From bd7f6e539e8d415655f558dbd3e314ea0def3f65 Mon Sep 17 00:00:00 2001
From: Jack Lloyd <jack@randombit.net>
Date: Sun, 3 Sep 2017 15:02:41 -0400
Subject: Add support for ARMv8 PMULL instruction

Contributed in GH #842
---
 src/lib/modes/aead/gcm/gcm.cpp         |  6 +++
 src/lib/modes/aead/gcm/pmull/info.txt  | 16 ++++++++
 src/lib/modes/aead/gcm/pmull/pmull.cpp | 71 ++++++++++++++++++++++++++++++++++
 src/lib/modes/aead/gcm/pmull/pmull.h   | 19 +++++++++
 4 files changed, 112 insertions(+)
 create mode 100644 src/lib/modes/aead/gcm/pmull/info.txt
 create mode 100644 src/lib/modes/aead/gcm/pmull/pmull.cpp
 create mode 100644 src/lib/modes/aead/gcm/pmull/pmull.h

diff --git a/src/lib/modes/aead/gcm/gcm.cpp b/src/lib/modes/aead/gcm/gcm.cpp
index 0d0cbff3c..f9ffdee97 100644
--- a/src/lib/modes/aead/gcm/gcm.cpp
+++ b/src/lib/modes/aead/gcm/gcm.cpp
@@ -14,6 +14,9 @@
 #if defined(BOTAN_HAS_GCM_CLMUL)
   #include <botan/internal/clmul.h>
   #include <botan/cpuid.h>
+#elif defined(BOTAN_HAS_GCM_PMULL)
+  #include <botan/internal/pmull.h>
+  #include <botan/cpuid.h>
 #endif
 
 namespace Botan {
@@ -25,6 +28,9 @@ void GHASH::gcm_multiply(secure_vector<uint8_t>& x) const
 #if defined(BOTAN_HAS_GCM_CLMUL)
    if(CPUID::has_clmul())
       return gcm_multiply_clmul(x.data(), m_H.data());
+#elif defined(BOTAN_HAS_GCM_PMULL)
+   if(CPUID::has_arm_pmull())
+      return gcm_multiply_pmull(x.data(), m_H.data());
 #endif
 
