The following uses the SLP node vectype rather than the vectype
stored in the DR group.
2021-09-17 Richard Biener <rguenther@suse.de>
* tree-vect-stmts.c (vectorizable_load): Use the vectype
from the SLP node.
gcc/fortran/ChangeLog:
* gfortran.h (gfc_omp_clauses): Add order_unconstrained.
* dump-parse-tree.c (show_omp_clauses): Dump it.
* openmp.c (gfc_match_omp_clauses): Match unconstrained/reproducible
modifiers to ordered(concurrent).
(OMP_DISTRIBUTE_CLAUSES): Accept ordered clause.
(resolve_omp_clauses): Reject ordered + order on same directive.
* trans-openmp.c (gfc_trans_omp_clauses, gfc_split_omp_clauses): Pass
on unconstrained modifier of ordered(concurrent).
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/order-5.f90: New test.
* gfortran.dg/gomp/order-6.f90: New test.
* gfortran.dg/gomp/order-7.f90: New test.
* gfortran.dg/gomp/order-8.f90: New test.
* gfortran.dg/gomp/order-9.f90: New test.
This removes adjusting alignment based on the vectorized accesses
and instead keeps what was set on the original access. The
code generating the actual accesses make sure to properly align
the vectorized accesses based on the generated pointer already
and the vectorizers alignment is always based of the desired
alignment of a vector type and thus will reset alignment to
unknown this way for example when doing strided accesses.
2021-09-20 Richard Biener <rguenther@suse.de>
* tree-vect-data-refs.c (vect_duplicate_ssa_name_ptr_info):
Do not compute alignment of the vectorized access here.
This properly marks the loop as for a runtime alias peel rather
than (pointlessly) going through DR_MISALIGNMENT.
2021-09-20 Richard Biener <rguenther@suse.de>
* tree-vect-data-refs.c (vect_enhance_data_refs_alignment):
Store -1 for runtime alias peeling iterations.
This obsoletes the 32bit hppa-hpux configurations which only support
STABS as debuginfo format.
2021-09-20 Richard Biener <rguenther@suse.de>
gcc/
* config.gcc: Obsolete hppa[12]*-*-hpux10* and hppa[12]*-*-hpux11*.
contrib/
* config-list.mk: --enable-obsolete for hppa2.0-hpux10.1 and
hppa2.0-hpux11.9.
When running the testsuite under Windows, we noticed failures in
testcase which attempt to match compiler error messages containing the
name of the executable.
For instance, gcc.dg/analyzer/signal-4a.c tries to match 'cc1:' which
obviously fails when the executable is called cc1.exe.
This patch removes the .exe suffix from various toolchain executables
to avoid this problem.
2021-09-08 Christophe Lyon <christophe.lyon@foss.st.com>
Torbjörn SVENSSON <torbjorn.svensson@st.com>
gcc/testsuite/
* lib/prune.exp (prune_gcc_output): Remove .exe suffix from
toolchain executables names.
'RESERVED_LOCATION_P' means 'UNKNOWN_LOCATION' or 'BUILTINS_LOCATION'.
We're using 'UNKNOWN_LOCATION' as a spare value for 'Empty', so should
ascertain that we don't use it as a key additionally. Similarly for
'BUILTINS_LOCATION' that we'd later like to use as a spare value for
'Deleted'.
As discussed in the source code comment added, for these we didn't have
stable behavior anyway.
Follow-up to r239175 (commit 88fa5555a3)
"On-demand locations within string-literals".
gcc/
* input.c (string_concat_db::record_string_concatenation)
(string_concat_db::get_string_concatenation): Skip for
'RESERVED_LOCATION_P'.
gcc/testsuite/
* gcc.dg/plugin/diagnostic-test-string-literals-1.c: Adjust
expected error diagnostics.
This adds the capability to analyze the dependence of mixed
pointer/array accesses. The example is from where using a masked
load/store creates the pointer-based access when an otherwise
unconditional access is array based. Other examples would include
accesses to an array mixed with accesses from inlined helpers
that work on pointers.
