[3/3] No need to vectorize simple only-live stmts

2016-06-03  Alan Hayward  <alan.hayward@arm.com>

[3/3] No need to vectorize simple only-live stmts

gcc/
* tree-vect-stmts.c (vect_stmt_relevant_p): Do not vectorize non live
relevant stmts which are simple and invariant.
* tree-vect-loop.c (vectorizable_live_operation): Check relevance
instead of simple and invariant

testsuite/
* gcc.dg/vect/vect-live-slp-5.c: Remove dg check.

From-SVN: r237065

[2/3] Vectorize inductions that are live after the loop

2016-06-03  Alan Hayward  <alan.hayward@arm.com>

[2/3] Vectorize inductions that are live after the loop

gcc/
* tree-vect-loop.c (vect_analyze_loop_operations): Allow live stmts.
(vectorizable_reduction): Check for new relevant state.
(vectorizable_live_operation): vectorize live stmts using
BIT_FIELD_REF.  Remove special case for gimple assigns stmts.
* tree-vect-stmts.c (is_simple_and_all_uses_invariant): New function.
(vect_stmt_relevant_p): Check for stmts which are only used live.
(process_use): Use of a stmt does not inherit it's live value.
(vect_mark_stmts_to_be_vectorized): Simplify relevance inheritance.
(vect_analyze_stmt): Check for new relevant state.
* tree-vectorizer.h (vect_relevant): New entry for a stmt which is used
outside the loop, but not inside it.

testsuite/
* gcc.dg/tree-ssa/pr64183.c: Ensure test does not vectorize.
* testsuite/gcc.dg/vect/no-scevccp-vect-iv-2.c: Remove xfail.
* gcc.dg/vect/vect-live-1.c: New test.
* gcc.dg/vect/vect-live-2.c: New test.
* gcc.dg/vect/vect-live-3.c: New test.
* gcc.dg/vect/vect-live-4.c: New test.
* gcc.dg/vect/vect-live-5.c: New test.
* gcc.dg/vect/vect-live-slp-1.c: New test.
* gcc.dg/vect/vect-live-slp-2.c: New test.
* gcc.dg/vect/vect-live-slp-3.c: New test.

From-SVN: r237064

[1/3] Split vect_get_vec_def_for_operand into two

2016-06-03  Alan Hayward  <alan.hayward@arm.com>

[1/3] Split vect_get_vec_def_for_operand into two

gcc/
* tree-vectorizer.h (vect_get_vec_def_for_operand_1): New
* tree-vect-stmts.c (vect_get_vec_def_for_operand_1): New
(vect_get_vec_def_for_operand): Split out code.

From-SVN: r237063

rs6000: Remove the ancient mfcr peepholes

These peepholes replace two mfcr;mask sequences by one mfcr;mask;mask
sequence.  On modern cpus, the original mfcr's were actually mfocrf,
but the new insn is an actual heavy-weight mfcr.  This is very bad
for performance.

The comment says there is a three cycle delay between two consecutive
mfcr insns.  This may have been true on rios, and it's true on 604,
but on 603, 750, 7400 it is just a single cycle (on 7450 it is two).

This is also a define_peephole, and we should get rid of those.

So this patch just removes the peepholes; the benefit is marginal at
best, and it so very hurts in other cases.

* config/rs6000/rs6000.md (define_peepholes for two mfcr's): Delete.

From-SVN: r237062

2016-06-03  Alan Hayward  <alan.hayward@arm.com>

gcc/
* tree-vect-stmts.c (vectorizable_call) Remove GOMP_SIMD_LANE code.

From-SVN: r237061

[ARM][obvious] Fix typos in *thumb1_mulsi3 comment

* config/arm/thumb1.md (*thumb1_mulsi3): Fix typos in comment.

From-SVN: r237058

* MAINTAINERS (Write After Approval): Add myself.

From-SVN: r237055

re PR middle-end/71387 (ICE in emit_move_insn, at expr.c:3418 with -Og)

PR middle-end/71387
* cgraph.c (cgraph_edge::redirect_call_stmt_to_callee): If redirecting
to noreturn e->callee->decl that has void return type and void
arguments, adjust gimple_call_fntype and remove lhs even if it had
previously addressable type.

* g++.dg/opt/pr71387.C: New test.

From-SVN: r237053

re PR tree-optimization/71328 (ice in verify_jump_thread)

PR tree-optimization/71328
* tree-ssa-threadupdate.c (duplicate_thread_path): Fix off-by-one
error when checking for a jump back onto the copied path.  */

PR tree-optimization/71328
* gcc.c-torture/compile/pr71328.c: New test.

From-SVN: r237052

re PR fortran/52393 (I/O: "READ format" statement with parenthesed default-char-expr)

2016-06-02  Jerry DeLisle  <jvdelisle@gcc.gnu.org>

PR fortran/52393
* gfortran.dg/fmt_read_3.f90: Fix typo.
* gfortran.dg/fmt_read_4.f90: New test.

From-SVN: r237051

Daily bump.

From-SVN: r237050

decl.c (xref_tag_1): Change pairs of errors to error + inform.

/cp
2016-06-02  Paolo Carlini  <paolo.carlini@oracle.com>

* decl.c (xref_tag_1): Change pairs of errors to error + inform.
(start_enum): Likewise.
* parser.c (cp_parser_class_head): Likewise.

/testsuite
2016-06-02  Paolo Carlini  <paolo.carlini@oracle.com>

* g++.dg/cpp0x/forw_enum10.C: Adjust for dg-message vs dg-error.
* g++.dg/cpp0x/forw_enum6.C: Likewise.
* g++.dg/cpp0x/forw_enum8.C: Likewise.
* g++.dg/cpp0x/override2.C: Likewise.
* g++.dg/parse/crash5.C: Likewise.
* g++.dg/parse/error16.C: Likewise.
* g++.dg/parse/error27.C: Likewise.
* g++.dg/template/qualttp15.C: Likewise.
* g++.dg/template/redecl4.C: Likewise.
* g++.old-deja/g++.other/crash39.C: Likewise.
* g++.old-deja/g++.other/struct1.C: Likewise.
* g++.old-deja/g++.pt/m9a.C: Likewise.
* g++.old-deja/g++.pt/memclass10.C: Likewise.

