update.
This changes the __numeric_traits primary template to assume its
argument is an integer type. For the three floating point types that are
supported by __numeric_traits_floating an explicit specialization of
__numeric_traits chooses the right base class.
This improves the failure mode for using __numeric_traits with an
unsupported type. Previously it would use __numeric_traits_floating as
the base class, and give somewhat obscure errors for trying to access
the static data members. Now it will use __numeric_traits_integer which
has a static_assert to check for supported types.
As a side effect of this change there is no need to instantiate
__conditional_type to decide which base class to use.
libstdc++-v3/ChangeLog:
* include/ext/numeric_traits.h (__numeric_traits): Change
primary template to always derive from __numeric_traits_integer.
(__numeric_traits<float>, __numeric_traits<double>)
(__numeric_traits<long double>): Add explicit specializations.
This fixes a bug in bitmap_list_view which could end up with
a NULL head->current which makes followup searches fail. Oops.
It also further optimizes the PRE DFS walk by removing useless
stuff and special-casing bitmaps with just one element for
EXECUTE_IF_AND_IN_BITMAP which makes a quite big difference.
2020-11-12 Richard Biener <rguenther@suse.de>
* bitmap.c (bitmap_list_view): Restore head->current.
* tree-ssa-pre.c (pre_expr_DFS): Elide expr_visited bitmap.
Special-case value expression bitmaps with one element.
(bitmap_find_leader): Likewise.
(sorted_array_from_bitmap_set): Elide expr_visited bitmap.
gcc/c-family
PR pch/86674
* c-pch.c (c_common_valid_pch): Use cpp_warning with CPP_W_INVALID_PCH
reason to fix -Werror=invalid-pch and -Wno-error=invalid-pch switches.
libcpp
PR pch/86674
* files.c (_cpp_find_file): Use CPP_DL_NOTE not CPP_DL_ERROR in call to
cpp_error.
The expression used to calculate the maximum value for an integer type
assumes that the number of bits in the value representation is always
sizeof(T) * CHAR_BIT. This is not true for the __int20 type on msp430,
which has only 20 bits in the value representation but 32 bits in the
object representation. This causes an integer overflow in a constant
expression, which is ill-formed.
This problem was already solved by DJ for std::numeric_limits<__int20>
by generalizing the helper macros to use a specified number of bits
instead of assuming sizeof(T) * CHAR_BIT. Then the INT_N_n types can
specify the number of bits using the __GLIBCXX_BITSIZE_INT_N_n macros
that the compiler defines.
I'm using a slightly different approach here. I've replaced the helper
macros entirely, and just expanded the calculations in the initializers
for the static data members. By reordering the data members we can reuse
__is_signed and __digits in the other initializers. This removes the
repetition of expanding __glibcxx_signed(T) and __glibcxx_digits(T)
multiple times in each initializer.
The __is_integer_nonstrict trait now defines a new constant, __width,
which is sizeof(T) * CHAR_BIT by default (defined as an enumerator so
that no storage is needed for a static data member). By specializing
__is_integer_nonstrict for the INT_N types that have padding bits, we
can provide the correct width via the __GLIBCXX_BITSIZE_INT_N_n macros.
libstdc++-v3/ChangeLog:
PR libstdc++/97798
* include/ext/numeric_traits.h (__glibcxx_signed)
(__glibcxx_digits, __glibcxx_min, __glibcxx_max): Remove
macros.
(__is_integer_nonstrict::__width): Define new constant.
(__numeric_traits_integer): Define constants in terms of each
other and __is_integer_nonstrict::__width, rather than the
removed macros.
(_GLIBCXX_INT_N_TRAITS): Macro to define explicit
specializations for non-standard integer types.
The following make sure to only iterate PRE insertion when
necessary - which is when AVAIL_OUT of a predecessor of a
block we already visited changed (that's backedge destinations).
