Fix an ICE in decompose_normal_address(), which cannot handle Altivec AND:
addresses, by disallowing them via implementing the target hook
rs6000_cannot_substitute_mem_equiv_p.
gcc/
PR rtl-optimization/93974
* config/rs6000/rs6000.c (TARGET_CANNOT_SUBSTITUTE_MEM_EQUIV_P): Define.
(rs6000_cannot_substitute_mem_equiv_p): New function.
gcc/testsuite/
PR rtl-optimization/93974
* g++.dg/pr93974.C: New test.
For symmetric transfers to work with C++20 coroutines, it is
currently necessary to tail call the callee coroutine from resume
method of the caller coroutine. However there are several targets
which don't support an indirect tail call to an arbitrary callee.
Unfortunately, the target 'function_ok_for_sibcall' is not usable
from the front end in all cases. While it is possible to add a new
hook to cover this circumstance, it is too late in the release
cycle to be sure of getting the setting correct for all targets.
So, this patch backs out the use of function_ok_for_sibcall () and
the mandate of CALL_EXPR_MUST_TAIL_CALL from the symmetric
transfer.
Targets that can make indirect tail calls to arbitrary callees will
still be able to make use of the symmetric transfer (without risking
overrunning the stack) for optimization levels >= 2.
The draft standard does not mandate unlimited symmetric transfers,
so removing this is a QOI issue (albeit an important one) rather
than a correctness one.
The test is moved and adjusted so that it can be opted into by any
target that supports the necessary tailcall.
gcc/cp/ChangeLog:
2020-04-16 Iain Sandoe <iain@sandoe.co.uk>
PR c++/94359
* coroutines.cc (build_actor_fn): Back out use of
targetm.function_ok_for_sibcall. Do not mark the resume
call as CALL_EXPR_MUST_TAIL_CALL.
gcc/testsuite/ChangeLog:
2020-04-16 Iain Sandoe <iain@sandoe.co.uk>
PR c++/94359
* g++.dg/coroutines/torture/symmetric-transfer-00-basic.C: Move..
* g++.dg/coroutines/symmetric-transfer-00-basic.C: ..here and
adjust to run at O2 for targets supporting the necessary tail
call.
The function (_Z4InitIN3VARIZ1qvEUlvE_EUlvE_EEbT_) that initializes
VAR has changed siganture. It was was weak / comdat [Linux] or
weak / global [Darwin] and now is text section local.
So: test that the symbol exists, is spelled the way intended by the
ABI mangling and is not weak (or global on Darwin).
gcc/testsuite/ChangeLog:
2020-04-16 Iain Sandoe <iain@sandoe.co.uk>
PR c++/94426
* g++.dg/cpp0x/lambda/pr94426-2.C: Adjust scan-asms to test
for the change on more platforms.
PR 93621 testcase makes redirect_call_stmt_to_callee wrongly assume
that a call statement needs redirecting but then rightly fails an
assert ensuring the call statement parameters have not already been
adjusted because they were already created adjusted as part of thunk
expansion.
The test fails because the decl in the call call statement is
different than the decl of the callee, because the latter was created
in save_inline_function_body. This patch adds a way to link these two
and detect the situation in redirect_call_stmt_to_callee.
2020-04-16 Martin Jambor <mjambor@suse.cz>
PR ipa/93621
* ipa-inline.h (ipa_saved_clone_sources): Declare.
* ipa-inline-transform.c (ipa_saved_clone_sources): New variable.
(save_inline_function_body): Link the new body holder with the
previous one.
* cgraph.c: Include ipa-inline.h.
(cgraph_edge::redirect_call_stmt_to_callee): Try to find the decl from
the statement in ipa_saved_clone_sources.
* cgraphunit.c: Include ipa-inline.h.
(expand_all_functions): Free ipa_saved_clone_sources.
testsuite/
* g++.dg/ipa/pr93621.C: New test.
