platform/upstream/gcc.git
4 years agoanalyzer: fix ICE on unreachable calls [PR 93947]
David Malcolm [Wed, 26 Feb 2020 14:43:57 +0000 (09:43 -0500)]
analyzer: fix ICE on unreachable calls [PR 93947]

PR analyzer/93947 reports an ICE at -O1 when attempting to analyze a
call that has been optimized away as unreachable.

The root cause is a NULL dereference due to the fndecl having a NULL
cgraph_node: the cgraph_node was created by
pass_build_cgraph_edges::execute, but was later removed by
symbol_table::remove_unreachable_nodes before the analyzer pass.

This patch fixes it by checking for NULL before handling the
cgraph_node.

The reproducer demonstrates a weakness in the analyzer's constraint
handling, where region_model::apply_constraints_for_gswitch fails
to spot when the cases fully cover the data type, and thus make the
default impossible.  For now this is xfail-ed in the testcase.

gcc/analyzer/ChangeLog:
PR analyzer/93947
* region-model.cc (region_model::get_fndecl_for_call): Gracefully
fail for fn_decls that don't have a cgraph_node.

gcc/testsuite/ChangeLog:
PR analyzer/93947
* gcc.dg/analyzer/torture/pr93947.c: New test.

4 years agoDaily bump.
GCC Administrator [Thu, 27 Feb 2020 00:16:39 +0000 (00:16 +0000)]
Daily bump.

4 years agoPPC64, fix documentation for __builtin_crypto_vpmsum* builtin functions.
Carl Love [Wed, 26 Feb 2020 23:21:20 +0000 (17:21 -0600)]
PPC64, fix documentation for __builtin_crypto_vpmsum* builtin functions.

PR target/91276 - Doc typos in __builtin_crypto_vpmsum*

gcc/ChangeLog:

2020-02-26  Carl Love  <cel@us.ibm.com>

PR target/91276
* doc/extend.texi (PowerPC AltiVec Built-in Functions available on
ISA 3.0): The builtin-function name __builtin_crypto_vpmsumb is only
for the vector unsigned short arguments.  It is also listed as the
name of the built-in for arguments vector unsigned short,
vector unsigned int and vector unsigned long long built-ins.  The
name of the builtins for these arguments should be:
__builtin_crypto_vpmsumh, __builtin_crypto_vpmsumw and
__builtin_crypto_vpmsumd respectively.

Signed-off-by: Carl Love <carll@us.ibm.com>
4 years agocoroutines: Amend parameter handling to match n4849.
Iain Sandoe [Wed, 26 Feb 2020 13:28:07 +0000 (13:28 +0000)]
coroutines: Amend parameter handling to match n4849.

In n4849 and preceding versions, [class.copy.elision] (1.3)
appears to confer additional permissions on coroutines to elide
parameter copies.

After considerable discussion on this topic by email and during
the February 2020 WG21 meeting, it has been determined that there
are no additional permissions applicable to coroutine parameter
copy elision.

The content of that clause in the standard is expected to be amended
eventually to clarify this.  Other than this, the handling of
parameter lifetimes is expected to be as per n4849:

 * A copy is made before the promise is constructed
 * If the promise CTOR uses the parms, then it should use the copy
   where appropriate.
 * The param copy lifetimes end after the promise is destroyed
   (during the coroutine frame destruction).
 * Otherwise, C++20 copy elision rules apply.

(as an aside) In practice, we expect that copy elision can only occur
when the coroutine body is fully inlined, possibly in conjunction with
heap allocation elision.

The patch:
 * Reorders the copying process to precede the promise CTOR and
    ensures the correct use.
 * Copies all params into the frame regardless of whether the coro
   body uses them (this is a bit unfortunate, and we should figure
   out an amendment for C++23).

gcc/cp/ChangeLog:

2020-02-26  Iain Sandoe  <iain@sandoe.co.uk>

* class.c (classtype_has_non_deleted_copy_ctor): New.
* coroutines.cc (struct param_info): Keep track of params
that are references, and cache the original type and whether
the DTOR is trivial.
(build_actor_fn): Handle param copies always, and adjust the
handling for references.
(register_param_uses): Only handle uses here.
(classtype_has_non_deleted_copy_ctor): New.
(morph_fn_to_coro): Adjust param copy handling to match n4849
by reordering ahead of the promise CTOR and always making a
frame copy, even if the param is unused in the coroutine body.
* cp-tree.h (classtype_has_non_deleted_copy_ctor): New.

gcc/testsuite/ChangeLog:

2020-02-26  Iain Sandoe  <iain@sandoe.co.uk>

* g++.dg/coroutines/coro1-refs-and-ctors.h: New.
* g++.dg/coroutines/torture/func-params-07.C: New test.
* g++.dg/coroutines/torture/func-params-08.C: New test.

4 years agolibgo: update to final Go1.14 release
Ian Lance Taylor [Wed, 26 Feb 2020 19:15:50 +0000 (11:15 -0800)]
libgo: update to final Go1.14 release

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/221158

4 years agors6000: Fix more testsuite fallout from rs6000_legitimate_address_p() fix. [PR93913]
Peter Bergner [Wed, 26 Feb 2020 17:58:08 +0000 (11:58 -0600)]
rs6000: Fix more testsuite fallout from rs6000_legitimate_address_p() fix. [PR93913]

PR target/93913
* gcc.target/powerpc/fold-vec-st-char.c (scan-assembler-times): Allow
stxv and stxvx instructions as well.
* gcc.target/powerpc/fold-vec-st-float.c: Likewise.
* gcc.target/powerpc/fold-vec-st-int.c: Likewise.
* gcc.target/powerpc/fold-vec-st-short.c: Likewise.

4 years agoc++: Some improvements to concept diagnostics
Patrick Palka [Thu, 20 Feb 2020 20:56:20 +0000 (15:56 -0500)]
c++: Some improvements to concept diagnostics

This patch improves our concept diagnostics in two ways.  First, it sets a more
precise location for the constraint expressions built in
finish_constraint_binary_op.  As a result, when a disjunction is unsatisfied we
now print e.g.

.../include/bits/range_access.h:467:2: note: neither operand of the disjunction is satisfied
  466 |  requires is_bounded_array_v<remove_reference_t<_Tp>> || __member_end<_Tp>
      |           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  467 |  || __adl_end<_Tp>
      |  ^~~~~~~~~~~~~~~~~

instead of

.../include/bits/range_access.h:467:2: note: neither operand of the disjunction is satisfied
  467 |  || __adl_end<_Tp>
      |  ^~

Second, this patch changes diagnose_atomic_constraint to print unsatisfied
atomic constraint expressions with their template arguments.  So e.g. we now
print

cpp2a/concepts-pr67719.C:9:8: note: the expression ‘(... &&(C<Tx>)()) [with Tx = {int, long int, void}]’ evaluated to ‘false’

instead of

cpp2a/concepts-pr67719.C:9:8: note: the expression ‘(... &&(C<Tx>)())’ evaluated to ‘false’

Tested on x86_64-pc-linux-gnu, and verified that all the diagnostics emitted in
our concept tests are no worse with this patch.

gcc/cp/ChangeLog:

* constraint.cc (finish_constraint_binary_op): Set expr's location range
to the range of its operands.
(satisfy_atom): Pass MAP instead of ARGS to diagnose_atomic_constraint.
(diagnose_trait_expr): Take the instantiated parameter mapping MAP
instead of the corresponding template arguments ARGS and adjust body
accordingly.
(diagnose_requires_expr): Likewise.
(diagnose_atomic_constraint): Likewise.  When printing an atomic
constraint expression, print the instantiated parameter mapping
alongside it.
* cxx-pretty-print.cc (cxx_pretty_printer::expression)
[NONTYPE_ARGUMENT_PACK]: Print braces around a NONTYPE_ARGUMENT_PACK.
(cxx_pretty_printer::type_id): Handle TYPE_ARGUMENT_PACK.

gcc/testsuite/ChangeLog:

* g++.dg/concepts/diagnostic2.C: New test.
* g++.dg/concepts/diagnostic3.C: New test.

4 years agoc++: Fix value-init crash in template [PR93676]
Marek Polacek [Wed, 19 Feb 2020 20:59:33 +0000 (15:59 -0500)]
c++: Fix value-init crash in template [PR93676]

Since <https://gcc.gnu.org/ml/gcc-patches/2015-02/msg00556.html> we
attempt to value-initialize in build_vec_init even when there's no
initializer but the type has a constexpr default constructor.  But
build_value_init doesn't work in templates, and build_vec_init
creates a lot of garbage that would not be used anyway, so don't
call it in a template.

PR c++/93676 - value-init crash in template.
* init.c (build_new_1): Don't call build_vec_init in a template.

* g++.dg/cpp0x/nsdmi-template19.C: New test.

4 years agolibstdc++: Fix use of inaccessible private member in split_view (PR93936)
Patrick Palka [Wed, 26 Feb 2020 13:38:06 +0000 (08:38 -0500)]
libstdc++: Fix use of inaccessible private member in split_view (PR93936)

We are calling _OuterIter::__current from _InnerIter::operator==, but the former
is private within this non-member friend.  Fix this by calling
_OuterIter::operator== instead, which does the right thing here.

libstdc++-v3/ChangeLog:

PR libstdc++/93936
* include/std/ranges (split_view::_InnerIter::operator==): Compare
the operands' _M_i rather than their _M_i.current().
* testsuite/std/ranges/adaptors/split.cc: Augment test.

4 years agolibstdc++: P1645R1 constexpr for <numeric> algorithms
Patrick Palka [Tue, 25 Feb 2020 19:40:41 +0000 (14:40 -0500)]
libstdc++: P1645R1 constexpr for <numeric> algorithms

This adds constexpr to 11 algorithms defined in <numeric> as per P1645R1.

libstdc++-v3/ChangeLog:

P1645R1 constexpr for <numeric> algorithms
* include/bits/stl_numeric.h (iota, accumulate, inner_product,
partial_sum, adjacent_difference): Make conditionally constexpr for
C++20.
* include/std/numeric (__cpp_lib_constexpr_numeric): Define this feature
test macro.
(reduce, transform_reduce, exclusive_scan, inclusive_scan,
transform_exclusive_scan, transform_inclusive_scan): Make conditionally
constexpr for C++20.
* include/std/version (__cpp_lib_constexpr_numeric): Define.
* testsuite/26_numerics/accumulate/constexpr.cc: New test.
* testsuite/26_numerics/adjacent_difference/constexpr.cc: Likewise.
* testsuite/26_numerics/exclusive_scan/constexpr.cc: Likewise.
* testsuite/26_numerics/inclusive_scan/constexpr.cc: Likewise.
* testsuite/26_numerics/inner_product/constexpr.cc: Likewise.
* testsuite/26_numerics/iota/constexpr.cc: Likewise.
* testsuite/26_numerics/partial_sum/constexpr.cc: Likewise.
* testsuite/26_numerics/reduce/constexpr.cc: Likewise.
* testsuite/26_numerics/transform_exclusive_scan/constexpr.cc: Likewise.
* testsuite/26_numerics/transform_inclusive_scan/constexpr.cc: Likewise.
* testsuite/26_numerics/transform_reduce/constexpr.cc: Likewise.

4 years agolibstdc++ Two simplifications for lexicographical_compare
Jonathan Wakely [Wed, 26 Feb 2020 15:19:44 +0000 (15:19 +0000)]
libstdc++ Two simplifications for lexicographical_compare

* include/bits/ranges_algo.h (__lexicographical_compare_fn): Declare
variables in smaller scope and avoid calling ranges::distance when we
know they are pointers. Remove statically-unreachable use of
__builtin_unreachable().
* include/bits/stl_algobase.h (__lexicographical_compare::__lc):
Define inline.

4 years agolibstdc++: Add __maybe_const_t and __maybe_empty_t aliases
Jonathan Wakely [Wed, 26 Feb 2020 15:19:43 +0000 (15:19 +0000)]
libstdc++: Add __maybe_const_t and __maybe_empty_t aliases

This introduces a couple of convenience alias templates to be used for
some repeated patterns using std::conditional_t.

