c-family: add support for per-location warning groups.

gcc/c-family/ChangeLog:

* c-common.c (c_wrap_maybe_const): Remove TREE_NO_WARNING.
(c_common_truthvalue_conversion): Replace direct uses of
TREE_NO_WARNING with warning_suppressed_p, suppress_warning, and
copy_no_warning.
(check_function_arguments_recurse): Same.
* c-gimplify.c (c_gimplify_expr): Same.
* c-warn.c (overflow_warning): Same.
(warn_logical_operator): Same.
(warn_if_unused_value): Same.
(do_warn_unused_parameter): Same.

c: add support for per-location warning groups.

gcc/ChangeLog:
* tree.h (warning_suppressed_at, copy_warning,
warning_suppressed_p, suppress_warning): New functions.

gcc/c/ChangeLog:

* c-decl.c (pop_scope): Replace direct uses of TREE_NO_WARNING with
warning_suppressed_p, suppress_warning, and copy_no_warning.
(diagnose_mismatched_decls): Same.
(duplicate_decls): Same.
(grokdeclarator): Same.
(finish_function): Same.
(c_write_global_declarations_1): Same.
* c-fold.c (c_fully_fold_internal): Same.
* c-parser.c (c_parser_expr_no_commas): Same.
(c_parser_postfix_expression): Same.
* c-typeck.c (array_to_pointer_conversion): Same.
(function_to_pointer_conversion): Same.
(default_function_array_conversion): Same.
(convert_lvalue_to_rvalue): Same.
(default_conversion): Same.
(build_indirect_ref): Same.
(build_function_call_vec): Same.
(build_atomic_assign): Same.
(build_unary_op): Same.
(c_finish_return): Same.
(emit_side_effect_warnings): Same.
(c_finish_stmt_expr): Same.
(c_omp_clause_copy_ctor): Same.

Add support for per-location warning groups.

gcc/ChangeLog:

* Makefile.in (OBJS-libcommon): Add diagnostic-spec.o.
* gengtype.c (open_base_files): Add diagnostic-spec.h.
* diagnostic-spec.c: New file.
* diagnostic-spec.h: New file.
* tree.h (no_warning, all_warnings, suppress_warning_at): New
declarations.
* warning-control.cc: New file.

Revert x86_order_regs_for_local_alloc changes in r12-1669.

Still put general regs as first alloca order.

gcc/ChangeLog:

PR target/101185
* config/i386/i386.c (x86_order_regs_for_local_alloc):
Revert r12-1669.

gcc/testsuite/ChangeLog

PR target/101185
* gcc.target/i386/bitwise_mask_op-3.c: Add xfail to
temporarily avoid regression, eventually xfail should be
removed.

Daily bump.

Add a testcase to confirm the equivalence's are being checked by EVRP.

* gcc.dg/tree-ssa/evrp30.c: New.

Only register relations on live edges

Register a relation on a conditional edge only if the LHS supports
this edge being taken.

gcc/
PR tree-optimization/101189
* gimple-range-fold.cc (fold_using_range::range_of_range_op): Pass
LHS range of condition to postfold routine.
(fold_using_range::postfold_gcond_edges): Only process the TRUE or
FALSE edge if the LHS range supports it being taken.
* gimple-range-fold.h (postfold_gcond_edges): Add range parameter.

gcc/testsuite/
* gcc.dg/tree-ssa/pr101189.c: New.

Fix relation query of equivalences.

When looking for relations between equivalencies, a typo was causing
the wrong bitmap to be checked. Effect was is missed them.
Plus don't dump blocks which don't exist.

* value-relation.cc (equiv_oracle::dump): Do not dump NULL blocks.
(relation_oracle::find_relation_block): Check correct bitmap.
(relation_oracle::dump): Do not dump NULL blocks.

Correctly unify recomputation with existing range.

When propagating the on-entry cache, new block ranges are calculated
by combining all the incoming edges and comparing to the old value.
When a recomputation was performed on an edge, it didn't take into account
that the value in the block may already be better than a potential
recompuation... Thus a worse values was sometimes propagated.
Fixed by simply calling the now correct range_on_edge the cache provides.

* gimple-range-cache.cc (ranger_cache::propagate_cache): Call
range_on_edge instead of manually calculating.

Fix comment typo.

* range-op.cc: Fix comment.

c++: alias CTAD and aggregate deduction cand [PR98832]

During alias CTAD, we're accidentally ignoring the aggregate deduction
candidate for the underlying template because this guide is added
separately via maybe_aggr_guide (which doesn't yet handle alias
templates) instead of via deduction_guides_for (which does).  This patch
makes maybe_aggr_guide handle alias templates in a manner similar to
deduction_guides_for.

PR c++/98832

gcc/cp/ChangeLog:

* pt.c (maybe_aggr_guide): Handle alias templates appropriately.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/class-deduction-alias9.C: New test.

c++: requires-expression folding [PR101182]

Here we're crashing because cp_fold_function walks into the (templated)
requirements of a requires-expression outside a template, but the
folding routines aren't prepared to handle templated trees.  This patch
fixes this by making cp_fold use evaluate_requires_expr to fold a
requires-expression as a whole, which also means we no longer need to
explicitly do so during gimplification.  (Note that we delay folding
of such requires-expressions for sake of better diagnostics when one is
used as the condition of a failed static_assert.)

PR c++/101182

gcc/cp/ChangeLog:

* constraint.cc (evaluate_requires_expr): Adjust function comment.
* cp-gimplify.c (cp_genericize_r) <case REQUIRES_EXPR>: Move to ...
(cp_fold) <case REQUIRES_EXPR>: ... here.

gcc/testsuite/ChangeLog:

* g++.dg/cpp2a/concepts-requires25.C: New test.

c: Fix up c_parser_has_attribute_expression [PR101176]

This function keeps src_range member of the result uninitialized, which at
least under valgrind can show up later when those uninitialized location_t's
can make it into the IL or location_t hash tables.

2021-06-24  Jakub Jelinek  <jakub@redhat.com>

PR c/101176
* c-parser.c (c_parser_has_attribute_expression): Set source range for
the result.

c: Fix C cast error-recovery [PR101171]

The following testcase ICEs during error-recovery, as build_c_cast calls
note_integer_operands on error_mark_node and that wraps it into
C_MAYBE_CONST_EXPR which is unexpected and causes ICE later on.
Seems most other callers of note_integer_operands check early if something
is error_mark_node and return before calling note_integer_operands on it.

The following patch fixes it by not calling on error_mark_node, another
possibility would be to handle error_mark_node in note_integer_operands and
just return it.

2021-06-24  Jakub Jelinek  <jakub@redhat.com>

PR c/101171
* c-typeck.c (build_c_cast): Don't call note_integer_operands on
error_mark_node.

* gcc.dg/pr101171.c: New test.

i386: Add pack/unpack patterns for 64bit vectors [PR89021]

2021-06-24  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/89021
* config/i386/i386-expand.c (ix86_expand_sse_unpack):
Handle V8QI and V4HI modes.
* config/i386/mmx.md (sse4_1_<any_extend:code>v4qiv4hi2):
New insn pattern.
(sse4_1_<any_extend:code>v4qiv4hi2): Ditto.
(mmxpackmode): New mode attribute.
(vec_pack_trunc_<mmxpackmode:mode>): New expander.
(mmxunpackmode): New mode attribute.
(vec_unpacks_lo_<mmxunpackmode:mode>): New expander.
(vec_unpacks_hi_<mmxunpackmode:mode>): Ditto.
(vec_unpacku_lo_<mmxunpackmode:mode>): Ditto.
(vec_unpacku_hi_<mmxunpackmode:mode>): Ditto.
* config/i386/i386.md (extsuffix): Move from ...
* config/i386/sse.md: ... here.

gcc/testsuite/

PR target/89021
* gcc.dg/vect/vect-nb-iter-ub-3.c (dg-additional-options):
Add --param vect-epilogues-nomask=0.
* gcc.target/i386/pr97249-1.c (foo): Add #pragma GCC unroll
to avoid loop vectorization.
(foo1): Ditto.
(foo2): Ditto.

libstdc++: Fix internal names: add missing underscores

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_math.h
(_GLIBCXX_SIMD_MATH_CALL2_): Rename arg2_ to __arg2.
(_GLIBCXX_SIMD_MATH_CALL3_): Rename arg2_ to __arg2 and arg3_ to
__arg3.

libstdc++: Ensure unrolled loops inline the lambda

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd.h (__execute_on_index_sequence)
(__execute_on_index_sequence_with_return)
(__call_with_n_evaluations, __call_with_subscripts): Add flatten
attribute.

libstdc++: Avoid raising fp exceptions in trunc, floor, and ceil

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_x86.h (_S_trunc, _S_floor)
(_S_ceil): Set bit 8 (_MM_FROUND_NO_EXC) on AVX and SSE4.1
roundp[sd] calls.

