libphobos: Move rt.sections modules to gcc.sections
These modules depend on a mixture between how the compiler emits
run-time module information, and what functions are exposed by the
platform to inquire about loaded global and thread-local data sections.
As the upstream implementation is written to work only with how the
reference D compiler writes out data, much of what is present does not
apply to the GCC D front-end. So it has been moved to a non-upstream
location in the source tree, where most of it will be rewritten once
each port has been completed.
The only tested module sections/elf_shared.d has been cleaned up so that
all deprecated declarations have been removed, as well as the brittle
module collision checking, which required bss_sections.c. All other
ports have been left unchanged apart from a commonizing of attributes.
libphobos/ChangeLog:
2019-04-13 Iain Buclaw <ibuclaw@gdcproject.org>
* libdruntime/Makefile.am (DRUNTIME_CSOURCES): Remove bss_sections.c.
(DRUNTIME_DSOURCES): Rename rt/sections_* modules to gcc/sections/*.
* libdruntime/Makefile.in: Regenerate.
* libdruntime/gcc/sections/android.d: New file.
* libdruntime/gcc/sections/elf_shared.d: New file.
* libdruntime/gcc/sections/osx.d: New file.
* libdruntime/gcc/sections/package.d: New file.
* libdruntime/gcc/sections/solaris.d: New file.
* libdruntime/gcc/sections/win32.d: New file.
* libdruntime/gcc/sections/win64.d: New file.
* libdruntime/rt/bss_section.c: Remove.
* libdruntime/rt/sections.d: Publicly import gcc.sections.
* libdruntime/rt/sections_android.d: Remove.
* libdruntime/rt/sections_elf_shared.d: Remove.
* libdruntime/rt/sections_osx.d: Remove.
* libdruntime/rt/sections_solaris.d: Remove.
* libdruntime/rt/sections_win32.d: Remove.
* libdruntime/rt/sections_win64.d: Remove.
ian [Sat, 13 Apr 2019 01:03:55 +0000 (01:03 +0000)]
compiler: improve type handling for string concat ops on constants
Resolve a small problem with concatenation of string constants: in a
string concat X + Y where X has named type and Y has abstract string
type, insure that the result has X's type, and disable folding if the
both sides have a concrete type that does not match.
jakub [Fri, 12 Apr 2019 21:45:54 +0000 (21:45 +0000)]
PR c/89933
c/
* c-decl.c (merge_decls): When newdecl's type is its main variant,
don't try to remove it from the variant list, but instead assert
it has no variants.
cp/
* decl.c (duplicate_decls): When newdecl's type is its main variant,
don't try to remove it from the variant list, but instead assert
it has no variants.
testsuite/
* c-c++-common/pr89933.c: New test.
PR c/88383 - ICE calling __builtin_has_attribute on a reference
PR c/89288 - ICE in tree_code_size, at tree.c:865
PR c/89798 - excessive vector_size silently accepted and truncated
PR c/89797 - ICE on a vector_size (1LU << 33) int variable
gcc/ChangeLog:
PR c/89797
* targhooks.c (default_vector_alignment): Avoid assuming
argument fits in SHWI.
* tree.h (TYPE_VECTOR_SUBPARTS): Avoid sign overflow in
a shift expression.
* doc/extend.texi (__builtin_has_attribute): Add a clarifying note.
gcc/c-family/ChangeLog:
PR c/88383
PR c/89288
PR c/89798
PR c/89797
* c-attribs.c (type_valid_for_vector_size): Detect excessively
large sizes.
(validate_attribute): Handle DECLs and expressions.
(has_attribute): Handle types referenced by expressions.
Avoid considering array attributes in ARRAY_REF expressions .
jason [Fri, 12 Apr 2019 16:25:59 +0000 (16:25 +0000)]
Avoid ICE on pmf{} in template.
Now that we return the original CONSTRUCTOR from finish_compound_literal,
the call to null_member_pointer_value_p in tsubst_copy_and_build was getting
confused because the CONSTRUCTOR was still empty rather than a valid PMF
value.
* call.c (null_member_pointer_value_p): Handle an empty CONSTRUCTOR
of PMF type.
jakub [Fri, 12 Apr 2019 16:20:21 +0000 (16:20 +0000)]
PR rtl-optimization/89965
* dce.c: Include rtl-iter.h.
(struct check_argument_load_data): New type.
