The grammar for a lambda context can include <source-name> 'M', and we
were adding the component that generated to the substitution table
twice. Just ignore the 'M' completely -- we'll already have done the
checks we need when we saw its predecessor. A prefix cannot be the
last component of a nested name, so we do not need to check for that
case (although we could if we wanted to be more lenient).
libiberty/
* cp-demangle.c (d_prefix): 'M' components are not
(re-)added to the substitution table.
* testsuite/demangle-expected: Add tests.
Richard Biener [Tue, 11 Oct 2022 09:34:55 +0000 (11:34 +0200)]
tree-optimization/107212 - SLP reduction of reduction paths
The following fixes an issue with how we handle epilogue generation
for SLP reductions of reduction paths where the actual live lanes
are not "canonical". We need to make sure to identify all live
lanes as reductions and thus have to iterate over all participating
SLP lanes when walking the reduction SSA use-def chain. Also the
previous attempt likely to mitigate such issue in
vectorizable_live_operation is misguided and has to be removed.
PR tree-optimization/107212
* tree-vect-loop.cc (vectorizable_reduction): Make sure to
set STMT_VINFO_REDUC_DEF for all live lanes in a SLP
reduction.
(vectorizable_live_operation): Do not pun to the SLP
node representative for reduction epilogue generation.
* gcc.dg/vect/pr107212-1.c: New testcase.
* gcc.dg/vect/pr107212-2.c: Likewise.
Andrew Stubbs [Mon, 29 Jun 2020 14:20:09 +0000 (15:20 +0100)]
amdgcn: Add vec_extract for partial vectors
Add vec_extract expanders for all valid pairs of vector types.
gcc/ChangeLog:
* config/gcn/gcn-protos.h (get_exec): Add prototypes for two variants.
* config/gcn/gcn-valu.md
(vec_extract<V_ALL:mode><V_ALL_ALT:mode>): New define_expand.
* config/gcn/gcn.cc (get_exec): Export the existing function. Add a
new overload variant.
Andrew Stubbs [Thu, 26 Mar 2020 21:22:45 +0000 (21:22 +0000)]
amdgcn: Resolve insn conditions at compile time
GET_MODE_NUNITS isn't a compile time constant, so we end up with many
impossible insns in the machine description. Adding MODE_VF allows the insns
to be eliminated completely.
gcc/ChangeLog:
* config/gcn/gcn-valu.md
(<cvt_name><VCVT_MODE:mode><VCVT_FMODE:mode>2<exec>): Use MODE_VF.
(<cvt_name><VCVT_FMODE:mode><VCVT_IMODE:mode>2<exec>): Likewise.
* config/gcn/gcn.h (MODE_VF): New macro.
Andre Vieira [Tue, 11 Oct 2022 09:49:27 +0000 (10:49 +0100)]
vect: Teach vectorizer how to handle bitfield accesses
gcc/ChangeLog:
* tree-if-conv.cc (if_convertible_loop_p_1): Move ordering of loop bb's from
here...
(tree_if_conversion): ... to here. Also call bitfield lowering when
appropriate.
(version_loop_for_if_conversion): Adapt to enable loop versioning when we only
need to lower bitfields.
(ifcvt_split_critical_edges): Relax condition of expected loop form as this is
checked earlier.
(get_bitfield_rep): New function.
(lower_bitfield): Likewise.
(bitfields_to_lower_p): Likewise.
(need_to_lower_bitfields): New global boolean.
(need_to_ifcvt): Likewise.
* tree-vect-data-refs.cc (vect_find_stmt_data_reference): Improve diagnostic
message.
* tree-vect-patterns.cc (vect_recog_temp_ssa_var): Add default value for last
parameter.
(vect_recog_bitfield_ref_pattern): New.
(vect_recog_bit_insert_pattern): New.
gcc/testsuite/ChangeLog:
* gcc.dg/vect/vect-bitfield-read-1.c: New test.
* gcc.dg/vect/vect-bitfield-read-2.c: New test.
* gcc.dg/vect/vect-bitfield-read-3.c: New test.
* gcc.dg/vect/vect-bitfield-read-4.c: New test.
* gcc.dg/vect/vect-bitfield-read-5.c: New test.
* gcc.dg/vect/vect-bitfield-read-6.c: New test.
* gcc.dg/vect/vect-bitfield-write-1.c: New test.
* gcc.dg/vect/vect-bitfield-write-2.c: New test.
* gcc.dg/vect/vect-bitfield-write-3.c: New test.
* gcc.dg/vect/vect-bitfield-write-4.c: New test.
* gcc.dg/vect/vect-bitfield-write-5.c: New test.
liuhongt [Mon, 10 Oct 2022 03:31:48 +0000 (11:31 +0800)]
Add define_insn_and_split to support general version of "kxnor".
For genereal_reg_operand, it will be splitted into xor + not.
For mask_reg_operand, it will be splitted with UNSPEC_MASK_OP just
like what we did for other logic operations.
The patch will optimize xor+not to kxnor when possible.
gcc/ChangeLog:
PR target/107093
* config/i386/i386.md (*notxor<mode>_1): New post_reload
define_insn_and_split.
(*notxorqi_1): Ditto.
Aldy Hernandez [Mon, 10 Oct 2022 18:42:10 +0000 (20:42 +0200)]
[PR107195] Set range to zero when nonzero mask is 0.
When solving 0 = _15 & 1, we calculate _15 as:
[irange] int [-INF, -2][0, +INF] NONZERO 0xfffffffe
The known value of _15 is [0, 1] NONZERO 0x1 which is intersected with
the above, yielding:
[0, 1] NONZERO 0x0
This eventually gets copied to a _Bool [0, 1] NONZERO 0x0.
This is problematic because here we have a bool which is zero, but
returns false for irange::zero_p, since the latter does not look at
nonzero bits. This causes logical_combine to assume the range is
not-zero, and all hell breaks loose.
I think we should just normalize a nonzero mask of 0 to [0, 0] at
creation, thus avoiding all this.
PR tree-optimization/107195
gcc/ChangeLog:
* value-range.cc (irange::set_range_from_nonzero_bits): Set range
to [0,0] when nonzero mask is 0.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/pr107195-1.c: New test.
* gcc.dg/tree-ssa/pr107195-2.c: New test.
Olivier Hainque [Fri, 3 Dec 2021 17:50:56 +0000 (17:50 +0000)]
Generic configury support for shared libs on VxWorks
This change adds the configury bits to activate the build of
shared libs on VxWorks ports configured with --enable-shared,
for libraries variants where this is generally supported (rtp,
code model !large - currently not compatible with -fPIC).
Set lt_cv_deplibs_check_method in libtool.m4, so the build of
libraries know how to establish dependencies. This is useful in
configurations such as aarch64 where proper support of LSE relies
on accurate dependency information between libstdc++ and libgcc_s
to begin with.
Regenerate configure scripts to reflect libtool.m4 change.
