The following teaches VN to handle reads from .MASK_STORE and
.LEN_STORE. For this push_partial_def is extended first for
convenience so we don't have to handle the full def case in the
caller (possibly other paths can be simplified then). Also
the partial definition stored value can have an offset applied
so we don't have to build a fake RHS when we register the pieces
of an existing store.
PR tree-optimization/106365
* tree-ssa-sccvn.cc (pd_data::rhs_off): New field determining
the offset to start encoding of RHS from.
(vn_walk_cb_data::vn_walk_cb_data): Initialize it.
(vn_walk_cb_data::push_partial_def): Allow the first partial
definition to be fully providing the def. Offset RHS
before encoding if requested.
(vn_reference_lookup_3): Initialize def_rhs everywhere.
Add support for .MASK_STORE and .LEN_STORE (partial) definitions.
* gcc.target/i386/vec-maskstore-vn.c: New testcase.
The following adds support for MASK_STORE, MASK_LOAD and friends
to call_may_clobber_ref_p and ref_maybe_used_by_call_p. Since
they all use a special argument to specify TBAA they are not really
suited for fnspec handling thus the manual support.
* tree-ssa-alias.cc (ref_maybe_used_by_call_p_1): Special-case
store internal functions and IFN_MASK_LOAD, IFN_LEN_LOAD
and IFN_MASK_LOAD_LANES.
(call_may_clobber_ref_p_1): Special-case IFN_MASK_STORE,
IFN_LEN_STORE and IFN_MASK_STORE_LANES.
This fixes the remaining problem reported in the PR, that the special
members should be trivial. This can be done by constraining the
non-trivial versions and adding defaulted overloads that will be used
when the union members are trivial.
Making these members trivial alters the argument passing ABI and so
isn't suitable for backporting to release branches.
libstdc++-v3/ChangeLog:
PR libstdc++/100823
* include/bits/stl_iterator.h (common_iterator): Define
destructor, copy constructor and move constructor as trivial
when the underlying types allow.
* testsuite/24_iterators/common_iterator/100823.cc: Check
triviality of special members.
This fixes the following conformance problems reported in the PR:
- Move constructor and move assignment should be defined.
- Copy assignment from a valueless object should be allowed.
Assignment is completely rewritten by this patch, as the previous
version had a number of problems. The converting assignment failed to
handle the case of assigning a new value to a valueless object, which
should work. It only accepted lvalue arguments, so wasn't usable to
implement the move assignment operator. Finally, it enforced the
precondition that the argument is not valueless, which is correct for
the converting assignment but not for the copy assignment.
A new _M_assign member is added to handle all cases of assignment
(copying from an lvalue, moving from an rvalue, and converting from a
different type). The not valueless precondition is checked in the
converting assignment before calling _M_assign, so isn't enforced for
copy and move assignment. The new function no longer uses a switch, so
handles valueless objects as the LHS or RHS of the assignment.
libstdc++-v3/ChangeLog:
PR libstdc++/100823
* include/bits/stl_iterator.h (common_iterator): Define move
constructor and move assignment operator.
(common_iterator::_M_assign): New function implementing
assignment.
(common_iterator::operator=): Use _M_assign.
(common_iterator::_S_valueless): New constant.
* testsuite/24_iterators/common_iterator/100823.cc: New test.
The noexcept-specifier for some std::common_iterator constructors was
incorrectly using an rvalue as the first argument of
std::is_nothrow_assignable_v. This gave the wrong answer for some types,
e.g. std::common_iterator<int*, S>, because an rvalue of scalar type
cannot be assigned to.
Also fix the friend declaration to use the same constraints as on the
definition of the class template. G++ fails to diagnose this error, due
to PR c++/96830.
Finally, the copy constructor was using std::move for its argument
in some cases, which should be removed.
libstdc++-v3/ChangeLog:
* include/bits/stl_iterator.h (common_iterator): Fix incorrect
uses of is_nothrow_assignable_v. Fix inconsistent constraints on
friend declaration. Do not move argument in copy constructor.
* testsuite/24_iterators/common_iterator/1.cc: Check for
noexcept constructibnle/assignable.
