| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
|
|
The AArch64 FEAT_FP8FMA extension introduces instructions for
multiply-add of vectors.
This patch introduces the following instructions:
1. {vmlalbq|vmlaltq}_f16_mf8_fpm.
2. {vmlalbq|vmlaltq}_lane{q}_f16_mf8_fpm.
3. {vmlallbbq|vmlallbtq|vmlalltbq|vmlallttq}_f32_mf8_fpm.
4. {vmlallbbq|vmlallbtq|vmlalltbq|vmlallttq}_lane{q}_f32_mf8_fpm.
It introduces the fp8fma flag.
gcc/ChangeLog:
* config/aarch64/aarch64-builtins.cc
(check_simd_lane_bounds): Add support for new unspecs.
(aarch64_expand_pragma_builtins): Add support for new unspecs.
* config/aarch64/aarch64-c.cc
(aarch64_update_cpp_builtins): New flags.
* config/aarch64/aarch64-option-extensions.def
(AARCH64_OPT_EXTENSION): New flags.
* config/aarch64/aarch64-simd-pragma-builtins.def
(ENTRY_FMA_FPM): Macro to declare fma intrinsics.
(REQUIRED_EXTENSIONS): Define to declare functions behind
command line flags.
* config/aarch64/aarch64-simd.md:
(@aarch64_<fpm_uns_op><VQ_HSF:mode><VQ_HSF:mode><V16QI_ONLY:mode><V16QI_ONLY:mode): Instruction pattern for fma intrinsics.
(@aarch64_<fpm_uns_op><VQ_HSF:mode><VQ_HSF:mode><V16QI_ONLY:mode><VB:mode><SI_ONLY:mode): Instruction pattern for fma intrinsics with lane.
* config/aarch64/aarch64.h
(TARGET_FP8FMA): New flag for fp8fma instructions.
* config/aarch64/iterators.md: New attributes and iterators.
* doc/invoke.texi: New flag for fp8fma instructions.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/simd/fma_fpm.c: New test.
|
|
The AArch64 FEAT_FP8DOT2 and FEAT_FP8DOT4 extension introduces
instructions for dot product of vectors.
This patch introduces the following intrinsics:
1. vdot{q}_{fp16|fp32}_mf8_fpm.
2. vdot{q}_lane{q}_{fp16|fp32}_mf8_fpm.
It introduces two flags: fp8dot2 and fp8dot4.
We had to add space for another type in aarch64_pragma_builtins_data
struct. The macros were updated to reflect that.
We added a new aarch64_builtin_signature variant, quaternary, and added
support for it in the functions aarch64_fntype and
aarch64_expand_pragma_builtin.
We added a new namespace, function_checker, to implement range checks
for functions defined using the new pragma approach. The old intrinsic
range checks will continue to work. All the new AdvSIMD intrinsics we
define that need lane checks should be using the function in this
namespace to implement the checks.
gcc/ChangeLog:
* config/aarch64/aarch64-builtins.cc
(ENTRY): Change to handle extra type.
(enum class): Added new variant.
(struct aarch64_pragma_builtins_data): Add support for another
type.
(aarch64_get_number_of_args): Handle new signature.
(require_integer_constant): New function to check whether the
operand is an integer constant.
(require_immediate_range): New function to validate index
ranges.
(check_simd_lane_bounds): New function to validate index
operands.
(aarch64_general_check_builtin_call): Call
function_checker::check-simd_lane_bounds.
(aarch64_expand_pragma_builtin): Handle new signature.
* config/aarch64/aarch64-c.cc
(aarch64_update_cpp_builtins): New flags.
* config/aarch64/aarch64-option-extensions.def
(AARCH64_OPT_EXTENSION): New flags.
* config/aarch64/aarch64-simd-pragma-builtins.def
(ENTRY_BINARY): Change to handle extra type.
(ENTRY_BINARY_FPM): Change to handle extra type.
(ENTRY_UNARY_FPM): Change to handle extra type.
(ENTRY_TERNARY_FPM_LANE): Macro to declare fpm ternary with
lane intrinsics.
(ENTRY_VDOT_FPM): Macro to declare vdot intrinsics.
(REQUIRED_EXTENSIONS): Define to declare functions behind
command line flags.
* config/aarch64/aarch64-simd.md:
(@aarch64_<fpm_uns_op><VHF:mode><VHF:mode><VB:mode><VB:mode>):
Instruction pattern for vdot2 intrinsics.
(@aarch64_<fpm_uns_op><VHF:mode><VHF:mode><VB:mode><VB2:mode><SI_ONLY:mode>):
Instruction pattern for vdot2 intrinsics with lane.
(@aarch64_<fpm_uns_op><VDQSF:mode><VDQSF:mode><VB:mode><VB:mode>):
Instruction pattern for vdot4 intrinsics.
(@aarch64_<fpm_uns_op><VDQSF:mode><VDQSF:mode><VB:mode><VB2:mode><SI_ONLY:mode>):
Instruction pattern for vdo4 intrinsics with lane.
* config/aarch64/aarch64.h
(TARGET_FP8DOT2): New flag for fp8dot2 instructions.
(TARGET_FP8DOT4): New flag for fp8dot4 instructions.
* config/aarch64/iterators.md: New attributes and iterators.
* doc/invoke.texi: New flag for fp8dot2 and fp8dot4
instructions.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/simd/vdot2_fpmdot.c: New test.
* gcc.target/aarch64/simd/vdot4_fpmdot.c: New test.
|
|
The AArch64 FEAT_FP8 extension introduces instructions for conversion
and scaling.
This patch introduces the following intrinsics:
1. vcvt{1|2}_{bf16|high_bf16|low_bf16}_mf8_fpm.
2. vcvt{q}_mf8_f16_fpm.
3. vcvt_{high}_mf8_f32_fpm.
4. vscale{q}_{f16|f32|f64}.
We introduced two aarch64_builtin_signatures enum variants, unary and
ternary, and added support for these variants in the functions
aarch64_fntype and aarch64_expand_pragma_builtin.
We added new simd_types for integers (s32, s32q, and s64q) and for
floating points (f8 and f8q).
Because we added support for fp8 intrinsics here, we modified the check
in acle/fp8.c that was checking that __ARM_FEATURE_FP8 macro is not
defined.
gcc/ChangeLog:
* config/aarch64/aarch64-builtins.cc
(ENTRY): Modified to support uses_fpmr flag.
(enum class): New variants to support new signatures.
(struct aarch64_pragma_builtins_data): Add a new boolean field,
uses_fpmr.
(aarch64_get_number_of_args): Helper function used in
aarch64_fntype and aarch64_expand_pragma_builtin.
(aarch64_fntype): Handle new signatures.
(aarch64_expand_pragma_builtin): Handle new signatures.
* config/aarch64/aarch64-c.cc
(aarch64_update_cpp_builtins): New flag for FP8.
* config/aarch64/aarch64-simd-pragma-builtins.def
(ENTRY_BINARY): Macro to declare binary intrinsics.
(ENTRY_TERNARY): Macro to declare ternary intrinsics.
(ENTRY_UNARY): Macro to declare unary intrinsics.
(ENTRY_VHSDF): Macro to declare binary intrinsics.
(ENTRY_VHSDF_VHSDI): Macro to declare binary intrinsics.
