public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [patch] PR51477
@ 2012-11-12  9:04 Steven Bosscher
  2012-11-12 19:48 ` Richard Henderson
  0 siblings, 1 reply; 2+ messages in thread
From: Steven Bosscher @ 2012-11-12  9:04 UTC (permalink / raw)
  To: GCC Patches; +Cc: Jakub Jelinek

[-- Attachment #1: Type: text/plain, Size: 792 bytes --]

Hello,

In this PR, a set to a global reg is removed by fast-dce and ud-dce.
Such sets should always be preserved.

The proposed fix is to make global regs always live. This is done in
the df-* parts of the patch from Jakub (the test case is also his).
In addition, ud-dce should consider sets of global regs inherently
necessary for DCE.

Bootstrapped and tested on x86_64-unknown-linux-gnu. OK for trunk?

Ciao!
Steven

gcc/
	PR rtl-optimization/51447
	* df-scan.c (df_get_entry_block_def_set): Add global regs to the set.
	* df-problems.c (df_lr_local_compute): Make global regs always live.

	PR rtl-optimization/51447
	* dce.c (deletable_insn_p): Make insns setting a global reg
	inherently necessary.

testsuite/
	PR rtl-optimization/51447
	* gcc.c-torture/execute/pr51447.c: New test.

[-- Attachment #2: PR51447.diff --]
[-- Type: application/octet-stream, Size: 3205 bytes --]

gcc/
	PR rtl-optimization/51447
	* df-scan.c (df_get_entry_block_def_set): Add global regs to the set.
	* df-problems.c (df_lr_local_compute): Make global regs always live.
	* dce.c (deletable_insn_p): Make insns setting a global reg
	inherently necessary.

testsuite/
	PR rtl-optimization/51447
	* gcc.c-torture/execute/pr51447.c: New test.

Index: df-scan.c
===================================================================
--- df-scan.c	(revision 193411)
+++ df-scan.c	(working copy)
@@ -3790,8 +3790,12 @@ df_get_entry_block_def_set (bitmap entry
   bitmap_clear (entry_block_defs);
 
   for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
-    if (FUNCTION_ARG_REGNO_P (i))
-      bitmap_set_bit (entry_block_defs, INCOMING_REGNO (i));
+    {
+      if (global_regs[i])
+	bitmap_set_bit (entry_block_defs, i);
+      if (FUNCTION_ARG_REGNO_P (i))
+	bitmap_set_bit (entry_block_defs, INCOMING_REGNO (i));
+    }
 
   /* The always important stack pointer.  */
   bitmap_set_bit (entry_block_defs, STACK_POINTER_REGNUM);
Index: df-problems.c
===================================================================
--- df-problems.c	(revision 193411)
+++ df-problems.c	(working copy)
@@ -931,7 +931,7 @@ df_lr_bb_local_compute (unsigned int bb_
 static void
 df_lr_local_compute (bitmap all_blocks ATTRIBUTE_UNUSED)
 {
-  unsigned int bb_index;
+  unsigned int bb_index, i;
   bitmap_iterator bi;
 
   bitmap_clear (&df->hardware_regs_used);
@@ -939,6 +939,11 @@ df_lr_local_compute (bitmap all_blocks A
   /* The all-important stack pointer must always be live.  */
   bitmap_set_bit (&df->hardware_regs_used, STACK_POINTER_REGNUM);
 
+  /* Global regs are always live, too.  */
+  for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+    if (global_regs[i])
+      bitmap_set_bit (&df->hardware_regs_used, i);
+
   /* Before reload, there are a few registers that must be forced
      live everywhere -- which might not already be the case for
      blocks within infinite loops.  */
Index: dce.c
===================================================================
--- dce.c	(revision 193411)
+++ dce.c	(working copy)
@@ -121,6 +121,12 @@ deletable_insn_p (rtx insn, bool fast, b
       && !insn_nothrow_p (insn))
     return false;
 
+  /* If INSN sets a global_reg, leave it untouched.  */
+  for (df_ref *def_rec = DF_INSN_DEFS (insn); *def_rec; def_rec++)
+    if (HARD_REGISTER_NUM_P (DF_REF_REGNO (*def_rec))
+	&& global_regs[DF_REF_REGNO (*def_rec)])
+      return false;
+
   body = PATTERN (insn);
   switch (GET_CODE (body))
     {
Index: testsuite/gcc.c-torture/execute/pr51447.c
===================================================================
--- testsuite/gcc.c-torture/execute/pr51447.c	(revision 0)
+++ testsuite/gcc.c-torture/execute/pr51447.c	(revision 0)
@@ -0,0 +1,27 @@
+/* PR rtl-optimization/51447 */
+
+extern void abort (void);
+
+#ifdef __x86_64__
+register void *ptr asm ("rbx");
+#else
+void *ptr;
+#endif
+
+int
+main (void)
+{
+  __label__ nonlocal_lab;
+  __attribute__((noinline, noclone)) void
+    bar (void *func)
+      {
+	ptr = func;
+	goto nonlocal_lab;
+      }
+  bar (&&nonlocal_lab);
+  return 1;
+nonlocal_lab:
+  if (ptr != &&nonlocal_lab)
+    abort ();
+  return 0;
+}

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

* Re: [patch] PR51477
  2012-11-12  9:04 [patch] PR51477 Steven Bosscher
@ 2012-11-12 19:48 ` Richard Henderson
  0 siblings, 0 replies; 2+ messages in thread
From: Richard Henderson @ 2012-11-12 19:48 UTC (permalink / raw)
  To: Steven Bosscher; +Cc: GCC Patches, Jakub Jelinek

On 11/12/2012 01:04 AM, Steven Bosscher wrote:
> gcc/
> 	PR rtl-optimization/51447
> 	* df-scan.c (df_get_entry_block_def_set): Add global regs to the set.
> 	* df-problems.c (df_lr_local_compute): Make global regs always live.
> 	* dce.c (deletable_insn_p): Make insns setting a global reg
> 	inherently necessary.
> 
> testsuite/
> 	PR rtl-optimization/51447
> 	* gcc.c-torture/execute/pr51447.c: New test.

Ok.


r~

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

end of thread, other threads:[~2012-11-12 19:48 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2012-11-12  9:04 [patch] PR51477 Steven Bosscher
2012-11-12 19:48 ` Richard Henderson

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).