platform/upstream/gcc.git
4 years agolibstdc++: Remove accidental -std=gnu++17 from test
Jonathan Wakely [Fri, 31 Jul 2020 17:02:10 +0000 (18:02 +0100)]
libstdc++: Remove accidental -std=gnu++17 from test

This was probably copied from a std::filesystem test and the -std option
wasn't removed.

libstdc++-v3/ChangeLog:

* testsuite/experimental/filesystem/filesystem_error/cons.cc:
Remove -std=gnu++17 option.

4 years agolibstdc++: Adjust tests that give different results in C++20
Jonathan Wakely [Fri, 31 Jul 2020 16:51:00 +0000 (17:51 +0100)]
libstdc++: Adjust tests that give different results in C++20

libstdc++-v3/ChangeLog:

* testsuite/20_util/is_aggregate/value.cc: Adjust for changes to
definition of aggregates in C++20.
* testsuite/20_util/optional/requirements.cc: Adjust for
defaulted comparisons in C++20.

4 years agolibstdc++: Add -Wno-deprecated for tests that warn in C++20
Jonathan Wakely [Fri, 31 Jul 2020 16:51:00 +0000 (17:51 +0100)]
libstdc++: Add -Wno-deprecated for tests that warn in C++20

libstdc++-v3/ChangeLog:

* testsuite/20_util/tuple/78939.cc: Suppress warnings about
deprecation of volatile-qualified structured bindings in C++20.
* testsuite/20_util/variable_templates_for_traits.cc: Likewise
for deprecation of is_pod in C++20

4 years agolibstdc++: Remove duplicate dg-do directive
Jonathan Wakely [Fri, 31 Jul 2020 16:51:00 +0000 (17:51 +0100)]
libstdc++: Remove duplicate dg-do directive

Also add an effective target to clarify it should only run for C++17 and
later.

libstdc++-v3/ChangeLog:

* testsuite/20_util/time_point_cast/rounding.cc: Remove
duplicate dg-do directive and add c++17 effective target.

4 years agod: Split up the grouped compilable and runnable tests.
Iain Buclaw [Tue, 21 Jul 2020 16:40:36 +0000 (18:40 +0200)]
d: Split up the grouped compilable and runnable tests.

The majority of tests in runnable are really compilable/ICE tests, and
have have dg-do adjusted where necessary.  Tests that had a dependency
on Phobos have also been reproduced and reduced with all imports
stripped from the test.

The end result is a collection of tests that only check the compiler bug
that was being fixed, rather than the library, and a reduction in time
spent running all tests.

gcc/testsuite/ChangeLog:

* gdc.dg/compilable.d: Removed.
* gdc.dg/gdc108.d: New test.
* gdc.dg/gdc115.d: New test.
* gdc.dg/gdc121.d: New test.
* gdc.dg/gdc122.d: New test.
* gdc.dg/gdc127.d: New test.
* gdc.dg/gdc131.d: New test.
* gdc.dg/gdc133.d: New test.
* gdc.dg/gdc141.d: New test.
* gdc.dg/gdc142.d: New test.
* gdc.dg/gdc15.d: New test.
* gdc.dg/gdc17.d: New test.
* gdc.dg/gdc170.d: New test.
* gdc.dg/gdc171.d: New test.
* gdc.dg/gdc179.d: New test.
* gdc.dg/gdc183.d: New test.
* gdc.dg/gdc186.d: New test.
* gdc.dg/gdc187.d: New test.
* gdc.dg/gdc19.d: New test.
* gdc.dg/gdc191.d: New test.
* gdc.dg/gdc194.d: New test.
* gdc.dg/gdc196.d: New test.
* gdc.dg/gdc198.d: New test.
* gdc.dg/gdc200.d: New test.
* gdc.dg/gdc204.d: New test.
* gdc.dg/gdc210.d: New test.
* gdc.dg/gdc212.d: New test.
* gdc.dg/gdc213.d: New test.
* gdc.dg/gdc218.d: New test.
* gdc.dg/gdc223.d: New test.
* gdc.dg/gdc231.d: New test.
* gdc.dg/gdc239.d: New test.
* gdc.dg/gdc24.d: New test.
* gdc.dg/gdc240.d: New test.
* gdc.dg/gdc241.d: New test.
* gdc.dg/gdc242a.d: New test.
* gdc.dg/gdc242b.d: New test.
* gdc.dg/gdc248.d: New test.
* gdc.dg/gdc250.d: New test.
* gdc.dg/gdc251.d: New test.
* gdc.dg/gdc253a.d: New test.
* gdc.dg/gdc253b.d: New test.
* gdc.dg/gdc255.d: New test.
* gdc.dg/gdc256.d: New test.
* gdc.dg/gdc261.d: New test.
* gdc.dg/gdc27.d: New test.
* gdc.dg/gdc273.d: New test.
* gdc.dg/gdc280.d: New test.
* gdc.dg/gdc284.d: New test.
* gdc.dg/gdc285.d: New test.
* gdc.dg/gdc286.d: New test.
* gdc.dg/gdc300.d: New test.
* gdc.dg/gdc309.d: New test.
* gdc.dg/gdc31.d: New test.
* gdc.dg/gdc35.d: New test.
* gdc.dg/gdc36.d: New test.
* gdc.dg/gdc37.d: New test.
* gdc.dg/gdc4.d: New test.
* gdc.dg/gdc43.d: New test.
* gdc.dg/gdc47.d: New test.
* gdc.dg/gdc51.d: New test.
* gdc.dg/gdc57.d: New test.
* gdc.dg/gdc66.d: New test.
* gdc.dg/gdc67.d: New test.
* gdc.dg/gdc71.d: New test.
* gdc.dg/gdc77.d: New test.
* gdc.dg/imports/gdc239.d: Remove phobos dependency.
* gdc.dg/imports/gdc241a.d: Updated imports.
* gdc.dg/imports/gdc241b.d: Likewise.
* gdc.dg/imports/gdc251a.d: Likewise.
* gdc.dg/imports/gdc253.d: Rename to...
* gdc.dg/imports/gdc253a.d: ...this.
* gdc.dg/imports/gdc253b.d: New.
* gdc.dg/imports/gdc36.d: New.
* gdc.dg/imports/runnable.d: Removed.
* gdc.dg/link.d: Removed.
* gdc.dg/runnable.d: Removed.
* gdc.dg/runnable2.d: Removed.
* gdc.dg/simd.d: Remove phobos dependency.

4 years agod: Fix regression, all 32-bit execution tests FAIL: internal error printing module...
Iain Buclaw [Fri, 31 Jul 2020 14:03:17 +0000 (16:03 +0200)]
d: Fix regression, all 32-bit execution tests FAIL: internal error printing module cycle

For 32-bit btr(), BIT_NOT_EXPR was being generated twice, something that
was not seen with the 64-bit variant.  Removed the second call to fix
the generated code.

gcc/d/ChangeLog:

PR d/96393
* intrinsics.cc (expand_intrinsic_bt): Don't generate BIT_NOT_EXPR for
btr32 intrinsic.

4 years agoSet and test no-warning bit to avoid -Wnonnull for synthesized expressions.
Martin Sebor [Fri, 31 Jul 2020 16:27:33 +0000 (10:27 -0600)]
Set and test no-warning bit to avoid -Wnonnull for synthesized expressions.

Resolves:
PR c++/96003 spurious -Wnonnull calling a member on the result of static_cast

gcc/c-family/ChangeLog:

PR c++/96003
* c-common.c (check_function_arguments_recurse): Return early when
no-warning bit is set.

gcc/cp/ChangeLog:

PR c++/96003
* class.c (build_base_path): Set no-warning bit on the synthesized
conditional expression in static_cast.

gcc/testsuite/ChangeLog:

PR c++/96003
* g++.dg/warn/Wnonnull7.C: New test.

4 years agolibstdc++: Fix use of newlocale in std:::from_chars
Jonathan Wakely [Fri, 31 Jul 2020 13:36:56 +0000 (14:36 +0100)]
libstdc++: Fix use of newlocale in std:::from_chars

libstdc++-v3/ChangeLog:

* src/c++17/floating_from_chars.cc (from_chars_impl): Use
LC_ALL_MASK not LC_ALL.

4 years agodebug/96383 - emit debug info for used external functions
Richard Biener [Thu, 30 Jul 2020 09:46:43 +0000 (11:46 +0200)]
debug/96383 - emit debug info for used external functions

This makes sure to emit full declaration DIEs including
formal parameters for used external functions.  This helps
debugging when debug information of the external entity is
not available and also helps external tools cross-checking
ABI compatibility which was the bug reporters use case.

For cc1 this affects debug information size as follows:

     VM SIZE                     FILE SIZE
 ++++++++++++++ GROWING       ++++++++++++++
  [ = ]       0 .debug_info   +1.63Mi  +1.3%
  [ = ]       0 .debug_str     +263Ki  +3.4%
  [ = ]       0 .debug_abbrev  +101Ki  +4.9%
  [ = ]       0 .debug_line   +5.71Ki  +0.0%
   +44%     +16 [Unmapped]        +48  +1.2%

 -------------- SHRINKING     --------------
  [ = ]       0 .debug_loc       -213  -0.0%
  -0.0%     -48 .text             -48  -0.0%
  [ = ]       0 .debug_ranges     -16  -0.0%

  -0.0%     -32 TOTAL         +1.99Mi  +0.6%

and DWARF compression via DWZ can only shave off minor bits
here.

Previously we emitted no DIEs for external functions at all
unless they were referenced via DW_TAG_GNU_call_site which
for some GCC revs caused a regular DIE to appear and since
GCC 4.9 only a stub without formal parameters.  This means
at -O0 we did not emit any DIE for external functions
but with optimization we emitted stubs.

2020-07-30  Richard Biener  <rguenther@suse.de>

PR debug/96383
* langhooks-def.h (lhd_finalize_early_debug): Declare.
(LANG_HOOKS_FINALIZE_EARLY_DEBUG): Define.
(LANG_HOOKS_INITIALIZER): Amend.
* langhooks.c: Include cgraph.h and debug.h.
(lhd_finalize_early_debug): Default implementation from
former code in finalize_compilation_unit.
* langhooks.h (lang_hooks::finalize_early_debug): Add.
* cgraphunit.c (symbol_table::finalize_compilation_unit):
Call the finalize_early_debug langhook.

gcc/c-family/
* c-common.h (c_common_finalize_early_debug): Declare.
* c-common.c: Include debug.h.
(c_common_finalize_early_debug): finalize_early_debug langhook
implementation generating debug for extern declarations.

gcc/c/
* c-objc-common.h (LANG_HOOKS_FINALIZE_EARLY_DEBUG):
Define to c_common_finalize_early_debug.

gcc/cp/
* cp-objcp-common.h (LANG_HOOKS_FINALIZE_EARLY_DEBUG):
Define to c_common_finalize_early_debug.

gcc/testsuite/
* gcc.dg/debug/dwarf2/pr96383-1.c: New testcase.
* gcc.dg/debug/dwarf2/pr96383-2.c: Likewise.

libstdc++-v3/
* testsuite/20_util/assume_aligned/3.cc: Use -g0.

4 years agoAmend match.pd syntax with force-simplified results
Richard Biener [Fri, 31 Jul 2020 12:24:26 +0000 (14:24 +0200)]
Amend match.pd syntax with force-simplified results

This adds a ! marker to result expressions that should simplify
(and if not fail the simplification).  This can for example be
used like

(simplify
  (plus (vec_cond:s @0 @1 @2) @3)
  (vec_cond @0 (plus! @1 @3) (plus! @2 @3)))

to make the simplification only apply in case both plus operations
in the result end up simplified to a simple operand.