(check_argument_load): New function.
(find_call_stack_args): Check for loads from stack slots still tracked
in sp_bytes and punt if any is found.
jakub [Fri, 12 Apr 2019 16:12:49 +0000 (16:12 +0000)]
PR translation/90041
* exgettext: Print MissingArgError, UnknownError or Warn
*.opt argument using error or warning instead of _ to mark it
as gcc-internal-format.
* c.opt (-fhandle-exceptions): Use %< and %> around option names
in the Warn diagnostics.
* params.def (PARAM_MAX_LTO_STREAMING_PARALLELISM): New parameter.
* lto.c (do_stream_out): rename to ...
(stream_out): ... this one; move original code to ...
(stream_out_partitions_1, stream_out_partitions): ... these new
functions.
(lto_wpa_write_files): Honnor lto_parallelism
r269586 changed the format of some warning messages.
Each switch in the warning message is now surrounded by single quotes.
This commit updates the regex's in arm.exp dejagnu files to match the
new format and remove the extra 20+ FAILs on excess errors that are
introduced on certain variations (e.g.
arm-eabi-aem/-marm/-march=armv7-a/-mfpu=vfpv3-d16/-mfloat-abi=softfp).
Regtested arm.exp with cross-compiler arm-none-eabi
gcc/testsuite/ChangeLog:
2019-04-12 Matthew Malcomson <matthew.malcomson@arm.com>
* g++.target/arm/arm.exp: Change format of default prune regex.
* gcc.target/arm/arm.exp: Change format of default prune regex.
PR target/87532
* config/rs6000/rs6000.c (rs6000_split_vec_extract_var): Use inner
mode of vector rather than mode of destination for move instruction.
* config/rs6000/vsx.md (*vsx_extract_<mode>_<VS_scalar>mode_var):
Use QI inner mode with V16QI vector mode.
S/390: Fix a problem with the bswap vector pattern
arch13 introduced instructions to perform vector element-wise byte
swaps on the way from or to memory. For a byte swap between vector
registers the vector permute instruction is required which needs a
permute pattern to be loaded into a vector register first.
With the current implementation there is a potential problem when the
decision for the reg-reg variant is made very late. This patch is
supposed to fix that.
With the patch the required permute pattern is generated already in
the expander and attached to the bswap pattern as USE operand. The
predicate in the insn_and_split pattern accepts it although the
permute constant as such is not a valid constant. For the reg-reg
variant only the vector register constraint is used for the permute
constant forcing LRA to a) push the constant into literal pool and b)
load the literal pool constant into a vector register.
gcc/ChangeLog:
2019-04-12 Andreas Krebbel <krebbel@linux.ibm.com>
* config/s390/predicates.md (permute_pattern_operand): New
predicate.
* config/s390/vector.md ("*vec_splats_bswap_vec<mode>"): Add USE
operand for the permute pattern.
("*vec_perm<mode>"): New insn definition.
("bswap<mode>"): Generate the permute pattern operand in the
expander and perform the operand reloads for pre arch13 level
already.
("*bswap<mode>_emu"): Rename to ...
("*bswap<mode>"): ... this. And make the splitter vxe2 only.
* config/s390/vx-builtins.md ("*vec_insert_and_zero_bswap<mode>"):
Add the USE operand for the permute pattern.
("*vec_set_bswap_vec<mode>"): Likewise.
jakub [Fri, 12 Apr 2019 07:28:35 +0000 (07:28 +0000)]
PR c/89946
* varasm.c (assemble_start_function): Don't use tree_fits_uhwi_p
and gcc_unreachable if it fails, just call tree_to_uhwi which
verifies that too. Test TREE_CHAIN instead of list_length > 1.
Start warning message with a lower-case letter. Formatting fixes.
c-family/
* c-attribs.c (handle_patchable_function_entry_attribute): Add
function comment. Warn if arguments of the attribute are not positive
integer constants.
testsuite/
* c-c++-common/pr89946.c: New test.
jakub [Fri, 12 Apr 2019 07:27:25 +0000 (07:27 +0000)]
PR rtl-optimization/90026
* cfgcleanup.c (try_optimize_cfg): When removing empty bb with no
successors, look for BARRIERs inside of the whole BB_FOOTER chain
rather than just at the start of it. If e->src BB_FOOTER is not NULL
in cfglayout mode, use emit_barrier_after_bb.