libstdc++: Fix condition when AVX512F ldexp implementation is used

This improves codegen of ldexp if AVX512VL is available.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_x86.h (_S_ldexp): The AVX512F
implementation doesn't require a _VecBltnBtmsk ABI tag, it
requires either a 64-Byte input (in which case AVX512F must be
available) or AVX512VL.

libstdc++: Minor simd_math cleanups

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_math.h: Undefine internal
macros after use.
(frexp): Move #if to a more sensible position and reformat
preceding code.
(logb): Call _SimdImpl::_S_logb for fixed_size instead of
duplicating the code here.
(modf): Simplify condition.

libstdc++: Remove incorrect fabs(simd) overload

fabs(int) returns double, this one didn't. This overload is not
specified in the Parallelism TS 2. Also remove the comment about labs
and llabs: it doesn't belong here.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_math.h (fabs): Remove
fabs(simd<integral>) overload.

libstdc++: Improve simd fixed_size codegen

Sometimes fixed_size objects will get unnecessarily copied on the stack.
The simd implementation should never pass _SimdTuple by value to avoid
requiring the optimizer to see through these copies.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_converter.h
(_SimdConverter::operator()): Pass _SimdTuple by const-ref.
* include/experimental/bits/simd_fixed_size.h
(_GLIBCXX_SIMD_FIXED_OP): Pass binary operator _SimdTuple
arguments by const-ref.
(_S_masked_unary): Pass _SimdTuple by const-ref.

libstdc++: Remove dead code in simd

This helper type became unused at some point.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd_fixed_size.h
(_AbisInSimdTuple): Removed.

libstdc++: Improve copysign(simd) codegen

This also resolves a test failure on aarch64 with -ffast-math and
fixed_size<N> with large N.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/experimental/bits/simd.h: Add missing operator~
overload for simd<floating-point> to __float_bitwise_operators.
* include/experimental/bits/simd_builtin.h
(_SimdImplBuiltin::_S_complement): Bitcast to int (and back) to
implement complement for floating-point vectors.
* include/experimental/bits/simd_fixed_size.h
(_SimdImplFixedSize::_S_copysign): New function, forwarding to
copysign implementation of _SimdTuple members.
* include/experimental/bits/simd_math.h (copysign): Call
_SimdImpl::_S_copysign for fixed_size arguments. Simplify
generic copysign implementation using the new ~ operator.

libstdc++: Fix typos and markdown errors in new simd/README.md

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* testsuite/experimental/simd/README.md: Fix typos.

libstdc++: Implement LWG 2762 for std::unique_ptr::operator*

The LWG issue proposes to add a conditional noexcept-specifier to
std::unique_ptr's dereference operator. The issue is currently in
Tentatively Ready status, but even if it isn't voted into the draft, we
can do it as a conforming extensions. This commit also adds a similar
noexcept-specifier to operator[] for the unique_ptr<T[], D> partial
specialization.

Also ensure that all dereference operators for shared_ptr are noexcept,
and adds tests for the std::optional accessors modified by the issue,
which were already noexcept in our implementation.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/shared_ptr_base.h (__shared_ptr_access::operator[]):
Add noexcept.
* include/bits/unique_ptr.h (unique_ptr::operator*): Add
conditional noexcept as per LWG 2762.
* testsuite/20_util/shared_ptr/observers/array.cc: Check that
dereferencing cannot throw.
* testsuite/20_util/shared_ptr/observers/get.cc: Likewise.
* testsuite/20_util/optional/observers/lwg2762.cc: New test.
* testsuite/20_util/unique_ptr/lwg2762.cc: New test.

Emit .file 0 directive earlier in DWARF 5

When the assembler supports it, the compiler automatically passes --gdwarf-5
to it, which has an interesting side effect: any assembly instruction prior
to the first .file directive defines a new line associated with .file 0 in
the .debug_line section and of course the numbering of these implicit lines
has nothing to do with that of the source code.  This can be problematic in
Ada when we do not generate .file/.loc directives for compiled-generated
functions to avoid too jumpy a debugging experience.

gcc/
* dwarf2out.c (dwarf2out_assembly_start): Emit .file 0 marker here..
(dwarf2out_finish): ...instead of here.

Fix --gdwarf-5 configure tests for Windows

The issues are that 1) they use readelf instead of objdump and 2) they use
ELF syntax in the assembly code.

gcc/
* configure.ac (--gdwarf-5 option): Use objdump instead of readelf.
(working --gdwarf-4/--gdwarf-5 for all sources): Likewise.
(--gdwarf-4 not refusing generated .debug_line): Adjust for Windows.
* configure: Regenerate.

Add cscope.out to git ignore.

ChangeLog:

* .gitignore: Add entry for cscope.out.

Merge vec_addsub patterns

This merges the vec_addsub<mode>3 patterns using a mode attribute
for the vec_merge merge operand.

2021-06-18  Richard Biener  <rguenther@suse.de>

* config/i386/sse.md (vec_addsubv4df3, vec_addsubv2df3,
vec_addsubv8sf3, vec_addsubv4sf3): Merge into ...
(vec_addsub<mode>3): ... using a new addsub_cst mode attribute.

Add x86 addsub SLP pattern

This addds SLP pattern recognition for the SSE3/AVX [v]addsubp{ds} v0, v1
instructions which compute { v0[0] - v1[0], v0[1], + v1[1], ... }
thus subtract, add alternating on lanes, starting with subtract.

It adds a corresponding optab and direct internal function,
vec_addsub$a3 and renames the existing i386 backend patterns to
the new canonical name.

The SLP pattern matches the exact alternating lane sequence rather
than trying to be clever and anticipating incoming permutes - we
could permute the two input vectors to the needed lane alternation,
do the addsub and then permute the result vector back but that's
only profitable in case the two input or the output permute will
vanish - something Tamars refactoring of SLP pattern recog should
make possible.

2021-06-17  Richard Biener  <rguenther@suse.de>

* config/i386/sse.md (avx_addsubv4df3): Rename to
vec_addsubv4df3.
(avx_addsubv8sf3): Rename to vec_addsubv8sf3.
(sse3_addsubv2df3): Rename to vec_addsubv2df3.
(sse3_addsubv4sf3): Rename to vec_addsubv4sf3.
* config/i386/i386-builtin.def: Adjust.
* internal-fn.def (VEC_ADDSUB): New internal optab fn.
* optabs.def (vec_addsub_optab): New optab.
* tree-vect-slp-patterns.c (class addsub_pattern): New.
(slp_patterns): Add addsub_pattern.
* tree-vect-slp.c (vect_optimize_slp): Disable propagation
across CFN_VEC_ADDSUB.
* tree-vectorizer.h (vect_pattern::vect_pattern): Make
m_ops optional.
* doc/md.texi (vec_addsub<mode>3): Document.

* gcc.target/i386/vect-addsubv2df.c: New testcase.
* gcc.target/i386/vect-addsubv4sf.c: Likewise.
* gcc.target/i386/vect-addsubv4df.c: Likewise.
* gcc.target/i386/vect-addsubv8sf.c: Likewise.
* gcc.target/i386/vect-addsub-2.c: Likewise.
* gcc.target/i386/vect-addsub-3.c: Likewise.

df: Fix up handling of paradoxical subregs in debug insns [PR101170]

The recent addition of gcc_assert (regno < endregno); triggers during
glibc build on m68k.
The problem is that RA decisions shouldn't depend on expressions in
DEBUG_INSNs and those expressions can contain paradoxical subregs of certain
pseudos.  If RA then decides to allocate the pseudo to a register
with very small hard register REGNO, we can trigger the new assert,
as (int) subreg_regno_offset may be negative on big endian and the small
REGNO + the negative offset can wrap around.

The following patch in that case records the range from the REGNO 0 to
endregno, before the addition of the assert as both regno and endregno are
unsigned it wouldn't record anything at all silently.

2021-06-24  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/101170
* df-scan.c (df_ref_record): For paradoxical big-endian SUBREGs
where regno + subreg_regno_offset wraps around use 0 as starting
regno.

* gcc.dg/pr101170.c: New test.

stor-layout: Avoid DECL_BIT_FIELD_REPRESENTATIVE with NULL TREE_TYPE [PR101172]

finish_bitfield_representative has an early out if the field after a
bitfield has error_mark_node type, but that early out leads to TREE_TYPE
of the DECL_BIT_FIELD_REPRESENTATIVE being NULL, which breaks assumptions
on code that uses the DECL_BIT_FIELD_REPRESENTATIVE during error-recovery.

The following patch instead sets TREE_TYPE of the representative to
error_mark_node, something the users can deal with better.  At this point
the representative can be set as DECL_BIT_FIELD_REPRESENTATIVE for multiple
bitfields, so making sure that we clear the DECL_BIT_FIELD_REPRESENTATIVE
instead would be harder (but doable, e.g. with the error_mark_node TREE_TYPE
set by this patch set some flag in the caller and if the flag is there, walk
all the fields once again and clear all DECL_BIT_FIELD_REPRESENTATIVE that
have error_mark_node TREE_TYPE).

