diff mbox

add reassociation width target function for power8

Message ID 1462378812.7759.114.camel@linux.vnet.ibm.com
State New
Headers show

Commit Message

Aaron Sawdey May 4, 2016, 4:20 p.m. UTC 
Hi,

This patch enables TARGET_SCHED_REASSOCIATION_WIDTH for power8 and up.
The widths returned are derived from testing with SPEC 2006 and some
simple tests on power8.

Bootstrapped and regtested on powerpc64le-unknown-linux-gnu, ok for
trunk?

2016-05-04  Aaron Sawdey <acsawdey@linux.vnet.ibm.com>

        * config/rs6000/rs6000.c (rs6000_reassociation_width): Add
        function for TARGET_SCHED_REASSOCIATION_WIDTH to enable
        parallel reassociation for power8 and forward.

Comments

Segher Boessenkool May 9, 2016, 4:09 p.m. UTC | #1 
Hi Aaron,

On Wed, May 04, 2016 at 11:20:12AM -0500, Aaron Sawdey wrote:
> This patch enables TARGET_SCHED_REASSOCIATION_WIDTH for power8 and up.
> The widths returned are derived from testing with SPEC 2006 and some
> simple tests on power8.
> 
> Bootstrapped and regtested on powerpc64le-unknown-linux-gnu, ok for
> trunk?
> 
> 2016-05-04  Aaron Sawdey <acsawdey@linux.vnet.ibm.com>

Two spaces before the email address.

>         * config/rs6000/rs6000.c (rs6000_reassociation_width): Add
>         function for TARGET_SCHED_REASSOCIATION_WIDTH to enable
>         parallel reassociation for power8 and forward.

This is okay for trunk.  Thanks,


Segher
diff mbox

Patch

Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c  (revision 235841)
+++ gcc/config/rs6000/rs6000.c  (working copy)
@@ -1755,6 +1755,9 @@ 
 #undef TARGET_CONDITIONAL_REGISTER_USAGE
 #define TARGET_CONDITIONAL_REGISTER_USAGE rs6000_conditional_register_usage

+#undef TARGET_SCHED_REASSOCIATION_WIDTH
+#define TARGET_SCHED_REASSOCIATION_WIDTH rs6000_reassociation_width
+
 #undef TARGET_TRAMPOLINE_INIT
 #define TARGET_TRAMPOLINE_INIT rs6000_trampoline_init

@@ -8633,6 +8636,40 @@ 
                                             true, worst_case);
 }

+/* Determine the reassociation width to be used in reassociate_bb.
+   This takes into account how many parallel operations we
+   can actually do of a given type, and also the latency.
+   P8:
+     int add/sub 6/cycle
+         mul 2/cycle
+     vect add/sub/mul 2/cycle
+     fp   add/sub/mul 2/cycle
+     dfp  1/cycle
+*/
+
+static int
+rs6000_reassociation_width (unsigned int opc ATTRIBUTE_UNUSED,
+                            enum machine_mode mode)
+{
+  switch (rs6000_cpu)
+    {
+    case PROCESSOR_POWER8:
+    case PROCESSOR_POWER9:
+      if (DECIMAL_FLOAT_MODE_P (mode))
+       return 1;
+      if (VECTOR_MODE_P (mode))
+       return 4;
+      if (INTEGRAL_MODE_P (mode))
+       return opc == MULT_EXPR ? 4 : 6;
+      if (FLOAT_MODE_P (mode))
+       return 4;
+      break;
+    default:
+      break;
+    }
+  return 1;
+}
+
 /* Change register usage conditional on target flags.  */
 static void
 rs6000_conditional_register_usage (void)