Remove a stray line.

From-SVN: r255558

PR c/81544 - attribute noreturn and warn_unused_result on the same function

PR c/81544 - attribute noreturn and warn_unused_result on the same function
accepted

Add tests missed in r255469.

From-SVN: r255557

altivec-12.c (main): Add tests for vec_avg.

gcc/testsuite/ChangeLog:

2017-12-07  Carl Love  <cel@us.ibm.com>

* gcc.target/powerpc/altivec-12.c (main):  Add tests for vec_avg.
* gcc.target/powerpc/builtins-1.c: Renamed builtins-1.h.
* gcc.target/powerpc/builtins-1.h: Add tests for vec_and,
vec_andc, vec_cpsgn, vec_mergeh,vec_mergel, vec_nor, vec_or,
vec_sel, vec_xor, vec_all_eq, vec_all_ge, vec_all_gt, vec_all_le,
vec_all_lt, vec_all_ne, vec_any_eq,vec_any_ge, vec_any_gt,vec_any_le,
vec_any_lt, vec_any_ne, vec_cntlz, vec_round, vec_rsqrt, vec_rsqrte,
vec_splat, vec_div, vec_mul,vec_ctf, vec_cts, vec_ctu.
* gcc.target/powerpc/builtins-1-be.c: New file, big endian
test file for tests in builtins-1.h.
* gcc.target/powerpc/builtins-1-le.c: New file, little endian
test file for tests in builtins-1.h.
* gcc.target/powerpc/builtins-2.c (main): Add tests for vec_ctf,
vec_cts, vec_ctu.
* gcc.target/powerpc/builtins-3-runnable.c (test_result_sp): Fix printf
conversion character.
  * gcc.target/powerpc/fold-vec-minmax-char.c (test1_min, test2_min,
test4_min, test1_max, test2_max): Add tests for vec_max and vec_min.
  * gcc.target/powerpc/fold-vec-minmax-int.c (test4_min, test5_min,
test7_min, test8_min): Add tests for vec_min.
  * gcc.target/powerpc/fold-vec-minmax-longlong.c (test7_max, test8_max,
test9_max, test10_max): Add tests for and vec_max.
  * gcc.target/powerpc/fold-vec-minmax-short.c (test4_min, test5_min,
tes7_min, test8_min): Add tests for and vec_min.
  * gcc.target/powerpc/p8vector-builtin-8.c (foo): Add test for vec_adds,
vec_subs, vec_sum2s, vec_sum4s.
  * gcc.target/powerpc/vsx-splats-p8.c: New file to test vec_splats.
* gcc.target/vsx-vector-5.c (arg1_tests_d, arg1_tests_d): Add
tests for vec_ceil, vec_floor, vec_nearbyint, vec_sqrt.
* gcc.target/vsx-vector-7.c (foo): Add tests for vec_sel.
* gcc.target/vsx-vector-abss.c: New file to test vec_abss.

From-SVN: r255556

altivec.h (vec_extract_fp32_from_shorth, [...]): Add #defines.

gcc/ChangeLog:

2017-12-11  Carl Love  <cel@us.ibm.com>

* config/rs6000/altivec.h (vec_extract_fp32_from_shorth,
vec_extract_fp32_from_shortl]): Add #defines.
* config/rs6000/rs6000-builtin.def (VSLDOI_2DI): Add macro expansion.
* config/rs6000/rs6000-c.c (ALTIVEC_BUILTIN_VEC_UNPACKH,
ALTIVEC_BUILTIN_VEC_UNPACKL, ALTIVEC_BUILTIN_VEC_AND,
ALTIVEC_BUILTIN_VEC_SLD, ALTIVEC_BUILTIN_VEC_SRL,
ALTIVEC_BUILTIN_VEC_SRO, ALTIVEC_BUILTIN_VEC_SLD,
ALTIVEC_BUILTIN_VEC_SLL): Add expansions.
* doc/extend.texi: Add documentation for the added builtins.

gcc/testsuite/ChangeLog:

2017-12-11  Carl Love  <cel@us.ibm.com>
* gcc.target/powerpc/altivec-7.c: Renamed altivec-7.h.
* gcc.target/powerpc/altivec-7.h (main): Add testcases for vec_unpackl.
Add dg-final tests for the instructions generated.
* gcc.target/powerpc/altivec-7-be.c: New file to test on big endian.
* gcc.target/powerpc/altivec-7-le.c: New file to test on little endian.
* gcc.target/powerpc/altivec-13.c (foo): Add vec_sld, vec_srl,
 vec_sro testcases. Add dg-final tests for the instructions generated.
* gcc.target/powerpc/builtins-3-p8.c (test_vsi_packs_vui,
test_vsi_packs_vsi, test_vsi_packs_vssi, test_vsi_packs_vusi,
test_vsi_packsu-vssi, test_vsi_packsu-vusi, test_vsi_packsu-vsll,
test_vsi_packsu-vull, test_vsi_packsu-vsi, test_vsi_packsu-vui): Add
testcases. Add dg-final tests for new instructions.
* gcc.target/powerpc/p8vector-builtin-2.c (vbschar_eq, vbchar_eq,
vuchar_eq, vbint_eq, vsint_eq, viint_eq, vuint_eq, vbool_eq, vbint_ne,
vsint_ne, vuint_ne, vbool_ne, vsign_ne, vuns_ne, vbshort_ne): Add
tests.
Add dg-final instruction tests.
* gcc.target/powerpc/vsx-vector-6.c: Renamed vsx-vector-6.h.
* gcc.target/powerpc/vsx-vector-6.h (vec_andc,vec_nmsub, vec_nmadd,
vec_or, vec_nor, vec_andc, vec_or, vec_andc, vec_msums): Add tests.
Add dg-final tests for the generated instructions.
* gcc.target/powerpc/builtins-3.c (test_sll_vsc_vsc_vsuc,
test_sll_vuc_vuc, test_sll_vsi_vsi_vuc, test_sll_vui_vui_vuc,
test_sll_vbll_vull, test_sll_vbll_vbll_vus, test_sll_vp_vp_vuc,
test_sll_vssi_vssi_vuc, test_sll_vusi_vusi_vuc, test_slo_vsc_vsc_vsc,
test_slo_vuc_vuc_vsc, test_slo_vsi_vsi_vsc, test_slo_vsi_vsi_vuc,
test_slo_vui_vui_vsc, test_slo_vui_vui_vuc, test_slo_vsll_slo_vsll_vsc,
test_slo_vsll_slo_vsll_vuc, test_slo_vull_slo_vull_vsc,
test_slo_vull_slo_vull_vuc, test_slo_vp_vp_vsc, test_slo_vp_vp_vuc,
test_slo_vssi_vssi_vsc, test_slo_vssi_vssi_vuc, test_slo_vusi_vusi_vsc,
test_slo_vusi_vusi_vuc, test_slo_vusi_vusi_vuc, test_slo_vf_vf_vsc,
test_slo_vf_vf_vuc, test_cmpb_float): Add tests.

From-SVN: r255555

[PR80693] drop value of parallel SETs dropped by combine

When combine drops a REG_UNUSED SET in a parallel, we have to clear
cached values, so that, even if the REGs remain used (e.g. because
they were referenced in the used SET_SRC), we will not use properties
of the dropped modified value as if they applied to the preserved
original one.

We fail to adjust REG_N_SETS.

for  gcc/ChangeLog

PR rtl-optimization/80693
PR rtl-optimization/81019
PR rtl-optimization/81020
* combine.c (distribute_notes): Reset any REG_UNUSED REGs that
are not mentioned in i3.  Place the REG_UNUSED note on i2,
possibly modified to REG_DEAD, if it did not originate in i3.

for  gcc/testsuite/ChangeLog

PR rtl-optimization/80693
PR rtl-optimization/81019
PR rtl-optimization/81020
* gcc.dg/pr80693.c: New.
* gcc.dg/pr81019.c: New.

From-SVN: r255554

recog.c (store_data_bypass_p_1): New function.

* recog.c (store_data_bypass_p_1): New function.
(store_data_bypass_p): Handle USE in a PARALLEL like CLOBBER.  Use
store_data_bypass_p_1 to avoid code duplication.  Formatting fixes.

From-SVN: r255553

ifcvt: Call fixup_partitions (PR83361)

After converting a conditional branch to an unconditional trap to a
conditional trap, if the original trap is still reachable from another
path, it may be that it is in a hot basic block  and only reachable from
cold blocks.  Fix that up.

PR rtl-optimization/83361
* ifcvt.c (if_convert): Call fixup_partitions.

gcc/testsuite/
PR rtl-optimization/83361
* gcc.dg/pr83361.c: New testcase.

From-SVN: r255550

rs6000.c (rs6000_gimple_fold_builtin): Add support for early folding of splat_u{8,16,32}.

