Add pragma(musttail) #4366
Add pragma(musttail) #4366
Conversation
I like that better - I assume GDC could implement it too. |
MrSmith33
force-pushed
the
musttail
branch
2 times, most recently
from
f5ba4a4 to
4da4c6e
Compare
| else if (ps.ident == Id.musttail) | ||
| { | ||
| pragmaMustTailSemantic(ps); | ||
| } |
There was a problem hiding this comment.
This is a frontend file too (everything in dmd/), so IN_LLVM again. Note that there's a IN_LLVM global/enum somewhere too, so that it's easy to use in if expressions.
And version (IN_LLVM) for pragmaMustTailSemantic() below too.
There was a problem hiding this comment.
What should non-ldc compilers do for this pragma? Will this file be syncronized with dmd at some point?
There was a problem hiding this comment.
Upstreaming our frontend diff is probably not realistic; some of it are hacks, and some others really totally LDC-specific and would just clutter the DMD frontend IMO. If another compiler supported it, we'd upstream a proper semantic for the new pragma though.
There was a problem hiding this comment.
Clutter is meaningless versus the asymptotic cost of complexity - it should be upstreamed. LDC is the real D compiler, the codebase should be honest about that
There was a problem hiding this comment.
Slightly less abstract: we should get a way upstream for there to no more IN_LLVM and so on other than registering a behaviour (and not "just override a weak symbol")
gen/funcgenstate.cpp
Outdated
| llvm::FunctionType *calleeType, | ||
| llvm::ArrayRef<llvm::Value *> args, | ||
| const char *name, bool isNothrow) { | ||
| const char *name, bool isNothrow, bool isMustTail) { |
There was a problem hiding this comment.
We should try to avoid changing the signature of this function.
I guess you could use the DValue returned by DtoCallFunction() like this:
DValue *bla = DtoCallFunction(…);
if (ce->isMustTail) { // ce is the CallExp*
if (auto ci = llvm::dyn_cast<llvm::CallInst>(DtoRVal(bla))) {
ci->setTailCallKind(llvm::CallInst::TCK_MustTail);
} else {
// report error
}
}There was a problem hiding this comment.
Hmm, that probably only works for trivial return value types. Another option: giving DtoCallFunction() an optional std::function param, which gets the LLCallBasePtr call as argument and gets a chance to manipulate it.
|
Wrt. a testcase, see e.g. https://github.com/ldc-developers/ldc/blob/master/tests/codegen/array_alloc_gh3041.d for a simple example checking the generated IR. The stuff in |
|
@ibuclaw Can GDC implement this too (or does it already have it?) ? |
MrSmith33
force-pushed
the
musttail
branch
3 times, most recently
from
628cb3b to
218d800
Compare
| ci->setTailCallKind(llvm::CallInst::TCK_MustTail); | ||
| } else { | ||
| if (!tf->isnothrow()) { | ||
| error(loc, "cannot perform tail-call - callee must be nothrow"); |
There was a problem hiding this comment.
can this error not be caught during semantic analysis?
There was a problem hiding this comment.
Where in semantic can I check this? I'm not really familiar with the code.
| if (!tf->isnothrow()) { | ||
| error(loc, "cannot perform tail-call - callee must be nothrow"); | ||
| } else { | ||
| error(loc, "cannot perform tail-call - no code like destructors or scope(exit) should run after the call"); |
There was a problem hiding this comment.
same as above, can this error not be caught during semantic analysis?
cc @ibuclaw |
Discussed in #4355
Open questions:
LDC_musttailwhat we want? Or we should usemusttail?