[PATCH]: Replace backend dataflow COMMITTED
Kenneth Zadeck
zadeck@naturalbridge.com
Wed Jan 11 13:04:00 GMT 2006
Kenneth Zadeck wrote:
> Ian,
>
> The attached patch and ChangeLog combine both patches mentioned here.
> I will wait for three days as you requested.
> thanks,
>
I received no other comments. I updated, bootstrapped and regression
tested again last night on i686-pc-linux-gnu.
Committed
Kenny
> Kenny
>
> Ian Lance Taylor wrote:
>> OK, I finally read through this patch, though I only skimmed most of
>> df-problems.c. I also read Ken's followup which incorporates Zdenek's
>> changes.
>>
>> I recommend that you look at my comments, and adjust the code. Then
>> repost the patch. Then if nobody has any comments for, say, three
>> days, please commit it.
>>
>> Thanks!
>>
>> Ian
>>
>> Daniel Berlin <dberlin@dberlin.org> writes:
>>
>>
>>> +/* Allocation for dataflow support routines.
>>> + Copyright (C) 1999, 2000, 2001, 2002, 2003, 2004, 2005
>>>
>>
>>
> All of the files touched have copyright changes.
> All of the lexigraphical changes asked for were applied uniformly to
> all of the df* files.
>> 2006 now, of course.
>>
>>
>>> +DF_SET_BLOCKS is an optional call used to define a region of the
>>> +program on which the analysis will be performed. The normal case is
>>> +to analyze the entire program and no call to df_set_blocks is made.
>>>
>>
>> Should this say "the entire function" rather than "the entire
>> program?" Or does it really scan the entire program somehow?
>>
>>
> done for all df* files.
>>> +Different optimizations have different needs. Ultimately, only
>>> +register allocation and schedulers should be using the bitmaps and the
>>> +rest of the backend should be upgraded to using and maintaining the
>>> +linked information.
>>
>> What bitmaps does this refer to? Is this paragraph essentially a
>> FIXME comment, or is it trying to say something about the dataflow
>> routines are used?
>>
>>
>>
> The latter. Comment fixed.
>>> +/****************************************************************************/
>>>
>>> +/* Functions to create, destroy and manipulate an instance of
>>> df. */
>>> +/****************************************************************************/
>>>
>>>
>>
>> The usual convention in gcc is not to use this kind of block comment,
>> but to instead simply group things in pages (as you have done). Not a
>> big deal, though.
>>
>>
> As per private comments, changed to match the style used in many tree
> optimizations.
>>> + (*df->problems_in_order[i]->problem->free_fun)
>>> (df->problems_in_order[i]); +
>>> + free (df);
>>> +}
>>>
>>
>> Why don't you have to free the bitmaps? And what about the def and
>> ref information?
>>
>>
> It is up to each problem to release it own bitmaps. This is what is
> done in the problem->free_fun.
>>> + /* Do not do DF_SCAN */
>>>
>>
>> Please put a period at the end of the sentence. Actually I don't
>> understand this comment anyhow.
>>
>>
> the scanning problem is special, and has it's own external calls to
> keep it up to date.
>>> + This will not work properly for RD and RU since these problems
>>> + depend on the bit vectors being ordered in a particular manner and
>>> + that will be too expensive to keep up to data during simple
>>> + modifications.
>>>
>>
>> Is it worth checking for this case?
>>
> I got rid of df_analyze_simple_change_some_blocks,
> df_analyze_simple_change_one_block and df_compact_blocks for now.
> They is used on the dataflow branch, but we have not been happy with
> them and I believe that we can make them completely unnecessary.
> There was a notion that we would try to keep the dataflow up to date
> as the compilation progressed from phase to phase as is done for the
> current dataflow. This has proven to be slower that just recreating
> it as we need it.
>
> The current dataflow (in flow) can be kept up to date during changes
> to the control flow graph because there was no notion that the
> incremental update actually had to produce precise and accurate
> information. If you wanted to insist that the information be precise,
> you had to update it from scratch. We have several functions to do
> this with df. We will wait to add them if they are actually necessary
> for the phase we are working on.
>>
>>> +/* Called from the rtl_compact_blocks to reorganize the problems basic
>>> + block info. */
>>>
>>
>> What does this mean? As far as I can see df_compact_blocks is not
>> called from anywhere?
