zosd.asm

一个TIDM642DSP处理器测试程序

      .global _osd
     
_osd:
      lddw .d2 *b4++,a29:a28
   || MV .S1 A6,a1
   || mv .L2 b4,b5
      lddw .d1 *a4++,a31:a30
   || lddw .d2 *b4++,b29:b28    ;++++1
   || zero .l1 a9
   || mv .s1 a4,a5
   
      lddw .d1 *a4++,b31:b30    ;++++1
   || SHR 	.S1 	A1,5,A1	
      SUB .S1 A1,2,A2	
      lddw .d2 *b4++,a21:a20
      lddw .d1 *a4++,a23:a22
   || lddw .d2 *b4++,b21:b20
      lddw .d1 *a4++,b23:b22
      
      ;Get a29,a28 (mask64_1)

   ||cmpeq4 .s1 a9,a29,a27  ;+++++0
      
      
      ;Get a31,a30 (src64_1);Get b29,b28(mask64_2)
 
      xpnd4  .m1 a27,a25     ;+++++0
   || cmpeq4 .s1 a9,a28,a26  ;+++++0
   || cmpeq4 .s2x a9,b29,b27 ;+++++1
   
  
      xpnd4  .m1 a26,a24     ;+++++0
   || xpnd4  .m2 b27,b25     ;+++++1
   || cmpeq4 .s2 a9,b28,b26  ;+++++1
   || cmpeq4 .s1 a9,a21,a19  ;+++++4
LOOP:     ;Get b31,b30 (src64_2)

      and .l1 a25,a31,a31     ;+++++0
   || xpnd4  .m2 b26,b24      ;+++++1 
   || xpnd4  .m1 a19,a17      ;+++++4
   || cmpeq4 .s1 a9,a20,a18   ;+++++5
   || add .d1 A2,1,A1	      ;exit count
   
      and .s1  a24,a30,a30    ;+++++0
   || and .l2 b25,b31,b31     ;+++++1
   || xpnd4  .m1 a18,a16      ;+++++5
   || cmpeq4 .s2 a9,b21,b19   ;+++++6
   ||[a1] lddw .d2 *b4++,a29:a28 ;up2
   || mv  .d1 a28,a7          ;temporary
   || mv  .l1 a29,a8          ;temporary
      or  .l1 a8,a31,a31     ;+++++0
   || and .d2 b24,b30,b30     ;+++++1
   || or  .l2 b29,b31,b31     ;+++++1 
   || xpnd4  .m2 b19,b17      ;+++++6
   || or  .s1 a7,a30,a30
   || cmpeq4 .s2 a9,b20,b18   ;+++++7
      
      or  .l2 b28,b30,b30     ;+++++1
   || and .s1 a17,a23,a23     ;+++++4  
   || and .l1  a16,a22,a22    ;+++++5
   || xpnd4  .m2 b18,b16      ;+++++7
   || stdw .d1 a31:a30,*a5++   ;up1
      or  .l1 a20,a22,a22     ;+++++5
   || and .l2 b17,b23,b23     ;+++++6
   || or  .s1 a21,a23,a23     ;+++++4
   ||[a1] lddw .d1 *a4++,a31:a30  ;loop+2    
   || [a1] lddw .d2 *b4++,b29:b28  ;loop+3
   || [!a1]  b .s2 b3             ;exit
      
      stdw .d1t2 b31:b30,*a5++    ;+++++1
   || or  .l2 b21,b23,b23         ;+++++6
   || and .s2 b16,b22,b22         ;+++++7
   || BDEC 	.S1  LOOP, A2         ;JUMP
   || [a1] lddw .d2 *b4++,a21:a20  ;loop+5
      [a1] lddw .d1 *a4++,b31:b30  ;loop+4 ;up3 
   || or  .l2 b20,b22,b22          ;+++++7
   || [a1] cmpeq4 .s1 a9,a29,a27   ;+++++0
   
      stdw .d1 a23:a22,*a5++       ;+++++5
   || [a1] lddw .d2 *b4++,b21:b20  ;loop+7
   
      stdw .d1 b23:b22,*a5++       ;+++++7 ;up4
  
     [a1]  lddw .d1 *a4++,a23:a22  ;loop+6
   ||[a1]  xpnd4  .m1 a27,a25
   ||[a1]  cmpeq4 .s1 a9,a28,a26
   ||[a1]  cmpeq4 .s2x a9,b29,b27  ;+++++1
   
       lddw .d1 *a4++,b23:b22  ;loop+8
   ||  xpnd4  .m1 a26,a24      ;+++++0
   ||  xpnd4  .m2 b27,b25      ;+++++1
   ||  cmpeq4 .s2 a9,b28,b26   ;+++++1
   ||  cmpeq4 .s1 a9,a21,a19   ;+++++4