The idea is quite simple and old - analyze the data-ref indices
as if the reference was pointer-based. The following change does
this by changing dr_analyze_indices to work on the indices
sub-structure and storing an alternate indices substructure in
each data reference. That alternate set of indices is analyzed
lazily by initialize_data_dependence_relation when it fails to
match-up the main set of indices of two data references.
initialize_data_dependence_relation is refactored into a head
and a tail worker and changed to work on one of the indices
structures and thus away from using DR_* access macros which
continue to reference the main indices substructure.
There are quite some vectorization and loop distribution opportunities
unleashed in SPEC CPU 2017, notably 520.omnetpp_r, 548.exchange2_r,
510.parest_r, 511.povray_r, 521.wrf_r, 526.blender_r, 527.cam4_r and
544.nab_r see amendments in what they report with -fopt-info-loop while
the rest of the specrate set sees no changes there. Measuring runtime
for the set where changes were reported reveals nothing off-noise
besides 511.povray_r which seems to regress slightly for me
(on a Zen2 machine with -Ofast -march=native).
2021-09-08 Richard Biener <rguenther@suse.de>
PR tree-optimization/65206
* tree-data-ref.h (struct data_reference): Add alt_indices,
order it last.
* tree-data-ref.c (free_data_ref): Release alt_indices.
(dr_analyze_indices): Work on struct indices and get DR_REF as tree.
(create_data_ref): Adjust.
(initialize_data_dependence_relation): Split into head
and tail. When the base objects fail to match up try
again with pointer-based analysis of indices.
* tree-vectorizer.c (vec_info_shared::check_datarefs): Do
not compare the lazily computed alternate set of indices.
* gcc.dg/torture/20210916.c: New testcase.
* gcc.dg/vect/pr65206.c: Likewise.
The patch at r12-3662-g5fee8a0a9223d factored the code for
printing the names of programes into a separate function.
However the moved editions that print out the names of the
assembler, linker (and dsymutil on Darwin) when those are
specified at configure-time were not adjusted accordingly,
leading to a bootstrap fail.
Fixed by testing specifically for execute OK, since we know
these are programs.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* gcc.c: Test for execute OK when we find the
programs for assembler linker and dsymutil and those
were specified at configure-time.
Resolves:
PR middle-end/102403 - ICE in init_from_control_deps, at gimple-predicate-analysis.cc:2364
gcc/ChangeLog:
PR middle-end/102403
* gimple-predicate-analysis.cc (predicate::init_from_control_deps):
Correct a function pre/postcondition.
gcc/testsuite/ChangeLog:
PR middle-end/102403
* gcc.dg/uninit-pr102403.c: New test.
* gcc.dg/uninit-pr102403-c2.c: New test.
These were used on older systems to equate the FAT libgcc_s
library to single-slice equivalents. Unused for any current
system and never emitted by GCC.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libgcc/ChangeLog:
* config/t-slibgcc-darwin: Delete unused code.
This reorganises the Darwin symbol vers files to include
the generic ones at the top level; allowing for arch ports
to override (via either exclusion or inclusion as needed).
We add an X86-specific vers file containing the new HF
symbols. Note that although Darwin does not use ELF-style
symbol versioning - the parser that produces the map can
consume it. Using the ELF-style description will help us
know at which rev the symbols were introduced.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libgcc/ChangeLog:
* config/i386/t-darwin: Add in a vers file for X86-specific
symbols.
* config/t-darwin: Add the generic symbol maps here...
* config/t-slibgcc-darwin: ... removing from here.
* config/i386/libgcc-darwin.ver: New file.
We want to override the libgcc2 generic version of these functions
for X86. First exclude the original and the add in the replacements.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
libgcc/ChangeLog:
* config/i386/t-softfp: Exclude libgcc2 versions of __divhc3
and __mulhc3.
We have a small unwinder shim that is only used for Darwin10
(and only then in quite specific cases). To avoid linking
this code for every executable or DSO, we can present the crt
as a convenience library (rather than a .o file).
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/ChangeLog:
* config/darwin.h (LINK_COMMAND_SPEC_A): Use Darwin10
unwinder shim as a convenience library.
libgcc/ChangeLog:
* config.host: Use convenience library for Darwin10
unwinder shim.
