gcc/ChangeLog

   1 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
   2             Alan Hayward  <alan.hayward@arm.com>
   3             David Sherwood  <david.sherwood@arm.com>
   4
   5         * coretypes.h (fixed_size_mode): Declare.
   6         (fixed_size_mode_pod): New typedef.
   7         * builtins.h (target_builtins::x_apply_args_mode)
   8         (target_builtins::x_apply_result_mode): Change type to
   9         fixed_size_mode_pod.
  10         * builtins.c (apply_args_size, apply_result_size, result_vector)
  11         (expand_builtin_apply_args_1, expand_builtin_apply)
  12         (expand_builtin_return): Update accordingly.
  13
  14 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
  15
  16         * cse.c (hash_rtx_cb): Hash only the encoded elements.
  17         * cselib.c (cselib_hash_rtx): Likewise.
  18         * expmed.c (make_tree): Build VECTOR_CSTs directly from the
  19         CONST_VECTOR encoding.
  20
  21 2017-01-03  Jakub Jelinek  <jakub@redhat.com>
  22             Jeff Law  <law@redhat.com>
  23
  24         PR target/83641
  25         * config/i386/i386.c (ix86_adjust_stack_and_probe_stack_clash): For
  26         noreturn probe, use gen_pop instead of ix86_emit_restore_reg_using_pop,
  27         only set RTX_FRAME_RELATED_P on both the push and pop if cfa_reg is sp
  28         and add REG_CFA_ADJUST_CFA notes in that case to both insns.
  29
  30         PR target/83641
  31         * config/i386/i386.c (ix86_adjust_stack_and_probe_stack_clash): Do not
  32         explicitly probe *sp in a noreturn function if there were any callee
  33         register saves or frame pointer is needed.
  34
  35 2018-01-03  Jakub Jelinek  <jakub@redhat.com>
  36
  37         PR debug/83621
  38         * cfgexpand.c (expand_debug_expr): Return NULL if mode is
  39         BLKmode for ternary, binary or unary expressions.
  40
  41         PR debug/83645
  42         * var-tracking.c (delete_vta_debug_insn): New inline function.
  43         (delete_vta_debug_insns): Add USE_CFG argument, if true, walk just
  44         insns from get_insns () to NULL instead of each bb separately.
  45         Use delete_vta_debug_insn.  No longer static.
  46         (vt_debug_insns_local, variable_tracking_main_1): Adjust
  47         delete_vta_debug_insns callers.
  48         * rtl.h (delete_vta_debug_insns): Declare.
  49         * final.c (rest_of_handle_final): Call delete_vta_debug_insns
  50         instead of variable_tracking_main.
  51
  52 2018-01-03  Martin Sebor  <msebor@redhat.com>
  53
  54         PR c/83559
  55         * doc/extend.texi (attribute const): Fix a typo.
  56         * ipa-pure-const.c ((warn_function_const, warn_function_pure): Avoid
  57         issuing -Wsuggest-attribute for void functions.
  58
  59 2018-01-03  Martin Sebor  <msebor@redhat.com>
  60
  61         * gimple-ssa-warn-restrict.c (builtin_memref::builtin_memref): Use
  62         offset_int::from instead of wide_int::to_shwi.
  63         (maybe_diag_overlap): Remove assertion.
  64         Use HOST_WIDE_INT_PRINT_DEC instead of %lli.
  65         * gimple-ssa-sprintf.c (format_directive): Same.
  66         (parse_directive): Same.
  67         (sprintf_dom_walker::compute_format_length): Same.
  68         (try_substitute_return_value): Same.
  69
  70 2017-01-03  Jeff Law  <law@redhat.com>
  71
  72         PR middle-end/83654
  73         * explow.c (anti_adjust_stack_and_probe_stack_clash): Test a
  74         non-constant residual for zero at runtime and avoid probing in
  75         that case.  Reorganize code for trailing problem to mirror handling
  76         of the residual.
  77
  78 2018-01-03  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
  79
  80         PR tree-optimization/83501
  81         * tree-ssa-strlen.c (get_string_cst): New.
  82         (handle_char_store): Call get_string_cst.
  83
  84 2018-01-03  Martin Liska  <mliska@suse.cz>
  85
  86         PR tree-optimization/83593
  87         * tree-ssa-strlen.c: Include tree-cfg.h.
  88         (strlen_check_and_optimize_stmt): Add new argument cleanup_eh.
  89         (strlen_dom_walker): Add new member variable m_cleanup_cfg.
  90         (strlen_dom_walker::strlen_dom_walker): Initialize m_cleanup_cfg
  91         to false.
  92         (strlen_dom_walker::before_dom_children): Call
  93         gimple_purge_dead_eh_edges. Dump tranformation with details
  94         dump flags.
  95         (strlen_dom_walker::before_dom_children): Update call by adding
  96         new argument cleanup_eh.
  97         (pass_strlen::execute): Return TODO_cleanup_cfg if needed.
  98
  99 2018-01-03  Martin Liska  <mliska@suse.cz>
 100
 101         PR ipa/83549
 102         * cif-code.def (VARIADIC_THUNK): New enum value.
 103         * ipa-fnsummary.c (compute_fn_summary): Do not inline variadic
 104         thunks.
 105
 106 2018-01-03  Jan Beulich  <jbeulich@suse.com>
 107
 108         * sse.md (mov<mode>_internal): Tighten condition for when to use
 109         vmovdqu<ssescalarsize> for TI and OI modes.
 110
 111 2018-01-03  Jakub Jelinek  <jakub@redhat.com>
 112
 113         Update copyright years.
 114
 115 2018-01-03  Martin Liska  <mliska@suse.cz>
 116
 117         PR ipa/83594
 118         * ipa-visibility.c (function_and_variable_visibility): Skip
 119         functions with noipa attribure.
 120
 121 2018-01-03  Jakub Jelinek  <jakub@redhat.com>
 122
 123         * gcc.c (process_command): Update copyright notice dates.
 124         * gcov-dump.c (print_version): Ditto.
 125         * gcov.c (print_version): Ditto.
 126         * gcov-tool.c (print_version): Ditto.
 127         * gengtype.c (create_file): Ditto.
 128         * doc/cpp.texi: Bump @copying's copyright year.
 129         * doc/cppinternals.texi: Ditto.
 130         * doc/gcc.texi: Ditto.
 131         * doc/gccint.texi: Ditto.
 132         * doc/gcov.texi: Ditto.
 133         * doc/install.texi: Ditto.
 134         * doc/invoke.texi: Ditto.
 135
 136 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 137
 138         * vector-builder.h (vector_builder::m_full_nelts): Change from
 139         unsigned int to poly_uint64.
 140         (vector_builder::full_nelts): Update prototype accordingly.
 141         (vector_builder::new_vector): Likewise.
 142         (vector_builder::encoded_full_vector_p): Handle polynomial full_nelts.
 143         (vector_builder::operator ==): Likewise.
 144         (vector_builder::finalize): Likewise.
 145         * int-vector-builder.h (int_vector_builder::int_vector_builder):
 146         Take the number of elements as a poly_uint64 rather than an
 147         unsigned int.
 148         * vec-perm-indices.h (vec_perm_indices::m_nelts_per_input): Change
 149         from unsigned int to poly_uint64.
 150         (vec_perm_indices::vec_perm_indices): Update prototype accordingly.
 151         (vec_perm_indices::new_vector): Likewise.
 152         (vec_perm_indices::length): Likewise.
 153         (vec_perm_indices::nelts_per_input): Likewise.
 154         (vec_perm_indices::input_nelts): Likewise.
