[PATCH 02/10] ARM: ftrace: document mcount formats

Rabin Vincent rabin at rab.in
Sat Feb 13 14:48:30 EST 2010


Add a comment describing the mcount variants and how the callsites look
like.

Signed-off-by: Rabin Vincent <rabin at rab.in>
---
 arch/arm/kernel/entry-common.S |   32 ++++++++++++++++++++++++++++++++
 1 files changed, 32 insertions(+), 0 deletions(-)

diff --git a/arch/arm/kernel/entry-common.S b/arch/arm/kernel/entry-common.S
index 0b042bd..d412d7c 100644
--- a/arch/arm/kernel/entry-common.S
+++ b/arch/arm/kernel/entry-common.S
@@ -92,6 +92,38 @@ ENDPROC(ret_from_fork)
 #define CALL(x) .long x
 
 #ifdef CONFIG_FUNCTION_TRACER
+/*
+ * When compiling with -pg, gcc inserts a call to the mcount routine at the
+ * start of every function.  In mcount, apart from the function's address (in
+ * lr), we need to get hold of the function's caller's address.
+ *
+ * Older GCCs (pre-4.4) inserted a call to a routine called mcount like this:
+ *
+ *	bl	mcount
+ *
+ * These versions have the limitation that in order for the mcount routine to
+ * be able to determine the function's caller's address, an APCS-style frame
+ * pointer (which is set up with something like the code below) is required.
+ *
+ *	mov     ip, sp
+ *	push    {fp, ip, lr, pc}
+ *	sub     fp, ip, #4
+ *
+ * With EABI, these frame pointers are not available unless -mapcs-frame is
+ * specified, and if building as Thumb-2, not even then.
+ *
+ * Newer GCCs (4.4+) solve this problem by introducing a new version of mcount,
+ * with call sites like:
+ *
+ *	push	{lr}
+ *	bl	__gnu_mcount_nc
+ *
+ * With these compilers, frame pointers are not necessary.
+ *
+ * With both the mcount types, we need to restore the original lr before
+ * returning.  In the __gnu_mcount_nc, version we're allowed to clobber ip.
+ * No other registers should be clobbered.
+ */
 #ifdef CONFIG_DYNAMIC_FTRACE
 ENTRY(mcount)
 	stmdb	sp!, {r0-r3, lr}
-- 
1.6.6




More information about the linux-arm-kernel mailing list