[LLVMdev] Cast failure in SelectionDAGBuilder

[Talin]

I'm trying to track down the problem with the assertion failure in
SelectionDAGBuilder.cpp. This is the code:

  *case* *Intrinsic*::gcroot:
    *if* (GFI) {
      *const* Value *Alloca = I.getArgOperand(0);
      *const* Constant *TypeMap = cast<Constant>(I.getArgOperand(1));
*      FrameIndexSDNode *FI =
cast<FrameIndexSDNode>(getValue(Alloca).getNode());*
GFI->addStackRoot(FI->getIndex(), TypeMap);
    }
    *return* 0

The cast<FrameIndexSDNode> is what's failing. Apparently the node isn't a
FrameIndexSDNode.

Now, we discussed a similar problem on this list earlier, and it was stated
that the cause was due to calls to llvm.gcroot() not being in the first
block. However, that's not the case this time - here's the function being
processed (according to I.Parent.Parent->dump()):

define internal %tart.core.Object*
@"tart.core.Object.coerce[int32](int32)->tart.core.Object"(i32 %value) gc
"tart-gc" {
prologue:
  call void @llvm.dbg.declare(metadata !{i32 %value}, metadata !48487)
  %gc_root = alloca %"tart.core.ValueRef[char]"*, !dbg !48488
  store %"tart.core.ValueRef[char]"* null, %"tart.core.ValueRef[char]"**
%gc_root
  %0 = bitcast %"tart.core.ValueRef[char]"** %gc_root to i8**, !dbg !48488
  call void @llvm.gcroot(i8** %0, i8* null), !dbg !48488
  br label %entry, !dbg !48488

entry:                                            ; preds = %prologue
  %ValueRef_new = call %"tart.core.ValueRef[char]"*
@"tart.core.ValueRef[int32].type.alloc"(), !dbg !48488
  store %"tart.core.ValueRef[char]"* %ValueRef_new,
%"tart.core.ValueRef[char]"** %gc_root, !dbg !48488
  %construct = call %tart.core.Hashable
@"tart.core.ValueRef[int32].construct(int32)"(%"tart.core.ValueRef[char]"*
%ValueRef_new, i32 %value), !dbg !48488
  store %"tart.core.ValueRef[char]"* null, %"tart.core.ValueRef[char]"**
%gc_root, !dbg !48488
  %upcast = getelementptr inbounds %"tart.core.ValueRef[char]"*
%ValueRef_new, i32 0, i32 0, i32 0, !dbg !48488
  ret %tart.core.Object* %upcast, !dbg !48488
}


As you can see, the only call to llvm.gcroot() is in fact in the first
block.

Now, here's the weird part: If I turn on optimization, the assertion failure
no longer happens. What's the difference? Here's what the function looks
like (I.Parent.Parent->dump()) with -O2:

define internal %7*
@"tart.reflect.Type[].iterate->tart.core.Iterator[tart.reflect.Type]"(%"tart.reflect.Type[]"*
%self) nounwind gc "tart-gc" {
prologue:
  %gc_root = alloca i8*, align 8
  store i8* null, i8** %gc_root
  call void @llvm.gcroot(i8** %gc_root, i8* null), !dbg !48487
  %new.i = call i8* @malloc(i64 32) nounwind, !dbg !48488
  %0 = bitcast i8* %new.i to %tart.core.TypeInfoBlock**
  store %tart.core.TypeInfoBlock* bitcast (%25*
@"tart.reflect.Type[].ArrayIterator.type.tib" to %tart.core.TypeInfoBlock*),
%tart.core.TypeInfoBlock** %0, align 8, !dbg !48488
  tail call void @llvm.dbg.declare(metadata !{null}, metadata !48489)
  %1 = getelementptr inbounds i8* %new.i, i64 16
  %2 = bitcast i8* %1 to %"tart.reflect.Type[]"**
  store %"tart.reflect.Type[]"* %self, %"tart.reflect.Type[]"** %2, align 8,
!dbg !48490
  %3 = getelementptr inbounds i8* %new.i, i64 24
  %4 = bitcast i8* %3 to i64*
  store i64 0, i64* %4, align 8, !dbg !48491
  %intf_ptr = bitcast i8* %new.i to %7*
  ret %7* %intf_ptr, !dbg !48487
}

OK, I'm looking at the crash in the debugger, and it looks like the NodeType
(getValue(Alloca).getNode().getOpcode()) is value 41. According to what the
debugger is telling me, that value corresponse to llvm::ISD::CopyFromReg.

When I debug the optimized version, the node value is 13
(llvm::ISD::FrameIndex) which seems right.

Does any of that make any sense?

-- 
-- Talin
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20101022/fd8fd645/attachment.html>