Enable some ipa-sra on fortran by allowing signature changes on functions
with "fn spec" attribute when ipa-modref is enabled. This is possible since ipa-modref
knows how to preserve things we trace in fnspec and fnspec generated by fortran forntend
are quite simple and can be analysed automatically now. To be sure I will also add
code that merge fnspec to parameters.
This unfortunately hits bug in ipa-param-manipulation when we remove parameter
that specifies size of variable length parameter. For this reason I added a hack
that prevent signature changes on such functions and will handle it incrementally.
I tried creating C testcase but it is blocked by another problem that we punt ipa-sra
on access attribute. This is optimization regression we ought to fix so I filled
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103223.
As a followup I will add code classifying the type attributes (we have just few) and
get stats on access attribute.
gcc/ChangeLog:
* ipa-fnsummary.c (compute_fn_summary): Do not give up on signature
changes on "fn spec" attribute; give up on varadic types.
* ipa-param-manipulation.c: Include attribs.h.
(build_adjusted_function_type): New parameter ARG_MODIFIED; if it is
true remove "fn spec" attribute.
(ipa_param_adjustments::build_new_function_type): Update.
(ipa_param_body_adjustments::modify_formal_parameters): update.
* ipa-sra.c: Include attribs.h.
(ipa_sra_preliminary_function_checks): Do not check for TYPE_ATTRIBUTES.
There's no need for two constructors, when we can do it all with one
that defaults to the common behavior:
path_range_query (bool resolve = true, gimple_ranger *ranger = NULL);
Tested on x86-64 Linux.
gcc/ChangeLog:
* gimple-range-path.cc (path_range_query::path_range_query): Merge
ctors.
(path_range_query::import_p): Move from header file.
(path_range_query::~path_range_query): Adjust for combined ctors.
* gimple-range-path.h: Merge ctors.
(path_range_query::import_p): Move to .cc file.
ipa-modref gets confused by EAF flags of memcpy becuase parameter 1 is
escaping but used only directly. In modref we do not track values saved to
memory and thus we clear all other flags on each store. This needs to also
happen when called function escapes parameter.
gcc/ChangeLog:
PR tree-optimization/103182
* ipa-modref.c (callee_to_caller_flags): Fix merging of flags.
(modref_eaf_analysis::analyze_ssa_name): Fix merging of flags.
Since r12-5056-g3439657b0286, there has been a regression in
test results; an additional 100 FAILs running the g++ and
libstdc++ testsuite on cris-elf, a newlib target. The
failures are linker errors, not finding a definition for
getentropy. It appears newlib has since 2017-12-03
declarations of getentropy and arc4random, and provides an
implementation of arc4random using getentropy, but provides no
definition of getentropy, not even a stub yielding ENOSYS.
This is similar to what it does for many other functions too.
While fixing newlib (like adding said stub) would likely help,
it still leaves older newlib releases hanging. Thankfully,
the libstdc++ configury test can be improved to try linking
where possible; using the bespoke GCC_TRY_COMPILE_OR_LINK
instead of AC_TRY_COMPILE. BTW, I see a lack of consistency;
some tests use AC_TRY_COMPILE and some GCC_TRY_COMPILE_OR_LINK
for no apparent reason, but this commit just amends
r12-5056-g3439657b0286.
libstdc++-v3:
PR libstdc++/103166
* acinclude.m4 (GLIBCXX_CHECK_GETENTROPY, GLIBCXX_CHECK_ARC4RANDOM):
Use GCC_TRY_COMPILE_OR_LINK instead of AC_TRY_COMPILE.
* configure: Regenerate.
Recently we changed the PROFILE_HOOK _mcount call to pass in the link
register as an argument. This actually does not work when the _mcount
call uses a PLT because the GOT register setup code ends up getting
inserted before the PROFILE_HOOK and clobbers the link register
argument.
These glibc tests are failing:
gmon/tst-gmon-pie-gprof
gmon/tst-gmon-static-gprof
This patch fixes this by saving the instruction that stores the Link
Register to the _mcount argument and then inserts the GOT register setup
instructions after that.