* include/std/ranges (__detail::__maybe_empty_t): Define new helper
alias.
(__detail::__maybe_const_t): Likewise.
(__adaptor::_RangeAdaptor): Use __maybe_empty_t.
(transform_view, take_view, take_while_view, elements_view): Use
__maybe_const_t.
(join_view, split_view): Use both.

4 years agoc++: Fix ICE with static_cast when converting from int[] [PR93862]
Marek Polacek [Mon, 24 Feb 2020 19:43:48 +0000 (14:43 -0500)]
c++: Fix ICE with static_cast when converting from int[] [PR93862]

This ICEs since my patch for P0388, which allowed conversions to arrays
of unknown bound, but not the reverse, so these two static_casts are
ill-formed.

[expr.static.cast]/3 says that "cv1 T1" and "cv2 T2" have to be
reference-compatible and the comment in build_static_cast_1 says it too
but then we actually use reference_related_p...  Fixed thus.

2020-02-26  Marek Polacek  <polacek@redhat.com>

PR c++/93862 - ICE with static_cast when converting from int[].
* call.c (reference_compatible_p): No longer static.
* cp-tree.h (reference_compatible_p): Declare.
* typeck.c (build_static_cast_1): Use reference_compatible_p instead
of reference_related_p.

* g++.dg/cpp0x/rv-cast7.C: New test.

4 years agoc++: Add test for DR 1423, Convertibility of nullptr to bool.
Marek Polacek [Mon, 24 Feb 2020 17:56:00 +0000 (12:56 -0500)]
c++: Add test for DR 1423, Convertibility of nullptr to bool.

DR 1423, which supersedes DR 654, says that you can't copy-init
a bool from a std::nullptr_t:

  bool b = nullptr;  // error

Conversely, it works with direct-initialization which is more
permissive than copy-initialization.

No code changes necessary since we handle it right.

2020-02-26  Marek Polacek  <polacek@redhat.com>

DR 1423, Convertibility of nullptr to bool.
* g++.dg/DRs/dr1423.C: New test.

4 years agoc++: Fix ICE with constexpr init and [[no_unique_address]] [PR93803]
Marek Polacek [Mon, 24 Feb 2020 14:19:01 +0000 (09:19 -0500)]
c++: Fix ICE with constexpr init and [[no_unique_address]] [PR93803]

Here we crash when constexpr-initializing a class member of empty class
type with [[no_unique_address]].  Without the attribute we would have
a ctor (that initializes bar) of the form

  { .D.2173 = { .x = {} } }

but with the attribute reduced_constant_expression_p gets

  { .x = {} }

That means that "idx != field" is true for the latter and we see that
foo, the base class of bar, is an empty class, so we want to look at
the next initializable field (since empty class fields may not have an
initializer).  But in this case there are no more, therefore accessing
DECL_CHAIN (field) crashes.  Long story short, we need to avoid a crash
on a null field when we're initializing a class that only contains an
empty base class.

While poking into this I discovered c++/93898, but that's a different
problem.

2020-02-26  Marek Polacek  <polacek@redhat.com>

PR c++/93803 - ICE with constexpr init and [[no_unique_address]].
* constexpr.c (reduced_constant_expression_p): Don't crash on a null
field.

* g++.dg/cpp2a/constexpr-init16.C: New test.
* g++.dg/cpp2a/constexpr-init17.C: New test.

4 years agodump load permutations and refcount per SLP node
Richard Biener [Wed, 26 Feb 2020 13:51:26 +0000 (14:51 +0100)]
dump load permutations and refcount per SLP node

This adjusts dumping as proved useful in debugging.

2020-02-26  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (vect_print_slp_tree): Also dump ref count
and load permutation.

4 years agooptabs: Don't use scalar conversions for vectors [PR93843]
Richard Sandiford [Tue, 25 Feb 2020 19:20:58 +0000 (19:20 +0000)]
optabs: Don't use scalar conversions for vectors [PR93843]

In this PR we had a conversion between two integer vectors that
both had scalar integer modes.  We then tried to implement the
conversion using the scalar optab for those modes, instead of
doing the conversion elementwise.

I wondered about letting through scalar modes for single-element
vectors, but I don't have any evidence that that's useful/necessary,
so it seemed better to keep things simple.

2020-02-26  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR middle-end/93843
* optabs-tree.c (supportable_convert_operation): Reject types with
scalar modes.

gcc/testsuite/
PR middle-end/93843
* gcc.dg/vect/pr93843-1.c: New test.
* gcc.dg/vect/pr93843-2.c: Likewise.

4 years agoanalyzer: improvements to logging/dumping
David Malcolm [Fri, 21 Feb 2020 14:28:55 +0000 (09:28 -0500)]
analyzer: improvements to logging/dumping

This patch adds various information to -fdump-analyzer and
-fdump-analyzer-stderr to make it easier to track down
problems with state explosions in the exploded_graph.

It logs the number of unprocessed nodes in the worklist, for
the case where the upper limit on exploded nodes is reached.

It prints:
[a] a bar chart showing the number of exploded nodes by function, and

[b] bar charts for each function showing the number of exploded nodes
    per supernode/BB, and

[c] bar charts for each function showing the number of excess exploded
    nodes per supernode/BB beyond the limit
    (--param=analyzer-max-enodes-per-program-point), where that limit
    was reached

I've found these helpful in finding exactly where we fail to consolidate
state, leading to state explosions and false negatives due to the
thresholds being reached.

The patch also adds a "superedge::dump" member function I found myself
needing.

gcc/ChangeLog:
* Makefile.in (ANALYZER_OBJS): Add analyzer/bar-chart.o.

gcc/analyzer/ChangeLog:
* bar-chart.cc: New file.
* bar-chart.h: New file.
* engine.cc: Include "analyzer/bar-chart.h".
(stats::log): Only log the m_num_nodes kinds that are non-zero.
(stats::dump): Likewise when dumping.
(stats::get_total_enodes): New.
(exploded_graph::get_or_create_node): Increment the per-point-data
m_excess_enodes when hitting the per-program-point limit on
enodes.
(exploded_graph::print_bar_charts): New.
(exploded_graph::log_stats): Log the number of unprocessed enodes
in the worklist.  Call print_bar_charts.
(exploded_graph::dump_stats): Print the number of unprocessed
enodes in the worklist.
* exploded-graph.h (stats::get_total_enodes): New decl.
(struct per_program_point_data): Add field m_excess_enodes.
(exploded_graph::print_bar_charts): New decl.
* supergraph.cc (superedge::dump): New.
(superedge::dump): New.
* supergraph.h (supernode::get_function): New.
(superedge::dump): New decl.
(superedge::dump): New decl.

4 years agotestsuite: Add a -O2 -fgimple testcase next to the -O2 -fno-tree-dse one [PR93820]
Jakub Jelinek [Wed, 26 Feb 2020 09:58:13 +0000 (10:58 +0100)]
testsuite: Add a -O2 -fgimple testcase next to the -O2 -fno-tree-dse one [PR93820]

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/93820
* gcc.dg/pr93820-2.c: New test.

4 years agostore-merging: Fix coalesce_immediate_stores [PR93820]
Jakub Jelinek [Wed, 26 Feb 2020 08:33:48 +0000 (09:33 +0100)]
store-merging: Fix coalesce_immediate_stores [PR93820]

The following testcase is miscompiled in 8+.
The problem is that check_no_overlap has a special case for INTEGER_CST
marked stores (i.e. stores of constants), if both all currenly merged stores
and the one under consideration for merging with them are marked that way,
it anticipates that other INTEGER_CST marked stores that overlap with those
and precede those (have smaller info->order) could be merged with those and
doesn't punt for them.
In PR86844 and PR87859 fixes I've then added quite large code that is
performed after check_no_overlap and tries to find out if we need and can
merge further INTEGER_CST marked stores, or need to punt.
Unfortunately, that code is there only in the overlapping case code and
the testcase below shows that we really need it even in the adjacent store
case.  After sort_by_bitpos we have:
bitpos width order rhs_code
96 32 3 INTEGER_CST
128 32 1 INTEGER_CST
128 128 2 INTEGER_CST
192 32 0 MEM_REF
Because of the missing PR86844/PR87859-ish code in the adjacent store
case, we merge the adjacent (memory wise) stores 96/32/3 and 128/32/1,
and then we consider the 128-bit store which is in program-order in between
them, but in this case we punt, because the merging would extend the
merged store region from bitpos 96 and 64-bits to bitpos 96 and 160-bits
and that has an overlap with an incompatible store (the MEM_REF one).
The problem is that we can't really punt this way, because the 128-bit
store is in between those two we've merged already, so either we manage
to merge even that one together with the others, or would need to avoid
already merging the 96/32/3 and 128/32/1 stores together.
Now, rather than copying around the PR86844/PR87859 code to the other spot,
we can actually just use the overlapping code, merge_overlapping is really
a superset of merge_into, so that is what the patch does.  If doing
adjacent store merge for rhs_code other than INTEGER_CST, I believe the
current code is already fine, check_no_overlap in that case doesn't make
the exception and will punt if there is some earlier (smaller order)
non-mergeable overlapping store.  There is just one case that could be
problematic, if the merged_store has BIT_INSERT_EXPRs in them and the
new store is a constant store (INTEGER_CST rhs_code), then check_no_overlap
would do the exception and still would allow the special case.  But we
really shouldn't have the special case in that case, so this patch also
changes check_no_overlap to just have a bool whether we should have the
special case or not.

Note, as I said in the PR, for GCC11 we could consider performing some kind
of cheap DSE during the store merging (perhaps guarded with flag_tree_dse).
And another thing to consider is only consider as problematic non-mergeable
stores that not only have order smaller than last_order as currently, but
also have order larger than first_order, as in this testcase if we actually
ignored (not merged with anything at all) the 192/32/0 store, because it is
not in between the other stores we'd merge, it would be fine to merge the
other 3 stores, though of course the testcase can be easily adjusted by
putting the 192/32 store after the 128/32 store and then this patch would be
still needed.  Though, I think I'd need more time thinking this over.

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/93820
* gimple-ssa-store-merging.c (check_no_overlap): Change RHS_CODE
argument to ALL_INTEGER_CST_P boolean.
(imm_store_chain_info::try_coalesce_bswap): Adjust caller.
(imm_store_chain_info::coalesce_immediate_stores): Likewise.  Handle
adjacent INTEGER_CST store into merged_store->only_constants like
overlapping one.

* gcc.dg/pr93820.c: New test.

4 years agoc++: Fix rejects-valid bug in cxx_eval_outermost_constant_expr [PR93905]
Jakub Jelinek [Wed, 26 Feb 2020 08:04:44 +0000 (09:04 +0100)]
c++: Fix rejects-valid bug in cxx_eval_outermost_constant_expr [PR93905]

Add testcase for a bug that has been just on the 8 branch.

2020-02-26  Jakub Jelinek  <jakub@redhat.com>

PR c++/93905
* g++.dg/cpp0x/pr93905.C: New test.

4 years agoDaily bump.
GCC Administrator [Wed, 26 Feb 2020 00:16:34 +0000 (00:16 +0000)]
Daily bump.

4 years agotypo fix: Fix probablity, becuse, sucessor and destinarion typos [PR93912]
Jakub Jelinek [Tue, 25 Feb 2020 21:10:48 +0000 (22:10 +0100)]
typo fix: Fix probablity, becuse, sucessor and destinarion typos [PR93912]

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR other/93912
* config/sh/sh.c (expand_cbranchdi4): Fix comment typo, probablity
-> probability.
* cfghooks.c (verify_flow_info): Likewise.
* predict.c (combine_predictions_for_bb): Likewise.
* bb-reorder.c (connect_better_edge_p): Likewise.  Fix comment typo,
sucessor -> successor.
(find_traces_1_round): Fix comment typo, destinarion -> destination.
* omp-expand.c (expand_oacc_for): Fix comment typo, sucessors ->
successors.
* tree-ssa-loop-ch.c (should_duplicate_loop_header_p): Fix dump
message typo, sucessors -> successors.
c/
* gimple-parser.c (c_parser_gimple_parse_bb_spec_edge_probability):
Rename last argument from probablity to probability.

