[LLVMdev] LLVM tries to remove labels used in blockaddress()

[Chris Lattner]

On Mar 15, 2010, at 10:07 AM, Bob Wilson wrote:
>>> An earlier revision simply generated asm-code, where the appropriate label was missing, thus causing gcc to fail when i wanted to compile the asm-file.
>> 
>> Here is a slightly reduced testcase:
>> 
>> define i8* @test1() nounwind {
>> entry:
>> 	ret i8* blockaddress(@test_fun, %test_label)
>> }
>> 
>> define i32 @test_fun() nounwind {
>> entry:
>> 	ret i32 -1
>> test_label:
>> 	br label %ret
>> ret:
>> 	ret i32 -1
>> }
>> 
>> The basic problem is that we codegen test1, which generates a reference to test_label, then we codegen test_func.  The optimization passes that run before the code generator (UnreachableBlockElim & CodeGenPrepare) zap dead blocks, so test_label was being deleted.
> 
> I'm not sure I understand completely.  We should be able to rely on the indirectbr label arguments to identify blocks that are live.  You can have a blockaddress anywhere, but indirectbr is only defined within a function so there's no ordering problem if you're looking at the indirectbr.

The problem is that test1 is codegen'd and the reference to test_label is emitted to the .s file.  By the time test_func is codegen'd, the block is deleted, so the definition isn't emitted.  I'll fix this.

>> The code generator doesn't want dead blocks coming into it for various reasons, so removing them is important.  I guess we'll have to do something like buffer up the unemitted labels and emit them at the end of the file (in a meaningless location).  This should provide correct code, but is somewhat gross.
>> 
>> Bob/Dan, do you guys have any other ideas on how to handle this?
> 
> Is the problem when you've got a blockaddress that references a dead label in a different function?  If that's the case, the blockaddress value can be an undef.

But the reference has already been emitted to the .s file.  I'll take care of this by just queuing up labels and emitting them at the end of the function they correspond to.

-Chris