For example:
main.c:
extern int e;
int f2(int a) {
return a + e;
}
int f1(int a) {
return f2 (a + a);
}
int main(int argc, char ** argv) {
return f1 (argc);
}
Compiled:
or1k-smh-linux-gnu-gcc -Wall -c -O2 -fPIC -pg -S main.c
Before Fix:
main:
l.addi r1, r1, -16
l.sw 8(r1), r2
l.sw 0(r1), r16
l.addi r2, r1, 16 # Keeping FP, but not needed
l.sw 4(r1), r18
l.sw 12(r1), r9
l.jal 8 # GOT Setup clobbers r9 (Link Register)
l.movhi r16, gotpchi(_GLOBAL_OFFSET_TABLE_-4)
l.ori r16, r16, gotpclo(_GLOBAL_OFFSET_TABLE_+0)
l.add r16, r16, r9
l.or r18, r3, r3
l.or r3, r9, r9 # This is not the original LR
l.jal plt(_mcount)
l.nop
l.jal plt(f1)
l.or r3, r18, r18
l.lwz r9, 12(r1)
l.lwz r16, 0(r1)
l.lwz r18, 4(r1)
l.lwz r2, 8(r1)
l.jr r9
l.addi r1, r1, 16
After the fix:
main:
l.addi r1, r1, -12
l.sw 0(r1), r16
l.sw 4(r1), r18
l.sw 8(r1), r9
l.or r18, r3, r3
l.or r3, r9, r9 # We now have r9 (LR) set early
l.jal 8 # Clobbers r9 (Link Register)
l.movhi r16, gotpchi(_GLOBAL_OFFSET_TABLE_-4)
l.ori r16, r16, gotpclo(_GLOBAL_OFFSET_TABLE_+0)
l.add r16, r16, r9
l.jal plt(_mcount)
l.nop
l.jal plt(f1)
l.or r3, r18, r18
l.lwz r9, 8(r1)
l.lwz r16, 0(r1)
l.lwz r18, 4(r1)
l.jr r9
l.addi r1, r1, 12
Fixes: 308531d148 ("or1k: Add return address argument to _mcount call")
gcc/ChangeLog:
* config/or1k/or1k-protos.h (or1k_profile_hook): New function.
* config/or1k/or1k.h (PROFILE_HOOK): Change macro to reference
new function or1k_profile_hook.
* config/or1k/or1k.c (struct machine_function): Add new field
set_mcount_arg_insn.
(or1k_profile_hook): New function.
(or1k_init_pic_reg): Update to inject pic rtx after _mcount arg
when profiling.
(or1k_frame_pointer_required): Frame pointer no longer needed
when profiling.
I introduced bug into find_func_aliases_for_call in handling pure functions.
Instead of reading global memory pure functions are believed to write global
memory. This results in misoptimization of the testcase at -O1.
The change to pta-callused.c updates the template for new behaviour of the
constraint generation. We copy nonlocal memory to calluse which is correct but
also not strictly necessary because later we take care to add nonlocal_p flag
manually.
gcc/ChangeLog:
PR tree-optimization/103209
* tree-ssa-structalias.c (find_func_aliases_for_call): Fix
use of handle_rhs_call
gcc/testsuite/ChangeLog:
PR tree-optimization/103209
* gcc.dg/tree-ssa/pta-callused.c: Update template.
* gcc.c-torture/execute/pr103209.c: New test.
PHIs must be resolved first while solving ranges in a block,
regardless of where they appear in the import bitmap. We went through
a similar exercise for the relational code, but missed these.
Tested on x86-64 & ppc64le Linux.
gcc/ChangeLog:
PR tree-optimization/103202
* gimple-range-path.cc
(path_range_query::compute_ranges_in_block): Solve PHI imports first.
Previous patches made some of the complex parts of the issue rate
code redundant.
gcc/
* config/aarch64/aarch64.c (aarch64_vector_op::n_advsimd_ops): Delete.
(aarch64_vector_op::m_seen_loads): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Don't push to
m_advsimd_ops.
(aarch64_vector_op::count_ops): Remove vectype and factor parameters.
Remove code that tries to predict different vec_flags from the
current loop's.
(aarch64_vector_costs::add_stmt_cost): Update accordingly.
Remove m_advsimd_ops handling.
Previously we tried to account for the different issue rates of
the various vector modes by guessing what the Advanced SIMD version
of an SVE loop would look like and what its issue rate was likely to be.
We'd then increase the cost of the SVE loop if the Advanced SIMD loop
might issue more quickly.
This patch moves that logic to better_main_loop_than_p, so that we
can compare loops side-by-side rather than having to guess. This also
means we can apply the issue rate heuristics to *any* vector loop
comparison, rather than just weighting SVE vs. Advanced SIMD.
The actual heuristics are otherwise unchanged. We're just
applying them in a different place.
gcc/
* config/aarch64/aarch64.c (aarch64_vector_costs::m_saw_sve_only_op)
(aarch64_sve_only_stmt_p): Delete.
(aarch64_vector_costs::prefer_unrolled_loop): New function,
extracted from adjust_body_cost.
(aarch64_vector_costs::better_main_loop_than_p): New function,
using heuristics extracted from adjust_body_cost and
adjust_body_cost_sve.
(aarch64_vector_costs::adjust_body_cost_sve): Remove
advsimd_cycles_per_iter and could_use_advsimd parameters.
Update after changes above.
(aarch64_vector_costs::adjust_body_cost): Update after changes above.
-mtune=neoverse-512tvb sets the likely SVE vector length to 128 bits,
but it also takes into account Neoverse V1, which is a 256-bit target.