4 years agoCorrect an attribute access example.
Martin Sebor [Tue, 25 Feb 2020 20:14:56 +0000 (13:14 -0700)]
Correct an attribute access example.

gcc/ChangeLog:
* doc/extend.texi (attribute access): Correct an example.

4 years agoaarch64: Add bfloat16 vldn/vstn intrinsics
Mihail Ionescu [Tue, 18 Feb 2020 14:29:47 +0000 (14:29 +0000)]
aarch64: Add bfloat16 vldn/vstn intrinsics

This patch adds the load/store bfloat16 intrinsics to the AArch64 back-end.
ACLE documents are at https://developer.arm.com/docs/101028/latest
ISA documents are at https://developer.arm.com/docs/ddi0596/latest

2020-02-25  Mihail Ionescu  <mihail.ionescu@arm.com>

gcc/
* config/aarch64/aarch64-builtins.c (aarch64_scalar_builtin_types):
Add simd_bf.
(aarch64_init_simd_builtin_scalar_types): Register simd_bf.
(VAR15, VAR16): New.
* config/aarch64/iterators.md (VALLDIF): Enable for V4BF and V8BF.
(VD): Enable for V4BF.
(VDC): Likewise.
(VQ): Enable for V8BF.
(VQ2): Likewise.
(VQ_NO2E): Likewise.
(VDBL, Vdbl): Add V4BF.
(V_INT_EQUIV, v_int_equiv): Add V4BF and V8BF.
* config/aarch64/arm_neon.h (bfloat16x4x2_t): New typedef.
(bfloat16x8x2_t): Likewise.
(bfloat16x4x3_t): Likewise.
(bfloat16x8x3_t): Likewise.
(bfloat16x4x4_t): Likewise.
(bfloat16x8x4_t): Likewise.
(vcombine_bf16): New.
(vld1_bf16, vld1_bf16_x2): New.
(vld1_bf16_x3, vld1_bf16_x4): New.
(vld1q_bf16, vld1q_bf16_x2): New.
(vld1q_bf16_x3, vld1q_bf16_x4): New.
(vld1_lane_bf16): New.
(vld1q_lane_bf16): New.
(vld1_dup_bf16): New.
(vld1q_dup_bf16): New.
(vld2_bf16): New.
(vld2q_bf16): New.
(vld2_dup_bf16): New.
(vld2q_dup_bf16): New.
(vld3_bf16): New.
(vld3q_bf16): New.
(vld3_dup_bf16): New.
(vld3q_dup_bf16): New.
(vld4_bf16): New.
(vld4q_bf16): New.
(vld4_dup_bf16): New.
(vld4q_dup_bf16): New.
(vst1_bf16, vst1_bf16_x2): New.
(vst1_bf16_x3, vst1_bf16_x4): New.
(vst1q_bf16, vst1q_bf16_x2): New.
(vst1q_bf16_x3, vst1q_bf16_x4): New.
(vst1_lane_bf16): New.
(vst1q_lane_bf16): New.
(vst2_bf16): New.
(vst2q_bf16): New.
(vst3_bf16): New.
(vst3q_bf16): New.
(vst4_bf16): New.
(vst4q_bf16): New.

gcc/testsuite/
* gcc.target/aarch64/advsimd-intrinsics/bf16_vstn.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/bf16_vldn.c: New test.

4 years agoaarch64: Add bfloat16 vdup and vreinterpret ACLE intrinsics
Mihail Ionescu [Tue, 18 Feb 2020 14:23:09 +0000 (14:23 +0000)]
aarch64: Add bfloat16 vdup and vreinterpret ACLE intrinsics

This patch adds support for the bf16 duplicate and reinterpret intrinsics.
ACLE documents are at https://developer.arm.com/docs/101028/latest
ISA documents are at https://developer.arm.com/docs/ddi0596/latest

2020-02-25  Mihail Ionescu  <mihail.ionescu@arm.com>

gcc/
* config/aarch64/iterators.md (VDQF_F16) Add V4BF and V8BF.
(VALL_F16): Likewise.
(VALLDI_F16): Likewise.
(Vtype): Likewise.
(Vetype): Likewise.
(vswap_width_name): Likewise.
(VSWAP_WIDTH): Likewise.
(Vel): Likewise.
(VEL): Likewise.
(q): Likewise.
* config/aarch64/arm_neon.h (vset_lane_bf16, vsetq_lane_bf16): New.
(vget_lane_bf16, vgetq_lane_bf16): New.
(vcreate_bf16): New.
(vdup_n_bf16, vdupq_n_bf16): New.
(vdup_lane_bf16, vdup_laneq_bf16): New.
(vdupq_lane_bf16, vdupq_laneq_bf16): New.
(vduph_lane_bf16, vduph_laneq_bf16): New.
(vreinterpret_bf16_u8, vreinterpretq_bf16_u8): New.
(vreinterpret_bf16_u16, vreinterpretq_bf16_u16): New.
(vreinterpret_bf16_u32, vreinterpretq_bf16_u32): New.
(vreinterpret_bf16_u64, vreinterpretq_bf16_u64): New.
(vreinterpret_bf16_s8, vreinterpretq_bf16_s8): New.
(vreinterpret_bf16_s16, vreinterpretq_bf16_s16): New.
(vreinterpret_bf16_s32, vreinterpretq_bf16_s32): New.
(vreinterpret_bf16_s64, vreinterpretq_bf16_s64): New.
(vreinterpret_bf16_p8, vreinterpretq_bf16_p8): New.
(vreinterpret_bf16_p16, vreinterpretq_bf16_p16): New.
(vreinterpret_bf16_p64, vreinterpretq_bf16_p64): New
(vreinterpret_bf16_f16, vreinterpretq_bf16_f16): New
(vreinterpret_bf16_f32, vreinterpretq_bf16_f32): New.
(vreinterpret_bf16_f64, vreinterpretq_bf16_f64): New.
(vreinterpretq_bf16_p128): New.
(vreinterpret_s8_bf16, vreinterpretq_s8_bf16): New.
(vreinterpret_s16_bf16, vreinterpretq_s16_bf16): New.
(vreinterpret_s32_bf16, vreinterpretq_s32_bf16): New.
(vreinterpret_s64_bf16, vreinterpretq_s64_bf16): New.
(vreinterpret_u8_bf16, vreinterpretq_u8_bf16): New.
(vreinterpret_u16_bf16, vreinterpretq_u16_bf16): New.
(vreinterpret_u32_bf16, vreinterpretq_u32_bf16): New.
(vreinterpret_u64_bf16, vreinterpretq_u64_bf16): New.
(vreinterpret_p8_bf16, vreinterpretq_p8_bf16): New.
(vreinterpret_p16_bf16, vreinterpretq_p16_bf16): New.
(vreinterpret_p64_bf16, vreinterpretq_p64_bf16): New.
(vreinterpret_f32_bf16, vreinterpretq_f32_bf16): New.
(vreinterpret_f64_bf16,vreinterpretq_f64_bf16): New.
(vreinterpret_f16_bf16,vreinterpretq_f16_bf16): New.
(vreinterpretq_p128_bf16): New.

gcc/testsuite/
* gcc.target/aarch64/advsimd-intrinsics/bf16_dup.c: New test.
* gcc.target/aarch64/advsimd-intrinsics/bf16_reinterpret.c: New test.

4 years agolibstdc++: LWG 3397 basic_istream_view::iterator should not provide iterator_category
Patrick Palka [Mon, 24 Feb 2020 22:13:40 +0000 (17:13 -0500)]
libstdc++: LWG 3397 basic_istream_view::iterator should not provide iterator_category

libstdc++-v3/ChangeLog:

LWG 3397 basic_istream_view::iterator should not provide
iterator_category
* include/std/ranges (basic_istream_view:_Iterator::iterator_category):
Rename to ...
(basic_istream_view:_Iterator::iterator_concept): ... this.
* testsuite/std/ranges/istream_view.cc: Augment test.

4 years agolibstdc++: LWG 3325 Constrain return type of transformation function for transform_view
Patrick Palka [Mon, 24 Feb 2020 21:21:55 +0000 (16:21 -0500)]
libstdc++: LWG 3325 Constrain return type of transformation function for transform_view

libstdc++-v3/ChangeLog:

LWG 3325 Constrain return type of transformation function for
transform_view
* include/std/ranges (transform_view): Constrain the return type of the
transformation function as per LWG 3325.
* testsuite/std/ranges/adaptors/lwg3325_neg.cc: New test.

4 years agolibstdc++: LWG 3313 join_view::iterator::operator-- is incorrectly constrained
Patrick Palka [Mon, 24 Feb 2020 21:09:48 +0000 (16:09 -0500)]
libstdc++: LWG 3313 join_view::iterator::operator-- is incorrectly constrained

libstdc++-v3/ChangeLog:

LWG 3313 join_view::_Iterator::operator-- is incorrectly constrained
* include/std/ranges (join_view::_Iterator::operator--): Require that
range_reference_t<_Base> models common_range.
* testsuite/std/ranges/adaptors/lwg3313_neg.cc: New test.

4 years agolibstdc++: LWG 3301 transform_view::iterator has incorrect iterator_category
Patrick Palka [Mon, 24 Feb 2020 21:38:07 +0000 (16:38 -0500)]
libstdc++: LWG 3301 transform_view::iterator has incorrect iterator_category

libstdc++-v3/ChangeLog:

LWG 3301 transform_view::_Iterator has incorrect iterator_category
* include/std/ranges (transform_view::_Iterator::_S_iter_cat): Adjust
determination of iterator_category as per LWG 3301.
* testsuite/std/ranges/adaptors/transform.cc: Augment test.

4 years agolibstdc++: LWG 3292 iota_view is under-constrained
Patrick Palka [Mon, 24 Feb 2020 22:01:29 +0000 (17:01 -0500)]
libstdc++: LWG 3292 iota_view is under-constrained

libstdc++-v3/ChangeLog:

LWG 3292 iota_view is under-constrained
* include/std/ranges (iota_view): Require that _Winc models semiregular
  as per LWG 3292.
* testsuite/std/ranges/iota/lwg3292_neg.cc: New test.

4 years agoarm: ACLE intrinsics for bfloat16 dot product
Dennis Zhang [Tue, 25 Feb 2020 17:38:00 +0000 (17:38 +0000)]
arm: ACLE intrinsics for bfloat16 dot product

This patch is part of a series adding support for Armv8.6-A features.
It adds intrinsics for brain half-precision float-point (BF16) dot
instructions with AdvSIMD support.

gcc/ChangeLog:

2020-02-25  Dennis Zhang  <dennis.zhang@arm.com>

* config/arm/arm_neon.h (vbfdot_f32, vbfdotq_f32): New
(vbfdot_lane_f32, vbfdotq_laneq_f32): New.
(vbfdot_laneq_f32, vbfdotq_lane_f32): New.
* config/arm/arm_neon_builtins.def (vbfdot): New entry.
(vbfdot_lanev4bf, vbfdot_lanev8bf): Likewise.
* config/arm/iterators.md (VSF2BF): New attribute.
* config/arm/neon.md (neon_vbfdot<VCVTF:mode>): New entry.
(neon_vbfdot_lanev4bf<VCVTF:mode>): Likewise.
(neon_vbfdot_lanev8bf<VCVTF:mode>): Likewise.

gcc/testsuite/ChangeLog:

2020-02-25  Dennis Zhang  <dennis.zhang@arm.com>

* gcc.target/arm/simd/bf16_dot_1.c: New test.
* gcc.target/arm/simd/bf16_dot_2.c: New test.
* gcc.target/arm/simd/bf16_dot_3.c: New test.

