Split postreload gcse from gcse.c

[Steven Bosscher]

And of course, I attached the wrong patch.  The larger the patch,
the more likely it is you attach the wrong one.
So again,

On Wednesday 18 August 2004 18:08, Steven Bosscher wrote:
> Hi,
>
> This patch splits postreload gcse from gcse.c to a new file
> called (surprise) postreload-gcse.c.
>
> To avoid having yet another pair of hash/equivalence functions,
> I've merged expr_hash and expr_equiv from gcse.c with canon_hash
> and exp_equiv from cse.c
>
> I've bootstrapped it on a number of targets, as have others, and
> we've also looked at performance.  This patch makes postreload
> gcse a lot faster, but the generated code is not better overall,
> unfortunately.
>
> For example, on x86 most test cases improve slightly, but eon
> takes a 7% hit.  On PowerPC, for which this pass was originally
> contributed, there's also little or no overall improvement due to
> postreload gcse, neither for compile time nor for the generated
> code.
>
> So what's left with this patch is:
> - splitting it off from gcse.c, which is a nice cleanup, and
> - merging a bunch of almost-duplicate functions.
>
> I've just finished bootstrapping the patch (needed re-doing after
> rth's RTX_UNCHANGING_P patch) and testing is in progress.
> OK if it passes?
>
> Gr.
> Steven
>
>
>
>
>         * Makefile.in (OBJS-common): Add postreload-gcse.c.
>         Add new postreload-gcse.o.
>         * cse.c (SAFE_HASH): Define as wrapper around safe_hash.
>         (lookup_as_function, insert, rehash_using_reg, use_related_value,
>         equiv_constant): Use SAFE_HASH instead of safe_hash.
>         (exp_equiv_p): Export.  Add for_gcse argument when comparing
>         for GCSE.
>         (lookup, lookup_for_remove, merge_equiv_classes, find_best_addr,
>         find_comparison_args, fold_rtx, cse_insn): Update callers.
>         (hash_rtx): New function derived from old canon_hash and bits
>         from gcse.c hash_expr_1.
>         (canon_hash_string): Rename to hash_rtx_string.
>         (canon_hash, safe_hash): Make static inline.  Call hash_rtx.
>         * cselib.c (hash_rtx): Rename to cselib_hash_rtx.
>         (cselib_lookup): Update this caller.
>         * gcse.c (modify_mem_list_set, canon_modify_mem_list_set):
>         Make static.
>         (hash_expr): Call hash_rtx.
>         (ldst_entry): Likewise.
>         (expr_equiv_p): Call exp_equiv_p.
>         (struct unoccr, hash_expr_1, hash_string_1, lookup_expr,
>         reg_used_on_edge, reg_set_between_after_reload_p,
>         reg_used_between_after_reload_p, get_avail_load_store_reg,
>         is_jump_table_basic_block, bb_has_well_behaved_predecessors,
>         get_bb_avail_insn, hash_scan_set_after_reload,
>         compute_hash_table_after_reload,
>         eliminate_partially_redundant_loads, gcse_after_reload,
>         get_bb_avail_insn, gcse_after_reload_main): Remove.
>         * postreload-gcse.c: New file, reincarnating most of the above.
>         * rtl.h (exp_equiv_p, hash_rtx): New prototypes.
>         (gcse_after_reload_main): Update prototype.
>         * timevar.def (TV_GCSE_AFTER_RELOAD): New timevar.
>         * passes.c (rest_of_handle_gcse2): Use it.


Gr.
Steven
-------------- next part --------------
A non-text attachment was scrubbed...
Name: postreload-gcse.diff
Type: text/x-diff
Size: 124820 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20040818/898a4074/attachment.bin>