On Thu, Jul 4, 2024 at 3:24 AM Eduard Zingerman <eddyz87@xxxxxxxxx> wrote:
>
> + for (j = 1; j <= spills_num; ++j) {
> + if ((insn - j)->off >= subprogs[cur_subprog].nocsr_stack_off ||
> + (insn + j)->off >= subprogs[cur_subprog].nocsr_stack_off) {
> + /* do a second visit of this instruction,
> + * so that verifier can inline it
> + */
> + i -= 1;
> + insn -= 1;
> + goto next_insn;
> + }
> + }
> +
> + /* apply the rewrite:
> + * *(u64 *)(r10 - X) = rY ; num-times
> + * call() -> call()
> + * rY = *(u64 *)(r10 - X) ; num-times
> + */
> + err = verifier_remove_insns(env, i + delta - spills_num, spills_num);
> + if (err)
> + return err;
> + err = verifier_remove_insns(env, i + delta - spills_num + 1, spills_num);
> + if (err)
> + return err;
> +
> + i += spills_num - 1;
> + /* ^ ^ do a second visit of this instruction,
> + * | '-- so that verifier can inline it
> + * '--------------- jump over deleted fills
> + */
> + delta -= 2 * spills_num;
> + insn = env->prog->insnsi + i + delta;
> + goto next_insn;
> + }
somewhere after spill/fill removal subprog->stack_depth
needs to be adjust to nocsr_stack_off,
otherwise extra stack space is wasted.
I couldn't find this logic in the patch.
Once the adjustment logic is done, pls add a selftest with
nocsr and may_goto, since may_goto processing is in the same
do_misc_fixups() loop and it needs to grow the stack while
spill/fill removal will shrink the stack.
