Re: Regressions problem (200 failures)

[Hans-Bernhard Broeker <broeker@physik.rwth-aachen.de>]

This is an answer to < 20000301123337B.mitchell@codesourcery.com >

[I've been pointed to this discussion by Eli Zaretskii, but I'm not on the
gdb mailing list, myself. I just read through it on the WWW archive. So
please, if you answer, Cc: to me, if possible. Thank you.]

The point raised in this discussion has, by coincidence, caused a similar
problem, with the DJGPP release version of GCC 2.95.2 and GDB 4.18. The
problem is that for very short functions (one, maybe two lines of actual
code, between the braces), gdb would not stop *at all* if you 'step' into
a function from outside, because of badly positioned line number debug
symbols.

Looking at the assembly generated by GCC, it turned out that the problem
lies in the way the function prologues and epilogues were written,
compared to earlier GCC releases. So, to answer one of the questions
raised in your discussion: to some extent, the prologue/epilogue have
indeed changed, since 1994. The whole method of outputting prologues has
been changed, since gcc-2.8.1, it seems, even though the typical set of
machine operations has stayed the same, for this platform. Originally,
prologues and epilogues were generated directly as assembly, by a
specialized function, i.e. they were not subject to RTL transformations.
Now, by default at least, they're generated as RTL, rather early in the
compilation, and subject to modification along with the 'real code'.

As to the question where the first line number label ought to be put, and
what line it should point, I think the behaviour of previous GCC/GDB
combinations was perfectly sane: the line number opcode is output right
after the prologue, and it points to the line the next machine instruction
originated from (initialization of an automatic variable, if present, an
executable statement otherwise). 

Opposed to this expected behaviour, gcc-2.95.2 outputs a line note
*before* the prologue (and one for the closing brace after the epilogue,
instead of before it, as it used to be). By disabling the RTL-style
prologue generating mechanism (undocumented GCC option
-mno-schedule-prologue), you get back the traditional behaviour.

Currently, the conclusion of discussion between me and Eli is that this
constitutes a bug in gcc-2.95.2. Wether or not that's still present in the
current snapshot remains to be checked. AFAICS, the GCC patch from Mark
Mitchell that caused all this hassle for your GDB testsuite meant to fix
that, but didn't work out as planned.

Hans-Bernhard Broeker (broeker@physik.rwth-aachen.de)
Even if all the snow were burnt, ashes would remain.