* config/t-darwin: Build Darwin10 unwinder shim as a
convenience library.
gcc/testsuite
* gfortran.dg/goacc/privatization-1-compute.f90: Make test names
unique.
* gfortran.dg/goacc/routine-external-level-of-parallelism-2.f:
Likewise.
LTO usage requires binutils 2.35 or newer due to
https://sourceware.org/PR25355.
This adds a note in the prerequisites page about it.
Ok?
gcc/ChangeLog:
* doc/install.texi: Add note about
binutils 2.35 is required for LTO usage.
So the problem here is that now the lto-plugin requires NM that works
with LTO to work so we need to pass down NM just like we do for ranlib
and ar.
OK? Bootstrapped and tested with --with-build-config=bootstrap-lto on aarch64-linux-gnu.
Note you need to use binutils 2.35 or later too due to ttps://sourceware.org/PR25355
(I will submit another patch to improve the installation instructions too).
config/ChangeLog:
PR bootstrap/102389
* bootstrap-lto-lean.mk: Handle NM like RANLIB AND AR.
* bootstrap-lto.mk: Likewise.
Every time we allocate a threading edge we push it onto the path in a
distinct step. There's no need to do this in two steps, and avoiding
this, keeps us from exposing the internals of the registry.
I've also did some tiny cleanups in thread_across_edge, most importantly
removing the bitmap in favor of an auto_bitmap.
There are no functional changes.
gcc/ChangeLog:
* tree-ssa-threadbackward.c
(back_threader_registry::register_path): Use push_edge.
* tree-ssa-threadedge.c
(jump_threader::thread_around_empty_blocks): Same.
(jump_threader::thread_through_normal_block): Same.
(jump_threader::thread_across_edge): Same. Also, use auto_bitmap.
Tidy up code.
* tree-ssa-threadupdate.c
(jt_path_registry::allocate_thread_edge): Remove.
(jt_path_registry::push_edge): New.
(dump_jump_thread_path): Make static.
* tree-ssa-threadupdate.h (allocate_thread_edge): Remove.
(push_edge): New.
The current 'fixed_host_execute' implementation fails on Darwin
platforms for a number of reasons:
1/ If the sub-process spawn fails (e.g. because of missing or mal-
formed params); rather than reporting the fail output into the
match stream, as indicated by the expect manual, it terminates
the script.
- We fix this by (a) checking that the executable is valid as well
as existing (b) we put the spawn into a catch block and report
a failure.
2/ There is no recovery path at all for a buffer-full case (and we
do see buffer-full events with the default sizes).
- Added by the patch here, however it is not as sophisticated as
the methods used by dejagnu internally. Here we set the process
to be "nowait" and then close the connection - with the intent
that this will terminate the spawned process.
3/ The expect logic assumes that 'Totals:' is a valid indicator
for the end of the spawned process output. This is not true
even for the default dejagnu header (there are a number of
additional reporting lines after). In addition to this, there
are some tests that intentionally produce more output after
the totals report (and there are tests that do not use that
mechanism at all).
The effect is the we might arrive at the "wait" for the spawned
process to finish - but that process might not have completed
all its output. For Darwin, at least that causes a deadlock
between expect and the spawnee - the latter is doing a non-
cancellable write and the former is waiting for the latter to
terminate. For some reason this does not seem to affect Linux
perhaps the pty implementation allows the write(s) are able to
proceed even though there is no reader.
- This is fixed by modifying the loop termination condition to be
either EOF (which will be the 'correct' condition) or a timeout
which would represent an error either in the runtime or in the
parsing of the output. As added precautions, we only try to
wait if there is a correcly-spawned process, and we are also
specific about which process we are waiting for.
4/ Darwin appears to have a bug in either the tcl or termios
'cooking' code that ocassionally inserts an additional CR char
into the stream - thus '\n' => '\r\r\n' instead of '\r\n'. The
original program output is correct (it only contains a single
\n) - the additional character is being inserted somewhere in
the translations applied before the output reaches expect.
The logic of this expect implementation does not tolerate single
\r or \n characters (it will fail with a timeout or buffer-full
if that occurs).
- This is fixed by having a line-end match that is adjusted for
Darwin.