From-SVN: r237046

microblaze.c: fix warnings

gcc/ChangeLog:
* config/microblaze/microblaze.c (get_branch_target): Add return
NULL_RTX for the non-CALL_P case.
(insert_wic_for_ilb_runout): Remove unused local "wic_addr1".
(insert_wic): Remove unused local "j".

From-SVN: r237044

re PR c++/71372 (C++ FE drops TREE_THIS_VOLATILE in cp_fold on all tcc_reference trees)

PR c++/71372
* cp-gimplify.c (cp_fold): For INDIRECT_REF, if the folded expression
is INDIRECT_REF or MEM_REF, copy over TREE_READONLY, TREE_SIDE_EFFECTS
and TREE_THIS_VOLATILE flags.  For ARRAY_REF and ARRAY_RANGE_REF, copy
over TREE_READONLY, TREE_SIDE_EFFECTS and TREE_THIS_VOLATILE flags
to the newly built tree.

* c-c++-common/pr71372.c: New test.

From-SVN: r237041

Fix display name of PRED_FORTRAN_FAIL_IO

* predict.def: Fix typo in PRED_FORTRAN_FAIL_IO display name.

From-SVN: r237040

source_date_epoch-1.c (main): Test __DATE__ and __TIME__ strings with __builtin_strcmp instead of printf and...

* gcc.dg/cpp/source_date_epoch-1.c (main): Test __DATE__ and
__TIME__ strings with __builtin_strcmp instead of printf and
dg-output.

From-SVN: r237039

Update TARGET_FUNCTION_INCOMING_ARG documentation

On x86, interrupt handlers are only called by processors which push
interrupt data onto stack at the address where the normal return address
is.  Since interrupt handlers must access interrupt data via pointers so
that they can update interrupt data, the pointer argument is passed as
"argument pointer - word".

TARGET_FUNCTION_INCOMING_ARG defines how callee sees its argument.
Normally it returns REG, NULL, or CONST_INT.  This patch adds arbitrary
address computation based on hard register, which can be forced into a
register, to the list.

When copying an incoming argument onto stack, assign_parm_setup_stack
has:

if (argument in memory)
  copy argument in memory to stack
else
  move argument to stack

Since an arbitrary address computation may be passed as an argument, we
change it to:

if (argument in memory)
  copy argument in memory to stack
else
  {
    if (argument isn't in register)
      force argument into a register
    move argument to stack
  }

* function.c (assign_parm_setup_stack): Force source into a
register if needed.
* target.def (function_incoming_arg): Update documentation to
allow arbitrary address computation based on hard register.
* doc/tm.texi: Regenerated.

Co-Authored-By: Julia Koval <julia.koval@intel.com>
From-SVN: r237037

Fix first match heuristics

* predict.c (combine_predictions_for_bb): Fix first match in
cases where a first predictor contains more than one occurence
in list of predictors.  Take the best value in such case.
* gcc.dg/predict-9.c: New test.

From-SVN: r237036

gcc-dg.exp (cleanup-after-saved-dg-test): Add missing global set_compiler_env_var and global saved_compiler_env_var.

* lib/gcc-dg.exp (cleanup-after-saved-dg-test): Add missing
global set_compiler_env_var and global saved_compiler_env_var.

From-SVN: r237035

[rtlanal] Fix rtl-optimization/71295

PR rtl-optimization/71295
* rtlanal.c (subreg_get_info): If taking a subreg at the requested
offset would go over the size of the inner mode reject it.

* gcc.c-torture/compile/pr71295.c: New test.

From-SVN: r237034

On behalf of Aaron Conole <aconole@redhat.com>

On behalf of  Aaron Conole  <aconole@redhat.com>
* libgcov-driver-system.c (__gcov_error_file): New.
(get_gcov_error_file): New.
(gcov_error): Use and set __gcov_error_file.
(gcov_error_exit): New.
* libgcov-driver.c (gcov_exit): Call gcov_error_exit.

From-SVN: r237033

sse.md (*vec_concatv4si): Use v=v,v instead of x=x,x and v=v,m instead of x=x,m.

* config/i386/sse.md (*vec_concatv4si): Use v=v,v instead of
x=x,x and v=v,m instead of x=x,m.

* gcc.target/i386/avx512vl-concatv4si-1.c: New test.

From-SVN: r237031

sse.md (*vec_concatv2si_sse4_1): Add avx512dq v=Yv,rm alternative.

* config/i386/sse.md (*vec_concatv2si_sse4_1): Add avx512dq v=Yv,rm
alternative.  Change x=x,x alternative to v=Yv,Yv and x=rm,C
alternative to v=rm,C.

* gcc.target/i386/avx512dq-concatv2si-1.c: New test.
* gcc.target/i386/avx512vl-concatv2si-1.c: New test.

From-SVN: r237030

sse.md (*vec_concatv2di): Add x86_avx512dq v=Yv,rm alternative.

* config/i386/sse.md (*vec_concatv2di): Add x86_avx512dq v=Yv,rm
alternative.  Change x=xm,C alternative to v=vm,C, x=x,x alternative
to v=Yv,Yv and x=x,m to v=v,m.  Use maybe_evex prefix attribute
instead of vex for the last two above mentioned alternatives.

* gcc.target/i386/avx512dq-concatv2di-1.c: New test.
* gcc.target/i386/avx512vl-concatv2di-1.c: New test.
* gcc.target/i386/sse2-init-v2di-2.c: Adjust expected vec_concatv2di
alternative number.

From-SVN: r237028

Fix fallout from: [ARM] PR target/70830: Avoid POP-{reglist}^ when returning from interrupt handlers

PR target/70830
* config/arm/arm.c (arm_output_multireg_pop): Guard "pop" on update.