2020-07-31  Richard Biener  <rguenther@suse.de>

* genmatch.c (expr::force_leaf): Add and initialize.
(expr::gen_transform): Honor force_leaf by passing
NULL as sequence argument to maybe_push_res_to_seq.
(parser::parse_expr): Allow ! marker on result expression
operations.
* doc/match-and-simplify.texi: Amend.

4 years agovect: Don't consider branch costs if no peeled iterations
Kewen Lin [Fri, 31 Jul 2020 12:49:39 +0000 (07:49 -0500)]
vect: Don't consider branch costs if no peeled iterations

Currently vectorizer cost modeling counts branch taken costs for
prologue and epilogue if the number of iterations is unknown.
But it isn't sensible if there are no peeled iterations.

This patch is to guard them under peel_iters_prologue > 0 or
peel_iters_epilogue > 0.

Bootstrapped/regtested on powerpc64le-linux-gnu and aarch64-linux-gnu.

gcc/ChangeLog:

* tree-vect-loop.c (vect_get_known_peeling_cost): Don't consider branch
taken costs for prologue and epilogue if they don't exist.
(vect_estimate_min_profitable_iters): Likewise.

gcc/testsuite/ChangeLog:

* gcc.target/aarch64/sve/cost_model_2.c: Adjust due to cost model
change.

4 years agolibstdc++: Add PR number to ChangeLog entry
Jonathan Wakely [Fri, 31 Jul 2020 13:15:43 +0000 (14:15 +0100)]
libstdc++: Add PR number to ChangeLog entry

4 years agoDo not allocate huge array in output_in_order.
Martin Liska [Thu, 30 Jul 2020 07:24:40 +0000 (09:24 +0200)]
Do not allocate huge array in output_in_order.

We noticed that when analyzing LTRANS memory peak that happens right
after the start:
https://gist.github.com/marxin/223890df4d8d8e490b6b2918b77dacad

In case of chrome, we have symtab->order == 200M, so we allocate
16B * 200M = 3.2GB.

gcc/ChangeLog:

* cgraph.h: Remove leading empty lines.
* cgraphunit.c (enum cgraph_order_sort_kind): Remove
ORDER_UNDEFINED.
(struct cgraph_order_sort): Add constructors.
(cgraph_order_sort::process): New.
(cgraph_order_cmp): New.
(output_in_order): Simplify and push nodes to vector.

4 years agomiddle-end/96369 - fix missed short-circuiting during range folding
Richard Biener [Fri, 31 Jul 2020 06:41:56 +0000 (08:41 +0200)]
middle-end/96369 - fix missed short-circuiting during range folding

This makes the special case of constant evaluated LHS for a
short-circuiting or/and explicit rather than doing range
merging and eventually exposing a side-effect that shouldn't be
evaluated.

2020-07-31  Richard Biener  <rguenther@suse.de>

PR middle-end/96369
* fold-const.c (fold_range_test): Special-case constant
LHS for short-circuiting operations.

* c-c++-common/pr96369.c: New testcase.

4 years agolibgcov: support overloaded malloc
Martin Liska [Tue, 2 Jun 2020 11:31:48 +0000 (13:31 +0200)]
libgcov: support overloaded malloc

gcc/ChangeLog:

* gcov-io.h (GCOV_PREALLOCATED_KVP): New.

libgcc/ChangeLog:

* libgcov-driver.c: Add __gcov_kvp_pool
and __gcov_kvp_pool_index variables.
* libgcov.h (allocate_gcov_kvp): New.
(gcov_topn_add_value): Use it.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-prof/indir-call-prof-malloc.c: New test.

4 years agoAArch64: Add attributes according to flags in built-in functions [PR94442]
xiezhiheng [Fri, 31 Jul 2020 08:52:25 +0000 (09:52 +0100)]
AArch64: Add attributes according to flags in built-in functions [PR94442]

2020-07-31  Zhiheng Xie  <xiezhiheng@huawei.com>

gcc/ChangeLog:

* config/aarch64/aarch64-builtins.c (aarch64_general_add_builtin):
Add new argument ATTRS.
(aarch64_call_properties): New function.
(aarch64_modifies_global_state_p): Likewise.
(aarch64_reads_global_state_p): Likewise.
(aarch64_could_trap_p): Likewise.
(aarch64_add_attribute): Likewise.
(aarch64_get_attributes): Likewise.
(aarch64_init_simd_builtins): Add attributes for each built-in function.

4 years agoImprove var-tracking dataflow iteration order
Richard Biener [Fri, 24 Jul 2020 11:44:09 +0000 (13:44 +0200)]
Improve var-tracking dataflow iteration order

This builds upon the rev_post_order_and_mark_dfs_back_seme improvements
and makes vt_find_locations iterate over the dataflow problems for
each toplevel SCC separately, improving memory locality and avoiding
to process nodes after the SCC before the SCC itself stabilized.

On the asan_interceptors.cc testcase this for example reduces the
number of visited blocks from 3751 to 2867.  For stage3-gcc
this reduces the number of visited blocks by ~4%.

2020-07-28  Richard Biener  <rguenther@suse.de>

PR debug/78288
* var-tracking.c (vt_find_locations): Use
rev_post_order_and_mark_dfs_back_seme and separately iterate
over toplevel SCCs.

4 years agoCompute RPO with adjacent SCC members, expose toplevel SCC extents
Richard Biener [Mon, 20 Jul 2020 11:38:16 +0000 (13:38 +0200)]
Compute RPO with adjacent SCC members, expose toplevel SCC extents

This produces a more optimal RPO order for iteration processing
by making sure that SCC exits are processed before SCC members
themselves..  This avoids iterating blocks unrelated to the current
iteration for RPO VN and has the chance to improve code-generation
for the non-iterative mode of RPO VN.  The patch also exposes toplevel
SCCs and gets rid of the ad-hoc max_rpo computation in RPO VN.

For simplicity it also removes the odd reverse ordering of the RPO
array returned from rev_post_order_and_mark_dfs_back_seme.

Overall reduction in the number of visited blocks isn't spectacular
for bootstrap (~2.5%) but single cases see up to a 10% reduction.

The same function can be used to optimize var-tracking iteration order
as seen in the followup.

2020-07-28  Richard Biener  <rguenther@suse.de>

* cfganal.h (rev_post_order_and_mark_dfs_back_seme): Adjust
prototype.
* cfganal.c (rpoamdbs_bb_data): New struct with pre BB data.
(tag_header): New helper.
(cmp_edge_dest_pre): Likewise.
(rev_post_order_and_mark_dfs_back_seme): Compute SCCs,
find SCC exits and perform a DFS walk with extra edges to
compute a RPO with adjacent SCC members when requesting an
iteration optimized order and populate the toplevel SCC array.
* tree-ssa-sccvn.c (do_rpo_vn): Remove ad-hoc computation
of max_rpo and fill it in from SCC extent info instead.

* gcc.dg/torture/20200727-0.c: New testcase.

4 years agoc++: decl_constant_value and unsharing [PR96197]
Patrick Palka [Fri, 31 Jul 2020 02:21:41 +0000 (22:21 -0400)]
c++: decl_constant_value and unsharing [PR96197]

In the testcase from the PR we're seeing excessive memory use (> 5GB)
during constexpr evaluation, almost all of which is due to the call to
decl_constant_value in the VAR_DECL/CONST_DECL branch of
cxx_eval_constant_expression.  We reach here every time we evaluate an
ARRAY_REF of a constexpr VAR_DECL, and from there decl_constant_value
makes an unshared copy of the VAR_DECL's initializer.  But unsharing
here is unnecessary because callers of cxx_eval_constant_expression
already unshare its result when necessary.

To fix this excessive unsharing, this patch adds a new defaulted
parameter unshare_p to decl_really_constant_value and
decl_constant_value so that callers can control whether to unshare.

As a simplification, we can also move the call to unshare_expr in
constant_value_1 outside of the loop, since doing unshare_expr on a
DECL_P is a no-op.

Now that we no longer unshare the result of decl_constant_value and
decl_really_constant_value from cxx_eval_constant_expression, memory use
during constexpr evaluation for the testcase from the PR falls from ~5GB
to 15MB according to -ftime-report.

gcc/cp/ChangeLog:

PR c++/96197
* constexpr.c (cxx_eval_constant_expression) <case CONST_DECL>:
Pass false to decl_constant_value and decl_really_constant_value
so that they don't unshare their result.
* cp-tree.h (decl_constant_value): New declaration with an added
bool parameter.
(decl_really_constant_value): Add bool parameter defaulting to
true to existing declaration.
* init.c (constant_value_1): Add bool parameter which controls
whether to unshare the initializer before returning.  Call
unshare_expr at most once.
(scalar_constant_value): Pass true to constant_value_1's new
bool parameter.
(decl_really_constant_value): Add bool parameter and forward it
to constant_value_1.
(decl_constant_value): Likewise, but instead define a new
overload with an added bool parameter.

gcc/testsuite/ChangeLog:

PR c++/96197
* g++.dg/cpp1y/constexpr-array8.C: New test.

4 years agoDaily bump.
GCC Administrator [Fri, 31 Jul 2020 00:16:26 +0000 (00:16 +0000)]
Daily bump.

4 years agod: Fix associative array literals that don't have alignment holes filled
Iain Buclaw [Sat, 18 Jul 2020 15:14:54 +0000 (17:14 +0200)]
d: Fix associative array literals that don't have alignment holes filled

Associative array literal keys with alignment holes are now filled using
memset() prior to usage, with LTR evaluation of side-effects enforced.

gcc/d/ChangeLog:

PR d/96152
* d-codegen.cc (build_array_from_exprs): New function.
* d-tree.h (build_array_from_exprs): Declare.
* expr.cc (ExprVisitor::visit (AssocArrayLiteralExp *)): Use
build_array_from_exprs to generate key and value arrays.

gcc/testsuite/ChangeLog:

PR d/96152
* gdc.dg/pr96152.d: New test.

4 years agod: Add -Wvarargs warning flag to the D front-end
Iain Buclaw [Thu, 16 Jul 2020 16:56:18 +0000 (18:56 +0200)]
d: Add -Wvarargs warning flag to the D front-end

The D front-end has C-style variadic functions and va_start/va_arg, so
it is right to also have warnings for inproper use.

gcc/d/ChangeLog:

PR d/96154
* gdc.texi (Warnings): Document -Wvarargs.
* lang.opt: Add -Wvarargs

gcc/testsuite/ChangeLog:

PR d/96154
* gdc.dg/pr96154a.d: New test.
* gdc.dg/pr96154b.d: New test.

4 years agod: Fix ICE in expand_intrinsic_vaarg
Iain Buclaw [Thu, 16 Jul 2020 16:34:18 +0000 (18:34 +0200)]
d: Fix ICE in expand_intrinsic_vaarg

Both intrinsics did not handle the case where the va_list object comes
from a ref parameter.

gcc/d/ChangeLog:

PR d/96140
* intrinsics.cc (expand_intrinsic_vaarg): Handle ref parameters as
arguments to va_arg().
(expand_intrinsic_vastart): Handle ref parameters as arguments to
va_start().

gcc/testsuite/ChangeLog:

PR d/96140
* gdc.dg/pr96140.d: New test.