(REQUIRED_EXTENSIONS): Define to declare functions behind
command line flags.
* config/aarch64/aarch64-simd.md
(@aarch64_<fpm_unary_bf_uns_op><V8BF_ONLY:mode><VB:mode>): Unary
pattern.
(@aarch64_<fpm_unary_hf_uns_op><V8HF_ONLY:mode><VB:mode>): Unary
pattern.
(@aarch64_lower_<fpm_unary_bf_uns_op><V8BF_ONLY:mode><V16QI_ONLY:mode>):
Unary pattern.
(@aarch64_lower_<fpm_unary_hf_uns_op><V8HF_ONLY:mode><V16QI_ONLY:mode>):
Unary pattern.
(@aarch64<fpm_uns_op><VB:mode><VCVTFPM:mode><VH_SF:mode>):
Binary pattern.
(@aarch64_<fpm_uns_op><V16QI_ONLY:mode><V8QI_ONLY:mode><V4SF_ONLY:mode><V4SF_ONLY:mode>):
Unary pattern.
(@aarch64_<fpm_uns_op><VHSDF:mode><VHSDI:mode>): Binary pattern.
* config/aarch64/iterators.md: New attributes and iterators.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/acle/fp8.c: Remove check that fp8 feature
macro doesn't exist.
* gcc.target/aarch64/simd/scale_fpm.c: New test.
* gcc.target/aarch64/simd/vcvt_fpm.c: New test.
|
|
This patch refactors the infrastructure for defining advsimd pragma
intrinsics, adding support for more flexible type and signature
handling in future SIMD extensions.
A new simd_type structure is introduced, which allows for consistent
mode and qualifier management across various advsimd operations.
gcc/ChangeLog:
* config/aarch64/aarch64-builtins.cc (ENTRY): Modify to
include modes and qualifiers for simd_type structure.
(ENTRY_VHSDF): Move to aarch64-builtins.cc to decouple.
(struct simd_type): New structure for managing mode and
qualifier combinations for SIMD types.
(struct aarch64_pragma_builtins_data): Replace mode with
simd_type to support multiple argument types for intrinsics.
(aarch64_fntype): Modify to handle different shapes type.
(aarch64_expand_pragma_builtin): Modify to handle different
shapes type.
* config/aarch64/aarch64-simd-pragma-builtins.def (ENTRY_BINARY):
Move from aarch64-builtins.cc.
(ENTRY_VHSDF): Move from aarch64-builtins.cc.
(REQUIRED_EXTENSIONS): New macro.
|
|
For cstorebf4 it uses comparison_operator for BFmode compare, which is
incorrect when directly uses ix86_expand_setcc as it does not canonicalize
the input comparison to correct the compare code by swapping operands.
The original code without AVX10.2 calls emit_store_flag_force, who
actually calls to emit_store_flags_1 and recurisive calls to this expander
again with swapped operand and flag.
Therefore, we can avoid do the redundant recurisive call by adjusting
the comparison_operator to ix86_fp_comparison_operator, and calls
ix86_expand_setcc directly.
gcc/ChangeLog:
PR target/117495
* config/i386/i386.md (cstorebf4): Use ix86_fp_comparison_operator
and calls ix86_expand_setcc directly.
gcc/testsuite/ChangeLog:
PR target/117495
* gcc.target/i386/pr117495.c: New test.
|
|
error: unrecognizable insn:
(insn 35 34 36 2 (set (subreg:RVVM1SF (reg/v:RVVM1x4SF 142 [ _r ]) 0)
(unspec:RVVM1SF [
(const_vector:RVVM1SF repeat [
(const_double:SF 0.0 [0x0.0p+0])
])
(reg:DI 0 zero)
(const_int 1 [0x1])
(reg:SI 66 vl)
(reg:SI 67 vtype)
] UNSPEC_TH_VWLDST)) -1
(nil))
during RTL pass: mode_sw
PR target/116591
gcc/ChangeLog:
* config/riscv/vector.md: Add restriction to call pred_th_whole_mov.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/xtheadvector/pr116591.c: New test.
|
|
This attempts to simplify and clean up our std::hash code. The primary
benefit is improved diagnostics for users when they do something wrong
involving std::hash or unordered containers. An additional benefit is
that for the unstable ABI (--enable-symvers=gnu-versioned-namespace) we
can reduce the memory footprint of several std::hash specializations.
In the current design, __hash_enum is a base class of the std::hash
primary template, but the partial specialization of __hash_enum for
non-enum types is disabled. This means that if a user forgets to
specialize std::hash for their class type (or forgets to use a custom
hash function for unordered containers) they get error messages about
std::__hash_enum not being constructible. This is confusing when there
is no enum type involved: why should users care about __hash_enum not
being constructible if they're not trying to hash enums?
This change makes the std::hash primary template only derive from
__hash_enum when the template argument type is an enum. Otherwise, it
derives directly from a new class template, __hash_not_enabled. This new
class template defines the deleted members that cause a given std::hash
specialization to be a disabled specialization (as per P0513R0). Now
when users try to use a disabled specialization, they get more
descriptive errors that mention __hash_not_enabled instead of
__hash_enum.
Additionally, adjust __hash_base to remove the deprecated result_type
and argument_type typedefs for C++20 and later.
In the current code we use a __poison_hash base class in the std::hash
specializations for std::unique_ptr, std::optional, and std::variant.
The primary template of __poison_hash has deleted special members, which
is used to conditionally disable the derived std::hash specialization.
This can also result in confusing diagnostics, because seeing "poison"
in an enabled specialization is misleading. Only some uses of
__poison_hash actually "poison" anything, i.e. cause a specialization to
be disabled. In other cases it's just an empty base class that does
nothing.
This change removes __poison_hash and changes the std::hash
specializations that were using it to conditionally derive from
__hash_not_enabled instead. When the std::hash specialization is
enabled, there is no more __poison_hash base class. However, to preserve
the ABI properties of those std::hash specializations, we need to
replace __poison_hash with some other empty base class. This is needed
because in the current code std::hash<std::variant<int, const int>> has
two __poison_hash<int> base classes, which must have unique addresses,
so sizeof(std::hash<std::variant<int, const int>>) == 2. To preserve
this unfortunate property, a new __hash_empty_base class is used as a
base class to re-introduce du0plicate base classes that increase the
class size. For the unstable ABI we don't use __hash_empty_base so the
std::hash<std::variant<T...>> specializations are always size 1, and
the class hierarchy is much simpler so will compile faster.
Additionally, remove the result_type and argument_type typedefs from all
disabled specializations of std::hash for std::unique_ptr,
std::optional, and std::variant. Those typedefs are useless for disabled
specializations, and although the standard doesn't say they must *not*
be present for disabled specializations, it certainly only requires them
for enabled specializations. Finally, for C++20 the typedefs are also
removed from enabled specializations of std::hash for std::unique_ptr,
std::optional, and std::variant.
libstdc++-v3/ChangeLog:
* doc/xml/manual/evolution.xml: Document removal of nested types
from std::hash specializations.
* doc/html/manual/api.html: Regenerate.
* include/bits/functional_hash.h (__hash_base): Remove
deprecated nested types for C++20.
(__hash_empty_base): Define new class template.
(__is_hash_enabled_for): Define new variable template.
(__poison_hash): Remove.
(__hash_not_enabled): Define new class template.