From-SVN: r237027

rs6000: Correct the "length" attribute for trunctddd2

* config/rs6000/dfp.md (trunctddd2): Correct the "length" attribute.

From-SVN: r237026

Daily bump.

From-SVN: r237025

Fix ChangeLog entry:

2016-05-24  Michael Meissner  <meissner@linux.vnet.ibm.com>

* config/rs6000/altivec.md (VNEG iterator): New iterator for
VNEGW/VNEGD instructions.
(p9_neg<mode>2): New insns for ISA 3.0 VNEGW/VNEGD.
(neg<mode>2): Add expander for V2DImode added in ISA 2.07, and
support for ISA 3.0 VNEGW/VNEGD instructions.

From-SVN: r237020

rl78.c: fix warning

gcc/ChangeLog:
* config/rl78/rl78.c (rl78_expand_prologue): Convert local
from int to unsigned.

From-SVN: r237017

re PR c/70688 (bogus OpenACC data clause errors involving reductions)

PR c/70688
* pr70688.c: New file.

From-SVN: r237011

libtool-version: Bump soversion.

2016-06-01  Matthias Klose  <doko@ubuntu.com>

        * libtool-version: Bump soversion.

From-SVN: r237007

re PR target/71186 (PowerPC64: Autovectorised code hits ICE with -O3 -mpower9 -mlra)

[gcc]
2016-05-31  Michael Meissner  <meissner@linux.vnet.ibm.com>

PR target/71186
* config/rs6000/vsx.md (xxspltib_<mode>_nosplit): Add alternatives
for loading up all 0's or all 1's.

[gcc/testsuite]
2016-05-31  Michael Meissner  <meissner@linux.vnet.ibm.com>

PR target/71186
* gcc.target/powerpc/pr71186.c: New test.

Index: gcc/config/rs6000/vsx.md
===================================================================
--- gcc/config/rs6000/vsx.md (.../svn+ssh://meissner@gcc.gnu.org/svn/gcc/trunk/gcc/config/rs6000) (revision 236935)
+++ gcc/config/rs6000/vsx.md (.../gcc/config/rs6000) (working copy)
@@ -776,8 +776,8 @@ (define_insn "xxspltib_v16qi"
   [(set_attr "type" "vecperm")])

 (define_insn "xxspltib_<mode>_nosplit"
-  [(set (match_operand:VSINT_842 0 "vsx_register_operand" "=wa")
- (match_operand:VSINT_842 1 "xxspltib_constant_nosplit" "wE"))]
+  [(set (match_operand:VSINT_842 0 "vsx_register_operand" "=wa,wa")
+ (match_operand:VSINT_842 1 "xxspltib_constant_nosplit" "jwM,wE"))]
   "TARGET_P9_VECTOR"
 {
   rtx op1 = operands[1];

[gcc]
2016-05-31  Michael Meissner  <meissner@linux.vnet.ibm.com>

* config/rs6000/vsx.md (vsx_splat_<mode>, V2DI/V2DF): Simplify
alternatives, eliminating preferred register class.  Add support
for the MTVSRDD instruction in ISA 3.0.
(vsx_splat_v4si_internal): Use splat_input_operand instead of
reg_or_indexed_operand.
(vsx_splat_v4sf_internal): Likewise.

[gcc/testsuite]
2016-05-31  Michael Meissner  <meissner@linux.vnet.ibm.com>

* gcc.target/powerpc/p9-splat-4.c: New test.

From-SVN: r237006

re PR fortran/52393 (I/O: "READ format" statement with parenthesed default-char-expr)

2016-06-01  Jerry DeLisle  <jvdelisle@gcc.gnu.org>

PR fortran/52393
* io.c (match_io): For READ, try to match a default character
expression. If found, set the dt format expression to this,
otherwise go back and try control list.

PR fortran/52393
* gfortran.dg/fmt_read_3.f90: New test.

From-SVN: r237003

sourcebuild.texi (arm_acq_rel): Document new effective target.

2016-06-01  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/
    * doc/sourcebuild.texi (arm_acq_rel): Document new effective target.

    gcc/testsuite/
    * lib/target-supports.exp (check_effective_target_arm_acq_rel): New
    procedure.

From-SVN: r237002

c-common.c (get_source_date_epoch): Rename to cb_get_source_date_epoch.

gcc/c-family/ChangeLog:

2016-05-13  Eduard Sanou  <dhole@openmailbox.org>

* c-common.c (get_source_date_epoch): Rename to
cb_get_source_date_epoch.
* c-common.c (cb_get_source_date_epoch): Use a single generic erorr
message when the parsing fails.  Use error_at instead of fatal_error.
* c-common.h (get_source_date_epoch): Rename to
cb_get_source_date_epoch.
* c-common.h (cb_get_source_date_epoch): Prototype.
* c-common.h (MAX_SOURCE_DATE_EPOCH): Define.
* c-common.h (c_omp_region_type): Remove trailing comma.
* c-lex.c (init_c_lex): Set cb->get_source_date_epoch callback.
* c-lex.c (c_lex_with_flags): Remove initialization of
pfile->source_date_epoch.

gcc/ChangeLog:

2016-05-13  Eduard Sanou  <dhole@openmailbox.org>

* doc/cppenv.texi: Note that the `%s` in `date` is a non-standard
extension.
* gcc.c (driver_handle_option): Call set_source_date_epoch_envvar.
* gcc.c (set_source_date_epoch_envvar): New function, sets
the SOURCE_DATE_EPOCH environment variable to the current time.

gcc/testsuite/ChangeLog:

2016-05-13  Eduard Sanou  <dhole@openmailbox.org>

* gcc.dg/cpp/source_date_epoch-1.c: New file, test the proper
behaviour of the macros __DATE__ and __TIME__ when SOURCE_DATE_EPOCH
env var is set.
* gcc.dg/cpp/source_date_epoch-2.c: New file, test the error output
when parsing the SOURCE_DATE_EPOCH env var, and make sure it is only
shown once.
* lib/gcc-dg.exp (dg-set-compiler-env-var): New function, set env vars
during compilation.
* lib/gcc-dg.exp (restore-compiler-env-var): New function, restore env
vars set by dg-set-compiler-env-var.

libcpp/ChangeLog:

2016-05-13  Eduard Sanou  <dhole@openmailbox.org>

* include/cpplib.h (cpp_callbacks): Add get_source_date_epoch
callback.
* include/cpplib.h (cpp_init_source_date_epoch): Remove prototype.
* init.c (cpp_init_source_date_epoch): Remove function.
* init.c (cpp_create_reader): Initialize pfile->source_date_epoch.
* internal.h (cpp_reader): Extend comment about source_date_epoch.
* macro.c (_cpp_builtin_macro_text): Use get_source_date_epoch
callback only once, read pfile->source_date_epoch on future passes.
Check that get_source_date_epoch callback is not NULL.

From-SVN: r237001

tree-vect-loop.c (vect_determine_vectorization_factor): Also compute the factor for live Phi nodes.

* tree-vect-loop.c (vect_determine_vectorization_factor): Also compute
the factor for live Phi nodes.

From-SVN: r236999

loop-dolop.c (doloop_optimize): Us likely max iteration bound.

* loop-dolop.c (doloop_optimize): Us likely max iteration bound.
* tree-parloops.c (parallelize_loops): likewise.
* tree-ssa-loop-unswitch.c (tree_unswitch_single_loop,
tree_unswitch_outer_loop): likewise.

From-SVN: r236998

re PR fortran/71156 (PURE interface/definition inconsistency: accepts invalid, rejects valid)

2016-06-01  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/71156
* decl.c (copy_prefix): Add checks that the module procedure
declaration prefixes are compliant with the interface. Invert
order of existing elemental and pure checks.
* resolve.c (resolve_fl_procedure): Invert order of elemental
and pure errors.

2016-06-01  Paul Thomas  <pault@gcc.gnu.org>

PR fortran/71156
* gfortran.dg/submodule_14.f08: Add missing recursive prefix
to the module procedure declaration.
* gfortran.dg/submodule_16.f08: New test.

From-SVN: r236996

re PR middle-end/71371 (ICE with OpenMP taskloop and addressable iterator)

PR middle-end/71371
* gimplify.c (gimplify_omp_for): Temporarily clear gimplify_omp_ctxp
around creation of the temporary.

* c-c++-common/gomp/pr71371.c: New test.

From-SVN: r236994

re PR tree-optimization/71366 (ICE on valid code at -O3 on x86_64-linux-gnu: Segmentation fault)

2016-06-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71366
* tree-ssa-loop-ivcanon.c (edges_to_remove): New global.
(unloop_loops): Move removing edges here ...
(try_unroll_loop_completely): ... from here.
(try_peel_loop): ... and here.
(tree_unroll_loops_completely_1): Track parent loops via
bitmap of header BBs.
(tree_unroll_loops_completely): Adjust for that.

* gcc.dg/torture/pr71366-1.c: New testcase.
* gcc.dg/torture/pr71366-2.c: Likewise.

From-SVN: r236993

altivec.h (vec_slv): New macro.

gcc/ChangeLog:

2016-06-01  Kelvin Nilsen  <kelvin@gcc.gnu.org>

* config/rs6000/altivec.h (vec_slv): New macro.
(vec_srv): New macro.
* config/rs6000/altivec.md (UNSPEC_VSLV): New value.
(UNSPEC_VSRV): New value.
(vslv): New insn.
(vsrv): New insn.
* config/rs6000/rs6000-builtin.def (vslv): New builtin definition.
(vsrv): New builtin definition.
* config/rs6000/rs6000-c.c (P9V_BUILTIN_VSLV): Macro expansion to
define argument types for new builtin.
(P9V_BUILTIN_VSRV): Macro expansion to define argument types for
new builtin.
* doc/extend.texi: Document the new vec_vslv and vec_srv built-in
functions.

gcc/testsuite/ChangeLog:

2016-06-01  Kelvin Nilsen  <kelvin@gcc.gnu.org>

* gcc.target/powerpc/vslv-0.c: New test.
* gcc.target/powerpc/vslv-1.c: New test.
* gcc.target/powerpc/vsrv-0.c: New test.
* gcc.target/powerpc/vsrv-1.c: New test.

From-SVN: r236992

re PR target/67310 ([PATCH] gcc 4.8.4 fails to compile with -march=native on VIA nano CPU)

PR target/67310
* config/i386/driver-i386.c (host_detect_local_cpu): Correctly
detect processor family for signature_CENTAUR_ebx.
<case PROCESSOR_I486>: Pass c3, winchip2 or winchip-c6 for
signature_CENTAUR_ebx.
<case PROCESSOR _PENTIUMPRO>: Pass c3-2 for signature_CENTAUR_ebx.
<default>: Pass x86-64 for has_longmode.

Co-Authored-By: Jocelyn Mayer <l_indien@magic.fr>
From-SVN: r236991

nvptx.c (nvptx_assemble_undefined_decl): Reject undefined weak.

* config/nvptx/nvptx.c (nvptx_assemble_undefined_decl): Reject
undefined weak.

testsuite/
* c-c++-common/torture/pr57945.c: Add expected PTX error.
* gcc.target/nvptx/weak.c: New.

