Following are a series of cleanups to the frange code in preparation
for a much more invasive patch rewriting the NAN and sign tracking
bits. Please be patient, as I'm trying to break everything up into
small chunks instead of dropping a mondo patch removing the NAN and
sign tristate handling.
No functional changes.
Regstrapped on x86-64 Linux, plus I ran selftests for
-ffinite-math-only.
gcc/ChangeLog:
* value-query.cc (range_query::get_tree_range): Remove check for overflow.
* value-range-pretty-print.cc (vrange_printer::visit): Move read
of type until after undefined_p is checked.
* value-range.cc (frange::set): Remove asserts for REAL_CST.
(frange::contains_p): Tidy up.
(range_tests_nan): Add comment.
* value-range.h (frange::type): Check for undefined_p.
(frange::set_undefined): Remove set of endpoints.
This patch adjusts OpenMP/OpenACC clause list handling in a couple of
places, in preparation for the following mapping-clause expansion rework
patch. Firstly mapping groups are removed as a whole in the C and C++
front-ends when an error is detected, which avoids leaving "badly-formed"
mapping clause groups in the list.
Secondly, reindexing of the omp_mapping_group hashmap (during
omp_build_struct_sibling_lists) has been reimplemented, fixing some
tricky corner-cases where mapping groups are removed from a list at the
same time as it is being reordered.
Thirdly, code to check if a different clause on the same directive maps
the whole of a struct that we have a component mapping for (for example)
has been outlined, removing a bit of code duplication.
2022-09-13 Julian Brown <julian@codesourcery.com>
gcc/
* gimplify.cc (omp_group_last): Allow GOMP_MAP_ATTACH_DETACH after
GOMP_MAP_STRUCT (for reindexing).
(omp_gather_mapping_groups): Reimplement using...
(omp_gather_mapping_groups_1): This new function. Stop processing at
GATHER_SENTINEL.
(omp_group_base): Allow GOMP_MAP_TO_PSET without any following node.
(omp_index_mapping_groups): Reimplement using...
(omp_index_mapping_groups_1): This new function. Handle
REINDEX_SENTINEL.
(omp_reindex_mapping_groups, omp_mapped_by_containing_struct): New
functions.
(omp_tsort_mapping_groups_1): Adjust handling of base group being the
same as current group. Use omp_mapped_by_containing_struct.
(omp_build_struct_sibling_lists): Use omp_mapped_by_containing_struct
and omp_reindex_mapping_groups. Robustify group deletion for reordered
lists.
(gimplify_scan_omp_clauses): Update calls to
omp_build_struct_sibling_lists.
gcc/c/
* c-typeck.cc (c_finish_omp_clauses): Remove whole mapping node group
on error.
gcc/cp/
* semantics.cc (finish_omp_clauses): Likewise.
This patch refactors struct sibling-list processing in gimplify.cc, and
adjusts some related mapping-clause processing in the Fortran FE and
omp-low.cc accordingly.
2022-09-13 Julian Brown <julian@codesourcery.com>
gcc/fortran/
* trans-openmp.cc (gfc_trans_omp_clauses): Don't create
GOMP_MAP_TO_PSET mappings for class metadata, nor GOMP_MAP_POINTER
mappings for POINTER_TYPE_P decls.
gcc/
* gimplify.cc (gimplify_omp_var_data): Remove GOVD_MAP_HAS_ATTACHMENTS.
(GOMP_FIRSTPRIVATE_IMPLICIT): Renumber.
(insert_struct_comp_map): Refactor function into...
(build_omp_struct_comp_nodes): This new function. Remove list handling
and improve self-documentation.
(extract_base_bit_offset): Remove BASE_REF, OFFSETP parameters. Move
code to strip outer parts of address out of function, but strip no-op
conversions.
(omp_mapping_group): Add DELETED field for use during reindexing.
(omp_strip_components_and_deref, omp_strip_indirections): New functions.
(omp_group_last, omp_group_base): Add GOMP_MAP_STRUCT handling.
(omp_gather_mapping_groups): Initialise DELETED field for new groups.
(omp_index_mapping_groups): Notice DELETED groups when (re)indexing.
(omp_siblist_insert_node_after, omp_siblist_move_node_after,
omp_siblist_move_nodes_after, omp_siblist_move_concat_nodes_after): New
helper functions.
(omp_accumulate_sibling_list): New function to build up GOMP_MAP_STRUCT
node groups for sibling lists. Outlined from gimplify_scan_omp_clauses.
(omp_build_struct_sibling_lists): New function.
(gimplify_scan_omp_clauses): Remove struct_map_to_clause,
struct_seen_clause, struct_deref_set. Call
omp_build_struct_sibling_lists as pre-pass instead of handling sibling
lists in the function's main processing loop.
(gimplify_adjust_omp_clauses_1): Remove GOVD_MAP_HAS_ATTACHMENTS
handling, unused now.
* omp-low.cc (scan_sharing_clauses): Handle pointer-type indirect
struct references, and references to pointers to structs also.
gcc/testsuite/
* g++.dg/goacc/member-array-acc.C: New test.
* g++.dg/gomp/member-array-omp.C: New test.
* g++.dg/gomp/target-3.C: Update expected output.
* g++.dg/gomp/target-lambda-1.C: Likewise.
* g++.dg/gomp/target-this-2.C: Likewise.
* c-c++-common/goacc/deep-copy-arrayofstruct.c: Move test from here.
* c-c++-common/gomp/target-50.c: New test.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/deep-copy-15.c: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-16.c: New test.
* testsuite/libgomp.oacc-c++/deep-copy-17.C: New test.