(__hash_enum): Remove partial specialization for non-enums.
(hash): Derive from __hash_not_enabled for non-enums, instead of
__hash_enum.
* include/bits/unique_ptr.h (__uniq_ptr_hash): Derive from
__hash_base. Conditionally derive from __hash_empty_base.
(__uniq_ptr_hash<>): Remove disabled specialization.
(hash): Do not derive from __hash_base unconditionally.
Conditionally derive from either __uniq_ptr_hash or
__hash_not_enabled.
* include/std/optional (__optional_hash_call_base): Remove.
(__optional_hash): Define new class template.
(hash): Derive from either
(hash): Conditionally derive from either __optional_hash or
__hash_not_enabled. Remove nested typedefs.
* include/std/variant (_Base_dedup): Replace __poison_hash with
__hash_empty_base.
(__variant_hash_call_base_impl): Remove.
(__variant_hash): Define new class template.
(hash): Conditionally derive from either __variant_hash or
__hash_not_enabled. Remove nested typedefs.
* testsuite/20_util/optional/hash.cc: Check whether nested types
are present.
* testsuite/20_util/variant/hash.cc: Likewise.
* testsuite/20_util/optional/hash_abi.cc: New test.
* testsuite/20_util/unique_ptr/hash/abi.cc: New test.
* testsuite/20_util/unique_ptr/hash/types.cc: New test.
* testsuite/20_util/variant/hash_abi.cc: New test.
|
|
We have two overloads of _M_find_before_node but they have quite
different performance characteristics, which isn't necessarily obvious.
The original version, _M_find_before_node(bucket, key, hash_code), looks
only in the specified bucket, doing a linear search within that bucket
for an element that compares equal to the key. This is the typical fast
lookup for hash containers, assuming the load factor is low so that each
bucket isn't too large.
The newer _M_find_before_node(key) was added in r12-6272-ge3ef832a9e8d6a
and could be naively assumed to calculate the hash code and bucket for
key and then call the efficient _M_find_before_node(bkt, key, code)
function. But in fact it does a linear search of the entire container.
This is potentially very slow and should only be used for a suitably
small container, as determined by the __small_size_threshold() function.
We don't even have a comment pointing out this O(N) performance of the
newer overload.
Additionally, the newer overload is only ever used in exactly one place,
which would suggest it could just be removed. However there are several
places that do the linear search of the whole container with an explicit
loop each time.
This adds a new member function, _M_locate, and uses it to replace most
uses of _M_find_node and the loops doing linear searches. This new
member function does both forms of lookup, the linear search for small
sizes and the _M_find_node(bkt, key, code) lookup within a single
bucket. The new function returns a __location_type which is a struct
that contains a pointer to the first node matching the key (if such a
node is present), or the hash code and bucket index for the key. The
hash code and bucket index allow the caller to know where a new node
with that key should be inserted, for the cases where the lookup didn't
find a matching node.
The result struct actually contains a pointer to the node *before* the
one that was located, as that is needed for it to be useful in erase and
extract members. There is a member function that returns the found node,
i.e. _M_before->_M_nxt downcast to __node_ptr, which should be used in
most cases.
This new function greatly simplifies the functions that currently have
to do two kinds of lookup and explicitly check the current size against
the small size threshold.
Additionally, now that try_emplace is defined directly in _Hashtable
(not in _Insert_base) we can use _M_locate in there too, to speed up
some try_emplace calls. Previously it did not do the small-size linear
search.
It would be possible to add a function to get a __location_type from an
iterator, and then rewrite some functions like _M_erase and
_M_extract_node to take a __location_type parameter. While that might be
conceptually nice, it wouldn't really make the code any simpler or more
readable than it is now. That isn't done in this change.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (__location_type): New struct.
(_M_locate): New member function.
(_M_find_before_node(const key_type&)): Remove.
(_M_find_node): Move variable initialization into condition.
(_M_find_node_tr): Likewise.
(operator=(initializer_list<T>), try_emplace, _M_reinsert_node)
(_M_merge_unique, find, erase(const key_type&)): Use _M_locate
for lookup.
|
|
I realised that _M_merge_unique and _M_merge_multi call extract(iter)
which then has to call _M_get_previous_node to iterate through the
bucket to find the node before the one iter points to. Since the merge
function is already iterating over the entire container, we had the
previous node a moment ago. Walking the whole bucket to find it again is
wasteful. We could just rewrite the loop in terms of node pointers
instead of iterators, and then call _M_extract_node directly. However,
this is only possible when the source container is the same type as the
destination, because otherwise we can't access the source's private
members (_M_before_begin, _M_begin, _M_extract_node etc.)
Add overloads of _M_merge_unique and _M_merge_multi that work with
source containers of the same type, to enable this optimization.
For both overloads of _M_merge_unique we can also remove the conditional
modifications to __n_elt and just consistently decrement it for every
element processed. Use a multiplier of one or zero that dictates whether
__n_elt is passed to _M_insert_unique_node or not. We can also remove
the repeated calls to size() and just keep track of the size in a local
variable.
Although _M_merge_unique and _M_merge_multi should be safe for
"self-merge", i.e. when doing c.merge(c), it's wasteful to search/insert
every element when we don't need to do anything. Add 'this == &source'
checks to the overloads taking an lvalue of the container's own type.
Because those checks aren't needed for the rvalue overloads, change
those to call the underlying _M_merge_xxx function directly instead of
going through the lvalue overload that checks the address.
I've also added more extensive tests for better coverage of the new
overloads added in this commit.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_M_merge_unique): Add overload for
merging from same type.
(_M_merge_unique<Compatible>): Simplify size tracking. Add
comment.
(_M_merge_multi): Add overload for merging from same type.
(_M_merge_multi<Compatible>): Add comment.
* include/bits/unordered_map.h (unordered_map::merge): Check for
self-merge in the lvalue overload. Call _M_merge_unique directly
for the rvalue overload.
(unordered_multimap::merge): Likewise.
* include/bits/unordered_set.h (unordered_set::merge): Likewise.
(unordered_multiset::merge): Likewise.
* testsuite/23_containers/unordered_map/modifiers/merge.cc:
Add more tests.
* testsuite/23_containers/unordered_multimap/modifiers/merge.cc:
Likewise.
* testsuite/23_containers/unordered_multiset/modifiers/merge.cc:
Likewise.
* testsuite/23_containers/unordered_set/modifiers/merge.cc:
Likewise.
|
|
This removes the overloaded _S_equals and _S_node_equals functions,
replacing them with 'if constexpr' in the handful of places they're
used.
libstdc++-v3/ChangeLog:
* include/bits/hashtable_policy.h (_Hashtable_base::_S_equals):
Remove.
(_Hashtable_base::_S_node_equals): Remove.
(_Hashtable_base::_M_key_equals_tr): Fix inaccurate
static_assert string.
(_Hashtable_base::_M_equals, _Hashtable_base::_M_equals_tr): Use
'if constexpr' instead of _S_equals.
(_Hashtable_base::_M_node_equals): Use 'if constexpr' instead of
_S_node_equals.
|
|
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_Hashtable): Remove _Equality base
class.
(_Hashtable::_M_equal): Define equality comparison here instead
of in _Equality::_M_equal.
