This teaches forwprop to rewrite more vector loads that are only
used in BIT_FIELD_REFs as scalar loads. This provides the
remaining uplift to SPEC CPU 2017 510.parest_r on Zen 2 which
has CPU gathers disabled.
In particular vector load + vec_unpack + bit-field-ref is turned
into (extending) scalar loads which avoids costly XMM/GPR
transitions. To not conflict with vector load + bit-field-ref
+ vector constructor matching to vector load + shuffle the
extended transform is only done after vector lowering.
2021-07-30 Richard Biener <rguenther@suse.de>
* tree-ssa-forwprop.c (pass_forwprop::execute): Split
out code to decompose vector loads ...
(optimize_vector_load): ... here. Generalize it to
handle intermediate widening and TARGET_MEM_REF loads
and apply it to loads with a supported vector mode as well.
The following avoids vectorizing MIN/MAX reductions on bools which,
when ending up as vector(2) <signed-boolean:64> would need to be
adjusted because of the sign change. The fix instead avoids any
reduction vectorization where the result isn't compatible
to the original scalar type since we don't compensate for that
either.
2021-08-04 Richard Biener <rguenther@suse.de>
PR tree-optimization/101756
* tree-vect-slp.c (vectorizable_bb_reduc_epilogue): Make sure
the result of the reduction epilogue is compatible to the original
scalar result.
* gcc.dg/vect/bb-slp-pr101756.c: New testcase.
When parsing default arguments, we need to temporarily clear parser->omp_declare_simd
and parser->oacc_routine, otherwise it can clash with further declarations
inside of e.g. lambdas inside of those default arguments.
2021-08-04 Jakub Jelinek <jakub@redhat.com>
PR c++/101759
* parser.c (cp_parser_default_argument): Temporarily override
parser->omp_declare_simd and parser->oacc_routine to NULL.
* g++.dg/gomp/pr101759.C: New test.
* g++.dg/goacc/pr101759.C: New test.
The file has two identical halves, seems like twice applied patch.
2021-08-04 Jakub Jelinek <jakub@redhat.com>
* gcc.c-torture/execute/ieee/pr29302-1.x: Undo doubly applied patch.
The define_peephole2 which is added by r12-2640-gf7bf03cf69ccb7dc
should only work on general registers, considering that x86 also
supports mov instructions between gpr, sse reg, mask reg, limiting the
peephole2 predicate to general_reg_operand.
gcc/ChangeLog:
PR target/101743
* config/i386/i386.md (peephole2): Refine predicate from
register_operand to general_reg_operand.
The file has two identical halves, seems like twice applied patch.
2021-08-04 Jakub Jelinek <jakub@redhat.com>
* config/t-slibgcc-fuchsia: Undo doubly applied patch.
This makes tail recursion optimization produce a loop structure
manually rather than relying on loop fixup. That also allows the
loop to be marked as finite (it would eventually blow the stack
if it were not).
2021-08-04 Richard Biener <rguenther@suse.de>
PR tree-optimization/101769
* tree-tailcall.c (eliminate_tail_call): Add the created loop
for the first recursion and return it via the new output parameter.
(optimize_tail_call): Pass through new output param.
(tree_optimize_tail_calls_1): After creating all latches,
add the created loop to the loop tree. Do not mark loops for fixup.
* g++.dg/tree-ssa/pr101769.C: New testcase.
Previous CLs add new language constructs in Go 1.17, specifically,
unsafe.Add, unsafe.Slice, and conversion from a slice to a pointer
to an array. This CL handles them in the escape analysis.
At the point of the escape analysis, unsafe.Add and unsafe.Slice
are still builtin calls, so just handle them in data flow.
Conversion from a slice to a pointer to an array has already been
lowered to a combination of compound expression, conditional
expression and slice info expressions, so handle them in the
escape analysis.
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/339671
The only different between selectnbrecv and selectnbrecv2 is the later
set the input pointer value by second return value from chanrecv.
So by making selectnbrecv return two values from chanrecv, we can get
rid of selectnbrecv2, the compiler can now call only selectnbrecv and
generate simpler code.
