[llvm] [LV] Introduce the EVLIVSimplify Pass for EVL-vectorized loops (PR #91796)

[Min-Yih Hsu via llvm-commits]

================
@@ -0,0 +1,296 @@
+//===------ 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/ValueTracking.h"
+#include "llvm/IR/IRBuilder.h"
+#include "llvm/IR/PatternMatch.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Pass.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"
+
+#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);
+};
+
+struct EVLIndVarSimplify : public LoopPass {
+  static char ID;
+
+  EVLIndVarSimplify() : LoopPass(ID) {
+    initializeEVLIndVarSimplifyPass(*PassRegistry::getPassRegistry());
+  }
+
+  bool runOnLoop(Loop *L, LPPassManager &LPM) override;
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<ScalarEvolutionWrapperPass>();
+    AU.setPreservesCFG();
+  }
+};
+} // 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;
+}
+
+// Remove the original induction variable if it's not used anywhere.
+static void tryCleanupOriginalIndVar(PHINode *OrigIndVar,
+                                     const InductionDescriptor &IVD) {
+  if (OrigIndVar->getNumIncomingValues() != 2)
+    return;
+  Value *InitValue = OrigIndVar->getIncomingValue(0);
+  Value *RecValue = OrigIndVar->getIncomingValue(1);
+  if (InitValue != IVD.getStartValue())
+    std::swap(InitValue, RecValue);
+
+  // If the only user of OrigIndVar is the one that produces RecValue, then we
+  // can safely remove it.
+  if (!OrigIndVar->hasOneUse() || OrigIndVar->user_back() != RecValue)
+    return;
+
+  LLVM_DEBUG(dbgs() << "Removed the original IndVar " << *OrigIndVar << "\n");
+  // Turn OrigIndVar into dead code by replacing all its uses by the initial
+  // value of this loop.
+  OrigIndVar->replaceAllUsesWith(InitValue);
+  OrigIndVar->eraseFromParent();
+}
+
+bool EVLIndVarSimplifyImpl::run(Loop &L) {
+  if (!EnableEVLIndVarSimplify)
+    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);
----------------
mshockwave wrote:

The SCEV of initial and final values are probably going to be gone, since they're only used in the ad-hoc check I wrote, which is subject to be replaced. Also, I don't quite understand what you mean by saying getBounds doesn't use SCEV, because judging from its code it's using ScalarEvolution to calculate the initial and final values. Could you elaborate a little more?

https://github.com/llvm/llvm-project/pull/91796