* include/bits/hashtable_policy.h (_Equality): Remove.
|
|
There's no reason to have a separate base class defining the insert
member functions now. They can all be moved into the _Hashtable class,
which simplifies them slightly.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_Hashtable): Remove inheritance from
__detail::_Insert and move its members into _Hashtable.
* include/bits/hashtable_policy.h (__detail::_Insert): Remove.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
Use scoped guard types to clean up if an exception is thrown. This
allows some try-catch blocks to be removed.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (operator=(const _Hashtable&)): Use
RAII instead of try-catch.
(_M_assign(_Ht&&, _NodeGenerator&)): Likewise.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
We can just use a cast to the appropriate type instead of calling a
function to do it. This gives the compiler less work to compile and
optimize, and at -O0 avoids a function call per element.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_Hashtable::__fwd_value_for):
Remove.
(_Hashtable::_M_assign): Use static_cast instead of
__fwd_value_for.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
This adds a convenient _M_assign overload for the common case where the
node generator is the _AllocNode type. Only two places need to call
_M_assign with a _ReuseOrAllocNode node generator, so all the other
calls to _M_assign can use the new overload instead of manually
constructing a node generator.
The _AllocNode::operator(Args&&...) function doesn't need to be a
variadic template. It is only ever called with a single argument of type
const value_type& or value_type&&, so could be simplified. That isn't
done in this commit.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_Hashtable): Remove typedefs for
node generators.
(_Hashtable::_M_assign(_Ht&&)): Add new overload.
(_Hashtable::operator=(initializer_list<value_type>)): Add local
typedef for node generator.
(_Hashtable::_M_assign_elements): Likewise.
(_Hashtable::operator=(const _Hashtable&)): Use new _M_assign
overload.
(_Hashtable(const _Hashtable&)): Likewise.
(_Hashtable(const _Hashtable&, const allocator_type&)):
Likewise.
(_Hashtable(_Hashtable&&, __node_alloc_type&&, false_type)):
Likewise.
* include/bits/hashtable_policy.h (_Insert): Remove typedef for
node generator.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
This reworks the internal member functions for erasure from
unordered containers, similarly to the earlier commit doing it for
insertion.
Instead of multiple overloads of _M_erase which are selected via tag
dispatching, the erase(const key_type&) member can use 'if constexpr' to
choose an appropriate implementation (returning after erasing a single
element for unique keys, or continuing to erase all equivalent elements
for non-unique keys).
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (_Hashtable::_M_erase): Remove
overloads for erasing by key, moving logic to ...
(_Hashtable::erase): ... here.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
This completely reworks the internal member functions for insertion into
unordered containers. Currently we use a mixture of tag dispatching (for
unique vs non-unique keys) and template specialization (for maps vs
sets) to correctly implement insert and emplace members.
This removes a lot of complexity and indirection by using 'if constexpr'
to select the appropriate member function to call.
Previously there were four overloads of _M_emplace, for unique keys and
non-unique keys, and for hinted insertion and non-hinted. However two of
those were redundant, because we always ignore the hint for unique keys
and always use a hint for non-unique keys. Those four overloads have
been replaced by two new non-overloaded function templates:
_M_emplace_uniq and _M_emplace_multi. The former is for unique keys and
doesn't take a hint, and the latter is for non-unique keys and takes a
hint.
In the body of _M_emplace_uniq there are special cases to handle
emplacing values from which a key_type can be extracted directly. This
means we don't need to allocate a node and construct a value_type that
might be discarded if an equivalent key is already present. The special
case applies when emplacing the key_type into std::unordered_set, or
when emplacing std::pair<cv key_type, X> into std::unordered_map, or
when emplacing two values into std::unordered_map where the first has
type cv key_type. For the std::unordered_set case, obviously if we're
inserting something that's already the key_type, we can look it up
directly. For the std::unordered_map cases, we know that the inserted
std::pair<const key_type, mapped_type> would have its first element
initialized from first member of a std::pair value, or from the first of
two values, so if that is a key_type, we can look that up directly.
All the _M_insert overloads used a node generator parameter, but apart
from the one case where _M_insert_range was called from
_Hashtable::operator=(initializer_list<value_type>), that parameter was
always the _AllocNode type, never the _ReuseOrAllocNode type. Because
operator=(initializer_list<value_type>) was rewritten in an earlier
commit, all calls to _M_insert now use _AllocNode, so there's no reason
to pass the generator as a template parameter when inserting.
The multiple overloads of _Hashtable::_M_insert can all be removed now,
because the _Insert_base::insert members now call either _M_emplace_uniq
or _M_emplace_multi directly, only passing a hint to the latter. Which
one to call is decided using 'if constexpr (__unique_keys::value)' so
there is no unnecessary code instantiation, and overload resolution is
much simpler.
The partial specializations of the _Insert class template can be
entirely removed, moving the minor differences in 'insert' member
functions into the common _Insert_base base class. The different
behaviour for maps and sets can be implemented using enable_if
constraints and 'if constexpr'. With the _Insert class template no
longer needed, the _Insert_base class template can be renamed to
_Insert. This is a minor simplification for the complex inheritance
hierarchy used by _Hashtable, removing one base class. It also means
one less class template instantiation, and no need to match the right
partial specialization of _Insert. The _Insert base class could be
removed entirely by moving all its 'insert' members into _Hashtable,
because without any variation in specializations of _Insert there is no
reason to use a base class to define those members. That is left for a
later commit.
Consistently using _M_emplace_uniq or _M_emplace_multi for insertion
means we no longer attempt to avoid constructing a value_type object to
find its key, removing the PR libstdc++/96088 optimizations. This fixes
the bugs caused by those optimizations, such as PR libstdc++/115285, but
causes regressions in the expected number of allocations and temporary
objects constructed for the PR 96088 tests. It should be noted that the
"regressions" in the 96088 tests put us exactly level with the number of
allocations done by libc++ for those same tests.
To mitigate this to some extent, _M_emplace_uniq detects when the
emplace arguments already contain a key_type (either as the sole
argument, for unordered_set, or as the first part of a pair of
arguments, for unordered_map). In that specific case we don't need to
allocate a node and construct a value type to check for an existing
element with equivalent key.
The remaining regressions in the number of allocations and temporaries
should be addressed separately, with more conservative optimizations
specific to std::string. That is not part of this commit.
libstdc++-v3/ChangeLog:
PR libstdc++/115285
* include/bits/hashtable.h (_Hashtable::_M_emplace): Replace
with _M_emplace_uniq and _M_emplace_multi.
(_Hashtable::_S_forward_key, _Hashtable::_M_insert_unique)
(_Hashtable::_M_insert_unique_aux, _Hashtable::_M_insert):
Remove.
* include/bits/hashtable_policy.h (_ConvertToValueType):
Remove.
(_Insert_base::_M_insert_range): Remove overload for unique keys
and rename overload for non-unique keys to ...
(_Insert_base::_M_insert_range_multi): ... this.
(_Insert_base::insert): Call _M_emplace_uniq or _M_emplace_multi
instead of _M_insert. Add insert overloads from _Insert.
(_Insert_base): Rename to _Insert.
(_Insert): Remove
* testsuite/23_containers/unordered_map/96088.cc: Adjust
expected number of allocations.
* testsuite/23_containers/unordered_set/96088.cc: Likewise.
|
|
Currently the _ReuseOrAllocNode::operator(Args&&...) function always
destroys the value stored in recycled nodes and constructs a new value.