4 years agolibstdc++: Make COW string use allocator_traits for nested types
Jonathan Wakely [Thu, 30 Jul 2020 19:58:09 +0000 (20:58 +0100)]
libstdc++: Make COW string use allocator_traits for nested types

When compiled as C++20 the COW std::string fails due to assuming that
the allocator always defines size_type and difference_type. That has
been incorrect since C++11, but we got away with it for specializations
using std::allocator until those members were removed in C++20.

libstdc++-v3/ChangeLog:

* include/bits/basic_string.h (size_type, difference_type):
Use allocator_traits to obtain the allocator's size_type and
difference_type.

4 years agolibstdc++: Check _GLIBCXX_USE_C99_STDLIB for strtof and strtold
Jonathan Wakely [Thu, 30 Jul 2020 19:55:56 +0000 (20:55 +0100)]
libstdc++: Check _GLIBCXX_USE_C99_STDLIB for strtof and strtold

On broken systems we only have strtod, not strtof and strtold. Just use
strtod for all types, even though that will produce incorrect results in
some cases.

Similarly, if _GLIBCXX_USE_C99_MATH is not defined then std::isinf won't
be declared. Just refer to it unqualified, which should find the C
library's isinf macro if that hasn't been #undef'd by <cmath>.

libstdc++-v3/ChangeLog:

* src/c++17/floating_from_chars.cc (from_chars_impl): Use
isinf unqualified.
[!_GLIBCXX_USE_C99_STDLIB]: Use strtod for float and long
double.

4 years agolibstdc++: Fix tests using wrong allocator type
Jonathan Wakely [Thu, 30 Jul 2020 17:41:47 +0000 (18:41 +0100)]
libstdc++: Fix tests using wrong allocator type

libstdc++-v3/ChangeLog:

* testsuite/23_containers/unordered_multiset/cons/noexcept_default_construct.cc:
Use allocator with the correct value type.
* testsuite/23_containers/unordered_set/cons/noexcept_default_construct.cc:
Likewise.

4 years ago[PATCH] RS6000 Add testlsbb by Byte operations
Will Schmidt [Thu, 21 May 2020 20:21:34 +0000 (15:21 -0500)]
[PATCH] RS6000 Add testlsbb by Byte operations

Add support for new instructions to test LSB by Byte.

2020-07-29  Will Schmidt  <will_schmidt@vnet.ibm.com>

gcc/ChangeLog:

* config/rs6000/altivec.h (vec_test_lsbb_all_ones): New define.
(vec_test_lsbb_all_zeros): New define.
* config/rs6000/rs6000-builtin.def (BU_P10_VSX_1): New built-in
handling macro.
(XVTLSBB_ZEROS, XVTLSBB_ONES): New builtin defines.
(xvtlsbb_all_zeros, xvtlsbb_all_ones): New builtin overloads.
* config/rs6000/rs6000-call.c (P10_BUILTIN_VEC_XVTLSBB_ZEROS,
P10_BUILTIN_VEC_XVTLSBB_ONES): New altivec_builtin_types entries.
* config/rs6000/rs6000.md (UNSPEC_XVTLSBB):  New unspec.
* config/rs6000/vsx.md (*xvtlsbb_internal): New instruction define.
(xvtlsbbo, xvtlsbbz): New instruction expands.

gcc/testsuite/ChangeLog:

* gcc.target/powerpc/lsbb-runnable.c: New test.
* gcc.target/powerpc/lsbb.c: New test.

4 years agoRISC-V: Add support for TLS stack protector canary access
Cooper Qu [Mon, 13 Jul 2020 08:15:08 +0000 (16:15 +0800)]
RISC-V: Add support for TLS stack protector canary access

gcc/
* config/riscv/riscv-opts.h (stack_protector_guard): New enum.
* config/riscv/riscv.c (riscv_option_override): Handle
the new options.
* config/riscv/riscv.md (stack_protect_set): New pattern to handle
flexible stack protector guard settings.
(stack_protect_set_<mode>): Ditto.
(stack_protect_test): Ditto.
(stack_protect_test_<mode>): Ditto.
* config/riscv/riscv.opt (mstack-protector-guard=,
mstack-protector-guard-reg=, mstack-protector-guard-offset=): New
options.
* doc/invoke.texi (Option Summary) [RISC-V Options]:
Add -mstack-protector-guard=, -mstack-protector-guard-reg=, and
-mstack-protector-guard-offset=.
(RISC-V Options): Ditto.

Signed-off-by: cooper <cooper.qu@linux.alibaba.com>
Signed-off-by: Guo Ren <guoren@linux.alibaba.com>
4 years agod: Inline bounds checking for simple array assignments.
Iain Buclaw [Sun, 19 Jul 2020 13:18:08 +0000 (15:18 +0200)]
d: Inline bounds checking for simple array assignments.

This optimizes the code generation of simple array assignments, inlining
the array bounds checking code so there is no reliance on the library
routine _d_arraycopy(), which also deals with postblit and copy
constructors for non-trivial arrays.

gcc/d/ChangeLog:

* expr.cc (ExprVisitor::visit (AssignExp *)): Inline bounds checking
for simple array assignments.

gcc/testsuite/ChangeLog:

* gdc.dg/array1.d: New test.

4 years agod: Refactor use of built-in memcmp/memcpy/memset into helper functions.
Iain Buclaw [Fri, 17 Jul 2020 15:20:02 +0000 (17:20 +0200)]
d: Refactor use of built-in memcmp/memcpy/memset into helper functions.

Generating calls to memset, memcpy, and memcmp is frequent enough that
it becomes beneficial to put them into their own routine.  All parts of
the front-end have been updated to call the new helper functions instead
of doing it themselves.

gcc/d/ChangeLog:

* d-codegen.cc (build_memcmp_call): New function.
(build_memcpy_call): New function.
(build_memset_call): New function.
(build_float_identity): Call build_memcmp_call.
(lower_struct_comparison): Likewise.
(build_struct_comparison): Likewise.
* d-tree.h (build_memcmp_call): Declare.
(build_memcpy_call): Declare.
(build_memset_call): Declare.
* expr.cc (ExprVisitor::visit (EqualExp *)): Call build_memcmp_call.
(ExprVisitor::visit (AssignExp *)): Call build_memset_call.
(ExprVisitor::visit (ArrayLiteralExp *)): Call build_memcpy_call.
(ExprVisitor::visit (StructLiteralExp *)): Call build_memset_call.

4 years agod: Move private functions out of ExprVisitor into local statics
Iain Buclaw [Thu, 16 Jul 2020 12:02:24 +0000 (14:02 +0200)]
d: Move private functions out of ExprVisitor into local statics

None of these functions need access to the context pointer of the
visitor class, so have been made free standing.

gcc/d/ChangeLog:

* expr.cc (needs_postblit): Move out of ExprVisitor as a static
function.  Update all callers.
(needs_dtor): Likewise.
(lvalue_p): Likewise.
(binary_op): Likewise.
(binop_assignment): Likewise.

4 years agolibstdc++: Fix test for old string ABI
Jonathan Wakely [Thu, 30 Jul 2020 15:04:59 +0000 (16:04 +0100)]
libstdc++: Fix test for old string ABI

The COW string doesn't accept const_iterator arguments in insert and
related member functions. Pass a mutable iterator instead.

libstdc++-v3/ChangeLog:

* testsuite/20_util/from_chars/4.cc: Pass non-const iterator
to string::insert.

4 years agoMAINTAINERS: Add myself for write after approval
Joe Ramsay [Thu, 30 Jul 2020 11:12:36 +0000 (11:12 +0000)]
MAINTAINERS: Add myself for write after approval

2020-07-30  Joe Ramsay  <joe.ramsay@arm.com>

* MAINTAINERS (Write After Approval): Add myself.

4 years agoRequire CET support only for the final GCC build
H.J. Lu [Wed, 15 Jul 2020 13:16:01 +0000 (06:16 -0700)]
Require CET support only for the final GCC build

With --enable-cet, require CET support only for the final GCC build.
Don't enable CET without CET support for non-bootstrap build, in stage1
nor for build support.

config/

PR bootstrap/96202
* cet.m4 (GCC_CET_HOST_FLAGS): Don't enable CET without CET
support in stage1 nor for build support.

gcc/

PR bootstrap/96202
* configure: Regenerated.

libbacktrace/

PR bootstrap/96202
* configure: Regenerated.

libcc1/

PR bootstrap/96202
* configure: Regenerated.

libcpp/

PR bootstrap/96202
* configure: Regenerated.

libdecnumber/

PR bootstrap/96202
* configure: Regenerated.

libiberty/

PR bootstrap/96202
* configure: Regenerated.

lto-plugin/

PR bootstrap/96202
* configure: Regenerated.

4 years agolibstdc++: cv bool can't be an integer-like type (LWG 3467)
Jonathan Wakely [Thu, 30 Jul 2020 11:23:55 +0000 (12:23 +0100)]
libstdc++: cv bool can't be an integer-like type (LWG 3467)

libstdc++-v3/ChangeLog:

* include/bits/iterator_concepts.h (__detail::__cv_bool): New
helper concept.
(__detail::__integral_nonbool): Likewise.
(__detail::__is_integer_like): Use __integral_nonbool.
* testsuite/std/ranges/access/lwg3467.cc: New test.

4 years agolibstdc++: Add options for ieee float to relevant tests
Jonathan Wakely [Thu, 30 Jul 2020 11:23:55 +0000 (12:23 +0100)]
libstdc++: Add options for ieee float to relevant tests

libstdc++-v3/ChangeLog:

* testsuite/20_util/from_chars/4.cc: Use dg-add-options ieee.
* testsuite/29_atomics/atomic_float/1.cc: Likewise.

4 years agolibstdc++: Make testsuite usable with -fno-exceptions
Jonathan Wakely [Thu, 30 Jul 2020 11:23:54 +0000 (12:23 +0100)]
libstdc++: Make testsuite usable with -fno-exceptions

Previously it was not possible to add -fno-exceptions to the testsuite
flags, because some files that are compiled by the v3-build_support
procedure failed with exceptions disabled.

This adjusts those files to still compile without exceptions (with
degraded functionality in some cases).

The sole testcase that explicitly checks for -fno-exceptions has also
been adjusted to use the more robust exceptions_enabled effective-target
keyword from gcc/testsuite/lib/target-supports.exp.

libstdc++-v3/ChangeLog:

* testsuite/23_containers/vector/bool/72847.cc: Use the
exceptions_enabled effective-target keyword instead of
checking for an explicit -fno-exceptions option.
* testsuite/util/testsuite_abi.cc (examine_symbol): Remove
redundant try-catch.
* testsuite/util/testsuite_allocator.h [!__cpp_exceptions]:
Do not define check_allocate_max_size and memory_resource.
* testsuite/util/testsuite_containers.h: Replace comment with
#error if wrong standard dialect used.
* testsuite/util/testsuite_shared.cc: Likewise.

4 years agod: Implement core.bitop.rol() and core.bitop.ror() as intrinsics.
Iain Buclaw [Wed, 15 Jul 2020 21:00:11 +0000 (23:00 +0200)]
d: Implement core.bitop.rol() and core.bitop.ror() as intrinsics.

gcc/d/ChangeLog:

* intrinsics.cc (expand_intrinsic_rotate): Add function.
(maybe_expand_intrinsic): Handle rol and ror intrinsics.
* intrinsics.def (ROL): Add intrinsic.
(ROL_TIARG): Add intrinsic.
(ROR): Add intrinsic.
(ROR_TIARG): Add intrinsic.

gcc/testsuite/ChangeLog:

* gdc.dg/intrinsics.d: Add ror and rol tests.

