Register allocator should be able to preserve all local values throughout their scopes in unoptimized code to aid debugger

[yurivict]

Bugzilla Link	18697
Version	trunk
OS	All
CC	@dwblaikie,@echristo,@hfinkel,@pogo59

[yurivict]

Currently gdb often refuses to print local values when they aren't used any more:
(gdb) p this
$11 =

This is because register allocator doesn't keep them after they are used last.

llvm should either alter register allocator behavior by default in unoptimized code, always preserving values until they go out of scope, or have a special option to do this.

This feature will be greatly appreciated by people who do debugging on llvm generated code.

[echristo]

I may take an option to do this, but as a default mechanism this isn't going to happen. I'll leave this open, but anything in this area will need to be discussed.

[dwblaikie]

I may take an option to do this, but as a default mechanism this isn't going
to happen.

Just to be clear, this doesn't sound like a "-g should change codegen" request - but a "pessimize -O0 a bit more than we already do".

I'm assuming you're saying it won't be a default because we couldn't accept such pessimism.

Might be right, though our current alloca approach is pretty heavy handed at -O0 already... the only things we don't alloca are a few special cases that fall through the cracks, not because we actually chose them for perf reasons.

Yuri - could you provide a specific reproduction. I cannot reproduce this in simple examples.

[pogo59]

This sounds like the extend-lifetimes work @SLTozer ?

[SLTozer]

Similar, although with the distinction that this seems to apply to unoptimized code, which at the moment we explicitly exclude in extend-lifetimes; theoretically though if we turned it on at O0 we could see improved liveness... depends on what the reproducer is though, since most of the time you should have 100% value availability at O0!