2021-06-24  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/101172
* stor-layout.c (finish_bitfield_representative): If nextf has
error_mark_node type, set repr type to error_mark_node too.

* gcc.dg/pr101172.c: New test.

IBM Z: Define NO_PROFILE_COUNTERS

s390 glibc does not need counters in the .data section, since it stores
edge hits in its own data structure.  Therefore counters only waste
space and confuse diffing tools (e.g. kpatch), so don't generate them.

gcc/ChangeLog:

* config/s390/s390.c (s390_function_profiler): Ignore labelno
parameter.
* config/s390/s390.h (NO_PROFILE_COUNTERS): Define.

gcc/testsuite/ChangeLog:

* gcc.target/s390/mnop-mcount-m31-mzarch.c: Adapt to the new
prologue size.
* gcc.target/s390/mnop-mcount-m64.c: Likewise.

Fix SLP permute propagation error

This fixes SLP permute propagation to not propagate across operations
that have different semantics on different lanes like for example
the recently added COMPLEX_ADD_ROT90.

2021-06-24  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (vect_optimize_slp): Do not propagate
across operations that have different semantics on different
lanes.

openmp: in_reduction clause support on target construct

This patch adds support for in_reduction clause on target construct, though
for now only for synchronous targets (without nowait clause).
The encountering thread in that case runs the target task and blocks until
the target region ends, so it is implemented by remapping it before entering
the target, initializing the private copy if not yet initialized for the
current thread and then using the remapped addresses for the mapping
addresses.
For nowait combined with in_reduction the patch contains a hack where the
nowait clause is ignored.  To implement it correctly, I think we would need
to create a new private variable for the in_reduction and initialize it before
doing the async target and adjust the map addresses to that private variable
and then pass a function pointer to the library routine with code where the callback
would remap the address to the current threads private variable and use in_reduction
combiner to combine the private variable we've created into the thread's copy.
The library would then need to make sure that the routine is called in some thread
participating in the parallel (and not in an unshackeled thread).

2021-06-24  Jakub Jelinek  <jakub@redhat.com>

gcc/
* tree.h (OMP_CLAUSE_MAP_IN_REDUCTION): Document meaning for OpenMP.
* gimplify.c (gimplify_scan_omp_clauses): For OpenMP map clauses
with OMP_CLAUSE_MAP_IN_REDUCTION flag partially defer gimplification
of non-decl OMP_CLAUSE_DECL.  For OMP_CLAUSE_IN_REDUCTION on
OMP_TARGET user outer_ctx instead of ctx for placeholders and
initializer/combiner gimplification.
* omp-low.c (scan_sharing_clauses): Handle OMP_CLAUSE_MAP_IN_REDUCTION
on target constructs.
(lower_rec_input_clauses): Likewise.
(lower_omp_target): Likewise.
* omp-expand.c (expand_omp_target): Temporarily ignore nowait clause
on target if in_reduction is present.
gcc/c-family/
* c-common.h (enum c_omp_region_type): Add C_ORT_TARGET and
C_ORT_OMP_TARGET.
* c-omp.c (c_omp_split_clauses): For OMP_CLAUSE_IN_REDUCTION on
combined target constructs also add map (always, tofrom:) clause.
gcc/c/
* c-parser.c (omp_split_clauses): Pass C_ORT_OMP_TARGET instead of
C_ORT_OMP for clauses on target construct.
(OMP_TARGET_CLAUSE_MASK): Add in_reduction clause.
(c_parser_omp_target): For non-combined target add
map (always, tofrom:) clauses for OMP_CLAUSE_IN_REDUCTION.  Pass
C_ORT_OMP_TARGET to c_finish_omp_clauses.
* c-typeck.c (handle_omp_array_sections): Adjust ort handling
for addition of C_ORT_OMP_TARGET and simplify, mapping clauses are
never present on C_ORT_*DECLARE_SIMD.
(c_finish_omp_clauses): Likewise.  Handle OMP_CLAUSE_IN_REDUCTION
on C_ORT_OMP_TARGET, set OMP_CLAUSE_MAP_IN_REDUCTION on
corresponding map clauses.
gcc/cp/
* parser.c (cp_omp_split_clauses): Pass C_ORT_OMP_TARGET instead of
C_ORT_OMP for clauses on target construct.
(OMP_TARGET_CLAUSE_MASK): Add in_reduction clause.
(cp_parser_omp_target): For non-combined target add
map (always, tofrom:) clauses for OMP_CLAUSE_IN_REDUCTION.  Pass
C_ORT_OMP_TARGET to finish_omp_clauses.
* semantics.c (handle_omp_array_sections_1): Adjust ort handling
for addition of C_ORT_OMP_TARGET and simplify, mapping clauses are
never present on C_ORT_*DECLARE_SIMD.
(handle_omp_array_sections): Likewise.
(finish_omp_clauses): Likewise.  Handle OMP_CLAUSE_IN_REDUCTION
on C_ORT_OMP_TARGET, set OMP_CLAUSE_MAP_IN_REDUCTION on
corresponding map clauses.
* pt.c (tsubst_expr): Pass C_ORT_OMP_TARGET instead of C_ORT_OMP for
clauses on target construct.
gcc/testsuite/
* c-c++-common/gomp/target-in-reduction-1.c: New test.
* c-c++-common/gomp/clauses-1.c: Add in_reduction clauses on
target or combined target constructs.
libgomp/
* testsuite/libgomp.c-c++-common/target-in-reduction-1.c: New test.
* testsuite/libgomp.c-c++-common/target-in-reduction-2.c: New test.
* testsuite/libgomp.c++/target-in-reduction-1.C: New test.
* testsuite/libgomp.c++/target-in-reduction-2.C: New test.

predcom: Refactor more by encapsulating global states

This patch is to encapsulate global states into a class and
making their accessors as member functions, remove some
consequent useless clean up code, and do some clean up with
RAII.

Bootstrapped/regtested on powerpc64le-linux-gnu P9,
x86_64-redhat-linux and aarch64-linux-gnu, also
bootstrapped on ppc64le P9 with bootstrap-O3 config.

gcc/ChangeLog:

* tree-predcom.c (class pcom_worker): New class.
(release_chain): Renamed to...
(pcom_worker::release_chain): ...this.
(release_chains): Renamed to...
(pcom_worker::release_chains): ...this.
(aff_combination_dr_offset): Renamed to...
(pcom_worker::aff_combination_dr_offset): ...this.
(determine_offset): Renamed to...
(pcom_worker::determine_offset): ...this.
(class comp_ptrs): New class.
(split_data_refs_to_components): Renamed to...
(pcom_worker::split_data_refs_to_components): ...this,
and update with class comp_ptrs.
(suitable_component_p): Renamed to...
(pcom_worker::suitable_component_p): ...this.
(filter_suitable_components): Renamed to...
(pcom_worker::filter_suitable_components): ...this.
(valid_initializer_p): Renamed to...
(pcom_worker::valid_initializer_p): ...this.
(find_looparound_phi): Renamed to...
(pcom_worker::find_looparound_phi): ...this.
(add_looparound_copies): Renamed to...
(pcom_worker::add_looparound_copies): ...this.
(determine_roots_comp): Renamed to...
(pcom_worker::determine_roots_comp): ...this.
(determine_roots): Renamed to...
(pcom_worker::determine_roots): ...this.
(single_nonlooparound_use): Renamed to...
(pcom_worker::single_nonlooparound_use): ...this.
(remove_stmt): Renamed to...
(pcom_worker::remove_stmt): ...this.
(execute_pred_commoning_chain): Renamed to...
(pcom_worker::execute_pred_commoning_chain): ...this.
(execute_pred_commoning): Renamed to...
(pcom_worker::execute_pred_commoning): ...this.
(struct epcc_data): New member worker.
(execute_pred_commoning_cbck): Call execute_pred_commoning
with pcom_worker pointer.
(find_use_stmt): Renamed to...
(pcom_worker::find_use_stmt): ...this.
(find_associative_operation_root): Renamed to...
(pcom_worker::find_associative_operation_root): ...this.
(find_common_use_stmt): Renamed to...
(pcom_worker::find_common_use_stmt): ...this.
(combinable_refs_p): Renamed to...
(pcom_worker::combinable_refs_p): ...this.
(reassociate_to_the_same_stmt): Renamed to...
(pcom_worker::reassociate_to_the_same_stmt): ...this.
(stmt_combining_refs): Renamed to...
(pcom_worker::stmt_combining_refs): ...this.
(combine_chains): Renamed to...
(pcom_worker::combine_chains): ...this.
(try_combine_chains): Renamed to...
(pcom_worker::try_combine_chains): ...this.
(prepare_initializers_chain): Renamed to...
(pcom_worker::prepare_initializers_chain): ...this.
(prepare_initializers): Renamed to...
(pcom_worker::prepare_initializers): ...this.
(prepare_finalizers_chain): Renamed to...
(pcom_worker::prepare_finalizers_chain): ...this.
(prepare_finalizers): Renamed to...
(pcom_worker::prepare_finalizers): ...this.
(tree_predictive_commoning_loop): Renamed to...
(pcom_worker::tree_predictive_commoning_loop): ...this, adjust
some calls and remove some cleanup code.
(tree_predictive_commoning): Adjusted to use pcom_worker instance.
(static variable looparound_phis): Remove.
(static variable name_expansions): Remove.

refactor SLP permute propagation

This refactors SLP permute propagation to record the outgoing permute
separately from the incoming/materialized one.  Instead of separate
arrays/bitmaps I've now created a struct to represent the state.

