nptx.c (nvptx_option_override): Only set flag_toplevel_reorder, if not explicitly specified.

* config/nvptx/nptx.c (nvptx_option_override): Only set
flag_toplevel_reorder, if not explicitly specified.  Set
flag_no_common, unless explicitly specified.

testsuite/
* gcc.target/nvptx/uninit-decl.c: Force common storage,  add
non-common cases.
* gcc.dg/tree-ssa/ssa-store-ccp-2.c: Add -fcommon.

From-SVN: r236532

jit: implement gcc_jit_rvalue_set_bool_require_tail_call

This implements the libgccjit support for must-tail-call via
a new:
  gcc_jit_rvalue_set_bool_require_tail_call
API entrypoint.

(I didn't implement a wrapper for this within the C++ bindings)

gcc/jit/ChangeLog:
* docs/topics/compatibility.rst: Add LIBGCCJIT_ABI_6.
* docs/topics/expressions.rst (Function calls): Add documentation
of gcc_jit_rvalue_set_bool_require_tail_call.
* docs/_build/texinfo/libgccjit.texi: Regenerate.
* jit-common.h (gcc::jit::recording::base_call): Add forward decl.
* jit-playback.c: Within namespace gcc::jit::playback...
(context::build_call) Add "require_tail_call" param and use it
to set CALL_EXPR_MUST_TAIL_CALL.
(context::new_call): Add "require_tail_call" param.
(context::new_call_through_ptr): Likewise.
* jit-playback.h: Within namespace gcc::jit::playback...
(context::new_call: Add "require_tail_call" param.
(context::new_call_through_ptr): Likewise.
(context::build_call): Likewise.
* jit-recording.c: Within namespace gcc::jit::recording...
(base_call::base_call): New constructor.
(base_call::write_reproducer_tail_call): New method.
(call::call): Update for inheritance from base_call.
(call::replay_into): Provide m_require_tail_call to call
to new_call.
(call::write_reproducer): Call write_reproducer_tail_call.
(call_through_ptr::call_through_ptr): Update for inheritance from
base_call.
(call_through_ptr::replay_into): Provide m_require_tail_call to call
to new_call_through_ptr.
(recording::call_through_ptr::write_reproducer): Call
write_reproducer_tail_call.
* jit-recording.h: Within namespace gcc::jit::recording...
(rvalue::dyn_cast_base_call): New virtual function.
(class base_call): New subclass of class rvalue.
(class call): Inherit from base_call rather than directly from
rvalue, moving get_precedence and m_args to base_call.
(class call_through_ptr): Likewise.
* libgccjit.c (gcc_jit_rvalue_set_bool_require_tail_call): New
function.
* libgccjit.h
(LIBGCCJIT_HAVE_gcc_jit_rvalue_set_bool_require_tail_call): New
macro.
(gcc_jit_rvalue_set_bool_require_tail_call): New function.
* libgccjit.map (LIBGCCJIT_ABI_6): New.
(gcc_jit_rvalue_set_bool_require_tail_call): Add.

gcc/testsuite/ChangeLog:
* jit.dg/all-non-failing-tests.h: Add
test-factorial-must-tail-call.c.
* jit.dg/test-error-impossible-must-tail-call.c: New test case.
* jit.dg/test-factorial-must-tail-call.c: New test case.

From-SVN: r236531

calls.c: fix warning on targets without REG_PARM_STACK_SPACE

gcc/ChangeLog:
* calls.c (can_implement_as_sibling_call_p): Mark param
reg_parm_stack_space with ATTRIBUTE_UNUSED.

From-SVN: r236527

re PR fortran/71204 (ICE with -O0 in expand_expr_real_1, at expr.c:9651)

PR fortran/71204
* frontend-passes.c (realloc_string_callback): Clear inserted_block
and changed_statement before calling create_var.

* gfortran.dg/pr71204.f90: New test.

From-SVN: r236525

i386.c (ix86_rtx_costs): Use IS_STACK_MODE when calculating cost of standard 80387 constants.

* gcc/config/i386/i386.c (ix86_rtx_costs) <case CONST_DOUBLE>:
Use IS_STACK_MODE when calculating cost of standard 80387 constants.
Fallthru to CONST_VECTOR case to calculate cost of standard SSE
constants.
<case CONST_WIDE_INT>: Calculate cost of (MEM (SYMBOL_REF)).
(ix86_legitimate_constant_p): Use CASE_CONST_SCALAR_INT
and CASE_CONST_ANY.

From-SVN: r236524

nvptx.md (sincossf3): New pattern.

gcc/
* config/nvptx/nvptx.md (sincossf3): New pattern.

gcc/testsuite/
* gcc.target/nvptx/sincos.c: New test.

From-SVN: r236523

re PR c++/70572 (ICE on code with decltype (auto) in digest_init_r, at cp/typeck2.c:1103 with -std=c++14)

/cp
2016-05-20  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/70572
* decl.c (cp_finish_decl): Check do_auto_deduction return value
and return immediately in case of erroneous code.

/testsuite
2016-05-20  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/70572
* g++.dg/cpp1y/auto-fn31.C: New.

From-SVN: r236522

Add -mgeneral-regs-only option

X86 Linux kernel is compiled only with integer instructions.  Currently,

-mno-sse -mno-mmx -mno-sse2 -mno-3dnow -mno-avx -mno-80387
-mno-fp-ret-in-387  -mskip-rax-setup

is used to compile kernel.  If we add another non-integer feature, it
has to be turned off.  We can add a -mgeneral-regs-only option, similar
to AArch64, to disable all non-integer features so that kernel doesn't
need a long list and the same option will work for future compilers.
It can also be used to compile interrupt handler.

gcc/

PR target/70738
* common/config/i386/i386-common.c
(OPTION_MASK_ISA_GENERAL_REGS_ONLY_UNSET): New.
(ix86_handle_option): Disable MPX, MMX, SSE and x87 instructions
for -mgeneral-regs-only.
* config/i386/i386.c (ix86_option_override_internal): Don't
enable x87 instructions if only the general registers are
allowed.
* config/i386/i386.opt: Add -mgeneral-regs-only.
* doc/invoke.texi: Document -mgeneral-regs-only.

gcc/testsuite/

PR target/70738
* gcc.target/i386/pr70738-1.c: Likewise.
* gcc.target/i386/pr70738-2.c: Likewise.
* gcc.target/i386/pr70738-3.c: Likewise.
* gcc.target/i386/pr70738-4.c: Likewise.
* gcc.target/i386/pr70738-5.c: Likewise.
* gcc.target/i386/pr70738-6.c: Likewise.
* gcc.target/i386/pr70738-7.c: Likewise.
* gcc.target/i386/pr70738-8.c: Likewise.
* gcc.target/i386/pr70738-9.c: Likewise.