* testsuite/libgomp.oacc-c-c++-common/deep-copy-arrayofstruct.c: Move
test to here, make "run" test.
After inlining and IPA transforms we run fixup_cfg to fixup CFG
effects in other functions. But that fails to clean abnormal
edges from non-pure/const calls which might no longer be necessary
when ->calls_setjmp is false. The following ensures this happens
and refactors things so we call EH/abnormal cleanup only on the
last stmt in a block.
PR tree-optimization/106938
* tree-cfg.cc (execute_fixup_cfg): Purge dead abnormal
edges for all last stmts in a block. Do EH cleanup
only on the last stmt in a block.
* gcc.dg/pr106938.c: New testcase.
This assert was put here to make sure that the legacy
get_value_range() wasn't being called on stuff that legacy couldn't
handle (floats, etc), because the result would ultimately be copied
into a value_range_equiv.
In this case, simplify_casted_cond() is calling it on an offset_type
which is neither an integer nor a pointer. However, range_of_expr
happily punted on it, and then the fallthru code set the range to
VARYING. As value_range_equiv can store VARYING types of anything
(including types it can't handle), this is fine.
The easiest thing to do is remove the assert. If someone from the non
legacy world tries to get a non integer/pointer range here, it's going
to blow up anyhow because the temporary in get_value_range is
int_range_max.
PR tree-optimization/106936
gcc/ChangeLog:
* value-query.cc (range_query::get_value_range): Remove assert.
gcc/testsuite/ChangeLog:
* g++.dg/tree-ssa/pr106936.C: New test.
This patch reimplements the omp_target_reorder_clauses function in
anticipation of supporting "deeper" struct mappings (that is, with
several structure dereference operators, or similar).
The idea is that in place of the (possibly quadratic) algorithm in
omp_target_reorder_clauses that greedily moves clauses containing
addresses that are subexpressions of other addresses before those other
addresses, we employ a topological sort algorithm to calculate a proper
order for map clauses. This should run in linear time, and hopefully
handles degenerate cases where multiple "levels" of indirect accesses
are present on a given directive.
The new method also takes care to keep clause groups together, addressing
the concerns raised in:
https://gcc.gnu.org/pipermail/gcc-patches/2021-May/570501.html
To figure out if some given clause depends on a base pointer in another
clause, we strip off the outer layers of the address expression, and check
(via a tree_operand_hash hash table we have built) if the result is a
"base pointer" as defined in OpenMP 5.0 (1.2.6 Data Terminology). There
are some subtleties involved, however:
- We must treat MEM_REF with zero offset the same as INDIRECT_REF.
This should probably be fixed in the front ends instead so we always
use a canonical form (probably INDIRECT_REF). The following patch
shows one instance of the problem, but there may be others:
https://gcc.gnu.org/pipermail/gcc-patches/2021-May/571382.html
- Mapping a whole struct implies mapping each of that struct's
elements, which may be base pointers. Because those base pointers
aren't necessarily explicitly referenced in the directive in question,
we treat the whole-struct mapping as a dependency instead.
2022-09-13 Julian Brown <julian@codesourcery.com>
gcc/
* gimplify.cc (is_or_contains_p, omp_target_reorder_clauses): Delete
functions.
(omp_tsort_mark): Add enum.
(omp_mapping_group): Add struct.
(debug_mapping_group, omp_get_base_pointer, omp_get_attachment,
omp_group_last, omp_gather_mapping_groups, omp_group_base,
omp_index_mapping_groups, omp_containing_struct,
omp_tsort_mapping_groups_1, omp_tsort_mapping_groups,
omp_segregate_mapping_groups, omp_reorder_mapping_groups): New
functions.
(gimplify_scan_omp_clauses): Call above functions instead of
omp_target_reorder_clauses, unless we've seen an error.
* omp-low.cc (scan_sharing_clauses): Avoid strict test if we haven't
sorted mapping groups.
gcc/testsuite/
* g++.dg/gomp/target-lambda-1.C: Adjust expected output.
* g++.dg/gomp/target-this-3.C: Likewise.
* g++.dg/gomp/target-this-4.C: Likewise.
My change to match.pd (that added the two simplifications this patch
touches) results in more |/^/& assignments with pointer arguments,
but since r12-1608 we reject pointer operands for BIT_NOT_EXPR.
Disallowing them for BIT_NOT_EXPR and allowing for BIT_{IOR,XOR,AND}_EXPR
leads to a match.pd maintainance nightmare (see one of the patches in the
PR), so either we want to allow pointer operand on BIT_NOT_EXPR (but then
we run into issues e.g. with the ranger which expects it can emulate
BIT_NOT_EXPR ~X as - 1 - X which doesn't work for pointers which don't
support MINUS_EXPR), or the following patch disallows pointer arguments
for all of BIT_{IOR,XOR,AND}_EXPR with the exception of BIT_AND_EXPR
with INTEGER_CST last operand (for simpler pointer realignment).
I had to tweak one reassoc optimization and the two match.pd
simplifications.
2022-09-14 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/106878
* tree-cfg.cc (verify_gimple_assign_binary): Disallow pointer,
reference or OFFSET_TYPE BIT_IOR_EXPR, BIT_XOR_EXPR or, unless
the second argument is INTEGER_CST, BIT_AND_EXPR.
* match.pd ((type) X op CST -> (type) (X op ((type-x) CST)),
(type) (((type2) X) op Y) -> (X op (type) Y)): Punt for
POINTER_TYPE_P or OFFSET_TYPE.
