Use new template parameters to replace usage of lambdas to move or not
tree values on copy.
libstdc++-v3/ChangeLog:
* include/bits/move.h (_GLIBCXX_FWDREF): New.
* include/bits/stl_tree.h: Adapt to use latter.
(_Rb_tree<>::_M_clone_node): Add _MoveValue template parameter.
(_Rb_tree<>::_M_mbegin): New.
(_Rb_tree<>::_M_begin): Use latter.
(_Rb_tree<>::_M_copy): Add _MoveValues template parameter.
* testsuite/23_containers/map/allocator/move_cons.cc: New test.
* testsuite/23_containers/multimap/allocator/move_cons.cc: New test.
* testsuite/23_containers/multiset/allocator/move_cons.cc: New test.
* testsuite/23_containers/set/allocator/move_cons.cc: New test.
Another remaining case is that we end up comparing calls with mismatching
number of parameters or with different permutations of them.
This is because we hash decls to nothing. This patch improves that by
hashing decls by their code and parm decls by indexes that are stable.
Also for defualt defs in SSA_NAMEs we can add the corresponding decl (that
is usually parm decls).
Still we could improve on this by hasing ssa names by their definit parameters
and possibly making maps of other decls and assigning them stable function
local IDs.
* ipa-icf-gimple.c (func_checker::hash_operand): Improve hashing of
decls.
one of common remaining reasons for ICF to fail after loading in fuction
body is mismatched type of automatic vairable. This is becuase
compatible_types_p resorts to checking TYPE_MAIN_VARIANTS for
euqivalence that prevents merging many TBAA compaitle cases. (And thus
is also not reflected by the hash extended by alias sets of accesses.)
Since in gimple
automatic variables are just blocks of memory I think we should only
check its size only. All accesses are matched when copmparing the actual
loads/stores.
I am not sure if we need to match types of other DECLs but I decided I can try
to be safe here: for PARM_DECl/RESUILT_DECL we match them anyway to be sure
that functions are ABI compatible. For CONST_DECL and readonly global
VAR_DECLs they are matched when comparing their constructors.
* ipa-icf-gimple.c (func_checker::compare_decl): Do not compare types
of local variables.
2020-11-10 Andrea Corallo <andrea.corallo@arm.com>
PR target/97726
* gcc.target/arm/simd/bf16_vldn_1.c: Relax regexps not to fail on
big endian.
* gcc.target/arm/simd/vldn_lane_bf16_1.c: Likewise
* gcc.target/arm/simd/vmmla_1.c: Add -mfloat-abi=hard flag.
This modifies vectorizable_slp_permutation to update the type of the children
of a perm node before trying to permute them. This allows us to be able to
permute invariant nodes.
This will be covered by test from the SLP pattern matcher.
gcc/ChangeLog:
* tree-vect-slp.c (vectorizable_slp_permutation): Update types on nodes
when needed.
Unlike the other headers that declare alias templates in namespace pmr,
<regex> includes <memory_resource>. That was done because the
pmr::string::const_iterator typedef requires pmr::string to be complete,
which requires pmr::polymorphic_allocator<char> to be complete.
By using __normal_iterator<const char*, pmr::string> instead of the
const_iterator typedef we can avoid the completeness requirement.
This makes <regex> smaller, by not requiring <memory_resource> and its
<shared_mutex> dependency, which depends on <chrono>. Backporting this
will also help with PR 97876, where <stop_token> ends up being needed by
<regex> via <memory_resource>.
libstdc++-v3/ChangeLog:
PR libstdc++/92546
* include/std/regex (pmr::smatch, pmr::wsmatch): Declare using
underlying __normal_iterator type, not nested typedef
basic_string::const_iterator.
This makes hybrid SLP discovery deal with stmts indirectly consumed
by SLP, for example via patterns. This means that all uses of a
stmt end up in SLP vectorized stmts.
This helps my prototype patches for PR97832 where I make SLP discovery
re-associate chains to make operands match. This ends up building
SLP computation nodes without 1:1 representatives in the scalar IL
and thus no scalar lane defs in SLP_TREE_SCALAR_STMTS. Nevertheless
all of the original scalar stmts are consumed so this represents
another kind of SLP pattern for the computation chain result.
2020-11-20 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (maybe_push_to_hybrid_worklist): New function.
(vect_detect_hybrid_slp): Use it. Perform a backward walk
over the IL.
It always annoyed me to see those empty SLP nodes in dumpfiles:
t.c:16:3: note: node 0x3a2a280 (max_nunits=1, refcnt=1)
t.c:16:3: note: { }
t.c:16:3: note: children 0x3a29db0 0x3a29e90
resulting from two-operator handling. The following makes
sure to also dump the operation template or VEC_PERM_EXPR.
2020-11-20 Richard Biener <rguenther@suse.de>
* tree-vect-slp.c (vect_print_slp_tree): Also dump
SLP_TREE_REPRESENTATIVE.
The following patch implements __builtin_clear_padding builtin that clears
the padding bits in object representation (but preserves value
representation). Inside of unions it clears only those padding bits that
are padding for all the union members (so that it never alters value
representation).
It handles trailing padding, padding in the middle of structs including
bitfields (PDP11 unhandled, I've never figured out how those bitfields
work), VLAs (doesn't handle variable length structures, but I think almost
nobody uses them and it isn't worth the extra complexity). For VLAs and
sufficiently large arrays it uses runtime clearing loop instead of emitting
straight-line code (unless arrays are inside of a union).
