Merge pull request #1221 from ashwinyes/develop_arm_softfp

arm: add support for softfp in arm vfp assembly files

Makefile.arm

@@ -1,5 +1,4 @@
-#ifeq logical or
-ifeq ($(CORE), $(filter $(CORE),CORTEXA9 CORTEXA15))
+ifeq ($(CORE), $(filter $(CORE),ARMV7 CORTEXA9 CORTEXA15))
 ifeq ($(OSNAME), Android)
 CCOMMON_OPT += -mfpu=neon  -march=armv7-a
 FCOMMON_OPT += -mfpu=neon  -march=armv7-a
@@ -9,28 +8,12 @@ FCOMMON_OPT += -mfpu=vfpv3 -march=armv7-a
 endif
 endif
 
-ifeq ($(CORE), ARMV7)
-ifeq ($(OSNAME), Android)
-ifeq ($(ARM_SOFTFP_ABI), 1)
-CCOMMON_OPT += -mfpu=neon -march=armv7-a
-FCOMMON_OPT += -mfpu=neon -march=armv7-a
-else
-CCOMMON_OPT += -mfpu=neon -march=armv7-a -Wl,--no-warn-mismatch
-FCOMMON_OPT += -mfpu=neon -march=armv7-a -Wl,--no-warn-mismatch
-endif
-else
-CCOMMON_OPT += -mfpu=vfpv3 -march=armv7-a
-FCOMMON_OPT += -mfpu=vfpv3 -march=armv7-a
-endif
-endif
-
 ifeq ($(CORE), ARMV6)
 CCOMMON_OPT += -mfpu=vfp -march=armv6
 FCOMMON_OPT += -mfpu=vfp -march=armv6
 endif
 
-
 ifeq ($(CORE), ARMV5)
-CCOMMON_OPT += -marm -march=armv5
-FCOMMON_OPT += -marm -march=armv5
+CCOMMON_OPT += -march=armv5
+FCOMMON_OPT += -march=armv5
 endif

Makefile.system

@@ -242,6 +242,10 @@ EXTRALIB	+= -lm
 NO_EXPRECISION = 1
 endif
 
+ifeq ($(OSNAME), Android)
+EXTRALIB	+= -lm
+endif
+
 ifeq ($(OSNAME), AIX)
 EXTRALIB	+= -lm
 endif
@@ -486,12 +490,10 @@ BINARY_DEFINED  = 1
 CCOMMON_OPT += -marm
 FCOMMON_OPT += -marm
 
+# If softfp abi is mentioned on the command line, force it.
 ifeq ($(ARM_SOFTFP_ABI), 1)
-CCOMMON_OPT += -mfloat-abi=softfp -DARM_SOFTFP_ABI
-FCOMMON_OPT += -mfloat-abi=softfp -DARM_SOFTFP_ABI
-else
-CCOMMON_OPT += -mfloat-abi=hard
-FCOMMON_OPT += -mfloat-abi=hard
+CCOMMON_OPT += -mfloat-abi=softfp
+FCOMMON_OPT += -mfloat-abi=softfp
 endif
 
 ifeq ($(OSNAME), Android)

common_arm.h

@@ -111,11 +111,6 @@ REALNAME:
 
 #define PROFCODE
 
-#ifdef __ARM_PCS
-//-mfloat-abi=softfp
-#define SOFT_FLOAT_ABI
-#endif
-
 #endif
 
 

kernel/arm/KERNEL.ARMV6

@@ -1,7 +1,5 @@
+include $(KERNELDIR)/KERNEL.ARMV5
 
-
-
-###############################################################################
 SAMAXKERNEL  = iamax_vfp.S
 DAMAXKERNEL  = iamax_vfp.S
 CAMAXKERNEL  = iamax_vfp.S
@@ -44,10 +42,10 @@ DAXPYKERNEL  = axpy_vfp.S
 CAXPYKERNEL  = axpy_vfp.S
 ZAXPYKERNEL  = axpy_vfp.S
 
-SCOPYKERNEL  = copy.c
-DCOPYKERNEL  = copy.c
-CCOPYKERNEL  = zcopy.c
-ZCOPYKERNEL  = zcopy.c
+SROTKERNEL   = rot_vfp.S
+DROTKERNEL   = rot_vfp.S
+CROTKERNEL   = rot_vfp.S
+ZROTKERNEL   = rot_vfp.S
 
 SDOTKERNEL   = sdot_vfp.S
 DDOTKERNEL   = ddot_vfp.S
@@ -59,16 +57,6 @@ DNRM2KERNEL  = nrm2_vfp.S
 CNRM2KERNEL  = nrm2_vfp.S
 ZNRM2KERNEL  = nrm2_vfp.S
 