5/ The default buffer size does seem to be too small in some cases
noting that GCC uses 10000 as the match buffer size and the
default is 2000.
- Fixed by increasing the size to 8192.
6/ There is a somewhat arbitrary dumping of output where we match
^$prefix\tSOMETHING... and then process the something. This
essentially allows the match to start at any place in the buffer
following any collection of non-line-end chars.
- Fixed by amending the match for 'general' lines to accommodate
these cases, and reporting such lines to the log. At least this
should allow debugging of any cases where output that should be
recognized is being dropped.
Signed-off-by: Iain Sandoe <iain@sandoe.co.uk>
gcc/testsuite/ChangeLog:
* jit.dg/jit.exp (fixed_local_execute): Amend the match and
exit conditions to cater for more platforms.
There was an inline extern declaration for dump_ranger that was a bit of
a hack. I've removed it in favor of an actual prototype. There are
also some trivial changes to the dumping code in the path solver.
gcc/ChangeLog:
* gimple-range-path.cc (path_range_query::path_range_query): Add
header.
(path_range_query::dump): Remove extern declaration of dump_ranger.
* gimple-range-trace.cc (dump_ranger): Add DEBUG_FUNCTION marker.
* gimple-range-trace.h (dump_ranger): Add prototype.
gcc/
* gcc.c (find_a_program): New function, factored out of...
(find_a_file): Here.
(execute): Use find_a_program when looking for programs rather
than find_a_file.
This patch tackles PR middle-end/88173 where the order of operands in
a comparison affects constant folding. As diagnosed by Jason Merrill,
"match.pd handles these comparisons very differently". The history is
that the middle end, typically canonicalizes comparisons to place
constants on the right, but when a comparison contains two constants
we need to check/transform both constants, i.e. on both the left and the
right. Hence the added lines below duplicate for @0 the same transform
applied a few lines above for @1.
Whilst preparing the testcase, I noticed that this transformation is
incorrectly disabled with -fsignaling-nans even when both operands are
known not be be signaling NaNs, so I've corrected that and added a
second test case. Unfortunately, c-c++-common/pr57371-4.c then starts
failing, as it doesn't distinguish QNaNs (which are quiet) from SNaNs
(which signal), so this patch includes a minor tweak to the expected
behaviour for QNaNs in that existing test.
2021-09-19 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR middle-end/88173
* match.pd (cmp @0 REAL_CST@1): When @0 is also REAL_CST, apply
the same transformations as to @1. For comparisons against NaN,
don't check HONOR_SNANS but confirm that neither operand is a
signaling NaN.
gcc/testsuite/ChangeLog
PR middle-end/88173
* c-c++-common/pr57371-4.c: Tweak/correct test case for QNaNs.
* g++.dg/pr88173-1.C: New test case.
* g++.dg/pr88173-2.C: New test case.
So this is a simple fix is to just add to the assert that
sclass and dclass are both greater than or equal to NO_REGS.
NO_REGS is documented as the first register class so it should
have the value of 0.
gcc/ChangeLog:
* lra-constraints.c (check_and_process_move): Assert
that dclass and sclass are greater than or equal to NO_REGS.
This patch adds handling for unconstrained and reproducible modifiers on
order(concurrent) clause. For all static schedules (including auto and
no schedule or dist_schedule clauses) I believe what we implement is
reproducible, so the patch doesn't do much beyond recognizing those.
Note, there is an OpenMP/spec issue that needs resolution on what
should happen with the dynamic schedules (whether it should be an error
to mix such clauses, or silently make it non-reproducible, and in which
exact cases), so it might need some follow-up.
Besides that, this patch allows order(concurrent) clause on the distribute
construct which is something also added in OpenMP 5.1, and finally
check the newly added restriction that at most one order clause
can appear on a construct.
The allowing of order clause on distribute has a side-effect that
order(concurrent) copyin(thrpriv) is no longer allowed on combined/composite
constructs with distribute parallel for{, simd} in it, previously the
order applied only to for/simd and so a threadprivate var could be seen
in the construct, but now it also applies to distribute and so on the parallel
we shouldn't refer to a threadprivate var.
2021-09-18 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_ORDER_UNCONSTRAINED): Define.
