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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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++-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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
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.
The current std::lock algorithm is the one called "persistent" in Howard
Hinnant's https://howardhinnant.github.io/dining_philosophers.html post.
While it tends to perform acceptably fast, it wastes a lot of CPU cycles
by continuously locking and unlocking the uncontended mutexes.
Effectively, it's a spin lock with no back-off.
This replaces it with the one Howard calls "smart and polite". It's
smart, because when a Mi.try_lock() call fails because mutex Mi is
contended, the algorithm reorders the mutexes until Mi is first, then
calls Mi.lock(), to block until Mi is no longer contended. It's
polite because it uses std::this_thread::yield() between the failed
Mi.try_lock() call and the Mi.lock() call. (In reality it uses
__gthread_yield() directly, because using this_thread::yield() would
require shuffling code around to avoid a circular dependency.)
This version of the algorithm is inspired by some hints from Howard, so
that it has strictly bounded stack usage. As the comment in the code
says:
// This function can recurse up to N levels deep, for N = 1+sizeof...(L1).
// On each recursion the lockables are rotated left one position,
// e.g. depth 0: l0, l1, l2; depth 1: l1, l2, l0; depth 2: l2, l0, l1.
// When a call to l_i.try_lock() fails it recurses/returns to depth=i
// so that l_i is the first argument, and then blocks until l_i is locked.
The 'i' parameter is the desired permuation of the lockables, and the
'depth' parameter is the depth in the call stack of the current
instantiation of the function template. If i == depth then the function
calls l0.lock() and then l1.try_lock()... for each lockable in the
parameter pack l1. If i > depth then the function rotates the lockables
to the left one place, and calls itself again to go one level deeper.
Finally, if i < depth then the function returns to a shallower depth,
equivalent to a right rotate of the lockables. When a call to
try_lock() fails, i is set to the index of the contended lockable, so
that the next call to l0.lock() will use the contended lockable as l0.
This commit also replaces the std::try_lock implementation details. The
new code is identical in behaviour, but uses a pair of constrained
function templates. This avoids instantiating a class template, and is a
litle simpler to call where used in std::__detail::__lock_impl and
std::try_lock.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/std/mutex (__try_to_lock): Move to __detail namespace.
(struct __try_lock_impl): Replace with ...
(__detail::__try_lock_impl<Idx>(tuple<Lockables...>&)): New
function templates to implement std::try_lock.
(try_lock): Use new __try_lock_impl.
(__detail::__lock_impl(int, int&, L0&, L1&...)): New function
template to implement std::lock.
(lock): Use __lock_impl.
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.
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.
PR target/88529
gcc/ChangeLog:
* df-scan.c (df_ref_record): Check that regno < endregno.
* function.c (assign_parms, expand_function_end): Do nothing with a
TYPE_EMPTY_P result.
gcc/testsuite/ChangeLog:
* g++.target/i386/empty-class1.C: New test.
Andrew MacLeod
Jeff Law
Uros Bizjak
Martin Liska
Patrick Palka
Jonathan Wakely
Andre Vehreschild
Jakub Jelinek
Kewen Lin
GCC Administrator
Sandra Loosemore
Thomas Rodgers
David Malcolm
Sergei Trofimovich
Richard Biener
Jojo R
Kito Cheng
liuhongt
Jason Merrill