From-SVN: r236520

This patch adds support for the signed and unsigned int versions of the...

This patch adds support for the signed and unsigned int versions of the
vec_addec altivec builtins from the Power Architecture 64-Bit ELF V2 ABI
OpenPOWER ABI for Linux Supplement (16 July 2015 Version 1.1). There are
many of the builtins that are missing and this is part of a series
of patches to add them.

There aren't instructions for the int versions of vec_addec so the
output code is built from other built-ins that do have instructions
which in this case is the following.

vec_addec (va, vb, carryv) == vec_or (vec_addc (va, vb),
vec_addc(vec_add(va, vb),
 vec_and (carryv, 0x1)))

The new test cases are executable tests which verify that the generated
code produces expected values. C macros were used so that the same
test case could be used for both the signed and unsigned versions. An
extra executable test case is also included to ensure that the modified
support for the __int128 versions of vec_addec is not broken. The same
test case could not be used for both int and __int128 because of some
differences in loading and storing the vectors.

Bootstrapped and tested on powerpc64le-unknown-linux-gnu and
powerpc64-unknown-linux-gnu with no regressions. Is this ok for trunk?

[gcc]

2016-05-20  Bill Seurer  <seurer@linux.vnet.ibm.com>

* config/rs6000/rs6000-builtin.def (vec_addec): Change vec_addec to a
special case builtin.
* config/rs6000/rs6000-c.c (altivec_resolve_overloaded_builtin): Add
support for ALTIVEC_BUILTIN_VEC_ADDEC.
* config/rs6000/rs6000.c (altivec_init_builtins): Add definition
for __builtin_vec_addec.

[gcc/testsuite]

2016-05-20  Bill Seurer  <seurer@linux.vnet.ibm.com>

* gcc.target/powerpc/vec-addec.c: New test.
* gcc.target/powerpc/vec-addec-int128.c: New test.

From-SVN: r236515

Implement CALL_EXPR_MUST_TAIL_CALL

This patch implements support for marking CALL_EXPRs
as being mandatory for tail-call-optimization. expand_call
tries harder to perform the optimization on such CALL_EXPRs,
and issues an error if it fails.

Currently this flag isn't accessible from any frontend,
so the patch uses a plugin for testing the functionality.

gcc/ChangeLog:
* calls.c (maybe_complain_about_tail_call): New function.
(initialize_argument_information): Call
maybe_complain_about_tail_call when clearing *may_tailcall.
(can_implement_as_sibling_call_p): Call
maybe_complain_about_tail_call when returning false.
(expand_call): Read CALL_EXPR_MUST_TAIL_CALL and, if set,
ensure try_tail_call is set.  Call maybe_complain_about_tail_call
if tail-call optimization fails.
* cfgexpand.c (expand_call_stmt): Initialize
CALL_EXPR_MUST_TAIL_CALL from gimple_call_must_tail_p.
* gimple-pretty-print.c (dump_gimple_call): Dump
gimple_call_must_tail_p.
* gimple.c (gimple_build_call_from_tree): Call
gimple_call_set_must_tail with the value of
CALL_EXPR_MUST_TAIL_CALL.
* gimple.h (enum gf_mask): Add GF_CALL_MUST_TAIL_CALL.
(gimple_call_set_must_tail): New function.
(gimple_call_must_tail_p): New function.
* print-tree.c (print_node): Update printing of TREE_STATIC
to reflect its use for CALL_EXPR_MUST_TAIL_CALL.
* tree-core.h (struct tree_base): Add MUST_TAIL_CALL to the
trailing comment listing applicable flags.
* tree.h (CALL_EXPR_MUST_TAIL_CALL): New macro.

gcc/testsuite/ChangeLog:
* gcc.dg/plugin/must-tail-call-1.c: New test case.
* gcc.dg/plugin/must-tail-call-2.c: New test case.
* gcc.dg/plugin/must_tail_call_plugin.c: New file.
* gcc.dg/plugin/plugin.exp (plugin_test_list): Add the above.

From-SVN: r236514

Introduce can_implement_as_sibling_call_p

gcc/ChangeLog:
* calls.c (expand_call): Move "Rest of purposes for tail call
optimizations to fail" to...
(can_implement_as_sibling_call_p): ...this new function, and
split into multiple "if" statements.

From-SVN: r236513

cfgloop.h (expected_loop_iterations_unbounded, [...]): Unconstify.

* cfgloop.h (expected_loop_iterations_unbounded,
expected_loop_iterations): Unconstify.
* cfgloopanal.c (expected_loop_iterations_unbounded): Sanity check the
profile with known upper bound; return 3 when profile is absent.
(expected_loop_iterations): Update.

From-SVN: r236511

loop-doloop.c (doloop_optimize): Use get_estimated_loop_iterations_int and get_max_loop_iterations_int.

* loop-doloop.c (doloop_optimize): Use get_estimated_loop_iterations_int
and get_max_loop_iterations_int.

From-SVN: r236509

prefetch-5.c: xfail.

* gcc.dg/tree-ssa/prefetch-5.c: xfail.
* tree-ssa-loop-niter.c (idx_infer_loop_bounds): We can not produce
realistic upper bounds here.

From-SVN: r236508

re PR libstdc++/71081 (experimental/memory_resource/1.cc run for targets without atomics)

2016-05-20  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    PR libstdc++/71081
    * testsuite/experimental/memory_resource/1.cc: Add required argument
    to dg-require-atomic-builtins.

From-SVN: r236507

re PR c++/71210 (internal compiler error: in assign_temp, at function.c:961)

PR c++/71210
* gimple-fold.c (gimple_fold_call): Do not remove lhs of noreturn
calls if the LHS is variable length or has addressable type.
If targets[0]->decl is a noreturn call with void return type and
zero arguments, adjust fntype and remove lhs in that case.

* g++.dg/opt/pr71210-1.C: New test.
* g++.dg/opt/pr71210-2.C: New test.

From-SVN: r236506

re PR target/29756 (SSE intrinsics hard to use without redundant temporaries appearing)

PR tree-optimization/29756
gcc.dg/tree-ssa/vector-6.c: Add -Wno-psabi -w to dg-options.
Add -msse2 for x86 and -maltivec for powerpc.  Use scan-tree-dump-times
only on selected targets where V4SImode vectors are known to be
supported.

