]> git.ipfire.org Git - thirdparty/gcc.git/blobdiff - gcc/ChangeLog
[arm] Early split addvdi4
[thirdparty/gcc.git] / gcc / ChangeLog
index 2424b977a5dd22c1702b8a6d2c22117bacb995e1..8275411b671eee980fa73744bc41df07de0eb6df 100644 (file)
@@ -1,3 +1,475 @@
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (arm_select_cc_mode): Allow either the first
+       or second operand of the PLUS inside a DImode equality test to be
+       sign-extend when selecting CC_Vmode.
+       * config/arm/arm.md (addvdi4): Early-split the operation into SImode
+       instructions.
+       (addsi3_cin_vout_reg, addsi3_cin_vout_imm, addsi3_cin_vout_0): New
+       expand patterns.
+       (addsi3_cin_vout_reg_insn, addsi3_cin_vout_imm_insn): New patterns.
+       (addsi3_cin_vout_0): Likewise.
+       (adddi3_compareV): Delete.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (addsi3_compareV_reg_nosum): New insn.
+       (addsi3_compareV_imm_nosum): New insn.  Also add peephole2 patterns
+       to transform this back into the summation version when that leads
+       to smaller code.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (addv<mode>4): Delete.
+       (addvsi4): New pattern.  Handle immediate values that the architecture
+       supports.
+       (addvdi4): New pattern.
+       (addsi3_compareV): Rename to ...
+       (addsi3_compareV_reg): ... this.  Add constraints for thumb2 variants
+       and use COMPARE rather than NE.
+       (addsi3_compareV_imm): New pattern.
+       * config/arm/arm.c (arm_select_cc_mode): Return CC_Vmode for
+       a signed-overflow check.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-modes.def (CC_ADC): New CC mode.
+       * config/arm/arm.c (arm_select_cc_mode): Detect selection of
+       CC_ADCmode.
+       (maybe_get_arm_condition_code): Handle CC_ADCmode.
+       * config/arm/arm.md (uaddvdi4): Early expansion of unsigned addition
+       with overflow.
+       (addsi3_cin_cout_reg, addsi3_cin_cout_imm, addsi3_cin_cout_0): New
+       expand patterns.
+       (addsi3_cin_cout_reg_insn, addsi3_cin_cout_0_insn): New insn patterns
+       (addsi3_cin_cout_imm_insn): Likewise.
+       (adddi3_compareC): Delete insn.
+       * config/arm/predicates.md (arm_carry_operation): Handle CC_ADCmode.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (adddi3): Call gen_addsi3_compare_op1.
+       * (uaddv<mode>4): Delete expansion pattern.
+       (uaddvsi4): New pattern.
+       (uaddvdi4): Likewise.
+       (addsi3_compareC): Delete pattern, change callers to use
+       addsi3_compare_op1.
+       (addsi3_compare_op1): No-longer anonymous.  Clean up constraints to
+       reduce the number of alternatives and re-work type attribute handling.
+       (addsi3_compare_op2): Clean up constraints to reduce the number of
+       alternatives and re-work type attribute handling.
+       (compare_addsi2_op0): Likewise.
+       (compare_addsi2_op1): Likewise.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-modes.def (CC_NCV, CC_CZ): Delete CC modes.
+       * config/arm/arm.c (arm_select_cc_mode): Remove old selection code
+       for DImode operands.
+       (arm_gen_dicompare_reg): Remove unreachable expansion code.
+       (maybe_get_arm_condition_code): Remove support for CC_CZmode and
+       CC_NCVmode.
+       * config/arm/arm.md (arm_cmpdi_insn): Delete.
+       (arm_cmpdi_unsigned): Delete.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (arm_const_double_prefer_rsbs_rsc): New function.
+       (arm_canonicalize_comparison): For GT/LE/GTU/GEU, use the constant
+       unchanged only if that will be cheaper.
+       (arm_select_cc_mode): Recognize a swapped comparison that will
+       be regenerated using RSBS or RSCS.  Relax restriction on selecting
+       CC_RSBmode.
+       (arm_gen_dicompare_reg): Handle LE/GT/LEU/GEU comparisons against
+       a constant.
+       (arm_gen_compare_reg): Handle compare (CONST, X) when the mode
+       is CC_RSBmode.
