[llvm] [ConstantRange] Improve `shlWithNoWrap` (PR #101800)

[Nikita Popov via llvm-commits]

================
@@ -1624,12 +1624,44 @@ ConstantRange ConstantRange::shlWithNoWrap(const ConstantRange &Other,
     return getEmpty();
 
   ConstantRange Result = shl(Other);
+  if (!NoWrapKind)
+    return Result;
 
-  if (NoWrapKind & OverflowingBinaryOperator::NoSignedWrap)
-    Result = Result.intersectWith(sshl_sat(Other), RangeType);
+  if (NoWrapKind & OverflowingBinaryOperator::NoSignedWrap) {
+    std::optional<unsigned> ShAmtBound;
+    if (isAllNonNegative())
+      ShAmtBound = getSignedMin().countLeadingZeros();
+    else if (isAllNegative())
+      ShAmtBound = getSignedMax().countLeadingOnes();
+    ConstantRange ShAmtRange = Other;
+    if (ShAmtBound)
+      ShAmtRange = ShAmtRange.intersectWith(
+          ConstantRange(APInt(getBitWidth(), 0),
+                        APInt(getBitWidth(), *ShAmtBound)),
+          Unsigned);
+    Result = Result.intersectWith(sshl_sat(ShAmtRange), RangeType);
+  }
 
-  if (NoWrapKind & OverflowingBinaryOperator::NoUnsignedWrap)
-    Result = Result.intersectWith(ushl_sat(Other), RangeType);
+  if (NoWrapKind & OverflowingBinaryOperator::NoUnsignedWrap) {
----------------
nikic wrote:

I think you should be able to skip the shl() call now for the nuw case, as the implementation is optimal by itself.

I'm also wondering whether it's easy to make the nuw+nsw case optimal. We effectively have one less bit to work with in that case.

https://github.com/llvm/llvm-project/pull/101800