-SROTKERNEL   = rot_vfp.S
-DROTKERNEL   = rot_vfp.S
-CROTKERNEL   = rot_vfp.S
-ZROTKERNEL   = rot_vfp.S
-
-SSCALKERNEL  = scal.c
-DSCALKERNEL  = scal.c
-CSCALKERNEL  = zscal.c
-ZSCALKERNEL  = zscal.c
-
 SSWAPKERNEL  = swap_vfp.S
 DSWAPKERNEL  = swap_vfp.S
 CSWAPKERNEL  = swap_vfp.S
@@ -84,26 +72,25 @@ DGEMVTKERNEL = gemv_t_vfp.S
 CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
-STRMMKERNEL	= strmm_kernel_4x2_vfp.S
-DTRMMKERNEL	= dtrmm_kernel_4x2_vfp.S
-CTRMMKERNEL	= ctrmm_kernel_2x2_vfp.S
-ZTRMMKERNEL	= ztrmm_kernel_2x2_vfp.S
-
 SGEMMKERNEL    = sgemm_kernel_4x2_vfp.S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
 SGEMMINCOPY    = sgemm_ncopy_4_vfp.S
 SGEMMITCOPY    = sgemm_tcopy_4_vfp.S
 SGEMMINCOPYOBJ = sgemm_incopy.o
 SGEMMITCOPYOBJ = sgemm_itcopy.o
+endif
 SGEMMONCOPY    = sgemm_ncopy_2_vfp.S
-SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
-SGEMMONCOPYOBJ =  sgemm_oncopy.o
-SGEMMOTCOPYOBJ =  sgemm_otcopy.o
+SGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
+SGEMMONCOPYOBJ = sgemm_oncopy.o
+SGEMMOTCOPYOBJ = sgemm_otcopy.o
 
 DGEMMKERNEL    = dgemm_kernel_4x2_vfp.S
+ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
 DGEMMINCOPY    = dgemm_ncopy_4_vfp.S
 DGEMMITCOPY    = dgemm_tcopy_4_vfp.S
 DGEMMINCOPYOBJ = dgemm_incopy.o
 DGEMMITCOPYOBJ = dgemm_itcopy.o
+endif
 DGEMMONCOPY    = dgemm_ncopy_2_vfp.S
 DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
 DGEMMONCOPYOBJ = dgemm_oncopy.o
@@ -121,26 +108,8 @@ ZGEMMOTCOPY    = zgemm_tcopy_2_vfp.S
 ZGEMMONCOPYOBJ = zgemm_oncopy.o
 ZGEMMOTCOPYOBJ = zgemm_otcopy.o
 
-STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
-
-DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
-
-CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
-
-ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
-
-
-
+STRMMKERNEL    = strmm_kernel_4x2_vfp.S
+DTRMMKERNEL    = dtrmm_kernel_4x2_vfp.S
+CTRMMKERNEL    = ctrmm_kernel_2x2_vfp.S
+ZTRMMKERNEL    = ztrmm_kernel_2x2_vfp.S
 

kernel/arm/KERNEL.ARMV7

@@ -1,91 +1,12 @@
-
-#################################################################################
-SAMAXKERNEL  = iamax_vfp.S
-DAMAXKERNEL  = iamax_vfp.S
-CAMAXKERNEL  = iamax_vfp.S
-ZAMAXKERNEL  = iamax_vfp.S
-
-SAMINKERNEL  = iamax_vfp.S
-DAMINKERNEL  = iamax_vfp.S
-CAMINKERNEL  = iamax_vfp.S
-ZAMINKERNEL  = iamax_vfp.S
-
-SMAXKERNEL   = iamax_vfp.S
-DMAXKERNEL   = iamax_vfp.S
-
-SMINKERNEL   = iamax_vfp.S
-DMINKERNEL   = iamax_vfp.S
-
-ISAMAXKERNEL = iamax_vfp.S
-IDAMAXKERNEL = iamax_vfp.S
-ICAMAXKERNEL = iamax_vfp.S
-IZAMAXKERNEL = iamax_vfp.S
-
-ISAMINKERNEL = iamax_vfp.S
-IDAMINKERNEL = iamax_vfp.S
-ICAMINKERNEL = iamax_vfp.S
-IZAMINKERNEL = iamax_vfp.S
-
-ISMAXKERNEL  = iamax_vfp.S
-IDMAXKERNEL  = iamax_vfp.S
-
-ISMINKERNEL  = iamax_vfp.S
-IDMINKERNEL  = iamax_vfp.S
-
-SSWAPKERNEL  = swap_vfp.S
-DSWAPKERNEL  = swap_vfp.S
-CSWAPKERNEL  = swap_vfp.S
-ZSWAPKERNEL  = swap_vfp.S
-
-SASUMKERNEL  = asum_vfp.S
-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
-
-SCOPYKERNEL  = copy.c
-DCOPYKERNEL  = copy.c
-CCOPYKERNEL  = zcopy.c
-ZCOPYKERNEL  = zcopy.c
-
-SDOTKERNEL   = sdot_vfp.S
-DDOTKERNEL   = ddot_vfp.S
-CDOTKERNEL   = cdot_vfp.S
-ZDOTKERNEL   = zdot_vfp.S
+include $(KERNELDIR)/KERNEL.ARMV6
 
