Merge pull request #761 from wernsaar/develop

Ref #740: all assembly codes now clear floating point register correctly

kernel/arm/KERNEL.ARMV6

@@ -1,26 +1,4 @@
-SGEMVNKERNEL = ../arm/gemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-CGEMVNKERNEL = ../arm/zgemv_n.c
-CGEMVTKERNEL = ../arm/zgemv_t.c
 
-DGEMVNKERNEL = ../arm/gemv_n.c
-DGEMVTKERNEL = ../arm/gemv_t.c
-
-CTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-CGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-CGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-CGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-
-#ZTRMMKERNEL	= ../generic/ztrmmkernel_2x2.c
-#ZGEMMKERNEL    = ../generic/zgemmkernel_2x2.c
-#ZGEMMONCOPY    = ../generic/zgemm_ncopy_2.c
-#ZGEMMOTCOPY    = ../generic/zgemm_tcopy_2.c
-
-
-#STRMMKERNEL	= ../generic/trmmkernel_2x2.c
-#SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
-#SGEMMONCOPY    =  ../generic/gemm_ncopy_2.c
-#SGEMMOTCOPY    =  ../generic/gemm_tcopy_2.c
 
 
 ###############################################################################
@@ -96,19 +74,19 @@ DSWAPKERNEL  = swap_vfp.S
 CSWAPKERNEL  = swap_vfp.S
 ZSWAPKERNEL  = swap_vfp.S
 
-# BAD SGEMVNKERNEL = gemv_n_vfp.S
-# BAD DGEMVNKERNEL = gemv_n_vfp.S
-# CGEMVNKERNEL = cgemv_n_vfp.S
+SGEMVNKERNEL = gemv_n_vfp.S
+DGEMVNKERNEL = gemv_n_vfp.S
+CGEMVNKERNEL = cgemv_n_vfp.S
 ZGEMVNKERNEL = zgemv_n_vfp.S
 
-# BAD SGEMVTKERNEL = gemv_t_vfp.S
-# BAD DGEMVTKERNEL = gemv_t_vfp.S
-# CGEMVTKERNEL = cgemv_t_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_4x2_vfp.S
 DTRMMKERNEL	= dtrmm_kernel_4x2_vfp.S
-#CTRMMKERNEL	= ctrmm_kernel_2x2_vfp.S
+CTRMMKERNEL	= ctrmm_kernel_2x2_vfp.S
 ZTRMMKERNEL	= ztrmm_kernel_2x2_vfp.S
 
 SGEMMKERNEL    = sgemm_kernel_4x2_vfp.S
@@ -131,9 +109,9 @@ DGEMMOTCOPY    = ../generic/gemm_tcopy_2.c
 DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 
-#CGEMMKERNEL    = cgemm_kernel_2x2_vfp.S
-#CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
-#CGEMMOTCOPY    = cgemm_tcopy_2_vfp.S
+CGEMMKERNEL    = cgemm_kernel_2x2_vfp.S
+CGEMMONCOPY    = cgemm_ncopy_2_vfp.S
+CGEMMOTCOPY    = cgemm_tcopy_2_vfp.S
 CGEMMONCOPYOBJ = cgemm_oncopy.o
 CGEMMOTCOPYOBJ = cgemm_otcopy.o
 

kernel/arm/KERNEL.ARMV7

@@ -1,8 +1,3 @@
-SGEMVNKERNEL = ../arm/gemv_n.c
-SGEMVTKERNEL = ../arm/gemv_t.c
-CGEMVNKERNEL = ../arm/zgemv_n.c
-CGEMVTKERNEL = ../arm/zgemv_t.c
-
 
 #################################################################################
 SAMAXKERNEL  = iamax_vfp.S
@@ -77,14 +72,14 @@ DSCALKERNEL  = scal.c
 CSCALKERNEL  = zscal.c
 ZSCALKERNEL  = zscal.c
 
-# BAD SGEMVNKERNEL = gemv_n_vfp.S
-DGEMVNKERNEL = gemv_n_vfp.S
-#CGEMVNKERNEL = cgemv_n_vfp.S
+SGEMVNKERNEL = gemv_n_vfpv3.S
+DGEMVNKERNEL = gemv_n_vfpv3.S
+CGEMVNKERNEL = cgemv_n_vfp.S
 ZGEMVNKERNEL = zgemv_n_vfp.S
 