For this testcase we ended up generating the invalid rtl:
(insn 10 9 11 2 (set (reg:VNx16BI 105)
(and:VNx16BI (xor:VNx16BI (reg:VNx8BI 103)
(reg:VNx16BI 104))
(reg:VNx16BI 104))) "/tmp/bar.c":9:12 -1
(nil))
Fixed by taking the VNx16BI lowpart. It's safe to do that here because
the gp (r104) masks out the extra odd-indexed bits.
2020-04-16 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR target/94606
* config/aarch64/aarch64.c (aarch64_expand_sve_const_pred_eor): Take
the VNx16BI lowpart of the recursively-generated constant.
gcc/testsuite/
PR target/94606
* gcc.dg/vect/pr94606.c: New test.
PR c++/94314
* cgraphclones.c (set_new_clone_decl_and_node_flags): Drop
DECL_IS_REPLACEABLE_OPERATOR during cloning.
* tree-ssa-dce.c (valid_new_delete_pair_p): New function.
(propagate_necessity): Check operator names.
PR c++/94314
* g++.dg/pr94314.C: Do not use dg-additional-options
and remove not needed stdio.h include.
* g++.dg/pr94314-2.C: Likewise.
* g++.dg/pr94314-3.C: Likewise.
* g++.dg/pr94314-4.C: New test.
Co-Authored-By: Jakub Jelinek <jakub@redhat.com>
Here we're ICE'ing in do_narrow during error-recovery, because ocp_convert
returns error_mark_node after it attempts to reduce a const decl to its
erroneous DECL_INITIAL via scalar_constant_value, and we later pass this
error_mark_node to fold_build2 which isn't prepared to handle error_mark_nodes.
We could fix this ICE in do_narrow by checking if ocp_convert returns
error_mark_node, but for the sake of consistency and for better error recovery
it seems it'd be preferable if ocp_convert didn't care that a const decl's
initializer is erroneous and would instead proceed as if the decl was not const,
which is the approach that this patch takes.
gcc/cp/ChangeLog:
PR c++/94475
* cvt.c (ocp_convert): If the result of scalar_constant_value is
erroneous, ignore it and use the original expression.
gcc/testsuite/ChangeLog:
PR c++/94475
* g++.dg/conversion/err-recover2.C: New test.
* g++.dg/diagnostic/pr84138.C: Remove now-bogus warning.
* g++.dg/warn/Wsign-compare-8.C: Remove now-bogus warning.
Hi,
in GCC 8 we changed -fprofile-generate=<path> to use mangled absolute paths in
the <path> directory. This was necessary to avoid clashes of files when gcc is
executed from different directories to build different sources of same
filename.
However this made it difficult to build projects on setups where instrumented
build is done in one directory, feedback build in different and possibly
training happens in yet another directory structure. This happens i.e. for
Firefox builds for month or two.
This patch adds -fprofile-prefix-path that can be used to inform gcc where the
root of build directory is and strip it form the gcda filenames.
This is similar to exisitng debug-prefix-map but without the map feature since
it seems useless for profile data.
We spent quite some time with Maritn Liska discussing options and found no
better solution. I was looking how this work on LLVM and they produce single
profdata file which is then transformed into kind of simple database by
llvmprofdata tool. This database keys functions by filename and symbol name.
If you arrane two files with same name define static variable with same symbol
name this gets messedup and result in wrong info. So I think this is not very
good solution and preffer the extra option.
Bootstrapped/regtested x86_64-linux. I plan to commit it later today if there
are no complains.
I suppose our manual could have some central section on profile feedback
explaining the whole setup at one place.
Honza
PR gcov-profile/93401
* common.opt (profile-prefix-path): New option.
* coverae.c: Include diagnostics.h.
(coverage_init): Strip profile prefix path.
* doc/invoke.texi (-fprofile-prefix-path): Document.
early-remat.c:process_block wasn't handling insns that set multiple
candidate registers, which led to an assertion failure at the end
of the main loop.
Instructions that set two pseudos aren't rematerialisation candidates in
themselves, but we still need to track them if another instruction that
sets the same register is a rematerialisation candidate.