* tree-pretty-print.c (dump_omp_clause): Print unconstrained:
for OMP_CLAUSE_ORDER_UNCONSTRAINED.
gcc/c-family/
* c-omp.c (c_omp_split_clauses): Split order clause also to
distribute construct. Copy over OMP_CLAUSE_ORDER_UNCONSTRAINED.
gcc/c/
* c-parser.c (c_parser_omp_clause_order): Parse unconstrained
and reproducible modifiers.
(OMP_DISTRIBUTE_CLAUSE_MASK): Add order clause.
gcc/cp/
* parser.c (cp_parser_omp_clause_order): Parse unconstrained
and reproducible modifiers.
(OMP_DISTRIBUTE_CLAUSE_MASK): Add order clause.
gcc/testsuite/
* c-c++-common/gomp/order-1.c (f2): Add tests for distribute
with order clause.
(f3): Remove.
* c-c++-common/gomp/order-2.c: Don't expect error for distribute
with order clause.
* c-c++-common/gomp/order-5.c: New test.
* c-c++-common/gomp/order-6.c: New test.
* c-c++-common/gomp/clause-dups-1.c (f1): Add tests for
duplicated order clause.
(f9): New function.
* c-c++-common/gomp/clauses-1.c (baz, bar): Don't mix copyin and
order(concurrent) clauses on the same composite construct combined
with distribute, instead split it into two tests, one without
copyin and one without order(concurrent). Add order(concurrent)
clauses to {,{,target} teams} distribute.
* g++.dg/gomp/attrs-1.C (baz, bar): Likewise.
* g++.dg/gomp/attrs-2.C (baz, bar): Likewise.
Besides conversion instructions, pass_rpad also handles scalar
sqrt/rsqrt/rcp/round instructions, while r12-3614 should only want to
handle conversion instructions, so fix it.
gcc/ChangeLog:
* config/i386/i386-features.c (remove_partial_avx_dependency):
Restrict TARGET_USE_VECTOR_FP_CONVERTS and
TARGET_USE_VECTOR_CONVERTS to conversion instructions only.
OpenMP 5.1 allows default(private) or default(firstprivate) even in C/C++,
but it behaves the same way as in Fortran only for variables not declared at
namespace or file scope. For the namespace/file scope variables it instead
behaves as default(none).
2021-09-18 Jakub Jelinek <jakub@redhat.com>
gcc/
* gimplify.c (omp_default_clause): For C/C++ default({,first}private),
if file/namespace scope variable doesn't have predetermined sharing,
treat it as if there was default(none).
gcc/c/
* c-parser.c (c_parser_omp_clause_default): Handle private and
firstprivate arguments, adjust diagnostics on unknown argument.
gcc/cp/
* parser.c (cp_parser_omp_clause_default): Handle private and
firstprivate arguments, adjust diagnostics on unknown argument.
* cp-gimplify.c (cxx_omp_finish_clause): Handle OMP_CLAUSE_PRIVATE.
gcc/testsuite/
* c-c++-common/gomp/default-2.c: New test.
* c-c++-common/gomp/default-3.c: New test.
* g++.dg/gomp/default-1.C: New test.
libgomp/
* testsuite/libgomp.c++/default-1.C: New test.
* testsuite/libgomp.c-c++-common/default-1.c: New test.
* libgomp.texi (OpenMP 5.1): Mark "private and firstprivate argument
to default clause in C and C++" as implemented.
I've been working on the resolution of CWG1835 by P1787, which among many
other things clarified that a name after -> or . is looked up first in the
class of the object expression even if it's dependent. This patch does not
make that change; this is a smaller change extracted from that work in
progress to make the lookup in the object type work better in cases where
unqualified lookup doesn't find anything.
Basically, if we see "t.foo::" we know that looking up foo in t needs to
find a type, so we build an implicit TYPENAME_TYPE for it.
This also implements the change from P1787 to assume that a name followed by
< in a type-only context names a template, since the less-than operator
can't appear in a type context. This makes some of the lines in dtor11.C
work.
I introduce the predicate 'dependentish_scope_p' for the case where the
current instantiation has dependent bases, so even though we can perform
name lookup, we can't conclude that a lookup failure is conclusive.
gcc/cp/ChangeLog:
* cp-tree.h (dependentish_scope_p): Declare.
