This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 2/2] S/390: Do not end groups after fallthru edge


> Can't we just set s390_sched_state to s390_last_sched_state in
> s390_sched_init.

Good idea, this simplifies the code quite a bit.

> Preserving the sched state across basic blocks for your case works
> only if the BBs are traversed with the fall through edges coming
> first. Is that the case? We probably should have a description for
> s390_last_sched_state stating this.

According to the documentation the basic blocks should be traversed in
"instruction stream" order, i.e. a fallthru edge will be considered
first as far as I understand it.

Regards
 Robin

--

gcc/ChangeLog:

2017-10-17  Robin Dapp  <rdapp@linux.vnet.ibm.com>

	* config/s390/s390.c (s390_bb_fallthru_entry_likely): New
	function.
	(s390_sched_init): Do not reset s390_sched_state if we entered
	the current basic block via a fallthru edge and all others are
	very unlikely.
diff --git a/gcc/config/s390/s390.c b/gcc/config/s390/s390.c
index c1a144e..dac81bc 100644
--- a/gcc/config/s390/s390.c
+++ b/gcc/config/s390/s390.c
@@ -83,6 +83,7 @@ along with GCC; see the file COPYING3.  If not see
 #include "symbol-summary.h"
 #include "ipa-prop.h"
 #include "ipa-fnsummary.h"
+#include "sched-int.h"
 
 /* This file should be included last.  */
 #include "target-def.h"
@@ -14346,6 +14347,28 @@ s390_z10_prevent_earlyload_conflicts (rtx_insn **ready, int *nready_p)
   ready[0] = tmp;
 }
 
+/* Returns TRUE if BB is entered via a fallthru edge and all other
+   incoming edges are less than unlikely.  */
+static bool
+s390_bb_fallthru_entry_likely (basic_block bb)
+{
+  edge e, fallthru_edge;
+  edge_iterator ei;
+
+  if (!bb)
+    return false;
+
+  fallthru_edge = find_fallthru_edge (bb->preds);
+  if (!fallthru_edge)
+    return false;
+
+  FOR_EACH_EDGE (e, ei, bb->preds)
+    if (e != fallthru_edge
+	&& e->probability >= profile_probability::unlikely ())
+      return false;
+
+  return true;
+}
 
 /* The s390_sched_state variable tracks the state of the current or
    the last instruction group.
@@ -14354,7 +14377,7 @@ s390_z10_prevent_earlyload_conflicts (rtx_insn **ready, int *nready_p)
    3     the last group is complete - normal insns
    4     the last group was a cracked/expanded insn */
 
-static int s390_sched_state;
+static int s390_sched_state = 0;
 
 #define S390_SCHED_STATE_NORMAL  3
 #define S390_SCHED_STATE_CRACKED 4
@@ -14764,7 +14787,17 @@ s390_sched_init (FILE *file ATTRIBUTE_UNUSED,
 {
   last_scheduled_insn = NULL;
   memset (last_scheduled_unit_distance, 0, MAX_SCHED_UNITS * sizeof (int));
-  s390_sched_state = 0;
+
+  /* If the next basic block is most likely entered via a fallthru edge
+     we keep the last sched state.  Otherwise we start a new group.
+     The scheduler traverses basic blocks in "instruction stream" ordering
+     so if we see a fallthru edge here, s390_sched_state will be of its
+     source block.  */
+  rtx_insn *insn = current_sched_info->prev_head
+    ? NEXT_INSN (current_sched_info->prev_head) : NULL;
+  basic_block bb = insn ? BLOCK_FOR_INSN (insn) : NULL;
+  if (!s390_bb_fallthru_entry_likely (bb))
+    s390_sched_state = 0;
 }
 
 /* This target hook implementation for TARGET_LOOP_UNROLL_ADJUST calculates

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]