diff mbox

[4/5] combine: distribute_log_links for PARALLELs of SETs

Message ID 07a14f524da96447dadd570ebf7ee568627aa611.1415984897.git.segher@kernel.crashing.org
State New
Headers show

Commit Message

Segher Boessenkool Nov. 14, 2014, 7:19 p.m. UTC 
Now that LOG_LINKS are per regno, we can distribute them on PARALLELs
just fine.  Do so.  This makes PARALLELs not lose their LOG_LINKS early
when e.g. a trivial reg-reg move is combined, so that they can be used
in more useful combinations as well.


2014-11-14  Segher Boessenkool  <segher@kernel.crashing.org>

gcc/
	* combine.c (distribute_links): Handle multiple SETs.

---
 gcc/combine.c | 52 +++++++++++++++++++++++++++++++++++++---------------
 1 file changed, 37 insertions(+), 15 deletions(-)

Comments

Jeff Law Nov. 25, 2014, 7:37 p.m. UTC | #1 
On 11/14/14 12:19, Segher Boessenkool wrote:
> Now that LOG_LINKS are per regno, we can distribute them on PARALLELs
> just fine.  Do so.  This makes PARALLELs not lose their LOG_LINKS early
> when e.g. a trivial reg-reg move is combined, so that they can be used
> in more useful combinations as well.
>
>
> 2014-11-14  Segher Boessenkool  <segher@kernel.crashing.org>
>
> gcc/
> 	* combine.c (distribute_links): Handle multiple SETs.
So the code in distribute_links implies that we're not going to see hard 
register SUBREGs, so ignore my concerns with the prior patch in this 
series WRT hard register SUBREGs.


This is OK once prereqs are approved.

You might consider pushing the two LOG_LINKs related patches forward 
independently of the patch to rip apart the PARALLELs.  Though I think 
that all of the patches are pretty close to being approved.  Your call.

Jeff
Segher Boessenkool Nov. 25, 2014, 9:54 p.m. UTC | #2 
On Tue, Nov 25, 2014 at 12:37:57PM -0700, Jeff Law wrote:
> This is OK once prereqs are approved.

That is only 3/5 now.

> You might consider pushing the two LOG_LINKs related patches forward 
> independently of the patch to rip apart the PARALLELs.  Though I think 
> that all of the patches are pretty close to being approved.  Your call.

I probably will; 2/5 (the "rip" patch) needs updating, and my patch stack
is much too high :-)  Not that [34]/5 do terribly much without 2/5.

Thanks for all the reviews,


Segher
diff mbox

Patch

diff --git a/gcc/combine.c b/gcc/combine.c
index cf184db..8c7f052 100644
--- a/gcc/combine.c
+++ b/gcc/combine.c
@@ -13832,24 +13832,46 @@  distribute_links (struct insn_link *links)
 
       next_link = link->next;
 
-      /* If the insn that this link points to is a NOTE or isn't a single
-	 set, ignore it.  In the latter case, it isn't clear what we
-	 can do other than ignore the link, since we can't tell which
-	 register it was for.  Such links wouldn't be used by combine
-	 anyway.
-
-	 It is not possible for the destination of the target of the link to
-	 have been changed by combine.  The only potential of this is if we
-	 replace I3, I2, and I1 by I3 and I2.  But in that case the
-	 destination of I2 also remains unchanged.  */
-
-      if (NOTE_P (link->insn)
-	  || (set = single_set (link->insn)) == 0)
+      /* If the insn that this link points to is a NOTE, ignore it.  */
+      if (NOTE_P (link->insn))
+	continue;
+
+      set = 0;
+      rtx pat = PATTERN (link->insn);
+      if (GET_CODE (pat) == SET)
+	set = pat;
+      else if (GET_CODE (pat) == PARALLEL)
+	{
+	  int i;
+	  for (i = 0; i < XVECLEN (pat, 0); i++)
+	    {
+	      set = XVECEXP (pat, 0, i);
+	      if (GET_CODE (set) != SET)
+		continue;
+
+	      reg = SET_DEST (set);
+	      while (GET_CODE (reg) == ZERO_EXTRACT
+		     || GET_CODE (reg) == STRICT_LOW_PART
+		     || GET_CODE (reg) == SUBREG)
+		reg = XEXP (reg, 0);
+
+	      if (!REG_P (reg))
+		continue;
+
+	      if (REGNO (reg) == link->regno)
+		break;
+	    }
+	  if (i == XVECLEN (pat, 0))
+	    continue;
+	}
+      else
 	continue;
 
       reg = SET_DEST (set);
-      while (GET_CODE (reg) == SUBREG || GET_CODE (reg) == ZERO_EXTRACT
-	     || GET_CODE (reg) == STRICT_LOW_PART)
+
+      while (GET_CODE (reg) == ZERO_EXTRACT
+	     || GET_CODE (reg) == STRICT_LOW_PART
+	     || GET_CODE (reg) == SUBREG)
 	reg = XEXP (reg, 0);
 
       /* A LOG_LINK is defined as being placed on the first insn that uses