2022-10-09 Olivier Hainque <hainque@adacore.com>
* libtool.m4 (*vxworks*): When enable_shared, set dynamic_linker
and friends for rtp !large. Assume the linker has the required
abilities and set lt_cv_deplibs_check_method.
gcc/
* config.gcc (*vxworks*): Add t-slibgcc fragment
if enable_shared.
libgcc/
* config.host (*vxworks*): When enable_shared, add
libgcc and crtstuff "shared" fragments for rtp except
large code model.
(aarch64*-wrs-vxworks7*): Remove t-slibgcc-libgcc from
the list of fragments.
Olivier Hainque [Fri, 18 Feb 2022 22:44:53 +0000 (22:44 +0000)]
Tigthen the addition of -lgcc_eh to vxworks_libgcc_spec
This change refines VXWORKS_LIBGCC_SPEC wrt the inclusion
of -lgcc_eh.
Unless the compiler features support for dual sjlj and
table based eh, libgcc_eh.a is available only with multilib
variants for which we build a shared lib (mrtp on VxWorks).
Rework logic to handle absence of libgcc_s
for -mrtp -mcmodel=large, using a conditional expr kind of
spec.
The gthread support in libgcc_eh might resort to libgcc
functions on some targets, e.g. cas synchronisation routines
on aarch64. Arrange to append -lgcc also after -lgcc_eh
in VXWORKS_LIBGCC_SPEC.
2022-10-09 Olivier Hainque <hainque@adacore.com>
gcc/
* config/vxworks.h (VX_LGCC_EH_SO0, VX_LGCC_EH_SO1): New
internal macros.
(VXWORKS_LIBGCC_SPEC): Use them and document.
Martin Liska [Tue, 11 Oct 2022 06:52:28 +0000 (08:52 +0200)]
ranger: add override keyword
Fixes the following clang warning:
gcc/gimple-range-op.cc:310:16: warning: 'fold_range' overrides a member function but is not marked 'override' [-Winconsistent-missing-override]
Eugene Rozenfeld [Mon, 10 Oct 2022 21:10:31 +0000 (14:10 -0700)]
Fix PR107193.
The bug was introduced in f30e9fd33e56a5a721346ea6140722e1b193db42.
A variable (cur_locus_e) was incorrectly declared inside a loop.
I also moved two other declarations (last and locus) down to make
the code more clear.
Tested on x86_64-pc-linux-gnu.
gcc/ChangeLog:
PR debug/107193
* tree-cfg.cc (assign_discriminators): Move declaration of cur_locus_e
out of the loop.
Jeff Law [Tue, 11 Oct 2022 04:44:26 +0000 (00:44 -0400)]
[PR rtl-optimization/107182] Clear EDGE_CROSSING for jump->ret optimization
When turning a jump to a return into a return, we need to clear EDGE_CROSSING
of the fallthru edge to prevent a checking failure.
I considered not applying the transformation when the edge has EDGE_CROSSING
set, but it still seems like we ought to eliminate the unnecessary jump in
that case.
gcc/
PR rtl-optimization/107182
* cfgrtl.cc (fixup_reorder_chain): When optimizing a jump to a
return, clear EDGE_CROSSING on the appropriate edge.
Nathan Sidwell [Mon, 10 Oct 2022 21:55:04 +0000 (17:55 -0400)]
c++: Lambda context mangling
VAR and FIELD decls can become part of a lambda context, when the
lambda is 'attached' to that entity (It's a C++20 ODR thing that was
discovered with modules, but is actually separate.) We were not
marking those decls as substitution candidates, leading to demangling
failures and variance from other compilers.
This patch bumps the ABI, and adds the contexts them to the
substitution table. This is the intent of the ABI.
The compiler neglected to notice that a conversion from a string
constant to a string type was a valid string constant.
No test case because this only caused a compiler failure when compiling
without optimization, which is not the normal case, and is not a case
that we test.
compiler: only build thunk struct type when it is needed
Instead of building the thunk struct type in the determine_types pass,
build it when we need it. That ensures that we are consistent in
determining whether an argument is constant.
We no longer need to add a field for a call to recover, as the
simplify_thunk_statements pass runs after the build_recover_thunks pass,
so the additional argument will already have been added to the call.
Marek Polacek [Thu, 29 Sep 2022 21:49:32 +0000 (17:49 -0400)]
c-family: ICE with [[gnu::nocf_check]] [PR106937]
When getting the name of an attribute, we ought to use
get_attribute_name, which handles both [[]] and __attribute__(())
forms. Failure to do so may result in an ICE, like here.
pp_c_attributes_display wasn't able to print the [[]] form of
attributes, so this patch teaches it to.
When printing a pointer to function with a standard attribute, the attribute
should be printed after the parameter-list. With this patch we print:
aka 'void (*)(int) [[gnu::nocf_check]]'
or, in C++ with noexcept:
aka 'void (*)(int) noexcept [[gnu::nocf_check]]'
pp_c_attributes has been unused since its introduction in r56273 so
this patch removes it.
PR c++/106937
gcc/c-family/ChangeLog:
* c-pretty-print.cc (pp_c_specifier_qualifier_list): Print only GNU
attributes here.
(c_pretty_printer::direct_abstract_declarator): Print the standard [[]]
attributes here.
(pp_c_attributes): Remove.
(pp_c_attributes_display): Print the [[]] form if appropriate. Use
get_attribute_name. Don't print a trailing space when printing the
[[]] form.
* c-pretty-print.h (pp_c_attributes): Remove.
gcc/cp/ChangeLog:
* error.cc: Include "attribs.h".
(dump_type_prefix): Print only GNU attributes here.
(dump_type_suffix): Print standard attributes here.
Andrea Corallo [Tue, 27 Sep 2022 14:20:28 +0000 (16:20 +0200)]
Don't ICE running selftests if errors were raised [PR99723]
Hi all
this is to address PR 99723.
In the PR GCC crashes as the initialization of common trees is not
performed as no compilation is happening, this is because we raise an
error earlier while processing the arch flags.
This patch changes the code to execute selftests only if no errors
where raised before.
Bootstrapped on aarch64, okay for trunk?
Best Regards
Andrea
2022-09-27 Andrea Corallo <andrea.corallo@arm.com>
PR other/99723
* toplev.cc (toplev::main): Don't run self tests in case of
previous error.
Kito Cheng [Fri, 7 Oct 2022 08:55:14 +0000 (16:55 +0800)]
RISC-V: Add riscv_vector.h wrapper in testsuite to prevent pull in stdint.h from C library
For RISC-V linux/glibc toolchain will got header file not found when including
stdint.h if multilib is not enabled, it because some header file will
try to include gnu/stubs-<ABI-NAME>.h from the system, however it only
generated when multilib enabled.
In order to prevent that, we introduce a wrapper for riscv_vector.h,
include stdint-gcc.h rather than the default stdint.h.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/base/riscv_vector.h: New.