The _ReuseOrAllocNode type is only ever used for implementing
assignment, either from another unordered container of the same type, or
from std::initializer_list<value_type>. Consequently, the parameter pack
Args only ever consists of a single parameter or type const value_type&
or value_type. We can replace the variadic parameter pack with a single
forwarding reference parameter, and when the value_type is assignable
from that type we can use assignment instead of destroying the existing
value and then constructing a new one.
Using assignment is typically only possible for sets, because for maps
the value_type is std::pair<const key_type, mapped_type> and in most
cases std::is_assignable_v<const key_type&, const key_type&> is false.
libstdc++-v3/ChangeLog:
* include/bits/hashtable_policy.h (_ReuseOrAllocNode::operator()):
Replace parameter pack with a single parameter. Assign to
existing value when possible.
* testsuite/23_containers/unordered_multiset/allocator/move_assign.cc:
Adjust expected count of operations.
* testsuite/23_containers/unordered_set/allocator/move_assign.cc:
Likewise.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
This replaces a call to _M_insert_range with open coding the loop. This
will allow removing the node generator parameter from _M_insert_range in
a later commit.
libstdc++-v3/ChangeLog:
* include/bits/hashtable.h (operator=(initializer_list)):
Refactor to not use _M_insert_range.
Reviewed-by: François Dumont <fdumont@gcc.gnu.org>
|
|
The system_time() function used the wrong element of the splits array.
Also add a comment about the units for time measurements.
libstdc++-v3/ChangeLog:
* testsuite/util/testsuite_performance.h (time_counter): Add
comment about times.
(time_counter::system_time): Use correct split value.
|
|
The results of 'make check-performance' are appended to the .sum file,
with no indication where one set of results ends and the next begins. We
could just remove the file when starting a new run, but appending makes
it a little easier to compare with previous runs, without having to copy
and store old files.
This adds a header containing a timestamp to the file when starting a
new run.
libstdc++-v3/ChangeLog:
* scripts/check_performance: Add timestamp to output file at
start of run.
|
|
With recent glibc releases the __gthread_active_p() function is always
true, so we always append "-thread" onto performance benchmark names.
Use the __gnu_cxx::__is_single_threaded() function instead.
libstdc++-v3/ChangeLog:
* testsuite/util/testsuite_performance.h: Use
__gnu_cxx::__is_single_threaded instead of __gthread_active_p().
|
|
This fixes some -Wdeprecated-declarations warnings.
libstdc++-v3/ChangeLog:
* testsuite/performance/ext/pb_ds/hash_int_erase_mem.cc: Replace
std::unary_function with result_type and argument_type typedefs.
* testsuite/util/performance/assoc/multimap_common_type.hpp:
Likewise.
|
|
The use of unnamed std::lock_guard temporaries was intentional here, as
they were used like barriers (but std::barrier isn't available until
C++20). But that gives nodiscard warnings, because unnamed temporary
locks are usually unintentional. Use named variables in new block scopes
instead.
libstdc++-v3/ChangeLog:
* testsuite/performance/20_util/memory_resource/pools.cc: Fix
-Wunused-value warnings about unnamed std::lock_guard objects.
|
|
There are some SVE intrinsics that support one set of suffixes for
one extension (E1, say) and another set of suffixes for another
extension (E2, say). It is usually the case that, mutatis mutandis,
E2 extends E1. Listing E1 first would then ensure that the manual
C overload would also require E1, making it suitable for resolving
both the E1 forms and, where appropriate, the E2 forms.
However, there was one exception: the I8MM, F32MM, and F64MM extensions
to SVE each added variants of svmmla, but there was no svmmla for SVE
itself. This was handled by adding an SVE entry for svmmla that only
defined the C overload; it had no variants of its own.
This situation occurs more often with upcoming patches. Rather than
keep adding these dummy entries, it seemed better to make the code
automatically compute the lowest common denominator for all definitions
that share the same C overload.
gcc/
* config/aarch64/aarch64-protos.h
(aarch64_required_extensions::common_denominator): New member
function.
* config/aarch64/aarch64-sve-builtins-base.def: Remove zero-variant
entry for mmla.
* config/aarch64/aarch64-sve-builtins-shapes.cc (mmla_def): Remove
support for it.
* config/aarch64/aarch64-sve-builtins.cc
(function_builder::add_overloaded): Relax the assert for duplicate
definitions and instead calculate the common denominator of all
requirements.
|
|
We currently support generating vectorized math calls to the AMD core
math library (ACML) (-mveclibabi=acml). That library is end-of-life and
its successor is the math library from AMD Optimizing CPU Libraries
(AOCL).
This patch adds support for AOCL (-mveclibabi=aocl). That significantly
broadens the range of vectorized math functions optimized for AMD CPUs
that GCC can generate calls to.
See the edit to invoke.texi for a complete list of added functions.
Compared to the list of functions in AOCL LibM docs I left out these
vectorized function families:
- sincos and all functions working with arrays ... Because these
functions have pointer arguments and that would require a bigger
rework of ix86_veclibabi_aocl(). Also, I'm not sure if GCC even ever
generates calls to these functions.
- linearfrac ... Because these functions are specific to the AMD
library. There's no equivalent glibc function nor GCC internal
function nor GCC built-in.
- powx, sqrt, fabs ... Because GCC doesn't vectorize these functions
into calls and uses instructions instead.
I also left amd_vrd2_expm1() (the AMD docs list the function but I
wasn't able to link calls to it with the current version of the
library).
gcc/ChangeLog:
PR target/56504
* config/i386/i386-options.cc (ix86_option_override_internal):
Add ix86_veclibabi_type_aocl case.
* config/i386/i386-options.h (ix86_veclibabi_aocl): Add extern
ix86_veclibabi_aocl().
* config/i386/i386-opts.h (enum ix86_veclibabi): Add
ix86_veclibabi_type_aocl into the ix86_veclibabi enum.
* config/i386/i386.cc (ix86_veclibabi_aocl): New function.
* config/i386/i386.opt: Add the 'aocl' type.
* doc/invoke.texi: Document -mveclibabi=aocl.
gcc/testsuite/ChangeLog:
PR target/56504
* gcc.target/i386/vectorize-aocl1.c: New test.
Signed-off-by: Filip Kastl <fkastl@suse.cz>
|
|
2024-11-13 John David Anglin <danglin@gcc.gnu.org>
gcc/ChangeLog:
PR target/117525
* config/pa/pa.md (fix_truncsfsi2): Remove inner `fix:SF`.
(fix_truncdfsi2, fix_truncsfdi2, fix_truncdfdi2,
fixuns_truncsfsi2, fixuns_truncdfsi2, fixuns_truncsfdi2,
fixuns_truncdfdi2): Likewise.
|
|
gcc/analyzer/ChangeLog:
* checker-path.cc (checker_path::debug): Explicitly use
global_dc's reference printer.
* diagnostic-manager.cc
(diagnostic_manager::prune_interproc_events): Likewise.
(diagnostic_manager::prune_system_headers): Likewise.
gcc/ChangeLog:
* diagnostic-path.cc (diagnostic_event::get_desc): Add param
"ref_pp" and use instead of global_dc.
(class path_label): Likewise, adding field m_ref_pp.