2021-06-23  Richard Biener  <rguenther@suse.de>

* tree-vect-slp.c (slpg_vertex): New struct.
(vect_slp_build_vertices): Adjust.
(vect_optimize_slp): Likewise.  Maintain an outgoing permute
and a materialized one.

tree-optimization/101105 - fix runtime alias test optimization

We were ignoring DR_STEP for VF == 1 which is OK only in case
the scalar order is preserved or both DR steps are the same.

2021-06-23  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101105
* tree-vect-data-refs.c (vect_prune_runtime_alias_test_list):
Only ignore steps when they are equal or scalar order is preserved.

* gcc.dg/torture/pr101105.c: New testcase.

i386: Add vashlm3/vashrm3/vlshrm3 to enable vectorization of vector shift vector. [PR98434]

Add expanders for vashl<VI12_AVX512BW>, vlshr<VI12_AVX512BW>,
vashr<VI1_AVX512BW> and vashr<v32hi,v16hi,v4di,v8di>.

Besides there's some assumption in expand_mult_const that mul and
add must be available at the same time, but for i386, addv8qi is
restricted under TARGET_64BIT, but mulv8qi not, that could cause ICE.
So restrict mulv8qi and shiftv8qi under TARGET_64BIT.

gcc/ChangeLog:

PR target/98434
* config/i386/i386-expand.c (ix86_expand_vec_interleave):
Adjust comments for ix86_expand_vecop_qihi2.
(ix86_expand_vecmul_qihi): Renamed to ..
(ix86_expand_vecop_qihi2): Adjust function prototype to
support shift operation, add static to definition.
(ix86_expand_vec_shift_qihi_constant): Add static to definition.
(ix86_expand_vecop_qihi): Call ix86_expand_vecop_qihi2 and
ix86_expand_vec_shift_qihi_constant.
* config/i386/i386-protos.h (ix86_expand_vecmul_qihi): Deleted.
(ix86_expand_vec_shift_qihi_constant): Deleted.
* config/i386/sse.md (VI12_256_512_AVX512VL): New mode
iterator.
(mulv8qi3): Call ix86_expand_vecop_qihi directly, add
condition TARGET_64BIT.
(mul<mode>3): Ditto.
(<insn><mode>3): Ditto.
(vlshr<mode>3): Extend to support avx512 vlshr.
(v<insn><mode>3): New expander for
vashr/vlshr/vashl.
(v<insn>v8qi3): Ditto.
(vashrv8hi3<mask_name>): Renamed to ..
(vashr<mode>3): And extend to support V16QImode for avx512.
(vashrv16qi3): Deleted.
(vashrv2di3<mask_name>): Extend expander to support avx512
instruction.

gcc/testsuite/ChangeLog:

PR target/98434
* gcc.target/i386/pr98434-1.c: New test.
* gcc.target/i386/pr98434-2.c: New test.
* gcc.target/i386/avx512vl-pr95488-1.c: Adjust testcase.

Daily bump.

c++: excessive instantiation during CTAD [PR101174]

We set DECL_CONTEXT on implicitly generated deduction guides so that
their access is consistent with that of the constructor.  But this
apparently leads to excessive instantiation in some cases, ultimately
because instantiation of a deduction guide should be independent of
instantiation of the resulting class specialization, but setting the
DECL_CONTEXT of the former to the latter breaks this independence.

To fix this, this patch makes push_access_scope handle artificial
deduction guides specifically rather than setting their DECL_CONTEXT
in build_deduction_guide.  We could alternatively make the class
befriend the guide via DECL_BEFRIENDING_CLASSES, but that wouldn't
be a complete fix and would break class-deduction-access3.C below
since friendship isn't transitive.

PR c++/101174

gcc/cp/ChangeLog:

* pt.c (push_access_scope): For artificial deduction guides,
set the access scope to that of the constructor.
(pop_access_scope): Likewise.
(build_deduction_guide): Don't set DECL_CONTEXT on the guide.

libstdc++-v3/ChangeLog:

* testsuite/23_containers/multiset/cons/deduction.cc:
Uncomment CTAD example that was rejected by this bug.
* testsuite/23_containers/set/cons/deduction.cc: Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction-access3.C: New test.
* g++.dg/cpp1z/class-deduction91.C: New test.

doc/lto.texi: List slim object format as the default

Slim LTO object files have been the default for quite a while, since:
  commit e9f67e625c2a4225a7169d7220dcb85b6fdd7ca9
  Author:     Jan Hubicka <hubicka@gcc.gnu.org>
  common.opt (ffat-lto-objects): Disable by default.

That commit did not update lto.texi, so do it now.

gcc/ChangeLog:

* doc/lto.texi (Design Overview): Update that slim objects are
the default.

Signed-off-by: Dimitar Dimitrov <dimitar@dinux.eu>

fortran/dump-parse-tree.c: Use proper enum type

gcc/fortran/ChangeLog:

* dump-parse-tree.c (show_omp_clauses): Fix enum type used
for dumping gfc_omp_defaultmap_category.

Do not enable pcrel-opt by default

SPEC2017 testing on p10 shows that this optimization does not have a
positive impact on performance. So we are no longer going to enable it
by default. The test cases for it needed to be updated so they always
enable it to test it.

gcc/

* config/rs6000/rs6000-cpus.def: Take OPTION_MASK_PCREL_OPT out
of OTHER_POWER10_MASKS so it will not be enabled by default.

gcc/testsuite/

* gcc.target/powerpc/pcrel-opt-inc-di.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-df.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-di.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-hi.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-qi.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-sf.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-si.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-ld-vector.c: Enable -mpcrel-opt to
test it.
* gcc.target/powerpc/pcrel-opt-st-df.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-di.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-hi.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-qi.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-sf.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-si.c: Enable -mpcrel-opt to test it.
* gcc.target/powerpc/pcrel-opt-st-vector.c: Enable -mpcrel-opt to
test it.

testsuite: add -fwrapv for 950704-1.c

gcc/testsuite
* gcc.c-torture/execute/950704-1.c: Add -fwrapv to avoid
undefined behavior.

libstdc++: Fix comment in chrono::year::is_leap()

libstdc++-v3/ChangeLog:

* include/std/chrono (chrono::year::is_leap()): Fix incorrect
logic in comment.

libstdc++: Document simd testsuite

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* testsuite/experimental/simd/README.md: New file.

libstdc++: Improve output verbosity options and default

For most uses --quiet was too quiet while the default was too noisy. Now
the default output, if stdout is a tty, shows the last successful test
on the same line. With --percentage it adds a percentage at the start of
the line. --percentage is not default because it requires more resources
and might not be 100% compatible to all environments.
If stdout is not a tty the default is quiet output like for dejagnu.

Additionally, argument parsing now recognizes contracted short options
which is easier to use with e.g. DRIVEROPTS=-pxk.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* testsuite/experimental/simd/driver.sh: Rewrite output
verbosity logic. Add -p/--percentage option. Allow -v/--verbose
to be used twice. Add -x and -o short options. Parse long
options with = instead of separating space generically. Parce
contracted short options. Make unrecognized options an error.
If same-line output is active, trap on EXIT to increment the
progress (only with --percentage), erase the line and print the
current status.
* testsuite/experimental/simd/generate_makefile.sh: Initialize
helper files for progress account keeping. Update help target
for changes to DRIVEROPTS.

libstdc++: Remove -fno-tree-vrp after PR98834 was resolved

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* testsuite/Makefile.am (check-simd): Remove -fno-tree-vrp flag
and associated warning.
* testsuite/Makefile.in: Regenerate.

libstdc++: More efficient std::chrono::year::leap

Simple change to std::chrono::year::is_leap. If a year is multiple of 100,
then it's divisible by 400 if and only if it's divisible by 16. The latter
allows for better code generation.

The expression is then either y%16 or y%4 which are both powers of two
and so it can be rearranged to use simple bitmask operations.

