a little faster mcl_fp_addNC96

src/low_armv7.s

@@ -1,7 +1,7 @@
 	.arch armv7-a
 
 	.global mcl_fp_addNC64
-	.global mcl_fp_addNC96_1
+	.global mcl_fp_addNC96
 	.global mcl_fp_addNC96_2
 
 	.align 2
@@ -15,33 +15,28 @@ mcl_fp_addNC64:
 
 
 	.align 2
-mcl_fp_addNC96_1:
+mcl_fp_addNC96:
 	push	{r4, lr}
-	ldm	r1, {r3, r12, lr}
-	ldm	r2, {r1, r4}
-	ldr	r2, [r2, #8]
-	adds	r1, r1, r3
-	adcs	r3, r4, r12
-	adc	r2, r2, lr
-	stm	r0, {r1, r3}
-	str	r2, [r0, #8]
-	pop	{r4, lr}
+	ldm		r1, {r1, r3, r12}
+	ldm		r2, {r2, r4, lr}
+	adds	r1, r1, r2
+	adcs	r3, r3, r4
+	adc		r12, r12, lr
+	stm		r0, {r1, r3, r12}
+	pop		{r4, lr}
 	bx		lr
 
+# slower
 	.align 2
 mcl_fp_addNC96_2:
-	ldr		r3, [r1]
-	ldr		r12, [r2]
+	ldr		r3, [r1], #4
+	ldr		r12, [r2], #4
 	adds	r3, r3, r12
-	str		r3, [r0]
-
-	ldr		r3, [r1, #4]
-	ldr		r12, [r2, #4]
-	adcs	r3, r3, r12
-	str		r3, [r0, #4]
+	str		r3, [r0], #4
 
-	ldr		r3, [r1, #8]
-	ldr		r12, [r2, #8]
+	ldm		r1, {r1, r3}
+	ldm		r2, {r2, r12}
+	adcs	r1, r1, r2
 	adcs	r3, r3, r12
-	str		r3, [r0, #8]
+	stm		r0, {r1, r3}
 	bx		lr

test/low_test.cpp

@@ -8,7 +8,7 @@
 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_1(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_addNC96_2(uint32_t *z, const uint32_t *x, const uint32_t *y);
 
 CYBOZU_TEST_AUTO(addNC64)
@@ -38,12 +38,12 @@ CYBOZU_TEST_AUTO(addNC)
 		rg.read(x, N);
 		rg.read(y, N);
 		low_add<N>(z, x, y);
-		mcl_fp_addNC96_1(w, x, y);
+		mcl_fp_addNC96(w, x, y);
 		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
 		mcl_fp_addNC96_2(w, x, y);
 		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
 	}
-	CYBOZU_BENCH("add96_1", mcl_fp_addNC96_1, x, x, y);
+	CYBOZU_BENCH("add96", mcl_fp_addNC96, x, x, y);
 	CYBOZU_BENCH("add96_2", mcl_fp_addNC96_2, x, x, y);
 }