re PR tree-optimization/82291 (wrong code at -O3 on x86_64-linux-gnu)

2017-09-22  Richard Biener  <rguenther@suse.de>

PR tree-optimization/82291
* tree-if-conv.c (predicate_mem_writes): Make sure to
remove writes in blocks predicated with false.

* gcc.dg/torture/pr82291.c: New testcase.

From-SVN: r253093

sese.c: Include cfganal.h.

2017-09-22  Richard Biener  <rguenther@suse.de>

* sese.c: Include cfganal.h.
(if_region_set_false_region): Remove.
(create_if_region_on_edge): Likewise.
(move_sese_in_condition): Re-implement without destroying
dominators.

From-SVN: r253092

re PR fortran/52832 ([F03] ASSOCIATE construct with proc-pointer selector is rejected)

2017-09-22  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/52832
* gfortran.dg/associate_31.f90 : Remove failing test.

From-SVN: r253091

graphite-isl-ast-to-gimple.c (translate_pending_phi_nodes): Verify both BBs contain loop PHI nodes before dispatching to copy_loop_phi_args.

2017-09-21  Richard Biener  <rguenther@suse.de>

* graphite-isl-ast-to-gimple.c (translate_pending_phi_nodes):
Verify both BBs contain loop PHI nodes before dispatching to
copy_loop_phi_args.
(graphite_regenerate_ast_isl): Do not recompute dominators,
do not verify three times.  Restructure for clarity.
* graphite-scop-detection.c (same_close_phi_node,
remove_duplicate_close_phi, make_close_phi_nodes_unique,
defined_in_loop_p, canonicalize_loop_closed_ssa,
canonicalize_loop_closed_ssa_form): Simplify, remove excess
checking and SSA rewrite, move to ...
* graphite.c: ... here.  Include ssa.h and tree-ssa-loop-manip.h.
(graphite_initialize): Do not pass in ctx, do not reset the
SCEV cache, compute only dominators.
(graphite_transform_loops): Allocate ISL ctx after
graphite_initialize.  Call canonicalize_loop_closed_ssa_form.
Maintain post-dominators only around build_scops.
* sese.c (if_region_set_false_region): Make static.  Free
and recompute dominators.
(move_sese_in_condition): Assert we don't get called with
post-dominators computed.
* sese.h (if_region_set_false_region): Remove.

From-SVN: r253090

sse.md ("mov<mode>_internal"): Use <sseinsnmode> mode attribute for TARGET_AVX512VL.

/gcc
        * config/i386/sse.md ("mov<mode>_internal"): Use <sseinsnmode>
mode attribute for TARGET_AVX512VL.

* config/i386/i386.opt (mprefer-avx256): New option.
* config/i386/i386.c (ix86_target_string): Add -mprefer-avx256
to flag_opts.
(ix86_preferred_simd_mode): Return 256-bit AVX modes
for TARGET_PREFER_AVX256.
* doc/invoke.texi (x86 Options): Document -mprefer-avx256.

/testsuite

        * gcc.target/i386/avx512f-constant-set.c: New test.

* g++.dg/ext/pr57362.C: Test __attribute__((target("prefer-avx256"))).
* gcc.target/i386/avx512f-prefer.c: New test.

From-SVN: r253089

Daily bump.

From-SVN: r253088

<commit missing files>

2017-09-13  Michael Meissner  <meissner@linux.vnet.ibm.com>

* gcc.target/powerpc/float128-fma1.c: New test.
* gcc.target/powerpc/float128-fma2.c: Likewise.
* gcc.target/powerpc/float128-sqrt1.c: Likewise.
* gcc.target/powerpc/float128-sqrt2.c: Likewise.

From-SVN: r253083

i386.c (ix86_adjust_stack_and_probe_stack_clash): Fix dump output if the only stack space is for pushed registers.

* config/i386/i386.c (ix86_adjust_stack_and_probe_stack_clash):
Fix dump output if the only stack space is for pushed registers.

* lib/target-supports.exp
(check_effective_target_frame_pointer_for_non_leaf): Add
case for x86 Solaris.

From-SVN: r253082

Fix spu build after r253072

2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* config/spu/spu.c (spu_sched_adjust_cost): Update after renaming
of insn_cost.

From-SVN: r253080

re PR go/82284 (go -version segfaults on big endian architectures)

PR go/82284
* elf.c (backtrace_initialize): Set pd.exe_filename.

From-SVN: r253078

re PR fortran/52832 ([F03] ASSOCIATE construct with proc-pointer selector is rejected)

2017-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/52832
* match.c (gfc_match_associate): Before failing the association
try again, allowing a proc pointer selector.

PR fortran/80120
PR fortran/81903
PR fortran/82121
* primary.c (gfc_match_varspec): Introduce 'tgt_expr', which
points to the associate selector, if any. Go through selector
references, after resolution for variables, to catch any full
or section array references. If a class associate name does
not have the same declared type as the selector, resolve the
selector and copy the declared type to the associate name.
Before throwing a no implicit type error, resolve all allowed
selector expressions, and copy the resulting typespec.

PR fortran/67543
* resolve.c (resolve_assoc_var): Selector must cannot be the
NULL expression and it must have a type.

PR fortran/78152
* resolve.c (resolve_symbol): Allow associate names to be
coarrays.

2017-09-21  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/78512
* gfortran.dg/associate_26.f90 : New test.

PR fortran/80120
* gfortran.dg/associate_27.f90 : New test.

PR fortran/81903
* gfortran.dg/associate_28.f90 : New test.

PR fortran/82121
* gfortran.dg/associate_29.f90 : New test.

PR fortran/67543
* gfortran.dg/associate_30.f90 : New test.

PR fortran/52832
* gfortran.dg/associate_31.f90 : New test.

From-SVN: r253077

PR c/81882 - attribute ifunc documentation uses invalid code

gcc/ChangeLog:

PR c/81882
* doc/extend.texi (attribute ifunc): Avoid relying on ill-formed
code (in C++) or code that triggers warnings.