Co-authored-by: Jonathan Wakely <jwakely@redhat.com>
Co-authored-by: Ulrich Drepper <drepper@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/chrono (chrono::year::is_leap()): Optimize.

tree-inline: Fix TREE_READONLY of parameter replacements

tree-inline leaves behind VAR_DECLs which are TREE_READONLY (because
they are copies of const parameters) but are written to because they
need to be initialized.  This patch resets the flag unconditionally so
that this does not happen.

There are other sources of variables which are incorrectly marked as
TREE_READOLY, but with this patch and a verifier catching them I can
at least compile the Ada run-time library.

gcc/ChangeLog:

2021-06-22  Richard Biener  <rguenther@suse.de>
    Martin Jambor  <mjambor@suse.cz>

* tree-inline.c (setup_one_parameter): Set TREE_READONLY of the
param replacement unconditionally.  Adjust comment.

Split gimple-range into gimple-range-fold and gimple-range.

Split the fold_using_range functions from gimple-range into gimple-range-fold.
Also move the gimple_range_calc* routines into gimple-range-gori.

* Makefile.in (OBJS): Add gimple-range-fold.o
* gimple-range-fold.cc: New.
* gimple-range-fold.h: New.
* gimple-range-gori.cc (gimple_range_calc_op1): Move to here.
(gimple_range_calc_op2): Ditto.
* gimple-range-gori.h: Move prototypes to here.
* gimple-range.cc: Adjust include files.
(fur_source:fur_source): Relocate to gimple-range-fold.cc.
(fur_source::get_operand): Ditto.
(fur_source::get_phi_operand): Ditto.
(fur_source::query_relation): Ditto.
(fur_source::register_relation): Ditto.
(class fur_edge): Ditto.
(fur_edge::fur_edge): Ditto.
(fur_edge::get_operand): Ditto.
(fur_edge::get_phi_operand): Ditto.
(fur_stmt::fur_stmt): Ditto.
(fur_stmt::get_operand): Ditto.
(fur_stmt::get_phi_operand): Ditto.
(fur_stmt::query_relation): Ditto.
(class fur_depend): Relocate to gimple-range-fold.h.
(fur_depend::fur_depend): Relocate to gimple-range-fold.cc.
(fur_depend::register_relation): Ditto.
(fur_depend::register_relation): Ditto.
(class fur_list): Ditto.
(fur_list::fur_list): Ditto.
(fur_list::get_operand): Ditto.
(fur_list::get_phi_operand): Ditto.
(fold_range): Ditto.
(adjust_pointer_diff_expr): Ditto.
(gimple_range_adjustment): Ditto.
(gimple_range_base_of_assignment): Ditto.
(gimple_range_operand1): Ditto.
(gimple_range_operand2): Ditto.
(gimple_range_calc_op1): Relocate to gimple-range-gori.cc.
(gimple_range_calc_op2): Ditto.
(fold_using_range::fold_stmt): Relocate to gimple-range-fold.cc.
(fold_using_range::range_of_range_op): Ditto.
(fold_using_range::range_of_address): Ditto.
(fold_using_range::range_of_phi): Ditto.
(fold_using_range::range_of_call): Ditto.
(fold_using_range::range_of_builtin_ubsan_call): Ditto.
(fold_using_range::range_of_builtin_call): Ditto.
(fold_using_range::range_of_cond_expr): Ditto.
(fold_using_range::range_of_ssa_name_with_loop_info): Ditto.
(fold_using_range::relation_fold_and_or): Ditto.
(fold_using_range::postfold_gcond_edges): Ditto.
* gimple-range.h: Add gimple-range-fold.h to include files. Change
GIMPLE_RANGE_STMT_H to GIMPLE_RANGE_H.
(gimple_range_handler): Relocate to gimple-range-fold.h.
(gimple_range_ssa_p): Ditto.
(range_compatible_p): Ditto.
(class fur_source): Ditto.
(class fur_stmt): Ditto.
(class fold_using_range): Ditto.
(gimple_range_calc_op1): Relocate to gimple-range-gori.h
(gimple_range_calc_op2): Ditto.

Do not continue propagating values which cannot be set properly.

If the on-entry cache cannot properly represent a range, do not continue
trying to propagate it.

PR tree-optimization/101148
PR tree-optimization/101014
* gimple-range-cache.cc (ranger_cache::ranger_cache): Adjust.
(ranger_cache::~ranger_cache): Adjust.
(ranger_cache::block_range): Check if propagation disallowed.
(ranger_cache::propagate_cache): Disallow propagation if new value
can't be stored properly.
* gimple-range-cache.h (ranger_cache::m_propfail): New member.

Adjust on_entry cache to indicate if the value was set properly.

* gimple-range-cache.cc (class ssa_block_ranges): Adjust prototype.
(sbr_vector::set_bb_range): Return true.
(class sbr_sparse_bitmap): Adjust.
(sbr_sparse_bitmap::set_bb_range): Return value.
(block_range_cache::set_bb_range): Return value.
(ranger_cache::propagate_cache): Use return value to print msg.
* gimple-range-cache.h (class block_range_cache): Adjust.

Dump should be read only. Do not trigger new lookups.

* gimple-range.cc (dump_bb): Use range_on_edge from the cache.

Use more logicals to eliminate useless test/compare instructions

gcc/
* config/h8300/logical.md (<code><mode>3<ccnz>): Use <cczn>
so this pattern can be used for test/compare removal.  Pass
current insn to compute_logical_op_length and output_logical_op.
* config/h8300/h8300.c (compute_logical_op_cc): Remove.
(h8300_and_costs): Add argument to compute_logical_op_length.
(output_logical_op): Add new argument.  Use it to determine if the
condition codes are used and adjust the output accordingly.
(compute_logical_op_length): Add new argument and update length
computations when condition codes are used.
* config/h8300/h8300-protos.h (compute_logical_op_length): Update
prototype.
(output_logical_op): Likewise.

i386: Add PPERM two-operand 64bit vector permutation [PR89021]

Add emulation of V8QI PPERM permutations for TARGET_XOP target.  Similar
to PSHUFB, the permutation is performed with V16QI PPERM instruction,
where selector is defined in V16QI mode with inactive elements set to 0x80.
Specific to two operand permutations is the remapping of elements from
the second operand (e.g. e[8] -> e[16]), as we have to account for the
inactive elements from the first operand.

2021-06-23  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/89021
* config/i386/i386-expand.c (expand_vec_perm_pshufb):
Handle 64bit modes for TARGET_XOP.  Use indirect gen_* functions.
* config/i386/mmx.md (mmx_ppermv64): New insn pattern.
* config/i386/i386.md (unspec): Move UNSPEC_XOP_PERMUTE from ...
* config/i386/sse.md (unspec): ... here.

arm: Revert partially ebd5e86c0f41dc1d692f9b2b68a510b1f6835a3e

PR target/98636

gcc/ChangeLog:

* optc-save-gen.awk: Put back arm_fp16_format to
checked_options.

c++: CTAD and deduction guide selection [PR86439]

During CTAD, we select the best viable deduction guide using
build_new_function_call, which performs overload resolution on the set
of candidate guides and then forms a call to the guide.  As the PR
points out, this latter step is unnecessary and occasionally incorrect
since a call to the selected guide may be ill-formed, or forming the
call may have side effects such as prematurely deducing the type of a {}.

So this patch introduces a specialized subroutine based on
build_new_function_call that stops short of building a call to the
selected function, and makes do_class_deduction use this subroutine
instead.  And since a call is no longer built, do_class_deduction
doesn't need to set tf_decltype or cp_unevaluated_operand anymore.

This change causes us to reject some container CTAD examples in the
libstdc++ testsuite due to deduction failure for {}, which AFAICT is the
correct behavior.  Previously in e.g. the first removed example

  std::map{{std::pair{1, 2.0}, {2, 3.0}, {3, 4.0}}, {}},

the type of the {} would get deduced to less<int> as a side effect of
forming a call to the chosen guide

  template<typename _Key, typename _Tp, typename _Compare = less<_Key>,
           typename _Allocator = allocator<pair<const _Key, _Tp>>>
      map(initializer_list<pair<_Key, _Tp>>,
          _Compare = _Compare(), _Allocator = _Allocator())
      -> map<_Key, _Tp, _Compare, _Allocator>;

which made later overload resolution for the constructor call
unambiguous.  Now, the type of the {} remains undeduced until
constructor overload resolution, and we complain about ambiguity
for the two equally good constructor candidates

  map(initializer_list<value_type>,
      const _Compare& = _Compare(),
      const allocator_type& = allocator_type())

  map(initializer_list<value_type>, const allocator_type&).

This patch fixes these problematic container CTAD examples by giving
the {} an appropriate concrete type.  Two of these adjusted CTAD
examples (one for std::set and one for std::multiset) end up triggering
an unrelated CTAD bug on trunk, PR101174, so these two adjusted examples
are commented out for now.