+       (maybe_get_arm_condition_code): CC_RSBmode now returns the same codes
+       as CCmode.
+       * config/arm/arm.md (rsb_imm_compare_scratch): New pattern.
+       (rscsi3_<CC_EXTEND>out_scratch): New pattern.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-modes.def (CC_NV, CC_B): New CC modes.
+       * config/arm/arm.c (arm_select_cc_mode): Recognize constructs that
+       need these modes.
+       (arm_gen_dicompare_reg): New code to early expand the sub-operations
+       of EQ, NE, LT, GE, LTU and GEU.
+       * config/arm/iterators.md (CC_EXTEND): New code attribute.
+       * config/arm/predicates.md (arm_adcimm_operand): New predicate..
+       * config/arm/arm.md (cmpsi3_carryin_<CC_EXTEND>out): New pattern.
+       (cmpsi3_imm_carryin_<CC_EXTEND>out): Likewise.
+       (cmpsi3_0_carryin_<CC_EXTEND>out): Likewise.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (cbranchdi4): Accept reg_or_int_operand for
+       operand 2.
+       (cstoredi4): Similarly, but for operand 3.
+       * config/arm/arm.c (arm_canoncialize_comparison): Allow
+       canonicalization of unsigned compares with a constant on Arm.
+       Prefer using const+1 and adjusting the comparison over swapping the
+       operands whenever the original constant was not valid.
+       (arm_gen_dicompare_reg): If Y is not a valid operand, force it to a
+       register here.
+       (arm_validize_comparison): Do not force invalid DImode operands to
+       registers here.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (arm_select_cc_mode): For DImode equality tests
+       return CC_Zmode if comparing against a constant where one word is
+       zero.
+       (arm_gen_compare_reg): Split DImode handling to ...
+       (arm_gen_dicompare_reg): ... here.  Handle equality comparisons
+       against simple constants.
+       * config/arm/arm.md (arm_cmpdi_zero): Delete pattern.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (subsi3_carryin_shift_alt): New pattern.
+       (rsbsi3_carryin_shift_alt): Likewise.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (negscc_borrow): New pattern.
+       (mov_negscc): Don't split if the insn would match negscc_borrow.
+       * config/arm/thumb2.md (thumb2_mov_negscc): Likewise.
+       (thumb2_mov_negscc_strict_it): Likewise.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (arm_insn_cost): New function.
+       (TARGET_INSN_COST): Override default definition.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (arm_rtx_costs_internal, case MINUS): Handle
+       borrow operations.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (strip_carry_operation): New function.
+       (arm_rtx_costs_internal, case PLUS): Handle addtion with carry-in
+       for SImode.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/predicates.md (arm_carry_operation): New special
+       predicate.
+       * config/arm/iterators.md (LTUGEU): Delete iterator.
+       (cnb): Delete code attribute.
+       (optab): Delete ltu and geu elements.
+       * config/arm/arm.md (addsi3_carryin): Renamed from
+       addsi3_carryin_<optab>.  Remove iterator and use arm_carry_operand.
+       (add0si3_carryin): Similarly, but from add0si3_carryin_<optab>.
+       (addsi3_carryin_alt2): Similarly, but from addsi3_carryin_alt2_<optab>.
+       (addsi3_carryin_clobercc): Similarly.
+       (addsi3_carryin_shift): Similarly.  Do not allow register shifts in
+       Thumb2 state.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (arm_subdi3): Delete insn.
+       (zextendsidi_negsi, negdi_extendsidi): Delete insn_and_split.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-modes.def (CC_RSB): New CC mode.
+       * config/arm/predicates.md (arm_borrow_operation): Handle CC_RSBmode.
+       * config/arm/arm.c (arm_select_cc_mode): Detect when we should
+       return CC_RSBmode.
+       (maybe_get_arm_condition_code): Handle CC_RSBmode.
+       * config/arm/arm.md (subsi3_carryin): Make this pattern available to
+       expand.
+       (subdi3): Rewrite to early-expand the sub-operations.
+       (rsb_im_compare): New pattern.
+       (negdi2): Delete.
+       (negdi2_insn): Delete.
+       (arm_negsi2): Correct type attribute to alu_imm.
+       (negsi2_0compare): New insn pattern.
