[PATCH 1/3] perf tests: Introduce perf_regs_load function on ARM64

Tue Apr 22 06:42:00 PDT 2014

Hi Jean,

Apologies for the delay on this.

On Tue, Mar 25, 2014 at 03:23:26PM +0000, Jean Pihet wrote:
> Hi Mark,
> 
> On 21 March 2014 16:11, Mark Rutland <mark.rutland at arm.com> wrote:
> > Hi Jean,
> >
> > On Wed, Mar 19, 2014 at 09:42:33AM +0000, Jean Pihet wrote:
> >> Introducing perf_regs_load function, which is going
> >> to be used for dwarf unwind test in following patches.
> >>
> >> It takes single argument as a pointer to the regs dump
> >> buffer and populates it with current registers values, as
> >> expected by the perf built-in unwinding test.
> >>
> >> Signed-off-by: Jean Pihet <jean.pihet at linaro.org>
> >> Cc: Steve Capper <steve.capper at linaro.org>
> >> Cc: Corey Ashford <cjashfor at linux.vnet.ibm.com>
> >> Cc: Frederic Weisbecker <fweisbec at gmail.com>
> >> Cc: Ingo Molnar <mingo at kernel.org>
> >> Cc: Namhyung Kim <namhyung at kernel.org>
> >> Cc: Paul Mackerras <paulus at samba.org>
> >> Cc: Peter Zijlstra <a.p.zijlstra at chello.nl>
> >> Cc: Arnaldo Carvalho de Melo <acme at ghostprotocols.net>
> >> Cc: David Ahern <dsahern at gmail.com>
> >> Cc: Jiri Olsa <jolsa at redhat.com>
> >> ---
> >>  tools/perf/arch/arm64/Makefile            |  1 +
> >>  tools/perf/arch/arm64/include/perf_regs.h |  2 ++
> >>  tools/perf/arch/arm64/tests/regs_load.S   | 39 +++++++++++++++++++++++++++++++
> >>  3 files changed, 42 insertions(+)
> >>  create mode 100644 tools/perf/arch/arm64/tests/regs_load.S
> >>
> >> diff --git a/tools/perf/arch/arm64/Makefile b/tools/perf/arch/arm64/Makefile
> >> index 67e9b3d..9b8f87e 100644
> >> --- a/tools/perf/arch/arm64/Makefile
> >> +++ b/tools/perf/arch/arm64/Makefile
> >> @@ -4,4 +4,5 @@ LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/dwarf-regs.o
> >>  endif
> >>  ifndef NO_LIBUNWIND
> >>  LIB_OBJS += $(OUTPUT)arch/$(ARCH)/util/unwind-libunwind.o
> >> +LIB_OBJS += $(OUTPUT)arch/$(ARCH)/tests/regs_load.o
> >>  endif
> >> diff --git a/tools/perf/arch/arm64/include/perf_regs.h b/tools/perf/arch/arm64/include/perf_regs.h
> >> index 2359546..1e052f1 100644
> >> --- a/tools/perf/arch/arm64/include/perf_regs.h
> >> +++ b/tools/perf/arch/arm64/include/perf_regs.h
> >> @@ -9,6 +9,8 @@
> >>  #define PERF_REG_IP  PERF_REG_ARM64_PC
> >>  #define PERF_REG_SP  PERF_REG_ARM64_SP
> >>
> >> +void perf_regs_load(u64 *regs);
> >> +
> >>  static inline const char *perf_reg_name(int id)
> >>  {
> >>       switch (id) {
> >> diff --git a/tools/perf/arch/arm64/tests/regs_load.S b/tools/perf/arch/arm64/tests/regs_load.S
> >> new file mode 100644
> >> index 0000000..92ab968
> >> --- /dev/null
> >> +++ b/tools/perf/arch/arm64/tests/regs_load.S
> >> @@ -0,0 +1,39 @@
> >> +#include <linux/linkage.h>
> >> +
> >> +/*
> >> + * Implementation of void perf_regs_load(u64 *regs);
> >> + *
> >> + * This functions fills in the 'regs' buffer from the actual registers values,
> >> + * in the way the perf built-in unwinding test expects them:
> >> + * - the PC at the time at the call to this function. Since this function
> >> + *   is called using a bl instruction, the PC value is taken from LR,
> >
> > Is it guaranteed that this function is always invoked with a branch with
> > link instruction, or is that just what current compiler versions are
> > doing? I couldn't see where we would get that guarantee from.
> The current compiler implements the call as a bl instruction.

While I don't think we can rely on the compiler using a bl to call the
function it shouldn't matter here if we only care about the LR value
being an address within the caller, as it doesn't look amenable to tail
call optimization.

> > If it is called with a branch with link, then the LR value will be the
> > PC at call time + 4, rather than just the exact PC at call time. If not
> > then we don't have a guaranteed relationship between the PC at call time
> > and the current LR value.
> >
> > If the only place that perf_regs_load is used is a single test which
> > doesn't care about the precise PC at the time of the call, then it's
> > probably OK to use the LR value, but we should be careful to document
> > what the faked-up PC actually is and how we expect it to be used.
> The code is only used by an unwinding test. The unwinding code
> resolves the function name from an address range found in the dwarf
> information so in principle it is ok to use the PC/LR at the time of
> the call to a function.
> 
> Is the comment above OK or do you want an update of the code as well?

If we just need an (arbitrary) address within the caller, a comment
update should be fine.

> >> + * - the current SP (not touched by this function),
> >> + * - the current value of LR is merely retrieved and stored because the
> >> + *   value before the call to this function is unknown at this time; it will
> >> + *   be unwound from the dwarf information in unwind__get_entries.
> >> + */
> >> +
> >> +.text
> >> +.type perf_regs_load,%function
> >> +ENTRY(perf_regs_load)
> >> +     stp x0,  x1,  [x0], #16 // store x0..x29
> >> +     stp x2,  x3,  [x0], #16
> >> +     stp x4,  x5,  [x0], #16
> >> +     stp x6,  x7,  [x0], #16
> >> +     stp x8,  x9,  [x0], #16
> >> +     stp x10, x11, [x0], #16
> >> +     stp x12, x13, [x0], #16
> >> +     stp x14, x15, [x0], #16
> >> +     stp x16, x17, [x0], #16
> >> +     stp x18, x19, [x0], #16
> >> +     stp x20, x21, [x0], #16
> >> +     stp x22, x23, [x0], #16
> >> +     stp x24, x25, [x0], #16
> >> +     stp x26, x27, [x0], #16
> >> +     stp x28, x29, [x0], #16
> >> +     mov x1,  sp
> >> +     stp x30, x1,  [x0], #16 // store lr and sp
> >> +     str x30, [x0]           // store pc as lr in order to skip the call
> >> +                             //  to this function
> >
> > It might be better to word this a "store the lr in place of the pc". To
> > me at least the current wording implies the opposite of what the code
> > seems to be doing.
> Ok the last comment can be updated.

Ok, cheers.

With those changes I think this looks fine.

Thanks,
Mark.