Re: Can we have egcs 1.0.3 please?

Re: Can we have egcs 1.0.3 please?

[H.J. Lu]

> 
> 
>   In message <m0yQNZt-00058vC@ocean.lucon.org>you write:
>   > There is a bad typo in libio. I have sent a patch twice.
> I asked Ulrich to take a look at the libio patch not too long
> ago.  I haven't heard back from him yet.
> 
> I'd really like everything going into the 1.0.3 tree to have
> been reviewed by at least one person other than the author.
> 
> Maybe if you explained what is wrong and how your patch fixes
> the problem, someone could try to evaluate the patch.  Instead
> you just said "it's a typo" with zero explanation about what's
> going on.

See

http://www.cygnus.com/ml/egcs-bugs/1998-Apr/0028.html
http://www.cygnus.com/ml/egcs-bugs/1998-Apr/0046.html

I am enclosing my original patch with the typo in it here. As you can
see, I changed "count" to "_gcount" in "if" without removing count.

-- 
H.J. Lu (hjl@gnu.org)
---
--- /home/work/misc/gnu/import/egcs/libio/isgetline.cc	Thu Aug 21 15:58:19 1997
+++ isgetline.cc	Mon Feb 16 10:48:27 1998
@@ -67,8 +68,9 @@
   if (ipfx1())
     {
       streambuf *sbuf = rdbuf();
-      long count = _IO_getline(sbuf, buf, len - 1, delim, -1);
-      if (count == 0 && sbuf->sgetc() == EOF)
+      int ch;
+      long count = _IO_getline_info(sbuf, buf, len - 1, delim, -1, &ch);
+      if (_gcount == 0 && ch == EOF)
 	set(ios::failbit|ios::eofbit);
       else
 	_gcount = count;

Re: Can we have egcs 1.0.3 please?

[Jim Wilson]

	Jim is correct, except for one minor nit.
	REG_N_REFS must never be zero if references to the register still
	remain in the insn chain. 

I was just making a general statement, to avoid a panic about a bug that
wasn't there.  It wasn't intended to be complete.

But, as long as we are nit picking, I'll add another clarification.

REG_N_REFS must never be two unless there are exactly two references to the
register.  Local-alloc treats this case specially, and will perform an invalid
optimization if REG_N_REFS is two but there are more than two references in
the RTL.  This is why combine deliberately sets REG_N_REFS to three in one
place.

Jim

Re: Can we have egcs 1.0.3 please?

[Jeffrey A Law]

[ Note The change in cc-line, this is suitable for the egcs
  discussion list :-) ]

  In message <199804182235.PAA27290@rtl.cygnus.com>you write:
  > There is no requirement that REG_N_REFS be exactly correct, and many
  > optimizations passes after flow don't bother to try to maintain it.
  > Thus the fact that REG_N_REFS is wrong is itself not necessarily a bug.
Jim is correct, except for one minor nit.

REG_N_REFS must never be zero if references to the register still
remain in the insn chain.  Other than that, REG_N_REFS can have
any value without effecting the correctness of the generated code.

Of course accurate REG_N_REFS is likely to lead to better register
allocation.

jeff