Reported-by: Christoph Müllner <christoph.muellner@vrull.eu> Tested-by: Christoph Müllner <christoph.muellner@vrull.eu> Reviewed-by: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
Aldy Hernandez [Mon, 10 Oct 2022 11:43:16 +0000 (13:43 +0200)]
Add frange::maybe_isnan (bool sign).
It is useful to know if there's the possiblity of a NAN with a given
sign. This is to complement maybe_isnan(void) which returns TRUE for a
NAN of any sign.
A follow-up patch implementing ABS will make use of this.
Aldy Hernandez [Mon, 10 Oct 2022 11:43:56 +0000 (13:43 +0200)]
Return non-legacy ranges in range.h.
int_range<1> is a legacy range (think anti ranges, legacy VRP, etc).
There is a penalty for converting anything built with <1> to
non-legacy. Since most of the uses of these functions are now ranger,
we can save a miniscule amount of time by converting them to
non-legacy.
Jonathan Wakely [Fri, 7 Oct 2022 11:18:57 +0000 (12:18 +0100)]
libstdc++: std::make_signed_t<cv bool> should be ill-formed
Currently we only reject std::make_signed_t<bool> but not cv bool.
Similarly for std::make_unsigned_t<cv bool>.
As well as making those ill-formed, this adds a requires-clause to the
make_signed and make_unsigned primary templates. This makes
non-integral, non-enum cases fail immediately with a clear error, rather
than giving an error about __make_signed_selector<T, false, false> being
incomplete.
libstdc++-v3/ChangeLog:
* include/std/type_traits (make_signed, make_unsigned): Add
specializations for cv bool. Add requires-clause for C++20 to
improve diagnostics for non-integral, non-enum cases.
* testsuite/20_util/make_signed/requirements/typedefs_neg.cc:
Check cv bool.
* testsuite/20_util/make_unsigned/requirements/typedefs_neg.cc:
Likewise.
* testsuite/24_iterators/range_access/range_access_cpp20_neg.cc:
Adjust expected errors for C++20 and later.
* testsuite/lib/prune.exp: Prune "in requirements [with ...]"
lines from diagnostics.
Jakub Jelinek [Mon, 10 Oct 2022 07:34:35 +0000 (09:34 +0200)]
openmp, fortran: Fix up IFN_ASSUME call
Like in other spots in trans-openmp.cc that create a TARGET_EXPR, the
slot has to be created with create_tmp_var_raw, because gfc_create_var
adds the var to BLOCK_VARS and that ICEs during expansion because
gimple_add_tmp_var_fn has:
gcc_assert (!DECL_CHAIN (tmp) && !DECL_SEEN_IN_BIND_EXPR_P (tmp));
assertion. Also, both C/C++ ensure the argument to IFN_ASSUME has
boolean_type_node, it is easier if Fortran does that too.
2022-10-10 Jakub Jelinek <jakub@redhat.com>
* trans-openmp.cc (gfc_trans_omp_assume): Use create_tmp_var_raw
instead of gfc_create_var for TARGET_EXPR slot creation. Create it
with boolean_type_node and convert.
arc: Use negative enter pattern instruction's offsets
The enter pattern instruction contains the necessary information for
the dwarf machinery to generate the appropriate dwarf code. This
patch is fixing the register offsets related to CFA, and adds a test.
gcc/
* config/arc/arc.cc (arc_save_callee_enter): Use negative offsets.
gcc/testsuite
* gcc.target/arc/enter-dw2-1.c: New file.
Manually expanding into 32-bit comparisons is much more efficient than
the default expansion into word-size comparisons. Note that word for PRU
is 8-bit.
PR target/106562
gcc/ChangeLog:
* config/pru/pru-protos.h (pru_noteq_condition): New
function declaration.
* config/pru/pru.cc (pru_noteq_condition): New function.
* config/pru/pru.md (cbranchdi4): Define new pattern.
gcc/testsuite/ChangeLog:
* gcc.target/pru/pr106562-1.c: New test.
* gcc.target/pru/pr106562-2.c: New test.
* gcc.target/pru/pr106562-3.c: New test.
* gcc.target/pru/pr106562-4.c: New test.
If the number of shift positions is a constant, then the DI shift
operation is expanded to a sequence of 2 to 4 machine instructions.
That is more efficient than the default action to call libgcc.
gcc/ChangeLog:
* config/pru/pru.md (lshrdi3): New expand pattern.
(ashldi3): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/pru/ashiftdi-1.c: New test.
* gcc.target/pru/lshiftrtdi-1.c: New test.
YunQiang Su [Tue, 2 Aug 2022 10:57:18 +0000 (10:57 +0000)]
MIPS: improve -march=native arch detection
If we cannot get info from options and cpuinfo, we try to get from:
1. getauxval(AT_BASE_PLATFORM), introduced since Linux 5.7
2. _MIPS_ARCH from host compiler.
mnan=2008 option is also used if __mips_nan2008__ is used.
This can fix the wrong loader usage on r5/r6 platform with
-march=native.
gcc/ChangeLog:
* config.gcc: set with_arch to default_mips_arch if no defined.
* config/mips/driver-native.cc (host_detect_local_cpu):
try getauxval(AT_BASE_PLATFORM) and _MIPS_ARCH, too.
pass -mnan=2008 if __mips_nan2008__ is defined.
* config.in: define HAVE_SYS_AUXV_H and HAVE_GETAUXVAL.
* configure.ac: detect sys/auxv.h and getauxval.
* configure: regenerated.
Jason Merrill [Sat, 17 Sep 2022 10:04:05 +0000 (12:04 +0200)]
c++: track whether we expect a TARGET_EXPR to be elided
A discussion at Cauldron made me think that with the formalization of copy
elision in C++17, we should be able to determine before optimization which
TARGET_EXPRs will become temporaries and which are initializers. This patch
implements that: we set TARGET_EXPR_ELIDING_P if it's used as an
initializer, and later check that we were right.
There's an exception in the cp_gimplify_expr check to allow extra
temporaries of non-addressable type: this is used by
gimplify_init_ctor_preeval to materialize subobjects of a CONSTRUCTOR on the
rhs of a MODIFY_EXPR rather than materializing the whole object. If the
type isn't addressable, there's no way for a program to tell the difference,
so this is a valid optimization.
I considered changing replace_placeholders_for_class_temp_r to check
TARGET_EXPR_ELIDING_P instead of potential_prvalue_result_of, but decided
that would be wrong: if we have an eliding TARGET_EXPR inside a non-eliding
one, we would miss replacing its placeholders.
gcc/cp/ChangeLog:
* cp-tree.h (TARGET_EXPR_ELIDING_P): New.
(unsafe_copy_elision_p, set_target_expr_eliding)
(cp_build_init_expr): Declare.
* call.cc (unsafe_copy_elision_p): No longer static.
(build_over_call, build_special_member_call)
(build_new_method_call): Use cp_build_init_expr.
* coroutines.cc (expand_one_await_expression)
(build_actor_fn, flatten_await_stmt, handle_nested_conditionals)
(await_statement_walker, morph_fn_to_coro): Use cp_build_init_expr.