-# BAD SGEMVTKERNEL = gemv_t_vfp.S
+SGEMVTKERNEL = gemv_t_vfp.S
 DGEMVTKERNEL = gemv_t_vfp.S
-#CGEMVTKERNEL = cgemv_t_vfp.S
+CGEMVTKERNEL = cgemv_t_vfp.S
 ZGEMVTKERNEL = zgemv_t_vfp.S
 
 STRMMKERNEL  =  strmm_kernel_4x4_vfpv3.S
@@ -92,24 +87,15 @@ DTRMMKERNEL  =  dtrmm_kernel_4x4_vfpv3.S
 CTRMMKERNEL  =  ctrmm_kernel_2x2_vfpv3.S
 ZTRMMKERNEL  =  ztrmm_kernel_2x2_vfpv3.S
 
-#SGEMMKERNEL    =  ../generic/gemmkernel_2x2.c
 SGEMMKERNEL    =  sgemm_kernel_4x4_vfpv3.S
-SGEMMINCOPY    =
-SGEMMITCOPY    =
 SGEMMONCOPY    =  sgemm_ncopy_4_vfp.S
 SGEMMOTCOPY    =  sgemm_tcopy_4_vfp.S
-SGEMMINCOPYOBJ =
-SGEMMITCOPYOBJ =
 SGEMMONCOPYOBJ =  sgemm_oncopy.o
 SGEMMOTCOPYOBJ =  sgemm_otcopy.o
 
 DGEMMKERNEL    =  dgemm_kernel_4x4_vfpv3.S
-DGEMMINCOPY    =
-DGEMMITCOPY    =
 DGEMMONCOPY    =  dgemm_ncopy_4_vfp.S
 DGEMMOTCOPY    =  dgemm_tcopy_4_vfp.S
-DGEMMINCOPYOBJ =
-DGEMMITCOPYOBJ =
 DGEMMONCOPYOBJ = dgemm_oncopy.o
 DGEMMOTCOPYOBJ = dgemm_otcopy.o
 

kernel/arm/cdot_vfp.S

@@ -185,14 +185,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r4, fp, #128
 	vstm	r4, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0                                          // clear floating point register
+        vmov    s0, r4
+	vmov	s1, s0
+	vmov	s2, s0
+	vmov	s3, s0
+
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
-	vsub.f32                s0 , s0 , s0
-	vsub.f32                s1 , s1 , s1
-	vsub.f32                s2 , s2 , s2
-	vsub.f32                s3 , s3 , s3
-
 	cmp	N, #0
 	ble	cdot_kernel_L999
 

kernel/arm/cgemm_kernel_2x2_vfp.S

@@ -57,6 +57,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -138,7 +142,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -340,7 +344,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -514,7 +518,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -681,7 +685,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -822,6 +826,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/cgemm_kernel_2x2_vfpv3.S

@@ -73,6 +73,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -550,7 +554,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -730,7 +734,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s24, s16
 	vmov.f32		s25, s16
@@ -879,6 +883,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/cgemv_n_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define ALPHA_I [fp, #-236]
 #define ALPHA_R [fp, #-244]
 
@@ -117,7 +121,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro INIT_F4
 
 	pld	[ YO, #Y_PRE ]
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
         vmov.f32                s10, s8
         vmov.f32                s11, s8
@@ -220,7 +224,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
 
 .endm
@@ -267,7 +271,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
         vmov.f32                s10, s8
         vmov.f32                s11, s8
@@ -384,7 +388,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-        vsub.f32                s8 , s8 , s8
+	flds			s8 , FP_ZERO
         vmov.f32                s9 , s8
 
 .endm
@@ -448,6 +452,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	cgemvn_kernel_L999
 

kernel/arm/cgemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -116,10 +120,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
-	vsub.f32	s14, s14, s14
-	vsub.f32	s15, s15, s15
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
+	vmov.f32	s14, s12
+	vmov.f32	s15, s12
 
 .endm
 
@@ -172,8 +176,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
 
 .endm
 
@@ -215,10 +219,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
-	vsub.f32	s14, s14, s14
-	vsub.f32	s15, s15, s15
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
+	vmov.f32	s14, s12
+	vmov.f32	s15, s12
 
 .endm
 
@@ -281,8 +285,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s12, s12, s12
-	vsub.f32	s13, s13, s13
+	flds		s12, FP_ZERO
+	vmov.f32	s13, s12
 
