This adds a simple reduction vectorization capability to the
non-loop vectorizer. Simple meaning it lacks any of the fancy
ways to generate the reduction epilogue but only supports
those we can handle via a direct internal function reducing
a vector to a scalar. One of the main reasons is to avoid
massive refactoring at this point but also that more complex
epilogue operations are hardly profitable.
Mixed sign reductions are for now fend off and I'm not finally
settled with whether we want an explicit SLP node for the
reduction epilogue operation. Handling mixed signs could be
done by multiplying with a { 1, -1, .. } vector. Fend off
are also reductions with non-internal operands (constants
or register parameters for example).
Costing is done by accounting the original scalar participating
stmts for the scalar cost and log2 permutes and operations for
the vectorized epilogue.
--
SPEC CPU 2017 FP with rate workload measurements show (picked
fastest runs of three) regressions for 507.cactuBSSN_r (1.5%),
508.namd_r (2.5%), 511.povray_r (2.5%), 526.blender_r (0.5) and
527.cam4_r (2.5%) and improvements for 510.parest_r (5%) and
538.imagick_r (1.5%). This is with -Ofast -march=znver2 on a Zen2.
Statistics on CPU 2017 shows that the overwhelming number of seeds
we find are reductions of two lanes (well - that's basically every
associative operation). That means we put a quite high pressure
on the SLP discovery process this way.
In total we find 583218 seeds we put to SLP discovery out of which
66205 pass that and only 6185 of those make it through
code generation checks. 796 of those are discarded because the reduction
is part of a larger SLP instance. 4195 of the remaining
are deemed not profitable to vectorize and 1194 are finally
vectorized. That's a poor 0.2% rate.
Of the 583218 seeds 486826 (83%) have two lanes, 60912 have three (10%),
28181 four (5%), 4808 five, 909 six and there are instances up to 120
lanes.
There's a set of 54086 candidate seeds we reject because
they contain a constant or invariant (not implemented yet) but still
have two or more lanes that could be put to SLP discovery.
2021-06-16 Richard Biener <rguenther@suse.de>
PR tree-optimization/54400
* tree-vectorizer.h (enum slp_instance_kind): Add
slp_inst_kind_bb_reduc.
(reduction_fn_for_scalar_code): Declare.
* tree-vect-data-refs.c (vect_slp_analyze_instance_dependence):
Check SLP_INSTANCE_KIND instead of looking at the
representative.
(vect_slp_analyze_instance_alignment): Likewise.
* tree-vect-loop.c (reduction_fn_for_scalar_code): Export.
* tree-vect-slp.c (vect_slp_linearize_chain): Split out
chain linearization from vect_build_slp_tree_2 and generalize
for the use of BB reduction vectorization.
(vect_build_slp_tree_2): Adjust accordingly.
(vect_optimize_slp): Elide permutes at the root of BB reduction
instances.
(vectorizable_bb_reduc_epilogue): New function.
(vect_slp_prune_covered_roots): Likewise.
(vect_slp_analyze_operations): Use them.
(vect_slp_check_for_constructors): Recognize associatable
chains for BB reduction vectorization.
(vectorize_slp_instance_root_stmt): Generate code for the
BB reduction epilogue.
* gcc.dg/vect/bb-slp-pr54400.c: New testcase.
The case of an initializer with side effects for a zero-length array seems
extremely unlikely, but we should still return the right type in that case.
PR c++/101029
gcc/cp/ChangeLog:
* init.c (build_vec_init): Preserve the type of base.
The gori engine can calculate outgoing ranges for exported values. This
change allows 1st degree recomputation. If a name is not exported from a
block, but one of the ssa_names used directly in computing it is, then
we can recompute the ssa_name on the edge using the edge values for its
operands.
* gimple-range-gori.cc (gori_compute::has_edge_range_p): Check with
may_recompute_p.
(gori_compute::may_recompute_p): New.
(gori_compute::outgoing_edge_range_p): Perform recomputations.
* gimple-range-gori.h (class gori_compute): Add prototype.
Range_on_edge was implemented in the cache to always return a range, but
only returned true when the edge actally changed the range.
Return true with any range that can be calculated.
* gimple-range-cache.cc (ranger_cache::range_on_edge): Always return
true when a range can be calculated.
* gimple-range.cc (gimple_ranger::dump_bb): Check has_edge_range_p.
