gcroot + `section not found for addresss ...' ???

Hi,

If I type

define i32 @main() gc "shadow-stack" {
entry:
  %0 = alloca i8*
  %1 = call i8* @malloc(i64 1)
  store i8* %1, i8** %0
  call void @llvm.gcroot(i8** %0, i8* null)
  ret i32 0
}

declare i8* @malloc(i64)

declare void @llvm.gcroot(i8**, i8*) nounwind

in test.ll

and then do

llc test.ll
gcc test.s

I get the error

ld: in /var/folders/Ea/EakEx6b+GBSo6cUn2AnFZk+++TI/-Tmp-//cc5d3RCx.o, section not found for address 0x64200E00000004CD

If I remove line containing malloc and the line after that, no such error happens. If I remove
the gcroot line then no such error happens.

Any hints as to what this error means would be great!

Cheers,
N

The linker is going off in the weeds trying to parse the dwarf unwind info. The CIE has:

Leh_frame_common_begin0:
  .long 0 ## CIE Identifier Tag
  .byte 1 ## DW_CIE_VERSION
  .asciz "zLR" ## CIE Augmentation
  .byte 1 ## CIE Code Alignment Factor
  .byte 120 ## CIE Data Alignment Factor
  .byte 16 ## CIE Return Address Column
  .byte 2 ## Augmentation Size
  .byte 16 ## LSDA Encoding = pcrel
  .byte 16 ## FDE Encoding = pcrel

But the FDE has:

  .byte 0 ## Augmentation size

The linker is expecting that since the CIE said there was an LSDA in each FDE that is pcrel encoded, that there would in fact be an LSDA in each FDE. But in this case, the compiler is trying to omit the LSDA by marking that the FDE has a zero sized augmentation data.

I'll have to research if this is a valid dwarf unwind optimization. Even if it is, it is tripping up the existing linker.

-Nick

This is an interesting problem. The GC code is being converted into 'invokes' instead of calls:

define i32 @main() gc "shadow-stack" {
entry:
  %gc_frame = alloca %gc_stackentry.main
  %gc_currhead = load %gc_stackentry** @llvm_gc_root_chain
  %gc_frame.map = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 0, i32 1
  store %gc_map* getelementptr inbounds (%gc_map.0* @__gc_main, i32 0, i32 0), %gc_map** %gc_frame.map
  %0 = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 1
  store i8* null, i8** %0
  %gc_frame.next = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 0, i32 0
  %gc_newhead = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 0
  store %gc_stackentry* %gc_currhead, %gc_stackentry** %gc_frame.next
  store %gc_stackentry* %gc_newhead, %gc_stackentry** @llvm_gc_root_chain
  %1 = invoke i8* @malloc(i64 1)
          to label %entry.cont unwind label %gc_cleanup

entry.cont: ; preds = %entry
  store i8* %1, i8** %0
  %gc_frame.next1 = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 0, i32 0
  %gc_savedhead = load %gc_stackentry** %gc_frame.next1
  store %gc_stackentry* %gc_savedhead, %gc_stackentry** @llvm_gc_root_chain
  ret i32 0

gc_cleanup: ; preds = %entry
  %gc_frame.next2 = getelementptr %gc_stackentry.main* %gc_frame, i32 0, i32 0, i32 0
  %gc_savedhead3 = load %gc_stackentry** %gc_frame.next2
  store %gc_stackentry* %gc_savedhead3, %gc_stackentry** @llvm_gc_root_chain
  unwind
}

However, there isn't a personality function associated here. This conflicts with the fact that there's an LSDA associated with the function. It's not really feasible that there would be an LSDA but no personality function. Also note that the landing pad of the 'invoke' doesn't include any of the exception handling intrinsics. During the DwarfEHPrepare pass, an @llvm.eh.exception() call is added (thus marking that as a "landing pad" in code-gen), but there isn't an equivalent @llvm.eh.selector call.

I don't know enough about GC stuff to say what should happen here. Though it looks like we at least need the gc_cleanup to generate an llvm.eh.selector call at the very least...The question becomes which personality function it should call, because that's language-specific.

-bw

Hi Bill,

However, there isn't a personality function associated here. This conflicts with the fact that there's an LSDA associated with the function. It's not really feasible that there would be an LSDA but no personality function.

GCC recently added support for this, i.e. a situation in which there are only
cleanups to be run: it uses the C personality function even if the language
being compiled is not C.

I don't know enough about GC stuff to say what should happen here. Though it looks like we at least need the gc_cleanup to generate an llvm.eh.selector call at the very least...The question becomes which personality function it should call, because that's language-specific.

One reason I added lowering for "unwind" in DwarfEHPrepare was to help handle
this situation: an invoke followed by some cleanup code followed by unwind. I
had GC lowering in mind. I think the only thing missing is having the code
generators use the C personality function by default if there is no personality
function, but as I didn't work on this for a while I don't recall exactly how
far I got.

Ciao, Duncan.

Hi Duncan,

If it's a matter of defaulting to the C personality function, then that should be a fairly easy change to make. I'll look into it, though it may take a backseat to other things that I have to do right now unless someone else wants to.

Nicolas,

Could you file a PR for this?

-bw

Hi Bill,

If it's a matter of defaulting to the C personality function, then that should be a fairly easy change to make. I'll look into it, though it may take a backseat to other things that I have to do right now unless someone else wants to.

I'll add this to the list of exception handling problems: it should be possible
to write this level of simple eh code without a selector, personality etc.

Ciao, Duncan.

Definitely agree. :slight_smile:

-bw