+       (negsi2_carryin): New insn pattern.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (addsi3_carryin_alt2): Use arm_not_operand for
+       operand 2.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (addsi3_carryin_shift_<optab>): Reorder operands
+       to match canonical form.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.md (zero_extend<mode>di2): Convert to define_expand.
+       (extend<mode>di2): Likewise.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-protos.h (arm_decompose_di_binop): New prototype.
+       * config/arm/arm.c (arm_decompose_di_binop): New function.
+       * config/arm/arm.md (adddi3): Also accept any const_int for op2.
+       If not generating Thumb-1 code, decompose the operation into 32-bit
+       pieces.
+       * add0si_carryin_<optab>: New pattern.
+
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+       * arm.md (adddi3): Only accept register operands.
+       (arm_adddi3): Convert to simple insn with no split.  Do not accept
+       constants.
+       (adddi_sesidi_di): Delete patern.
+       (adddi_zesidi_di): Likewise.
+       (uaddv<mode>4): Use LTU as condition for branch.
+       (adddi3_compareV): Convert to simple insn with no split.
+       (addsi3_compareV_upper): Delete pattern.
+       (adddi3_compareC): Convert to simple insn with no split.  Correct
+       flags setting expression.
+       (addsi3_compareC_upper): Delete pattern.
+       (addsi3_compareC): Correct flags setting expression.
+       (subdi3_compare1): Convert to simple insn with no split.
+       (subsi3_carryin_compare): Delete pattern.
+       (arm_subdi3): Convert to simple insn with no split.
+       (subdi_zesidi): Delete pattern.
+       (subdi_di_sesidi): Delete pattern.
+       (subdi_zesidi_di): Delete pattern.
+       (subdi_sesidi_di): Delete pattern.
+       (subdi_zesidi_zesidi): Delete pattern.
+       (negvdi3): Use s_register_operand.
+       (negdi2_compare): Convert to simple insn with no split.
+       (negdi2_insn): Likewise.
+       (negsi2_carryin_compare): Delete pattern.
+       (negdi_zero_extendsidi): Delete pattern.
+       (arm_cmpdi_insn): Convert to simple insn with no split.
+       (negdi2): Don't call gen_negdi2_neon.
+       * config/arm/neon.md (adddi3_neon): Delete pattern.
+       (subdi3_neon): Delete pattern.
+       (negdi2_neon): Delete pattern.
+       (splits for negdi2_neon): Delete splits.
+
+2019-10-18  Jakub Jelinek  <jakub@redhat.com>
+
+       PR middle-end/92153
+       * ggc-page.c (release_pages): Read g->alloc_size before free rather
+       than after it.
+
+2019-10-18  Andre Vieira  <andre.simoesdiasvieira@arm.com>
+
+       * config/arm/t-multilib: Add rule to regenerate mutlilib header file
+       with any change to t-multilib, t-aprofile and t-rmprofile.  Also add
+       new multilib variants and new mappings.
+
+2019-10-18  Georg-Johann Lay  <avr@gjlay.de>
+
+       PR target/86040
+       * config/avr/avr.c (avr_out_lpm): Do not shortcut-return.
+
+2019-10-18  Prathamesh Kulkarni  <prathamesh.kulkarni@linaro.org>
+           Richard Sandiford  <richard.sandiford@arm.com>
+
+       PR target/86753
+       * tree-vectorizer.h (scalar_cond_masked_key): New struct,
+       and define hashmap traits for it.
+       (loop_vec_info::scalar_cond_masked_set): New member.
+       (vect_record_loop_mask): Adjust prototype.
+       * tree-vectorizer.c (scalar_cond_masked_key::get_cond_ops_from_tree):
+       Implement method.
+       * tree-vect-loop.c (vectorizable_reduction): Pass NULL as last arg to
+       vect_record_loop_mask.
+       (vectorizable_live_operation): Likewise.
+       (vect_record_loop_mask): New param scalar_mask. Add entry
+       cond, loop_mask to scalar_cond_masked_set if scalar_mask is non NULL.
+       * tree-vect-stmts.c (check_load_store_masking): New param scalar_mask.
+       Pass it as last arg to vect_record_loop_mask.
+       (vectorizable_call): Pass scalar_mask as last arg to
+       vect_record_loop_mask.
+       (vectorizable_store): Likewise.
+       (vectorizable_load): Likewise.
