New attempt: subword-based DCE, PR42575

PR42575 is an example where we're creating a dead store into a subword
of a multiword pseudo register, which then messes up register
allocation.  I tried to fix this earlier, and it was another issue where
I feel the patch review process has so far failed.

The patch below replaces the overengineered, broken and disabled
byte-level DCE with a simpler word-level DCE, run from lower-subreg.
Previously I made the mistake of letting myself get influenced by
worries about compile-time performance which are IMO completely
overblown.  There are a number of reasons why compile-time is a
non-issue in this case, never mind the general observation that
compilation is embarrassingly parallel in practice:

* the code is guarded by a test in lower-subreg which verifies that
multiword pseudos exist in the function; that alone makes the
compile-time impact exactly nil for the majority of code
* it's replacing an existing, more expensive pass which was approved
(although later disabled)

Because I took this concern too seriously, the previous patch was a weak
basic-block local DCE that was fast and fixed the testcase, but had no
other value.  This time, I've simply implemented it in the way I think
it should be done, taking the existing byte-level DCE and reducing it to
a fairly simple pass that only operates on two-word pseudos.

This triggers in several testcases other than the one in PR42575, for
example in crafty or the XFS filesystem (tested with an ARM compiler).
This demonstrates that it fixes a real-world, general problem that can
occur for more than one reason.

 ./spec2k/186.crafty/run/00000004/attacks.s
-  0 .text         00000b28  00000000  00000000  00000034  2**2
+  0 .text         00000a40  00000000  00000000  00000034  2**2
 ./spec2k/186.crafty/run/00000004/swap.s
-  0 .text         000005a0  00000000  00000000  00000034  2**2
+  0 .text         00000500  00000000  00000000  00000034  2**2

That seems like a large difference - I've looked at RTL dumps, and from
what I've seen, what it's doing appears sane.

Last time, Eric posted counter-patches which tried hackish ways of
tweaking the generated RTL so that the problem is hidden for the PR42575
testcase; both of these made code generation worse on average and failed
to fix the more general issue of dead stores into subwords.  One of them
contained a useful idea, which is to allow lower-subreg to split regs
even if they are stored in full in a non-decomposable context.  I've
implemented this in a more general way, but it can't be used just yet
because a number of x86 testcases require that certain DImode regs are
_not_ split (e.g. because they are copied from an MMX hard register).
It'll require somewhat more work to add the necessary cost stuff to
lower-subreg.

Hence, I think this new DCE patch should go in instead.  Bootstrapped
and regression tested on i686-linux.  Ok?

Bernd
PR rtl-optimization/42575
	* dce.c (word_dce_process_block): Renamed from byte_dce_process_block.
	Argument AU removed.  All callers changed.  Ignore artificial refs.
	Use return value of df_word_lr_simulate_defs to decide whether an insn
	is necessary.
	(fast_dce): Rename arg to WORD_LEVEL.
	(run_word_dce): Renamed from rest_of_handle_fast_byte_dce.  No longer
	static.
	(pass_fast_rtl_byte_dce): Delete.
	* dce.h (run_word_dce): Declare.
	* df-core.c (df_print_word_regset): Renamed from df_print_byteregset.
	All callers changed.  Simplify code to only deal with two-word regs.
	* df.h (DF_WORD_LR): Renamed from DF_BYTE_LR.
	(DF_WORD_LR_BB_INFO): Renamed from DF_BYTE_LR_BB_INFO.
	(DF_WORD_LR_IN): Renamed from DF_BYTE_LR_IN.
	(DF_WORD_LR_OUT): Renamed from DF_BYTE_LR_OUT.
	(struct df_word_lr_bb_info): Renamed from df_byte_lr_bb_info.
	(df_word_lr_mark_ref): Declare.
	(df_word_lr_add_problem, df_word_lr_mark_ref, df_word_lr_simulate_defs,
	df_word_lr_simulate_uses): Declare or rename from byte variants.
	(df_byte_lr_simulate_artificial_refs_at_top,
	df_byte_lr_simulate_artificial_refs_at_end, df_byte_lr_get_regno_start,
	df_byte_lr_get_regno_len): Delete declarations.
	(df_word_lr_get_bb_info): Rename from df_byte_lr_get_bb_info.
	* df-problems.c (df_word_lr_problem_data): Renamed from
	df_byte_lr_problem_data, all members deleted except for
	WORD_LR_BITMAPS, which is renamed from BYTE_LR_BITMAPS.  Uses changed.
	(df_word_lr_expand_bitmap, df_byte_lr_simulate_artificial_refs_at_top,
	df_byte_lr_simulate_artificial_refs_at_end, df_byte_lr_get_regno_start,
	df_byte_lr_get_regno_len, df_byte_lr_check_regs,
	df_byte_lr_confluence_0): Delete functions.
	(df_word_lr_free_bb_info): Renamed from df_byte_lr_free_bb_info; all
	callers changed.
	(df_word_lr_alloc): Renamed from df_byte_lr_alloc; all callers changed.
	Don't initialize members that were deleted, don't try to discover data
	about registers.  Ignore hard regs.
	(df_word_lr_reset): Renamed from df_byte_lr_reset; all callers changed.
	(df_word_lr_mark_ref): New function.
	(df_word_lr_bb_local_compute): Renamed from
	df_byte_bb_lr_local_compute; all callers changed.  Use
	df_word_lr_mark_ref.  Assert that artificial refs don't include
	pseudos.  Ignore hard registers.
	(df_word_lr_local_compute): Renamed from df_byte_lr_local_compute.
	Assert that exit block uses don't contain pseudos.
	(df_word_lr_init): Renamed from df_byte_lr_init; all callers changed.
	(df_word_lr_confluence_n): Renamed from df_byte_lr_confluence_n; all
	callers changed.  Ignore hard regs.
	(df_word_lr_transfer_function): Renamed from
	df_byte_lr_transfer_function; all callers changed.
	(df_word_lr_free): Renamed from df_byte_lr_free; all callers changed.
	(df_word_lr_top_dump): Renamed from df_byte_lr_top_dump; all callers
	changed.
	(df_word_lr_bottom_dump): Renamed from df_byte_lr_bottom_dump; all
	callers changed.
	(problem_WORD_LR): Renamed from problem_BYTE_LR; uses changed;
	confluence operator 0 set to NULL.
	(df_word_lr_add_problem): Renamed from df_byte_lr_add_problem; all
	callers changed.
	(df_word_lr_simulate_defs): Renamed from df_byte_lr_simulate_defs.
	Return bool, true if bitmap changed or insn otherwise necessary.
	All callers changed.  Simplify using df_word_lr_mark_ref.
	(df_word_lr_simulate_uses): Renamed from df_byte_lr_simulate_uses;
	all callers changed.  Simplify using df_word_lr_mark_ref.
	* lower-subreg.c: Include "dce.h"
	(decompose_multiword_subregs): Call run_word_dce if df available.
	* Makefile.in (lower-subreg.o): Adjust dependencies.
	* timevar.def (TV_DF_WORD_LR): Renamed from TV_DF_BYTE_LR.

New attempt: subword-based DCE, PR42575

Commit Message

Comments

Patch