4 years agod: Refactor matching and lowering of intrinsic functions.
Iain Buclaw [Tue, 14 Jul 2020 16:45:42 +0000 (18:45 +0200)]
d: Refactor matching and lowering of intrinsic functions.

Intrinsics are now matched explicitly, rather than through a common
alias where there are multiple overrides for a common intrinsic.
Where there is a corresponding DECL_FUNCTION_CODE, that is now stored in
the D intrinsic array.  All run-time std.math intrinsics have been
removed, as the library implementation already forwards to core.math.

gcc/d/ChangeLog:

* d-tree.h (DEF_D_INTRINSIC): Rename second argument from A to B.
* intrinsics.cc (intrinsic_decl): Add built_in field.
(DEF_D_INTRINSIC): Rename second argument from ALIAS to BUILTIN.
(maybe_set_intrinsic): Handle new intrinsic codes.
(expand_intrinsic_bt): Likewise.
(expand_intrinsic_checkedint): Likewise.
(expand_intrinsic_bswap): Remove.
(expand_intrinsic_sqrt): Remove.
(maybe_expand_intrinsic): Group together intrinsic cases that map
directly to gcc built-ins.
* intrinsics.def (DEF_D_BUILTIN): Rename second argument from A to B.
Update all callers to pass equivalent DECL_FUNCTION_CODE.
(DEF_CTFE_BUILTIN): Likewise.
(STD_COS): Remove intrinsic.
(STD_FABS): Remove intrinsic.
(STD_LDEXP): Remove intrinsic.
(STD_RINT): Remove intrinsic.
(STD_RNDTOL): Remove intrinsic.
(STD_SIN): Remove intrinsic.
(STD_SQRTF): Remove intrinsic.
(STD_SQRT): Remove intrinsic.
(STD_SQRTL): Remove intrinsic.

gcc/testsuite/ChangeLog:

* gdc.dg/intrinsics.d: New test.

4 years agotree-optimization/96370 - make reassoc expr rewrite more robust
Richard Biener [Thu, 30 Jul 2020 08:24:42 +0000 (10:24 +0200)]
tree-optimization/96370 - make reassoc expr rewrite more robust

In the face of the more complex tricks in reassoc with respect
to negate processing it can happen that the expression rewrite
is fooled to recurse on a leaf and pick up a bogus expression
code.  The following patch makes the expression rewrite more
robust in providing the expression code to it directly since
it is the same for all operations in a chain.

2020-07-30  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96370
* tree-ssa-reassoc.c (rewrite_expr_tree): Add operation
code parameter and use it instead of picking it up from
the stmt that is being rewritten.
(reassociate_bb): Pass down the operation code.

* gcc.dg/pr96370.c: New testcase.

4 years agonvptx: Provide vec_set<mode> and vec_extract<vmode><mode> patterns
Roger Sayle [Thu, 30 Jul 2020 08:23:38 +0000 (10:23 +0200)]
nvptx: Provide vec_set<mode> and vec_extract<vmode><mode> patterns

This patch provides standard vec_extract and vec_set patterns to the
nvptx backend, to extract an element from a PTX vector and set an
element of a PTX vector respectively.  PTX vectors (I hesitate to
call them SIMD vectors) may contain up to four elements, so vector
modes up to size four are supported by this patch even though the
nvptx backend currently only allows V2SI and V2DI, i.e. two out
of the ten possible vector modes.

As an example of the improvement, the following C function:

typedef int __v2si __attribute__((__vector_size__(8)));
int foo (__v2si arg) { return arg[0]+arg[1]; }

previously generated this code using a shift:

  mov.u64     %r25, %ar0;
  ld.v2.u32   %r26, [%r25];
  mov.b64     %r28, %r26;
  shr.s64     %r30, %r28, 32;
  cvt.u32.u32 %r31, %r26.x;
  cvt.u32.u64 %r32, %r30;
  add.u32     %value, %r31, %r32;

but with this patch now generates:

  mov.u64   %r25, %ar0;
  ld.v2.u32 %r26, [%r25];
  mov.u32   %r28, %r26.x;
  mov.u32   %r29, %r26.y;
  add.u32   %value, %r28, %r29;

I've implemented these getters and setters as their own instructions
instead of attempting the much more intrusive patch of changing the
backend's definition of register_operand.  Given the limited utility
of PTX vectors, I'm not convinced that attempting to support them as
operands in every instruction would be worth the effort involved.

This patch has been tested on nvptx-none hosted on x86_64-pc-linux-gnu
with "make" and "make check" with no new regressions.

2020-07-15  Roger Sayle  <roger@nextmovesoftware.com>
    Tom de Vries  <tdevries@suse.de>

gcc/ChangeLog:

* config/nvptx/nvptx.md (nvptx_vector_index_operand): New predicate.
(VECELEM): New mode attribute for a vector's uppercase element mode.
(Vecelem): New mode attribute for a vector's lowercase element mode.
(*vec_set<mode>_0, *vec_set<mode>_1, *vec_set<mode>_2)
(*vec_set<mode>_3): New instructions.
(vec_set<mode>): New expander to generate one of the above insns.
(vec_extract<mode><Vecelem>): New instruction.

gcc/testsuite/ChangeLog:

* gcc.target/nvptx/v2si-vec-set-extract.c: New test.

4 years agoTune memcpy and memset for Zen cores.
Martin Liska [Mon, 1 Jun 2020 11:21:40 +0000 (13:21 +0200)]
Tune memcpy and memset for Zen cores.

Based on the collected numbers in PR95435, I suggest the following
tuning changes:

gcc/ChangeLog:

PR target/95435
* config/i386/x86-tune-costs.h: Use libcall for large sizes for
-m32. Start using libcall from 128+ bytes.

4 years agoRe-format zen memcpy/memset costs.
Martin Liska [Mon, 1 Jun 2020 09:21:33 +0000 (11:21 +0200)]
Re-format zen memcpy/memset costs.

The patch improves readability of the memcpy and memset
expansion strategies.

gcc/ChangeLog:

* config/i386/x86-tune-costs.h: Change code formatting.

4 years agoc++: overload sets and placeholder return type [PR64194]
Patrick Palka [Thu, 30 Jul 2020 02:06:44 +0000 (22:06 -0400)]
c++: overload sets and placeholder return type [PR64194]

In the testcase below, template argument deduction for the call
g(id<int>) goes wrong because the functions in the overload set id<int>
each have a yet-undeduced auto return type, and this undeduced return
type makes try_one_overload fail to match up any of the overloads with
g's parameter type, leading to g's template argument going undeduced and
to the overload set going unresolved.

This patch fixes this issue by performing return type deduction via
instantiation before doing try_one_overload, in a manner similar to what
resolve_address_of_overloaded_function does.

gcc/cp/ChangeLog:

PR c++/64194
* pt.c (resolve_overloaded_unification): If the function
template specialization has a placeholder return type,
then instantiate it before attempting unification.

gcc/testsuite/ChangeLog:

PR c++/64194
* g++.dg/cpp1y/auto-fn60.C: New test.

4 years agoc++: alias_ctad_tweaks and constrained dguide [PR95486]
Patrick Palka [Thu, 30 Jul 2020 02:06:41 +0000 (22:06 -0400)]
c++: alias_ctad_tweaks and constrained dguide [PR95486]

In the below testcase, we're ICEing from alias_ctad_tweaks ultimately
because the implied deduction guide for X's user-defined constructor
already has constraints associated with it.  We then carry over these
constraints to 'fprime', the overlying deduction guide for the alias
template Y, via tsubst_decl from alias_ctad_tweaks.  Later in
alias_ctad_tweaks we call get_constraints followed by set_constraints
without doing remove_constraints in between, which triggers the !found
assert in set_constraints.

This patch fixes this issue by adding an intervening call to
remove_constraints.

gcc/cp/ChangeLog:

PR c++/95486
* pt.c (alias_ctad_tweaks): Call remove_constraints before
calling set_constraints.

gcc/testsuite/ChangeLog:

PR c++/95486
* g++.dg/cpp2a/class-deduction-alias3.C: New test.

4 years agoc++: abbreviated function template friend matching [PR96106]
Patrick Palka [Thu, 30 Jul 2020 02:06:36 +0000 (22:06 -0400)]
c++: abbreviated function template friend matching [PR96106]

In the below testcase, duplicate_decls wasn't merging the tsubsted
friend declaration for 'void add(auto)' with its definition, because
reduce_template_parm_level (during tsubst_friend_function) lost the
DECL_VIRTUAL_P flag on the auto's invented template parameter, which
caused template_heads_equivalent_p to deem the two template heads as not
equivalent in C++20 mode.

This patch makes reduce_template_parm_level carry over the
DECL_VIRTUAL_P flag from the original TEMPLATE_PARM_DECL.

gcc/cp/ChangeLog:

PR c++/96106
* pt.c (reduce_template_parm_level): Propagate DECL_VIRTUAL_P
from the original TEMPLATE_PARM_DECL to the new lowered one.

gcc/testsuite/ChangeLog:

PR c++/96106
* g++.dg/concepts/abbrev7.C: New test.

4 years agoc++: constraints and explicit instantiation [PR96164]
Patrick Palka [Thu, 30 Jul 2020 02:06:33 +0000 (22:06 -0400)]
c++: constraints and explicit instantiation [PR96164]

When considering to instantiate a member of a class template as part of
an explicit instantiation of the class template, we need to first check
the member's constraints before proceeding with the instantiation of the
member.

gcc/cp/ChangeLog:

PR c++/96164
* constraint.cc (constraints_satisfied_p): Return true if
!flags_concepts.
* pt.c (do_type_instantiation): Update a paragraph taken from
[temp.explicit] to reflect the latest specification.  Don't
instantiate a member with unsatisfied constraints.

gcc/testsuite/ChangeLog:

PR c++/96164
* g++.dg/cpp2a/concepts-explicit-inst5.C: New test.

4 years agoDaily bump.
GCC Administrator [Thu, 30 Jul 2020 00:16:24 +0000 (00:16 +0000)]
Daily bump.

4 years agonvptx: Support floating point reciprocal instructions
Roger Sayle [Tue, 28 Jul 2020 13:55:47 +0000 (15:55 +0200)]
nvptx: Support floating point reciprocal instructions

The following patch addds support for PTX's rcp.rn.f32 and rcp.rn.f64
instructions.  Note that the "rcp.rn" forms of this instruction
calculate the fully IEEE compliant result for the reciprocal, unlike
the rcp.approx variants that just provide fast approximations.

This patch has been tested on nvptx-none hosted on x86_64-pc-linux-gnu
with "make" and "make check" with no new regressions.

2020-07-12  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog:

* config/nvptx/nvptx.md (recip<mode>2): New instruction.

gcc/testsuite/ChangeLog:

* gcc.target/nvptx/recip-1.c: New test.

4 years agoUpdate gcc ja.po, sv.po.
Joseph Myers [Wed, 29 Jul 2020 19:32:40 +0000 (19:32 +0000)]
Update gcc ja.po, sv.po.

* ja.po, sv.po: Update.

4 years agolibstdc++: Review _Local_iterator/_Local_const_iterator implementations.
François Dumont [Mon, 20 Jan 2020 18:23:09 +0000 (19:23 +0100)]
libstdc++: Review _Local_iterator/_Local_const_iterator implementations.