* cp-gimplify.cc (cp_gimplify_init_expr)
(cp_gimplify_expr): Check TARGET_EXPR_ELIDING_P.
(cp_fold_r): Propagate it.
(cp_fold): Use cp_build_init_expr.
* decl.cc (check_initializer): Use cp_build_init_expr.
* except.cc (build_throw): Use cp_build_init_expr.
* init.cc (get_nsdmi): Call set_target_expr_eliding.
(perform_member_init, expand_default_init, expand_aggr_init_1)
(build_new_1, build_vec_init): Use cp_build_init_expr.
* method.cc (do_build_copy_constructor): Use cp_build_init_expr.
* semantics.cc (simplify_aggr_init_expr, finalize_nrv_r)
(finish_omp_reduction_clause): Use cp_build_init_expr.
* tree.cc (build_target_expr): Call set_target_expr_eliding.
(bot_manip): Copy TARGET_EXPR_ELIDING_P.
* typeck.cc (cp_build_modify_expr): Call set_target_expr_eliding.
(check_return_expr): Use cp_build_modify_expr.
* typeck2.cc (split_nonconstant_init_1)
(split_nonconstant_init): Use cp_build_init_expr.
(massage_init_elt): Call set_target_expr_eliding.
(process_init_constructor_record): Clear TARGET_EXPR_ELIDING_P on
unsafe copy elision.
(set_target_expr_eliding, cp_build_init_expr): New.
Marek Polacek [Wed, 5 Oct 2022 19:51:30 +0000 (15:51 -0400)]
c++: fixes for derived-to-base reference binding [PR107085]
This PR reports that
struct Base {};
struct Derived : Base {};
static_assert(__reference_constructs_from_temporary(Base const&, Derived));
doesn't pass, which it should: it's just like
const Base& b(Derived{});
where we bind 'b' to the Base subobject of a temporary object of type
Derived. The ck_base conversion didn't have ->need_temporary_p set because
we didn't need to create a temporary object just for the base, but the whole
object is a temporary so we're still binding to a temporary. Since the
Base subobject is an xvalue, a new function is introduced.
Set discriminators for call stmts on the same line within the same basic block.
Call statements are possible split points of a basic block so they may end up
in different basic blocks by the time pass_ipa_auto_profile executes.
This change will also simplify call site lookups since now location with discriminator
will uniquely identify the call site (no callee function name is needed).
Qing Zhao [Fri, 7 Oct 2022 14:59:01 +0000 (14:59 +0000)]
Use array_at_struct_end_p in __builtin_object_size [PR101836]
Use array_at_struct_end_p to determine whether the trailing array
of a structure is flexible array member in __builtin_object_size.
gcc/ChangeLog:
PR tree-optimization/101836
* tree-object-size.cc (addr_object_size): Use array_at_struct_end_p
to determine a flexible array member reference.
gcc/testsuite/ChangeLog:
PR tree-optimization/101836
* gcc.dg/pr101836.c: New test.
* gcc.dg/pr101836_1.c: New test.
* gcc.dg/pr101836_2.c: New test.
* gcc.dg/pr101836_3.c: New test.
* gcc.dg/pr101836_4.c: New test.
* gcc.dg/pr101836_5.c: New test.
* gcc.dg/strict-flex-array-2.c: New test.
* gcc.dg/strict-flex-array-3.c: New test.
Qing Zhao [Fri, 7 Oct 2022 14:58:20 +0000 (14:58 +0000)]
Add a new option -fstrict-flex-arrays[=n] and new attribute strict_flex_array
Add the following new option -fstrict-flex-arrays[=n] and a corresponding
attribute strict_flex_array to GCC:
'-fstrict-flex-arrays'
Control when to treat the trailing array of a structure as a flexible array
member for the purpose of accessing the elements of such an array.
The positive form is equivalent to '-fstrict-flex-arrays=3', which is the
strictest. A trailing array is treated as a flexible array member only when
it declared as a flexible array member per C99 standard onwards.
The negative form is equivalent to '-fstrict-flex-arrays=0', which is the
least strict. All trailing arrays of structures are treated as flexible
array members.
'-fstrict-flex-arrays=LEVEL'
Control when to treat the trailing array of a structure as a flexible array
member for the purpose of accessing the elements of such an array. The value
of LEVEL controls the level of strictness
The possible values of LEVEL are the same as for the
'strict_flex_array' attribute (*note Variable Attributes::).
You can control this behavior for a specific trailing array field
of a structure by using the variable attribute 'strict_flex_array'
attribute (*note Variable Attributes::).
'strict_flex_array (LEVEL)'
The 'strict_flex_array' attribute should be attached to the trailing
array field of a structure. It controls when to treat the trailing array
field of a structure as a flexible array member for the purposes of accessing
the elements of such an array. LEVEL must be an integer betwen 0 to 3.
LEVEL=0 is the least strict level, all trailing arrays of
structures are treated as flexible array members. LEVEL=3 is the
strictest level, only when the trailing array is declared as a
flexible array member per C99 standard onwards ('[]'), it is
treated as a flexible array member.
There are two more levels in between 0 and 3, which are provided to
support older codes that use GCC zero-length array extension
('[0]') or one-element array as flexible array members('[1]'): When
LEVEL is 1, the trailing array is treated as a flexible array member
when it is declared as either '[]', '[0]', or '[1]'; When
LEVEL is 2, the trailing array is treated as a flexible array member
when it is declared as either '[]', or '[0]'.
This attribute can be used with or without the
'-fstrict-flex-arrays'. When both the attribute and the option
present at the same time, the level of the strictness for the
specific trailing array field is determined by the attribute.
gcc/c-family/ChangeLog:
* c-attribs.cc (handle_strict_flex_array_attribute): New function.
(c_common_attribute_table): New item for strict_flex_array.
* c.opt: (fstrict-flex-arrays): New option.
(fstrict-flex-arrays=): New option.
gcc/c/ChangeLog:
* c-decl.cc (flexible_array_member_type_p): New function.
(one_element_array_type_p): Likewise.
(zero_length_array_type_p): Likewise.
(add_flexible_array_elts_to_size): Call new utility
routine flexible_array_member_type_p.
(is_flexible_array_member_p): New function.
(finish_struct): Set the new DECL_NOT_FLEXARRAY flag.
gcc/cp/ChangeLog:
* module.cc (trees_out::core_bools): Stream out new bit
decl_not_flexarray.
(trees_in::core_bools): Stream in new bit decl_not_flexarray.
gcc/ChangeLog:
* doc/extend.texi: Document strict_flex_array attribute.
* doc/invoke.texi: Document -fstrict-flex-arrays[=n] option.
* print-tree.cc (print_node): Print new bit decl_not_flexarray.
* tree-core.h (struct tree_decl_common): New bit field
decl_not_flexarray.
* tree-streamer-in.cc (unpack_ts_decl_common_value_fields): Stream
in new bit decl_not_flexarray.
