[Lldb-commits] [PATCH] Profile Assembly Until Ret Instruction
endlessroad at google.com
Wed Jul 30 14:52:54 PDT 2014
Thanks Jason! That's a very informative post, clarify things a lot :-)
Well I have to admit that my patch is specifically for certain kind of
functions, and now I see that's not the general case.
I did some experiment with gdb. gdb uses CFI for frame 0, either x86 or
x86_64. It looks for FDE of frame 0, and do CFA calculations according to
- For compiler generated functions: I think there are 2 usage scenarios for
frame 0: breakpoint and signal.
- Breakpoints are usually at source line boundary instead of
instruction boundary, and generally we won't be caught at stack pointer
changing locations, so CFI is still valid.
- For signal, synchronous unwind table may not be sufficient here. But
only stack changing instructions will cause incorrect CFA calculation, so
it' not always the case.
- For hand written assembly functions: from what I've seen, most of the
time CFI is present and actually asynchronous.
So it seems that in most cases, even with only synchronous unwind table,
CFI is still correct.
I believe we can trust eh_frame for frame 0 and use assembly profiling as
fallback. If both failed, maybe code owner should use
On Tue, Jul 29, 2014 at 4:59 PM, Jason Molenda <jmolenda at apple.com> wrote:
> It was a tricky one and got lost in the shuffle of a busy week. I was
> always reluctant to try profiling all the instructions in a function. On
> x86, compiler generated code (gcc/clang anyway) is very simplistic about
> setting up the stack frame at the start and only having one epilogue - so
> anything fancier risked making mistakes and could possibly have a
> performance impact as we run functions through the disassembler.
> For hand-written assembly functions (which can be very creative with their
> prologue/epilogue and where it is placed), my position is that they should
> write eh_frame instructions in their assembly source to tell lldb where to
> find things. There is one or two libraries on Mac OS X where we break the
> "ignore eh_frame for the currently executing function" because there are
> many hand-written assembly functions in there and the eh_frame is going to
> beat our own analysis.
> After I wrote the x86 unwinder, Greg and Caroline implemented the arm
> unwinder where it emulates every instruction in the function looking for
> prologue/epilogue instructions. We haven't seen it having a particularly
> bad impact performance-wise (lldb only does this disassembly for functions
> that it finds on stacks during an execution run, and it saves the result so
> it won't re-compute it for a given function). The clang armv7 codegen
> often has mid-function epilogues (early returns) which definitely
> complicated things and made it necessary to step through the entire
> function bodies. There's a bunch of code I added to support these
> mid-function epilogues - I have to save the register save state when I see
> an instruction which looks like an epilogue, and when I see the final ret
> instruction (aka restoring the saved lr contents into pc), I re-install the
> register save state from before the epilogue started.
> These things always make me a little nervous because the instruction
> analyzer obviously is doing a static analysis so it knows nothing about
> flow control. Tong's patch stops when it sees the first CALL instruction -
> but that's not right, that's just solving the problem for his particular
> function which doesn't have any CALL instructions before his prologue. :)
> You could imagine a function which saves a couple of registers, calls
> another function, then saves a couple more because it needs more scratch
> If we're going to change to profiling deep into the function -- and I'm
> not opposed to doing that, it's been fine on arm -- we should just do the
> entire function I think.
> Another alternative would be to trust eh_frame on x86_64 at frame 0. This
> is one of those things where there's not a great solution. The unwind
> instructions in eh_frame are only guaranteed to be accurate for synchronous
> unwinds -- that is, they are only guaranteed to be accurate at places where
> an exception could be thrown - at call sites. So for instances, there's no
> reason why the compiler has to describe the function prologue instructions
> at all. There's no requirement that the eh_frame instructions describe the
> epilogue instructions. The information about spilled registers only needs
> to be emitted where we could throw an exception, or where a callee could
> throw an exception.
> clang/gcc both emit detailed instructions for the prologue setup. But for
> i386 codegen if the compiler needs to access some pc-relative data, it will
> do a "call next-instruction; pop %eax" to get the current pc value.
> (x86_64 has rip-relative addressing so this isn't needed) If you're
> debugging -fomit-frame-pointer code, that means your CFA is expressed in
> terms of the stack pointer and the stack pointer just changed mid-function
> --- and eh_frame instructions don't describe this.
> The end result: If you want accurate unwinds 100% of the time, you can't
> rely on the unwind instructions from eh_frame. But they'll get you
> accurate unwinds 99.9% of the time ... also, last I checked, neither clang
> nor gcc describe the epilogue instructions.
> In *theory* the unwind instructions from the DWARF debug_frame section
> should be asynchronous -- they should describe how to find the CFA address
> for every instruction in the function. Which makes sense - you want
> eh_frame to be compact because it's bundled into the executable, so it
> should only have the information necessary for exception handling and you
> can put the verbose stuff in debug_frame DWARF for debuggers. But instead
> (again, last time I checked), the compilers put the exact same thing in
> debug_frame even if you use the -fasynchronous-unwind-tables (or whatever
> that switch was) option.
> So I don't know, maybe we should just start trusting eh_frame at frame 0
> and write off those .1% cases where it isn't correct instead of trying to
> get too fancy with the assembly analysis code.
> > On Jul 29, 2014, at 4:17 PM, Todd Fiala <tfiala at google.com> wrote:
> > Hey Jason,
> > Do you have any feedback on this?
> > Thanks!
> > -Todd
> > On Fri, Jul 25, 2014 at 1:42 PM, Tong Shen <endlessroad at google.com>
> > Sorry, wrong version of patch...
> > On Fri, Jul 25, 2014 at 1:41 PM, Tong Shen <endlessroad at google.com>
> > Hi Molenda, lldb-commits,
> > For now, x86 assembly profiler will stop after 10 "non-prologue"
> instructions. In practice it may not be sufficient. For example, we have a
> hand-written assembly function, which have hundreds of instruction before
> actual (stack-adjusting) prologue instructions.
> > One way is to change the limit to 1000; but there will always be
> functions that break the limit :-) I believe the right thing to do here is
> parsing all instructions before "ret"/"call" as prologue instructions.
> > Here's what I changed:
> > - For "push %rbx" and "mov %rbx, -8(%rbp)": only add first row for that
> register. They may appear multiple times in function body. But as long as
> one of them appears, first appearance should be in prologue(If it's not in
> prologue, this function will not use %rbx, so these 2 instructions should
> not appear at all).
> > - Also monitor "add %rsp 0x20".
> > - Remove non prologue instruction count.
> > - Add "call" instruction detection, and stop parsing after it.
> > Thanks.
> > --
> > Best Regards, Tong Shen
> > --
> > Best Regards, Tong Shen
> > _______________________________________________
> > lldb-commits mailing list
> > lldb-commits at cs.uiuc.edu
> > http://lists.cs.uiuc.edu/mailman/listinfo/lldb-commits
> > --
> > Todd Fiala | Software Engineer | tfiala at google.com |
Best Regards, Tong Shen
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the lldb-commits