Doing so fixes various false positives from
-Wanalyzer-tainted-array-index at -O1 and above (e.g. seen on the
Linux kernel)
gcc/analyzer/ChangeLog:
PR analyzer/106373
* sm-taint.cc (taint_state_machine::on_condition): Potentially
update the state of the RHS as well as the LHS.
gcc/testsuite/ChangeLog:
PR analyzer/106373
* gcc.dg/analyzer/torture/taint-read-index-3.c: New test.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
gcc/fortran/ChangeLog:
PR fortran/101330
* openmp.cc (gfc_match_iterator): Remove left-over code from
development that could lead to a crash on invalid input.
gcc/testsuite/ChangeLog:
PR fortran/101330
* gfortran.dg/gomp/affinity-clause-7.f90: New test.
The testcase in the PR demonstrates how it is possible for one
__builtin_setjmp_receiver label to appear in
nonlocal_goto_handler_labels list twice (after the block with
__builtin_setjmp_setup referring to it was duplicated).
remove_node_from_insn_list did not account for this possibility and
removed only the first copy from the list. Add an assert verifying that
duplicates are not present.
To avoid adding a label to the list twice, move registration of the
label from __builtin_setjmp_setup handling to __builtin_setjmp_receiver.
gcc/ChangeLog:
PR rtl-optimization/101347
* builtins.cc (expand_builtin) [BUILT_IN_SETJMP_SETUP]: Move
population of nonlocal_goto_handler_labels from here ...
(expand_builtin) [BUILT_IN_SETJMP_RECEIVER]: ... to here.
* rtlanal.cc (remove_node_from_insn_list): Verify that a
duplicate is not present in the remainder of the list.
This function remains unused since remove_node_from_insn_list was cloned
from it.
gcc/ChangeLog:
* rtl.h (remove_node_from_expr_list): Remove declaration.
* rtlanal.cc (remove_node_from_expr_list): Remove (no uses).
The following adjusts vectorizer code generation to avoid splitting
out address increments for invariant addresses which causes objects
to get TREE_ADDRESSABLE when not necessary.
* tree-vect-data-refs.cc (bump_vector_ptr): Return an
invariant updated address when the input was invariant.
__builtin_cexpi can't be vectorized since there's gap between it and
vectorized sincos version(In libmvec, it passes a double and two
double pointer and returns nothing.) And it will lose some
vectorization opportunity if sin & cos are optimized to cexpi before
vectorizer.
I'm trying to add vect_recog_cexpi_pattern to split cexpi to sin and
cos, but it failed vectorizable_simd_clone_call since NULL is returned
by cgraph_node::get (fndecl). So alternatively, the patch try to move
pass_cse_sincos after vectorizer, just before pas_cse_reciprocals.
Also original pass_cse_sincos additionaly expands pow&cabs, this patch
split that part into a separate pass named pass_expand_powcabs which
remains the old pass position.
gcc/ChangeLog:
* passes.def: (Split pass_cse_sincos to pass_expand_powcabs
and pass_cse_sincos, and move pass_cse_sincos after vectorizer).
* timevar.def (TV_TREE_POWCABS): New timevar.
* tree-pass.h (make_pass_expand_powcabs): Split from pass_cse_sincos.
* tree-ssa-math-opts.cc (gimple_expand_builtin_cabs): Ditto.
(class pass_expand_powcabs): Ditto.
(pass_expand_powcabs::execute): Ditto.
(make_pass_expand_powcabs): Ditto.
(pass_cse_sincos::execute): Remove pow/cabs expand part.
(make_pass_cse_sincos): Ditto.
gcc/testsuite/ChangeLog:
* gcc.dg/pow-sqrt-synth-1.c: Adjust testcase.
2022-07-20 Richard Biener <richard.guenther@gmail.com>
Hongtao Liu <hongtao.liu@intel.com>
gcc/ChangeLog:
PR tree-optimization/106010
* tree-complex.cc (init_dont_simulate_again): Lower complex
type move.
(expand_complex_move): Also expand COMPLEX_CST for rhs.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr106010-1a.c: New test.
* gcc.target/i386/pr106010-1b.c: New test.
* gcc.target/i386/pr106010-1c.c: New test.
* gcc.target/i386/pr106010-2a.c: New test.
* gcc.target/i386/pr106010-2b.c: New test.
* gcc.target/i386/pr106010-2c.c: New test.
* gcc.target/i386/pr106010-3a.c: New test.
* gcc.target/i386/pr106010-3b.c: New test.