From-SVN: r253076

[demangler PATCH]: Revert and update generic lambda demangling

https://gcc.gnu.org/ml/gcc-patches/2017-09/msg01482.html
PR demangler/82195
* cp-demangle.c (d_name): Revert addition of 'toplevel' parm.
(has_return_type): Recurse for DEMANGLE_COMPONENT_LOCAL_NAME.
(d_encoding): Revert d_name change.  Use is_fnqual_component_type
to strip modifiers that do not belong.
(d_special_name, d_class_enum_type): Revert d_name call change.
(d_expresion_1): Commonize DEMANGLE_COMPONENT_UNARY building.
(d_local_name): Revert parsing of a function type.
(d_print_comp_inner): An inner LOCAL_NAME might contain a
TEMPLATE.
* testsuite/demangle-expected: Add & adjust tests

From-SVN: r253075

* stor-layout.c (bit_from_pos): Do not distribute the conversion.

From-SVN: r253074

re PR testsuite/78421 (vect-strided-a-u8-i2-gap.c fails on armeb)

2017-09-21  Tamar Christina  <tamar.christina@arm.com>

PR testsuite/78421
* lib/target-supports.exp (check_effective_target_vect_hw_misalign):
Invert arm check.

From-SVN: r253073

Rename existing insn_cost to insn_sched_cost

haifa-sched exports an insn_cost function, but it is only used in a
few places and specialised to scheduling.  This patch renames it to
insn_sched_cost.

* haifa-sched.c: Rename insn_cost to insn_sched_cost.
* sched-rgn.c: Ditto.
* sel-sched-ir.c: Ditto.

From-SVN: r253072

openmp.c (gfc_match_oacc_wait): Don't restrict wait directive arguments to constant integers.

gcc/fortran/
* openmp.c (gfc_match_oacc_wait): Don't restrict wait directive
arguments to constant integers.

gcc/testsuite/
* gfortran.dg/goacc/wait.f90: New test.

From-SVN: r253071

toplev: read from /dev/urandom only when needed

* toplev.h (set_random_seed): Adjust return type.
* toplev.c (init_local_tick): Move eager initialization of random_seed
to get_random_seed.  Adjust comment.
(init_random_seed): Inline to get_random_seed, delete.
(get_random_seed): Initialize random_seed lazily.
(set_random_seed): Do not return previous value.
(print_switch_value): Do not call get_random_seed.

From-SVN: r253069

ipa: fix dumping with deleted multiversioning nodes

2017-09-21  Evgeny Kudryashov <kudryashov@ispras.ru>

* cgraph.c (delete_function_version): New, broken out from...
(cgraph_node::delete_function_version): ...here.  Rename to
cgraph_node::delete_function_version_by_decl.  Update all uses.
(cgraph_node::remove): Call delete_function_version.

From-SVN: r253066

re PR sanitizer/81715 (asan-stack=1 redzone allocation is too inflexible)

PR sanitizer/81715
* tree-inline.c (expand_call_inline): Emit clobber stmts for
VAR_DECLs to which addressable non-volatile parameters are mapped
and for id->retvar after the return value assignment.  Clear
id->retval and id->retbnd after inlining.

* g++.dg/tree-ssa/pr8781.C (noop): Change argument type from
const predicate to const predicate & to avoid UB.
* g++.dg/opt/pr81715.C: New test.

From-SVN: r253065

re PR tree-optimization/82276 (-O2: ICE: SSA corruption during RTL pass: expand; at tree-ssa-coalesce.c:1010)

2017-09-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/82276
PR tree-optimization/82244
* tree-vrp.c (build_assert_expr_for): Set
SSA_NAME_OCCURS_IN_ABNORMAL_PHI if the variable we assert on
has it set.
(remove_range_assertions): Revert earlier change.

* gcc.dg/torture/pr82276.c: New testcase.

From-SVN: r253062

PR71951: Fix unwinding with -fomit-frame-pointer

As described in PR71951, if libgcc is built with -fomit-frame-pointer,
unwinding crashes, for example while doing a backtrace.  The underlying
reason is the Dwarf unwinder does not setup the frame pointer register
in the initialization code.  When later unwinding a function that uses
the frame pointer, it tries to read FP using _Unwind_GetGR, and this
crashes if has never restored FP.  To unwind correctly the first frame
must save and restore FP (it is unwound in a special way so that it
uses SP instead of FP).  This is done by adding -fno-omit-frame-pointer.

    gcc/
PR target/71951
* config/aarch64/aarch64.h (LIBGCC2_UNWIND_ATTRIBUTE): Define.

From-SVN: r253061

graphite-isl-ast-to-gimple.c (graphite_regenerate_ast_isl): Restore valid IL after code generation errors.

2017-09-21  Richard Biener  <rguenther@suse.de>

* graphite-isl-ast-to-gimple.c (graphite_regenerate_ast_isl):
Restore valid IL after code generation errors.
* graphite.c (graphite_transform_loops): Diagnose code
generation issues as MSG_MISSED_OPTIMIZATION and continue
with processing SCOPs.

From-SVN: r253060

Make more use of simplify_gen_binary

This patch replaces various places that previously used:

  if (GET_CODE (y) == CONST_INT)
    ... plus_constant (..., x, [-]INTVAL (y)) ...
  else
    ... gen_rtx_PLUS/MINUS (..., x, y) ...

with single calls to simplify_gen_binary.  This allows
them to handle polynomial integers as well as constants.

2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* calls.c (compute_argument_addresses): Use simplify_gen_binary
rather than choosing between plus_constant and gen_rtx_<CODE>.
* expr.c (emit_push_insn): Likewise.
(expand_expr_real_2): Likewise.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r253059

Add copy_rtx call to RTL loop unroller

This is needed if the step is an unshared constant, like many
(const ...)s are.  Without this patch, libgfortran would fail
to build for SVE.

2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* loop-unroll.c (split_iv): Call copy_rtx on the step.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r253058

Fix unguarded uses of tree_to_uhwi

This patch uses tree_fits_uhwi_p to protect a previously unguarded
use of tree_to_uhwi.  Previously we would ICE for variable-sized types.

2017-09-20  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree.c (find_atomic_core_type): Check tree_fits_uhwi_p before
calling tree_to_uhwi.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r253057

Tighten tree-ssa-ccp.c:get_value_for_expr condition

bit_value_unop and bit_value_binop require constant values
to be INTEGER_CSTs:

  gcc_assert ((rval.lattice_val == CONSTANT
               && TREE_CODE (rval.value) == INTEGER_CST)
              || wi::sext (rval.mask, TYPE_PRECISION (TREE_TYPE (rhs))) == -1);

However, when deciding whether to record a constant value,
the for_bits_p handling in get_value_for_expr used a negative
test for ADDR_EXPR:

  else if (is_gimple_min_invariant (expr)
           && (!for_bits_p || TREE_CODE (expr) != ADDR_EXPR))

This patch uses a positive test for INTEGER_CST instead.

