Ping! Unreviewed patch (hot/cold partitioning update)
Caroline Tice
ctice@apple.com
Mon Aug 16 21:52:00 GMT 2004
This is the rest of the hot/cold partitioning update patch, again. I
have removed the tiny part that Mark
Mitchell already approved and the objectionable change to
config/pa/som.h that John Anglin pointed out
was incorrect. That leave 1189 lines of patch to be reviewed. Since
it's already been a month since my
original patch submission (this is the THIRD ping), I would really
appreciate somebody reviewing it before
it becomes hopelessly out of sync with the current source tree. Thank
you VERY much.
-- Caroline Tice
ctice@apple.com
From my original submission (on July 15):
This patch fixes what I consider to be the worst of the problems with
the hot/cold partitioning optimization. The
issues it does not address are: 1). making partitioning work correctly
with dwarf debugging turned on. That will
be the topic for the next hot/cold partitioning patch. What this patch
does is warns users that "-g" doesn't work
with hot/cold partitioning (if their debug format is DWARF), and
changes the partitioning option to just
basic block reordering. 2). Some efficiency concerns. Those can also
be addressed in a later patch. I really
wanted to focus this patch on what I consider to be correctness issues.
The problems that this patch does attempt to address/fix are:
- Fix places where adding in_unlikely_executed_text to the enum data
type "in_section" threw off switch
case statements.
- Make it work correctly (by turning it off) for functions where user
specifies "__attribute__ section"
- Make it work correctly (by turning it off) for linkonce sections
- Make it work correctly with -ffunction-sections flag
- Make it output correct cold section labels
- Make it set section alignments at start of function rather than on
each section switch
- Undo some changes to original assembly code generation
- Turn off hot/cold partitioning in the presence of DWARF debugging
(for the moment)
- Turn off hot/cold partitioning for architectures that do not support
named sections
- Use variables rather than constants for cold section labels and names
(to work correctly with
-ffunction-sections, among other things)
I have tested this patch on Apple G4 & G5 running apple-darwin, and on
an i386 running Linux. The tests
include doing a profiledbootstrap with hot/cold partitioning turned on
(and debugging turned off on the i386);
running DejaGnu tests; and running all the SPECInt tests (with
partitioning) that the mainline compiler can pass
(without partitioning).
I would appreciate it if people with experience with the following
config files could make sure I didn't do
the wrong thing in there by accident:
gcc/config/arm/pe.h
gcc/config/i386/cygming.h
gcc/config/mcore/mcore.h
Is this patch okay to commit to gcc 3.5 mainline?
Updated ChangeLog Entry:
2004-08-16 Caroline Tice <ctice@apple.com>
* Makefile.in (STAGEFEEDBACK_FLAGS_TO_PASS) Add
"-freorder-blocks-and-partition" to the flags used in second
stage of profiledbootstrap.
* bb-reorder.c (push_to_next_round_p): Add new variable,
next_round_is_last; set and use variable to make sure, when
partitioning, that the last trace construction round consists
of all (and only) cold basic blocks.
(rotate_loop): Don't copy blocks that end in a section
crossing jump.
(copy_bb): Correctly initialize "partition" of duplicated bb.
(add_unlikely_executed_notes): Add a comment.
(find_rarely_executed_basic_blocks_and_crossing_edges): Modify
to make sure, if function contains hot blocks, that the
successors of ENTRY_BLOCK_PTR are hot; also, only look for
crossing edges if the architecture supports named sections.
(mark_bb_for_unlikely_executed_section): Modify to always
insert the NOTE_INSN_UNLIKELY_EXECUTED_CODE immediately after
the basic block note insn.
(fix_crossing_unconditional_branches): Remove extra space.
(fix_edges_for_rarely_executed_code): Modify to only do
partitioning work if the architecture supports named sections.
(reorder_basic_blocks): Modify to only add
NOTE_INSN_UNLIKELY_EXECUTED_CODE notes if the architecture
supports named sections.
* c-common.c (handle_section_attribute): Initialize new global
variable, user_defined_section_attribute, to true if user has
specified one.
* cfgcleanup.c (try_forward_edges): Modify to not attempt to
forward edges that cross section boundaries.
* cfglayout.c (fixup_reorder_chain): Modify to only fix up
partitioning information if the architecture supports named
sections.
* cfgrtl.c (target.h): Add statement to include this.
(rtl_split_block): Make sure newly created bb gets correct
partition.
(try_redirect_by_replacing_jump): Make sure redirection isn't
attempting to cross section boundaries.
(force_nonfallthru_and_redirect): Only do partition fix up if
architecture supports named sections.
(rtl_split_edge): Make sure newly created bb ends up in
correct partition.
(commit_one_edge_insertion): Remove code that incorrectly
updated basic block partition; Make sure partition fix up only
happens if architecture supports named sections and it's not
already done.
(rtl_verify_flow_info_1): Fix if-condition on test/error
condition that fallthru edges are not allowed to cross section
boundaries.
* defaults.h (NORMAL_TEXT_SECTION_NAME): Remove this.
* final.c (final_scan_insn): Remove redundant test from
if-statement;
change calls to text_section into calls to function_section;
add code
to only to partitioning fix up if architecture supports named
sections.
* ifcvt.c (find_if_case_1): Make sure newly created bb has
correct
partition.
(if_convert): Add targetm.have_named_sections to test.
* output.h (unlikely_section_label): Extern declaration for new
global
variable.
(unlikely_text_section_name): Likewise.
* opts.c (decode_options): If both partitioning and DWARF
debugging
are turned on, issue a warning that this doesn't work, and
change
partitiong to basic block reordering (without hot/cold
partitions).
* passes.c (rest_of_handle_final): Re-set new global variable,
user_defined_section_attribute, to false.
(rest_of_compilation): Change options for calling partitioning
function: Don't call if the user defined the section attribute,
and
don't call if DECL_ONE_ONLY is true for the current function.
* predict.c (choose_function_section): Return immediately if we
are doing hot/cold partitioning (i.e. let the basic block
partitioning
determine where the function belongs).
* reg-stack.c (emit_swap_insn): Add condition to step over
NOTE_INSN_UNLIKELY_EXECUTED_CODE notes.
* toplev.c (user_defined_section_attribute): New global
variable.
* toplev.h (user_defined_section_attribute): Extern declaration
for new
global variable.
* varasm.c (unlikely_section_label): New global variable.
(unlikely_text_section_name): New global variable.
(unlikely_text_section): Add code to initialize
unlikely_text_section_name
if necessary; modify to use unlikely_text_section_name and
unlikely_section_label; also to use named_section properly.
(in_unlikely_text_section): Modify to work correctly with
named_section
and to use unlikely_text_section_name.
(named_section): Add code to work properly with cold section.
(function_section): Clean up if-statement.
* config/darwin.c (darwin_asm_named_section): Return to
original code,
removing use of SECTION_FORMAT_STRING.
* config/arm/pe.h (switch_to_section): Add case for
in_unlikely_executed_text to switch statement.
* config/i386/cygming.h (switch_to_section): Likewise.
* config/i386/darwin.h (NORMAL_TEXT_SECTION_NAME): Remove.
(SECTION_FORMAT_STRING): Likewise.
* config/mcore/mcore.h (switch_to_section): Likewise.
* config/rs6000/darwin.h (NORMAL_TEXT_SECTION_NAME): Remove.
(SECTION_FORMAT_STRING): Remove.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: gcc5-hot-cold5f.txt
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20040816/1dd971f9/attachment.txt>
-------------- next part --------------
More information about the Gcc-patches
mailing list