arm: add softfp support in kernel/arm/axpy_vfp.S

kernel/arm/KERNEL.ARMV6

@@ -37,6 +37,11 @@ DASUMKERNEL  = asum_vfp.S
 CASUMKERNEL  = asum_vfp.S
 ZASUMKERNEL  = asum_vfp.S
 
+SAXPYKERNEL  = axpy_vfp.S
+DAXPYKERNEL  = axpy_vfp.S
+CAXPYKERNEL  = axpy_vfp.S
+ZAXPYKERNEL  = axpy_vfp.S
+
 SGEMMKERNEL    = ../generic/gemmkernel_4x2.c
 ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
 SGEMMINCOPY    = sgemm_ncopy_4_vfp.S
@@ -76,11 +81,6 @@ ZGEMMOTCOPYOBJ = zgemm_otcopy.o
 
 ifeq ($(ARM_ABI),hard)
 
-SAXPYKERNEL  = axpy_vfp.S
-DAXPYKERNEL  = axpy_vfp.S
-CAXPYKERNEL  = axpy_vfp.S
-ZAXPYKERNEL  = axpy_vfp.S
-
 SDOTKERNEL   = sdot_vfp.S
 DDOTKERNEL   = ddot_vfp.S
 CDOTKERNEL   = cdot_vfp.S

kernel/arm/axpy_vfp.S

@@ -38,18 +38,52 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#ifndef ARM_SOFTFP_ABI
-//hard	
-#define	OLD_INC_X	[fp, #0 ]
-#define	OLD_Y		[fp, #4 ]
-#define	OLD_INC_Y	[fp, #8 ]
-#else
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(COMPLEX)
+
+#if !defined(DOUBLE)
+#define	OLD_ALPHA	r3
 #define	OLD_X		[fp, #0 ]
 #define	OLD_INC_X	[fp, #4 ]
 #define	OLD_Y		[fp, #8 ]
 #define	OLD_INC_Y	[fp, #12 ]
+#else
+#define	OLD_ALPHA	[fp, #0]
+#define	OLD_X		[fp, #8 ]
+#define	OLD_INC_X	[fp, #12 ]
+#define	OLD_Y		[fp, #16 ]
+#define	OLD_INC_Y	[fp, #20 ]
 #endif
-	
+
+#else //COMPLEX
+
+#if !defined(DOUBLE)
+#define	OLD_ALPHAR	r3
+#define	OLD_ALPHAI	[fp, #0 ]
+#define	OLD_X		[fp, #4 ]
+#define	OLD_INC_X	[fp, #8 ]
+#define	OLD_Y		[fp, #12 ]
+#define	OLD_INC_Y	[fp, #16 ]
+#else
+#define	OLD_ALPHAR	[fp, #0]
+#define	OLD_ALPHAI	[fp, #8]
+#define	OLD_X		[fp, #16 ]
+#define	OLD_INC_X	[fp, #20 ]
+#define	OLD_Y		[fp, #24 ]
+#define	OLD_INC_Y	[fp, #28 ]
+#endif
+
+#endif //!defined(COMPLEX)
+
+#else //__ARM_PCS_VFP
+
+#define	OLD_INC_X	[fp, #0 ]
+#define	OLD_Y		[fp, #4 ]
+#define	OLD_INC_Y	[fp, #8 ]
+
+#endif //!defined(__ARM_PCS_VFP)
+
 #define	N	r0
 #define Y	r1
 #define	INC_X	r2
@@ -370,13 +404,28 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         add     fp, sp, #8
 	sub     sp, sp, #STACKSIZE                              // reserve stack
 
-#ifdef ARM_SOFTFP_ABI
-#ifndef DOUBLE
-	vmov	s0, r3     //move alpha to s0
+#if !defined(__ARM_PCS_VFP)
+#if !defined(COMPLEX)
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHA
 	ldr	X, OLD_X
+#else
+	vldr	d0, OLD_ALPHA
+	ldr	X, OLD_X
+#endif
+#else //COMPLEX
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHAR
+	vldr	s1, OLD_ALPHAI
+	ldr	X, OLD_X
+#else
+	vldr	d0, OLD_ALPHAR
+	vldr	d1, OLD_ALPHAI
+	ldr     X, OLD_X
+#endif
 #endif
 #endif
-	
+
 	ldr    INC_X , OLD_INC_X
 	ldr         Y, OLD_Y
 	ldr    INC_Y , OLD_INC_Y