 SNRM2KERNEL  = nrm2_vfpv3.S
 DNRM2KERNEL  = nrm2_vfpv3.S
 CNRM2KERNEL  = nrm2_vfpv3.S
 ZNRM2KERNEL  = nrm2_vfpv3.S
 
-SROTKERNEL   = rot_vfp.S
-DROTKERNEL   = rot_vfp.S
-CROTKERNEL   = rot_vfp.S
-ZROTKERNEL   = rot_vfp.S
-
-SSCALKERNEL  = scal.c
-DSCALKERNEL  = scal.c
-CSCALKERNEL  = zscal.c
-ZSCALKERNEL  = zscal.c
-
 SGEMVNKERNEL = gemv_n_vfpv3.S
 DGEMVNKERNEL = gemv_n_vfpv3.S
-CGEMVNKERNEL = cgemv_n_vfp.S
-ZGEMVNKERNEL = zgemv_n_vfp.S
-
-SGEMVTKERNEL = gemv_t_vfp.S
-DGEMVTKERNEL = gemv_t_vfp.S
-CGEMVTKERNEL = cgemv_t_vfp.S
-ZGEMVTKERNEL = zgemv_t_vfp.S
-
-STRMMKERNEL  =  strmm_kernel_4x4_vfpv3.S
-DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
-CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
-ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 
 SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
 SGEMMONCOPY    =  sgemm_ncopy_4_vfp.S
@@ -100,35 +21,10 @@ DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 
 CGEMMKERNEL    = cgemm_kernel_2x2_vfpv3.S
-CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
-CGEMMOTCOPY    = cgemm_tcopy_2_vfp.S
-CGEMMONCOPYOBJ = cgemm_oncopy.o
-CGEMMOTCOPYOBJ = cgemm_otcopy.o
-
 ZGEMMKERNEL    = zgemm_kernel_2x2_vfpv3.S
-ZGEMMONCOPY    = zgemm_ncopy_2_vfp.S
-ZGEMMOTCOPY    = zgemm_tcopy_2_vfp.S
-ZGEMMONCOPYOBJ = zgemm_oncopy.o
-ZGEMMOTCOPYOBJ = zgemm_otcopy.o
-
-STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
-STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
-STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
-STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
-
-DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
-
-CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
-
-ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
-ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
-ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
-ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 
+STRMMKERNEL  =  strmm_kernel_4x4_vfpv3.S
+DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
+CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
+ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 

kernel/arm/asum_vfp.S

@@ -475,6 +475,14 @@ asum_kernel_L999:
 	vadd.f32	s0 , s0, s1				// set return value
 #endif
 
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	r0, s0
+#else
+	vmov	r0, r1, d0
+#endif
+#endif
+
 	bx	lr
 
 	EPILOGUE

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

kernel/arm/cdot_vfp.S

@@ -41,8 +41,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define	N	r0
 #define	X	r1
 #define	INC_X	r2
-#define	OLD_Y	r3
-
 
 /******************************************************
 * [fp, #-128] - [fp, #-64] is reserved
@@ -50,7 +48,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 * registers
 *******************************************************/
 
-#define OLD_INC_Y	[fp, #4 ]
+#if !defined(__ARM_PCS_VFP)
+#define	OLD_RETURN_ADDR	r0
+#define	OLD_N		r1
+#define	OLD_X		r2
+#define	OLD_INC_X	r3
+#define	OLD_Y		[fp, #0 ]
+#define	OLD_INC_Y	[fp, #4 ]
+#define	RETURN_ADDR	r8
+#else
+#define	OLD_Y		r3
+#define	OLD_INC_Y	[fp, #0 ]
+#endif
 
 #define I	r5
 #define Y	r6
@@ -179,7 +188,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	.align 5
 
 	push	{r4 - r9, fp}
-	add	fp, sp, #24
+	add	fp, sp, #28
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
 	sub	r4, fp, #128
@@ -191,8 +200,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	vmov	s2, s0
 	vmov	s3, s0
 
