fast version of mcl_fp_addNC64

src/low_armv7.s

@@ -1,22 +1,9 @@
 	.arch armv7-a
 
 	.global mcl_fp_addNC64
-	.global mcl_fp_addNC64_2
 
 	.align 2
 mcl_fp_addNC64:
-	ldrd	r2, [r2]
-	stmfd	sp!, {r4, r5}
-	ldrd	r4, [r1]
-	adds	r2, r2, r4
-	adc	r3, r3, r5
-	ldmfd	sp!, {r4, r5}
-	strd	r2, [r0]
-	bx	lr
-
-# slow
-	.align 2
-mcl_fp_addNC64_2:
 	ldm	r1, {r3, r12}
 	ldm	r2, {r1, r2}
 	adds	r1, r1, r3

test/low_test.cpp

@@ -8,7 +8,6 @@
 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_addNC64_2(uint32_t *z, const uint32_t *x, const uint32_t *y);
 
 CYBOZU_TEST_AUTO(addNC)
 {
@@ -23,12 +22,7 @@ CYBOZU_TEST_AUTO(addNC)
 		low_add<N>(z, x, y);
 		mcl_fp_addNC64(w, x, y);
 		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
-		mcl_fp_addNC64_2(w, x, y);
-		CYBOZU_TEST_EQUAL_ARRAY(z, w, N);
 	}
 	CYBOZU_BENCH("add1", mcl_fp_addNC64, x, x, y);
-	CYBOZU_BENCH("add2", mcl_fp_addNC64_2, x, x, y);
-	CYBOZU_BENCH("add1", mcl_fp_addNC64, x, x, y);
-	CYBOZU_BENCH("add2", mcl_fp_addNC64_2, x, x, y);
 }