With aarch64_track_speculation, we had extra code to do exactly what the
!strong_zero_p path already did. The rest is reducing code duplication.
* config/aarch64/aarch64 (aarch64_split_compare_and_swap): Disable
strong_zero_p for aarch64_track_speculation; unify some code paths;
use aarch64_gen_compare_reg instead of open-coding.
From-SVN: r275966
This pattern will only be used with the __sync functions, because
we do not yet have a bare TImode atomic load.
* config/aarch64/aarch64.c (aarch64_gen_compare_reg): Add support
for NE comparison of TImode values.
(aarch64_emit_load_exclusive): Add support for TImode.
(aarch64_emit_store_exclusive): Likewise.
(aarch64_split_compare_and_swap): Disable strong_zero_p for TImode.
* config/aarch64/atomics.md (@atomic_compare_and_swap<ALLI_TI>):
Change iterator from ALLI to ALLI_TI.
(@atomic_compare_and_swap<JUST_TI>): New.
(@atomic_compare_and_swap<JUST_TI>_lse): New.
(aarch64_load_exclusive_pair): New.
(aarch64_store_exclusive_pair): New.
* config/aarch64/iterators.md (JUST_TI): New.
From-SVN: r275965
2019-09-19 Feng Xue <fxue@os.amperecomputing.com>
* ipa-fnsummary.c (set_cond_stmt_execution_predicate): Do not compute
trivial predicate for condition branch.
(set_switch_stmt_execution_predicate): Do not compute trivial predicate
for switch case.
(compute_bb_predicates): Update predicate based on post-dominating
relationship.
(analyze_function_body): Calculate post-dominating information.
2019-09-19 Feng Xue <fxue@os.amperecomputing.com>
* gcc.dg/ipa/pr91089.c: Add a new function and pattern.
From-SVN: r275963
For conditional reductions, the "then" value needs to be the candidate
value calculated by this iteration while the "else" value needs to be
the result carried over from previous iterations. If the COND_EXPR
is the other way around, we need to swap it.
2019-09-19 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* tree-vectorizer.h (vectorizable_condition): Take an int
reduction index instead of a boolean flag.
* tree-vect-stmts.c (vectorizable_condition): Likewise.
Swap the "then" and "else" values for EXTRACT_LAST_REDUCTION
reductions if the reduction accumulator is the "then" rather
than the "else" value.
(vect_analyze_stmt): Update call accordingly.
(vect_transform_stmt): Likewise.
* tree-vect-loop.c (vectorizable_reduction): Likewise,
asserting that the index is > 0.
From-SVN: r275962
2019-09-19 Richard Biener <rguenther@suse.de>
PR tree-optimization/91812
* tree-ssa-phiprop.c (propagate_with_phi): Do not replace
volatile loads.
* gcc.dg/torture/pr91812.c: New testcase.
From-SVN: r275960
This patch rewrites the way simplify_subreg handles constants.
It uses similar native_encode/native_decode routines to the
tree-level handling of VIEW_CONVERT_EXPR, meaning that we can
move between rtx constants and the target memory image of them.
The main point of this patch is to support subregs of constant-length
vectors for VLA vectors, beyond the very simple cases that were already
handled. Many of the new tests failed before the patch for variable-
length vectors.
The boolean side is tested more by the upcoming SVE ACLE work.
2019-09-19 Richard Sandiford <richard.sandiford@arm.com>
gcc/
* defaults.h (TARGET_UNIT): New macro.
(target_unit): New type.
* rtl.h (native_encode_rtx, native_decode_rtx)
(native_decode_vector_rtx, subreg_size_lsb): Declare.
(subreg_lsb_1): Turn into an inline wrapper around subreg_size_lsb.
* rtlanal.c (subreg_lsb_1): Delete.
(subreg_size_lsb): New function.
* simplify-rtx.c: Include rtx-vector-builder.h
(simplify_immed_subreg): Delete.