After my patch for PR91706, or before that with the qualified call,
tsubst_baselink returned a BASELINK with BASELINK_BINFO indicating a base of
a still-dependent derived class. We need to look up the relevant base binfo
in the substituted class.
PR c++/101078
PR c++/91706
gcc/cp/ChangeLog:
* pt.c (tsubst_baselink): Update binfos in non-dependent case.
gcc/testsuite/ChangeLog:
* g++.dg/template/access39.C: New test.
Enable_new_values takes a boolean, returning the old value. The constructor
for ranger_cache initialized the m_new_value_p field by calling this routine
and ignorng the result. This potentially loads the old value uninitialized.
* gimple-range-cache.cc (ranger_cache::ranger_cache): Initialize
m_new_value_p directly.
The patch for 96391 changed linemap_compare_locations to give up on
comparing locations from macro expansions if we don't have column
information. But in this testcase, the BOILERPLATE macro is multiple lines
long, so we do want to compare locations within the macro. So this patch
moves the LINE_MAP_MAX_LOCATION_WITH_COLS check inside the block, to use it
for failing gracefully.
PR c++/100796
PR preprocessor/96391
libcpp/ChangeLog:
* line-map.c (linemap_compare_locations): Only use comparison with
LINE_MAP_MAX_LOCATION_WITH_COLS to avoid abort.
gcc/testsuite/ChangeLog:
* g++.dg/plugin/location-overflow-test-pr100796.c: New test.
* g++.dg/plugin/plugin.exp: Run it.
In addition to V8QI permutations, several other missing permutations are
added for 64bit vector modes for TARGET_SSSE3 and TARGET_SSE4_1 targets.
2021-06-16 Uroš Bizjak <ubizjak@gmail.com>
gcc/
PR target/89021
* config/i386/i386-expand.c (expand_vec_perm_2perm_pblendv):
Handle 64bit modes for TARGET_SSE4_1.
(expand_vec_perm_pshufb2): Handle 64bit modes for TARGET_SSSE3.
(expand_vec_perm_even_odd_pack): Handle V4HI mode.
(expand_vec_perm_even_odd_1) <case E_V4HImode>: Expand via
expand_vec_perm_pshufb2 for TARGET_SSSE3 and via
expand_vec_perm_even_odd_pack for TARGET_SSE4_1.
* config/i386/mmx.md (mmx_packusdw): New insn pattern.
In r12-1486-gcb326a6442f09cb36b05ce556fc91e10bfeb0cf6 I changed
__decay_copy to be a function object of unnamed class type. This causes
problems when importing the library headers:
error: conflicting global module declaration 'constexpr const std::ranges::__cust_access::<unnamed struct> std::ranges::__cust_access::__decay_copy'
The fix is to use a named struct instead of an anonymous one.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/iterator_concepts.h (__decay_copy): Name type.
In r12-1489-g8b93548778a487f31f21e0c6afe7e0bde9711fc4 I made the
[range.access] CPO types final and non-addressable. Tim Song pointed out
this is wrong. Only the [range.iter.ops] functions should be final and
non-addressable. Revert the changes to the [range.access] objects.
Signed-off-by: Jonathan Wakely <jwakely@redhat.com>
libstdc++-v3/ChangeLog:
* include/bits/ranges_base.h (ranges::begin, ranges::end)
(ranges::cbegin, ranges::cend, ranges::rbeing, ranges::rend)
(ranges::crbegin, ranges::crend, ranges::size, ranges::ssize)
(ranges::empty, ranges::data, ranges::cdata): Remove final
keywords and deleted operator& overloads.
* testsuite/24_iterators/customization_points/iter_move.cc: Use
new is_customization_point_object function.
* testsuite/24_iterators/customization_points/iter_swap.cc:
Likewise.
* testsuite/std/concepts/concepts.lang/concept.swappable/swap.cc:
Likewise.
* testsuite/std/ranges/access/begin.cc: Likewise.
* testsuite/std/ranges/access/cbegin.cc: Likewise.
* testsuite/std/ranges/access/cdata.cc: Likewise.
* testsuite/std/ranges/access/cend.cc: Likewise.
* testsuite/std/ranges/access/crbegin.cc: Likewise.
* testsuite/std/ranges/access/crend.cc: Likewise.
* testsuite/std/ranges/access/data.cc: Likewise.
* testsuite/std/ranges/access/empty.cc: Likewise.
* testsuite/std/ranges/access/end.cc: Likewise.
* testsuite/std/ranges/access/rbegin.cc: Likewise.