* tree-streamer-out.cc (pack_ts_decl_common_value_fields): Stream
out new bit decl_not_flexarray.
* tree.cc (array_at_struct_end_p): Update it with the new bit field
decl_not_flexarray.
* tree.h (DECL_NOT_FLEXARRAY): New flag.
gcc/testsuite/ChangeLog:
* g++.dg/strict-flex-array-1.C: New test.
* gcc.dg/strict-flex-array-1.c: New test.
Olivier Hainque [Fri, 7 Oct 2022 10:12:26 +0000 (10:12 +0000)]
Specialize paths to version.h in _vxworks-versions.h
The _vxworks-versions.h runtime file helps us control
the compilation of some library components depending on
the OS version extracted out of a system header.
The system header name is "version.h", and gcc has a
"version.h" file of its own.
gcc's version.h is now generated and the current
instance instead of the OS one, resulting in build failures
from
This change introduces a twist in the way
_vxworks-versions.h #includes version.h, using a relative
path with components that will match in the OS include
dirs only.
The actual relative path is conditioned on _VSB_CONFIG_FILE
to accommodate a change in the include dirs organisation between
VxWorks 6 and 7.
2022-10-07 Olivier Hainque <hainque@adacore.com>
gcc/
* config/vxworks/_vxworks-versions.h: Use OS specific
paths in #include of version.h.
David Malcolm [Fri, 7 Oct 2022 16:41:59 +0000 (12:41 -0400)]
analyzer: extract bits from integer constants [PR105783]
Fix a false positive from -Wanalyzer-null-dereference due to -fanalyzer
failing to grok the value of a particular boolean field initialized to a
constant.
gcc/analyzer/ChangeLog:
PR analyzer/105783
* region-model.cc (selftest::get_bit): New function.
(selftest::test_bits_within_svalue_folding): New.
(selfftest::analyzer_region_model_cc_tests): Call it.
* svalue.cc (constant_svalue::maybe_fold_bits_within): Handle the
case of extracting a single bit.
gcc/testsuite/ChangeLog:
PR analyzer/105783
* gcc.dg/analyzer/pr105783.c: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
Now we have templated lambdas, we can have variadic template lambdas,
and this leads to lambda signatures containing parameter packs. But
just like 'auto' inside such a signature, we don't have a containing
template, and thus fail. The fix is to check is_lambda_arg, just as
for a template parameter. This allows us to demangle g++'s manglings
of such lambdas.
It's not a totally accurate demangling, because we don't mangle the
template head (that's a separate issue), but it is better than failing
to demangle.
Due to the way we print subexprs, we add an unnecessary parens around
the argument of the pack. That's an orthogonal problem, for which the
solution is to have better knowledge of operator precedence.
libiberty/
* cp-demangle.c (d_print_comp_inner): Allow parameter packs
in a lambda signature.
* testsuite/demangle-expected: Add tests.
Patrick Palka [Fri, 7 Oct 2022 16:01:58 +0000 (12:01 -0400)]
c++ modules: ICE with bitfield in class template
According to grokbitfield, DECL_BIT_FIELD_REPRESENTATIVE contains the
width of the bitfield until we layout the class type (after which it'll
contain a decl). Thus for a bitfield in a class template it'll always
be the width, and this patch makes us avoid ICEing from mark_class_def
in this case.
gcc/cp/ChangeLog:
* module.cc (trees_out::mark_class_def): Guard against
DECL_BIT_FIELD_REPRESENTATIVE not being a decl.
Jason Merrill [Fri, 7 Oct 2022 01:10:52 +0000 (21:10 -0400)]
c++: catch parm initialization tweak
We want to push the INIT_EXPR inside the CLEANUP_POINT_EXPR for the same
reason we want to push it into the MUST_NOT_THROW_EXPR: any cleanups follow
the initialization.
gcc/cp/ChangeLog:
* init.cc (expand_default_init): Also push the INIT_EXPR inside a
CLEANUP_POINT_EXPR.
Martin Liska [Wed, 5 Oct 2022 09:15:36 +0000 (11:15 +0200)]
IPA: support -flto + -flive-patching=inline-clone
There's no fundamental reason why -flive-patching=inline-clone can't
coexist with -flto. Yes, one can theoretically have many more clone
function that includes a live patch. It is pretty much the same
as in-module inlining.
gcc/ChangeLog:
* opts.cc (finish_options): Print sorry message only
for -flive-patching=inline-only-static.
gcc/testsuite/ChangeLog:
* gcc.dg/live-patching-2.c: Update scanned pattern.
* gcc.dg/live-patching-5.c: New test.
Patrick Palka [Fri, 7 Oct 2022 13:32:45 +0000 (09:32 -0400)]
c++ modules: static var in inline function [PR104433]
The below testcase fails to link with the error
undefined reference to `f()::y'
ultimately because during stream out for the static VAR_DECL y we
override DECL_EXTERNAL to true, which later during IPA confuses
symbol_table::remove_unreachable_nodes into thinking it's safe
to not emit the symbol.
The streaming code here already avoids overriding DECL_EXTERNAL for
inline vars and functions, so it seems natural to extend this to
static vars from an inline function.
PR c++/104433
gcc/cp/ChangeLog:
* module.cc (trees_out::core_bools): Don't override
DECL_EXTERNAL to true for static variables from an inline
function.
gcc/testsuite/ChangeLog:
* g++.dg/modules/static-2_a.H: New test.
* g++.dg/modules/static-2_b.C: New test.
Jason Merrill [Wed, 5 Oct 2022 15:50:59 +0000 (11:50 -0400)]
gimplify: prevent some C++ temporary elision
In this testcase, we were optimizing away the temporary for f(), but
C++17 and above are clear that there is a temporary, and because its
destructor has visible side-effects we can't optimize it away under the
as-if rule. So disable this optimization for TREE_ADDRESSABLE type.
I moved the declaration of volatile_p after the call to
gimple_fold_indirect_ref_rhs to minimize indentation changes; I don't see
any way the value of that flag could be affected by the call.
gcc/ChangeLog:
* gimplify.cc (gimplify_modify_expr_rhs): Don't optimize
x = *(A*)&<expr> to x = <expr> for a TREE_ADDRESSABLE type.
autopar performs insertion of stores, eventually requiring a
virtual loop PHI and assorted LC PHI adjustments which we intend
to do once after the pass finishes. But we also perform intermediate
update_ssa after loop duplication which can lose this fact. The
following forces renaming of the virtual operand before the final
SSA update to fix that. It also removes the explicit update_ssa
call from the gimple_duplicate_sese_tail utility as has been done
for all other such utilities and instead performs the SSA update
from autopar.
PR tree-optimization/107153
* tree-cfg.cc (gimple_duplicate_sese_tail): Do not update
SSA form here.
* tree-parloops.cc (gen_parallel_loop): Update SSA form
after to-exit-first transform, no PHI insertion is necessary.
(pass_parallelize_loops::execute): Force re-write of the
virtual operand SSA web.