 155         * vec-perm-indices.c (vec_perm_indices::new_vector): Take the
 156         number of elements per input as a poly_uint64 rather than an
 157         unsigned int.  Use the original encoding for variable-length
 158         vectors, rather than clamping each individual element.
 159         For the second and subsequent elements in each pattern,
 160         clamp the step and base before clamping their sum.
 161         (vec_perm_indices::series_p): Handle polynomial element counts.
 162         (vec_perm_indices::all_in_range_p): Likewise.
 163         (vec_perm_indices_to_tree): Likewise.
 164         (vec_perm_indices_to_rtx): Likewise.
 165         * tree-vect-stmts.c (vect_gen_perm_mask_any): Likewise.
 166         * tree-vector-builder.c (tree_vector_builder::new_unary_operation)
 167         (tree_vector_builder::new_binary_operation): Handle polynomial
 168         element counts.  Return false if we need to know the number
 169         of elements at compile time.
 170         * fold-const.c (fold_vec_perm): Punt if the number of elements
 171         isn't known at compile time.
 172
 173 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 174
 175         * vec-perm-indices.h (vec_perm_builder): Change element type
 176         from HOST_WIDE_INT to poly_int64.
 177         (vec_perm_indices::element_type): Update accordingly.
 178         (vec_perm_indices::clamp): Handle polynomial element_types.
 179         * vec-perm-indices.c (vec_perm_indices::series_p): Likewise.
 180         (vec_perm_indices::all_in_range_p): Likewise.
 181         (tree_to_vec_perm_builder): Check for poly_int64 trees rather
 182         than shwi trees.
 183         * vector-builder.h (vector_builder::stepped_sequence_p): Handle
 184         polynomial vec_perm_indices element types.
 185         * int-vector-builder.h (int_vector_builder::equal_p): Likewise.
 186         * fold-const.c (fold_vec_perm): Likewise.
 187         * optabs.c (shift_amt_for_vec_perm_mask): Likewise.
 188         * tree-vect-generic.c (lower_vec_perm): Likewise.
 189         * tree-vect-slp.c (vect_transform_slp_perm_load): Likewise.
 190         * config/aarch64/aarch64.c (aarch64_evpc_tbl): Cast d->perm
 191         element type to HOST_WIDE_INT.
 192
 193 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 194             Alan Hayward  <alan.hayward@arm.com>
 195             David Sherwood  <david.sherwood@arm.com>
 196
 197         * alias.c (addr_side_effect_eval): Take the size as a poly_int64
 198         rather than an int.  Use plus_constant.
 199         (memrefs_conflict_p): Take the sizes as poly_int64s rather than ints.
 200         Take the offset "c" as a poly_int64 rather than a HOST_WIDE_INT.
 201
 202 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 203             Alan Hayward  <alan.hayward@arm.com>
 204             David Sherwood  <david.sherwood@arm.com>
 205
 206         * calls.c (emit_call_1, expand_call): Change struct_value_size from
 207         a HOST_WIDE_INT to a poly_int64.
 208
 209 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 210             Alan Hayward  <alan.hayward@arm.com>
 211             David Sherwood  <david.sherwood@arm.com>
 212
 213         * calls.c (load_register_parameters): Cope with polynomial
 214         mode sizes.  Require a constant size for BLKmode parameters
 215         that aren't described by a PARALLEL.  If BLOCK_REG_PADDING
 216         forces a parameter to be padded at the lsb end in order to
 217         fill a complete number of words, require the parameter size
 218         to be ordered wrt UNITS_PER_WORD.
 219
 220 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 221             Alan Hayward  <alan.hayward@arm.com>
 222             David Sherwood  <david.sherwood@arm.com>
 223
 224         * reload1.c (spill_stack_slot_width): Change element type
 225         from unsigned int to poly_uint64_pod.
 226         (alter_reg): Treat mode sizes as polynomial.
 227
 228 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 229             Alan Hayward  <alan.hayward@arm.com>
 230             David Sherwood  <david.sherwood@arm.com>
 231
 232         * reload.c (complex_word_subreg_p): New function.
 233         (reload_inner_reg_of_subreg, push_reload): Use it.
 234
 235 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 236             Alan Hayward  <alan.hayward@arm.com>
 237             David Sherwood  <david.sherwood@arm.com>
 238
 239         * lra-constraints.c (process_alt_operands): Reject matched
 240         operands whose sizes aren't ordered.
 241         (match_reload): Refer to this check here.
 242
 243 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 244             Alan Hayward  <alan.hayward@arm.com>
 245             David Sherwood  <david.sherwood@arm.com>
 246
 247         * builtins.c (expand_ifn_atomic_compare_exchange_into_call): Assert
 248         that the mode size is in the set {1, 2, 4, 8, 16}.
 249
 250 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 251             Alan Hayward  <alan.hayward@arm.com>
 252             David Sherwood  <david.sherwood@arm.com>
 253
 254         * var-tracking.c (adjust_mems): Treat mode sizes as polynomial.
 255         Use plus_constant instead of gen_rtx_PLUS.
 256
 257 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 258             Alan Hayward  <alan.hayward@arm.com>
 259             David Sherwood  <david.sherwood@arm.com>
 260
 261         * config/cr16/cr16-protos.h (cr16_push_rounding): Declare.
 262         * config/cr16/cr16.h (PUSH_ROUNDING): Move implementation to...
 263         * config/cr16/cr16.c (cr16_push_rounding): ...this new function.
 264         * config/h8300/h8300-protos.h (h8300_push_rounding): Declare.
 265         * config/h8300/h8300.h (PUSH_ROUNDING): Move implementation to...
 266         * config/h8300/h8300.c (h8300_push_rounding): ...this new function.
 267         * config/i386/i386-protos.h (ix86_push_rounding): Declare.
 268         * config/i386/i386.h (PUSH_ROUNDING): Move implementation to...
 269         * config/i386/i386.c (ix86_push_rounding): ...this new function.
 270         * config/m32c/m32c-protos.h (m32c_push_rounding): Take and return
 271         a poly_int64.
 272         * config/m32c/m32c.c (m32c_push_rounding): Likewise.
 273         * config/m68k/m68k-protos.h (m68k_push_rounding): Declare.
 274         * config/m68k/m68k.h (PUSH_ROUNDING): Move implementation to...
 275         * config/m68k/m68k.c (m68k_push_rounding): ...this new function.
 276         * config/pdp11/pdp11-protos.h (pdp11_push_rounding): Declare.
 277         * config/pdp11/pdp11.h (PUSH_ROUNDING): Move implementation to...
 278         * config/pdp11/pdp11.c (pdp11_push_rounding): ...this new function.
 279         * config/stormy16/stormy16-protos.h (xstormy16_push_rounding): Declare.
 280         * config/stormy16/stormy16.h (PUSH_ROUNDING): Move implementation to...
 281         * config/stormy16/stormy16.c (xstormy16_push_rounding): ...this new
 282         function.
 283         * expr.c (emit_move_resolve_push): Treat the input and result
 284         of PUSH_ROUNDING as a poly_int64.
 285         (emit_move_complex_push, emit_single_push_insn_1): Likewise.
 286         (emit_push_insn): Likewise.
 287         * lra-eliminations.c (mark_not_eliminable): Likewise.
 288         * recog.c (push_operand): Likewise.
 289         * reload1.c (elimination_effects): Likewise.
 290         * rtlanal.c (nonzero_bits1): Likewise.
 291         * calls.c (store_one_arg): Likewise.  Require the padding to be
 292         known at compile time.
 293
 294 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 295             Alan Hayward  <alan.hayward@arm.com>
 296             David Sherwood  <david.sherwood@arm.com>
 297
 298         * expr.c (emit_single_push_insn_1): Treat mode sizes as polynomial.