* testsuite/std/ranges/access/rend.cc: Likewise.
* testsuite/std/ranges/access/size.cc: Likewise.
* testsuite/std/ranges/access/ssize.cc: Likewise.
* testsuite/util/testsuite_iterators.h
(is_customization_point_object): New function.
Model the zero-high-half semantics of the narrowing arithmetic Neon
instructions in the aarch64_<sur><addsub>hn<mode> RTL pattern.
Modeling these semantics allows for better RTL combinations while
also removing some register allocation issues as the compiler now
knows that the operation is totally destructive.
Add new tests to narrow_zero_high_half.c to verify the benefit of
this change.
gcc/ChangeLog:
2021-06-14 Jonathan Wright <jonathan.wright@arm.com>
* config/aarch64/aarch64-simd.md (aarch64_<sur><addsub>hn<mode>):
Change to an expander that emits the correct instruction
depending on endianness.
(aarch64_<sur><addsub>hn<mode>_insn_le): Define.
(aarch64_<sur><addsub>hn<mode>_insn_be): Define.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/narrow_zero_high_half.c: Add new tests.
Split the aarch64_<su>qmovn<mode> pattern into separate scalar and
vector variants. Further split the vector RTL pattern into big/
little endian variants that model the zero-high-half semantics of the
underlying instruction. Modeling these semantics allows for better
RTL combinations while also removing some register allocation issues
as the compiler now knows that the operation is totally destructive.
Add new tests to narrow_zero_high_half.c to verify the benefit of
this change.
gcc/ChangeLog:
2021-06-14 Jonathan Wright <jonathan.wright@arm.com>
* config/aarch64/aarch64-simd-builtins.def: Split generator
for aarch64_<su>qmovn builtins into scalar and vector
variants.
* config/aarch64/aarch64-simd.md (aarch64_<su>qmovn<mode>_insn_le):
Define.
(aarch64_<su>qmovn<mode>_insn_be): Define.
(aarch64_<su>qmovn<mode>): Split into scalar and vector
variants. Change vector variant to an expander that emits the
correct instruction depending on endianness.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/narrow_zero_high_half.c: Add new tests.
Split the aarch64_sqmovun<mode> pattern into separate scalar and
vector variants. Further split the vector pattern into big/little
endian variants that model the zero-high-half semantics of the
underlying instruction. Modeling these semantics allows for better
RTL combinations while also removing some register allocation issues
as the compiler now knows that the operation is totally destructive.
Add new tests to narrow_zero_high_half.c to verify the benefit of
this change.
gcc/ChangeLog:
2021-06-14 Jonathan Wright <jonathan.wright@arm.com>
* config/aarch64/aarch64-simd-builtins.def: Split generator
for aarch64_sqmovun builtins into scalar and vector variants.
* config/aarch64/aarch64-simd.md (aarch64_sqmovun<mode>):
Split into scalar and vector variants. Change vector variant
to an expander that emits the correct instruction depending
on endianness.
(aarch64_sqmovun<mode>_insn_le): Define.
(aarch64_sqmovun<mode>_insn_be): Define.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/narrow_zero_high_half.c: Add new tests.
Modeling the zero-high-half semantics of the XTN narrowing
instruction in RTL indicates to the compiler that this is a totally
destructive operation. This enables more RTL simplifications and also
prevents some register allocation issues.
Add new tests to narrow_zero_high_half.c to verify the benefit of
this change.
gcc/ChangeLog:
2021-06-11 Jonathan Wright <jonathan.wright@arm.com>
* config/aarch64/aarch64-simd.md (aarch64_xtn<mode>_insn_le):
Define - modeling zero-high-half semantics.
(aarch64_xtn<mode>): Change to an expander that emits the
appropriate instruction depending on endianness.
(aarch64_xtn<mode>_insn_be): Define - modeling zero-high-half
semantics.
(aarch64_xtn2<mode>_le): Rename to...
(aarch64_xtn2<mode>_insn_le): This.
(aarch64_xtn2<mode>_be): Rename to...
(aarch64_xtn2<mode>_insn_be): This.
(vec_pack_trunc_<mode>): Emit truncation instruction instead
of aarch64_xtn.
* config/aarch64/iterators.md (Vnarrowd): Add Vnarrowd mode
attribute iterator.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/narrow_zero_high_half.c: Add new tests.