* tree-ssa-reassoc.cc (optimize_range_tests_cmp_bitwise): For
pointers cast them to pointer sized integers first.
* gcc.c-torture/compile/pr106878.c: New test.
The following avoids creating BIT_FIELD_REF of bitfields in
update-address-taken. The patch doesn't implement punning to
a full precision integer type but leaves a comment according to
that.
PR tree-optimization/106934
* tree-ssa.cc (non_rewritable_mem_ref_base): Avoid BIT_FIELD_REFs
of bitfields.
(maybe_rewrite_mem_ref_base): Likewise.
* gfortran.dg/pr106934.f90: New testcase.
in vectorizable_nonlinear_induction, r13-2503-gc13223b790bbc5 prevent variable peeling by
only checking LOOP_VINFO_MASK_SKIP_NITERS (loop_vinfo). But when
"!vect_use_loop_mask_for_alignment_p (loop_vinfo) &&
LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0", vectorizer will
still do variable peeling for epilog, and it hits gcc_assert in
vect_peel_nonlinear_iv_init.
gcc/ChangeLog:
PR tree-optimization/106905
* tree-vect-loop.cc (vectorizable_nonlinear_induction): Return
false when !vect_use_loop_mask_for_alignment_p (loop_vinfo) &&
LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo) < 0.
gcc/testsuite/ChangeLog:
* gcc.target/i386/pr106905.c: New test.
* gcc.target/ia64/pr106905.c: New test.
When the status wrapper is used, the gluefile need to be prefixed with
-Wl, in order for the test cases to have the dump files with the
expected names.
2022-09-14 Torbjörn SVENSSON <torbjorn.svensson@foss.st.com>
gcc/testsuite/
PR target/95720
* lib/g++.exp: Moved gluefile block to after flags have been
prefixed for the target_compile call.
* lib/gcc.exp: Likewise.
* lib/wrapper.exp: Reset adjusted state flag.
Co-Authored-By: Yvan ROUX <yvan.roux@foss.st.com>
Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com>
This patch resolves PR target/106877 an ICE-on-invalid inline-asm
regression. An innocent upstream change means that the test case
from PR inline-asm/84683 now hits a different assert in reg-stack.cc's
move_for_stack_reg. Fixed by duplicating Jakub's solution to PR 84683
https://gcc.gnu.org/pipermail/gcc-patches/2018-March/495193.html at
this second (similar) gcc_assert.
2022-09-13 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR target/106877
* reg-stack.cc (move_for_stack_reg): Check for any_malformed_asm
in gcc_assert.
gcc/testsuite/ChangeLog
PR target/106877
* g++.dg/ext/pr106877.C: New test case.
While icv_addr[1] = false; assignments where icv_addr has void *
element type is correct and matches how it is used (in those cases
the void * pointer is then cast to bool and used that way), there is no
reason not to add explicit (void *) casts there which are there already
for (void *) true. And, there is in fact even no point in actually
doing those stores at all because we set that pointer to NULL a few
lines earlier. So, this patch adds the explicit casts and then
comments those out to show intent.
2022-09-13 Jakub Jelinek <jakub@redhat.com>
PR libgomp/106906
* env.c (get_icv_member_addr): Cast false to void * before assigning
it to icv_addr[1], and comment the whole assignment out.
libstdc++-v3/ChangeLog:
* include/std/ranges (join_view::_Iterator::_M_satisfy):
Adjust resetting _M_inner as per LWG 3569.
(join_view::_Iterator::_M_inner): Wrap in std::optional
as per LWG 3569.
(join_view::_Iterator::_Iterator): Relax constraints as
per LWG 3569.
(join_view::_Iterator::operator*): Adjust as per LWG 3569.
(join_view::_Iterator::operator->): Likewise.
(join_view::_Iterator::operator++): Likewise.
(join_view::_Iterator::operator--): Likewise.
(join_view::_Iterator::iter_move): Likewise.
(join_view::_Iterator::iter_swap): Likewise.
* testsuite/std/ranges/adaptors/join.cc (test14): New test.
I noticed compiling e.g. std/ranges/adaptors/join.cc with
-D_GLIBCXX_DEBUG -Wsystem-headers -Wall gives the warning:
gcc/libstdc++-v3/include/debug/safe_iterator.h:477:9: warning: suggest parentheses around ‘&&’ within ‘||’ [-Wparentheses]
libstdc++-v3/ChangeLog:
* include/debug/safe_iterator.h (_GLIBCXX_DEBUG_VERIFY_OPERANDS):
Add parentheses to avoid -Wparentheses warning.
Since some callers need the complain parameter but not the object
parameter, let's introduce and use an overload of cxx_constant_value
that omits the latter.
gcc/cp/ChangeLog:
* cp-tree.h (cxx_constant_value): Define two-parameter version
that omits the object parameter.
* decl.cc (build_explicit_specifier): Omit NULL_TREE object
argument to cxx_constant_value.
* except.cc (build_noexcept_spec): Likewise.
* pt.cc (expand_integer_pack): Likewise.
(fold_targs_r): Likewise.
* semantics.cc (finish_if_stmt_cond): Likewise.
It looks like we aren't respecting SFINAE for:
* an invalid/non-constant conditional explicit-specifier
* a non-constant conditional noexcept-specifier
* a non-constant argument to __integer_pack
This patch fixes these in the usual way, by passing complain and
propagating error_mark_node appropriately.
gcc/cp/ChangeLog:
* decl.cc (build_explicit_specifier): Pass complain to
cxx_constant_value.
* except.cc (build_noexcept_spec): Likewise.
* pt.cc (expand_integer_pack): Likewise.
