Handle exceptions from throw_...

[maleadt]

Apparently LLVM managed to optimize many throws away, which breaks now that we don't inline everything anymore. Let's try to work around some.

[maleadt]

So this has been pretty painful. ptxas is really buggy, as it turns out.

[maleadt]

This passes CuArrays tests again. Summarized:we now have a couple of LLVM passes that rewrite IR to make it GPU compatible:

• replace throw_XXX functions (relies on the name, to be replaced with overdubbing of Core.throw)

julia> kernel(x) = x[2] = 1
kernel (generic function with 1 method)

julia> @cuda kernel(CuArray{Int}(1)); CUDAdrv.synchronize()
ERROR: a boundserror exception occurred

• avoid trap/unreachable that cause miscompilations (Shared memory + multiple function exits cause invalid results #4, we should really use LLVM for this)

[vchuravy]

This reminds me of twice_return in Julia codegen which needs special handling as well (I recall there being at least two instances).

c.f. JuliaLang/julia@6a1cc96

[maleadt]

Similar indeed. A proper fix seems tricky though, because then you'd need LLVM to reason about the thread mask and only return when all threads are active (assuming this issue is related due to exiting from a function during divergence... but shmem seems related too). It looks like CuArrays mapreduce triggers this quite easily though, so let's hope that helps us spot any changes in upstream irgen (ie. patterns in Julia-generated IR that generate a CFG ptxas doesn't like) before running into this again.