From-SVN: r236505

PR71206: inconsistent types after match.pd transformation

2016-05-20  Marc Glisse  <marc.glisse@inria.fr>

PR tree-optimization/71079
PR tree-optimization/71206
gcc/
* match.pd ((X ^ Y) ^ (X ^ Z)): Convert the arguments.

gcc/testsuite/
* gcc.dg/tree-ssa/pr71206.c: New testcase.

From-SVN: r236504

Add test-cases mentioned in r236502.

From-SVN: r236503

tree-vectorizer.c (get_vec_alignment_for_decl): New static function.

2016-05-20  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>

* tree-vectorizer.c (get_vec_alignment_for_decl): New static function.
(get_vec_alignment_for_array_decl): Likewise.
(get_vec_alignment_for_record_decl): Likewise.
(increase_alignment::execute): Move code to find alignment to
get_vec_alignment_for_array_decl and call get_vec_alignment_for_decl.
(type_align_map): New hash_map.

testsuite/
* gcc.dg/vect/section-anchors-vect-70.c: New test-case.
* gcc.dg/vect/section-anchors-vect-71.c: Likewise.
* gcc.dg/vect/section-anchors-vect-72.c: Likewise.

From-SVN: r236502

re PR target/29756 (SSE intrinsics hard to use without redundant temporaries appearing)

2016-05-20  Richard Guenther  <rguenther@suse.de>

PR tree-optimization/29756
* tree.def (BIT_INSERT_EXPR): New tcc_expression tree code.
* expr.c (expand_expr_real_2): Handle BIT_INSERT_EXPR.
* fold-const.c (operand_equal_p): Likewise.
(fold_ternary_loc): Add constant folding of BIT_INSERT_EXPR.
* gimplify.c (gimplify_expr): Handle BIT_INSERT_EXPR.
* tree-inline.c (estimate_operator_cost): Likewise.
* tree-pretty-print.c (dump_generic_node): Likewise.
* tree-ssa-operands.c (get_expr_operands): Likewise.
* cfgexpand.c (expand_debug_expr): Likewise.
* gimple-pretty-print.c (dump_ternary_rhs): Likewise.
* gimple.c (get_gimple_rhs_num_ops): Handle BIT_INSERT_EXPR.
* tree-cfg.c (verify_gimple_assign_ternary): Verify BIT_INSERT_EXPR.

* tree-ssa.c (non_rewritable_lvalue_p): We can rewrite
vector inserts using BIT_FIELD_REF or MEM_REF on the lhs.
(execute_update_addresses_taken): Do it.

* gcc.dg/tree-ssa/vector-6.c: New testcase.

From-SVN: r236501

re PR target/71185 (ice in gimplify_modify_expr, at gimplify.c:4873)

2016-05-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71185
* tree-ssa-loop-prefetch.c (gather_memory_references): Drop
register operations.

* gcc.dg/tree-ssa/pr71185.c: New testcase.

From-SVN: r236499

tree-if-conv.c (add_bb_predicate_gimplified_stmts): Use gimple_seq_add_seq_without_update.

2016-05-20  Richard Biener  <rguenther@suse.de>

* tree-if-conv.c (add_bb_predicate_gimplified_stmts): Use
gimple_seq_add_seq_without_update.
(release_bb_predicate): Assert we have no operands to free.
(if_convertible_loop_p_1): Calculate post dominators later.
Do not free BB predicates here.
(combine_blocks): Do not recompute BB predicates.
(version_loop_for_if_conversion): Save BB predicates around
loop versioning.

* gcc.dg/tree-ssa/ifc-cd.c: Adjust.

From-SVN: r236498

Change ENABLE_VALGRIND_CHECKING to

* config.in: Regenerated.
* configure: Likewise.
* configure.ac: Handle --enable-valgrind-annotations.
* lex.c (new_buff): Use ENABLE_VALGRIND_ANNOTATIONS instead
of ENABLE_VALGRIND_CHECKING.
(_cpp_free_buff): Likewise.

From-SVN: r236496

Daily bump.

From-SVN: r236495

function: Restructure *logue insertion

This patch restructures how the prologues/epilogues are inserted.  Sibcalls
that run without prologue are now handled in shrink-wrap.c; it communicates
what is already handled by setting the EDGE_IGNORE flag.  The
try_shrink_wrapping function then doesn't need to be passed the bb_flags
anymore.

* function.c (make_epilogue_seq): Remove epilogue_end parameter.
(thread_prologue_and_epilogue_insns): Remove bb_flags.  Restructure
code.  Ignore sibcalls on EDGE_IGNORE edges.
* shrink-wrap.c (handle_simple_exit): New function.  Set EDGE_IGNORE
on edges for sibcalls that run without prologue.  The rest of the
function is combined from...
(fix_fake_fallthrough_edge): ... this, and ...
(try_shrink_wrapping): ... a part of this.  Remove the bb_with
function argument, make it a local variable.

From-SVN: r236491

cygming.h (DWARF2_UNWIND_INFO): Allow --disable-sjlj-exceptions for TARGET_BI_ARCH to select DWARF-2 EH for...

2016-05-19  Sandra Loosemore  <sandra@codesourcery.com>

gcc/
* config/i386/cygming.h (DWARF2_UNWIND_INFO): Allow
--disable-sjlj-exceptions for TARGET_BI_ARCH to select DWARF-2 EH
for 32-bit mode and SEH for 64-bit.
* config/i386/mingw32.h (SHARED_LIBGCC_UNDEFS_SPEC): Handle
TARGET_64BIT_DEFAULT.

libgcc/
* config.host [x86_64-*-cygwin*]: Handle tmake_eh_file for mixed
dw2/seh configuration.
[x86_64-*-mingw*]: Likewise.

From-SVN: r236490

PR c/71171: Fix uninitialized source_range in c_parser_postfix_expression

A common way for a c_expr to have an uninitialized src_range is in
error-handling, where the "value" field is set to error_mark_node
without touching the src_range, leading to complaints from valgrind.

This patch introduces a new method c_expr::set_error which sets
the value to error_mark_node whilst initializing the src_range to
UNKNOWN_LOCATION.

This fixes the valgrind issue seen in PR c/71171, along with various
other related issues seen when running the testsuite using the
checker patch I posted here:
  https://gcc.gnu.org/ml/gcc-patches/2015-12/msg00887.html
(this checker still doesn't fully work yet, but it seems to be good
for easily detecting these issues without needing Valgrind).

gcc/c/ChangeLog:
PR c/71171
* c-parser.c (c_parser_generic_selection): Use c_expr::set_error
in error-handling.
(c_parser_postfix_expression): Likewise.
* c-tree.h (c_expr::set_error): New method.
* c-typeck.c (parser_build_binary_op): In error-handling, ensure
that result's range is initialized.

