I'm trying to fix that bug: https://llvm.org/bugs/show_bug.cgi?id=20049
It turns out this is the kind of optimization that I really need, as when
it isn't done, all kind of other optimizations opportunities down the road
are not realized as they are not exposed.
I have no idea where to start digging for this. I assume there is some
kind of interaction between memory dependency and alias analysis that is
overly conservative. Can someone gives me some infos on how the 2 interact
together ? What is the code that is supposed to remove these kind of loads ?
First, it looks like you're using an older version of LLVM (the load
syntax has changed). That's definitely not recommended since it will
greatly limit your choices when encountering an issue like this.
I just reused an old code sample. The problem still exists in recent
version of LLVM.
Second, the aliasing problem has to do with the effects of the
"allocmemory" callsite on a memory location associated with an earlier
alloca. There's nothing that prevents your allocmemory function from
writing to global state. You have to teach the alias analysis that an
unescaped noalias pointer can't alias the global state allocmemory might
access. Slightly surprised we don't get this today, but oh well. Take a
look at the isNonEscapingLocalObject predicate in BasicAA. Then look at
"getModRefInfo(ImmutableCallSite CS, const MemoryLocation &Loc)". Double
check to make sure this is the one that MDA actually calls.
I don't think this is the problem. When there is only 2 calls to
allocmemory, loads are optimized away as expected. But it seems that the
analysis is confused with 3+ calls.
Third, you might want to take a look the new inaccessiblememonly
attribute. Depending on how you're modelling your allocation, this might
help resolve the aliasing problem in different way. However, be aware that
this is *very* new. In particular, I triggered an optimizer bug by adding
the new attribute to your particular example. 
Fourth, have you considered implementing a simple escape analysis pass? I
notice that the store-load forwarding would just fall out once you removed
the last allocation. I believe the fixed point would then become a
function which returns the constant answer and does nothing else.
Yes I did. More specifically, there is a pass that try to recognize memory
allocation like calls and optimize them, which right now have libc and some
overloads of operator new hardcoded in it but not much more. This could be
greatly improved IMO, to be language agnostic (and better support C++ in
the process), but really, I do think relying on this to get the load
eliminated, when noalias already provide this information to the optimizer,
is overkill.
Fifth, some pointers on debugging this yourself. GVN (which is the one
which does complicated *-load forwarding) calls MDA
(MemoryDepedenceAnalysis). Using the appropriate -debug-only options will
likely get you to the right area. You can also consider using the
MemDepPrinter pass to bypass GVN. I don't know of a way to issue raw AA
queries for testing. That would be useful, but I don't know that we have
it.
Hope that helps.
Yes, thanks. I'll probably come back with more question once I've played
with these options a bit. Sorry for the late answer, I was mostly off the
grid the past 2 weeks.
You might also consider using a variant of your allocation function which takes a pointer to the global state it needs to modify. Doing this would allow you to use argmemonly to restrict the aliasing while still allowing whole program optimization. I haven’t tried this in practice, but it seems like it would probably work…