>>
> ditto.
>>
>>> +void +df_compact_blocks (struct df *df)
>>> +{
>>> + int i, p;
>>> + basic_block bb;
>>> + void ** problem_temps;
>>> + int size = last_basic_block * sizeof (void *);
>>> + problem_temps = xmalloc (size);
>>> +
>>> + for (p = 0; p < df->num_problems_defined; p++)
>>> + {
>>> + struct dataflow * dflow = df->problems_in_order[p];
>>> + if (*dflow->problem->free_bb_fun)
>>> + {
>>> + df_grow_bb_info (dflow);
>>> + memcpy (problem_temps, dflow->block_info, size);
>>> +
>>> + /* Copy the bb info from the problem tmps to the proper
>>> + place in the block_info vector. Null out the copied
>>> + item. */
>>> + i = NUM_FIXED_BLOCKS;
>>> + FOR_EACH_BB (bb) + {
>>> + df_set_bb_info (dflow, i, problem_temps[bb->index]);
>>> + problem_temps[bb->index] = NULL;
>>> + i++;
>>> + }
>>> + memset (dflow->block_info + i, 0, +
>>> (last_basic_block - i) * sizeof (void *));
>>> +
>>> + /* Free any block infos that were not copied (and NULLed).
>>> + These are from orphaned blocks. */
>>> + for (i = NUM_FIXED_BLOCKS; i<last_basic_block; i++)
>>> + {
>>> + if (problem_temps[i])
>>> + (*dflow->problem->free_bb_fun)(dflow, problem_temps[i]);
>>> + }
>>> + }
>>> + }
>>> +
>>> + free (problem_temps);
>>> +
>>> + i = NUM_FIXED_BLOCKS;
>>> + FOR_EACH_BB (bb) + {
>>> + BASIC_BLOCK (i) = bb;
>>> + bb->index = i;
>>> + i++;
>>> + }
>>> +
>>> + gcc_assert (i == n_basic_blocks);
>>> +
>>> + for (; i<last_basic_block; i++)
>>> + BASIC_BLOCK (i) = NULL;
>>> +}
>>>
>>
>> Should we just call the existing compact_blocks routine here?
>>
>>
>>
> deleted, ditto
>>> +/* Create new references of type DF_REF_TYPE for each part of
>>> register REG
>>> + at address LOC within INSN of BB. */
>>> +static void
>>> +df_ref_record (struct dataflow *dflow, rtx reg, rtx *loc,
>>> + basic_block bb, rtx insn, + enum df_ref_type
>>> ref_type, + enum df_ref_flags ref_flags, + bool
>>> record_live)
>>> +{
>>> + unsigned int regno;
>>> + struct df * df = dflow->df;
>>> +
>>> + gcc_assert (REG_P (reg) || GET_CODE (reg) == SUBREG);
>>> +
>>> + /* For the reg allocator we are interested in some SUBREG rtx's,
>>> but not
>>> + all. Notably only those representing a word extraction from a
>>> multi-word
>>> + reg. As written in the docu those should have the form
>>> + (subreg:SI (reg:M A) N), with size(SImode) > size(Mmode).
>>> + XXX Is that true? We could also use the global word_mode
>>> variable. */
>>> + if ((df->flags & DF_SUBREGS) == 0
>>> + && GET_CODE (reg) == SUBREG
>>> + && (GET_MODE_SIZE (GET_MODE (reg)) < GET_MODE_SIZE (word_mode)
>>> + || GET_MODE_SIZE (GET_MODE (reg))
>>> + >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (reg)))))
>>> + {
>>> + loc = &SUBREG_REG (reg);
>>> + reg = *loc;
>>> + ref_flags |= DF_REF_STRIPPED;
>>> + }
>>>
>>
>> I gather that what you are looking for here is cases where a SUBREG is
>> pulling out one (or more) registers from a multi-register value. It's
>> a little more complicated than this, because not all registers have
>> the same size. If you have a hard register here, you should be
>> looking at hard_regno_nregs. If you have a pseudo-register, you
>> should use REGMODE_NATURAL_SIZE.
>>
>> But I see that this is from the existing df.c code. Hmmm. Maybe we
>> can fix this later.
>>
>>
> Danny is out of town over the weekend and needed more info before
> attacking this. We will take you up on your offer to fix this next
> week in a follow up patch.
