[LV] Introduce the EVLIVSimplify Pass for EVL-vectorized loops

llvm/include/llvm/CodeGen/EVLIndVarSimplify.h

@@ -0,0 +1,31 @@
+//===- EVLIndVarSimplify.h - Optimize vectorized loops w/ EVL IV-*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass optimizes a vectorized loop with canonical IV to using EVL-based
+// IV if it was tail-folded by predicated EVL.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CODEGEN_EVLINDVARSIMPLIFY_H
+#define LLVM_CODEGEN_EVLINDVARSIMPLIFY_H
+
+#include "llvm/Analysis/LoopAnalysisManager.h"
+#include "llvm/IR/PassManager.h"
+
+namespace llvm {
+class Loop;
+class LPMUpdater;
+
+/// Turn vectorized loops with canonical induction variables into loops that
+/// only use a single EVL-based induction variable.
+struct EVLIndVarSimplifyPass : public PassInfoMixin<EVLIndVarSimplifyPass> {
+  PreservedAnalyses run(Loop &L, LoopAnalysisManager &LAM,
+                        LoopStandardAnalysisResults &AR, LPMUpdater &U);
+};
+} // namespace llvm
+#endif

llvm/lib/CodeGen/CMakeLists.txt

@@ -54,6 +54,7 @@ add_llvm_component_library(LLVMCodeGen
   EarlyIfConversion.cpp
   EdgeBundles.cpp
   EHContGuardCatchret.cpp
+  EVLIndVarSimplify.cpp
   ExecutionDomainFix.cpp
   ExpandLargeDivRem.cpp
   ExpandLargeFpConvert.cpp