 299         Use plus_constant instead of gen_rtx_PLUS.
 300
 301 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 302             Alan Hayward  <alan.hayward@arm.com>
 303             David Sherwood  <david.sherwood@arm.com>
 304
 305         * auto-inc-dec.c (set_inc_state): Take the mode size as a poly_int64
 306         rather than an int.
 307
 308 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 309             Alan Hayward  <alan.hayward@arm.com>
 310             David Sherwood  <david.sherwood@arm.com>
 311
 312         * expr.c (expand_expr_real_1): Use tree_to_poly_uint64
 313         instead of int_size_in_bytes when handling VIEW_CONVERT_EXPRs
 314         via stack temporaries.  Treat the mode size as polynomial too.
 315
 316 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 317             Alan Hayward  <alan.hayward@arm.com>
 318             David Sherwood  <david.sherwood@arm.com>
 319
 320         * expr.c (expand_expr_real_2): When handling conversions involving
 321         unions, apply tree_to_poly_uint64 to the TYPE_SIZE rather than
 322         multiplying int_size_in_bytes by BITS_PER_UNIT.  Treat GET_MODE_BISIZE
 323         as a poly_uint64 too.
 324
 325 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 326             Alan Hayward  <alan.hayward@arm.com>
 327             David Sherwood  <david.sherwood@arm.com>
 328
 329         * rtlanal.c (subreg_get_info): Handle polynomial mode sizes.
 330
 331 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 332             Alan Hayward  <alan.hayward@arm.com>
 333             David Sherwood  <david.sherwood@arm.com>
 334
 335         * combine.c (can_change_dest_mode): Handle polynomial
 336         REGMODE_NATURAL_SIZE.
 337         * expmed.c (store_bit_field_1): Likewise.
 338         * expr.c (store_constructor): Likewise.
 339         * emit-rtl.c (validate_subreg): Operate on polynomial mode sizes
 340         and polynomial REGMODE_NATURAL_SIZE.
 341         (gen_lowpart_common): Likewise.
 342         * reginfo.c (record_subregs_of_mode): Likewise.
 343         * rtlanal.c (read_modify_subreg_p): Likewise.
 344
 345 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 346             Alan Hayward  <alan.hayward@arm.com>
 347             David Sherwood  <david.sherwood@arm.com>
 348
 349         * internal-fn.c (expand_vector_ubsan_overflow): Handle polynomial
 350         numbers of elements.
 351
 352 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 353             Alan Hayward  <alan.hayward@arm.com>
 354             David Sherwood  <david.sherwood@arm.com>
 355
 356         * match.pd: Cope with polynomial numbers of vector elements.
 357
 358 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 359             Alan Hayward  <alan.hayward@arm.com>
 360             David Sherwood  <david.sherwood@arm.com>
 361
 362         * fold-const.c (fold_indirect_ref_1): Handle polynomial offsets
 363         in a POINTER_PLUS_EXPR.
 364
 365 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 366             Alan Hayward  <alan.hayward@arm.com>
 367             David Sherwood  <david.sherwood@arm.com>
 368
 369         * omp-simd-clone.c (simd_clone_subparts): New function.
 370         (simd_clone_init_simd_arrays): Use it instead of TYPE_VECTOR_SUBPARTS.
 371         (ipa_simd_modify_function_body): Likewise.
 372
 373 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 374             Alan Hayward  <alan.hayward@arm.com>
 375             David Sherwood  <david.sherwood@arm.com>
 376
 377         * tree-vect-generic.c (nunits_for_known_piecewise_op): New function.
 378         (expand_vector_piecewise): Use it instead of TYPE_VECTOR_SUBPARTS.
 379         (expand_vector_addition, add_rshift, expand_vector_divmod): Likewise.
 380         (expand_vector_condition, vector_element): Likewise.
 381         (subparts_gt): New function.
 382         (get_compute_type): Use subparts_gt.
 383         (count_type_subparts): Delete.
 384         (expand_vector_operations_1): Use subparts_gt instead of
 385         count_type_subparts.
 386
 387 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 388             Alan Hayward  <alan.hayward@arm.com>
 389             David Sherwood  <david.sherwood@arm.com>
 390
 391         * tree-vect-data-refs.c (vect_no_alias_p): Replace with...
 392         (vect_compile_time_alias): ...this new function.  Do the calculation
 393         on poly_ints rather than trees.
 394         (vect_prune_runtime_alias_test_list): Update call accordingly.
 395
 396 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 397             Alan Hayward  <alan.hayward@arm.com>
 398             David Sherwood  <david.sherwood@arm.com>
 399
 400         * tree-vect-slp.c (vect_build_slp_tree_1): Handle polynomial
 401         numbers of units.
 402         (vect_schedule_slp_instance): Likewise.
 403
 404 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 405             Alan Hayward  <alan.hayward@arm.com>
 406             David Sherwood  <david.sherwood@arm.com>
 407
 408         * tree-vect-slp.c (vect_get_and_check_slp_defs): Reject
 409         constant and extern definitions for variable-length vectors.
 410         (vect_get_constant_vectors): Note that the number of units
 411         is known to be constant.
 412
 413 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 414             Alan Hayward  <alan.hayward@arm.com>
 415             David Sherwood  <david.sherwood@arm.com>
 416
 417         * tree-vect-stmts.c (vectorizable_conversion): Treat the number
 418         of units as polynomial.  Choose between WIDE and NARROW based
 419         on multiple_p.
 420
 421 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 422             Alan Hayward  <alan.hayward@arm.com>
 423             David Sherwood  <david.sherwood@arm.com>
 424
 425         * tree-vect-stmts.c (simd_clone_subparts): New function.
 426         (vectorizable_simd_clone_call): Use it instead of TYPE_VECTOR_SUBPARTS.
 427
 428 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 429             Alan Hayward  <alan.hayward@arm.com>
 430             David Sherwood  <david.sherwood@arm.com>
 431
 432         * tree-vect-stmts.c (vectorizable_call): Treat the number of
 433         vectors as polynomial.  Use build_index_vector for
 434         IFN_GOMP_SIMD_LANE.
 435
 436 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 437             Alan Hayward  <alan.hayward@arm.com>
 438             David Sherwood  <david.sherwood@arm.com>
 439
 440         * tree-vect-stmts.c (get_load_store_type): Treat the number of
 441         units as polynomial.  Reject VMAT_ELEMENTWISE and VMAT_STRIDED_SLP
 442         for variable-length vectors.
 443         (vectorizable_mask_load_store): Treat the number of units as
 444         polynomial, asserting that it is constant if the condition has
 445         already been enforced.
 446         (vectorizable_store, vectorizable_load): Likewise.
 447
 448 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 449             Alan Hayward  <alan.hayward@arm.com>
 450             David Sherwood  <david.sherwood@arm.com>
 451
 452         * tree-vect-loop.c (vectorizable_live_operation): Treat the number
 453         of units as polynomial.  Punt if we can't tell at compile time
 454         which vector contains the final result.
 455
 456 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 457             Alan Hayward  <alan.hayward@arm.com>
 458             David Sherwood  <david.sherwood@arm.com>
 459
 460         * tree-vect-loop.c (vectorizable_induction): Treat the number
 461         of units as polynomial.  Punt on SLP inductions.  Use an integer
 462         VEC_SERIES_EXPR for variable-length integer reductions.  Use a
 463         cast of such a series for variable-length floating-point
 464         reductions.
 465
 466 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 467             Alan Hayward  <alan.hayward@arm.com>
 468             David Sherwood  <david.sherwood@arm.com>
 469
 470         * tree.h (build_index_vector): Declare.
 471         * tree.c (build_index_vector): New function.
