[llvm] [LoopIdiom] Use HashRecognize to optimize CRC (PR #143208)

[Piotr Fusik via llvm-commits]

================
@@ -1473,6 +1491,160 @@ bool LoopIdiomRecognize::avoidLIRForMultiBlockLoop(bool IsMemset,
   return false;
 }
 
+bool LoopIdiomRecognize::optimizeCRCLoop(const PolynomialInfo &Info) {
+  // FIXME: Hexagon has a special HexagonLoopIdiom that optimizes CRC using
+  // carry-less multiplication instructions, which is more efficient than our
+  // Sarwate table-lookup optimization. Hence, until we're able to emit
+  // target-specific instructions for Hexagon, subsuming HexagonLoopIdiom,
+  // disable the optimization for Hexagon.
+  Module &M = *CurLoop->getHeader()->getModule();
+  Triple TT(M.getTargetTriple());
+  if (TT.getArch() == Triple::hexagon)
+    return false;
+
+  // First, create a new GlobalVariable corresponding to the
+  // Sarwate-lookup-table.
+  Type *CRCTy = Info.LHS->getType();
+  unsigned CRCBW = CRCTy->getIntegerBitWidth();
+  std::array<Constant *, 256> CRCConstants;
+  transform(HashRecognize::genSarwateTable(Info.RHS, Info.ByteOrderSwapped),
+            CRCConstants.begin(),
+            [CRCTy](const APInt &E) { return ConstantInt::get(CRCTy, E); });
+  Constant *ConstArray =
+      ConstantArray::get(ArrayType::get(CRCTy, 256), CRCConstants);
+  GlobalVariable *GV =
----------------
pfusik wrote:

What do you mean by "identical" ? Constant arrays cannot be merged if the program compares array addresses. So I'd expect that's an option. Also, same-translation-unit case would be nice to test, in case someone implements CRC as a C macro.

https://github.com/llvm/llvm-project/pull/143208