The way I think this can be used for atomics is e.g. if the structures
are power of two sized and small enough that we use the hw atomics
for say compare_exchange __builtin_clear_padding could be called first on
the address of expected and desired arguments (for desired only if we want
to ensure that most of the time the atomic memory will have padding bits
cleared), then perform the weak cmpxchg and if that fails, we got the
value from the atomic memory; we can call __builtin_clear_padding on a copy
of that and then compare it with expected, and if it is the same with the
padding bits masked off, we can use the original with whatever random
padding bits in it as the new expected for next cmpxchg.
__builtin_clear_padding itself is not atomic and therefore it shouldn't
be called on the atomic memory itself, but compare_exchange*'s expected
argument is a reference and normally the implementation may store there
the current value from memory, so padding bits can be cleared in that,
and desired is passed by value rather than reference, so clearing is fine
too.
When using libatomic, we can use it either that way, or add new libatomic
APIs that accept another argument, pointer to the padding bit bitmask,
and construct that in the template as
alignas (_T) unsigned char _mask[sizeof (_T)];
std::memset (_mask, ~0, sizeof (_mask));
__builtin_clear_padding ((_T *) _mask);
which will have bits cleared for padding bits and set for bits taking part
in the value representation. Then libatomic could internally instead
of using memcmp compare
for (i = 0; i < N; i++) if ((val1[i] & mask[i]) != (val2[i] & mask[i]))
2020-11-20 Jakub Jelinek <jakub@redhat.com>
PR libstdc++/88101
gcc/
* builtins.def (BUILT_IN_CLEAR_PADDING): New built-in function.
* gimplify.c (gimplify_call_expr): Rewrite single argument
BUILT_IN_CLEAR_PADDING into two-argument variant.
* gimple-fold.c (clear_padding_unit, clear_padding_buf_size): New
const variables.
(struct clear_padding_struct): New type.
(clear_padding_flush, clear_padding_add_padding,
clear_padding_emit_loop, clear_padding_type,
clear_padding_union, clear_padding_real_needs_padding_p,
clear_padding_type_may_have_padding_p,
gimple_fold_builtin_clear_padding): New functions.
(gimple_fold_builtin): Handle BUILT_IN_CLEAR_PADDING.
* doc/extend.texi (__builtin_clear_padding): Document.
gcc/c-family/
* c-common.c (check_builtin_function_arguments): Handle
BUILT_IN_CLEAR_PADDING.
gcc/testsuite/
* c-c++-common/builtin-clear-padding-1.c: New test.
* c-c++-common/torture/builtin-clear-padding-1.c: New test.
* c-c++-common/torture/builtin-clear-padding-2.c: New test.
* c-c++-common/torture/builtin-clear-padding-3.c: New test.
* c-c++-common/torture/builtin-clear-padding-4.c: New test.
* c-c++-common/torture/builtin-clear-padding-5.c: New test.
* g++.dg/torture/builtin-clear-padding-1.C: New test.
* g++.dg/torture/builtin-clear-padding-2.C: New test.
* gcc.dg/builtin-clear-padding-1.c: New test.
The documentation for POST_MODIFY says:
Currently, the compiler can only handle second operands of the
form (plus (reg) (reg)) and (plus (reg) (const_int)), where
the first operand of the PLUS has to be the same register as
the first operand of the *_MODIFY.
The following testcase ICEs, because combine just attempts to simplify
things and ends up with
(post_modify (reg1) (plus (mult (reg2) (const_int 4)) (reg1))
but the target predicates accept it, because they only verify
that POST_MODIFY's second operand is PLUS and the second operand
of the PLUS is a REG.
The following patch fixes this by performing further verification that
the POST_MODIFY is in the form it should be.
2020-11-20 Jakub Jelinek <jakub@redhat.com>
PR target/97528
* config/arm/arm.c (neon_vector_mem_operand): For POST_MODIFY, require
first POST_MODIFY operand is a REG and is equal to the first operand
of PLUS.
* gcc.target/arm/pr97528.c: New test.
There is a loophole in new string store merging support added recently:
it does not check that the stores are consecutive, which is obviously
required if you want to concatenate them... Simple fix attached, the
nice thing being that it can fall back to the regular processing if
any hole is detected in the series of stores, thanks to the handling
of STRING_CST by native_encode_expr.
gcc/ChangeLog:
* gimple-ssa-store-merging.c (struct merged_store_group): Add
new 'consecutive' field.
(merged_store_group): Set it to true.
(do_merge): Set it to false if the store is not consecutive and
set string_concatenation to false in this case.
(merge_into): Call do_merge on entry.
(merge_overlapping): Likewise.
gcc/testsuite/ChangeLog:
* gnat.dg/opt90a.adb: New test.
* gnat.dg/opt90b.adb: Likewise.
* gnat.dg/opt90c.adb: Likewise.
* gnat.dg/opt90d.adb: Likewise.
* gnat.dg/opt90e.adb: Likewise.
* gnat.dg/opt90a_pkg.ads: New helper.
* gnat.dg/opt90b_pkg.ads: Likewise.
* gnat.dg/opt90c_pkg.ads: Likewise.
* gnat.dg/opt90d_pkg.ads: Likewise.
* gnat.dg/opt90e_pkg.ads: Likewise.
after fixing few issues I gotto stage where 1.4M icf mismatches are due to
comparing two gimple clobber. The problem is that operand_equal_p match
clobber
case CONSTRUCTOR:
/* In GIMPLE empty constructors are allowed in initializers of
aggregates. */
return !CONSTRUCTOR_NELTS (arg0) && !CONSTRUCTOR_NELTS (arg1);
But this happens too late after comparing its types (that are not very relevant
for memory store).