(tsubst_function_decl): Propagate error_mark_node returned
from build_explicit_specifier.
gcc/testsuite/ChangeLog:
* g++.dg/cpp1z/noexcept-type26.C: New test.
* g++.dg/cpp2a/explicit19.C: New test.
* g++.dg/ext/integer-pack6.C: New test.
As PR104482 shown, it's one regression about the handlings when
the argument number is more than the one of built-in function
prototype. The new bif support only catches the case that the
argument number is less than the one of function prototype, but
it misses the case that the argument number is more than the one
of function prototype. Because it uses "n != expected_args",
n is updated in
for (n = 0; !VOID_TYPE_P (TREE_VALUE (fnargs)) && n < nargs;
fnargs = TREE_CHAIN (fnargs), n++)
, it's restricted to be less than or equal to expected_args with
the guard !VOID_TYPE_P (TREE_VALUE (fnargs)), so it's wrong.
The fix is to use nargs instead, also move the checking hunk's
location ahead to avoid useless further scanning when the counts
mismatch.
PR target/104482
gcc/ChangeLog:
* config/rs6000/rs6000-c.cc (altivec_resolve_overloaded_builtin): Fix
the equality check for argument number, and move this hunk ahead.
gcc/testsuite/ChangeLog:
* gcc.target/powerpc/pr104482.c: New test.
PR105485 exposes that new builtin function framework doesn't handle
unresolved overloaded builtin function well. With new builtin
function support, we don't have builtin info for any overloaded
rs6000_gen_builtins enum, since they are expected to be resolved to
one specific instance. So when function rs6000_gimple_fold_builtin
faces one unresolved overloaded builtin, the access for builtin info
becomes out of bound and gets ICE then.
We should not try to fold one unresolved overloaded builtin there
and as the previous support we should emit one error message during
expansion phase like "unresolved overload for builtin ...".
PR target/105485
gcc/ChangeLog:
* config/rs6000/rs6000-builtin.cc (rs6000_gimple_fold_builtin): Add
the handling for unresolved overloaded builtin function.
(rs6000_expand_builtin): Likewise.
gcc/testsuite/ChangeLog:
* g++.target/powerpc/pr105485.C: New test.
Commit r12-6679-g7ca1582ca60dc8 made vectorizer accept one
unroll factor to be applied to vectorization factor when
vectorizing the main loop, it would be suggested by target
when doing costing.
This patch introduces function determine_suggested_unroll_factor
for rs6000 port, to make it be able to suggest the unroll factor
for a given loop being vectorized. Referring to aarch64 port
and basing on the analysis on SPEC2017 performance evaluation
results, it mainly considers these aspects:
1) unroll option and pragma which can disable unrolling for the
given loop;
2) simple hardware resource model with issued non memory access
vector insn per cycle;
3) aggressive heuristics when iteration count is unknown:
- reduction case to break cross iteration dependency;
- emulated gather load;
4) estimated iteration count when iteration count is unknown;
With this patch, SPEC2017 performance evaluation results on
Power8/9/10 are listed below (speedup pct.):
* Power10
- O2: all are neutral (excluding some noises);
- Ofast: 510.parest_r +6.67%, the others are neutral
(use ... for the followings);
- Ofast + unroll: 510.parest_r +5.91%, ...
- Ofast + LTO + PGO: 510.parest_r +3.00%, ...
- Ofast + cheap vect cost: 510.parest_r +6.23%, ...
- Ofast + very-cheap vect cost: all are neutral;
* Power9
- Ofast: 510.parest_r +8.73%, 538.imagick_r +11.18%
(likely noise), 500.perlbench_r +1.84%, ...
* Power8
- Ofast: 510.parest_r +5.43%, ...;
This patch also introduces one documented parameter
rs6000-vect-unroll-limit= similar to what aarch64 proposes,
by evaluating on P8/P9/P10, the default value 4 is slightly
better than the other choices like 2 and 8.
It also parameterizes two other values as undocumented
parameters for future tweaking. One parameter is
rs6000-vect-unroll-issue, it's to simply model hardware
resource for non memory access vector instructions to avoid
excessive unrolling, initially I tried to use the value in
the hook rs6000_issue_rate, but the evaluation showed it's
bad, so I evaluated different values 2/4/6/8 on P8/P9/P10 at
Ofast, the results showed the default value 4 is good enough
on these different architectures. For a record, choice 8
could make 510.parest_r's gain become smaller or gone on
P8/P9/P10; choice 6 could make 503.bwaves_r degrade by more
than 1% on P8/P10; and choice 2 could make 538.imagick_r
degrade by 3.8%. The other parameter is
rs6000-vect-unroll-reduc-threshold. It's mainly inspired by
510.parest_r and tweaked as it, evaluating with different
values 0/1/2/3 for the threshold, it showed value 1 is the
best choice. For a record, choice 0 could make 525.x264_r
degrade by 2% and 527.cam4_r degrade by 2.95% on P10,
548.exchange2_r degrade by 1.41% and 527.cam4_r degrade by
2.54% on P8; choice 2 and bigger values could make
510.parest_r's gain become smaller.
gcc/ChangeLog:
* config/rs6000/rs6000.cc (class rs6000_cost_data): Add new members
m_nstores, m_reduc_factor, m_gather_load and member function
determine_suggested_unroll_factor.
(rs6000_cost_data::update_target_cost_per_stmt): Update for m_nstores,
m_reduc_factor and m_gather_load.
(rs6000_cost_data::determine_suggested_unroll_factor): New function.
(rs6000_cost_data::finish_cost): Use determine_suggested_unroll_factor.
