[PATCH] Fix RTL DSE compile time hog (PR rtl-optimization/48141)
Richard Guenther
richard.guenther@gmail.com
Wed Mar 16 13:31:00 GMT 2011
On Wed, Mar 16, 2011 at 2:27 PM, Kenneth Zadeck
<zadeck@naturalbridge.com> wrote:
> i think that i sympathize with richard's "do we really need another parm?"
> whine. I would have just nailed it in without the parm. I think that it
> is unreasonable to expect that a program with basic blocks this big needs
> the full range of optimizations performed.
>
> It is perfectly reasonable to have the quality of the optimization degrade
> as a function of size.
>
> If you want to tie the trigger to something, then leave out the parm and tie
> it -O3.
I think having --param limits for this kind of stuff is exactly the right
thing to do. We should try to improve both documentation (list
those params in a separate section) and diagnostics.
Richard.
>
> On 03/16/2011 07:59 AM, Richard Guenther wrote:
>>
>> On Wed, Mar 16, 2011 at 12:52 PM, Jakub Jelinek<jakub@redhat.com> wrote:
>>>
>>> On Wed, Mar 16, 2011 at 09:46:20AM +0100, Paolo Bonzini wrote:
>>>>
>>>> On 03/16/2011 12:12 AM, Kenneth Zadeck wrote:
>>>>>
>>>>> so how much time does this save?
>>>>>
>>>>> I agree that this is a useful simplification, but it seems unlikely to
>>>>> be that important in real code.
>>>>> it seems like the 5000 store test would in general provide a better
>>>>> safety valve.
>>>>
>>>> I think having both is a good idea.
>>>
>>> Here is the second patch, ok for both to trunk if this one passes
>>> bootstrap/regtest?
>>>
>>> With just this patch alone on the testcase with the default
>>> --param=max-dse-active-local-stores=5000 cc1 spends 18.9 seconds
>>> in DSE1+DSE2, with =1000 just 4 seconds and with =10 0.4 seconds.
>>> With the earlier patch in addition to this one the time in DSE1+DSE2
>>> is 0.3 seconds no matter what the parameter is.
>>
>> That suggests we can avoid the new param? (btw, we should start
>> thinking of emitting a diagnostic if we run into such artificial limits)
>>
>> The earlier patch is ok for all branches.
>>
>> Thanks,
>> Richard.
>>
>>> 2011-03-16 Jakub Jelinek<jakub@redhat.com>
>>>
>>> PR rtl-optimization/48141
>>> * params.def (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES): New.
>>> * dse.c: Include params.h.
>>> (active_local_stores_len): New variable.
>>> (add_wild_read, dse_step1): Clear it when setting
>>> active_local_stores
>>> to NULL.
>>> (record_store, check_mem_read_rtx): Decrease it when removing
>>> from the chain.
>>> (scan_insn): Likewise. Increase it when adding to chain, if it
>>> reaches PARAM_MAX_DSE_ACTIVE_LOCAL_STORES limit, set to 1 and
>>> set active_local_stores to NULL before the addition.
>>> * Makefile.in (dse.o): Depend on $(PARAMS_H).
>>>
>>> --- gcc/params.def.jj 2011-02-15 15:42:27.000000000 +0100
>>> +++ gcc/params.def 2011-03-16 12:20:16.000000000 +0100
>>> @@ -698,6 +698,12 @@ DEFPARAM(PARAM_MAX_SCHED_READY_INSNS,
>>> "The maximum number of instructions ready to be issued to be
>>> considered by the scheduler during the first scheduling pass",
>>> 100, 0, 0)
>>>
>>> +/* This is the maximum number of active local stores RTL DSE will
>>> consider. */
>>> +DEFPARAM (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES,
>>> + "max-dse-active-local-stores",
>>> + "Maximum number of active local stores in RTL dead store
>>> elimination",
>>> + 5000, 0, 0)
>>> +
>>> /* Prefetching and cache-optimizations related parameters. Default
>>> values are
>>> usually set by machine description. */
>>>
>>> --- gcc/dse.c.jj 2011-02-15 15:42:26.000000000 +0100
>>> +++ gcc/dse.c 2011-03-16 12:34:18.000000000 +0100
>>> @@ -47,6 +47,7 @@ along with GCC; see the file COPYING3.
>>> #include "optabs.h"
>>> #include "dbgcnt.h"
>>> #include "target.h"
>>> +#include "params.h"
>>>
>>> /* This file contains three techniques for performing Dead Store
>>> Elimination (dse).