Existing tests showed the need for this once polynomial constants
are added.

2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-ssa-ccp.c (get_value_for_expr): Use a positive test for
INTEGER_CST rather than a negative test for ADDR_EXPR.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r253056

Tweak

From-SVN: r253055

Add missing int_cst_rangeN checks to tree-vrp.c

The BIT_AND_EXPR handling in extract_range_from_binary_expr_1
was using value_range_constant_singleton without first checking
whether the range was a constant.  The earlier handling was correctly
guarded:

  /* If either input range contains only non-negative values
     we can truncate the result range maximum to the respective
     maximum of the input range.  */
  if (int_cst_range0 && tree_int_cst_sgn (vr0.min) >= 0)
    wmax = wi::min (wmax, vr0.max, TYPE_SIGN (expr_type));
  if (int_cst_range1 && tree_int_cst_sgn (vr1.min) >= 0)
    wmax = wi::min (wmax, vr1.max, TYPE_SIGN (expr_type));

so this patch uses the same guards again.

Existing tests showed the need for this once polynomial constants
are added.

2017-09-21  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-vrp.c (extract_range_from_binary_expr_1): Check
int_cst_rangeN before calling value_range_constant_singleton (&vrN).

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r253054

re PR tree-optimization/71351 (ICE: Segmentation fault (graphite))

2017-09-21  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71351
* graphite-isl-ast-to-gimple.c (translate_isl_ast_to_gimple::
graphite_create_new_loop_guard): Remove, fold remaining parts
into caller ...
(translate_isl_ast_node_for): ... here and simplify.

* gfortran.dg/graphite/pr71351.f90: New testcase.
* gfortran.dg/graphite/interchange-3.f90: Adjust.

From-SVN: r253052

Fix tests for std::clamp

* testsuite/25_algorithms/clamp/1.cc: Fix order of arguments and
expected results when using predicate defining reverse order.
* testsuite/25_algorithms/clamp/constexpr.cc: Likewise.

From-SVN: r253051

re PR target/82260 ([x86] Unnecessary use of 8-bit registers with -Os.  slightly slower and larger code)

PR target/82260
* config/i386/i386.md (*movqi_internal): Replace (=q,q) alternative
with (=Q,Q), (=R,R) and (=r,r) alternatives, only enable the
latter two for 64-bit, renumber alternatives, for -Os imov (=q,n)
alternative always use QI mode, for -Os imov (=R,R) alternative
always use SI mode, for imov (=Q,Q) or (=r,r) alternatives
ignore -Os.

* gcc.target/i386/pr82260-1.c: New test.
* gcc.target/i386/pr82260-2.c: New test.

From-SVN: r253050

s390.c (MIN_UNROLL_PROBES): Define.

* config/s390/s390.c (MIN_UNROLL_PROBES): Define.
(allocate_stack_space): New function, partially extracted from
s390_emit_prologue.
(s390_emit_prologue): Track offset to most recent stack probe.
Code to allocate space moved into allocate_stack_space.
Dump actions when no stack is allocated.
(s390_prologue_plus_offset): New function.
(s390_emit_stack_probe): Likewise.

* gcc.dg/stack-check-5.c:  Add argument for s390.
* lib/target-supports.exp:
(check_effective_target_supports_stack_clash_protection): Enable for
s390/s390x targets.

Co-Authored-By: Jeff Law <law@redhat.com>
From-SVN: r253049

[C++ PATCH] class member ordering

https://gcc.gnu.org/ml/gcc-patches/2017-09/msg01426.html
* name-lookup.c (member_name_cmp): Use DECL_UID for final
ordering.

From-SVN: r253048

enable handling of -gno- command-line options as negated prefixes

This patch that adds -g to the set of negatable prefixes along with -f,
-m and -W.  Besides the mapping from -gno- to negated -g in option_map
and adding g to the [fmW] matches for negatable options, I had to
introduce gno- as an remapping prefix, for the option searching
machinery to backtrack to and recognize as a remapping prefix, instead
of backtracking to -g and stopping at it as if no-* was its Joined
argument.  Adding such remapping prefixes to preempt further
backtracking can be accomplished by introducing the prefix as an
Undocumented option with a Joined argument and without Driver, Target,
Common, or any language-specific option.  Whenever we match such a fake
options prefix, we abandon further backtracking (it matches, after all),
but find_opt returns the same code it would if it hadn't found any
match, so that we resort to option mapping.

I've arranged for such remapping prefixes to not be considered when
looking for and suggesting a correct spelling for misspelled options.
While testing that, I found a few -W-started options that were not
marked as RejectNegative but should (-Wno-a, is not something we'd like
to suggest ;-)  I've also marked as such -g-started options that
it makes no sense to negate, and removed the explicit -gno- ones,
allowing their opposites to be negated.

for  gcc/ChangeLog

* common.opt (Wa, Wl, Wp, g, gz=): Add
RejectNegative.
(gno-column-info): Remove.
(gcolumn-info): Drop RejectNegative.
(gno-): New prefix.
(gno-record-gcc-switches): Remove.
(grecord-gcc-switches): Drop RejectNegative.
(gno-split-dwarf): Remove.
(gsplit-dwarf): Drop RejectNegative.
(gno-strict-dwarf): Remove.
(gstrict-dwarf): Drop RejectNegative.
* config/darwin.opt (gfull, gused): Add RejectNegative.
* dwarf2out.c (gen_producer_string): Drop
gno-record-gcc-switches handler.
* optc-gen.awk: Add g to prefixes with negative forms.
* opts-common.c (remapping_prefix_p): New.
(find_opt): Check it.
(generate_canonical_option): Test g prefix.
(option_map): Add -gno- mapping.
(add_misspelling_candidates): Check remapping_prefix_p.

for  gcc/ada/ChangeLog

* gcc-interface/lang.opt (gant, gnatO, gnat): Add
        RejectNegative.

for  gcc/c-family/ChangeLog

* c.opt (gen-decls): Add RejectNegative.

From-SVN: r253047

Daily bump.

From-SVN: r253046

libgo: fix build when using -enable-static=no

    With -enable-static=no we don't build non-pic objects, but libgotool.a
    is built from non-pic objects.  Build the packages that go into
    libgotool.a in static mode in all cases.

    Also ensure that internal test packages are built, since nothing
    explicitly depended on them.

