Re: [PATCH] [i386] Fix _mm256_zeroupper to notify LRA that vzeroupper will kill sse registers. [PR target/82735]

[Hongtao Liu <crazylht@gmail.com>]

On Fri, May 14, 2021 at 10:27 AM Hongtao Liu <crazylht@gmail.com> wrote:
>
> On Thu, May 13, 2021 at 7:52 PM Richard Sandiford
> <richard.sandiford@arm.com> wrote:
> >
> > Jakub Jelinek <jakub@redhat.com> writes:
> > > On Thu, May 13, 2021 at 12:32:26PM +0100, Richard Sandiford wrote:
> > >> Jakub Jelinek <jakub@redhat.com> writes:
> > >> > On Thu, May 13, 2021 at 11:43:19AM +0200, Uros Bizjak wrote:
> > >> >> > >   Bootstrapped and regtested on X86_64-linux-gnu{-m32,}
> > >> >> > >   Ok for trunk?
> > >> >> >
> > >> >> > Some time ago a support for CLOBBER_HIGH RTX was added (and later
> > >> >> > removed for some reason). Perhaps we could resurrect the patch for the
> > >> >> > purpose of ferrying 128bit modes via vzeroupper RTX?
> > >> >>
> > >> >> https://gcc.gnu.org/legacy-ml/gcc-patches/2017-11/msg01325.html
> > >> >
> > >> > https://gcc.gnu.org/legacy-ml/gcc-patches/2019-09/msg01468.html
> > >> > is where it got removed, CCing Richard.
> > >>
> > >> Yeah.  Initially clobber_high seemed like the best appraoch for
> > >> handling the tlsdesc thing, but in practice it was too difficult
> > >> to shoe-horn the concept in after the fact, when so much rtl
> > >> infrastructure wasn't prepared to deal with it.  The old support
> > >> didn't handle all cases and passes correctly, and handled others
> > >> suboptimally.
> > >>
> > >> I think it would be worth using the same approach as
> > >> https://gcc.gnu.org/legacy-ml/gcc-patches/2019-09/msg01466.html for
> > >> vzeroupper: represent the instructions as call_insns in which the
> > >> call has a special vzeroupper ABI.  I think that's likely to lead
> > >> to better code than clobber_high would (or at least, it did for tlsdesc).
>
> From an implementation perspective， I guess you're meaning we should
> implement TARGET_INSN_CALLEE_ABI and TARGET_FNTYPE_ABI in the i386
> backend.
>
When I implemented the vzeroupper pattern as call_insn and defined
TARGET_INSN_CALLEE_ABI for it, I got several failures. they're related
to 2 parts

1. requires_stack_frame_p return true for vzeroupper which should be false.
2. in subst_stack_regs, vzeroupper shouldn't kill arguments

I've tried a rough patch like below, it works for those failures,
unfortunately, I don't have an arm machine to test, so I want to ask
would the below change break something in the arm backend?

modified   gcc/reg-stack.c
@@ -174,6 +174,7 @@
 #include "reload.h"
 #include "tree-pass.h"
 #include "rtl-iter.h"
+#include "function-abi.h"

 #ifdef STACK_REGS

@@ -2385,7 +2386,7 @@ subst_stack_regs (rtx_insn *insn, stack_ptr regstack)
   bool control_flow_insn_deleted = false;
   int i;

-  if (CALL_P (insn))
+  if (CALL_P (insn) && insn_callee_abi (insn).id () == 0)
     {
       int top = regstack->top;

modified   gcc/shrink-wrap.c
@@ -58,7 +58,12 @@ requires_stack_frame_p (rtx_insn *insn,
HARD_REG_SET prologue_used,
   unsigned regno;

   if (CALL_P (insn))
-    return !SIBLING_CALL_P (insn);
+    {
+      if (insn_callee_abi (insn).id() != 0)
+ return false;
+      else
+ return !SIBLING_CALL_P (insn);
+    }

   /* We need a frame to get the unique CFA expected by the unwinder.  */
   if (cfun->can_throw_non_call_exceptions && can_throw_internal (insn))
> > >
> > > Perhaps a magic call_insn that is split post-reload into a normal insn
> > > with the sets then?
> >
> > I'd be tempted to treat it is a call_insn throughout.  The unspec_volatile
> > means that we can't move the instruction, so converting a call_insn to an
> > insn isn't likely to help from that point of view.  The sets are also
> > likely to be handled suboptimally compared to the more accurate register
> > information attached to the call: all code that handles calls has to be
> > prepared to deal with partial clobbers, whereas most code dealing with
> > sets will assume that the set does useful work, and that the rhs of the
> > set is live.
> >
> > Thanks,
> > Richard
> >
>
>
> --
> BR,
> Hongtao



-- 
BR,
Hongtao