* ifcvt.c (noce_get_alt_condition): If the condition is a compare
against a constant, try to adjust the compare to have the desired
constant in it so that min/max optimizations happen more often.
From-SVN: r44435
* basic-block.h (redirect_edge_and_branch_force,
redirect_edge_and_branch, block_label, forwarder_block_p): Declare.
* flow.c (redirect_edge_and_branch_force,
redirect_edge_and_branch, block_label, forwarder_block_p): Make global.
(redirect_edge_and_branch_force): Fix copying of lifeness information.
(block_label): Handle EXIT_BLOCK_PTR by returning NULL.
* ifcvt.c (dead_or_predictable): Take BB as an new destionation
instead of label; update CFG after transformation.
(find_if_case_1): Update call, use redirect_edge_and_branch_force
for finishing the transformation; handle even case where ELSE
does not follow THEN.
(find_if_case_2): Update call of dead_or_predictable; simplify
CFG update.
* emit-rtl.c (split_branch_probability): New global variable.
(try_split): Take care to set split_branch_probability and
create REG_BR_PROB note for new jump insns.
* md.texi (define_split): Document new feature.
* i386.c (ix86_split_fp_branch): Redistribute branch probability notes.
From-SVN: r44249
* predict.def: New file.
* predict.h: New file.
* predict.c: Include predict.h.
(predictor_info): New structure and array.
(predict_edge, predict_insn): Rewrite; make global.
(predict_edge_def, predict_insn_def): New global functions.
(dump_prediction, combine_predictions_for_insns): New
static functions.
(estimate_probability): Change calls to predict_edge
to predict_edge_def and calls to predict_insn to predict_insn_def;
combine probabilities.
(expected_value_to_br_prob): Behave as predictor.
* rtl.h (REG_BR_PRED): Update coment.
(invert_br_probabilities): New.
* jump.c: Include predict.h
(duplicate_loop_exit_test): Drop branch prediction notes.
(invert_jump): Use invert_br_probabilities.
* Makefile.in: Add dependancy on predict.h for jump and predict pass.
* ifcvt.c (dead_or_predicable): Use invert_br_probabilities.
* combine.c (distribute_notes): Handle BR_PRED.
From-SVN: r43115
* ifcvt.c (merge_if_block): Use any_uncondjump_p, not simplejump_p
to determine if a block has an unconditional jump at its end.
* jump.c (jump_optimize_1): When converting a conditional jump
to an unconditional jump, build the jump using gen_jump instead
of just modify operands in-place.
From-SVN: r41939
2001-04-03 Jakub Jelinek <jakub@redhat.com>
David S. Miller <davem@pierdol.cobaltmicro.com>
Andrew MacLeod <amacleod@redhat.com>
Use byte offsets in SUBREGs instead of words.
* alias.c (nonlocal_mentioned_p): Use subreg_regno function.
* caller-save.c (mark_set_regs): Change callers of subreg_hard_regno
to pass new argument.
(add_stored_regs): Use subreg_regno_offset function.
* calls.c (expand_call): For non-paradoxical SUBREG take endianess
into account.
(precompute_arguments): Use gen_lowpart_SUBREG.
* combine.c (try_combine): Replace explicit XEXP with SUBREG_REG.
(combine_simplify_rtx): Rework to use SUBREG_BYTE.
(simplify_set): Rework to use SUBREG_BYTE.
(expand_field_assignment): Use SUBREG_BYTE.
(make_extraction): Use SUBREG_BYTE.
(if_then_else_cond): Use SUBREG_BYTE.
(apply_distributive_law): Use SUBREG_BYTE and fixup subreg comments.
(gen_lowpart_for_combine): Compute full byte offset.
* cse.c (mention_regs): Use SUBREG_BYTE.
(remove_invalid_subreg_refs): Rework to use SUBREG_BYTE.
(canon_hash): Use SUBREG_BYTE.
(fold_rtx): Pass SUBREG_BYTE div UNITS_PER_WORD to operand_subword.
(gen_lowpart_if_possible): Formatting.
* dbxout.c (dbxout_symbol_location): Compute SUBREG hard regnos
correctly.