2020-04-16 Richard Sandiford <richard.sandiford@arm.com>
gcc/
PR rtl-optimization/94605
* early-remat.c (early_remat::process_block): Handle insns that
set multiple candidate registers.
gcc/testsuite/
PR rtl-optimization/94605
* gcc.target/aarch64/sve/pr94605.c: New test.
This removes { dg-final { scan-tree-dump "tiled" "graphite" } } scans
from graphite tests that pass/fail dependent on the ISL version used.
Note all scans match the actually dumped "not tiled" messages with
ISL 0.12 and not the alternative "tiled by <number>". With ISL
0.22 neither is printed because the tiling infrastructure doesn't
expect the new schedule layout (though looking at a few cases it
looks superior with more dimensions marked as permutable.
Anyway, the scans have nothing to do with interchange and just
add to testsuite noise.
2020-04-16 Richard Biener <rguenther@suse.de>
* gcc.dg/graphite/interchange-1.c: Remove scan for tiled.
* gcc.dg/graphite/interchange-10.c: Likewise.
* gcc.dg/graphite/interchange-11.c: Likewise.
* gcc.dg/graphite/interchange-3.c: Likewise.
* gcc.dg/graphite/interchange-4.c: Likewise.
* gcc.dg/graphite/interchange-7.c: Likewise.
* gcc.dg/graphite/interchange-9.c: Likewise.
* gcc.dg/graphite/uns-interchange-9.c: Likewise.
* gfortran.dg/graphite/interchange-3.f90: Likewise.
This adjusts emit_move_multi_word to handle moves into paradoxical
subregs parts that are not there and adjusts lower-subregs
CLOBBER resolving to deal with those as well.
2020-04-16 Richard Biener <rguenther@suse.de>
PR middle-end/94614
* expr.c (emit_move_multi_word): Do not generate code when
the destination part is undefined_operand_subword_p.
* lower-subreg.c (resolve_clobber): Look through a paradoxica
subreg.
As Iain reported, my change broke the case when one has bison >= 3,
but make decides there is no reason to regenerate plural.c, unfortunately
that seems to be a scenario I haven't tested. The problem is that
the pregenerated plural.c has been generated with bison 1.35, but when
config.h says HAVE_BISON3, the code assumes it is the bison3 variant.
What used to work fine is when one has bison >= 3 and plural.c has been
regenerated (e.g. do touch intl/plural.y and it will work), or when
one doesn't have any bison (then nothing is regenerated, but HAVE_BISON3
isn't defined either), or when one has bison < 3 and doesn't need to
regenerate, or when one has bison < 3 and it is regenerated.
The following patch fixes this, by killing the HAVE_BISON3 macro from
config.h, and instead remembering the fact whether plural.c has been created
with bison < 3 or bison >= 3 in a separate new plural-config.h header.
The way this works:
- user doesn't have bison
- user has bison >= 3, but intl/{plural-config.h,plural.c} aren't older than intl/plural.y
- user has bison < 3, but intl/{plural-config.h,plural.c} aren't older than intl/plural.y
pregenerated !USE_BISON3 plural.c and plural-config.h from source
dir is used, nothing in the objdir
- user has bison >= 3 and intl/plural.y is newer
Makefile generates plural.c and USE_BISON3 plural-config.h in the
objdir, which is then used in preference to srcdir copies
- user has bison < 3 and intl/plural.y is newer
Makefile generates plural.c and !USE_BISON3 plural-config.h in the
objdir, which is then used in preference to srcdir copies
I have tested all these cases and make all-yes worked in all the cases.
If one uses the unsupported ./configure where srcdir == objdir, I guess
(though haven't tested) that it should still work, just it would be nice
if such people didn't try to check in the plural{.c,-config.h} they have
regenerated.
What doesn't work, but didn't work before either (just tested gcc-9 branch
too) is when one doesn't have bison and plural.y is newer than plural.c.
Don't do that ;)
2020-04-16 Jakub Jelinek <jakub@redhat.com>
PR bootstrap/92008
intl/
* configure.ac: Remove HAVE_BISON3 AC_DEFINE.
* Makefile.in (HEADERS): Add plural-config.h.
