[llvm-dev] [RFC] Adding CPS call support

Kavon Farvardin via llvm-dev llvm-dev at lists.llvm.org
Mon Apr 17 08:30:18 PDT 2017


There is a need for dedicated continuation-passing style (CPS) calls in LLVM to
support functional languages. Herein I describe the problem and propose a
solution. Feedback and/or tips are greatly appreciated, as our goal is to
implement these changes so they can be merged into LLVM trunk.


Implementations of functional languages like Haskell and ML (e.g., GHC and
Manticore) use a continuation-passing style (CPS) transformation in order to
manage the call stack explicitly. This is done prior to generating LLVM IR, so
the implicit call stack within LLVM is not used for call and return.

When making a non-tail call while in CPS, we initialize a stack frame for the
return through our own stack pointer, and then pass that stack pointer to the
callee when we jump to it. It is here when we run into a problem in LLVM.

Consider the following CPS call to @bar and how it will return:


define void @foo (i8** %sp, ...) {
    ; ...
    ; finish stack frame by writing return address
  %retAddr = blockaddress(@foo, %retpt)
  store i8* %retAddr, i8** %sp
    ; jump to @bar
  tail call void @bar(i8** %sp, ...)

 retpt: 				; <- how can @bar "call" %retpt?
   %sp2 = ???
   %val = ???
   ; ...


 define void @bar (i8** %sp, ...) {
 	  ; perform a return
 	%retAddr0 = load i8*, i8** %sp
 	%retAddr1 = bitcast i8* %retAddr0 to void (i8**, i64)*
 	%val = bitcast i64 1 to i64
 	  ; jump back to %retpt in @foo, passing %sp and %val
 	tail call void %retAddr1(i8** %sp, i64 %val) 


There is currently no way to jump back to %retpt from another function, as block
addresses have restricted usage in LLVM [1]. Our main difficulty is that we
cannot jump to a block address without knowing its calling convention, i.e., the
particular machine registers (or memory locations) that the block expects
incoming values to be passed in.

The workaround we have been using in GHC for LLVM is to break apart every
function, placing the code for the continuation of each call into a new
function. We do this only so that we can store a function pointer instead of a
block address to our stack. This particularly gross transformation inhibits
optimizations in both GHC and LLVM, and we would like to remove the need for it.


I believe the lowest-impact method of fixing this problem with LLVM is the 

First, we add a special 'cps' call instruction marker to be used on non-tail
calls. Then, we use a specialized calling convention for these non-tail calls,
which fix the returned values to specific locations in the machine code [2].

To help illustrate what's going on, let's rewrite the above example using the
proposed 'cps' call:


define { ... } @foo (i8** %sp, ...) {
    ; ...
    ; finish stack frame by writing return address
  %retAddr = blockaddress(@foo, %retpt)
  store i8* %retAddr, i8** %sp
    ; jump to @bar
  %retVals = cps call ghccc {i8**, i64} @bar (i8** %sp, ...)
  br label %retpt

   %sp2 = extractvalue {i8**, i64} %retVals, 0
   %val = extractvalue {i8**, i64} %retVals, 1
   ; ...


 define {i8**, i64} @bar (i8** %sp, ...) { 
 	  ; perform a return
 	%retAddr0 = load i8*, i8** %sp
 	%retAddr1 = bitcast i8* %retAddr0 to {i8**, i64} (i8**, i64)*
 	%val = bitcast i64 1 to i64
 	  ; jump back to %retpt in @foo, passing %sp and %val
 	tail call ghccc void %retAddr1(i8** %sp, i64 %val)

 	unreachable   ; <- ideally this would be our terminator,
 	              ; but 'unreachable' breaks TCO, so we will
 	              ; emit a ret of the struct "returned" by the call.


The important point here is that the 'cps' marked call will lower to a jump. The
'cps' call marker means that the callee knows how to return using the arguments
explicitly passed to it, i.e., the stack pointer %sp. The callee cannot use a
'ret' instruction if it is 'cps' called.

Either before or during 'cps' call lowering, any instructions following the
'cps' call to @bar are sunk into the the block %retpt, and the unconditional
branch to %retpt is deleted/ignored. We include that branch to preserve 
control-flow information for LLVM IR optimization passes.

The 'extractvalue' instructions are what ensure the calling convention of
%retpt, since the fields of the struct %retVals are returned in physical
registers dictated by the (modified) ghccc convention. Those same physical
registers are used by the ghccc tail call in @bar when it jumps back to %retpt.
So, the call & return convention of ghccc ensures that everything matches up.

Interaction with LLVM

(1) Caller-saved Values

One may wonder how this would work if there are caller-saved values of the 'cps'
call. But, in our first example, which closely matches what CPS code looks like,
the call to @bar was in tail position. Thus, in the second example, there are no
caller-saved values for the 'cps' call to @bar, as all live values were passed
as arguments in the call.

This caller-saved part is a bit subtle. It works fine in my experience [2] when
@bar is a function not visible to LLVM. My impression is that even if @bar is
visible to LLVM, there is still no issue, but if you can think of any corner
cases that would be great!

(2) Inlining

My gut feeling is that we cannot inline a 'cps' marked call-site without more
effort. This is because we might end up with something odd like this once the
dust settles:

    %retAddr = blockaddress(@foo, %retpt)
    %retAddr1 = bitcast i8* %retAddr to {i8**, i64} (i8**, i64)*
    tail call ghccc %retAddr1 ( %sp, ... )

We could add a pass that turns the above sequence into just an unconditional
branch to %retpt, using a phi-node to replace each 'extractvalue' instruction in
that block.

I'm not sure whether inlining in LLVM is important for us yet, as we tend to
inline quite a lot before generating LLVM IR. I don't think this additional fix-
up pass would be too difficult to implement if it's desired.

Implementation Sketch and Conclusion

My current plan is to add this special lowering of 'cps' calls during the
translation from LLVM IR to the SelectionDAG. I welcome any suggestions or tips
on the best way to approach this. An important goal for us is to merge this into
trunk since we do not want to bundle a special version of LLVM with GHC.

Please let me know soon if you have any objections to this feature.

Thanks for reading,


[1] http://llvm.org/docs/LangRef.html#blockaddress
[2] http://kavon.farvard.in/papers/ml16-cwc-llvm.pdf

More information about the llvm-dev mailing list