I'm curious why __builtin_return_address() doesn't supports other arguments than 0 in ARM ?
It's a problem that somehow you can't deduce calling function address from the stack of ARM ?
Or something else ?
Thanks
Add a comment
|
2 Answers
According to this post <http://codingrelic.geekhold.com/2009/05/pre-mortem-backtracing.html>,
Also on some architectures, including my beloved MIPS, only
__builtin_return_address(0)works. MIPS has no frame pointer, making it difficult to walk back up the stack. Frame 0 can use the return address register directly. If ARM also does not have a frame pointer, this would explain the limitation.
See also http://gcc.gnu.org/onlinedocs/gcc/Return-Address.html.
3 Comments
Simon Richter
On the ARM, the return address is passed in register
R14, and it is the duty of the callee to save it when calling another function. So even with a frame pointer, there is no guarantee that the return address is ever stored on the stack.
R.. GitHub STOP HELPING ICE
Indeed, when the return address is callee-saved rather than saved on the stack by the call instruction, it's impossible in general to find it. There should be a way using dwarf2 unwind/debugging data, but that would require
__builtin_return_address to be a call to a heavy weight unwind library call rather than a trivial builtin...
Agnius Vasiliauskas
BTW, I solved this stack tracing problem on ARM by using -finstrument-functions which are called on every function entry/exit. There is overhead of course, but it is acceptable for me. (And there is
no_instrument_function attribute where max speed of call is required...)Backtrace on ARM is hard. The Glibc backtrace function does work these days, but you need an up to date compiler/glibc, and you need to have built everything with -funwind-tables. GDB also has trouble without debug info.
1 Comment
jfritz42
Thanks for mentioning -funwind-tables! My backtraces on ARM were always depth 1 until I enabled this compiler flag. Using GCC 4.3.2.