This is the gofrontend version of https://golang.org/cl/292890.
Reviewed-on: https://go-review.googlesource.com/c/gofrontend/+/339529
It is the prefix of the "es" and "eI" constraints.
2021-08-03 Segher Boessenkool <segher@kernel.crashing.org>
* config/rs6000/constraints.md: Remove "e" from the list of available
constraint characters.
The histogram value for indirect calls was incorrectly set up.
That is fixed now.
With this change the tree-prof tests checking indirect call inlining with AutoFDO
in gcc.dg and g++.dg are passing.
Resolves:
PR gcov-profile/71672 - inlining indirect calls does not work with autofdo
gcc/ChangeLog:
PR gcov-profile/71672
* auto-profile.c (afdo_indirect_call): Fix setup of the historgram value for indirect calls.
* create_gcov tool doesn't currently support dwarf 5 so I made a change in profopt.exp
to pass -gdwarf-4 when compiling the binary to profile.
* I updated the invocation of create_gcov in profopt.exp to pass -gcov_version=2.
I recently made a change to create_gcov to support version 2:
https://github.com/google/autofdo/pull/117 .
* I removed useless -o perf.data from the invocation of gcc-auto-profile in
target-supports.exp.
These changes contribute to fixing PR gcov-profile/71672.
gcc/testsuite/ChangeLog:
* lib/profopt.exp: Pass gdwarf-4 when compiling test to profile; pass -gcov_version=2.
* lib/target-supports.exp: Remove unnecessary -o perf.data passed to gcc-auto-profile.
indir-call-prof-2.c has -fno-early-inlining but AutoFDO can't work without
early inlining (it needs to match the inlining of the profiled binary).
I changed profopt.exp to always pass -fearly-inlining for AutoFDO.
With that change the indirect call inlining in indir-call-prof-2.c happens in the early inliner
so I changed the dg-final-use-autofdo.
Contributes to fixing PR gcov-profile/71672
gcc/testsuite/ChangeLog:
* gcc.dg/tree-prof/indir-call-prof-2.c: Fix dg-final-use-autofdo.
* lib/profopt.exp: Pass -fearly-inlining when compiling with AutoFDO.
* Changed several tests to use -fdump-ipa-afdo-optimized instead of -fdump-ipa-afdo
in dg-options so that the expected output can be found
* Increased the number of iterations in several tests so that perf can have
enough sampling events
Contributes to fixing PR gcov-profile/71672.
gcc/testsuite/ChangeLog:
* g++.dg/tree-prof/indir-call-prof.C: Fix options, increase the number of iterations.
* g++.dg/tree-prof/morefunc.C: Fix options, increase the number of iterations.
* g++.dg/tree-prof/reorder.C: Fix options, increase the number of iterations.
* gcc.dg/tree-prof/indir-call-prof-2.c: Fix options, increase the number of iterations.
* gcc.dg/tree-prof/indir-call-prof.c: Fix options.
Resolves:
PR testsuite/101688 - g++.dg/warn/Wstringop-overflow-4.C fails on 32-bit archs with new jump threader
gcc/testsuite:
PR testsuite/101688
* g++.dg/warn/Wstringop-overflow-4.C: Disable a test case in ILP32.
Copy the test for _mm_minpos_epu16 from
gcc/testsuite/gcc.target/i386/sse4_1-phminposuw.c, with
a few adjustments:
- Adjust the dejagnu directives for powerpc platform.
- Make the data not be monotonically increasing,
such that some of the returned values are not
always the first value (index 0).
- Create a list of input data testing various scenarios
including more than one minimum value and different
orders and indices of the minimum value.
- Fix a masking issue where the index was being truncated
to 2 bits instead of 3 bits, which wasn't found because
all of the returned indices were 0 with the original
generated data.
- Support big-endian.
2021-08-03 Paul A. Clarke <pc@us.ibm.com>
gcc/testsuite
* gcc.target/powerpc/sse4_1-phminposuw.c: Copy from
gcc/testsuite/gcc.target/i386, adjust dg directives to suit,
make more robust.
