Re: PATCH for tree-to-rtl expansion of C functions

[Arpad Beszedes <beszedes at cc dot u-szeged dot hu>]

> This patch looks incorrect.  First if there are new testsuite failures
> you need to modify the testcases so they no longer fail.  E.g. it
> should be possible by checking "&& ! defined (__OPTIMIZE_SIZE__)" in
> stdio-opt-1.c.

Ok, I will create a patch for the testcase as well.

> Second, you seem to not only disable the fprintf->fwrite
> transformation, but also fprintf->fputs transformation.  The latter
> shouldn't cause any extra arguments or registers used, in fact it may
> decrease the number for fprintf(s, "%s", "foo") -> fputs("foo", s)
> because the "%s" no longer needs to be passed to the stdio function.

Ok, I will not modify c-common.c. However, I made some measurements on 
our testbed and surprisingly the overall result was that preventing 
fprintf->fputs transformation produced significant size reduction (with 
transformation 0.3% vs. 0.5% without). One cause for this was the 
opposite order of passing arguments to fprintf and fputs, which cause 
other instructions and registers to be reorganized.

Regards,
  Arpad.