+#if !defined(__ARM_PCS_VFP)
+	mov	RETURN_ADDR, OLD_RETURN_ADDR
+	mov	N, OLD_N
+	mov	X, OLD_X
+	mov	INC_X, OLD_INC_X
+	ldr	Y, OLD_Y
+	ldr	INC_Y, OLD_INC_Y
+#else
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
+#endif
 
 	cmp	N, #0
 	ble	cdot_kernel_L999
@@ -265,7 +283,6 @@ cdot_kernel_S10:
 
 
 cdot_kernel_L999:
-
 	sub	r3, fp, #128
 	vldm	r3, { s8 - s15}					// restore floating point registers
 
@@ -276,8 +293,11 @@ cdot_kernel_L999:
 	vadd.f32	s0 , s0, s2
 	vsub.f32	s1 , s1, s3
 #endif
+#if !defined(__ARM_PCS_VFP)
+	vstm	RETURN_ADDR, {s0 - s1}
+#endif
 
-	sub	sp, fp, #24
+	sub	sp, fp, #28
 	pop	{r4 - r9, fp}
 	bx	lr
 

kernel/arm/cgemm_kernel_2x2_vfp.S

@@ -64,9 +64,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	r3
+#define OLD_ALPHAI_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define B	[fp, #12 ]
+#define C	[fp, #16 ]
+#define OLD_LDC	[fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -816,6 +825,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/cgemm_kernel_2x2_vfpv3.S

@@ -80,9 +80,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	r3
+#define OLD_ALPHAI_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define B	[fp, #12 ]
+#define C	[fp, #16 ]
+#define OLD_LDC	[fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -873,6 +882,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/cgemv_n_vfp.S

@@ -38,11 +38,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR	r3
+#define OLD_ALPHAI	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define OLD_LDA		[fp, #8 ]
+#define X		[fp, #12 ]
+#define OLD_INC_X	[fp, #16 ]
+#define Y		[fp, #20 ]
+#define OLD_INC_Y	[fp, #24 ]
+#else
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+#endif
+
 #define OLD_A		r3
 #define	OLD_M		r0
 
@@ -462,6 +474,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	cgemvn_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	s0, OLD_ALPHAR
+	vldr	s1, OLD_ALPHAI
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_M, M
 	vstr    s0 , ALPHA_R

kernel/arm/cgemv_t_vfp.S

@@ -38,11 +38,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR	r3
+#define OLD_ALPHAI	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define OLD_LDA		[fp, #8 ]
+#define X		[fp, #12 ]
+#define OLD_INC_X	[fp, #16 ]
+#define Y		[fp, #20 ]
+#define OLD_INC_Y	[fp, #24 ]
+#else
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+#endif
+
 #define OLD_A		r3
 #define	OLD_N		r1
 
@@ -359,6 +371,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	OLD_N, #0
 	ble	cgemvt_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	s0, OLD_ALPHAR
+	vldr	s1, OLD_ALPHAI
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_N, N
 

kernel/arm/ctrmm_kernel_2x2_vfp.S

@@ -67,10 +67,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	r3
+#define OLD_ALPHAI_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define B	[fp, #12 ]
+#define C	[fp, #16 ]
+#define OLD_LDC	[fp, #20 ]
+#define OFFSET  [fp, #24 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -826,6 +836,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/ctrmm_kernel_2x2_vfpv3.S

@@ -66,10 +66,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	r3
+#define OLD_ALPHAI_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define B	[fp, #12 ]
+#define C	[fp, #16 ]
+#define OLD_LDC	[fp, #20 ]
+#define OFFSET  [fp, #24 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -846,6 +856,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/ddot_vfp.S

@@ -246,6 +246,9 @@ ddot_kernel_L999:
 	vldm	r3, { d8 - d15}					// restore floating point registers
 
 	vadd.f64	d0 , d0, d1				// set return value
+#if !defined(__ARM_PCS_VFP)
+	vmov	r0, r1, d0
+#endif
 	sub	sp, fp, #24
 	pop	{r4 - r9, fp}
 	bx	lr

kernel/arm/dgemm_kernel_4x2_vfp.S

@@ -62,10 +62,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
-
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #12 ]
+#define B	[fp, #16 ]
+#define C	[fp, #20 ]
+#define OLD_LDC	[fp, #24 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -429,6 +436,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/dgemm_kernel_4x4_vfpv3.S

@@ -79,9 +79,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #12 ]
+#define B	[fp, #16 ]
+#define C	[fp, #20 ]
+#define OLD_LDC	[fp, #24 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -878,6 +886,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/dtrmm_kernel_4x2_vfp.S

@@ -65,10 +65,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-276 ]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #12 ]
+#define B	[fp, #16 ]
+#define OLD_C	[fp, #20 ]
+#define OLD_LDC	[fp, #24 ]
+#define OFFSET  [fp, #28 ]
+#else
 #define B	[fp, #4 ]
 #define OLD_C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -404,6 +413,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/dtrmm_kernel_4x4_vfpv3.S