In the context of ipa-icf we do not really need to match RHS of gimple clobbers:
it is enough to know that the LHS stores can be considered equivalent.
I this added logic to hash them all the same way and compare using
TREE_CLOBBER_P flag. I see other option in extending operand_equal_p
in fold-const to handle them more generously or making stmt hash and compare
to skip comparing/hashing RHS of gimple_clobber_p.
* ipa-icf-gimple.c (func_checker::hash_operand): Hash gimple clobber.
(func_checker::operand_equal_p): Special case gimple clobber.
The patch introduces absM named pattern to generate optimal insn sequence
for CMOVE_TARGET targets. Currently, the expansion goes through neg+max
optabs, and the following code is generated:
movl %edi, %eax
negl %eax
cmpl %edi, %eax
cmovl %edi, %eax
This sequence is unoptimal in two ways. a) The compare instruction is
not needed, since NEG insn sets the sign flag based on the result.
The CMOV can use sign flag to select between negated and original value:
movl %edi, %eax
negl %eax
cmovs %edi, %eax
b) On some targets, CMOV is undesirable due to its performance issues.
In addition to TARGET_EXPAND_ABS bypass, the patch introduces STV
conversion of abs RTX to use PABS SSE insn:
vmovd %edi, %xmm0
vpabsd %xmm0, %xmm0
vmovd %xmm0, %eax
The patch changes compare mode of NEG instruction to CCGOCmode,
which is the same mode as the mode of SUB instruction. IOW, sign bit
becomes usable.
Also, the mode iterator of <maxmin:code><mode>3 pattern is changed
to SWI48x instead of SWI248. The purpose of maxmin expander is to
prepare max/min RTX for STV to eventually convert them to SSE PMAX/PMIN
instructions, in order to *avoid* CMOV insns with general registers.
2020-11-20 Uroš Bizjak <ubizjak@gmail.com>
gcc/
PR target/97873
* config/i386/i386.md (*neg<mode>2_2): Rename from
"*neg<mode>2_cmpz". Use CCGOCmode instead of CCZmode.
(*negsi2_zext): Rename from *negsi2_cmpz_zext.
Use CCGOCmode instead of CCZmode.
(*neg<mode>_ccc_1): New insn pattern.
(*neg<dwi>2_doubleword): Use *neg<mode>_ccc_1.
(abs<mode>2): Add FLAGS_REG clobber.
Use TARGET_CMOVE insn predicate.
(*abs<mode>2_1): New insn_and_split pattern.
(*absdi2_doubleword): Ditto.
(<maxmin:code><mode>3): Use SWI48x mode iterator.
(*<maxmin:code><mode>3): Use SWI48 mode iterator.
* config/i386/i386-features.c
(general_scalar_chain::compute_convert_gain): Handle ABS code.
(general_scalar_chain::convert_insn): Ditto.
(general_scalar_to_vector_candidate_p): Ditto.
gcc/testsuite/
PR target/97873
* gcc.target/i386/pr97873.c: New test.
* gcc.target/i386/pr97873-1.c: New test.
Eric reported that the --enable-link-serialization changes seemed to
cause the binaries to be always relinked, for example from the
gcc/ directory of the build tree:
make
[relink of gnat1, brig1, cc1plus, d21, f951, go1, lto1, ...]
make
[relink of gnat1, brig1, cc1plus, d21, f951, go1, lto1, ...]
Furthermore as reported in PR, it can cause problems during make install
where make install rebuilds the binaries again.
The problem is that for make .PHONY targets are just
"rebuilt" always, so it is very much undesirable for the cc1plus$(exeext)
etc. dependencies to include .PHONY targets, but I was using
them - cc1plus.prev which would depend on some *.serial and
e.g. cc1.serial depending on c and c depending on cc1$(exeext).
The following patch rewrites this so that *.serial and *.prev aren't
.PHONY targets, but instead just make variables.
I was worried that the order in which the language makefile fragments are
included (which is quite random, what order we get from the filesystem
matching */config-lang.in) would be a problem but it seems to work fine
- as it uses make = rather than := variables, later definitions are just
fine for earlier uses as long as the uses aren't needed during the
makefile parsing, but only in the dependencies of make targets and in
their commands.
2020-11-20 Jakub Jelinek <jakub@redhat.com>
PR other/97911
gcc/
* configure.ac: In SERIAL_LIST use lang words without .serial
suffix. Change $lang.prev from a target to variable and instead
of depending on *.serial expand to the *.serial variable if
the word is in the SERIAL_LIST at all, otherwise to nothing.
* configure: Regenerated.
gcc/c/
* Make-lang.in (c.serial): Change from goal to a variable.
(.PHONY): Drop c.serial.
gcc/ada/
* gcc-interface/Make-lang.in (ada.serial): Change from goal to a
variable.
(.PHONY): Drop ada.serial and ada.prev.
(gnat1$(exeext)): Depend on $(ada.serial) rather than ada.serial.
gcc/brig/
* Make-lang.in (brig.serial): Change from goal to a variable.
(.PHONY): Drop brig.serial and brig.prev.
(brig1$(exeext)): Depend on $(brig.serial) rather than brig.serial.
gcc/cp/
* Make-lang.in (c++.serial): Change from goal to a variable.
(.PHONY): Drop c++.serial and c++.prev.
(cc1plus$(exeext)): Depend on $(c++.serial) rather than c++.serial.
gcc/d/
* Make-lang.in (d.serial): Change from goal to a variable.
