On Fri, May 29, 2026 at 4:51 AM Hongtao Liu <[email protected]> wrote:
>
> On Thu, May 28, 2026 at 5:59 PM Uros Bizjak <[email protected]> wrote:
> >
> > On Thu, May 28, 2026 at 11:48 AM Jakub Jelinek <[email protected]> wrote:
> > >
> > > On Thu, May 28, 2026 at 11:31:27AM +0200, Uros Bizjak wrote:
> > > > Hm, OTOH, is this condition correct:
> > > >
> > > > case TYPE_LEA:
> > > > if (TARGET_APX_NDD && <nf_applied>)
> > > > return "%{nf%} add{<imodesuffix>}\t{%2, %1, %0|%0, %1, %2}";
> > > > else
> > > > return "#";
> > > >
> > > > Should it be "TARGET_APX_NDD || <nf_applied>"?
> > >
> > > The pattern is weird, but TARGET_APX_NDD || <nf_applied> would be wrong.
> > > Without TARGET_APX_NDD there is no 3 operand add, so we can't use that.
> > > Of course, question is if there will be CPUs with just APX-NF and not
> > > APX-NDD or with just APX-NDD and not APX-NF.
> > >
> > > I'd say it is a bad idea to have the "lea" variant come before the
> > > 3 apx_ndd variants, if the lea variant came last, then at least the r <-
> > > r, l
> > > case would be covered for TARGET_APX_NDD by the earlier (but now later)
> > > r <- rje, r variant and could then emit either the 3 operand add or 3
> > > operand add with {nf}. Anyway, with the current position it could also be
> > > if (TARGET_APX_NDD)
> > > return "<nf_prefix>add{<imodesuffix>}\t{%2, %1, %0|%0, %1, %2}";
> > > else
> > > {
> > > operands[3] = gen_rtx_PLUS (<MODE>mode, operands[1], operands[2]);
> > > return "lea{<imodesuffix>}\t{%E3, %0|%0, %E3}";
> > > }
> >
> > I think that the correct solution is to put LEA at the end. LEA
> > alternative is intended for non-NDD cases to allow two-register
> > operation, but NDD should override it, w/ or w/o NF. Let's ask
> > Hongtao.
> LEA and NF ADD sometimes have the same rtx pattern, we put the LEA
> pattern first so that NF ADD rtx pattern can be recognized as LEA when
> possible. Both have similar performance, but LEA uses less encoding
> space (especially without EGPR - legacy LEA is smaller than NF ADD).
> Similar reasons, we want NDD ADD to be splitted back to LEA.
LEA is not fully compatible with ADD because %rsp can't be used as an
index register (this is a general limitation of address expression
which LEA inherits). As shown by Jakub, there are some problems when
instructions are shadowing each other. In this particular case, the
splitter at i386.md::7215 is not triggered for some reason, so we just
skip the split and emit LEA by hand.
Also, this opens a question - is this part in x86_validate_address_register:
/* Don't allow SUBREGs that span more than a word. It can
lead to spill failures when the register is one word out
of a two word structure. */
if (GET_MODE_SIZE (mode) > UNITS_PER_WORD)
return NULL_RTX;
still needed, or is it here only because of some old reload limitation?
Uros.