* config/rs6000/rs6000.opt (rs6000-vect-unroll-limit): New parameter.
(rs6000-vect-unroll-issue): Likewise.
(rs6000-vect-unroll-reduc-threshold): Likewise.
* doc/invoke.texi (rs6000-vect-unroll-limit): Document new parameter.
The following makes sure to clear the CTRL altering flag when
folding emits a __builitin_unreachable in place of a virtual call
which now might become a trap.
PR middle-end/106909
* gimple-fold.cc (gimple_fold_call): Clear the ctrl-altering
flag of a unreachable call.
The following avoids setting and not clearing an auto_bb_flag
on EXIT_BLOCK which we don't verify for such stale flags but
dump_bb_info still asserts on them.
PR tree-optimization/106913
* tree-ssa-uninit.cc (warn_uninitialized_vars): Do not set
ft_reachable on EXIT_BLOCK.
This patch fixes various issues around the handling of vectors
and (particularly) vector structures with +nosimd. Previously,
passing and returning structures would trigger an ICE, since:
* we didn't allow the structure modes to be stored in FPRs
* we didn't provide +nosimd move patterns
* splitting the moves into word-sized pieces (the default
strategy without move patterns) doesn't work because the
registers are doubleword sized.
The patch is a bit of a hodge-podge since a lot of the handling of
moves, register costs, and register legitimacy is so interconnected.
It didn't seem feasible to split things further.
Some notes:
* The patch recognises vector and tuple modes based on TARGET_FLOAT
rather than TARGET_SIMD, and instead adds TARGET_SIMD to places
that really do need the vector ISA. This is necessary for the
modes to be handled correctly in register arguments and returns.
* The 64-bit (DREG) STP peephole required TARGET_SIMD but the
LDP peephole didn't. I think the LDP one is right, since
DREG moves could involve GPRs as well as FPRs.
* The patch keeps the existing choices of instructions for
TARGET_SIMD, just in case they happen to be better than FMOV
on some uarches.
* Before the patch, +nosimd Q<->Q moves of 128-bit scalars went via
a GPR, thanks to a secondary reload pattern. This approach might
not be ideal, but there's no reason that 128-bit vectors should
behave differently from 128-bit scalars. The patch therefore
extends the current scalar approach to vectors.
* Multi-vector LD1 and ST1 require TARGET_SIMD, so the TARGET_FLOAT
structure moves need to use LDP/STP and LDR/STR combinations
instead. That's also what we do for big-endian even with
TARGET_SIMD, so most of the code was already there. The patterns
for structures of 64-bit vectors are identical, but the patterns
for structures of 128-bit vectors need to cope with the lack of
128-bit Q<->Q moves.
It isn't feasible to move multi-vector tuples via GPRs, so the
patch moves them via memory instead. This contaminates the port
with its first secondary memory reload.
gcc/
* config/aarch64/aarch64.cc (aarch64_classify_vector_mode): Use
TARGET_FLOAT instead of TARGET_SIMD.
(aarch64_vectorize_related_mode): Restrict ADVSIMD handling to
TARGET_SIMD.
(aarch64_hard_regno_mode_ok): Don't allow tuples of 2 64-bit vectors
in GPRs.
(aarch64_classify_address): Treat little-endian structure moves
like big-endian for TARGET_FLOAT && !TARGET_SIMD.
(aarch64_secondary_memory_needed): New function.
(aarch64_secondary_reload): Handle 128-bit Advanced SIMD vectors
in the same way as TF, TI and TD.
(aarch64_rtx_mult_cost): Restrict ADVSIMD handling to TARGET_SIMD.
(aarch64_rtx_costs): Likewise.
(aarch64_register_move_cost): Treat a pair of 64-bit vectors
separately from a single 128-bit vector. Handle the cost implied
by aarch64_secondary_memory_needed.
(aarch64_simd_valid_immediate): Restrict ADVSIMD handling to
TARGET_SIMD.
(aarch64_expand_vec_perm_const_1): Likewise.
(TARGET_SECONDARY_MEMORY_NEEDED): New macro.
* config/aarch64/iterators.md (VTX): New iterator.
* config/aarch64/aarch64.md (arches): Add fp_q as a synonym of simd.
(arch_enabled): Adjust accordingly.
(@aarch64_reload_mov<TX:mode>): Extend to...
(@aarch64_reload_mov<VTX:mode>): ...this.
* config/aarch64/aarch64-simd.md (mov<mode>): Require TARGET_FLOAT
rather than TARGET_SIMD.
(movmisalign<mode>): Likewise.
(load_pair<DREG:mode><DREG2:mode>): Likewise.
(vec_store_pair<DREG:mode><DREG2:mode>): Likewise.
(load_pair<VQ:mode><VQ2:mode>): Likewise.
(vec_store_pair<VQ:mode><VQ2:mode>): Likewise.
(@aarch64_split_simd_mov<mode>): Likewise.
(aarch64_get_low<mode>): Likewise.
(aarch64_get_high<mode>): Likewise.
(aarch64_get_half<mode>): Likewise. Canonicalize to a move for
lowpart extracts.
(*aarch64_simd_mov<VDMOV:mode>): Require TARGET_FLOAT rather than
TARGET_SIMD. Use different w<-w and r<-w instructions for
!TARGET_SIMD. Disable immediate moves for !TARGET_SIMD but
add an alternative specifically for w<-Z.
(*aarch64_simd_mov<VQMOV:mode>): Require TARGET_FLOAT rather than
TARGET_SIMD. Likewise for the associated define_splits. Disable
FPR moves and immediate moves for !TARGET_SIMD but add an alternative
specifically for w<-Z.
