This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH 000/236] Introduce rtx subclasses
- From: David Malcolm <dmalcolm at redhat dot com>
- To: gcc-patches at gcc dot gnu dot org
- Cc: David Malcolm <dmalcolm at redhat dot com>
- Date: Wed, 6 Aug 2014 13:19:39 -0400
- Subject: [PATCH 000/236] Introduce rtx subclasses
- Authentication-results: sourceware.org; auth=none
This is the patch series I spoke about at Cauldron in the talk
"A proposal for typesafe RTL"; slides here:
http://dmalcolm.fedorapeople.org/presentations/cauldron-2014/rtl
They can also be seen at:
https://dmalcolm.fedorapeople.org/gcc/patch-backups/rtx-classes/v20/
The aim of the patch series is to improve the type-safety and
readability of the backend by introducing subclasses of rtx (actually
rtx_def) for *instructions*, and also for EXPR_LIST, INSN_LIST, SEQUENCE.
That way we can document directly in the code the various places that
manipulate insn chains vs other kinds of rtx node.
An example of a bug detected using this approach: in mn10300.c there
was dead code of the form:
if (GET_CODE (insn) == PARALLEL)
insn = XVECEXP (insn, 0, 0);
where the test should really have been on "PATTERN (insn)", not "insn":
if (GET_CODE (PATTERN (insn)) == PARALLEL)
insn = XVECEXP (PATTERN (insn), 0, 0);
[as discussed in https://gcc.gnu.org/ml/gcc/2014-07/msg00078.html]
The class hierarchy looks like this (using indentation to show
inheritance, and indicating the invariants):
class rtx_def;
class rtx_expr_list; /* GET_CODE (X) == EXPR_LIST */
class rtx_insn_list; /* GET_CODE (X) == INSN_LIST */
class rtx_sequence; /* GET_CODE (X) == SEQUENCE */
class rtx_insn; /* INSN_CHAIN_CODE_P (GET_CODE (X)) */
class rtx_real_insn; /* INSN_P (X) */
class rtx_debug_insn; /* DEBUG_INSN_P (X) */
class rtx_nonjump_insn; /* NONJUMP_INSN_P (X) */
class rtx_jump_insn; /* JUMP_P (X) */
class rtx_call_insn; /* CALL_P (X) */
class rtx_jump_table_data; /* JUMP_TABLE_DATA_P (X) */
class rtx_barrier; /* BARRIER_P (X) */
class rtx_code_label; /* LABEL_P (X) */
class rtx_note; /* NOTE_P (X) */
The patch series converts roughly 4300 places in the code from using
rtx to the more concrete rtx_insn *, in such places as:
* the core types within basic blocks
* hundreds of function params, struct fields, etc. e.g. within
register allocators, schedulers
* "insn" and "curr_insn" within .md files (peephole, attributes,
define_bypass guards)
* insn_t in sel-sched-ir.h
* Target hooks: updated params of 25 of them
* Debug hooks: "label" and "var_location"
etc
The patch series also contains some cleanups using inline methods:
* being able to get rid of this boilerplate everywhere that jump tables
are handled:
if (GET_CODE (PATTERN (table)) == ADDR_VEC)
vec = XVEC (PATTERN (table), 0);
else
vec = XVEC (PATTERN (table), 1);
in favor of a helper method (inlined):
vec = table->get_labels ();
* having a subclass for EXPR_LIST allows for replacing this kind of
thing:
for (x = forced_labels; x; x = XEXP (x, 1))
if (XEXP (x, 0))
set_label_offsets (XEXP (x, 0), NULL_RTX, 1);
with the following, which captures that it's an EXPR_LIST, and makes
it clearer that we're simply walking a singly-linked list:
for (rtx_expr_list *x = forced_labels; x; x = x->next ())
if (x->element ())
set_label_offsets (x->element (), NULL_RTX, 1);
There are some surface details to the patches:
* class names. The subclass names correspond to the lower_case name
from rtl.def, with an "rtx_" prefix. "rtx_insn" and "rtx_real_insn"
don't correspond to concrete node kinds, and hence I had to invent
the names. (In an earlier version of the patches these were
"rtx_base_insn" and "rtx_insn" respectively, but the former occurred
much more than the latter and so it seemed better to use the shorter
spelling for the common case).
* there's a NULL_RTX define in rtl.h. In an earlier version of the
patch I added a NULL_INSN define, but in this version I simply use
NULL, and I'm in two minds about whether a NULL_INSN is desirable
(would we have a NULL_FOO for all of the subclasses?). I like having
a strong distinction between arbitrary RTL nodes vs instructions,
so maybe there's a case for NULL_INSN, but not for the subclasses?
* I added an "rtx_real_insn" subclass for the INSN_P predicate, adding
the idea of a PATTERN, a basic_block, and a location - but I hardly
use this anywhere. That said, it seems to be a real concept in the
code, so I added it.
* "pointerness" of the types. "rtx" is a typedef to "rtx_def *" i.e.
there's an implicit pointer. In the discussion about using C++
classes for gimple statements:
https://gcc.gnu.org/ml/gcc-patches/2014-04/msg01427.html
Richi said:
> To followup myself here, it's because 'tree' is a typedef to a pointer
> and thus 'const tree' is different from 'const tree_node *'.
>
> Not sure why we re-introduced the 'mistake' of making 'tree' a pointer
> when we introduced 'gimple'. If we were to make 'gimple' the class
> type itself we can use gimple *, const gimple * and also const gimple &
> (when a NULL pointer is not expected).
So in the following patches the pointerness is explicit: the patches
refer to:
rtx_insn *insn;
rather than just:
rtx_insn insn;
and so one can write:
const rtx_insn *insn
and the "constness" applies to the insn, not to the pointer.
