Re: [patch/idea] Add register scrambling to testsuite

[Matheus Castanho <msc@linux.ibm.com>]

Hi DJ,

DJ Delorie via Libc-alpha <libc-alpha@sourceware.org> writes:

> Florian Weimer <fweimer@redhat.com> writes:
>> This funcationality is likely more easily written as an assembler
>> routine that expects a function pointer and closure argument, performs
>> the register clobbers, and then calls the function with the closure
>> argument.
>
> I thought of putting the scramble just before the do_test calls, but
> there were more than one of those, and for the cases I'm trying to
> avoid, it doesn't matter when the regs are clobbered.
>

I was actually working on a patch similar to this one, but looks like
you beat me to it =). And you version seems cleaner.

You could embed your call to support_scramble_registers() in the CALL
macro on test-string.h:

#define CALL(impl, ...) \
  ({ support_scramble_registers (); (* (proto_t) (impl)->fn) (__VA_ARGS__); })

This way it would apply for all optimized function calls without
requiring many more changes. It also wouldn't affect the string
benchtests, since they use a separate CALL definition from
bench-string.h

> Calling scramble once is enough to more accurately simulate a "busy app"
> where gcc would have eventually filled all the registers with something,
> which is what happened in the bug I referenced.  A closure-based asm
> routine would make the work much more complicated, for no real benefit.
>
> I've seen testsuites where each call-under-test was wrapped in a routine
> that called it multiple times, permuting the "untouched" registers each
> time.  I don't think we need to go that far.

--
Matheus Castanho