(.y.c): Also create plural-config.h.
(dcigettext.o loadmsgcat.o plural.o plural-exp.o): Also depend
on plural-config.h.
(plural-config.h): Depend on plural.c.
* plural-exp.h: Include plural-config.h. Use USE_BISON3 instead
of HAVE_BISON3.
* plural.y: Use USE_BISON3 instead of HAVE_BISON3.
* configure: Regenerated.
* plural.c: Regenerated.
* config.h.in: Regenerated.
* plural-config.h: Generated.
contrib/
* gcc_update: Add intl/plural.y dependency for intl/plural-config.h.
get_ref_base_and_extent recognizes ARRAY_REFs with variable index but
into arrays of length one as constant offset accesses. However,
max_size in such cases is extended to span the whole element. This
confuses SRA verification when SRA also builds its (total
scalarization) access structures to describe fields under such array -
get_ref_base_and_extent returns different size and max_size for them.
Fixed by not performing the check for total scalarization accesses.
The subsequent check then had to be changed to use size and not
max_size too, which meant it has to be skipped when the access
structure describes a genuine variable array access.
Bootstrapped and tested on x86_64-linux.
2020-04-16 Martin Jambor <mjambor@suse.cz>
PR tree-optimization/94598
* tree-sra.c (verify_sra_access_forest): Fix verification of total
scalarization accesses under access to one-element arrays.
testsuite/
* gcc.dg/tree-ssa/pr94598.c: New test.
This test got committed into a spot where nothing actually tests it.
As there is no main, I assume it was meant to be gcc.c-torture/compile/
test and the test PASSes after moving there (both x86_64-linux and
i686-linux). Though, it passed before the PR92372 fixes too.
2020-04-16 Jakub Jelinek <jakub@redhat.com>
PR ipa/92372
* gcc.c-torture/pr92372.c: Move ...
* gcc.c-torture/compile/pr92372.c: ... here.
GCC 4.2 (but I think not the latest tip of GCC 4.2 branch) has broken value
initialization, see PR33916. The following patch provides a workaround for
that. Tested with GCC 4.2 on a reduced testcase I've distilled from the
assign_param_data_one class which has been miscompiled the same.
2020-04-16 Jakub Jelinek <jakub@redhat.com>
PR bootstrap/89494
* function.c (assign_parm_find_data_types): Add workaround for
BROKEN_VALUE_INITIALIZATION compilers.
bison 3 apparently made a backwards incompatible change, dropped
YYLEX_PARAM/YYPARSE_PARAM support and instead needs %param or %lex-param
and %parse-param. Furthermore, there is no easy way to conditionalize
on bison version in the *.y files.
While e.g. glibc bumped bison requirement and just has the bison 3
compatible version, Richi said there are still systems with older bison
where we want to build gcc.
So, this patch instead determines during configure bison version, and
depending on that when building plural.c (if building it at all) tweaks
what is passed over to bison if needed.
Tested with both bison 3 and bison 1.35, in each case with reconfiguring
intl and building with make all-yes (as in my setup intl isn't normally
used).
2020-04-16 Jakub Jelinek <jakub@redhat.com>
PR bootstrap/92008
* configure.ac: Add check for bison >= 3, AC_DEFINE HAVE_BISON3
and AC_SUBST BISON3_YES and BISON3_NO.
* Makefile.in (.y.c): Prefix $(YACC) invocation with @BISON3_NO@,
add @BISON3_YES@ prefixed rule to adjust the *.y source using sed
and adjust output afterwards.
* plural-exp.h (PLURAL_PARSE): If HAVE_BISON3 is defined, use
struct parse_args * type for arg instead of void *.
* plural.y: Add magic /* BISON3 ... */ comments with bison >= 3
directives.
(YYLEX_PARAM, YYPARSE_PARAM): Don't define if HAVE_BISON3 is defined.
(yylex, yyerror): Adjust prototypes and definitions if HAVE_BISON3
is defined.
* plural.c: Regenerated.
* config.h.in: Regenerated.
* configure: Regenerated.
This adds the SSA name version to the gdb pretty-printing of SSA names.