(native_encode_rtx, native_decode_vector_rtx, native_decode_rtx)
(simplify_const_vector_byte_offset, simplify_const_vector_subreg): New
functions.
(simplify_subreg): Use them.
(test_vector_subregs_modes, test_vector_subregs_repeating)
(test_vector_subregs_fore_back, test_vector_subregs_stepped)
(test_vector_subregs): New functions.
(test_vector_ops): Call test_vector_subregs for integer vector
modes with at least 2 elements.
From-SVN: r275959
Emit DW_AT_GNU_bias with -fgnat-encodings=gdb. gdb implements this,
but not the encoded variant.
2019-09-19 Tom Tromey <tromey@adacore.com>
gcc/ada/
* gcc-interface/misc.c (gnat_get_type_bias): Return the bias
when -fgnat-encodings=gdb.
gcc/testsuite/
* gnat.dg/bias1.adb: New testcase.
From-SVN: r275958
In cases where pragma Annotate accepts a string literal as an argument,
we now also accept a concatenation of string literals.
2019-09-19 Steve Baird <baird@adacore.com>
gcc/ada/
* sem_prag.adb (Preferred_String_Type): A new function. Given an
expression, determines whether the preference rules defined for
the third-and-later arguments of pragma Annotate suffice to
determine the type of the expression. If so, then the preferred
type is returned; if not then Empty is returned. Handles
concatenations.
(Analyze_Pragma): Replace previous code, which dealt only with
string literals, with calls to the new Preferred_String_Type
function, which also handles concatenations.
* doc/gnat_rm/implementation_defined_pragmas.rst: Update
documentation for pragma Annotate.
* gnat_rm.texi: Regenerate.
gcc/testsuite/
* gnat.dg/annotation1.adb: New testcase.
From-SVN: r275957
This fixes a recent code quality regression for targets that do not
require the strict alignment of memory accesses: the compiler would
generate a useless temporary for a slice of an array component in an
overaligned record type.
Running these commands:
gcc -c p.adb -gnatws -gnatD
grep loop p.adb.dg
On the following sources:
procedure P (N : Positive) is
type Rec1 is record
I : Integer;
end record;
type Arr is array (Positive range <>) of Rec1;
type Rec2 is record
A : Arr (1 .. 128);
end record;
for Rec2'Alignment use 8;
procedure Proc (A : Arr) is
begin
null;
end;
R : Rec2;
begin
Proc (R.A (1 .. N));
end;
Should execute silently.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_util.adb (Is_Possibly_Unaligned_Slice): Do not return true
on pure alignment considerations if the target does not require
the strict alignment of memory accesses.
From-SVN: r275956
This prevents the object code from reading too many bytes from the
source for a copy operation involving a private discriminated record
type with default discriminants and generated for the assignment of an
aggregate to a variable or the initialization of a constant.
The front-end already knows that it needs to convert the operation
involving the aggregate into individual assignments if the type of the
aggregate has mutable components, but it would not do so if this type is
private, which does not change anything for code generation.
Running these commands:
gnatmake -q p -g -fsanitize=address
p
On the following sources:
with Q; use Q;
procedure P is
type Rec is record
A : T;
end record;
C : constant Rec := Rec'(A => Default_T);
begin
null;
end;
package Q is
type T is private;
Default_T : constant T;
private
A : constant := 170;
B : constant := 8192;
type A_Index is range 1 .. A;
type B_Index is range 1 .. B;
type A_Array is array (A_Index) of Boolean;
type B_Array is array (B_Index) of Boolean;
type Data_Type is (A_Type, B_Type);
type T (Discriminant : Data_Type := A_Type) is record
case Discriminant is
when A_Type =>
Field_A : A_Array;
when B_Type =>
Field_B : B_Array;
end case;
end record;
Default_T : constant T :=
T'(Discriminant => A_Type, Field_A => (others => True));
end Q;
Should execute silently.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_aggr.adb (Has_Mutable_Components): Look at the underlying
type of components to find out whether they are mutable.
