I looked at this a little more.
I propose we remove the call to compile under gcl (and only gcl).
Further we change coerce-float-fun so that (coerce-float-fun f)
returns (symbol-function f) if f is a compiled maxima function.
There is a special check in coerce-float-fun for gcl. If f is a
compiled function it does not return (symbol-function f) under gcl but
it does return (symbol-function f) for other lisps. I think this is
because (compile nil (symbol-function f)) does not work under gcl but
it does under other lisps. Since this only happens in quadpack
functions I think this check was added to make quadpack work under
gcl. If we don't compile the integrand under gcl, this special check
in not needed anymore.
With this something the user can do
f(x) := block([], mode_declare(x, float), 1/(x+2+2^-5)/sqrt(1-x^2))$
compile(f)$
quad_qags(f, x, -1, 1)
and it should be just as fast as if we called compile on the
integrand. Also plotting should be faster if we plot compiled
functions.
Andrej