From-SVN: r236990

re PR tree-optimization/71261 (Trunk GCC hangs on knl and broadwell targets)

2016-06-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71261
* tree-vect-patterns.c (check_bool_pattern): Gather a hash-set
of stmts successfully put in the bool pattern.  Remove
single-use restriction.
(adjust_bool_pattern_cast): Add cast at the use site via the
pattern def sequence.
(adjust_bool_pattern): Remove recursion, maintain a hash-map
of patterned defs.  Use the pattern def seqence instead of
multiple independent patterns.
(sort_after_uid): New qsort compare function.
(adjust_bool_stmts): New function to process stmts in the bool
pattern in IL order.
(vect_recog_bool_pattern): Adjust.
* tree-if-conv.c (ifcvt_split_def_stmt): Remove.
(ifcvt_walk_pattern_tree): Likewise.
(stmt_is_root_of_bool_pattern): Likewise.
(ifcvt_repair_bool_pattern): Likewise.
(tree_if_conversion): Do not call ifcvt_repair_bool_pattern.

* gcc.dg/torture/vect-bool-1.c: New testcase.

From-SVN: r236989

parse.c (case_decl): Move ST_OMP_* to ...

* parse.c (case_decl): Move ST_OMP_* to ...
(case_omp_decl): ... here, new macro.
(verify_st_order): For case_omp_decl, complain about
p->state >= ORDER_EXEC, but don't change p->state otherwise.

* gfortran.dg/gomp/order-1.f90: New test.
* gfortran.dg/gomp/order-2.f90: New test.

From-SVN: r236987

loop-unroll.c (decide_unroll_constant_iterations, [...]): Use likely upper bounds.

* loop-unroll.c (decide_unroll_constant_iterations,
decide_unroll_runtime_iterations, decide_unroll_stupid): Use
likely upper bounds.
* loop-iv.c (find_simple_exit): Dump likely upper bounds.

* gcc.dg/unroll-6.c: Update template.
* gcc.dg/unroll-7.c: New testcase.
* gcc.dg/unroll-8.c: New testcase.

From-SVN: r236986

Remove the unused OMP_CLAUSE_DEVICE_RESIDENT

gcc/
* tree-core.h (enum omp_clause_code): Remove
OMP_CLAUSE_DEVICE_RESIDENT.  Adjust all users.

From-SVN: r236985

[ARM] Use proper output modifier for DImode register in store exclusive patterns

* config/arm/sync.md (arm_store_exclusive<mode>):
Use 'H' output modifier on operands[2] rather than creating a new
entry in out-of-bounds memory of the operands array.
(arm_store_release_exclusivedi): Likewise.

From-SVN: r236984

[ARM] Tie operand 1 to operand 0 in AESMC pattern when fusing AES/AESMC

* config/arm/arm.c (arm_fusion_enabled_p): New function.
* config/arm/arm-protos.h (arm_fusion_enabled_p): Declare prototype.
* config/arm/crypto.md (crypto_<crypto_pattern>, CRYPTO_UNARY):
Add "=w,0" alternative.  Enable it when AES/AESMC fusion is enabled.

From-SVN: r236982

tree-vect-loop.c (vect_determine_vectorization_factor): Also take into account live statements for mask producers.

* tree-vect-loop.c (vect_determine_vectorization_factor): Also take
into account live statements for mask producers.

From-SVN: r236981

re PR tree-optimization/71311 (spec2006 test case 416.gamess fails since r235817)

2016-06-01  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71311
* match.pd (@0 < @1 && @0 < @2 -> @0 < min(@1,@2)): Add :c and
restrict to non-INTEGER_CST @0.

From-SVN: r236980

match.pd ((A & B) - (A & ~B) -> B - (A ^ B)): Add missing :c.

2016-06-01  Richard Biener  <rguenther@suse.de>

* match.pd ((A & B) - (A & ~B) -> B - (A ^ B)): Add missing :c.
(relational patterns): Use :c to avoid pattern duplications.

From-SVN: r236979

re PR c++/66635 (g++ ICE when compiling constexpr code with comma operator)

2016-06-01  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/66635
* g++.dg/cpp0x/constexpr-ice16.C: New.

From-SVN: r236978

genmatch.c (comparison_code_p): New predicate.

2016-06-01  Richard Biener  <rguenther@suse.de>

* genmatch.c (comparison_code_p): New predicate.
(swap_tree_comparison): New function.
(commutate): Add for_vec parameter to append new for entries.
Support commutating relational operators by swapping it alongside
operands.
(lower_commutative): Adjust.
(dt_simplify::gen): Do not pass artificial operators to gen
functions.
(decision_tree::gen): Do not add artificial operators as parameters.
(parser::parse_expr): Verify operator commutativity when :c is
applied.  Allow :C to override this.
* match.pd: Adjust patterns to use :C instead of :c where required.

From-SVN: r236977

re PR ada/71358 (GNAT.Command_Line.Getopt fails if there are no switches)

PR ada/71358
* g-comlin.adb (Display_Section_Help): Do not dereference
Config.Switches if it's null.
(Getopt): Likewise.

From-SVN: r236975

re PR tree-optimization/71314 (test case gcc.dg/tree-ssa/ssa-thread-14.c fails starting with its introduction in r235653)

Fix PR tree-optimization/71314

gcc/testsuite/ChangeLog:

PR tree-optimization/71314
* gcc.dg/tree-ssa/ssa-thread-14.c: Adjust target selector.  Pass
-mbranch-cost=2.

From-SVN: r236974

re PR tree-optimization/71077 (gcc -lto raises ICE)

Fix PR tree-optimization/71077

gcc/ChangeLog:

PR tree-optimization/71077
* tree-ssa-threadedge.c (simplify_control_stmt_condition_1): In
the combining step, use boolean_false_node and boolean_true_node
as the designated false/true return values.

gcc/testsuite/ChangeLog:

PR tree-optimization/71077
* gcc.dg/tree-ssa/pr71077.c: New test.

From-SVN: r236973

Daily bump.

From-SVN: r236972

predict-lop-exit-1.C: Update template for new predictor name.

* g++.d/predict-lop-exit-1.C: Update template for new predictor name.
* g++.d/predict-lop-exit-2.C: Update template for new predictor name.
* g++.d/predict-lop-exit-2.C: Update template for new predictor name.