Jonathan Wakely [Thu, 6 Oct 2022 17:41:22 +0000 (18:41 +0100)]
libstdc++: Add --disable-libstdcxx-hosted as an alias for hosted-libstdcxx
Most libstdc++ configure args are of the form --enable-libstdcxx-xxx but
the option to build freestanding is --disable-hosted-libstdcxx. If you
accidentally type --disable-libstdcxx-hosted then it will be ignored.
This adds --disable-libstdcxx-hosted as an alias for the existing arg,
so it works whichever way you say it. If both args are used explicitly
and their values do not agree, configure fails with an error.
libstdc++-v3/ChangeLog:
* acinclude.m4 (GLIBCXX_ENABLE_HOSTED): Add libstdcxx-hosted
enable arg as an alias for hosted-libstdcxx enable arg.
* configure: Regenerate.
Martin Liska [Fri, 7 Oct 2022 08:22:24 +0000 (10:22 +0200)]
fix clang warnings
Fixes:
gcc/c-family/name-hint.h:109:66: warning: unqualified call to 'std::move' [-Wunqualified-std-cast-call]
gcc/config/i386/i386-expand.cc:1351:9: warning: argument 'operands' of type 'rtx[3]' (aka 'rtx_def *[3]') with mismatched bound [-Warray-parameter]
gcc/config/i386/i386.cc:15635:8: warning: argument 'operands' of type 'rtx[2]' (aka 'rtx_def *[2]') with mismatched bound [-Warray-parameter]
gcc/cp/module.cc:17482:51: warning: argument 'counts' of type 'unsigned int[8]' with mismatched bound [-Warray-parameter]
gcc/cp/module.cc:17508:37: warning: argument 'counts' of type 'unsigned int[8]' with mismatched bound [-Warray-parameter]
gcc/cp/name-lookup.cc:6385:16: warning: unqualified call to 'std::move' [-Wunqualified-std-cast-call]
gcc/c-family/ChangeLog:
* name-hint.h: Use std::move.
gcc/ChangeLog:
* config/i386/i386-protos.h (ix86_binary_operator_ok): Add array
size to function parameter.
(ix86_unary_operator_ok): Likewise.
gcc/cp/ChangeLog:
* module.cc (enum module_state_counts): Use array size.
* name-lookup.cc (class namespace_limit_reached): Likewise.
(class module_state): Move up in the file.
Jakub Jelinek [Fri, 7 Oct 2022 07:01:04 +0000 (09:01 +0200)]
c++: Improve handling of foreigner namespace attributes
In some cases we want to look up or remove both standard
attributes and attributes from gnu namespace but not others.
This patch arranges for ATTR_NS of "" to stand for ATTR_NS
NULL or "gnu", so that we don't need 2 separate calls, and
introduces is_attribute_namespace_p function which allows
testing the namespace of an attribute similar way.
The patch also uses the new lookup_attribute overload and extra
tests to avoid emitting weird warnings on foreign namespace attributes
which we should just ignore (perhaps with a warning), but shouldn't
imply any meaning to them just because they have a name matching some
standard or gnu attribute name.
2022-10-07 Jakub Jelinek <jakub@redhat.com>
gcc/
* attribs.h (is_attribute_namespace_p): New inline function.
(lookup_attribute): Document meaning of ATTR_NS equal to "".
* attribs.cc (remove_attribute): Use is_attribute_namespace_p.
(private_lookup_attribute): For ATTR_NS "" match either standard
attribute or "gnu" namespace one.
gcc/c-family/
* c-common.cc (attribute_fallthrough_p): Lookup fallthrough attribute
only in gnu namespace or as standard attribute, treat fallthrough
attributes in other namespaces like any other unknown attribute.
gcc/cp/
* parser.cc (cp_parser_check_std_attribute): Only do checks if
attribute is a standard attribute or in gnu namespace and only
lookup other attributes in those namespaces.
* cp-gimplify.cc (lookup_hotness_attribute): Adjust function comment.
Only return true for standard attribute or gnu namespace attribute.
(remove_hotness_attribute): Only remove hotness attributes when
they are standard or in gnu namespace, implement it in a single
loop rather than former 4 now 8 remove_attribute calls.
gcc/testsuite/
* g++.dg/cpp1z/fallthrough2.C: New test.
* g++.dg/cpp2a/attr-likely7.C: New test.
Jakub Jelinek [Fri, 7 Oct 2022 06:59:05 +0000 (08:59 +0200)]
fixincludes: Deal also with the _Float128x cases [PR107059]
On Wed, Sep 28, 2022 at 08:19:43PM +0200, Jakub Jelinek via Gcc-patches wrote:
> Another case are the following 3 snippets:
> # if !__GNUC_PREREQ (7, 0) || defined __cplusplus
> # error "_Float128X supported but no constant suffix"
> # else
> # define __f128x(x) x##f128x
> # endif
> ...
> # if !__GNUC_PREREQ (7, 0) || defined __cplusplus
> # error "_Float128X supported but no complex type"
> # else
> # define __CFLOAT128X _Complex _Float128x
> # endif
> ...
> # if !__GNUC_PREREQ (7, 0) || defined __cplusplus
> # error "_Float128x supported but no type"
> # endif
> but as no target has _Float128x right now and don't see it
> coming soon, it isn't a big deal (on the glibc side it is of
> course ok to adjust those).
This incremental patch deals handles the above 3 cases, so we
fixinclude what glibc itself changed too.
Jakub Jelinek [Fri, 7 Oct 2022 06:56:04 +0000 (08:56 +0200)]
fixincludes: Fix up powerpc floatn.h tweaks [PR107059]
On Wed, Sep 28, 2022 at 12:23:31AM +0000, Joseph Myers wrote:
> In general the changes match those made by fixincludes, though I think
> the ones in sysdeps/powerpc/bits/floatn.h, where the header tests
> __LDBL_MANT_DIG__ == 113 or uses #elif, wouldn't match the existing
> fixincludes patterns.
You're right, missed that.
The header has:
/* Defined to a complex binary128 type if __HAVE_FLOAT128 is 1. */
# if __HAVE_FLOAT128
# if __LDBL_MANT_DIG__ == 113 && defined __cplusplus
typedef long double _Float128;
# define __CFLOAT128 _Complex long double
# elif !__GNUC_PREREQ (7, 0) || defined __cplusplus
/* The type _Float128 exist for powerpc only since GCC 7.0. */
typedef __float128 _Float128;
/* Add a typedef for older GCC and C++ compilers which don't natively support
_Complex _Float128. */
typedef _Complex float __cfloat128 __attribute__ ((__mode__ (__KC__)));
# define __CFLOAT128 __cfloat128
# else
# define __CFLOAT128 _Complex _Float128
# endif
# endif
and my current rules don't do anything about that.
The following patch fixes that.