From-SVN: r236488

re PR c++/71075 (Broken diagnostic: 'integer_cst' not supported by 'dump_decl')

PR c++/71075
* pt.c (unify_template_argument_mismatch): Use %qE instead of %qD.

* g++.dg/diagnostic/pr71075.C: New test.

From-SVN: r236487

Fix handling of non-dependent calls with default template args.

PR c++/10200
* pt.c (fn_type_unification): Add outer template args if needed.
(type_unification_real): Handle getting full args.

From-SVN: r236486

PR c++/71184: Fix NULL dereference in cp_parser_operator

The source-range handling for the array form of operator
new/delete erroneously assumed that the "]" was present,
leading to a dereference of NULL when it's absent.

Fix it thusly.

gcc/cp/ChangeLog:
PR c++/71184
* parser.c (cp_parser_operator): For array new/delete, check that
cp_parser_require returned a non-NULL token before dereferencing
it.

gcc/testsuite/ChangeLog:
PR c++/71184
* g++.dg/pr71184.C: New test case.

From-SVN: r236483

Makefile.in (GTFILES): Add cilk.h and cilk-common.c.

2016-05-16  Ryan Burn  <contact@rnburn.com>

  * Makefile.in (GTFILES): Add cilk.h and cilk-common.c.
  * gengtype.c (open_base_files): Add cilk.h to ifiles.

From-SVN: r236482

sched-deps.c (sched_analyze_2): Also force pending loads from memory.

* sched-deps.c (sched_analyze_2) <case TRAP_IF>: Also
force pending loads from memory.

From-SVN: r236481

darn-0.c: New test.

gcc/testsuite/ChangeLog:

2016-05-19  Kelvin Nilsen  <kelvin@gcc.gnu.org>

* gcc.target/powerpc/darn-0.c: New test.
* gcc.target/powerpc/darn-1.c: New test.
* gcc.target/powerpc/darn-2.c: New test.

gcc/ChangeLog:

2016-05-19  Kelvin Nilsen  <kelvin@gcc.gnu.org>

* config/rs6000/altivec.md (UNSPEC_DARN): New unspec constant.
(UNSPEC_DARN_32): New unspec constant.
(UNSPEC_DARN_RAW): New unspec constant.
(darn_32): New instruction.
(darn_raw): New instruction.
(darn): New instruction.
* config/rs6000/rs6000-builtin.def (RS6000_BUILTIN_0): Add
support and documentation for this macro.
(BU_P9_MISC_1): New macro definition.
(BU_P9_64BIT_MISC_0): New macro definition.
(BU_P9_MISC_0): New macro definition.
(darn_32): New builtin definition.
(darn_raw): New builtin definition.
(darn): New builtin definition.
* config/rs6000/rs6000.c: Add #define RS6000_BUILTIN_0 and #undef
RS6000_BUILTIN_0 directives to surround each occurrence of
#include "rs6000-builtin.def".
(rs6000_builtin_mask_calculate): Add in the RS6000_BTM_MODULO and
RS6000_BTM_64BIT flags to the returned mask, depending on
configuration.
(def_builtin): Correct an error in the assignments made to the
debugging variable attr_string.
(rs6000_expand_builtin): Add support for no-operand built-in
functions.
(builtin_function_type): Remove fatal_error assertion that is no
longer valid.
(rs6000_common_init_builtins): Add support for no-operand built-in
functions.
* config/rs6000/rs6000.h (RS6000_BTM_MODULO): New macro
definition.
(RS6000_BTM_PURE): Enhance comment to clarify intent of this flag
definition.
(RS6000_BTM_64BIT): New macro definition.
* doc/extend.texi: Document __builtin_darn (void),
__builtin_darn_raw (void), and __builtin_darn_32 (void) built-in
functions.

From-SVN: r236480

tree-vect-loop.c (vect_analyze_loop_2): Use also max_loop_iterations_int.

* tree-vect-loop.c (vect_analyze_loop_2): Use also
max_loop_iterations_int.

From-SVN: r236478

re PR tree-optimization/71031 (ICE in extract_range_from_binary_expr_1, at tree-vrp.c:2535 w/ -Os)

PR tree-optimization/71031
* tree-vrp.c (extract_range_from_binary_expr_1): Turn assert into a
condition and adjust the code a bit.

* gcc.dg/tree-ssa/vrp100.c: New test.

From-SVN: r236477

Fix memory leak in tree-vect-stmts.c

* tree-vect-stmts.c (vectorizable_simd_clone_call): Utilize
auto_vec instead of vec.

From-SVN: r236472

Fix memory leak in tree-parloops.c

* tree-parloops.c (oacc_entry_exit_ok): Release a vector.

From-SVN: r236471

Fix memory leak in tree-if-conv.c

* tree-if-conv.c (ifcvt_repair_bool_pattern): Utilize auto_vecs.

From-SVN: r236470

Fix memory leak in ipa-pure-const

* ipa-pure-const.c (set_function_state): Remove an existing
funct_state.
(remove_node_data): Do not free it as it's released
in set_function_state.

From-SVN: r236469

Fix memory leak in tree-vect-slp.c

* tree-vect-slp.c (vect_attempt_slp_rearrange_stmts): Release
bitmap.

From-SVN: r236468

Fix memory leak in omp-simd-clone.c

* omp-simd-clone.c (simd_clone_adjust): Release vector.

From-SVN: r236467

Fix leak in gcc/tree-ssa-reassoc.c.

* tree-ssa-reassoc.c (eliminate_duplicate_pair): Truncate
an auto_vec instead of re-creating it.

From-SVN: r236466

Fix leak in tree-ssa-loop-prefetch.c

* tree-ssa-loop-prefetch.c (determine_loop_nest_reuse): Use
auto_vec instead of vec.

From-SVN: r236465

LTO: display file name if LTO version check fails

* lto-section-in.c (lto_get_section_data): Call
lto_check_version with additional argument.
* lto-streamer.c (lto_check_version): Add new argument.
* lto-streamer.h (lto_check_version): Likewise.

From-SVN: r236464

[ARM] Fix costing of sign-extending load in rtx costs

* config/arm/arm.c (arm_new_rtx_costs, SIGN_EXTEND case):
Don't add cost of inner memory when handling sign-extended
loads.

From-SVN: r236461

cse.c (cse_main): Free dominance info.

gcc/

* cse.c (cse_main): Free dominance info.
(rest_of_handle_cse): Don't free dominance info.
(rest_of_handle_cse2): Likewise.
(rest_of_handle_cse_after_global_opts): Likewise.

gcc/testsuite/

* gcc.dg/pr71148.c: New test.

