On Sat, Sep 29, 2012 at 10:26 PM, Steven Bosscher <[email protected]> wrote:
> Hi Vlad,
>
> Thanks for the testing and the logs. You must have good hardware, your
> timings are all ~3 times faster than mine :-)
>
> On Sat, Sep 29, 2012 at 3:01 AM, Vladimir Makarov <[email protected]> wrote:
>> ----------------------------------32-bit------------------------------------
>> Reload:
>> 581.85user 29.91system 27:15.18elapsed 37%CPU (0avgtext+0avgdata
>> LRA:
>> 629.67user 24.16system 24:31.08elapsed 44%CPU (0avgtext+0avgdata
>
> This is a ~8% slowdown.
>
>
>> ----------------------------------64-bit:-----------------------------------
>> Reload:
>> 503.26user 36.54system 30:16.62elapsed 29%CPU (0avgtext+0avgdata
>> LRA:
>> 598.70user 30.90system 27:26.92elapsed 38%CPU (0avgtext+0avgdata
>
> This is a ~19% slowdown
I think both measurements run into swap (low CPU utilization), from the LRA
numbers I'd say that LRA uses less memory but the timings are somewhat
useless with the swapping.
>> Here is the numbers for PR54146 on the same machine with -O1 only for
>> 64-bit (compiler reports error for -m32).
>
> Right, the test case is for 64-bits only, I think it's preprocessed
> code for AMD64.
>
>> Reload:
>> 350.40user 21.59system 17:09.75elapsed 36%CPU (0avgtext+0avgdata
>> LRA:
>> 468.29user 21.35system 15:47.76elapsed 51%CPU (0avgtext+0avgdata
>
> This is a ~34% slowdown.
>
> To put it in another perspective, here are my timings of trunk vs lra
> (both checkouts done today):
>
> trunk:
> integrated RA : 181.68 (24%) usr 1.68 (11%) sys 183.43
> (24%) wall 643564 kB (20%) ggc
> reload : 11.00 ( 1%) usr 0.18 ( 1%) sys 11.17 (
> 1%) wall 32394 kB ( 1%) ggc
> TOTAL : 741.64 14.76 756.41
> 3216164 kB
>
> lra branch:
> integrated RA : 174.65 (16%) usr 1.33 ( 8%) sys 176.33
> (16%) wall 643560 kB (20%) ggc
> reload : 399.69 (36%) usr 2.48 (15%) sys 402.69
> (36%) wall 41852 kB ( 1%) ggc
> TOTAL :1102.06 16.05 1120.83
> 3231738 kB
>
> That's a 49% slowdown. The difference is completely accounted for by
> the timing difference between reload and LRA.
> (Timings done on gcc17, which is AMD Opteron(tm) Processor 8354 with
> 15GB ram, so swapping is no issue.)
>
> It looks like the reload timevar is used for LRA. Why not have
> multiple timevars, one per phase of LRA? Sth like the patch below
> would be nice. This gives me the following timings:
>
> integrated RA : 189.34 (16%) usr 1.84 (11%) sys 191.18
> (16%) wall 643560 kB (20%) ggc
> LRA non-specific : 59.82 ( 5%) usr 0.22 ( 1%) sys 60.12 (
> 5%) wall 18202 kB ( 1%) ggc
> LRA virtuals eliminatenon: 56.79 ( 5%) usr 0.03 ( 0%) sys 56.80 (
> 5%) wall 19223 kB ( 1%) ggc
> LRA reload inheritance : 6.41 ( 1%) usr 0.01 ( 0%) sys 6.42 (
> 1%) wall 1665 kB ( 0%) ggc
> LRA create live ranges : 175.30 (15%) usr 2.14 (13%) sys 177.44
> (15%) wall 2761 kB ( 0%) ggc
> LRA hard reg assignment : 130.85 (11%) usr 0.20 ( 1%) sys 131.17
> (11%) wall 0 kB ( 0%) ggc
> LRA coalesce pseudo regs: 2.54 ( 0%) usr 0.00 ( 0%) sys 2.55 (
> 0%) wall 0 kB ( 0%) ggc
> reload : 6.73 ( 1%) usr 0.20 ( 1%) sys 6.92 (
> 1%) wall 0 kB ( 0%) ggc
>
> so the LRA "slowness" (for lack of a better word) appears to be due to
> scalability problems in all sub-passes.