* gcc.target/i386/pr106010-3c.c: New test.
* gcc.target/i386/pr106010-4a.c: New test.
* gcc.target/i386/pr106010-4b.c: New test.
* gcc.target/i386/pr106010-4c.c: New test.
* gcc.target/i386/pr106010-5a.c: New test.
* gcc.target/i386/pr106010-5b.c: New test.
* gcc.target/i386/pr106010-5c.c: New test.
* gcc.target/i386/pr106010-6a.c: New test.
* gcc.target/i386/pr106010-6b.c: New test.
* gcc.target/i386/pr106010-6c.c: New test.
* gcc.target/i386/pr106010-7a.c: New test.
* gcc.target/i386/pr106010-7b.c: New test.
* gcc.target/i386/pr106010-7c.c: New test.
* gcc.target/i386/pr106010-8a.c: New test.
* gcc.target/i386/pr106010-8b.c: New test.
* gcc.target/i386/pr106010-8c.c: New test.
* gcc.target/i386/pr106010-9a.c: New test.
* gcc.target/i386/pr106010-9b.c: New test.
* gcc.target/i386/pr106010-9c.c: New test.
* gcc.target/i386/pr106010-9d.c: New test.
Doing so speeds up -fanalyzer from taking over 4 hours to under a
minute on the Linux kernel's sound/soc/codecs/cs47l90.c
gcc/analyzer/ChangeLog:
PR analyzer/106359
* region.h (string_region::tracked_p): New.
* store.cc (binding_cluster::binding_cluster): Move here from
store.h. Add assertion that base_region is tracked_p.
* store.h (binding_cluster::binding_cluster): Move to store.cc.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
Join nodes which carry outgoing ranges on incoming edges are uncommon,
but can still be resolved by setting the dominator range, and then
calculating incoming edges. Avoid doing so if one of the incoing edges
is not dominated by the same dominator.
* gimple-range-cache.cc (ranger_cache::range_from_dom): Check
for incoming ranges on join nodes and add to worklist.
Avoid calling range_of_dom recursively by putting all nodes to be
calculated on the worklist, and figure out which kind they are
when removed from the list.
* gimple-range-cache.cc (ranger_cache::resolve_dom): New.
(ranger_cache::range_from_dom): Put all nodes to be calculated
in the worklist and resolve after the dom walk.
* gimple-range-cache.h (resolve_dom): New prototype.
gcc/fortran/ChangeLog:
PR fortran/103590
* resolve.cc (find_array_spec): Change function result to bool to
enable error recovery. Generate error message for invalid array
reference of non-array entity instead of an internal error.
(gfc_resolve_ref): Use function result from find_array_spec for
error recovery.
gcc/testsuite/ChangeLog:
PR fortran/103590
* gfortran.dg/associate_54.f90: Adjust.
* gfortran.dg/associate_59.f90: New test.
Add testsuite/testsuite_string.h header to help testing __gnu_debug::basic_string like
std::basic_string depending on _GLIBCXX_DEBUG.
Add using of base type methods in __gnu_debug::basic_string to make use of the method
overloads when there is no debug version.
Fix _GLIBCXX_DEBUG_PEDANTIC assertions in <debug/string>. This header has to be used directly
like __gnu_debug::string, it is not included by _GLIBCXX_DEBUG. It means that
_GLIBCXX_DEBUG_PEDANTIC is not considered to define __glibcxx_check_string and
__glibcxx_check_string_len which are then empty macros. Now those macros are defined
directly in <debug/string> and properly consider _GLIBCXX_DEBUG_PEDANTIC.
libstdc++-v3/ChangeLog:
* include/debug/debug.h [_GLIBCXX_DEBUG](__glibcxx_requires_string): Define
using _GLIBCXX_DEBUG_PEDASSERT.
[_GLIBCXX_DEBUG](__glibcxx_requires_string_len): Likewise.
* include/debug/macros.h
(__glibcxx_check_string, __glibcxx_check_string_len): Move...
* include/debug/string
(__glibcxx_check_string, __glibcxx_check_string_len): ...here. And define depending
on _GLIBCXX_DEBUG_PEDANTIC no matter if _GLIBCXX_DEBUG is defined.
Add using of std::string find, rfind, find_first_of, find_last_of, find_first_not_of
and find_last_not_of. Remove debug implementations having no debug assertion.