To not regress this also makes sure to locally iterate insertion
since even topological sort of expressions isn't enough to
guarantee we get all opportunities of a block in one iteration.
This avoids costly re-compute of the topologically sorted expression
array (more micro-optimization is possible here).
2020-11-12 Richard Biener <rguenther@suse.de>
* tree-ssa-pre.c (bitmap_value_replace_in_set): Return
whether we have changed anything.
(do_pre_regular_insertion): Get topologically sorted array
of expressions from caller.
(do_pre_partial_partial_insertion): Likewise.
(insert): Compute topologically sorted arrays of expressions
here and locally iterate actual insertion. Iterate only
when AVAIL_OUT of an already visited block source changed.
This patch adds a missing not to the SVE2 BCAX (Bitwise clear and
exclusive or) pattern, fixing the PR. Since SVE doesn't have an
unpredicated not instruction, we need to use a (vacuously) predicated
not here.
To ensure that the predicate is instantiated correctly (to all 1s) for
the intrinsics, we pull out a separate expander from the define_insn.
From the ISA reference [1]:
> Bitwise AND elements of the second source vector with the
> corresponding inverted elements of the third source vector, then
> exclusive OR the results with corresponding elements of the first
> source vector.
[1] : https://developer.arm.com/docs/ddi0602/g/a64-sve-instructions-alphabetic-order/bcax-bitwise-clear-and-exclusive-or
gcc/ChangeLog:
PR target/97730
* config/aarch64/aarch64-sve2.md (@aarch64_sve2_bcax<mode>):
Change to define_expand, add missing (trivially-predicated) not
rtx to fix wrong code bug.
(*aarch64_sve2_bcax<mode>): New.
gcc/testsuite/ChangeLog:
PR target/97730
* gcc.target/aarch64/sve2/bcax_1.c (OP): Add missing bitwise not
to match correct bcax semantics.
* gcc.dg/vect/pr97730.c: New test.
This fixes the postorder compute for the case of multiple
expression leaders for a value.
2020-11-12 Richard Biener <rguenther@suse.de>
PR tree-optimization/97806
* tree-ssa-pre.c (pre_expr_DFS): New overload for visiting
values, visiting all leaders for a value. Use a bitmap
for visited values.
(sorted_array_from_bitmap_set): Walk over values and adjust.
* gcc.dg/pr97806.c: New testcase.
As the testcase shows, CLEANUP_POINT_EXPR (and I think TRY_FINALLY_EXPR too)
suffer from the same problem that I was trying to fix in
r10-3597-g1006c9d4395a939820df76f37c7b085a4a1a003f
for CLEANUP_STMT, namely that if in the middle of the body expression of
those stmts is e.g. return stmt, goto, break or continue (something that
changes *jump_target and makes it start skipping stmts), we then skip the
cleanups too, which is not appropriate - the cleanups were either queued up
during the non-skipping execution of the body (for CLEANUP_POINT_EXPR), or
for TRY_FINALLY_EXPR are relevant already after entering the body block.
> Would it make sense to always use a NULL jump_target when evaluating
> cleanups?
I was afraid of that, especially for TRY_FINALLY_EXPR, but it seems that
during constexpr evaluation the cleanups will most often be just very simple
destructor calls (or calls to cleanup attribute functions).
Furthermore, for neither of these 3 tree codes we'll reach that code if
jump_target && *jump_target initially (there is a return NULL_TREE much
earlier for those except for trees that could embed labels etc. in it and
clearly these 3 don't count in that).
2020-11-12 Jakub Jelinek <jakub@redhat.com>
PR c++/97790
* constexpr.c (cxx_eval_constant_expression) <case CLEANUP_POINT_EXPR,
case TRY_FINALLY_EXPR, case CLEANUP_STMT>: Don't pass jump_target to
cxx_eval_constant_expression when evaluating the cleanups.
* g++.dg/cpp2a/constexpr-dtor9.C: New test.