@@ -66,10 +66,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-276 ]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	[fp, #4]
+#define OLD_A_SOFTFP	[fp, #12 ]
+#define B	[fp, #16 ]
+#define OLD_C	[fp, #20 ]
+#define OLD_LDC	[fp, #24 ]
+#define OFFSET  [fp, #28 ]
+#else
 #define B	[fp, #4 ]
 #define OLD_C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -846,6 +855,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/gemv_n_vfp.S

@@ -38,11 +38,36 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(DOUBLE)
+#define OLD_ALPHA	r3
+#define OLD_A_SOFTFP	[fp, #0 ]
+#define OLD_LDA		[fp, #4 ]
+#define X		[fp, #8 ]
+#define OLD_INC_X	[fp, #12 ]
+#define Y		[fp, #16 ]
+#define OLD_INC_Y	[fp, #20 ]
+#else
+#define OLD_ALPHA	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define OLD_LDA		[fp, #12]
+#define X		[fp, #16]
+#define OLD_INC_X	[fp, #20]
+#define Y		[fp, #24]
+#define OLD_INC_Y	[fp, #28]
+#endif
+
+#else
+
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+
+#endif
+
 #define OLD_A		r3
 #define	OLD_M		r0
 
@@ -508,6 +533,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	gemvn_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHA
+#else
+	vldr	d0, OLD_ALPHA
+#endif
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_M, M
 

kernel/arm/gemv_n_vfpv3.S

@@ -38,25 +38,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#ifndef ARM_SOFTFP_ABI
-//hard
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
-#define OLD_A		r3
-#else
-#define OLD_A_SOFTFP	[fp, #0 ]
-#define	OLD_LDA		[fp, #4 ]
-#define	X		[fp, #8 ]
-#define	OLD_INC_X	[fp, #12 ]
-#define	Y		[fp, #16 ]
-#define	OLD_INC_Y	[fp, #20 ]
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(DOUBLE)
 #define OLD_ALPHA	r3
-#define OLD_A		r3
+#define OLD_A_SOFTFP	[fp, #0 ]
+#define OLD_LDA		[fp, #4 ]
+#define X		[fp, #8 ]
+#define OLD_INC_X	[fp, #12 ]
+#define Y		[fp, #16 ]
+#define OLD_INC_Y	[fp, #20 ]
+#else
+#define OLD_ALPHA	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define OLD_LDA		[fp, #12]
+#define X		[fp, #16]
+#define OLD_INC_X	[fp, #20]
+#define Y		[fp, #24]
+#define OLD_INC_Y	[fp, #28]
+#endif
+
+#else
+
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+
 #endif
 
+#define OLD_A		r3
 #define	OLD_M		r0
 
 #define AO1	r0
@@ -565,18 +577,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	gemvn_kernel_L999
 
-#ifndef DOUBLE
-#ifdef ARM_SOFTFP_ABI
-
-	vmov	s0, 	OLD_ALPHA
-	ldr	OLD_A,	OLD_A_SOFTFP
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHA
+#else
+	vldr	d0, OLD_ALPHA
 #endif
+	ldr	OLD_A, OLD_A_SOFTFP
 #endif
 
 	str	OLD_A, A
 	str	OLD_M, M
-	
-	
+
 	ldr    INC_X , OLD_INC_X
 	ldr    INC_Y , OLD_INC_Y
 

kernel/arm/gemv_t_vfp.S

@@ -38,25 +38,37 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#ifndef ARM_SOFTFP_ABI
-//hard abi	
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
-#define OLD_A		r3
-#else
-#define OLD_A_SOFTFP	[fp, #0 ]
-#define	OLD_LDA		[fp, #4 ]
-#define	X		[fp, #8 ]
-#define	OLD_INC_X	[fp, #12 ]
-#define	Y		[fp, #16 ]
-#define	OLD_INC_Y	[fp, #20 ]
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(DOUBLE)
 #define OLD_ALPHA	r3
-#define OLD_A		r3
+#define OLD_A_SOFTFP	[fp, #0 ]
+#define OLD_LDA		[fp, #4 ]
+#define X		[fp, #8 ]
+#define OLD_INC_X	[fp, #12 ]
+#define Y		[fp, #16 ]
+#define OLD_INC_Y	[fp, #20 ]
+#else
+#define OLD_ALPHA	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define OLD_LDA		[fp, #12]
+#define X		[fp, #16]
+#define OLD_INC_X	[fp, #20]
+#define Y		[fp, #24]
+#define OLD_INC_Y	[fp, #28]
 #endif
 
+#else
+
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+
+#endif
+
+#define OLD_A		r3
 #define	OLD_N		r1
 