The removal of the check programs in libphobos means that all tests
will now be compiled one at a time, causing a notable slowdown in
comparison with the previous method of testing.
libphobos/ChangeLog:
2019-04-12 Iain Buclaw <ibuclaw@gdcproject.org>
* testsuite/Makefile.am:
* testsuite/Makefile.am (AUTOMAKE_OPTIONS): Remove dejagnu.
(RUNTEST): Remove variable.
(RUNTESTDEFAULTFLAGS, check_p_subno, check_p_numbers0,
check_p_numbers1, check_p_numbers2, check_p_numbers3,
check_p_numbers4, check_p_numbers5, check_p_numbers6, check_p_numbers,
check_p_subdirs, check_DEJAGNU_libphobos_targets): New variables.
(site.exp, %/site.exp, check-DEJAGNU, check-am, clean-local): New
rules written so that all the *.exp files are ran parallelized.
(CLEANFILES): Add *.exe, *.o, and site.exp.
* testsuite/Makefile.in: Regenerate.
Currently, the druntime and phobos unittests are compiled as a separate
check program, then ran by the libphobos.unittest/unittest.exp script.
As PR d/89255 notes, this process lacks proper multilib handling.
As a first step, a new internal option that instructs the compiler to
put the reference to all unittest functions in another symbol has been
added. This will allow each module to be compiled separately as a
standalone program using dg-runtest without running into collisions in
the D runtime module registry.
gcc/d/ChangeLog:
2019-04-12 Iain Buclaw <ibuclaw@gdcproject.org>
* d-tree.h (DECL_IN_UNITTEST_CONDITION_P): Define.
* decl.cc (DeclVisitor): Add in_version_unittest_ field.
(DeclVisitor::visit(ConditionalDeclaration)): New override.
(DeclVisitor::visit(FuncDeclaration)): Set
DECL_IN_UNITTEST_CONDITION_P.
* lang.opt (-fbuilding-libphobos-tests): Add option.
* modules.cc (current_testing_module): New static variable.
(build_module_tree): Generate second moduleinfo symbol to hold
reference to unittests if flag_building_libphobos_tests.
(register_module_decl): Check DECL_IN_UNITTEST_CONDITION_P to decide
which moduleinfo the decl should be registered against.
Adds new virtual isVersionCondition, this is so that in the code
generation pass, a ConditionDeclaration's condition can be identified
without requiring a Visitor function.
This also replaces calls to __TBB_ASSERT so that there are two macro
definitions provided by c++config -
__PSTL_ASSERT(_Condition)
__PSTL_ASSERT_MSG(_Condition, _Message)
* include/bits/c++config:
Add definition for __PSTL_ASSERT.
Add definition for __PSTL_ASSERT_MSG.
* include/pstl/algorithm_impl.h: Replace use of assert().
* include/pstl/numeric_impl.h: Replace use of assert().
* include/pstl/parallel_backend_tbb.h:
Replace use of assert().
Replace use of __TBB_ASSERT().
* include/pstl/parallel_backend_utils.h: Replace use of assert().
PR translation/89939
* frontend-passes.c (B_ERROR): Delete macro.
(C_ERROR): Delete macro.
(B_ERROR_1): New macro.
(C_ERROR_1): New macro.
(C_ERROR_2): New macro.
(inline_matmul_assign): Use new macros.
(call_external_blas): Likewise.
PR libstdc++/90046 fix build failure on epiphany-elf
The epiphany-elf target aligns structs to 8 bytes, which causes the
static_assert(alignof(_Chunk) == 1) to fail.
Instead of requiring _Chunks to be positionable at any alignment, ensure
new buffers are aligned to alignof(_Chunk). Because the buffer size is a
power of two, we know that both the buffer size and sizeof(_Chunk) are
multiples of alignof(_Chunk). So is p is aligned to alignof(_Chunk) then
so is (p + size - sizeof(_Chunk)). So just ensure the new buffer is
aligned to at least alignof(_Chunk), which should already be true
because the caller requests at least alignof(max_align_t).
PR libstdc++/90046
* src/c++17/memory_resource.cc
(monotonic_buffer_resource::_Chunk::allocate): Increase alignment if
needed to allow placing a _Chunk at the end of the buffer.