    static const uint64_t R = 0xE100000000000000;
diff --git a/src/lib/modes/aead/gcm/pmull/info.txt b/src/lib/modes/aead/gcm/pmull/info.txt
new file mode 100644
index 000000000..711587568
--- /dev/null
+++ b/src/lib/modes/aead/gcm/pmull/info.txt
@@ -0,0 +1,16 @@
+<defines>
+GCM_PMULL -> 20170903
+</defines>
+
+<arch>
+arm64
+</arch>
+
+<cc>
+gcc:4.9
+clang:3.8
+</cc>
+
+<header:internal>
+pmull.h
+</header:internal>
diff --git a/src/lib/modes/aead/gcm/pmull/pmull.cpp b/src/lib/modes/aead/gcm/pmull/pmull.cpp
new file mode 100644
index 000000000..54e841650
--- /dev/null
+++ b/src/lib/modes/aead/gcm/pmull/pmull.cpp
@@ -0,0 +1,71 @@
+/*
+* Contributed by Jeffrey Walton
+*
+* Botan is released under the Simplified BSD License (see license.txt)
+*/
+
+#include <botan/internal/pmull.h>
+#include <arm_neon.h>
+
+namespace Botan {
+
+BOTAN_FUNC_ISA("+crypto")
+void gcm_multiply_pmull(uint8_t x[16], const uint8_t H[16])
+   {
+   /*
+   * Implementing GCM on ARMv8, http://conradoplg.cryptoland.net/files/2010/12/gcm14.pdf
+   */
+
+   const uint64x2_t a64 = vreinterpretq_u64_u8(vcombine_u8(vrev64_u8(vld1_u8(x+8)), vrev64_u8(vld1_u8(x))));
+   const uint64x2_t b64 = vreinterpretq_u64_u8(vcombine_u8(vrev64_u8(vld1_u8(H+8)), vrev64_u8(vld1_u8(H))));
+
+   uint64x2_t T0, T1, T2, T3, T4, T5;
+
+   T0 = (uint64x2_t)vmull_p64(vgetq_lane_u64(a64, 0), vgetq_lane_u64(b64, 0));
+   T1 = (uint64x2_t)vmull_p64(vgetq_lane_u64(a64, 1), vgetq_lane_u64(b64, 0));
+   T2 = (uint64x2_t)vmull_p64(vgetq_lane_u64(a64, 0), vgetq_lane_u64(b64, 1));
+   T3 = (uint64x2_t)vmull_p64(vgetq_lane_u64(a64, 1), vgetq_lane_u64(b64, 1));
+
+   T1 = veorq_u64(T1, T2);
+   T2 = vreinterpretq_u64_u8(vextq_u8(vdupq_n_u8(0), vreinterpretq_u8_u64(T1), 8));
+   T1 = vreinterpretq_u64_u8(vextq_u8(vreinterpretq_u8_u64(T1), vdupq_n_u8(0), 8));
+   T0 = veorq_u64(T0, T2);
+   T3 = veorq_u64(T3, T1);
+
+   T4 = vshrq_n_u64(T0, 31);
+   T0 = vshlq_n_u64(T0, 1);
+
+   T5 = vshrq_n_u64(T3, 31);
+   T3 = vshlq_n_u64(T3, 1);
+
+   T2 = vreinterpretq_u64_u8(vextq_u8(vreinterpretq_u8_u64(T4), vdupq_n_u8(0), 12));
+   T5 = vreinterpretq_u64_u8(vextq_u8(vdupq_n_u8(0), vreinterpretq_u8_u64(T5), 12));
+   T4 = vreinterpretq_u64_u8(vextq_u8(vdupq_n_u8(0), vreinterpretq_u8_u64(T4), 12));
+   T0 = vorrq_u64(T0, T4);
+   T3 = vorrq_u64(T3, T5);
+   T3 = vorrq_u64(T3, T2);
+
+   T4 = vreinterpretq_u64_u32(vshlq_n_u32(vreinterpretq_u32_u64(T0), 31));
+   T5 = vreinterpretq_u64_u32(vshlq_n_u32(vreinterpretq_u32_u64(T0), 30));
+   T2 = vreinterpretq_u64_u32(vshlq_n_u32(vreinterpretq_u32_u64(T0), 25));
+
+   T4 = veorq_u64(T4, T5);
+   T4 = veorq_u64(T4, T2);
+   T5 = vreinterpretq_u64_u8(vextq_u8(vreinterpretq_u8_u64(T4), vdupq_n_u8(0), 4));
+   T3 = veorq_u64(T3, T5);
+   T4 = vreinterpretq_u64_u8(vextq_u8(vdupq_n_u8(0), vreinterpretq_u8_u64(T4), 4));
+   T0 = veorq_u64(T0, T4);
+   T3 = veorq_u64(T3, T0);
+
+   T4 = vreinterpretq_u64_u32(vshrq_n_u32(vreinterpretq_u32_u64(T0), 1));
+   T1 = vreinterpretq_u64_u32(vshrq_n_u32(vreinterpretq_u32_u64(T0), 2));
+   T2 = vreinterpretq_u64_u32(vshrq_n_u32(vreinterpretq_u32_u64(T0), 7));
+   T3 = veorq_u64(T3, T1);
+   T3 = veorq_u64(T3, T2);
+   T3 = veorq_u64(T3, T4);
+
+   vst1_u8(x+0, vrev64_u8(vreinterpret_u8_u64(vget_high_u64(T3))));
+   vst1_u8(x+8, vrev64_u8(vreinterpret_u8_u64(vget_low_u64(T3))));
+   }
+
+}
diff --git a/src/lib/modes/aead/gcm/pmull/pmull.h b/src/lib/modes/aead/gcm/pmull/pmull.h
new file mode 100644
index 000000000..04afddfbf
--- /dev/null
+++ b/src/lib/modes/aead/gcm/pmull/pmull.h
@@ -0,0 +1,19 @@
+/*
+* PMULL hook
+* (C) 2017 Jack Lloyd
+*
+* Botan is released under the Simplified BSD License (see license.txt)
+*/
+
+#ifndef BOTAN_GCM_PMULL_H__
+#define BOTAN_GCM_PMULL_H__
+
+#include <botan/types.h>
+
+namespace Botan {
+
+void gcm_multiply_pmull(uint8_t x[16], const uint8_t H[16]);
+
+}
+
+#endif
-- 
cgit v1.2.3