But we could go either way here: the class could be "rtx_insn_def",
with "rtx_insn" a typedef to an "rtx_insn_def *" etc with:
class rtx_def;
class rtx_expr_list_def; /* GET_CODE (X) == EXPR_LIST */
class rtx_insn_list_def; /* GET_CODE (X) == INSN_LIST */
class rtx_sequence_def; /* GET_CODE (X) == SEQUENCE */
class rtx_insn_def; /* INSN_CHAIN_CODE_P (GET_CODE (X)) */
class rtx_real_insn_def; /* INSN_P (X) */
class rtx_debug_insn_def; /* DEBUG_INSN_P (X) */
class rtx_nonjump_insn_def; /* NONJUMP_INSN_P (X) */
class rtx_jump_insn_def; /* JUMP_P (X) */
class rtx_call_insn_def; /* CALL_P (X) */
class rtx_jump_table_data_def; /* JUMP_TABLE_DATA_P (X) */
class rtx_barrier_def; /* BARRIER_P (X) */
class rtx_code_label_def; /* LABEL_P (X) */
class rtx_note_def; /* NOTE_P (X) */
and a family of typedefs of pointers to the classes:
typedef rtx_def *rtx;
typedef rtx_expr_list_def *rtx_expr_list;
typedef rtx_insn_list_def *rtx_insn_list;
typedef rtx_sequence_def *rtx_sequence;
typedef rtx_insn_def *rtx_insn;
typedef rtx_real_insn_def *rtx_real_insn;
typedef rtx_debug_insn_def *rtx_debug_insn;
typedef rtx_nonjump_insn_def *rtx_nonjump_insn;
typedef rtx_jump_insn_def *rtx_jump_insn;
typedef rtx_call_insn_def *rtx_call_insn;
typedef rtx_jump_table_data_def *rtx_jump_table_data;
typedef rtx_barrier_def *rtx_barrier;
typedef rtx_code_label_def *rtx_code_label;
typedef rtx_note_def *rtx_note;
* Should as_a <rtx_insn *> accept a NULL pointer? It's possible to make
the is_a_helper cope with NULL, but this adds an extra conditional.
I instead added an as_a_nullable<> cast, so that you can explicitly
add a check against NULL before checking the code of the rtx node.
But maybe it's cleaner to simply have is_a_helper<rtx_insn *> cope
with NULL?
Some deeper questions:
* should rtx_insn eventually be a separate class from rtx, separating
insn chain nodes from rtx nodes? I don't know if that's a worthwhile
longterm goal, but this patch series gets us somewhere closer to
being able to achieve that. (actually getting there would be a much
more invasive set of patches).
To keep this reviewable, and to try to mitigate bitrot, I've chopped it up
into numerous relatively small patches. The aim is that at every patch,
the code correctly builds on all supported configurations. That said,
there's a complicated dependency graph of types in gcc's code, so to
tame that, the patch series is divided into 6 phases:
* phase 1 adds "scaffolding": in various places, strengthen the return
types from internal APIs and macros so as to promise an rtx_insn *
rather than a plain rtx. For example, the DEP_PRO/DEP_CON macros
in sched-int.h which lookup fields within struct _dep become inline
functions that return rtx_insn * (using checked casts). In this way,
stronger type information can be used by subsequent patches whilst
avoiding the chicken-and-egg issue since writes to the fields can
still be arbitrary rtx nodes, according to the type system at least.
* phase 2: an alphabetical tour of the backend: a series of patches,
from alias.c through web.c, each patch touching one file,
strengthening the types within it as much as we can at that point.
The patches sometimes make use of the alphabetic ordering in order to
use APIs that have already been strengthened to work on rtx_insn.
* phase 3: similar to phase 2, but for the various config
subdirectories.
* phase 4: tears down the scaffolding, replacing checked casts as much
as possible by strengthening core fields of core types. For example,
we eventually reach the point in sched-int.h where the fields "pro"
and "con" within struct _dep can become rtx_insn *, and hence
DEP_PRO/DEP_CON can be converted back to plain macros, without needing
the checked cast.
* phase 5: all of the above was for instructions; this phase adds three
subclasses for other node kinds. I experimented with subclasses for
various node kinds; these three seemed most appropriate: EXPR_LIST,
INSN_LIST, SEQUENCE. I kept these as a separate phase as Jeff asked
me to separate them from the instruction patches, to avoid
complicating things, but I think these three are also a worthwhile
cleanup with relatively little complexity.
* phase 6: (new since my Cauldron talk): this phase freely uses
EXPR_LIST, INSN_LIST, SEQUENCE to do cleanups that were awkward
without them. In particular, by the end of this phase, NEXT_INSN()
and PREV_INSN() require rtx_insn * rather than plain rtx.
Correctness
===========
In theory, these patches should not affect the outwardly-visible behavior
of the compiler, merely enable various kinds of errors to be more easily
detected when the compiler is built, and to improve the readability of the
code.
The patches (and my control for testing) has been on top of r211061, which
being from 2014-05-29 is now two months old, but hopefully is still
reviewable; naturally I'll perform bootstrap and regression testing for
each patch in turn before committing.
I've successfully bootsrapped & regression-tested the end-result of the
attached patches on:
* x86_64 Linux (Fedora 20)
* powerpc64-unknown-linux-gnu (Fedora 18; gcc110 in buildfarm)
* s390x-ibm-linux-gnu (RHEL 7)
I've also repeatedly bootstrapped this on x86_64 at many stages of
assembling the patch series.