* pt.c (dependentish_scope_p): New.
* parser.c (cp_parser_lookup_name): Return a TYPENAME_TYPE
for lookup of a type in a dependent object.
(cp_parser_template_id): Handle TYPENAME_TYPE.
(cp_parser_template_name): If we're looking for a type,
a name followed by < names a template.
gcc/testsuite/ChangeLog:
* g++.dg/template/dtor5.C: Adjust expected error.
* g++.dg/cpp23/lookup2.C: New test.
* g++.dg/template/dtor11.C: New test.
gcc/fortran/ChangeLog:
PR fortran/102366
* trans-decl.c (gfc_finish_var_decl): Disable the warning message
for variables moved from stack to static storange if they are
declared in the main, but allow the move to happen.
gcc/testsuite/ChangeLog:
PR fortran/102366
* gfortran.dg/pr102366.f90: New test.
All path::iterator operations are non-throwing.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/fs_path.h (path::iterator): Add noexcept to all
member functions and friend functions.
(distance): Add noexcept.
(advance): Add noexcept and inline.
* include/experimental/bits/fs_path.h (path::iterator):
Add noexcept to all member functions.
Also rename the test so it actually runs.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
PR libstdc++/102270
* include/std/tuple (_Tuple_impl): Add constexpr to constructor
missed in previous patch.
* testsuite/20_util/tuple/cons/102270.C: Moved to...
* testsuite/20_util/tuple/cons/102270.cc: ...here.
* testsuite/util/testsuite_allocator.h (SimpleAllocator): Add
constexpr to constructor so it can be used for C++20 tests.
This is an optimisation for middle-end worker-partitioning support (used
to support multiple workers on AMD GCN). At present, barriers may be
emitted in cases where they aren't needed and cannot be optimised away.
This patch stops the extraneous barriers from being emitted in the
first place.
One exception to the above (where the barrier is still needed) is for
predicated blocks of code that perform a write to gang-private shared
memory from one worker. We must execute a barrier before other workers
read that shared memory location.
gcc/
* config/gcn/gcn.c (gimple.h): Include.
(gcn_fork_join): Emit barrier for worker-level joins.
* omp-oacc-neuter-broadcast.cc (find_local_vars_to_propagate): Add
writes_gang_private bitmap parameter. Set bit for blocks
containing gang-private variable writes.
(worker_single_simple): Don't emit barrier after predicated block.
(worker_single_copy): Don't emit barrier if we're not broadcasting
anything and the block contains no gang-private writes.
(neuter_worker_single): Don't predicate blocks that only contain
NOPs or internal marker functions. Pass has_gang_private_write
argument to worker_single_copy.
(oacc_do_neutering): Add writes_gang_private bitmap handling.
This patch implements an algorithm to lay out local data-share (LDS)
space. It currently works for AMD GCN. At the moment, LDS is used for
three things:
1. Gang-private variables
2. Reduction temporaries (accumulators)
3. Broadcasting for worker partitioning
After the patch is applied, (2) and (3) are placed at preallocated
locations in LDS, and (1) continues to be handled by the backend (as it
is at present prior to this patch being applied). LDS now looks like this:
+--------------+ (gang-private size + 1024, = 1536)
| free space |
| ... |
| - - - - - - -|
| worker bcast |
+--------------+
| reductions |
+--------------+ <<< -mgang-private-size=<number> (def. 512)
| gang-private |
| vars |
+--------------+ (32)
| low LDS vars |
+--------------+ LDS base
So, gang-private space is fixed at a constant amount at compile time
(which can be increased with a command-line switch if necessary
for some given code). The layout algorithm takes out a slice of the
remainder of usable space for reduction vars, and uses the rest for
worker partitioning.
The partitioning algorithm works as follows.
1. An "adjacency" set is built up for each basic block that might
do a broadcast. This is calculated by starting at each such block,
and doing a recursive DFS walk over successors to find the next
block (or blocks) that *also* does a broadcast
(dfs_broadcast_reachable_1).
