The fix "[omp, ftracer] Don't duplicate blocks in SIMT region" adds iteration
over insns in ignore_bb_p, which makes it more expensive.
Counteract this by piggybacking the computation of can_duplicate_bb_p onto
count_insns, which is called at the start of ftracer.
Bootstrapped and reg-tested on x86_64-linux.
gcc/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
* tracer.c (count_insns): Rename to ...
(analyze_bb): ... this.
(cache_can_duplicate_bb_p, cached_can_duplicate_bb_p): New function.
(ignore_bb_p): Use cached_can_duplicate_bb_p.
(tail_duplicate): Call cache_can_duplicate_bb_p.
Factor out can_duplicate_bb_p out of ignore_bb_p.
Also factor out can_duplicate_insn_p and can_duplicate_bb_no_insn_iter_p to
expose the parts of can_duplicate_bb_p that are per-bb and per-insn.
Bootstrapped and reg-tested on x86_64-linux.
gcc/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
* tracer.c (can_duplicate_insn_p, can_duplicate_bb_no_insn_iter_p)
(can_duplicate_bb_p): New function, factored out of ...
(ignore_bb_p): ... here.
In commit ef275d1f20 I implemented the
wrong resolution of LWG 3474. This removes the deduction guide and
alters the views::join factory to create the right type explicitly.
libstdc++-v3/ChangeLog:
* include/std/ranges (join_view): Remove deduction guide.
(views::join): Add explicit template argument list to prevent
deducing the wrong type.
* testsuite/std/ranges/adaptors/join.cc: Move test for LWG 3474
here, from ...
* testsuite/std/ranges/adaptors/join_lwg3474.cc: Removed.
As written in the comment, tree-ssa-math-opts.c wouldn't create a DIVMOD
ifn call for division + modulo by constant for the fear that during
expansion we could generate better code for those cases.
If the divisoris a power of two, that is certainly the case always,
but otherwise expand_divmod can punt in many cases, e.g. if the division
type's precision is above HOST_BITS_PER_WIDE_INT, we don't even call
choose_multiplier, because it works on HOST_WIDE_INTs (true, something
we should fix eventually now that we have wide_ints), or if pre/post shift
is larger than BITS_PER_WORD.
So, the following patch recognizes DIVMOD with constant last argument even
when it is unclear if expand_divmod will be able to optimize it, and then
during DIVMOD expansion if the divisor is constant attempts to expand it as
division + modulo and if they actually don't contain any libcalls or
division/modulo, they are kept as is, otherwise that sequence is thrown away
and divmod optab or libcall is used.
2020-10-06 Jakub Jelinek <jakub@redhat.com>
PR rtl-optimization/97282
* tree-ssa-math-opts.c (divmod_candidate_p): Don't return false for
constant op2 if it is not a power of two and the type has precision
larger than HOST_BITS_PER_WIDE_INT or BITS_PER_WORD.
* internal-fn.c (contains_call_div_mod): New function.
(expand_DIVMOD): If last argument is a constant, try to expand it as
TRUNC_DIV_EXPR followed by TRUNC_MOD_EXPR, but if the sequence
contains any calls or {,U}{DIV,MOD} rtxes, throw it away and use
divmod optab or divmod libfunc.
* gcc.target/i386/pr97282.c: New test.
This ICEs because node->alias_target is (not yet) a FUNCTION_DECL, but
IDENTIFIER_NODE.
I guess we should retry the discovery before LTO streaming out, the reason
to do it this early is that it can affect the gimplification and omp lowering.
2020-10-06 Jakub Jelinek <jakub@redhat.com>
PR middle-end/97289
* omp-offload.c (omp_discover_declare_target_tgt_fn_r): Only follow
node->alias_target if it is a FUNCTION_DECL.
* c-c++-common/gomp/pr97289.c: New test.
This patch rearranges feature bits for MVE and FP to implement the
following flags for -mcpu=cortex-m55.
- +nomve: equivalent to armv8.1-m.main+fp.dp+dsp.
- +nomve.fp: equivalent to armv8.1-m.main+mve+fp.dp (+dsp is implied by +mve).
- +nofp: equivalent to armv8.1-m.main+mve (+dsp is implied by +mve).
- +nodsp: equivalent to armv8.1-m.main+fp.dp.
Combinations of the above:
- +nomve+nofp: equivalent to armv8.1-m.main+dsp.
- +nodsp+nofp: equivalent to armv8.1-m.main.
Due to MVE and FP sharing vfp_base, some new syntax was required in the CPU
description to implement the concept of 'implied bits'. These are non-named
features added to the ISA late, depending on whether one or more features which
depend on them are present. This means vfp_base can be present when only one of
MVE and FP is removed, but absent when both are removed.
gcc/ChangeLog:
2020-07-31 Joe Ramsay <joe.ramsay@arm.com>
* config/arm/arm-cpus.in:
(ALL_FPU_INTERNAL): Remove vfp_base.
(VFPv2): Remove vfp_base.
(MVE): Remove vfp_base.