This patch adds this VF (VL) factor to aarch64_vec_op_count.
gcc/
* config/aarch64/aarch64.c (aarch64_vec_op_count::m_vf_factor):
New member variable.
(aarch64_vec_op_count::aarch64_vec_op_count): Add a parameter for it.
(aarch64_vec_op_count::vf_factor): New function.
(aarch64_vector_costs::aarch64_vector_costs): When costing for
neoverse-512tvb, pass a vf_factor of 2 for the Neoverse V1 version
of an SVE loop.
(aarch64_vector_costs::adjust_body_cost): Read the vf factor
instead of hard-coding 2.
This patch just moves the main cycle estimation routines
into aarch64_vec_op_count.
gcc/
* config/aarch64/aarch64.c
(aarch64_vec_op_count::rename_cycles_per_iter): New function.
(aarch64_vec_op_count::min_nonpred_cycles_per_iter): Likewise.
(aarch64_vec_op_count::min_pred_cycles_per_iter): Likewise.
(aarch64_vec_op_count::min_cycles_per_iter): Likewise.
(aarch64_vec_op_count::dump): Move earlier in file. Dump the
above properties too.
(aarch64_estimate_min_cycles_per_iter): Delete.
(adjust_body_cost): Use aarch64_vec_op_count::min_cycles_per_iter
instead of aarch64_estimate_min_cycles_per_iter. Rely on the dump
routine to print CPI estimates.
(adjust_body_cost_sve): Likewise. Use the other functions above
instead of doing the work inline.
-mtune=neoverse-512tvb uses two issue rates, one for Neoverse V1
and one with more generic parameters. We use both rates when
making a choice between scalar, Advanced SIMD and SVE code.
Previously we calculated the Neoverse V1 issue rates from the
more generic issue rates, but by removing m_scalar_ops and
(later) m_advsimd_ops, it becomes easier to track multiple
issue rates directly.
This patch therefore converts m_ops and (temporarily) m_advsimd_ops
into arrays.
gcc/
* config/aarch64/aarch64.c (aarch64_vec_op_count): Allow default
initialization.
(aarch64_vec_op_count::base_issue_info): Remove handling of null
issue_infos.
(aarch64_vec_op_count::simd_issue_info): Likewise.
(aarch64_vec_op_count::sve_issue_info): Likewise.
(aarch64_vector_costs::m_ops): Turn into a vector.
(aarch64_vector_costs::m_advsimd_ops): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Add entries to
the vectors based on aarch64_tune_params.
(aarch64_vector_costs::analyze_loop_vinfo): Update the pred_ops
of all entries in m_ops.
(aarch64_vector_costs::add_stmt_cost): Call count_ops for all
entries in m_ops.
(aarch64_estimate_min_cycles_per_iter): Remove issue_info
parameter and get the information from the ops instead.
(aarch64_vector_costs::adjust_body_cost_sve): Take a
aarch64_vec_issue_info instead of a aarch64_vec_op_count.
(aarch64_vector_costs::adjust_body_cost): Update call accordingly.
Exit earlier if m_ops is empty for either cost structure.
Now that vector finish_costs is passed the associated scalar costs,
we can record the scalar issue information while computing the scalar
costs, rather than trying to estimate it while computing the vector
costs.
This simplifies things a little, but the main motivation is to improve
accuracy.
gcc/
* config/aarch64/aarch64.c (aarch64_vector_costs::m_scalar_ops)
(aarch64_vector_costs::m_sve_ops): Replace with...
(aarch64_vector_costs::m_ops): ...this.
(aarch64_vector_costs::analyze_loop_vinfo): Update accordingly.
(aarch64_vector_costs::adjust_body_cost_sve): Likewise.
(aarch64_vector_costs::aarch64_vector_costs): Likewise.
Initialize m_vec_flags here rather than in add_stmt_cost.
(aarch64_vector_costs::count_ops): Test for scalar reductions too.
Allow vectype to be null.
(aarch64_vector_costs::add_stmt_cost): Call count_ops for scalar
code too. Don't require vectype to be nonnull.
(aarch64_vector_costs::adjust_body_cost): Take the loop_vec_info
and scalar costs as parameters. Use the scalar costs to determine
the cycles per iteration of the scalar loop, then multiply it
by the estimated VF.
(aarch64_vector_costs::finish_cost): Update call accordingly.
This patch gets the floatness of a memory access from the data
reference rather than the vectype. This makes it more suitable
for use in scalar costing code.
gcc/
* config/aarch64/aarch64.c (aarch64_dr_type): New function.
(aarch64_vector_costs::count_ops): Use it rather than the
vectype to determine floatness.
This patch gets the scalar mode of a reduction operation from the
gimple stmt rather than the vectype. This makes it more suitable
for use in scalar costs.
gcc/
* config/aarch64/aarch64.c (aarch64_sve_in_loop_reduction_latency):
Remove vectype parameter and get floatness from the type of the
stmt lhs instead.
