expressly use fld.d/fst.d for floating point registers instead of LD/ST macros

kernel/loongarch64/cgemm_kernel_16x4_lasx.S

@@ -196,17 +196,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     SDARG      $r25,   $sp,   16
     SDARG      $r26,   $sp,   24
     SDARG      $r27,   $sp,   32
-    ST         $f23,   $sp,   40
-    ST         $f24,   $sp,   48
-    ST         $f25,   $sp,   56
-    ST         $f26,   $sp,   64
-    ST         $f27,   $sp,   72
-    ST         $f28,   $sp,   80
-    ST         $f29,   $sp,   88
-    ST         $f30,   $sp,   96
-    ST         $f31,   $sp,   104
-    ST         ALPHA_R,$sp,   112
-    ST         ALPHA_I,$sp,   120
+    fst.d         $f23,   $sp,   40
+    fst.d         $f24,   $sp,   48
+    fst.d         $f25,   $sp,   56
+    fst.d         $f26,   $sp,   64
+    fst.d         $f27,   $sp,   72
+    fst.d         $f28,   $sp,   80
+    fst.d         $f29,   $sp,   88
+    fst.d         $f30,   $sp,   96
+    fst.d         $f31,   $sp,   104
+    fst.d         ALPHA_R,$sp,   112
+    fst.d         ALPHA_I,$sp,   120
 
     xvldrepl.w  VALPHAR, $sp, 112
     xvldrepl.w  VALPHAI, $sp, 120
@@ -3741,17 +3741,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     LDARG      $r25,   $sp,   16
     LDARG      $r26,   $sp,   24
     LDARG      $r27,   $sp,   32
-    LD         $f23,   $sp,   40
-    LD         $f24,   $sp,   48
-    LD         $f25,   $sp,   56
-    LD         $f26,   $sp,   64
-    LD         $f27,   $sp,   72
-    LD         $f28,   $sp,   80
-    LD         $f29,   $sp,   88
-    LD         $f30,   $sp,   96
-    LD         $f31,   $sp,   104
+    fld.d         $f23,   $sp,   40
+    fld.d         $f24,   $sp,   48
+    fld.d         $f25,   $sp,   56
+    fld.d         $f26,   $sp,   64
+    fld.d         $f27,   $sp,   72
+    fld.d         $f28,   $sp,   80
+    fld.d         $f29,   $sp,   88
+    fld.d         $f30,   $sp,   96
+    fld.d         $f31,   $sp,   104
 
     addi.d     $sp,    $sp,   128
     jirl       $r0,    $r1,   0x0
 
-    EPILOGUE
+    EPILOGUE

kernel/loongarch64/dgemm_kernel_16x4.S

@@ -1098,16 +1098,16 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     SDARG    $r25,  $sp,   16
     SDARG    $r26,  $sp,   24
     SDARG    $r27,  $sp,   32
-    ST       $f23,  $sp,   40
-    ST       $f24,  $sp,   48
-    ST       $f25,  $sp,   56
-    ST       $f26,  $sp,   64
-    ST       $f27,  $sp,   72
-    ST       $f28,  $sp,   80
-    ST       $f29,  $sp,   88
-    ST       $f30,  $sp,   96
-    ST       $f31,  $sp,   104
-    ST       ALPHA, $sp,   112
+    fst.d       $f23,  $sp,   40
+    fst.d       $f24,  $sp,   48
+    fst.d       $f25,  $sp,   56
+    fst.d       $f26,  $sp,   64
+    fst.d       $f27,  $sp,   72
+    fst.d       $f28,  $sp,   80
+    fst.d       $f29,  $sp,   88
+    fst.d       $f30,  $sp,   96
+    fst.d       $f31,  $sp,   104
+    fst.d       ALPHA, $sp,   112
 
 #if defined (TRMMKERNEL) && !defined(LEFT)
     sub.d   OFF,   ZERO,  OFFSET
@@ -3504,15 +3504,15 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     LDARG    $r25,  $sp,   16
     LDARG    $r26,  $sp,   24
     LDARG    $r27,  $sp,   32
-    LD       $f23,  $sp,   40
-    LD       $f24,  $sp,   48
-    LD       $f25,  $sp,   56
-    LD       $f26,  $sp,   64
-    LD       $f27,  $sp,   72
-    LD       $f28,  $sp,   80
-    LD       $f29,  $sp,   88
-    LD       $f30,  $sp,   96
-    LD       $f31,  $sp,   104
+    fld.d       $f23,  $sp,   40
+    fld.d       $f24,  $sp,   48
+    fld.d       $f25,  $sp,   56
+    fld.d       $f26,  $sp,   64
+    fld.d       $f27,  $sp,   72
+    fld.d       $f28,  $sp,   80
+    fld.d       $f29,  $sp,   88
+    fld.d       $f30,  $sp,   96
+    fld.d       $f31,  $sp,   104
     addi.d   $sp,   $sp,   120
 
     jirl    $r0, $r1, 0x0

kernel/loongarch64/zgemm_kernel_8x4_lasx.S

@@ -196,17 +196,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     SDARG      $r25,   $sp,   16
     SDARG      $r26,   $sp,   24
     SDARG      $r27,   $sp,   32
-    ST         $f23,   $sp,   40
-    ST         $f24,   $sp,   48
-    ST         $f25,   $sp,   56
-    ST         $f26,   $sp,   64
-    ST         $f27,   $sp,   72
-    ST         $f28,   $sp,   80
-    ST         $f29,   $sp,   88
-    ST         $f30,   $sp,   96
-    ST         $f31,   $sp,   104
-    ST         ALPHA_R,$sp,   112
-    ST         ALPHA_I,$sp,   120
+    fst.d         $f23,   $sp,   40
+    fst.d         $f24,   $sp,   48
+    fst.d         $f25,   $sp,   56
+    fst.d         $f26,   $sp,   64
+    fst.d         $f27,   $sp,   72
+    fst.d         $f28,   $sp,   80
+    fst.d         $f29,   $sp,   88
+    fst.d         $f30,   $sp,   96
+    fst.d         $f31,   $sp,   104
+    fst.d         ALPHA_R,$sp,   112
+    fst.d         ALPHA_I,$sp,   120
 
     xvldrepl.d  VALPHAR, $sp, 112
     xvldrepl.d  VALPHAI, $sp, 120
@@ -3529,17 +3529,17 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
     LDARG      $r25,   $sp,   16
     LDARG      $r26,   $sp,   24
     LDARG      $r27,   $sp,   32
-    LD         $f23,   $sp,   40
-    LD         $f24,   $sp,   48
-    LD         $f25,   $sp,   56
-    LD         $f26,   $sp,   64
-    LD         $f27,   $sp,   72
-    LD         $f28,   $sp,   80
-    LD         $f29,   $sp,   88
-    LD         $f30,   $sp,   96
-    LD         $f31,   $sp,   104
+    fld.d         $f23,   $sp,   40
+    fld.d         $f24,   $sp,   48
+    fld.d         $f25,   $sp,   56
+    fld.d         $f26,   $sp,   64
+    fld.d         $f27,   $sp,   72
+    fld.d         $f28,   $sp,   80
+    fld.d         $f29,   $sp,   88
+    fld.d         $f30,   $sp,   96
+    fld.d         $f31,   $sp,   104
 
     addi.d     $sp,    $sp,   128
     jirl       $r0,    $r1,   0x0
 
-    EPILOGUE
+    EPILOGUE