* dwarf2out.c (is_pseudo_reg): Fixup explicit XEXP into SUBREG_REG
(mem_loc_descriptor): Fixup explicit XEXP into SUBREG_REG
(loc_descriptor): Fixup explicit XEXP into SUBREG_REG
* dwarfout.c (is_pseudo_reg): Fixup explicit XEXP into SUBREG_REG
(output_mem_loc_descriptor): Fixup explicit XEXP into SUBREG_REG
(output_loc_descriptor): Fixup explicit XEXP into SUBREG_REG
* emit-rtl.c (gen_rtx_SUBREG): New function, used to verify
certain invariants about SUBREGs the compiler creates.
(gen_lowpart_SUBREG): New function.
(subreg_hard_regno): New function to get the final register number.
(gen_lowpart_common): Use SUBREG_BYTE.
(gen_imagpart): Spacing nits.
(subreg_realpart_p): Use SUBREG_BYTE.
(gen_highpart): Use SUBREG_BYTE.
(subreg_lowpart_p): Always compute endian corrected goal offset,
even at the byte level, then compare against that.
(constant_subword): New function, pulled out all constant cases
from operand_subword and changed second argument name to offset.
(operand_subword): Detect non REG/SUBREG/CONCAT/MEM cases early
and call constant_subword to do the work. Return const0_rtx if
looking for a word outside of OP.
(operand_subword_force): Change second arg name to offset.
* expmed.c (store_bit_field): Use SUBREG_BYTE.
(store_split_bit_field): Use SUBREG_BYTE.
(extract_bit_field): Use SUBREG_BYTE.
(extract_split_bit_field): Use SUBREG_BYTE.
(expand_shift): Use SUBREG_BYTE.
* expr.c (store_expr, expand_expr): Use gen_lowpart_SUBREG.
* final.c (alter_subreg) Use subreg_hard_regno and SUBREG_BYTE.
* flow.c (set_noop_p): Use SUBREG_BYTE.
(mark_set_1): Remove ALTER_HARD_SUBREG. Use subreg_regno_offset instead.
* function.c (fixup_var_refs_1): Fixup explicit XEXP into a SUBREG_REG.
(fixup_memory_subreg): Use SUBREG_BYTE and remove byte endian
correction code.
(optimize_bit_field): Use SUBREG_BYTE.
(purge_addressof_1): Use SUBREG_BYTE.
(purge_single_hard_subreg_set): Use subreg_regno_offset function.
(assign_params): Mark arguments SUBREG_PROMOTED_VAR_P if they are
actually promoted by the caller and PROMOTE_FOR_CALLS_ONLY is true.
* gengenrtl.c (special_rtx): Add SUBREG.
* global.c (mark_reg_store): Use SUBREG_BYTE.
(set_preference): Rework to use subreg_regno_offset and SUBREG_BYTE.
* ifcvt (noce_emit_move_insn): Use SUBREG_BYTE.
* integrate.c (copy_rtx_and_substitute): Use SUBREG_BYTE and make sure
final byte offset is congruent to subreg's mode size.
(subst_constants): Use SUBREG_BYTE.
(mark_stores): Use subreg_regno_offset function.
* jump.c (rtx_renumbered_equal_p, true_regnum): Use subreg_regno_offset
function and SUBREG_BYTE.
* local-alloc.c (combine_regs): Use subreg_regno_offset function.
(reg_is_born): Use subreg_hard_regno.
* recog.c (valid_replace_rtx_1): Use SUBREG_BYTE and remove byte
endian correction code. Don't combine subregs unless resulting
offset aligns with type. Fix subreg constant extraction for DImode.
Simplify SUBREG of VOIDmode CONST_DOUBLE.
(general_operand): Remove dead mode_altering_drug code.
(indirect_operand): Use SUBREG_BYTE.
(constrain_operands): Use subreg_regno_offset function.
* reg-stack.c (get_true_reg): Use subreg_regno_offset function.
* regmove.c (regmove_optimize): Use SUBREG_BYTE.
(optimize_reg_copy_3): Use gen_lowpart_SUBREG.
* regs.h (REG_SIZE): Allow target to override.
(REGMODE_NATURAL_SIZE): New macro which target can override.
* reload.c (reload_inner_reg_of_subreg): subreg_regno should be used
on the entire subreg rtx.
(push_reload): Use SUBREG_BYTE in comments and code.
(find_dummy_reload): Use subreg_regno_offset. Only adjust offsets
for hard registers inside subregs.