>>
>>> +/* A set to a non-paradoxical SUBREG for which the number of
>>> word_mode units
>>> + covered by the outer mode is smaller than that covered by the
>>> inner mode,
>>> + is a read-modify-write operation.
>>> + This function returns true iff the SUBREG X is such a SUBREG. */
>>> +bool
>>> +df_read_modify_subreg_p (rtx x)
>>> +{
>>> + unsigned int isize, osize;
>>> + if (GET_CODE (x) != SUBREG)
>>> + return false;
>>> + isize = GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)));
>>> + osize = GET_MODE_SIZE (GET_MODE (x));
>>> + return (isize > osize && isize > UNITS_PER_WORD);
>>> +}
>>>
>>
>> Here again we should probably consider hard_regno_nregs and
>> REGMODE_NATURAL_SIZE.
>>
>>
> We will do this later.
>>> + case ASM_OPERANDS:
>>> + case UNSPEC_VOLATILE:
>>> + case TRAP_IF:
>>> + case ASM_INPUT:
>>> + {
>>> + /* Traditional and volatile asm instructions must be considered
>>> to use
>>> + and clobber all hard registers, all pseudo-registers and all of
>>> + memory. So must TRAP_IF and UNSPEC_VOLATILE operations.
>>> +
>>> + Consider for instance a volatile asm that changes the fpu
>>> rounding
>>> + mode. An insn should not be moved across this even if it
>>> only uses
>>> + pseudo-regs because it might give an incorrectly rounded
>>> result.
>>> +
>>> + For now, just mark any regs we can find in ASM_OPERANDS as
>>> + used. */
>>> +
>>> + /* For all ASM_OPERANDS, we must traverse the vector of input
>>> operands.
>>> + We can not just fall through here since then we would be
>>> confused
>>> + by the ASM_INPUT rtx inside ASM_OPERANDS, which do not indicate
>>> + traditional asms unlike their normal usage. */
>>> + if (code == ASM_OPERANDS)
>>> + {
>>> + int j;
>>> +
>>> + for (j = 0; j < ASM_OPERANDS_INPUT_LENGTH (x); j++)
>>> + df_uses_record (dflow, &ASM_OPERANDS_INPUT (x, j),
>>> + DF_REF_REG_USE, bb, insn, 0);
>>> + return;
>>> + }
>>> + break;
>>>
>>
>> Please add a note to this comment explaining where and how ASMs are
>> handled.
>>
>>
>>> +/* Process INSN recursively, and return true if it contains an asm
>>> + instruction. */
>>> +static bool
>>> +df_insn_contains_asm (rtx *loc, rtx insn)
>>>
>>
>> Please rewrite this function to use for_each_rtx.
>>
>>
> done
>>> + /* The stack ptr is used (honorarily) by a CALL insn. */
>>> + x = df_reg_use_gen (STACK_POINTER_REGNUM);
>>> + df_uses_record (dflow, &XEXP (x, 0), DF_REF_REG_USE, bb,
>>> insn, + 0);
>>>
>> The call to df_reg_use_gen is useless, and just wastes a little bit of
>> memory until the next gc. Just use
>> regno_reg_rtx[STACK_POINTER_REGNUM].
>>
>>
> done
>>> + if (global_regs[i])
>>> + {
>>> + x = df_reg_use_gen (i);
>>> + df_uses_record (dflow, &XEXP (x, 0),
>>> + DF_REF_REG_USE, bb, insn, + 0);
>>>
>>
>> This df_reg_use_gen is also useless.
>>
>>
>>> + EXECUTE_IF_SET_IN_BITMAP (df_invalidated_by_call, 0, ui, bi)
>>> + {
>>> + x = df_reg_def_gen (false, ui);
>>> + df_def_record_1 (dflow, x, bb, insn, DF_REF_CLOBBER, false);
>>> + }
>>>
>>
>> This df_reg_def_gen call is not 100% useless, but it could be cleaned
>> up with a minor change to df_def_record_1.