    Reviewed-on: https://go-review.googlesource.com/65050

From-SVN: r253042

Fix C++ testsuite fallout from r252976.

gcc/testsuite/ChangeLog:

PR c/81854
* g++.dg/ext/attr-ifunc-2.C: Correct type errors.
* g++.dg/ext/attr-ifunc-4.C: Ditto.
* lib/target-supports.exp: Adjust for C++.

From-SVN: r253041

Fix testsuite fallout from r252976.

gcc/testsuite/ChangeLog:

PR c/81854
* gcc.dg/attr-ifunc-1.c: Correct type errors.
* gcc.dg/attr-ifunc-2.c: Ditto.
* gcc.dg/attr-ifunc-3.c: Ditto.
* gcc.dg/attr-ifunc-4.c: Ditto.
* gcc.dg/attr-ifunc-5.c: Ditto.
* gcc.dg/pr81854.c: Require ifunc support.
* gcc/testsuite/lib/target-supports.exp: Correct type error.

From-SVN: r253037

powerpcspe.c (rs6000_expand_prologue): Fix thinko in stack clash protection support.

* config/powerpcspe/powerpcspe.c (rs6000_expand_prologue): Fix
thinko in stack clash protection support.

From-SVN: r253036

PR libstdc++/79162 Fix std::string regression due to LWG 2946 (old ABI)

PR libstdc++/79162
* include/bits/basic_string.h [!_GLIBCXX_USE_CXX11_ABI]
(basic_string::_If_sv): Remove from the overload set when the
argument is derived from basic_string.

From-SVN: r253035

explow.c (compute_stack_clash_protection_loop_data): Use CONST_INT_P instead of explicit test.

* explow.c (compute_stack_clash_protection_loop_data): Use
CONST_INT_P instead of explicit test.  Verify object is a
CONST_INT_P before looking at INTVAL.
(anti_adjust_stack_and_probe_stack_clash): Use CONST_INT_P
instead of explicit test.

* gcc.target/i386/stack-check-11.c: Update test and regexp
so that it works for both i?86 and x86_64.

From-SVN: r253034

rs6000: Don't touch below the stack pointer (PR77687)

With the 32-bit SVR4 ABI we don't have a red zone, so we have to restore
the callee-saved registers before we restore the stack pointer.

The previous fix for this PR failed in two ways, for huge frames: first,
we use a negative offset from r11 in that case, so the (mem:BLK 11) access
does no good; second, sched does not handle accesses to mem:BLK correctly
in this case (does not make dependencies).

This patch fixes it by doing a store to (mem:BLK (scratch)) instead.
This means no unrelated (not to stack) loads/stores can be moved over the
stack restore either, but so be it.

PR target/77687
* config/rs6000/rs6000.md (stack_restore_tie): Store to a scratch
address instead of to r1 and r11.

gcc/testsuite/
PR target/77687
* gcc.target/powerpc/pr77687.c: New testcase.

From-SVN: r253033

re PR sanitizer/77631 (no symbols in backtrace shown by ASan when debug info is split)

PR sanitizer/77631
Support for external debug info.
* elf.c: Include <errno.h>, <sys/stat.h>, <unistd.h>.
(S_ISLNK): Define if not defined.
(xstrnlen): Define if strnlen is not available.
(b_elf_note): Define type.
(NT_GNU_BUILD_ID): Define macro.
(elf_crc32, elf_crc32_file): New static functions.
(elf_is_symlink, elf_readlink): New static functions.
(elf_open_debugfile_by_buildid): New static function.
(elf_try_debugfile): New static function.
(elf_find_debugfile_by_debuglink): New static function.
(elf_open_debugfile_by_debuglink): New static function.
(elf_add): Add filename and debuginfo parameters.  Adjust all
callers.  Look for external debug info notes, and try to fetch
debug info from external file.
(struct phdr_data): Add exe_filename field.
(phdr_callback): Pass filename to elf_add.
(backtrace_initialize): Add filename parameter.
* internal.h (backtrace_initialize): Add filename parameter.
* fileline.c (fileline_initialize): Pass filename to
backtrace_initialize.
* pecoff.c (fileline_initialize): Add unused filename parameter.
* unknown.c (fileline_initialize): Likewise.
* xcoff.c (fileline_initialize): Likewise.
* configure.ac: Check for objcopy --add-gnu-debuglink.
* Makefile.am (dtest): New test target.
* configure, Makefile.in: Rebuild.

Co-Authored-By: Denis Khalikov <d.khalikov@partner.samsung.com>
From-SVN: r253032

P0409R2 - allow lambda capture [=, this]

P0409R2 - allow lambda capture [=, this]
* parser.c (cp_parser_lambda_introducer): For cxx2a don't pedwarn on
redundant [=, this].

* g++.dg/cpp1z/lambda-this1.C: Don't expect error for c++2a on [=, this]
capture.  Add further tests.
* g++.dg/cpp0x/lambda/lambda-capture-redundancy.C: Don't expect error
for c++2a on [=, this] capture.

From-SVN: r253030

cplusplus.C: Test that __cplusplus is equal to 201703L.

* g++.dg/cpp1z/cplusplus.C: Test that __cplusplus is equal to 201703L.
* g++.dg/cpp1z/cplusplus_1z.C: New test.

From-SVN: r253029

PR libstdc++/79162 Fix std::string regression due to LWG 2946

PR libstdc++/79162
* include/bits/basic_string.h (basic_string::_If_sv): Remove from the
overload set when the argument is derived from basic_string.
* testsuite/21_strings/basic_string/cons/char/moveable2_c++17.cc: New
test.
* testsuite/21_strings/basic_string/cons/wchar_t/moveable2_c++17.cc:
New test.

From-SVN: r253024

syscall: enable ParseDirent for AIX

    Reviewed-on: https://go-review.googlesource.com/64990

From-SVN: r253022

syscall: workaround for getsockname bug in AIX

    Reviewed-on: https://go-review.googlesource.com/64552

From-SVN: r253021

Fix failing C++17 test

* testsuite/24_iterators/range_access_cpp17.cc: Fix order of dg-do
and dg-options directives. Fix invalid test.

From-SVN: r253019

PR libstdc++/81469 deprecate std::uncaught_exception for C++17

PR libstdc++/81469
* libsupc++/exception (uncaught_exception): Deprecate for C++17.
* testsuite/18_support/exception_ptr/62258.cc: Add -Wno-deprecated.
* testsuite/18_support/uncaught_exception/14026.cc: Likewise.