4 years agolibstdc++: Remove __memmove wrapper for constexpr algorithms
Jonathan Wakely [Tue, 25 Feb 2020 14:16:42 +0000 (14:16 +0000)]
libstdc++: Remove __memmove wrapper for constexpr algorithms

The mutating sequence algorithms std::copy, std::copy_backward,
std::move and std::move_backward conditionally use __builtin_memmove
for trivially copyable types. However, because memmove isn't usable in
constant expressions the use of __builtin_memmove is wrapped in a
__memmove function which replaces __builtin_memmove with a handwritten
loop when std::is_constant_evaluated() is true.

This means we have a manual loop for non-trivially copyable cases, and a
different manual loop for trivially copyable but constexpr cases. The
latter loop has incorrect semantics for the {copy,move}_backward cases
and so isn't used for them. Until earlier today the latter loop also had
incorrect semantics for the std::move cases, trying to move from const
rvalues.

The approach taken by this patch is to remove the __memmove function
entirely and use the original (and correct) manual loops for the
constexpr cases as well as the non-trivially copyable cases. This was
already done for move_backward and copy_backward, but was incorrectly
turning copy_backward into move_backward, by failing to use the _IsMove
constant to select the right specialization. This patch also fixes that.

* include/bits/ranges_algobase.h (__copy_or_move): Do not use memmove
during constant evaluation. Call __builtin_memmove directly instead of
__memmove.
(__copy_or_move_backward): Likewise.
* include/bits/stl_algobase.h (__memmove): Remove.
(__copy_move<M, true, random_access_iterator_tag>::__copy_m)
(__copy_move_backward<M, true, random_access_iterator_tag>::__copy_m):
Use __builtin_memmove directly instead of __memmove.
(__copy_move_a2): Do not use memmove during constant evaluation.
(__copy_move_backward_a2): Use _IsMove constant to select correct
__copy_move_backward specialization.
* testsuite/25_algorithms/copy_backward/constexpr.cc: Check for copies
begin turned into moves during constant evaluation.

4 years agoFix ChangeLog date
Jonathan Wakely [Tue, 25 Feb 2020 16:58:52 +0000 (16:58 +0000)]
Fix ChangeLog date

4 years ago[ARM] Fix -mpure-code for v6m
Christophe Lyon [Tue, 25 Feb 2020 15:54:14 +0000 (15:54 +0000)]
[ARM] Fix -mpure-code for v6m

When running the testsuite with -fdisable-rtl-fwprop2 and -mpure-code
for cortex-m0, I noticed that some testcases were failing because we
still generate "ldr rX, .LCY", which is what we want to avoid with
-mpure-code. This is latent since a recent improvement in fwprop
(PR88833).

In this patch I change the thumb1_movsi_insn pattern so that it emits
the desired instruction sequence when arm_disable_literal_pool is set.

To achieve that, I introduce a new required_for_purecode attribute to
enable the corresponding alternative in thumb1_movsi_insn and take the
actual instruction sequence length into account.

gcc/ChangeLog:

2020-02-13  Christophe Lyon  <christophe.lyon@linaro.org>

* config/arm/arm.md (required_for_purecode): New attribute.
(enabled): Handle required_for_purecode.
* config/arm/thumb1.md (thumb1_movsi_insn): Add alternative to
work with -mpure-code.

4 years agocombine: Fix find_split_point handling of constant store into ZERO_EXTRACT [PR93908]
Jakub Jelinek [Tue, 25 Feb 2020 12:56:47 +0000 (13:56 +0100)]
combine: Fix find_split_point handling of constant store into ZERO_EXTRACT [PR93908]

git is miscompiled on s390x-linux with -O2 -march=zEC12 -mtune=z13.
I've managed to reduce it into the following testcase.  The problem is that
during combine we see the s->k = -1; bitfield store and change the SET_SRC
from a pseudo into a constant:
(set (zero_extract:DI (mem/j:HI (plus:DI (reg/v/f:DI 60 [ s ])
                (const_int 10 [0xa])) [0 +0 S2 A16])
        (const_int 2 [0x2])
        (const_int 7 [0x7]))
    (const_int -1 [0xffffffffffffffff]))
This on s390x with the above option isn't recognized as valid instruction,
so find_split_point decides to handle it as IOR or IOR/AND.
src is -1, mask is 3 and pos is 7.
src != mask (this is also incorrect, we want to set all (both) bits in the
bitfield), so we go for IOR/AND, but instead of trying
mem = (mem & ~0x180) | ((-1 << 7) & 0x180)
we actually try
mem = (mem & ~0x180) | (-1 << 7)
and that is further simplified into:
mem = mem | (-1 << 7)
aka
mem = mem | 0xff80
which doesn't set just the 2-bit bitfield, but also many other bitfields
that shouldn't be touched.
We really should do:
mem = mem | 0x180
instead.
The problem is that we assume that no bits but those low len (2 here) will
be set in the SET_SRC, but there is nothing that can prevent that, we just
should ignore the other bits.

The following patch fixes it by masking src with mask, this way already
the src == mask test will DTRT, and as the code for or_mask uses
gen_int_mode, if the most significant bit is set after shifting it left by
pos, it will be properly sign-extended.

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/93908
* combine.c (find_split_point): For store into ZERO_EXTRACT, and src
with mask.

* gcc.c-torture/execute/pr93908.c: New test.

4 years agolibstdc++: Add test accidentally left out of previous commit
Jonathan Wakely [Tue, 25 Feb 2020 12:42:03 +0000 (12:42 +0000)]
libstdc++: Add test accidentally left out of previous commit

* testsuite/25_algorithms/move_backward/93872.cc: Add test left out of
previous commit.

4 years agolibstdc++: Fix regression in std::move algorithm (PR 93872)
Jonathan Wakely [Tue, 25 Feb 2020 12:21:44 +0000 (12:21 +0000)]
libstdc++: Fix regression in std::move algorithm (PR 93872)

The std::move and std::move_backward algorithms dispatch to the
std::__memmove helper when appropriate. That function uses a
pointer-to-const for the source values, preventing them from being
moved. The two callers of that function have the same problem.

Rather than altering __memmove and its callers to work with const or
non-const source pointers, this takes a more conservative approach of
casting away the const at the point where we want to do a move
assignment. This relies on the fact that we only use __memmove when the
type is trivially copyable, so we know the move assignment doesn't alter
the source anyway.

PR libstdc++/93872
* include/bits/stl_algobase.h (__memmove): Cast away const before
doing move assignment.
* testsuite/25_algorithms/move/93872.cc: New test.
* testsuite/25_algorithms/move_backward/93872.cc: New test.

4 years agoFix link failure with debug info in LTO mode
Eric Botcazou [Tue, 25 Feb 2020 11:34:00 +0000 (12:34 +0100)]
Fix link failure with debug info in LTO mode

This fixes a regression whereby the program fails to link with debug
info in LTO mode because of an undefined reference to a symbol coming
from the object files containing the early debug info.

* dwarf2out.c (dwarf2out_size_function): Run in early-DWARF mode.

4 years agotestcase for last_vuse in FRE
Richard Biener [Tue, 25 Feb 2020 11:19:33 +0000 (12:19 +0100)]
testcase for last_vuse in FRE

This adds a testcase for some basic FRE functionality.

2020-02-25  Richard Biener  <rguenther@suse.de>

* gcc.dg/tree-ssa/ssa-fre-86.c: New testcase.

4 years agodoc: minor --enable-checking wording fixes
Roman Zhuykov [Tue, 25 Feb 2020 11:08:00 +0000 (14:08 +0300)]
doc: minor --enable-checking wording fixes

gcc/ChangeLog:
 doc/install.texi (--enable-checking): Adjust wording.

4 years agotree-optimization/93868 copy SLP tree before re-arranging stmts
Richard Biener [Tue, 25 Feb 2020 09:31:16 +0000 (10:31 +0100)]
tree-optimization/93868 copy SLP tree before re-arranging stmts

This avoids altering possibly shared SLP subtrees when attempting
to get rid of permutations in SLP reductions by copying the SLP
subtree before re-arranging stmts in it.

2020-02-25  Richard Biener  <rguenther@suse.de>

PR tree-optimization/93868
* tree-vect-slp.c (slp_copy_subtree): New function.
(vect_attempt_slp_rearrange_stmts): Copy the SLP tree before
re-arranging stmts in it.

* gcc.dg/torture/pr93868.c: New testcase.

4 years agopass_manager: Fix ICE with -fdump-passes -fdisable-tree-* [PR93874]
Jakub Jelinek [Tue, 25 Feb 2020 09:21:34 +0000 (10:21 +0100)]
pass_manager: Fix ICE with -fdump-passes -fdisable-tree-* [PR93874]

dump_passes pushes a dummy function for which it evaluates the gates
and checks whether the pass is enabled or disabled.
Unfortunately, if any -fdisable-*-*/-fenable-*-* options were seen,
we ICE during is_pass_explicitly_enabled_or_disabled because slot
is non-NULL then and the code will do:
  cgraph_uid = func ? cgraph_node::get (func)->get_uid () : 0;
but the dummy function doesn't have a cgraph node.

So, either we need to create and then remove a cgraph node for the dummy
function like the following patch, or function.c would need to export the
in_dummy_function flag (or have some way to query that flag from other TUs)
and we'd need to check it in is_pass_explicitly_enabled_or_disabled.

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/93874
* passes.c (pass_manager::dump_passes): Create a cgraph node for the
dummy function and remove it at the end.

* gcc.dg/pr93874.c: New test.

4 years agoDead code in fortran/simplify.c
Steven G. Kargl [Tue, 25 Feb 2020 08:25:33 +0000 (09:25 +0100)]
Dead code in fortran/simplify.c

* simplify.c (degrees_f): Remove unused code.

4 years agotestsuite: Fix recently added ipa testcases [PR93763]
Jakub Jelinek [Tue, 25 Feb 2020 08:05:57 +0000 (09:05 +0100)]
testsuite: Fix recently added ipa testcases [PR93763]

Seems the test has been badly reduced (if the original doesn't emit
warnings, it is always better in the reduction script avoid introducing new
ones).
Also, the g++.dg/ipa/ test fails with -std=c++98 because it is written in
C++11.

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR ipa/93763
* gcc.dg/ipa/pr93763.c: Adjust the test so that it compiles without
warnings and still ICEs before the ipa-cp.c fix.
* g++.dg/ipa/pr93763.C: Require c++11 effective target.

4 years agoc: Small diagnostics tweak - add missing ? after did you mean [PR93858]
Jakub Jelinek [Tue, 25 Feb 2020 08:04:39 +0000 (09:04 +0100)]
c: Small diagnostics tweak - add missing ? after did you mean [PR93858]

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR c/93858
* c-pragma.c (handle_pragma_diagnostic): Add missing ? after
"did you mean" hint in diagnostics.

4 years agoFix typo: paramter -> parameter [PR93864]
Jakub Jelinek [Tue, 25 Feb 2020 08:02:57 +0000 (09:02 +0100)]
Fix typo: paramter -> parameter [PR93864]

2020-02-25  Jakub Jelinek  <jakub@redhat.com>

PR translation/93864
* config/lm32/lm32.c (lm32_setup_incoming_varargs): Fix comment typo
paramter -> parameter.
* config/aarch64/aarch64.c (aarch64_is_extend_from_extract): Likewise.
* ipa-prop.h (struct ipa_agg_replacement_value): Likewise.

* intrinsic.texi (CO_BROADCAST): Fix typo, paramter -> parameter.
* trans-array.c (gfc_allocate_pdt_comp, gfc_deallocate_pdt_comp,
gfc_check_pdt_dummy): Fix comment typo paramter -> parameter.

* objc.dg/encode-2.m: Fix comment typo paramter -> parameter.
* obj-c++.dg/encode-4.mm: Likewise.
* gfortran.dg/data_array_5.f90: Likewise.
* gcc.dg/decl-1.c: Likewise.

4 years agodoc: properly describe --enable-checking behavior
Roman Zhuykov [Tue, 25 Feb 2020 07:15:49 +0000 (10:15 +0300)]
doc: properly describe --enable-checking behavior

This patch rewords the whole description to fix minor issues:
 - documents 'gimple' and 'types' checks,
 - clarifies what happens when option is used without '=list',
 - fixes inaccurate wrong wording about release snapshots,
 - describes that release checks can only de disabled explicitly.

gcc/ChangeLog:
* doc/install.texi (--enable-checking): Properly document current
behavior.
(--enable-stage1-checking): Minor clarification about bootstrap.