(operands_match_p): Use subreg_regno_offset.
(find_reloads): Use SUBREG_BYTE and only advance offset for subregs
containing hard regs.
(find_reload_toplev): Use SUBREG_BYTE. Remove byte endian
corrections when fixing up MEM subregs.
(find_reloads_address_1): Use SUBREG_BYTE, subreg_regno, and
subreg_regno_offset where appropriate.
(find_reloads_subreg_address): Use SUBREG_BYTE. Remove
byte endian corrections when fixing up MEM subregs.
(subst_reloads): When combining two subregs, make sure final
offset is congruent to subreg's mode size.
(find_replacement): Use SUBREG_BYTE and subreg_regno_offset.
(refers_to_regno_for_reload_p): Use subreg_regno.
(reg_overlap_mentioned_for_reload_p): Use subreg_regno_offset.
* reload1.c (eliminate_regs) Use SUBREG_BYTE. Remove byte endian
correction code for memory subreg fixups.
(forget_old_reload_1): Use subreg_regno_offset.
(choose_reload_regs): Use subreg_regno.
(emit_input_reload_insns): Use SUBREG_BYTE.
(reload_combine_note_store): Use subreg_regno_offset.
(move2add_note_store): Use subreg_regno_offset.
* resource.c (update_live_status, mark_referenced_resources): Use
subreg_regno function.
(mark_set_resources): Use subreg_regno function.
* rtl.h (SUBREG_WORD): Rename to SUBREG_BYTE.
(subreg_regno_offset, subreg_regno): Define prototypes.
(subreg_hard_regno, constant_subword, gen_rtx_SUBREG): Newi functions.
(gen_lowpart_SUBREG): Add prototype.
* rtl.texi (subreg): Update to reflect new byte offset representation.
Add mentioning of the effect that BYTES_BIG_ENDIAN has on subregs now.
* rtlanal.c (refers_to_regno_p): Use subreg_regno.
(reg_overlap_mentioned_p): Use subreg_regno.
(replace_regs); Make sure final offset of combined subreg is
congruent to size of subreg's mode.
(subreg_regno_offset): New function.
(subreg_regno): New function.
* sched-vis.c (print_value): Change SUBREG_WORD to SUBREG_BYTE.
* sdbout.c (sdbout_symbol): Compute offset using alter_subreg.
* stmt.c (expand_anon_union_decl): Use gen_lowpart_SUBREG.
* tm.texi (ALTER_HARD_SUBREG): Remove, it is now dead.
(SUBREG_REGNO_OFFSET): Describe SUBREG_REGNO_OFFSET overrides.
* config/a29k/a29k.c (gpc_reg_operand): Use subreg_regno.
(a29k_get_reloaded_address): Use SUBREG_BYTE.
(print_operand): Use SUBREG_BYTE.
* config/alpha/alpha.c (print_operand_address): Use SUBREG_BYTE.
* config/arm/arm.c (arm_reload_in_hi): Use SUBREG_BYTE.
(arm_reload_out_hi): Use SUBREG_BYTE.
* config/d30v/d30v.c (d30v_split_double): Use subreg_regno_offset
instead of SUBREG_WORD.
(d30v_print_operand_memory_reference): Use subreg_regno_offset.
* config/dsp16xx/dsp16xx.md (extendqihi2, zero_extendqihi2): Fix
SUBREG creation to use byte offset.
* config/h8300/h8300.md (Unnamed HImode zero extraction and 16bit
inverted load insns): Fix explicit rtl subregs to use byte
offsets.
* config/i370/i370.md (cmpstrsi, movstrsi, mulsi3, divsi3,
udivsi3, umodsi3): Generate SUBREGs with byte offsets.
* config/i860/i860.c (single_insn_src_p): Use SUBREG_BYTE.
* config/i860/i860.md (mulsi3_big): Fixup explicit SUBREGs in rtl
to use byte offsets.
(unnamed fmlow.dd insn): Fixup SUBREGS to use byte offsets.
* config/i960/i960.md (extendhisi2): Generate SUBREGs with byte
offsets, also make sure it is congruent to SUBREG's mode size.
(extendqisi2, extendqihi2, zero_extendhisi2, zero_extendqisi2,
unnamed ldob insn): Generate SUBREGs with byte offset.