_Local_iterator_base inherits _Node_iterator_base and so share the same
comparison operators. It avoids to expose special method _M_curr for debug
mode to compare such iterators.

libstdc++-v3/ChangeLog:

* include/bits/hashtable_policy.h (_Node_iterator_base()): New.
(operator==(const _Node_iterator_base&, const _Node_iterator_base&)):
Make hidden friend.
(operator!=(const _Node_iterator_base&, const _Node_iterator_base&)):
Make hidden friend.
(_Local_iterator_base<>): Inherits _Node_iterator_base.
(_Local_iterator_base<>::_M_cur): Remove.
(_Local_iterator_base<>::_M_curr()): Remove.
(operator==(const _Local_iterator_base&, const _Local_iterator_base&)):
Remove.
(operator!=(const _Local_iterator_base&, const _Local_iterator_base&)):
Remove.
* include/debug/unordered_map (unordered_map<>::_M_invalidate): Adapt.
(unordered_multimap<>::_M_invalidate): Adapt.
* include/debug/unordered_set (unordered_set<>::_M_invalidate): Adapt.
(unordered_multiset<>::_M_invalidate): Adapt.

4 years agoc++: Implement C++20 implicit move changes. [PR91427]
Jason Merrill [Tue, 21 Jul 2020 04:19:49 +0000 (00:19 -0400)]
c++: Implement C++20 implicit move changes. [PR91427]

P1825R0 extends the C++11 implicit move on return by removing the
constraints on the called constructor: previously, it needed to take an
rvalue reference to the type of the returned variable.  The paper also
allows move on throw of parameters and implicit move of rvalue references.

Discussion on the CWG reflector about how to avoid breaking the PR91212 test
in the new model settled on the model of doing only a single overload
resolution, with the variable treated as an xvalue that can bind to
non-const lvalue references.  So this patch implements that approach.  The
implementation does not use the existing LOOKUP_PREFER_RVALUE flag, but
instead sets a flag on the representation of the static_cast turning the
variable into an xvalue.

For the time being I'm limiting the new semantics to C++20 mode; since it
was moved as a DR, we will probably want to apply the change to other
standard modes as well once we have a better sense of the impact on existing
code, probably in GCC 12.

gcc/cp/ChangeLog:

PR c++/91427
* cp-tree.h (IMPLICIT_RVALUE_P): New.
(enum cp_lvalue_kind_flags): Add clk_implicit_rval.
(implicit_rvalue_p, set_implicit_rvalue_p): New.
* call.c (reference_binding): Check clk_implicit_rval.
(build_over_call): Adjust C++20 implicit move.
* coroutines.cc (finish_co_return_stmt): Simplify implicit move.
* except.c (build_throw): Adjust C++20 implicit move.
* pt.c (tsubst_copy_and_build) [STATIC_CAST_EXPR]: Propagate
IMPLICIT_RVALUE_P.
* tree.c (lvalue_kind): Set clk_implicit_rval.
* typeck.c (treat_lvalue_as_rvalue_p): Overhaul.
(maybe_warn_pessimizing_move): Adjust.
(check_return_expr): Adjust C++20 implicit move.

gcc/testsuite/ChangeLog:

PR c++/91427
* g++.dg/coroutines/co-return-syntax-10-movable.C: Extend.
* g++.dg/cpp0x/Wredundant-move1.C: Adjust for C++20.
* g++.dg/cpp0x/Wredundant-move7.C: Adjust for C++20.
* g++.dg/cpp0x/Wredundant-move9.C: Adjust for C++20.
* g++.dg/cpp0x/elision_neg.C: Adjust for C++20.
* g++.dg/cpp0x/move-return2.C: Adjust for C++20.
* g++.dg/cpp0x/ref-qual20.C: Adjust for C++20.
* g++.dg/cpp2a/implicit-move1.C: New test.
* g++.dg/cpp2a/implicit-move2.C: New test.
* g++.dg/cpp2a/implicit-move3.C: New test.

4 years agoc++: Avoid calling const copy ctor on implicit move. [PR91212]
Jason Merrill [Wed, 29 Jul 2020 04:57:40 +0000 (00:57 -0400)]
c++: Avoid calling const copy ctor on implicit move. [PR91212]

Our implementation of C++11 implicit move was wrong for return; we didn't
actually hit the check for the type of the first parameter of the selected
constructor, because we didn't see LOOKUP_PREFER_RVALUE set properly.

Fixing that to look at the right flags fixed the issue for this testcase,
but broke implicit move for a by-value converting constructor (PR58051).  I
think this was not allowed in C++17, but it is allowed under the implicit
move changes from C++20, and those changes were voted to apply as a DR to
earlier standards as well, so I don't want to break it now.

So after fixing the flags check I changed the test to allow value
parameters.

gcc/cp/ChangeLog:

PR c++/91212
* call.c (build_over_call): Don't call a const ref
overload for implicit move.

gcc/testsuite/ChangeLog:

PR c++/91212
* g++.dg/cpp0x/move-return3.C: New test.

4 years agoOpenMP: Handle order(concurrent) clause in gfortran
Tobias Burnus [Wed, 29 Jul 2020 16:37:13 +0000 (18:37 +0200)]
OpenMP: Handle order(concurrent) clause in gfortran

gcc/fortran/ChangeLog:

* dump-parse-tree.c (show_omp_clauses): Handle order(concurrent).
* gfortran.h (struct gfc_omp_clauses): Add order_concurrent.
* openmp.c (enum omp_mask1, OMP_DO_CLAUSES, OMP_SIMD_CLAUSES):
Add OMP_CLAUSE_ORDER.
* trans-openmp.c (gfc_trans_omp_clauses, gfc_split_omp_clauses):
Handle order(concurrent) clause.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/order-3.f90: New test.
* gfortran.dg/gomp/order-4.f90: New test.

4 years agopreprocessor: Teach traditional about has_include [PR95889]
Tiziano Müller [Wed, 29 Jul 2020 14:28:23 +0000 (07:28 -0700)]
preprocessor: Teach traditional about has_include [PR95889]

Traditional cpp (used by fortran) didn;t know about the new
__has_include__ implementation.  Hey, since when did traditional cpp
grow __has_include__? That wasn't in knr!

libcpp/
* init.c (builtin_array): Add xref comment.
* traditional.c (fun_like_macro): Add HAS_INCLUDE codes.
gcc/testsuite/
* c-c++-common/cpp/has-include-1-traditional.c: New.

4 years agotestsuite: libstdc++ atomic_float/value_init.cc requires libatomic
David Edelsohn [Tue, 28 Jul 2020 20:22:04 +0000 (16:22 -0400)]
testsuite: libstdc++ atomic_float/value_init.cc requires libatomic

atomic_float/value_init.cc requires libatomic for some targets, i.e., when
it tries to perform an atomic operation with a 64 bit floating point
double type on a 32 bit target.  This patch adds AIX and Darwin to the
list of targets that require the libatomic option and adds the option to
the atomic_float/value_init.cc testcase.

libstdc++-v3/ChangeLog:

2020-07-28  David Edelsohn  <dje.gcc@gmail.com>
    Jonathan Wakely  <jwakely@redhat.com>
    Rainer Orth  <ro@CeBiTec.Uni-Bielefeld.DE>

* testsuite/lib/dg-options.exp (add_options_for_libatomic): Add
target powerpc-ibm-aix* and powerpc*-*-darwin*.
* testsuite/29_atomics/atomic_float/value_init.cc: Add options
for libatomic.

4 years agoDon't make -gsplit-dwarf imply -g
Fangrui Song [Wed, 13 May 2020 15:27:29 +0000 (08:27 -0700)]
Don't make -gsplit-dwarf imply -g

-gsplit-dwarf introduces order dependency: it overrides previous -g0 and -g1.

Don't imply -g so that it can be plugged into a build without worrying
that unnecessary debugging information may be generated.

2020-05-13  Fangrui Song  <maskray@google.com>

PR debug/95096
* opts.c (common_handle_option): Don't make -gsplit-dwarf imply -g.
* doc/invoke.texi (-gsplit-dwarf): Update documentation.

4 years agoOpenMP: Permit in Fortran omp target data without map
Tobias Burnus [Wed, 29 Jul 2020 13:09:55 +0000 (15:09 +0200)]
OpenMP: Permit in Fortran omp target data without map

gcc/fortran/ChangeLog:

* openmp.c (resolve_omp_clauses): Permit 'omp target data' without
map if use_device_{addr,ptr} is present.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/map-3.f90: New test.
* gfortran.dg/gomp/map-4.f90: New test.

4 years agoarm: Enable no-writeback vldr.16/vstr.16.
Joe Ramsay [Wed, 29 Jul 2020 13:04:28 +0000 (14:04 +0100)]
arm: Enable no-writeback vldr.16/vstr.16.

There was previously no way to specify that a register operand cannot
have any writeback modifiers, and as a result the argument to vldr.16
and vstr.16 could be erroneously output with post-increment. This
change adds a constraint which forbids all writeback, and
selects it in the relevant case for vldr.16 and vstr.16

Bootstrapped on arm-linux, gcc and CMSIS-DSP testsuites are clean.
Is this patch OK for trunk? If yes, please commit on my behalf as I
don't
have commit rights.

gcc/ChangeLog:

* config/arm/arm-protos.h (arm_coproc_mem_operand_no_writeback):
Declare prototype.
(arm_mve_mode_and_operands_type_check): Declare prototype.
* config/arm/arm.c (arm_coproc_mem_operand): Refactor to use
_arm_coproc_mem_operand.
(arm_coproc_mem_operand_wb): New function to cover full, limited
and no writeback.
(arm_coproc_mem_operand_no_writeback): New constraint for memory
operand with no writeback.
(arm_print_operand): Extend 'E' specifier for memory operand
that does not support writeback.
(arm_mve_mode_and_operands_type_check): New constraint check for
MVE memory operands.
* config/arm/constraints.md: Add Uj constraint for VFP vldr.16
and vstr.16.
* config/arm/vfp.md (*mov_load_vfp_hf16): New pattern for
vldr.16.
(*mov_store_vfp_hf16): New pattern for vstr.16.
(*mov<mode>_vfp_<mode>16): Remove MVE moves.

4 years agogcc-changelog: fix combining of arguments.
Martin Liska [Wed, 29 Jul 2020 12:13:42 +0000 (14:13 +0200)]
gcc-changelog: fix combining of arguments.

contrib/ChangeLog:

2020-07-29  Martin Liska  <mliska@suse.cz>

* git-backport.py: fix how are ChangeLog paths combined.

4 years agoFortran : Don't warn for LOGICAL kind conversion PR96319
Mark Eggleston [Mon, 27 Jul 2020 14:28:50 +0000 (15:28 +0100)]
Fortran  : Don't warn for LOGICAL kind conversion PR96319

LOGICAL values will always fit regardless of kind so there
is no need for warnings.

2020-07-29  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/fortran/

PR fortran/96319
* intrinsic.c (gfc_convert_type_warn):  Add check for
LOGICAL type so that warnings are not output.

2020-07-29  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/96319
* gfortran.dg/pr96319.f90: New test.

4 years agotree-optimization/96349 - avoid abnormal coalescing issues in loop split
Richard Biener [Tue, 28 Jul 2020 07:45:52 +0000 (09:45 +0200)]
tree-optimization/96349 - avoid abnormal coalescing issues in loop split

This avoids splitting a loop when the entry value of a loop PHI is
involved with abnormal coalescing.

2020-07-28  Richard Biener  <rguenther@suse.de>

PR tree-optimization/96349
* tree-ssa-loop-split.c (stmt_semi_invariant_p_1): When the
condition runs into a loop PHI with an abnormal entry value give up.

