[llvm-dev] (LLD / lto ) How to avoid GOT code generation

[Fāng-ruì Sòng via llvm-dev]

On Sat, Jan 9, 2021 at 10:45 PM Fāng-ruì Sòng <maskray at google.com> wrote:
>
> On Sat, Jan 9, 2021 at 12:03 PM Moshtaghi, Alireza via llvm-dev
> <llvm-dev at lists.llvm.org> wrote:
> >
> > Hi
> >
> > I’m generating a special shared object in x86_64 large memory model (-mcmodel=large).
> >
> > For this object, I pass -z notext to lld to leave the text relocations alone and normally it builds and works with no problem.
> >
> > But when I add link time optimization, lld generates R_X86_64_GOTOFF64 which stops itself from linking. (Note that when I objdump -r on my elf objects without lto, this relocation is not generated; but for some reason lld decides to treat the non pic lto object as a pic object and generate got/pic code)
> >
> > How can I tell lld to not generate GOT / PIC code ?
> >
> >
> >
> > Here is the example:
> >
> > When I compile as follows, my shared lib is generated and works fine
> >
> > clang -mcmodel=large -o sample.o -c sample.c.  # -flto=thin fails
> >
> > ld.lld -Bshareable -z notext -o out.so sample.o
> >
> >
> >
> > but when I compile with -flto=thin, lld errors that R_X86_64_GOTOFF64 can’t be used against foo and foo_call and wants me to compile with -fPIC but I don’t want to use PIC and GOT
> >
> >
> >
> > sample.c :
> >
> > extern int foo;
> >
> > int* bar = &foo;
> >
> >
> >
> > int foo_call (int, int *);
> >
> >
> >
> > int foo_call (int a, int *b) {
> >
> >   return a+ *b;
> >
> > }
> >
> >
> >
> > int dummy (void) {
> >
> >   int *fooptr = &foo;
> >
> >   return foo_call (1, fooptr);
> >
> > }
>
> clang -fpic -flto -mcmodel=large -o sample.o -c sample.c will work.
>
> Your compile mode is -fno-pic (Clang default for Linux), which can
> only be linked in -no-pie mode.
> -fpie is compatible with -no-pie and -pie.
> -fpic is compatible with -no-pie, -pie and -shared (-Bsharable).
>
> -fno-pic + -shared can sometimes work, work in more cases with -z
> notext, but still not always.
>
> For this case: the formula for R_X86_64_GOTOFF64 is S+A-GOT where S
> represents the symbol value.
> Since foo is undefined, the relocation cannot be resolved at link
> time. LLD does not produce a dynamic relocation
> because R_X86_64_GOTOFF64 is not a generally acceptable dynamic
> relocation type by ld.so implementations.

I'd say R_X86_64_GOTOFF64 is a direct access relocation type, not a
GOT indirection relocation type.

A GOT indirection relocation type should compute to a place in GOT
(e.g. G+A, G+GOT+A-P), but R_X86_64_GOTOFF64 (S+A-GOT) is different.
"GOT" refers to the address of _GLOBAL_OFFSET_TABLE_.

The large code model needs the GOT base ("GOT" in x86-64 psABI:
"Represents the address of the global offset table") as a fixed place
in the text segment to
compute the addresses of symbols. The GOT base is not used in the GOT
indirection manner.