(vfp_base): Redefine as implied bit dependent on MVE or FP
(cortex-m55): Add flags to disable MVE, MVE FP, FP and DSP extensions.
* config/arm/arm.c (arm_configure_build_target): Add implied bits to ISA.
* config/arm/parsecpu.awk:
(gen_isa): Print implied bits and their dependencies to ISA header.
(gen_data): Add parsing for implied feature bits.
gcc/testsuite/ChangeLog:
* gcc.target/arm/cortex-m55-nodsp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nodsp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nodsp-nofp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nofp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nofp-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nofp-nomve-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nomve-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nomve-flag-softfp.c: New test.
* gcc.target/arm/cortex-m55-nomve.fp-flag-hard.c: New test.
* gcc.target/arm/cortex-m55-nomve.fp-flag-softfp.c: New test.
* gcc.target/arm/multilib.exp: Add tests for -mcpu=cortex-m55.
The implementation of Stream_from_file mishandled several cases:
* It reversed the check for whether bytes were already available in
the peek buffer.
* It considered positive return values from lseek to be an error, when
only a -1 return value indicates an error.
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/259437
This avoids unnecessary instantiations of std::numeric_limits or
inclusion of <limits> when a more lightweight alternative would work.
Some uses can be replaced with __gnu_cxx::__int_traits and some can just
use size_t(-1) directly where SIZE_MAX is needed.
libstdc++-v3/ChangeLog:
* include/bits/regex.h: Use __int_traits<int> instead of
std::numeric_limits<int>.
* include/bits/uniform_int_dist.h: Use __int_traits<T>::__max
instead of std::numeric_limits<T>::max().
* include/bits/hashtable_policy.h: Use size_t(-1) instead of
std::numeric_limits<size_t>::max().
* include/std/regex: Include <ext/numeric_traits.h>.
* include/std/string_view: Use typedef for __int_traits<int>.
* src/c++11/hashtable_c++0x.cc: Use size_t(-1) instead of
std::numeric_limits<size_t>::max().
* testsuite/std/ranges/iota/96042.cc: Include <limits>.
* testsuite/std/ranges/iota/difference_type.cc: Likewise.
* testsuite/std/ranges/subrange/96042.cc: Likewise.
When adding new features to <numeric> I included the required headers
adjacent to the new code. This cleans it up by moving all the includes
to the start of the file.
libstdc++-v3/ChangeLog:
* include/std/numeric: Move all #include directives to the top
of the header.
* testsuite/26_numerics/gcd/gcd_neg.cc: Adjust dg-error line
numbers.
* testsuite/26_numerics/lcm/lcm_neg.cc: Likewise.
These are cleanups so that multi-range union/intersect doesn't
have to deal with legacy code. Instead, these should be done in
legacy mode.
gcc/ChangeLog:
* value-range.cc (irange::legacy_intersect): Only handle
legacy ranges.
(irange::legacy_union): Same.
(irange::union_): When unioning legacy with non-legacy,
first convert to legacy and do everything in legacy mode.
(irange::intersect): Same, but for intersect.
* range-op.cc (range_tests): Adjust for above changes.
This patch imports three fixes from the ranger branch:
1. Fold division by zero into varying instead of undefined.
This provides compatibility with existing stuff on trunk.
2. Solver changes for lshift.
This should not affect anything on trunk, as it only involves
the GORI solver which is yet to be contributed.
3. Preserve existing behavior for ABS([-MIN,-MIN]).
This is actually unrepresentable, but trunk has traditionally
treated this as [-MIN,-MIN] so this patch just syncs range-ops
with the rest of trunk.
gcc/ChangeLog:
* range-op.cc (operator_div::wi_fold): Return varying for
division by zero.
(class operator_rshift): Move class up.
(operator_abs::wi_fold): Return [-MIN,-MIN] for ABS([-MIN,-MIN]).
(operator_tests): Adjust tests.
> See my comment above for Martins attempts to improve things. I don't
> really want to try decide what to do with those late diagnostic IL
> printing but my commit was blamed for showing target-mem-ref unsupported.
>
> I don't have much time to spend to think what to best print and what not,
> but yes, printing only the MEM_REF part is certainly imprecise.
Here is an updated version of the patch that prints TARGET_MEM_REF the way
it should be printed - as C representation of what it actually means.
Of course it would be better to have the original expressions, but with the
late diagnostics we no longer have them.
2020-10-05 Richard Biener <rguenther@suse.de>
Jakub Jelinek <jakub@redhat.com>
PR c++/97197
gcc/cp/
* error.c (dump_expr): Handle TARGET_MEM_REF.
gcc/c-family/
* c-pretty-print.c: Include langhooks.h.
(c_pretty_printer::postfix_expression): Handle TARGET_MEM_REF as
expression.
(c_pretty_printer::expression): Handle TARGET_MEM_REF as
unary_expression.
(c_pretty_printer::unary_expression): Handle TARGET_MEM_REF.
std::allocator and std::pmr::polymorphic_allocator should throw
std::bad_array_new_length from their allocate member functions if the
number of bytes required cannot be represented in std::size_t.
libstdc++-v3/ChangeLog:
* config/abi/pre/gnu.ver: Add new symbol.