 .endm
 
@@ -345,6 +349,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	cgemvt_kernel_L999
 

kernel/arm/ctrmm_kernel_2x2_vfp.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -136,7 +141,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -301,10 +306,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -318,10 +323,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmias CO1, { s4 - s7 }
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s12
 	FMAC_I1 s5 , s0 , s13
@@ -343,7 +348,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -490,8 +495,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -500,8 +505,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmias CO1, { s4 - s5 }
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s12
 	FMAC_I1 s5 , s0 , s13
@@ -519,7 +524,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -663,10 +668,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
-	vsub.f32	s6, s6, s6
-	vsub.f32	s7, s7, s7
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
+	vmov.f32	s6, s4
+	vmov.f32	s7, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -689,7 +694,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -795,8 +800,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	flds		s0, ALPHA_R
 	flds		s1, ALPHA_I
 
-	vsub.f32	s4, s4, s4
-	vsub.f32	s5, s5, s5
+	flds		s4, FP_ZERO
+	vmov.f32	s5, s4
 
 	FMAC_R1 s4 , s0 , s8
 	FMAC_I1 s5 , s0 , s9
@@ -831,6 +836,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/ctrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -351,7 +355,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -529,7 +533,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -706,7 +710,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16 , FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s24, s16
 	vmov.f32		s25, s16
@@ -852,6 +856,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 4 * 2
 	str	r3, LDC

kernel/arm/dgemm_kernel_4x2_vfp.S

@@ -56,8 +56,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
+
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
 #define OLD_LDC	[fp, #12 ]
@@ -85,7 +90,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -173,7 +178,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -233,7 +238,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d12, d8
 
 .endm
@@ -283,7 +288,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -338,7 +343,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -380,7 +385,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8, FP_ZERO
 
 .endm
 
@@ -433,6 +438,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/dtrmm_kernel_4x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -90,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -165,7 +169,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -220,7 +224,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d12, d8
 
 .endm
@@ -268,7 +272,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9, d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -318,7 +322,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -357,7 +361,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 
 .endm
 
@@ -409,6 +413,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/dtrmm_kernel_4x4_vfpv3.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -89,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -386,7 +391,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -468,7 +473,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 	vmov.f64		d24, d16
 	vmov.f64		d28, d16
@@ -527,7 +532,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -601,7 +606,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -656,7 +661,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d20, d16
 
 .endm
@@ -699,7 +704,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -753,7 +758,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 
 .endm
@@ -794,7 +799,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 
 .endm
 
@@ -850,6 +855,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #3					// ldc = ldc * 8
 	str	r3, LDC

kernel/arm/gemv_n_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define M	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -79,7 +83,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	pld     [ YO , #Y_PRE ]
 	pld     [ YO , #Y_PRE+32 ]
 
-	vsub.f64	d8 , d8 , d8
+	fldd		d8 , FP_ZERO
 	vmov.f64	d9  , d8
 	vmov.f64	d10 , d8
 	vmov.f64	d11 , d8
@@ -158,7 +162,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 
 .endm
 
@@ -185,7 +189,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 	vmov.f64	d13 , d12
 	vmov.f64	d14 , d12
 	vmov.f64	d15 , d12
@@ -245,7 +249,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d12 , d12 , d12
+	fldd		d12 , FP_ZERO
 
 .endm
 
@@ -279,7 +283,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	pld     [ YO , #Y_PRE ]
 
-	vsub.f32	s8 , s8 , s8
+	flds		s8 , FP_ZERO
 	vmov.f32	s9  , s8
 	vmov.f32	s10 , s8
 	vmov.f32	s11 , s8
@@ -357,7 +361,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 
 .endm
 
@@ -384,7 +388,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 	vmov.f32	s13 , s12
 	vmov.f32	s14 , s12
 	vmov.f32	s15 , s12
@@ -445,7 +449,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s12 , s12 , s12
+	flds		s12 , FP_ZERO
 
 .endm
 
@@ -494,6 +498,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	gemvn_kernel_L999
 

kernel/arm/gemv_n_vfpv3.S

@@ -62,6 +62,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define M	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 
 #define X_PRE	64
 #define Y_PRE	0
@@ -79,7 +83,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	pld     [ YO , #Y_PRE ]
 	pld     [ YO , #Y_PRE+32 ]
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 	vmov.f64	d25 , d24
 	vmov.f64	d26 , d24
 	vmov.f64	d27 , d24
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 
 .endm
 