(arch64_in_loop_reduction_latency): Likewise.
(aarch64_detect_vector_stmt_subtype): Update caller.
(aarch64_vector_costs::count_ops): Likewise.
Later patches make aarch64 use the new vector hooks. We then
only need to track one set of ops for each aarch64_vector_costs
structure. This in turn means that it's more convenient to merge
aarch64_sve_op_count and aarch64_vec_op_count.
The patch also adds issue info and vec flags to aarch64_vec_op_count,
so that the structure is more self-descriptive. This simplifies some
things later.
gcc/
* config/aarch64/aarch64.c (aarch64_sve_op_count): Fold into...
(aarch64_vec_op_count): ...this. Add a constructor.
(aarch64_vec_op_count::vec_flags): New function.
(aarch64_vec_op_count::base_issue_info): Likewise.
(aarch64_vec_op_count::simd_issue_info): Likewise.
(aarch64_vec_op_count::sve_issue_info): Likewise.
(aarch64_vec_op_count::m_issue_info): New member variable.
(aarch64_vec_op_count::m_vec_flags): Likewise.
(aarch64_vector_costs): Add a constructor.
(aarch64_vector_costs::m_sve_ops): Change type to aarch64_vec_op_count.
(aarch64_vector_costs::aarch64_vector_costs): New function.
Initialize m_scalar_ops, m_advsimd_ops and m_sve_ops.
(aarch64_vector_costs::count_ops): Remove vec_flags and
issue_info parameters, using the new aarch64_vec_op_count
functions instead.
(aarch64_vector_costs::add_stmt_cost): Update call accordingly.
(aarch64_sve_op_count::dump): Fold into...
(aarch64_vec_op_count::dump): ..here.
For tests like:
int res[2];
void
f1 (int x, int y)
{
res[0] = res[1] = x + y;
}
we generated:
add w0, w0, w1
adrp x1, .LANCHOR0
add x2, x1, :lo12:.LANCHOR0
str w0, [x1, #:lo12:.LANCHOR0]
str w0, [x2, 4]
ret
Using [x1, #:lo12:.LANCHOR0] for the first store prevented the
two stores being recognised as a pair. However, the MEM_EXPR
and MEM_OFFSET information tell us that the MEMs really are
consecutive. The peehole2 context then guarantees that the
first address is equivalent to [x2, 0].
While there: the reg_mentioned_p tests for loads were probably correct,
but seemed a bit indirect. We're matching two consecutive loads,
so the thing we need to test is that the second MEM in the original
sequence doesn't depend on the result of the first load in the
original sequence.
gcc/
* config/aarch64/aarch64.c: Include tree-dfa.h.
(aarch64_check_consecutive_mems): New function that takes MEM_EXPR
and MEM_OFFSET into account.
(aarch64_swap_ldrstr_operands): Use it.
(aarch64_operands_ok_for_ldpstp): Likewise. Check that the
address of the second memory doesn't depend on the result of
the first load.
gcc/testsuite/
* gcc.target/aarch64/stp_1.c: New test.
gcc/fortran/ChangeLog:
* parse.c (decode_omp_directive): Fix permitting 'nowait' for some
combined directives, add missing 'omp end ... loop'.
(gfc_ascii_statement): Fix ST_OMP_END_TEAMS_LOOP result.
* openmp.c (resolve_omp_clauses): Add missing combined loop constructs
case values to the 'if(directive-name: ...)' check.
* trans-openmp.c (gfc_split_omp_clauses): Put nowait on target if
first leaf construct accepting it.
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/unexpected-end.f90: Update dg-error.
* gfortran.dg/gomp/clauses-1.f90: New test.
* gfortran.dg/gomp/nowait-2.f90: New test.
* gfortran.dg/gomp/nowait-3.f90: New test.
My recent libgomp change apparently broke libgomp build for gcn offloading.
The problem is that gcn, unlike nvptx, doesn't override teams.c source file
and the patch I've committed assumed all the non-LIBGOMP_USE_PTHREADS targets
do not use it. My understanding is that gcn included omp_get_num_teams
and omp_get_team_num definitions in both icv-device.o and teams.o,
with the definitions only in the former working correctly.
This patch brings gcn into sync with how nvptx does it, that teams.c
is overridden, provides a dummy GOMP_teams_reg and omp_get_{num_teams,team_num}
definitions and icv-device.c doesn't provide those.
2021-11-12 Jakub Jelinek <jakub@redhat.com>
PR target/103201
* config/gcn/icv-device.c (omp_get_num_teams, omp_get_team_num): Move
to ...
* config/gcn/teams.c: ... here. New file.
This patch converts one more open coded construction of a
DEBUG_EXPR_DECL to a call of build_debug_expr_decl that I missed in my
previous patch befause it happens to be in the Fortran front-end.
gcc/fortran/ChangeLog:
2021-11-11 Martin Jambor <mjambor@suse.cz>
* trans-types.c (gfc_get_array_descr_info): Use build_debug_expr_decl
instead of building DEBUG_EXPR_DECL manually.