llvm/lib/CodeGen/EVLIndVarSimplify.cpp

@@ -0,0 +1,236 @@
+//===------ EVLIndVarSimplify.cpp - Optimize vectorized loops w/ EVL IV----===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass optimizes a vectorized loop with canonical IV to using EVL-based
+// IV if it was tail-folded by predicated EVL.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/EVLIndVarSimplify.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/IVDescriptors.h"
+#include "llvm/Analysis/LoopInfo.h"
+#include "llvm/Analysis/LoopPass.h"
+#include "llvm/Analysis/ScalarEvolution.h"
+#include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/Analysis/ValueTracking.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/MathExtras.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Scalar/LoopPassManager.h"
+#include "llvm/Transforms/Utils/Local.h"
+#include "llvm/Transforms/Utils/ScalarEvolutionExpander.h"
+
+#define DEBUG_TYPE "evl-iv-simplify"
+
+using namespace llvm;
+
+STATISTIC(NumEliminatedCanonicalIV, "Number of canonical IVs we eliminated");
+
+static cl::opt<bool> EnableEVLIndVarSimplify(
+    "enable-evl-indvar-simplify",
+    cl::desc("Enable EVL-based induction variable simplify Pass"), cl::Hidden,
+    cl::init(true));
+
+namespace {
+struct EVLIndVarSimplifyImpl {
+  ScalarEvolution &SE;
+
+  explicit EVLIndVarSimplifyImpl(LoopStandardAnalysisResults &LAR)
+      : SE(LAR.SE) {}
+
+  explicit EVLIndVarSimplifyImpl(ScalarEvolution &SE) : SE(SE) {}
+
+  // Returns true if modify the loop.
+  bool run(Loop &L);
+};
+} // anonymous namespace
+
+static uint32_t getVFFromIndVar(const SCEV *Step, const Function &F) {
+  if (!Step)
+    return 0U;
+
+  // Looking for loops with IV step value in the form of `(<constant VF> x
+  // vscale)`.
+  if (auto *Mul = dyn_cast<SCEVMulExpr>(Step)) {
+    if (Mul->getNumOperands() == 2) {
+      const SCEV *LHS = Mul->getOperand(0);
+      const SCEV *RHS = Mul->getOperand(1);
+      if (auto *Const = dyn_cast<SCEVConstant>(LHS)) {
+        uint64_t V = Const->getAPInt().getLimitedValue();
+        if (isa<SCEVVScale>(RHS) && llvm::isUInt<32>(V))
+          return static_cast<uint32_t>(V);
+      }
+    }
+  }
+
+  // If not, see if the vscale_range of the parent function is a fixed value,
+  // which makes the step value to be replaced by a constant.
+  if (F.hasFnAttribute(Attribute::VScaleRange))
+    if (auto *ConstStep = dyn_cast<SCEVConstant>(Step)) {
+      APInt V = ConstStep->getAPInt().abs();
+      ConstantRange CR = llvm::getVScaleRange(&F, 64);
+      if (const APInt *Fixed = CR.getSingleElement()) {
+        V = V.zextOrTrunc(Fixed->getBitWidth());
+        uint64_t VF = V.udiv(*Fixed).getLimitedValue();
+        if (VF && llvm::isUInt<32>(VF) &&
+            // Make sure step is divisible by vscale.
+            V.urem(*Fixed).isZero())
+          return static_cast<uint32_t>(VF);
+      }
+    }
+
+  return 0U;
+}
+
+bool EVLIndVarSimplifyImpl::run(Loop &L) {
+  if (!EnableEVLIndVarSimplify)
+    return false;
+
+  BasicBlock *LatchBlock = L.getLoopLatch();
+  ICmpInst *OrigLatchCmp = L.getLatchCmpInst();
+  if (!LatchBlock || !OrigLatchCmp)
+    return false;
+
+  InductionDescriptor IVD;
+  PHINode *IndVar = L.getInductionVariable(SE);
+  if (!IndVar || !L.getInductionDescriptor(SE, IVD)) {
+    LLVM_DEBUG(dbgs() << "Cannot retrieve IV from loop " << L.getName()
+                      << "\n");
+    return false;
+  }
+
+  BasicBlock *InitBlock, *BackEdgeBlock;
+  if (!L.getIncomingAndBackEdge(InitBlock, BackEdgeBlock)) {
+    LLVM_DEBUG(dbgs() << "Expect unique incoming and backedge in "
+                      << L.getName() << "\n");
+    return false;
+  }
+
+  // Retrieve the loop bounds.
+  std::optional<Loop::LoopBounds> Bounds = L.getBounds(SE);
+  if (!Bounds) {
+    LLVM_DEBUG(dbgs() << "Could not obtain the bounds for loop " << L.getName()
+                      << "\n");
+    return false;
+  }
+  Value *CanonicalIVInit = &Bounds->getInitialIVValue();
+  Value *CanonicalIVFinal = &Bounds->getFinalIVValue();
+
+  const SCEV *StepV = IVD.getStep();
+  uint32_t VF = getVFFromIndVar(StepV, *L.getHeader()->getParent());
+  if (!VF) {
+    LLVM_DEBUG(dbgs() << "Could not infer VF from IndVar step '" << *StepV
+                      << "'\n");
+    return false;
+  }
+  LLVM_DEBUG(dbgs() << "Using VF=" << VF << " for loop " << L.getName()
+                    << "\n");
+
+  // Try to find the EVL-based induction variable.
+  using namespace PatternMatch;
+  BasicBlock *BB = IndVar->getParent();
+
+  Value *EVLIndVar = nullptr;
+  Value *RemTC = nullptr;
+  Value *TC = nullptr;
+  auto IntrinsicMatch = m_Intrinsic<Intrinsic::experimental_get_vector_length>(
+      m_Value(RemTC), m_SpecificInt(VF),
+      /*Scalable=*/m_SpecificInt(1));
+  for (auto &PN : BB->phis()) {
+    if (&PN == IndVar)
+      continue;
+
+    // Check 1: it has to contain both incoming (init) & backedge blocks
+    // from IndVar.
+    if (PN.getBasicBlockIndex(InitBlock) < 0 ||
+        PN.getBasicBlockIndex(BackEdgeBlock) < 0)
+      continue;
+    // Check 2: EVL index is always increasing, thus its inital value has to be
+    // equal to either the initial IV value (when the canonical IV is also
+    // increasing) or the last IV value (when canonical IV is decreasing).
+    Value *Init = PN.getIncomingValueForBlock(InitBlock);
+    using Direction = Loop::LoopBounds::Direction;
+    switch (Bounds->getDirection()) {
+    case Direction::Increasing:
+      if (Init != CanonicalIVInit)
+        continue;
+      break;
+    case Direction::Decreasing:
+      if (Init != CanonicalIVFinal)
+        continue;
+      break;
+    case Direction::Unknown:
+      // To be more permissive and see if either the initial or final IV value
+      // matches PN's init value.
+      if (Init != CanonicalIVInit && Init != CanonicalIVFinal)
+        continue;
+      break;
+    }
+    Value *RecValue = PN.getIncomingValueForBlock(BackEdgeBlock);
+    assert(RecValue);
+
+    LLVM_DEBUG(dbgs() << "Found candidate PN of EVL-based IndVar: " << PN
+                      << "\n");
+
+    // Check 3: Pattern match to find the EVL-based index and total trip count
+    // (TC).
+    if (match(RecValue,
+              m_c_Add(m_ZExtOrSelf(IntrinsicMatch), m_Specific(&PN))) &&
+        match(RemTC, m_Sub(m_Value(TC), m_Specific(&PN)))) {
+      EVLIndVar = RecValue;
+      break;
+    }
+  }
+
+  if (!EVLIndVar || !TC)
+    return false;
+
+  LLVM_DEBUG(dbgs() << "Using " << *EVLIndVar << " for EVL-based IndVar\n");
+
+  // Create an EVL-based comparison and replace the branch to use it as
+  // predicate.
+
+  // Loop::getLatchCmpInst check at the beginning of this function has ensured
+  // that latch block ends in a conditional branch.
+  auto *LatchBranch = cast<BranchInst>(LatchBlock->getTerminator());
+  assert(LatchBranch->getNumSuccessors() == 2);
+  ICmpInst::Predicate Pred;
+  if (LatchBranch->getSuccessor(0) == L.getHeader())
+    Pred = ICmpInst::ICMP_ULT;
+  else
+    Pred = ICmpInst::ICMP_UGE;
+
+  IRBuilder<> Builder(OrigLatchCmp);
+  auto *NewLatchCmp = Builder.CreateICmp(Pred, EVLIndVar, TC);
+  OrigLatchCmp->replaceAllUsesWith(NewLatchCmp);
+
+  // llvm::RecursivelyDeleteDeadPHINode only deletes cycles whose values are
+  // not used outside the cycles. However, in this case the now-RAUW-ed
+  // OrigLatchCmp will be consied a use outside the cycle while in reality it's
+  // practically dead. Thus we need to remove it before calling
+  // RecursivelyDeleteDeadPHINode.
+  (void)RecursivelyDeleteTriviallyDeadInstructions(OrigLatchCmp);
+  if (llvm::RecursivelyDeleteDeadPHINode(IndVar))
+    LLVM_DEBUG(dbgs() << "Removed original IndVar\n");
+
+  ++NumEliminatedCanonicalIV;
+
+  return true;
+}
+
+PreservedAnalyses EVLIndVarSimplifyPass::run(Loop &L, LoopAnalysisManager &LAM,
+                                             LoopStandardAnalysisResults &AR,
+                                             LPMUpdater &U) {
+  if (EVLIndVarSimplifyImpl(AR).run(L))
+    return PreservedAnalyses::allInSet<CFGAnalyses>();
+  return PreservedAnalyses::all();
+}