 472         * tree-vect-loop.c (get_initial_defs_for_reduction): Treat the number
 473         of units as polynomial, forcibly converting it to a constant if
 474         vectorizable_reduction has already enforced the condition.
 475         (vect_create_epilog_for_reduction): Likewise.  Use build_index_vector
 476         to create a {1,2,3,...} vector.
 477         (vectorizable_reduction): Treat the number of units as polynomial.
 478         Choose vectype_in based on the largest scalar element size rather
 479         than the smallest number of units.  Enforce the restrictions
 480         relied on above.
 481
 482 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 483             Alan Hayward  <alan.hayward@arm.com>
 484             David Sherwood  <david.sherwood@arm.com>
 485
 486         * tree-vect-data-refs.c (vector_alignment_reachable_p): Treat the
 487         number of units as polynomial.
 488
 489 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 490             Alan Hayward  <alan.hayward@arm.com>
 491             David Sherwood  <david.sherwood@arm.com>
 492
 493         * target.h (vector_sizes, auto_vector_sizes): New typedefs.
 494         * target.def (autovectorize_vector_sizes): Return the vector sizes
 495         by pointer, using vector_sizes rather than a bitmask.
 496         * targhooks.h (default_autovectorize_vector_sizes): Update accordingly.
 497         * targhooks.c (default_autovectorize_vector_sizes): Likewise.
 498         * config/aarch64/aarch64.c (aarch64_autovectorize_vector_sizes):
 499         Likewise.
 500         * config/arc/arc.c (arc_autovectorize_vector_sizes): Likewise.
 501         * config/arm/arm.c (arm_autovectorize_vector_sizes): Likewise.
 502         * config/i386/i386.c (ix86_autovectorize_vector_sizes): Likewise.
 503         * config/mips/mips.c (mips_autovectorize_vector_sizes): Likewise.
 504         * omp-general.c (omp_max_vf): Likewise.
 505         * omp-low.c (omp_clause_aligned_alignment): Likewise.
 506         * optabs-query.c (can_vec_mask_load_store_p): Likewise.
 507         * tree-vect-loop.c (vect_analyze_loop): Likewise.
 508         * tree-vect-slp.c (vect_slp_bb): Likewise.
 509         * doc/tm.texi: Regenerate.
 510         * tree-vectorizer.h (current_vector_size): Change from an unsigned int
 511         to a poly_uint64.
 512         * tree-vect-stmts.c (get_vectype_for_scalar_type_and_size): Take
 513         the vector size as a poly_uint64 rather than an unsigned int.
 514         (current_vector_size): Change from an unsigned int to a poly_uint64.
 515         (get_vectype_for_scalar_type): Update accordingly.
 516         * tree.h (build_truth_vector_type): Take the size and number of
 517         units as a poly_uint64 rather than an unsigned int.
 518         (build_vector_type): Add a temporary overload that takes
 519         the number of units as a poly_uint64 rather than an unsigned int.
 520         * tree.c (make_vector_type): Likewise.
 521         (build_truth_vector_type): Take the number of units as a poly_uint64
 522         rather than an unsigned int.
 523
 524 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 525             Alan Hayward  <alan.hayward@arm.com>
 526             David Sherwood  <david.sherwood@arm.com>
 527
 528         * target.def (get_mask_mode): Take the number of units and length
 529         as poly_uint64s rather than unsigned ints.
 530         * targhooks.h (default_get_mask_mode): Update accordingly.
 531         * targhooks.c (default_get_mask_mode): Likewise.
 532         * config/i386/i386.c (ix86_get_mask_mode): Likewise.
 533         * doc/tm.texi: Regenerate.
 534
 535 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 536             Alan Hayward  <alan.hayward@arm.com>
 537             David Sherwood  <david.sherwood@arm.com>
 538
 539         * omp-general.h (omp_max_vf): Return a poly_uint64 instead of an int.
 540         * omp-general.c (omp_max_vf): Likewise.
 541         * omp-expand.c (omp_adjust_chunk_size): Update call to omp_max_vf.
 542         (expand_omp_simd): Handle polynomial safelen.
 543         * omp-low.c (omplow_simd_context): Add a default constructor.
 544         (omplow_simd_context::max_vf): Change from int to poly_uint64.
 545         (lower_rec_simd_input_clauses): Update accordingly.
 546         (lower_rec_input_clauses): Likewise.
 547
 548 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 549             Alan Hayward  <alan.hayward@arm.com>
 550             David Sherwood  <david.sherwood@arm.com>
 551
 552         * tree-vectorizer.h (vect_nunits_for_cost): New function.
 553         * tree-vect-loop.c (vect_model_reduction_cost): Use it.
 554         * tree-vect-slp.c (vect_analyze_slp_cost_1): Likewise.
 555         (vect_analyze_slp_cost): Likewise.
 556         * tree-vect-stmts.c (vect_model_store_cost): Likewise.
 557         (vect_model_load_cost): Likewise.
 558
 559 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 560             Alan Hayward  <alan.hayward@arm.com>
 561             David Sherwood  <david.sherwood@arm.com>
 562
 563         * tree-vect-slp.c (vect_record_max_nunits, vect_build_slp_tree_1)
 564         (vect_build_slp_tree_2, vect_build_slp_tree): Change max_nunits
 565         from an unsigned int * to a poly_uint64_pod *.
 566         (calculate_unrolling_factor): New function.
 567         (vect_analyze_slp_instance): Use it.  Track polynomial max_nunits.
 568
 569 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 570             Alan Hayward  <alan.hayward@arm.com>
 571             David Sherwood  <david.sherwood@arm.com>
 572
 573         * tree-vectorizer.h (_slp_instance::unrolling_factor): Change
 574         from an unsigned int to a poly_uint64.
 575         (_loop_vec_info::slp_unrolling_factor): Likewise.
 576         (_loop_vec_info::vectorization_factor): Change from an int
 577         to a poly_uint64.
 578         (MAX_VECTORIZATION_FACTOR): Bump from 64 to INT_MAX.
 579         (vect_get_num_vectors): New function.
 580         (vect_update_max_nunits, vect_vf_for_cost): Likewise.
 581         (vect_get_num_copies): Use vect_get_num_vectors.
 582         (vect_analyze_data_ref_dependences): Change max_vf from an int *
 583         to an unsigned int *.
 584         (vect_analyze_data_refs): Change min_vf from an int * to a
 585         poly_uint64 *.
 586         (vect_transform_slp_perm_load): Take the vf as a poly_uint64 rather
 587         than an unsigned HOST_WIDE_INT.
 588         * tree-vect-data-refs.c (vect_analyze_possibly_independent_ddr)
 589         (vect_analyze_data_ref_dependence): Change max_vf from an int *
 590         to an unsigned int *.
 591         (vect_analyze_data_ref_dependences): Likewise.
 592         (vect_compute_data_ref_alignment): Handle polynomial vf.
 593         (vect_enhance_data_refs_alignment): Likewise.
 594         (vect_prune_runtime_alias_test_list): Likewise.
 595         (vect_shift_permute_load_chain): Likewise.
 596         (vect_supportable_dr_alignment): Likewise.
 597         (dependence_distance_ge_vf): Take the vectorization factor as a
 598         poly_uint64 rather than an unsigned HOST_WIDE_INT.
 599         (vect_analyze_data_refs): Change min_vf from an int * to a
 600         poly_uint64 *.
 601         * tree-vect-loop-manip.c (vect_gen_scalar_loop_niters): Take
 602         vfm1 as a poly_uint64 rather than an int.  Make the same change
 603         for the returned bound_scalar.
 604         (vect_gen_vector_loop_niters): Handle polynomial vf.