 #define M	r0
@@ -518,11 +530,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	OLD_N, #0
 	ble	gemvt_kernel_L999
 
-#ifndef DOUBLE
-#ifdef ARM_SOFTFP_ABI
-	vmov	s0,	OLD_ALPHA
-	ldr	OLD_A,	OLD_A_SOFTFP
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHA
+#else
+	vldr	d0, OLD_ALPHA
 #endif
+	ldr	OLD_A, OLD_A_SOFTFP
 #endif
 
 	str	OLD_A, A

kernel/arm/gemv_t_vfpv3.S

@@ -38,11 +38,36 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(DOUBLE)
+#define OLD_ALPHA	r3
+#define OLD_A_SOFTFP	[fp, #0 ]
+#define OLD_LDA		[fp, #4 ]
+#define X		[fp, #8 ]
+#define OLD_INC_X	[fp, #12 ]
+#define Y		[fp, #16 ]
+#define OLD_INC_Y	[fp, #20 ]
+#else
+#define OLD_ALPHA	[fp, #0 ]
+#define OLD_A_SOFTFP	[fp, #8 ]
+#define OLD_LDA		[fp, #12]
+#define X		[fp, #16]
+#define OLD_INC_X	[fp, #20]
+#define Y		[fp, #24]
+#define OLD_INC_Y	[fp, #28]
+#endif
+
+#else
+
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+
+#endif
+
 #define OLD_A		r3
 #define	OLD_N		r1
 
@@ -476,6 +501,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	OLD_N, #0
 	ble	gemvt_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	s0, OLD_ALPHA
+#else
+	vldr	d0, OLD_ALPHA
+#endif
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_N, N
 

kernel/arm/nrm2_vfp.S

@@ -573,6 +573,13 @@ nrm2_kernel_L999:
 #else
 	vsqrt.f32	s1, s1
 	vmul.f32	s0, s0, s1
+#endif
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vmov	r0, s0
+#else
+	vmov	r0, r1, d0
+#endif
 #endif
 
 	bx	lr

kernel/arm/nrm2_vfpv3.S

@@ -503,8 +503,13 @@ nrm2_kernel_L999:
 #else
 	vsqrt.f32	s1, s1
 	vmul.f32	s0, s0, s1
-#ifdef ARM_SOFTFP_ABI
-	vmov		r0, s0
+#endif
+
+#if !defined(__ARM_PCS_VFP)
+#if defined(DOUBLE)
+	vmov	r0, r1, d0
+#else
+	vmov	r0, s0
 #endif
 #endif
 

kernel/arm/rot_vfp.S

@@ -40,6 +40,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define	OLD_INC_Y	[fp, #0 ]
 
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+#define	OLD_C		[fp, #4]
+#define	OLD_S		[fp, #8]
+#else
+#define	OLD_C		[fp, #8]
+#define	OLD_S		[fp, #16]
+#endif
+#endif
 
 #define	N	r0
 #define X	r1
@@ -462,7 +471,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         add     fp, sp, #8
 
 	ldr    INC_Y , OLD_INC_Y
-
+#if !defined(__ARM_PCS_VFP)
+#if !defined(DOUBLE)
+	vldr	s0, OLD_C
+	vldr	s1, OLD_S
+#else
+	vldr	d0, OLD_C
+	vldr	d1, OLD_S
+#endif
+#endif
 
 	cmp	N, #0
 	ble	rot_kernel_L999

kernel/arm/sdot_vfp.S

@@ -329,20 +329,19 @@ sdot_kernel_L999:
 	vldm	r3, { s8 - s15}					// restore floating point registers
 
 #if	defined(DSDOT)
-
 	vadd.f64	d0 , d0, d1				// set return value
-
-#ifdef ARM_SOFTFP_ABI
-	vmov	r0, r1, d0
+#else
+	vadd.f32	s0 , s0, s1				// set return value
 #endif
 
+#if !defined(__ARM_PCS_VFP)
+#if defined(DSDOT)
+	vmov	r0, r1, d0
 #else
-
-	vadd.f32	s0 , s0, s1				// set return value
-#ifdef ARM_SOFTFP_ABI
 	vmov	r0, s0
 #endif
 #endif
+
 	sub	sp, fp, #24
 	pop	{r4 - r9, fp}
 	bx	lr

kernel/arm/sgemm_kernel_4x2_vfp.S

@@ -62,9 +62,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -416,6 +424,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/sgemm_kernel_4x4_vfpv3.S

@@ -58,14 +58,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define	OLD_M	r0
 #define	OLD_N	r1
 #define	OLD_K	r2
-
-#ifdef ARM_SOFTFP_ABI
-#define OLD_ALPHA r3
-//#define OLD_A	
-#else //hard
 #define	OLD_A	r3
 #define OLD_ALPHA s0
-#endif
 
