Module Name: src
Committed By: riastradh
Date: Thu Sep 10 11:30:28 UTC 2020
Modified Files:
src/sys/crypto/aes/arch/arm: aes_neon_32.S
Log Message:
aes neon: Hoist dsbd/dsbe address calculation out of loop.
To generate a diff of this commit:
cvs rdiff -u -r1.9 -r1.10 src/sys/crypto/aes/arch/arm/aes_neon_32.S
Please note that diffs are not public domain; they are subject to the
copyright notices on the relevant files.
Modified files:
Index: src/sys/crypto/aes/arch/arm/aes_neon_32.S
diff -u src/sys/crypto/aes/arch/arm/aes_neon_32.S:1.9 src/sys/crypto/aes/arch/arm/aes_neon_32.S:1.10
--- src/sys/crypto/aes/arch/arm/aes_neon_32.S:1.9 Thu Sep 10 11:30:08 2020
+++ src/sys/crypto/aes/arch/arm/aes_neon_32.S Thu Sep 10 11:30:28 2020
@@ -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 @@ ENTRY(aes_neon_dec1)
/*
* 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 @@ ENTRY(aes_neon_dec1)
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 @@ ENTRY(aes_neon_dec1)
_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 @@ ENTRY(aes_neon_dec1)
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