* include/bits/functexcept.h (__throw_bad_array_new_length):
Declare new function.
* include/ext/malloc_allocator.h (malloc_allocator::allocate):
Throw bad_array_new_length for impossible sizes (LWG 3190).
* include/ext/new_allocator.h (new_allocator::allocate):
Likewise.
* include/std/memory_resource (polymorphic_allocator::allocate)
(polymorphic_allocator::allocate_object): Use new function,
__throw_bad_array_new_length.
* src/c++11/functexcept.cc (__throw_bad_array_new_length):
Define.
* testsuite/20_util/allocator/lwg3190.cc: New test.
As IFN_GOMP_SIMT_XCHG_* are part of the group marked by
IFN_GOMP_SIMT_ENTER_ALLOC/IFN_GOMP_SIMT_EXIT, handle them conservatively
in ignore_bb_p.
Build on x86_64-linux with nvptx accelerator, tested with libgomp.
gcc/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
* tracer.c (ignore_bb_p): Ignore GOMP_SIMT_XCHG_*.
My change to namespace-scope spell corrections ignored the issue that
different targets might have different builtins, and therefore perturb
iteration order. This fixes it by using an intermediate array of
identifier, which we sort before considering.
gcc/cp/
* name-lookup.c (maybe_add_fuzzy_decl): New.
(maybe_add_fuzzy_binding): New.
(consider_binding_level): Use intermediate sortable vector for
namespace bindings.
gcc/testsuite/
* c-c++-common/spellcheck-reserved.c: Restore diagnostic.
This patch adds vendor and part numbers which were missing from the
initial entry for Neoverse V1 in AArch32 GCC.
gcc/ChangeLog:
* config/arm/arm-cpus.in (neoverse-v1): Add missing vendor and
part numbers.
In commit ab3f4b27ab "[omp, ftracer] Don't duplicate blocks in SIMT region" I
added a comment in ignore_bb_p suggesting a reordering of SIMT_VOTE_ANY and
SIMT_EXIT, which is not possible since VOTE_ANY may have data dependencies to
storage that is deallocated by SIMT_EXIT.
I've now opened a PR (PR97291) to describe the problem the reordering was
intended to fix.
Remove the incorrect suggestion.
gcc/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
* tracer.c (ignore_bb_p): Remove incorrect suggestion.
As Jonathan Wakely pointed out[1], my change in commit
f9ddb696a2 should have been rounding to
the target clock duration type rather than the input clock duration type
in __atomic_futex_unsigned::_M_load_when_equal_until just as (e.g.)
condition_variable does.
As well as fixing this, let's create a rather contrived test that fails
with the previous code, but unfortunately only when run on a machine
with an uptime of over 208.5 days, and even then not always.
[1] https://gcc.gnu.org/pipermail/libstdc++/2020-September/051004.html
libstdc++-v3/ChangeLog:
PR libstdc++/91486
* include/bits/atomic_futex.h:
(__atomic_futex_unsigned::_M_load_when_equal_until): Use target
clock duration type when rounding.
* testsuite/30_threads/async/async.cc (test_pr91486_wait_for):
Rename from test_pr91486.
(float_steady_clock): New class for test.
(test_pr91486_wait_until): New test.
Commit 53ad6b1979 split the implementation
of std::chrono::__detail::ceil so that when compiling for C++17 and
later std::chrono::ceil is used but when compiling for earlier versions
a separate implementation is used to comply with C++11's limited
constexpr rules. Let's run the equivalent of the existing
std::chrono::ceil test cases on std::chrono::__detail::ceil too to make
sure that it doesn't get broken.
libstdc++-v3/ChangeLog:
* testsuite/20_util/duration_cast/rounding_c++11.cc: Copy
rounding.cc and alter to support compilation for C++11 and to
test std::chrono::__detail::ceil.
> This breaks ia64:
>
> In file included from ./tm.h:23,
> from ../../gcc/gencheck.c:23:
> ./options.h:7816:40: error: ISO C++ forbids zero-size array 'explicit_mask' [-Werror=pedantic]
> 7816 | unsigned HOST_WIDE_INT explicit_mask[0];
> | ^
> ./options.h:7816:26: error: zero-size array member 'cl_target_option::explicit_mask' not at end of 'struct cl_target_option' [-Werror=pedantic]
> 7816 | unsigned HOST_WIDE_INT explicit_mask[0];
> | ^~~~~~~~~~~~~
> ./options.h:7812:16: note: in the definition of 'struct cl_target_option'
> 7812 | struct GTY(()) cl_target_option
> | ^~~~~~~~~~~~~~~~
Oops, sorry.
The following patch should fix that and should also fix streaming of the
new explicit_mask_* members.
2020-10-05 Jakub Jelinek <jakub@redhat.com>
* opth-gen.awk: Don't emit explicit_mask array if n_target_explicit
is equal to n_target_explicit_mask.