 605         (vect_do_peeling): Likewise.  Update call to
 606         vect_gen_scalar_loop_niters and handle polynomial bound_scalars.
 607         (vect_gen_vector_loop_niters_mult_vf): Assert that the vf must
 608         be constant.
 609         * tree-vect-loop.c (vect_determine_vectorization_factor)
 610         (vect_update_vf_for_slp, vect_analyze_loop_2): Handle polynomial vf.
 611         (vect_get_known_peeling_cost): Likewise.
 612         (vect_estimate_min_profitable_iters, vectorizable_reduction): Likewise.
 613         (vect_worthwhile_without_simd_p, vectorizable_induction): Likewise.
 614         (vect_transform_loop): Likewise.  Use the lowest possible VF when
 615         updating the upper bounds of the loop.
 616         (vect_min_worthwhile_factor): Make static.  Return an unsigned int
 617         rather than an int.
 618         * tree-vect-slp.c (vect_attempt_slp_rearrange_stmts): Cope with
 619         polynomial unroll factors.
 620         (vect_analyze_slp_cost_1, vect_analyze_slp_instance): Likewise.
 621         (vect_make_slp_decision): Likewise.
 622         (vect_supported_load_permutation_p): Likewise, and polynomial
 623         vf too.
 624         (vect_analyze_slp_cost): Handle polynomial vf.
 625         (vect_slp_analyze_node_operations): Likewise.
 626         (vect_slp_analyze_bb_1): Likewise.
 627         (vect_transform_slp_perm_load): Take the vf as a poly_uint64 rather
 628         than an unsigned HOST_WIDE_INT.
 629         * tree-vect-stmts.c (vectorizable_simd_clone_call, vectorizable_store)
 630         (vectorizable_load): Handle polynomial vf.
 631         * tree-vectorizer.c (simduid_to_vf::vf): Change from an int to
 632         a poly_uint64.
 633         (adjust_simduid_builtins, shrink_simd_arrays): Update accordingly.
 634
 635 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 636             Alan Hayward  <alan.hayward@arm.com>
 637             David Sherwood  <david.sherwood@arm.com>
 638
 639         * match.pd: Handle bit operations involving three constants
 640         and try to fold one pair.
 641
 642 2018-01-03  Richard Sandiford  <richard.sandiford@linaro.org>
 643
 644         * tree-vect-loop-manip.c: Include gimple-fold.h.
 645         (slpeel_make_loop_iterate_ntimes): Add step, final_iv and
 646         niters_maybe_zero parameters.  Handle other cases besides a step of 1.
 647         (vect_gen_vector_loop_niters): Add a step_vector_ptr parameter.
 648         Add a path that uses a step of VF instead of 1, but disable it
 649         for now.
 650         (vect_do_peeling): Add step_vector, niters_vector_mult_vf_var
 651         and niters_no_overflow parameters.  Update calls to
 652         slpeel_make_loop_iterate_ntimes and vect_gen_vector_loop_niters.
 653         Create a new SSA name if the latter choses to use a ste other
 654         than zero, and return it via niters_vector_mult_vf_var.
 655         * tree-vect-loop.c (vect_transform_loop): Update calls to
 656         vect_do_peeling, vect_gen_vector_loop_niters and
 657         slpeel_make_loop_iterate_ntimes.
 658         * tree-vectorizer.h (slpeel_make_loop_iterate_ntimes, vect_do_peeling)
 659         (vect_gen_vector_loop_niters): Update declarations after above changes.
 660
 661 2018-01-02  Michael Meissner  <meissner@linux.vnet.ibm.com>
 662
 663         * config/rs6000/rs6000.md (floor<mode>2): Add support for IEEE
 664         128-bit round to integer instructions.
 665         (ceil<mode>2): Likewise.
 666         (btrunc<mode>2): Likewise.
 667         (round<mode>2): Likewise.
 668
 669 2018-01-02  Aaron Sawdey  <acsawdey@linux.vnet.ibm.com>
 670
 671         * config/rs6000/rs6000-string.c (expand_block_move): Allow the use of
 672         unaligned VSX load/store on P8/P9.
 673         (expand_block_clear): Allow the use of unaligned VSX
 674         load/store on P8/P9.
 675
 676 2018-01-02  Bill Schmidt  <wschmidt@linux.vnet.ibm.com>
 677
 678         * config/rs6000/rs6000-p8swap.c (swap_feeds_both_load_and_store):
 679         New function.
 680         (rs6000_analyze_swaps): Mark a web unoptimizable if it contains a
 681         swap associated with both a load and a store.
 682
 683 2018-01-02  Andrew Waterman  <andrew@sifive.com>
 684
 685         * config/riscv/linux.h (ICACHE_FLUSH_FUNC): New.
 686         * config/riscv/riscv.md (clear_cache): Use it.
 687
 688 2018-01-02  Artyom Skrobov  <tyomitch@gmail.com>
 689
 690         * web.c: Remove out-of-date comment.
 691
 692 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 693
 694         * expr.c (fixup_args_size_notes): Check that any existing
 695         REG_ARGS_SIZE notes are correct, and don't try to re-add them.
 696         (emit_single_push_insn_1): Move stack_pointer_delta adjustment to...
 697         (emit_single_push_insn): ...here.
 698
 699 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 700
 701         * rtl.h (CONST_VECTOR_ELT): Redefine to const_vector_elt.
 702         (const_vector_encoded_nelts): New function.
 703         (CONST_VECTOR_NUNITS): Redefine to use GET_MODE_NUNITS.
 704         (const_vector_int_elt, const_vector_elt): Declare.
 705         * emit-rtl.c (const_vector_int_elt_1): New function.
 706         (const_vector_elt): Likewise.
 707         * simplify-rtx.c (simplify_immed_subreg): Avoid taking the address
 708         of CONST_VECTOR_ELT.
 709
 710 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 711
 712         * expr.c: Include rtx-vector-builder.h.
 713         (const_vector_mask_from_tree): Use rtx_vector_builder and operate
 714         directly on the tree encoding.
 715         (const_vector_from_tree): Likewise.
 716         * optabs.c: Include rtx-vector-builder.h.
 717         (expand_vec_perm_var): Use rtx_vector_builder and create a repeating
 718         sequence of "u" values.
 719         * vec-perm-indices.c: Include rtx-vector-builder.h.
 720         (vec_perm_indices_to_rtx): Use rtx_vector_builder and operate
 721         directly on the vec_perm_indices encoding.
 722
 723 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 724
 725         * doc/rtl.texi (const_vector): Describe new encoding scheme.
 726         * Makefile.in (OBJS): Add rtx-vector-builder.o.
 727         * rtx-vector-builder.h: New file.
 728         * rtx-vector-builder.c: Likewise.
 729         * rtl.h (rtx_def::u2): Add a const_vector field.
 730         (CONST_VECTOR_NPATTERNS): New macro.
 731         (CONST_VECTOR_NELTS_PER_PATTERN): Likewise.
 732         (CONST_VECTOR_DUPLICATE_P): Likewise.
 733         (CONST_VECTOR_STEPPED_P): Likewise.
 734         (CONST_VECTOR_ENCODED_ELT): Likewise.
 735         (const_vec_duplicate_p): Check for a duplicated vector encoding.
 736         (unwrap_const_vec_duplicate): Likewise.
 737         (const_vec_series_p): Check for a non-duplicated vector encoding.
 738         Say that the function only returns true for integer vectors.
 739         * emit-rtl.c: Include rtx-vector-builder.h.
 740         (gen_const_vec_duplicate_1): Delete.
 741         (gen_const_vector): Call gen_const_vec_duplicate instead of
 742         gen_const_vec_duplicate_1.