From-SVN: r236460

[ARM] PR target/71056: Don't use vectorized builtins when NEON is not available

PR target/71056
* config/arm/arm-builtins.c (arm_builtin_vectorized_function): Return
NULL_TREE early if NEON is not available.  Remove now redundant check
in ARM_CHECK_BUILTIN_MODE.

* gcc.target/arm/pr71056.c: New test.

From-SVN: r236459

decl.c (finish_enum_value_list): Use the specified mode.

2016-05-19  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        * decl.c (finish_enum_value_list): Use the specified mode.

testsuite:
2016-05-19  Bernd Edlinger  <bernd.edlinger@hotmail.de>

        * c-c++-common/pr69669.c: Check the used mode.

From-SVN: r236458

re PR sanitizer/64354 (no preprocessor symbol __SANITIZE_THREAD__)

2016-05-19  Maxim Ostapenko  <m.ostapenko@samsung.com>

PR sanitizer/64354
* cppbuiltin.c (define_builtin_macros_for_compilation_flags): Add new
builtin __SANITIZE_THREAD__ macros for fsanitize=thread switch.
* doc/cpp.texi: Document new macros.

* c-c++-common/tsan/sanitize-thread-macro.c: New test.

From-SVN: r236457

Move ChangeLog entry to the right file.

PR other/61321
PR other/61233
* demangle.h (enum demangle_component_type)
<DEMANGLE_COMPONENT_CONVERSION>: New value.

From-SVN: r236449

Remove libiberty/ path prefixes.

From-SVN: r236448

re PR tree-optimization/69848 (poor vectorization of a loop from SPEC2006 464.h264ref)

PR tree-optimization/69848
* tree-vect-loop.c (vectorizable_reduction): Don't factor
comparison expr out of VEC_COND_EXPR for COND_REDUCTION.

From-SVN: r236447

re PR c++/70498 (Libiberty Demangler segfaults (3))

PR c++/70498
* cp-demangle.c (d_expression_1): Formatting fix.

From-SVN: r236445

Fix up 2 ChangeLog entries.

From-SVN: r236443

function: A fix for my previous commit

It failed for targets that have an eh_return pattern with a splitter
gated by epilogue_done.

        * function.c (thread_prologue_and_epilogue_insn): Move the
        "goto epilogue_done" one block later.

From-SVN: r236441

re PR tree-optimization/70729 (Loop marked with omp simd pragma is not vectorized)

2016-05-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/70729
* passes.def: Move LIM pass before PRE.  Remove no longer
required copyprop and move first DCE out of the loop pipeline.

* gcc.dg/autopar/outer-6.c: Adjust to avoid redundant store.
* gcc.dg/graphite/scop-18.c: Likewise.
* gcc.dg/pr41783.c: Disable LIM.
* gcc.dg/tree-ssa/loadpre10.c: Likewise.
* gcc.dg/tree-ssa/loadpre23.c: Likewise.
* gcc.dg/tree-ssa/loadpre24.c: Likewise.
* gcc.dg/tree-ssa/loadpre25.c: Likewise.
* gcc.dg/tree-ssa/loadpre4.c: Likewise.
* gcc.dg/tree-ssa/loadpre8.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-16.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-18.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-20.c: Likewise.
* gcc.dg/tree-ssa/ssa-pre-3.c: Likewise.
* gfortran.dg/pr42108.f90: Likewise.

From-SVN: r236440

PR driver/69265: add hint for options with misspelled arguments

opts-common.c's cmdline_handle_error handles invalid arguments
for options with CL_ERR_ENUM_ARG by building a string listing the
valid arguments.  By also building a vec of valid arguments, we
can use find_closest_string and provide a hint if we see a close
misspelling.

gcc/ChangeLog:
PR driver/69265
* Makefile.in (GCC_OBJS): Move spellcheck.o to...
(OBJS-libcommon-target): ...here.
* opts-common.c: Include spellcheck.h.
(cmdline_handle_error): Build a vec of valid options and use it
to suggest provide hints for misspelled arguments.

gcc/testsuite/ChangeLog:
PR driver/69265
* gcc.dg/spellcheck-options-11.c: New test case.

From-SVN: r236439

Daily bump.

From-SVN: r236438

cplus-dem.c (enum type_kind_t): Add tk_rvalue_reference constant.

        * cplus-dem.c (enum type_kind_t): Add tk_rvalue_reference
        constant.
        (demangle_template_value_parm): Handle tk_rvalue_reference
        type kind.
        (do_type): Support 'O' type id (rvalue references).

        * testsuite/demangle-expected: Add tests.

From-SVN: r236434

re PR c++/71100 (Internal compiler error while calling a pointer to member function that throws)

PR c++/71100
* cgraph.c (cgraph_edge::redirect_call_stmt_to_callee): Don't drop
lhs if it has TREE_ADDRESSABLE type.

* g++.dg/opt/pr71100.C: New test.

From-SVN: r236430

re PR target/71145 (Alpha: Error: No lda !gpdisp!278 was found)

PR target/71145
* config/alpha/alpha.md (trap): Add (use (reg:DI 29)).
(*exception_receiver_1): Return "#" for TARGET_EXPLICIT_RELOCS.

From-SVN: r236423

pt.c (value_dependent_expression_p): Tweak new cases to better match the wording in the standard.

* pt.c (value_dependent_expression_p): Tweak new cases to better
match the wording in the standard.

From-SVN: r236421

Allow constant global VAR_DECLs in constant jump functions

2016-05-18  Martin Jambor  <mjambor@suse.cz>

PR ipa/69708
* ipa-cp.c (ipa_get_jf_pass_through_result): Allow non-ip constant
input for NOP_EXPR pass-through functions.
* ipa-prop.c (ipa_compute_jump_functions_for_edge): Allow
aggregate global constant VAR_DECLs in constant jump functions.

testsuite/
* gcc.dg/ipa/iinline-cstagg-2.c: New test.
* gcc.dg/ipa/ipcp-cstagg-5.c: Likewise.
* gcc.dg/ipa/ipcp-cstagg-6.c: Likewise.
* gcc.dg/ipa/ipcp-cstagg-7.c: Likewise.

From-SVN: r236418

Const parameters are always unmodified

2016-05-18  Martin Jambor  <mjambor@suse.cz>

PR ipa/69708
* ipa-prop.c (parm_preserved_before_stmt_p): Return true for loads
from TREE_READONLY parameters.