(aarch64_simd_mov_from_<mode>high): Require TARGET_FLOAT rather than
TARGET_SIMD. Restrict the existing alternatives to TARGET_SIMD
but add a new r<-w one for !TARGET_SIMD.
(*aarch64_get_high<mode>): New pattern.
(load_pair_lanes<mode>): Require TARGET_FLOAT rather than TARGET_SIMD.
(store_pair_lanes<mode>): Likewise.
(*aarch64_combine_internal<mode>): Likewise. Restrict existing
w<-w, w<-r and w<-m alternatives to TARGET_SIMD but add a new w<-r
alternative for !TARGET_SIMD.
(*aarch64_combine_internal_be<mode>): Likewise.
(aarch64_combinez<mode>): Require TARGET_FLOAT rather than TARGET_SIMD.
Remove bogus arch attribute.
(*aarch64_combinez_be<mode>): Likewise.
(@aarch64_vec_concat<mode>): Require TARGET_FLOAT rather than
TARGET_SIMD.
(aarch64_combine<mode>): Likewise.
(aarch64_rev_reglist<mode>): Likewise.
(mov<mode>): Likewise.
(*aarch64_be_mov<VSTRUCT_2D:mode>): Extend to TARGET_FLOAT &&
!TARGET_SIMD, regardless of endianness. Extend associated
define_splits in the same way, both for this pattern and the
ones below.
(*aarch64_be_mov<VSTRUCT_2Qmode>): Likewise. Restrict w<-w
alternative to TARGET_SIMD.
(*aarch64_be_movoi): Likewise.
(*aarch64_be_movci): Likewise.
(*aarch64_be_movxi): Likewise.
(*aarch64_be_mov<VSTRUCT_4QD:mode>): Extend to TARGET_FLOAT
&& !TARGET_SIMD, regardless of endianness. Restrict w<-w alternative
to TARGET_SIMD for tuples of 128-bit vectors.
(*aarch64_be_mov<VSTRUCT_4QD:mode>): Likewise.
* config/aarch64/aarch64-ldpstp.md: Remove TARGET_SIMD condition
from DREG STP peephole. Change TARGET_SIMD to TARGET_FLOAT in
the VQ and VP_2E LDP and STP peepholes.
gcc/testsuite/
* gcc.target/aarch64/ldp_stp_20.c: New test.
* gcc.target/aarch64/ldp_stp_21.c: Likewise.
* gcc.target/aarch64/ldp_stp_22.c: Likewise.
* gcc.target/aarch64/ldp_stp_23.c: Likewise.
* gcc.target/aarch64/ldp_stp_24.c: Likewise.
* gcc.target/aarch64/movv16qi_1.c (gpr_to_gpr): New function.
* gcc.target/aarch64/movv8qi_1.c (gpr_to_gpr): Likewise.
* gcc.target/aarch64/movv16qi_2.c: New test.
* gcc.target/aarch64/movv16qi_3.c: Likewise.
* gcc.target/aarch64/movv2di_1.c: Likewise.
* gcc.target/aarch64/movv2x16qi_1.c: Likewise.
* gcc.target/aarch64/movv2x8qi_1.c: Likewise.
* gcc.target/aarch64/movv3x16qi_1.c: Likewise.
* gcc.target/aarch64/movv3x8qi_1.c: Likewise.
* gcc.target/aarch64/movv4x16qi_1.c: Likewise.
* gcc.target/aarch64/movv4x8qi_1.c: Likewise.
* gcc.target/aarch64/movv8qi_2.c: Likewise.
* gcc.target/aarch64/movv8qi_3.c: Likewise.
* gcc.target/aarch64/vect_unary_2.c: Likewise.
The ls64-related move expanders and splits required TARGET_SIMD.
That isn't necessary, since the 64-byte values are stored entirely
in GPRs. (The associated define_insn was already correct.)
I wondered about moving the patterns to aarch64.md, but it wasn't
clear-cut.
gcc/
* config/aarch64/aarch64-simd.md (movv8di): Remove TARGET_SIMD
condition. Likewise for the related define_split. Tweak formatting.
gcc/testsuite/
* gcc.target/aarch64/acle/ls64_asm_2.c: New test.
I misplaced one remark into 'gcn' instead of 'nvptx' in
commit r13-2625-g6b43f556f392a7165582aca36a19fe7389d995b2
libgomp/ChangeLog:
* libgomp.texi (gcn): Move misplaced -march=sm_30 remark to ...
(nvptx): ... here.
The functions
abstract_virtuals_error
cxx_constant_value
get_target_expr
instantiate_non_dependent_expr
require_complete_type
are each just a non-SFINAE-enabled wrapper for the corresponding
SFINAE-enabled version that's suffixed by '_sfinae'. But this suffix is
at best redundant since a 'complain' parameter already broadly conveys
that a function is SFINAE-enabled, and having two such versions of a
function is less concise than just using a default argument for 'complain'
(and arguably no less mistake prone).
So this patch squashes the two versions of each of the above functions
by adding a default 'complain' argument to the SFINAE-enabled version
whose '_sfinae' suffix we then remove.
gcc/cp/ChangeLog:
* call.cc (build_conditional_expr): Adjust calls to
'_sfinae'-suffixed functions.
(build_temp): Likewise.
(convert_like_internal): Likewise.
(convert_arg_to_ellipsis): Likewise.
(build_over_call): Likewise.
(build_cxx_call): Likewise.
(build_new_method_call): Likewise.
* constexpr.cc (cxx_eval_outermost_constant_expr): Likewise.