(zero_extendqihi2): SUBREG's are byte offsets.
* config/m68hc11/m68hc11.c (m68hc11_gen_lowpart): Use SUBREG_BYTE.
(m68hc11_gen_highpart): Use SUBREG_BYTE.
* config/m68k/m68k.md (zero_extendhisi2, zero_extendqihi2,
zero-extendqisi2): Generate SUBREGs with byte offset.
(umulsidi3, mulsidi3, subreghi1ashrdi_const32,
subregsi1ashrdi_const32, subreg1lshrdi_const32): Fixup explicit
subregs in rtl to use byte offsets.
* config/m88k/m88k.md (extendsidi2): fixup subregs to use byte offset.
* config/mips/mips.c (mips_move_1word): Use subreg_regno_offset.
(mips_move_2words): Use subreg_regno_offset.
(mips_secondary_reload_class): Use subreg_regno_offset.
* config/mips/mips.md (DImode plus, minus, move, and logical op
splits): Fixup explicit subregs in rtl to use byte offsets.
* config/mn10200/mn10200.c (print_operand): Use subreg_regno function.
* config/mn10300/mn10300.c (print_operand): Use subreg_regno function.
* config/ns32k/ns32k.md (udivmoddisi4): Fix explicit subregs in
rtl to use byte offsets.
* config/pa/pa.c (emit_move_sequence): Use SUBREG_BYTE.
* config/pa/pa.md (floatunssisf2, floatunssidf2, mulsi3): fix explicit
subregs to use byte offsets.
* config/pdp11/pdp11.md (zero_extendhisi2, modhi3, modhi3+1):
Fixup explicit subregs in rtl to use byte offsets.
* config/romp/romp.c (memory_offset_in_range_p): Use SUBREG_BYTE
and remove byte endian correction code.
* config/sh/sh.c (output_movedouble): Use subreg_regno.
(gen_ashift_hi): Use SUBREG_BYTE.
(regs_used): Use subreg_regno_offset.
(machine_dependent_reorg): Use subreg_regno_offset.
* config/sh/sh.h (INDEX_REGISTER_RTX_P): Use SUBREG_BYTE.
* config/sh/sh.md (DImode and DFmode move splits): Use subreg_regno.
(movdf_i4): Subregs are byte offsets now.
* config/sparc/sparc.c (ultra_find_type): Use SUBREG_BYTE.
* config/sparc/sparc.h (ALTER_HARD_SUBREG): Removed.
(REGMODE_NATURAL_SIZE): Override.
(REG_SIZE): For SUBREG check float mode on SUBREG_REG's mode.
* config/sparc/sparc.md (TFmode move splits): Generate SUBREGs
with byte offsets.
(zero_extendhisi2, zero_extendqidi2_insn, extendhisi2,
extendqihi2, sign_extendqihi2_insn, sign_extendqisi2_insn,
extendqidi2): Generate SUBREGs with byte offsets, also make sure
it is congruent to SUBREG's mode size.
(smulsi3_highpart_v8plus): Fix explicit subregs in rtl to use byte
offsets.
(cmp_siqi_trunc, cmp_siqi_trunc_set, cmp_diqi_trunc,
cmp_diqi_trunc_set, lshrdi3_v8plus+1, lshrdi3_v8plus+2,
lshrdi3_v8plus+3, lshrdi3_v8plus+4): Use proper
SUBREG_BYTE offset for non-paradoxical subregs in patterns.
* config/v850/v850.c (print_operand, output_move_double): Use
subreg_regno function.
Co-Authored-By: Andrew MacLeod <amacleod@redhat.com>
Co-Authored-By: David S. Miller <davem@pierdol.cobaltmicro.com>
From-SVN: r41058
* ifcvt.c (noce_emit_move_insn): New.
(noce_try_store_flag, noce_try_store_flag_constants,
noce_try_store_flag_inc, noce_try_store_flag_mask,
noce_try_cmove, noce_try_cmove_arith, noce_try_minmax,
noce_try_abs): Use it.
(noce_process_if_block): Likewise.
For STRICT_LOW_PART, take mode from its SUBREG.
* gcc.c-torture/compile/20010329-1.c: New test.
From-SVN: r41001
While building global lifetime information, keep track of which registers are
only conditionally set. Use this to force rescans in more cases where they
are needed.