* predict.def (PRED_LOOP_EXTRA_EXIT): Define.
* predict.c (predict_iv_comparison): Also check PRED_LOOP_EXTRA_EXIT.
(predict_extra_loop_exits): Use PRED_LOOP_EXTRA_EXIT instead of
PRED_LOOP_EXIT.

From-SVN: r236968

invoke.texi (-frename-registers): Drop -fpeel-loops from list of flags impliying the register renaming.

* doc/invoke.texi (-frename-registers): Drop -fpeel-loops from list
of flags impliying the register renaming.
* toplev.c (process_options): Do not imply flag_rename_registers with
loop peeling.

From-SVN: r236966

* pt.c (instantiate_decl): Avoid recalculation.

From-SVN: r236947

PR c++/60095 - partial specialization of variable templates

PR c++/69515
PR c++/69009
* pt.c (instantiate_template_1): Don't put the partial
specialization in DECL_TI_TEMPLATE.
(partial_specialization_p, impartial_args): Remove.
(regenerate_decl_from_template): Add args parm.
(instantiate_decl): Look up the partial specialization again.

From-SVN: r236946

PR c++/71227 - specializing hidden friend

* pt.c (check_explicit_specialization): Give better diagnostic about
specializing a hidden friend.

From-SVN: r236941

re PR ada/71317 (Compiling gnat for rtems fails at s-taprop.adb)

* s-osinte-kfreebsd-gnu.ads (clock_getres): Define.
(Get_Page_Size): Remove duplicate and return int.

PR ada/71317
* s-osinte-rtems.ads (clock_getres): Define.
(Get_Page_Size): Remove duplicate and return int.

From-SVN: r236938

re PR c++/71248 (crash on in-class initializer of array of pointer to member)

/cp
2016-05-31  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71248
* decl.c (check_static_variable_definition): Use DECL_SOURCE_LOCATION
to obtain correct locations; avoid redundant diagnostics on
out-of-class definitions.

/testsuite
2016-05-31  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71248
* g++.dg/cpp0x/pr71248.C: New.
* g++.dg/cpp0x/auto7.C: Test column numbers too.
* g++.dg/cpp0x/constexpr-static8.C: Likewise.
* g++.dg/init/new37.C: Likewise.
* g++.dg/template/static1.C: Likewise.
* g++.dg/template/static2.C: Likewise.

From-SVN: r236931

sh.h (ASM_OUTPUT_SYMBOL_REF): Remove macro and use the default implementation.

gcc/
* config/sh/sh.h (ASM_OUTPUT_SYMBOL_REF): Remove macro and use the
default implementation.

From-SVN: r236930

* dwarf2out.c (cur_line_info_table): Add GTY marker.

From-SVN: r236929

Remove SH5 target regs.

gcc/
* config/sh/constraints.md (b): Remove constraint.
* config/sh/predicates.md (arith_reg_operand): Remove TARGET_REGISTER_P.
* config/sh/sh-modes.def (PDI): Remove.
* config/sh/sh.c (sh_target_reg_class,
sh_optimize_target_register_callee_saved): Remove functions.
(sh_option_override): Don't set MASK_SAVE_ALL_TARGET_REGS.
(sh_expand_epilogue): Update comment.
(sh_hard_regno_mode_ok, sh_register_move_cost, calc_live_regs,
sh_secondary_reload): Remove TARGET_REGS related code.
* config/sh/sh.h (FIRST_TARGET_REG, LAST_TARGET_REG,
TARGET_REGISTER_P): Remove macros.
(SH_DBX_REGISTER_NUMBER, REG_ALLOC_ORDER): Remove target regs.
* config/sh/sh.md (PR_MEDIA_REG, T_MEDIA_REG, FR23_REG, TR0_REG,
TR1_REG, TR2_REG): Remove constants.
* config/sh/sh.opt (SAVE_ALL_TARGET_REGS): Remove.

From-SVN: r236928

Simplify SH DImode add, sub, neg patterns.

* config/sh/sh.md (adddi3, subdi3, negdi2, abs<mode>2): Remove
define_expand patterns.
(adddi3_compact): Rename to adddi3.
(subdi3_compact): Rename to subdi3.
(*negdi2): Rename to negdi2.
(*abs<mode>2): Rename to abs<mode>2.

From-SVN: r236927

Fix wrong-code issues of RX atomic operations.

gcc/
* config/rx/rx.md (FETCHOP_NO_MINUS): New code iterator.
(atomic_<fetchop_name>_fetchsi): Extract minus operator into ...
(atomic_sub_fetchsi): ... this new pattern.
(mvtc): Add CC_REG clobber.

From-SVN: r236926

Scan %\[re\]di instead of %rdi

Since x32 may use 32-bit register in address, we should scan %\[re\]di
instead of %rdi.

* gcc.target/i386/avx512vl-vbroadcast-3.c: Scan %\[re\]di
instead of %rdi.
* gcc.target/i386/avx512vl-vcvtps2ph-3.c: Likewise.

From-SVN: r236925

gimplify.c (gimplify_switch_expr): Also handle GIMPLE_TRY.

* gimplify.c (gimplify_switch_expr): Also handle GIMPLE_TRY.

* c-c++-common/Wswitch-unreachable-3.c: New test.
* g++.dg/warn/Wswitch-unreachable-1.C: New test.

From-SVN: r236924

[AArch64] Use aarch64_fusion_enabled_p to check for insn fusion capabilities

* config/aarch64/aarch64.c (aarch_macro_fusion_pair_p): Use
aarch64_fusion_enabled_p to check for fusion capabilities.

From-SVN: r236921

re PR tree-optimization/71352 (ICE at -O1 and above on x86_64-linux-gnu: in zero_one_operation, at tree-ssa-reassoc.c:1251)

2016-05-31  Richard Biener  <rguenther@suse.de>

PR tree-optimization/71352
* tree-ssa-reassoc.c (zero_one_operation): Handle op equal to
minus one and a negate.

* gcc.dg/tree-ssa/reassoc-45.c: New testcase.

