While HLS does take C/C++ as an input, it typically requires some amount of hardware specification to successfully generate HDL. MATLAB Coder and Simulink Coder produce code intended for C/C++ compilers. So the code contains constructs that C/C++ compilers know how to implement in software but is too ambiguous for an HLS tool to automatically implement in fixed-resource hardware. Examples include unbounded loops, dynamically-sized arrays, dynamic memory allocation, pointers, and recursion.
Even in simpler cases where the generated C/C++ can run through HLS, it can be challenging for an HLS tool to generate efficient hardware without some amount of hardware specification. For instance whether a loop is completely unrolled into parallel hardware threads vs not unrolled to conserve hardware resources, or some balance of the two. Other examples include FIR filter architectures, complex multiplication implementations in FFTs, insertion of registers to break up long timing paths, and optimizing fixed-point data types for efficient resource usage.
Most HLS tools support SystemC® (IEEE Std. 1666™-2011) for adding these hardware constructs to C code in so you can simulate this functionality before running it through HLS. Some HLS tools let you describe hardware implementations as non-executable compiler directives, or “pragmas”, which results in this new functionality being added during HLS so simulation results often differ between pre- and post-HLS. But manually editing C/C++ generated from Simulink or MATLAB introduces manual coding error risks and guarantees that the functionality will differ from what you simulated in MATLAB or Simulink.
(Update March 2022) Starting with R2022a, you can generate synthesizable SystemC code from streaming MATLAB algorithms https://www.mathworks.com/help/hdlcoder/systemc-code-generation-from-matlab.html