[Mesa-dev] [Mesa-stable] [PATCH] i965: Fix JIP to properly skip over unrelated control flow.

[Kenneth Graunke]

On Friday, November 20, 2015 05:36:35 PM Francisco Jerez wrote:
> Kenneth Graunke <kenneth at whitecape.org> writes:
> 
> > On Friday, November 20, 2015 02:38:10 PM Francisco Jerez wrote:
> >> Kenneth Graunke <kenneth at whitecape.org> writes:
> >> 
> >> > On Thursday, November 19, 2015 02:05:44 PM Kenneth Graunke wrote:
> >> >> We've apparently always been botching JIP for sequences such as:
> >> >> 
> >> >>    do
> >> >>        cmp.f0.0 ...
> >> >>        (+f0.0) break
> >> >>        ...
> >> >>        if
> >> >>           ...
> >> >>        else
> >> >>           ...
> >> >>        endif
> >> >>        ...
> >> >>    while
> >> >> 
> >> >> Normally, UIP is supposed to point to the final destination of the jump,
> >> >> while in nested control flow, JIP is supposed to point to the end of the
> >> >> current nesting level.  It essentially bounces out of the current nested
> >> >> control flow, to an instruction that has a JIP which bounces out another
> >> >> level, and so on.
> >> >> 
> >> >> In the above example, when setting JIP for the BREAK, we call
> >> >> brw_find_next_block_end(), which begins a search after the BREAK for the
> >> >> next ENDIF, ELSE, WHILE, or HALT.  It ignores the IF and finds the ELSE,
> >> >> setting JIP there.
> >> >> 
> >> >> This makes no sense at all.  The break is supposed to skip over the
> >> >> whole if/else/endif block entirely.  They have a sibling relationship,
> >> >> not a nesting relationship.
> >> >> 
> >> >> This patch fixes brw_find_next_block_end() to track depth as it does
> >> >> its search, and ignore anything not at depth 0.  So when it sees the
> >> >> IF, it ignores everything until after the ENDIF.  That way, it finds
> >> >> the end of the right block.
> >> >> 
> >> >> Caught while debugging a tessellation shader - no apparent effect on
> >> >> Piglit.  I did look for actual applications that were affected, and
> >> >> found that GLBenchmark Manhattan had a BREAK with a bogus JIP.
> >> >> 
> >> >> Cc: mesa-stable at lists.freedesktop.org
> >> >> Signed-off-by: Kenneth Graunke <kenneth at whitecape.org>
> >> >
> >> > I tried pretty hard to produce a Piglit test that showed an actual
> >> > problem from doing this wrong - and I wasn't able to.
> >> >
> >> > It seems it just steps through some extra instructions which do
> >> > nothing, and is pretty harmless.
> >> >
> >> From my understanding of how control flow is implemented, jumping to the
> >> ENDIF instruction of an inactive IF-ENDIF construct (or similarly to the
> >> WHILE instruction of an inactive loop) is fully equivalent to jumping to
> >> the same point of an active (i.e. properly nested) but non-diverging
> >> IF-ENDIF construct, and will behave the same: It will have no effect on
> >> the current per-channel enables (because the IP of the ENDIF instruction
> >> won't match the UIP value present at the top of the stack), and for that
> >> reason will go on and jump to the instruction pointed to by the JIP
> >> value of the ENDIF, which will be another ENDIF/WHILE/HALT instruction
> >> closer to the right ENDIF/WHILE instruction that closes the current
> >> block.
> >> 
> >> > So I don't think this should actually go to stable after all.
> >> >
> >> Yeah, seems pretty harmless -- If it weren't harmeless you'd also need
> >> to apply a similar fix to WHILE loops, but I don't think you do.
> >
> > Thanks, Curro!  I appreciate you confirming the theory :)
> >
> > I haven't pushed the patch yet, so would you like to add a Reviewed-by?
> >
> > What WHILE fix are you thinking of?  We may as well get it right, even
> > if it is harmless.  At a cursory glance, I didn't see anything wrong,
> > as it uses the loop stack.  But I might've missed something...
> 
> Hmm, does it?  AFAICT brw_find_next_block_end() will just give you the
> first WHILE instruction it finds even if the corresponding loop starts
> after the given offset.  Not that it matters much anyway.

Ahh.  I thought you meant the code for patching WHILE's JIP was
similarly broken.  But you're right, I think as is,


DO
   CMP.f0.0 ...
   (+f0.0) BREAK

   ...

   DO
      ...
   WHILE

   ...
WHILE

will make the BREAK point to the first WHILE instead of the second.
The obvious fix would be to handle DO just like I handle IF.
Unfortunately, there is no DO instruction...and at this point, the
loop stack is gone.

I suppose one fix would be to make brw_find_next_block_end() only
accept a WHILE as the end of the block if the WHILE's jump target
is before our starting offset.  In other words, if the WHILE's
top-of-the-loop is after our BREAK/CONTINUE, it is obviously the
wrong loop.  (This works because WHILE's jumps are set as soon as
we emit the WHILE, so they're already in place by brw_set_uip_jip.)

I think that would work, but it would at least need a comment.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: This is a digitally signed message part.
URL: <http://lists.freedesktop.org/archives/mesa-dev/attachments/20151120/9ae06a85/attachment.sig>