[llvm] [SCEV] Infer loop max trip count from memory accesses (PR #70361)

[Shilei Tian via llvm-commits]

================
@@ -8191,6 +8204,133 @@ ScalarEvolution::getSmallConstantTripMultiple(const Loop *L,
   return getSmallConstantTripMultiple(L, ExitCount);
 }
 
+/// Collect all load/store instructions that must be executed in every iteration
+/// of loop \p L .
+static void
+collectExecLoadStoreInsideLoop(const Loop *L, DominatorTree &DT,
+                               SmallVector<Instruction *, 4> &MemInsts) {
+  // It is difficult to tell if the load/store instruction is executed on every
+  // iteration inside an irregular loop.
+  if (!L->isLoopSimplifyForm() || !L->isInnermost())
+    return;
+
+  // FIXME: To make the case more typical, we only analyze loops that have one
+  // exiting block and the block must be the latch. It is easier to capture
+  // loops with memory access that will be executed in every iteration.
+  const BasicBlock *LoopLatch = L->getLoopLatch();
+  assert(LoopLatch && "normal form loop doesn't have a latch");
+  if (L->getExitingBlock() != LoopLatch)
+    return;
+
+  // We will not continue if sanitizer is enabled.
+  const Function *F = LoopLatch->getParent();
+  if (F->hasFnAttribute(Attribute::SanitizeAddress) ||
+      F->hasFnAttribute(Attribute::SanitizeThread) ||
+      F->hasFnAttribute(Attribute::SanitizeMemory) ||
+      F->hasFnAttribute(Attribute::SanitizeHWAddress) ||
+      F->hasFnAttribute(Attribute::SanitizeMemTag))
+    return;
+
+  for (auto *BB : L->getBlocks()) {
+    // We need to make sure that max execution time of MemAccessBB in loop
+    // represents latch max excution time. The BB below should be skipped:
+    //            Entry
+    //              │
+    //        ┌─────▼─────┐
+    //        │Loop Header◄─────┐
+    //        └──┬──────┬─┘     │
+    //           │      │       │
+    //  ┌────────▼──┐ ┌─▼─────┐ │
+    //  │MemAccessBB│ │OtherBB│ │
+    //  └────────┬──┘ └─┬─────┘ │
+    //           │      │       │
+    //         ┌─▼──────▼─┐     │
+    //         │Loop Latch├─────┘
+    //         └────┬─────┘
+    //              ▼
+    //             Exit
+    if (!DT.dominates(BB, LoopLatch))
+      continue;
+
+    for (Instruction &I : *BB) {
+      if (isa<LoadInst>(&I) || isa<StoreInst>(&I))
+        MemInsts.push_back(&I);
+    }
+  }
+}
+
+/// Return a SCEV representing the memory size of pointer \p V .
+static const SCEV *getCertainSizeOfMem(const SCEV *V, Type *RTy,
+                                       const DataLayout &DL,
+                                       const TargetLibraryInfo &TLI,
+                                       ScalarEvolution *SE) {
+  const SCEVUnknown *PtrBase = dyn_cast<SCEVUnknown>(V);
+  if (!PtrBase)
+    return nullptr;
+  Value *Ptr = PtrBase->getValue();
+  uint64_t Size = 0;
+  if (!llvm::getObjectSize(Ptr, Size, DL, &TLI))
+    return nullptr;
+  return SE->getConstant(RTy, Size);
+}
+
+/// Check if we can make sure that the indices of a GEP instruction will not
+/// wrap.
+static bool checkIndexWrap(Value *Ptr, ScalarEvolution *SE) {
+  auto *PtrGEP = dyn_cast<GetElementPtrInst>(Ptr);
+  if (!PtrGEP)
+    return false;
+  for (Value *Index : PtrGEP->indices()) {
+    Value *V = Index;
+    if (isa<ZExtInst>(V) || isa<SExtInst>(V))
+      V = cast<Instruction>(Index)->getOperand(0);
+    auto *SCEV = SE->getSCEV(V);
+    if (isa<SCEVCouldNotCompute>(SCEV))
+      return false;
+    auto *AryExpr = dyn_cast<SCEVNAryExpr>(SE->getSCEV(V));
+    if (!AryExpr)
+      continue;
+    if (!AryExpr->hasNoSelfWrap())
+      return false;
+  }
+  return true;
+}
+
+const SCEV *
+ScalarEvolution::getConstantMaxTripCountFromMemAccess(const Loop *L) {
+  SmallVector<Instruction *, 4> MemInsts;
+  collectExecLoadStoreInsideLoop(L, DT, MemInsts);
+
+  SmallVector<const SCEV *> InferCountColl;
+  const DataLayout &DL = getDataLayout();
+
+  for (Instruction *I : MemInsts) {
+    Value *Ptr = getLoadStorePointerOperand(I);
+    assert(Ptr && "empty pointer operand");
+    if (!checkIndexWrap(Ptr, this))
+      continue;
+    auto *AddRec = dyn_cast<SCEVAddRecExpr>(getSCEV(Ptr));
+    if (!AddRec || !AddRec->isAffine())
----------------
shiltian wrote:

Okay, yeah, that makes sense. However, with what we have right now, we will get the max trip count 257, which is out of the range of `%iv`. That might be the reason that the initial patch has the "weird" get small count of index thing.

https://github.com/llvm/llvm-project/pull/70361