It would be nice to see if LRA just has a larger constant cost factor
compared to reload or if it has bigger complexity.
> The code size changes are impressive, but I think that this kind of
> slowdown should be addressed before making LRA the default for any
> target.
Certainly if it shows bigger complexity, not sure for the constant factor
(but for sure improvements are welcome).
I suppose there is the option to revert back to reload by default for
x86_64 as well for 4.8, right? That is, do both reload and LRA
co-exist for each target or is it a definite decision target by target?
Thanks,
Richard.
> Ciao!
> Steven
>
>
>
>
> Index: lra-assigns.c
> ===================================================================
> --- lra-assigns.c (revision 191858)
> +++ lra-assigns.c (working copy)
> @@ -1261,6 +1261,8 @@ lra_assign (void)
> bitmap_head insns_to_process;
> bool no_spills_p;
>
> + timevar_push (TV_LRA_ASSIGN);
> +
> init_lives ();
> sorted_pseudos = (int *) xmalloc (sizeof (int) * max_reg_num ());
> sorted_reload_pseudos = (int *) xmalloc (sizeof (int) * max_reg_num ());
> @@ -1312,5 +1314,6 @@ lra_assign (void)
> free (sorted_pseudos);
> free (sorted_reload_pseudos);
> finish_lives ();
> + timevar_pop (TV_LRA_ASSIGN);
> return no_spills_p;
> }
> Index: lra.c
> ===================================================================
> --- lra.c (revision 191858)
> +++ lra.c (working copy)
> @@ -2193,6 +2193,7 @@ lra (FILE *f)
>
> lra_dump_file = f;
>
> + timevar_push (TV_LRA);
>
> init_insn_recog_data ();
>
> @@ -2271,6 +2272,7 @@ lra (FILE *f)
> to use a constant pool. */
> lra_eliminate (false);
> lra_inheritance ();
> +
> /* We need live ranges for lra_assign -- so build them. */
> lra_create_live_ranges (true);
> live_p = true;
> @@ -2343,6 +2345,8 @@ lra (FILE *f)
> #ifdef ENABLE_CHECKING
> check_rtl (true);
> #endif
> +
> + timevar_pop (TV_LRA);
> }
>
> /* Called once per compiler to initialize LRA data once. */
> Index: lra-eliminations.c
> ===================================================================
> --- lra-eliminations.c (revision 191858)
> +++ lra-eliminations.c (working copy)
> @@ -1297,6 +1297,8 @@ lra_eliminate (bool final_p)
> struct elim_table *ep;
> int regs_num = max_reg_num ();
>
> + timevar_push (TV_LRA_ELIMINATE);
> +
> bitmap_initialize (&insns_with_changed_offsets, ®_obstack);
> if (final_p)
> {
> @@ -1317,7 +1319,7 @@ lra_eliminate (bool final_p)
> {
> update_reg_eliminate (&insns_with_changed_offsets);
> if (bitmap_empty_p (&insns_with_changed_offsets))
> - return;
> + goto lra_eliminate_done;
> }
> if (lra_dump_file != NULL)
> {
> @@ -1349,4 +1351,7 @@ lra_eliminate (bool final_p)
> process_insn_for_elimination (insn, final_p);
> }
> bitmap_clear (&insns_with_changed_offsets);
> +
> +lra_eliminate_done:
> + timevar_pop (TV_LRA_ELIMINATE);
> }
> Index: lra-lives.c
> ===================================================================
> --- lra-lives.c (revision 191858)
> +++ lra-lives.c (working copy)