 743         (const_vec_series_p_1): Operate directly on the CONST_VECTOR encoding.
 744         (gen_const_vec_duplicate): Use rtx_vector_builder.
 745         (gen_const_vec_series): Likewise.
 746         (gen_rtx_CONST_VECTOR): Likewise.
 747         * config/powerpcspe/powerpcspe.c: Include rtx-vector-builder.h.
 748         (swap_const_vector_halves): Take an rtx pointer rather than rtx.
 749         Build a new vector rather than modifying a CONST_VECTOR in-place.
 750         (handle_special_swappables): Update call accordingly.
 751         * config/rs6000/rs6000-p8swap.c: Include rtx-vector-builder.h.
 752         (swap_const_vector_halves): Take an rtx pointer rather than rtx.
 753         Build a new vector rather than modifying a CONST_VECTOR in-place.
 754         (handle_special_swappables): Update call accordingly.
 755
 756 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 757
 758         * simplify-rtx.c (simplify_const_binary_operation): Use
 759         CONST_VECTOR_ELT instead of XVECEXP.
 760
 761 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 762
 763         * tree-cfg.c (verify_gimple_assign_ternary): Allow the size of
 764         the selector elements to be different from the data elements
 765         if the selector is a VECTOR_CST.
 766         * tree-vect-stmts.c (vect_gen_perm_mask_any): Use a vector of
 767         ssizetype for the selector.
 768
 769 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 770
 771         * optabs.c (shift_amt_for_vec_perm_mask): Try using series_p
 772         before testing each element individually.
 773         * tree-vect-generic.c (lower_vec_perm): Likewise.
 774
 775 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 776
 777         * selftest.h (selftest::vec_perm_indices_c_tests): Declare.
 778         * selftest-run-tests.c (selftest::run_tests): Call it.
 779         * vector-builder.h (vector_builder::operator ==): New function.
 780         (vector_builder::operator !=): Likewise.
 781         * vec-perm-indices.h (vec_perm_indices::series_p): Declare.
 782         (vec_perm_indices::all_from_input_p): New function.
 783         * vec-perm-indices.c (vec_perm_indices::series_p): Likewise.
 784         (test_vec_perm_12, selftest::vec_perm_indices_c_tests): Likewise.
 785         * fold-const.c (fold_ternary_loc): Use tree_to_vec_perm_builder
 786         instead of reading the VECTOR_CST directly.  Detect whether both
 787         vector inputs are the same before constructing the vec_perm_indices,
 788         and update the number of inputs argument accordingly.  Use the
 789         utility functions added above.  Only construct sel2 if we need to.
 790
 791 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 792
 793         * optabs.c (expand_vec_perm_var): Use an explicit encoding for
 794         the broadcast of the low byte.
 795         (expand_mult_highpart): Use an explicit encoding for the permutes.
 796         * optabs-query.c (can_mult_highpart_p): Likewise.
 797         * tree-vect-loop.c (calc_vec_perm_mask_for_shift): Likewise.
 798         * tree-vect-stmts.c (perm_mask_for_reverse): Likewise.
 799         (vectorizable_bswap): Likewise.
 800         * tree-vect-data-refs.c (vect_grouped_store_supported): Use an
 801         explicit encoding for the power-of-2 permutes.
 802         (vect_permute_store_chain): Likewise.
 803         (vect_grouped_load_supported): Likewise.
 804         (vect_permute_load_chain): Likewise.
 805
 806 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 807
 808         * vec-perm-indices.h (vec_perm_indices_to_tree): Declare.
 809         * vec-perm-indices.c (vec_perm_indices_to_tree): New function.
 810         * tree-ssa-forwprop.c (simplify_vector_constructor): Use it.
 811         * tree-vect-slp.c (vect_transform_slp_perm_load): Likewise.
 812         * tree-vect-stmts.c (vectorizable_bswap): Likewise.
 813         (vect_gen_perm_mask_any): Likewise.
 814
 815 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 816
 817         * int-vector-builder.h: New file.
 818         * vec-perm-indices.h: Include int-vector-builder.h.
 819         (vec_perm_indices): Redefine as an int_vector_builder.
 820         (auto_vec_perm_indices): Delete.
 821         (vec_perm_builder): Redefine as a stand-alone class.
 822         (vec_perm_indices::vec_perm_indices): New function.
 823         (vec_perm_indices::clamp): Likewise.
 824         * vec-perm-indices.c: Include fold-const.h and tree-vector-builder.h.
 825         (vec_perm_indices::new_vector): New function.
 826         (vec_perm_indices::new_expanded_vector): Update for new
 827         vec_perm_indices class.
 828         (vec_perm_indices::rotate_inputs): New function.
 829         (vec_perm_indices::all_in_range_p): Operate directly on the
 830         encoded form, without computing elided elements.
 831         (tree_to_vec_perm_builder): Operate directly on the VECTOR_CST
 832         encoding.  Update for new vec_perm_indices class.
 833         * optabs.c (expand_vec_perm_const): Create a vec_perm_indices for
 834         the given vec_perm_builder.
 835         (expand_vec_perm_var): Update vec_perm_builder constructor.
 836         (expand_mult_highpart): Use vec_perm_builder instead of
 837         auto_vec_perm_indices.
 838         * optabs-query.c (can_mult_highpart_p): Use vec_perm_builder and
 839         vec_perm_indices instead of auto_vec_perm_indices.  Use a single
 840         or double series encoding as appropriate.
 841         * fold-const.c (fold_ternary_loc): Use vec_perm_builder and
 842         vec_perm_indices instead of auto_vec_perm_indices.
 843         * tree-ssa-forwprop.c (simplify_vector_constructor): Likewise.
 844         * tree-vect-data-refs.c (vect_grouped_store_supported): Likewise.
 845         (vect_permute_store_chain): Likewise.
 846         (vect_grouped_load_supported): Likewise.
 847         (vect_permute_load_chain): Likewise.
 848         (vect_shift_permute_load_chain): Likewise.
 849         * tree-vect-slp.c (vect_build_slp_tree_1): Likewise.
 850         (vect_transform_slp_perm_load): Likewise.
 851         (vect_schedule_slp_instance): Likewise.
 852         * tree-vect-stmts.c (perm_mask_for_reverse): Likewise.
 853         (vectorizable_mask_load_store): Likewise.
 854         (vectorizable_bswap): Likewise.
 855         (vectorizable_store): Likewise.
 856         (vectorizable_load): Likewise.
 857         * tree-vect-generic.c (lower_vec_perm): Use vec_perm_builder and
 858         vec_perm_indices instead of auto_vec_perm_indices.  Use
 859         tree_to_vec_perm_builder to read the vector from a tree.
 860         * tree-vect-loop.c (calc_vec_perm_mask_for_shift): Take a
 861         vec_perm_builder instead of a vec_perm_indices.
 862         (have_whole_vector_shift): Use vec_perm_builder and
 863         vec_perm_indices instead of auto_vec_perm_indices.  Leave the
 864         truncation to calc_vec_perm_mask_for_shift.
 865         (vect_create_epilog_for_reduction): Likewise.
 866         * config/aarch64/aarch64.c (expand_vec_perm_d::perm): Change
 867         from auto_vec_perm_indices to vec_perm_indices.
 868         (aarch64_expand_vec_perm_const_1): Use rotate_inputs on d.perm
 869         instead of changing individual elements.
 870         (aarch64_vectorize_vec_perm_const): Use new_vector to install
 871         the vector in d.perm.
 872         * config/arm/arm.c (expand_vec_perm_d::perm): Change
 873         from auto_vec_perm_indices to vec_perm_indices.
 874         (arm_expand_vec_perm_const_1): Use rotate_inputs on d.perm
 875         instead of changing individual elements.