@@ -175,7 +179,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S8
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 	vmov.f64	d25 , d24
 	vmov.f64	d26 , d24
 	vmov.f64	d27 , d24
@@ -269,7 +273,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d24 , d24 , d24
+	fldd		d24 , FP_ZERO
 
 .endm
 
@@ -302,7 +306,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	pld     [ YO , #Y_PRE ]
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 	vmov.f32	s25 , s24
 	vmov.f32	s26 , s24
 	vmov.f32	s27 , s24
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 
 .endm
 
@@ -396,7 +400,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S8
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 	vmov.f32	s25 , s24
 	vmov.f32	s26 , s24
 	vmov.f32	s27 , s24
@@ -489,7 +493,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s24 , s24 , s24
+	flds		s24 , FP_ZERO
 
 .endm
 
@@ -538,6 +542,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s31 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	gemvn_kernel_L999
 

kernel/arm/gemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -75,8 +79,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f64	d2 , d2 , d2
-	vsub.f64	d3 , d3 , d3
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -123,7 +127,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d2 , d2 , d2
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -160,8 +165,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f64	d2 , d2 , d2
-	vsub.f64	d3 , d3 , d3
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -224,7 +229,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d2 , d2 , d2
+	fldd		d2, FP_ZERO
+	vmov.f64	d3 , d2 
 
 .endm
 
@@ -276,8 +282,9 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f32	s2 , s2 , s2
-	vsub.f32	s3 , s3 , s3
+	flds		s2 , FP_ZERO
+	vmov.f32	s3 , s2 
+
 
 .endm
 
@@ -321,7 +328,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f32	s2 , s2 , s2
+	flds		s2 , FP_ZERO
 
 .endm
 
@@ -356,8 +363,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f32	s2 , s2 , s2
-	vsub.f32	s3 , s3 , s3
+	flds		s2 , FP_ZERO
+	vmov.f32	s3 , s2 
 
 .endm
 
@@ -418,7 +425,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f32	s2 , s2 , s2
+	flds		s2 , FP_ZERO
 
 .endm
 
@@ -488,6 +495,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	gemvt_kernel_L999
 

kernel/arm/nrm2_vfpv3.S

@@ -405,12 +405,15 @@ KERNEL_S1_END_\@:
 	.align 5
 
 #if defined(DOUBLE)
-	vsub.f64                d0 , d0 , d0		// scale=0.0
+	movs			r12 , #0
+	vmov.f32		s0 , r12		// scale=0.0
+	vcvt.f64.f32    d0, s0
 	vmov.f64		d1 , #1.0		// ssq=1.0
 	vmov.f64		d7 , d1			// value 1.0
 	vmov.f64		d6 , d0			// value 0.0
 #else
-	vsub.f32                s0 , s0 , s0		// scale=0.0
+	movs			r12 , #0
+	vmov.f32		s0 , r12		// scale=0.0
 	vmov.f32		s1 , #1.0		// ssq=1.0
 	vmov.f32		s7 , s1			// value 1.0
 	vmov.f32		s6 , s0			// value 0.0

kernel/arm/sgemm_kernel_4x2_vfp.S

@@ -56,6 +56,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -85,7 +89,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -161,7 +165,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -221,7 +225,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s12, s8
 
 .endm
@@ -271,7 +275,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -326,7 +330,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -368,7 +372,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8, FP_ZERO
 
 .endm
 
@@ -421,6 +425,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/sgemm_kernel_4x4_vfpv3.S

@@ -73,7 +73,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
-#define ALPHA	[fp, #-280]
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, #-240]
+#define FP_ZERO_1 [fp, #-236]
+
+#define ALPHA   [fp, #-280]
 
 #define B	[fp, #4 ]
 #define C	[fp, #8 ]
@@ -102,7 +106,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -349,7 +353,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -443,7 +447,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s20, s16
 	vmov.f32		s24, s16
 	vmov.f32		s28, s16
@@ -506,7 +510,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -590,7 +594,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -651,7 +655,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s20, s16
 
 .endm
@@ -696,7 +700,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -755,7 +759,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 	vmov.f32		s17, s16
 
 .endm
@@ -799,7 +803,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			s16, FP_ZERO
 