Fix issues mentioned in the PR.
PR libbacktrace/103167
libbacktrace/ChangeLog:
* elf.c (elf_uncompress_lzma_block): Cast to unsigned int.
(elf_uncompress_lzma): Likewise.
* xztest.c (test_samples): memcpy only if v > 0.
This patch implements relaxing the requirements when a map with the implicit
attribute encounters an overlapping existing map. As the OpenMP 5.0 spec
describes on page 320, lines 18-27 (and 5.1 spec, page 352, lines 13-22):
"If a single contiguous part of the original storage of a list item with an
implicit data-mapping attribute has corresponding storage in the device data
environment prior to a task encountering the construct that is associated with
the map clause, only that part of the original storage will have corresponding
storage in the device data environment as a result of the map clause."
2021-11-12 Chung-Lin Tang <cltang@codesourcery.com>
include/ChangeLog:
* gomp-constants.h (GOMP_MAP_FLAG_SPECIAL_3): Define special bit macro.
(GOMP_MAP_IMPLICIT): New special map kind bits value.
(GOMP_MAP_FLAG_SPECIAL_BITS): Define helper mask for whole set of
special map kind bits.
(GOMP_MAP_IMPLICIT_P): New predicate macro for implicit map kinds.
gcc/ChangeLog:
* tree.h (OMP_CLAUSE_MAP_RUNTIME_IMPLICIT_P): New access macro for
'implicit' bit, using 'base.deprecated_flag' field of tree_node.
* tree-pretty-print.c (dump_omp_clause): Add support for printing
implicit attribute in tree dumping.
* gimplify.c (gimplify_adjust_omp_clauses_1):
Set OMP_CLAUSE_MAP_RUNTIME_IMPLICIT_P to 1 if map clause is implicitly
created.
(gimplify_adjust_omp_clauses): Adjust place of adding implicitly created
clauses, from simple append, to starting of list, after non-map clauses.
* omp-low.c (lower_omp_target): Add GOMP_MAP_IMPLICIT bits into kind
values passed to libgomp for implicit maps.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/target-implicit-map-1.c: New test.
* c-c++-common/goacc/combined-reduction.c: Adjust scan test pattern.
* c-c++-common/goacc/firstprivate-mappings-1.c: Likewise.
* c-c++-common/goacc/mdc-1.c: Likewise.
* g++.dg/goacc/firstprivate-mappings-1.C: Likewise.
libgomp/ChangeLog:
* target.c (gomp_map_vars_existing): Add 'bool implicit' parameter, add
implicit map handling to allow a "superset" existing map as valid case.
(get_kind): Adjust to filter out GOMP_MAP_IMPLICIT bits in return value.
(get_implicit): New function to extract implicit status.
(gomp_map_fields_existing): Adjust arguments in calls to
gomp_map_vars_existing, and add uses of get_implicit.
(gomp_map_vars_internal): Likewise.
* testsuite/libgomp.c-c++-common/target-implicit-map-1.c: New test.
This replaces the printf used by failed debug assertions with fprintf,
so we can write to stderr.
To avoid including <stdio.h> the assert function is moved into the
library. To avoid programs using a vague linkage definition of the old
inline function, the function is renamed. Code compiled with old
versions of GCC might still call the old function, but code compiled
with the newer GCC will call the new function and write to stderr.
libstdc++-v3/ChangeLog:
PR libstdc++/59675
* acinclude.m4 (libtool_VERSION): Bump version.
* config/abi/pre/gnu.ver (GLIBCXX_3.4.30): Add version and
export new symbol.
* configure: Regenerate.
* include/bits/c++config (__replacement_assert): Remove, declare
__glibcxx_assert_fail instead.
* src/c++11/debug.cc (__glibcxx_assert_fail): New function to
replace __replacement_assert, writing to stderr instead of
stdout.
* testsuite/util/testsuite_abi.cc: Update latest version.
The KIND argument of the INDEX intrinsic is a compile time constant
that is used at compile time only to resolve to a kind-specific library
function. That argument is otherwise completely ignored at runtime, and there is
no code generated for it as the library procedure has no kind argument.
This confuses the scalarizer which expects to see every argument
of elemental functions used when calling a procedure.
This change removes the argument from the scalarization lists
at the beginning of the scalarization process, so that the argument
is completely ignored.
This also reverts the existing workaround
(commit d09847357b except for its testcase).
PR fortran/97896
gcc/fortran/ChangeLog:
* intrinsic.c (add_sym_4ind): Remove.
(add_functions): Use add_sym4 instead of add_sym4ind.
Don’t special case the index intrinsic.
* iresolve.c (gfc_resolve_index_func): Use the individual arguments
directly instead of the full argument list.
* intrinsic.h (gfc_resolve_index_func): Update the declaration
accordingly.