 /******************************************************
 * [fp, #-128] - [fp, #-64] is reserved
@@ -77,10 +71,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define M	[fp, #-256 ]
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
-
-#ifndef ARM_SOFTFP_ABI
 #define A	[fp, #-268 ]
-#endif
 
 #define FP_ZERO [fp, #-240]
 #define FP_ZERO_0 [fp, #-240]
@@ -88,17 +79,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA   [fp, #-280]
 
-#ifdef ARM_SOFTFP_ABI
-#define A	[fp, #4 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
 #define B	[fp, #8 ]
 #define C	[fp, #12 ]
 #define OLD_LDC	[fp, #16 ]
-#else //hard
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #endif
-	
+
 #define I	r0
 #define J	r1
 #define L	r2
@@ -867,16 +859,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K
-	
-#ifdef ARM_SOFTFP_ABI
-	str	OLD_ALPHA, ALPHA
-#else //hard
 	str	OLD_A, A
 	vstr	OLD_ALPHA, ALPHA
-#endif
+
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 

kernel/arm/strmm_kernel_4x2_vfp.S

@@ -65,10 +65,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-276 ]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define OLD_C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#define OFFSET  [fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define OLD_C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -395,6 +404,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/strmm_kernel_4x4_vfpv3.S

@@ -64,10 +64,19 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define ALPHA	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHA_SOFTFP	r3
+#define OLD_A_SOFTFP	[fp, #4 ]
+#define B	[fp, #8 ]
+#define C	[fp, #12 ]
+#define OLD_LDC	[fp, #16 ]
+#define OFFSET  [fp, #20 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -782,6 +791,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vmov	OLD_ALPHA, OLD_ALPHA_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/swap_vfp.S

@@ -38,9 +38,43 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
+#if !defined(__ARM_PCS_VFP)
+
+#if !defined(COMPLEX)
+
+#if !defined(DOUBLE)
+#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_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_X		[fp, #4 ]
+#define	OLD_INC_X	[fp, #8 ]
+#define	OLD_Y		[fp, #12 ]
+#define	OLD_INC_Y	[fp, #16 ]
+#else
+#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(__ARM_PCS_VFP)
+
+#else
 #define	OLD_INC_X	[fp, #0 ]
 #define	OLD_Y		[fp, #4 ]
 #define	OLD_INC_Y	[fp, #8 ]
+#endif
 
 
 #define	N	r0
@@ -229,6 +263,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	push    {r4 , fp}
         add     fp, sp, #8
 
+#if !defined(__ARM_PCS_VFP)
+	ldr	X, OLD_X
+#endif
 	ldr    INC_X , OLD_INC_X
 	ldr         Y, OLD_Y
 	ldr    INC_Y , OLD_INC_Y

kernel/arm/zdot_vfp.S

@@ -41,8 +41,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define	N	r0
 #define	X	r1
 #define	INC_X	r2
-#define	OLD_Y	r3
-
 
 /******************************************************
 * [fp, #-128] - [fp, #-64] is reserved
@@ -50,7 +48,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 * registers
 *******************************************************/
 
-#define OLD_INC_Y	[fp, #4 ]
+#if !defined(__ARM_PCS_VFP)
+#define	OLD_RETURN_ADDR	r0
+#define	OLD_N		r1
+#define	OLD_X		r2
+#define	OLD_INC_X	r3
+#define	OLD_Y		[fp, #0 ]
+#define	OLD_INC_Y	[fp, #4 ]
+#define	RETURN_ADDR	r8
+#else
+#define	OLD_Y		r3
+#define	OLD_INC_Y	[fp, #0 ]
+#endif
 
 #define I	r5
 #define Y	r6
@@ -181,7 +190,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	.align 5
 
 	push	{r4 - r9, fp}
-	add	fp, sp, #24
+	add	fp, sp, #28
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
 	sub	r4, fp, #128
@@ -194,9 +203,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vcvt.f64.f32    d2, s0
         vcvt.f64.f32    d3, s0
 
+#if !defined(__ARM_PCS_VFP)
+	mov	RETURN_ADDR, OLD_RETURN_ADDR
+	mov	N, OLD_N
+	mov	X, OLD_X
+	mov	INC_X, OLD_INC_X
+	ldr	Y, OLD_Y
+	ldr	INC_Y, OLD_INC_Y
+#else
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
-
+#endif
 
 	cmp	N, #0
 	ble	zdot_kernel_L999
@@ -280,8 +297,11 @@ zdot_kernel_L999:
 	vadd.f64	d0 , d0, d2
 	vsub.f64	d1 , d1, d3
 #endif
+#if !defined(__ARM_PCS_VFP)
+	vstm	RETURN_ADDR, {d0 - d1}
+#endif
 