PR c++/86439

gcc/cp/ChangeLog:

* call.c (print_error_for_call_failure): Constify 'args' parameter.
(perform_dguide_overload_resolution): Define.
* cp-tree.h: (perform_dguide_overload_resolution): Declare.
* pt.c (do_class_deduction): Use perform_dguide_overload_resolution
instead of build_new_function_call.  Don't use tf_decltype or
set cp_unevaluated_operand.  Remove unnecessary NULL_TREE tests.

libstdc++-v3/ChangeLog:

* testsuite/23_containers/map/cons/deduction.cc: Replace ambiguous
CTAD examples.
* testsuite/23_containers/multimap/cons/deduction.cc: Likewise.
* testsuite/23_containers/multiset/cons/deduction.cc: Likewise.
Mention one of the replaced examples is broken due to PR101174.
* testsuite/23_containers/set/cons/deduction.cc: Likewise.
* testsuite/23_containers/unordered_map/cons/deduction.cc: Replace
ambiguous CTAD examples.
* testsuite/23_containers/unordered_multimap/cons/deduction.cc:
Likewise.
* testsuite/23_containers/unordered_multiset/cons/deduction.cc:
Likewise.
* testsuite/23_containers/unordered_set/cons/deduction.cc: Likewise.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1z/class-deduction88.C: New test.
* g++.dg/cpp1z/class-deduction89.C: New test.
* g++.dg/cpp1z/class-deduction90.C: New test.

i386: Prevent unwanted combine from LZCNT to BSR [PR101175]

The current RTX pattern for BSR allows combine pass to convert LZCNT insn
to BSR. Note that the LZCNT has a defined behavior to return the operand
size when operand is zero, where BSR has not.

Add a BSR specific setting of zero-flag to RTX pattern of BSR insn
in order to avoid matching unwanted combinations.

2021-06-23  Uroš Bizjak  <ubizjak@gmail.com>

gcc/
PR target/101175
* config/i386/i386.md (bsr_rex64): Add zero-flag setting RTX.
(bsr): Ditto.
(*bsrhi): Remove.
(clz<mode>2): Update RTX pattern for additions.

gcc/testsuite/

PR target/101175
* gcc.target/i386/pr101175.c: New test.

libstdc++: Avoid "__lockable" name defined as macro by newlib

libstdc++-v3/ChangeLog:

* include/std/mutex (__detail::__try_lock_impl): Rename
parameter to avoid clashing with newlib's __lockable macro.
(try_lock): Add 'inline' specifier.
* testsuite/17_intro/names.cc: Add check for __lockable.
* testsuite/30_threads/try_lock/5.cc: Add options for pthreads.

fortran: Fix deref of optional in gen. code. [PR100337]

gcc/fortran/ChangeLog:

PR fortran/100337
* trans-intrinsic.c (conv_co_collective): Check stat for null ptr
before dereferrencing.

gcc/testsuite/ChangeLog:

PR fortran/100337
* gfortran.dg/coarray_collectives_17.f90: New test.

openmp: Fix up *_reduction clause handling with UDRs on PARM_DECLs [PR101167]

The following testcase FAILs, because the UDR combiner is invoked incorrectly.
lower_omp_rec_clauses expects that when it sets
DECL_VALUE_EXPR/DECL_HAS_VALUE_EXPR_P
for both the placeholder and the var that everything will be properly
regimplified, but as the variable in question is a PARM_DECL rather than
VAR_DECL, lower_omp_regimplify_p doesn't say that it should be regimplified
and so it is not.

2021-06-23  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/101167
* omp-low.c (lower_omp_regimplify_p): Regimplify also PARM_DECLs
and RESULT_DECLs that have DECL_HAS_VALUE_EXPR_P set.

* testsuite/libgomp.c-c++-common/task-reduction-15.c: New test.

contrib: add git-commit-mklog wrapper

contrib/ChangeLog:

* gcc-git-customization.sh: Use the new wrapper.
* git-commit-mklog.py: New file.
* prepare-commit-msg: Support GCC_MKLOG_ARGS.

rs6000: Fix typos in float128 ISA3.1 support

The recent float128 ISA3.1 support (r12-1340) has some typos,
it makes the libgcc build fail if it's with one binutils
(assembler) which doesn't support Power10 insns.  The error
looks like:

Error: invalid switch -mpower10
Error: unrecognized option -mpower10
... [...libgcc/shared-object.mk:14: float128-p10.o] Error 1

What this patch does are:
  - fix test target typo libgcc_cv_powerpc_3_1_float128_hw
    (written wrongly as libgcc_cv_powerpc_float128_hw, so it's
     going to build ISA3.1 stuffs just when detecting ISA3.0).
  - fix test used for libgcc_cv_powerpc_3_1_float128_hw check.
  - fix test option used for libgcc_cv_powerpc_3_1_float128_hw
    check.
  - remove the ISA3.1 related contents from t-float128-hw.
  - add new macro FLOAT128_HW_INSNS_ISA3_1 to differentiate
    ISA3.1 content from ISA3.0 part in ifunc support.

Bootstrapped/regtested on:
  - powerpc64le-linux-gnu P10
  - powerpc64le-linux-gnu P9 (w/i and w/o p10 supported as)
  - powerpc64-linux-gnu P8 (w/i and w/o p10 supported as)

libgcc/ChangeLog:

* configure: Regenerate.
* configure.ac (test for libgcc_cv_powerpc_3_1_float128_hw): Fix
typos among the name, CFLAGS and the test.
* config/rs6000/t-float128-hw (fp128_3_1_hw_funcs, fp128_3_1_hw_src,
fp128_3_1_hw_static_obj, fp128_3_1_hw_shared_obj, fp128_3_1_hw_obj):
Remove.
* config/rs6000/t-float128-p10-hw (FLOAT128_HW_INSNS): Append
macro FLOAT128_HW_INSNS_ISA3_1.
(FP128_3_1_CFLAGS_HW): Fix option typo.
* config/rs6000/float128-ifunc.c (SW_OR_HW_ISA3_1): Guard this with
FLOAT128_HW_INSNS_ISA3_1.
(__floattikf_resolve): Likewise.
(__floatuntikf_resolve): Likewise.
(__fixkfti_resolve): Likewise.
(__fixunskfti_resolve): Likewise.
(__floattikf): Likewise.
(__floatuntikf): Likewise.
(__fixkfti): Likewise.
(__fixunskfti): Likewise.

Daily bump.

libstdc++: Simplify std::try_lock and std::lock further

The std::try_lock and std::lock algorithms can use iteration instead of
recursion when all lockables have the same type and can be held by an
array of unique_lock<L> objects.

By making this change to __detail::__try_lock_impl it also benefits
__detail::__lock_impl, which uses it. For std::lock we can just put the
iterative version directly in std::lock, to avoid making any call to
__detail::__lock_impl.

Signed-off-by: Matthias Kretz <m.kretz@gsi.de>
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
Co-authored-by: Matthias Kretz <m.kretz@gsi.de>
libstdc++-v3/ChangeLog:

* include/std/mutex (lock): Replace recursion with iteration
when lockables all have the same type.
(__detail::__try_lock_impl): Likewise. Pass lockables as
parameters, instead of a tuple. Always lock the first one, and
recurse for the rest.
(__detail::__lock_impl): Adjust call to __try_lock_impl.
(__detail::__try_to_lock): Remove.
* testsuite/30_threads/lock/3.cc: Check that mutexes are locked.
* testsuite/30_threads/lock/4.cc: Also test non-heterogeneous
arguments.
* testsuite/30_threads/unique_lock/cons/60497.cc: Also check
std::try_lock.
* testsuite/30_threads/try_lock/5.cc: New test.

libstdc++: Remove garbage collection support for C++23 [P2186R2]

This removes the non-functional garbage colection support from <memory>,
as proposed for C++23 by P2186R2.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/std/memory (declare_reachable, undeclare_reachable)
(declare_no_pointers, undeclare_no_pointers, get_pointer_safety)
(pointer_safety): Only define for C++11 to C++20 inclusive.
* testsuite/20_util/pointer_safety/1.cc: Do not run for C++23.

libstdc++: Implement LWG 3422 for std::seed_seq

This ensures that the std::seed_seq initializer-list constructor will
not be used for list-initialization unless the initializers in the list
are integers. This allows list-initialization syntax to be used with a
pair of pointers and for that to use the appropriate constructor.

Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:

* include/bits/random.h (seed_seq): Constrain initializer-list
constructor.
* include/bits/random.tcc (seed_seq): Add template parameter.
* testsuite/26_numerics/random/seed_seq/cons/default.cc: Check
for noexcept.
* testsuite/26_numerics/random/seed_seq/cons/initlist.cc: Check
constraints.

