For Sumatra targets, we wanted to experiment with an infrastructure for supporting deoptimization, that is, transferring execution from the compiled code running on the GPU back to the equivalent bytecodes being run through the interpreter on the CPU.   The following describes some experiments with deoptimization using the HSAIL Backend which are now available on the graal trunk.

Some reasons for deoptimization on the GPU are:

• When compile-time assumptions are violated, we can "trap" to the interpreter (this relies on the fact that the interpreter can handle anything).  In addition, we have a way of handling certain hopefully rare events, such as throwing exceptions back to the CPU, which might be difficult to implement completely from the GPU.  If statistics shows that such events are not actually "rare", we can
  
  • decide that in the future this particular lambda is not a good candidate for offload.
  • or in some cases we might be able to recompile and generate new code for the GPU.
• compiled code running on the GPU might get to a point where it needs the CPU to do something before the GPU can make further progress.  For example, if we are supporting allocation on the GPU, we could get to a point where we cannot allocate any new object until a GC happens.  If the target does not have an easy way to spin and wait for the CPU to do the GC, one way to support this is to deoptimize.  The interpreter will let the GC happen and then continue executing bytecodes from the point of the deoptimization, including finishing the allocation.

Implementation Notes

Compile Time

We use the graal compiler to generate the hsail code.  The graal compiler has a mature infrastructure for supporting deoptimization and still achieving good code quality.  See http://design.cs.iastate.edu/vmil/2013/papers/p04-Duboscq.pdf.  Basically the compiler keeps track of the deoptimization state at each deopt point, and from that we can tell what HSAIL registers need to be saved, which registers contain oops, etc.

Execution Time

When we dispatch a kernel to the GPU, we need to finish the execution of all the workitems even if one or more of the workitems deoptimize.  The kernel code can be executed across a possibly very large number of workitems, each of which can have its own state.  The non-deoptimizing workitems can finish as they normally would but we need to be able to save the state of the deoptimizing workitems.   We don't know how many workitems are going to need to deoptimize and need to save their state, yet we want to avoid having to allocate state-saving space for the entire possibly very large range of workitems.

A particular HSA target has a maximum number of workitems that can be executing concurrently.  So to save space, we need only allocate state-saving space for this maximum number of possible concurrent workitems.  We do this by having deopting workitems set a "deopt happened" flag, and having each workitem check this "deopt happened" flag.  Future workitems that see the "deopt happened" flag as true just set a simple flag indicating they never ran and exit immediately, thus not needing to save any additional state.  Currently the never_ran flag array is one byte per workitem.  We are looking at ways to make this smaller but HSA devices have a lot of freedom in how they schedule workitems which makes it difficult to figure out the never-rans.

Workitems that deopt atomically bump an index saying where they should store their deopt data.  The deopt data consists of

• the workitem id, which is a linear index across the range and unique for each workitem
• the deoptimization actionAndReason, just as in cpu deoptimizations
• the first HSAILFrame

An HSAILFrame consists of:

• the deopt Id or "pc" offset where the deopt occurred
• number of 32-bit s registers saved
• number of 64-bit d registers saved
• number of stack-slot variables saved (work in progress)
• space for saving the s and d registers and stack slot variables
• in the current hsail frame we save a bitmap of the d registers that contain oops.  This will be replaced later with a data structure for each kernel that will map a deopt ID to an oopMap for that deoptID.

To avoid code bloat, we currently have one deopt exit point per kernel, and in that deopt exit code we save the union of the actual registers that are live at any of the infopoints.

When the GPU dispatch completes, each workitem will have either finished normally, deopted, or not run at all.  Still in the VM, we check if there were any deopts.  If not, we know everything completed normally and we can just return back to java.  However, if there was at least one deopt then

• for the workitems that finished normally, there is nothing to do
  
  
• if there are any deopted workitems, we want to run each deopting workitem thru the interpreter starting from the byte code index of the deoptimization point.  Note that it is possible for different workitems to have different deoptimization "PCs".    We currently re-dispatch each of the deopting workitems sequentially although other policies are clearly possible.
  • Note: Getting to the interpreter from the saved hsail state goes first through some special compiled host trampoline code infrastructure that Gilles Duboscq designed.  The trampoline host code takes the deoptId and a pointer to the saved hsail frame as input and then immediately deoptimizes just as any host compiled code would deoptimize.
    
    
• for each never-ran workitem, we can just run it from the beginning of the kernel method, just making sure we pass the arguments and the appropriate workitem id for each one.  Again, we currently do this sequentially although other policies are possible.

GC Considerations

Currently, the normal kernel dispatch runs in "thread in VM" mode and thus does not need to worry about moving Oops.  However, each time a deopting workitem is run through the interpreter or each time a never-ran workitem is run we are back in Java mode which can cause GCs.  So for each saved hsail frame, we need to know which of the saved state contain oops and make sure those locations are updated in the face of GC.  The current strategy of copying oops into an Object array supplied by the java side will be replaced later with an oops_do type of strategy.