[llvm] [InstCombine] Improve shamt range calculation (PR #72535)

[Nikita Popov via llvm-commits]

================
@@ -962,16 +963,13 @@ static bool setShiftFlags(BinaryOperator &I, const SimplifyQuery &Q) {
   }
 
   // Compute what we know about shift count.
-  KnownBits KnownCnt =
-      computeKnownBits(I.getOperand(1), Q.DL, /*Depth*/ 0, Q.AC, Q.CxtI, Q.DT);
-  // If we know nothing about shift count or its a poison shift, we won't be
-  // able to prove anything so return before computing shift amount.
-  if (KnownCnt.isUnknown())
-    return false;
+  ConstantRange KnownCnt = computeConstantRangeIncludingKnownBits(
+      I.getOperand(1), /* ForSigned */ false, Q);
   unsigned BitWidth = KnownCnt.getBitWidth();
-  APInt MaxCnt = KnownCnt.getMaxValue();
-  if (MaxCnt.uge(BitWidth))
-    return false;
+  // Since shift produces a poison value if RHS is equal to or larger than the
+  // bit width, we can safely assume that RHS is less than the bit width.
+  APInt MaxCnt(BitWidth, BitWidth - 1);
+  MaxCnt = APIntOps::umin(MaxCnt, KnownCnt.getUnsignedMax());
----------------
nikic wrote:

```suggestion
  uint64_t MaxCnt = KnownCnt.getUnsignedMax().getLimitedValue(BitWidth - 1);
```
I think in this case it's easier to switch to working on plain integer?

https://github.com/llvm/llvm-project/pull/72535