* testsuite/util/testsuite_string.h: New file. Provides __gnu_test::string and
__gnu_test::wtring which definition depends on _GLIBCXX_DEBUG.
* testsuite/21_strings/basic_string/debug/find1_neg.cc: New test case.
* testsuite/21_strings/basic_string/debug/find2_neg.cc: New test case.
* testsuite/21_strings/basic_string/operations/find/char/1.cc:
Include <testsuite_string.h> and use __gnu_test::string.
* testsuite/21_strings/basic_string/operations/find/char/2.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/char/3.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/char/4.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/char/5.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/char/6.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/wchar_t/1.cc:
Include <testsuite_string.h> and use __gnu_test::wstring.
* testsuite/21_strings/basic_string/operations/find/wchar_t/2.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/wchar_t/3.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/wchar_t/4.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/wchar_t/5.cc: Likewise.
* testsuite/21_strings/basic_string/operations/find/wchar_t/6.cc: Likewise.
In case of l/rvalue or cv-qual mismatch during reference binding, we
try to give more helpful diagnostics by computing a bad conversion that
allows the mismatch. But in doing so, we may end up considering and
instantiating a conversion function that could induce a hard error and
in turn cause us to reject otherwise valid code. We could just give up
on producing a better diagnostic here, but ideally we'd preserve the
better diagnostics for invalid code while avoiding unnecessary template
instantiations for valid code.
To that end, this patch adapts the bad conversion shortcutting mechanism
from r12-3346-g47543e5f9d1fc5 to additionally handle this situation.
The main observation from there is that during overload resolution, if we
know we have a strictly viable candidate then we don't need to distinguish
between an unviable and non-strictly viable candidate. Thus we don't
need to distinguish between an invalid and bad conversion either, which
is what this patch exploits. Of course, we don't know whether we have a
strictly viable candidate until after the fact, so we still need to
remember when we deferred distinguishing between an invalid and bad
conversion. This patch adds a special conversion kind ck_deferred_bad
for this purpose.
PR c++/94894
PR c++/105766
PR c++/106201
gcc/cp/ChangeLog:
* call.cc (enum conversion_kind): Add ck_deferred_bad enumerator.
(has_next): Return false for it.
(reference_binding): Return a ck_deferred_bad conversion instead
of an actual bad conversion when LOOKUP_SHORTCUT_BAD_CONVS is set.
Remove now obsolete early exit for the incomplete TO case.
(implicit_conversion_1): Don't mask out LOOKUP_SHORTCUT_BAD_CONVS.
(add_function_candidate): Set LOOKUP_SHORTCUT_BAD_CONVS iff
shortcut_bad_convs.
(missing_conversion_p): Also return true for a ck_deferred_bad
conversion.
* cp-tree.h (LOOKUP_SHORTCUT_BAD_CONVS): Define.
gcc/testsuite/ChangeLog:
* g++.dg/conversion/ref8.C: New test.
* g++.dg/conversion/ref9.C: New test.
When a returns_twice call has an associated abnormal edge, the edge
corresponds to the "second return" from the call. It wouldn't make sense
if any executable statements appeared between the call and the
destination of the edge (they wouldn't be re-executed upon the "second
return"), so verify that.
gcc/ChangeLog:
* tree-cfg.cc (gimple_verify_flow_info): Check placement of
returns_twice calls.
A returns_twice call may have associated abnormal edges that correspond
to the "second return" from the call. If the call is duplicated, the
copies of those edges also need to be abnormal, but e.g. tracer does not
enforce that. Just prohibit the (unlikely to be useful) duplication.
gcc/ChangeLog:
* cfghooks.cc (duplicate_block): Expand comment.
* tree-cfg.cc (gimple_can_duplicate_bb_p): Reject blocks with
calls that may return twice.
gcc/ChangeLog:
* tree-ssa-sink.cc (select_best_block): Punt if selected block
has incoming abnormal edges.
gcc/testsuite/ChangeLog:
* gcc.dg/setjmp-7.c: New test.
The new built-ins need to be detectable using __has_builtin, and the
library should use that to check for them.