+       (vectorizable_condition): Check if another part of vectorized code
+       applies loop_mask to condition or to it's inverse, and if yes,
+       apply loop_mask to result of vector comparison.
+
+2019-10-17  John David Anglin  <danglin@gcc.gnu.org>
+
+       * config/pa/pa.c (pa_output_indirect_call): Fix typos in last change.
+
+2019-10-18  Jakub Jelinek  <jakub@redhat.com>
+
+       PR tree-optimization/92056
+       * tree-ssa-strlen.c (determine_min_objsize): Call init_object_sizes
+       before calling compute_builtin_object_size.
+
+2019-10-17  Iain Sandoe  <iain@sandoe.co.uk>
+
+       PR target/65342
+       * config/rs6000/darwin.md (movdi_low, movsi_low_st): Delete.
+       (movdi_low_st): Delete.
+       * config/rs6000/rs6000.c
+       (darwin_rs6000_legitimate_lo_sum_const_p): New.
+       (mem_operand_gpr): Validate Mach-O LO_SUM cases separately.
+       * config/rs6000/rs6000.md (movsi_low): Delete.
+
+2019-10-17  Jason Merrill  <jason@redhat.com>
+
+       * gimplify.h (get_initialized_tmp_var): Add default argument to
+       post_p.
+       * gimplify.c (gimplify_self_mod_expr, gimplify_omp_atomic): Remove
+       NULL post_p argument.
+       * targhooks (std_gimplify_va_arg_expr): Likewise.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       * tree-vectorizer.h (_stmt_vec_info::cond_reduc_code): Remove.
+       (STMT_VINFO_VEC_COND_REDUC_CODE): Likewise.
+       * tree-vectorizer.c (vec_info::new_stmt_vec_info): Do not
+       initialize STMT_VINFO_VEC_COND_REDUC_CODE.
+       * tree-vect-loop.c (vect_is_simple_reduction): Set
+       STMT_VINFO_REDUC_CODE.
+       (vectorizable_reduction): Remove dead and redundant code, use
+       STMT_VINFO_REDUC_CODE instead of STMT_VINFO_VEC_COND_REDUC_CODE.
+
+2019-10-17  Georg-Johann Lay  <avr@gjlay.de>
+
+       Fix breakage introduced by r276985.
+
+       * config/avr/avr.c (avr_option_override): Remove set of
+       PARAM_ALLOW_STORE_DATA_RACES.
+       * common/config/avr/avr-common.c (avr_option_optimization_table)
+       [OPT_LEVELS_ALL]: Turn on -fallow-store-data-races.
+
+2019-10-17  H.J. Lu  <hongjiu.lu@intel.com>
+
+       * config/i386/i386.h (processor_costs): Add clear_ratio.
+       (CLEAR_RATIO): Remove MIN and use ix86_cost->clear_ratio.
+       * config/i386/x86-tune-costs.h: Set clear_ratio to the minimum
+       of 6 and move_ratio in all cost models.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       * tree-vect-loop.c (check_reduction_path): Compute reduction
+       operation here.
+       (vect_is_simple_reduction): Remove special-case of single-stmt
+       reduction path detection.
+
+2019-10-17  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm-cpus.in (marvel-pj4): Add +fp to the architecture.
+
+2019-10-17  Yuliang Wang  <yuliang.wang@arm.com>
+
+       * config/aarch64/aarch64-sve2.md (aarch64_sve2_eor3<mode>)
+       (aarch64_sve2_nor<mode>, aarch64_sve2_nand<mode>)
+       (aarch64_sve2_bsl<mode>, aarch64_sve2_nbsl<mode>)
+       (aarch64_sve2_bsl1n<mode>, aarch64_sve2_bsl2n<mode>):
+       New combine patterns.
+       * config/aarch64/iterators.md (BSL_DUP): New int iterator for the
+       above.
+       (bsl_1st, bsl_2nd, bsl_dup, bsl_mov): Attributes for the above.
+
+2019-10-17  Aldy Hernandez  <aldyh@redhat.com>
+
+       * tree-vrp.c (value_range_base::dump): Display +INF for both
+       pointers and integers when appropriate.
+
+2019-10-17  Andre Vieira  <andre.simoesdiasvieira@arm.com>
+
+       * tree-vect-loop.c (vect_analyze_loop_2): Use same condition to decide
+       when to use versioning threshold.