Add tests to verify that Neon narrowing-shift instructions clear the
top half of the result vector. It is sufficient to show that a
subsequent combine with a zero-vector is optimized away - leaving
just the narrowing-shift instruction.
gcc/testsuite/ChangeLog:
2021-06-15 Jonathan Wright <jonathan.wright@arm.com>
* gcc.target/aarch64/narrow_zero_high_half.c: New test.
When SRA transforms an assignment where the RHS is an aggregate decl
that it creates replacements for, the (least efficient) fallback
method of dealing with them is to store all the replacements back into
the original decl and then let the original assignment takes its
course.
That of course should not need to be done for TREE_READONLY bases
which cannot change contents. The SRA code handled this situation
only for DECL_IN_CONSTANT_POOL const decls, this patch modifies the
check so that it tests for TREE_READONLY and I also looked at all
other callers of generate_subtree_copies and added checks to another
one dealing with the same exact situation and one which deals with it
in a non-assignment context.
gcc/ChangeLog:
2021-06-11 Martin Jambor <mjambor@suse.cz>
PR tree-optimization/100453
* tree-sra.c (create_access): Disqualify any const candidates
which are written to.
(sra_modify_expr): Do not store sub-replacements back to a const base.
(handle_unscalarized_data_in_subtree): Likewise.
(sra_modify_assign): Likewise. Earlier, use TREE_READONLy test
instead of constant_decl_p.
gcc/testsuite/ChangeLog:
2021-06-10 Martin Jambor <mjambor@suse.cz>
PR tree-optimization/100453
* gcc.dg/tree-ssa/pr100453.c: New test.
I've noticed this test now on various arches sometimes FAILs, sometimes
PASSes (the line 12 test in particular).
The problem is that a = 0; initialization in the caller no longer happens
before the f(&a) call as what the argument points to is only used in
debug info.
Making the function noipa forces the caller to initialize it and still
tests what the test wants to test, namely that we don't consider *p as
valid location for the c variable at line 18 (after it has been overwritten
with *p = 1;).
2021-06-16 Jakub Jelinek <jakub@redhat.com>
* gcc.dg/guality/pr49888.c (f): Use noipa attribute instead of
noinline, noclone.
The following testcase is miscompiled on x86_64-linux, the bitfield store
is implemented as a RMW 64-bit operation at d+24 when the d variable has
size of only 28 bytes and scheduling moves in between the R and W part
a store to a different variable that happens to be right after the d
variable.
The reason for this is that we weren't creating
DECL_BIT_FIELD_REPRESENTATIVEs for bitfields in unions.
The following patch does create them, but treats all such bitfields as if
they were in a structure where the particular bitfield is the only field.
2021-06-16 Jakub Jelinek <jakub@redhat.com>
PR middle-end/101062
* stor-layout.c (finish_bitfield_representative): For fields in unions
assume nextf is always NULL.
(finish_bitfield_layout): Compute bit field representatives also in
unions, but handle it as if each bitfield was the only field in the
aggregate.
* gcc.dg/pr101062.c: New test.
When we face a sm_ord vs sm_unord for the same ref during
store sequence merging we assert that the ref is already marked
unsupported. But it can be that it will only be marked so
during the ongoing merging so instead of asserting mark it here.
Also apply some optimization to not waste resources to search
for already unsupported refs.
2021-06-16 Richard Biener <rguenther@suse.de>
PR tree-optimization/101088
* tree-ssa-loop-im.c (sm_seq_valid_bb): Only look for
supported refs on edges. Do not assert same ref but
different kind stores are unsuported but mark them so.
(hoist_memory_references): Only look for supported refs
on exits.
* gcc.dg/torture/pr101088.c: New testcase.
This patch tackles PR46235 to improve the code generated for bit tests
on x86_64 by making more use of the bt instruction. Currently, GCC emits
bt instructions when followed by condition jumps (thanks to Uros' splitters).
This patch adds splitters in i386.md, to catch the cases where bt is followed
by a conditional move (as in the original report), or by a setc/setnc (as in
comment 5 of the Bugzilla PR).