(event_range::event_range): Add param "ref_pp" and pass to
m_path_label.
(path_summary::path_summary): Add param "ref_pp" and pass to
event_range ctor.
(diagnostic_text_output_format::print_path): Pass *pp to
path_summary ctor.
(selftest::test_empty_path): Pass *event_pp to pass_summary ctor.
(selftest::test_intraprocedural_path): Likewise.
(selftest::test_interprocedural_path_1): Likewise.
(selftest::test_interprocedural_path_2): Likewise.
(selftest::test_recursion): Likewise.
(selftest::test_control_flow_1): Likewise.
(selftest::test_control_flow_2): Likewise.
(selftest::test_control_flow_3): Likewise.
(selftest::assert_cfg_edge_path_streq): Likewise.
(selftest::test_control_flow_5): Likewise.
(selftest::test_control_flow_6): Likewise.
* diagnostic-path.h (diagnostic_event::get_desc): Add param
"ref_pp".
* lazy-diagnostic-path.cc (selftest::test_intraprocedural_path):
Pass *event_pp to get_desc.
* simple-diagnostic-path.cc (selftest::test_intraprocedural_path):
Likewise.
Signed-off-by: David Malcolm <dmalcolm@redhat.com>
|
|
This patch transforms the following POW calls to equivalent LDEXP calls, as
discussed in PR57492:
powi (powof2, i) -> ldexp (1.0, i * log2 (powof2))
powof2 * ldexp (x, i) -> ldexp (x, i + log2 (powof2))
a * ldexp(1., i) -> ldexp (a, i)
This is especially helpful for SVE architectures as LDEXP calls can be
implemented using the FSCALE instruction, as seen in the following patch:
https://gcc.gnu.org/g:9b2915d95d855333d4d8f66b71a75f653ee0d076
SPEC2017 was run with this patch, while there are no noticeable improvements,
there are no non-noise regressions either.
The patch was bootstrapped and regtested on aarch64-linux-gnu, no regression.
Signed-off-by: Soumya AR <soumyaa@nvidia.com>
gcc/ChangeLog:
PR target/57492
* match.pd: Added patterns to fold calls to pow to ldexp and optimize
specific ldexp calls.
gcc/testsuite/ChangeLog:
PR target/57492
* gcc.dg/tree-ssa/ldexp.c: New test.
* gcc.dg/tree-ssa/pow-to-ldexp.c: New test.
|
|
This patch adds test cases for the Function Multi-Versioning (FMV)
feature for RISC-V, which reuses the existing test cases from the
aarch64 and ported them to RISC-V.
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/testsuite/ChangeLog:
* g++.target/riscv/mv-symbols1.C: New test.
* g++.target/riscv/mv-symbols2.C: New test.
* g++.target/riscv/mv-symbols3.C: New test.
* g++.target/riscv/mv-symbols4.C: New test.
* g++.target/riscv/mv-symbols5.C: New test.
* g++.target/riscv/mvc-symbols1.C: New test.
* g++.target/riscv/mvc-symbols2.C: New test.
* g++.target/riscv/mvc-symbols3.C: New test.
* g++.target/riscv/mvc-symbols4.C: New test.
|
|
TARGET_GET_FUNCTION_VERSIONS_DISPATCHER
This patch implements the TARGET_GENERATE_VERSION_DISPATCHER_BODY and
TARGET_GET_FUNCTION_VERSIONS_DISPATCHER for RISC-V. This is used to
generate the dispatcher function and get the dispatcher function for
function multiversioning.
This patch copies many codes from commit 0cfde688e213 ("[aarch64]
Add function multiversioning support") and modifies them to fit the
RISC-V port. A key difference is the data structure of feature bits in
RISC-V C-API is a array of unsigned long long, while in AArch64 is not
a array. So we need to generate the array reference for each feature
bits element in the dispatcher function.
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* config/riscv/riscv.cc (add_condition_to_bb): New function.
(dispatch_function_versions): New function.
(get_suffixed_assembler_name): New function.
(make_resolver_func): New function.
(riscv_generate_version_dispatcher_body): New function.
(riscv_get_function_versions_dispatcher): New function.
(TARGET_GENERATE_VERSION_DISPATCHER_BODY): Implement it.
(TARGET_GET_FUNCTION_VERSIONS_DISPATCHER): Implement it.
|
|
This patch implements the TARGET_MANGLE_DECL_ASSEMBLER_NAME for RISC-V.
This is used to add function multiversioning suffixes to the assembler
name.
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* config/riscv/riscv.cc
(riscv_mangle_decl_assembler_name): New function.
(TARGET_MANGLE_DECL_ASSEMBLER_NAME): Define.
|
|
TARGET_OPTION_FUNCTION_VERSIONS
This patch implements TARGET_COMPARE_VERSION_PRIORITY and
TARGET_OPTION_FUNCTION_VERSIONS for RISC-V.
The TARGET_COMPARE_VERSION_PRIORITY is implemented to compare the
priority of two function versions based on the rules defined in the
RISC-V C-API Doc PR #85:
https://github.com/riscv-non-isa/riscv-c-api-doc/pull/85/files#diff-79a93ca266139524b8b642e582ac20999357542001f1f4666fbb62b6fb7a5824R721
If multiple versions have equal priority, we select the function with
the most number of feature bits generated by
riscv_minimal_hwprobe_feature_bits. When it comes to the same number of
feature bits, we diff two versions and select the one with the least
significant bit set. Since a feature appears earlier in the feature_bits
might be more important to performance.
The TARGET_OPTION_FUNCTION_VERSIONS is implemented to check whether the
two function versions are the same. This Implementation reuses the code
in TARGET_COMPARE_VERSION_PRIORITY and check it returns 0, which means
the equal priority.
Co-Developed-by: Hank Chang <hank.chang@sifive.com>
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* config/riscv/riscv.cc
(parse_features_for_version): New function.
(compare_fmv_features): New function.
(riscv_compare_version_priority): New function.
(riscv_common_function_versions): New function.
(TARGET_COMPARE_VERSION_PRIORITY): Implement it.
(TARGET_OPTION_FUNCTION_VERSIONS): Implement it.
|
|
This patch implements the TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P for
RISC-V. This hook is used to process attribute
((target_version ("..."))).
As it is the first patch which introduces the target_version attribute,
we also set TARGET_HAS_FMV_TARGET_ATTRIBUTE to 0 to use "target_version"
for function versioning.
Co-Developed-by: Hank Chang <hank.chang@sifive.com>
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* config/riscv/riscv-protos.h
(riscv_process_target_attr): Remove as it is not used.
(riscv_option_valid_version_attribute_p): Declare.
(riscv_process_target_version_attr): Declare.
* config/riscv/riscv-target-attr.cc
(riscv_target_attrs): Renamed from riscv_attributes.
(riscv_target_version_attrs): New attributes for target_version.
(riscv_process_one_target_attr): New arguments to select attrs.
(riscv_process_target_attr): Likewise.
(riscv_option_valid_attribute_p): Likewise.
(riscv_process_target_version_attr): New function.
(riscv_option_valid_version_attribute_p): New function.
* config/riscv/riscv.cc
(TARGET_OPTION_VALID_VERSION_ATTRIBUTE_P): Implement it.
