add mcl_fp_addNC{160,192,224}

src/low_armv7.s

@@ -72,6 +72,65 @@ cl_fp_addNC128_2:
 	pop		{r4, r5, r6, lr}
 	bx		lr
 
+	.globl	mcl_fp_addNC160
+	.align	2
+mcl_fp_addNC160:
+	push	{r4, lr}
+	ldm		r1!, {r3, r4}
+	ldm		r2!, {r12, lr}
+	adds	r3, r3, r12
+	adcs	r4, r4, lr
+	stm		r0!, {r3, r4}
+	ldm		r1, {r1, r3, r4}
+	ldm		r2, {r2, r12, lr}
+	adcs	r1, r1, r2
+	adcs	r3, r3, r12
+	adcs	r4, r4, lr
+	stm		r0, {r1, r3, r4}
+	pop		{r4, lr}
+	bx		lr
+
+	.globl	mcl_fp_addNC192
+	.align	2
+mcl_fp_addNC192:
+	push	{r4, r5, r6, lr}
+	ldm		r1!, {r3, r4, r5}
+	ldm		r2!, {r6, r12, lr}
+	adds	r3, r3, r6
+	adcs	r4, r4, r12
+	adcs	r5, r5, lr
+	stm		r0!, {r3, r4, r5}
+
+	ldm		r1, {r3, r4, r5}
+	ldm		r2, {r6, r12, lr}
+	adcs	r3, r3, r6
+	adcs	r4, r4, r12
+	adcs	r5, r5, lr
+	stm		r0, {r3, r4, r5}
+	pop		{r4, r5, r6, lr}
+	bx		lr
+
+	.globl	mcl_fp_addNC224
+	.align	2
+mcl_fp_addNC224:
+	push	{r4, r5, r6, lr}
+	ldm		r1!, {r3, r4, r5}
+	ldm		r2!, {r6, r12, lr}
+	adds	r3, r3, r6
+	adcs	r4, r4, r12
+	adcs	r5, r5, lr
+	stm		r0!, {r3, r4, r5}
+
+	ldm		r1, {r1, r3, r4, r5}
+	ldm		r2, {r2, r6, r12, lr}
+	adcs	r1, r1, r2
+	adcs	r3, r3, r6
+	adcs	r4, r4, r12
+	adcs	r5, r5, lr
+	stm		r0, {r1, r3, r4, r5}
+	pop		{r4, r5, r6, lr}
+	bx		lr
+
 	.globl	mcl_fp_addNC256
 	.align	2
 mcl_fp_addNC256:
@@ -84,12 +143,12 @@ mcl_fp_addNC256:
 	adcs	r6, r6, lr
 	stm		r0!, {r3, r4, r5, r6}
 
-	ldm		r1!, {r3, r4, r5, r6}
-	ldm		r2!, {r7, r8, r12, lr}
+	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}
+	stm		r0, {r3, r4, r5, r6}
 	pop		{r4, r5, r6, r7, r8, lr}
 	bx		lr

test/low_test.cpp

@@ -11,6 +11,9 @@ 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_addNC160(uint32_t *z, const uint32_t *x, const uint32_t *y);
+extern "C" void mcl_fp_addNC192(uint32_t *z, const uint32_t *x, const uint32_t *y);
+extern "C" void mcl_fp_addNC224(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);
 extern "C" void add_test(uint32_t *z, const uint32_t *x, const uint32_t *y);
 
@@ -22,6 +25,9 @@ void addNC(uint32_t *z, const uint32_t *x, const uint32_t *y);
 DEF_ADD(64)
 DEF_ADD(96)
 DEF_ADD(128)
+DEF_ADD(160)
+DEF_ADD(192)
+DEF_ADD(224)
 DEF_ADD(256)
 
 #define CAT(S, BIT) "S##BIT"
@@ -48,8 +54,11 @@ void benchAdd()
 CYBOZU_TEST_AUTO(addNC64) { benchAdd<64>(); }
 CYBOZU_TEST_AUTO(addNC96) { benchAdd<96>(); }
 CYBOZU_TEST_AUTO(addNC128) { benchAdd<128>(); }
+CYBOZU_TEST_AUTO(addNC160) { benchAdd<160>(); }
+CYBOZU_TEST_AUTO(addNC192) { benchAdd<192>(); }
+CYBOZU_TEST_AUTO(addNC224) { benchAdd<224>(); }
 CYBOZU_TEST_AUTO(addNC256) { benchAdd<256>(); }
-#if 1
+#if 0
 CYBOZU_TEST_AUTO(addNC)
 {
 	using namespace mcl::fp;