From-SVN: r253018

Deprecate nested types in std::hash

* include/bits/c++config (_GLIBCXX17_DEPRECATED): Define.
* include/bits/functional_hash.h (__hash_base::result_type)
(__hash_base::argument_type): Add _GLIBCXX17_DEPRECATED.
* include/std/optional (hash<optional<T>>::result_type)
(hash<optional<T>>::argument_type): Add deprecated attribute.
(__is_fast_hash<hash<optional<T>>>): Add partial specialization.
* include/std/variant (hash<variant<Types...>>::result_type)
(hash<variant<Types...>>::argument_type): Add deprecated attribute.
(__is_fast_hash<hash<variant<Types...>>>): Add partial specialization.

From-SVN: r253017

internal,net,os,runtime,syscall: fixes for AIX following update to go1.9

    Reviewed-on: https://go-review.googlesource.com/64551

From-SVN: r253016

config.gcc: Support "knm".

gcc/

        * config.gcc: Support "knm".
        * config/i386/driver-i386.c (host_detect_local_cpu): Detect "knm".
        * config/i386/i386-c.c (ix86_target_macros_internal): Handle
        PROCESSOR_KNM.
        * config/i386/i386.c (m_KNM): Define.
        (processor_target_table): Add "knm".
        (PTA_KNM): Define.
        (ix86_option_override_internal): Add "knm".
        (ix86_issue_rate): Add PROCESSOR_KNM.
        (ix86_adjust_cost): Ditto.
        (ia32_multipass_dfa_lookahead): Ditto.
        (get_builtin_code_for_version): Handle PROCESSOR_KNM.
        (fold_builtin_cpu): Add M_INTEL_KNM.
        * config/i386/i386.h (processor_costs): Define TARGET_KNM.
        (processor_type): Add PROCESSOR_KNM.
         * config/i386/x86-tune.def: Add m_KNM.
        * doc/invoke.texi: Add knm as x86 -march=/-mtune= CPU type.

libgcc/
        * config/i386/cpuinfo.h (processor_types): Add INTEL_KNM.
        * config/i386/cpuinfo.c (get_intel_cpu): Detect Knights Mill.

gcc/testsuite/

        * gcc.target/i386/builtin_target.c: Test knm.
        * gcc.target/i386/funcspec-56.inc: Test arch=knm.

From-SVN: r253013

Revert my 2017-09-19 patch.

From-SVN: r253012

Remove non-standard std::copy_exception function

* libsupc++/exception_ptr.h (copy_exception): Remove deprecated
non-standard function.

From-SVN: r253011

PR libstdc++/82262 fix std::hash<std::optional<const T>>

PR libstdc++/82262
* include/std/optional (__optional_hash_call_base): Add template
parameter for remove_const_t<_Tp> and use it consistently.
* testsuite/20_util/optional/hash.cc: Test optional<const T>.

From-SVN: r253010

re PR tree-optimization/80213 (ICE in check_loop_closed_ssa_use, at tree-ssa-loop-manip.c:704)

2017-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/80213
* graphite-scop-detection.c (trivially_empty_bb_p): Labels
are allowed in empty BBs as well.
(canonicalize_loop_closed_ssa): Also look for other complex
edges.
(scop_detection::get_sese): Include the loop-closed PHI block
in loop SESEs.
(scop_detection::merge_sese): Remove code adding extra blocks.
(scop_detection::region_has_one_loop): Adjust for get_sese changes.
(build_scops): Assert the final returned scop is invalid.

From-SVN: r253008

re PR tree-optimization/77362 ([graphite] ICE in sese_build_liveouts_use w/ -O2 -floop-nest-optimize)

2017-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/77362
* gcc.dg/graphite/pr77362.c: New testcase.

From-SVN: r253006

re PR tree-optimization/82264 (ICE in vn_phi_lookup at gcc/tree-ssa-sccvn.c:3125)

2017-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/82264
* tree-ssa-sccvn.c (vn_phi_eq): Use safe_dyn_cast to check
for GIMPLE_CONDs.
(vn_phi_lookup): Likewise.
(vn_phi_insert): Likewise.

* gcc.dg/torture/pr82264.c: New testcase.

From-SVN: r253005

dwarf2out.c (tree_add_const_value_attribute): For INTEGER_CST that fits into uhwi or shwi...

* dwarf2out.c (tree_add_const_value_attribute): For INTEGER_CST
that fits into uhwi or shwi, add DW_AT_const_value regardless
of early_dwarf without going through RTL, using add_AT_unsigned
or add_AT_int.

* g++.dg/debug/dwarf2/template-params-13.C: New test.
* g++.dg/debug/dwarf2/template-params-5.C: Adjust regexps so that
it doesn't match newlines.
* g++.dg/debug/dwarf2/template-params-3.C: Likewise.
* g++.dg/debug/dwarf2/template-func-params-3.C: Likewise.
* g++.dg/debug/dwarf2/lang-cpp98.C: Likewise.
* g++.dg/debug/dwarf2/template-func-params-2.C: Likewise.
* g++.dg/debug/dwarf2/template-func-params-1.C: Likewise.
* g++.dg/debug/dwarf2/template-func-params-5.C: Likewise.
* g++.dg/debug/dwarf2/template-params-1.C: Likewise.
* g++.dg/debug/dwarf2/template-params-2.C: Likewise.
* g++.dg/debug/dwarf2/lang-cpp14.C: Likewise.
* g++.dg/debug/dwarf2/lang-cpp11.C: Likewise.

From-SVN: r253004

dwarf2out.c (DEBUG_LTO_DWO_INFO_SECTION): Reorder defines.

* dwarf2out.c (DEBUG_LTO_DWO_INFO_SECTION): Reorder defines.
(DEBUG_LTO_ABBREV_SECTION): Likewise.
(DEBUG_LTO_MACINFO_SECTION): Likewise.
(DEBUG_MACRO_SECTION): Likewise.
(DEBUG_LTO_MACRO_SECTION): Likewise.
(DEBUG_STR_DWO_SECTION): Likewise.
(DEBUG_LTO_STR_DWO_SECTION): Likewise.
(DEBUG_LTO_LINE_SECTION): Drop .dwo suffix from the name.
(DEBUG_LTO_DWO_LINE_SECTION): Define.
(DEBUG_LTO_LINE_STR_SECTION): Define.
(init_sections_and_labels): Initialize debug_line_str_section
variable.  Initialize debug_loc_section for -gdwarf-5 to
DEBUG_LOCLISTS_SECTION.  Formatting fixes.