(monotonic_buffer_resource::_M_new_buffer): Remove static_assert.
jakub [Thu, 11 Apr 2019 15:28:18 +0000 (15:28 +0000)]
PR translation/90035
* parser.h (struct cp_parser): Add
type_definition_forbidden_message_arg member.
* parser.c (cp_debug_parser): Print it.
(cp_parser_check_type_definition): Pass
parser->type_definition_forbidden_message_arg as second argument to
error.
(cp_parser_has_attribute_expression, cp_parser_sizeof_operand): Set
parser->type_definition_forbidden_message_arg and use G_() with
%qs for parser->type_definition_forbidden_message instead of
building untranslatable message using concat.
PR tree-optimization/90020
* tree-ssa-sccvn.c (vn_reference_may_trap): New function.
* tree-ssa-sccvn.h (vn_reference_may_trap): Declare.
* tree-ssa-pre.c (compute_avail): Use it to not put
possibly trapping references after a call that might not
return into EXP_GEN.
* gcse.c (compute_hash_table_work): Do not elide
marking a block containing a call if the call might not
return.
Update documentation regarding bogus memory leaks in libstdc++
* doc/xml/faq.xml: Add information about emergency EH pool.
* doc/xml/manual/debug.xml: Update list of memory debugging tools.
Move outdated information on mt_allocator to a separate section.
* doc/xml/manual/evolution.xml: Clarify that GLIBCXX_FORCE_NEW
doesn't affect the default allocator.
Fix a stack exhaustion bug in libiberty's demangler when decoding a pathalogically constructed mangled name.
PR 89394
* cp-demangle.c (cplus_demangle_fill_name): Reject negative
lengths.
(d_count_templates_scopes): Replace num_templates and num_scopes
parameters with a struct d_print_info pointer parameter. Adjust
body of the function accordingly. Add recursion counter and check
that the recursion limit is not reached.
(d_print_init): Pass dpi parameter to d_count_templates_scopes.
Reset recursion counter afterwards, unless the recursion limit was
reached.
jakub [Wed, 10 Apr 2019 07:27:20 +0000 (07:27 +0000)]
PR c++/90010
* gimple-ssa-sprintf.c (target_to_host): Fix handling of targstr
with strlen in between hostsz-3 and hostsz-1 inclusive when no
translation is needed, and when translation is needed, only append
... if the string length is hostsz or more bytes long. Avoid using
strncpy or strcat.
* include/std/variant: Adjust whitespace. Add comments.
(_Multi_array): Leave primary template undefined.
(_Multi_array<_Tp>): Define partial specialization for base case of
recursion.
(__gen_vtable_impl, __gen_vtable): Remove redundant && from type
which is always a reference.
(__gen_vtable::_S_apply()): Remove function, inline body into
default member initializer.
* testsuite/20_util/variant/visit.cc: Test with noncopyable types.
The __visitor_result_type helper didn't use std::invoke and so didn't
compile when the visitor was a pointer-to-member rather than a function
object. Use std::invoke_result instead.
* include/std/variant (__variant_idx_cookie): Add member type.
(__visitor_result_type): Remove.
(__do_visit): Use invoke_result instead of __visitor_result_type.
* testsuite/20_util/variant/visit.cc: New test.
The "*neon_mov<mode>" patterns for 128 bit sized quantities uses the "Dn"
constraint to match vmov.f32 and vmov.i<vec-width> patterns.
This constraint boils down to using the `neon_immediate_valid` function.
Once the constraint has matched, the output C statement asserts that function
passes.
The output C statement calls `neon_immediate_valid` with the mode taken from the
iterator, while the constraint takes the mode from the operand.
This can cause a discrepency when the operand is a CONST_INT, as the constraint
passes VOIDmode which `neon_immediate_valid` treats as DImode, while the C
statement passes the mode of the iterator which can be TImode.
When this happens, the `neon_immediate_valid` can fail in the second call (if
e.g. the CONST_INT is a valid immediate in DImode but not TImode) which would
trigger the assertion.
The testcase added with this patch triggers this when compiled with an arm cross
compiler using the command line below.
gcc -march=armv8-a -c neon-immediate-timode.c -O1 -mfloat-abi=hard -mfpu=neon-fp-armv8
This patch splits the original "Dn" constraint into three new constraints, "DN"
for TImode CONST_INT, "Dn" for DImode CONST_INT, and "Dm" for CONST_VECTOR.