There are 204 configurations in the list in contrib/config-list.mk, however
12 failed to work for a control build; so I patched my config-list.mk
accordingly:
# Which of the above are known to currently not work?
KNOWN_BROKEN=
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55143
KNOWN_BROKEN += alpha64-dec-vms alpha-dec-vms
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55144
KNOWN_BROKEN += bfin-linux-uclibc cris-linux crisv32-linux
# Discussion at: https://gcc.gnu.org/ml/gcc/2013-11/msg00574.html
# but this doesn't seem to have been filed in BZ
KNOWN_BROKEN += i686-interix3OPT-enable-obsolete
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47098
KNOWN_BROKEN += i686-openbsd3.0
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55143
KNOWN_BROKEN += ia64-hp-vms
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=61287
KNOWN_BROKEN += nios2-elf nios2-linux-gnu
# See e.g. https://www.mail-archive.com/gcc@gcc.gnu.org/msg70568.html
# but this doesn't seem to have been filed in BZ
KNOWN_BROKEN += vax-openbsd
# https://gcc.gnu.org/bugzilla/show_bug.cgi?id=48904
KNOWN_BROKEN += x86_64-knetbsd-gnu
LIST= $(filter-out $(KNOWN_BROKEN),$(FULL_LIST))
With that, I've successfully built the patches on 193 configurations, both
the end state of the patch series, and at many places along the way.
I've manually verified the build for nios2-elf *without* ada, to
work around PR61287.
Hence I believe that this should continue to build on every supported
target.
Performance
===========
I tested the performance with --enable-checking=release using
two large files (kdecore.cc, bigcode.c), comparing a control build
to a patched build.
There were no significant differences in compilation time:
Compilation of kdecore.cc at -O3 with -g for x86_64-unknown-linux-gnu: usr
control: [47.58, 47.99, 47.6, 47.8, 47.74, 47.76, 47.81, 47.9, 47.91, 47.99, 48.14, 47.77, 47.65, 47.86, 47.96, 47.74, 48.09, 47.97, 47.86, 47.79, 48.17, 47.76, 47.88, 47.85, 48.24, 48.01, 47.98, 47.91, 48.12, 47.7, 47.75, 47.69, 47.83, 47.82, 47.76, 48.05, 47.85, 48.05, 48.25, 47.95]
experiment: [47.65, 47.68, 47.67, 47.73, 47.91, 48.27, 47.83, 48.0, 47.95, 47.75, 47.72, 47.81, 47.98, 48.36, 47.67, 47.72, 47.81, 47.83, 47.89, 47.67, 47.72, 47.74, 47.79, 47.77, 47.67, 48.78, 47.88, 47.76, 47.96, 47.71, 47.87, 47.77, 47.82, 47.74, 47.77, 48.28, 47.78, 47.62, 47.67, 47.72]
Min: 47.580000 -> 47.620000: 1.00x slower
Avg: 47.888250 -> 47.843000: 1.00x faster
Not significant
Stddev: 0.16618 -> 0.22717: 1.3670x larger
Timeline: http://goo.gl/ikKUwD
Compilation of kdecore.cc at -O3 with -g for x86_64-unknown-linux-gnu: sys
control: [6.29, 6.04, 6.31, 6.15, 6.24, 6.19, 6.2, 6.18, 6.0, 6.15, 6.12, 6.34, 6.26, 6.09, 6.24, 6.16, 6.0, 6.25, 6.19, 6.2, 6.23, 6.24, 6.19, 6.22, 6.22, 6.26, 6.18, 6.11, 6.16, 6.24, 6.19, 6.23, 6.33, 6.14, 6.21, 6.2, 6.23, 6.14, 6.23, 6.13]
experiment: [6.3, 6.19, 6.21, 6.21, 6.21, 6.25, 6.19, 6.1, 6.13, 6.12, 6.19, 6.09, 6.31, 6.23, 6.2, 6.22, 6.21, 6.22, 6.09, 6.15, 6.14, 6.26, 6.09, 6.11, 6.3, 6.19, 6.14, 6.14, 6.14, 6.17, 6.12, 6.11, 6.24, 6.26, 6.21, 6.17, 6.16, 6.18, 6.19, 6.15]
Min: 6.000000 -> 6.090000: 1.01x slower
Avg: 6.192000 -> 6.182250: 1.00x faster
Not significant
Stddev: 0.07653 -> 0.05925: 1.2918x smaller
Timeline: http://goo.gl/3r9lQJ
Compilation of kdecore.cc at -O3 with -g for x86_64-unknown-linux-gnu: wall
control: [54.16, 54.21, 54.1, 54.12, 54.16, 54.13, 54.19, 54.26, 54.09, 54.32, 54.44, 54.29, 54.09, 54.13, 54.38, 54.08, 54.27, 54.41, 54.22, 54.16, 54.58, 54.17, 54.24, 54.25, 54.64, 54.44, 54.34, 54.2, 54.46, 54.13, 54.12, 54.1, 54.35, 54.14, 54.15, 54.43, 54.26, 54.37, 54.66, 54.26]
experiment: [54.21, 54.24, 54.09, 54.3, 54.46, 54.86, 54.34, 54.39, 54.4, 54.17, 54.16, 54.16, 54.48, 54.82, 54.11, 54.13, 54.25, 54.32, 54.24, 54.08, 54.11, 54.19, 54.07, 54.06, 54.16, 55.23, 54.21, 54.1, 54.34, 54.13, 54.18, 54.11, 54.25, 54.23, 54.17, 54.64, 54.15, 53.99, 54.05, 54.06]
Min: 54.080000 -> 53.990000: 1.00x faster
Avg: 54.262500 -> 54.266000: 1.00x slower
Not significant
Stddev: 0.15337 -> 0.24797: 1.6169x larger
Timeline: http://goo.gl/Gbk0RB
Compilation of kdecore.cc at -O3 with -g for x86_64-unknown-linux-gnu: ggc
