[llvm] [AArch64] Add MATCH loops to LoopIdiomVectorizePass (PR #101976)

[Ricardo Jesus via llvm-commits]

================
@@ -939,3 +988,361 @@ void LoopIdiomVectorize::transformByteCompare(GetElementPtrInst *GEPA,
       report_fatal_error("Loops must remain in LCSSA form!");
   }
 }
+
+bool LoopIdiomVectorize::recognizeFindFirstByte() {
+  // Currently the transformation only works on scalable vector types, although
+  // there is no fundamental reason why it cannot be made to work for fixed
+  // vectors too.
+  if (!TTI->supportsScalableVectors() || DisableFindFirstByte)
+    return false;
+
+  // Define some constants we need throughout.
+  BasicBlock *Header = CurLoop->getHeader();
+  LLVMContext &Ctx = Header->getContext();
+
+  // We are expecting the blocks below. For now, we will bail out for almost
+  // anything other than this.
+  //
+  // Header:
+  //   %14 = phi ptr [ %24, %OuterBB ], [ %3, %Header.preheader ]
+  //   %15 = load i8, ptr %14, align 1
+  //   br label %MatchBB
+  //
+  // MatchBB:
+  //   %20 = phi ptr [ %7, %Header ], [ %17, %InnerBB ]
+  //   %21 = load i8, ptr %20, align 1
+  //   %22 = icmp eq i8 %15, %21
+  //   br i1 %22, label %ExitSucc, label %InnerBB
+  //
+  // InnerBB:
+  //   %17 = getelementptr inbounds i8, ptr %20, i64 1
+  //   %18 = icmp eq ptr %17, %10
+  //   br i1 %18, label %OuterBB, label %MatchBB
+  //
+  // OuterBB:
+  //   %24 = getelementptr inbounds i8, ptr %14, i64 1
+  //   %25 = icmp eq ptr %24, %6
+  //   br i1 %25, label %ExitFail, label %Header
+
+  // We expect the four blocks above, which include one nested loop.
+  if (CurLoop->getNumBackEdges() != 1 || CurLoop->getNumBlocks() != 4 ||
+      CurLoop->getSubLoops().size() != 1)
+    return false;
+
+  auto *InnerLoop = CurLoop->getSubLoops().front();
+  PHINode *IndPhi = dyn_cast<PHINode>(&Header->front());
+  if (!IndPhi || IndPhi->getNumIncomingValues() != 2)
+    return false;
+
+  // Check instruction counts.
+  auto LoopBlocks = CurLoop->getBlocks();
+  if (LoopBlocks[0]->sizeWithoutDebug() > 3 ||
+      LoopBlocks[1]->sizeWithoutDebug() > 4 ||
+      LoopBlocks[2]->sizeWithoutDebug() > 3 ||
+      LoopBlocks[3]->sizeWithoutDebug() > 3)
+    return false;
+
+  // Check that no instruction other than IndPhi has outside uses.
----------------
rj-jesus wrote:

It wasn't but is now, thanks.

https://github.com/llvm/llvm-project/pull/101976