+
+2019-10-17  Andre Vieira  <andre.simoesdiasvieira@arm.com>
+
+       * tree-vect-loop.c (determine_peel_for_niter): New function contained
+       outlined code from ...
+       (vect_analyze_loop_2): ... here.
+
+2019-10-17  Andre Vieira  <andre.simoesdiasvieira@arm.com>
+
+       * tree-vect-loop.c (vect_transform_loop): Move code from here...
+       * tree-vect-loop-manip.c (vect_loop_versioning): ... to here.
+       * tree-vectorizer.h (vect_loop_versioning): Remove unused parameters.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       * tree-vect-loop.c (needs_fold_left_reduction_p): Export.
+       (vect_is_simple_reduction): Move all validity checks ...
+       (vectorizable_reduction): ... here.  Compute whether we
+       need a fold-left reduction here.
+       * tree-vect-patterns.c (vect_reassociating_reduction_p): Merge
+       both overloads, check needs_fold_left_reduction_p directly.
+       * tree-vectorizer.h (needs_fold_left_reduction_p): Declare.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       * tree-ssa-pre.c (create_component_ref_by_pieces_1): Fix
+       TARGET_MEM_REF creation.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       PR tree-optimization/92129
+       * tree-vect-loop.c (vectorizable_reduction): Also fail
+       on GIMPLE_SINGLE_RHS.
+
+2019-10-17  Jakub Jelinek  <jakub@redhat.com>
+
+       PR tree-optimization/92056
+       * tree-object-size.c (cond_expr_object_size): Return early if then_
+       processing resulted in unknown size.
+
+       PR tree-optimization/92115
+       * tree-ssa-ifcombine.c (ifcombine_ifandif): Force condition into
+       temporary if it could trap.
+
+2019-10-17  Richard Biener  <rguenther@suse.de>
+
+       PR debug/91887
+       * dwarf2out.c (gen_formal_parameter_die): Also try to match
+       context_die against a DW_TAG_GNU_formal_parameter_pack parent.
+
+2019-10-16  Jakub Jelinek  <jakub@redhat.com>
+
+       * tree-ssa-strlen.c (maybe_invalidate): Use
+       HOST_WIDE_INT_PRINT_UNSIGNED instead of "%zu".
+
+2019-10-16  Andrew Burgess  <andrew.burgess@embecosm.com>
+           Jim Wilson  <jimw@sifive.com>
+
+       * config/riscv/riscv.h (REG_CLASS_CONTENTS): Add argument passing
+       regs to SIBCALL_REGS.
+       * config/riscv/riscv.c (riscv_regno_to_class): Change argument
+       passing regs to SIBCALL_REGS.
+
+2019-10-16  Martin Sebor  <msebor@redhat.com>
+
+       PR tree-optimization/83821
+       * tree-ssa-strlen.c (maybe_invalidate): Add argument.  Consider
+       the length of a string when available.
+       (handle_builtin_memset) Add argument.
+       (handle_store, strlen_check_and_optimize_call): Same.
+       (check_and_optimize_stmt): Same.  Pass it to callees.
+
+2019-10-16  Martin Sebor  <msebor@redhat.com>
+
+       PR tree-optimization/91996
+       * tree-ssa-strlen.c (maybe_warn_pointless_strcmp): Improve location
+       information.
+       (compare_nonzero_chars): Add an overload.
+       (count_nonzero_bytes): Add an argument.  Call overload above.
+       Handle non-constant lengths in some range.
+       (handle_store): Add an argument.
+       (check_and_optimize_stmt): Pass an argument to handle_store.
+
+2019-10-16  Richard Earnshaw  <rearnsha@arm.com>
+
+       * config/arm/arm.c (neon_valid_immediate): Clear bytes before use.
+
+2019-10-16  Mihailo Stojanovic  <mistojanovic@wavecomp.com>
+
+       * config/mips/mips.c (mips_expand_builtin_insn): Force the
+       operands which correspond to the same input-output register to
+       have the same pseudo assigned to them.
+
 2019-10-16  Ilya Leoshkevich  <iii@linux.ibm.com>
 
        * cfgrtl.c (find_partition_fixes): Remove bbs_in_cold_partition.