control: [1264522.0, 1264525.0, 1264523.0, 1264524.0, 1264522.0, 1264520.0, 1264524.0, 1264517.0, 1264523.0, 1264529.0, 1264518.0, 1264522.0, 1264531.0, 1264519.0, 1264524.0, 1264518.0, 1264533.0, 1264522.0, 1264524.0, 1264529.0, 1264515.0, 1264525.0, 1264526.0, 1264523.0, 1264523.0, 1264531.0, 1264526.0, 1264524.0, 1264518.0, 1264521.0, 1264522.0, 1264520.0, 1264523.0, 1264520.0, 1264525.0, 1264532.0, 1264524.0, 1264524.0, 1264519.0, 1264525.0]
experiment: [1264530.0, 1264523.0, 1264523.0, 1264524.0, 1264531.0, 1264523.0, 1264528.0, 1264519.0, 1264516.0, 1264524.0, 1264522.0, 1264524.0, 1264522.0, 1264525.0, 1264520.0, 1264522.0, 1264518.0, 1264518.0, 1264519.0, 1264523.0, 1264522.0, 1264517.0, 1264528.0, 1264525.0, 1264519.0, 1264522.0, 1264522.0, 1264519.0, 1264519.0, 1264523.0, 1264532.0, 1264526.0, 1264526.0, 1264520.0, 1264525.0, 1264530.0, 1264529.0, 1264525.0, 1264527.0, 1264523.0]
Mem max: 1264533.000 -> 1264532.000: 1.0000x smaller
Usage over time: http://goo.gl/OPqL9P
Compilation of big-code.c at -O3 with -g for x86_64-unknown-linux-gnu: usr
control: [36.16, 36.13, 36.13, 36.33, 36.16, 36.12, 36.07, 36.19, 36.11, 36.58, 36.14, 36.53, 36.19, 36.17, 36.16, 36.19, 36.22, 36.34, 36.14, 36.13, 36.12, 36.38, 36.17, 36.28, 36.38, 36.15, 36.13, 36.16, 36.23, 36.17, 36.16, 36.07, 36.24, 36.17, 36.11, 36.14, 36.14, 36.12, 36.15, 36.21]
experiment: [36.15, 36.09, 36.16, 36.23, 36.14, 36.11, 36.16, 36.16, 36.15, 36.14, 36.2, 36.2, 36.22, 36.15, 36.14, 36.23, 36.23, 36.15, 36.19, 36.09, 36.12, 36.26, 36.18, 36.14, 36.23, 36.18, 36.19, 36.13, 36.42, 36.15, 36.16, 36.18, 36.19, 36.12, 36.16, 36.3, 36.1, 36.13, 36.19, 36.3]
Min: 36.070000 -> 36.090000: 1.00x slower
Avg: 36.196750 -> 36.178000: 1.00x faster
Not significant
Stddev: 0.11136 -> 0.06366: 1.7494x smaller
Timeline: http://goo.gl/TbdnnN
Compilation of big-code.c at -O3 with -g for x86_64-unknown-linux-gnu: sys
control: [1.31, 1.3, 1.33, 1.42, 1.35, 1.34, 1.34, 1.32, 1.32, 1.27, 1.32, 1.29, 1.3, 1.3, 1.29, 1.33, 1.3, 1.3, 1.28, 1.29, 1.3, 1.39, 1.31, 1.36, 1.32, 1.29, 1.33, 1.33, 1.29, 1.29, 1.31, 1.36, 1.25, 1.29, 1.33, 1.33, 1.38, 1.28, 1.39, 1.27]
experiment: [1.24, 1.31, 1.27, 1.34, 1.39, 1.35, 1.27, 1.26, 1.28, 1.38, 1.32, 1.29, 1.25, 1.26, 1.28, 1.27, 1.3, 1.26, 1.29, 1.31, 1.35, 1.27, 1.34, 1.28, 1.22, 1.28, 1.43, 1.33, 1.28, 1.29, 1.26, 1.33, 1.28, 1.34, 1.32, 1.26, 1.41, 1.35, 1.3, 1.34]
Min: 1.250000 -> 1.220000: 1.02x faster
Avg: 1.317500 -> 1.304500: 1.01x faster
Not significant
Stddev: 0.03607 -> 0.04701: 1.3032x larger
Timeline: http://goo.gl/ocXV2i
Compilation of big-code.c at -O3 with -g for x86_64-unknown-linux-gnu: wall
control: [37.59, 37.54, 37.57, 37.86, 37.62, 37.57, 37.52, 37.63, 37.54, 37.96, 37.57, 37.93, 37.61, 37.58, 37.56, 37.63, 37.63, 37.75, 37.54, 37.54, 37.53, 37.89, 37.59, 37.75, 37.8, 37.55, 37.58, 37.6, 37.63, 37.57, 37.58, 37.54, 37.59, 37.57, 37.55, 37.58, 37.62, 37.51, 37.65, 37.59]
experiment: [37.5, 37.5, 37.53, 37.68, 37.64, 37.57, 37.54, 37.53, 37.54, 37.62, 37.63, 37.6, 37.58, 37.52, 37.53, 37.6, 37.65, 37.52, 37.59, 37.51, 37.58, 37.64, 37.63, 37.53, 37.56, 37.57, 37.73, 37.57, 37.82, 37.55, 37.52, 37.62, 37.59, 37.58, 37.6, 37.67, 37.62, 37.6, 37.6, 37.75]
Min: 37.510000 -> 37.500000: 1.00x faster
Avg: 37.625250 -> 37.592750: 1.00x faster
Not significant
Stddev: 0.11404 -> 0.06943: 1.6425x smaller
Timeline: http://goo.gl/WT5UGE
Compilation of big-code.c at -O3 with -g for x86_64-unknown-linux-gnu: ggc
control: [657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0]
experiment: [657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0, 657274.0]
Mem max: 657274.000 -> 657274.000: no change
Usage over time: http://goo.gl/ogWEOG
As for the performance of a regular build i.e. with as_a<>
checks *enabled*; looking at the wallclock time taken for a bootstrap and
regression test, for my s390 builds (with -j3) I saw:
s390 control:
"make" time: 68 mins
"make check" time: 122 mins
total time: 190 mins
s390 experiment:
"make" time: 70 mins
"make check" time: 126 mins
total time: 196 mins
showing a 3% increase, presumably due to the as_a and as_a_nullable
checks.
i.e. a release build shows no change in performance; a debug build shows
a 3% increase in time taken to bootstrap and regression test. I believe
the debug build could be sped up with further patches to eliminate the
checked casts.