>>
>>
>>> + if (CALL_P (insn))
>>> + {
>>> + rtx note;
>>> +
>>> + /* We do not record hard registers clobbered by the call,
>>> + since there are awfully many of them and "defs" created
>>> + through them are not interesting (since no use can be legally
>>> + reached by them). So we must just make sure we include
>>> them when
>>> + computing kill bitmaps. */
>>> +
>>> + /* There may be extra registers to be clobbered. */
>>> + for (note = CALL_INSN_FUNCTION_USAGE (insn);
>>> + note;
>>> + note = XEXP (note, 1))
>>> + if (GET_CODE (XEXP (note, 0)) == CLOBBER)
>>> + df_defs_record (dflow, XEXP (note, 0), bb, insn);
>>> + }
>>> + }
>>>
>>
>> I don't follow the large comment here. The code above already called
>> df_defs_record_1 for everything in df_invalidated_by_call. What is it
>> that you are not recording?
>>
>>
> adjusted as per offline comments
>>
>>> +#ifdef EH_USES
>>> + if ((df->flags & DF_HARD_REGS)
>>> + && has_eh_preds (bb))
>>> + {
>>> + unsigned int i;
>>> + for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
>>> + if (EH_USES (i))
>>> + {
>>> + rtx use = df_reg_use_gen (i);
>>> + df_ref_create_structure (dflow, regno_reg_rtx[i],
>>> + &XEXP (use, 0), NULL, EXIT_BLOCK_PTR,
>>> + DF_REF_REG_USE, +
>>> DF_REF_ARTIFICIAL | DF_REF_AT_TOP);
>>> + }
>>> + }
>>> +#endif
>>>
>>
>> I don't think this is right (though maybe I said otherwise on IRC).
>> The only target which actually defines EH_USES is ia64. I think that
>> if EH_USES is set, it means that the register must be valid when an
>> exception is taken. Therefore what matters for EH_USES is not
>> has_eh_preds, but has_eh_succs.
>>
>>
> adjusted as per offline comments.
> ------------------------------------------------------------------------
>
> 2005-01-06 Danny Berlin <dberlin@dberlin.org>
> Kenneth Zadeck <zadeck@naturalbridge.com>
>
> * df.h (DF_SCAN, DF_RU, DF_RD, DF_LR, DF_UR, DF_UREC, DF_CHAIN,
> DF_RI, DF_LAST_PROBLEM_PLUS1, DF_DU_CHAIN, DF_UD_CHAIN,
> DF_REF_TYPE_NAMES, DF_HARD_REGS, DF_EQUIV_NOTES, DF_SUBREGS,
> DF_SCAN_BB_INFO, DF_RU_BB_INFO, DF_RD_BB_INFO, DF_LR_BB_INFO,
> DF_UR_BB_INFO, DF_UREC_BB_INFO, DF_LIVE_IN, DF_LIVE_OUT,
> DF_RA_LIVE_IN, DF_RA_LIVE_OUT, DF_UPWARD_LIVE_IN,
> DF_UPWARD_LIVE_OUT, DF_REF_REAL_REG, DF_REF_REGNO,
> DF_REF_REAL_LOC, DF_REF_REG, DF_REF_LOC, DF_REF_BB, DF_REF_BBNO,
> DF_REF_INSN, DF_REF_INSN_UID, DF_REF_TYPE, DF_REF_CHAIN,
> DF_REF_ID, DF_REF_FLAGS, DF_REF_NEXT_REG, DF_REF_PREV_REG,
> DF_REF_NEXT_REF, DF_REF_DATA, DF_REF_REG_DEF_P, DF_REF_REG_USE_P,
> DF_REF_REG_MEM_STORE_P, DF_REF_REG_MEM_LOAD_P, DF_REF_REG_MEM_P,
> DF_DEFS_SIZE, DF_DEFS_GET, DF_DEFS_SET, DF_USES_SIZE, DF_USES_GET,
> DF_USES_SET, DF_REG_SIZE, DF_REG_DEF_GET, DF_REG_DEF_SET,
> DF_REG_USE_GET, DF_REG_USE_SET, DF_REGNO_FIRST_DEF,
> DF_REGNO_LAST_USE, DF_INSN_SIZE, DF_INSN_GET, DF_INSN_SET,
> DF_INSN_CONTAINS_ASM, DF_INSN_LUID, DF_INSN_DEFS, DF_INSN_USES,
> DF_INSN_UID_GET, DF_INSN_UID_LUID, DF_INSN_UID_DEFS,
> DF_INSN_UID_USES, DF_SCAN_INITIAL, DF_SCAN_GLOBAL,
> DF_SCAN_POST_ALLOC): New macros.