(gdb) p (tree)$1
$5 = <ssa_name 0x7ffff68435a0 323>
2020-04-16 Richard Biener <rguenther@suse.de>
* gdbhooks.py (TreePrinter): Print SSA_NAME_VERSION of SSA_NAME
nodes.
This snippet has been copied from the non-structured binding declaration
parsing later in the function, and while for non-structured bindings
it can be followed by comma or semicolon, structured bindings may be
only followed by semicolon.
Or, do we want to have a different message for the case when there is
a comma (and keep this corrected one only if there is something else)
that would explain better what is the bug (or add a fix-it hint)?
Marek said in the PR that clang++ reports
error: decomposition declaration must be the only declaration in its group
There is another thing Marek noted (though, something for different spot),
that diagnostic for auto x(1), [e,f] = test2; could also use a clearer
wording like the above (or a fix-it hint), but the question is if we should
assume [ after , as a structured binding or if we should do some tentative
parsing first to figure out if it looks like a structured binding.
2020-04-16 Jakub Jelinek <jakub@redhat.com>
PR c++/94571
* parser.c (cp_parser_simple_declaration): Fix up a pasto in
diagnostics.
* g++.dg/cpp1z/decomp51.C: New test.
PR go/94607
* go-gcc.cc (class Gcc_backend): Define builtin_const,
builtin_noreturn, builtin_novops.
(Gcc_backend::define_builtin): Change const_p and noreturn_p
parameters to a single flags parameter. Change all callers.
(Gcc_backend::Gcc_backend): Pass novops for prefetch.
Some more C++20 changes from P1614R2, "The Mothership has Landed".
This removes all redundant equality and inequality operators in the
Utilities clause, as they can be synthesized from the remaining equality
operators.
It also removes the single redundant operator in the Localization
clause, because it didn't seem worth doing in a separate commit.
* include/bits/allocator.h (operator!=): Do not define for C++20.
* include/bits/locale_classes.h (operator!=): Likewise.
* include/bits/std_function.h (operator==(nullptr_t, const function&))
(operator!=(const function&, nullptr_t))
(operator!=(nullptr_t, const function&)): Likewise.
* include/ext/bitmap_allocator.h (operator!=): Likewise.
* include/ext/debug_allocator.h (operator!=): Likewise.
* include/ext/extptr_allocator.h (operator!=): Likewise.
* include/ext/malloc_allocator.h (operator!=): Likewise.
* include/ext/mt_allocator.h (operator!=): Likewise.
* include/ext/new_allocator.h (operator!=): Likewise.
* include/ext/pool_allocator.h (operator!=): Likewise.
* include/ext/throw_allocator.h (operator!=): Likewise.
* include/std/bitset (bitset::operator!=): Likewise.
* include/std/memory_resource (operator!=): Likewise.
* include/std/scoped_allocator (operator!=): Likewise.
Another C++20 change from P1614R2, "The Mothership has Landed".
* include/std/typeindex (operator<=>): Define for C++20.
* testsuite/20_util/typeindex/comparison_operators_c++20.cc: New test.
GCC 4.2 due to PR33916 miscompiles temp_dump_context ctor, because it doesn't
zero initialize the whole dump_context temporary on which it runs the static
get method and during destruction of the temporary an uninitialized pointer
is deleted.
More recent GCC versions properly zero initialize it and ideally optimize away
the construction/destruction of the temporary, as it isn't used for anything,
but there is no reason to create the temporary, static member functions can
be called without an associated object.
2020-04-15 Gustavo Romero <gromero@linux.ibm.com>
PR bootstrap/89494
* dumpfile.c (selftest::temp_dump_context::temp_dump_context):
Don't construct a dump_context temporary to call static method.
2020-04-15 Andrea Corallo <andrea.corallo@arm.com>
* config/aarch64/falkor-tag-collision-avoidance.c
(valid_src_p): Check for aarch64_address_info type before
accessing base field.
