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

[Ricardo Jesus via llvm-commits]

================
@@ -939,3 +988,432 @@ 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. We also need to know the target's minimum page size in order to
+  // generate runtime memory checks to ensure the vector version won't fault.
+  if (!TTI->supportsScalableVectors() || !TTI->getMinPageSize().has_value() ||
+      DisableFindFirstByte)
+    return false;
+
+  // Define some constants we need throughout.
+  BasicBlock *Header = CurLoop->getHeader();
+  LLVMContext &Ctx = Header->getContext();
+
+  // We are expecting the four blocks defined below: Header, MatchBB, InnerBB,
+  // and OuterBB. For now, we will bail our for almost anything else. The Four
+  // blocks contain 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.
+  for (BasicBlock *BB : LoopBlocks)
+    for (Instruction &I : *BB)
+      if (&I != IndPhi)
+        for (User *U : I.users())
+          if (!CurLoop->contains(cast<Instruction>(U)))
+            return false;
+
+  // Match the branch instruction in the header. We are expecting an
+  // unconditional branch to the inner loop.
+  //
+  // Header:
+  //   %14 = phi ptr [ %24, %OuterBB ], [ %3, %Header.preheader ]
+  //   %15 = load i8, ptr %14, align 1
+  //   br label %MatchBB
+  BasicBlock *MatchBB;
+  if (!match(Header->getTerminator(), m_UnconditionalBr(MatchBB)) ||
+      !InnerLoop->contains(MatchBB))
+    return false;
+
+  // MatchBB should be the entrypoint into the inner loop containing the
+  // comparison between a search element and a needle.
+  //
+  // 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
+  BasicBlock *ExitSucc, *InnerBB;
+  Value *LoadSearch, *LoadNeedle;
+  CmpPredicate MatchPred;
+  if (!match(MatchBB->getTerminator(),
+             m_Br(m_ICmp(MatchPred, m_Value(LoadSearch), m_Value(LoadNeedle)),
+                  m_BasicBlock(ExitSucc), m_BasicBlock(InnerBB))) ||
+      MatchPred != ICmpInst::ICMP_EQ || !InnerLoop->contains(InnerBB))
+    return false;
+
+  // We expect outside uses of `IndPhi' in ExitSucc (and only there).
+  for (User *U : IndPhi->users())
+    if (!CurLoop->contains(cast<Instruction>(U))) {
+      auto *PN = dyn_cast<PHINode>(U);
+      if (!PN || PN->getParent() != ExitSucc)
+        return false;
+    }
+
+  // Match the loads and check they are simple.
+  Value *Search, *Needle;
+  if (!match(LoadSearch, m_Load(m_Value(Search))) ||
+      !match(LoadNeedle, m_Load(m_Value(Needle))) ||
+      !cast<LoadInst>(LoadSearch)->isSimple() ||
+      !cast<LoadInst>(LoadNeedle)->isSimple())
+    return false;
+
+  // Check we are loading valid characters.
+  Type *CharTy = LoadSearch->getType();
+  if (!CharTy->isIntegerTy() || LoadNeedle->getType() != CharTy)
+    return false;
+
+  // Pick the vectorisation factor based on CharTy, work out the cost of the
+  // match intrinsic and decide if we should use it.
+  // Note: For the time being we assume 128-bit vectors.
+  unsigned VF = 128 / CharTy->getIntegerBitWidth();
+  SmallVector<Type *> Args = {
+      ScalableVectorType::get(CharTy, VF), FixedVectorType::get(CharTy, VF),
+      ScalableVectorType::get(Type::getInt1Ty(Ctx), VF)};
+  IntrinsicCostAttributes Attrs(Intrinsic::experimental_vector_match, Args[2],
+                                Args);
+  if (TTI->getIntrinsicInstrCost(Attrs, TTI::TCK_SizeAndLatency) > 4)
+    return false;
+
+  // The loads come from two PHIs, each with two incoming values.
+  PHINode *PSearch = dyn_cast<PHINode>(Search);
+  PHINode *PNeedle = dyn_cast<PHINode>(Needle);
+  if (!PSearch || PSearch->getNumIncomingValues() != 2 || !PNeedle ||
+      PNeedle->getNumIncomingValues() != 2)
+    return false;
+
+  // One PHI comes from the outer loop (PSearch), the other one from the inner
+  // loop (PNeedle). PSearch effectively corresponds to IndPhi.
+  if (InnerLoop->contains(PSearch))
+    std::swap(PSearch, PNeedle);
+  if (PSearch != &Header->front() || PNeedle != &MatchBB->front())
+    return false;
+
+  // The incoming values of both PHI nodes should be a gep of 1.
+  Value *SearchStart = PSearch->getIncomingValue(0);
+  Value *SearchIndex = PSearch->getIncomingValue(1);
+  if (CurLoop->contains(PSearch->getIncomingBlock(0)))
+    std::swap(SearchStart, SearchIndex);
+
+  Value *NeedleStart = PNeedle->getIncomingValue(0);
+  Value *NeedleIndex = PNeedle->getIncomingValue(1);
+  if (InnerLoop->contains(PNeedle->getIncomingBlock(0)))
+    std::swap(NeedleStart, NeedleIndex);
+
+  // Match the GEPs.
+  if (!match(SearchIndex, m_GEP(m_Specific(PSearch), m_One())) ||
+      !match(NeedleIndex, m_GEP(m_Specific(PNeedle), m_One())))
+    return false;
+
+  // Check the GEPs result type matches `CharTy'.
+  GetElementPtrInst *GEPSearch = cast<GetElementPtrInst>(SearchIndex);
+  GetElementPtrInst *GEPNeedle = cast<GetElementPtrInst>(NeedleIndex);
+  if (GEPSearch->getResultElementType() != CharTy ||
+      GEPNeedle->getResultElementType() != CharTy)
+    return false;
+
+  // InnerBB should increment the address of the needle pointer.
+  //
+  // InnerBB:
+  //   %17 = getelementptr inbounds i8, ptr %20, i64 1
+  //   %18 = icmp eq ptr %17, %10
+  //   br i1 %18, label %OuterBB, label %MatchBB
+  BasicBlock *OuterBB;
+  Value *NeedleEnd;
+  if (!match(InnerBB->getTerminator(),
+             m_Br(m_SpecificICmp(ICmpInst::ICMP_EQ, m_Specific(GEPNeedle),
+                                 m_Value(NeedleEnd)),
+                  m_BasicBlock(OuterBB), m_Specific(MatchBB))) ||
+      !CurLoop->contains(OuterBB))
+    return false;
+
+  // OuterBB should increment the address of the search element pointer.
+  //
+  // OuterBB:
+  //   %24 = getelementptr inbounds i8, ptr %14, i64 1
+  //   %25 = icmp eq ptr %24, %6
+  //   br i1 %25, label %ExitFail, label %Header
+  BasicBlock *ExitFail;
+  Value *SearchEnd;
+  if (!match(OuterBB->getTerminator(),
+             m_Br(m_SpecificICmp(ICmpInst::ICMP_EQ, m_Specific(GEPSearch),
+                                 m_Value(SearchEnd)),
+                  m_BasicBlock(ExitFail), m_Specific(Header))))
----------------
rj-jesus wrote:

Thanks, done in [`@find_first_of_i8_multi_exit`](https://github.com/llvm/llvm-project/pull/101976/commits/6d1755189549f3f6dd804d98810cbc374cd087a8#diff-f256205074dbd8200178875fe608f7daf7a0871a9e4bea54d8215524588aca76R325).

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