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 <acsaw...@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.
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)
--
Aaron Sawdey, Ph.D. acsaw...@linux.vnet.ibm.com
050-2/C113 (507) 253-7520 home: 507/263-0782
IBM Linux Technology Center - PPC Toolchain