* config/riscv/riscv.h (TARGET_HAS_FMV_TARGET_ATTRIBUTE): Define
it to 0 to use "target_version" for function versioning.
|
|
This patch implements the riscv_minimal_hwprobe_feature_bits feature
for the RISC-V target. The feature bits are defined in the
libgcc/config/riscv/feature_bits.c to provide bitmasks of ISA extensions
that defined in RISC-V C-API. Thus, we need a function to generate the
feature bits for IFUNC resolver to dispatch between different functions
based on the hardware features.
The minimal feature bits means to use the earliest extension appeard in
the Linux hwprobe to cover the given ISA string. To allow older kernels
without some implied extensions probe to run the FMV dispatcher
correctly.
For example, V implies Zve32x, but Zve32x appears in the Linux kernel
since v6.11. If we use isa string directly to generate FMV dispatcher
with functions with "arch=+v" extension, since we have V implied the
Zve32x, FMV dispatcher will check if the Zve32x extension is supported
by the host. If the Linux kernel is older than v6.11, the FMV dispatcher
will fail to detect the Zve32x extension even it already implies by the
V extension, thus making the FMV dispatcher fail to dispatch the correct
function.
Thus, we need to generate the minimal feature bits to cover the given
ISA string to allow the FMV dispatcher to work correctly on older
kernels.
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* common/config/riscv/riscv-common.cc
(RISCV_EXT_BITMASK): New macro.
(struct riscv_ext_bitmask_table_t): New struct.
(riscv_minimal_hwprobe_feature_bits): New function.
* common/config/riscv/riscv-ext-bitmask.def: New file.
* config/riscv/riscv-subset.h (GCC_RISCV_SUBSET_H): Include
riscv-feature-bits.h.
(riscv_minimal_hwprobe_feature_bits): Declare the function.
* config/riscv/riscv-feature-bits.h: New file.
|
|
This patch adds the priority syntax parser to support the Function
Multi-Versioning (FMV) feature in RISC-V. This feature allows users to
specify the priority of the function version in the attribute syntax.
Chnages based on RISC-V C-API PR:
https://github.com/riscv-non-isa/riscv-c-api-doc/pull/85
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* config/riscv/riscv-target-attr.cc
(riscv_target_attr_parser::handle_priority): New function.
(riscv_target_attr_parser::update_settings): Update priority
attribute.
* config/riscv/riscv.opt: Add TargetVariable riscv_fmv_priority.
|
|
Some architectures may use ',' in the attribute string, but it is not
used as the separator for different targets. To avoid conflict, we
introduce a new macro TARGET_CLONES_ATTR_SEPARATOR to separate different
clones.
As an example, according to RISC-V C-API Specification [1], RISC-V allows
',' in the attribute string in the "arch=" option to specify one more
ISA extensions in the same target function, which conflict with the
default separator to separate different clones. This patch introduces
TARGET_CLONES_ATTR_SEPARATOR for RISC-V and choose '#' as the separator,
since '#' is not allowed in the target_clones option string.
[1] https://github.com/riscv-non-isa/riscv-c-api-doc/blob/c6c5d6d9cf96b342293315a5dff3d25e96ef8191/src/c-api.adoc#__attribute__targetattr-string
Signed-off-by: Yangyu Chen <cyy@cyyself.name>
gcc/ChangeLog:
* defaults.h (TARGET_CLONES_ATTR_SEPARATOR): Define new macro.
* multiple_target.cc (get_attr_str): Use
TARGET_CLONES_ATTR_SEPARATOR to separate attributes.
(separate_attrs): Likewise.
(expand_target_clones): Likewise.
* attribs.cc (attr_strcmp): Likewise.
(sorted_attr_string): Likewise.
* tree.cc (get_target_clone_attr_len): Likewise.
* config/riscv/riscv.h (TARGET_CLONES_ATTR_SEPARATOR): Define
TARGET_CLONES_ATTR_SEPARATOR for RISC-V.
* doc/tm.texi: Document TARGET_CLONES_ATTR_SEPARATOR.
* doc/tm.texi.in: Likewise.
|
|
2024-11-14 Paul Thomas <pault@gcc.gnu.org>
gcc/fortran
PR fortran/105054
* resolve.cc (get_temp_from_expr): If the pointer function has
a deferred character length, generate a new deferred charlen
for the temporary.
gcc/testsuite/
PR fortran/105054
* gfortran.dg/ptr_func_assign_6.f08: New test.
|
|
Add warnings for the use of zero as a null pointer constant to the C FE.
PR c/117059
gcc/c-family/ChangeLog:
* c.opt (Wzero-as-null-pointer-constant): Enable for C and ObjC.
gcc/c/ChangeLog:
* c-typeck.cc (parse_build_binary_op): Add warning.
(build_conditional_expr): Add warning.
(convert_for_assignment): Add warning.
gcc/ChangeLog:
* doc/invoke.texi (Wzero-as-null-pointer-constant): Adapt
description.
gcc/testsuite/ChangeLog:
* gcc.dg/Wzero-as-null-pointer-constant.c: New test.
Suggested-by: Alejandro Colomar <alx@kernel.org>
Acked-by: Alejandro Colomar <alx@kernel.org>
Reviewed-by: Joseph Myers <josmyers@redhat.com>
|
|
C23 roughly says that {d,D}{32,64,128} floating point constant suffixes
are alternate spellings of {df,dd,dl} suffixes in annex H.
So, the following patch allows that alternate spelling.
Or is it intentional it isn't enabled and we need to do everything in
there first before trying to define __STDC_IEC_60559_DFP__?
Like add support for _Decimal32x and _Decimal64x types (including
the d32x and d64x suffixes) etc.
2024-11-13 Jakub Jelinek <jakub@redhat.com>
libcpp/
* expr.cc (interpret_float_suffix): Handle d32 and D32 suffixes
for C like df, d64 and D64 like dd and d128 and D128 like
dl.
gcc/c-family/
* c-lex.cc (interpret_float): Subtract 3 or 4 from copylen
rather than 2 if last character of CPP_N_DFLOAT is a digit.
gcc/testsuite/
* gcc.dg/dfp/c11-constants-3.c: New test.
* gcc.dg/dfp/c11-constants-4.c: New test.
* gcc.dg/dfp/c23-constants-3.c: New test.
* gcc.dg/dfp/c23-constants-4.c: New test.
|
|
The following patch implements the C2Y N3298 paper Introduce complex literals
by providing different (or no) diagnostics on imaginary constants (except
for integer ones).
For _DecimalN constants we don't support _Complex _DecimalN and error on any
i/j suffixes mixed with DD/DL/DF, so nothing changed there.
2024-11-13 Jakub Jelinek <jakub@redhat.com>
PR c/117029
libcpp/
* include/cpplib.h (struct cpp_options): Add imaginary_constants
member.
* init.cc (struct lang_flags): Add imaginary_constants bitfield.
(lang_defaults): Add column for imaginary_constants.
(cpp_set_lang): Copy over imaginary_constants.
* expr.cc (cpp_classify_number): Diagnose CPP_N_IMAGINARY
non-CPP_N_FLOATING constants differently for C.
gcc/testsuite/
* gcc.dg/cpp/pr7263-3.c: Adjust expected diagnostic wording.
* gcc.dg/c23-imaginary-constants-1.c: New test.
* gcc.dg/c23-imaginary-constants-2.c: New test.
* gcc.dg/c23-imaginary-constants-3.c: New test.