From-SVN: r253003

re PR c++/64954 (GCC incorrectly rejects constexpr variable initialization.)

2017-09-20  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/64954
* g++.dg/cpp0x/constexpr-ctor20.C: New.

From-SVN: r253002

graphite-sese-to-poly.c (extract_affine): Properly handle POINTER_PLUS_EXPR, BIT_NOT_EXPR and conversion to signed.

2017-09-20  Richard Biener  <rguenther@suse.de>

* graphite-sese-to-poly.c (extract_affine): Properly handle
POINTER_PLUS_EXPR, BIT_NOT_EXPR and conversion to signed.

From-SVN: r253001

re PR tree-optimization/81373 (Graphite ICE in ssa_default_def at gcc/tree-dfa.c:305)

2017-09-20  Richard Biener  <rguenther@suse.de>

PR tree-optimization/81373
* graphite-scop-detection.c (build_cross_bb_scalars_def):
Force SESE live-out defs to be handled even if they are
scev_analyzable_p.

* gcc.dg/graphite/pr81373.c: New testcase.

From-SVN: r253000

combine-stack-adj.c (combine_stack_adjustments_for_block): Do nothing for stack adjustments with REG_STACK_CHECK.

* combine-stack-adj.c (combine_stack_adjustments_for_block): Do
nothing for stack adjustments with REG_STACK_CHECK.
* sched-deps.c (parse_add_or_inc): Reject insns with
REG_STACK_CHECK from dependency breaking.
* config/i386/i386.c (pro_epilogue_adjust_stack): Return insn.
(ix86_adjust_satck_and_probe_stack_clash): Add REG_STACK_NOTEs.
* reg-notes.def (STACK_CHECK): New note.

* gcc.target/i386/stack-check-11.c: New test.

From-SVN: r252999

i386.c (ix86_adjust_stack_and_probe_stack_clash): New.

* config/i386/i386.c (ix86_adjust_stack_and_probe_stack_clash): New.
(ix86_expand_prologue): Dump stack clash info as needed.
Call ix86_adjust_stack_and_probe_stack_clash as needed.

* gcc.dg/stack-check-4.c: New test.
* gcc.dg/stack-check-5.c: New test.
* gcc.dg/stack-check-6.c: New test.
* gcc.dg/stack-check-6a.c: New test.
* gcc.dg/stack-check-7.c: New test.
* gcc.dg/stack-check-8.c: New test.
* gcc.dg/stack-check-9.c: New test.
* gcc.dg/stack-check-10.c: New test.
* lib/target-supports.exp
(check_effective_target_supports_stack_clash_protection): Enable for
x86 and x86_64 targets.

From-SVN: r252998

function.c (dump_stack_clash_frame_info): New function.

        * function.c (dump_stack_clash_frame_info): New function.
        * function.h (dump_stack_clash_frame_info): Prototype.
        (enum stack_clash_probes): New enum.

From-SVN: r252997

alpha.c (alpha_expand_prologue): Also check flag_stack_clash_protection.

* config/alpha/alpha.c (alpha_expand_prologue): Also check
flag_stack_clash_protection.
* config/arm/arm.c (arm_compute_static_chain_stack_bytes): Likewise.
(arm_expand_prologue, thumb1_expand_prologue): Likewise.
(arm_frame_pointer_required): Likewise.
* config/ia64/ia64.c (ia64_compute_frame_size): Likewise.
(ia64_expand_prologue): Likewise.
* config/mips/mips.c (mips_expand_prologue): Likewise.
* config/powerpcspe/powerpcspe.c (rs6000_expand_prologue): Likewise.
* config/sparc/sparc.c (sparc_expand_prologue): Likewise.
(sparc_flat_expand_prologue): Likewise.
* config/spu/spu.c (spu_expand_prologue): Likewise.

From-SVN: r252996

explow.c: Include "params.h".

2017-09-18  Jeff Law  <law@redhat.com>

* explow.c: Include "params.h".
(anti_adjust_stack_and_probe_stack_clash): New function.
(get_stack_check_protect): Likewise.
(compute_stack_clash_protection_loop_data): Likewise.
(emit_stack_clash_protection_loop_start): Likewise.
(emit_stack_clash_protection_loop_end): Likewise.
(allocate_dynamic_stack_space): Use get_stack_check_protect.
Use anti_adjust_stack_and_probe_stack_clash.
* explow.h (compute_stack_clash_protection_loop_data): Prototype.
(emit_stack_clash_protection_loop_start): Likewise.
(emit_stack_clash_protection_loop_end): Likewise.
* rtl.h (get_stack_check_protect): Prototype.
* target.def (stack_clash_protection_final_dynamic_probe): New hook.
* targhooks.c (default_stack_clash_protection_final_dynamic_probe): New.
* targhooks.h (default_stack_clash_protection_final_dynamic_probe):
Prototype.
* doc/tm.texi.in (TARGET_STACK_CLASH_PROTECTION_FINAL_DYNAMIC_PROBE):
Add @hook.
* doc/tm.texi: Rebuilt.
* config/aarch64/aarch64.c (aarch64_expand_prologue): Use
get_stack_check_protect.
* config/alpha/alpha.c (alpha_expand_prologue): Likewise.
* config/arm/arm.c (arm_expand_prologue): Likewise.
(arm_frame_pointer_required): Likewise.
* config/i386/i386.c (ix86_expand_prologue): Likewise.
* config/ia64/ia64.c (ia64_expand_prologue): Likewise.
* config/mips/mips.c (mips_expand_prologue): Likewise.
* config/powerpcspe/powerpcspe.c (rs6000_emit_prologue): Likewise.
* config/rs6000/rs6000.c (rs6000_emit_prologue): Likewise.
* config/sparc/sparc.c (sparc_expand_prologue): Likewise.
(sparc_flat_expand_prologue): Likewise.

* gcc.dg/stack-check-3.c: New test.

From-SVN: r252995

common.opt (-fstack-clash-protection): New option.