This fixes an error with Clang when C++23 is enabled.
gcc/cp/ChangeLog:
* cp-objcp-common.cc (names_builtin_p): Return true for
RID_REF_CONSTRUCTS_FROM_TEMPORARY and
RID_REF_CONVERTS_FROM_TEMPORARY.
libstdc++-v3/ChangeLog:
* include/std/type_traits (__cpp_lib_reference_from_temporary)
(reference_constructs_from_temporary)
(reference_converts_from_temporary): Only define when the
built-ins are available.
GCC does not support in-tree builds at the moment, so .gitignore
concealing artifacts of accidental in-tree ./configure run may cause
confusion. Un-ignore config.h, which is known to break the build.
ChangeLog:
* .gitignore: Do not ignore config.h.
The optimizations from commit r12-8175-ga54137c88061c7 introduced a
clang integer sanitizer error.
Fix this with an explicit static_cast, similar to the fix for PR 96766.
libstdc++-v3/ChangeLog:
* include/std/charconv (__from_chars_alnum_to_val): Replace
implicit conversion from int to unsigned char with explicit
cast.
PR analyzer/106321 reports false positives from
-Wanalyzer-tainted-array-index on switch statements, seen e.g.
in the Linux kernel in drivers/vfio/pci/vfio_pci_core.c, where
vfio_pci_core_ioctl has:
| 744 | switch (info.index) {
| | ~~~~~~ ~~~~~~~~~~
| | | |
| | | (8) ...to here
| | (9) following ‘case 0 ... 5:’ branch...
|......
| 751 | case VFIO_PCI_BAR0_REGION_INDEX ... VFIO_PCI_BAR5_REGION_INDEX:
| | ~~~~
| | |
| | (10) ...to here
and then a false complaint about "use of attacker-controlled value
‘info.index’ in array lookup without upper-bounds checking", where
info.index has clearly had its bounds checked by the switch/case.
It turns out that when I rewrote switch handling for the analyzer in
r12-3101-g8ca7fa84a3af35, I removed notifications to state machines
about the constraints on cases.
This patch fixes that oversight by adding a new on_bounded_ranges vfunc
for region_model_context, called on switch statement edges, which calls
a new state_machine vfunc. It implements it for the "taint" state
machine, so that it updates the "has bounds" flags at out-edges for
switch statements, based on whether the bounds from the edge appear to
actually constrain the switch index.
gcc/analyzer/ChangeLog:
PR analyzer/106321
* constraint-manager.h (bounded_ranges::get_count): New.
(bounded_ranges::get_range): New.
* engine.cc (impl_region_model_context::on_bounded_ranges): New.
* exploded-graph.h (impl_region_model_context::on_bounded_ranges):
New decl.
* region-model.cc (region_model::apply_constraints_for_gswitch):
Potentially call ctxt->on_bounded_ranges.
* region-model.h (region_model_context::on_bounded_ranges): New
vfunc.
(noop_region_model_context::on_bounded_ranges): New.
(region_model_context_decorator::on_bounded_ranges): New.
* sm-taint.cc: Include "analyzer/constraint-manager.h".
(taint_state_machine::on_bounded_ranges): New.
* sm.h (state_machine::on_bounded_ranges): New.
gcc/testsuite/ChangeLog:
PR analyzer/106321
* gcc.dg/analyzer/torture/taint-read-index-2.c: Add test coverage
for switch statements.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
I found this logging tweak very helpful when working on
PR analyzer/106284.
gcc/analyzer/ChangeLog:
* engine.cc (exploded_graph::process_node): Show any description
of the out-edge when logging it for consideration.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
The poll() function is not always available in RTEMS.
libstdc++-v3/ChangeLog:
* configure: Regnerate.
* configure.ac (newlib, *-rtems*): Remove HAVE_POLL.
get_memory_rtx tries hard to come up with a MEM_EXPR to record
in the memory attributes but in the last fallback fails to properly
account for an unknown offset and thus, as visible in this testcase,
incorrect alignment computed from set_mem_attributes. The following
rectifies both parts.
PR middle-end/106331
* builtins.cc (get_memory_rtx): Compute alignment from
the original address and set MEM_OFFSET to unknown when
we create a MEM_EXPR from the base object of the address.
* gfortran.dg/pr106331.f90: New testcase.
The dwarf2out map of tree to symbol + offset is populated too early
when streaming in trees so that when WPA tree merging decides to
recycle them the mapping prevails and if we are unlucky the same
address is used for another tree with a symbol + offset DIE to
record. The following mitigates the resulting ICE by relaxing the
assert, allowing re-use of a slot during WPA. Delaying the register
would be better but it's already somewhat hairy and uglifying this
further doesn't look too important right now.
