[gcc r12-2946] Improve autoprefetcher heuristic (partly fix regression in PR91598)

[gcc r12-2946] Improve autoprefetcher heuristic (partly fix regression in PR91598)

[Maxim Kuvyrkov]

https://gcc.gnu.org/g:6d527883072ce96a33169036fca7740172223b52

commit r12-2946-g6d527883072ce96a33169036fca7740172223b52
Author: Maxim Kuvyrkov <maxim.kuvyrkov@linaro.org>
Date:   Thu Aug 29 15:21:36 2019 +0000

    Improve autoprefetcher heuristic (partly fix regression in PR91598)
    
            PR rtl-optimization/91598
            * haifa-sched.c (autopref_rank_for_schedule): Prioritize "irrelevant"
            insns after memory reads and before memory writes.

Diff:
---
 gcc/haifa-sched.c | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/gcc/haifa-sched.c b/gcc/haifa-sched.c
index 26d11279551..5048dd4a5a3 100644
--- a/gcc/haifa-sched.c
+++ b/gcc/haifa-sched.c
@@ -5683,9 +5683,16 @@ autopref_rank_for_schedule (const rtx_insn *insn1, const rtx_insn *insn2)
       int irrel2 = data2->status == AUTOPREF_MULTIPASS_DATA_IRRELEVANT;
 
       if (!irrel1 && !irrel2)
+	/* Sort memory references from lowest offset to the largest.  */
 	r = data1->offset - data2->offset;
-      else
+      else if (write)
+	/* Schedule "irrelevant" insns before memory stores to resolve
+	   as many producer dependencies of stores as possible.  */
 	r = irrel2 - irrel1;
+      else
+	/* Schedule "irrelevant" insns after memory reads to avoid breaking
+	   memory read sequences.  */
+	r = irrel1 - irrel2;
     }
 
   return r;