* common.opt (-fstack-clash-protection): New option.
* flag-types.h (enum stack_check_type): Note difference between
-fstack-check= and -fstack-clash-protection.
* params.def (PARAM_STACK_CLASH_PROTECTION_GUARD_SIZE): New PARAM.
(PARAM_STACK_CLASH_PROTECTION_PROBE_INTERVAL): Likewise.
* toplev.c (process_options): Issue warnings/errors for cases
not handled with -fstack-clash-protection.
* doc/invoke.texi (-fstack-clash-protection): Document new option.
(-fstack-check): Note additional problem with -fstack-check=generic.
Note that -fstack-check is primarily for Ada and refer users
to -fstack-clash-protection for stack-clash-protection.
Document new params for stack clash protection.

* gcc.dg/stack-check-2.c: New test.
* lib/target-supports.exp
(check_effective_target_supports_stack_clash_protection): New function.
(check_effective_target_frame_pointer_for_non_leaf): Likewise.
(check_effective_target_caller_implicit_probes): Likewise.

From-SVN: r252994

runtime: restore "goroutine in C code" message

    In the 1.9 upgrade I took out the word "goroutine" from a traceback
    showing a goroutine running in C code, to let TestCgoNumGoroutine
    pass.  However, it turns out that some code is actually checking for
    that string; for example,
    https://github.com/grpc/grpc-go/blob/master/test/leakcheck/leakcheck.go#L44
    So keep the message the same, and change the test.

    Reviewed-on: https://go-review.googlesource.com/64850

From-SVN: r252991

Daily bump.

From-SVN: r252990

i386.md (*scc_bt<mode>): New insn_and_split pattern.

* config/i386/i386.md (*scc_bt<mode>): New insn_and_split pattern.
(*scc_bt<mode>_1): Ditto.
(*scc_bt<mode>_mask): Ditto.

testsuite/ChangeLog:

* gcc.target/i386/bt-5.c: New test.
* gcc.target/i386/bt-6.c: Ditto.
* gcc.target/i386/bt-mask-3.c: Ditto.
* gcc.target/i386/bt-mask-4.c: Ditto.

From-SVN: r252986

pr31146-2.C: Only do scan-tree-dump for c++14_down.

* g++.dg/tree-ssa/pr31146-2.C: Only do scan-tree-dump for c++14_down.
* g++.dg/tree-ssa/pr41428.C: Likewise.
* g++.dg/expr/bool1.C: Only do dg-do compile instead of dg-do run for
c++17 and up.
* g++.dg/expr/bool3.C: Likewise.
* g++.dg/expr/bitfield5.C: Likewise.
* g++.old-deja/g++.jason/bool5.C: Likewise.

From-SVN: r252985

i386.c (ix86_split_long_move): Do not handle address used for LEA in a special way.

* config/i386/i386.c (ix86_split_long_move): Do not handle
address used for LEA in a special way.

From-SVN: r252984

PR libstdc++/71500 restore C++11 compatibility in <regex>

PR libstdc++/71500
* include/bits/regex_executor.tcc
(_Backref_matcher<BidIt, regex_traits<C>>::_M_apply): Use
std::__equal4 instead of C++14 4-iterator overloads of std::equal.
* include/bits/stl_algobase.h (__equal4): New functions implementing
4-iterator overloads of std::equal for use in C++11.
(equal(It1, It1, It2, It2), equal(It1, It1, It2, It2, BinaryPred)):
Move function bodies to new __equal4 functions.
* testsuite/28_regex/simple_c++11.cc: New.

From-SVN: r252981

Fix typo in simplify_binary_operation_1

* simplify-rtx.c (simplify_binary_operation_1): Fix typo in comment.

From-SVN: r252979

PR libstdc++/82254 fix std::is_nothrow_invocable_r w.r.t throwing conversions

PR libstdc++/82254
* include/std/type_traits (__is_invocable): Add partial specialization
for INVOKE<void> case and remove is_void<R> check from partial
specialization for INVOKE<R> case.
(__is_nt_invocable_impl): New helper for is_nothrow_invocable_r.
(is_nothrow_invocable_r): Use __is_nt_invocable_impl.
* testsuite/20_util/is_nothrow_invocable/value.cc: Add tests for
conversions that can throw or fail to convert. Use static assert
strings to explain negative results.
* testsuite/20_util/is_nothrow_invocable/value_ext.cc: Use
is_nothrow_constructible in is_nt_invocable_conv.

From-SVN: r252977

PR c/81854 - weak alias of an incompatible symbol accepted

gcc/ChangeLog:

PR c/81854
* cgraphunit.c (handle_alias_pairs): Reject aliases between functions
of incompatible types.

gcc/testsuite/ChangeLog:

PR c/81854
* gcc.dg/pr81854.c: New test.
* g++.dg/ext/attr-ifunc-5.C: New test.
* g++.dg/ext/attr-ifunc-1.C: Adjust.
* g++.dg/ext/attr-ifunc-2.C: Same.
* g++.dg/ext/attr-ifunc-3.C: Same.
* g++.dg/ext/attr-ifunc-4.C: Same.
* g++.old-deja/g++.abi/vtable2.C: Same.
* gcc.dg/attr-ifunc-1.c: Same.

From-SVN: r252976

rs6000.c (rs6000_gimple_fold_builtin): Add handling for early folding of vector loads (ALTIVEC_BUILTIN_LVX_*).

[gcc]

2017-09-19  Will Schmidt  <will_schmidt@vnet.ibm.com>

    * config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add handling
    for early folding of vector loads (ALTIVEC_BUILTIN_LVX_*).
    * config/rs6000/rs6000-c.c (altivec_resolve_overloaded_builtin):
    Remove obsoleted code for handling ALTIVEC_BUILTIN_VEC_LD.

From-SVN: r252975

fold-vec-ld-misc.c: New.

[gcc/testsuite]

2017-09-19  Will Schmidt  <will_schmidt@vnet.ibm.com>

        * gcc.target/powerpc/fold-vec-ld-misc.c: New.

From-SVN: r252974

re PR tree-optimization/82244 (-O2: ICE: tree check: expected ssa_name, have integer_cst in replace_uses_by, at tree-cfg.c:1904)

2017-09-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/82244
* tree-vrp.c (remove_range_assertions): Do not propagate
a constant to abnormals but replace the assert with a copy.

* gcc.dg/torture/pr82244.c: New testcase.

From-SVN: r252973

lra: make reload_pseudo_compare_func a proper comparator

PR rtl-optimization/57878
PR rtl-optimization/68988
* lra-assigns.c (reload_pseudo_compare_func): Remove fragmentation
avoidance test involving non_reload_pseudos.  Move frequency test
below the general fragmentation avoidance test.