(.PHONY): Drop d.serial and d.prev.
(d21$(exeext)): Depend on $(d.serial) rather than d.serial.
gcc/fortran/
* Make-lang.in (fortran.serial): Change from goal to a variable.
(.PHONY): Drop fortran.serial and fortran.prev.
(f951$(exeext)): Depend on $(fortran.serial) rather than
fortran.serial.
gcc/go/
* Make-lang.in (go.serial): Change from goal to a variable.
(.PHONY): Drop go.serial and go.prev.
(go1$(exeext)): Depend on $(go.serial) rather than go.serial.
gcc/jit/
* Make-lang.in (jit.serial): Change from goal to a
variable.
(.PHONY): Drop jit.serial and jit.prev.
($(LIBGCCJIT_FILENAME)): Depend on $(jit.serial) rather than
jit.serial.
gcc/lto/
* Make-lang.in (lto1.serial, lto2.serial): Change from goals to
variables.
(.PHONY): Drop lto1.serial, lto2.serial, lto1.prev and lto2.prev.
($(LTO_EXE)): Depend on $(lto1.serial) rather than lto1.serial.
($(LTO_DUMP_EXE)): Depend on $(lto2.serial) rather than lto2.serial.
gcc/objc/
* Make-lang.in (objc.serial): Change from goal to a variable.
(.PHONY): Drop objc.serial and objc.prev.
(cc1obj$(exeext)): Depend on $(objc.serial) rather than objc.serial.
gcc/objcp/
* Make-lang.in (obj-c++.serial): Change from goal to a variable.
(.PHONY): Drop obj-c++.serial and obj-c++.prev.
(cc1objplus$(exeext)): Depend on $(obj-c++.serial) rather than
obj-c++.serial.
This patch is to fix insn type of p8_mtvsrd_df from mfvsr to mtvsr,
in order to align with the other places using mtvsrd.
gcc/ChangeLog:
* config/rs6000/rs6000.md (p8_mtvsrd_df): Fix insn type.
As mentioned in the PR, the previous PR91029 patch was testing
op2 >= 0 which is unnecessary, even negative op2 values will work the same,
furthermore, from if a % b > 0 we can deduce a > 0 rather than just a >= 0
(0 % b would be 0), and it actually valid even for other constants than 0,
a % b > 5 means a > 5 (a % b has the same sign as a and a in [0, 5] would
result in a % b in [0, 5]. Also, we can deduce a range for the other
operand, if we know
a % b >= 20, then b must be (in absolute value for signed modulo) > 20,
for a % [0, 20] the result would be [0, 19].
2020-11-19 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/91029
* range-op.cc (operator_trunc_mod::op1_range): Don't require signed
types, nor require that op2 >= 0. Implement (a % b) >= x && x > 0
implies a >= x and (a % b) <= x && x < 0 implies a <= x.
(operator_trunc_mod::op2_range): New method.
* gcc.dg/tree-ssa/pr91029-1.c: New test.
* gcc.dg/tree-ssa/pr91029-2.c: New test.
When shifting outside the valid range of [0, precision-1], we can
choose to process just the valid ones since the rest is undefined.
this allows us to produce results for x << [0,2][+INF, +INF] by discarding
the invalid ranges and processing just [0,2].
gcc/
PR tree-optimization/93781
* range-op.cc (get_shift_range): Rename from
undefined_shift_range_check and now return valid shift ranges.
(operator_lshift::fold_range): Use result from get_shift_range.
(operator_rshift::fold_range): Ditto.
gcc/testsuite/
* gcc.dg/tree-ssa/pr93781-1.c: New.
* gcc.dg/tree-ssa/pr93781-2.c: New.
* gcc.dg/tree-ssa/pr93781-3.c: New.
This exposes the template specialization table, so the modules
machinery may access it. The hashed entity (tmpl, args & spec) is
available, along with a hash table walker. We also need a way of
finding or inserting entries, along with some bookkeeping fns to deal
with the instantiation and (partial) specialization lists.
gcc/cp/
* cp-tree.h (struct spec_entry): Moved from pt.c.
(walk_specializations, match_mergeable_specialization)
(get_mergeable_specialization_flags)
(add_mergeable_specialization): Declare.
* pt.c (struct spec_entry): Moved to cp-tree.h.
(walk_specializations, match_mergeable_specialization)
(get_mergeable_specialization_flags)
(add_mergeable_specialization): New.
Since glibc 2.27 the pthread_self symbol has been defined in libc rather
than libpthread. Because we only call pthread_self through a weak alias
it's possible for statically linked executables to end up without a
definition of pthread_self. This crashes when trying to call an
undefined weak symbol.
We can use the __GLIBC_PREREQ version check to detect the version of
glibc where pthread_self is no longer in libpthread, and call it
directly rather than through the weak reference.
It would be better to check for pthread_self in libc during configure
instead of hardcoding the __GLIBC_PREREQ check. That would be
complicated by the fact that prior to glibc 2.27 libc.a didn't have the
pthread_self symbol, but libc.so.6 did. The configure checks would need
to try to link both statically and dynamically, and the result would
depend on whether the static libc.a happens to be installed during
configure (which could vary between different systems using the same
version of glibc). Doing it properly is left for a future date, as that
will be needed anyway after glibc moves all pthread symbols from
libpthread to libc. When that happens we should revisit the whole
approach of using weak symbols for pthread symbols.
For the purposes of std::this_thread::get_id() we call
pthread_self() directly when using glibc 2.27 or later. Otherwise, if
__gthread_active_p() is true then we know the libpthread symbol is
available so we call that. Otherwise, we are single-threaded and just
use ((__gthread_t)1) as the thread ID.