Add a naive implementation of the subject x86 intrinsic to
ease porting.
2021-08-03 Paul A. Clarke <pc@us.ibm.com>
gcc
* config/rs6000/smmintrin.h (_mm_minpos_epu16): New.
In C++17 the out-of-class definitions for static constexpr variables are
redundant, because they are implicitly inline. This change avoids
"redundant redeclaration" warnings from -Wsystem-headers -Wdeprecated.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/random.tcc (linear_congruential_engine): Do not
define static constexpr members when they are implicitly inline.
* include/std/ratio (ratio, __ratio_multiply, __ratio_divide)
(__ratio_add, __ratio_subtract): Likewise.
* include/std/type_traits (integral_constant): Likewise.
* testsuite/26_numerics/random/pr60037-neg.cc: Adjust dg-error
line number.
This adds a partial specialization of allocator_traits, similar to what
was already done for std::allocator. This means that most uses of
polymorphic_allocator via the traits can avoid the metaprogramming
overhead needed to deduce the properties from polymorphic_allocator.
In addition, I'm changing polymorphic_allocator::delete_object to invoke
the destructor (or pseudo-destructor) directly, rather than calling
allocator_traits::destroy, which calls polymorphic_allocator::destroy
(which is deprecated). This is observable if a user has specialized
allocator_traits<polymorphic_allocator<Foo>> and expects to see its
destroy member function called. I consider explicit specializations of
allocator_traits to be wrong-headed, and this use case seems unnecessary
to support. So delete_object just invokes the destructor directly.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/std/memory_resource (polymorphic_allocator::delete_object):
Call destructor directly instead of using destroy.
(allocator_traits<polymorphic_allocator<T>>): Define partial
specialization.
The std::random_shuffle algorithm was removed in C++14 (without
deprecation). This adds the deprecated attribute for C++14 and later, so
that users are warned they should not be using it in those dialects.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* doc/xml/manual/evolution.xml: Document deprecation.
* doc/html/*: Regenerate.
* include/bits/c++config (_GLIBCXX14_DEPRECATED): Define.
(_GLIBCXX14_DEPRECATED_SUGGEST): Define.
* include/bits/stl_algo.h (random_shuffle): Deprecate for C++14
and later.
* testsuite/25_algorithms/headers/algorithm/synopsis.cc: Adjust
for C++11 and C++14 changes to std::random_shuffle and
std::shuffle.
* testsuite/25_algorithms/random_shuffle/1.cc: Add options to
use deprecated algorithms.
* testsuite/25_algorithms/random_shuffle/59603.cc: Likewise.
* testsuite/25_algorithms/random_shuffle/moveable.cc: Likewise.
* testsuite/25_algorithms/random_shuffle/requirements/explicit_instantiation/2.cc:
Likewise.
* testsuite/25_algorithms/random_shuffle/requirements/explicit_instantiation/pod.cc:
Likewise.
This change adds options to tests that explicitly use deprecated
features, so that -D_GLIBCXX_USE_DEPRECATED=0 can be used to run the
rest of the testsuite. The tests that explicitly/intentionally use
deprecated features will still be able to use them, but they can be
disabled for the majority of tests.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* testsuite/23_containers/forward_list/operations/3.cc:
Use lambda instead of std::bind2nd.
* testsuite/20_util/function_objects/binders/3113.cc: Add
options for testing deprecated features.
* testsuite/20_util/pair/cons/99957.cc: Likewise.
* testsuite/20_util/shared_ptr/assign/auto_ptr.cc: Likewise.
* testsuite/20_util/shared_ptr/assign/auto_ptr_neg.cc: Likewise.
* testsuite/20_util/shared_ptr/assign/auto_ptr_rvalue.cc:
Likewise.
* testsuite/20_util/shared_ptr/cons/43820_neg.cc: Likewise.
* testsuite/20_util/shared_ptr/cons/auto_ptr.cc: Likewise.
* testsuite/20_util/shared_ptr/cons/auto_ptr_neg.cc: Likewise.
* testsuite/20_util/shared_ptr/creation/dr925.cc: Likewise.
