Re: [Qemu-devel] [PATCH] target-arm: Move VLD/VST multiple into helper functions

[Aurelien Jarno <aurelien@aurel32.net>]

On Mon, May 02, 2011 at 05:01:24PM +0100, Peter Maydell wrote:
> On 20 April 2011 15:52, Peter Maydell <peter.maydell@linaro.org> wrote:
> > Move VLD/VST multiple into helper functions, as some cases can
> > generate more TCG ops than the maximum per-instruction limit
> > and certainly more than the recommended 20.
> 
> I've had a review comment in private email that this patch
> slows down ffmpeg by a little over 4% (in a linux-user execution
> run where about 5% of instructions were an affected vld/vst). So
> perhaps the right approach is to increase the max-ops-per-insn limit
> and update the guidance in tcg/README instead?
> 

Does this patch fixes a real issue (ie most probably a crash), or it is
just to make the arm target compliant with the README?

Two remarks there:
- The guidance in tcg/README is probably true for complex helpers in the
  sense of doing a lot of logic/arithmetic operations for which the TCG
  instruction set if not really rich and for which the compiler can do
  a lot better. When the instruction is emulated mostly by load/store 
  ops, there isn't much possible optimizations left.
- Your patch calls the slow version of the _mmu functions, which doesn't
  use the QEMU TLB. It is therefore normal to expect a slow down.
  Unfortunately the fast version of _mmu functions are not really usable
  from an helper, or at least they need a lot of code to convert an mmu
  index to the right function. It is something I have on my TODO list
  for some months, haven't found time to look at it.

-- 
Aurelien Jarno                          GPG: 1024D/F1BCDB73
aurelien@aurel32.net                 http://www.aurel32.net