From-SVN: r236417

Indirect inlining of targets from references of global constants

2016-05-18  Martin Jambor  <mjambor@suse.cz>

PR ipa/69708
* cgraph.h (cgraph_indirect_call_info): New field
guaranteed_unmodified.
* ipa-cp.c (ipa_get_indirect_edge_target_1): Also pass parameter value
to ipa_find_agg_cst_for_param, check guaranteed_unmodified when
appropriate.
* ipa-inline-analysis.c (evaluate_conditions_for_known_args): Also
pass the parameter value to ipa_find_agg_cst_for_param.
* ipa-prop.c (ipa_load_from_parm_agg): New parameter
guaranteed_unmodified, store AA results there instead of bailing out
if present.
(ipa_note_param_call): Also initialize guaranteed_unmodified flag.
(ipa_analyze_indirect_call_uses): Also set guaranteed_unmodified flag.
(find_constructor_constant_at_offset): New function.
(ipa_find_agg_cst_from_init): Likewise.
(ipa_find_agg_cst_for_param): Also seearch for aggregate values in
static initializers of contants, report back through a new paameter
from_global_constant if that was the case.
(try_make_edge_direct_simple_call): Also pass parameter value to
ipa_find_agg_cst_for_param, check guaranteed_unmodified when
appropriate.
(ipa_write_indirect_edge_info): Stream new flag guaranteed_unmodified.
(ipa_read_indirect_edge_info): Likewise.
* ipa-prop.h (ipa_find_agg_cst_for_param): Update declaration.
(ipa_load_from_parm_agg): Likewise.

testsuite/
* gcc.dg/ipa/iinline-cstagg-1.c: New test.
* gcc.dg/ipa/ipcp-cstagg-1.c: Likewise.
* gcc.dg/ipa/ipcp-cstagg-2.c: Likewise.
* gcc.dg/ipa/ipcp-cstagg-3.c: Likewise.
* gcc.dg/ipa/ipcp-cstagg-4.c: Likewise.

From-SVN: r236416

re PR c++/69793 (ICE on invalid code in "cp_lexer_peek_nth_token")

/cp
2016-05-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/69793
* parser.c (cp_parser_template_id): Don't call cp_lexer_peek_nth_token
when the previous cp_lexer_peek_token returns CPP_EOF.

/testsuite
2016-05-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/69793
* g++.dg/template/crash122.C: New.

From-SVN: r236414

re PR target/71161 (Lots of ASAN and libgo runtime FAILs after r236090)

PR target/71161
* elf.c (phdr_callback) [__i386__]: Add
__attribute__((__force_align_arg_pointer__)).

From-SVN: r236397

[Patch, lra] Guard in_class_p with REG_P check

gcc/

PR rtl-optimization/71150
* lra-constraint (process_addr_reg): Guard "in_class_p" with REG_P
check.

From-SVN: r236396

re PR c++/70466 ([ICE on invalid code in tree check: expected constructor, have parm_decl in convert_like_real, at cp/call.c:6371 with -std=c++11)

/cp
2016-05-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/70466
* call.c (convert_like_real): Check that we are actually converting
from an init list.

/testsuite
2016-05-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/70466
* g++.dg/template/pr70466-1.C: New.
* g++.dg/template/pr70466-2.C: Likewise.

From-SVN: r236395

re PR target/70915 (Improve loading 0/-1 in VSX registers on PowerPC)

[gcc]
2016-05-18  Michael Meissner  <meissner@linux.vnet.ibm.com>

PR target/70915
* config/rs6000/constraints.md (wE constraint): New constraint
for a vector constant that can be loaded with XXSPLTIB.
(wM constraint): New constraint for a vector constant of a 1's.
(wS constraint): New constraint for a vector constant that can be
loaded with XXSPLTIB and a vector sign extend instruction.
* config/rs6000/predicates.md (xxspltib_constant_split): New
predicates for wE/wS constraints.
(xxspltib_constant_nosplit): Likewise.
(easy_vector_constant): Add support for constants that can be
loaded via XXSPLTIB.
(all_ones_constant): New predicate for vector constant with all
1's set.
(splat_input_operand): Add support for ISA 3.0 word splat
operations.
* config/rs6000/rs6000.c (xxspltib_constant_p): New function to
return if a constant can be loaded with the ISA 3.0 XXSPLTIB
instruction and possibly with a sign extension.
(output_vec_const_move): Add support for XXSPLTIB. If we are
loading up 0/-1 into Altivec registers, prefer using VSPLTISW
instead of XXLXOR/XXLORC.
(rs6000_expand_vector_init): Add support for ISA 3.0 word splat
operations.
(rs6000_legitimize_reload_address): Likewise.
(rs6000_output_move_128bit): Use output_vec_const_move to emit
constants.
* config/rs6000/vsx.md (VSX_M): Add TImode (if -mvsx-timode) and
combine VSX_M and VSX_M2 into one iterator.
(VSX_M2): Likewise.
(VSINT_84): New iterators for loading constants with XXSPLTIB.
(VSINT_842): Likewise.
(UNSPEC_VSX_SIGN_EXTEND): New UNSPEC.
(xxspltib_v16qi): New insns to load up constants with the ISA 3.0
XXSPLTIB instruction.
(xxspltib_<mode>_nosplit): Likewise.
(xxspltib_<mode>_split): New insn to load up constants with
XXSPLTIB and a sign extend instruction.
(vsx_mov<mode>): Replace single move that handled all vector types
with separate 32-bit and 64-bit moves.  Combine the movti_<bit>
moves (when -mvsx-timode is in effect) into the main vector
moves.  Eliminate separate moves for <VSr> <VSa>, where the
preferred register class (<VSr>) is listed first, and the
secondary register class (<VSa>) is listed second with a '?' to
discourage use.  Prefer loading 0/-1 in any VSX register for ISA
3.0, and Altivec registers for ISA 2.06/2.07 (PR target/70915) so
that if the register was involved in a slow operation, the
clear/set operation does not wait for the slow operation to
finish.  Adjust the length attributes for 32-bit mode.  Use
rs6000_output_move_128bit and drop the use of the string
instructions for 32-bit movti when -mvsx-timode is in effect.  Use
spacing so that the alternatives and attributes don't generate
long lines, and put things in columns, so that it is easier to
match up the operands and attributes with the insn alternatives.
(vsx_mov<mode>_64bit): Likewise.
(vsx_mov<mode>_32bit): Likewise.
(vsx_movti_64bit): Fold movti into normal vector moves.
(vsx_movti_32bit): Likewise.
(vsx_splat_<mode>, V4SI/V4SF modes): Add support for ISA 3.0 word
spat instructions.
(vsx_splat_v4si_internal): Likewise.
(vsx_splat_v4sf_internal): Likewise.
(vector fusion peepholes): Use VSX_M instead of VSX_M2.
(vsx_sign_extend_qi_<mode>): New ISA 3.0 instructions to sign
extend vector elements.
(vsx_sign_extend_hi_<mode>): Likewise.
(vsx_sign_extend_si_v2di): Likewise.
* config/rs6000/rs6000-protos.h (xxspltib_constant_p): Add
declaration.
* doc/md.texi (PowerPC constraints): Document the wE, wM, and wS
constraints.  Add trailing period to wL documentation.