llvm/lib/Passes/PassBuilder.cpp

@@ -86,6 +86,7 @@
 #include "llvm/CodeGen/DwarfEHPrepare.h"
 #include "llvm/CodeGen/EarlyIfConversion.h"
 #include "llvm/CodeGen/EdgeBundles.h"
+#include "llvm/CodeGen/EVLIndVarSimplify.h"
 #include "llvm/CodeGen/ExpandLargeDivRem.h"
 #include "llvm/CodeGen/ExpandLargeFpConvert.h"
 #include "llvm/CodeGen/ExpandMemCmp.h"

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -24,6 +24,7 @@
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/Analysis/ScopedNoAliasAA.h"
 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
+#include "llvm/CodeGen/EVLIndVarSimplify.h"
 #include "llvm/CodeGen/GlobalMergeFunctions.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/Pass.h"
@@ -1273,6 +1274,8 @@ void PassBuilder::addVectorPasses(OptimizationLevel Level,
   FPM.addPass(LoopVectorizePass(
       LoopVectorizeOptions(!PTO.LoopInterleaving, !PTO.LoopVectorization)));
 
+  FPM.addPass(createFunctionToLoopPassAdaptor(EVLIndVarSimplifyPass()));
+
   FPM.addPass(InferAlignmentPass());
   if (IsFullLTO) {
     // The vectorizer may have significantly shortened a loop body; unroll

llvm/lib/Passes/PassRegistry.def

@@ -656,6 +656,7 @@ LOOP_ANALYSIS("should-run-extra-simple-loop-unswitch",
 #endif
 LOOP_PASS("canon-freeze", CanonicalizeFreezeInLoopsPass())
 LOOP_PASS("dot-ddg", DDGDotPrinterPass())
+LOOP_PASS("evl-iv-simplify", EVLIndVarSimplifyPass())
 LOOP_PASS("guard-widening", GuardWideningPass())
 LOOP_PASS("extra-simple-loop-unswitch-passes",
               ExtraLoopPassManager<ShouldRunExtraSimpleLoopUnswitch>())