[Bug sanitizer/112708] "gcc -fsanitize=address" produces wrong debug info for variables in function prologue

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112708

--- Comment #9 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Bruno Haible from comment #8)
> (In reply to Richard Biener from comment #7)
> > It's -fvar-tracking, not -fvar-tracking-assignments.  At -O0 debug info
> > during the prologue is unreliable without that.
> 
> Then how about enabling -fvar-tracking automatically when "-g" or "-ggdb" is
> requested and the debug format supports it?

There are major problems with that.  var-tracking is very compile time
intensive,
so it would significantly slow down -O0 compilation.
And, most of the time at -O0 variables live in their memory slots, so that
var-tracking time would be also wasted.
The only things which need var-tracking at -O0 are:
1) register variables (or whatever else doesn't get an allocated stack slot)
2) variables at the start and end of their lifetime, where they are not yet in
   their stack slot or no longer in them (so, often function prologues and
epilogues,
   or say for VLAs before everything is set up etc.)
So, as written in other PRs, for -O0 we'd likely want to invent some cheaper
var-tracking variant which would only track the above 2 and not waste time on
the rest.

> The documentation
> https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Debugging-Options.html says:
> "It is enabled by default when compiling with optimization (-Os, -O, -O2,
> …), debugging information (-g) and the debug info format supports it."
> 
> Why not also enable it when compiling _without_ optimization?
> 
> I'm not using "-Og" because the documentation
> https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Optimize-Options.html says
> "-Og enables all -O1 optimization flags except for those that may interfere
> with debugging", but what I want is optimal debugging and don't want to
> spend CPU cycles on optimization.

As for -Og, -Og doesn't perform too many optimizations and in function
prologues code is usually better debuggable than with -O0.  But there is one
major obstackle also recorded in various PRs, we don't artificially make all
vars used at the end of their
scope, which means that if some var is unused after certain point in a
function, it might no longer be available and e.g. DWARF5 DW_OP_entry_value
used as the last resort; but that doesn't work always.

For the -fsanitize=address case, the thing is that the marking of end of
prologue is
something we do during function prologue generation in pro_and_epilogue (or
debugger does that from scanning the function IL).  While the asan "prologue"
is something emitted during RTL expansion and so for pro_and_epilogue seen as
part of function body.