Splitting things up this way requires using one extra alternative in the
"*neon_mov<mode>" patterns, but makes it clear from the constraint what mode is
being used.
We also remove the behaviour of treating VOIDmode as DImode in
`neon_valid_immediate` since the original "Dn" constraint was the only place
that functionality was used. VOIDmode is now never passed to that function.
An assertion has been added to the function to ensure this problem is caught
earlier on.
Bootstrapped on arm-none-linux-gnueabihf
Regtested on cross-compiler arm-none-eabi
gcc/ChangeLog:
2019-04-09 Matthew Malcomson <matthew.malcomson@arm.com>
PR target/90024
* config/arm/arm.c (neon_valid_immediate): Disallow VOIDmode parameter.
* config/arm/constraints.md (Dm, DN, Dn): Split previous Dn constraint
into three.
* config/arm/neon.md (*neon_mov<mode>): Account for TImode and DImode
differences directly.
(*smax<mode>3_neon, vashl<mode>3, vashr<mode>3_imm): Use Dm constraint.
gcc/testsuite/ChangeLog:
2019-04-09 Matthew Malcomson <matthew.malcomson@arm.com>
PR target/90024
* gcc.dg/torture/neon-immediate-timode.c: New test.
jakub [Tue, 9 Apr 2019 10:26:13 +0000 (10:26 +0000)]
PR tree-optimization/89998
* gimple-ssa-sprintf.c (try_substitute_return_value): Use lhs type
instead of integer_type_node if possible, don't add ranges if return
type is not compatible with int.
* gimple-fold.c (gimple_fold_builtin_sprintf,
gimple_fold_builtin_snprintf): Use lhs type instead of hardcoded
integer_type_node.
* gcc.c-torture/compile/pr89998-1.c: New test.
* gcc.c-torture/compile/pr89998-2.c: New test.
2019-04-08 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vect-data-refs.c (vect_get_smallest_scalar_type): Always
use gimple_expr_type for load and store calls. Skip over the
condition argument in a conditional internal function.
Protect use of TREE_INT_CST_LOW.
jakub [Tue, 9 Apr 2019 06:38:07 +0000 (06:38 +0000)]
PR target/90015
* config/riscv/riscv.c (riscv_get_interrupt_type): Fix comment typo.
(riscv_merge_decl_attributes): Fix typo in diagnostics. Remove
trailing period from it too.
ian [Tue, 9 Apr 2019 04:36:51 +0000 (04:36 +0000)]
compiler: sort packages in export data more deterministically
We can have multiple packages with the same name, so also sort by pkgpath.
To avoid an inconsistent sort, sort by symbol and pointer address if
we somehow get two different packages with the same name and pkgpath.
ville [Mon, 8 Apr 2019 19:45:48 +0000 (19:45 +0000)]
Fix visit<R> for variant.
* include/std/variant (__do_visit): Add a template parameter
for enforcing same return types for visit.
(__gen_vtable_impl): Likewise.
(_S_apply_single_alt): Adjust.
(__visit_invoke_impl): New. Handle casting to void.
(__do_visit_invoke): New. Enforces same return types.
(__do_visit_invoke_r): New. Converts return types.
(__visit_invoke): Adjust.
(__gen_vtable): Add a template parameter for enforcing
same return types for visit.
* testsuite/20_util/variant/visit_r.cc: Add a test for a visitor with
different return types.
* testsuite/20_util/variant/visit_neg.cc: New. Ensures that
visitors with different return types don't accidentally
compile with regular visitation.
The fma_forest, fma_root_node and func_fma_steering classes lack a
copy constructor. However, they contain pointers to allocated memory
so this omission can be regarded as poor style. We don't need to copy
such objects, so declare the copy constructor private to inhibit
accidental copying.
2019-04-08 Andrea Corallo <andrea.corallo@arm.com>
jakub [Mon, 8 Apr 2019 12:35:22 +0000 (12:35 +0000)]
PR rtl-optimization/89865
* config/i386/i386.md
(SWI12 peephole for mem {+,-,&,|,^}= x; mem != 0): Fix up operand
numbers not to clash with the additional operands[4].
(peepholes for mem {+,-,&,|,^}= x; mem != 0): New peephole2s
with extra register copy in the middle.
* gcc.target/i386/pr49095.c: Adjust number of expected RMW spots
on ia32.