I've run additionally
MACRO_LIST=`pwd`/../gcc/macro_list TARGET_MACHINE=x86_64-pc-linux-gnu \
../fixincludes/fixinc.sh /tmp/include-fixed \
`echo /usr/src/libc | sed -e :a -e 's,[^/]*/\.\.\/,,' -e ta`
in the builddir/fixincludes directory where /usr/src/libc is latest glibc
trunk checkout and seems the remaining defined __cplusplus cases in the floatn.h
and floatn-common.h headers are ok or acceptable.
The remaining cases are:
#if __GNUC_PREREQ (7, 0) && !defined __cplusplus
# define __HAVE_FLOATN_NOT_TYPEDEF 1
#else
# define __HAVE_FLOATN_NOT_TYPEDEF 0
#endif
which is IMHO ok because this is only used in tgmath.h or tgmath-like math.h
stuff which is C only, as C++ doesn't have _Generic.
Another case are the following 3 snippets:
# if !__GNUC_PREREQ (7, 0) || defined __cplusplus
# error "_Float128X supported but no constant suffix"
# else
# define __f128x(x) x##f128x
# endif
...
# if !__GNUC_PREREQ (7, 0) || defined __cplusplus
# error "_Float128X supported but no complex type"
# else
# define __CFLOAT128X _Complex _Float128x
# endif
...
# if !__GNUC_PREREQ (7, 0) || defined __cplusplus
# error "_Float128x supported but no type"
# endif
but as no target has _Float128x right now and don't see it
coming soon, it isn't a big deal (on the glibc side it is of
course ok to adjust those).
OT, besides floatn.h and floatn-common.h headers, the only
one remaining in /tmp/include-fixed is sysdeps/arm/unwind.h, perhaps
-#if defined(linux) || defined(__NetBSD__)
+#if defined(__linux__) || defined(__NetBSD__)
should be done in that header (and libgcc/config/arm/unwind-arm.h
too).
2022-10-07 Jakub Jelinek <jakub@redhat.com>
PR bootstrap/107059
* inclhack.def (glibc_cxx_floatn_2): Handle #elif the same as #if.
(glibc_cxx_floatn_4): New.
* fixincl.x: Regenerated.
* tests/base/bits/floatn.h: Regenerated.
Olivier Hainque [Thu, 3 Mar 2022 16:58:51 +0000 (16:58 +0000)]
Downgrade DWARF_VERSION_DEFAULT to 3 for VxWorks >= 7
Using 4 as the DWARF_VERSION_DEFAULT value for VxWorks observably
still hangs recent system debuggers on tbreak for some contructs,
and downgrading to 3 improves the situation.
2022-03-06 Olivier Hainque <hainque@adacore.com>
gcc/
* config/vxworks.h (DWARF_VERSION_DEFAULT): Adjust from
4 to 3 for VxWorks >= 7.
Introduce DWARF_VERSION_DEFAULT and redefine for VxWorks
This change introduces a target overridable macro to replace
the hardcoded value used to initialize dwarf_version from common.opt.
The main advantage compared to special code in a target
override_options hook is that redefinitions by target config files
are visible by both the compiler proper and by the driver, which
might refer to dwarf_version in ASM_DEBUG_SPECs and friends.
This is useful at least on VxWorks, where we usually need to
default to dwarf 4 or even 2 to accommodate non-gdb debuggers
provided by the environment, including for assembly sources
used in libgcc for some ports (witnessed with lse.S on aarch64).
2022-02-28 Olivier Hainque <hainque@adacore.com>
gcc/
* defaults.h (DWARF_DEFAULT_VERSION): Define if not
defined already.
* common.opt (gdwarf-): Use it.
* doc/tm.texi.in (DWARF_DEFAULT_VERSION): Document.
* doc/tm.texi: Update accordingly.
* config/vxworks.h (DWARF_DEFAULT_VERSION): Redefine.
* config/vxworks.cc: Remove code setting dwarf_version, now
handled by the DWARF_DEFAULT_VERSION redefinition.
Jason Merrill [Fri, 30 Sep 2022 14:04:22 +0000 (10:04 -0400)]
c++: fix broken conversion in coroutines
You can't use CONVERT_EXPR to convert between two class types.
VIEW_CONVERT_EXPR takes liberties with the C++ type system, but is probably
safe in this context. Let's also only use it when the type isn't already
what we want.
Aldy Hernandez [Thu, 6 Oct 2022 17:11:08 +0000 (19:11 +0200)]
[PR107170] Avoid copying incompatible types in legacy VRP.
Legacy VRP is calling ranger deep inside the bowels, and then trying to
copy an incompatible type. My previous patch in this area assumed that
the only possibility out of vr_values::get_value_range for an
unsupported type was VARYING, but UNDEFINED can also be returned.
PR tree-optimization/107170
gcc/ChangeLog:
* vr-values.cc (vr_values::range_of_expr): Do not die on
unsupported types.
David Malcolm [Thu, 6 Oct 2022 19:46:49 +0000 (15:46 -0400)]
analyzer: fix another ICE in PR 107158
I overreduced PR analyzer/107158 in r13-3096-gef878564140cbc, and there
was another ICE in the original reproducer, which this patch fixes.
gcc/analyzer/ChangeLog:
PR analyzer/107158
* store.cc (store::replay_call_summary_cluster): Eliminate
special-casing of RK_HEAP_ALLOCATED in favor of sharing code with
RK_DECL, avoiding an ICE due to attempting to bind a
compound_svalue into a binding_cluster when an svalue in the
summary cluster converts to a compound_svalue in the caller.
gcc/testsuite/ChangeLog:
PR analyzer/107158
* gcc.dg/analyzer/call-summaries-pr107158-2.c: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
Joseph Myers [Thu, 6 Oct 2022 14:26:21 +0000 (14:26 +0000)]
c: C2x typeof
C2x adds typeof as a standard feature. In general this follows
existing GNU C semantics very closely, but there are various ways in
which the implementation involves more than simply enabling the
keyword for C2x:
* As well as typeof, there is a typeof_unqual variant, which removes
all qualifiers and _Atomic from the resulting type (whereas typeof
preserves qualifiers and _Atomic on qualified or atomic (lvalue or
type name) operands).
* The typeof keyword is disabled by -fno-asm, so enabling it for C2x
needs to be implemented in a way that preserves the disabling by
-fno-asm for older standard versions (which having -fno-asm having
no effect on it in C2x mode). This is done via adding a new D_EXT11
mask (which is also where the C++ front-end change comes from, to
handle D_EXT11 appropriately there for -fno-asm and
-fno-gnu-keywords).
* GNU typeof treats the noreturn property of a function (as specified
in standard C with _Noreturn or [[noreturn]]) as being part of the
type of a pointer to function, but it is not part of the type in
standard terms. Thus a special case is needed in the typeof
implementation, just like in the _Generic implementation, to avoid
treating it as a type for standard typeof. It seems plausible this
is being used when copying the type of one object to another using
typeof, so the existing semantics are preserved for __typeof__, and
for typeof in pre-C2x modes, while typeof for C2x or later has the
standard semantics.