2. The adjacency set is inverted to get adjacent predecessor blocks also.
3. Blocks that will perform a broadcast are sorted by size of that
broadcast: the biggest blocks are handled first.
4. A splay tree structure is used to calculate the spans of LDS memory
that are already allocated by the blocks adjacent to this one
(merge_ranges{,_1}.
5. The current block's broadcast space is allocated from the first free
span not allocated in the splay tree structure calculated above
(first_fit_range). This seems to work quite nicely and efficiently
with the splay tree structure.
6. Continue with the next-biggest broadcast block until we're done.
In this way, "adjacent" broadcasts will not use the same piece of
LDS memory.
PR96334 "openacc: Unshare reduction temporaries for GCN" got merged in:
The GCN backend uses tree nodes like MEM((__lds TYPE *) <constant>)
for reduction temporaries. Unlike e.g. var decls and SSA names, these
nodes cannot be shared during gimplification, but are so in some
circumstances. This is detected when appropriate --enable-checking
options are used. This patch unshares such nodes when they are reused
more than once.
gcc/
* config/gcn/gcn-protos.h
(gcn_goacc_create_worker_broadcast_record): Update prototype.
* config/gcn/gcn-tree.c (gcn_goacc_get_worker_red_decl): Use
preallocated block of LDS memory. Do not cache/share decls for
reduction temporaries between invocations.
(gcn_goacc_reduction_teardown): Unshare VAR on second use.
(gcn_goacc_create_worker_broadcast_record): Add OFFSET parameter
and return temporary LDS space at that offset. Return pointer in
"sender" case.
* config/gcn/gcn.c (acc_lds_size, gang_private_hwm, lds_allocs):
New global vars.
(ACC_LDS_SIZE): Define as acc_lds_size.
(gcn_init_machine_status): Don't initialise lds_allocated,
lds_allocs, reduc_decls fields of machine function struct.
(gcn_option_override): Handle default size for gang-private
variables and -mgang-private-size option.
(gcn_expand_prologue): Use LDS_SIZE instead of LDS_SIZE-1 when
initialising M0_REG.
(gcn_shared_mem_layout): New function.
(gcn_print_lds_decl): Update comment. Use global lds_allocs map and
gang_private_hwm variable.
(TARGET_GOACC_SHARED_MEM_LAYOUT): Define target hook.
* config/gcn/gcn.h (machine_function): Remove lds_allocated,
lds_allocs, reduc_decls. Add reduction_base, reduction_limit.
* config/gcn/gcn.opt (gang_private_size_opt): New global.
(mgang-private-size=): New option.
* doc/tm.texi.in (TARGET_GOACC_SHARED_MEM_LAYOUT): Place
documentation hook.
* doc/tm.texi: Regenerate.
* omp-oacc-neuter-broadcast.cc (targhooks.h, diagnostic-core.h):
Add includes.
(build_sender_ref): Handle sender_decl being pointer.
(worker_single_copy): Add PLACEMENT and ISOLATE_BROADCASTS
parameters. Pass placement argument to
create_worker_broadcast_record hook invocations. Handle
sender_decl being pointer and isolate_broadcasts inserting extra
barriers.
(blk_offset_map_t): Add typedef.
(neuter_worker_single): Add BLK_OFFSET_MAP parameter. Pass
preallocated range to worker_single_copy call.
(dfs_broadcast_reachable_1): New function.
(idx_decl_pair_t, used_range_vec_t): New typedefs.
(sort_size_descending): New function.
(addr_range): New class.
(splay_tree_compare_addr_range, splay_tree_free_key)
(first_fit_range, merge_ranges_1, merge_ranges): New functions.
(execute_omp_oacc_neuter_broadcast): Rename to...
(oacc_do_neutering): ... this. Add BOUNDS_LO, BOUNDS_HI
parameters. Arrange layout of shared memory for broadcast
operations.
(execute_omp_oacc_neuter_broadcast): New function.
(pass_omp_oacc_neuter_broadcast::gate): Remove num_workers==1
handling from here. Enable pass for all OpenACC routines in order
to call shared memory-layout hook.
* target.def (create_worker_broadcast_record): Add OFFSET
parameter.
(shared_mem_layout): New hook.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/broadcast-many.c: Update.