> (df_flow_dir, df_ref_type, df_ref_flags, df_alloc_function,
> df_free_bb_function, df_local_compute_function, df_init_function,
> df_dataflow_function, df_confluence_function_0,
> df_confluence_function_n, df_transfer_function,
> df_finalizer_function, df_free_function, df_dump_problem_function,
> df_problem, dataflow, df_insn_info, df_reg_info, df_ref, df_link,
> df_ref_info, df, df_map, df_scan_bb_info, df_ru_bb_info,
> df_ru_bb_info, df_rd_bb_info, df_lr_bb_info, df_ur_bb_info,
> df_urec_bb_info, ) New types.
> (df_invalidated_by_call, df_all_hard_regs, df_state) New public
> variables.
> (df_init, df_add_problem, df_set_blocks, df_finish, df_analyze,
> df_analyze_simple_change_some_blocks,
> df_analyze_simple_change_one_block, df_compact_blocks,
> df_bb_replace, df_bb_regno_last_use_find,
> df_bb_regno_first_def_find, df_bb_regno_last_def_find,
> df_insn_regno_def_p, df_find_def, df_find_use,
> df_iterative_dataflow, df_dump, df_chain_dump, df_refs_chain_dump,
> df_regs_chain_dump, df_insn_debug, df_insn_debug_regno,
> df_regno_debug, df_ref_debug, debug_df_insn, debug_df_regno,
> debug_df_reg, debug_df_defno, debug_df_useno, debug_df_ref,
> debug_df_chain, df_get_dependent_problem, df_chain_create,
> df_chain_unlink, df_chain_copy, df_get_live_in, df_get_live_out,
> df_grow_bb_info, df_chain_dump, df_print_bb_index,
> df_ru_add_problem, df_ru_get_bb_info, df_rd_add_problem,
> df_rd_get_bb_info, df_lr_add_problem, df_lr_get_bb_info,
> df_ur_add_problem, df_ur_get_bb_info, df_urec_add_problem,
> df_urec_get_bb_info, df_chain_add_problem, df_ri_add_problem,
> df_reg_lifetime, df_scan_get_bb_info, df_scan_add_problem,
> df_rescan_blocks, df_ref_create, df_get_artificial_defs,
> df_get_artificial_uses, df_reg_chain_create, df_reg_chain_unlink,
> df_ref_remove, df_insn_refs_delete, df_refs_delete,
> df_reorganize_refs, df_set_state, df_hard_reg_init,
> df_read_modify_subreg_p) New public functions.
> * df-core.c: The core dataflow solver and glue routines for rtl
> dataflow.
> (df_init, df_add_problem, df_set_blocks, df_finish,
> df_hybrid_search_forward, df_hybrid_search_backward,
> df_iterative_dataflow, df_prune_to_subcfg, df_analyze_problem,
> df_analyze, df_get_bb_info, df_set_bb_info, df_bb_replace,
> df_bb_regno_last_use_find, df_bb_regno_first_def_find,
> df_bb_regno_last_def_find, df_insn_regno_def_p, df_find_def,
> df_reg_defined, df_find_use, df_reg_used, df_dump,
> df_refs_chain_dump, df_regs_chain_dump, df_insn_debug,
> df_insn_debug_regno, df_regno_debug, df_ref_debug, debug_df_insn,
> debug_df_reg, debug_df_regno, debug_df_ref debug_df_defno,
> debug_df_useno, reset_df_after_reload): New functions.
> * df-scan.c: The scanning fuctions, once in df.c, completely
> rewritten so that they now fully model the functionality of
> register usage at the backend.
> (df_scan_free_internal, df_scan_get_bb_info, df_scan_set_bb_info,
> df_scan_free_bb_info, df_scan_alloc, df_scan_free, df_scan_dump,
> df_scan_add_problem, df_grow_reg_info, df_grow_ref_info,
> df_grow_insn_info, df_rescan_blocks, df_ref_create,
> df_get_artificial_defs, df_get_artificial_uses,
> df_reg_chain_create, df_ref_unlink, df_reg_chain_unlink,
> df_ref_remove, df_insn_create_insn_record, df_insn_refs_delete,
> df_refs_delete, df_reorganize_refs, df_set_state,
> df_ref_create_structure, df_ref_record, df_read_modify_subreg_p,
> df_def_record_1, df_defs_record, df_uses_record,
> df_insn_contains_asm_1, df_insn_contains_asm, df_insn_refs_record,
> df_has_eh_preds, df_bb_refs_record, df_refs_record, df_mark_reg,
> df_record_exit_block_uses, df_hard_reg_init): New functions.