From-SVN: r275954
This prevents the compiler from issuing a bogus error about a constant
whose full declaration appears too late, if it is declared in a nested
generic package and instantiated in another nested instantiation, when
the instantiations are done in a unit withed from the main unit and
containing an inlined subprogram, and cross-unit inlining is enabled.
It turns out that, under these very peculiar conditions, the compiler
ends up instantiating the body of the generic package twice, which leads
to various semantic errors, in particular for declarations of constants.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* sem_ch12.adb (Instantiate_Package_Body): Check that the body
has not already been instantiated when the body of the parent
was being loaded.
gcc/testsuite/
* gnat.dg/inline21.adb, gnat.dg/inline21_g.ads,
gnat.dg/inline21_h.adb, gnat.dg/inline21_h.ads,
gnat.dg/inline21_q.ads: New testcase.
From-SVN: r275953
This prevents the compiler from issuing a bogus error about the
visibility of an operator in an instantiation of a nested generic
package which is itself used as an actual of an instantiation of another
generic package, when the instantiations are done in a unit withed from
the main unit and containing an inlined subprogram, and cross-unit
inlining is enabled.
In most cases, the compiler does not check the visibility of operators
in an instantiation context because this has already been done when the
generic package has been analyzed. However, there are exceptions like
the actuals of an instantiation of a generic child unit which is done
as a compilation unit and the In_Instance predicate has a special check
for these cases.
This check would incorrectly trigger here and needs to be tightened.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* sem_util.adb (In_Instance): Test whether the current unit has
been analyzed instead of being on the scope stack to detect the
case of actuals of an instantiation of a generic child unit done
as a compilation unit.
gcc/testsuite/
* gnat.dg/inline20.adb, gnat.dg/inline20_g.adb,
gnat.dg/inline20_g.ads, gnat.dg/inline20_h.ads,
gnat.dg/inline20_i.ads, gnat.dg/inline20_q-io.ads,
gnat.dg/inline20_q.ads, gnat.dg/inline20_r.ads: New testcase.
From-SVN: r275952
New routine to create 2 connected sockets. This routine is analog of the
UNIX system call socketpair. On UNIX platforms it is implemented on the
base of socketpair. On other platforms it is implemented by conecting
network sockets over loopback interface.
2019-09-19 Dmitriy Anisimkov <anisimko@adacore.com>
gcc/ada/
* libgnat/g-socket.ads, libgnat/g-socket.adb
(Create_Socket_Pair): New routine.
* libgnat/g-socthi.ads (OS_Has_Socketpair): Boolean constant.
(C_Socketpair): New imported routine.
* libgnat/g-socthi__mingw.ads, libgnat/g-socthi__vxworks.ads
(Default_Socket_Pair_Family): New constant.
(C_Socketpair): New routine.
* libgnat/g-socthi__mingw.adb, libgnat/g-socthi__vxworks.adb
(C_Socketpair): Is separated in anouther file.
* libgnat/g-sthcso.adb (C_Socketpair): Non UNIX implementation.
* libgnat/g-stsifd__sockets.adb: Reuse C_Socketpair.
From-SVN: r275951
A constant of pointer type is considered as mutable in SPARK, according
to SPARK RM 3.10, but this should be based on the declared type of the
constant instead of its underlying type.
There is no impact on compilation hence no test.
2019-09-19 Yannick Moy <moy@adacore.com>
gcc/ada/
* sem_prag.adb (Analyze_Depends_In_Decl_Part): Simplify previous
test.
From-SVN: r275950
2019-09-19 Ed Schonberg <schonberg@adacore.com>
gcc/ada/
* sem_ch7.adb (Install_Parent_Private_Declarations): If a
generic child unit is instantiated within a sibling child unit,
the analysis of its private part requires that the private part
of the ancestor be available, as is done when the context
includes an instance of the ancestor.
gcc/testsuite/
* gnat.dg/generic2-child.ads, gnat.dg/generic2-io_any.adb,
gnat.dg/generic2-io_any.ads, gnat.dg/generic2.ads: New testcase.