This patch fixes an ICE we were seeing due to a missing vec_duplicate pattern.
gcc/ChangeLog:
2020-04-15 Andre Vieira <andre.simoesdiasvieira@arm.com>
* config/arm/mve.md (mve_vec_duplicate<mode>): New pattern.
(V_sz_elem2): Remove unused mode attribute.
gcc/testsuite/ChangeLog:
2020-04-15 Andre Vieira <andre.simoesdiasvieira@arm.com>
Srinath Parvathaneni <srinath.parvathaneni@arm.com>
* gcc.target/arm/mve/intrinsics/mve_vec_duplicate.c: New test.
Without disabling this, the pattern can try and move DImode values
between floating point registers and general registers.
The constraints on this pattern can't handle that, and reload goes into
an infinite loop.
This was the cause of a testsuite failure in cde_v_1_mve.c, which is now
gone.
A DImode move for MVE now uses the `movdi_vfp` pattern, which is the same
pattern used for such a move when MVE is not available but the target has
TARGET_HARD_FLOAT.
Testing done:
Bootstrapped and regtested on arm-none-linux-gnueabihf
regtested on arm-none-eabi
gcc/ChangeLog:
2020-04-15 Matthew Malcomson <matthew.malcomson@arm.com>
* config/arm/arm.md (arm_movdi): Disallow for MVE.
As reported in the PR, GCC 10 (and also 9.3.1 but not 9.3.0) fails to build
when using older binutils which lack LSE support, because those instructions
are used in libgcc.
Thanks to Kyrylo's hint, the following patches (hopefully) allow it to build
even with older binutils by using .inst directive if LSE support isn't
available in the assembler.
2020-04-15 Jakub Jelinek <jakub@redhat.com>
PR target/93053
* configure.ac (LIBGCC_CHECK_AS_LSE): Add HAVE_AS_LSE checking.
* config/aarch64/lse.S: Include auto-target.h, if HAVE_AS_LSE
is not defined, use just .arch armv8-a.
(B, M, N, OPN): Define.
(COMMENT): New .macro.
(CAS, CASP, SWP, LDOP): Use .inst directive if HAVE_AS_LSE is not
defined. Otherwise, move the operands right after the glue? and
comment out operands where the macros are used.
* configure: Regenerated.
* config.in: Regenerated.
This change started with a bugreport about a typo in one requires testcase
(diagnosed with -Wunknown-pragmas only), but following discussion lead to
noting that we do not diagnose restriction that requires directives in
C/C++ may only appear at file or namespace scope; and several our tests
violated that.
2020-04-15 Jakub Jelinek <jakub@redhat.com>
PR c/94593
* c-parser.c (c_parser_pragma) <case PRAGMA_OMP_REQUIRES>: Reject
requires directive when not at file scope.
* parser.c (cp_parser_pragma) <case PRAGMA_OMP_REQUIRES>: Reject
requires directive when not at file or namespace scope.
* c-c++-common/gomp/requires-1.c: Fix a typo, requries -> requires.
Move directives to file scope.
(i): Remove.
* c-c++-common/gomp/requires-2.c: Move directives to file scope.
(i, foo): Remove.
* c-c++-common/gomp/requires-4.c: Move directives to file scope.
* c-c++-common/gomp/atomic-19.c: Move requires directive to file scope.
* c-c++-common/gomp/atomic-20.c: Likewise.
* c-c++-common/gomp/atomic-21.c: Likewise.
* c-c++-common/gomp/atomic-22.c: Likewise.
* gcc.dg/gomp/requires-1.c: New test.
* g++.dg/gomp/requires-1.C: New test.
* g++.dg/gomp/requires-2.C: New test.
* g++.dg/gomp/atomic-18.C: Move requires directive to file scope.
This makes same_type_for_tbaa_p conservative in the same way
get_alias_set is about void * which we allow to alias all other
pointers.
2020-04-15 Richard Biener <rguenther@suse.de>
PR middle-end/94539
* tree-ssa-alias.c (same_type_for_tbaa): Defer to
alias_sets_conflict_p for pointers.
* gcc.dg/alias-14.c: Make dg-do run.