[gcc/testsuite]
2016-05-18  Michael Meissner  <meissner@linux.vnet.ibm.com>

* gcc.target/powerpc/p9-splat-1.c: New tests for ISA 3.0 word
splat operations and the XXSPLTIB instruction.
* gcc.target/powerpc/p9-splat-2.c: Likewise.
* gcc.target/powerpc/p9-splat-3.c: Likewise.
* gcc.target/powerpc/pr47755.c: Allow vspltisw in addition to
xxlxor to clear a register.

From-SVN: r236394

To...

To: gcc-patches@gcc.gnu.org
Subject: PR 71020: Handle abnormal PHIs in tree-call-cdce.c
From: Richard Sandiford <richard.sandiford@arm.com>
Gcc: private.sent
--text follows this line--
The PR is about a case where tree-call-cdce.c causes two abnormal
PHIs for the same variable to be live at the same time, leading to
a coalescing failure.  It seemed like getting rid of these kinds of
input would be generally useful, so I added a utility to tree-dfa.c.

Tested on x86_64-linux-gnu.

gcc/
PR middle-end/71020
* tree-dfa.h (replace_abnormal_ssa_names): Declare.
* tree-dfa.c (replace_abnormal_ssa_names): New function.
* tree-call-cdce.c: Include tree-dfa.h.
(can_guard_call_p): New function, extracted from...
(can_use_internal_fn): ...here.
(shrink_wrap_one_built_in_call_with_conds): Remove failure path
and return void.
(shrink_wrap_one_built_in_call): Likewise.
(use_internal_fn): Likewise.
(shrink_wrap_conditional_dead_built_in_calls): Update accordingly
and return void.  Call replace_abnormal_ssa_names.
(pass_call_cdce::execute): Check can_guard_call_p during the
initial walk.  Assume shrink_wrap_conditional_dead_built_in_calls
will always change something.

gcc/testsuite/
* gcc.dg/torture/pr71020.c: New test.

From-SVN: r236393

Respect --param ipa-max-agg-items=0

2016-05-18  Martin Jambor  <mjambor@suse.cz>

PR ipa/70646
* ipa-prop.c (determine_locally_known_aggregate_parts): Bail out early
if parameter PARAM_IPA_MAX_AGG_ITEMS is zero.

From-SVN: r236390

[PR 70646] Store size to inlining predicate conditions

2016-05-18  Martin Jambor  <mjambor@suse.cz>

PR ipa/70646
* ipa-inline.h (condition): New field size.
* ipa-inline-analysis.c (add_condition): New parameter SIZE, use it
for comaprison and store it into the new condition.
(evaluate_conditions_for_known_args): Use condition size to check
access sizes for all but CHANGED conditions.
(unmodified_parm_1): New parameter size_p, store access size into it.
(unmodified_parm): Likewise.
(unmodified_parm_or_parm_agg_item): Likewise.
(eliminated_by_inlining_prob): Pass NULL to unmodified_parm as size_p.
(set_cond_stmt_execution_predicate): Extract access sizes and store
them to conditions.
(set_switch_stmt_execution_predicate): Likewise.
(will_be_nonconstant_expr_predicate): Likewise.
(will_be_nonconstant_predicate): Likewise.
(inline_read_section): Stream condition size.
(inline_write_summary): Likewise.

testsuite/
* gcc.dg/ipa/pr70646.c: New test.

From-SVN: r236389

vget_lane.c: Add fp16 tests.

* gcc.target/aarch64/advsimd-intrinsics/vget_lane.c: Add fp16 tests.

From-SVN: r236388

vreinterpret.c: Add missing tests for vreinterpretq_p{8,16}.

* gcc.target/aarch64/advsimd-intrinsics/vreinterpret.c: Add missing tests for vreinterpretq_p{8,16}.

From-SVN: r236387

vsli_n.c: Add checks for vsliq_n_s64 and vsliq_n_u64.

* gcc.target/aarch64/advsimd-intrinsics/vsli_n.c: Add checks for vsliq_n_s64 and vsliq_n_u64.

From-SVN: r236385

arm-neon-ref.h (CHECK, [...]): Print which type was checked.

* gcc.target/aarch64/advsimd-intrinsics/arm-neon-ref.h (CHECK,
CHECK_FP, CHECK_CUMULATIVE_SAT): Print which type was checked.

From-SVN: r236384

vmul.c: Remove useless #ifdef.

* gcc.target/aarch64/advsimd-intrinsics/vmul.c: Remove useless #ifdef.
* gcc.target/aarch64/advsimd-intrinsics/vshl.c: Likewise.
* gcc.target/aarch64/advsimd-intrinsics/vtst.c: Likewise.

From-SVN: r236383

vreinterpret.c: Fix typo in comment.

* gcc.target/aarch64/advsimd-intrinsics/vreinterpret.c: Fix typo in comment.

From-SVN: r236382

Fix ChangeLog date.

From-SVN: r236381

noplt_3.c: Scan for "br\t".

* gcc.target/aarch64/noplt_3.c: Scan for "br\t".
* gcc.target/aarch64/tail_indirect_call_1.c: Scan for "br\t",
"blr\t" and switch to scan-assembler-times.

From-SVN: r236377

tree-ssa-loop-im.c (determine_max_movement): Properly add condition cost to PHI cost instead of total_cost.

2016-05-18  Richard Biener  <rguenther@suse.de>

* tree-ssa-loop-im.c (determine_max_movement): Properly add
condition cost to PHI cost instead of total_cost.

From-SVN: r236376

Set DECL_PT_UID for merged variables in IPA ICF (PR70856).

PR fortran/70856
* ipa-icf.c (sem_variable::merge): Set DECL_PT_UID for
merged variables.

From-SVN: r236375

lto-streamer.h (LTO_major_version): Bump to 6.

2016-05-18  Richard Biener  <rguenther@suse.de>

* lto-streamer.h (LTO_major_version): Bump to 6.

From-SVN: r236374

function: Factor out make_*logue_seq

Make new functions make_split_prologue_seq, make_prologue_seq, and
make_epilogue_seq.

* function.c (make_split_prologue_seq, make_prologue_seq,
make_epilogue_seq): New functions, factored out from...
(thread_prologue_and_epilogue_insns): Here.