[gcc]

2017-12-11  Will Schmidt  <will_schmidt@vnet.ibm.com>

* config/rs6000/rs6000.c (rs6000_gimple_fold_builtin): Add support for
early folding of splat_u{8,16,32}.

[testsuite]

2017-12-11  Will Schmidt  <will_schmidt@vnet.ibm.com>

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

From-SVN: r255549

intrinsic.texi (MAXLOC): Update documentation for character arrays and KIND argument.

2017-12-11  Thomas Koenig  <tkoenig@gcc.gnu.org>

* intrinsic.texi (MAXLOC): Update documentation for
character arrays and KIND argument.
(MINLOC): Likewise.

From-SVN: r255546

aarch64.c (aarch64_print_operand): Don't start output_operand_lossage first argument with capital letter.

* config/aarch64/aarch64.c (aarch64_print_operand): Don't start
output_operand_lossage first argument with capital letter.
(aarch64_override_options): Don't start error and sorry first argument
with capital letter.

From-SVN: r255544

re PR fortran/83379 (Intermittent failure of min/maxval_char*.f90 test cases)

2017-12-11  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/83379
* gfortran.dg/minval_char_1.f90: Correct logic for mask
calculation.
* gfortran.dg/minval_char_2.f90: Likewise.
* gfortran.dg/maxval_char_1.f90: Likewise.
* gfortran.dg/maxval_char_2.f90: Likewise.

From-SVN: r255543

re PR fortran/45689 ([F03] Missing transformational intrinsic in the trans_func_f2003 list)

2017-12-11  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/45689
* simplify.c (min_max_choose): Add prototype.
(gfc_count): Format correctly.
(simplify_transformation): Pass array argument to init_result_expr.
(gfc_simplify_minval_maxval): Remove.
(gfc_min): New function.
(gfc_simplify_minval): Call simplify_transformation.
(gfc_max): New function.
(gfc_simplify_maxval): Call simplify_transformation.

2017-12-11  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/45689
* gfortran.dg/minval_parameter_1.f90: New test.
* gfortran.dg/maxval_parameter_1.f90: New test.

From-SVN: r255542

Fix stack overflow with autofdo (PR83355)

g++.dg/bprob* is failing currently with autofdo.

Running in gdb shows that there is a very deep recursion in get_index_by_decl until it
overflows the stack.

gcc/:
2017-12-11  Andi Kleen  <ak@linux.intel.com>

PR gcov-profile/83355
* auto-profile.c (string_table::get_index_by_decl): Don't
recurse when abstract origin points to itself.

From-SVN: r255540

re PR tree-optimization/83320 (Mismatched free() / delete / delete []  in gimple-loop-interchange.cc:948)

PR tree-optimization/83320
* gimple-loop-interchange.cc (free_data_refs_with_aux): Use delete.
(prune_datarefs_not_in_loop): Ditto.

From-SVN: r255539

Daily bump.

From-SVN: r255537

* ChangeLog: Fix whitespace some more.

From-SVN: r255534

ChangeLog: Fix whitespace.

* ChangeLog: Fix whitespace.
* testsuite/ChangeLog: Ditto.

From-SVN: r255533

re PR fortran/53478 (gfortran segfaults when module name clashes with C binding name of procedure)

2017-12-10  Dominique d'Humieres  <dominiq@lps.ens.fr>

PR fortran/53478
* gfortran.h (gfc_find_case_gsymbol): New prototype.
* symbol.c (gfc_find_case_gsymbol): New procedure, case
insensistive version of gfc_find_gsymbol.
* resolve.c (resolve_common_blocks): Use it.
Replace %s with %qs where needed.

* gfortran.dg/binding_label_tests_4.f03: Update dg-error.
* gfortran.dg/binding_label_tests_6.f03: Likewise.
* gfortran.dg/binding_label_tests_7.f03: Likewise.
* gfortran.dg/binding_label_tests_8.f03: Likewise.
* gfortran.dg/binding_label_tests_10_main.f03: Likewise.
* gfortran.dg/binding_label_tests_11_main.f03: Likewise.
* gfortran.dg/binding_label_tests_13_main.f03: Likewise.
* gfortran.dg/test_common_binding_labels_3_main.f03: Likewise.
* gfortran.dg/binding_label_tests_29.f90: New test.

From-SVN: r255530

* doc/install.texi (Specific): Tweak link to mkssoftware.com.

From-SVN: r255529

re PR tree-optimization/83337 (ICE at -O3 x86_64-linux-gnu: in interpret_rhs_expr, at tree-scalar-evolution.c:1775)

PR tree-optimization/83337
* gimple-loop-interchange.cc (compute_access_stride): Handle bitfield DRs
properly.

* gcc.dg/tree-ssa/loop-interchange-14.c: New test.
* gcc.dg/tree-ssa/loop-interchange-15.c: New test.

From-SVN: r255528

Daily bump.

From-SVN: r255527

re PR fortran/82934 (Segfault on assumed character length in allocate)

2017-12-09  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/82934
PR fortran/83318
* match.c (gfc_match_allocate): Enforce F2008:C631.

2017-12-09  Steven G. Kargl  <kargl@gcc.gnu.org>

PR fortran/82934
PR fortran/83318
* gfortran.dg/allocate_assumed_charlen_2.f90: new test.

From-SVN: r255524

re PR lto/83338 (SPEC CPU2017 510.parest_r ICE)

PR tree-optimization/83338
* tree-vect-stmts.c (vectorizable_operation): Handle POINTER_DIFF_EXPR
vectorization as MINUS_EXPR with a subsequent VIEW_CONVERT_EXPR from
vector of unsigned integers to vector of signed integers.

* gcc.dg/vect/pr83338.c: New test.

From-SVN: r255523

re PR fortran/83316 (ICE: minval/maxval and characters)

2017-12-09  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/83316
* arith.c (gfc_character2character): New function.
* arith.h: Add prototype.
* simplify.c (gfc_convert_constant): Handle character type.

2017-12-09  Thomas Koenig  <tkoenig@gcc.gnu.org>

PR fortran/83316
* gfortran.dg/minval_char_5.f90: New test.

From-SVN: r255522

Add .type and .size directives to riscv libgcc functions.

libgcc/
* config/riscv/div.S: Use FUNC_* macros.
* config/riscv/muldi3.S, config/riscv/multi3.S: Likewise
* config/riscv/save-restore.S: Likewise.
* config/riscv/riscv-asm.h: New.

From-SVN: r255521

Daily bump.

From-SVN: r255520

[multiple changes]

2017-12-08  Vladimir Makarov  <vmakarov@redhat.com>

PR rtl-optimization/83317
* lra-constraints.c (process_address_1): Add insn code check.

2017-12-08  Jakub Jelinek  <jakub@redhat.com>

PR rtl-optimization/83317
* gcc.target/i386/pr83317.c: New test.

From-SVN: r255517

re PR rtl-optimization/81595 (ICE: verify_flow_info failed (error: multiple hot/cold transitions found))

PR rtl-optimization/81595
* gcc.c-torture/compile/pr81595.c: New test.

From-SVN: r255516

re PR tree-optimization/83323 (186.crafty miscompares)

Fix PR83323

* gimple-loop-jam (unroll_jam_possible_p): Correct test for
head-controlled loops and loop BBs.
* common.opt (funroll-and-jam): Remove, instead ...
(floop-unroll-and-jam): ... reuse this option.
* opts.c (default_options_table): Use OPT_floop_unroll_and_jam.
* doc/invoke.texi (-funroll-and-jam): Move docu to ...
(-floop-unroll-and-jam): ... this option.

testsuite/
* gcc.dg/pr83323.c: New test.
* gcc.dg/unroll-and-jam.c: Use -floop-unroll-and-jam.

From-SVN: r255514

ipa-polymorphic-call.c (noncall_stmt_may_be_vtbl_ptr_store): Fix a comment typo, get_base_ref_and_offset -> get_ref_base_and_extent.

* ipa-polymorphic-call.c (noncall_stmt_may_be_vtbl_ptr_store): Fix
a comment typo, get_base_ref_and_offset -> get_ref_base_and_extent.
* ipa-prop.c (stmt_may_be_vtbl_ptr_store): Likewise.

From-SVN: r255513

re PR middle-end/81782 (Yet another -Wmaybe-uninitialized false positive with empty array)

2017-12-08  Richard Biener  <rguenther@suse.de>

PR middle-end/81782
* tree-ssa-uninit.c (warn_uninitialized_vars): Properly
handle accesses outside of zero-sized vars.

* gcc.dg/uninit-pr81782.c: New testcase.

From-SVN: r255512

Disable hwcap on Solaris in g++.dg/opt/pr83252.C

* g++.dg/opt/pr83252.C: Add -mclear-hwcap on *-*-solaris*.