* gcc.dg/torture/pr96349.c: New testcase.

4 years agomore SCEV cache clearing
Richard Biener [Wed, 29 Jul 2020 10:30:59 +0000 (12:30 +0200)]
more SCEV cache clearing

This fixes two more places, in loop interchange and in the
vectorizer where the SCEV verifier sees stale entries.

2020-07-29  Richard Biener  <rguenther@suse.de>

* tree-vectorizer.c (vectorize_loops): Reset the SCEV
cache if we removed any SIMD UID SSA defs.
* gimple-loop-interchange.cc (pass_linterchange::execute):
Reset the scev cache if we interchanged a loop.

4 years agotree-optimization/95679 - properly signal changes from propagate_into_phi_args
Richard Biener [Wed, 29 Jul 2020 07:59:01 +0000 (09:59 +0200)]
tree-optimization/95679 - properly signal changes from propagate_into_phi_args

This restores a lost setting of something_changed with the
recent refactoring of the substitute and fold engine.  The
reported ICE in the PR was meanwhile mitigated in other ways
but the issue can still result in missed optimizations via
failed runs of CFG cleanup.

2020-07-29  Richard Biener  <rguenther@suse.de>

PR tree-optimization/95679
* tree-ssa-propagate.h
(substitute_and_fold_engine::propagate_into_phi_args): Return
whether anything changed.
* tree-ssa-propagate.c
(substitute_and_fold_engine::propagate_into_phi_args): Likewise.
(substitute_and_fold_dom_walker::before_dom_children): Update
something_changed.

4 years agogcc/fortran/module.c: Fix indentation
Tobias Burnus [Wed, 29 Jul 2020 10:26:12 +0000 (12:26 +0200)]
gcc/fortran/module.c: Fix indentation

gcc/fortran/ChangeLog:

* module.c (mio_symbol_attribute): Fix indent of previous
commit.

4 years agoFortran : ICE in gfc_conv_scalarized_array_ref PR53298
Mark Eggleston [Fri, 17 Jul 2020 13:22:48 +0000 (14:22 +0100)]
Fortran  : ICE in gfc_conv_scalarized_array_ref PR53298

When an array of characters is an argument to a subroutine and
is accessed using (:)(1:) an ICE occurs.  The upper bound of the
substring does not have an expression and such should not have
a Scalarization State structure added to the Scalarization State
chain.

2020-07-29  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/fortran/

PR fortran/53298
* trans-array.c (gfc_walk_array_ref): If ref->ss.end is set
call gfc_get_scalar_ss.

2020-07-29  Mark Eggleston  <markeggleston@gcc.gnu.org>

gcc/testsuite/

PR fortran/53298
* gfortran.dg/pr53298.f90: New test.

4 years agolibstdc++: Fix unordered containers move constructors noexcept qualification
François Dumont [Mon, 20 Jan 2020 18:15:43 +0000 (19:15 +0100)]
libstdc++: Fix unordered containers move constructors noexcept qualification

_Hashtable move constructor is wrongly qualified as noexcept(true) regardless of
_Equal and _H1 copy constructor qualifications.
_Hashtable allocator-aware move constructor is missing its noexcept
qualification like the depending unordered containers ones.

libstdc++-v3/ChangeLog:

* include/bits/hashtable.h
(_Hashtable(_Hashtable&& __ht, __node_alloc_type&& __a, true_type)):
Add noexcept qualification.
(_Hashtable(_Hashtable&&)): Fix noexcept qualification.
(_Hashtable(_Hashtable&&, const allocator_type&)): Add noexcept
qualification.
* include/bits/unordered_map.h
(unordered_map(unordered_map&&, const allocator_type&)): Add noexcept
qualification.
(unordered_multimap(unordered_multimap&&, const allocator_type&)):
Likewise.
* include/bits/unordered_set.h
(unordered_set(unordered_set&&, const allocator_type&)): Likewise.
(unordered_multiset(unordered_multiset&&, const allocator_type&)):
Likewise.
* include/debug/unordered_map
(unordered_map(unordered_map&&, const allocator_type&)): Likewise.
(unordered_multimap(unordered_multimap&&, const allocator_type&)):
Likewise.
* include/debug/unordered_set
(unordered_set(unordered_set&&, const allocator_type&)): Likewise.
(unordered_multiset(unordered_multiset&&, const allocator_type&)):
Likewise.
* testsuite/23_containers/unordered_map/allocator/default_init.cc:
New test.
* testsuite/23_containers/unordered_map/cons/noexcept_default_construct.cc:
New test.
* testsuite/23_containers/unordered_map/cons/noexcept_move_construct.cc:
New test.
* testsuite/23_containers/unordered_map/modifiers/move_assign.cc:
New test.
* testsuite/23_containers/unordered_multimap/cons/noexcept_default_construct.cc:
New test.
* testsuite/23_containers/unordered_multimap/cons/noexcept_move_construct.cc:
New test.
* testsuite/23_containers/unordered_multiset/cons/noexcept_default_construct.cc:
New test.
* testsuite/23_containers/unordered_multiset/cons/noexcept_move_construct.cc:
New test.
* testsuite/23_containers/unordered_set/allocator/default_init.cc:
New test.
* testsuite/23_containers/unordered_set/cons/noexcept_default_construct.cc:
New test.
* testsuite/23_containers/unordered_set/cons/noexcept_move_construct.cc:
New test.

4 years agovect: Fix infinite loop while determining peeling amount
Stefan Schulze Frielinghaus [Wed, 22 Jul 2020 07:27:49 +0000 (09:27 +0200)]
vect: Fix infinite loop while determining peeling amount

This is a follow up to commit 5c9669a0e6c respectively discussion
https://gcc.gnu.org/pipermail/gcc-patches/2020-June/549132.html

In case a type has a lower alignment than its size, ensure that we
advance.  For example, on s390x we have for a long double an alignment
constraint of 8 bytes whereas the size is 16 bytes.  Increasing the loop
variable only by TARGET_ALIGN / DR_SIZE which equals zero in case of a
long double results in an infinite loop.  This is fixed by ensuring that
we advance at least by one.

gcc/ChangeLog:

* tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
Ensure that loop variable npeel_tmp advances in each iteration.

4 years agoOpenMP: Add 'omp requires' to Fortran (mostly parsing)
Tobias Burnus [Wed, 29 Jul 2020 08:37:44 +0000 (10:37 +0200)]
OpenMP: Add 'omp requires' to Fortran (mostly parsing)

gcc/fortran/ChangeLog:

* gfortran.h (enum gfc_statement): Add ST_OMP_REQUIRES.
(enum gfc_omp_requires_kind): New.
(enum gfc_omp_atomic_op): Add GFC_OMP_ATOMIC_ACQ_REL.
(struct gfc_namespace): Add omp_requires and omp_target_seen.
(gfc_omp_requires_add_clause,
(gfc_check_omp_requires): New.
* match.h (gfc_match_omp_requires): New.
* module.c (enum ab_attribute, attr_bits): Add omp requires clauses.
(mio_symbol_attribute): Read/write them.
* openmp.c (gfc_check_omp_requires, (gfc_omp_requires_add_clause,
gfc_match_omp_requires): New.
(gfc_match_omp_oacc_atomic): Use requires's default mem-order.
* parse.c (decode_omp_directive): Match requires, set omp_target_seen.
(gfc_ascii_statement): Handle ST_OMP_REQUIRES.
* trans-openmp.c (gfc_trans_omp_atomic): Handle GFC_OMP_ATOMIC_ACQ_REL.

gcc/testsuite/ChangeLog:

* gfortran.dg/gomp/requires-1.f90: New test.
* gfortran.dg/gomp/requires-2.f90: New test.
* gfortran.dg/gomp/requires-3.f90: New test.
* gfortran.dg/gomp/requires-4.f90: New test.
* gfortran.dg/gomp/requires-5.f90: New test.
* gfortran.dg/gomp/requires-6.f90: New test.
* gfortran.dg/gomp/requires-7.f90: New test.
* gfortran.dg/gomp/requires-8.f90: New test.
* gfortran.dg/gomp/requires-9.f90: New test.

4 years agoconfig/mmix/mmix.h (NO_FUNCTION_CSE): Define to 1.
Hans-Peter Nilsson [Wed, 29 Jul 2020 00:46:09 +0000 (02:46 +0200)]
config/mmix/mmix.h (NO_FUNCTION_CSE): Define to 1.

The tests gcc.dg/tree-ssa/loop-1.c and gcc.dg/weak/typeof-2.c
assume this setting and are as a consequence riddled with
exceptions for targets that actually do yield better code when
calling through a register rather than repeatedly the same
symbol.  Nonetheless, defining it makes sense for MMIX.

(Even better IMHO, this macro shouldn't exist and instead rtx
costs be used or perhaps just a target-specific default of
flag_no_function_cse.)

gcc:
* config/mmix/mmix.h (NO_FUNCTION_CSE): Define to 1.

4 years agommix.h (ASM_OUTPUT_EXTERNAL): Define to default_elf_asm_output_external.
Hans-Peter Nilsson [Wed, 29 Jul 2020 00:29:54 +0000 (02:29 +0200)]
mmix.h (ASM_OUTPUT_EXTERNAL): Define to default_elf_asm_output_external.

Whoops.  When un-disabling visibility support for mmix, I missed that
some of the newly enabled tests were FAILs, for not emitting .hidden
for references to external declarations.  This takes care of
gcc.dg/visibility-14.c .. -19.c, and gcc.dg/visibility-23.c.

gcc:
* config/mmix/mmix.h (ASM_OUTPUT_EXTERNAL): Define to
default_elf_asm_output_external.

4 years agoDaily bump.
GCC Administrator [Wed, 29 Jul 2020 00:16:24 +0000 (00:16 +0000)]
Daily bump.

4 years agoipa/96291: don't crash on unoptimized lto functions
Sergei Trofimovich [Sat, 25 Jul 2020 18:26:50 +0000 (19:26 +0100)]
ipa/96291: don't crash on unoptimized lto functions

In PR ipa/96291 the test contained an SCC with one
unoptimized function. This tricked ipa-cp into NULL dereference.

has_undead_caller_from_outside_scc_p() did not take into account
that unoptimized funtions don't have IPA summary analysis. And
dereferenced NULL pointer causing an ICE.

gcc/
PR ipa/96291
* ipa-cp.c (has_undead_caller_from_outside_scc_p): Consider
unoptimized callers as undead.

gcc/testsuite/
PR ipa/96291
* gcc.dg/lto/pr96291_0.c: New testcase.
* gcc.dg/lto/pr96291_1.c: Support file.
* gcc.dg/lto/pr96291_2.c: Likewise.
* gcc.dg/lto/pr96291.h: Likewise.

4 years agomiddle-end: Parity and popcount folding optimizations.
Roger Sayle [Tue, 28 Jul 2020 21:55:12 +0000 (22:55 +0100)]
middle-end: Parity and popcount folding optimizations.

This patch implements several constant folding optimizations
for __builtin_parity and friends.  We canonicalize popcount(x)&1
as parity(x) in gimple, and potentially convert back again when
we expand to RTL.  parity(~x) is simplified to parity(x), which
is true for all integer modes with an even number of bits.
But probably most usefully, parity(x)^parity(y) can be simplified
to a parity(x^y), requiring only a single libcall or popcount.

This patch optimizes popcount and parity of an argument known to have
at most a single bit set, to be that single bit.  Hence, popcount(x&8)
is simplified to (x>>3)&1.   This generalizes the existing optimization
of popcount(x&1) being simplified to x&1, which is cleaned up with
this patch.