>
> * df-problems.c: Seven concrete dataflow problems that use the
> scanning in df-scan.c and are solved by the engine in df-core.c.
> (df_get_dependent_problem, df_chain_create, df_chain_unlink,
> df_chain_copy, df_get_live_in, df_get_live_out, df_grow_bb_info,
> df_chain_dump, df_print_bb_index, df_ref_bitmap, df_set_seen,
> df_unset_seen, df_ru_get_bb_info, df_ru_set_bb_info,
> df_ru_free_bb_info, df_ru_alloc,
> df_ru_bb_local_compute_process_def,
> df_ru_bb_local_compute_process_use, df_ru_bb_local_compute,
> df_ru_local_compute, df_ru_init_solution, df_ru_confluence_n,
> df_ru_transfer_function, df_ru_free, df_ru_dump,
> df_ru_add_problem, df_rd_get_bb_info, df_rd_set_bb_info,
> df_rd_free_bb_info, df_rd_alloc,
> df_rd_bb_local_compute_process_def, df_rd_bb_local_compute,
> df_rd_local_compute, df_rd_init_solution, df_rd_confluence_n,
> df_rd_transfer_function, df_rd_free, df_rd_dump,
> df_rd_add_problem, df_lr_get_bb_info, df_lr_set_bb_info,
> df_lr_free_bb_info, df_lr_alloc, df_lr_bb_local_compute,
> df_lr_local_compute, df_lr_init, df_lr_confluence_0,
> df_lr_confluence_n, df_lr_transfer_function, df_lr_free,
> df_lr_dump, df_lr_add_problem, df_ur_get_bb_info,
> df_ur_set_bb_info, df_ur_free_bb_info, df_ur_alloc,
> df_ur_bb_local_compute, df_ur_local_compute, df_ur_init,
> df_ur_local_finalize, df_ur_confluence_n, df_ur_transfer_function,
> df_ur_free, df_ur_dump, df_ur_add_problem, df_urec_get_bb_info,
> df_urec_set_bb_info, df_urec_free_bb_info, df_urec_alloc,
> df_urec_mark_reg_change, df_urec_check_earlyclobber,
> df_urec_mark_reg_use_for_earlyclobber,
> df_urec_mark_reg_use_for_earlyclobber_1, df_urec_bb_local_compute,
> df_urec_local_compute, df_urec_init, df_urec_local_finalize,
> df_urec_confluence_n, df_urec_transfer_function, df_urec_free,
> df_urec_dump, df_urec_add_problem, df_chain_alloc,
> df_chain_create_bb_process_use, df_chain_create_bb,
> df_chain_finalize, df_chain_free, df_chains_dump,
> df_chain_add_problem, df_ri_alloc, df_ri_bb_compute,
> df_ri_compute, df_ri_free, df_ri_dump, df_ri_add_problem,
> df_reg_lifetime): New functions.
> * df.c: Deleted file.
> * ddg.c (create_ddg_dep_no_link, build_inter_loop_deps): Made code
> consistent with new df api.
> * modulo-sched.c (sms_schedule, rest_of_handle_sms,
> rest_of_handle_sms): Ditto.
> * web.c (unionfind_union, union_defs, entry_register, web_main):
> Ditto.
> * loop_invariant.c (invariant_for_use, hash_invariant_expr_1,
> invariant_expr_equal_p, find_defs, check_dependencies,
> find_invariant_insn, find_invariants_to_move, move_invariant_reg,
> free_inv_motion_data, move_loop_invariants): Ditto.
> * sched-deps.c (sched_analyze_1): Ditto.
>
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dataflow18a.diff.gz
Type: application/x-gzip
Size: 74136 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060111/296e75f6/attachment.bin>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: dataflow18.changelog.gz
Type: application/x-gzip
Size: 2279 bytes
Desc: not available
URL: <http://gcc.gnu.org/pipermail/gcc-patches/attachments/20060111/296e75f6/attachment-0001.bin>
More information about the Gcc-patches
mailing list