4 years agoanalyzer: fix -fdump-analyzer
David Malcolm [Mon, 24 Feb 2020 22:02:11 +0000 (17:02 -0500)]
analyzer: fix -fdump-analyzer

This patch fixes a bug with -fdump-analyzer, which is meant to write
purely a dumpfile, but was erroneously sending part of the dump to
stderr.

gcc/analyzer/ChangeLog:
* engine.cc (exploded_graph::get_or_create_node): Dump the
program_state to the pp, rather than to stderr.

4 years agoUpdate gcc de.po.
Joseph Myers [Tue, 25 Feb 2020 01:16:02 +0000 (01:16 +0000)]
Update gcc de.po.

* de.po: Update.

4 years agoDaily bump.
GCC Administrator [Tue, 25 Feb 2020 00:16:33 +0000 (00:16 +0000)]
Daily bump.

4 years agoanalyzer: disable the "taint" checker by default
David Malcolm [Fri, 21 Feb 2020 15:50:16 +0000 (10:50 -0500)]
analyzer: disable the "taint" checker by default

PR analyzer/93032 tracks a false negative where we fail to report
FILE * leaks within zlib/contrib/minizip/mztools.c.

The underlying issue is a combinatorial explosion of states within the
exploded graph.  In particular, the state of the "taint" checker is
exploding, leading to the analyzer bailing out.

I have a patch kit under construction that fixes the state explosion
issue enough for the "file" checker to report the leaks, but doing so
requires disabling the "taint" checker.  Given that the latter is more
of a proof-of-concept, this patch disables it by default, to stop it
breaking the other checkers.

gcc/analyzer/ChangeLog:
PR analyzer/93032
* sm.cc (make_checkers): Require the "taint" checker to be
explicitly enabled.

gcc/ChangeLog:
PR analyzer/93032
* doc/invoke.texi (-Wnanalyzer-tainted-array-index): Note that
-fanalyzer-checker=taint is also required.
(-fanalyzer-checker=): Note that providing this option enables the
given checker, and doing so may be required for checkers that are
disabled by default.

gcc/testsuite/ChangeLog:
PR analyzer/93032
* gcc.dg/analyzer/pr93382.c: Add "-fanalyzer-checker=taint".
* gcc.dg/analyzer/taint-1.c: Likewise.

4 years agoanalyzer: fix ICE with OFFSET_TYPE [PR 93899]
David Malcolm [Mon, 24 Feb 2020 17:12:28 +0000 (12:12 -0500)]
analyzer: fix ICE with OFFSET_TYPE [PR 93899]

PR analyzer/93899 reports an ICE within make_region_for_type when
handling a param of type OFFSET_TYPE within
exploded_graph::add_function_entry.

This patch fixes the ICE by further generalizing the "give up on this
tree code" logic from
  r10-6667-gf76a88ebf089871dcce215aa0cb1956ccc060895
     for PR analyzer/93388 and
  r10-6695-g2e6233935c77b56a68e939c629702f960b8e6fb2
     for PR analyzer/93778
by replacing the gcc_unreachable in make_region_for_type with a return
of NULL, and handling this in add_region_for_type by notifying the ctxt.

Doing so means that numerous places that create regions now need to have
a context passed to them, so most of the patch is churn involved in
passing a context around to where it's needed.

gcc/analyzer/ChangeLog:
PR analyzer/93899
* engine.cc
(impl_region_model_context::impl_region_model_context): Add logger
param.
* engine.cc (exploded_graph::add_function_entry): Create an
impl_region_model_context and pass it to the push_frame call.
Bail if the resulting state is invalid.
(exploded_graph::build_initial_worklist): Likewise.
(exploded_graph::build_initial_worklist): Handle the case where
add_function_entry fails.
* exploded-graph.h
(impl_region_model_context::impl_region_model_context): Add logger
param.
* region-model.cc (map_region::get_or_create): Add ctxt param and
pass it to add_region_for_type.
(map_region::can_merge_p): Pass NULL as a ctxt to call to
get_or_create.
(array_region::get_element): Pass ctxt to call to get_or_create.
(array_region::get_or_create): Add ctxt param and pass it to
add_region_for_type.
(root_region::push_frame): Pass ctxt to get_or_create calls.
(region_model::get_lvalue_1): Likewise.
(region_model::make_region_for_unexpected_tree_code): Assert that
ctxt is non-NULL.
(region_model::get_rvalue_1): Pass ctxt to get_svalue_for_fndecl
and get_svalue_for_label calls.
(region_model::get_svalue_for_fndecl): Add ctxt param and pass it
to get_region_for_fndecl.
(region_model::get_region_for_fndecl): Add ctxt param and pass it
to get_or_create.
(region_model::get_svalue_for_label): Add ctxt param and pass it
to get_region_for_label.
(region_model::get_region_for_label): Add ctxt param and pass it
to get_region_for_fndecl and get_or_create.
(region_model::get_field_region): Add ctxt param and pass it to
get_or_create_view and get_or_create.
(make_region_for_type): Replace gcc_unreachable with return NULL.
(region_model::add_region_for_type): Add ctxt param.  Handle a
return of NULL from make_region_for_type by calling
make_region_for_unexpected_tree_code.
(region_model::get_or_create_mem_ref): Pass ctxt to calls to
get_or_create_view.
(region_model::get_or_create_view): Add ctxt param and pass it to
add_region_for_type.
(selftest::test_state_merging): Pass ctxt to get_or_create_view.
* region-model.h (region_model::get_or_create): Add ctxt param.
(region_model::add_region_for_type): Likewise.
(region_model::get_svalue_for_fndecl): Likewise.
(region_model::get_svalue_for_label): Likewise.
(region_model::get_region_for_fndecl): Likewise.
(region_model::get_region_for_label): Likewise.
(region_model::get_field_region): Likewise.
(region_model::get_or_create_view): Likewise.

gcc/testsuite/ChangeLog:
PR analyzer/93899
* g++.dg/analyzer/pr93899.C: New test.

4 years agointernal/poll: add hurd build tag
Ian Lance Taylor [Mon, 24 Feb 2020 17:27:03 +0000 (09:27 -0800)]
internal/poll: add hurd build tag

Patch from Svante Signell.

Fixes GCC PR go/93900

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/220592

4 years agoRemove a hunk duplicated during a merge.
Martin Sebor [Mon, 24 Feb 2020 17:18:11 +0000 (10:18 -0700)]
Remove a hunk duplicated during a merge.

gcc/cp/ChangeLog:
* parser.c (cp_parser_check_class_key): Remove a duplicate hunk
of code.

4 years agoPR c++/93804 - exempt extern C headers from -Wredundant-tags
Martin Sebor [Mon, 24 Feb 2020 17:14:16 +0000 (10:14 -0700)]
PR c++/93804 - exempt extern C headers from -Wredundant-tags

gcc/cp/ChangeLog:

PR c++/93804
* parser.c (cp_parser_check_class_key): Avoid issuing -Wredundant-tags
in shared C/C++ code in headers.

gcc/testsuite/ChangeLog:

PR c++/93804
* g++.dg/warn/Wredundant-tags-4.C: New test.
* g++.dg/warn/Wredundant-tags-5.C: New test.
* g++.dg/warn/Wredundant-tags-5.h: New test.

4 years agoanalyzer: eliminate irrelevant control-flow edges from paths
David Malcolm [Sat, 22 Feb 2020 00:25:40 +0000 (19:25 -0500)]
analyzer: eliminate irrelevant control-flow edges from paths

Paths emitted by the analyzer can be quite verbose at the default of
-fanalyzer-verbosity=2.

Consider the double-free in this example:

  #include <stdlib.h>

  int foo ();
  int bar ();

  void test (int a, int b, int c)
  {
    void *p = malloc (1024);
    while (a)
      foo ();
    if (b)
      foo ();
    else
      bar ();
    if (c)
      free (p);
    free (p);
  }

Previously, the analyzer would emit a checker_path containing all
control-flow information on the exploded_path leading to the
double-free:

  test.c: In function 'test':
  test.c:17:3: warning: double-'free' of 'p' [CWE-415] [-Wanalyzer-double-free]
     17 |   free (p);
        |   ^~~~~~~~
    'test': events 1-9
      |
      |    8 |   void *p = malloc (1024);
      |      |             ^~~~~~~~~~~~~
      |      |             |
      |      |             (1) allocated here
      |    9 |   while (a)
      |      |         ~
      |      |         |
      |      |         (2) following 'false' branch (when 'a == 0')...
      |   10 |     foo ();
      |   11 |   if (b)
      |      |      ~
      |      |      |
      |      |      (3) ...to here
      |      |      (4) following 'false' branch (when 'b == 0')...
      |......
      |   14 |     bar ();
      |      |     ~~~~~~
      |      |     |
      |      |     (5) ...to here
      |   15 |   if (c)
      |      |      ~
      |      |      |
      |      |      (6) following 'true' branch (when 'c != 0')...
      |   16 |     free (p);
      |      |     ~~~~~~~~
      |      |     |
      |      |     (7) ...to here
      |      |     (8) first 'free' here
      |   17 |   free (p);
      |      |   ~~~~~~~~
      |      |   |
      |      |   (9) second 'free' here; first 'free' was at (8)
      |

despite the fact that only the "if (c)" is relevant to triggering the
double-free.

This patch implements pruning of control flow events at
-fanalyzer-verbosity=2, based on reachability information within the
exploded_graph.
The diagnostic_manager pre-computes reachability information about
which exploded_nodes can reach the exploded_node of the diagnostic,
and uses this to prune irrelvent control flow edges.

The patch also adds a -fanalyzer-verbosity=3 to preserve these edges,
so that the "show me everything" debugging level becomes
-fanalyzer-verbosity=4.

With these changes, the "while (a)" and "if (b)" edges are pruned from
the above example, leading to:

  test.c: In function 'test':
  test.c:17:3: warning: double-'free' of 'p' [CWE-415] [-Wanalyzer-double-free]
     17 |   free (p);
        |   ^~~~~~~~
    'test': events 1-5
      |
      |    8 |   void *p = malloc (1024);
      |      |             ^~~~~~~~~~~~~
      |      |             |
      |      |             (1) allocated here
      |......
      |   15 |   if (c)
      |      |      ~
      |      |      |
      |      |      (2) following 'true' branch (when 'c != 0')...
      |   16 |     free (p);
      |      |     ~~~~~~~~
      |      |     |
      |      |     (3) ...to here
      |      |     (4) first 'free' here
      |   17 |   free (p);
      |      |   ~~~~~~~~
      |      |   |
      |      |   (5) second 'free' here; first 'free' was at (4)
      |

The above example is gcc.dg/analyzer/edges-2.c.

gcc/analyzer/ChangeLog:
* checker-path.cc (superedge_event::should_filter_p): Update
filter for empty descriptions to cover verbosity level 3 as well
as 2.
* diagnostic-manager.cc: Include "analyzer/reachability.h".
(class path_builder): New class.
(diagnostic_manager::emit_saved_diagnostic): Create a path_builder
and pass it to build_emission_path, rather passing eg; similarly
for add_events_for_eedge and ext_state.
(diagnostic_manager::build_emission_path): Replace "eg" param
with a path_builder, pass it to add_events_for_eedge.
(diagnostic_manager::add_events_for_eedge): Replace ext_state
param with path_builder; pass it to add_events_for_superedge.
(diagnostic_manager::significant_edge_p): New.
(diagnostic_manager::add_events_for_superedge): Add path_builder
param.  Reject insignificant edges at verbosity levels below 3.
(diagnostic_manager::prune_for_sm_diagnostic): Update highest
verbosity level to 4.
* diagnostic-manager.h (class path_builder): New forward decl.
(diagnostic_manager::build_emission_path): Replace "eg" param
with a path_builder.
(diagnostic_manager::add_events_for_eedge): Replace ext_state
param with path_builder.
(diagnostic_manager::significant_edge_p): New.
(diagnostic_manager::add_events_for_superedge): Add path_builder
param.
* reachability.h: New file.

gcc/ChangeLog:
* doc/invoke.texi (-fanalyzer-verbosity=): "2" only shows
significant control flow events; add a "3" which shows all
control flow events; the old "3" becomes "4".

gcc/testsuite/ChangeLog:
* gcc.dg/analyzer/analyzer-verbosity-2a.c: New test.
* gcc.dg/analyzer/analyzer-verbosity-3.c: New test, based on
analyzer-verbosity-2.c
* gcc.dg/analyzer/analyzer-verbosity-3a.c: New test.
* gcc.dg/analyzer/edges-1.c: New test.
* gcc.dg/analyzer/edges-2.c: New test.
* gcc.dg/analyzer/file-paths-1.c: Add -fanalyzer-verbosity=3.

