[llvm] [InstCombine] Rotate transformation port from SelectionDAG to InstCombine (PR #160628)

[Yingwei Zheng via llvm-commits]

================
@@ -2405,6 +2405,23 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
               matchBSwapOrBitReverse(*II, /*MatchBSwaps*/ true,
                                      /*MatchBitReversals*/ true))
         return BitOp;
+
+      // R = fshl(X, X, C2)
+      // fshl(R, R, C1) --> fshl(X, X, (C1 + C2) % bitsize)
+      Value *InnerOp;
+      Constant *ShAmtInnerC;
+      if (match(Op0, m_FShl(m_Value(InnerOp), m_Deferred(InnerOp),
+                            m_ImmConstant(ShAmtInnerC))) &&
+          Op0 == Op1) {
+        APInt Sum =
+            ShAmtC->getUniqueInteger() + ShAmtInnerC->getUniqueInteger();
----------------
dtcxzyw wrote:

> With this change, will this allow rotates with contant vectors such that something like this'll be allowed: <1, 2, 3> + <5, 6, 7> = <6, 8, 10>?

No. I mean, it is enough to just handle scalar + vector splat cases. See also https://llvm.org/docs/InstCombineContributorGuide.html#guidelines-for-reviewers. If you want to handle non-splat cases, use `ConstantFoldBinaryOpOperands` instead.


https://github.com/llvm/llvm-project/pull/160628