From: Richard Earnshaw <rearnsha@arm.com>
Date: Fri, 18 Oct 2019 19:03:03 +0000 (+0000)
Subject: [arm] Reduce cost of insns that are simple reg-reg moves.
X-Git-Url: https://git.libre-soc.org/?a=commitdiff_plain;h=24d28a8778f4c7efcebea6c785806082eeacf9fe;p=gcc.git

[arm] Reduce cost of insns that are simple reg-reg moves.

Consider this sequence during combine:

Trying 18, 7 -> 22:
   18: r118:SI=r122:SI
      REG_DEAD r122:SI
    7: r114:SI=0x1-r118:SI-ltu(cc:CC_RSB,0)
      REG_DEAD r118:SI
      REG_DEAD cc:CC_RSB
   22: r1:SI=r114:SI
      REG_DEAD r114:SI
Failed to match this instruction:
(set (reg:SI 1 r1 [+4 ])
    (minus:SI (geu:SI (reg:CC_RSB 100 cc)
            (const_int 0 [0]))
        (reg:SI 122)))
Successfully matched this instruction:
(set (reg:SI 114)
    (geu:SI (reg:CC_RSB 100 cc)
        (const_int 0 [0])))
Successfully matched this instruction:
(set (reg:SI 1 r1 [+4 ])
    (minus:SI (reg:SI 114)
        (reg:SI 122)))
allowing combination of insns 18, 7 and 22
original costs 4 + 4 + 4 = 12
replacement costs 8 + 4 = 12

The costs are all correct, but we really don't want this combination
to take place.  The original costs contain an insn that is a simple
move of one pseudo register to another and it is extremely likely that
register allocation will eliminate this insn entirely.  On the other
hand, the resulting sequence really does expand into a sequence that
costs 12 (ie 3 insns).

We don't want to prevent combine from eliminating such moves, as this
can expose more combine opportunities, but we shouldn't rate them as
profitable in themselves.  We can do this be adjusting the costs
slightly so that the benefit of eliminating such a simple insn is
reduced.

We only do this before register allocation; after allocation we give
such insns their full cost.

	* config/arm/arm.c (arm_insn_cost): New function.
	(TARGET_INSN_COST): Override default definition.

From-SVN: r277174
---

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index e40511bf148..91b80e52f50 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,8 @@
+2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
+
+	* config/arm/arm.c (arm_insn_cost): New function.
+	(TARGET_INSN_COST): Override default definition.
+
 2019-10-18  Richard Earnshaw  <rearnsha@arm.com>
 
 	* config/arm/arm.c (arm_rtx_costs_internal, case MINUS): Handle
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index fd742226f16..bdea81d29d4 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -181,6 +181,7 @@ static bool arm_have_conditional_execution (void);
 static bool arm_cannot_force_const_mem (machine_mode, rtx);
 static bool arm_legitimate_constant_p (machine_mode, rtx);
 static bool arm_rtx_costs (rtx, machine_mode, int, int, int *, bool);
+static int arm_insn_cost (rtx_insn *, bool);
 static int arm_address_cost (rtx, machine_mode, addr_space_t, bool);
 static int arm_register_move_cost (machine_mode, reg_class_t, reg_class_t);
 static int arm_memory_move_cost (machine_mode, reg_class_t, bool);
@@ -510,6 +511,8 @@ static const struct attribute_spec arm_attribute_table[] =
 #define TARGET_RTX_COSTS arm_rtx_costs
 #undef  TARGET_ADDRESS_COST
 #define TARGET_ADDRESS_COST arm_address_cost
+#undef TARGET_INSN_COST
+#define TARGET_INSN_COST arm_insn_cost
 
 #undef TARGET_SHIFT_TRUNCATION_MASK
 #define TARGET_SHIFT_TRUNCATION_MASK arm_shift_truncation_mask
@@ -11486,6 +11489,24 @@ arm_rtx_costs (rtx x, machine_mode mode ATTRIBUTE_UNUSED, int outer_code,
   return result;
 }
 
+static int
+arm_insn_cost (rtx_insn *insn, bool speed)
+{
+  int cost;
+
+  /* Don't cost a simple reg-reg move at a full insn cost: such moves
+     will likely disappear during register allocation.  */
+  if (!reload_completed
+      && GET_CODE (PATTERN (insn)) == SET
+      && REG_P (SET_DEST (PATTERN (insn)))
+      && REG_P (SET_SRC (PATTERN (insn))))
+    return 2;
+  cost = pattern_cost (PATTERN (insn), speed);
+  /* If the cost is zero, then it's likely a complex insn.  We don't want the
+     cost of these to be less than something we know about.  */
+  return cost ? cost : COSTS_N_INSNS (2);
+}
+
 /* All address computations that can be done are free, but rtx cost returns
    the same for practically all of them.  So we weight the different types
    of address here in the order (most pref first):