[PATCH] D15911: Switch Action and ActionList over to using std::shared_ptr.

Justin Lebar via cfe-commits cfe-commits at lists.llvm.org
Tue Jan 5 22:37:47 PST 2016


jlebar added a comment.

The main reason I want this is for CUDA.  The way CUDA compilation will work, once I finish my patch, is: For each GPU arch, we compile device code to assembly (ptx) and then assemble the ptx into an object file (cubin).  We then pass the cubins *and* ptx files to nVidia's fatbinary program, which "links" them into one blob.  We then embed the blob in the host code.

So in this scheme, our Action graph is a DAG: The ptx is used as an input both to the cubin and fatbin actions.  Making this DAG work in the absence of this patch is...pretty unpleasant.


http://reviews.llvm.org/D15911





More information about the cfe-commits mailing list