Patterns zero_extendhisi2, zero_extendqisi2 and extendhisi2_internal can
load value from memory, but they don't treat volatile memory correctly.
Add %v1 before load instructions to emit 'memw' instruction when
-mserialize-volatile is in effect.
2020-04-14 Max Filippov <jcmvbkbc@gmail.com>
gcc/
* config/xtensa/xtensa.md (zero_extendhisi2, zero_extendqisi2)
(extendhisi2_internal): Add %v1 before the load instructions.
gcc/testsuite/
* gcc.target/xtensa/pr94584.c: New test.
My "simplification" of std::compare_three_way's constraints in commit
f214ffb336d582a66149068a2a96b7fcf395b5de was incorrect, because
std::three_way_comparable_with imposes additional restrictions beyond
the <=> expression being valid.
* libsupc++/compare (compare_three_way): Fix constraint so that
BUILTIN-PTR-THREE-WAY does not require three_way_comparable_with.
* testsuite/18_support/comparisons/object/builtin-ptr-three-way.cc:
New test.
This also implements the proposed resolution to LWG issue 3247, so that
the ill-formed <=> expression with nullptr is not used.
PR libstdc++/94562
* include/bits/shared_ptr.h (operator<=>): Define for C++20.
* include/bits/shared_ptr_base.h (operator<=>): Likewise.
* include/bits/unique_ptr.h (operator<=>): Add inline specifier.
* testsuite/20_util/shared_ptr/comparison/cmp_c++20.cc: New test.
* testsuite/20_util/shared_ptr/comparison/less.cc: Do not expect
std::less<A*> to be used when comparing std::shared_ptr<A> objects in
C++20.
For symmetric transfers to work with C++20 coroutines, it is
currently necessary to tail call the callee coroutine from resume
method of the caller coroutine. The current codegen marks these
resume calls as "MUST_TAIL_CALL" to indicate that the tail call is
required for correctness.
Unfortunately, several targets have ABI constraints that prevent
an indirect tail-call, which results in the PRs compile error.
The change here tests the target sibcall hook for the resume
expression and only marks it as requiring a tail call if that's
supported.
This doesn't fix the underlying problem; that really a solution is
needed to allow the tail-calls (or equivalent) to take place - but
that will be deferred until next stage 1.
gcc/cp/ChangeLog:
2020-04-14 Iain Sandoe <iain@sandoe.co.uk>
PR c++/94359
* coroutines.cc (build_actor_fn): Check that the target can
support the resume tailcall before mandating it.
gcc/testsuite/ChangeLog:
2020-04-14 Iain Sandoe <iain@sandoe.co.uk>
PR c++/94359
* g++.dg/coroutines/torture/symmetric-transfer-00-basic.C:
Expect a run fail for targets without arbitrary indirect
tail-calls.
For rs6000 target, it is not valid to make PC-relative
references to TLS symbols. So addr_to_insn_form() needs
to check if things are TLS before returning PC-rel forms.
2020-04-14 Aaron Sawdey <acsawdey@linux.ibm.com>
PR target/94542
* config/rs6000/rs6000.c (address_to_insn_form): Do not attempt to
use PC-relative addressing for TLS references.
This fixes a garbled concepts diagnostic by moving the handling of DECLTYPE_TYPE
from pp_cxx_type_specifier_seq to cxx_pretty_printer::simple_type_specifier, a
move which also seems to be more consistent with the language grammar.
This patch also fixes pretty printing of rvalue reference types via
cxx_pretty_printer::type_id, which eventually calls pp_c_pointer which currently
doesn't distinguish between lvalue and rvalue references.
gcc/c-family/ChangeLog:
PR c++/85278
* c-pretty-print.c (pp_c_pointer) <case REFERENCE_TYPE>: Print a double
ampersand if it's an rvalue reference type.
gcc/cp/ChangeLog:
PR c++/85278
* cxx-pretty-print.c (cxx_pretty_printer:simple_type_specifier)
<case DECLTYPE_TYPE>: Handle DECLTYPE_TYPE here instead of ...
(pp_cxx_type_specifier_seq) <case DECLTYPE_TYPE>: ... here.