From-SVN: r275949
Traversal functions as defined in SPARK RM 3.10 should not be inlined
for analysis in GNATprove, as this changes the ownership behavior.
Disable the inlining performed in GNATprove on functions which could be
interpreted as such.
There is no impact on compilation and thus no test.
2019-09-19 Yannick Moy <moy@adacore.com>
gcc/ada/
* inline.adb (Can_Be_Inlined_In_GNATprove_Mode): Add special
case for traversal functions.
From-SVN: r275948
Now that SPARK supports access types, global constants of access type
may appear as outputs of a subprogram, with the meaning that the
underlying memory can be modified (see SPARK RM 3.10).
2019-09-19 Yannick Moy <moy@adacore.com>
gcc/ada/
* sem_prag.adb (Analyze_Global_In_Decl_Part): Do not issue an
error when a constant of an access type is used as output in a
Global contract.
(Analyze_Depends_In_Decl_Part): Do not issue an error when a
constant of an access type is used as output in a Depends
contract.
gcc/testsuite/
* gnat.dg/global2.adb, gnat.dg/global2.ads: New testcase.
From-SVN: r275947
This fixes a segfault at run time for the call to a local subprogram
through an access value if the type of this access value is derived
from an initial access-to-subprogram type and the access value was
originally obtained with the initial type.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* sem_ch3.adb (Build_Derived_Access_Type): If this is an access-
to-subprogram type, copy Can_Use_Internal_Rep from the parent.
gcc/testsuite/
* gnat.dg/access9.adb: New testcase.
From-SVN: r275945
Unit sem_spark was implementing the borrow-checker for the support of
ownership pointers in SPARK. It has been moved to gnat2why codebase to
facilitate its evolution and allow the more powerful flow analysis to
provide its results for better analysis on pointers.
2019-09-19 Yannick Moy <moy@adacore.com>
gcc/ada/
* gcc-interface/Make-lang.in: Remove references to sem_spark.
* sem_spark.adb, sem_spark.ads: Remove unit.
From-SVN: r275944
GNAT implements Machine_Rounding as an alias for Rounding but, whereas
the implementation of the latter is in line when possible, that of the
former is always out of line, which is not aligned with the intent of
the Ada RM.
This changes the compiler to using for Machine_Rounding the same in line
implementation as Rounding when possible.
Running these commands:
gcc -c f.adb -gnatD
grep system f.adb.dg
On the following sources:
function F (Val : Float) return Integer is
begin
return Integer (Float'Machine_Rounding (Val));
end;
Should execute silently.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_attr.adb (Is_Inline_Floating_Point_Attribute): Treat
Machine_Rounding as an alias for Rounding.
* sem_res.adb (Simplify_Type_Conversion): Likewise.
From-SVN: r275943
In the case of GNAT-LLVM, the GNAT FE no longer does expansion of
up-level references identified by the subprogram unnesting machinery
into activation record references. This is now only done by the FE when
generating C code. This expansion is already taken care of by the
gnat-llvm middle phase, so there's no benefit to also doing it in the
front end.
2019-09-19 Gary Dismukes <dismukes@adacore.com>
gcc/ada/
* exp_unst.adb (Unnest_Subprogram): Bypass the transformation of
up-level references unless Opt.Generate_C_Code is enabled.
From-SVN: r275942
In the general case, the comparison for equality of array objects is
implemented by a local function that contains, among other things, a
loop running over the elements, comparing them one by one and exiting
as soon as an element is not the same in the two array objects.
For the specific case of constrained 2-element arrays, this is rather
heavy and unnecessarily obfuscates the control flow of the program,
so this change implements a simple conjunction of comparisons for it.
