[PATCH] D130433: [InstCombine] Add fold for redundant sign bits count comparison

[Alexander Shaposhnikov via Phabricator via llvm-commits]

alexander-shaposhnikov added inline comments.


================
Comment at: llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp:1688-1689
+  unsigned BitWidth = XType->getScalarSizeInBits();
+  if (Shift == 0 || Shift > BitWidth || PowerOf2.logBase2() < 2 ||
+      PowerOf2.logBase2() > BitWidth - 2)
+    return nullptr;
----------------
spatel wrote:
> spatel wrote:
> > spatel wrote:
> > > I think this is still over-specified. It seems all we need is a non-zero shift and a power-of-2 compare constant:
> > > https://alive2.llvm.org/ce/z/adaQ_n
> > Sorry - I botched that code. We do have to guard against overflow when we increment the cmp constant:
> > https://alive2.llvm.org/ce/z/eEUfx3
> We don't need to guard against a poison shift (Shift > BitWidth), and the lower pow2 bound is not needed? 
> 
> Please add a test like this:
> ```
> define i1 @src(i8 %x) {
>   %y = ashr i8 %x, 5
>   %z = xor i8 %y, %x
>   %c = icmp ult i8 %z, 2
>   ret i1 %c
> }
> ```
> We should get that:
> https://alive2.llvm.org/ce/z/KemU48
> 
> Also, "u< 1" will be reduced to "== 0", so we probably want to canonicalize that pattern (and its inverse) too:
> https://alive2.llvm.org/ce/z/ou5byR
> ...but that should be another patch.
@spatel - you are right, I put in place those constraints because the earlier version of the alive proof required them.



Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D130433/new/

https://reviews.llvm.org/D130433