Summary
=======
OK for trunk?
Dave
Patch list and overall diffstat follow:
David Malcolm (236):
Convert lab_rtx_for_bb from pointer_map_t to pointer_map<rtx>
JUMP_LABEL is not always a LABEL
config/mn10300: Fix missing PATTERN in PARALLEL handling
PHASE 1: Initial "scaffolding" commits
Introduce as_a_nullable
Introduce rtx_insn subclass of rtx_def
New function: for_each_rtx_in_insn
Split BB_HEAD et al into BB_HEAD/SET_BB_HEAD variants
Replace BB_HEAD et al macros with functions
Split NEXT_INSN/PREV_INSN into lvalue and rvalue forms
Replace PREV_INSN et al macros with functions
Convert DF_REF_INSN to a function for now
DEP_PRO/DEP_CON scaffolding
VINSN_INSN_RTX scaffolding
BB_NOTE_LIST scaffolding
BND_TO scaffolding
Add subclasses for the various kinds of instruction
Strengthen return types of various {next|prev}_*insn from rtx to
rtx_insn *
Strengthen return type of gen_label_rtx
Return rtx_insn from get_insns/get_last_insn
entry_of_function returns an insn
Make tablejump_p accept a rtx_jump_table_data **
delete_trivially_dead_insns works on insns
last_call_insn returns an rtx_call_insn *
make_insn_raw returns an rtx_insn
bb_note returns a rtx_note *
asan_emit_stack_protection returns an insn
cfgexpand.c: Use rtx_insn
rtl_data.x_parm_birth_insn is an insn
Convert various rtx to rtx_note *
emit_jump_table_data returns an rtx_jump_table_data *
emit_* functions return rtx_insn
emit_move et al return rtx_insn *
next_cc0_user and prev_cc0_setter scaffolding
Return types of unlink_insn_chain and duplicate_insn_chain
get_last_bb_insn returns an rtx_insn
sel_bb_{head|end} return rtx_insn
find_first_parameter_load returns an rtx_insn
create_insn_rtx_from_pattern and create_copy_of_insn_rtx return
rtx_insn
Use rtx_insn internally within generated functions
Debug hooks: use rtx_insn and rtx_code_label
try_split returns an rtx_insn
peephole returns an rtx_insn
Pass "insn" as an rtx_insn within generated get_attr_ fns in
insn-attrtab.c
define_bypass guard functions take a pair of rtx_insn
delete_related_insns returns an rtx_insn
PHASE 2: Per-file commits in main source directory
alias.c: Use rtx_insn
asan.c: strengthen some rtx locals
auto-inc-dec.c: strengthen various rtx to rtx_insn *
bb-reorder.c: Use rtx_insn
bt-load.c: Use rtx_insn
builtins.c: strengthen various rtx to rtx_insn * and other subclasses
calls.c: Use rtx_insn
caller-save.c: Use rtx_insn
cfgbuild.c: Use rtx_insn
cfgcleanup.c: Use rtx_insn (also touches basic-block.h and ifcvt.c)
cfgloop.c: Use rtx_insn
cfgloopanal.c: Use rtx_insn
cfgrtl.c: Use rtx subclasses
combine.c: Use rtx_insn
combine-stack-adj.c: Use rtx_insn
compare-elim.c: Use rtx_insn
cprop.c: Use rtx_insn
cse.c: Use rtx_insn
dce.c: Use rtx subclasses
ddg: Use rtx_insn
df-*.c: Use rtx_insn
dwarf2cfi.c: Use rtx_insn
dwarf2out.c: Use rtx_insn
except.*: Use rtx_insn (also touches function.h)
explow.c: Use rtx_insn and rtx_code_label
expmed.c: Use rtx_insn and rtx_code_label
expr.c: Use rtx_insn and rtx_code_label
final.c: Use rtx_insn (also touches output.c and config/arc/arc.c)
function.c: Use rtx_insn
fwprop.c: Use rtx_insn
genpeep.c: peephole requires an rtx_insn
gcse.c: Use rtx_insn
haifa-sched.c: Use rtx_insn
hw-doloop: Use rtx_insn (touches config/bfin/bfin.c)
ifcvt.c: Use rtx_insn
init-regs.c: rtx_insn
internal-fn.c: Use rtx_insn and rtx_code_label
ira: Use rtx_insn in various places
jump.c: Use rtx_insn in a few places (also touches rtl.h and
cfgexpand.c)
loop-doloop.c: Use rtx_insn in a few places
loop-invariant.c: Use rtx_insn in various places
loop-iv.c: Use rtx_insn (also touches cfgloop.h and loop-unroll.c)
loop-unroll.c: Use rtx_insn (also touches basic-block.h)
lower-subreg.c: Use rtx_insn
lra: use rtx_insn
mode-switching.c: Use rtx_insn
get_ebb_head_tail works with rtx_insn
modulo-sched.c: Use rtx_insn in various places
optabs.c: Use rtx_insn and rtx_code_label
postreload-gcse.c: Use rtx_insn in various places