From-SVN: r255511

Prevent SRA from removing type changing assignment

2017-12-08  Martin Jambor  <mjambor@suse.cz>

PR tree-optimization/83141
* tree-sra.c (contains_vce_or_bfcref_p): Move up in the file, also
test for MEM_REFs implicitely changing types with padding.  Remove
inline keyword.
(build_accesses_from_assign): Added contains_vce_or_bfcref_p checks.

testsuite/
* gcc.dg/tree-ssa/pr83141.c: New test.
* gcc.dg/guality/pr54970.c: XFAIL tests querying a[0].

From-SVN: r255510

arc.c (arc_attribute_table): Add exclusions to the comment.

* config/arc/arc.c (arc_attribute_table): Add exclusions to
the comment.
* config/avr/avr.c (avr_attribute_table): Likewise.
* config/msp430/msp430.c (msp430_attribute_table): Likewise.
* config/rl78/rl78.c (rl78_attribute_table): Likewise.
* config/nds32/nds32.c (nds32_attribute_table): Likewise.
* config/darwin.h (SUBTARGET_ATTRIBUTE_TABLE): Initialize new member
of struct attribute_spec.
* config/i386/cygming.h (SUBTARGET_ATTRIBUTE_TABLE): Likewise.
ada/
* gcc-interface/utils.c (gnat_internal_attribute_table): Add
exclusions to the comment.
brig/
* brig-lang.c (brig_attribute_table): Fix up comment.

From-SVN: r255509

re PR target/82960 (spu_machine_dependent_reorg does not handle jump_table_data insn)

gcc/
PR target/82960
* config/spu/spu.c (pad_bb): Only check INSN_CODE when INSN_P is true.

From-SVN: r255508

re PR tree-optimization/83609 (ICE in read_complex_part at gcc/expr.c:3202)

* profile-count.c (profile_count::from_gcov_type): Move from
profile-count.h; handle overflow.
* profile-count. (profile_count::from_gcov_type): Move offline.

PR middle-end/83609
* gcc.c-torture/compile/pr83069.c: New testcase.

From-SVN: r255507

combine: Fix PR83304

In PR83304 two insns are combined, where the I2 uses a register that
has a REG_DEAD note on an insn after I2 but before I3.  In such a case
move_deaths should move that death note.  But move_deaths only looks
at the reg_stat[regno].last_death insn, and that field can be zeroed
out (previously, use_crosses_set_p would prevent the combination in
this case).

If the last_death field is zero it means "unknown", not "no death", so
we have to find if there is a REG_DEAD note.

PR rtl-optimization/83304
* combine.c (move_deaths): If we do not know where a register died,
search for it.

From-SVN: r255506

gimple-loop-interchange.cc (tree_loop_interchange::interchange): Provide -fopt-info-loop feedback when we interchange in a nest.

2017-12-08  Richard Biener  <rguenther@suse.de>

* gimple-loop-interchange.cc (tree_loop_interchange::interchange):
Provide -fopt-info-loop feedback when we interchange in a nest.

From-SVN: r255505

[arm] PR target/83206: Make native driver select fp-capable armv6 cores

A quirk in the historical naming of some ARMv6 products means that the
main CPU name implies the presence or otherwise of the floating point unit.
This causes problems when using -mfpu=auto with -mcpu=native: the driver is
picking a CPU that does not support a floating-point unit, even though
one may well exist.

This patch addresses this by selecting the FP-capable names so that FP
instructions will be generated if the other options suggest this is
permitted.

Note that a more complete fix is really needed here to look up the
FP/simd capabilities and append the appropriate capability extensions.
This will be the subject of some follow-up patches.

* config/arm/driver-arm.c (arm_cpu_table): Use fp-capable product names
for armv6 ARM CPU IDs.

From-SVN: r255504

[arm] Don't strip off all architecture features from -march passed to assembler

When GCC invokes the assembler it generates a sanitized version of the
user-specified -march option to pass through, since the assembler does
not understand all the new FPU-related architectural options.
Unfortunately it goes too far and strips off all the architectural
extensions, including some that are unrelated to the -mfpu variant
selected.

Again, this doesn't really matter when compiling C code because the
compiler will override the command-line specified architecture with
directives in the assembly file itself, but when using the compiler
driver to invoke the assembler the only indiciation of the desired
architecture might come from the command line.

We fix this by adjusting the canonicalization pass to remove any
option that only specifies features that can be expressed by -mfpu
(any that go beyond that are already supported by the assembler).  We
do have to take care to re-order the options, though as the assembler
expects feature options to be in a canonical order (unlike the
compiler, where ordering is handled left-to-right: there's only a
difference if there are negation options, but a canonicalized
architecture string shouldn't have any of those).  We do this by
recording which options we need and then sorting the final list
alphabetically.

* common/config/arm/arm-common.c: Include <algorithm>.
(INCLUDE_VECTOR): Define.
(compare_opt_names): New function.
(arm_rewrite_selected_arch): Only strip out extensions that can be
expressed through -mfpu.  Sort the remaining extensions
alphabetically.

From-SVN: r255503

[arm] Generate a -mfpu= option for passing to the assembler

When gcc runs with the new -mfpu=auto option (either explicitly or
when that's the default behaviour) then this option is not passed
through to the assembler as we cannot rely on the assembler
understanding it (currently it doesn't understand it at all, but in
future that might change).  That means that the assembler falls back
to its builtin default, which may not correspond to what the user
expected based on the command-line options they passed.

Normally that wouldn't matter because assembler files generated by the
compiler will contain explicit directives that set the FPU type
directly and override any internal defaults; but when the compiler
driver is used to invoke the assembler directly (because the source
file ends in .s or .S) then this might cause a problem if that assumes
the FPU matches the compiler.

To address this, this patch makes the driver construct a -mfpu= option
for the assembler in the same way as the compiler generates an
internal .fpu directive.  As mentioned, this makes no difference if
the assembler file explicitly overrides the command line options, but
helps in the case where this is implicit.

* config/arm/arm.h (arm_asm_auto_mfpu): Declare.
(ASM_CPU_SPEC_FUNCTIONS): Add new rule asm_auto_mfpu.
(ASM_CPU_SPEC): Use it if -mfpu is set to auto.
* common/config/arm/arm-common.c (arm_asm_auto_mfpu): New function.
-- This line, and those below, will be ignored--

M    gcc/ChangeLog
M    gcc/common/config/arm/arm-common.c
M    gcc/config/arm/arm.h

From-SVN: r255502

re PR bootstrap/81470 (Bootstrap comparison failures in gcc/ada)

2017-06-08  Tristan Gingold  <gindold@adacore.com>

PR ada/81470
* dwarf2out.c (dwarf2out_do_cfi_startproc): Only emit
.cfi_personality or .cfi_lsda if the eh data format is dwarf2.

From-SVN: r255501

Initialize new member of SOLARIS_ATTRIBUTE_TABLE

* config/sol2.h (SOLARIS_ATTRIBUTE_TABLE): Initialize new member
of struct attribute_spec.

From-SVN: r255500

re PR tree-optimization/81303 (410.bwaves regression caused by r249919)

2017-12-08  Richard Biener  <rguenther@suse.de>

PR tree-optimization/81303
* gfortran.dg/pr81303.f: New testcase.
* gfortran.dg/vect/pr81303.f: Likewise.

From-SVN: r255499

Enable VNNI support [5/5]

gcc/
* config/i386/avx512vnniintrin.h (_mm512_dpwssds_epi32,
_mm512_mask_dpwssds_epi32, _mm512_maskz_dpwssds_epi32): New intrinsics.
* config/i386/avx512vnnivlintrin.h (_mm256_dpwssds_epi32,
_mm256_mask_dpwssds_epi32, _mm256_maskz_dpwssds_epi32,
_mm_dpwssds_epi32, _mm_mask_dpwssds_epi32,
_mm_maskz_dpwssds_epi32): Ditto.

gcc/testsuite/
* gcc.target/i386/avx512f-vnni-1.c: Add checks for vdpdwssds.
* gcc.target/i386/avx512vl-vnni-1.c: Ditto.
* gcc.target/i386/avx512f-vpdpwssds-2.c: New test.
* gcc.target/i386/avx512vl-vpdpwssds-2.c: Ditto.

From-SVN: r255498

re PR tree-optimization/81303 (410.bwaves regression caused by r249919)

2017-12-08  Richard Biener  <rguenther@suse.de>

PR tree-optimization/81303
* tree-vect-stmts.c (vect_is_simple_cond): For invariant
conditions try to create a comparison vector type matching
the data vector type.
(vectorizable_condition): Adjust.
* tree-vect-patterns.c (vect_recog_mask_conversion_pattern):
Leave invariant conditions alone in case we can vectorize those.

* gcc.target/i386/vectorize9.c: New testcase.
* gcc.target/i386/vectorize10.c: New testcase.