(cxx_pretty_printer::direct_abstract_declarator) <case DECLTYPE_TYPE>:
New no-op case.
gcc/testsuite/ChangeLog:
PR c++/85278
* g++.dg/concepts/diagnostic9.C: New test.
IPA-SRA can segfault when processing a call to an internal function
because such calls do not have corresponding call graphs and should be
treated like memory accesses anyway, which what the patch below does.
The patch makes an attempt to differentiate between reads and writes,
although for things passed by value it does not make any difference.
It treats all arguments of functions denoted with internal_store_fn_p
as written to, even though in practice only some are, but for IPA-SRA
purposes, actions needed to be taken when processing a read are also
always performed when analyzing a write, so the code is just slightly
pessimistic. But not as pessimistic as bailing out on any internal
call immediately.
I have LTO bootstrapped and tested the patch on x86_64-linux and
normally bootstrapped and tested it on aarch64-linux, although one
which is not SVE capable. I would appreciate testing on such machine
too - as well as a small testcase that would follow all relevant
conventions in gcc.target/aarch64/sve.
2020-04-14 Martin Jambor <mjambor@suse.cz>
PR ipa/94434
* ipa-sra.c: Include internal-fn.h.
(enum isra_scan_context): Update comment.
(scan_function): Treat calls to internal_functions like loads or stores.
When evaluating the initializer of 'a' in the following example
struct A {
A() = default; A(const A&);
A *p = this;
};
constexpr A foo() { return {}; }
constexpr A a = foo();
the PLACEHOLDER_EXPR for 'this' in the aggregate initializer returned by foo
gets resolved to the RESULT_DECL of foo. But due to guaranteed RVO, the 'this'
should really be resolved to '&a'.
Fixing this properly by immediately resolving 'this' and PLACEHOLDER_EXPRs to
the ultimate object under construction would in general mean that we would no
longer be able to cache constexpr calls for which RVO possibly applies, because
the result of the call may now depend on the ultimate object under construction.
So as a mostly correct stopgap solution that retains cachability of RVO'd
constexpr calls, this patch fixes this issue by rewriting all occurrences of the
RESULT_DECL in the result of a constexpr function call with the current object
under construction, after the call returns. This means the 'this' pointer
during construction of the temporary will still point to the temporary object
instead of the ultimate object, but besides that this approach seems
functionally equivalent to the proper approach.
gcc/cp/ChangeLog:
PR c++/94034
* constexpr.c (replace_result_decl_data): New struct.
(replace_result_decl_data_r): New function.
(replace_result_decl): New function.
(cxx_eval_call_expression): Use it.
* tree.c (build_aggr_init_expr): Set the location of the AGGR_INIT_EXPR
to that of its initializer.
gcc/testsuite/ChangeLog:
PR c++/94034
* g++.dg/cpp0x/constexpr-empty15.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi6a.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi6b.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi7a.C: New test.
* g++.dg/cpp1y/constexpr-nsdmi7b.C: New test.
In this PR the testcase ICEs because a BIT_INSERT_EXPR whose replaced bits are
not fully inside the container is generated. A size check is added to avoid
this kind of ICE.
gcc/ChangeLog:
PR tree-optimization/94574
* tree-ssa.c (non_rewritable_lvalue_p): Add size check when analyzing
whether a vector-insert is rewritable using a BIT_INSERT_EXPR.
gcc/testsuite/ChangeLog:
PR tree-optimization/94574
* gcc.dg/pr94574.c: New test.
Even though ix86_hard_regno_mode_ok doesn't allow xmm16-xmm31 nor
ymm16-ymm31 in 128/256 bit modes when AVX512VL is disabled, LRA can
still generate reg to reg moves with xmm16-xmm31 and ymm16-ymm31 in
128/256 bit modes. Remove mode size check in ix86_get_ssemov.
gcc/
PR target/94561
* config/i386/i386.c (ix86_get_ssemov): Remove mode size check.
gcc/testsuite/
PR target/94561
* gcc.target/i386/pr94561.c: New test.