postreload.c: Use rtx_insn (also touches rtl.h and cprop.c)
predict.*: Use rtx_insn (also touches function.c and
config/cris/cris.c)
print-rtl.c: Use rtx_insn for various debug_ functions (also touches
config/rs6000/rs6000.c)
recog.c: Use rtx_insn
ree.c: Use rtx_insn
reg-stack.c: Use rtx_insn
regcprop.c: Use rtx_insn
reginfo.c: Use rtx_insn (also touches rtl.h)
regrename.c: Use rtx_insn
regstat.c: Use rtx_insn
reload: Use rtx_insn (also touches caller-save.c and config/arc/arc)
resource.c: Use rtx_insn
rtlanal.c: Use rtx_insn
sched-deps.c: Use rtx_insn
sched-ebb.c: Use rtx_insn (requires touching sched-int.h and
config/c6x/c6x.c)
sched-rgn.c: Use rtx_insn in a couple of places
sel-sched.c: Use rtx_insn
sel-sched-ir.c: Use rtx_insn
shrink-wrap.*: Use rtx_insn (touches config/i386/i386.c)
stack-ptr-mod.c: Use rtx_insn
stmt.c: Use rtx_insn
store-motion.c: Use rtx_insn
valtrack.c: Use rtx_insn
varasm.c: Use rtx_insn
var-tracking.c: Use rtx_insn
web.c: Use rtx_insn
PHASE 3: Per-config subdir commits
config/aarch64/aarch64.c: Use rtx_insn
config/alpha/alpha.c: Use rtx_insn
config/arc: Use rtx_insn
config/arm: Use rtx_insn and rtx_code_label
config/avr: Use rtx_insn
config/bfin: Use rtx_insn
config/c6x: Use rtx_insn
config/epiphany: Use rtx_insn
config/h8300: Use rtx_insn
config/i386/i386.c: Use rtx_code_label
config/i386/i386: Use rtx_insn
config/ia64/ia64.c: Use rtx_insn
config/iq2000: Use rtx_insn
config/m68k: Use rtx_insn
config/mep: Use rtx_insn and rtx_code_label
config/microblaze/microblaze.c: Use rtx_insn and rtx_code_label
config/mips: Use rtx_insn and rtx_code_label
config/nds32: Use rtx_insn
config/pa: Use rtx_insn
config/picochip: Use rtx_insn
config/rs6000: Use rtx_insn
config/rx: Use rtx_insn
config/s390: Use rtx_insn and rtx_code_label
config/score/score.c: Use rtx_insn
config/sh: Use rtx_insn and rtx_code_label
config/sparc: Use rtx_insn
config/spu/spu.c: Use rtx_insn
config/tilegx: Use rtx_insn
config/tilepro: Use rtx_insn
config/v850: Use rtx_insn
config/xtensa: Use rtx_insn and rtx_code_label
PHASE 4: Removal of scaffolding
struct eh_landing_pad_d: field "landing_pad" is an rtx_code_label
Remove BB_FOOTER scaffolding
Convert edge_def.insns.r to rtx_insn *
function.c and shrink-wrap.*: more rtx_insn
reorder_insns requires rtx_insn *
delete_insn_and_edges takes an rtx_insn *
unshare_all_rtl_again takes an rtx_insn *
Add rtx_jump_table_data::get_labels method
struct haifa_sched_info: prev_head and next_tail
shorten_branches takes an rtx_insn
final accepts an rtx_insn
final_start_function takes an rtx_insn
Strengthen haifa_sched_info callbacks and 3 scheduler hooks
Eliminate BB_NOTE_LIST scaffolding
du_chain.insn is an rtx_insn
sel-sched-ir.h: Make ilist_t work on insn_t rather than rtx
insn_t becomes an rtx_insn *
Remove VINSN_INSN_RTX scaffolding
Remove DEP_PRO/CON scaffolding
cselib and incdec
Tighten up params of create_basic_block_structure
Remove BB_HEAD, BB_END, BB_HEADER scaffolding
cselib_record_sets_hook takes an rtx_insn
Params of add_insn and unlink_insn_chain
Strengthen fields in struct sequence_stack and struct emit_status
get_last_insn_anywhere returns an rtx_insn
Strengthen various insn emission functions
Use rtx_insn in more places in sel-sched.c
Use rtx_insn in more places in fwprop.c
Various condition-handling calls
duplicate_insn_chain accepts rtx_insn
Use rtx_insn in more places in haifa-sched.c
Various scheduling strengthenings
Remove insn_addresses_new from 'various scheduling strengthenings'
Remove DF_REF_INSN scaffolding
Tweak to dse.c
cselib (also touches sched-deps.c)
Use rtx_insn for various target.def hooks
Convert PATTERN from a macro to a pair of inline functions
Convert various INSN accessors in rtl.h to inline functions
Tweak to ira-lives.c
PHASE 5: Additional rtx subclasses
Introduce rtx_insn_list subclass of rtx_def
Use rtx_insn_list in various places
Introduce rtx_sequence subclass of rtx_def
dwarf2cfi.c: Use rtx_sequence
