[LLVMdev] Removing AvailableExternal values in GlobalDCE (was Re: RFC: ThinLTO Impementation Plan)

[Teresa Johnson]

On Thu, Jun 4, 2015 at 5:33 PM, Reid Kleckner <rnk at google.com> wrote:
> On Thu, Jun 4, 2015 at 5:17 PM, Teresa Johnson <tejohnson at google.com> wrote:
>>
>> Agreed. Although I assume you mean invoke the new pass under a
>> ThinLTO-only option so that avail extern are not dropped in the
>> compile pass before the LTO link?
>
>
> No, this pass would actually be an improvement to the standard -O2 pipeline.
> The special case is the regular LTO pass pipeline, which wants to run
> GlobalDCE but doesn't want to drop available_externally function definitions
> until after linker-stage inlining.

Ok. It looks to me like the LTO compile step with -O2 -flto -c uses
this same -O2 optimization pipeline as without LTO. Clang communicates
the fact that we are doing an LTO compile to llvm via the
-emit-llvm-bc flag, which just tells it to emit bitcode and exit
before codegen. So I would either need to key off of that or setup a
new flag to indicate to llvm that we are doing an LTO -c compile. Or
is there some other way that I am missing?

Incidentally, we'll also want to skip this new pass and keep any
referenced avail extern functions in the ThinLTO -c compile step for
the same reasons (and there are no imported functions yet at that
point).

Teresa

>
> Consider this test case:
>
> declare void @blah()
> define i32 @main() {
>   call void @foo()
>   ret i32 0
> }
> define available_externally void @foo() noinline {
>   call void @bar()
>   ret void
> }
> define linkonce_odr void @bar() noinline {
>   call void @blah()
>   ret void
> }
>
> If I run opt -O2 on this and feed it to llc, it actually generates code for
> bar, even though there are no actual references to bar in the final code:
>
> main:                                   # @main
>         pushq   %rax
>         callq   foo
>         xorl    %eax, %eax
>         popq    %rdx
>         retq
>
> bar:                                    # @bar
>         jmp     blah                    # TAILCALL
>
> This corner case happens to come up organically with dllimported classes,
> which is why I happened to think of it. :) I'm happy with a flag to
> globaldce for LTO and the original patch, honestly.



-- 
Teresa Johnson | Software Engineer | tejohnson at google.com | 408-460-2413