 .endm
 
@@ -856,6 +860,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/strmm_kernel_4x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA	[fp, #-276 ]
 
 #define B	[fp, #4 ]
@@ -90,7 +94,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -156,7 +160,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s12, s8
 	vmov.f32		s13, s8
@@ -211,7 +215,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s12, s8
 
 .endm
@@ -259,7 +263,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9, s8
 	vmov.f32		s10, s8
 	vmov.f32		s11, s8
@@ -309,7 +313,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 	vmov.f32		s9 , s8
 
 .endm
@@ -348,7 +352,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s8 , s8 , s8
+	flds			s8 , FP_ZERO
 
 .endm
 
@@ -400,6 +404,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/strmm_kernel_4x4_vfpv3.S

@@ -58,6 +58,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define K	[fp, #-264 ]
 #define A	[fp, #-268 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA	[fp, #-280]
 
 #define B	[fp, #4 ]
@@ -88,7 +92,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -322,7 +326,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -405,7 +409,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x4
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 	vmov.f32		s24, s16
 	vmov.f32		s28, s16
@@ -464,7 +468,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -538,7 +542,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s20, s16
 	vmov.f32		s21, s16
@@ -593,7 +597,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s20, s16
 
 .endm
@@ -636,7 +640,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT4x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 	vmov.f32		s18, s16
 	vmov.f32		s19, s16
@@ -690,7 +694,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 	vmov.f32		s17, s16
 
 .endm
@@ -731,7 +735,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f32		s16 , s16 , s16
+	flds			S16, FP_ZERO
 
 .endm
 
@@ -787,6 +791,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { s8 - s31} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #2					// ldc = ldc * 4
 	str	r3, LDC

kernel/arm/zdot_vfp.S

@@ -187,13 +187,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r4, fp, #128
 	vstm	r4, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0                                          // clear floating point register
+        vmov    s0, r4
+        vcvt.f64.f32    d0, s0
+        vcvt.f64.f32    d1, s0
+        vcvt.f64.f32    d2, s0
+        vcvt.f64.f32    d3, s0
+
 	mov	Y, OLD_Y
 	ldr	INC_Y, OLD_INC_Y
 
-	vsub.f64                d0 , d0 , d0
-	vsub.f64                d1 , d1 , d1
-	vsub.f64                d2 , d2 , d2
-	vsub.f64                d3 , d3 , d3
 
 	cmp	N, #0
 	ble	zdot_kernel_L999

kernel/arm/zgemm_kernel_2x2_vfp.S

@@ -57,6 +57,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -131,7 +135,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -383,7 +387,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -557,7 +561,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -724,7 +728,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -869,6 +873,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/zgemm_kernel_2x2_vfpv3.S

@@ -73,6 +73,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-240]
+#define FP_ZERO_0 [fp, # -240]
+#define FP_ZERO_1 [fp, # -236]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -147,7 +151,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -404,7 +408,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -586,7 +590,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -766,7 +770,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16, FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d24, d16
 	vmov.f64		d25, d16
@@ -915,6 +919,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/zgemv_n_vfp.S

@@ -59,6 +59,11 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
+
 #define ALPHA_I [fp, #-236]
 #define ALPHA_R [fp, #-244]
 
@@ -117,7 +122,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 .macro INIT_F4
 
 	pld	[ YO, #Y_PRE ]
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
         vmov.f64                d10, d8
         vmov.f64                d11, d8
@@ -222,7 +227,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
 
 .endm
@@ -269,7 +274,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S4
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
         vmov.f64                d10, d8
         vmov.f64                d11, d8
@@ -386,7 +391,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-        vsub.f64                d8 , d8 , d8
+	fldd			d8, FP_ZERO
         vmov.f64                d9 , d8
 
 .endm
@@ -450,6 +455,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	OLD_M, #0
 	ble	zgemvn_kernel_L999
 

kernel/arm/zgemv_t_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 #define I	r12
 
+#define FP_ZERO [fp, #-228]
+#define FP_ZERO_0 [fp, #-228]
+#define FP_ZERO_1 [fp, #-224]
+
 #define N	[fp, #-252 ]
 #define A	[fp, #-256 ]
 
@@ -117,10 +121,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F2
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
-	vsub.f64	d14, d14, d14
-	vsub.f64	d15, d15, d15
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
+	vmov.f64	d14, d12
+	vmov.f64	d15, d12
 