(cxx_constant_value_sfinae): Rename to ...
(cxx_constant_value): ... this. Document its default arguments.
(fold_non_dependent_expr): Adjust function comment.
* cp-tree.h (instantiate_non_dependent_expr_sfinae): Rename to ...
(instantiate_non_dependent_expr): ... this. Give its 'complain'
parameter a default argument.
(get_target_expr_sfinae, get_target_expr): Likewise.
(require_complete_type_sfinae, require_complete_type): Likewise.
(abstract_virtuals_error_sfinae, abstract_virtuals_error):
Likewise.
(cxx_constant_value_sfinae, cxx_constant_value): Likewise.
* cvt.cc (build_up_reference): Adjust calls to '_sfinae'-suffixed
functions.
(ocp_convert): Likewise.
* decl.cc (build_explicit_specifier): Likewise.
* except.cc (build_noexcept_spec): Likewise.
* init.cc (build_new_1): Likewise.
* pt.cc (expand_integer_pack): Likewise.
(instantiate_non_dependent_expr_internal): Adjust function
comment.
(instantiate_non_dependent_expr): Rename to ...
(instantiate_non_dependent_expr_sfinae): ... this. Document its
default argument.
(tsubst_init): Adjust calls to '_sfinae'-suffixed functions.
(fold_targs_r): Likewise.
* semantics.cc (finish_compound_literal): Likewise.
(finish_decltype_type): Likewise.
(cp_build_bit_cast): Likewise.
* tree.cc (build_cplus_new): Likewise.
(get_target_expr): Rename to ...
(get_target_expr_sfinae): ... this. Document its default
argument.
* typeck.cc (require_complete_type): Rename to ...
(require_complete_type_sfinae): ... this. Document its default
argument.
(cp_build_array_ref): Adjust calls to '_sfinae'-suffixed
functions.
(convert_arguments): Likewise.
(cp_build_binary_op): Likewise.
(build_static_cast_1): Likewise.
(cp_build_modify_expr): Likewise.
(convert_for_initialization): Likewise.
* typeck2.cc (abstract_virtuals_error): Rename to ...
(abstract_virtuals_error_sfinae): ... this. Document its default
argument.
(build_functional_cast_1): Adjust calls to '_sfinae'-suffixed
functions.
Here we're neglecting to clear cp_unevaluated_operand when substituting
into the arguments of the alias template-id 'skip<(T(), 0), T>' with T=A,
which means cp_unevaluated_operand remains set during mark_used for
A::A() and so we don't synthesize it. Later constant evaluation for
the substituted template argument '(A(), 0)' (from coerce_template_parms)
fails with "'constexpr A::A()' used before its definition" since it was
never synthesized.
This doesn't happen with a class template because tsubst_aggr_type
clears cp_unevaluated_operand during substitution thereof. But since
template arguments are generally manifestly constant-evaluated, which in
turn are evaluated even in an unevaluated operand, we should be clearing
cp_unevaluated_operand more broadly whenever substituting into any set
of template arguments. To that end this patch makes us clear it during
tsubst_template_args.
PR c++/101906
gcc/cp/ChangeLog:
* pt.cc (tsubst_template_args): Set cp_evaluated here.
(tsubst_aggr_type): Not here.
gcc/testsuite/ChangeLog:
* g++.dg/template/evaluated1.C: New test.
* g++.dg/template/evaluated1a.C: New test.
* g++.dg/template/evaluated1b.C: New test.
* g++.dg/template/evaluated1c.C: New test.
As with PR105623, we need to call mark_single_function sooner to
resolve the type of a BASELINK.
PR c++/106893
PR c++/90451
gcc/cp/ChangeLog:
* decl.cc (cp_finish_decl): Call mark_single_function.
gcc/testsuite/ChangeLog:
* g++.dg/cpp1y/auto-fn65.C: New test.
We already know to treat a variable of array-of-unknown-bound type as
dependent, we should do the same for arr{}.
PR c++/93259
gcc/cp/ChangeLog:
* pt.cc (type_dependent_expression_p): Treat a compound
literal of array-of-unknown-bound type like a variable.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/initlist-array17.C: New test.
This patch implements new target hook TARGET_CONSTANT_OK_FOR_CPROP_P in
order to exclude CONST_INTs that cannot fit into a MOVI machine instruction
from cprop.
gcc/ChangeLog:
* config/xtensa/xtensa.cc (TARGET_CONSTANT_OK_FOR_CPROP_P):
New macro definition.
(xtensa_constant_ok_for_cprop_p):
Implement the hook as mentioned above.
Although PR106320 affected only the 10 and 11 branches, and the testcase
from there is already correctly accepted on trunk and the 12 branch, we
still should add the testcase to trunk/12 too for inter-branch consistency.
PR libstdc++/106320
libstdc++-v3/ChangeLog:
* testsuite/std/ranges/adaptors/join.cc (test13): New test.
We can't use the type of an array variable directly if we haven't deduced
its length yet.
PR c++/106567
gcc/cp/ChangeLog:
* lambda.cc (type_deducible_expression_p): Check
array_of_unknown_bound_p.
gcc/testsuite/ChangeLog:
* g++.dg/cpp0x/lambda/lambda-array4.C: New test.
Since C++11 relaxed the requirement for template arguments to have
external linkage, it's possible to get -Wsubobject-linkage warnings
without using any anonymous namespaces. This confuses users when they
get diagnostics that refer to an anonymous namespace that doesn't exist
in their code.
This changes the diagnostic to say "has internal linkage" for C++11 and
later, if the type isn't actually a member of the anonymous namespace.