From-SVN: r236920

builtins.c (java_builtins): Use popcount* and bswap* builtins to implement bitCount() and reverseBytes()...

2016-05-31  Roger Sayle  <roger@nextmovesoftware.com>

gcc/java:
* builtins.c (java_builtins): Use popcount* and bswap* builtins to
implement bitCount() and reverseBytes() methods in java.lang.Integer
and friends.
(initialize_builtins): Annotate math builtins with ECF_LEAF.  Call
define_builtin for the new popcount* and bswap* builtins.

libjava:
* testsuite/libjava.lang/BuiltinBitCount.java: New test case.
* testsuite/libjava.lang/BuiltinReverseBytes.java: Likewise.

From-SVN: r236919

Don't needlessly clear xmemdup allocated memory.

* xmemdup.c (xmemdup): Use xmalloc rather than xcalloc.

From-SVN: r236917

[AArch64] Remove aarch64_simd_attr_length_move

* config/aarch64/aarch64.c (aarch64_simd_attr_length_move): Delete.
* config/aarch64/aarch64-protos.h (aarch64_simd_attr_length_move):
Delete prototype.
* config/aarch64/iterators.md (insn_count): Add descriptive comment.
* config/aarch64/aarch64-simd.md (*aarch64_mov<mode>, VSTRUCT modes):
Remove use of aarch64_simd_attr_length_move, set length attribute
directly.
(*aarch64_be_movoi): Likewise.
(*aarch64_be_movci): Likewise.
(*aarch64_be_movxi): Likewise.

From-SVN: r236916

loop-init.c (gate): Do not enale RTL loop unroller with -fpeel-loops.

* loop-init.c (gate): Do not enale RTL loop unroller with -fpeel-loops.
It no longer does that.
* toplev.c (process_options): Do not enable flag_web with -fpeel-loops.

From-SVN: r236915

[PATCH] AARCH64: Remove spurious attribute __unused__ from NEON intrinsic

gcc/

* config/aarch64/arm_neon.h (vdupb_laneq_s8): Remove spurious
attribute __unused__.

From-SVN: r236914

arm-protos.h (arm_arch_thumb1): Declare.

2016-05-31  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/
    * config/arm/arm-protos.h (arm_arch_thumb1): Declare.
    * config/arm/arm.c (arm_arch_thumb1): Define.
    (arm_option_override): Initialize arm_arch_thumb1.
    * config/arm/arm.h (arm_arch_thumb1): Declare.
    (TARGET_ARM_ARCH_ISA_THUMB): Use arm_arch_thumb to determine if target
    support Thumb-1 ISA.

    gcc/testsuite/
    * gcc.target/arm/armv5_thumb_isa.c: New test.

From-SVN: r236913

Make analyze_brprob.py executable.

From-SVN: r236911

AVX-512. Limit constraint for scalar operand in split to AVX-512VL.

PR target/71346
gcc/
* config/i386/sse.md (define_insn_and_split "*vec_extractv4sf_0"): Use
`Yv' for scalar operand.
testsuite/
* gcc.target/i386/pr71346.c: New test.

From-SVN: r236909

Handle 3-arg phi in copy_bb_and_scalar_dependences

2016-05-31  Tom de Vries  <tom@codesourcery.com>

PR tree-optimization/69068
* graphite-isl-ast-to-gimple.c (copy_bb_and_scalar_dependences): Handle
phis with more than two args.

* gcc.dg/graphite/pr69068.c: New test.

From-SVN: r236907

Daily bump.

From-SVN: r236906

PR c++/71306 - bogus -Wplacement-new with an array element

gcc/cp/ChangeLog:
2016-05-27  Martin Sebor  <msebor@redhat.com>

PR c++/71306
* init.c (warn_placement_new_too_small): Handle placement new arguments
that are elements of arrays more carefully.  Remove a pointless loop.

gcc/testsuite/ChangeLog:
2016-05-27  Martin Sebor  <msebor@redhat.com>

PR c++/71306
* g++.dg/warn/Wplacement-new-size-3.C: New test.

From-SVN: r236902

re PR c++/71349 (Combined async target clause parsing issues)

PR c++/71349
* c-parser.c (c_parser_omp_for): Don't disallow nowait clause
when combined with target construct.

* parser.c (cp_parser_omp_for): Don't disallow nowait clause
when combined with target construct.
(cp_parser_omp_parallel): Pass cclauses == NULL as last argument
to cp_parser_omp_all_clauses.

* c-omp.c (c_omp_split_clauses): Put OMP_CLAUSE_DEPEND to
C_OMP_CLAUSE_SPLIT_TARGET.  Put OMP_CLAUSE_NOWAIT to
C_OMP_CLAUSE_SPLIT_TARGET if combined with target construct,
instead of C_OMP_CLAUSE_SPLIT_FOR.

* c-c++-common/gomp/clauses-1.c (bar): Add dd argument.  Add
nowait depend(inout: dd[0]) clauses where permitted.

From-SVN: r236900

config.gcc: Move hard float support for arm*hf*-*-freebsd* into armv6*-*-freebsd* for...

2016-05-30  Andreas Tobler  <andreast@gcc.gnu.org>

    * config.gcc: Move hard float support for arm*hf*-*-freebsd* into
    armv6*-*-freebsd* for FreeBSD 11. Eliminate the arm*hf*-*-freebsd*
    target.

From-SVN: r236898

config.gcc (sparc*-*-*): Support cpu_32, cpu_64, tune_32 and tune_64.

2016-05-30  Jose E. Marchesi  <jose.marchesi@oracle.com>

* config.gcc (sparc*-*-*): Support cpu_32, cpu_64, tune_32 and
tune_64.
* doc/install.texi (--with-cpu-32, --with-cpu-64): Document
support on SPARC.
* config/sparc/linux64.h (OPTION_DEFAULT_SPECS): Add entries for
cpu_32, cpu_64, tune_32 and tune_64.
* config/sparc/sol2.h (OPTION_DEFAULT_SPECS): Likewise.