 .endm
 
@@ -173,8 +177,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_F1
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
 
 .endm
 
@@ -216,10 +220,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S2
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
-	vsub.f64	d14, d14, d14
-	vsub.f64	d15, d15, d15
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
+	vmov.f64	d14, d12
+	vmov.f64	d15, d12
 
 .endm
 
@@ -282,8 +286,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT_S1
 
-	vsub.f64	d12, d12, d12
-	vsub.f64	d13, d13, d13
+	fldd		d12, FP_ZERO
+	vmov.f64	d13, d12
 
 .endm
 
@@ -346,6 +350,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
         vstm    r12, { s8 - s15 }                                 // store floating point registers
 #endif
 
+        movs    r12, #0
+        str     r12, FP_ZERO
+        str     r12, FP_ZERO_1
+
 	cmp	M, #0
 	ble	zgemvt_kernel_L999
 

kernel/arm/ztrmm_kernel_2x2_vfp.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-232]
+#define FP_ZERO_0 [fp, #-232]
+#define FP_ZERO_1 [fp, #-228]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -140,7 +144,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -356,10 +360,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -373,10 +377,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmiad CO1, { d4 - d7 }
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d12
 	FMAC_I1 d5 , d0 , d13
@@ -398,7 +402,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d12, d8
 	vmov.f64		d13, d8
@@ -545,8 +549,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -555,8 +559,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 	fstmiad CO1, { d4 - d5 }
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d12
 	FMAC_I1 d5 , d0 , d13
@@ -574,7 +578,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 	vmov.f64		d10, d8
 	vmov.f64		d11, d8
@@ -718,10 +722,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
-	vsub.f64	d6, d6 , d6
-	vsub.f64	d7, d7 , d7
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
+	vmov.f64	d6 , d4
+	vmov.f64	d7 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -744,7 +748,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d8 , d8 , d8
+	fldd			d8 , FP_ZERO
 	vmov.f64		d9 , d8
 
 .endm
@@ -850,8 +854,8 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	fldd		d0, ALPHA_R
 	fldd		d1, ALPHA_I
 
-	vsub.f64	d4, d4 , d4
-	vsub.f64	d5, d5 , d5
+	fldd		d4 , FP_ZERO
+	vmov.f64	d5 , d4
 
 	FMAC_R1 d4 , d0 , d8
 	FMAC_I1 d5 , d0 , d9
@@ -888,6 +892,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

kernel/arm/ztrmm_kernel_2x2_vfpv3.S

@@ -59,6 +59,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define N	[fp, #-260 ]
 #define K	[fp, #-264 ]
 
+#define FP_ZERO [fp, #-236]
+#define FP_ZERO_0 [fp, #-236]
+#define FP_ZERO_1 [fp, #-232]
+
 #define ALPHA_I	[fp, #-272]
 #define ALPHA_R	[fp, #-280]
 
@@ -134,7 +138,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -388,7 +392,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x2
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d20, d16
 	vmov.f64		d21, d16
@@ -566,7 +570,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT2x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d18, d16
 	vmov.f64		d19, d16
@@ -743,7 +747,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 
 .macro INIT1x1
 
-	vsub.f64		d16 , d16 , d16
+	fldd			d16 , FP_ZERO
 	vmov.f64		d17, d16
 	vmov.f64		d24, d16
 	vmov.f64		d25, d16
@@ -889,6 +893,10 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 	sub	r3, fp, #128
 	vstm	r3, { d8 - d15} 				// store floating point registers
 
+        movs    r4, #0
+        str     r4, FP_ZERO
+        str     r4, FP_ZERO_1
+
 	ldr	r3, OLD_LDC
 	lsl	r3, r3, #4					// ldc = ldc * 8 * 2
 	str	r3, LDC

lapack-netlib/TESTING/nep.in

@@ -10,7 +10,7 @@ NEP:  Data file for testing Nonsymmetric Eigenvalue Problem routines
   0   5   7    3  200             Values of INIBL  (nibble crossover point)
   1   2   4    2    1             Values of ISHFTS (number of simultaneous shifts)
   0   1   2    0    1             Values of IACC22 (select structured matrix multiply: 0, 1 or 2)
-20.0                              Threshold value
+30.0                              Threshold value
 T                                 Put T to test the error exits
 1                                 Code to interpret the seed
 NEP  21