* It turns out that, even after Martin Uecker's changes in this area,
there were still cases where rvalues could wrongly have a qualified
or atomic type in GCC. This applied to ++ and -- increment and
decrement expressions, and also to calls to functions returning an
atomic type. (For the latter, the working draft doesn't actually
explicitly exclude the function call expression having an atomic
type, but given all the changes that have gone into C17 and C2x to
avoid rvalues ever having qualified types, and given that
lvalue-to-rvalue conversion removes both qualifiers and _Atomic, it
seems unlikely that this (or casts, where GCC already removes
_Atomic) is actually intended as a route to allow an
_Atomic-qualified rvalue; I've noted this to raise as an NB comment
on the CD ballot.)
Bootstrapped with no regressions for x86_64-pc-linux-gnu. OK to
commit (C+
gcc/
* doc/invoke.texi (-fno-asm): Update description of effects on
typeof keyword.
gcc/c-family/
* c-common.cc (c_common_reswords): Mark typeof as D_EXT11. Add
typeof_unqual.
* c-common.h (enum rid): Add RID_TYPEOF_UNQUAL.
(D_EXT11): New macro. Values of subsequent macros updated.
gcc/c/
* c-parser.cc (c_parse_init): Add D_EXT11 to mask if flag_no_asm
and not C2x.
(c_keyword_starts_typename, c_token_starts_declspecs)
(c_parser_declspecs, c_parser_objc_selector): Handle
RID_TYPEOF_UNQUAL.
(c_parser_typeof_specifier): Handle RID_TYPEOF_UNQUAL.
Distinguish typeof for C2x from __typeof__ for all standard
versions and typeof before C2x.
* c-typeck.cc (build_function_call_vec): Use unqualified version
of non-void return type.
(build_unary_op): Use unqualified type for increment and
decrement.
Patrick Palka [Thu, 6 Oct 2022 14:04:52 +0000 (10:04 -0400)]
c++: remove optimize_specialization_lookup_p
Roughly speaking, optimize_specialization_lookup_p returns true for a
non-template member function of a class template, e.g.
template<class T> struct A { int f(); };
The idea behind the optimization guarded by this predicate is that if
we want to look up the specialization A<T>::f [with T=int], then we can
just do a name lookup for f in A<int> and avoid having to add a
spec_entry for f in the decl_specializations table.
But the benefit of this optimization seems questionable because in
order to do the name lookup we first need to look up A<T> [with T=int]
in the type_specializations table, which is as expensive as the
decl_specializations lookup we're avoiding. And according to some
experiments (using stdc++.h, range-v3 and libstdc++ tests) the compiler
is slightly (<1%) _faster_ if we disable this optimization.
Additionally, this optimization means we won't record an explicit
specialization in decl_specializations for such a template either, which
is an unfortunate inconsistency that apparently breaks the below modules
testcase.
So since this optimization doesn't improve performance, and complicates
the explicit specialization story which causes issues with modules, this
patch proposes to remove it.
gcc/cp/ChangeLog:
* pt.cc (optimize_specialization_lookup_p): Remove.
(retrieve_specialization): Assume the above returns false
and simplify accordingly.
(register_specialization): Likewise.
gcc/testsuite/ChangeLog:
* g++.dg/modules/indirect-3_b.C: Expect that the entity
foo::TPL<0>::frob is tagged as a specialization instead
of as a declaration.
* g++.dg/modules/tpl-spec-8_a.H: New test.
* g++.dg/modules/tpl-spec-8_b.C: New test.
Eric Botcazou [Thu, 6 Oct 2022 13:13:50 +0000 (15:13 +0200)]
Fix wrong code generated by unroll-and-jam pass
There is a loophole in the unroll-and-jam pass that can quickly result in
wrong code generation. The code reads:
if (!compute_data_dependences_for_loop (outer, true, &loop_nest,
&datarefs, &dependences))
{
if (dump_file && (dump_flags & TDF_DETAILS))
fprintf (dump_file, "Cannot analyze data dependencies\n");
free_data_refs (datarefs);
free_dependence_relations (dependences);
continue;
}
but compute_data_dependences_for_loop may return true even if the analysis
is reported as failing by compute_affine_dependence for a dependence pair:
(compute_affine_dependence
ref_a: data[_14], stmt_a: data[_14] = i_59;
ref_b: data[_14], stmt_b: data[_14] = i_59;
Data ref a:
Data ref b:
affine dependence test not usable: access function not affine or constant.
) -> dependence analysis failed
Note that this is a self-dependence pair and the code for them reads:
/* Nothing interesting for the self dependencies. */
if (dra == drb)
continue;
This means that the pass may reorder "complex" accesses to the same memory
location in successive iterations, which is OK for reads but not for writes.
gcc/
* gimple-loop-jam.cc (tree_loop_unroll_and_jam): Bail out for a self
dependency that is a write-after-write if the access function is not
affine or constant.
gcc/testsuite/
* gcc.c-torture/execute/20221006-1.c: New test.
Richard Biener [Thu, 6 Oct 2022 09:48:03 +0000 (11:48 +0200)]
middle-end/107115 - avoid bogus redundant store removal during RTL expansion
The following preserves the (premature) redundant store removal
done in store_expr by appropriately guarding it with
mems_same_for_tbaa_p. The testcase added needs scheduling disabled
for now since there's a similar bug there still present.
PR middle-end/107115
* expr.cc (store_expr): Check mems_same_for_tbaa_p before
eliding a seemingly redundant store.
Kyrylo Tkachov [Thu, 6 Oct 2022 11:09:28 +0000 (12:09 +0100)]
aarch64: Remove redundant zero-extends with LDAR
Like other loads in AArch64, the LDARB,LDARH,LDAR instructions clear out the top part of their
destination register and we can thus avoid having to explicitly zero-extend it.
We were missing a combine pattern that this patch adds.
For one of the examples in the testcase we generated:
load_uint8_t_ext_uint16_t:
adrp x0, .LANCHOR0
add x0, x0, :lo12:.LANCHOR0
ldarb w0, [x0]
and w0, w0, 255
ret
but now generate:
load_uint8_t_ext_uint16_t:
adrp x0, .LANCHOR0
add x0, x0, :lo12:.LANCHOR0
ldarb w0, [x0]
ret
Bootstrapped and tested on aarch64-none-linux-gnu.
gcc/ChangeLog:
* config/aarch64/atomics.md (*atomic_load<ALLX:mode>_zext<SD_HSDI:mode>):
New pattern.
Philipp Tomsich [Sun, 7 Aug 2022 22:30:52 +0000 (00:30 +0200)]
aarch64: update Ampere-1 core definition
This brings the extensions detected by -mcpu=native on Ampere-1 systems
in sync with the defaults generated for -mcpu=ampere1.
Note that some early kernel versions on Ampere1 may misreport the
presence of PAUTH and PREDRES (i.e., -mcpu=native will add 'nopauth'
and 'nopredres').
projects / thirdparty / gcc.git / log