* trans-decl.c (gfc_get_extern_function_decl): Don’t modify the
list of arguments in the case of the index intrinsic.
* trans-array.h (gfc_get_intrinsic_for_expr,
gfc_get_proc_ifc_for_expr): New.
* trans-array.c (gfc_get_intrinsic_for_expr,
arg_evaluated_for_scalarization): New.
(gfc_walk_elemental_function_args): Add intrinsic procedure
as argument. Count arguments. Check arg_evaluated_for_scalarization.
* trans-intrinsic.c (gfc_walk_intrinsic_function): Update call.
* trans-stmt.c (get_intrinsic_for_code): New.
(gfc_trans_call): Update call.
gcc/testsuite/ChangeLog:
* gfortran.dg/index_5.f90: New.
The following patch implements what I've been talking about earlier,
honor that for explicit num_teams clause we create at least the
lower-bound (if not specified, upper-bound) teams in the league.
For host fallback, it still means we only have one thread doing all the
teams, sequentially one after another.
For PTX and GCN, I think the new teams-2.c test and maybe teams-4.c too
will or might fail.
For these offloads, I think it is ok to remove symbols no longer used
from libgomp.a.
If num_teams_lower is bigger than the provided num_blocks or num_workgroups,
we should arrange for gomp_num_teams_var to be num_teams_lower - 1,
stop using the %ctaid.x or __builtin_gcn_dim_pos (0) for omp_get_team_num ()
and instead use for it some .shared var that GOMP_teams4 initializes to
%ctaid.x or __builtin_gcn_dim_pos (0) when first and for !first
increment that by num_blocks or num_workgroups each time and only
return false when we are above num_teams_lower.
Any help with actually implementing this for the 2 architectures highly
appreciated.
2021-11-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-builtins.def (BUILT_IN_GOMP_TEAMS): Remove.
(BUILT_IN_GOMP_TEAMS4): New.
* builtin-types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
* omp-low.c (lower_omp_teams): Use GOMP_teams4 instead of
GOMP_teams, pass to it also num_teams lower-bound expression
or a dup of upper-bound if it is missing and a flag whether
it is the first call or not.
gcc/fortran/
* types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
libgomp/
* libgomp_g.h (GOMP_teams4): Declare.
* libgomp.map (GOMP_5.1): Export GOMP_teams4.
* target.c (GOMP_teams4): New function.
* config/nvptx/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* config/gcn/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* testsuite/libgomp.c/teams-4.c (main): Expect exactly 2
teams instead of <= 2.
* testsuite/libgomp.c-c++-common/teams-2.c: New test.
The following fixes a missed valueization when simplifying
a MEM[&...] combination during valueization.
2021-11-12 Richard Biener <rguenther@suse.de>
PR tree-optimization/103204
* tree-ssa-sccvn.c (valueize_refs_1): Re-valueize the
top operand after folding in an address.
* gcc.dg/torture/pr103204.c: New testcase.
The idea is for opcodes to be able to see whether bfd is compiled
for 64-bit. A lot of --enable-targets=all libopcodes is wasted space
if bfd can't load 64-bit target object files.
* Makefile.def (configure-opcodes): Depend on configure-bfd.
* Makefile.in: Regenerate.
This implements P1004R2 ("Making std::vector constexpr") for C++20.
For now, debug mode vectors are not supported in constant expressions.
To make that work we might need to disable all attaching/detaching of
safe iterators. That can be fixed later.
Co-authored-by: Josh Marshall <joshua.r.marshall.1991@gmail.com>
libstdc++-v3/ChangeLog:
* include/bits/alloc_traits.h (_Destroy): Make constexpr for
C++20 mode.
* include/bits/allocator.h (__shrink_to_fit::_S_do_it):
Likewise.
* include/bits/stl_algobase.h (__fill_a1): Declare _Bit_iterator
overload constexpr for C++20.
* include/bits/stl_bvector.h (_Bit_type, _S_word_bit): Move out
of inline namespace.
(_Bit_reference, _Bit_iterator_base, _Bit_iterator)
(_Bit_const_iterator, _Bvector_impl_data, _Bvector_base)
(vector<bool, A>>): Add constexpr to every member function.
(_Bvector_base::_M_allocate): Initialize storage during constant
evaluation.
(vector<bool, A>::_M_initialize_value): Use __fill_bvector_n
instead of memset.
(__fill_bvector_n): New helper function to replace memset during
constant evaluation.
* include/bits/stl_uninitialized.h (__uninitialized_copy<false>):
Move logic to ...
(__do_uninit_copy): New function.
(__uninitialized_fill<false>): Move logic to ...
(__do_uninit_fill): New function.
(__uninitialized_fill_n<false>): Move logic to ...
(__do_uninit_fill_n): New function.
(__uninitialized_copy_a): Add constexpr. Use __do_uninit_copy.
