[llvm-branch-commits] [llvm] release/19.x: [SLP]Check that operand of abs does not overflow before making it part of minbitwidth transformation (PR #113146)

[Nikita Popov via llvm-branch-commits]

================
@@ -15440,9 +15440,25 @@ bool BoUpSLP::collectValuesToDemote(
                 MaskedValueIsZero(I->getOperand(1), Mask, SimplifyQuery(*DL)));
       });
     };
+    auto AbsChecker = [&](unsigned BitWidth, unsigned OrigBitWidth) {
+      assert(BitWidth <= OrigBitWidth && "Unexpected bitwidths!");
+      return all_of(E.Scalars, [&](Value *V) {
+        auto *I = cast<Instruction>(V);
+        unsigned SignBits = OrigBitWidth - BitWidth;
+        APInt Mask = APInt::getBitsSetFrom(OrigBitWidth, BitWidth - 1);
+        unsigned Op0SignBits =
+            ComputeNumSignBits(I->getOperand(0), *DL, 0, AC, nullptr, DT);
+        return SignBits <= Op0SignBits &&
+               ((SignBits != Op0SignBits &&
+                 !isKnownNonNegative(I->getOperand(0), SimplifyQuery(*DL))) ||
----------------
nikic wrote:

I don't understand this part of the condition. What is the meaning of `SignBits != Op0SignBits && !isKnownNonNegative`? Should this be `SignBits != Op0SignBits || isKnownNonNegative`?

Though I'm not really sure why we'd be interested in handling the case where the abs is known non-negative (including also the MaskedValueIsZero check below): If it's non-negative, wouldn't we expect the abs to fold away anyway?

https://github.com/llvm/llvm-project/pull/113146