On Mon, Mar 19, 2018 at 01:31:30PM -0700, Matthias Kaehlcke wrote:
> > The ORC unwinder relies on objtool, which reverse engineers the compiled
> > code. This is objtool's first exposure to clang, so I'm not at all
> > surprised if it's getting confused.
> >
> > Send me one of the .o files and I can take a quick look to see how bad
> > it is, but I'm guessing it's going to be a lot of work to make objtool
> > compatible with clang (and unfortunately I won't have the bandwidth to
> > work on that in the near term.)
> >
> > In the meantime I'd recommend that you use frame pointers (and
> > CONFIG_STACK_VALIDATION=n) for clang-compiled kernels.
>
> Thanks for your assessment!
>
> dvo_ch7017.o is attached.
Here's a (surprisingly easy) fix for this particular issue, though I'd
be shocked if there weren't a bunch more issues lurking elsewhere. Let
me know how it goes.
BTW, one thing I noticed in the .o file is that most of the functions'
stacks are aligned to 16 bytes. It might be worth checking if the clang
-mstack-alignment=8 option is getting set, and if so, if it's working
properly. Otherwise, with aligned stacks, the frame pointer is forced,
which defeats most of the benefits of ORC.
diff --git a/tools/objtool/check.c b/tools/objtool/check.c
index 92b6a2c21631..f02df714c18e 100644
--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -1386,6 +1386,17 @@ static int update_insn_state(struct instruction *insn,
struct insn_state *state)
state->vals[op->dest.reg].offset =
-state->stack_size;
}
+ else if (op->src.reg == CFI_BP && op->dest.reg ==
CFI_SP &&
+ cfa->base == CFI_BP) {
+
+ /*
+ * mov %rbp, %rsp
+ *
+ * Restore the original stack pointer (clang).
+ */
+ state->stack_size = -state->regs[CFI_BP].offset;
+ }
+
else if (op->dest.reg == cfa->base) {
/* mov %reg, %rsp */