* optc-save-gen.awk: Compute has_target_explicit_mask and if false,
don't emit code iterating over explicit_mask array elements. Stream
also explicit_mask_* target members.
I've noticed a -Wnarrowing warning on gimple-ssa-store-merging.c, this
change fixes that up.
2020-10-05 Jakub Jelinek <jakub@redhat.com>
* gimple-ssa-store-merging.c
(imm_store_chain_info::output_merged_store): Use ~0U instead of ~0 in
unsigned int array initializer.
When running the libgomp testsuite on x86_64-linux with nvptx accelerator on
the test-case included in this patch, we run into:
...
FAIL: libgomp.fortran/pr95654.f90 -O3 -fomit-frame-pointer -funroll-loops \
-fpeel-loops -ftracer -finline-functions execution test
...
The test-case is a minimal version of this FAIL:
...
FAIL: libgomp.fortran/pr66199-5.f90 -O3 -fomit-frame-pointer -funroll-loops \
-fpeel-loops -ftracer -finline-functions execution test
...
but that one has stopped failing at commit c2ebf4f10d "openmp: Add support
for non-rect simd and improve collapsed simd support".
The problem is that ftracer duplicates a block containing GOMP_SIMT_VOTE_ANY.
That is, before ftracer we have (dropping the GOMP_SIMT_ prefix):
...
bb4(ENTER_ALLOC)
*----------+
| \
| \
| v
| *
v bb8
*<------------*
bb5(VOTE_ANY)
*-------------+
| |
| |
| |
| |
| v
| *
v bb7(XCHG_IDX)
*<------------*
bb6(EXIT)
...
The XCHG_IDX internal-fn does inter-SIMT-lane communication, which for nvptx
maps onto shfl, an operator which has the requirement that the warp executing
the operator is convergent. The warp diverges at bb4, and
reconverges at bb5, and does not diverge by going to bb7, so the shfl is
indeed executed by a convergent warp.
After ftracer, we have:
...
bb4(ENTER_ALLOC)
*----------+
| \
| \
| \
| \
v v
* *
bb5(VOTE_ANY) bb8(VOTE_ANY)
* *
|\ /|
| \ +--------+ |
| \/ |
| /\ |
| / +----------v
|/ *
v bb7(XCHG_IDX)
*<--------------*
bb6(EXIT)
...
The warp diverges again at bb5, but does not reconverge again before bb6, so
the shfl is executed by a divergent warp, which causes the FAIL.
Fix this by making ftracer ignore blocks containing ENTER_ALLOC, VOTE_ANY and
EXIT, effectively treating the SIMT region conservatively.
An argument can be made that the test needs to be added in a more
generic place, like gimple_can_duplicate_bb_p or some such, and that ftracer
then needs to use the generic test. But that's a discussion with a much
broader scope, so I'm leaving that for another patch.
Bootstrapped and reg-tested on x86_64-linux.
Build on x86_64-linux with nvptx accelerator, tested with libgomp.
gcc/ChangeLog:
PR fortran/95654
* tracer.c (ignore_bb_p): Ignore GOMP_SIMT_ENTER_ALLOC,
GOMP_SIMT_VOTE_ANY and GOMP_SIMT_EXIT.
libgomp/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
PR fortran/95654
* testsuite/libgomp.fortran/pr95654.f90: New test.
The optional KIND argument to the MINLOC/MAXLOC intrinsic must not be
passed to the library function, as the kind conversion of the result
is treated explicitly elsewhere.
gcc/fortran/ChangeLog:
PR fortran/97272
* trans-intrinsic.c (strip_kind_from_actual): Helper function for
removal of KIND argument.
(gfc_conv_intrinsic_minmaxloc): Ignore KIND argument here, as it
is treated elsewhere.
gcc/testsuite/ChangeLog:
PR fortran/97272
* gfortran.dg/pr97272.f90: New test.
In recent Technology Levels of AIX 7.2, new "#ifdef __cplusplus" have been
added. Thus, the aix_malloc fix was applied in wrong locations. This patch
increases the context to avoid this.
fixincludes/ChangeLog:
2020-10-03 Clément Chigot <clement.chigot@atos.net>
* inclhack.def (aix_malloc): Add more context to select.
* fixincl.x: Regenerate.
* tests/base/malloc.h: Update expected results.
Seems I've missed that set_option has special treatment for
CLVC_BIT_CLEAR/CLVC_BIT_SET.
Which means I'll need to change the generic handling, so that for
global_options_set elements mentioned in CLVC_BIT_* options are treated
differently, instead of using the accumulated bitmasks they'll need to use
their specific bitmask variables during the option saving/restoring.
Here is a patch that implements that.
2020-10-03 Jakub Jelinek <jakub@redhat.com>
* opth-gen.awk: For variables referenced in Mask and InverseMask,
don't use the explicit_mask bitmask array, but add separate
explicit_mask_* members with the same types as the variables.