 876         (arm_vectorize_vec_perm_const): Use new_vector to install
 877         the vector in d.perm.
 878         * config/powerpcspe/powerpcspe.c (rs6000_expand_extract_even):
 879         Update vec_perm_builder constructor.
 880         (rs6000_expand_interleave): Likewise.
 881         * config/rs6000/rs6000.c (rs6000_expand_extract_even): Likewise.
 882         (rs6000_expand_interleave): Likewise.
 883
 884 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 885
 886         * optabs-query.c (can_vec_perm_var_p): Check whether lowering
 887         to qimode could truncate the indices.
 888         * optabs.c (expand_vec_perm_var): Likewise.
 889
 890 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
 891
 892         * Makefile.in (OBJS): Add vec-perm-indices.o.
 893         * vec-perm-indices.h: New file.
 894         * vec-perm-indices.c: Likewise.
 895         * target.h (vec_perm_indices): Replace with a forward class
 896         declaration.
 897         (auto_vec_perm_indices): Move to vec-perm-indices.h.
 898         * optabs.h: Include vec-perm-indices.h.
 899         (expand_vec_perm): Delete.
 900         (selector_fits_mode_p, expand_vec_perm_var): Declare.
 901         (expand_vec_perm_const): Declare.
 902         * target.def (vec_perm_const_ok): Replace with...
 903         (vec_perm_const): ...this new hook.
 904         * doc/tm.texi.in (TARGET_VECTORIZE_VEC_PERM_CONST_OK): Replace with...
 905         (TARGET_VECTORIZE_VEC_PERM_CONST): ...this new hook.
 906         * doc/tm.texi: Regenerate.
 907         * optabs.def (vec_perm_const): Delete.
 908         * doc/md.texi (vec_perm_const): Likewise.
 909         (vec_perm): Refer to TARGET_VECTORIZE_VEC_PERM_CONST.
 910         * expr.c (expand_expr_real_2): Use expand_vec_perm_const rather than
 911         expand_vec_perm for constant permutation vectors.  Assert that
 912         the mode of variable permutation vectors is the integer equivalent
 913         of the mode that is being permuted.
 914         * optabs-query.h (selector_fits_mode_p): Declare.
 915         * optabs-query.c: Include vec-perm-indices.h.
 916         (selector_fits_mode_p): New function.
 917         (can_vec_perm_const_p): Check whether targetm.vectorize.vec_perm_const
 918         is defined, instead of checking whether the vec_perm_const_optab
 919         exists.  Use targetm.vectorize.vec_perm_const instead of
 920         targetm.vectorize.vec_perm_const_ok.  Check whether the indices
 921         fit in the vector mode before using a variable permute.
 922         * optabs.c (shift_amt_for_vec_perm_mask): Take a mode and a
 923         vec_perm_indices instead of an rtx.
 924         (expand_vec_perm): Replace with...
 925         (expand_vec_perm_const): ...this new function.  Take the selector
 926         as a vec_perm_indices rather than an rtx.  Also take the mode of
 927         the selector.  Update call to shift_amt_for_vec_perm_mask.
 928         Use targetm.vectorize.vec_perm_const instead of vec_perm_const_optab.
 929         Use vec_perm_indices::new_expanded_vector to expand the original
 930         selector into bytes.  Check whether the indices fit in the vector
 931         mode before using a variable permute.
 932         (expand_vec_perm_var): Make global.
 933         (expand_mult_highpart): Use expand_vec_perm_const.
 934         * fold-const.c: Includes vec-perm-indices.h.
 935         * tree-ssa-forwprop.c: Likewise.
 936         * tree-vect-data-refs.c: Likewise.
 937         * tree-vect-generic.c: Likewise.
 938         * tree-vect-loop.c: Likewise.
 939         * tree-vect-slp.c: Likewise.
 940         * tree-vect-stmts.c: Likewise.
 941         * config/aarch64/aarch64-protos.h (aarch64_expand_vec_perm_const):
 942         Delete.
 943         * config/aarch64/aarch64-simd.md (vec_perm_const<mode>): Delete.
 944         * config/aarch64/aarch64.c (aarch64_expand_vec_perm_const)
 945         (aarch64_vectorize_vec_perm_const_ok): Fuse into...
 946         (aarch64_vectorize_vec_perm_const): ...this new function.
 947         (TARGET_VECTORIZE_VEC_PERM_CONST_OK): Delete.
 948         (TARGET_VECTORIZE_VEC_PERM_CONST): Redefine.
 949         * config/arm/arm-protos.h (arm_expand_vec_perm_const): Delete.
 950         * config/arm/vec-common.md (vec_perm_const<mode>): Delete.
 951         * config/arm/arm.c (TARGET_VECTORIZE_VEC_PERM_CONST_OK): Delete.
 952         (TARGET_VECTORIZE_VEC_PERM_CONST): Redefine.
 953         (arm_expand_vec_perm_const, arm_vectorize_vec_perm_const_ok): Merge
 954         into...
 955         (arm_vectorize_vec_perm_const): ...this new function.  Explicitly
 956         check for NEON modes.
 957         * config/i386/i386-protos.h (ix86_expand_vec_perm_const): Delete.
 958         * config/i386/sse.md (VEC_PERM_CONST, vec_perm_const<mode>): Delete.
 959         * config/i386/i386.c (ix86_expand_vec_perm_const_1): Update comment.
 960         (ix86_expand_vec_perm_const, ix86_vectorize_vec_perm_const_ok): Merge
 961         into...
 962         (ix86_vectorize_vec_perm_const): ...this new function.  Incorporate
 963         the old VEC_PERM_CONST conditions.
 964         * config/ia64/ia64-protos.h (ia64_expand_vec_perm_const): Delete.
 965         * config/ia64/vect.md (vec_perm_const<mode>): Delete.
 966         * config/ia64/ia64.c (ia64_expand_vec_perm_const)
 967         (ia64_vectorize_vec_perm_const_ok): Merge into...
 968         (ia64_vectorize_vec_perm_const): ...this new function.
 969         * config/mips/loongson.md (vec_perm_const<mode>): Delete.
 970         * config/mips/mips-msa.md (vec_perm_const<mode>): Delete.
 971         * config/mips/mips-ps-3d.md (vec_perm_constv2sf): Delete.
 972         * config/mips/mips-protos.h (mips_expand_vec_perm_const): Delete.
 973         * config/mips/mips.c (mips_expand_vec_perm_const)
 974         (mips_vectorize_vec_perm_const_ok): Merge into...
 975         (mips_vectorize_vec_perm_const): ...this new function.
 976         * config/powerpcspe/altivec.md (vec_perm_constv16qi): Delete.
 977         * config/powerpcspe/paired.md (vec_perm_constv2sf): Delete.
 978         * config/powerpcspe/spe.md (vec_perm_constv2si): Delete.
 979         * config/powerpcspe/vsx.md (vec_perm_const<mode>): Delete.
 980         * config/powerpcspe/powerpcspe-protos.h (altivec_expand_vec_perm_const)
 981         (rs6000_expand_vec_perm_const): Delete.
 982         * config/powerpcspe/powerpcspe.c (TARGET_VECTORIZE_VEC_PERM_CONST_OK):
 983         Delete.
 984         (TARGET_VECTORIZE_VEC_PERM_CONST): Redefine.
 985         (altivec_expand_vec_perm_const_le): Take each operand individually.
 986         Operate on constant selectors rather than rtxes.
 987         (altivec_expand_vec_perm_const): Likewise.  Update call to
 988         altivec_expand_vec_perm_const_le.
 989         (rs6000_expand_vec_perm_const): Delete.