2020-07-28  Roger Sayle  <roger@nextmovesoftware.com>
    Richard Biener  <rguenther@suse.de>

gcc/ChangeLog
* match.pd (popcount(x)&1 -> parity(x)): New simplification.
(parity(~x) -> parity(x)): New simplification.
(parity(x)^parity(y) -> parity(x^y)): New simplification.
(parity(x&1) -> x&1): New simplification.
(popcount(x) -> x>>C): New simplification.

gcc/testsuite/ChangeLog
* gcc.dg/fold-popcount-5.c: New test.
* gcc.dg/fold-parity-1.c: Likewise.
* gcc.dg/fold-parity-2.c: Likewise.
* gcc.dg/fold-parity-3.c: Likewise.
* gcc.dg/fold-parity-4.c: Likewise.
* gcc.dg/fold-parity-5.c: Likewise.

4 years agoc++: Set more DECL_CONTEXTs
Nathan Sidwell [Tue, 28 Jul 2020 15:57:36 +0000 (08:57 -0700)]
c++: Set more DECL_CONTEXTs

I discovered we were not setting DECL_CONTEXT in a few cases, and
grokfndecl's control flow wasn't making it clear that we were doing it
in all cases.

gcc/cp/
* cp-gimplify.c (cp_genericize_r): Set IMPORTED_DECL's context.
* cp-objcp-common.c (cp_pushdecl): Set decl's context.
* decl.c (grokfndecl): Make DECL_CONTEXT setting clearer.

4 years agoc++: better fixup_type_variants
Nathan Sidwell [Tue, 28 Jul 2020 15:49:13 +0000 (08:49 -0700)]
c++: better fixup_type_variants

fixup_type_variants was almost doing all that finish_struct needs.
May as well make it do it all.

gcc/cp/
* class.c (fixup_type_variants): Copy TYPE_SIZE and
TYPE_SIZE_UINIT.
(finish_struct): Call it.

4 years agoc++: tree dump indentation
Nathan Sidwell [Tue, 28 Jul 2020 15:33:48 +0000 (08:33 -0700)]
c++: tree dump indentation

We were always forcing an indent, even if there was nothing to indent.
Fixed thusly.

gcc/cp/
* ptree.c (cxx_print_decl): Better indentation.

4 years agotestsuite: Fix spello
Nathan Sidwell [Tue, 28 Jul 2020 15:31:19 +0000 (08:31 -0700)]
testsuite: Fix spello

'Patterns' has one 'r'.

gcc/testsuite/
* lib/options.exp: Fix spello

4 years agoc++: Fix up cp_lexer_safe_previous_token [PR96328]
Jakub Jelinek [Tue, 28 Jul 2020 13:41:30 +0000 (15:41 +0200)]
c++: Fix up cp_lexer_safe_previous_token [PR96328]

The following testcase ICEs, because cp_lexer_safe_previous_token calls
cp_lexer_previous_token and that ICEs, because all tokens in the lexer
buffer before the current one (CPP_EOF) have been purged.

cp_lexer_safe_previous_token is used in the context where it is ok if it
punts, so the patch changes the function so that it doesn't assert there is
some previous token, but instead returns NULL like in other cases where it
punts.

In addition to this, in the last hunk it does a micro-optimization, don't
call the potentially expensive function if it will not need the result,
instead check the least expensive condition first.

And the middle hunk is a similar change from Mark's version of the patch,
to use the safe variant in there because it is again just about a hint
and it is better not to provide the hint than to ICE, though we don't have a
testcase that would ICE.

2020-07-28  Jakub Jelinek  <jakub@redhat.com>
    Mark Wielaard  <mark@klomp.org>

PR c++/96328
* parser.c (cp_lexer_safe_previous_token): Don't call
cp_lexer_previous_token, instead inline it by hand and return NULL
instead of failing assertion if all previous tokens until the first
one are purged.
(cp_parser_error_1): Optimize - only call cp_lexer_safe_previous_token
if token->type is CPP_NAME.  Use cp_lexer_safe_previous_token instead
of cp_lexer_previous_token for the missing_token_desc != RT_NONE
case too.

* g++.dg/diagnostic/pr96328.C: New test.

Co-Authored-By: Mark Wielaard <mark@klomp.org>
4 years agolibcpp: Fix up raw string literal parsing error-recovery [PR96323]
Jakub Jelinek [Tue, 28 Jul 2020 13:40:15 +0000 (15:40 +0200)]
libcpp: Fix up raw string literal parsing error-recovery [PR96323]

For (invalid) newline inside of the raw string literal delimiter, doing
continue means we skip the needed processing of newlines.  Instead of
duplicating that, this patch just doesn't continue for those.

2020-07-28  Jakub Jelinek  <jakub@redhat.com>

PR preprocessor/96323
* lex.c (lex_raw_string): For c == '\n' don't continue after reporting
an prefix delimiter error.

* c-c++-common/cpp/pr96323.c: New test.

4 years agonvptx: Support 16-bit shifts and extendqihi2
Tom de Vries [Tue, 28 Jul 2020 13:13:54 +0000 (15:13 +0200)]
nvptx: Support 16-bit shifts and extendqihi2

Add support for 16-bits shifts and for sign extension from 8 bits to
16 bits.

This patch has been tested on nvptx-none with no new regressions.

2020-07-28  Roger Sayle  <roger@nextmovesoftware.com>
    Tom de Vries  <tdevries@suse.de>

gcc/ChangeLog:

* config/nvptx/nvptx.md (extendqihi2): New instruction.
(ashl<mode>3, ashr<mode>3, lshr<mode>3): Support HImode.

gcc/testsuite/ChangeLog:

* gcc.target/nvptx/cvt.c: New test.
* gcc.target/nvptx/shift16.c: New test.

4 years agopreprocessor: Simplify read_main
Nathan Sidwell [Tue, 28 Jul 2020 12:49:35 +0000 (05:49 -0700)]
preprocessor: Simplify read_main

We can always use the final map to get the return value, rather than
conditionally only when there was an immediate line directive.

libcpp/
* init.c (cpp_read_main_file): Always use the last map for
the return value.

4 years agolibstdc++: Do not over-size hashtable buckets on range insertion
François Dumont [Mon, 20 Jan 2020 18:01:18 +0000 (19:01 +0100)]
libstdc++: Do not over-size hashtable buckets on range insertion

We used to consider range size on insertion but on unique keys container
not all range values might be inserted resulting in over-sizing. In this
case we just consider user reservation and if none then the container will
adapt to actually inserted elements.

libstdc++-v3/ChangeLog:

* include/bits/hashtable.h
(_Hashtable<>(_InputIterator, _InputIterator, size_t, const _H1&,
const _H2&, const _Hash&, const _Equal&, const _ExtractKey&,
const allocator_type&, true_type)): New.
(_Hashtable<>(_InputIterator, _InputIterator, size_t, const _H1&,
const _H2&, const _Hash&, const _Equal&, const _ExtractKey&,
const allocator_type&, false_type)): New.
(_Hashtable<>(_InputIterator, _InputIterator, size_t, const _H1&,
const _H2&, const _Hash&, const _Equal&, const _ExtractKey&,
const allocator_type&)): Delegate to latters.
(operator=(initializer_list<value_type>)): Rehash if too small.
(_M_insert(_Arg&&, const _NodeGenerator&, true_type)): Remove
size_t len parameter.
* include/bits/hashtable_policy.h (_Insert_base<>::_M_insert_range):
Do not try to get input range distance.
* testsuite/23_containers/unordered_set/cons/bucket_hint.cc: New test.
* testsuite/23_containers/unordered_set/modifiers/insert.cc: New test.

4 years agotestsuite: Fix up nontype-subob1.C for targets with short int, int or long long ssize_t
Jakub Jelinek [Tue, 28 Jul 2020 09:32:36 +0000 (11:32 +0200)]
testsuite: Fix up nontype-subob1.C for targets with short int, int or long long ssize_t

2020-07-28  Jakub Jelinek  <jakub@redhat.com>

* g++.dg/cpp2a/nontype-subob1.C: Allow s1, i1 or x1 instead of l1 for
targets with short, int or long long ssize_t.

4 years agoexpander: Fix ICE in maybe_warn_rdwr_sizes [PR96335]
Jakub Jelinek [Tue, 28 Jul 2020 09:08:29 +0000 (11:08 +0200)]
expander: Fix ICE in maybe_warn_rdwr_sizes [PR96335]

The following testcase ICEs in maybe_warn_rdwr_sizes.  The problem is that
the caller uses its fndecl and fntype variables to fill up rdwr_map, and
the fntype in that case is a prototype with the access attribute and all
the checks needed for that performed.  But the maybe_warn_rdwr_sizes
function tries to rediscover fndecl/fntype itself and does it differently
from how the caller did (for fndecl get_callee_fndecl and fntype from that
FUNCTION_DECL, otherwise sets fntype to CALL_EXPR_FN's type).

On the testcase, get_callee_fndecl does find a FUNCTION_DECL because
it does STRIP_NOPS in between.

Instead of trying to rediscover those, this patch just passes them down,
like is done in several other functions.

2020-07-28  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/96335
* calls.c (maybe_warn_rdwr_sizes): Add FNDECL and FNTYPE arguments,
instead of trying to rediscover them in the body.
(initialize_argument_information): Adjust caller.

* gcc.dg/pr96335.c: New test.

4 years agovect: Refactor peel_iters_{pro,epi}logue cost modeling
Kewen Lin [Tue, 28 Jul 2020 02:30:26 +0000 (21:30 -0500)]
vect: Refactor peel_iters_{pro,epi}logue cost modeling

This patch is to refactor the existing peel_iters_prologue and
peel_iters_epilogue cost model handlings, by following the structure
below suggested by Richard Sandiford:

  - calculate peel_iters_prologue
  - calculate peel_iters_epilogue
  - add costs associated with peel_iters_prologue
  - add costs associated with peel_iters_epilogue
  - add costs related to branch taken/not_taken.

Bootstrapped/regtested on aarch64-linux-gnu.

gcc/ChangeLog:

* tree-vect-loop.c (vect_get_known_peeling_cost): Factor out some code
to determine peel_iters_epilogue to...
(vect_get_peel_iters_epilogue): ...this new function.
(vect_estimate_min_profitable_iters): Refactor cost calculation on
peel_iters_prologue and peel_iters_epilogue.

4 years agoDaily bump.
GCC Administrator [Tue, 28 Jul 2020 00:16:25 +0000 (00:16 +0000)]
Daily bump.

4 years agocompiler,runtime: pass only ptr and len to some runtime calls
Ian Lance Taylor [Mon, 27 Jul 2020 22:19:54 +0000 (15:19 -0700)]
compiler,runtime: pass only ptr and len to some runtime calls

This ports https://golang.org/cl/227163 to the Go frontend.
This is a step toward moving up to the go1.15rc1 release.

Original CL description:

    cmd/compile,runtime: pass only ptr and len to some runtime calls

    Some runtime calls accept a slice, but only use ptr and len.
    This change modifies most such routines to accept only ptr and len.

    After this change, the only runtime calls that accept an unnecessary
    cap arg are concatstrings and slicerunetostring.
    Neither is particularly common, and both are complicated to modify.

    Negligible compiler performance impact. Shrinks binaries a little.
    There are only a few regressions; the one I investigated was
    due to register allocation fluctuation.