* testsuite/20_util/unique_ptr/cons/auto_ptr.cc: Likewise.
* testsuite/20_util/unique_ptr/cons/auto_ptr_neg.cc: Likewise.
* testsuite/ext/pb_ds/example/priority_queue_erase_if.cc:
Likewise.
* testsuite/ext/pb_ds/example/priority_queue_split_join.cc:
Likewise.
* testsuite/lib/dg-options.exp (dg_add_options_using-deprecated):
New proc.
This reduces the size of <regex> a little. This is one of the largest
and slowest headers in the library.
By using <bits/stl_algobase.h> and <bits/stl_algo.h> instead of
<algorithm> we don't need to parse all the parallel algorithms and
std::ranges:: algorithms that are not needed by <regex>. Similarly, by
using <bits/stl_tree.h> and <bits/stl_map.h> instead of <map> we don't
need to parse the definition of std::multimap.
The _State_info type is not movable or copyable, so doesn't need to use
std::unique_ptr<bool[]> to manage a bitset, we can just delete it in the
destructor. It would use a lot less space if we used a bitset instead,
but that would be an ABI break. We could do it for the versioned
namespace, but this patch doesn't do so. For future reference, using
vector<bool> would work, but would increase sizeof(_State_info) by two
pointers, because it's three times as large as unique_ptr<bool[]>. We
can't use std::bitset because the length isn't constant. We want a
bitset with a non-constant but fixed length.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/regex_executor.h (_State_info): Replace
unique_ptr<bool[]> with array of bool.
* include/bits/regex_executor.tcc: Likewise.
* include/bits/regex_scanner.tcc: Replace std::strchr with
__builtin_strchr.
* include/std/regex: Replace standard headers with smaller
internal ones.
* testsuite/28_regex/traits/char/lookup_classname.cc: Include
<string.h> for strlen.
* testsuite/28_regex/traits/char/lookup_collatename.cc:
Likewise.
std::wstring_convert and std::wbuffer_convert types are not copyable or
movable, and store a plain pointer without a deleter. That means a much
simpler type that just uses delete in its destructor can be used instead
of std::unique_ptr.
That avoids including and parsing all of <bits/unique_ptr.h> in every
header that includes <locale>. It also avoids instantiating
unique_ptr<C> and std::tuple<C*, default_delete<C>> when the conversion
utilities are used.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/locale_conv.h (__detail::_Scoped_ptr): Define new
RAII class template.
(wstring_convert, wbuffer_convert): Use __detail::_Scoped_ptr
instead of unique_ptr.
This patch adds an option to tune for Neoverse cores that have
a total vector bandwidth of 512 bits (4x128 for Advanced SIMD
and a vector-length-dependent equivalent for SVE). This is intended
to be a compromise between tuning aggressively for a single core like
Neoverse V1 (which can be too narrow) and tuning for AArch64 cores
in general (which can be too wide).
-mcpu=neoverse-512tvb is equivalent to -mcpu=neoverse-v1
-mtune=neoverse-512tvb.
gcc/
* doc/invoke.texi: Document -mtune=neoverse-512tvb and
-mcpu=neoverse-512tvb.
* config/aarch64/aarch64-cores.def (neoverse-512tvb): New entry.
* config/aarch64/aarch64-tune.md: Regenerate.
* config/aarch64/aarch64.c (neoverse512tvb_sve_vector_cost)
(neoverse512tvb_sve_issue_info, neoverse512tvb_vec_issue_info)
(neoverse512tvb_vector_cost, neoverse512tvb_tunings): New structures.
(aarch64_adjust_body_cost_sve): Handle -mtune=neoverse-512tvb.
(aarch64_adjust_body_cost): Likewise.
The AArch64 vector costs try to take issue rates into account.
However, when vectorising an outer loop, we lumped the inner
and outer operations together, which is somewhat meaningless.
This patch restricts the heuristic to the inner loop.
gcc/
* config/aarch64/aarch64.c (aarch64_add_stmt_cost): Only
record issue information for operations that occur in the
innermost loop.
