[src/trunk]: src/sys/crypto/aes/arch/arm aes neon: Hoist dsbd/dsbe address ca...

[riastradh <riastradh%NetBSD.org@localhost>]

details:   https://anonhg.NetBSD.org/src/rev/bf5f7ba31933
branches:  trunk
changeset: 954757:bf5f7ba31933
user:      riastradh <riastradh%NetBSD.org@localhost>
date:      Thu Sep 10 11:30:28 2020 +0000

description:
aes neon: Hoist dsbd/dsbe address calculation out of loop.

diffstat:

 sys/crypto/aes/arch/arm/aes_neon_32.S |  15 ++++++++++-----
 1 files changed, 10 insertions(+), 5 deletions(-)

diffs (57 lines):

diff -r e02a1bdd47dd -r bf5f7ba31933 sys/crypto/aes/arch/arm/aes_neon_32.S
--- a/sys/crypto/aes/arch/arm/aes_neon_32.S     Thu Sep 10 11:30:08 2020 +0000
+++ b/sys/crypto/aes/arch/arm/aes_neon_32.S     Thu Sep 10 11:30:28 2020 +0000
@@ -1,4 +1,4 @@
-/*     $NetBSD: aes_neon_32.S,v 1.9 2020/09/10 11:30:08 riastradh Exp $        */
+/*     $NetBSD: aes_neon_32.S,v 1.10 2020/09/10 11:30:28 riastradh Exp $       */
 
 /*-
  * Copyright (c) 2020 The NetBSD Foundation, Inc.
@@ -28,7 +28,7 @@
 
 #include <arm/asm.h>
 
-RCSID("$NetBSD: aes_neon_32.S,v 1.9 2020/09/10 11:30:08 riastradh Exp $")
+RCSID("$NetBSD: aes_neon_32.S,v 1.10 2020/09/10 11:30:28 riastradh Exp $")
 
        .fpu    neon
 
@@ -431,6 +431,9 @@
 
        /*
         * r3: 3 & ~(nrounds - 1)
+        * r4: dsbd
+        * r5: dsbe
+        * r6,r8,r10,ip: temporaries
         * q0={d0-d1}: x/ak
         * q1={d2-d3}: 0x0f0f...
         * q2={d4-d5}: lo/k/j/io
@@ -488,6 +491,10 @@
        add     r4, ip, #(dsb9 - .Lconstants)
        vld1.8  {q4-q5}, [r4 :256]      /* q4 := dsb9[0], q5 := dsb9[1] */
 
+       /* r4 := dsbd, r5 := dsbe */
+       add     r4, ip, #(dsbd - .Lconstants)
+       add     r5, ip, #(dsbe - .Lconstants)
+
        /* q0 := rk[0] + diptlo(lo) + dipthi(hi) */
        veor    q0, q14, q2
        veor    q0, q0, q3
@@ -496,7 +503,6 @@
 
        _ALIGN_TEXT
 1:     /* load dsbd */
-       add     r4, ip, #(dsbd - .Lconstants)
        vld1.8  {q8-q9}, [r4 :256]      /* q8 := dsbd[0], q9 := dsbd[1] */
 
        vld1.8  {q14}, [r0 :128]!       /* q14 = *rk++ */
@@ -522,8 +528,7 @@
        veor    q0, q0, q13
 
        /* load dsbe */
-       add     r4, ip, #(dsbe - .Lconstants)
-       vld1.8  {q8-q9}, [r4 :256]!     /* q8 := dsbe[0], q9 := dsbe[1] */
+       vld1.8  {q8-q9}, [r5 :256]      /* q8 := dsbe[0], q9 := dsbe[1] */
 
        /* q0 := x(mc) + dsbb_0(io) + dsbb_1(jo) */
        vtbl.8  d28, {q0}, d30