4 years agoc++: Fix ICE with -Wmismatched-tags [PR93869]
Marek Polacek [Fri, 21 Feb 2020 17:58:04 +0000 (12:58 -0500)]
c++: Fix ICE with -Wmismatched-tags [PR93869]

This is a crash in cp_parser_check_class_key:
  tree type_decl = TYPE_MAIN_DECL (type);
  tree name = DECL_NAME (type_decl); // HERE
because TYPE_MAIN_DECL of type was null as it's not a class type.
Instead of checking CLASS_TYPE_P we should simply check class_key
a bit earlier (in this case it was typename_type).

2020-02-24  Marek Polacek  <polacek@redhat.com>

PR c++/93869 - ICE with -Wmismatched-tags.
* parser.c (cp_parser_check_class_key): Check class_key earlier.

* g++.dg/warn/Wmismatched-tags-2.C: New test.

4 years agoortran: ICE using SHAPE with FINDLOC PR93835
Mark Eggleston [Mon, 24 Feb 2020 15:40:03 +0000 (15:40 +0000)]
ortran: ICE using SHAPE with FINDLOC PR93835

The expression representing the array returned by SHAPE does not
have its shape defined. An ICE occurs when FINDLOC attempts to
use the shape of the array.  Add shape to expression before returning
from SHAPE.

Whitespace issues identified by Steven G. Kargl  <kargl@gcc.gnu.org>
have also been fixed.

gcc/fortran/ChangeLog

PR fortran/93835
* simplify.c (simplify_findloc_nodim) : Fix whitespace issues.
(gfc_simplify_shape) : Create and initialise one shape value
for the result expression. Set shape value with the rank of
the source array.

gcc/testsuite/ChangeLog

PR fortran/93835
* gfortran.dg/pr77351.f90 : Check for one error instead of two.
* gfortran.dg/pr93835.f08 : New test.

4 years agoc++: Fix ICE with ill-formed array list-initialization [PR93712]
Marek Polacek [Thu, 13 Feb 2020 19:05:51 +0000 (14:05 -0500)]
c++: Fix ICE with ill-formed array list-initialization [PR93712]

My P0388R4 patch changed build_array_conv to create an identity
conversion at the start of the conversion chain and now we crash
in convert_like_real:

 7457     case ck_identity:
 7458       if (BRACE_ENCLOSED_INITIALIZER_P (expr))
 7459         {
 7460           int nelts = CONSTRUCTOR_NELTS (expr);
 7461           if (nelts == 0)
 7462             expr = build_value_init (totype, complain);
 7463           else if (nelts == 1)
 7464             expr = CONSTRUCTOR_ELT (expr, 0)->value;
 7465           else
 7466             gcc_unreachable ();  // HERE
 7467         }

in a test like this

  int f (int const (&)[2])
  { return f({1, "M"}); }

Instead of creating a ck_identity at the start of the conversion chain,
so that conv_get_original_expr can be used with a ck_aggr, let's set
u.expr for a ck_aggr, and adjust next_conversion not to try to see
what's next in the chain if it gets a ck_aggr.

2020-02-24  Marek Polacek  <polacek@redhat.com>

PR c++/93712 - ICE with ill-formed array list-initialization.
* call.c (next_conversion): Return NULL for ck_aggr.
(build_aggr_conv): Set u.expr instead of u.next.
(build_array_conv): Likewise.
(build_complex_conv): Likewise.
(conv_get_original_expr): Handle ck_aggr.

* g++.dg/cpp0x/initlist-array11.C: New test.

4 years agolibstdc++: Add missing bits of P0896R4 pertaining to [back|front]_insert_iterator
Patrick Palka [Mon, 24 Feb 2020 12:59:08 +0000 (07:59 -0500)]
libstdc++: Add missing bits of P0896R4 pertaining to [back|front]_insert_iterator

This adds some missing pieces of the Ranges TS that make back_insert_iterator and
front_insert_iterator conform to the new output_iterator requirements.

It also fixes a bug in ranges::__copy_or_move and
ranges::__copy_or_move_backward in which we were inspecting the iter_value_t of
the output iterator, but output iterators such as back_insert_iterator and
front_insert_iterator whose value_type = void do not have an iter_value_t
according to [readable.traits] p4.  The entire __use_memmove condition should
probably be rewritten, but the simplest fix for now is to inspect the
iterator_traits of the output iterator instead.

libstdc++-v3/ChangeLog:

PR libstdc++/93884
* include/bits/ranges_algobase.h (__copy_or_move,
__copy_or_move_backward): Don't inspect the iter_value_t of the output
iterator, instead inspect its iterator_traits directly.
* include/bits/stl_iterator.h (back_insert_iterator::container):
Conditionally initialize.
(back_insert_iterator::difference_type): Conditionally define.
(back_insert_iterator::back_insert_iterator): Conditionally define this
default constructor.
(front_insert_iterator::container): Conditionally initialize.
(front_insert_iterator::difference_type): Conditionally define.
(front_insert_iterator::front_insert_iterator): Conditionally define
this default constructor.
* 24_iterators/back_insert_iterator/pr93884.cc: New test.
* 24_iterators/front_insert_iterator/pr93884.cc: New test.

4 years agolibstdc++: P0769R2 Add shift to <algorithm>
Patrick Palka [Fri, 21 Feb 2020 18:55:01 +0000 (13:55 -0500)]
libstdc++: P0769R2 Add shift to <algorithm>

This patch adds std::shift_left and std::shift_right as per P0769R2.  Alhough
these are STL-style algos, this patch places them in <bits/ranges_algo.h>
because they make use of some functions in the ranges namespace that are more
easily reachable from <bits/ranges_algo.h> than from <bits/stl_algo.h>, namely
ranges::next.  In order to place these algos in <bits/stl_algo.h>, we would need
to include <bits/range_access.h> from <bits/stl_algo.h> which would undesirably
increase the size of <bits/stl_algo.h>.

libstdc++-v3/ChangeLog:

P0769R2 Add shift to <algorithm>
* include/bits/ranges_algo.h (shift_left, shift_right): New.
* testsuite/25_algorithms/shift_left/1.cc: New test.
* testsuite/25_algorithms/shift_right/1.cc: New test.

4 years agointernal/syscall/unix: add hurd build tag
Ian Lance Taylor [Mon, 24 Feb 2020 14:38:11 +0000 (06:38 -0800)]
internal/syscall/unix: add hurd build tag

Patch from Svante Signell.

Fixes GCC PR go/93900

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/220589

4 years agofortran: ICE in gfc_conv_constant_to_tree PR93604
Mark Eggleston [Mon, 24 Feb 2020 14:33:35 +0000 (14:33 +0000)]
fortran: ICE in gfc_conv_constant_to_tree PR93604

Using a BOZ constant in a structure constructor in a data statement
resulted in an ICE. Output a "BOZ literal constant cannot appear in
a structure constructor" error message instead.

Original patch provided by Steven G. Kargl  <kargl@gcc.gnu.org>.

Test case added later.

gcc/fortran/ChangeLog

PR fortran/93604
* decl.c (gfc_match_data) : Check whether the data expression
is a derived type and is a constructor. If a BOZ constant
is encountered in the constructor output an error and return
MATCH_ERROR.

gcc/testsuite/ChangeLog

PR fortran/93604
* gfortran.dg/pr93604.f90 : New test.

4 years agoc++: P1937R2 - Fixing inconsistencies between const{expr,eval} functions
Jakub Jelinek [Mon, 24 Feb 2020 14:23:23 +0000 (15:23 +0100)]
c++: P1937R2 - Fixing inconsistencies between const{expr,eval} functions

The following patch implements my understanding of P1937R2, though I wonder
if https://eel.is/c++draft/expr.const#14.example-1 shouldn't have been
also either removed or adjusted by the P1937R2 paper.

2020-02-24  Jakub Jelinek  <jakub@redhat.com>

P1937R2 - Fixing inconsistencies between const{expr,eval} functions
* call.c (build_over_call): Don't evaluate immediate functions in
unevaluated operands.

* g++.dg/ext/consteval1.C: Change dg-{message,error} into dg-bogus.
* g++.dg/cpp2a/consteval6.C: Likewise.
* g++.dg/cpp2a/consteval3.C: Change dg-error for unevaluated operands
into dg-bogus.

4 years agolibstdc++: Fix noexcept-specifier for istream_iterator
Jonathan Wakely [Mon, 24 Feb 2020 14:22:21 +0000 (14:22 +0000)]
libstdc++: Fix noexcept-specifier for istream_iterator

Somehow I missed that the _M_value member can throw on construction.

* include/bits/stream_iterator.h (istream_iterator(default_sentinel_t)):
Make noexcept-specifier conditional.
* testsuite/24_iterators/istream_iterator/cons/sentinel.cc: Check
noexcept-specifier.

4 years agoc++: Fix C++20 variadic lambda init-capture grammar.
Jason Merrill [Mon, 24 Feb 2020 01:52:41 +0000 (20:52 -0500)]
c++: Fix C++20 variadic lambda init-capture grammar.

The grammar for variadic init-capture was fixed at the Prague C++ meeting
where we finalized C++20.

gcc/cp/ChangeLog
2020-02-24  Jason Merrill  <jason@redhat.com>

P0780R2: Resolve lambda init-capture pack grammar.
* parser.c (cp_parser_lambda_introducer): Expect &...x=y rather than
...&x=y.

4 years agolibstdc++: Add default_sentinel support to stream iterators
Jonathan Wakely [Mon, 24 Feb 2020 13:11:31 +0000 (13:11 +0000)]
libstdc++: Add default_sentinel support to stream iterators

Missing pieces of P0896R4 "The One Ranges Proposal" for C++20.

* include/bits/stream_iterator.h (istream_iterator(default_sentinel_t)):
Add constructor.
(operator==(istream_iterator, default_sentinel_t)): Add operator.
(ostream_iterator::difference_type): Define to ptrdiff_t for C++20.
* include/bits/streambuf_iterator.h
(istreambuf_iterator(default_sentinel_t)): Add constructor.
(operator==(istreambuf_iterator, default_sentinel_t)): Add operator.
* testsuite/24_iterators/istream_iterator/cons/sentinel.cc:
New test.
* testsuite/24_iterators/istream_iterator/sentinel.cc: New test.
* testsuite/24_iterators/istreambuf_iterator/cons/sentinel.cc:
New test.
* testsuite/24_iterators/istreambuf_iterator/sentinel.cc: New test.

4 years agoPR78353: Fix testcases
Christophe Lyon [Mon, 24 Feb 2020 13:01:52 +0000 (13:01 +0000)]
PR78353: Fix testcases

Skip the test if arm7a is not supported at link time. This is the case
if the toolchain targets an M-profile CPU by default and does not have
A-profile multilib: the link step fails because it tries to mix
M-profile startup files with A-profile testcase.

2020-02-24  Christophe Lyon  <christophe.lyon@linaro.org>

PR lto/78353
* gcc.target/arm/pr78353-1.c: Add arm_arch_v7a_multilib effective
target.
* gcc.target/arm/pr78353-2.c: Likewise.