The issue-based vector costs currently assume that a multiply-add
sequence can be implemented using a single instruction. This is
generally true for scalars (which have a 4-operand instruction)
and SVE (which allows the output to be tied to any input).
However, for Advanced SIMD, multiplying two values and adding
an invariant will end up being a move and an MLA.
The only target to use the issue-based vector costs is Neoverse V1,
which would generally prefer SVE in this case anyway. I therefore
don't have a self-contained testcase. However, the distinction
becomes more important with a later patch.
gcc/
* config/aarch64/aarch64.c (aarch64_multiply_add_p): Add a vec_flags
parameter. Detect cases in which an Advanced SIMD MLA would almost
certainly require a MOV.
(aarch64_count_ops): Update accordingly.
When the vectoriser scalarises a strided store, it counts one
scalar_store for each element plus one vec_to_scalar extraction
for each element. However, extracting element 0 is free on AArch64,
so it should have zero cost.
I don't have a testcase that requires this for existing -mtune
options, but it becomes more important with a later patch.
gcc/
* config/aarch64/aarch64.c (aarch64_is_store_elt_extraction): New
function, split out from...
(aarch64_detect_vector_stmt_subtype): ...here.
(aarch64_add_stmt_cost): Treat extracting element 0 as free.
This patch adds tuning fields for the total cost of a gather load
instruction. Until now, we've costed them as one scalar load
per element instead. Those scalar_load-based values are also
what the patch uses to fill in the new fields for existing
cost structures.
gcc/
* config/aarch64/aarch64-protos.h (sve_vec_cost):
Add gather_load_x32_cost and gather_load_x64_cost.
* config/aarch64/aarch64.c (generic_sve_vector_cost)
(a64fx_sve_vector_cost, neoversev1_sve_vector_cost): Update
accordingly, using the values given by the scalar_load * number
of elements calculation that we used previously.
(aarch64_detect_vector_stmt_subtype): Use the new fields.
This patch splits the SVE-specific part of aarch64_adjust_body_cost
out into its own subroutine, so that a future patch can call it
more than once. I wondered about using a lambda to avoid having
to pass all the arguments, but in the end this way seemed clearer.
gcc/
* config/aarch64/aarch64.c (aarch64_adjust_body_cost_sve): New
function, split out from...
(aarch64_adjust_body_cost): ...here.
This patch adds a simple fixed-point class for holding fractional
cost values. It can exactly represent the reciprocal of any
single-vector SVE element count (including the non-power-of-2 ones).
This means that it can also hold 1/N for all N in [1, 16], which should
be enough for the various *_per_cycle fields.
For now the assumption is that the number of possible reciprocals
is fixed at compile time and so the class should always be able
to hold an exact value.
The class uses a uint64_t to hold the fixed-point value, which means
that it can hold any scaled uint32_t cost. Normally we don't worry
about overflow when manipulating raw uint32_t costs, but just to be
on the safe side, the class uses saturating arithmetic for all
operations.
As far as the changes to the cost routines themselves go:
- The changes to aarch64_add_stmt_cost and its subroutines are
just laying groundwork for future patches; no functional change
intended.
- The changes to aarch64_adjust_body_cost mean that we now
take fractional differences into account.
gcc/
* config/aarch64/fractional-cost.h: New file.
* config/aarch64/aarch64.c: Include <algorithm> (indirectly)
and cost_fraction.h.
(vec_cost_fraction): New typedef.
(aarch64_detect_scalar_stmt_subtype): Use it for statement costs.
(aarch64_detect_vector_stmt_subtype): Likewise.
(aarch64_sve_adjust_stmt_cost, aarch64_adjust_stmt_cost): Likewise.
(aarch64_estimate_min_cycles_per_iter): Use vec_cost_fraction
for cycle counts.
(aarch64_adjust_body_cost): Likewise.
(aarch64_test_cost_fraction): New function.
(aarch64_run_selftests): Call it.
The tuning structures have an sve_width field that specifies the
number of bits in an SVE vector (or SVE_NOT_IMPLEMENTED if not
applicable). This patch turns the field into a bitmask so that
it can specify multiple widths at the same time. For now we
always treat the mininum width as the likely width.