From-SVN: r255497

Enable VNNI support [4/5]

gcc/
* config/i386/avx512vnniintrin.h (_mm512_dpwssd_epi32,
_mm512_mask_dpwssd_epi32, _mm512_maskz_dpwssd_epi32): New intrinsics.
* config/i386/avx512vnnivlintrin.h (_mm256_dpwssd_epi32,
_mm256_mask_dpwssd_epi32, _mm256_maskz_dpwssd_epi32, _mm_dpwssd_epi32,
_mm_mask_dpwssd_epi32, _mm_maskz_dpwssd_epi32): Ditto.

gcc/testsuite/
* gcc.target/i386/avx512f-vnni-1.c: Add vdpwssd checks.
* gcc.target/i386/avx512vl-vnni-1.c: Ditto.
* gcc.target/i386/avx512f-vpdpwssd-2.c: New.
* gcc.target/i386/avx512vl-vpdpwssd-2.c: Ditto.

From-SVN: r255496

Enable VAES support [2/5]

gcc/
* config.gcc: Add vaesintrin.h.
* config/i386/i386-builtin-types.def (V64QI_FTYPE_V64QI_V64QI): New type.
* config/i386/i386-builtin.def (__builtin_ia32_vaesdec_v16qi,
__builtin_ia32_vaesdec_v32qi, __builtin_ia32_vaesdec_v64qi): New builtins.
* config/i386/i386.c (ix86_expand_args_builtin): Handle new type.
* config/i386/immintrin.h: Include vaesintrin.h.
* config/i386/sse.md (vaesdec_<mode>): New pattern.
* config/i386/vaesintrin.h (_mm256_aesdec_epi128, _mm512_aesdec_epi128,
_mm_aesdec_epi128): New intrinsics.

gcc/testsuite/
* gcc.target/i386/avx512-check.h: Handle bit_VAES.
* gcc.target/i386/avx512f-aesdec-2.c: New test.
* gcc.target/i386/avx512fvl-vaes-1.c: Ditto.
* gcc.target/i386/avx512vl-aesdec-2.c: Ditto.
* gcc.target/i386/i386.exp (check_effective_target_avx512vaes): New.

From-SVN: r255495

const-addr-1.c: New.

2017-12-07  Sandra Loosemore  <sandra@codesourcery.com>

gcc/testsuite/
* gcc.target/nios2/const-addr-1.c: New.
* gcc.target/nios2/const-addr-2.c: New.
* gcc.target/nios2/const-addr-3.c: New.

From-SVN: r255494

nios2.md (ld<bhw_uns>io): Add splitter for memory operand.

2017-12-07  Sandra Loosemore  <sandra@codesourcery.com>

gcc/
* config/nios2/nios2.md (ld<bhw_uns>io): Add splitter for memory
operand.
(ld<bh>io_signed): Likewise.
(st<bhw>io): Likewise.
* config/nios2/predicates.md (ldstio_memory_operand): Allow
SMALL_INT12 constant integer operand.

From-SVN: r255493

nios2.c (nios2_symbolic_constant_allowed): Rename to...

2017-12-07  Sandra Loosemore  <sandra@codesourcery.com>

gcc/
* config/nios2/nios2.c (nios2_symbolic_constant_allowed):
Rename to...
(nios2_large_constant_allowed): ...this.  Adjust uses.
(nios2_plus_symbolic_constant_p): Rename to...
(nios2_plus_large_constant_p): ...this.  Adjust uses.
(nios2_legitimate_address_p): Correct CONST_INT handling.
(nios2_symbolic_memory_operand_p): Rename to...
(nios2_large_constant_memory_operand_p): ...this.  Adjust uses.
(nios2_large_constant_p): Check for large constant integers too.
(nios2_split_large_constant): Handle constant integers.
(nios2_split_symbolic_memory_operand): Rename to...
(nios2_split_large_constant_memory_operand): ...this.  Adjust uses.
(nios2_legitimize_constant_address): Handle constant integers.
(r0rel_constant_p): Handle small constant integers.
(nios2_print_operand_address): Handle r0-relative integer addresses.
* config/nios2/nios2-protos.h: Adjust for renamed functions.
* config/nios2/nios2.md: Adjust for renamed functions.

From-SVN: r255492

Add srodata section support to riscv port.

2017-12-07  Andrew Waterman  <andrew@sifive.com>
gcc/
* config/riscv/riscv.c (TARGET_ASM_SELECT_SECTION): New define.
(TARGET_HAVE_SRODATA_SECTION): New define.
(riscv_select_section): New function.

From-SVN: r255491

Daily bump.

From-SVN: r255490

re PR target/83252 (Wrong code with "-march=skylake-avx512 -O3")

PR target/83252
* gcc.target/i386/i386.exp (check_effective_target_bmi2): Moved to ...
* lib/target-supports.exp (check_effective_target_bmi2): ... here.  Guard with
i?86-*-* x86_64-*-*.
* g++.dg/opt/pr83252.C: New test.

From-SVN: r255487

re PR target/81906 (Calls to rint() wrongly optimized away starting in g++ 6)

PR target/81906
* config/i386/i386.c (ix86_expand_rint): Handle flag_rounding_math.

* gcc.target/i386/pr81906.c: New test.

From-SVN: r255486

re PR tree-optimization/83075 (Invalid strncpy optimization)

PR tree-optimization/83075
* gcc.dg/tree-ssa/strncpy-2.c: Use size_t instead of unsigned, add
separate function with noipa attribute to also verify behavior when
optimizers don't know the sizes and aliasing, verify resulting sizes
and array content.  Add -Wstringop-overflow to dg-options.
* gcc.dg/tree-ssa/strncat.c: Likewise.

From-SVN: r255485

Improve fstack_protector effective target

Effective target fstack_protector fails to return an error for
newlib-based target (such as arm-none-eabi targets) which does not
support stack protector. This is due to the test being too simplist for
stack protection code to be generated by GCC: it does not contain a
local buffer and does not read unknown input.

This commit adds a small local buffer with a copy of the filename to
trigger stack protector code to be generated. The filename is used
instead of the full path so as to ensure the size will fit in the local
buffer.

2017-12-07  Thomas Preud'homme  <thomas.preudhomme@arm.com>

    gcc/testsuite/
    * lib/target-supports.exp (check_effective_target_fstack_protector):
    Copy filename in local buffer to trigger stack protection.

From-SVN: r255484

re PR libstdc++/83120 (ext/special_functions/hyperg failure on AIX)

        PR libstdc++/83120
        * testsuite/ext/special_functions/hyperg/check_value.cc: Add
        -ffp-contract=off. Add -ffloat-store for m68* and ia32.
        * testsuite/tr1/5_numerical_facilities/special_functions/17_hyperg/check_value.cc: Same.

From-SVN: r255483

Allow _Alignas in compound literals (C11 DR#444).

C11 DR#444 dealt with how C11 intended to allow alignment specifiers
on struct and union members, but failed to include that in the syntax.
The final resolution of that DR also allows alignment specifiers in
type names in compound literals (in order to apply an increased
alignment to the unnamed object created by the compound literal), but
not other cases of type names.

This patch implements allowing alignment specifiers in compound
literals and adds tests for the resolution of the DR (including that
they are allowed on struct and union members, which GCC already
implemented).  Because the parser has to parse the parenthesized type
name of a compound literal before it can tell that it's a compound
literal (rather than, depending on the context, a cast expression or
sizeof (type-name) or _Alignof (type-name)), this means _Alignas
specifiers are allowed syntactically in those contexts and then an
error is given once it's known to be an invalid use (whereas _Alignas
specifiers are disallowed syntactically in other contexts where type
names can occur and a compound literal is not possible).

Bootstrapped with no regressions on x86_64-pc-linux-gnu.

gcc/c:
* c-decl.c (build_compound_literal): Add parameter alignas_align
and set alignment of decl if nonzero.
* c-parser.c (c_keyword_starts_typename): Allow RID_ALIGNAS.
(c_parser_declspecs): Allow RID_ALIGNAS to follow a type, like a
qualifier.
(c_parser_struct_declaration): Update syntax comment.
(c_parser_type_name): Add alignas_ok argument and pass it to
c_parser_declspecs.
(c_parser_cast_expression): Pass true to c_parser_type_name and
give error if a cast used an _Alignas specifier.
(c_parser_sizeof_expression): Pass true to c_parser_type_name and
give error if sizeof (type-name) used an _Alignas specifier.
(c_parser_alignof_expression): Pass true to c_parser_type_name and
give error if _Alignof (type-name) used an _Alignas specifier.
(c_parser_postfix_expression_after_paren_type): Check specified
alignment for a compound literal and pass it to
build_compound_literal.
* c-parser.h (c_parser_type_name): Update prototype.
* c-tree.h (build_compound_literal): Update prototype.

gcc/testsuite:
* gcc.dg/c11-align-7.c, gcc.dg/c11-align-8.c,
gcc.dg/c11-align-9.c, gcc.dg/gnu11-align-1.c: New tests.
* gcc.dg/c11-align-5.c (test): Update expected error for sizeof
case.

