From a499a1aa9354b076bba85eabb9de98969caed99c Mon Sep 17 00:00:00 2001
From: MITSUNARI Shigeo <herumi@nifty.com>
Date: Fri, 24 Jun 2016 16:43:36 +0900
Subject: [PATCH] mcl_fp_addNC128 for arm

---
 src/low_armv7.s   | 21 +++++++++++++++++++++
 test/low_test.cpp | 12 ++++++++----
 2 files changed, 29 insertions(+), 4 deletions(-)

diff --git a/src/low_armv7.s b/src/low_armv7.s
index 7d252e4..7c70f86 100644
--- a/src/low_armv7.s
+++ b/src/low_armv7.s
@@ -57,3 +57,24 @@ mcl_fp_addNC128:
 	pop		{r4, lr}
 	bx		lr
 
+	.globl	mcl_fp_addNC256
+	.align	2
+mcl_fp_addNC256:
+	push	{r4, r5, r6, r7, r8, lr}
+	ldm		r1!, {r3, r4, r5, r6}
+	ldm		r2!, {r7, r8, r12, lr}
+	adds	r3, r3, r7
+	adcs	r4, r4, r8
+	adcs	r5, r5, r12
+	adcs	r6, r6, lr
+	stm		r0!, {r3, r4, r5, r6}
+
+	ldm		r1!, {r3, r4, r5, r6}
+	ldm		r2!, {r7, r8, r12, lr}
+	adcs	r3, r3, r7
+	adcs	r4, r4, r8
+	adcs	r5, r5, r12
+	adcs	r6, r6, lr
+	stm		r0!, {r3, r4, r5, r6}
+	pop		{r4, r5, r6, r7, r8, lr}
+	bx		lr
diff --git a/test/low_test.cpp b/test/low_test.cpp
index 80ee474..478620b 100644
--- a/test/low_test.cpp
+++ b/test/low_test.cpp
@@ -11,6 +11,7 @@ cybozu::XorShift rg;
 extern "C" void mcl_fp_addNC64(uint32_t *z, const uint32_t *x, const uint32_t *y);
 extern "C" void mcl_fp_addNC96(uint32_t *z, const uint32_t *x, const uint32_t *y);
 extern "C" void mcl_fp_addNC128(uint32_t *z, const uint32_t *x, const uint32_t *y);
+extern "C" void mcl_fp_addNC256(uint32_t *z, const uint32_t *x, const uint32_t *y);
 
 template<size_t N>
 void addNC(uint32_t *z, const uint32_t *x, const uint32_t *y);
@@ -20,6 +21,7 @@ void addNC(uint32_t *z, const uint32_t *x, const uint32_t *y);
 DEF_ADD(64)
 DEF_ADD(96)
 DEF_ADD(128)
+DEF_ADD(256)
 
 #define CAT(S, BIT) "S##BIT"
 
@@ -45,11 +47,12 @@ void benchAdd()
 CYBOZU_TEST_AUTO(addNC64) { benchAdd<64>(); }
 CYBOZU_TEST_AUTO(addNC96) { benchAdd<96>(); }
 CYBOZU_TEST_AUTO(addNC128) { benchAdd<128>(); }
+CYBOZU_TEST_AUTO(addNC256) { benchAdd<256>(); }
 #if 0
-CYBOZU_TEST_AUTO(addNC128)
+CYBOZU_TEST_AUTO(addNC)
 {
 	using namespace mcl::fp;
-	const size_t bit = 128;
+	const size_t bit = 256;
 	const size_t N = bit / UnitBitSize;
 	Unit x[N], y[N];
 	for (int i = 0; i < 10; i++) {
@@ -60,11 +63,12 @@ CYBOZU_TEST_AUTO(addNC128)
 		low_add<N>(z, x, y);
 		addNC<bit>(w, x, y);
 		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
-		mcl_fp_addNC128_2(w, x, y);
+		mcl_fp_addNC256_2(w, x, y);
 		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
 	}
 	std::string name = "name" + cybozu::itoa(bit);
 	CYBOZU_BENCH(name.c_str(), addNC<bit>, x, x, y);
-	CYBOZU_BENCH("ad128", mcl_fp_addNC128_2, x, x, y);
+	CYBOZU_BENCH("ad128", mcl_fp_addNC256_2, x, x, y);
 }
 #endif
+