[PATCH] S/390: Fix PR89952 incorrect CFI

Robin Dapp rdapp@linux.vnet.ibm.com
Thu Apr 18 14:55:00 GMT 2019


> +   Establish an ANTI dependency between r11 and r15 restores from FPRs
> +   to prevent the instructions scheduler from reordering them since
> +   this would break CFI.  No further handling in the sched_reorder
> +   hook is required since the r11 and r15 restore will never appear in
> +   the same ready list with that change.  */
> +  if (r11_restore == NULL || r15_restore == NULL)
> +    return;
> +  add_dependence (r11_restore, r15_restore, REG_DEP_ANTI);
> +}

an anti dependency seems an odd choice because r15 would be restored
before r11 (according to the patch's changes in
s390_restore_gprs_from_fprs) and the CFI note for r11 reads
from/requires r15.  This would rather indicate a data dependency from
r15 to r11 but I understand that you don't want the scheduler to assume
there is a real dependency including effects on insn priority and latency.

I looked in haifa-sched.c and sched-ebb.c for methods to prevent
reordering two "special" instructions but did not find anything. There
is of course code to prevent scheduling stack-referencing insns across a
stack pointer restore but it doesn't seem to allow adding more

I came across add_deps_for_risky_insns, though, which also uses anti
dependencies, apparently for a similar reason, so maybe that's the
normal and accepted way to do it.


More information about the Gcc-patches mailing list