https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116398
Richard Sandiford <rsandifo at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |rsandifo at gcc dot
gnu.org
--- Comment #28 from Richard Sandiford <rsandifo at gcc dot gnu.org> ---
I'm back from holiday, so taking.
(In reply to Segher Boessenkool from comment #26)
> So, the one thing I really worry about a bit: will everything still work if
> we can lose some log_links in some (unusual) cases? Or does anything rely on
> that never happening.
Yeah, that was my concern at first as well. But I don't think the lack of a
log link enables any transformations that would otherwise not occur. All
processing of log links seems to handle the absence of a link for a particular
register in a conservative way.
Even the initial set-up errs on the side of dropping links, since for example
create_log_links has:
/* flow.c claimed:
We don't build a LOG_LINK for hard registers contained
in ASM_OPERANDs. If these registers get replaced,
we might wind up changing the semantics of the insn,
even if reload can make what appear to be valid
assignments later. */
if (regno < FIRST_PSEUDO_REGISTER
&& asm_noperands (PATTERN (use_insn)) >= 0)
continue;
which excludes combinations by dropping log links, rather than during
try_combine. And:
/* If this register is being initialized using itself, and the
register is uninitialized in this basic block, and there are
no LOG_LINKS which set the register, then part of the
register is uninitialized. In that case we can't assume
anything about the number of nonzero bits.
??? We could do better if we checked this in
reg_{nonzero_bits,num_sign_bit_copies}_for_combine. Then we
could avoid making assumptions about the insn which initially
sets the register, while still using the information in other
insns. We would have to be careful to check every insn
involved in the combination. */
if (insn
&& reg_referenced_p (x, PATTERN (insn))
&& !REGNO_REG_SET_P (DF_LR_IN (BLOCK_FOR_INSN (insn)),
REGNO (x)))
{
struct insn_link *link;
FOR_EACH_LOG_LINK (link, insn)
if (dead_or_set_p (link->insn, x))
break;
if (!link)
{
rsp->nonzero_bits = GET_MODE_MASK (mode);
rsp->sign_bit_copies = 1;
return;
}
}
treats the lack of a log link is a possible sign of uninitialised data, but
that would be a missed optimisation rather than a correctness issue.