PR lto/106334
* dwarf2out.cc (dwarf2out_register_external_die): Allow
map entry re-use during WPA.
This middle-end patch resolves PR c/106264 which is a spurious warning
regression caused by the tree-level expansion of modf, frexp and remquo
producing "expression has no-effect" when the built-in function's result
is ignored. When these built-ins were first expanded at tree-level,
fold_builtin_n would blindly set TREE_NO_WARNING for all built-ins. Now
that we're more discerning, we should precisely call suppress_warning
selectively on those COMPOUND_EXPRs that need them.
2022-07-19 Roger Sayle <roger@nextmovesoftware.com>
Richard Biener <rguenther@suse.de>
gcc/ChangeLog
PR c/106264
* builtins.cc (fold_builtin_frexp): Call suppress_warning on
COMPOUND_EXPR to silence spurious warning if result isn't used.
(fold_builtin_modf): Likewise.
(do_mpfr_remquo): Likewise.
gcc/testsuite/ChangeLog
PR c/106264
* gcc.dg/pr106264.c: New test case.
When we manage to allocate a buffer of the expected size we can simplify the code to
perform the expected algorithm.
libstdc++-v3/ChangeLog:
* include/bits/stl_algo.h
(__merge_adaptive): Adapt to merge only when buffer is large enough..
(__merge_adaptive_resize): New, adapt merge when buffer is too small.
(__inplace_merge): Adapt, use latter.
(__stable_sort_adaptive): Adapt to sort only when buffer is large enough.
(__stable_sort_adaptive_resize): New, adapt sort when buffer is too small.
(__stable_sort): Adapt, use latter.
Whenever a relation is added, register_transitive is always called.
If neither operand was in a relation before, or this is not a new
relation, then there is no need to register transitives.
PR tree-optimization/106280
* value-relation.cc (dom_oracle::register_relation): Register
transitives only when it is possible for there to be one.
(dom_oracle::set_one_relation): Return NULL if this is an
existing relation.
Add missing index references for the `-mrelax' and `-mriscv-attribute'
invocation options.
gcc/
* doc/invoke.texi (RISC-V Options): Add index references for
`mrelax' and `mriscv-attribute'.
Add missing second space around the `-mstack-protector-guard-reg='
invocation option.
gcc/
* doc/invoke.texi (Option Summary): Add missing second space
around `-mstack-protector-guard-reg='.
Correct the name of the `-mriscv-attribute' invocation option, including
a typo in the negated form.
gcc/
* doc/invoke.texi (Option Summary): Fix `-mno-riscv-attribute'.
(RISC-V Options): Likewise, and `-mriscv-attribute'.
Add mtune and mcpu options for ARCHS release 310a type CPU. The
mtune=release31a is designed to be used as an alternative to the
mcpu=hs4x_rel31 option.
ARCHS4x release 31a uses DSP instructions which are implemented a bit
different than mpy9. Hence, use safer mpy2 option.
gcc/
* config/arc/arc-arch.h (arc_tune_attr): Add
ARC_TUNE_ARCHS4X_REL31A variant.
* config/arc/arc.cc (arc_override_options): Tune options for
release 310a.
(arc_sched_issue_rate): Use correct enum.
(arc600_corereg_hazard): Textual change.
(arc_hazard): Add release 310a tunning.
* config/arc/arc.md (tune): Update and take into consideration new
tune option.
(tune_dspmpy): Likewise.
(tune_store): New attribute.
* config/arc/arc.opt (mtune): New tune option.
* config/arc/arcHS4x.md (hs4x_brcc0, hs4x_brcc1): New cpu units.
(hs4x_brcc_op): New instruction rezervation.
(hs4x_data_store_1_op): Likewise.
* config/arc/arc-cpus.def (hs4x_rel31): New cpu variant.
* config/arc/arc-tables.opt: Regenerate.
* config/arc/t-multilib: Likewise.
* doc/invoke.texi (ARC): Update mcpu and tune sections.
