From patchwork Thu Aug 19 09:33:32 2010
Content-Type: text/plain; charset="utf-8"
MIME-Version: 1.0
Content-Transfer-Encoding: 7bit
Subject: prevent extraneous nop past sibling call on sparc
Date: Wed, 18 Aug 2010 23:33:32 -0000
From: Olivier Hainque <hainque@adacore.com>
X-Patchwork-Id: 62125
Message-Id: <20100819093332.GA9139@cardhu.act-europe.fr>
To: gcc-patches@gcc.gnu.org
Cc: hainque@adacore.com

Hello,

The sparc back-end emits an extra nop past calls at the very end of a
function, to prevent the return address from falling within a different
function, which causes trouble to some backtrace computation engines.

Such a nop is pointless for a sibling call since the return address is
setup explicitly, and this patch is a suggestion to remove it for this
particular case. We found this useful in the course of the development
of a machine code coverage analyzer.

OK ?

Thanks in advance,

Olivier

2010-08-19  Olivier Hainque  <hainque@adacore.com>

	* config/sparc/sparc.c (sparc_asm_function_epilogue): Don't output
	an extra nop past a sibling call at the very end.

	testsuite/
	* gcc.target/sparc/sibcall-dslot.c: New testcase.


Index: config/sparc/sparc.c
===================================================================
--- config/sparc/sparc.c	(revision 163326)
+++ config/sparc/sparc.c	(working copy)
@@ -4527,11 +4527,11 @@
 static void
 sparc_asm_function_epilogue (FILE *file, HOST_WIDE_INT size ATTRIBUTE_UNUSED)
 {
-  /* If code does not drop into the epilogue, we have to still output
-     a dummy nop for the sake of sane backtraces.  Otherwise, if the
-     last two instructions of a function were "call foo; dslot;" this
-     can make the return PC of foo (i.e. address of call instruction
-     plus 8) point to the first instruction in the next function.  */
+  /* If the last two instructions of a function are "call foo; dslot;"
+     the return address might point to the first instruction in the next
+     function and we have to output a dummy nop for the sake of sane
+     backtraces in such cases.  This is pointless for sibling calls since
+     the return address is explicitly adjusted.  */
 
   rtx insn, last_real_insn;
 
@@ -4543,7 +4543,8 @@
       && GET_CODE (PATTERN (last_real_insn)) == SEQUENCE)
     last_real_insn = XVECEXP (PATTERN (last_real_insn), 0, 0);
 
-  if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN)
+  if (last_real_insn && GET_CODE (last_real_insn) == CALL_INSN
+      && !SIBLING_CALL_P (last_real_insn))
     fputs("\tnop\n", file);
 
   sparc_output_deferred_case_vectors ();
Index: testsuite/gcc.target/sparc/sibcall-dslot.c
===================================================================
--- testsuite/gcc.target/sparc/sibcall-dslot.c	(revision 0)
+++ testsuite/gcc.target/sparc/sibcall-dslot.c	(revision 0)
@@ -0,0 +1,11 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+
+extern int one ();
+
+int some_n ()
+{
+    return one ();
+}
+
+/* { dg-final { scan-assembler-not "nop" } } */