From-SVN: r255482

[AArch64] Fix ICEs in aarch64_print_operand

Three related regression fixes:

- We can't use asserts like:

    gcc_assert (GET_MODE_SIZE (mode) == 16);

  in aarch64_print_operand because it could trigger for invalid user input.

- The output_operand_lossage in aarch64_print_address_internal:

    output_operand_lossage ("invalid operand for '%%%c'", op);

  wasn't right because "op" is an rtx_code enum rather than the
  prefix character.

- aarch64_print_operand_address shouldn't call output_operand_lossage
  (because it doesn't have a prefix code) but instead fall back to
  output_addr_const.

2017-12-05  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* config/aarch64/aarch64.c (aarch64_print_address_internal): Return
a bool success value.  Don't call output_operand_lossage here.
(aarch64_print_ldpstp_address): Return a bool success value.
(aarch64_print_operand_address): Call output_addr_const if
aarch64_print_address_internal fails.
(aarch64_print_operand): Don't assert that the mode is 16 bytes for
'y'; call output_operand_lossage instead.  Call output_operand_lossage
if aarch64_print_ldpstp_address fails.

gcc/testsuite/
* gcc.target/aarch64/asm-2.c: New test.
* gcc.target/aarch64/asm-3.c: Likewise.

From-SVN: r255481

Make more use of VECTOR_CST_ENCODED_ELT

This patch makes various bits of code operate directly on the new
VECTOR_CST encoding, instead of using VECTOR_CST_ELT on all elements
of the vector.

Previous patches handled operations that produce a new VECTOR_CST,
while this patch handles things like predicates.  It also makes
print_node dump the encoding instead of the full vector that
the encoding represents.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree-vector-builder.h
(tree_vector_builder::binary_encoded_nelts): Declare.
* tree-vector-builder.c
(tree_vector_builder::binary_encoded_nelts): New function.
* fold-const.c (negate_expr_p): Likewise.
(operand_equal_p, fold_checksum_tree): Likewise.
* tree-loop-distribution.c (const_with_all_bytes_same): Likewise.
* tree.c (integer_zerop, integer_onep, integer_all_onesp, real_zerop)
(real_onep, real_minus_onep, add_expr, initializer_zerop): Likewise.
(uniform_vector_p): Likewise.
* varasm.c (const_hash_1, compare_constant): Likewise.
* tree-ssa-ccp.c: Include tree-vector-builder.h.
(valid_lattice_transition): Operate directly on the VECTOR_CST
encoding.
* ipa-icf.c: Include tree-vector-builder.h.
(sem_variable::equals): Operate directly on the VECTOR_CST encoding.
* print-tree.c (print_node): Print encoding of VECTOR_CSTs.

From-SVN: r255480

Make build_vector static

The only remaining uses of build_vector are in the selftests in tree.c.
This patch makes it static and moves it to the selftest part of the file.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree.c (build_vector): Delete.
* tree.h (build_vector): Make static and move into the self-testing
block.

From-SVN: r255479

Make gimple_build_vector take a tree_vector_builder

This patch changes gimple_build_vector so that it takes a
tree_vector_builder instead of a size and a vector of trees.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* vector-builder.h (vector_builder::derived): New const overload.
(vector_builder::elt): New function.
* tree-vector-builder.h (tree_vector_builder::type): New function.
(tree_vector_builder::apply_step): Declare.
* tree-vector-builder.c (tree_vector_builder::apply_step): New
function.
* gimple-fold.h (tree_vector_builder): Declare.
(gimple_build_vector): Take a tree_vector_builder instead of a
type and vector of elements.
* gimple-fold.c (gimple_build_vector): Likewise.
* tree-vect-loop.c (get_initial_def_for_reduction): Update call
accordingly.
(get_initial_defs_for_reduction): Likewise.
(vectorizable_induction): Likewise.

From-SVN: r255478

Use tree_vector_builder::new_binary_operation for folding

This patch makes fold-const.c operate directly on the VECTOR_CST
encoding when folding an operation that has two VECTOR_CST inputs.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* tree-vector-builder.h
(tree_vector_builder::new_binary_operation): Declare.
* tree-vector-builder.c
(tree_vector_builder::new_binary_operation): New function.
* fold-const.c (fold_relational_const): Use it.
(const_binop): Likewise.  Check that both input vectors have
the same number of elements, thus excluding things like WIDEN_SUM.
Check whether it is possible to operate directly on the encodings
of stepped inputs.

From-SVN: r255477

Use tree_vector_builder::new_unary_operation for folding

This patch makes fold-const.c operate directly on the VECTOR_CST
encoding when folding an operation that has a single VECTOR_CST input.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* fold-const.c (fold_negate_expr_1): Use tree_vector_builder and
new_unary_operation, operating only on the encoded elements.
(const_unop): Likewise.
(exact_inverse): Likewise.
(distributes_over_addition_p): New function.
(const_binop): Use tree_vector_builder and new_unary_operation
for combinations of VECTOR_CST and INTEGER_CST.  Operate only
on the encoded elements unless the encoding is strided and the
operation does not distribute over addition.
(fold_convert_const):  Use tree_vector_builder and
new_unary_operation.  Operate only on the encoded elements
for truncating integer conversions, or for non-stepped encodings.

From-SVN: r255476

Use tree_vector_builder instead of build_vector

This patch switches most build_vector calls over to tree_vector_builder,
using explicit encodings where appropriate.  Later patches handle
the remaining uses of build_vector.

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* config/sparc/sparc.c: Include tree-vector-builder.h.
(sparc_fold_builtin): Use tree_vector_builder instead of build_vector.
* expmed.c: Include tree-vector-builder.h.
(make_tree): Use tree_vector_builder instead of build_vector.
* fold-const.c: Include tree-vector-builder.h.
(const_binop): Use tree_vector_builder instead of build_vector.
(const_unop): Likewise.
(native_interpret_vector): Likewise.
(fold_vec_perm): Likewise.
(fold_ternary_loc): Likewise.
* gimple-fold.c: Include tree-vector-builder.h.
(gimple_fold_stmt_to_constant_1): Use tree_vector_builder instead
of build_vector.
* tree-ssa-forwprop.c: Include tree-vector-builder.h.
(simplify_vector_constructor): Use tree_vector_builder instead
of build_vector.
* tree-vect-generic.c: Include tree-vector-builder.h.
(add_rshift): Use tree_vector_builder instead of build_vector.
(expand_vector_divmod): Likewise.
(optimize_vector_constructor): Likewise.
* tree-vect-loop.c: Include tree-vector-builder.h.
(vect_create_epilog_for_reduction): Use tree_vector_builder instead
of build_vector.  Explicitly use a stepped encoding for
{ 1, 2, 3, ... }.
* tree-vect-slp.c: Include tree-vector-builder.h.
(vect_get_constant_vectors): Use tree_vector_builder instead
of build_vector.
(vect_transform_slp_perm_load): Likewise.
(vect_schedule_slp_instance): Likewise.
* tree-vect-stmts.c: Include tree-vector-builder.h.
(vectorizable_bswap): Use tree_vector_builder instead of build_vector.
(vect_gen_perm_mask_any): Likewise.
(vectorizable_call): Likewise.  Explicitly use a stepped encoding.
* tree.c: (build_vector_from_ctor): Use tree_vector_builder instead
of build_vector.
(build_vector_from_val): Likewise.  Explicitly use a duplicate
encoding.

From-SVN: r255475

New VECTOR_CST layout

This patch uses a simple compression scheme to represent the contents
of a VECTOR_CST using its leading elements.  There are three formats:

1) a repeating sequence of N values.  This is encoded using the first
   N elements.

2) a "foreground" sequence of N values inserted at the beginning of
   a "background" repeating sequence of N values, such as:
   { 1, 2, 0, 0, 0, 0, ... }.  This is encoded using the first 2*N
   elements.

2) a "foreground" sequence of N values inserted at the beginning of
   a "background" repeating sequence of N interleaved linear series,
   such as: { 0, 0, 8, 10, 9, 11, 10, 12, ... }.  This is encoded
   using the first 3*N elements.  In practice the foreground values
   are often part of the same series as the background values,
   such as: { 1, 11, 2, 12, 3, 13, ... }.

This reduces the amount of work involved in processing simple
vector constants and means that the encoding extends naturally
to variable-length vectors.

