On Mon, Aug 14, 2017 at 1:11 PM, Dibyendu Majumdar
<mobile@xxxxxxxxxxxxxxx> wrote:
> I assume that you have not yet applied this approach to Sparse-LLVM
> and tested that it works for var arg calls? I would suggest completing
> the changes so that it can be tested thoroughly.
No. this is just one night's hack to make the plumbing part work.
Provide a size for the VALUE pseudo. I haven't have a chance to
integrate into sparse-llvm.
> The push approach has the benefit that it makes clear what arguments
> are being passed in a call. Previously this information was not
> revealed in the Sparse linearized dump. It would be a shame to lose
> that information - perhaps you could still dump the arguments along
> with a call?
Are you sure?
Here is what I have for RC5:
./test-linearize cse.c
.L414:
load.64 %r1052 <- 0[insn]
call.64 %r1053 <- try_to_cse, %arg1, %r1044(last), %r1052
store.64 %r1053 -> 0[insn]
phisrc.64 %phi474 <- %r1053
br .L413
You can see exactly what is the calling arguments. There is an pseudo_list
for the arguments.
struct /* call */ {
pseudo_t func;
struct pseudo_list *arguments; <==============
struct symbol *fntype;
};
The only thing you want, which is not there is the VALUE pseudo_size.
I add that in this patch.
I haven't take a deep look into the push instruction patch yet. I image this is
less impact on the IR and get the job done. With less redundant information
in the IR. The instruction stream is more compact as well.
Chris
--
To unsubscribe from this list: send the line "unsubscribe linux-sparse" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
