Sorting through all of the discussions would be difficult, as the ideas have morphed over the years. Also, some of the discussion took place offline at various LLVM dev conferences.
I can summarize the main points here:
The biggest improvement in GC would be to allow SSA values to be declared as GC roots - currently only alloca values, that is, values in memory, can be GC roots. This means that frontends need to copy traceable values to memory before calling any function that might trigger a collection. This makes the resulting IR complex and inefficient.
Chris’s proposal is that we add metadata to the Pointer type (perhaps using the address space field) to indicate that a pointer is a GC root. Note that this would not be a replacement for the existing llvm.gcroot intrinsic, but rather an addition to it - that is, you could choose to use either the Pointer attribute or the llvm.gcroot intrinsic. The reason for this is that for some languages, there are non-pointer GC roots (my language has discriminated unions that sometimes contains pointers and sometimes not.) These frontends would need to continue to use the existing intrinsics, but fortunately such cases are fairly rare and this is not a terrible burden. For most languages, the Pointer attribute would be a much easier way to deal with roots, since the “root-ness” of a value is simply a property of it’s type, and so gets carried along with the value automatically.
Thus, in this new scheme, Pointer types which were marked as roots could be traced regardless of whether they lived within a memory location or a register. These could be used to generate either stack maps or register maps. Note that these maps are language-specific and are implemented via a language-specific plugin, so you would need to update the current API to deal with items in registers. (An alternative plan is to allow the frontend to request via a flag that roots be copied to local variables at each safe point, so that register maps would not be needed. This alternate plan might be a good first step, and then move on to the more difficult problem of register maps if the first step is successful.)
The reason why this is difficult is that the presence of garbage collection roots has a major impact on the optimizer and backend code generators - certain optimizations can cause the stack map to be incorrect, so these optimizations must be prevented or compensated for. However, an advantage is that variables which are optimized away no longer need to be included in stack maps - something that is not possible with the current approach.
One other limitation of the Pointer approach over the existing llvm.gcroot system, is that the latter allows complex metadata to be associated with each root. This is useful in languages that don’t use tagged objects, that is, the type of every object it known at compile time. However, creating a metadata pointer for every Pointer type would be expensive, so the Pointer roots would only be used for languages which use tagged objects - which is fortunately most languages that use GC.
An even more ambitious plan would be to allow structs in SSA values to be declared as roots, which would be useful for languages like mine. We wouldn’t use register maps for these, since a struct might get “sliced” into multiple registers. Instead, the code generator would automatically spill the struct value to memory during a safe point, and otherwise treat the struct like the existing llvm.gcroot() intrinsic does. Note that this is a much harder problem, and would not be needed for languages like Java or Python where objects are always passed by reference. I wouldn’t expect an initial implementation to attempt to tackle this harder problem.
That would be the biggest improvement that I can think of. There are a few other minor improvements that I would also like to see:
One would be a set of intrinsics that would allow efficient iteration of stack frames in an efficient manner. The existing LLVM stack frame intrinsics are inefficient and cannot be relied upon in many cases. Basically you’d want 3 intrinsics, which would work on all supported platforms: The first would return an opaque reference to the current stack frame; The second would take a stack frame as it’s argument and return a pointer to it’s parent stack frame; And the third would take a stack frame argument and return a base address for the local variables of that frame. The lanuguage specific runtime would then use this base address, along with the generated stack maps, to access all of the stack roots for that frame.
An example of how this stack walking is done can be seen here: http://code.google.com/p/tart/source/browse/trunk/runtime/lib/gc_common.cpp#155 However, this code only works on x86 - the intrinsics that I envision would work on a much wider set of backend targets.
Note that these items are just a tiny part of a complete collector, however, the design of LLVM is that each language is supposed to implement its own collector, and LLVM only supplies the parts that need to be integrated into the code generator.
I can also suggest ways to test the new features without having to build a complete garbage collector. For example, one can create a trivial stack walker that merely counts the number of non-null root pointers, and write various unit tests that verify that the results are as expected.