one more patch for PR64317

one more patch for PR64317

[Vladimir Makarov]

[-- Attachment #1: Type: text/plain, Size: 1039 bytes --]

   The following patch improves inheritance for PR64317 testcase

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64317

   I ran a lot SPEC2000 benchmarks to get better default parameter value 
for EBB_PROBABILITY_CUTOFF in LRA inheritance. The new default parameter 
value improves SPECInt2000 by 0.4% on x86-64 without changing SPECFP2000 
rate.  The code size changes are insignificant (0.002% increase for 
SPECInt and 0.01% decrease for SPECFP).

The patch was bootstrapped and tested on x86-64.

Committed as rev.221070.

2015-02-27  Vladimir Makarov  <vmakarov@redhat.com>

         PR target/64317
         * params.def (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF): New.
         * params.h (LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF): New.
         * lra-constraints.c (EBB_PROBABILITY_CUTOFF): Use
         LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF.
         (lra_inheritance): Use '<' instead of '<=' for
         EBB_PROBABILITY_CUTOFF.
         * doc/invoke.texi (lra-inheritance-ebb-probability-cutoff):
         Document change.


[-- Attachment #2: pr64317-2.patch --]
[-- Type: text/plain, Size: 3364 bytes --]

Index: params.def
===================================================================
--- params.def	(revision 220916)
+++ params.def	(working copy)
@@ -836,6 +836,11 @@ DEFPARAM (PARAM_LRA_MAX_CONSIDERED_RELOA
 	  "The max number of reload pseudos which are considered during spilling a non-reload pseudo",
 	  500, 0, 0)
 
+DEFPARAM (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF,
+	  "lra-inheritance-ebb-probability-cutoff",
+	  "Minimal fall-through edge probability in percentage used to add BB to inheritance EEB in LRA",
+	  40, 0, 100)
+
 /* Switch initialization conversion will refuse to create arrays that are
    bigger than this parameter times the number of switch branches.  */
 
Index: params.h
===================================================================
--- params.h	(revision 220916)
+++ params.h	(working copy)
@@ -202,6 +202,8 @@ extern void init_param_values (int *para
   PARAM_VALUE (PARAM_IRA_LOOP_RESERVED_REGS)
 #define LRA_MAX_CONSIDERED_RELOAD_PSEUDOS \
   PARAM_VALUE (PARAM_LRA_MAX_CONSIDERED_RELOAD_PSEUDOS)
+#define LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF \
+  PARAM_VALUE (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF)
 #define SWITCH_CONVERSION_BRANCH_RATIO \
   PARAM_VALUE (PARAM_SWITCH_CONVERSION_BRANCH_RATIO)
 #define LOOP_INVARIANT_MAX_BBS_IN_LOOP \
Index: lra-constraints.c
===================================================================
--- lra-constraints.c	(revision 220916)
+++ lra-constraints.c	(working copy)
@@ -154,6 +154,7 @@
 #include "df.h"
 #include "ira.h"
 #include "rtl-error.h"
+#include "params.h"
 #include "lra-int.h"
 
 /* Value of LRA_CURR_RELOAD_NUM at the beginning of BB of the current
@@ -5694,7 +5695,8 @@ inherit_in_ebb (rtx_insn *head, rtx_insn
 /* This value affects EBB forming.  If probability of edge from EBB to
    a BB is not greater than the following value, we don't add the BB
    to EBB.  */
-#define EBB_PROBABILITY_CUTOFF ((REG_BR_PROB_BASE * 50) / 100)
+#define EBB_PROBABILITY_CUTOFF \
+  ((REG_BR_PROB_BASE * LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF) / 100)
 
 /* Current number of inheritance/split iteration.  */
 int lra_inheritance_iter;
@@ -5740,7 +5742,7 @@ lra_inheritance (void)
 	  e = find_fallthru_edge (bb->succs);
 	  if (! e)
 	    break;
-	  if (e->probability <= EBB_PROBABILITY_CUTOFF)
+	  if (e->probability < EBB_PROBABILITY_CUTOFF)
 	    break;
 	  bb = bb->next_bb;
 	}
Index: doc/invoke.texi
===================================================================
--- doc/invoke.texi	(revision 220916)
+++ doc/invoke.texi	(working copy)
@@ -10570,6 +10570,14 @@ by this parameter.  The default value of
 the minimal number of registers needed by typical instructions.
 This value is the best found from numerous experiments.
 