From-SVN: r236897

re PR c++/71238 (Undeclared function message imprecisely points to error column)

/cp
2016-05-30  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71238
* lex.c (unqualified_name_lookup_error): Take a location too.
(unqualified_fn_lookup_error): Take a cp_expr.
* cp-tree.h (unqualified_name_lookup_error,
unqualified_fn_lookup_error): Adjust declarations.
* semantics.c (perform_koenig_lookup): Adjust
unqualified_fn_lookup_error call, pass the location of
the identifier too as part of a cp_expr.

/testsuite
2016-05-30  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71238
* g++.dg/parse/pr71238.C: New.
* g++.dg/concepts/friend1.C: Test column numbers too.
* g++.dg/cpp0x/initlist31.C: Likewise.
* g++.dg/cpp0x/pr51420.C: Likewise.
* g++.dg/cpp0x/udlit-declare-neg.C: Likewise.
* g++.dg/cpp0x/udlit-member-neg.C: Likewise.
* g++.dg/ext/builtin3.C: Likewise.
* g++.dg/lookup/friend12.C: Likewise.
* g++.dg/lookup/friend7.C: Likewise.
* g++.dg/lookup/koenig1.C: Likewise.
* g++.dg/lookup/koenig5.C: Likewise.
* g++.dg/lookup/used-before-declaration.C: Likewise.
* g++.dg/overload/koenig1.C: Likewise.
* g++.dg/template/crash65.C: Likewise.
* g++.dg/template/friend57.C: Likewise.
* g++.dg/warn/Wshadow-5.C: Likewise.
* g++.dg/warn/Wunused-8.C: Likewise.
* g++.old-deja/g++.bugs/900211_01.C: Likewise.
* g++.old-deja/g++.jason/lineno5.C: Likewise.
* g++.old-deja/g++.jason/member.C: Likewise.
* g++.old-deja/g++.jason/report.C: Likewise.
* g++.old-deja/g++.jason/scoping12.C: Likewise.
* g++.old-deja/g++.law/visibility20.C: Likewise.
* g++.old-deja/g++.ns/koenig5.C: Likewise.
* g++.old-deja/g++.other/static5.C: Likewise.
* g++.old-deja/g++.pt/overload2.C: Likewise.

From-SVN: r236896

sync.md (mfence_nosse): Use "lock orl $0, -4(%esp)".

* config/i386/sync.md (mfence_nosse): Use "lock orl $0, -4(%esp)".

From-SVN: r236895

Don't cause ICEs when auto profile file is not found with checking

Currently, on a checking enabled compiler when -fauto-profile does
not find the profile feedback file it errors out with assertation
failures. Add proper errors for this case.

gcc/:

2016-05-30  Andi Kleen  <ak@linux.intel.com>

* auto-profile.c (read_profile): Replace asserts with errors
when file does not exist.
* gcov-io.c (gcov_read_words): Dito.

From-SVN: r236894

tree-cfg.c (print_loop): Print likely upper bounds.

* tree-cfg.c (print_loop): Print likely upper bounds.

From-SVN: r236893

invoke.texi (-fpeel-loops,-O3): Update documentation.

* doc/invoke.texi (-fpeel-loops,-O3): Update documentation.
* opts.c (default_options): Enable peel loops at -O3.
* tree-ssa-loop-ivcanon.c (peeled_loops): New static var.
(try_peel_loop): Do not re-peel already peeled loops;
use likely upper bounds; fix profile updating.
(pass_complete_unroll::execute): Initialize peeled_loops.

* gcc.dg/tree-ssa/peel1.c: New testcase.
* gcc.dg/tree-ssa/pr61743-1.c: Disable loop peeling.
* gcc.dg/tree-ssa/pr61743-2.c: Disable loop peeling.

From-SVN: r236892

Add profiling support for IVOPTS

* tree-ssa-loop-ivopts.c (get_computation_cost_at): Scale
computed costs by frequency of BB they belong to.
(get_scaled_computation_cost_at): New function.

From-SVN: r236888

Fix pr69067.c testcase

2016-05-30  Tom de Vries  <tom@codesourcery.com>

* gcc.dg/graphite/pr69067.c (main): Remove superfluous argument in call
to ce.

From-SVN: r236886

re PR c++/71099 (Misleading diagnostic message with 'virtual' used in out-of-line definitions of class template member functions)

/cp
2016-05-30  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71099
* parser.c (cp_parser_function_specifier_opt): Use current_class_type
to improve the diagnostic about wrong uses of 'virtual'.

/testsuite
2016-05-30  Paolo Carlini  <paolo.carlini@oracle.com>

PR c++/71099
* g++.dg/parse/virtual1.C: New.

From-SVN: r236885

args.h (clear_non_sret_int_hardware_registers): Use correct register when clearing %edx.

* gcc.target/i386/iamcu/args.h (clear_non_sret_int_hardware_registers):
Use correct register when clearing %edx.

From-SVN: r236883

match.pd: optimize unsigned mul overflow check

gcc/
2016-05-28  Alexander Monakov  <amonakov@ispras.ru>
            Marc Glisse  <marc.glisse@inria.fr>

PR tree-optimization/71289
* match.pd (-1 / B < A, A > -1 / B): New transformations.

gcc/testsuite/
2016-05-28  Alexander Monakov  <amonakov@ispras.ru>

PR tree-optimization/71289
* gcc.dg/pr71289.c: New test.

Co-Authored-By: Marc Glisse <marc.glisse@inria.fr>
From-SVN: r236882

* tree-vect-loop.c (vect_transform_loop): Update likely bounds.

From-SVN: r236880

tree-ssa-loop-ivcanon.c (try_peel_loop): Correctly set wont_exit for peeled copies...

* tree-ssa-loop-ivcanon.c (try_peel_loop): Correctly set wont_exit
for peeled copies; avoid underflow when updating estimates; correctly
scale loop profile.

From-SVN: r236878