(__uninitialized_move_a, __uninitialized_move_if_noexcept_a):
Add constexpr.
(__uninitialized_fill_a): Add constexpr. Use __do_uninit_fill.
(__uninitialized_fill_n_a): Add constexpr. Use
__do_uninit_fill_n.
(__uninitialized_default_n, __uninitialized_default_n_a)
(__relocate_a_1, __relocate_a): Add constexpr.
* include/bits/stl_vector.h (_Vector_impl_data, _Vector_impl)
(_Vector_base, vector): Add constexpr to every member function.
(_Vector_impl::_S_adjust): Disable ASan annotation during
constant evaluation.
(_Vector_base::_S_use_relocate): Disable bitwise-relocation
during constant evaluation.
(vector::_Temporary_value): Use a union for storage.
* include/bits/vector.tcc (vector, vector<bool>): Add constexpr
to every member function.
* include/std/vector (erase_if, erase): Add constexpr.
* testsuite/23_containers/headers/vector/synopsis.cc: Add
constexpr for C++20 mode.
* testsuite/23_containers/vector/bool/cmp_c++20.cc: Change to
compile-only test using constant expressions.
* testsuite/23_containers/vector/bool/capacity/29134.cc: Adjust
namespace for _S_word_bit.
* testsuite/23_containers/vector/bool/modifiers/insert/31370.cc:
Likewise.
* testsuite/23_containers/vector/cmp_c++20.cc: Likewise.
* testsuite/23_containers/vector/cons/89164.cc: Adjust errors
for C++20 and move C++17 test to ...
* testsuite/23_containers/vector/cons/89164_c++17.cc: ... here.
* testsuite/23_containers/vector/bool/capacity/constexpr.cc: New test.
* testsuite/23_containers/vector/bool/cons/constexpr.cc: New test.
* testsuite/23_containers/vector/bool/element_access/constexpr.cc: New test.
* testsuite/23_containers/vector/bool/modifiers/assign/constexpr.cc: New test.
* testsuite/23_containers/vector/bool/modifiers/constexpr.cc: New test.
* testsuite/23_containers/vector/bool/modifiers/swap/constexpr.cc: New test.
* testsuite/23_containers/vector/capacity/constexpr.cc: New test.
* testsuite/23_containers/vector/cons/constexpr.cc: New test.
* testsuite/23_containers/vector/data_access/constexpr.cc: New test.
* testsuite/23_containers/vector/element_access/constexpr.cc: New test.
* testsuite/23_containers/vector/modifiers/assign/constexpr.cc: New test.
* testsuite/23_containers/vector/modifiers/constexpr.cc: New test.
* testsuite/23_containers/vector/modifiers/swap/constexpr.cc: New test.
Since r12-5072 made _Safe_container::operator=(const _Safe_container&)
protected, the debug containers no longer compile in C++98 mode. They
have user-provided copy assignment operators in C++98 mode, and they
assign each base class in turn. The 'this->_M_safe() = __x' expressions
fail, because calling a protected member function is only allowed via
'this'. They could be fixed by using this->_Safe::operator=(__x) but a
simpler solution is to just remove the user-provided assignment
operators and let the compiler define them (as we do for C++11 and
later, by defining them as defaulted).
The only change needed for that to work is to define the _Safe_vector
copy assignment operator in C++98 mode, so that the implicit
__gnu_debug::vector::operator= definition will call it, instead of
needing to call _M_update_guaranteed_capacity() manually.
libstdc++-v3/ChangeLog:
* include/debug/deque (deque::operator=(const deque&)): Remove
definition.
* include/debug/list (list::operator=(const list&)): Likewise.
* include/debug/map.h (map::operator=(const map&)): Likewise.
* include/debug/multimap.h (multimap::operator=(const multimap&)):
Likewise.
* include/debug/multiset.h (multiset::operator=(const multiset&)):
Likewise.
* include/debug/set.h (set::operator=(const set&)): Likewise.
* include/debug/string (basic_string::operator=(const basic_string&)):
Likewise.
* include/debug/vector (vector::operator=(const vector&)):
Likewise.
(_Safe_vector::operator=(const _Safe_vector&)): Define for
C++98 as well.
All users of path_range_query are currently allocating a gimple_ranger
only to pass it to the query object. It's tidier to just do it from
path_range_query if no ranger was passed.
Tested on x86-64 Linux.
gcc/ChangeLog:
* gimple-range-path.cc (path_range_query::path_range_query): New
ctor without a ranger.
(path_range_query::~path_range_query): Free ranger if necessary.
(path_range_query::range_on_path_entry): Adjust m_ranger for pointer.
(path_range_query::ssa_range_in_phi): Same.
(path_range_query::compute_ranges_in_block): Same.
(path_range_query::compute_imports): Same.
(path_range_query::compute_ranges): Same.
(path_range_query::range_of_stmt): Same.
(path_range_query::compute_outgoing_relations): Same.