Fortran: fix sm computation in CFI_allocate [PR93524]

This patch fixes a bug in setting the step multiplier field in the
C descriptor for array dimensions > 2.

2021-06-21  Sandra Loosemore  <sandra@codesourcery.com>
    Tobias Burnus  <tobias@codesourcery.com>

libgfortran/
PR fortran/93524
* runtime/ISO_Fortran_binding.c (CFI_allocate): Fix
sm computation.

gcc/testsuite/
PR fortran/93524
* gfortran.dg/pr93524.c: New.
* gfortran.dg/pr93524.f90: New.

libstdc++: Fix for deadlock in std::counting_semaphore [PR100806]

libstdc++-v3/ChangeLog:
PR libstdc++/100806
* include/bits/semaphore_base.h (__atomic_semaphore::_M_release):
Force _M_release() to wake all waiting threads.
* testsuite/30_threads/semaphore/100806.cc: New test.

analyzer: fix ICE on malloc/alloca param type mismatch [PR101143]

gcc/analyzer/ChangeLog:
PR analyzer/101143
* region-model.cc (compat_types_p): New function.
(region_model::create_region_for_heap_alloc): Convert assertion to
an error check.
(region_model::create_region_for_alloca): Likewise.

gcc/testsuite/ChangeLog:
PR analyzer/101143
* gcc.dg/analyzer/pr101143.c: New test.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>

docs: drop unbalanced parenthesis in rtl.texi

gcc/ChangeLog:

* doc/rtl.texi: drop unbalanced parenthesis.

middle-end/101156 - remove not working optimization in gimplification

This removes a premature and not working optimization from the
gimplifier.  When gimplification is requested not to produce a SSA
name we try to avoid generating a copy when we did so anyway but
instead replace the LHS of its definition.  But that only works in
case there are no uses of the SSA name already which is something
we cannot easily check, so the following removes said optimization.

Statistics on the whole bootstrap shows we hit this optimization
only for libiberty/cp-demangle.c and overall we have 21652112
gimplifications where just 240 copies are elided.  Preserving
the optimization would require scanning the original expression
and the pre and post sequences for SSA names and uses, that seems
excessive to avoid these 240 copies.

2021-06-22  Richard Biener  <rguenther@suse.de>

PR middle-end/101156
* gimplify.c (gimplify_expr): Remove premature incorrect
optimization.

* gcc.dg/pr101156.c: New testcase.

testsuite: Add testcase for recently fixed PR [PR101159]

On Tue, Jun 22, 2021 at 11:00:51AM +0200, Richard Biener wrote:
> 2021-06-22  Richard Biener  <rguenther@suse.de>
>
>       PR tree-optimization/101159
>       * tree-vect-patterns.c (vect_recog_popcount_pattern): Add
>       missing NULL vectype check.

The following patch adds the testcase for it, IMHO it can't hurt and
from my experience testcases often trigger other bugs later on (rather
than the original bugs reappearing, though even that happens),
and also fixes a couple of typos in the new function.

2021-06-22  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/101159
* tree-vect-patterns.c (vect_recog_popcount_pattern): Fix some
comment typos.

* gcc.c-torture/compile/pr101159.c: New test.

expand: Fix up empty class return optimization [PR101160]

On Mon, Jun 14, 2021 at 11:24:22PM -0400, Jason Merrill via Gcc-patches wrote:
> The x86_64 psABI says that an empty class isn't passed or returned in memory or
> registers, so we shouldn't set %eax in this function.  Is this a reasonable
> place to implement that?  Another possibility would be to remove the hack to
> prevent i386.c:function_value_64 from returning NULL in this case and fix the
> callers to deal, but that seems like more work.
>
> The df-scan hunk catches the case where we look at a 0-length reg and build
> a range the length of unsigned int, which happened before I changed
> assign_parms to match expand_function_end.

The assign_params change unfortunately breaks e.g. the following testcase.
The problem is that some passes (e.g. subreg lowering but assign_parms
comments also talk about delayed slot scheduling) rely on crtl->return_rtx
not to contain pseudo registers, and the assign_parms change results
in the pseudo in there not being replaced with a hard register.

The following patch instead clears the crtl->return_rtx if a function
returns TYPE_EMPTY_P structure, that way (use (pseudo)) is not emitted
into the IL and it is treated like more like functions returning void.

I've also changed the effective target on the empty-class1.C testcase, so
that it doesn't fail on x86_64-linux with -m32 testing.

2021-06-22  Jakub Jelinek  <jakub@redhat.com>

PR middle-end/101160
* function.c (assign_parms): For decl_result with TYPE_EMPTY_P type
clear crtl->return_rtx instead of keeping it referencing a pseudo.

* g++.target/i386/empty-class1.C: Require lp64 effective target
instead of x86_64-*-*.
* g++.target/i386/empty-class2.C: New test.

fold-const: Return corresponding integral type for OFFSET_TYPE in range_check_type [PR101162]

Andrew's recent r12-1608-g2f1686ff70b25fceb04ca2ffc0a450fb682913ef change
to fail verification on various unary and binary operations with OFFSET_TYPE
revealed that e.g. switchconv happily performs multiplications and additions
in OFFSET_TYPE.

2021-06-22  Jakub Jelinek  <jakub@redhat.com>
    Andrew Pinski  <apinski@marvell.com>

PR tree-optimization/101162
* fold-const.c (range_check_type): Handle OFFSET_TYPE like pointer
types.

* g++.dg/opt/pr101162.C: New test.

Add relational self-tests.

* range-op.cc (range_relational_tests): New.
(range_op_tests): Call range_relational_tests.

Add relation between LHS and op1 for casts and copies.

* range-op.cc (operator_cast::lhs_op1_relation): New.
(operator_identity::lhs_op1_relation): Mew.

Add relation effects between operands to MINUS_EXPR.

* range-op.cc (operator_minus::op1_op2_relation_effect): New.

Add relations between LHS and op1/op2 for PLUS_EXPR.

* range-op.cc (operator_plus::lhs_op1_relation): New.
(operator_plus::lhs_op2_relation): New.

Add relational support to fold_using_range

Enable a relation oracle in ranger, and add full range-op relation support
to fold_using_range.

* gimple-range-cache.cc (ranger_cache::ranger_cache): Create a
relation_oracle if dominators exist.
(ranger_cache::~ranger_cache): Dispose of oracle.
(ranger_cache::dump_bb): Dump oracle.
* gimple-range.cc (fur_source::fur_source): New.
(fur_source::get_operand): Use mmeber query.
(fur_source::get_phi_operand): Use member_query.
(fur_source::query_relation): New.
(fur_source::register_dependency): Delete.
(fur_source::register_relation): New.
(fur_edge::fur_edge): Adjust.
(fur_edge::get_phi_operand): Fix comment.
(fur_edge::query): Delete.
(fur_stmt::fur_stmt): Adjust.
(fur_stmt::query): Delete.
(fur_depend::fur_depend): Adjust.
(fur_depend::register_relation): New.
(fur_depend::register_relation): New.
(fur_list::fur_list): Adjust.
(fur_list::get_operand): Use member query.
(fold_using_range::range_of_range_op): Process and query relations.
(fold_using_range::range_of_address): Adjust dependency call.
(fold_using_range::range_of_phi): Ditto.
(gimple_ranger::gimple_ranger): New.  Use ranger_ache oracle.
(fold_using_range::relation_fold_and_or): New.
(fold_using_range::postfold_gcond_edges): New.
* gimple-range.h (class gimple_ranger): Adjust.
(class fur_source): Adjust members.
(class fur_stmt): Ditto.
(class fold_using_range): Ditto.

Add relational support to range-op.

This patch integrates relations with range-op functionality so that any
known relations can be used to help reduce or resolve ranges.
Initially handle  EQ_EXPR, NE_EXPR, LE_EXPR, LT_EXPR, GT_EXPR and GE_EXPR.

* range-op.cc (range_operator::wi_fold): Apply relation effect.
(range_operator::fold_range): Adjust and apply relation effect.
(*::fold_range): Add relation parameters.
(*::op1_range): Ditto.
(*::op2_range): Ditto.
(range_operator::lhs_op1_relation): New.
(range_operator::lhs_op2_relation): New.
(range_operator::op1_op2_relation): New.
(range_operator::op1_op2_relation_effect): New.
(relop_early_resolve): New.
(operator_equal::op1_op2_relation): New.
(operator_equal::fold_range): Call relop_early_resolve.
(operator_not_equal::op1_op2_relation): New.
(operator_not_equal::fold_range): Call relop_early_resolve.
(operator_lt::op1_op2_relation): New.
(operator_lt::fold_range): Call relop_early_resolve.
(operator_le::op1_op2_relation): New.
(operator_le::fold_range): Call relop_early_resolve.
(operator_gt::op1_op2_relation): New.
(operator_gt::fold_range): Call relop_early_resolve.
(operator_ge::op1_op2_relation): New.
(operator_ge::fold_range): Call relop_early_resolve.
* range-op.h (class range_operator): Adjust parameters and methods.