With this patch, the function in the original PR
int foo(int a, int x, int y) {
if (a & (1 << x))
return a;
return 1;
}
which with -O2 on mainline generates:
foo: movl %edi, %eax
movl %esi, %ecx
sarl %cl, %eax
testb $1, %al
movl $1, %eax
cmovne %edi, %eax
ret
now generates:
foo: btl %esi, %edi
movl $1, %eax
cmovc %edi, %eax
ret
Likewise, IsBitSet1 and IsBitSet2 (from comment 5)
bool IsBitSet1(unsigned char byte, int index) {
return (byte & (1<<index)) != 0;
}
bool IsBitSet2(unsigned char byte, int index) {
return (byte >> index) & 1;
}
Before:
movzbl %dil, %eax
movl %esi, %ecx
sarl %cl, %eax
andl $1, %eax
ret
After:
movzbl %dil, %edi
btl %esi, %edi
setc %al
ret
According to Agner Fog, SAR/SHR r,cl takes 2 cycles on skylake,
where BT r,r takes only one, so the performance improvements on
recent hardware may be more significant than implied by just
the reduced number of instructions. I've avoided transforming cases
(such as btsi_setcsi) where using bt sequences may not be a clear
win (over sarq/andl).
2010-06-15 Roger Sayle <roger@nextmovesoftware.com>
gcc/ChangeLog
PR rtl-optimization/46235
* config/i386/i386.md: New define_split for bt followed by cmov.
(*bt<mode>_setcqi): New define_insn_and_split for bt followed by setc.
(*bt<mode>_setncqi): New define_insn_and_split for bt then setnc.
(*bt<mode>_setnc<mode>): New define_insn_and_split for bt followed
by setnc with zero extension.
gcc/testsuite/ChangeLog
PR rtl-optimization/46235
* gcc.target/i386/bt-5.c: New test.
* gcc.target/i386/bt-6.c: New test.
* gcc.target/i386/bt-7.c: New test.
As the following testcase shows, libffi didn't handle properly
classify_arguments of structures at byte offsets not divisible by
UNITS_PER_WORD. The following patch adjusts it to match what
config/i386/ classify_argument does for that and also ports the
PR38781 fix there (the second chunk).
This has been committed to upstream libffi already:
5651bea284
2021-06-16 Jakub Jelinek <jakub@redhat.com>
* src/x86/ffi64.c (classify_argument): For FFI_TYPE_STRUCT set words
to number of words needed for type->size + byte_offset bytes rather
than just type->size bytes. Compute pos before the loop and check
total size of the structure.
* testsuite/libffi.call/nested_struct12.c: New test.
gcc/ada/
* sem_util.adb (Is_Volatile_Function): Follow the exact wording
of SPARK (regarding volatile functions) and Ada (regarding
protected functions).
gcc/ada/
* sem_util.adb (Is_OK_Volatile_Context): All references to
volatile objects are legal in preanalysis.
(Within_Volatile_Function): Previously it was wrongly called on
Empty entities; now it is only called on E_Return_Statement,
which allow the body to be greatly simplified.
gcc/ada/
* sem_res.adb (Set_Slice_Subtype): Revert special-case
introduced previously, which is not needed as Itypes created for
slices are precisely always used.
gcc/ada/
* urealp.adb (Scale): Change first paramter to Uint and adjust.
(Equivalent_Decimal_Exponent): Pass U.Den directly to Scale.
* libgnat/s-exponr.adb (Negative): Rename to...
(Safe_Negative): ...this and change its lower bound.
(Exponr): Adjust to above renaming and deal with Integer'First.
gcc/ada/
* sem_res.adb (Flag_Effectively_Volatile_Objects): Detect also
allocators within restricted contexts and not just entity names.
(Resolve_Actuals): Remove duplicated code for detecting
restricted contexts; it is now exclusively done in
Is_OK_Volatile_Context.
(Resolve_Entity_Name): Adapt to new parameter of
Is_OK_Volatile_Context.
* sem_util.ads, sem_util.adb (Is_OK_Volatile_Context): Adapt to
handle contexts both inside and outside of subprogram call
actual parameters.
(Within_Subprogram_Call): Remove; now handled by
Is_OK_Volatile_Context itself and its parameter.
gcc/ada/
* checks.adb (Apply_Scalar_Range_Check): Fix handling of check depending
on the parameter passing mechanism. Grammar adjustment ("has"
=> "have").
(Parameter_Passing_Mechanism_Specified): Add a hyphen in a comment.
gcc/ada/
* sem_res.adb (Is_Assignment_Or_Object_Expression): Whitespace
cleanup.
(Is_Attribute_Expression): Prevent AST climbing from going to
the root of the compilation unit.
gcc/ada/
* ghost.adb: Add another special case where full analysis is
needed. This bug is due to quirks in the way
Mark_And_Set_Ghost_Assignment works (it happens very early,
before name resolution is done).