public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [PATCH] Punt on hard reg setters in modulo sched (PR target/83507)
@ 2018-01-09  0:01 Jakub Jelinek
  2018-01-09  6:57 ` Richard Biener
  0 siblings, 1 reply; 2+ messages in thread
From: Jakub Jelinek @ 2018-01-09  0:01 UTC (permalink / raw)
  To: Richard Biener, Jeff Law, Vladimir Makarov; +Cc: gcc-patches

Hi!

SMS is done before reload and assumes it can extend the lifetime by making
copies of the destination registers without trying to re-recognize the
instructions etc.  That can work only with pseudos, for some hard registers
there isn't even a valid move instruction to pseudo (e.g. the powerpc CA
register), for others it can be very much undesirable.

Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

2018-01-09  Jakub Jelinek  <jakub@redhat.com>

	PR target/83507
	* modulo-sched.c (schedule_reg_moves): Punt if we'd need to move
	hard registers.  Formatting fixes.

	* gcc.dg/sms-13.c: New test.

--- gcc/modulo-sched.c.jj	2018-01-03 10:19:55.710534046 +0100
+++ gcc/modulo-sched.c	2018-01-08 16:53:01.232465389 +0100
@@ -687,9 +687,9 @@ schedule_reg_moves (partial_schedule_ptr
       rtx set = single_set (u->insn);
       
       /* Skip instructions that do not set a register.  */
-      if ((set && !REG_P (SET_DEST (set))))
+      if (set && !REG_P (SET_DEST (set)))
         continue;
- 
+
       /* Compute the number of reg_moves needed for u, by looking at life
 	 ranges started at u (excluding self-loops).  */
       distances[0] = distances[1] = false;
@@ -743,7 +743,10 @@ schedule_reg_moves (partial_schedule_ptr
       first_move += ps->g->num_nodes;
 
       /* Generate each move.  */
-      old_reg = prev_reg = SET_DEST (single_set (u->insn));
+      old_reg = prev_reg = SET_DEST (set);
+      if (HARD_REGISTER_P (old_reg))
+	return false;
+
       for (i_reg_move = 0; i_reg_move < nreg_moves; i_reg_move++)
 	{
 	  ps_reg_move_info *move = ps_reg_move (ps, first_move + i_reg_move);
--- gcc/testsuite/gcc.dg/sms-13.c.jj	2018-01-08 17:00:16.427534076 +0100
+++ gcc/testsuite/gcc.dg/sms-13.c	2018-01-08 16:56:15.704496101 +0100
@@ -0,0 +1,32 @@
+/* PR target/83507 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fmodulo-sched -fno-tree-ter -fno-tree-coalesce-vars" } */
+
+void
+foo (unsigned short int x, unsigned char y)
+{
+  unsigned char *a = &y;
+  unsigned short int b;
+  int c;
+
+  while (y < 3)
+    {
+      if (x != 0)
+        ++y;
+      ++y;
+    }
+
+  for (c = 0; c < 5; ++c)
+    {
+      int d = 1;
+      d += b > x;
+      y &= d;
+    }
+
+  do
+    {
+      c += y;
+      x = c;
+    }
+  while (x != 0);
+}

	Jakub

^ permalink raw reply	[flat|nested] 2+ messages in thread

* Re: [PATCH] Punt on hard reg setters in modulo sched (PR target/83507)
  2018-01-09  0:01 [PATCH] Punt on hard reg setters in modulo sched (PR target/83507) Jakub Jelinek
@ 2018-01-09  6:57 ` Richard Biener
  0 siblings, 0 replies; 2+ messages in thread
From: Richard Biener @ 2018-01-09  6:57 UTC (permalink / raw)
  To: Jakub Jelinek, Jeff Law, Vladimir Makarov; +Cc: gcc-patches

On January 9, 2018 12:59:04 AM GMT+01:00, Jakub Jelinek <jakub@redhat.com> wrote:
>Hi!
>
>SMS is done before reload and assumes it can extend the lifetime by
>making
>copies of the destination registers without trying to re-recognize the
>instructions etc.  That can work only with pseudos, for some hard
>registers
>there isn't even a valid move instruction to pseudo (e.g. the powerpc
>CA
>register), for others it can be very much undesirable.
>
>Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

OK. 

Richard. 

>2018-01-09  Jakub Jelinek  <jakub@redhat.com>
>
>	PR target/83507
>	* modulo-sched.c (schedule_reg_moves): Punt if we'd need to move
>	hard registers.  Formatting fixes.
>
>	* gcc.dg/sms-13.c: New test.
>
>--- gcc/modulo-sched.c.jj	2018-01-03 10:19:55.710534046 +0100
>+++ gcc/modulo-sched.c	2018-01-08 16:53:01.232465389 +0100
>@@ -687,9 +687,9 @@ schedule_reg_moves (partial_schedule_ptr
>       rtx set = single_set (u->insn);
>       
>       /* Skip instructions that do not set a register.  */
>-      if ((set && !REG_P (SET_DEST (set))))
>+      if (set && !REG_P (SET_DEST (set)))
>         continue;
>- 
>+
>    /* Compute the number of reg_moves needed for u, by looking at life
> 	 ranges started at u (excluding self-loops).  */
>       distances[0] = distances[1] = false;
>@@ -743,7 +743,10 @@ schedule_reg_moves (partial_schedule_ptr
>       first_move += ps->g->num_nodes;
> 
>       /* Generate each move.  */
>-      old_reg = prev_reg = SET_DEST (single_set (u->insn));
>+      old_reg = prev_reg = SET_DEST (set);
>+      if (HARD_REGISTER_P (old_reg))
>+	return false;
>+
>       for (i_reg_move = 0; i_reg_move < nreg_moves; i_reg_move++)
> 	{
> 	  ps_reg_move_info *move = ps_reg_move (ps, first_move + i_reg_move);
>--- gcc/testsuite/gcc.dg/sms-13.c.jj	2018-01-08 17:00:16.427534076
>+0100
>+++ gcc/testsuite/gcc.dg/sms-13.c	2018-01-08 16:56:15.704496101 +0100
>@@ -0,0 +1,32 @@
>+/* PR target/83507 */
>+/* { dg-do compile } */
>+/* { dg-options "-O2 -fmodulo-sched -fno-tree-ter
>-fno-tree-coalesce-vars" } */
>+
>+void
>+foo (unsigned short int x, unsigned char y)
>+{
>+  unsigned char *a = &y;
>+  unsigned short int b;
>+  int c;
>+
>+  while (y < 3)
>+    {
>+      if (x != 0)
>+        ++y;
>+      ++y;
>+    }
>+
>+  for (c = 0; c < 5; ++c)
>+    {
>+      int d = 1;
>+      d += b > x;
>+      y &= d;
>+    }
>+
>+  do
>+    {
>+      c += y;
>+      x = c;
>+    }
>+  while (x != 0);
>+}
>
>	Jakub

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2018-01-09  6:56 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2018-01-09  0:01 [PATCH] Punt on hard reg setters in modulo sched (PR target/83507) Jakub Jelinek
2018-01-09  6:57 ` Richard Biener

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).