An undesirable consequence of this change is that code compiled prior to
the change might inline the old definition of this_thread::get_id()
which always returns (__gthread_t)1 in a program that isn't linked to
libpthread. Code compiled after the change will use pthread_self() and
so get a real TID. That could result in the main thread having different
thread::id values in different translation units. This seems acceptable,
as there are not expected to be many uses of thread::id in programs
that aren't linked to libpthread.
An earlier version of this patch also changed __gthread_self() to use
__GLIBC_PREREQ(2, 27) and only use the weak symbol for older glibc. Tha
might still make sense to do, but isn't needed by libstdc++ now.
libstdc++-v3/ChangeLog:
PR libstdc++/95989
* config/os/gnu-linux/os_defines.h (_GLIBCXX_NATIVE_THREAD_ID):
Define new macro to get reliable thread ID.
* include/bits/std_thread.h: (this_thread::get_id): Use new
macro if it's defined.
* testsuite/30_threads/jthread/95989.cc: New test.
* testsuite/30_threads/this_thread/95989.cc: New test.
This patch exposes the constexpr hash table so that the modules
machinery can save and load constexpr bodies. While there I noticed
that we could do a little constification of the hasher and comparator
functions. Also combine the saving machinery to a single function
returning void -- nothing ever looked at its return value.
gcc/cp/
* cp-tree.h (struct constexpr_fundef): Moved from constexpr.c.
(maybe_save_constexpr_fundef): Declare.
(register_constexpr_fundef): Take constexpr_fundef object, return
void.
* decl.c (mabe_save_function_definition): Delete, functionality
moved to maybe_save_constexpr_fundef.
(emit_coro_helper, finish_function): Adjust.
* constexpr.c (struct constexpr_fundef): Moved to cp-tree.h.
(constexpr_fundef_hasher::equal): Constify.
(constexpr_fundef_hasher::hash): Constify.
(retrieve_constexpr_fundef): Make non-static.
(maybe_save_constexpr_fundef): Break out checking and duplication
from ...
(register_constexpr_fundef): ... here. Just register the constexpr.
* fold-const.c (operand_compare::operand_equal_p): Fix thinko in
COMPONENT_REF handling and guard types_same_for_odr by
virtual_method_call_p.
(operand_compare::hash_operand): Likewise.
The C and C++ FEs handle zero sized arrays differently, C uses
NULL TYPE_MAX_VALUE on non-NULL TYPE_DOMAIN on complete ARRAY_TYPEs
with bitsize_zero_node TYPE_SIZE, while C++ FE likes to set
TYPE_MAX_VALUE to the largest value (and min to the lowest).
Martin has used array_type_nelts in get_parm_array_spec where the
function on the C form of [0] arrays returns error_mark_node and the code
crashes soon afterwards. The following patch teaches array_type_nelts about
this (e.g. dwarf2out already handles that as [0]). While it will change
what is_empty_type returns for certain types (e.g. struct S { int a[0]; };),
as those types occupy zero bits in C, it should make an ABI difference.
So, the tree.c change makes the c-decl.c code handle the [0] arrays
like any other constant extents, and the c-decl.c change just makes sure
that if we'd run into error_mark_node e.g. from the VLA expressions, we
don't crash on those.
2020-11-19 Jakub Jelinek <jakub@redhat.com>
PR c/97860
* tree.c (array_type_nelts): For complete arrays with zero min
and NULL max and zero size return -1.
* c-decl.c (get_parm_array_spec): Bail out of nelts is
error_operand_p.
* gcc.dg/pr97860.c: New test.
Since my r11-3092 the following is rejected with -std=c++20:
struct T { explicit T(); };
void fn(int n) {
new T[1]();
}
with "would use explicit constructor 'T::T()'". It is because since
that change we go into the P1009 block in build_new (array_p is false,
but nelts is non-null and we're in C++20). Since we only have (), we
build a {} and continue to build_new_1, which then calls build_vec_init
and then we error because the {} isn't CONSTRUCTOR_IS_DIRECT_INIT.
For (), which is value-initializing, we want to do what we were doing
before: pass empty init and let build_value_init take care of it.
For various reasons I wanted to dig a little bit deeper into this,
and as a result, I'm adding a test for [expr.new]/24 (and checked that
out current behavior matches clang++).
gcc/cp/ChangeLog:
PR c++/97523
* init.c (build_new): When value-initializing an array new,
leave the INIT as an empty vector.
gcc/testsuite/ChangeLog:
PR c++/97523
* g++.dg/expr/anew5.C: New test.
* g++.dg/expr/anew6.C: New test.
Unfortunately, the otherwise beautiful
for (constructor_elt &elt : *CONSTRUCTOR_ELTS (init))
is not immune to an empty constructor, so we have to check
CONSTRUCTOR_ELTS first.
gcc/cp/ChangeLog:
PR c++/97895
* pt.c (do_auto_deduction): Don't crash when the constructor has
zero elements.
gcc/testsuite/ChangeLog:
PR c++/97895
* g++.dg/cpp0x/auto54.C: New test.
this adds configure tests for features that modules can take advantage
of -- and if they are not present has reduced or fallback functionality.
gcc/
* configure.ac: Add tests for fstatat, sighandler_t, O_CLOEXEC,
unix-domain and ipv6 sockets.
* config.in: Rebuilt.
* configure: Rebuilt.
It turns out there are legitimate cases for the new decl to not have
lang-specific.