From-SVN: r236373

function: Do the CLEANUP_EXPENSIVE after shrink-wrapping, not before

We should do CLEANUP_EXPENSIVE after shrink-wrapping, because shrink-
wrapping creates constructs that CLEANUP_EXPENSIVE can optimise, and
nothing runs CLEANUP_EXPENSIVE later.

* function.c (rest_of_handle_thread_prologue_and_epilogue): Call
cleanup_cfg with CLEANUP_EXPENSIVE after shrink-wrapping instead
of before.  Add a comment.

From-SVN: r236372

tree-ssa-loop-ivopts.c (get_computation_cost_at): Check inv expression pointer, not pointer to the pointer.

* tree-ssa-loop-ivopts.c (get_computation_cost_at): Check inv
expression pointer, not pointer to the pointer.

From-SVN: r236371

[testsuite] Guard float64_t with __aarch64__

gcc/testsuite/
        * gcc.target/aarch64/advsimd-intrinsics/arm-neon-ref.h: Guard float64_t
        with __aarch64__.
        * gcc.target/aarch64/advsimd-intrinsics/vfms_vfma_n.c: Guard variable
        declarations under __aarch64__ and __ARM_FEATURE_FMA.

From-SVN: r236370

sse.md (pbroadcast_evex_isa): New mode attr.

* config/i386/sse.md (pbroadcast_evex_isa): New mode attr.
(avx2_pbroadcast<mode>): Add another alternative with v instead
of x constraints in it, using <pbroadcast_evex_isa> isa.
(avx2_pbroadcast<mode>_1): Similarly, add two such alternatives.

* gcc.target/i386/avx512bw-vpbroadcast-1.c: New test.
* gcc.target/i386/avx512bw-vpbroadcast-2.c: New test.
* gcc.target/i386/avx512bw-vpbroadcast-3.c: New test.
* gcc.target/i386/avx512vl-vpbroadcast-1.c: New test.
* gcc.target/i386/avx512vl-vpbroadcast-2.c: New test.
* gcc.target/i386/avx512vl-vpbroadcast-3.c: New test.

From-SVN: r236369

sse.md (<ssse3_avx2>_palignr<mode>): Use constraint x instead of v in second alternative, add avx512bw alternative.

* config/i386/sse.md (<ssse3_avx2>_palignr<mode>): Use
constraint x instead of v in second alternative, add avx512bw
alternative.

* gcc.target/i386/avx512vl-vpalignr-3.c: New test.
* gcc.target/i386/avx512bw-vpalignr-3.c: New test.

From-SVN: r236368

sse.md (<ssse3_avx2>_pshufb<mode>3<mask_name>): Use constraint x instead of v in second alternative, add avx512bw alternative.

* config/i386/sse.md (<ssse3_avx2>_pshufb<mode>3<mask_name>): Use
constraint x instead of v in second alternative, add avx512bw
alternative.

* gcc.target/i386/avx512vl-vpshufb-3.c: New test.
* gcc.target/i386/avx512bw-vpshufb-3.c: New test.

From-SVN: r236367

sse.md (*<ssse3_avx2>_pmulhrsw<mode>3<mask_name>): Use constraint x instead of v in second alternative, add avx512bw alternative.

* config/i386/sse.md (*<ssse3_avx2>_pmulhrsw<mode>3<mask_name>): Use
constraint x instead of v in second alternative, add avx512bw
alternative.

* gcc.target/i386/avx512vl-vpmulhrsw-3.c: New test.
* gcc.target/i386/avx512bw-vpmulhrsw-3.c: New test.

From-SVN: r236366

sse.md (avx2_pmaddubsw256, [...]): Add avx512bw alternative.

* config/i386/sse.md (avx2_pmaddubsw256, ssse3_pmaddubsw128): Add
avx512bw alternative.

* gcc.target/i386/avx512bw-vpmaddubsw-3.c: New test.

From-SVN: r236365

i386. Extend static buffers. Fix SF mode operand constraint to `Yv' in storehpd pattern.

gcc/
        * gcc/config/i386/sse.md (define_insn "*andnot<mode>3"): Extend static
array to 128 chars.
(define_insn "*andnottf3"): Ditto.
(define_insn "*<code><mode>3"/any_logic): Ditto.
(define_insn "*<code>tf3"/any_logic): Ditto.
(define_insn "sse2_storehpd"): Use Yv constraint for scalar
operand to block AVX-512VL insn variant emit when it is not enabled.

From-SVN: r236364

AVX-512. Use `Yv' for SF mode operand in vec_concat pattern.

gcc/
* config/i386/sse.md (define_insn "*vec_concatv2sf_sse4_1"): Use 'Yv'
constraint fot SF mode.

From-SVN: r236363

Fix patterns to enable sse-14.c to compile with -masm=intel.

gcc/
        * config/i386/sse.md (define_insn "srcp14<mode>"): Use proper operand
modifiers.
(define_insn "rsqrt14<mode>"): Ditto.
(define_insn "<mask_codefor>avx512dq_cvtps2qqv2di<mask_name>"): Ditto.
(define_insn "<fixsuffix>fix_truncv2sfv2di2<mask_name>"): Ditto.
(define_insn "avx512f_<code>v8div16qi2_mask_store"): Ditto.
(define_insn "vec_set_hi_<mode><mask_name>"): Ditto.
(define_insn "<mask_codefor>avx512dq_broadcast<mode><mask_name>"):
Ditto.
(define_insn "*avx512f_gatherdi<mode>"): Ditto.
(define_insn "*avx512f_scatterdi<mode>"): Ditto.
* config/i386/i386.c (ix86_print_operand): Expand check for size
override codes for Intel syntax.

Co-Authored-By: Kirill Yukhin <kirill.yukhin@intel.com>
From-SVN: r236362

re PR tree-optimization/71168 (ICE in find_uses_to_rename_use (tree-ssa-loop-manip.c:379))

2016-05-18  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71168
* tree-loop-distribution.c (distribute_loop): Move *destroy_p
initialization earlier.

* gcc.dg/torture/pr71168.c: New testcase.

From-SVN: r236361

[Patch AArch64] Simplify reduc_plus_scal_v2[sd]f sequence

gcc/

* config/aarch64/aarch64-simd.md
(aarch64_reduc_plus_internal<mode>): Rename to...
(reduc_plus_scal): ...This, and remove previous implementation.

From-SVN: r236360

Adding the testcase which was not addaed as part of r236356.

gcc/testsuite/ChangeLog:

2016-05-17  Kugan Vivekanandarajah  <kuganv@linaro.org>

PR middle-end/63586
* gcc.dg/tree-ssa/pr63586-2.c: New test.
* gcc.dg/tree-ssa/pr63586.c: New test.

From-SVN: r236359