[PATCH] D42818: [InstCombine][ValueTracking] Match non-uniform constant power-of-two vectors

[Sanjay Patel via Phabricator via llvm-commits]

spatel added inline comments.


================
Comment at: include/llvm/IR/PatternMatch.h:296
           return this->isValue(CI->getValue());
+      if (const auto *CDV = dyn_cast<ConstantDataVector>(V)) {
+        unsigned E = CDV->getNumElements();
----------------
Is there a reason to limit the matcher to ConstantDataVector? If not, we should generalize to:

```
        // Non-splat vector constant: check each element for a match.
        unsigned NumElts = V->getType()->getVectorNumElements();
        assert(NumElts != 0 && "Constant vector with no elements?");
        for (unsigned i = 0; i != NumElts; ++i) {
          Constant *Elt = C->getAggregateElement(i);
          if (!Elt)
            return false;
          if (isa<UndefValue>(Elt))
            continue;
          auto *CI = dyn_cast<ConstantInt>(Elt);
          if (!CI || !this->isValue(CI->getValue()))
            return false;
        }
        return true;

```

The difference is that this matcher can handle weird types (and undef elts):

```
define <3 x i30> @test_weird_vec_type_nonsplat_pow2(<3 x i30> %a0) {
  %1 = urem <3 x i30> %a0, <i30 1, i30 2, i30 8>
  ret <3 x i30> %1
}
```

Side note: I copied that loop from a chunk of InstCombineCompares.cpp, and it's not the first time. We should make that available from one place, so we don't keep copying code.


Repository:
  rL LLVM

https://reviews.llvm.org/D42818