Signed-off-by: Claudiu Zissulescu <claziss@gmail.com>
When r7-6373-g40b6bff965d004 fixed a costing issue it failed to
make the logic symmetric which means that we now fuse
normal vs. builtin when the cost model says so but we don't fuse
builtin vs. normal. The following fixes that, also allowing
the cost model to decide to fuse two builtin partitions as otherwise
an intermediate non-builtin can result in a partial merge as well.
* tree-loop-distribution.cc (loop_distribution::distribute_loop):
When computing cost-based merging do not disregard builtin
classified partitions in some cases.
* gcc.dg/tree-ssa/ldist-24.c: XFAIL.
* gcc.dg/tree-ssa/ldist-36.c: Adjust expected outcome.
The ARC soft udivmodsi4 algorithm and as well as using umodsi3
for reduced register set configurations are wrong.
libgcc/
* config/arc/lib2funcs.c (udivmodsi4): Update AND mask.
* config/arc/lib1funcs.S (umodsi3): Don't use it for RF16
configurations.
This patch extends the fix for PR106253 to AArch32. As with AArch64,
we were using ACLE intrinsics to vectorise scalar built-ins, even
though the two sometimes have different ECF_* flags. (That in turn
is because the ACLE intrinsics should follow the instruction semantics
as closely as possible, whereas the scalar built-ins follow language
specs.)
The patch also removes the copysignf built-in, which only existed
for this purpose and wasn't a “real” arm_neon.h built-in.
Doing this also has the side-effect of enabling vectorisation of
rint and roundeven. Logically that should be a separate patch,
but making it one would have meant adding a new int iterator
for the original set of instructions and then removing it again
when including new functions.
I've restricted the bswap tests to little-endian because we end
up with excessive spilling on big-endian. E.g.:
sub sp, sp, #8
vstr d1, [sp]
vldr d16, [sp]
vrev16.8 d16, d16
vstr d16, [sp]
vldr d0, [sp]
add sp, sp, #8
@ sp needed
bx lr
Similarly, the copysign tests require little-endian because on
big-endian we unnecessarily load the constant from the constant pool:
vldr.32 s15, .L3
vdup.32 d0, d7[1]
vbsl d0, d2, d1
bx lr
.L3:
.word -2147483648
gcc/
PR target/106253
* config/arm/arm-builtins.cc (arm_builtin_vectorized_function):
Delete.
* config/arm/arm-protos.h (arm_builtin_vectorized_function): Delete.
* config/arm/arm.cc (TARGET_VECTORIZE_BUILTIN_VECTORIZED_FUNCTION):
Delete.
* config/arm/arm_neon_builtins.def (copysignf): Delete.
* config/arm/iterators.md (nvrint_pattern): New attribute.
* config/arm/neon.md (<NEON_VRINT:nvrint_pattern><VCVTF:mode>2):
New pattern.
(l<NEON_VCVT:nvrint_pattern><su_optab><VCVTF:mode><v_cmp_result>2):
Likewise.
(neon_copysignf<mode>): Rename to...
(copysign<mode>3): ...this.
gcc/testsuite/
PR target/106253
* gcc.target/arm/vect_unary_1.c: New test.
* gcc.target/arm/vect_binary_1.c: Likewise.
The stack pointer adjustment in interrupt epilogue is happening after
restoring the ZOL registers which is wrong. Fixing this.
gcc/
* config/arc/arc.cc (arc_expand_epilogue): Adjust the frame
pointer first when in interrupts.
gcc/testsuite/
* gcc.target/arc/interrupt-13.c: New file.
Signed-off-by: Claudiu Zissulescu <claziss@gmail.com>
loop distribution currently cannot handle the situation when the
last partition is a builtin but there's a common reduction in all
partitions (like the final IV value). The following lifts this
restriction by making the last non-builtin partition provide the
definitions for the loop-closed PHI nodes. Since we have heuristics
in place to avoid code generating builtins last writing a testcase
is difficult (but I ran into a case with other pending patches that
made the heuristic ineffective). What's remaining is the inability
to preserve common reductions when all partitions could be builtins
(in some cases final value replacement could come to the rescue here).
* tree-loop-distribution.cc (copy_loop_before): Add
the ability to replace the original LC PHI defs.
(generate_loops_for_partition): Pass through a flag
whether to redirect original LC PHI defs.
(generate_code_for_partition): Likewise.
(loop_distribution::distribute_loop): Compute the partition
that should provide the LC PHI defs for common reductions
and pass that down.