> @@ -962,6 +962,8 @@ lra_create_live_ranges (bool all_p)
> basic_block bb;
> int i, hard_regno, max_regno = max_reg_num ();
>
> + timevar_push (TV_LRA_CREATE_LIVE_RANGES);
> +
> complete_info_p = all_p;
> if (lra_dump_file != NULL)
> fprintf (lra_dump_file,
> @@ -1016,6 +1018,7 @@ lra_create_live_ranges (bool all_p)
> sparseset_free (pseudos_live_through_setjumps);
> sparseset_free (pseudos_live);
> compress_live_ranges ();
> + timevar_pop (TV_LRA_CREATE_LIVE_RANGES);
> }
>
> /* Finish all live ranges. */
> Index: timevar.def
> ===================================================================
> --- timevar.def (revision 191858)
> +++ timevar.def (working copy)
> @@ -223,10 +223,16 @@ DEFTIMEVAR (TV_REGMOVE , "
> DEFTIMEVAR (TV_MODE_SWITCH , "mode switching")
> DEFTIMEVAR (TV_SMS , "sms modulo scheduling")
> DEFTIMEVAR (TV_SCHED , "scheduling")
> -DEFTIMEVAR (TV_IRA , "integrated RA")
> -DEFTIMEVAR (TV_RELOAD , "reload")
> +DEFTIMEVAR (TV_IRA , "integrated RA")
> +DEFTIMEVAR (TV_LRA , "LRA non-specific")
> +DEFTIMEVAR (TV_LRA_ELIMINATE , "LRA virtuals eliminatenon")
> +DEFTIMEVAR (TV_LRA_INHERITANCE , "LRA reload inheritance")
> +DEFTIMEVAR (TV_LRA_CREATE_LIVE_RANGES, "LRA create live ranges")
> +DEFTIMEVAR (TV_LRA_ASSIGN , "LRA hard reg assignment")
> +DEFTIMEVAR (TV_LRA_COALESCE , "LRA coalesce pseudo regs")
> +DEFTIMEVAR (TV_RELOAD , "reload")
> DEFTIMEVAR (TV_RELOAD_CSE_REGS , "reload CSE regs")
> -DEFTIMEVAR (TV_GCSE_AFTER_RELOAD , "load CSE after reload")
> +DEFTIMEVAR (TV_GCSE_AFTER_RELOAD , "load CSE after reload")
> DEFTIMEVAR (TV_REE , "ree")
> DEFTIMEVAR (TV_THREAD_PROLOGUE_AND_EPILOGUE, "thread pro- & epilogue")
> DEFTIMEVAR (TV_IFCVT2 , "if-conversion 2")
> Index: lra-coalesce.c
> ===================================================================
> --- lra-coalesce.c (revision 191858)
> +++ lra-coalesce.c (working copy)
> @@ -221,6 +221,8 @@ lra_coalesce (void)
> bitmap_head involved_insns_bitmap, split_origin_bitmap;
> bitmap_iterator bi;
>
> + timevar_push (TV_LRA_COALESCE);
> +
> if (lra_dump_file != NULL)
> fprintf (lra_dump_file,
> "\n********** Pseudos coalescing #%d: **********\n\n",
> @@ -371,5 +373,6 @@ lra_coalesce (void)
> free (sorted_moves);
> free (next_coalesced_pseudo);
> free (first_coalesced_pseudo);
> + timevar_pop (TV_LRA_COALESCE);
> return coalesced_moves != 0;
> }
> Index: lra-constraints.c
> ===================================================================
> --- lra-constraints.c (revision 191858)
> +++ lra-constraints.c (working copy)
> @@ -4859,6 +4859,8 @@ lra_inheritance (void)
> basic_block bb, start_bb;
> edge e;
>
> + timevar_push (TV_LRA_INHERITANCE);
> +
> lra_inheritance_iter++;
> if (lra_dump_file != NULL)
> fprintf (lra_dump_file, "\n********** Inheritance #%d: **********\n\n",
> @@ -4907,6 +4909,8 @@ lra_inheritance (void)
> bitmap_clear (&live_regs);
> bitmap_clear (&check_only_regs);
> free (usage_insns);
> +
> + timevar_pop (TV_LRA_INHERITANCE);
> }
>
> ^L