>>> @@ -387,6 +388,7 @@ static alloc_pool insn_info_pool;
>>> /* The linked list of stores that are under consideration in this
>>> basic block. */
>>> static insn_info_t active_local_stores;
>>> +static int active_local_stores_len;
>>>
>>> struct bb_info
>>> {
>>> @@ -947,6 +949,7 @@ add_wild_read (bb_info_t bb_info)
>>> }
>>> insn_info->wild_read = true;
>>> active_local_stores = NULL;
>>> + active_local_stores_len = 0;
>>> }
>>>
>>>
>>> @@ -1538,6 +1541,7 @@ record_store (rtx body, bb_info_t bb_inf
>>> {
>>> insn_info_t insn_to_delete = ptr;
>>>
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = ptr->next_local_store;
>>> else
>>> @@ -2074,6 +2078,7 @@ check_mem_read_rtx (rtx *loc, void *data
>>> if (dump_file)
>>> dump_insn_info ("removing from active", i_ptr);
>>>
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = i_ptr->next_local_store;
>>> else
>>> @@ -2163,6 +2168,7 @@ check_mem_read_rtx (rtx *loc, void *data
>>> if (dump_file)
>>> dump_insn_info ("removing from active", i_ptr);
>>>
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = i_ptr->next_local_store;
>>> else
>>> @@ -2222,6 +2228,7 @@ check_mem_read_rtx (rtx *loc, void *data
>>> if (dump_file)
>>> dump_insn_info ("removing from active", i_ptr);
>>>
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = i_ptr->next_local_store;
>>> else
>>> @@ -2426,6 +2433,7 @@ scan_insn (bb_info_t bb_info, rtx insn)
>>> if (dump_file)
>>> dump_insn_info ("removing from active", i_ptr);
>>>
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = i_ptr->next_local_store;
>>> else
>>> @@ -2453,6 +2461,12 @@ scan_insn (bb_info_t bb_info, rtx insn)
>>> fprintf (dump_file, "handling memset as BLKmode
>>> store\n");
>>> if (mems_found == 1)
>>> {
>>> + if (active_local_stores_len++
>>> +>= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES))
>>> + {
>>> + active_local_stores_len = 1;
>>> + active_local_stores = NULL;
>>> + }
>>> insn_info->next_local_store = active_local_stores;
>>> active_local_stores = insn_info;
>>> }
>>> @@ -2496,6 +2510,12 @@ scan_insn (bb_info_t bb_info, rtx insn)
>>> it as cannot delete. This simplifies the processing later. */
>>> if (mems_found == 1)
>>> {
>>> + if (active_local_stores_len++
>>> +>= PARAM_VALUE (PARAM_MAX_DSE_ACTIVE_LOCAL_STORES))
>>> + {
>>> + active_local_stores_len = 1;
>>> + active_local_stores = NULL;
>>> + }
>>> insn_info->next_local_store = active_local_stores;
>>> active_local_stores = insn_info;
>>> }
>>> @@ -2534,6 +2554,7 @@ remove_useless_values (cselib_val *base)
>>>
>>> if (del)
>>> {
>>> + active_local_stores_len--;
>>> if (last)
>>> last->next_local_store = insn_info->next_local_store;
>>> else
>>> @@ -2584,6 +2605,7 @@ dse_step1 (void)
>>> = create_alloc_pool ("cse_store_info_pool",
>>> sizeof (struct store_info), 100);
>>> active_local_stores = NULL;
>>> + active_local_stores_len = 0;
>>> cselib_clear_table ();
>>>
>>> /* Scan the insns. */
>>> --- gcc/Makefile.in.jj 2011-02-02 16:30:50.000000000 +0100
>>> +++ gcc/Makefile.in 2011-03-16 12:26:12.000000000 +0100
>>> @@ -3070,7 +3070,7 @@ dse.o : dse.c $(CONFIG_H) $(SYSTEM_H) co
>>> $(TREE_H) $(TM_P_H) $(REGS_H) hard-reg-set.h $(FLAGS_H) insn-config.h
>>> \
>>> $(RECOG_H) $(EXPR_H) $(DF_H) cselib.h $(DBGCNT_H) $(TIMEVAR_H) \
>>> $(TREE_PASS_H) alloc-pool.h $(ALIAS_H) dse.h $(OPTABS_H) $(TARGET_H) \
>>> - $(BITMAP_H)
>>> + $(BITMAP_H) $(PARAMS_H)
>>> fwprop.o : fwprop.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) $(RTL_H)
>>> \
>>> $(DIAGNOSTIC_CORE_H) insn-config.h $(RECOG_H) $(FLAGS_H) $(OBSTACK_H)
>>> $(BASIC_BLOCK_H) \
>>> output.h $(DF_H) alloc-pool.h $(TIMEVAR_H) $(TREE_PASS_H) $(TARGET_H)
>>> \
>>>
>>>
>>> Jakub
>>>
>
More information about the Gcc-patches
mailing list