An alternative would have been to add extra fields, which would
have coped correctly with non-power-of-2 widths. However,
we're very far from supporting constant non-power-of-2 vectors
in GCC, so I think the non-power-of-2 case will in reality always
have to be hidden behind VLA.
gcc/
* config/aarch64/aarch64-protos.h (tune_params::sve_width): Turn
into a bitmask.
* config/aarch64/aarch64.c (aarch64_cmp_autovec_modes): Update
accordingly.
(aarch64_estimated_poly_value): Likewise. Use the least significant
set bit for the minimum and likely values. Use the most significant
set bit for the maximum value.
gcc/ChangeLog:
* config/i386/sse.md (cond_<insn><mode>): New expander.
(cond_mul<mode>): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/i386/cond_op_addsubmul_d-1.c: New test.
* gcc.target/i386/cond_op_addsubmul_d-2.c: New test.
* gcc.target/i386/cond_op_addsubmul_q-1.c: New test.
* gcc.target/i386/cond_op_addsubmul_q-2.c: New test.
* gcc.target/i386/cond_op_addsubmul_w-1.c: New test.
* gcc.target/i386/cond_op_addsubmul_w-2.c: New test.
Appending to a string variable with `+=' is a bashism and does not work in
strict POSIX shells like dash. This results in the extra compilation flags not
to be set correctly. This patch replaces the `+=' syntax with a simple string
interpolation to append to the `EXTRA_CXXFLAGS' variable.
libsanitizer/ChangeLog
PR sanitizer/101111
* configure.tgt: Fix bashism in setting of `EXTRA_CXXFLAGS'.
The following testcase ICEs because DECL_FUNCTION_CODE asserts the builtin
is BUILT_IN_NORMAL, but it sees a backend (MD) builtin instead.
The FE, normal and MD builtin numbers overlap, so one should always
check what kind of builtin it is before looking at specific codes.
On the other side, region-model.cc has:
if (fndecl_built_in_p (callee_fndecl, BUILT_IN_NORMAL)
&& gimple_builtin_call_types_compatible_p (call, callee_fndecl))
switch (DECL_UNCHECKED_FUNCTION_CODE (callee_fndecl))
which IMO should use DECL_FUNCTION_CODE instead, it checked first it is
a normal builtin...
2021-08-03 Jakub Jelinek <jakub@redhat.com>
PR analyzer/101721
* sm-malloc.cc (known_allocator_p): Only check DECL_FUNCTION_CODE on
BUILT_IN_NORMAL builtins.
* gcc.dg/analyzer/pr101721.c: New test.
As mentioned in [1], there is one pre-existing issue before
the refactoring of FOR_EACH_LOOP_FN. The macro will always
set the given LOOP as NULL at the end of iterating unless
there is some early break inside, obviously there is no
early break and dloop will be set as NULL after the loop
iterating. It's kept as NULL after the factoring.
I tried to debug the test case gcc.dg/graphite/pr83359.c
with commit 555758de90 (also reproduced the ICE with
555758de90074~), and noticed the compilation of the test
case only covers the hunk:
else
{
moved_orig_loop_num[dloop->orig_loop_num] = -1;
dloop->orig_loop_num = 0;
}
it doesn't touch the if condition hunk to increase
"moved_orig_loop_num[dloop->orig_loop_num]". So the
following hunk guarded with
if (moved_orig_loop_num[orig_loop_num] == 2)
using dloop for dereference doesn't get executed. It
explains why the problem doesn't get exposed before.
By looking to the code using dloop, I think it's a copy
paste typo, the modified assertion codes have the same
words as the above condition check. In that context, the
expected original number has been assigned to variable
orig_loop_num by extracting from the arg0 of the call
IFN_LOOP_DIST_ALIAS.
[1] https://gcc.gnu.org/pipermail/gcc-patches/2021-July/576367.html
gcc/ChangeLog:
* tree-cfg.c (move_sese_region_to_fn): Fix typos on dloop.
