[PATCH]: Remove insn scanning from hot/cold partitioning...

[Caroline Tice <ctice at apple dot com>]

This patch should address all the remaining issues with hot/cold 
partitioning EXCEPT for making
it work with dwarf debug format (that will be forthcoming soon).  I 
believe this patch is appropriate for
stage 3, as it is a bug/feature fixing patch, but if I am mistaken, I 
apologize.

This patch includes all of the following unreviewed patch  (from Aug. 
31), which is required to make
everything work:

http://gcc.gnu.org/ml/gcc-patches/2004-08/msg02673.html
"Change function_section to allow partitioning w/o insn stream scanning"

The patch I am currently submitting fixes or addresses the following 
concerns with respect to hot/cold
basic block partitioning.  If anybody believes there is an issue or 
concern that needs to be addressed
which I have overlooked, please let me know:

 - Make sure there is at most one cluster of hot blocks and one cluster 
of cold blocks within a function,
    with a single boundary between them.
- Use a single NOTE_INSN_UNLIKELY_EXECUTED_CODE note to mark the 
beginning of the cold
   section (rather than attaching the note to each cold block).
- Remove some redundant test conditions within the cfg optimizations 
for determining when to disallow the
  optimizations.
- Remove code for scanning instruction stream to determine which 
section code belongs in (rely on
  single boundary instead).
- Make sure cold section label prints out in correct place.

[ From previous unreviewed patch]:
  - Split function_section into two functions, and change calls to 
function_section where appropriate.


I have tested this code on both a G4 running powerpc-apple-darwin, and 
an x86 running Linux.  It has
passed a profiledbootstrap (with partitioning turned on), DejaGnu 
tests, SPEC benchmark tests (with
partitioning turned on), and my own small partitioning test on both 
architectures.

Is this okay to commit to FSF mainline?

-- Caroline Tice
ctice@apple.com

2004-09-16  Caroline Tice  <ctice@apple.com>

        * Makefile.in (varasm.o): Add basic-block.h to list of 
requirements.
        (bb-reorder.o): Add errors.h to list of requirements.
        * bb-reorder.c (errors.h): Add new include.
        (struct bbro_basic_block_data_def): Add new field, in_trace.
        (add_unlikely_executed_notes):  Remove function.
        (mark_bb_for_unlikely_executed_section): Remove function.
        (add_remaining_hot_blocks_to_heap): New function.
        (insert_cold_section_boundary_note): New function.
        (verify_hot_cold_block_grouping): New function.
        (find_traces): Add and initialize local variable, 
last_hot_round;
        add call to add_remaining_hot_blocks_to_heap (when appropriate).
        (find_traces_1_round) : Add code to update new struct field;
        correct trace_length where it was incorrect before (after call 
to
        copy_bb).
        (bb_to_key):  Add code to force first block in function to be 
first
        block in traces, even when the block is cold.
        (fix_edges_for_rarely_executed_code):  Update comment describing
        function.
        (reorder_basic_blocks):  Add code to initialize new field 
(in_trace);
        remove call to add_unlikely_executed_notes; add call to
        verify_hot_cold_block_grouping.
        (partition_hot_cold_basic_blocks):  Update function comment.
        * cfgcleanup.c (try_simplify_condjump): Remove redundacy from 
condition.
        (try_forward_edges): Likewise.
        (merge_blocks_move_predecessor_nojumps): Likewise.
        (merge_blocks_move_successor_nojumps): Likewise.
        (merge_blocks_move): Likewise.
        (try_crossjump_bb): Likewise.
        * cfglayout.c (update_unlikely_executed_notes): Remove function.
        (fixup_reorder_chain): Remove code for adding 
UNLIKELY_EXECUTED_CODE
        notes to cold bb's; remove call to 
update_unlikely_executed_notes.
        * cfglayout.h (scan_ahead_for_unlikely_executed_note): Remove
        function declaration.
        * cfgrtl.c (can_delete_note_p): Remove UNLIKELY_EXECUTED_CODE 
note
        from consideration.
        (rtl_can_merge_blocks): Remove redundancy from condition.
        (try_redirect_by_replacing_jump): Likewise.
        (force_nonfallthru_and_redirect): Remove code for adding
        UNLIKELY_EXECUTED_CODE notes to cold bb's.
        (commit_one_edge_insertion): Likewise.
        (cfg_layout_can_merge_blocks_p): Remove redundancy from 
condition.
        * dbxout.c (FORCE_TEXT): Replace function_section with
        current_function_section.
        * except.c (output_function_esception_table): Likewise.
        * final.c (profile_function): Likewise.
        (scan_ahead_for_unlikely_executed_note): Remove function.
        (final_scan_insn):  Remove calls to 
scan_ahead_for_unlikely_executed_note,
        and related code for switching to cold section, except for the
        single time NOTE_INSN_UNLIKELY_EXECUTED_CODE may be encountered;
        replace appropriate calls to function_section with calls to
        current_function_section.
        * ifcvt.c (find_if_case_1): Remove redundancy from condition.
        (find_if_case_2): Likewise.
        * opts.c (decode_options): Change warnings about hot/cold 
partitioning
        options to informs (so as not to cause bootstrap failures).
        * output.h (current_function_section): Add new function decl.
        (insert_cold_section_boundary_note): Likewise.
        * reg-stack.c (emit_swap_insn): Remove UNLIKELY_EXECUTED_CODE 
note insn
        from consideration.
        * varasm.c (basic-block.h): Add new include.
        (initialize_cold_section_name): New function.
        (last_text_section): New global variable.
        (last_text_section_name): New global variable.
        (text_section): Update last_text_section.
        (unlikely_text_section): Replace code to determine cold section 
name with
        call to initialize_cold_section_name; Add code to update
        last_text_section.
        (named_section_real): Update last_text_section and
        last_text_section_name, if appropriate.
        (function_sectin): Remove code that scans ahead for 
UNLIKELY_EXECUTED_CODE
        notes.
        (current_function_section): New function.
        (assemble_start_function): Don't print out cold section name 
while
        aligning it; set DECL_SECTION_NAME to cold section name if 
entire
        function is cold; initialize last_text_section; add call to
        insert_cold_section_boundary_note.
        * config/x390/s390.c (s390_select_rtx_section): Replace 
function_section
        with current_function_section.
        * config/sparc/sparc.c (sparc_output_deferred_case_vectors): 
Likewise.
        * config/stormy16/stormy16.c (stormy_16_output_addr_vec): 
Likewise.
        * config/xtensa/xtensa.c (override_options): Turn off hot/cold
        partitioning for this architecture.

Attachment: gcc5-hot-cold4.txt
Description: Text document