Why does the generated code incorporate inefficient bit shifting operations in Real-Time Workshop Embedded Coder 5.0.1 (R2007b+)?

I noticed that when multiplying a signal by a power of 2 (via a Gain block), the generated code shows:

 (val << 14) >> 12

instead of:

 val << 2

To reproduce the issue, generate code for model, mux_to_array.mdl, found in the attachments. The generated code has the following line:

 mux_to_array_Y.Out1 = (int16_T)((((mux_to_array_U.In1 << 2U) << 14) >> 13) * 3);

Answer 1

[MathWorks Support Team]

The generated code is not incorrect. Notice the following comments in the generated code:

    96       /* Outport: '<Root>/Out 1' incorporates:
    97        *  Gain: '<Root>/Gain 1'
    98        *  Gain: '<Root>/Gain 2'
    99        *  Gain: '<Root>/Gain 5'
   100        *  Gain: '<Root>/Gain 6'
   101        *  Inport: '<Root>/In 1'
   102        *  SignalConversion: '<Root>/TmpHiddenBufferAtGain 1Inport1'
   103        */
   104       mux_to_array_Y.Out1 = (int16_T)((((mux_to_array_U.In1 << 2U) << 14) >> 13) *
   105         3);

Each gain is an operation y = fn(u). These functions are being folded to make one line of code equivalent to

y = f1(f2(f3(f4(u))))

Hence the reason for the behavior observed. Most likely, the shift operations will be optimized in the C compiler anyway.

This expression folding of gains has, however, been enhanced in R2008a. So in R2008a and later releases, the generated code appears as follows:

 mux_to_array_Y.Out1 = (int16_T)((mux_to_array_U.In1 << 3U) * 3);