* optc-save-gen.awk: Save, restore, compare and hash the separate
explicit_mask_* members.
this patch implements tracking of access ranges. This is only applied when
base pointer is an arugment. Incrementally i will extend it to also track
TBAA basetype so we can disambiguate ranges for accesses to same basetype
(which makes is quite bit more effective). For this reason i track the access
offset separately from parameter offset (the second track combined adjustments
to the parameter). This is I think last feature I would like to add to the
memory access summary this stage1.
Further work will be needed to opitmize the summary and merge adjacent
range/make collapsing more intelingent (so we do not lose track that often),
but I wanted to keep basic patch simple.
According to the cc1plus stats:
Alias oracle query stats:
refs_may_alias_p: 64108082 disambiguations, 74386675 queries
ref_maybe_used_by_call_p: 142319 disambiguations, 65004781 queries
call_may_clobber_ref_p: 23587 disambiguations, 29420 queries
nonoverlapping_component_refs_p: 0 disambiguations, 38117 queries
nonoverlapping_refs_since_match_p: 19489 disambiguations, 55748 must overlaps, 76044 queries
aliasing_component_refs_p: 54763 disambiguations, 755876 queries
TBAA oracle: 24184658 disambiguations 56823187 queries
16260329 are in alias set 0
10617146 queries asked about the same object
125 queries asked about the same alias set
0 access volatile
3960555 are dependent in the DAG
1800374 are aritificially in conflict with void *
Modref stats:
modref use: 10656 disambiguations, 47037 queries
modref clobber: 1473322 disambiguations, 1961464 queries
5027242 tbaa queries (2.563005 per modref query)
649087 base compares (0.330920 per modref query)
PTA query stats:
pt_solution_includes: 977385 disambiguations, 13609749 queries
pt_solutions_intersect: 1032703 disambiguations, 13187507 queries
Which should still compare with
https://gcc.gnu.org/pipermail/gcc-patches/2020-September/554930.html
there is about 2% more load disambiguations and 3.6% more store that is not
great, but the TBAA part helps noticeably more and also this should help
with -fno-strict-aliasing.
I plan to work on improving param tracking too.
Bootstrapped/regtested x86_64-linux with the other changes, OK?
2020-10-02 Jan Hubicka <hubicka@ucw.cz>
* ipa-modref-tree.c (test_insert_search_collapse): Update andling
of accesses.
(test_merge): Likewise.
* ipa-modref-tree.h (struct modref_access_node): Add offset, size,
max_size, parm_offset and parm_offset_known.
(modref_access_node::useful_p): Constify.
(modref_access_node::range_info_useful_p): New predicate.
(modref_access_node::operator==): New.
(struct modref_parm_map): New structure.
(modref_tree::merge): Update for racking parameters)
* ipa-modref.c (dump_access): Dump new fields.
(get_access): Fill in new fields.
(merge_call_side_effects): Update handling of parm map.
(write_modref_records): Stream new fields.
(read_modref_records): Stream new fields.
(compute_parm_map): Update for new parm map.
(ipa_merge_modref_summary_after_inlining): Update.
(modref_propagate_in_scc): Update.
* tree-ssa-alias.c (modref_may_conflict): Handle known ranges.
When Andrew Macleod investigated the recent rs6000 bootstrap failure,
he suggested a clean up of the headers in rs6000.c and rs6000-call.c.
It now is recommended to include ssa.h instead of the individual headers.
This also ensures that value-range.h is included and in the correct order
so that the tree-ssa-propagate.h inclusion of value-query.h and its
dependencies are satisfied.
Bootstrapped on powerpc-ibm-aix7.2.0.0 and powerpc64le-linux.
gcc/ChangeLog:
2020-10-02 David Edelsohn <dje.gcc@gmail.com>
Andrew MacLeod <amacleod@redhat.com>
* config/rs6000/rs6000.c: Include ssa.h. Reorder some headers.
* config/rs6000/rs6000-call.c: Same.
No one is interested in the mangled name of the C++20 template parameter
object for a class NTTP. So instead of printing
required for the satisfaction of ‘positive<T::ratio>’ [with T = X<::_ZTAXtl5ratioLin1ELi2EEE>]
let's print
required for the satisfaction of ‘positive<T::ratio>’ [with T = X<{-1, 2}>]
I don't think adding a test is necessary for this.
gcc/cp/ChangeLog:
PR c++/97014
* cxx-pretty-print.c (pp_cxx_template_argument_list): If the
argument is template_parm_object_p, print its DECL_INITIAL.
This fixes a linker error for older ARM cores without 64-bit atomics.
I think the { dg-add-options libatomic } is no longer needed, but it's
harmless to keep it there.
libstdc++-v3/ChangeLog:
* testsuite/29_atomics/atomic_float/value_init.cc: Use float
instead of double so that __atomic_load_8 isn't needed.
This test was supposed to verify that when __libc_single_threaded is
available we successfully detect recursive static initialization even
when linked to libpthread. But I forgot to that when recursive init is
detected, we terminate, and so the test fails.
This adds a terminate handler that exits cleanly, so the test passes
when recursive init is detected.
libstdc++-v3/ChangeLog:
* testsuite/18_support/96817.cc: Use terminate handler that
calls _Exit(0).
Here's the patch to remove DECL_ANTICIPATED, and with it hiddenness is
managed entirely in the symbol table. Sadly I couldn't get rid of the
actual field without more investigation -- it's repurposed for
OMP_PRIVATIZED_MEMBER. It looks like a the VAR-related flags in
lang_decl_base are not completely orthogonal, so perhaps some can be
turned into an enumeration or something. But that's more than I want
to do right now.
DECL_FRIEND_P Is still slightly suspect as it appears to mean more
than just in-class definition. However, I'm leaving that for now.
gcc/cp/
* cp-tree.h (lang_decl_base): anticipated_p is not used for
anticipatedness.
(DECL_ANTICIPATED): Delete.
* decl.c (duplicate_decls): Delete DECL_ANTICIPATED_management,
use was_hidden.
(cxx_builtin_function): Drop DECL_ANTICIPATED setting.
(xref_tag_1): Drop DECL_ANTICIPATED assert.
* name-lookup.c (name_lookup::adl_class_only): Drop
DECL_ANTICIPATED check.
(name_lookup::search_adl): Always dedup.
(anticipated_builtin_p): Reimplement.
(do_pushdecl): Drop DECL_ANTICIPATED asserts & update.
(lookup_elaborated_type_1): Drop DECL_ANTICIPATED update.
(do_pushtag): Drop DECL_ANTICIPATED setting.
* pt.c (push_template_decl): Likewise.
(tsubst_friend_class): Likewise.
libcc1/
* libcp1plugin.cc (libcp1plugin.cc): Drop DECL_ANTICIPATED test.
For 'no such binding' errors, we iterate over binding levels to find a
close match. At the namespace level we were using DECL_ANTICIPATED to
skip undeclared builtins. But (a) there are other unnameable things
there and (b) decl-anticipated is about to go away. This changes the
namespace scanning to iterate over the hash table, and look at
non-hidden bindings. This does mean we look at fewer strings
(hurrarh), but the order we meet them is somewhat 'random'. Our
distance measure is not very fine grained, and a couple of testcases
change their suggestion. I notice for the c/c++ common one, we now
match the output of the C compiler. For the other one we think 'int'
and 'int64_t' have the same distance from 'int64', and now meet the
former first. That's a little unfortunate. If it's too problematic I
suppose we could sort the strings via an intermediate array before
measuring distance.
gcc/cp/
* name-lookup.c (consider_decl): New, broken out of ...
(consider_binding_level): ... here. Iterate the hash table for
namespace bindings.
gcc/testsuite/
* c-c++-common/spellcheck-reserved.c: Adjust diagnostic.
* g++.dg/spellcheck-typenames.C: Adjust diagnostic.
ctor_omit_inherited_parms was being somewhat abused. What I'd missed
is that it checks for a base-dtor name, before proceeding with the
check. But we ended up passing it that during cloning before we'd
completed the cloning. It was also using DECL_ORIGIN to get to the
in-charge ctor, but we sometimes zap DECL_ABSTRACT_ORIGIN, and it ends
up processing the incoming function -- which happens to work. so,
this breaks out a predicate that expects to get the incharge ctor, and
will tell you whether its base ctor will need to omit the parms. We
call that directly during cloning.
Then the original fn is essentially just a wrapper, but uses
DECL_CLONED_FUNCTION to get to the in-charge ctor. That uncovered
abuse in add_method, which was happily passing TEMPLATE_DECLs to it.
Let's not do that. add_method itself contained a loop mostly
containing an 'if (nomatch) continue' idiom, except for a final 'if
(match) {...}' check, which itself contained instances of the former
idiom. I refactored that to use the former idiom throughout. In
doing that I found a place where we'd issue an error, but then not
actually reject the new member.
gcc/cp/
* cp-tree.h (base_ctor_omit_inherited_parms): Declare.
* class.c (add_method): Refactor main loop, only pass fns to
ctor_omit_inherited_parms.
(build_cdtor_clones): Rename bool parms.
(clone_cdtor): Call base_ctor_omit_inherited_parms.
* method.c (base_ctor_omit_inherited_parms): New, broken out of
...
(ctor_omit_inherited_parms): ... here, call it with
DECL_CLONED_FUNCTION.
gcc/testsuite/
* g++.dg/inherit/pr97268.C: New.
A previous patch in the series has taught IPA-CP to identify the
important cloning opportunities in 548.exchange2_r as worthwhile on
their own, but the optimization is still prevented from taking place
because of the overall unit-growh limit. This patches raises that
limit so that it takes place and the benchmark runs 30% faster (on AMD
Zen2 CPU at least).
Before this patch, IPA-CP uses the following formulae to arrive at the
overall_size limit:
base = MAX(orig_size, param_large_unit_insns)
unit_growth_limit = base + base * param_ipa_cp_unit_growth / 100
since param_ipa_cp_unit_growth has default 10, param_large_unit_insns
has default value 10000.
The problem with exchange2 (at least on zen2 but I have had a quick
look on aarch64 too) is that the original estimated unit size is 10513
and so param_large_unit_insns does not apply and the default limit is
therefore 11564 which is good enough only for one of the ideal 8
clonings, we need the limit to be at least 16291.
