Re: Benchmarking temporary Lisp objects [Was: Re: [RFC] temporary Lisp_Strings]

[Dmitry Antipov]

On 09/05/2014 11:15 AM, Paul Eggert wrote:

> I like the idea of simpler and cleaner, but your patch's implementations of
> build_local_vector and build_local_string use alloca expressions as function
> arguments

I don't do that - I'm using 1st arg (result) as a temporary variable,
assuming that Lisp_Object is always wide enough to hold an address:

#define build_local_vector(obj, size, init)                             \
  (MAX_ALLOCA < (size) * word_size + header_size                        \
    ? obj = Fmake_vector (make_number (size), (init))                   \
    : (obj = XIL ((uintptr_t) alloca                                    \ here
                 ((size) * word_size + header_size)),                   \
       obj = local_vector_init ((uintptr_t) XLI (obj), (size), (init))))

IIUC the compiler should know about alloca limitations and so should not
try to (mis)optimize the code above to:

#define build_local_vector(obj, size, init)                             \
  (MAX_ALLOCA < (size) * word_size + header_size                        \
    ? obj = Fmake_vector (make_number (size), (init))                   \
      obj = local_vector_init                                           \
       ((uintptr_t) alloca ((size) * word_size + header_size),          \
        (size), (init)))

Dmitry