Running these commands:
gcc -c p.ads -O -gnatD
grep loop p.ads.dg
On the following sources:
package P is
type Rec is record
Re : Float;
Im : Float;
end record;
type Arr is array (1 .. 2) of Rec;
function Equal (A, B : Arr) return Boolean is (A = B);
end P;
Should execute silently.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_ch4.adb (Expand_Array_Equality): If optimization is
enabled, generate a simple conjunction of comparisons for the
specific case of constrained 1-dimensional 2-element arrays.
Fix formatting.
From-SVN: r275941
Routines Homonym_Number and Get_Homonym_Number were exactly the same,
except for minor style differences. Keep the one in Exp_Util; remove the
one in Exp_Dbug. No test attached, because semantics is unaffected.
2019-09-19 Piotr Trojanek <trojanek@adacore.com>
gcc/ada/
* exp_dbug.ads, exp_dbug.adb (Get_Homonym_Number): Remove.
(Append_Homonym_Number): Use Homonym_Number instead of
Get_Homonym_Number.
* exp_util.ads, exp_util.adb (Homonym_Number): Mirror style of
the removed Get_Homonym_Number routine, i.e. initialize local
objects at declaration and refine the type of result.
* sem_util.adb (Add_Homonym_Suffix): Use Homonym_Number instead
of Get_Homonym_Number.
From-SVN: r275940
This patch fixes a compiler abort on a dynamic predicate applied to the
full view of a type in a generic package declaration, when the
expression for the predicate is a conditionql expression that contains
references to components of the full view of the type.
2019-09-19 Ed Schonberg <schonberg@adacore.com>
gcc/ada/
* sem_ch13.adb (Check_Aspect_At_End_Of_Declarations): Simplify
handling of expressions in predicates when the context is a
generic unit.
gcc/testsuite/
* gnat.dg/predicate14.adb, gnat.dg/predicate14.ads: New
testcase.
From-SVN: r275939
If a for loop starts with "for X in F (...)'Range loop", where F is a
function returning an unconstrained array, then memory is leaked. This
patch fixes that bug.
Running these commands:
gnatmake -q -f main.adb
main
On the following sources:
with Text_IO; use Text_IO;
package P is
function Get_Objects return String;
end P;
package body P is
function Get_Objects return String is
begin
return "xyzzy";
end Get_Objects;
end P;
with Text_IO; use Text_IO;
pragma Warnings (Off, "an internal GNAT unit");
with System.Secondary_Stack;
pragma Warnings (On, "an internal GNAT unit");
with P; use P;
procedure Main is
Max_Iterations : constant Integer := 1_000;
procedure Leak_Call is
begin
for Id in Get_Objects'Range loop
null;
end loop;
end Leak_Call;
procedure SS_Info is new System.Secondary_Stack.SS_Info
(Text_IO.Put_Line);
begin
for Iteration in 1 .. Max_Iterations loop
Leak_Call;
end loop;
SS_Info;
end Main;
Should produce the following output:
Secondary Stack information:
Total size : 10240 bytes
Current allocated space : 0 bytes
Number of Chunks : 1
Default size of Chunks : 10240
2019-09-19 Bob Duff <duff@adacore.com>
gcc/ada/
* sem_attr.adb (Resolve_Attribute): Make sure the secondary
stack is properly managed in the case of a 'Range attribute in a
loop.
From-SVN: r275938
This fixes a spurious type mismatch failure reported between formal and
actual of a call to a subprogram that comes from the instantiation of a
child generic unit that itself contains an instantiation of a slibling
child generic unit, when the parent is itself a generic unit with
private part. The regression was introduced by a recent change made to
clear the Is_Generic_Actual_Type on the implicit full view built when a
generic package is instantiated on a private type.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* sem_ch12.adb (Restore_Private_Views): Comment out new code
that clear the Is_Generic_Actual_Type also on the full view.
gcc/testsuite/
* gnat.dg/generic_inst13.adb,
gnat.dg/generic_inst13_pkg-nested_g.ads,
gnat.dg/generic_inst13_pkg-ops_g.ads,
gnat.dg/generic_inst13_pkg.ads: New testcase.
