[PATCH] x86: Add support for frame pointer less mcount

[Andi Kleen <andi at firstfloor dot org>]

Traditionally -pg profiling required enabling the frame pointer in the compiler,
because the mcount entry point is defined using the frame pointer.

Recent gcc has new -mfentry option that uses an alternative entry point and
does not require frame pointer. The new entry point is called __fentry__

This has advantages on systems where frame pointer is relatively expensive,
like Atom CPUs.

This patch adds __fentry__ support to x86 glibc. It just copies the existing
mcount entry points and adapts them to be frame pointer less.

I added a new GLIBC 2.13 namespace for this, because it's a new symbol.

-Andi

* 2010-07-17    Andi Kleen <ak@linux.intel.com>
    
    	* sysdeps/i386/i386-mcount.S (__fentry__): Add.
    	* sysdeps/x86_64/_mcount.S (__fentry__): Add.
    	* stdlib/Versions (__fentry__): Add for GLIBC 2.13
    	* Versions.def (GLIBC_2.13): Add.

diff --git a/Versions.def b/Versions.def
index eab006b..0ccda50 100644
--- a/Versions.def
+++ b/Versions.def
@@ -29,6 +29,7 @@ libc {
   GLIBC_2.10
   GLIBC_2.11
   GLIBC_2.12
+  GLIBC_2.13
 %ifdef USE_IN_LIBIO
   HURD_CTHREADS_0.3
 %endif
diff --git a/stdlib/Versions b/stdlib/Versions
index 3e7b892..2aa396e 100644
--- a/stdlib/Versions
+++ b/stdlib/Versions
@@ -100,6 +100,9 @@ libc {
   GLIBC_2.10 {
     quick_exit; __cxa_at_quick_exit;
   }
+  GLIBC_2.13 {
+    __fentry__;
+  }
   GLIBC_PRIVATE {
     # functions which have an additional interface since they are
     # are cancelable.
diff --git a/sysdeps/i386/i386-mcount.S b/sysdeps/i386/i386-mcount.S
index 8b11adb..ebb659f 100644
--- a/sysdeps/i386/i386-mcount.S
+++ b/sysdeps/i386/i386-mcount.S
@@ -53,3 +53,28 @@ C_LABEL(_mcount)
 
 #undef mcount
 weak_alias (_mcount, mcount)
+
+	/* Same as above, but doesn't require a frame pointer */
+	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(__fentry__)
+	ASM_TYPE_DIRECTIVE(C_SYMBOL_NAME(__fentry__), @function)
+	.align ALIGNARG(4)
+C_LABEL(__fentry__)
+	/* Save the caller-clobbered registers.  */
+	pushl %eax
+	pushl %ecx
+	pushl %edx
+
+	movl 12(%esp), %edx
+	movl 16(%esp), %eax
+
+	/* No need to access the PLT or GOT, __mcount_internal is an
+	   internal function and we can make a relative call.  */
+	call C_SYMBOL_NAME(__mcount_internal)
+
+	/* Pop the saved registers.  Please note that `mcount_nofp' has no
+	   return value.  */
+	popl %edx
+	popl %ecx
+	popl %eax
+	ret
+	ASM_SIZE_DIRECTIVE(C_SYMBOL_NAME(__fentry__))
diff --git a/sysdeps/x86_64/_mcount.S b/sysdeps/x86_64/_mcount.S
index c005932..1571a4d 100644
--- a/sysdeps/x86_64/_mcount.S
+++ b/sysdeps/x86_64/_mcount.S
@@ -65,3 +65,41 @@ C_LABEL(_mcount)
 
 #undef mcount
 weak_alias (_mcount, mcount)
+
+	ASM_GLOBAL_DIRECTIVE C_SYMBOL_NAME(__fentry__)
+	ASM_TYPE_DIRECTIVE(C_SYMBOL_NAME(__fentry__), @function)
+	.align ALIGNARG(4)
+C_LABEL(__fentry__)
+	/* Allocate space for 7 registers.  */
+	subq	$56,%rsp
+	movq	%rax,(%rsp)
+	movq	%rcx,8(%rsp)
+	movq	%rdx,16(%rsp)
+	movq	%rsi,24(%rsp)
+	movq	%rdi,32(%rsp)
+	movq	%r8,40(%rsp)
+	movq	%r9,48(%rsp)
+
+	/* Setup parameter for __mcount_internal.  */
+	/* selfpc is the return address on the stack.  */
+	movq	56(%rsp),%rsi
+	/* caller is the return address above it */
+	movq	64(%rsp),%rdi
+#ifdef PIC
+	call C_SYMBOL_NAME(__mcount_internal)@PLT
+#else
+	call C_SYMBOL_NAME(__mcount_internal)
+#endif
+	/* Pop the saved registers.  Please note that `__fentry__' has no
+	   return value.  */
+	movq	48(%rsp),%r9
+	movq	40(%rsp),%r8
+	movq	32(%rsp),%rdi
+	movq	24(%rsp),%rsi
+	movq	16(%rsp),%rdx
+	movq	8(%rsp),%rcx
+	movq	(%rsp),%rax
+	addq	$56,%rsp
+	ret
+
+	ASM_SIZE_DIRECTIVE(C_SYMBOL_NAME(__fentry__))



-- 
ak@linux.intel.com -- Speaking for myself only.