4 years agolibstdc++: enable_view has false positives (LWG 3326)
Jonathan Wakely [Mon, 24 Feb 2020 11:45:20 +0000 (11:45 +0000)]
libstdc++: enable_view has false positives (LWG 3326)

* include/std/ranges (__deep_const_range, __enable_view_impl): Remove.
(ranges::enable_view): Simplify (LWG 3326).
* include/bits/range_access.h (ranges::enable_view): Declare.
* include/bits/regex.h (__enable_view_impl): Remove partial
specialization.
* include/bits/stl_multiset.h (__enable_view_impl): Likewise.
* include/bits/stl_set.h (__enable_view_impl): Likewise.
* include/bits/unordered_set.h (__enable_view_impl): Likewise.
* include/debug/multiset.h (__enable_view_impl): Likewise.
* include/debug/set.h (__enable_view_impl): Likewise.
* include/debug/unordered_set (__enable_view_impl): Likewise.
* include/experimental/string_view (ranges::enable_view): Define
partial specialization.
* include/std/span (ranges::enable_view): Likewise.
* include/std/string_view (ranges::enable_view): Likewise.
* testsuite/std/ranges/view.cc: Check satisfaction of updated concept.

4 years agosccvn: Handle bitfields in push_partial_def [PR93582]
Jakub Jelinek [Mon, 24 Feb 2020 11:56:39 +0000 (12:56 +0100)]
sccvn: Handle bitfields in push_partial_def [PR93582]

The following patch adds support for bitfields to push_partial_def.
Previously pd.offset and pd.size were counted in bytes and maxsizei
in bits, now everything is counted in bits.

Not really sure how much of the further code can be outlined and moved, e.g.
the full def and partial def code doesn't have pretty much anything in
common (the partial defs case basically have some load bit range and a set
of store bit ranges that at least partially overlap and we need to handle
all the different cases, like negative pd.offset or non-negative, little vs.
bit endian, size so small that we need to preserve original bits on both
sides of the byte, size that fits or is too large.
Perhaps the storing of some value into a middle of existing buffer (i.e.
what push_partial_def now does in the loop) could, but the candidate for
sharing would be most likely store-merging rather than the other spots in
sccvn, and I think it is better not to touch store-merging at this stage.

Yes, I've thought about trying to do everything in place, but the code is
quite hard to understand and get right already now and if we tried to do the
optimize on the fly, it would need more special cases and would for gcov
coverage need more testcases to cover it.  Most of the time the sizes will
be small.  Furthermore, for bitfields native_encode_expr stores actually
number of bytes in the mode and not say actual bitsize rounded up to bytes,
so it wouldn't be just a matter of saving/restoring bytes at the start and
end, but we might need even 7 further bytes e.g. for __int128 bitfields.
Perhaps we could have just a fast path for the case where everything is byte
aligned and (for integral types the mode bitsize is equal to the size too)?

2020-02-24  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/93582
* tree-ssa-sccvn.c (vn_walk_cb_data::push_partial_def): Consider
pd.offset and pd.size to be counted in bits rather than bytes, add
support for maxsizei that is not a multiple of BITS_PER_UNIT and
handle bitfield stores and loads.
(vn_reference_lookup_3): Don't call ranges_known_overlap_p with
uncomparable quantities - bytes vs. bits.  Allow push_partial_def
on offsets/sizes that aren't multiple of BITS_PER_UNIT and adjust
pd.offset/pd.size to be counted in bits rather than bytes.
Formatting fix.  Rename shadowed len variable to buflen.

* gcc.dg/tree-ssa/pr93582-4.c: New test.
* gcc.dg/tree-ssa/pr93582-5.c: New test.
* gcc.dg/tree-ssa/pr93582-6.c: New test.
* gcc.dg/tree-ssa/pr93582-7.c: New test.
* gcc.dg/tree-ssa/pr93582-8.c: New test.

4 years agoOpenACC tile clause – apply exit/cycle checks (PR 93552)
Tobias Burnus [Mon, 24 Feb 2020 11:18:04 +0000 (12:18 +0100)]
OpenACC tile clause – apply exit/cycle checks (PR 93552)

        PR fortran/93552
        * match.c (match_exit_cycle): With OpenACC, check the kernels loop
        directive and tile clause as well.

        PR fortran/93552
        * gfortran.dg/goacc/tile-4.f90: New.

4 years agoPR47785: Add support for handling Xassembler/Wa options with LTO.
Prathamesh Kulkarni [Mon, 24 Feb 2020 06:25:45 +0000 (11:55 +0530)]
PR47785: Add support for handling Xassembler/Wa options with LTO.

2020-02-24  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
    Kugan Vivekandarajah  <kugan.vivekanandarajah@linaro.org>

PR driver/47785
* gcc.c (putenv_COLLECT_AS_OPTIONS): New function.
(driver::main): Call putenv_COLLECT_AS_OPTIONS.
* opts-common.c (parse_options_from_collect_gcc_options): New function.
(prepend_xassembler_to_collect_as_options): Likewise.
* opts.h (parse_options_from_collect_gcc_options): Declare prototype.
(prepend_xassembler_to_collect_as_options): Likewise.
* lto-opts.c (lto_write_options): Stream assembler options
in COLLECT_AS_OPTIONS.
* lto-wrapper.c (xassembler_options_error): New static variable.
(get_options_from_collect_gcc_options): Move parsing options code to
parse_options_from_collect_gcc_options and call it.
(merge_and_complain): Validate -Xassembler options.
(append_compiler_options): Handle OPT_Xassembler.
(run_gcc): Append command line -Xassembler options to
collect_gcc_options.
* doc/invoke.texi: Add documentation about using Xassembler
options with LTO.

testsuite/
* gcc.target/arm/pr78353-1.c: New test.
* gcc.target/arm/pr78353-2.c: Likewise.

4 years agoRISC-V: Adjust floating point code gen for LTGT compare
Kito Cheng [Mon, 24 Feb 2020 16:54:21 +0000 (10:54 -0600)]
RISC-V: Adjust floating point code gen for LTGT compare

 - Using gcc.dg/torture/pr91323.c as testcase, so no new testcase
   introduced.

 - We use 3 eq compare for LTGT compare before, in order to prevent exception
   flags setting when any input is NaN.

 - According latest GCC document LTGT and discussion on pr91323
   LTGT should signals on NaNs, like GE/GT/LE/LT.

 - So we expand (LTGT a b) to ((LT a b) | (GT a b)) for fit the document.

 - Tested rv64gc/rv32gc bare-metal/linux on qemu and
   rv64gc on HiFive unleashed board with linux.

ChangeLog

gcc/

Kito Cheng  <kito.cheng@sifive.com>

* config/riscv/riscv.c (riscv_emit_float_compare): Change the code gen
for LTGT.
(riscv_rtx_costs): Update cost model for LTGT.

4 years agoDaily bump.
GCC Administrator [Mon, 24 Feb 2020 00:16:40 +0000 (00:16 +0000)]
Daily bump.

4 years agoChanging cost propagation and ordering colorable bucket heuristics for PR93564.
Vladimir N. Makarov [Sun, 23 Feb 2020 21:20:05 +0000 (16:20 -0500)]
Changing cost propagation and ordering colorable bucket heuristics for PR93564.

2020-02-23  Vladimir Makarov  <vmakarov@redhat.com>

PR rtl-optimization/93564
* ira-color.c (struct update_cost_queue_elem): New member start.
(queue_update_cost, get_next_update_cost): Add new arg start.
(allocnos_conflict_p): New function.
(update_costs_from_allocno): Add new arg conflict_cost_update_p.
Add checking conflicts with allocnos_conflict_p.
(update_costs_from_prefs, restore_costs_from_copies): Adjust
update_costs_from_allocno calls.
(update_conflict_hard_regno_costs): Add checking conflicts with
allocnos_conflict_p.  Adjust calls of queue_update_cost and
get_next_update_cost.
(assign_hard_reg): Adjust calls of queue_update_cost.  Add
debugging print.
(bucket_allocno_compare_func): Restore previous version.

4 years agoAdd missing closing parenthises in error message.
Thomas König [Sun, 23 Feb 2020 16:22:26 +0000 (17:22 +0100)]
Add missing closing parenthises in error message.

2020-02-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/93889
* interface.c (compare_parameter): Fix error message.

4 years agoFix error message.
Thomas König [Sun, 23 Feb 2020 16:04:06 +0000 (17:04 +0100)]
Fix error message.

2020-02-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/93890
* interface.c: Replace "can not" by "cannot" and remove trailing
space.

2020-02-23  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/93890
* gfortran.dg/argument_checking_24.f90: Correct test case.

4 years agoPatch for PR57710
Paul Thomas [Sun, 23 Feb 2020 10:27:37 +0000 (10:27 +0000)]
Patch for PR57710

4 years agoDaily bump.
GCC Administrator [Sun, 23 Feb 2020 00:16:44 +0000 (00:16 +0000)]
Daily bump.

4 years agolibatomic: Fix last change [PR55930]
Jakub Jelinek [Sat, 22 Feb 2020 18:55:09 +0000 (19:55 +0100)]
libatomic: Fix last change [PR55930]

2020-02-22  Jakub Jelinek  <jakub@redhat.com>

PR other/55930
* Makefile.am (M_DEPS): Guard the empty definition with
@AMDEP_FALSE@ rather than @AMDEP_TRUE@.
* Makefile.in: Regenerated.

4 years agoc++: Use %qs in diagnostic message [PR93882]
Marek Polacek [Sat, 22 Feb 2020 16:53:45 +0000 (11:53 -0500)]
c++: Use %qs in diagnostic message [PR93882]

A tweak for translators, as requested in the PR.

2020-02-22  Marek Polacek  <polacek@redhat.com>

PR c++/93882
* decl.c (grokdeclarator): Use %qs in a diagnostic message.

4 years agoHonor --disable-dependency-tracking in libatomic
Richarde Purdie [Sat, 22 Feb 2020 15:13:13 +0000 (10:13 -0500)]
Honor --disable-dependency-tracking in libatomic

PR other/55930
* Makefile.am (M_DEPS): Honor -disable-dependency-tracking.
* Makefile.in: Regenerated.

4 years agoDaily bump.
GCC Administrator [Sat, 22 Feb 2020 00:16:30 +0000 (00:16 +0000)]
Daily bump.

4 years agoFix handling of floating-point homogeneous aggregates.
John David Anglin [Fri, 21 Feb 2020 23:30:24 +0000 (23:30 +0000)]
Fix handling of floating-point homogeneous aggregates.

2020-02-21  John David Anglin  <danglin@gcc.gnu.org>

* gcc/config/pa/pa.c (pa_function_value): Fix check for word and
double-word size when handling aggregate return values.
* gcc/config/pa/som.h (ASM_DECLARE_FUNCTION_NAME): Fix to indicate
that homogeneous SFmode and DFmode aggregates are passed and returned
in general registers.

4 years agoi18n: Fix translation of --help [PR93759]
Jakub Jelinek [Fri, 21 Feb 2020 21:01:03 +0000 (22:01 +0100)]
i18n: Fix translation of --help [PR93759]

The first two hunks make sure we actually translate what has been marked
for translation, i.e. the cl_options[...].help strings, rather than those
strings ammended in various ways, like:
_("%s  Same as %s."), help, ...
or
"%s  %s", help, _(use_diagnosed_msg)

The exgettext changes attempt to make sure that the cl_options[...].help
strings are marked as no-c-format, because otherwise if they happen
to contain a % character, such as the 90% substring, they will be marked
as c-format, which they aren't.

2020-02-21  Jakub Jelinek  <jakub@redhat.com>

PR translation/93759
* opts.c (print_filtered_help): Translate help before appending
messages to it rather than after that.

* exgettext: For *.opt help texts, use __opt_help_text("...")
rather than _("...") in the $emsg file and pass options that
say that this implies no-c-format.