2017-12-07  Richard Sandiford  <richard.sandiford@arm.com>

gcc/
* doc/generic.texi (VECTOR_CST): Describe new representation of
vector constants.
* vector-builder.h: New file.
* tree-vector-builder.h: Likewise.
* tree-vector-builder.c: Likewise.
* Makefile.in (OBJS): Add tree-vector-builder.o.
* tree.def (VECTOR_CST): Update comment to refer to generic.texi.
* tree-core.h (tree_base): Add a vector_cst field to the u union.
(tree_vector): Change the number of elements to
vector_cst_encoded_nelts.
* tree.h (VECTOR_CST_NELTS): Redefine using TYPE_VECTOR_SUBPARTS.
(VECTOR_CST_ELTS): Delete.
(VECTOR_CST_ELT): Redefine using vector_cst_elt.
(VECTOR_CST_LOG2_NPATTERNS, VECTOR_CST_NPATTERNS): New macros.
(VECTOR_CST_NELTS_PER_PATTERN, VECTOR_CST_DUPLICATE_P): Likewise.
(VECTOR_CST_STEPPED_P, VECTOR_CST_ENCODED_ELTS): Likewise.
(VECTOR_CST_ENCODED_ELT): Likewise.
(vector_cst_encoded_nelts): New function.
(make_vector): Take the values of VECTOR_CST_LOG2_NPATTERNS and
VECTOR_CST_NELTS_PER_PATTERN as arguments.
(vector_cst_int_elt, vector_cst_elt): Declare.
* tree.c: Include tree-vector-builder.h.
(tree_code_size): Abort if passed VECTOR_CST.
(tree_size): Update for new VECTOR_CST layout.
(make_vector): Take the values of VECTOR_CST_LOG2_NPATTERNS and
VECTOR_CST_NELTS_PER_PATTERN as arguments.
(build_vector): Use tree_vector_builder.
(vector_cst_int_elt, vector_cst_elt): New functions.
(drop_tree_overflow): For VECTOR_CST, drop the TREE_OVERFLOW from the
encoded elements and then create the vector in the canonical form.
(check_vector_cst, check_vector_cst_duplicate, check_vector_cst_fill)
(check_vector_cst_stepped, test_vector_cst_patterns): New functions.
(tree_c_tests): Call test_vector_cst_patterns.
* lto-streamer-out.c (DFS::DFS_write_tree_body): Handle the new
VECTOR_CST fields.
(hash_tree): Likewise.
* tree-streamer-out.c (write_ts_vector_tree_pointers): Likewise.
(streamer_write_tree_header): Likewise.
* tree-streamer-in.c (lto_input_ts_vector_tree_pointers): Likewise.
(streamer_alloc_tree): Likewise.  Update call to make_vector.
* fold-const.c (fold_ternary_loc): Avoid using VECTOR_CST_ELTS.

gcc/lto/
* lto.c (compare_tree_sccs_1): Compare the new VECTOR_CST flags.

From-SVN: r255474

Add underscores to local vars in selftest macros

Allows functions doing selftests to take parameters with names like
"actual".

2017-12-07  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
* selftest.h (ASSERT_TRUE_AT, ASSERT_FALSE_AT, ASSERT_EQ_AT)
(ASSERT_NE, ASSERT_PRED1): Add underscores to local variable names
* selftest-rtl.h (ASSERT_RTX_EQ, ASSERT_RTX_PTR_EQ): Likewise.

From-SVN: r255473

re PR tree-optimization/81303 (410.bwaves regression caused by r249919)

PR tree-optimization/81303
* Makefile.in (gimple-loop-interchange.o): New object file.
* common.opt (floop-interchange): Reuse the option from graphite.
* doc/invoke.texi (-floop-interchange): Ditto.  New document for
-floop-interchange and mention it for -O3.
* opts.c (default_options_table): Enable -floop-interchange at -O3.
* gimple-loop-interchange.cc: New file.
* params.def (PARAM_LOOP_INTERCHANGE_MAX_NUM_STMTS): New parameter.
(PARAM_LOOP_INTERCHANGE_STRIDE_RATIO): New parameter.
* passes.def (pass_linterchange): New pass.
* timevar.def (TV_LINTERCHANGE): New time var.
* tree-pass.h (make_pass_linterchange): New declaration.
* tree-ssa-loop-ivcanon.c (create_canonical_iv): Change to external
interchange.  Record IV before/after increment in new parameters.
* tree-ssa-loop-ivopts.h (create_canonical_iv): New declaration.
* tree-vect-loop.c (vect_is_simple_reduction): Factor out reduction
path check into...
(check_reduction_path): ...New function here.
* tree-vectorizer.h (check_reduction_path): New declaration.

gcc/testsuite
* gcc.dg/tree-ssa/loop-interchange-1.c: New test.
* gcc.dg/tree-ssa/loop-interchange-1b.c: New test.
* gcc.dg/tree-ssa/loop-interchange-2.c: New test.
* gcc.dg/tree-ssa/loop-interchange-3.c: New test.
* gcc.dg/tree-ssa/loop-interchange-4.c: New test.
* gcc.dg/tree-ssa/loop-interchange-5.c: New test.
* gcc.dg/tree-ssa/loop-interchange-6.c: New test.
* gcc.dg/tree-ssa/loop-interchange-7.c: New test.
* gcc.dg/tree-ssa/loop-interchange-8.c: New test.
* gcc.dg/tree-ssa/loop-interchange-9.c: New test.
* gcc.dg/tree-ssa/loop-interchange-10.c: New test.
* gcc.dg/tree-ssa/loop-interchange-11.c: New test.
* gcc.dg/tree-ssa/loop-interchange-12.c: New test.
* gcc.dg/tree-ssa/loop-interchange-13.c: New test.

Co-Authored-By: Richard Biener <rguenther@suse.de>
From-SVN: r255472

re PR target/83252 (Wrong code with "-march=skylake-avx512 -O3")

2017-12-07  Vladimir Makarov  <vmakarov@redhat.com>

PR target/83252
PR rtl-optimization/80818
* lra.c (add_regs_to_insn_regno_info): Make a hard reg in CLOBBER
always early clobbered.
* lra-lives.c (process_bb_lives): Check input hard regs for early
clobbered non-operand hard reg.

From-SVN: r255471

re PR middle-end/83164 (internal compiler error: verify_gimple failed)

PR middle-end/83164
* tree-cfg.c (verify_gimple_assign_binary): Don't require
types_compatible_p, just that TYPE_MODE is the same.

* gcc.c-torture/compile/pr83164.c: New test.

From-SVN: r255470

PR c/81544 - attribute noreturn and warn_unused_result on the same function accepted

PR c/81544 - attribute noreturn and warn_unused_result on the same function accepted
PR c/81566 - invalid attribute aligned accepted on functions

gcc/ada/ChangeLog:

PR c/81544
* gcc-interface/utils.c (gnat_internal_attribute_table): Initialize
new member of struct attribute_spec.

gcc/c/ChangeLog:

PR c/81544
* c-decl.c (c_decl_attributes): Look up existing declaration and
pass it to decl_attributes.

gcc/c-family/ChangeLog:

PR c/81544
PR c/81566
* c-attribs.c (attr_aligned_exclusions): New array.
(attr_alloc_exclusions, attr_cold_hot_exclusions): Same.
(attr_common_exclusions, attr_const_pure_exclusions): Same.
(attr_gnu_inline_exclusions, attr_inline_exclusions): Same.
(attr_noreturn_exclusions, attr_returns_twice_exclusions): Same.
(attr_warn_unused_result_exclusions): Same.
(handle_hot_attribute, handle_cold_attribute): Simplify.
(handle_const_attribute): Warn on function returning void.
(handle_pure_attribute): Same.
(handle_aligned_attribute): Diagnose conflicting attribute
specifications.
* c-warn.c (diagnose_mismatched_attributes): Simplify.

gcc/cp/ChangeLog:

PR c/81544
* cp-tree.h (decls_match): Add default argument.
* decl.c (decls_match): Avoid calling into the target back end
and triggering an error.
* decl2.c (cplus_decl_attributes): Look up existing declaration and
pass it to decl_attributes.
* tree.c (cxx_attribute_table): Initialize new member of struct
attribute_spec.

gcc/fortran/ChangeLog:

PR c/81544
* f95-lang.c (gfc_attribute_table): Initialize new member of struct
attribute_spec.

gcc/lto/ChangeLog:

PR c/81544
* lto-lang.c (lto_attribute_table): Initialize new member of struct
attribute_spec.

gcc/ChangeLog:

PR c/81544
* attribs.c (empty_attribute_table): Initialize new member of
struct attribute_spec.
(decl_attributes): Add argument.  Handle mutually exclusive
combinations of attributes.
(selftests::test_attribute_exclusions): New function.
(selftests::attribute_c_tests): Ditto.
* attribs.h (decl_attributes): Add default argument.
* selftest.h (attribute_c_tests): Declare.
* selftest-run-tests.c (selftest::run_tests): Call attribute_c_tests.
* tree-core.h (attribute_spec::exclusions, exclude): New type and
member.
* doc/extend.texi (Common Function Attributes): Update const and pure.

gcc/testsuite/ChangeLog:

PR c/81544
* c-c++-common/Wattributes-2.c: New test.
* c-c++-common/Wattributes.c: New test.
* c-c++-common/attributes-3.c: Adjust.
* gcc.dg/Wattributes-6.c: New test.
* gcc.dg/Wattributes-7.c: New test.
* gcc.dg/attr-noinline.c
* gcc.dg/pr44964.c: Same.
* gcc.dg/torture/pr42363.c: Same.
* gcc.dg/tree-ssa/ssa-ccp-2.c: Same.