PR c++/97905
gcc/cp/
* decl.c (duplicate_decls): Relax new assert.
gcc/testsuite/
* g++.dg/lookup/pr97905.C: New.
Add builtins for HALT and LMBD, per Texas Instruments document
SPRUHV7C. Use the new LMBD pattern to define an expand for clz.
Binutils [1] and sim [2] support for LMBD instruction are merged now.
[1] https://sourceware.org/pipermail/binutils/2020-October/113901.html
[2] https://sourceware.org/pipermail/gdb-patches/2020-November/173141.html
gcc/ChangeLog:
* config/pru/alu-zext.md: Add lmbd patterns for zero_extend
variants.
* config/pru/pru.c (enum pru_builtin): Add HALT and LMBD.
(pru_init_builtins): Ditto.
(pru_builtin_decl): Ditto.
(pru_expand_builtin): Ditto.
* config/pru/pru.h (CLZ_DEFINED_VALUE_AT_ZERO): Define PRU
value for CLZ with zero value parameter.
* config/pru/pru.md: Add halt, lmbd and clz patterns.
* doc/extend.texi: Document PRU builtins.
gcc/testsuite/ChangeLog:
* gcc.target/pru/halt.c: New test.
* gcc.target/pru/lmbd.c: New test.
Currently we have three vector cost models: cheap, dynamic and
unlimited. -O2 -ftree-vectorize uses “cheap” by default, but that's
still relatively aggressive about peeling and aliasing checks,
and can lead to significant code size growth.
This patch adds an even more conservative choice, which for lack of
imagination I've called “very cheap”. It only allows vectorisation
if the vector code entirely replaces the scalar code. It also
requires one iteration of the vector loop to pay for itself,
regardless of how often the loop iterates. (If the vector loop
needs multiple iterations to be beneficial then things are
probably too close to call, and the conservative thing would
be to stick with the scalar code.)
The idea is that this should be suitable for -O2, although the patch
doesn't change any defaults itself.
I tested this by building and running a bunch of workloads for SVE,
with three options:
(1) -O2
(2) -O2 -ftree-vectorize -fvect-cost-model=very-cheap
(3) -O2 -ftree-vectorize [-fvect-cost-model=cheap]
All three builds used the default -msve-vector-bits=scalable and
ran with the minimum vector length of 128 bits, which should give
a worst-case bound for the performance impact.
The workloads included a mixture of microbenchmarks and full
applications. Because it's quite an eclectic mix, there's not
much point giving exact figures. The aim was more to get a general
impression.
Code size growth with (2) was much lower than with (3). Only a
handful of tests increased by more than 5%, and all of them were
microbenchmarks.
In terms of performance, (2) was significantly faster than (1)
on microbenchmarks (as expected) but also on some full apps.
Again, performance only regressed on a handful of tests.
As expected, the performance of (3) vs. (1) and (3) vs. (2) is more
of a mixed bag. There are several significant improvements with (3)
over (2), but also some (smaller) regressions. That seems to be in
line with -O2 -ftree-vectorize being a kind of -O2.5.
The patch reorders vect_cost_model so that values are in order
of increasing aggressiveness, which makes it possible to use
range checks. The value 0 still represents “unlimited”,
so “if (flag_vect_cost_model)” is still a meaningful check.
gcc/
* doc/invoke.texi (-fvect-cost-model): Add a very-cheap model.
* common.opt (fvect-cost-model=): Add very-cheap as a possible option.
(fsimd-cost-model=): Likewise.
(vect_cost_model): Add very-cheap.
* flag-types.h (vect_cost_model): Add VECT_COST_MODEL_VERY_CHEAP.
Put the values in order of increasing aggressiveness.
* tree-vect-data-refs.c (vect_enhance_data_refs_alignment): Use
range checks when comparing against VECT_COST_MODEL_CHEAP.
(vect_prune_runtime_alias_test_list): Do not allow any alias
checks for the very-cheap cost model.
* tree-vect-loop.c (vect_analyze_loop_costing): Do not allow
any peeling for the very-cheap cost model. Also require one
iteration of the vector loop to pay for itself.
gcc/testsuite/
* gcc.dg/vect/vect-cost-model-1.c: New test.
* gcc.dg/vect/vect-cost-model-2.c: Likewise.
* gcc.dg/vect/vect-cost-model-3.c: Likewise.
* gcc.dg/vect/vect-cost-model-4.c: Likewise.
* gcc.dg/vect/vect-cost-model-5.c: Likewise.
* gcc.dg/vect/vect-cost-model-6.c: Likewise.
These tests use std::this_thread::sleep_for without including <thread>.
libstdc++-v3/ChangeLog:
* testsuite/30_threads/async/async.cc: Include <thread>.
* testsuite/30_threads/future/members/93456.cc: Likewise.
Add an initial cost table for Cortex-A76 - this is copied from
cotexa57_extra_costs but updated based on the Optimization Guide.
Use the new cost table on all Neoverse tunings and ensure the tunings
are consistent for all. As a result more compact code is generated
with more combined shift+alu operations. Eg. -mcpu=cortex-a76 will now
merge the shifts in:
int f(int x, int y) { return (x & y << 3) * (x | y << 3); }
and w2, w0, w1, lsl 3
orr w0, w0, w1, lsl 3
mul w0, w2, w0
ret
SPEC2017 codesize improves by 0.02% and SPECINT2017 shows 0.24% gain.
2020-11-18 Wilco Dijkstra <wdijkstr@arm.com>
gcc/
* config/aarch64/aarch64.c (neoversen1_tunings): Use new
cortexa76_extra_costs.