From-SVN: r275935
This patch fixes a bug where an array object initialized with a
concatenation, and that has an aspect_specification for Alignment,
causes the compiler goes into an infinite loop.
2019-09-19 Bob Duff <duff@adacore.com>
gcc/ada/
* exp_ch3.adb (Rewrite_As_Renaming): Return False if there are
any aspect specifications, because otherwise Insert_Actions
blows up.
gcc/testsuite/
* gnat.dg/concat3.adb: New testcase.
From-SVN: r275934
This fixes a regression introduced by the previous change that improved
the handling of explicit by-reference mechanism. For the very specific
case of a component of a bit-packed array, the front-end still needs to
insert a copy around the call because this is where the rewriting into
the sequence of mask-and-shifts is done for the code generator.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_ch6.adb (Add_Simple_Call_By_Copy_Code): Add
Bit_Packed_Array parameter and documet it. Always insert a copy
if it is set True.
(Expand_Actuals): Adjust the calls to
Add_Simple_Call_By_Copy_Code.
gcc/testsuite/
* gnat.dg/pack26.adb: New testcase.
From-SVN: r275933
This patch fixes a bug in which if a symbol is not found, gnatxref can
sometimes enter an infinite loop. No impact on compilation.
2019-09-19 Bob Duff <duff@adacore.com>
gcc/ada/
* xref_lib.adb (Get_Symbol_Name): If we reach EOF in the first
loop without finding the symbol, return "???". Otherwise, it's
an infinite loop.
(Parse_EOL): Assert that we're not already at EOF. Remove
processing of LF/CR -- there are no operating systems that use
that.
From-SVN: r275932
This improves the handling of an explicit by-reference passing mechanism
specified by means of the GNAT pragma Export_Function. This device sort
of circumvents the rules of the language for the by-reference passing
mechanism and it's then up to the programmer to ensure that the actual
parameter is addressable; if it is not, the compiler will generate a
temporary around the call, thus effectively passing the actual by copy.
It turns out that the compiler was too conservative when determining
whether the actual parameter is addressable, in particular if it's a
component of a record type subject to a representation clause.
The change effectively moves this computation from the front-end to the
back-end, which has much more information on the layout and alignment
of types and thus can be less conservative.
2019-09-19 Eric Botcazou <ebotcazou@adacore.com>
gcc/ada/
* exp_ch6.adb (Is_Legal_Copy): Also return false for an aliased
formal and a formal passed by reference in convention Ada. Add
missing guard to the existing test on Is_Valued_Procedure.
From-SVN: r275931
2019-09-19 Richard Biener <rguenther@suse.de>
* tree-parloops.c (parloops_is_slp_reduction): Do not set
LOOP_VINFO_OPERANDS_SWAPPED.
(parloops_is_simple_reduction): Likewise.
* tree-vect-loop.c (_loop_vec_info::_loop_vec_info): Do not
initialize operands_swapped.
(_loop_vec_info::~_loop_vec_info): Do not re-canonicalize stmts.
(vect_is_slp_reduction): Do not swap operands.
* tree-vectorizer.h (_loop_vec_info::operands_swapped): Remove.
(LOOP_VINFO_OPERANDS_SWAPPED): Likewise.
From-SVN: r275928
Extend pass rpad to handle avx512f vcvtusi2ss vcvtusi2ss
538.imagick_r improved by 4% with single copy run on SKYLAKE workstation.
gcc/
* config/i386/i386.md
(*floatuns<SWI48:mode><MODEF:mode>2_avx512):
Add avx_partial_xmm_update.
gcc/testsuie
* gcc.target/i386/pr87007-3.c: New test.
From-SVN: r275926
PR target/91683
* config/riscv/riscv-protos.h (riscv_split_symbol): New bool parameter.
(riscv_move_integer): Likewise.