I would like to raise param_ipa_cp_unit_growth a little bit more soon
too, but most certainly not to 55. Therefore, the large_unit must be
increased. In this patch, I decided to decouple the inlining and
ipa-cp large-unit parameters. It also makes sense because IPA-CP uses
it only at -O3 while inlining also at -O2 (IIUC). But if we agree we
can try raising param_large_unit_insns to 13-14 thousand
"instructions," perhaps it is not necessary. But then again, it may
make sense to actually increase the IPA-CP limit further.
I plan to experiment with IPA-CP tuning on a larger set of programs.
Meanwhile, mainly to address the 548.exchange2_r regression, I'm
suggesting this simple change.
gcc/ChangeLog:
2020-09-07 Martin Jambor <mjambor@suse.cz>
* params.opt (ipa-cp-large-unit-insns): New parameter.
* ipa-cp.c (get_max_overall_size): Use the new parameter.
When experimenting with IPA-CP parameters, especially when looking
into exchange2_r, it has been very useful to know what the value of
overall_size is at different stages of the decision process. This
patch therefore adds it to the generated dumps.
gcc/ChangeLog:
2020-09-07 Martin Jambor <mjambor@suse.cz>
* ipa-cp.c (estimate_local_effects): Add overeall_size to dumped
string.
(decide_about_value): Add dumping new overall_size.
This patch enhances the ability of IPA to reason under what conditions
loops in a function have known iteration counts or strides because it
replaces single predicates which currently hold conjunction of
predicates for all loops with vectors capable of holding multiple
predicates, each with a cumulative frequency of loops with the
property.
This second property is then used by IPA-CP to much more aggressively
boost its heuristic score for cloning opportunities which make
iteration counts or strides of frequent loops compile time constant.
gcc/ChangeLog:
2020-09-03 Martin Jambor <mjambor@suse.cz>
* ipa-fnsummary.h (ipa_freqcounting_predicate): New type.
(ipa_fn_summary): Change the type of loop_iterations and loop_strides
to vectors of ipa_freqcounting_predicate.
(ipa_fn_summary::ipa_fn_summary): Construct the new vectors.
(ipa_call_estimates): New fields loops_with_known_iterations and
loops_with_known_strides.
* ipa-cp.c (hint_time_bonus): Multiply param_ipa_cp_loop_hint_bonus
with the expected frequencies of loops with known iteration count or
stride.
* ipa-fnsummary.c (add_freqcounting_predicate): New function.
(ipa_fn_summary::~ipa_fn_summary): Release the new vectors instead of
just two predicates.
(remap_hint_predicate_after_duplication): Replace with function
remap_freqcounting_preds_after_dup.
(ipa_fn_summary_t::duplicate): Use it or duplicate new vectors.
(ipa_dump_fn_summary): Dump the new vectors.
(analyze_function_body): Compute the loop property vectors.
(ipa_call_context::estimate_size_and_time): Calculate also
loops_with_known_iterations and loops_with_known_strides. Adjusted
dumping accordinly.
(remap_hint_predicate): Replace with function
remap_freqcounting_predicate.
(ipa_merge_fn_summary_after_inlining): Use it.
(inline_read_section): Stream loopcounting vectors instead of two
simple predicates.
(ipa_fn_summary_write): Likewise.
* params.opt (ipa-max-loop-predicates): New parameter.
* doc/invoke.texi (ipa-max-loop-predicates): Document new param.
gcc/testsuite/ChangeLog:
2020-09-03 Martin Jambor <mjambor@suse.cz>
* gcc.dg/ipa/ipcp-loophint-1.c: New test.
A subsequent patch adds another two estimates that the code in
ipa_call_context::estimate_size_and_time computes, and the fact that
the function has a special output parameter for each thing it computes
would make it have just too many. Therefore, this patch collapses all
those ouptut parameters into one output structure.
gcc/ChangeLog:
2020-09-02 Martin Jambor <mjambor@suse.cz>
* ipa-inline-analysis.c (do_estimate_edge_time): Adjusted to use
ipa_call_estimates.
(do_estimate_edge_size): Likewise.
(do_estimate_edge_hints): Likewise.
* ipa-fnsummary.h (struct ipa_call_estimates): New type.
(ipa_call_context::estimate_size_and_time): Adjusted declaration.
(estimate_ipcp_clone_size_and_time): Likewise.
* ipa-cp.c (hint_time_bonus): Changed the type of the second argument
to ipa_call_estimates.
(perform_estimation_of_a_value): Adjusted to use ipa_call_estimates.
(estimate_local_effects): Likewise.
* ipa-fnsummary.c (ipa_call_context::estimate_size_and_time): Adjusted
to return estimates in a single ipa_call_estimates parameter.
(estimate_ipcp_clone_size_and_time): Likewise.
Hi,
as we discussed with Honza on the mailin glist last week, making
cached call context structure distinct from the normal one may make it
clearer that the cached data need to be explicitely deallocated.