From-SVN: r37801
* ifcvt.c (merge_if_block): Be prepared for JOIN to have no
remaining edges.
(find_if_block): Allow THEN with no outgoing edges.
* flow.c (merge_blocks_nomove): Remove a barrier not following
a jump as well.
From-SVN: r34317
* ifcvt.c (cond_exec_process_if_block): Do nothing if the last insn
in the block is more than just a simple conditional branch.
(noce_process_if_block): Similarly.
From-SVN: r34287
* ifcvt.c (seq_contains_jump): New.
(noce_try_store_flag_constants): Use it to fail conversion.
(noce_try_store_flag_inc, noce_try_store_flag_mask): Likewise.
From-SVN: r34179
* ifcvt.c (cond_exec_process_insns): Drop USE and CLOBBER insns
that get in the way after reload.
(cond_exec_process_if_block): Skip a label heading THEN block.
From-SVN: r33923
* Makefile.in (final.o): Depend on BASIC_BLOCK_H.
* final.c (final_end_function): Use app_disable. Rearrange note
handling into a switch. Emit deleted labels.
(output_asm_label): Generate label strings for deleted labels.
* flow.c (tail_recursion_label_list): New.
(find_basic_blocks_1): Set label_value_list directly. Collect list
of tail recursion labels from call_placeholders. Don't add deleted
labels to the label value list.
(cleanup_cfg): Use free_EXPR_LIST_list.
(flow_delete_insn_chain): Turn non-removable labels into notes.
(flow_delete_block): Don't disable deleting the block because of
a non-removable label.
(tail_recursion_label_p): New.
(merge_blocks_move_predecessor_nojumps): Don't disable the merge
because of a label.
(merge_blocks_move_successor_nojumps): Likewise. Also move a
jump table.
(merge_blocks): Disable a merge because of tail recursion labels.
* ifcvt.c (merge_if_block): Don't disable a merge because of a label.
Use a more accurate measure of not merging the join block.
(find_if_block): Don't disable conversion because of a label.
(find_if_case_1, find_if_case_2): Likewise.
* jump.c (duplicate_loop_exit_test): Preserve the kind of list
element when copying.
(squeeze_notes): Also leave EH notes.
(mark_jump_label): Ignore deleted labels. Use an INSN_LIST for
REG_LABEL notes.
(delete_insn): Preserve LABEL_NAME in NOTE_SOURCE_FILE when
deleting a label.
* print-rtl.c (print_rtx): Print NOTE_SOURCE_FILE for
NOTE_INSN_DELETED_LABEL. Print `[# deleted]' for a label_ref
referring to a deleted label. Convert tail handling to a switch.
* rtl.def (CODE_LABEL): Rearrange elements to be compatible with NOTE
for NOTE_INSN_DELETED_LABEL.
(NOTE): Fix commentary.
* rtl.h (REG_LABEL): Update commentary wrt INSN_LIST.
(REG_CC_SETTER, REG_CC_USER, REG_LIBCALL): Likewise.
(CODE_LABEL_NUMBER, LABEL_NAME): Update index.
(LABEL_NUSES, LABEL_REFS): Likewise.
* unroll.c (copy_loop_body): Don't copy NOTE_INSN_DELETED_LABEL.
From-SVN: r33876
* ifcvt.c (cond_exec_process_insns): New argument prob_val.
Attach it to call insns.
(cond_exec_process_if_block): Track probability for true and
false branches.
(dead_or_predicable): Likewise.
From-SVN: r33791
* ifcvt.c (noce_process_if_block): Fail the conversion if X is
referenced bewteen the condition and the jump. Don't delete
anything but the jump.
From-SVN: r33563
* ifcvt.c: New file.
* Makefile.in (OBJS): Add it.
(ifcvt.o): New target.
* jump.c (jump_optimize_1): Remove all code related to if-conversion,
and conditional arithmetic.
(find_insert_position): Remove.
* timevar.def (TV_IFCVT, TV_IFCVT2): New.
* toplev.c (DFI_ce, DFI_ce2): New.
(dump_file): Add ce and ce2 dumps.
(rest_of_compilation): Run if_convert a couple o times. Set
cse_not_expected after cse2. Don't set no_new_pseudos until
after sched1 or recompute_reg_usage.
From-SVN: r33547
