package/gcc/6.3.0/831-ARM-PR-target-70473-Reduce-size-of-Cortex-A8-automat.patch

   1 From 07453188b18a2cbe815917dbb893b4ebfdd793b6 Mon Sep 17 00:00:00 2001
   2 From: ktkachov <ktkachov@138bc75d-0d04-0410-961f-82ee72b054a4>
   3 Date: Fri, 26 Aug 2016 10:23:20 +0000
   4 Subject: [PATCH] [ARM] PR target/70473: Reduce size of Cortex-A8 automaton
   5
   6         PR target/70473
   7         * config/arm/cortex-a8-neon.md (cortex_a8_vfp_muld): Reduce
   8         reservation duration to 15 cycles.
   9         (cortex_a8_vfp_macs): Likewise.
  10         (cortex_a8_vfp_macd): Likewise.
  11         (cortex_a8_vfp_divs): Likewise.
  12         (cortex_a8_vfp_divd): Likewise.
  13
  14 git-svn-id: svn+ssh://gcc.gnu.org/svn/gcc/trunk@239772
  15 138bc75d-0d04-0410-961f-82ee72b054a4
  16
  17 [Backport from gcc upstream commit 79f5d5e3cd5a668d5ecb2b6b0cce80ef5706ac53.]
  18 Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
  19 ---
  20  gcc/config/arm/cortex-a8-neon.md | 14 +++++++++-----
  21  1 file changed, 9 insertions(+), 5 deletions(-)
  22
  23 diff --git a/gcc/config/arm/cortex-a8-neon.md b/gcc/config/arm/cortex-a8-neon.md
  24 index 45f861f..b16c299 100644
  25 --- a/gcc/config/arm/cortex-a8-neon.md
  26 +++ b/gcc/config/arm/cortex-a8-neon.md
  27 @@ -357,30 +357,34 @@
  28         (eq_attr "type" "fmuls"))
  29    "cortex_a8_vfp,cortex_a8_vfplite*11")
  30
  31 +;; Don't model a reservation for more than 15 cycles as this explodes the
  32 +;; state space of the automaton for little gain.  It is unlikely that the
  33 +;; scheduler will find enough instructions to hide the full latency of the
  34 +;; instructions.
  35  (define_insn_reservation "cortex_a8_vfp_muld" 17
  36    (and (eq_attr "tune" "cortexa8")
  37         (eq_attr "type" "fmuld"))
  38 -  "cortex_a8_vfp,cortex_a8_vfplite*16")
  39 +  "cortex_a8_vfp,cortex_a8_vfplite*15")
  40
  41  (define_insn_reservation "cortex_a8_vfp_macs" 21
  42    (and (eq_attr "tune" "cortexa8")
  43         (eq_attr "type" "fmacs,ffmas"))
  44 -  "cortex_a8_vfp,cortex_a8_vfplite*20")
  45 +  "cortex_a8_vfp,cortex_a8_vfplite*15")
  46
  47  (define_insn_reservation "cortex_a8_vfp_macd" 26
  48    (and (eq_attr "tune" "cortexa8")
  49         (eq_attr "type" "fmacd,ffmad"))
  50 -  "cortex_a8_vfp,cortex_a8_vfplite*25")
  51 +  "cortex_a8_vfp,cortex_a8_vfplite*15")
  52
  53  (define_insn_reservation "cortex_a8_vfp_divs" 37
  54    (and (eq_attr "tune" "cortexa8")
  55         (eq_attr "type" "fdivs, fsqrts"))
  56 -  "cortex_a8_vfp,cortex_a8_vfplite*36")
  57 +  "cortex_a8_vfp,cortex_a8_vfplite*15")
  58
  59  (define_insn_reservation "cortex_a8_vfp_divd" 65
  60    (and (eq_attr "tune" "cortexa8")
  61         (eq_attr "type" "fdivd, fsqrtd"))
  62 -  "cortex_a8_vfp,cortex_a8_vfplite*64")
  63 +  "cortex_a8_vfp,cortex_a8_vfplite*15")
  64
  65  ;; Comparisons can actually take 7 cycles sometimes instead of four,
  66  ;; but given all the other instructions lumped into type=ffarith that
  67 --
  68 2.7.4
  69