[llvm] [InstCombine] Improve bitfield addition (PR #77184)

[Simon Pilgrim via llvm-commits]

================
@@ -3379,6 +3379,149 @@ Value *InstCombinerImpl::foldAndOrOfICmps(ICmpInst *LHS, ICmpInst *RHS,
   return foldAndOrOfICmpsUsingRanges(LHS, RHS, IsAnd);
 }
 
+struct BitFieldAddBitMask {
+  const APInt *Lower;
+  const APInt *Upper;
+};
+struct BitFieldOptBitMask {
+  const APInt *Lower;
+  const APInt *Upper;
+  const APInt *New;
+};
+struct BitFieldAddInfo {
+  Value *X;
+  Value *Y;
+  bool opt;
+  union {
+    BitFieldAddBitMask AddMask;
+    BitFieldOptBitMask OptMask;
+  };
+};
+
+static Value *foldBitFieldArithmetic(BinaryOperator &I,
----------------
RKSimon wrote:

To address the reported issues, this has been written in terms of improving bitfield math, but I'm not certain if we should be addressing this in terms of more generic canonicalizations or not - are we likely to hit sub-parts of this in other places do you think?

https://github.com/llvm/llvm-project/pull/77184