diff mbox

[3/9] dce: Don't dead-code delete separately wrapped restores

Message ID 4af162bf22c8a405100c31f57ac2c3b0de9b721e.1470015604.git.segher@kernel.crashing.org
State New
Headers show

Commit Message

Segher Boessenkool Aug. 1, 2016, 1:42 a.m. UTC 
Deleting restores (before a noreturn) that are dead confuses dwarf2cfi.

2016-06-07  Segher Boessenkool  <segher@kernel.crashing.org>

	* dce.c (delete_unmarked_insns): Don't delete instructions with
	a REG_CFA_RESTORE note.
---
 gcc/dce.c | 9 +++++++++
 1 file changed, 9 insertions(+)

Comments

Jeff Law Sept. 8, 2016, 5:50 p.m. UTC | #1 
On 07/31/2016 07:42 PM, Segher Boessenkool wrote:
> Deleting restores (before a noreturn) that are dead confuses dwarf2cfi.
>
> 2016-06-07  Segher Boessenkool  <segher@kernel.crashing.org>
>
> 	* dce.c (delete_unmarked_insns): Don't delete instructions with
> 	a REG_CFA_RESTORE note.
I don't really understand this one.  Why is the restore marked dead and 
why doesn't that happen for normal epilogues?  Something wonky seems to 
be going on here.

jeff
Segher Boessenkool Sept. 9, 2016, 3:51 p.m. UTC | #2 
On Thu, Sep 08, 2016 at 11:50:56AM -0600, Jeff Law wrote:
> On 07/31/2016 07:42 PM, Segher Boessenkool wrote:
> >Deleting restores (before a noreturn) that are dead confuses dwarf2cfi.
> >
> >2016-06-07  Segher Boessenkool  <segher@kernel.crashing.org>
> >
> >	* dce.c (delete_unmarked_insns): Don't delete instructions with
> >	a REG_CFA_RESTORE note.
> I don't really understand this one.  Why is the restore marked dead and 
> why doesn't that happen for normal epilogues?  Something wonky seems to 
> be going on here.

Because it's not behind a NOTE_INSN_EPILOGUE_BEGIN it is not treated
specially by DCE, like insns in "normal" epilogues are.

It is marked dead because it *is* dead: the value in the callee-save
register is not used by anything anymore (noreturn!)


Segher
Jeff Law Sept. 9, 2016, 4:25 p.m. UTC | #3 
On 09/09/2016 09:51 AM, Segher Boessenkool wrote:
> On Thu, Sep 08, 2016 at 11:50:56AM -0600, Jeff Law wrote:
>> On 07/31/2016 07:42 PM, Segher Boessenkool wrote:
>>> Deleting restores (before a noreturn) that are dead confuses dwarf2cfi.
>>>
>>> 2016-06-07  Segher Boessenkool  <segher@kernel.crashing.org>
>>>
>>> 	* dce.c (delete_unmarked_insns): Don't delete instructions with
>>> 	a REG_CFA_RESTORE note.
>> I don't really understand this one.  Why is the restore marked dead and
>> why doesn't that happen for normal epilogues?  Something wonky seems to
>> be going on here.
>
> Because it's not behind a NOTE_INSN_EPILOGUE_BEGIN it is not treated
> specially by DCE, like insns in "normal" epilogues are.
Ohhhhh!  While I don't see the special handling of things after 
NOTE_INSN_EPILOGUE_BEG, I do see this in df-scan.c:

   if (targetm.have_epilogue () && epilogue_completed)
     {
       /* Mark all call-saved registers that we actually used.  */
       for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
         if (df_regs_ever_live_p (i) && !LOCAL_REGNO (i)
             && !TEST_HARD_REG_BIT (regs_invalidated_by_call, i))
           bitmap_set_bit (exit_block_uses, i);
     }

Which then presumably gets used elsewhere when solving problems.

I wonder if we need some kind of similar mechanism to make those 
registers appear to be live at the component epilogues.

Jeff
diff mbox

Patch

diff --git a/gcc/dce.c b/gcc/dce.c
index ea3fb00..d510287 100644
--- a/gcc/dce.c
+++ b/gcc/dce.c
@@ -587,6 +587,15 @@  delete_unmarked_insns (void)
 	  if (!dbg_cnt (dce))
 	    continue;
 
+	  if (crtl->shrink_wrapped_separate
+	      && find_reg_note (insn, REG_CFA_RESTORE, NULL))
+	    {
+	      if (dump_file)
+		fprintf (dump_file, "DCE: NOT deleting insn %d, it's a "
+				    "callee-save restore\n", INSN_UID (insn));
+	      continue;
+	    }
+
 	  if (dump_file)
 	    fprintf (dump_file, "DCE: Deleting insn %d\n", INSN_UID (insn));