except.c: Use rtx_sequence
final.c: Use rtx_sequence
function.c: Use rtx_sequence
jump.c: Use rtx_sequence
reorg.c: Use rtx_sequence
resource.c: Use rtx_sequence
sched-vis.c: Use rtx_sequence
varasm.c: Use rtx_sequence
Introduce rtx_expr_list subclass of rtx_def
Use rtx_expr_list for expr_status.x_forced_labels
rtl_data.x_nonlocal_goto_handler_labels becomes an rtx_expr_list
rtl_data.x_stack_slot_list becomes an rtx_expr_list
Use rtx_expr_list in various places
PHASE 6: Use extra rtx_def subclasses
Add JUMP_LABEL_AS_INSN
Use rtx subclasses in more places in reorg.c
Make SET_NEXT_INSN/SET_PREV_INSN require an rtx_insn
Strengthen return_label and naked_return_label to rtx_code_label *
Add insn method to rtx_expr_list
Use rtx_insn in more places in dwarf2cfi.c
inside_basic_block_p requires a const rtx_insn *
insn_current_reference_address takes an rtx_insn
Work towards NEXT_INSN/PREV_INSN requiring insns as their params
Delete find_last_value
find_first_parameter_load params and return type
tablejump_p takes an rtx_insn
NEXT_INSN and PREV_INSN take a const rtx_insn
Make INSN_HAS_LOCATION require an rtx_insn
Make insn_addresses_new require an rtx_insn
Use rtx_insn in various places in resource.[ch]
dfa_clear_single_insn_cache takes an rtx_insn
Strengthen params to active_insn_between
Make next_insn and previous_insn require an rtx_insn *
END OF PATCHES: Delete rtx-classes-status.txt
gcc/alias.c | 3 +-
gcc/asan.c | 11 +-
gcc/asan.h | 4 +-
gcc/auto-inc-dec.c | 54 +--
gcc/basic-block.h | 31 +-
gcc/bb-reorder.c | 30 +-
gcc/bt-load.c | 37 +-
gcc/builtins.c | 47 +-
gcc/caller-save.c | 18 +-
gcc/calls.c | 47 +-
gcc/cfgbuild.c | 44 +-
gcc/cfgcleanup.c | 98 +++--
gcc/cfgexpand.c | 75 ++--
gcc/cfgloop.c | 2 +-
gcc/cfgloop.h | 9 +-
gcc/cfgloopanal.c | 8 +-
gcc/cfgrtl.c | 365 ++++++++--------
gcc/combine-stack-adj.c | 49 ++-
gcc/combine.c | 464 ++++++++++----------
gcc/compare-elim.c | 19 +-
gcc/config/aarch64/aarch64.c | 25 +-
gcc/config/alpha/alpha.c | 69 +--
gcc/config/arc/arc-protos.h | 16 +-
gcc/config/arc/arc.c | 109 ++---
gcc/config/arc/arc.md | 8 +-
gcc/config/arc/constraints.md | 2 +-
gcc/config/arm/arm-protos.h | 6 +-
gcc/config/arm/arm.c | 137 +++---
gcc/config/avr/avr-log.c | 2 +-
gcc/config/avr/avr-protos.h | 69 +--
gcc/config/avr/avr.c | 153 +++----
gcc/config/bfin/bfin-protos.h | 2 +-
gcc/config/bfin/bfin.c | 96 +++--
gcc/config/c6x/c6x-protos.h | 4 +-
gcc/config/c6x/c6x.c | 206 ++++-----
gcc/config/cris/cris.c | 11 +-
gcc/config/epiphany/epiphany-protos.h | 6 +-
gcc/config/epiphany/epiphany.c | 26 +-
gcc/config/epiphany/mode-switch-use.c | 2 +-
gcc/config/epiphany/resolve-sw-modes.c | 5 +-
gcc/config/frv/frv.c | 61 +--
gcc/config/h8300/h8300-protos.h | 4 +-
gcc/config/h8300/h8300.c | 19 +-
gcc/config/i386/i386-protos.h | 14 +-
gcc/config/i386/i386.c | 278 ++++++------
gcc/config/i386/winnt.c | 2 +-
gcc/config/ia64/ia64.c | 193 +++++----
gcc/config/iq2000/iq2000-protos.h | 11 +-
gcc/config/iq2000/iq2000.c | 24 +-
gcc/config/iq2000/iq2000.md | 4 +-
gcc/config/m32c/m32c.c | 4 +-
gcc/config/m32r/m32r.c | 4 +-
gcc/config/m68k/m68k-protos.h | 7 +-
gcc/config/m68k/m68k.c | 43 +-
gcc/config/mcore/mcore-protos.h | 2 +-
gcc/config/mcore/mcore.c | 28 +-
gcc/config/mep/mep-protos.h | 18 +-
gcc/config/mep/mep.c | 262 ++++++------
gcc/config/microblaze/microblaze.c | 28 +-
gcc/config/microblaze/microblaze.md | 4 +-
gcc/config/mips/mips-protos.h | 16 +-
gcc/config/mips/mips.c | 237 ++++++-----
gcc/config/mips/mips.md | 2 +-
gcc/config/mn10300/mn10300.c | 12 +-
gcc/config/nds32/nds32-protos.h | 2 +-
gcc/config/nds32/nds32.c | 6 +-
gcc/config/pa/pa-protos.h | 40 +-
gcc/config/pa/pa.c | 81 ++--
gcc/config/picochip/picochip-protos.h | 3 +-
gcc/config/picochip/picochip.c | 59 +--
gcc/config/rs6000/rs6000-protos.h | 4 +-
gcc/config/rs6000/rs6000.c | 82 ++--
gcc/config/rx/rx-protos.h | 2 +-
gcc/config/rx/rx.c | 2 +-
gcc/config/s390/s390-protos.h | 8 +-