    Passes 'go test -race std cmd', modulo golang/go#38265 and golang/go#38266.
    Wow, does that take a long time to run.

    file      before    after     Δ       %
    compile   19655024  19655152  +128    +0.001%
    cover     5244840   5236648   -8192   -0.156%
    dist      3662376   3658280   -4096   -0.112%
    link      6680056   6675960   -4096   -0.061%
    pprof     14789844  14777556  -12288  -0.083%
    test2json 2824744   2820648   -4096   -0.145%
    trace     11647876  11639684  -8192   -0.070%
    vet       8260472   8256376   -4096   -0.050%
    total     115163736 115118808 -44928  -0.039%

For golang/go#36890

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/245099

4 years agocompiler: for package-scope "a = b; b = x" just set "a = x"
Ian Lance Taylor [Mon, 27 Jul 2020 21:22:55 +0000 (14:22 -0700)]
compiler: for package-scope "a = b; b = x" just set "a = x"

This avoids requiring an init function to initialize the variable.
This can only be done if x is a static initializer.

The go1.15rc1 runtime package relies on this optimization.
The package has a variable "var maxSearchAddr = maxOffAddr".
The maxSearchAddr variable is used by code that runs before package
initialization is complete.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/245098

4 years agogcc.dg/torture/pr39074-2.c, pr39074.c, pta-callused-1.c: Adjust for mmix.
Hans-Peter Nilsson [Mon, 27 Jul 2020 23:41:47 +0000 (01:41 +0200)]
gcc.dg/torture/pr39074-2.c, pr39074.c, pta-callused-1.c: Adjust for mmix.

These FAILs for mmix showed up as regressions for me due to a
flaw in the btest-gcc.sh test-results-accounting: a bug was
recently fixed regarding the naming of dump-files so the names
are again correct.  To wit, parts of the tests that were
UNRESOLVED, due to missing dump-files, and ignored in the
presence of other parts (execution, excess errors) PASSing,
became FAIL, trumping the PASSing parts of the tests.

As in a recent patch, the variables are "privatized" using
ASM_PN_FORMAT for MMIX and the lines to match look like:
 y::0_1 = { i }
 y::0_1, points-to NULL, points-to vars: { D.1465 } (nonlocal, escaped)
instead of e.g. for cris-elf:
 y.0_1 = { i }
 y.0_1, points-to NULL, points-to vars: { D.1433 } (nonlocal, escaped)
Also checked that the general pattern still matches for cris-elf.

gcc/testsuite:
* gcc.dg/torture/pr39074-2.c: Adjust for mmix.
* gcc.dg/torture/pr39074.c, gcc.dg/torture/pta-callused-1.c: Ditto.

4 years agoUpdate gcc .po files.
Joseph Myers [Mon, 27 Jul 2020 22:51:58 +0000 (22:51 +0000)]
Update gcc .po files.

* be.po, da.po, de.po, el.po, es.po, fi.po, fr.po, hr.po, id.po,
ja.po, nl.po, ru.po, sr.po, sv.po, tr.po, uk.po, vi.po, zh_CN.po,
zh_TW.po: Update.

4 years agocommon: Use strcmp to compare location file names
Nathan Sidwell [Mon, 27 Jul 2020 20:05:28 +0000 (13:05 -0700)]
common: Use strcmp to compare location file names

The logic to figure out where a missing #include should be inserted
uses pointer equality to check filenames -- the routine even says so.
But cpplib makes no such guarantee.  It happens to be true for input
that it preprocesses[* see line zero below], but is not true for
source that has already been preprocessed -- all those '# ...' line
directives produce disctinct filename strings.  That renders using
-fdirectives-only as a prescanning stage (as I understand some people
do), broken.

This patch changes to string comparisons, and explicitly rejects any
line-zero location map that occurs at the beginning of a file.  The
very first map of a file has a different string to the remaining maps,
and we never tripped on that because of the pointer comparison.  The
second testcase deploys -save-temps to cause an intermediate
preprocessed output that is read back.

gcc/c-family/
* c-common.c (try_to_locate_new_include_insertion_point): Use
strcmp, not pointer equality.
gcc/testsuite/
* g++.dg/lookup/missing-std-include-10.h: New.
* g++.dg/lookup/missing-std-include-10.C: New.
* g++.dg/lookup/missing-std-include-11.C: New.

4 years agoc++: Name as_base type
Nathan Sidwell [Wed, 22 Jul 2020 19:00:02 +0000 (12:00 -0700)]
c++: Name as_base type

The as-base type never got a name.  For modules I needed to give it a
name to serialize properly, and it's useful when debugging the
compiler, so we may as well have it on trunk.  There's also a bug
where its fields can have NSDMIs from the main class.  This happens to
be silent on trunk, but can be a GC leak where we retain a deferred
parse node there. (On modules it blows up, because we're not prepared
to serialize deferred parse nodes, as they should never survive
parsing.

gcc/cp/
* cp-tree.h (enum cp_tree_index): Add CPTI_AS_BASE_IDENTIFIER.
(as_base_identifier): Define.
* decl.c (initialize_predifined_identifiers): Initialize as_base
identifier.
* class.c (layout_class_type): Name the as-base type.  Zap
NSDMI its fields may have.

4 years agoDiagnose just-past-the-end references for minor array bounds.
Martin Sebor [Mon, 27 Jul 2020 19:54:50 +0000 (13:54 -0600)]
Diagnose just-past-the-end references for minor array bounds.

Resolves:
PR tree-optimization/84079 - missing -Warray-bounds taking the address of past-the-end element of a multidimensional array

gcc/ChangeLog:

PR tree-optimization/84079
* gimple-array-bounds.cc (array_bounds_checker::check_addr_expr):
Only allow just-past-the-end references for the most significant
array bound.

gcc/testsuite/ChangeLog:

PR tree-optimization/84079
* gcc.dg/Warray-bounds-62.c: New test.

4 years agolibstdc++: Review _Hashtable count, equal_range _M_erase(false_type,) code
François Dumont [Mon, 20 Jan 2020 07:24:47 +0000 (08:24 +0100)]
libstdc++: Review _Hashtable count, equal_range _M_erase(false_type,) code

Simplify operator[] implementation using find method. Review several
_Hashtable method implementations to limit the computation of bucket index.
Introduce _M_update_bbegin to simplify code.

libstdc++-v3/ChangeLog:

* include/bits/hashtable_policy.h (_Map_base<>::at): Use
_Hashtable<>::find.
(_Hashtable_base<>::_Equal_hash_code<>::_S_node_equals):New.
(_Hashtable_base<>::_M_node_equals): New, use latter.
(_Equality<_Key, _Value, _Alloc, _ExtractKey, _Equal, _H1, _H2, _Hash,
_RehashPolicy, false>::_M_equal): Adapt to use latter.
* include/bits/hashtable.h (_Hashtable<>::_M_update_bbegin): New.
(_Hashtable<>::_M_assign): Use latter.
(_Hashtable<>::_M_move_assign): Likewise.
(_Hashtable<>(_Hashtable<>&&)): Likewise.
(_Hashtable<>(_Hashtable<>&&, const allocator_type&)): Likewise.
(_Hashtable<>::swap): Likewise.
(_Hashtable<>::find): Build iterator directly from _M_find_node result.
(_Hashtable<>::count): Use _Hashtable<>::find.
(_Hashtable<>::equal_range): Likewise.
(_Hashtable<>::_M_erase(false_type, const key_type&)): Use
_M_node_equals.

4 years agodriver: fix a problem with implementation of -falign-foo=0 [PR96247]
Hu Jiangping [Mon, 27 Jul 2020 17:22:23 +0000 (18:22 +0100)]
driver: fix a problem with implementation of -falign-foo=0 [PR96247]

This patch makes the -falign-foo=0 work as described in the
documentation. Thanks for all the suggestions.

Changelog:
2020-07-27  Hu Jiangping  <hujiangping@cn.fujitsu.com>

PR driver/96247
* opts.c (check_alignment_argument): Set the -falign-Name
on/off flag on and set the -falign-Name string value null,
when the command-line specified argument is zero.

4 years agocompiler: scan all function literals for escape analysis
Ian Lance Taylor [Sun, 26 Jul 2020 18:08:13 +0000 (11:08 -0700)]
compiler: scan all function literals for escape analysis

We were scanning only function literals with closures, but not all
function literals have closures.

Discovered because compiler failed building 1.15rc1, as there is a
function literal in the runtime package (p1 in hexdumpWords) that has
no closure and, without escape analysis, was forcing a variable to the
heap which is not permitted in the runtime.

Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/244802

4 years agoopenacc: Deep copy attach/detach should not affect reference counts
Julian Brown [Thu, 18 Jun 2020 12:11:08 +0000 (05:11 -0700)]
openacc: Deep copy attach/detach should not affect reference counts

Attach and detach operations are not supposed to affect structural or
dynamic reference counts for OpenACC. Previously they did so, which led to
subtle problems in some circumstances. We can avoid reference-counting
attach/detach operations by extending and slightly repurposing the
do_detach field in target_var_desc. It is now called is_attach to better
reflect its new role.

2020-07-27  Julian Brown  <julian@codesourcery.com>
    Thomas Schwinge  <thomas@codesourcery.com>

libgomp/
* libgomp.h (struct target_var_desc): Rename do_detach field to
is_attach.
* oacc-mem.c (goacc_exit_datum_1): Add assert.  Don't set finalize for
GOMP_MAP_FORCE_DETACH. Update checking to use is_attach field.
(goacc_enter_data_internal): Don't affect reference counts
for attach mappings.
(goacc_exit_data_internal): Don't affect reference counts for detach
mappings.
* target.c (gomp_map_vars_existing): Don't affect reference counts for
attach mappings.
(gomp_map_vars_internal): Set renamed is_attach flag unconditionally to
mark attach mappings.
(gomp_unmap_vars_internal): Use is_attach flag to prevent affecting
reference count for attach mappings.
* testsuite/libgomp.oacc-c-c++-common/mdc-refcount-1.c: New test.
* testsuite/libgomp.oacc-c-c++-common/mdc-refcount-2.c: New test.
* testsuite/libgomp.oacc-c-c++-common/mdc-refcount-2.c: New test.
* testsuite/libgomp.oacc-fortran/deep-copy-6-no_finalize.F90: Mark
test as shouldfail.
* testsuite/libgomp.oacc-fortran/deep-copy-6.f90: Adjust to fail
gracefully in no-finalize mode.

Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
4 years agolibstdc++: Make std::from_chars always round to nearest
Jonathan Wakely [Mon, 27 Jul 2020 14:51:16 +0000 (15:51 +0100)]
libstdc++: Make std::from_chars always round to nearest

Also fix the tests that fail on targets without uselocale.

libstdc++-v3/ChangeLog:

* src/c++17/floating_from_chars.cc (from_chars_impl): Ensure
that FE_NEAREST is used.
* testsuite/20_util/from_chars/4.cc: Do not use if constexpr in
a { target c++14 } test.
[!_GLIBCXX_HAVE_USELOCALE]: Disable all tests.
* testsuite/20_util/from_chars/5.cc [!_GLIBCXX_HAVE_USELOCALE]:
Likewise.
* testsuite/20_util/from_chars/6.cc: New test.

4 years agoexpr: build string_constant only for a char type
Martin Liska [Mon, 27 Jul 2020 10:30:24 +0000 (12:30 +0200)]
expr: build string_constant only for a char type

gcc/ChangeLog:

PR tree-optimization/96058
* expr.c (string_constant): Build string_constant only
for a type that has same precision as char_type_node
and is an integral type.