4 years agolra: Stop registers being incorrectly marked live v2 [PR92989]
Richard Sandiford [Wed, 19 Feb 2020 12:59:17 +0000 (12:59 +0000)]
lra: Stop registers being incorrectly marked live v2 [PR92989]

This PR is about a case in which the clobbers at the start of
an EH receiver can lead to registers becoming unnecessarily
live in predecessor blocks.  My first attempt at fixing this
made sure that we update the bb liveness info based on the
real live set:

  http://gcc.gnu.org/g:e648e57efca6ce6d751ef8c2038608817b514fb4

But it turns out that the clobbered registers were also added to
the "gen" set of LRA's private liveness problem, where "gen" in
this context means "generates a requirement for a live value".
So the clobbered registers could still end up live via that
mechanism instead.

This patch therefore reverts the patch above and takes the other
approach floated in the original patch description: model the full
clobber by making the registers live and then dead again.

There's no specific need to revert the original patch, since the
code should no longer be sensitive to the order of the bb liveness
update and the modelling of the clobber.  But given that there's
no specific need to keep the original patch either, it seemed better
to restore the code to the more well-tested order.

Tested on aarch64-linux-gnu and x86_64-linux-gnu.  OK to install?

Richard

2020-02-19  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
PR rtl-optimization/PR92989
* lra-lives.c (process_bb_lives): Restore the original order
of the bb liveness update.  Call make_hard_regno_dead for each
register clobbered at the start of an EH receiver.

4 years agoDo not propagate self-dependent value (PR ipa/93763) (ChangeLog)
Jeff Law [Fri, 21 Feb 2020 20:24:27 +0000 (13:24 -0700)]
Do not propagate self-dependent value (PR ipa/93763) (ChangeLog)

            PR ipa/93763
            * ipa-cp.c (self_recursively_generated_p): Mark self-dependent value as
            self-recursively generated.

4 years agoDo not propagate self-dependent value (PR ipa/93763)
Feng Xue [Mon, 17 Feb 2020 09:07:04 +0000 (17:07 +0800)]
Do not propagate self-dependent value (PR ipa/93763)

        PR ipa/93763
        * ipa-cp.c (self_recursively_generated_p): Mark self-dependent value as
        self-recursively generated.

4 years agoDarwin: Fix wrong quoting on an error string (PR93860).
Iain Sandoe [Fri, 21 Feb 2020 20:12:39 +0000 (20:12 +0000)]
Darwin: Fix wrong quoting on an error string (PR93860).

The quotes should surround all of the literal content from the pragma
that has incorrect usage.

2020-02-21  Iain Sandoe  <iain@sandoe.co.uk>

PR target/93860
* config/darwin-c.c (pop_field_alignment): Adjust quoting of
error string.

4 years agoPR c++/93753 - ICE on a flexible array followed by a member in an anonymous struct...
Martin Sebor [Fri, 21 Feb 2020 17:45:56 +0000 (10:45 -0700)]
PR c++/93753 - ICE on a flexible array followed by a member in an anonymous struct with an initializer

gcc/cp/ChangeLog:

PR gcov-profile/93753
* class.c (check_flexarrays): Tighten up a test for potential members
of anonymous structs or unions.

gcc/testsuite/ChangeLog:

PR gcov-profile/93753
* g++.dg/ext/flexary36.C: New test.
* g++.dg/lto/pr93166_0.C: Make struct with flexarray valid.

4 years agolibstdc++: Define <=> for tuple, optional and variant
Jonathan Wakely [Fri, 21 Feb 2020 12:02:15 +0000 (12:02 +0000)]
libstdc++: Define <=> for tuple, optional and variant

Another piece of P1614R2.

* include/std/optional (operator<=>(optional<T>, optional<U>))
(operator<=>(optional<T>, nullopt), operator<=>(optional<T>, U)):
Define for C++20.
* include/std/tuple (__tuple_cmp): New helper function for <=>.
(operator<=>(tuple<T...>, tuple<U>...)): Define for C++20.
* include/std/variant (operator<=>(variant<T...>, variant<T...>))
(operator<=>(monostate, monostate)): Define for C++20.
* testsuite/20_util/optional/relops/three_way.cc: New test.
* testsuite/20_util/tuple/comparison_operators/three_way.cc: New test.
* testsuite/20_util/variant/89851.cc: Move to ...
* testsuite/20_util/variant/relops/89851.cc: ... here.
* testsuite/20_util/variant/90008.cc: Move to ...
* testsuite/20_util/variant/relops/90008.cc: ... here.
* testsuite/20_util/variant/relops/three_way.cc: New test.

4 years ago[PATCH, GCC/ARM] Add MVE target check to sourcebuild.texi
Mihail Ionescu [Fri, 21 Feb 2020 15:26:06 +0000 (15:26 +0000)]
[PATCH, GCC/ARM] Add MVE target check to sourcebuild.texi

Follow up to: https://gcc.gnu.org/ml/gcc-patches/2020-02/msg01109.html

Committed as obvious.

gcc/ChangeLog:

2020-02-21  Mihail Ionescu  <mihail.ionescu@arm.com>

* doc/sourcebuild.texi (arm_v8_1m_mve_ok):
Document new target supports option.

4 years agoarm: ACLE I8MM multiply-accumulate
Dennis Zhang [Fri, 21 Feb 2020 15:36:13 +0000 (15:36 +0000)]
arm: ACLE I8MM multiply-accumulate

This patch adds intrinsics for matrix multiply-accumulate instructions
including vmmlaq_s32, vmmlaq_u32, and vusmmlaq_s32.

gcc/ChangeLog:

2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>

* config/arm/arm_neon.h (vmmlaq_s32, vmmlaq_u32, vusmmlaq_s32): New.
* config/arm/arm_neon_builtins.def (smmla, ummla, usmmla): New.
* config/arm/iterators.md (MATMUL): New iterator.
(sup): Add UNSPEC_MATMUL_S, UNSPEC_MATMUL_U, and UNSPEC_MATMUL_US.
(mmla_sfx): New attribute.
* config/arm/neon.md (neon_<sup>mmlav16qi): New.
* config/arm/unspecs.md (UNSPEC_MATMUL_S, UNSPEC_MATMUL_U): New.
(UNSPEC_MATMUL_US): New.

gcc/testsuite/ChangeLog:

2020-02-21  Dennis Zhang  <dennis.zhang@arm.com>

* gcc.target/arm/simd/vmmla_1.c: New test.

4 years agotestsuite: Add -fcommon to gcc.target/i386/pr69052.c
Uros Bizjak [Fri, 21 Feb 2020 15:23:30 +0000 (16:23 +0100)]
testsuite: Add -fcommon to gcc.target/i386/pr69052.c

This testcase is susceptible to memory location details and start to fail
with default to -fno-common.  Use -fcommon to set expected testing conditions.

* gcc.target/i386/pr69052.c: Require target ia32.
(dg-options): Add -fcommon and remove -pie.

4 years ago[PATCH, GCC/ARM] Fix MVE scalar shift tests
Mihail Ionescu [Fri, 21 Feb 2020 15:21:23 +0000 (15:21 +0000)]
[PATCH, GCC/ARM] Fix MVE scalar shift tests

*** gcc/ChangeLog ***

2020-02-21  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>

* config/arm/arm.md: Prevent scalar shifts from being
used when big endian is enabled.

*** gcc/testsuite/ChangeLog ***

2020-02-21  Mihail-Calin Ionescu  <mihail.ionescu@arm.com>

* gcc.target/arm/armv8_1m-shift-imm-1.c: Add MVE target checks.
* gcc.target/arm/armv8_1m-shift-reg-1.c: Likewise.
* lib/target-supports.exp
(check_effective_target_arm_v8_1m_mve_ok_nocache): New.
(check_effective_target_arm_v8_1m_mve_ok): New.
(add_options_for_v8_1m_mve): New.

4 years agotestsuite: Require vect_mutiple_sizes for scan-tree-dump in vect-epilogues.c
Uros Bizjak [Fri, 21 Feb 2020 15:12:57 +0000 (16:12 +0100)]
testsuite: Require vect_mutiple_sizes for scan-tree-dump in vect-epilogues.c

Default testsuite flags do not enable V8QI (MMX) vector mode for
32bit x86 targets.  Require vect_multiple_sizes effective target in
scan-tree-dump to avoid "LOOP EPILOGUE VECTORIZED" failure.

* gcc.dg/vect/vect-epilogues.c (scan-tree-dump): Require
vect_mutiple_sizes effective target.

4 years agoAdapt libgomp acc_get_property.f90 test
Frederik Harwath [Fri, 21 Feb 2020 14:26:02 +0000 (15:26 +0100)]
Adapt libgomp acc_get_property.f90 test

The commit r10-6721-g8d1a1cb1b816381bf60cb1211c93b8eba1fe1472 has changed
the name of the type that is used for the return value of the Fortran
acc_get_property function without adapting the test acc_get_property.f90.

2020-02-21  Frederik Harwath  <frederik@codesourcery.com>

* testsuite/libgomp.oacc-fortran/acc_get_property.f90: Adapt to
changes from 2020-02-19, i.e. use integer(c_size_t) instead of
integer(acc_device_property) for the type of the return value of
acc_get_property.

4 years agotree-optimization: fix access path oracle on mismatched array refs [PR93586]
Jan Hubicka [Fri, 21 Feb 2020 14:36:00 +0000 (15:36 +0100)]
tree-optimization: fix access path oracle on mismatched array refs [PR93586]

nonoverlapping_array_refs_p is not supposed to give meaningful results when
bases of ref1 and ref2 are not same or completely disjoint and here it is
called on c[0][j_2][0] and c[0][1] so bases in sence of this functions are
"c[0][j_2]" and "c[0]" which do partially overlap.  nonoverlapping_array_refs
however walks pair of array references and in this case it misses to note the
fact that if it walked across first mismatched pair it is no longer safe to
compare rest.

The reason why it continues matching is because it hopes it will
eventually get pair of COMPONENT_REFs from types of same size and use
TBAA to conclude that their addresses must be either same or completely
disjoint.

This patch makes the loop to terminate early but popping all the
remaining pairs so walking can continue.  We could re-synchronize on
arrays of same size with TBAA but this is bit fishy (because we try to
support some sort of partial array overlaps) and hard to implement
(because of zero sized arrays and VLAs) so I think it is not worth the
effort.

In addition I notied that the function is not !flag_strict_aliasing safe
and added early exits on places we set seen_unmatched_ref_p since later
we do not check that in:

       /* If we skipped array refs on type of different sizes, we can
   no longer be sure that there are not partial overlaps.  */
       if (seen_unmatched_ref_p
    && !operand_equal_p (TYPE_SIZE (type1), TYPE_SIZE (type2), 0))
  {
    ++alias_stats
      .nonoverlapping_refs_since_match_p_may_alias;
}

   PR tree-optimization/93586
* tree-ssa-alias.c (nonoverlapping_array_refs_p): Finish array walk
after mismatched array refs; do not sure type size information to
recover from unmatched referneces with !flag_strict_aliasing_p.

* gcc.dg/torture/pr93586.c: New testcase.

4 years agoamdgcn: Use correct offset mode for gather/scatter
Andrew Stubbs [Thu, 20 Feb 2020 14:44:04 +0000 (14:44 +0000)]
amdgcn: Use correct offset mode for gather/scatter

The scatter/gather pattern names changed for GCC 10, but I hadn't noticed.
This switches the patterns to the new offset mode scheme.

2020-02-21  Andrew Stubbs  <ams@codesourcery.com>

gcc/
* config/gcn/gcn-valu.md (gather_load<mode>): Rename to ...
(gather_load<mode>v64si): ... this and set operand 2 to V64SI.
(scatter_store<mode>): Rename to ...
(scatter_store<mode>v64si): ... this and set operand 1 to V64SI.
(scatter<mode>_exec): Delete. Move contents ...
(mask_scatter_store<mode>): ... here, and rename that to ...
(mask_gather_load<mode>v64si): ... this. Set operand 2 to V64SI.
Remove mode conversion.
(mask_gather_load<mode>): Rename to ...
(mask_scatter_store<mode>v64si): ... this. Set operand 1 to V64SI.
Remove mode conversion.
* config/gcn/gcn.c (gcn_expand_scaled_offsets): Remove mode conversion.