Re: BUG with libio/dbz/rdbz test case on m88k-motorola-sysv3

[Manfred Hollstein <manfred at s-direktnet dot de>]

On Fri, 14 August 1998, 22:03:21, law@hurl.cygnus.com wrote:

 > 
 >   In message <13779.63584.537611.578811@slsvhmt>you write:
 >   > It's   function      "bytemap"  in   libio/dbz/dbz.c    which    get's
 >   > mis-compiled. Fortunately, the diffs in  the generated assembler  code
 >   > are quite small (note,  the no-gcse stuff works  OK); perhaps it'll be
 >   > an indication for you.
 >   > 
 > Is this happening on the mainline tree or on the egcs-1.1 branch?

This is on the egcs-1.1 branch.

 > 
 > The reason I ask is no matter what I do on the branch I end up with
 > different code.  Only slightly, but different and that makes the
 > analysis phase much more difficult/impossible.  Actually, I just tried
 > with the mainline sources and they don't seem to do anything different.
 > 
 > Your testcase was also incomplete -- it references undefined types,
 > a macro or two, etc.  So I may have screwed up something while trying
 > to make a testcase that I could actually compile.

Ah yes, I've actually copied dbz.c into two files dbz1.c and dbz2.c,
tampering with the "static" vs. "extern" stuff and ifdef'ed the
functions one by one.

 > 
 > At this point, there's not much I can do.  From the debugging dumps
 > gcse is doing the right thing, so if -fno-gcse is indeed causing the
 > problem to go away, then the real bug is later in the compiler and
 > just happens to be exposed by gcse.
 > 
 > Or maybe your stage3 compiler is being mis-compiled which in turn
 > causes the testcase to be mis-compiled.
 > 
 > In fact, you might try building the testcase with the stage1 and
 > stage3 compilers to see if they differ.

I'll try that on Monday; unfortunately (no, fortunately!) I'll be on
vacation beginning on Tuesday for the next 4 weeks without access to the
88k machine - I'll see if I can do something with a cross compiler.

 > 
 > jeff

manfred