From-SVN: r252972

validity_check.adb: Minor dg-options simplification.

gcc/testsuite/

* gnat.dg/validity_check.adb: Minor dg-options simplification.

From-SVN: r252971

re PR tree-optimization/69728 (internal compiler error: in outer_projection_mupa, at graphite-sese-to-poly.c:1175)

2017-09-19  Richard Biener  <rguenther@suse.de>

PR tree-optimization/69728
* graphite-sese-to-poly.c (schedule_error): New global.
(add_loop_schedule): Handle empty domain by failing the
schedule.
(build_original_schedule): Handle schedule_error.

* gfortran.dg/graphite/pr69728.f90: New testcase.
* gcc.dg/graphite/pr69728.c: Likewise.

From-SVN: r252968

Fix some gotools testing problems

* Makefile.am (GOTESTFLAGS): New variable.
(check-runtime): Pass it to gotest.
(check-go-tools): Sort summary.
(check-runtime): Likewise.
(check-cgo-test): Likewise.
(check-carchive-test): Likewise.
(check): Likewise.
* Makefile.in: Regenerate.

From-SVN: r252963

graphite-scop-detection.c (scop_detection::can_represent_loop): Do not iterate to sibling loops but only to siblings of inner loops.

2017-09-19  Richard Biener  <rguenther@suse.de>

* graphite-scop-detection.c (scop_detection::can_represent_loop):
Do not iterate to sibling loops but only to siblings of inner
loops.

From-SVN: r252962

Daily bump.

From-SVN: r252961

re PR c++/80947 (Different visibility for the lambda and its capture list members with -fvisibility=hidden)

2017-09-18  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/80947
* g++.dg/cpp0x/lambda/lambda-80947.C: New.

From-SVN: r252957

runtime: always initialize str field in __go_string_slice result

    Reviewed-on: https://go-review.googlesource.com/64110

From-SVN: r252953

PR libstdc++/81468 fix test for duration conversions

PR libstdc++/81468
* testsuite/20_util/duration/cons/dr1177.cc: Fix incorrect test and
improve static assertion messages.

From-SVN: r252950

re PR target/81613 (FAIL: gfortran.dg/intrinsic_modulo_1.f90   -O3 -g  execution test)

PR target/81613
* config/m68k/m68k.md (moveq feeding equality comparison): Check
that the registers are different.

From-SVN: r252949

i386.c (fold_builtin_cpu): Add M_AMDFAM17H to processor_model and "amdfam17h" to arch_names_table.

* config/i386/i386.c (fold_builtin_cpu): Add M_AMDFAM17H
to processor_model and "amdfam17h" to arch_names_table.
* doc/extend.texi (__builtin_cpu_is): Document amdfam17h CPU name.

From-SVN: r252948

re PR c/82234 (__builtin_shuffle is not in the keyword index)

PR c/82234
* doc/extend.texi: Add @findex entry for __builtin_shuffle.

From-SVN: r252947

PR c++/82069 - ICE with lambda in template

* semantics.c (process_outer_var_ref): Check uses_template_parms
instead of any_dependent_template_arguments_p.

From-SVN: r252936

Move computation of SLP_TREE_NUMBER_OF_VEC_STMTS

Previously SLP_TREE_NUMBER_OF_VEC_STMTS was calculated while scheduling
an SLP tree after analysis, but sometimes it can be useful to know the
value during analysis too.  This patch moves the calculation to
vect_slp_analyze_node_operations instead.

2017-09-18  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-vectorizer.h (vect_slp_analyze_operations): Replace parameters
with a vec_info *.
* tree-vect-loop.c (vect_analyze_loop_operations): Update call
accordingly.
* tree-vect-slp.c (vect_slp_analyze_node_operations): Add vec_info *
parameter.  Set SLP_TREE_NUMBER_OF_VEC_STMTS here rather than in
vect_schedule_slp_instance.
(vect_slp_analyze_operations): Replace parameters with a vec_info *.
Update call to vect_slp_analyze_node_operations.  Simplify return
value.
(vect_slp_analyze_bb_1): Update call accordingly.
(vect_schedule_slp_instance): Remove vectorization_factor parameter.
Don't calculate SLP_TREE_NUMBER_OF_VEC_STMTS here.
(vect_schedule_slp): Update call accordingly.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r252935

Fix an SVE failure in the Fortran matmul* tests

The vectoriser was calling vect_get_smallest_scalar_type without
having proven that the type actually is a scalar.  This seems to
be the intended behaviour: the ultimate test of whether the type
is interesting (and hence scalar) is whether an associated vector
type exists, but this is only tested later.

The patch simply makes the function cope gracefully with non-scalar
inputs.

2017-09-18  Richard Sandiford  <richard.sandiford@linaro.org>
    Alan Hayward  <alan.hayward@arm.com>
    David Sherwood  <david.sherwood@arm.com>

gcc/
* tree-vect-data-refs.c (vect_get_smallest_scalar_type): Cope
with types that aren't in fact scalar.

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r252934

Include phis in SLP unrolling calculation

Without this we'd pick an unrolling factor based purely on longs,
ignoring the ints.  It's posssible that vect_get_smallest_scalar_type
should also handle shifts, but I think we'd still want this as a
belt-and-braces fix.

2017-09-18  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree-vect-slp.c (vect_record_max_nunits): New function,
split out from...
(vect_build_slp_tree_1): ...here.
(vect_build_slp_tree_2): Call it for phis too.

gcc/testsuite/
* gcc.dg/vect/slp-multitypes-13.c: New test.

From-SVN: r252933

Fix vectorizable_mask_load_store handling of invariant masks

vectorizable_mask_load_store was not passing the required mask type to
vect_get_vec_def_for_operand.  This doesn't matter for masks that are
defined in the loop, since their STMT_VINFO_VECTYPE will be what we need
anyway.  But it's not possible to tell which mask type the caller needs
when looking at an invariant scalar boolean.  As the comment above the
function says:

   In case OP is an invariant or constant, a new stmt that creates a vector def
   needs to be introduced.  VECTYPE may be used to specify a required type for
   vector invariant.

This fixes the attached testcase for SVE.

2017-09-18  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree-vect-stmts.c (vectorizable_mask_load_store): Pass mask_vectype
to vect_get_vec_def_for_operand when getting the mask operand.

gcc/testsuite/
* gfortran.dg/vect/mask-store-1.f90: New test.

From-SVN: r252932