(neoversev1_tunings): Likewise.
(neoversen2_tunines): Likewise.
* config/arm/aarch-cost-tables.h (cortexa76_extra_costs):
add new costs.
Improve the inline memcpy expansion. Use integer load/store for copies <= 24
bytes instead of SIMD. Set the maximum copy to expand to 256 by default,
except that -Os or no Neon expands up to 128 bytes. When using LDP/STP of
Q-registers, also use Q-register accesses for the unaligned tail, saving 2
instructions (eg. all sizes up to 48 bytes emit exactly 4 instructions).
Cleanup code and comments.
The codesize gain vs the GCC10 expansion is 0.05% on SPECINT2017.
2020-11-03 Wilco Dijkstra <wdijkstr@arm.com>
gcc/
* config/aarch64/aarch64.c (aarch64_expand_cpymem): Cleanup code and
comments, tweak expansion decisions and improve tail expansion.
We need to include limits.h (or <climits>) in adaint.c because of LLONG_MIN.
gcc/ada/ChangeLog:
PR ada/97805
* adaint.c: Include climits in C++ and limits.h otherwise.
This adds the capability to locate the main file on the user or system
include paths. That's extremely useful to users building header
units. Searching has to be requiested (plain header-unit compilation
will not search). Also, to make include_next work as expected when
building a header unit, we add a mechanism to retrofit a non-searched
source file as one on the include path.
libcpp/
* include/cpplib.h (enum cpp_main_search): New.
(struct cpp_options): Add main_search field.
(cpp_main_loc): Declare.
(cpp_retrofit_as_include): Declare.
* internal.h (struct cpp_reader): Add main_loc field.
(_cpp_in_main_source_file): Not main if main is a header.
* init.c (cpp_read_main_file): Use main_search option to locate
main file. Set main_loc
* files.c (cpp_retrofit_as_include): New.
This makes it possible to use std::thread without including the whole of
<thread>. It also makes this_thread::get_id() and this_thread::yield()
available even when there is no gthreads support (e.g. when GCC is built
with --disable-threads or --enable-threads=single).
In order for the std:🧵:id return type of this_thread::get_id() to
be defined, std:thread itself is defined unconditionally. However the
constructor that creates new threads is not defined for single-threaded
builds. The thread::join() and thread::detach() member functions are
defined inline for single-threaded builds and just throw an exception
(because we know the thread cannot be joinable if the constructor that
creates joinable threads doesn't exit).
The thread::hardware_concurrency() member function is also defined
inline and returns 0 (as suggested by the standard when the value "is
not computable or well-defined").
The main benefit for most targets is that other headers such as <future>
do not need to include the whole of <thread> just to be able to create a
std::thread. That avoids including <stop_token> and std::jthread where
not required. This is another partial fix for PR 92546.
This also means we can use this_thread::get_id() and this_thread::yield()
in <stop_token> instead of using the gthread functions directly. This
removes some preprocessor conditionals, simplifying the code.
libstdc++-v3/ChangeLog:
PR libstdc++/92546
* include/Makefile.am: Add new <bits/std_thread.h> header.
* include/Makefile.in: Regenerate.
* include/std/future: Include new header instead of <thread>.
* include/std/stop_token: Include new header instead of
<bits/gthr.h>.
(stop_token::_S_yield()): Use this_thread::yield().
(_Stop_state_t::_M_requester): Change type to std:🧵:id.
(_Stop_state_t::_M_request_stop()): Use this_thread::get_id().
(_Stop_state_t::_M_remove_callback(_Stop_cb*)): Likewise.
Use __is_single_threaded() to decide whether to synchronize.
* include/std/thread (thread, operator==, this_thread::get_id)
(this_thread::yield): Move to new header.
(operator<=>, operator!=, operator<, operator<=, operator>)
(operator>=, hash<thread::id>, operator<<): Define even when
gthreads not available.
* src/c++11/thread.cc: Include <memory>.
* include/bits/std_thread.h: New file.
(thread, operator==, this_thread::get_id, this_thread::yield):
Define even when gthreads not available.
[!_GLIBCXX_HAS_GTHREADS] (thread::join, thread::detach)
(thread::hardware_concurrency): Define inline.
I recently added overflow checks to src/c++11/futex.cc for PR 93456, but
then changed the type of the timespec for PR 93421. This meant the
overflow checks were no longer using the right range, because the
variable being written to might be smaller than time_t.
This introduces new typedef that corresponds to the tv_sec member of the
struct being passed to the syscall, and uses that typedef in the range
checks.
libstdc++-v3/ChangeLog:
PR libstdc++/93421
PR libstdc++/93456
* src/c++11/futex.cc (syscall_time_t): New typedef for
the type of the syscall_timespec::tv_sec member.
(relative_timespec, _M_futex_wait_until)
(_M_futex_wait_until_steady): Use syscall_time_t in overflow
checks, not time_t.
In preparing module patch 7 I realized there was a cleanup I could
make to simplify it. This is that cleanup. Also, when doing the
cleanup I noticed some macros had been turned into inline functions,
but not renamed to the preprocessors internal namespace
(_cpp_$INTERNAL rather than cpp_$USER). Thus, this renames those
functions, deletes an internal field of the file structure, and
determines whether we're in the main file by comparing to
pfile->main_file, the _cpp_file of the main file.
libcpp/
* internal.h (cpp_in_system_header): Rename to ...
(_cpp_in_system_header): ... here.
(cpp_in_primary_file): Rename to ...
