[llvm] [LoopIdiomRecognize] Enable clmul optimization for CRC loops (PR #203405)

[Ramkumar Ramachandra via llvm-commits]

================
@@ -1665,24 +1649,105 @@ bool LoopIdiomRecognize::optimizeCRCLoop(const PolynomialInfo &Info) {
     Indexer = Info.ByteOrderSwapped ? HiIdx(Builder, Indexer, "indexer.hi")
                                     : LoByte(Builder, Indexer, "indexer.lo");
 
-    // Always index into a GEP using the index type.
-    Indexer = Builder.CreateZExt(
-        Indexer, SE->getDataLayout().getIndexType(GV->getType()),
-        "indexer.ext");
-
-    // CRCTableLd = CRCTable[(iv'th byte of data) ^ (top|bottom) byte of CRC].
-    Value *CRCTableGEP =
-        Builder.CreateInBoundsGEP(CRCTy, GV, Indexer, "tbl.ptradd");
-    Value *CRCTableLd = Builder.CreateLoad(CRCTy, CRCTableGEP, "tbl.ld");
-
-    // CRCNext = (CRC (<<|>>) 8) ^ CRCTableLd, or simply CRCTableLd in case of
-    // CRC-8.
-    Value *CRCNext = CRCTableLd;
-    if (CRCBW > 8) {
-      Value *CRCShift = Info.ByteOrderSwapped
-                            ? Builder.CreateShl(CRC, 8, "crc.be.shift")
-                            : Builder.CreateLShr(CRC, 8, "crc.le.shift");
-      CRCNext = Builder.CreateXor(CRCShift, CRCTableLd, "crc.next");
+    // Compute the next value of the CRC register based on Indexer. This could
+    // either be a table load or an on-the-fly computation, depending on whether
+    // llvm.clmul is fast on the target.
----------------
artagnon wrote:

> // Compute the next value of the CRC register based on Indexer. This could
> // either be a table load or an on-the-fly computation, depending on whether
> // llvm.clmul is fast on the target.

Not sure what Barrett reduction or clmul has to do with "on-the-fly computation".

https://github.com/llvm/llvm-project/pull/203405