 990         (rs6000_vectorize_vec_perm_const_ok): Delete.
 991         (rs6000_vectorize_vec_perm_const): New function.
 992         (rs6000_do_expand_vec_perm): Take a vec_perm_builder instead of
 993         an element count and rtx array.
 994         (rs6000_expand_extract_even): Update call accordingly.
 995         (rs6000_expand_interleave): Likewise.
 996         * config/rs6000/altivec.md (vec_perm_constv16qi): Delete.
 997         * config/rs6000/paired.md (vec_perm_constv2sf): Delete.
 998         * config/rs6000/vsx.md (vec_perm_const<mode>): Delete.
 999         * config/rs6000/rs6000-protos.h (altivec_expand_vec_perm_const)
1000         (rs6000_expand_vec_perm_const): Delete.
1001         * config/rs6000/rs6000.c (TARGET_VECTORIZE_VEC_PERM_CONST_OK): Delete.
1002         (TARGET_VECTORIZE_VEC_PERM_CONST): Redefine.
1003         (altivec_expand_vec_perm_const_le): Take each operand individually.
1004         Operate on constant selectors rather than rtxes.
1005         (altivec_expand_vec_perm_const): Likewise.  Update call to
1006         altivec_expand_vec_perm_const_le.
1007         (rs6000_expand_vec_perm_const): Delete.
1008         (rs6000_vectorize_vec_perm_const_ok): Delete.
1009         (rs6000_vectorize_vec_perm_const): New function.  Remove stray
1010         reference to the SPE evmerge intructions.
1011         (rs6000_do_expand_vec_perm): Take a vec_perm_builder instead of
1012         an element count and rtx array.
1013         (rs6000_expand_extract_even): Update call accordingly.
1014         (rs6000_expand_interleave): Likewise.
1015         * config/sparc/sparc.md (vec_perm_constv8qi): Delete in favor of...
1016         * config/sparc/sparc.c (sparc_vectorize_vec_perm_const): ...this
1017         new function.
1018         (TARGET_VECTORIZE_VEC_PERM_CONST): Redefine.
1019
1020 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
1021
1022         * optabs.c (expand_vec_perm_1): Assert that SEL has an integer
1023         vector mode and that that mode matches the mode of the data
1024         being permuted.
1025         (expand_vec_perm): Split handling of non-CONST_VECTOR selectors
1026         out into expand_vec_perm_var.  Do all CONST_VECTOR handling here,
1027         directly using expand_vec_perm_1 when forcing selectors into
1028         registers.
1029         (expand_vec_perm_var): New function, split out from expand_vec_perm.
1030
1031 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
1032
1033         * optabs-query.h (can_vec_perm_p): Delete.
1034         (can_vec_perm_var_p, can_vec_perm_const_p): Declare.
1035         * optabs-query.c (can_vec_perm_p): Split into...
1036         (can_vec_perm_var_p, can_vec_perm_const_p): ...these two functions.
1037         (can_mult_highpart_p): Use can_vec_perm_const_p to test whether a
1038         particular selector is valid.
1039         * tree-ssa-forwprop.c (simplify_vector_constructor): Likewise.
1040         * tree-vect-data-refs.c (vect_grouped_store_supported): Likewise.
1041         (vect_grouped_load_supported): Likewise.
1042         (vect_shift_permute_load_chain): Likewise.
1043         * tree-vect-slp.c (vect_build_slp_tree_1): Likewise.
1044         (vect_transform_slp_perm_load): Likewise.
1045         * tree-vect-stmts.c (perm_mask_for_reverse): Likewise.
1046         (vectorizable_bswap): Likewise.
1047         (vect_gen_perm_mask_checked): Likewise.
1048         * fold-const.c (fold_ternary_loc): Likewise.  Don't take
1049         implementations of variable permutation vectors into account
1050         when deciding which selector to use.
1051         * tree-vect-loop.c (have_whole_vector_shift): Don't check whether
1052         vec_perm_const_optab is supported; instead use can_vec_perm_const_p
1053         with a false third argument.
1054         * tree-vect-generic.c (lower_vec_perm): Use can_vec_perm_const_p
1055         to test whether the constant selector is valid and can_vec_perm_var_p
1056         to test whether a variable selector is valid.
1057
1058 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
1059
1060         * optabs-query.h (can_vec_perm_p): Take a const vec_perm_indices *.
1061         * optabs-query.c (can_vec_perm_p): Likewise.
1062         * fold-const.c (fold_vec_perm): Take a const vec_perm_indices &
1063         instead of vec_perm_indices.
1064         * tree-vectorizer.h (vect_gen_perm_mask_any): Likewise,
1065         (vect_gen_perm_mask_checked): Likewise,
1066         * tree-vect-stmts.c (vect_gen_perm_mask_any): Likewise,
1067         (vect_gen_perm_mask_checked): Likewise,
1068
1069 2018-01-02  Richard Sandiford  <richard.sandiford@linaro.org>
1070
1071         * optabs-query.h (qimode_for_vec_perm): Declare.
1072         * optabs-query.c (can_vec_perm_p): Split out qimode search to...
1073         (qimode_for_vec_perm): ...this new function.
1074         * optabs.c (expand_vec_perm): Use qimode_for_vec_perm.
1075
1076 2018-01-02  Aaron Sawdey  <acsawdey@linux.vnet.ibm.com>
1077
1078         * rtlanal.c (canonicalize_condition): Return 0 if final rtx
1079         does not have a conditional at the top.
1080
1081 2018-01-02  Richard Biener  <rguenther@suse.de>
1082
1083         * ipa-inline.c (big_speedup_p): Fix expression.
1084
1085 2018-01-02  Jan Hubicka  <hubicka@ucw.cz>
1086
1087         PR target/81616
1088         * config/i386/x86-tune-costs.h: Increase cost of integer load costs
1089         for generic 4->6.
1090
1091 2018-01-02  Jan Hubicka  <hubicka@ucw.cz>
1092
1093         PR target/81616
1094         Generic tuning.
1095         * x86-tune-costs.h (generic_cost): Reduce cost of FDIV 20->17,
1096         cost of sqrt 20->14, DIVSS 18->13, DIVSD 32->17, SQRtSS 30->14
1097         and SQRTsD 58->18, cond_not_taken_branch_cost. 2->1. Increase
1098         cond_taken_branch_cost 3->4.
1099
1100 2018-01-01  Jakub Jelinek  <jakub@redhat.com>
1101
1102         PR tree-optimization/83581
1103         * tree-loop-distribution.c (pass_loop_distribution::execute): Return
1104         TODO_cleanup_cfg if any changes have been made.
1105
1106         PR middle-end/83608
1107         * expr.c (store_expr_with_bounds): Use simplify_gen_subreg instead of
1108         convert_modes if target mode has the right side, but different mode
1109         class.
1110
1111         PR middle-end/83609
1112         * expr.c (expand_assignment): Fix up a typo in simplify_gen_subreg
1113         last argument when extracting from CONCAT.  If either from_real or
1114         from_imag is NULL, use expansion through memory.  If result is not
1115         a CONCAT and simplify_gen_subreg fails, try to simplify_gen_subreg
1116         the parts directly to inner mode, if even that fails, use expansion
1117         through memory.
1118
1119         PR middle-end/83623
1120         * expmed.c (expand_shift_1): For 2-byte rotates by BITS_PER_UNIT,
1121         check for bswap in mode rather than HImode and use that in expand_unop
1122         too.
1123 \f
1124 Copyright (C) 2018 Free Software Foundation, Inc.
1125
1126 Copying and distribution of this file, with or without modification,
1127 are permitted in any medium without royalty provided the copyright
1128 notice and this notice are preserved.