From-SVN: r255469

re PR target/82641 (Unable to enable crc32 for a certain function with target attribute on ARM (aarch32))

2017-12-07  Tamar Christina  <tamar.christina@arm.com>

PR target/82641
* config/arm/arm.c (INCLUDE_STRING): Define.
(arm_last_printed_arch_string, arm_last_printed_fpu_string): New.
(arm_declare_function_name): Conservatively emit .arch, .arch_extensions
and .fpu.

gcc/testsuite/
2017-12-07  Tamar Christina  <tamar.christina@arm.com>

PR target/82641
* gcc.target/arm/pragma_arch_attribute_2.c: New.
* gcc.target/arm/pragma_arch_attribute_2.c: New.
* gcc.target/arm/pragma_arch_attribute_3.c: New.
* gcc.target/arm/pragma_fpu_attribute.c: New.
* gcc.target/arm/pragma_fpu_attribute_2.c: New.

From-SVN: r255468

Add unroll and jam pass

* gimple-loop-jam.c: New file.
* Makefile.in (OBJS): Add gimple-loop-jam.o.
* common.opt (funroll-and-jam): New option.
* opts.c (default_options_table): Add unroll-and-jam at -O3.
* params.def (PARAM_UNROLL_JAM_MIN_PERCENT): New param.
(PARAM_UNROLL_JAM_MAX_UNROLL): Ditto.
* passes.def: Add pass_loop_jam.
* timevar.def (TV_LOOP_JAM): Add.
* tree-pass.h (make_pass_loop_jam): Declare.
* cfgloop.c (flow_loop_tree_node_add): Add AFTER argument.
* cfgloop.h (flow_loop_tree_node_add): Adjust declaration.
* cfgloopmanip.c (duplicate_loop): Add AFTER argument, adjust call
to flow_loop_tree_node_add.
(duplicate_subloops, copy_loops_to): Append to sibling list.
* cfgloopmanip.h: (duplicate_loop): Adjust declaration.
* doc/invoke.texi (-funroll-and-jam): Document new option.
(unroll-jam-min-percent, unroll-jam-max-unroll): Document new params.

testsuite/
* gcc.dg/unroll-and-jam.c: New test.

From-SVN: r255467

re PR tree-optimization/83296 (missing -Wstringop-overflow due to missing range info for MAX_EXPR)

2017-12-07  Richard Biener  <rguenther@suse.de>

PR tree-optimization/83296
PR tree-optimization/67769
* tree-ssa-phiopt.c (conditional_replacement): Do not reset
flow sensitive info in an unrelated BB.
(value_replacement): Use reset_flow_sensitive_info.
(minmax_replacement): Reset flow sensitive info on the def
we move.  Do not reset flow sensitive info in the whole BB
we move the stmt to.
(abs_replacement): Likewise.

* g++.dg/warn/Wstringop-overflow-1.C: New testcase.

From-SVN: r255466

rs6000: Initialise rs6000_cpu correctly (PR43871)

Finally, set rs6000_cpu based not on -mtune=, but only -mcpu= and the
defaults.

PR target/43871
* config/rs6000/rs6000.c (rs6000_option_override_internal): Set
rs6000_cpu to the given -mcpu=, or to the default processor.

From-SVN: r255464

rs6000: Don't use rs6000_cpu_attr directly

In some places in the code we currently use rs6000_cpu_attr, although
that is a cast from rs6000_tune.  This patch changes things to be more
regular and clearer.  Now rs6000_cpu_attr is not used any more, and
it is deleted.

* config/rs6000/rs6000.h (rs6000_cpu_attr): Delete.
* config/rs6000/rs6000.c (rs6000_variable_issue_1): Use rs6000_tune
instead of rs6000_cpu_attr.
(rs6000_adjust_cost): Ditto.
(is_microcoded_insn): Ditto.
(rs6000_adjust_priority): Ditto.
(rs6000_issue_rate): Ditto.
(rs6000_use_sched_lookahead): Ditto.
(rs6000_use_sched_lookahead_guard): Ditto.
(rs6000_sched_reorder): Ditto.
(force_new_group): Ditto.
* config/rs6000/rs6000.md (cpu attribute): Ditto.
(group_ending_nop): Ditto.

From-SVN: r255463

rs6000: Split rs6000_cpu

This splits rs6000_cpu into rs6000_cpu and rs6000_tune.  Both are still
initialised identically, so there is no behaviour change.

Now rs6000_cpu should be set by -mcpu=, and rs6000_tune by -mtune= (but
both still are set by -mtune=, fixed in a later patch in the series).

Also change rs6000_cpu to rs6000_tune in the appropriate cases (i.e.
most, but not those that enable using new insns).

* config/rs6000/rs6000.opt (rs6000_tune): New variable.
* config/rs6000/rs6000.c (rs6000_option_override_internal): Also set
rs6000_tune.  Use rs6000_tune instead of rs6000_cpu where appropriate.
(rs6000_loop_align): Use rs6000_tune instead of rs6000_cpu where
appropriate.
(rs6000_reassociation_width): Ditto.
(rs6000_emit_epilogue): Ditto.
(rs6000_adjust_cost): Ditto.
(is_microcoded_insn): Ditto.
(is_cracked_insn): Ditto.
(rs6000_adjust_priority): Ditto.
(rs6000_sched_reorder): Ditto.
(rs6000_sched_reorder2): Ditto.
(insn_must_be_first_in_group): Ditto.
(insn_must_be_last_in_group): Ditto.
(rs6000_register_move_cost): Ditto.
* config/rs6000/rs6000.h (rs6000_cpu_attr): Use rs6000_tune instead of
rs6000_cpu.

From-SVN: r255462

Enable VAES support [2/5]

gcc/
* config.gcc: Add vaesintrin.h.
* config/i386/i386-builtin-types.def (V64QI_FTYPE_V64QI_V64QI): New type.
* config/i386/i386-builtin.def (__builtin_ia32_vaesdec_v16qi,
__builtin_ia32_vaesdec_v32qi, __builtin_ia32_vaesdec_v64qi): New builtins.
* config/i386/i386.c (ix86_expand_args_builtin): Handle new type.
* config/i386/immintrin.h: Include vaesintrin.h.
* config/i386/sse.md (vaesdec_<mode>): New pattern.
* config/i386/vaesintrin.h (_mm256_aesdec_epi128, _mm512_aesdec_epi128,
_mm_aesdec_epi128): New intrinsics.

gcc/testsuite/
* gcc.target/i386/avx512-check.h: Handle bit_VAES.
* gcc.target/i386/avx512f-aesdec-2.c: New test.
* gcc.target/i386/avx512fvl-vaes-1.c: Ditto.
* gcc.target/i386/avx512vl-aesdec-2.c: Ditto.
* gcc.target/i386/i386.exp (check_effective_target_avx512vaes): New.

From-SVN: r255461

Daily bump.

From-SVN: r255460

re PR tree-optimization/69224 (-Warray-bounds false positive with -O3 and struct pointer parameter)

PR tree-optimization/69224
PR tree-optimization/80907
PR tree-optimization/82286
* gcc.dg/pr69224.c: New test.
* gcc.dg/pr80907.c: New test.
* gcc.dg/pr82286.c: New test.

From-SVN: r255457

re PR c++/80259 (ICE deleting friend function)

PR c++/80259
* decl2.c (grokfield): Diagnose = delete redefinition of a friend.

* g++.dg/cpp0x/pr80259.C: New test.

From-SVN: r255456

Correct argument to targetm.calls.promote_prototypes.

* call.c (convert_for_arg_passing): Pass NULL_TREE to
targetm.calls.promote_prototypes.
(type_passed_as): Likewise.

From-SVN: r255455

PR c++/82115 - ICE with variable initialized with its own address.

* pt.c (value_dependent_expression_p): Add lval parameter.  Don't
consider DECL_INITIAL if it's true.

From-SVN: r255454

C/C++: don't suggest implementation names as spelling fixes (PR c/83236)

gcc/c-family/ChangeLog:
PR c/83236
* c-common.c (selftest::c_family_tests): Call
selftest::c_spellcheck_cc_tests.
* c-common.h (selftest::c_spellcheck_cc_tests): New decl.
* c-spellcheck.cc: Include "selftest.h".
(name_reserved_for_implementation_p): New function.
(should_suggest_as_macro_p): New function.
(find_closest_macro_cpp_cb): Move the check for NT_MACRO to
should_suggest_as_macro_p and call it.
(selftest::test_name_reserved_for_implementation_p): New function.
(selftest::c_spellcheck_cc_tests): New function.
* c-spellcheck.h (name_reserved_for_implementation_p): New decl.

gcc/c/ChangeLog:
PR c/83236
* c-decl.c (lookup_name_fuzzy): Don't suggest names that are
reserved for use by the implementation.

gcc/cp/ChangeLog:
PR c/83236
* name-lookup.c (consider_binding_level): Don't suggest names that
are reserved for use by the implementation.

gcc/testsuite/ChangeLog:
PR c/83236
* c-c++-common/spellcheck-reserved.c: New test case.