(_cpp_in_main_source_file): ... here. Compare main_file equality
and check main_search value.
* lex.c (maybe_va_opt_error, _cpp_lex_direct): Adjust for rename.
* macro.c (_cpp_builtin_macro_text): Likewise.
(replace_args): Likewise.
* directives.c (do_include_next): Likewise.
(do_pragma_once, do_pragma_system_header): Likewise.
* files.c (struct _cpp_file): Delete main_file field.
(pch_open): Check pfile->main_file equality.
(make_cpp_file): Drop cpp_reader parm, don't set main_file.
(_cpp_find_file): Adjust.
(_cpp_stack_file): Check pfile->main_file equality.
(struct report_missing_guard_data): Add cpp_reader field.
(report_missing_guard): Check pfile->main_file equality.
(_cpp_report_missing_guards): Adjust.
This fixes a typo in the TREE_CODE compare which should
compare against TYPE_DECL, not TYPE_NAME.
2020-11-19 Richard Biener <rguenther@suse.de>
* fold-const.c (operand_compare::hash_operand): Fix typo.
This refactors things so assigned ranks are dumped and the cache
is consistently used also for PHIs.
2020-11-19 Richard Biener <rguenther@suse.de>
* tree-ssa-reassoc.c (get_rank): Refactor to consistently
use the cache and dump ranks assigned.
* fold-const.c (operand_compare::operand_equal_p): More OBJ_TYPE_REF
matching to correct place; drop OEP_ADDRESS_OF for TOKEN, OBJECT and
class.
(operand_compare::hash_operand): Hash ODR type for OBJ_TYPE_REF.
Add aarch64 vec_widen_lshift_lo/hi patterns and fix bug it triggers in
mid-end. This pattern takes one vector with N elements of size S, shifts
each element left by the element width and stores the results as N
elements of size 2*s (in 2 result vectors). The aarch64 backend
implements this with the shll,shll2 instruction pair.
gcc/ChangeLog:
* config/aarch64/aarch64-simd.md: Add vec_widen_lshift_hi/lo<mode>
patterns.
* tree-vect-stmts.c (vectorizable_conversion): Fix for widen_lshift
case.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/vect-widen-lshift.c: New test.
Add widening add, subtract patterns to tree-vect-patterns. Update the
widened code of patterns that detect PLUS_EXPR to also detect
WIDEN_PLUS_EXPR. These patterns take 2 vectors with N elements of size
S and perform an add/subtract on the elements, storing the results as N
elements of size 2*S (in 2 result vectors). This is implemented in the
aarch64 backend as addl,addl2 and subl,subl2 respectively. Add aarch64
tests for patterns.
gcc/ChangeLog:
* doc/generic.texi: Document new widen_plus/minus_lo/hi tree codes.
* doc/md.texi: Document new widenening add/subtract hi/lo optabs.
* expr.c (expand_expr_real_2): Add widen_add, widen_subtract cases.
* optabs-tree.c (optab_for_tree_code): Add case for widening optabs.
* optabs.def (OPTAB_D): Define vectorized widen add, subtracts.
* tree-cfg.c (verify_gimple_assign_binary): Add case for widening adds,
subtracts.
* tree-inline.c (estimate_operator_cost): Add case for widening adds,
subtracts.
* tree-vect-generic.c (expand_vector_operations_1): Add case for
widening adds, subtracts
* tree-vect-patterns.c (vect_recog_widen_add_pattern): New recog
pattern.
(vect_recog_widen_sub_pattern): New recog pattern.
(vect_recog_average_pattern): Update widened add code.
(vect_recog_average_pattern): Update widened add code.
* tree-vect-stmts.c (vectorizable_conversion): Add case for widened add,
subtract.
(supportable_widening_operation): Add case for widened add, subtract.
* tree.def
(WIDEN_PLUS_EXPR): New tree code.
(WIDEN_MINUS_EXPR): New tree code.
(VEC_WIDEN_ADD_HI_EXPR): New tree code.
(VEC_WIDEN_PLUS_LO_EXPR): New tree code.
(VEC_WIDEN_MINUS_HI_EXPR): New tree code.
(VEC_WIDEN_MINUS_LO_EXPR): New tree code.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/vect-widen-add.c: New test.
* gcc.target/aarch64/vect-widen-sub.c: New test.
Add widening add and subtract patterns to the aarch64
backend. These allow taking vectors of N elements of size S
and performing and add/subtract on the high or low half
widening the resulting elements and storing N/2 elements of size 2*S.
These correspond to the addl,addl2,subl,subl2 instructions.
gcc/ChangeLog:
* config/aarch64/aarch64-simd.md: New patterns
vec_widen_saddl_lo/hi_<mode>.
We need to fold the stmt to canonicalize MEM_REFs which means
we're back to using replace_uses_by. Which means we need dominators
to not require a CFG cleanup upthread.
2020-11-19 Richard Biener <rguenther@suse.de>
PR tree-optimization/97901
* tree-ssa-propagate.c (clean_up_loop_closed_phi): Compute
dominators and use replace_uses_by.
* gcc.dg/torture/pr97901.c: New testcase.
François Dumont
Jan Hubicka
Andrew MacLeod
Andrea Corallo
Nathan Sidwell
Tamar Christina
Jonathan Wakely
Richard Biener
Jakub Jelinek
Eric Botcazou
Uros Bizjak
Kewen Lin
Martin Uecker
GCC Administrator
Marek Polacek
Dimitar Dimitrov
Richard Sandiford
Wilco Dijkstra
Joel Hutton