gcc/config/s390/s390.c | 186 ++++----
gcc/config/score/score.c | 5 +-
gcc/config/sh/sh-protos.h | 23 +-
gcc/config/sh/sh.c | 279 ++++++------
gcc/config/sh/sh.md | 11 +-
gcc/config/sh/sh_optimize_sett_clrt.cc | 12 +-
gcc/config/sh/sh_treg_combine.cc | 35 +-
gcc/config/sparc/sparc-protos.h | 20 +-
gcc/config/sparc/sparc.c | 68 +--
gcc/config/spu/spu.c | 67 +--
gcc/config/spu/spu.md | 4 +-
gcc/config/stormy16/stormy16.c | 6 +-
gcc/config/tilegx/tilegx-protos.h | 8 +-
gcc/config/tilegx/tilegx.c | 93 ++--
gcc/config/tilepro/tilepro-protos.h | 8 +-
gcc/config/tilepro/tilepro.c | 89 ++--
gcc/config/v850/v850-protos.h | 2 +-
gcc/config/v850/v850.c | 30 +-
gcc/config/xtensa/xtensa-protos.h | 2 +-
gcc/config/xtensa/xtensa.c | 23 +-
gcc/coretypes.h | 19 +
gcc/cprop.c | 54 +--
gcc/cse.c | 121 +++---
gcc/cselib.c | 24 +-
gcc/cselib.h | 12 +-
gcc/dbxout.c | 8 +-
gcc/dce.c | 46 +-
gcc/ddg.c | 28 +-
gcc/ddg.h | 8 +-
gcc/debug.c | 9 +-
gcc/debug.h | 8 +-
gcc/df-core.c | 24 +-
gcc/df-problems.c | 67 +--
gcc/df-scan.c | 49 +--
gcc/df.h | 64 +--
gcc/doc/tm.texi | 42 +-
gcc/dse.c | 26 +-
gcc/dwarf2cfi.c | 79 ++--
gcc/dwarf2out.c | 30 +-
gcc/emit-rtl.c | 588 ++++++++++++++-----------
gcc/emit-rtl.h | 12 +-
gcc/except.c | 80 ++--
gcc/except.h | 4 +-
gcc/explow.c | 30 +-
gcc/expmed.c | 61 +--
gcc/expr.c | 84 ++--
gcc/expr.h | 8 +-
gcc/final.c | 168 ++++----
gcc/function.c | 93 ++--
gcc/function.h | 26 +-
gcc/fwprop.c | 40 +-
gcc/gcse.c | 108 ++---
gcc/genattr.c | 10 +-
gcc/genattrtab.c | 13 +-
gcc/genautomata.c | 12 +-
gcc/genconditions.c | 2 +-
gcc/genemit.c | 8 +-
gcc/gengenrtl.c | 4 +-
gcc/genoutput.c | 4 +-
gcc/genpeep.c | 5 +-
gcc/genrecog.c | 29 +-
gcc/haifa-sched.c | 572 +++++++++++++------------
gcc/hooks.c | 16 +-
gcc/hooks.h | 8 +-
gcc/hw-doloop.c | 13 +-
gcc/hw-doloop.h | 6 +-
gcc/ifcvt.c | 218 +++++-----
gcc/init-regs.c | 4 +-
gcc/insn-addr.h | 2 +-
gcc/internal-fn.c | 33 +-
gcc/ira-build.c | 10 +-
gcc/ira-conflicts.c | 10 +-
gcc/ira-costs.c | 13 +-
gcc/ira-emit.c | 27 +-
gcc/ira-int.h | 8 +-
gcc/ira-lives.c | 8 +-
gcc/ira.c | 64 +--
gcc/ira.h | 2 +-
gcc/is-a.h | 24 ++
gcc/jump.c | 65 +--
gcc/lists.c | 70 +--
gcc/loop-doloop.c | 19 +-
gcc/loop-invariant.c | 27 +-
gcc/loop-iv.c | 42 +-
gcc/loop-unroll.c | 59 +--
gcc/lower-subreg.c | 49 ++-
gcc/lra-assigns.c | 2 +-
gcc/lra-coalesce.c | 21 +-
gcc/lra-constraints.c | 138 +++---
gcc/lra-eliminations.c | 12 +-
gcc/lra-int.h | 34 +-
gcc/lra-lives.c | 2 +-
gcc/lra-spills.c | 9 +-
gcc/lra.c | 84 ++--
gcc/mode-switching.c | 21 +-
gcc/modulo-sched.c | 64 +--
gcc/optabs.c | 161 ++++---
gcc/output.h | 18 +-
gcc/postreload-gcse.c | 66 +--
gcc/postreload.c | 56 +--
gcc/predict.c | 18 +-
gcc/predict.h | 2 +-
gcc/print-rtl.c | 24 +-
gcc/recog.c | 39 +-
gcc/recog.h | 2 +-
gcc/ree.c | 45 +-
gcc/reg-stack.c | 89 ++--
gcc/regcprop.c | 20 +-
gcc/reginfo.c | 10 +-
gcc/regrename.c | 22 +-
gcc/regrename.h | 2 +-
gcc/regstat.c | 4 +-
gcc/reload.c | 39 +-
gcc/reload.h | 14 +-
gcc/reload1.c | 217 +++++-----
gcc/reorg.c | 463 ++++++++++----------
gcc/resource.c | 91 ++--
gcc/resource.h | 8 +-
gcc/rtl.h | 757 +++++++++++++++++++++++++++------
gcc/rtlanal.c | 136 +++---
gcc/sched-deps.c | 215 +++++-----
gcc/sched-ebb.c | 60 +--
gcc/sched-int.h | 147 +++----
gcc/sched-rgn.c | 110 ++---
gcc/sched-vis.c | 25 +-
gcc/sdbout.c | 6 +-
gcc/sel-sched-dump.c | 8 +-
gcc/sel-sched-dump.h | 2 +-
gcc/sel-sched-ir.c | 135 +++---
gcc/sel-sched-ir.h | 88 ++--
gcc/sel-sched.c | 89 ++--
gcc/shrink-wrap.c | 29 +-
gcc/shrink-wrap.h | 14 +-
gcc/stack-ptr-mod.c | 2 +-
gcc/stmt.c | 4 +-
gcc/store-motion.c | 70 +--
gcc/target.def | 58 +--
gcc/targhooks.c | 2 +-
gcc/targhooks.h | 2 +-
gcc/tree-cfg.c | 14 +
gcc/tree-outof-ssa.c | 4 +-
gcc/tree-ssa-loop-ivopts.c | 8 +-
gcc/valtrack.c | 13 +-
gcc/valtrack.h | 2 +-
gcc/var-tracking.c | 59 +--
gcc/varasm.c | 11 +-
gcc/vmsdbgout.c | 4 +-
gcc/web.c | 4 +-
223 files changed, 6733 insertions(+), 5524 deletions(-)
--
1.8.5.3