* gcc.dg/c23-imaginary-constants-4.c: New test.
* gcc.dg/c23-imaginary-constants-5.c: New test.
* gcc.dg/c23-imaginary-constants-6.c: New test.
* gcc.dg/c23-imaginary-constants-7.c: New test.
* gcc.dg/c23-imaginary-constants-8.c: New test.
* gcc.dg/c23-imaginary-constants-9.c: New test.
* gcc.dg/c23-imaginary-constants-10.c: New test.
* gcc.dg/c2y-imaginary-constants-1.c: New test.
* gcc.dg/c2y-imaginary-constants-2.c: New test.
* gcc.dg/c2y-imaginary-constants-3.c: New test.
* gcc.dg/c2y-imaginary-constants-4.c: New test.
* gcc.dg/c2y-imaginary-constants-5.c: New test.
* gcc.dg/c2y-imaginary-constants-6.c: New test.
* gcc.dg/c2y-imaginary-constants-7.c: New test.
* gcc.dg/c2y-imaginary-constants-8.c: New test.
* gcc.dg/c2y-imaginary-constants-9.c: New test.
* gcc.dg/c2y-imaginary-constants-10.c: New test.
* gcc.dg/c2y-imaginary-constants-11.c: New test.
* gcc.dg/c2y-imaginary-constants-12.c: New test.
|
|
This patch uses the FSCALE instruction provided by SVE to implement the
standard ldexp family of functions.
Currently, with '-Ofast -mcpu=neoverse-v2', GCC generates libcalls for the
following code:
float
test_ldexpf (float x, int i)
{
return __builtin_ldexpf (x, i);
}
double
test_ldexp (double x, int i)
{
return __builtin_ldexp(x, i);
}
GCC Output:
test_ldexpf:
b ldexpf
test_ldexp:
b ldexp
Since SVE has support for an FSCALE instruction, we can use this to process
scalar floats by moving them to a vector register and performing an fscale call,
similar to how LLVM tackles an ldexp builtin as well.
New Output:
test_ldexpf:
fmov s31, w0
ptrue p7.b, vl4
fscale z0.s, p7/m, z0.s, z31.s
ret
test_ldexp:
sxtw x0, w0
ptrue p7.b, vl8
fmov d31, x0
fscale z0.d, p7/m, z0.d, z31.d
ret
This is a revision of an earlier patch, and now uses the extended definition of
aarch64_ptrue_reg to generate predicate registers with the appropriate set bits.
The patch was bootstrapped and regtested on aarch64-linux-gnu, no regression.
OK for mainline?
Signed-off-by: Soumya AR <soumyaa@nvidia.com>
gcc/ChangeLog:
PR target/111733
* config/aarch64/aarch64-sve.md
(ldexp<mode>3): Added a new pattern to match ldexp calls with scalar
floating modes and expand to the existing pattern for FSCALE.
* config/aarch64/iterators.md:
(SVE_FULL_F_SCALAR): Added an iterator to match all FP SVE modes as well
as their scalar equivalents.
(VPRED): Extended the attribute to handle GPF_HF modes.
* internal-fn.def (LDEXP): Changed macro to incorporate ldexpf16.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/sve/fscale.c: New test.
|
|
This patch fixs https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117483
If prev and next satisfy the following rules, we should forbid the case
(next.get_sew() < prev.get_sew() && (!next.get_ta() || !next.get_ma()))
in the compatible function max_sew_overlap_and_next_ratio_valid_for_prev_sew_p.
Otherwise, the tail elements of next will be polluted.
DEF_SEW_LMUL_RULE (ge_sew, ratio_and_ge_sew, ratio_and_ge_sew,
max_sew_overlap_and_next_ratio_valid_for_prev_sew_p,
always_false, use_max_sew_and_lmul_with_next_ratio)
Passed the rv64gcv full regression test.
Signed-off-by: Li Xu <xuli1@eswincomputing.com>
PR target/117483
gcc/ChangeLog:
* config/riscv/riscv-vsetvl.cc: Fix bug.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/pr117483.c: New test.
|
|
This is a rewrite of a patch originally from Xianmiao Qu. Xianmiao
noticed that the costs we compute for LO_SUM expressions was incorrect.
Essentially we costed based solely on the first input to the LO_SUM.
In a LO_SUM, the first input is almost always going to be a REG and thus
isn't interesting. The second argument is almost always going to be
some kind of symbolic operand, which is much more interesting from a
costing standpoint.
The right way to fix this is to sum the cost of the two operands. I've
verified this produces the same code as Xianmiao's Qu's original patch.
This has been tested on rv32 and rv64 in my tester. It missed today's
bootstrap of riscv64 though :( Naturally I'll wait on the pre-commit CI
tester to render a verdict, but I don't expect any problems.
-- From Xianmiao Qu's original submission --
Currently, the cost of the LO_SUM expression is based on
the cost of calculating the first subexpression. When the
first subexpression is a register, the cost result will
be zero. It seems a bit unreasonable for a SET expression
to have a zero cost when its source is LO_SUM. Moreover,
having a cost of zero for the expression will lead the
loop invariant pass to calculate its benefits of being
moved outside the loop as zero, thus preventing the
out-of-loop placement of the loop invariant.
As an example, consider the following test case:
long a;
long b[];
long *c;
foo () {
for (;;)
*c = b[a];
}
When compiling with -march=rv64gc -mabi=lp64d -Os, the following code is
generated:
.cfi_startproc
lui a5,%hi(c)
ld a4,%lo(c)(a5)
lui a2,%hi(b)
lui a1,%hi(a)
.L2:
ld a5,%lo(a)(a1)
addi a3,a2,%lo(b)
slli a5,a5,3
add a5,a5,a3
ld a5,0(a5)
sd a5,0(a4)
j .L2
After adjust the cost of the LO_SUM expression, the instruction addi will be
moved outside the loop:
.cfi_startproc
lui a5,%hi(c)
ld a3,%lo(c)(a5)
lui a4,%hi(b)
lui a2,%hi(a)
addi a4,a4,%lo(b)
.L2:
ld a5,%lo(a)(a2)
slli a5,a5,3
add a5,a5,a4
ld a5,0(a5)
sd a5,0(a3)
j .L2
gcc/
* config/riscv/riscv.cc (riscv_rtx_costs): Correct costing of LO_SUM
expressions.
Co-authored-by: Jeff Law <jlaw@ventanamicro.com>
|
|
This reverts commit de3b277247ce98d189f121155b75f490725a42f6.
|
|
-maddress-mode=long. [PR 117418]
-maddress-mode=long let Pmode = DI_mode, so zero extend 32-bit address to
64-bit and uses a 64-bit register as a pointer for avoid raise an ICE.
gcc/ChangeLog:
PR target/117418
* config/i386/i386-expand.cc (ix86_expand_builtin): Convert
pointer's mode according to Pmode.
gcc/testsuite/ChangeLog:
PR target/117418
* gcc.target/i386/pr117418-1.c: New test.
|
|
|
|
This reverts commit 69bd93c167fefbdff0cb88614275358b7a2b2941.
|
|
The target-attr-norelax.c testcase was failing due to the redundant "\t"
check in the assembly output, and forgot to skip the check for lto build
in the testcase.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/target-attr-norelax.c: Fix testcase.
|