From-SVN: r255453

Move macro-spellchecking code from "gcc" to new files in c-family

The code for spellchecking macros really belongs in c-family, rather
than in gcc/spellcheck-tree.c, so this patch moves it there.

gcc/ChangeLog:
* Makefile.in (C_COMMON_OBJS): Add c-family/c-spellcheck.o.
* spellcheck-tree.c (find_closest_macro_cpp_cb): Move to
c-family/c-spellcheck.cc.
(best_macro_match::best_macro_match): Likewise.
* spellcheck-tree.h
(struct edit_distance_traits<cpp_hashnode *>): Move to
c-family/c-spellcheck.h.
(class best_macro_match): Likewise.

gcc/c-family/ChangeLog:
* c-spellcheck.cc: New file, taken from macro-handling code in
spellcheck-tree.c.
* c-spellcheck.h: New file, taken from macro-handling code in
spellcheck-tree.h.

gcc/c/ChangeLog:
* c-decl.c: Include "c-family/c-spellcheck.h".

gcc/cp/ChangeLog:
* name-lookup.c: Include "c-family/c-spellcheck.h".

From-SVN: r255452

re PR tree-optimization/83293 (ICE: in gsi_insert_seq_nodes_after,  at gimple-iterator.c:278)

PR tree-optimization/83293
* gimple-ssa-strength-reduction.c (insert_initializers): Use
GSI_NEW_STMT instead of GSI_SAME_STMT in gsi_insert_after that
might insert into empty bb.

* g++.dg/torture/pr83293.C: New test.

From-SVN: r255451

PR testsuite/83303 - FAIL: g++.dg/opt/new1.C on arm-none-eabi

PR testsuite/83303 - FAIL: g++.dg/opt/new1.C on arm-none-eabi
(extra -Walloc-size-larger-than warning

* g++.dg/opt/new1.C: Prune warning from test output.

From-SVN: r255450

re PR sanitizer/81281 (UBSAN: false positive, dropped promotion to long type.)

PR sanitizer/81281
* match.pd ((T)(P + A) - (T)P -> (T) A): Split into separate
simplify for plus with :c added, and pointer_plus without that.
((T)P - (T)(P + A) -> -(T) A): Likewise.  If type is integral
with undefined overflow and the conversion is not widening,
perform negation in utype and only convert to type afterwards.
((T)(P + A) - (T)(P + B) -> (T)A - (T)B): Split into separate
simplify for plus with :c added, and pointer_plus without that.
If type is integral with undefined overflow and the conversion is
not widening, perform minus in utype and only convert to type
afterwards.  Move the last pointer_diff_expr simplify into the
two outermost ifs.

* gcc.c-torture/execute/pr81281.c: New test.
* gcc.dg/pr81281-1.c: New test.
* gcc.dg/pr81281-2.c: New test.
* g++.dg/ubsan/pr81281.C: New test.
* g++.dg/ubsan/pr81281-aux.cc: New test.

From-SVN: r255449

PR tree-optimization/82646 - bogus -Wstringop-overflow with -D_FORTIFY_SOURCE=2 on strncpy with range to a member array

gcc/ChangeLog:

PR tree-optimization/82646
* builtins.c (maybe_emit_chk_warning): Use size as the bound for
strncpy, not maxlen.

gcc/testsuite/ChangeLog:

PR tree-optimization/82646
* gcc.dg/builtin-stringop-chk-1.c: Adjust.
* gcc.dg/builtin-stringop-chk-9.c: New test.
* g++.dg/ext/strncpy-chk1.C: Adjust.

From-SVN: r255448

invoke.texi (-Wstringop-truncation): Mention attribute nonstring.

gcc/ChangeLog:

* doc/invoke.texi (-Wstringop-truncation): Mention attribute
nonstring.

From-SVN: r255447

PR tree-optimization/83075 - Invalid strncpy optimization

gcc/ChangeLog:

PR tree-optimization/83075
* tree-ssa-strlen.c (handle_builtin_stxncpy): Avoid assuming
strncat/strncpy don't change length of source string.

gcc/testsuite/ChangeLog:

PR tree-optimization/83075
* gcc.dg/tree-ssa/strncat.c: New test.
* gcc.dg/tree-ssa/strncpy-2.c: Same.

From-SVN: r255446

libgo: remove -fplan9-extensions from CFLAGS

    Remove -fplan9-extensions from the CFLAGS used for libgo (no
    longer needed since the runtime was converted from C to Go).

    Reviewed-on: https://go-review.googlesource.com/82177

From-SVN: r255445

pr41305.C: Refine test option.

gcc/testsuite
* g++.dg/graphite/pr41305.C: Refine test option.
* gcc.dg/graphite/pr42205-1.c: Ditto.
* gcc.dg/graphite/pr42205-2.c: Ditto.
* gcc.dg/graphite/pr42211.c: Ditto.
* gcc.dg/graphite/pr46185.c: Ditto.
* gcc.dg/graphite/pr46966.c: Ditto.
* gcc.dg/graphite/pr59817-1.c: Ditto.
* gcc.dg/graphite/pr59817-2.c: Ditto.
* gcc.dg/graphite/pr60740.c: Ditto.
* gcc.dg/graphite/pr60785.c: Ditto.
* gcc.dg/graphite/pr68715-2.c: Ditto.
* gcc.dg/graphite/pr68715.c: Ditto.
* gcc.dg/graphite/pr70045.c: Ditto.
* gfortran.dg/graphite/pr14741.f90: Ditto.
* gfortran.dg/graphite/pr40982.f90: Ditto.
* gfortran.dg/graphite/pr42285.f90: Ditto.
* gfortran.dg/graphite/pr42334-1.f: Ditto.
* gfortran.dg/graphite/pr42334.f90: Ditto.
* gfortran.dg/graphite/pr43349.f: Ditto.
* gfortran.dg/graphite/pr59817.f: Ditto.

From-SVN: r255444

[ARM] testsuite: force hardfp in addr-modes-float.c

gcc/testsuite/ChangeLog:

<date>  Charles Baylis  <charles.baylis@linaro.org>

* gcc.target/arm/addr-modes-float.c: Place dg-add-options after
        dg-require-effective-target.
        (ATTR): New define.
        (POST_STORE): Pass ATTR as 2nd argument.
        (POST_LOAD): Likewise.
        (POST_STORE_VEC): Likewise.
        * gcc.target/arm/addr-modes-int.c (ATTR): New define.
        (PRE_STORE): Pass ATTR as 2nd argument.
        (POST_STORE): Likewise.
        (PRE_LOAD): Likewise.
        (POST_LOAD): Likewise.
        * gcc.target/arm/addr-modes.h:  (PRE_STORE): New parameter.
        (POST_STORE): Likewise.
        (POST_STORE_VEC): Likewise.
        (PRE_LOAD): Likewise.
        (POST_LOAD): Likewise.
        (POST_LOAD_VEC): Likewise.

From-SVN: r255443

compiler: handle set-and-use-temp in nilcheck code

    Change the code in Unary_expression::do_get_backend that introduces
    explicit nil checks for dereference operations to special case
    set-and-use-temporary expressions. For this case it is better to
    generate an explicit reference of the temp in the final conditional
    (avoids introducing tree sharing).

    Reviewed-on: https://go-review.googlesource.com/81915

From-SVN: r255442

re PR ada/66205 (gnatbind generates invalid code when finalization is enabled in restricted runtime)

PR ada/66205
* bindgen.adb (Gen_AdaFinal): If the restriction No_Task_Termination is
set, generate a null body.

From-SVN: r255441

Revert
2017-11-29  Martin Aberg  <maberg@gaisler.com>

* config/sparc/sparc.md (divdf3_fix): Add NOP and adjust length
to prevent b2bst errata sequence.
(sqrtdf2_fix): Likewise.

From-SVN: r255439

re PR tree-optimization/81945 (ICE in operator[], at vec.h:749)

PR tree-optimization/81945
* cfgloop.h (FOR_EACH_LOOP_FN): Use FN instead of hardcoding fn.
* tree-cfg.c (move_sese_region_to_fn): If any of the loops moved
to dest_cfun has orig_loop_num set, either remap it to the new
loop number if the loop got moved too, or clear it.

* gcc.dg/graphite/pr81945.c: New test.

From-SVN: r255438

Daily bump.

From-SVN: r255437