-	sub	sp, fp, #24
+	sub	sp, fp, #28
 	pop	{r4 - r9, fp}
 	bx	lr
 

kernel/arm/zgemm_kernel_2x2_vfp.S

@@ -64,9 +64,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	[fp, #4]
+#define OLD_ALPHAI_SOFTFP	[fp, #12]
+#define OLD_A_SOFTFP	[fp, #20 ]
+#define B	[fp, #24 ]
+#define C	[fp, #28 ]
+#define OLD_LDC	[fp, #32 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -863,6 +872,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/zgemm_kernel_2x2_vfpv3.S

@@ -80,9 +80,18 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	[fp, #4]
+#define OLD_ALPHAI_SOFTFP	[fp, #12]
+#define OLD_A_SOFTFP	[fp, #20 ]
+#define B	[fp, #24 ]
+#define C	[fp, #28 ]
+#define OLD_LDC	[fp, #32 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -909,6 +918,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/zgemv_n_vfp.S

@@ -38,11 +38,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR	[fp, #0 ]
+#define OLD_ALPHAI	[fp, #8 ]
+#define OLD_A_SOFTFP	[fp, #16]
+#define OLD_LDA		[fp, #20]
+#define X		[fp, #24]
+#define OLD_INC_X	[fp, #28]
+#define Y		[fp, #32]
+#define OLD_INC_Y	[fp, #36]
+#else
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+#endif
+
 #define OLD_A		r3
 #define	OLD_M		r0
 
@@ -465,6 +477,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	N, #0
 	ble	zgemvn_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	d0, OLD_ALPHAR
+	vldr	d1, OLD_ALPHAI
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_M, M
 	vstr    d0 , ALPHA_R

kernel/arm/zgemv_t_vfp.S

@@ -38,11 +38,23 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define STACKSIZE 256
 
-#define	OLD_LDA		[fp, #0 ]
-#define	X		[fp, #4 ]
-#define	OLD_INC_X	[fp, #8 ]
-#define	Y		[fp, #12 ]
-#define	OLD_INC_Y	[fp, #16 ]
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR	[fp, #0 ]
+#define OLD_ALPHAI	[fp, #8 ]
+#define OLD_A_SOFTFP	[fp, #16]
+#define OLD_LDA		[fp, #20]
+#define X		[fp, #24]
+#define OLD_INC_X	[fp, #28]
+#define Y		[fp, #32]
+#define OLD_INC_Y	[fp, #36]
+#else
+#define OLD_LDA		[fp, #0 ]
+#define X		[fp, #4 ]
+#define OLD_INC_X	[fp, #8 ]
+#define Y		[fp, #12 ]
+#define OLD_INC_Y	[fp, #16 ]
+#endif
+
 #define OLD_A		r3
 #define	OLD_N		r1
 
@@ -360,6 +372,12 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	cmp	OLD_N, #0
 	ble	zgemvt_kernel_L999
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	d0, OLD_ALPHAR
+	vldr	d1, OLD_ALPHAI
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
+
 	str	OLD_A, A
 	str	OLD_N, N
 

kernel/arm/ztrmm_kernel_2x2_vfp.S

@@ -66,10 +66,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	[fp, #4]
+#define OLD_ALPHAI_SOFTFP	[fp, #12]
+#define OLD_A_SOFTFP	[fp, #20 ]
+#define B	[fp, #24 ]
+#define C	[fp, #28 ]
+#define OLD_LDC	[fp, #32 ]
+#define OFFSET  [fp, #36 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -882,6 +892,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K

kernel/arm/ztrmm_kernel_2x2_vfpv3.S

@@ -66,10 +66,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
+#if !defined(__ARM_PCS_VFP)
+#define OLD_ALPHAR_SOFTFP	[fp, #4]
+#define OLD_ALPHAI_SOFTFP	[fp, #12]
+#define OLD_A_SOFTFP	[fp, #20 ]
+#define B	[fp, #24 ]
+#define C	[fp, #28 ]
+#define OLD_LDC	[fp, #32 ]
+#define OFFSET  [fp, #36 ]
+#else
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
 #define OFFSET  [fp, #16 ]
+#endif
 
 #define I	r0
 #define J	r1
@@ -883,6 +893,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	add	fp, sp, #24
 	sub	sp, sp, #STACKSIZE				// reserve stack
 
+#if !defined(__ARM_PCS_VFP)
+	vldr	OLD_ALPHA_R, OLD_ALPHAR_SOFTFP
+	vldr	OLD_ALPHA_I, OLD_ALPHAI_SOFTFP
+	ldr	OLD_A, OLD_A_SOFTFP
+#endif
 	str	OLD_M, M
 	str	OLD_N, N
 	str	OLD_K, K