[llvm-dev] RFC: Killing undef and spreading poison

[Friedman, Eli via llvm-dev]

On 11/8/2016 3:32 PM, Sanjoy Das wrote:
> Hi Nuno, Chandler,
>
> Nuno Lopes via llvm-dev wrote:
> > This program stores 8 bits, and leaves the remaining 24 bits
> > uninitialized. It then loads 16 bits, half initialized to %v, half
> > uninitialized. SROA transforms the above function to:
> >
> > define i16 @g(i8 %in) {
> >    %v = add nsw i8 127, %in
> >    %1 = zext i8 %v to i16
> >    %2 = shl i16 %1, 8
> >    %3 = and i16 undef, 255
> >    %4 = or i16 %3, %2
> >    ret i16 %4
> > }
>
> This program above returns i16 poison only if "shl i16 poison, 8" is a
> full value poison.  Whether that is the case today or not is anybody's
> guess (as Eli says, we should not rely too much on semantics that are
> known to be broken), so the important question is whether "shl i16
> poison, 8" _can_ be defined to be poison only on the higher 8 bits in
> a world with bitwise poison.  If we could make that happen, we'd also
> get some other fringe benefits, like ComputeKnownBits would be able to
> fearlessly return "the low 8 bits of `shl i16 %val, 8` are 0" as
> opposed to "the low 8 bits of `shl i16 %val, 8` are 0 or poison".
>
> One problem with shl always generating non-poison zeroes in the lower
> bits is that it then is not equivalent to multiplication.  This breaks
> optimizations like:
>
>   %t = ...
>   %q = shl i32 %t, 6
>
> =/=>
>
>   %t = ...
>   %q = mul i32 %t, 64
>
> and (probably more meaningful)
>
>   %t = ...
>   %q0 = shl i32 %t, 6
>   %q1 = mul i32 %q0, 3
>
> =/=>
>
>   %t = ...
>   ;; %q0 = shl i32 %t, 6
>   %q1 = mul i32 %q0, (3 * 64)
>
>
> and (probably even worse) it also breaks reassociation:
>
>   %t = ...
>   %q0 = shl i32 %t, %P
>   %q1 = mul i32 %q0, %Q
>
> (No other uses of %q0, but there are uses of %q1)
>
> =/=>
>
>   %t = ...
>   %q0 = mul i32 %t, %Q
>   %q1 = shl i32 %q1, %P
>
>
> I'm especially worried about this last re-association example since
> SCEV does things like this internally when canonicalizing expressions,
> and I'm not convinced that we can change it to tiptoe around these
> problematic cases.

Well, we could say non-nsw add and mul are actually "bitwise" 
operations, so "mul i32 poison, 2" is guaranteed to have its bottom bit 
zero (but "mul nsw i32 poison, 2" is poison).  Of course, there's a 
limit to how far we can go in that direction, or we just end up with the 
old notion of undef.  Off the top of my head, I'm not exactly sure where 
that line is.

-Eli

-- 
Employee of Qualcomm Innovation Center, Inc.
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project