* config/riscv/riscv.c (riscv_split_integer): Pass FALSE for new
riscv_move_integer arg.
(riscv_legitimize_move): Likewise.
(riscv_force_temporary): New parameter in_splitter. Don't call
force_reg if true.
(riscv_unspec_offset_high): Pass FALSE for new riscv_force_temporary
arg.
(riscv_add_offset): Likewise.
(riscv_split_symbol): New parameter in_splitter. Pass to
riscv_force_temporary.
(riscv_legitimize_address): Pass FALSE for new riscv_split_symbol
arg.
(riscv_move_integer): New parameter in_splitter. New local
can_create_psuedo. Don't call riscv_split_integer or force_reg when
in_splitter TRUE.
(riscv_legitimize_const_move): Pass FALSE for new riscv_move_integer,
riscv_split_symbol, and riscv_force_temporary args.
* config/riscv/riscv.md (low<mode>+1): Pass TRUE for new
riscv_move_integer arg.
(low<mode>+2): Pass TRUE for new riscv_split_symbol arg.
From-SVN: r275925
This CL serves as part of an initial change for enabling gollvm
building on arm64 linux, the rest of the change will be covered by
another one to the gollvm repo.
Incorporate type definition of 'uint128' to 'runtime' and 'syscall'
packges, the change is not specific to arm64 linux but made available
for all platforms.
Verified by building and unit-testing gollvm on linux x86-64 and arm64.
Verified by building and checking gccgo on linux x86-64 and arm64.
Fixesgolang/go#33711
Change-Id: I4720c7d810cfd4ef720962fb4104c5641b2459c0
From-SVN: r275919
We currently use default mid-end expanders for logical DImode operations.
These split operations without first splitting off complex immediates or
memory operands. The resulting expansions are non-optimal and allow for
fewer LDRD/STRD opportunities. So add back explicit expanders which ensure
memory operands and immediates are handled more efficiently.
gcc/
PR target/91738
* config/arm/arm.md (<logical_op>di3): Expand explicitly.
(one_cmpldi2): Likewise.
* config/arm/arm.c (const_ok_for_dimode_op): Return true if one
of the constant parts is simple.
* config/arm/iterators.md (LOGICAL): Add new code iterator.
(logical_op): Add new code attribute.
(logical_OP): Likewise.
* config/arm/predicates.md (arm_anddi_operand): Add predicate.
(arm_iordi_operand): Add predicate.
(arm_xordi_operand): Add predicate.
From-SVN: r275907
On Skylake, we should move integer register to SSE register without
going through memory. This patch restores Skylake SImode hard register
store cost to 6.
gcc/
PR target/90878
* config/i386/x86-tune-costs.h (skylake_cost): Restore SImode
hard register store cost to 6.
gcc/testsuite/
PR target/90878
* gcc.target/i386/pr90878.c: New test.
From-SVN: r275906
On Skylake, SImode store cost isn't less than half cost of 128-bit vector
store. This patch increases Skylake SImode pseudo register store cost to
make it the same as QImode and HImode.
gcc/
PR target/91446
* config/i386/x86-tune-costs.h (skylake_cost): Increase SImode
pseudo register store cost from 3 to 6 to make it the same as
QImode and HImode.
gcc/testsuite/
PR target/91446
* gcc.target/i386/pr91446.c: New test.
From-SVN: r275905
Cleanup the various highpart multiply patterns using iterators.
As a result the signed and unsigned variants and the pre-Armv6
multiply operand constraints are all handled in a single pattern
and simple expander.
gcc/
* config/arm/arm.md (smulsi3_highpart): Use <US> and <SE> iterators.
(smulsi3_highpart_nov6): Remove pattern.
(smulsi3_highpart_v6): Likewise.
(umulsi3_highpart): Likewise.
(umulsi3_highpart_nov6): Likewise.
(umulsi3_highpart_v6): Likewise.
(<US>mull_high): Add new combined multiply pattern.
From-SVN: r275899
Richard Henderson