This patch does that division. It is not mandatory for the overall
main goals of the patch set and can be dropped if deemed superfluous.
gcc/ChangeLog:
2020-09-02 Martin Jambor <mjambor@suse.cz>
* ipa-fnsummary.h (ipa_cached_call_context): New forward declaration
and class.
(class ipa_call_context): Make friend ipa_cached_call_context. Moved
methods duplicate_from and release to it too.
* ipa-fnsummary.c (ipa_call_context::duplicate_from): Moved to class
ipa_cached_call_context.
(ipa_call_context::release): Likewise, removed the parameter.
* ipa-inline-analysis.c (node_context_cache_entry): Change the type of
ctx to ipa_cached_call_context.
(do_estimate_edge_time): Remove parameter from the call to
ipa_cached_call_context::release.
Hi,
this large patch is mostly mechanical change which aims to replace
uses of separate vectors about known scalar values (usually called
known_vals or known_csts), known aggregate values (known_aggs), known
virtual call contexts (known_contexts) and known value
ranges (known_value_ranges) with uses of either new type
ipa_call_arg_values or ipa_auto_call_arg_values, both of which simply
contain these vectors inside them.
The need for two distinct comes from the fact that when the vectors
are constructed from jump functions or lattices, we really should use
auto_vecs with embedded storage allocated on stack. On the other hand,
the bundle in ipa_call_context can be allocated on heap when in cache,
one time for each call_graph node.
ipa_call_context is constructible from ipa_auto_call_arg_values but
then its vectors must not be resized, otherwise the vectors will stop
pointing to the stack ones. Unfortunately, I don't think the
structure embedded in ipa_call_context can be made constant because we
need to manipulate and deallocate it when in cache.
gcc/ChangeLog:
2020-09-01 Martin Jambor <mjambor@suse.cz>
* ipa-prop.h (ipa_auto_call_arg_values): New type.
(class ipa_call_arg_values): Likewise.
(ipa_get_indirect_edge_target): Replaced vector arguments with
ipa_call_arg_values in declaration. Added an overload for
ipa_auto_call_arg_values.
* ipa-fnsummary.h (ipa_call_context): Removed members m_known_vals,
m_known_contexts, m_known_aggs, duplicate_from, release and equal_to,
new members m_avals, store_to_cache and equivalent_to_p. Adjusted
construcotr arguments.
(estimate_ipcp_clone_size_and_time): Replaced vector arguments
with ipa_auto_call_arg_values in declaration.
(evaluate_properties_for_edge): Likewise.
* ipa-cp.c (ipa_get_indirect_edge_target): Adjusted to work on
ipa_call_arg_values rather than on separate vectors. Added an
overload for ipa_auto_call_arg_values.
(devirtualization_time_bonus): Adjusted to work on
ipa_auto_call_arg_values rather than on separate vectors.
(gather_context_independent_values): Adjusted to work on
ipa_auto_call_arg_values rather than on separate vectors.
(perform_estimation_of_a_value): Likewise.
(estimate_local_effects): Likewise.
(modify_known_vectors_with_val): Adjusted both variants to work on
ipa_auto_call_arg_values and rename them to
copy_known_vectors_add_val.
(decide_about_value): Adjusted to work on ipa_call_arg_values rather
than on separate vectors.
(decide_whether_version_node): Likewise.
* ipa-fnsummary.c (evaluate_conditions_for_known_args): Likewise.
(evaluate_properties_for_edge): Likewise.
(ipa_fn_summary_t::duplicate): Likewise.
(estimate_edge_devirt_benefit): Adjusted to work on
ipa_call_arg_values rather than on separate vectors.
(estimate_edge_size_and_time): Likewise.
(estimate_calls_size_and_time_1): Likewise.
(summarize_calls_size_and_time): Adjusted calls to
estimate_edge_size_and_time.
(estimate_calls_size_and_time): Adjusted to work on
ipa_call_arg_values rather than on separate vectors.
(ipa_call_context::ipa_call_context): Construct from a pointer to
ipa_auto_call_arg_values instead of inividual vectors.
(ipa_call_context::duplicate_from): Adjusted to access vectors within
m_avals.
(ipa_call_context::release): Likewise.
(ipa_call_context::equal_to): Likewise.
(ipa_call_context::estimate_size_and_time): Adjusted to work on
ipa_call_arg_values rather than on separate vectors.
(estimate_ipcp_clone_size_and_time): Adjusted to work with
ipa_auto_call_arg_values rather than on separate vectors.
(ipa_merge_fn_summary_after_inlining): Likewise. Adjusted call to
estimate_edge_size_and_time.
(ipa_update_overall_fn_summary): Adjusted call to
estimate_edge_size_and_time.
* ipa-inline-analysis.c (do_estimate_edge_time): Adjusted to work with
ipa_auto_call_arg_values rather than with separate vectors.
(do_estimate_edge_size): Likewise.
(do_estimate_edge_hints): Likewise.
* ipa-prop.c (ipa_auto_call_arg_values::~ipa_auto_call_arg_values):
New destructor.