Making that distinction involved renaming the current decl_anon_ns_mem_p to
something that better expresses its semantics.
For C++98 template arguments declared with 'static' are ill-formed
anyway, so the only way this warning can arise is via anonymous
namespaces. That means the existing wording is accurate for C++98 and so
we can keep it.
PR c++/86491
gcc/cp/ChangeLog:
* decl2.cc (constrain_class_visibility): Adjust wording of
-Wsubobject-linkage for cases where anonymous
namespaces aren't used.
* tree.cc (decl_anon_ns_mem_p): Now only true for actual anonymous
namespace members, rename old semantics to...
(decl_internal_context_p): ...this.
* cp-tree.h, name-lookup.cc, pt.cc: Adjust.
gcc/testsuite/ChangeLog:
* g++.dg/warn/anonymous-namespace-3.C: Use separate dg-warning
directives for C++98 and everything else.
* g++.dg/warn/Wsubobject-linkage-5.C: New test.
The <stdatomic.h> macro ATOMIC_VAR_INIT, previously declared obsolete,
is removed completely in C2x; disable it for C2x in GCC's
implementation. (Although ATOMIC_* are reserved names for this
header, disabling the macro for C2x still seems appropriate.)
Bootstrapped with no regressions for x86_64-pc-linux-gnu.
gcc/
* ginclude/stdatomic.h [defined __STDC_VERSION__ &&
__STDC_VERSION__ > 201710L] (ATOMIC_VAR_INIT): Do not define.
gcc/testsuite/
* gcc.dg/atomic/c2x-stdatomic-var-init-1.c: New test.
Reverse offload requests at least -misa=sm_35; with this patch, a warning
instead of an error is shown, still permitting reverse offload for all
other configured device types. This is achieved by not calling
GOMP_offload_register_ver (and stopping generating pointless 'static const char'
variables, once known.)
The tool_name as progname changes adds "nvptx " and "gcn " to the
"mkoffload: warning/error:" diagnostic.
gcc/ChangeLog:
* config/nvptx/mkoffload.cc (process): Replace a fatal_error by
a warning + not enabling offloading if -misa=sm_30 prevents
reverse offload.
(main): Use tool_name as progname for diagnostic.
* config/gcn/mkoffload.cc (main): Likewise.
libgomp/ChangeLog:
* libgomp.texi (Offload-Target Specifics: nvptx): Document
that reverse offload requires >= -march=sm_35.
* testsuite/libgomp.c-c++-common/requires-4.c: Build for nvptx
with -misa=sm_35.
* testsuite/libgomp.c-c++-common/requires-5.c: Likewise.
* testsuite/libgomp.c-c++-common/requires-6.c: Likewise.
* testsuite/libgomp.c-c++-common/reverse-offload-1.c: Likewise.
* testsuite/libgomp.fortran/reverse-offload-1.f90: Likewise.
* testsuite/libgomp.c/reverse-offload-sm30.c: New test.
We should avoid pessimizing the signbit when it's already correct. In
this particular case we were trying to change the signbit to "unknown",
when it was obviously negative.
This test is actually slated for removal with my upcoming revamp of the
signbit and NAN tracking, per the conversations regarding tristate. The
signbit will be removed in favor of keeping track of it in the range
itself, and NAN will just be a pair of boolean flags. However, I don't
plan to disturb the tree until after Cauldron.
Tested on x86-64 Linux including mpfr tests. Also tested selftests with
-ffinite-math-only on x86-64 as well as on a cross to pdp11-aout.
gcc/ChangeLog:
* value-range.cc (frange::set_signbit): Avoid changing sign when
already in the correct sign.
xtensa may use up to 4 registers to return a value from a function, but
recognition of only one register in the xtensa_function_value_regno_p
and missing untyped_call pattern result in that only one register is
saved by the __builtin_apply and returned by the __builtin_apply_return.
gcc/
* config/xtensa/xtensa.cc (xtensa_function_value_regno_p):
Recognize all 4 return registers.
* config/xtensa/xtensa.h (GP_RETURN_REG_COUNT): New definition.
* config/xtensa/xtensa.md (untyped_call): New pattern.
When testing a cross toolchain outside the build tree, the binary name
for gcov is prefixed with the triplet.
gcc/testsuite/ChangeLog:
* g++.dg/gcov/gcov.exp: Respect triplet when looking for gcov.
* gcc.misc-tests/gcov.exp: Likewise.
Signed-off-by: Torbjörn SVENSSON <torbjorn.svensson@foss.st.com>
The following patch is partially a workaround for bogus warnings
when the compiler isn't able to fold _M_disjunct call into constant
false, but also an optimization attempt - assuming _M_disjunct (__s)
is rare, the patch should shrink code size for the common case and
use library or for non-standard instantiations an out of line
function to handle the rare case.
2022-09-12 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/105329
* acinclude.m4 (libtool_VERSION): Change to 6:31:0.
* config/abi/pre/gnu.ver (GLIBCXX_3.4.21): Don't export
std::basic_string methods with name length of 15.
(GLIBCXX_3.4.31): Export std::basic_string::_M_replace_cold.
* testsuite/util/testsuite_abi.cc (check_version): Handle
GLIBCXX_3.4.31.
* include/bits/basic_string.h (std::basic_string::_M_replace_cold):
Declare.
* include/bits/basic_string.tcc (std::basic_string::_M_replace_cold):
Define and export even for C++20.
(std::basic_string::_M_replace): Use __builtin_expect, outline
the overlapping case to _M_replace_cold.
* configure: Regenerated.