Initial value-relation code.

This code provides a both an equivalence and relation oracle which can be
accessed via a range_query object.  This initial code drop includes the
oracles and access them, but does not utilize them yet.

* Makefile.in (OBJS): Add value-relation.o.
* gimple-range.h: Adjust include files.
* tree-data-ref.c: Adjust include file order.
* value-query.cc (range_query::get_value_range): Default to no oracle.
(range_query::query_relation): New.
(range_query::query_relation): New.
* value-query.h (class range_query): Adjust.
* value-relation.cc: New.
* value-relation.h: New.

tree-optimization/101151 - fix irreducible region check for sinking

The check whether two blocks are in the same irreducible region
and thus post-dominance checks being unreliable was incomplete
since an irreducible region can contain reducible sub-regions but
if one block is in the irreducible part and one not the check
still doesn't work as expected.

2021-06-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101151
* tree-ssa-sink.c (statement_sink_location): Expand irreducible
region check.

* gcc.dg/torture/pr101151.c: New testcase.

RISC-V: Add tune info for T-HEAD C906.

gcc/
* config/riscv/riscv.c (thead_c906_tune_info): New.
(riscv_tune_info_table): Use new tune.

testuite: Add pthread check to dg-module-cmi for omp module testing

gcc/testsuite:

* g++.dg/modules/omp-1_a.C: Check pthread is available for
dg-module-cmi.
* g++.dg/modules/omp-2_a.C: Ditto.

tree-optimization/101158 - adjust SLP call matching sequence

This moves the check for same operands after verifying we're
facing compatible calls.

2021-06-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101158
* tree-vect-slp.c (vect_build_slp_tree_1): Move same operand
checking after checking for matching operation.

* gfortran.dg/pr101158.f90: New testcase.

tree-optimization/101159 - fix missing NULL check in popcount pattern

This fixes a missing check for a NULL vectype in the new popcount
pattern.

2021-06-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101159
* tree-vect-patterns.c (vect_recog_popcount_pattern): Add
missing NULL vectype check.

tree-optimization/101154 - fix out-of bound access in SLP

This fixes an out-of-bound access of matches.

2021-06-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/101154
* tree-vect-slp.c (vect_build_slp_tree_2): Fix out-of-bound access.

i386: Use xor to write zero to memory with -Os even for more than 4 stores [PR11877]

> > 2021-06-20  Roger Sayle  <roger@nextmovesoftware.com>
> >
> > gcc/ChangeLog
> >         PR target/11877
> >         * config/i386/i386.md: New define_peephole2s to shrink writing
> >         1, 2 or 4 consecutive zeros to memory when optimizing for size.

It unfortunately doesn't extend well to larger memory clearing.
Consider e.g.
void
foo (int *p)
{
  p[0] = 0;
  p[7] = 0;
  p[23] = 0;
  p[41] = 0;
  p[48] = 0;
  p[59] = 0;
  p[69] = 0;
  p[78] = 0;
  p[83] = 0;
  p[89] = 0;
  p[98] = 0;
  p[121] = 0;
  p[132] = 0;
  p[143] = 0;
  p[154] = 0;
}
where with the patch we emit:
        xorl    %eax, %eax
        xorl    %edx, %edx
        xorl    %ecx, %ecx
        xorl    %esi, %esi
        xorl    %r8d, %r8d
        movl    %eax, (%rdi)
        movl    %eax, 28(%rdi)
        movl    %eax, 92(%rdi)
        movl    %eax, 164(%rdi)
        movl    %edx, 192(%rdi)
        movl    %edx, 236(%rdi)
        movl    %edx, 276(%rdi)
        movl    %edx, 312(%rdi)
        movl    %ecx, 332(%rdi)
        movl    %ecx, 356(%rdi)
        movl    %ecx, 392(%rdi)
        movl    %ecx, 484(%rdi)
        movl    %esi, 528(%rdi)
        movl    %esi, 572(%rdi)
        movl    %r8d, 616(%rdi)
Here is an incremental patch that emits:
        xorl    %eax, %eax
        movl    %eax, (%rdi)
        movl    %eax, 28(%rdi)
        movl    %eax, 92(%rdi)
        movl    %eax, 164(%rdi)
        movl    %eax, 192(%rdi)
        movl    %eax, 236(%rdi)
        movl    %eax, 276(%rdi)
        movl    %eax, 312(%rdi)
        movl    %eax, 332(%rdi)
        movl    %eax, 356(%rdi)
        movl    %eax, 392(%rdi)
        movl    %eax, 484(%rdi)
        movl    %eax, 528(%rdi)
        movl    %eax, 572(%rdi)
        movl    %eax, 616(%rdi)
instead.

2021-06-22  Jakub Jelinek  <jakub@redhat.com>

PR target/11877
* config/i386/i386-protos.h (ix86_last_zero_store_uid): Declare.
* config/i386/i386-expand.c (ix86_last_zero_store_uid): New variable.
* config/i386/i386.c (ix86_expand_prologue): Clear it.
* config/i386/i386.md (peephole2s for 1/2/4 stores of const0_rtx):
Remove "" from match_operand.  Emit new insns using emit_move_insn and
set ix86_last_zero_store_uid to INSN_UID of the last store.
Add peephole2s for 1/2/4 stores of const0_rtx following previous
successful peep2s.

* gcc.target/i386/pr11877-2.c: New test.

Remove my Write After Approval entry.

ChangeLog:

* MAINTAINERS: Remove my Write After Approval entry.

contrib: fix a flake8 issue

contrib/ChangeLog:

* mklog.py: Fix flake8 issue.

autofdo: Bump AUTO_PROFILE_VERSION.

gcc/ChangeLog:

* auto-profile.c (AUTO_PROFILE_VERSION): Bump as string format
was changed.

gcov: update comment about padding

gcc/ChangeLog:

* gcov-io.h: Remove padding entries.

Add vect_recog_popcount_pattern to handle mismatch between the vectorized popcount IFN and scalar popcount builtin.

The patch remove those pro- and demotions when backend support direct
optab.

For i386: it enables vectorization for vpopcntb/vpopcntw and optimized
for vpopcntq.

gcc/ChangeLog:

PR tree-optimization/97770
* tree-vect-patterns.c (vect_recog_popcount_pattern):
New.
(vect_recog_func vect_vect_recog_func_ptrs): Add new pattern.

gcc/testsuite/ChangeLog:

PR tree-optimization/97770
* gcc.target/i386/avx512bitalg-pr97770-1.c: Remove xfail.
* gcc.target/i386/avx512vpopcntdq-pr97770-1.c: Remove xfail.

Optimize vpexpand* to mask mov when mask have all ones in it's lower part (including 0 and -1).

gcc/ChangeLog:

PR target/100267
* config/i386/i386-builtin.def (BDESC): Adjust builtin name.
* config/i386/sse.md (<avx512>_expand<mode>_mask): Rename to ..
(expand<mode>_mask): this ..
(*expand<mode>_mask): New pre_reload splitter to transform
v{,p}expand* to vmov* when mask is zero, all ones, or has all
ones in it's lower part, otherwise still generate
v{,p}expand*.

gcc/testsuite/ChangeLog:

PR target/100267
* gcc.target/i386/avx512bw-pr100267-1.c: New test.
* gcc.target/i386/avx512bw-pr100267-b-2.c: New test.
* gcc.target/i386/avx512bw-pr100267-d-2.c: New test.
* gcc.target/i386/avx512bw-pr100267-q-2.c: New test.
* gcc.target/i386/avx512bw-pr100267-w-2.c: New test.
* gcc.target/i386/avx512f-pr100267-1.c: New test.
* gcc.target/i386/avx512f-pr100267-pd-2.c: New test.
* gcc.target/i386/avx512f-pr100267-ps-2.c: New test.
* gcc.target/i386/avx512vl-pr100267-1.c: New test.
* gcc.target/i386/avx512vl-pr100267-pd-2.c: New test.
* gcc.target/i386/avx512vl-pr100267-ps-2.c: New test.
* gcc.target/i386/avx512vlbw-pr100267-1.c: New test.
* gcc.target/i386/avx512vlbw-pr100267-b-2.c: New test.
* gcc.target/i386/avx512vlbw-pr100267-d-2.c: New test.
* gcc.target/i386/avx512vlbw-pr100267-q-2.c: New test.
* gcc.target/i386/avx512vlbw-pr100267-w-2.c: New test.

Fix ICE for vpexpand*.

gcc/ChangeLog

PR target/100310
* config/i386/i386-expand.c
(ix86_expand_special_args_builtin): Keep constm1_operand only
if it satisfies insn's operand predicate.

gcc/testsuite/ChangeLog

PR target/100310
* gcc.target/i386/pr100310.c: New test.

Daily bump.