* gimple-range-path.h (class path_range_query): New ctor.
* tree-ssa-loop-ch.c (ch_base::copy_headers): Remove gimple_ranger
as path_range_query allocates one.
* tree-ssa-threadbackward.c (class back_threader): Remove m_ranger.
(back_threader::~back_threader): Same.
We have much more thorough restrictions, that are shared between both
threader implementations, in the registry. I've been meaning to
remove the backward threader one, since it's only purpose was reducing
the search space. Previously there was a small time penalty for its
removal, but with the various patches in the past month, it looks like
the removal is a wash performance wise.
This catches 8 more jump threads in the backward threader in my suite.
Presumably, because we disallowed all loop crossing, whereas the
registry restrictions allow some crossing (if we exit the loop, etc).
Tested on x86-64 Linux.
gcc/ChangeLog:
* tree-ssa-threadbackward.c
(back_threader_profitability::profitable_path_p): Remove loop
crossing restriction.
Fix the Create_func_descriptors pass to traverse the subexpressions of
the function in a Call_expression. There are no subexpressions in the
normal case of calling a function a method directly, but there are
subexpressions when in code like F().M() when F returns an interface type.
Forgetting to traverse the function subexpressions was almost entirely
hidden by the fact that we also created the necessary thunks in
Bound_method_expression::do_flatten and
Interface_field_reference_expression::do_get_backend. However, when
the thunks were created there, they did not go through the
order_evaluations pass. This almost always worked, but failed in the
case in which the function being thunked returned multiple results, as
order_evaluations takes the necessary step of moving the
Call_expression into its own statement, and that would not happen when
order_evaluations was not called. Avoid hiding errors like this by
changing those methods to only lookup the previously created thunk,
rather than creating it if it was not already created.
The test case for this is https://golang.org/cl/363156.
Fixes https://golang.org/issue/49512
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/363274
Calling the placement version of ::operator new "implicitly creates
objects in the returned region of storage" as per [intro.object]. This
allows the returned memory to be used as storage for implicit-lifetime
types (including arrays) without additional action by the caller. This
is required by the proposed resolution of LWG 3147.
libstdc++-v3/ChangeLog:
* include/std/memory_resource (memory_resource::allocate):
Implicitly create objects in the returned storage.
This function only exists to avoid an error in the debug mode vector, so
doesn't need to be public.
libstdc++-v3/ChangeLog:
* include/bits/stl_bvector.h (vector<bool>::data()): Give
protected access, and delete for C++11 and later.
As discussed on the mailing list the template actually tests for missed
optimization where we fail to pragate size of an array. We no longer miss this
after modref improvements.
gcc/testsuite/ChangeLog:
2021-11-11 Jan Hubicka <hubicka@ucw.cz>
* gfortran.dg/inline_matmul_17.f90: Fix template
We newly can handle some extra cases, for example:
struct a {int a,b,c;};
__attribute__ ((noinline))
int init (struct a *a)
{
a->a=1;
a->b=2;
a->c=3;
}
int const_fn ()
{
struct a a;
init (&a);
return a.a + a.b + a.c;
}
Here pure/const stops on the fact that const_fn calls non-const init, while
modref knows that the memory it initializes is local to const_fn.
I ended up reordering passes so early modref is done after early pure-const
mostly to avoid need to change testsuite which greps for const functions
being detects in pure-const. Stil some testuiste compensation is needed.
gcc/ChangeLog:
2021-11-11 Jan Hubicka <hubicka@ucw.cz>
* ipa-modref.c (analyze_function): Do pure/const discovery, return
true on success.
(pass_modref::execute): If pure/const is discovered fixup cfg.
(ignore_edge): Do not ignore pure/const edges.
(modref_propagate_in_scc): Do pure/const discovery, return true if
cdtor was promoted pure/const.
(pass_ipa_modref::execute): If needed remove unreachable functions.
* ipa-pure-const.c (warn_function_noreturn): Fix whitespace.
(warn_function_cold): Likewise.
(skip_function_for_local_pure_const): Move earlier.
(ipa_make_function_const): Break out from ...
(ipa_make_function_pure): Break out from ...
(propagate_pure_const): ... here.
(pass_local_pure_const::execute): Use it.
* ipa-utils.h (ipa_make_function_const): Declare.
(ipa_make_function_pure): Declare.
* passes.def: Move early modref after pure-const.
gcc/testsuite/ChangeLog:
2021-11-11 Jan Hubicka <hubicka@ucw.cz>
* c-c++-common/tm/inline-asm.c: Disable pure-const.
* g++.dg/ipa/modref-1.C: Update template.
* gcc.dg/tree-ssa/modref-11.c: Disable pure-const.
* gcc.dg/tree-ssa/modref-14.c: New test.
* gcc.dg/tree-ssa/modref-8.c: Do not optimize sibling calls.
* gfortran.dg/do_subscript_3.f90: Add -O0.