+@item lra-inheritance-ebb-probability-cutoff
+LRA tries to reuse values reloaded in registers in subsequent insns.
+This optimization is called inheritance.  EBB is used as a region to
+do this optimization.  The parameter defines a minimal fall-through
+edge probability in percentage used to add BB to inheritance EBB in
+LRA.  The default value of the parameter is 40.  The value was chosen
+from numerous runs of SPEC2000 on x86-64.
+
 @item loop-invariant-max-bbs-in-loop
 Loop invariant motion can be very expensive, both in compilation time and
 in amount of needed compile-time memory, with very large loops.  Loops

Re: one more patch for PR64317

[Bernhard Reutner-Fischer]

On February 27, 2015 11:03:14 PM GMT+01:00, Vladimir Makarov <vmakarov@redhat.com> wrote:
>   The following patch improves inheritance for PR64317 testcase
>
>https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64317
>
> I ran a lot SPEC2000 benchmarks to get better default parameter value 
>for EBB_PROBABILITY_CUTOFF in LRA inheritance. The new default
>parameter 
>value improves SPECInt2000 by 0.4% on x86-64 without changing
>SPECFP2000 
>rate.  The code size changes are insignificant (0.002% increase for 
>SPECInt and 0.01% decrease for SPECFP).
>
>The patch was bootstrapped and tested on x86-64.
>
>Committed as rev.221070.


+DEFPARAM (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF,
+	  "lra-inheritance-ebb-probability-cutoff",
+	  "Minimal fall-through edge probability in percentage used to add BB to inheritance EEB in LRA",
+	  40, 0, 100)

s/EEB/EBB/
?
Thanks,

Re: one more patch for PR64317

[Vladimir Makarov]

On 2015-02-27 5:30 PM, Bernhard Reutner-Fischer wrote:
> On February 27, 2015 11:03:14 PM GMT+01:00, Vladimir Makarov <vmakarov@redhat.com> wrote:
>> Committed as rev.221070.
>
> +DEFPARAM (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF,
> +	  "lra-inheritance-ebb-probability-cutoff",
> +	  "Minimal fall-through edge probability in percentage used to add BB to inheritance EEB in LRA",
> +	  40, 0, 100)
>
> s/EEB/EBB/
>
>
Fixed.  Thanks for pointing this out.

Re: one more patch for PR64317

[Sandra Loosemore]

On 02/27/2015 03:03 PM, Vladimir Makarov wrote:
>    The following patch improves inheritance for PR64317 testcase
>
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64317
>
> [snip]
> Index: params.def
> ===================================================================
> --- params.def	(revision 220916)
> +++ params.def	(working copy)
> @@ -836,6 +836,11 @@ DEFPARAM (PARAM_LRA_MAX_CONSIDERED_RELOA
>  	  "The max number of reload pseudos which are considered during spilling a non-reload pseudo",
>  	  500, 0, 0)
>
> +DEFPARAM (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF,
> +	  "lra-inheritance-ebb-probability-cutoff",
> +	  "Minimal fall-through edge probability in percentage used to add BB to inheritance EEB in LRA",

s/EEB/EBB/?

I can't say I understand what either of those abbreviations means....

-Sandra

Re: one more patch for PR64317

[Vladimir Makarov]

On 2015-02-27 8:06 PM, Sandra Loosemore wrote:
> On 02/27/2015 03:03 PM, Vladimir Makarov wrote:
>>    The following patch improves inheritance for PR64317 testcase
>>
>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64317
>>
>> [snip]
>> Index: params.def
>> ===================================================================
>> --- params.def    (revision 220916)
>> +++ params.def    (working copy)
>> @@ -836,6 +836,11 @@ DEFPARAM (PARAM_LRA_MAX_CONSIDERED_RELOA
>>        "The max number of reload pseudos which are considered during 
>> spilling a non-reload pseudo",
>>        500, 0, 0)
>>
>> +DEFPARAM (PARAM_LRA_INHERITANCE_EBB_PROBABILITY_CUTOFF,
>> +      "lra-inheritance-ebb-probability-cutoff",
>> +      "Minimal fall-through edge probability in percentage used to 
>> add BB to inheritance EEB in LRA",
>
> s/EEB/EBB/?
>
> I can't say I understand what either of those abbreviations means....
>
EBB is an extended basic block.
EEB is a typo.
Sorry for using the abbreviation but it is a way to keep attribute name 
shorter (although it is already too long).