An inlined cast from int32_t to float32_t may get an incorrect optimization

XMLWordPrintable

    • Type: Bug
    • Resolution: Fixed
    • Priority: Medium
    • Code Generation Tools
    • CODEGEN-12270
    • Hide
      C2000_22.6.0.LTS
      C2000_22.6.1.LTS
      C2000_18.12.2.LTS
      C2000_21.6.0.LTS
      C2000_20.2.0.LTS
      Show
      C2000_22.6.0.LTS C2000_22.6.1.LTS C2000_18.12.2.LTS C2000_21.6.0.LTS C2000_20.2.0.LTS
    • Hide
      C2000_22.6.2.LTS*
      C2000_NEXT*
      C2000_21.6.2.LTS*
      Show
      C2000_22.6.2.LTS* C2000_NEXT* C2000_21.6.2.LTS*
    • default
    • Hide
      For the example case in description, replacing the inlined call to ftol() with a cast to float32_t was enough to avoid the bug. Also, encountering the bug depends on the right code shape and context in order for the incorrect optimization to be used:

      From this:
          float32_t flt = ftol(global_int32);
      to this:
          float32_t flt = (float32_t)global_int32;
      Show
      For the example case in description, replacing the inlined call to ftol() with a cast to float32_t was enough to avoid the bug. Also, encountering the bug depends on the right code shape and context in order for the incorrect optimization to be used: From this:     float32_t flt = ftol(global_int32); to this:     float32_t flt = (float32_t)global_int32;

      An inlined cast from int32_t to float32_t may get an incorrect optimization.

      For below cutdown C code:

      inline static float32_t ftol(int32_t input){
  return (float32_t)input;
}
void foo() 
{
    global_int32 = bar(...);
    ... more code needed for context
    float32_t flt = ftol(global_int32);
}

      The correct generated assembly should be:

              MOVL      XAR4,#3671270        
        ...
        MOVL      XAR4,*+XAR4[0]      
        ...
        MOV32     R0H,XAR4 
        ...
        MOVL      @||global_int32||,XAR4 
        ...
        I32TOF32  R0H,R0H      
        ...
        MOV32     @||flt||,R0H

      However, an incorrect optimization may generate below which is invalid since the memory load was based on XAR4 which is the same register:

              MOVL      XAR4,#3671270 
        ...
        MOVL      XAR4,*+XAR4[0]
        ...
        MOVL      @||global_int32||,XAR4 
        ...
        MOV32     R0H,*+XAR4[0]  ; <--- incorrect
      ; above should be:  MOV32 R0H,XAR4
        ...
        I32TOF32  R0H,R0H        
        ...
        MOV32     @||flt||,R0H

      Above incorrect optimization can run with --opt_level=off,0,1,2,3,4

            Assignee:
            TI User
            Reporter:
            TI User
            Votes:
            0 Vote for this issue
            Watchers:
            2 Start watching this issue

              Created:
              Updated:
              Resolved:

                Connection: Intermediate to External PROD System
                EXTSYNC-4482 - An inlined cast from int32_t to flo...
                SYNCHRONIZED
                • Last Sync Date: