Merge pull request #2312 from martin-frbg/power8be

Further Power8 big-endian corrections

Author: martin-frbg

kernel/power/KERNEL.POWER8

@@ -89,14 +89,30 @@ ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
 #SMINKERNEL   = ../arm/min.c
 #DMINKERNEL   = ../arm/min.c
 #
+ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
 ISAMAXKERNEL = isamax_power8.S
+else
+ISAMAXKERNEL = isamax.c
+endif
 IDAMAXKERNEL = idamax.c
+ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
 ICAMAXKERNEL = icamax_power8.S
+else
+ICAMAXKERNEL = icamax.c
+endif
 IZAMAXKERNEL = izamax.c
 #
+ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
 ISAMINKERNEL = isamin_power8.S
+else
+ISAMINKERNEL = isamin.c
+endif
 IDAMINKERNEL = idamin.c
+ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
 ICAMINKERNEL = icamin_power8.S
+else
+ICAMINKERNEL = icamin.c
+endif
 IZAMINKERNEL = izamin.c
 #
 #ISMAXKERNEL  = ../arm/imax.c
@@ -112,7 +128,11 @@ ZASUMKERNEL  = zasum.c
 #
 SAXPYKERNEL  = saxpy.c
 DAXPYKERNEL  = daxpy.c
+ifneq ($(__BYTE_ORDER__),$(__ORDER_BIG_ENDIAN__))
 CAXPYKERNEL  = caxpy_power8.S
+else
+CAXPYKERNEL  = caxpy.c
+endif
 ZAXPYKERNEL  = zaxpy.c
 #
 SCOPYKERNEL  = scopy.c

kernel/power/caxpy_power8.S

@@ -12,11 +12,12 @@
 
 	PROLOGUE
 
-caxpy_k:
 .LCF0:
 0:	addis 2,12,.TOC.-.LCF0@ha
 	addi 2,2,.TOC.-.LCF0@l
+#if _CALL_ELF ==2
 	.localentry	caxpy_k,.-caxpy_k
+#endif
 	mr. 7,3
 	ble 0,.L33
 	cmpdi 7,9,1
@@ -515,7 +516,9 @@ caxpy_k:
 	b .L13
 	.long 0
 	.byte 0,0,0,0,0,4,0,0
+#if _CALL_ELF ==2
 	.size	caxpy_k,.-caxpy_k
+#endif
 	.section	.rodata
 	.align 4
 	.set	.LANCHOR0,. + 0

kernel/power/icamin_power8.S

@@ -11,11 +11,12 @@
 
 	PROLOGUE
 
-icamin_k:
 .LCF0:
 0:	addis 2,12,.TOC.-.LCF0@ha
 	addi 2,2,.TOC.-.LCF0@l
+#if _CALL_ELF ==2
 	.localentry	icamin_k,.-icamin_k
+#endif
 	mr. 9,3
 	ble 0,.L25
 	cmpdi 7,5,0
@@ -388,7 +389,9 @@ icamin_k:
 	b .L21
 	.long 0
 	.byte 0,0,0,0,0,1,0,0
+#if _CALL_ELF ==2
 	.size	icamin_k,.-icamin_k
+#endif
 	.section	.rodata.cst16,"aM",@progbits,16
 	.align 4
 .LC2:

kernel/power/idamax.c

@@ -324,15 +324,15 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 
     if (inc_x == 1) {
 
-#if defined(_CALL_ELF) && (_CALL_ELF == 2)
         BLASLONG n1 = n & -32;
-        if (n1 > 0) {
+#if defined(_CALL_ELF) && (_CALL_ELF == 2)
+	if (n1 > 0) {
 
             max = diamax_kernel_32(n1, x, &maxf);
 
             i = n1;
         }
-#endif
+#endif	
         while (i < n) {
             if (ABS(x[i]) > maxf) {
                 max = i;

kernel/power/idamin.c

@@ -328,13 +328,12 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x) {
 
 #if defined(_CALL_ELF) && (_CALL_ELF == 2)
         BLASLONG n1 = n & -32;
-        if (n1 > 0) {
+	if (n1 > 0) {
 
             min = diamin_kernel_32(n1, x, &minf);
             i = n1;
         }
 #endif
- 
         while (i < n) {
             if (ABS(x[i]) < minf) {
                 min = i;

kernel/power/isamax_power8.S

@@ -12,11 +12,12 @@
 
 	PROLOGUE
 
-isamax_k:
 .LCF0:
 0:	addis 2,12,.TOC.-.LCF0@ha
 	addi 2,2,.TOC.-.LCF0@l
+#if _CALL_ELF ==2
 	.localentry	isamax_k,.-isamax_k
+#endif
 	mr. 11,3
 	ble 0,.L36
 	cmpdi 7,5,0
@@ -397,7 +398,9 @@ isamax_k:
 	b .L61
 	.long 0
 	.byte 0,0,0,0,0,1,0,0
+#if _CALL_ELF ==2
 	.size	isamax_k,.-isamax_k
+#endif
 	.section	.rodata.cst16,"aM",@progbits,16
 	.align 4
 .LC2:

kernel/power/isamin_power8.S

@@ -11,11 +11,12 @@
 
 	PROLOGUE
 
-isamin_k:
 .LCF0:
 0:	addis 2,12,.TOC.-.LCF0@ha
 	addi 2,2,.TOC.-.LCF0@l
+#if _CALL_ELF ==2
 	.localentry	isamin_k,.-isamin_k
+#endif
 	mr. 11,3
 	ble 0,.L36
 	cmpdi 7,5,0
@@ -380,7 +381,9 @@ isamin_k:
 	b .L35
 	.long 0
 	.byte 0,0,0,0,0,1,0,0
+#if _CALL_ELF ==2
 	.size	isamin_k,.-isamin_k
+#endif
 	.section	.rodata.cst16,"aM",@progbits,16
 	.align 4
 .LC2:

kernel/power/izamin.c

@@ -316,14 +316,14 @@ BLASLONG CNAME(BLASLONG n, FLOAT *x, BLASLONG inc_x)
         minf = CABS1(x,0); //index will not be incremented
 
 #if defined(_CALL_ELF) && (_CALL_ELF == 2)
-        BLASLONG n1 = n & -16;
+	BLASLONG n1 = n & -16;
         if (n1 > 0) {
 
             min = ziamin_kernel_16_TUNED(n1, x, &minf);
             i = n1;
             ix = n1 << 1;
         }
-#endif      
+#endif 
 
         while(i < n)
         {