[llvm] [LV] Add initial legality checks for early exit loops with side effects (PR #145663)

[Florian Hahn via llvm-commits]

================
@@ -1813,6 +1840,113 @@ bool LoopVectorizationLegality::isVectorizableEarlyExitLoop() {
                        "backedge taken count: "
                     << *SymbolicMaxBTC << '\n');
   UncountableExitingBB = SingleUncountableExitingBlock;
+  UncountableExitWithSideEffects = HasSideEffects;
+  return true;
+}
+
+bool LoopVectorizationLegality::canUncountableExitConditionLoadBeMoved(
+    BasicBlock *ExitingBlock) {
+  LoadInst *CriticalUncountableExitConditionLoad = nullptr;
+
+  // Try to find a load in the critical path for the uncountable exit condition.
+  // This is currently matching about the simplest form we can, expecting
+  // only one in-loop load, the result of which is directly compared against
+  // a loop-invariant value.
+  // FIXME: We're insisting on a single use for now, because otherwise we will
+  // need to make PHI nodes for other users. That can be done once the initial
+  // transform code lands.
+  auto *Br = cast<BranchInst>(ExitingBlock->getTerminator());
+
+  using namespace llvm::PatternMatch;
+  Value *L = nullptr;
+  Value *R = nullptr;
+  if (!match(Br->getCondition(),
+             m_OneUse(m_ICmp(m_OneUse(m_Value(L)), (m_Value(R)))))) {
+    reportVectorizationFailure(
+        "Early exit loop with store but no supported condition load",
+        "NoConditionLoadForEarlyExitLoop", ORE, TheLoop);
+    return false;
+  }
+
+  // FIXME: Don't rely on operand ordering for the comparison.
+  if (!TheLoop->isLoopInvariant(R)) {
+    reportVectorizationFailure(
+        "Early exit loop with store but no supported condition load",
+        "NoConditionLoadForEarlyExitLoop", ORE, TheLoop);
+    return false;
+  }
+
+  if (auto *Load = dyn_cast<LoadInst>(L)) {
+    // Make sure that the load address is not loop invariant; we want an
+    // address calculation that we can rotate to the next vector iteration.
+    const SCEV *PtrScev = PSE.getSE()->getSCEV(Load->getPointerOperand());
+    if (PSE.getSE()->isLoopInvariant(PtrScev, TheLoop)) {
----------------
fhahn wrote:

It would probably be safer to check for !AddRec here instead of bailing out on `isLoopInvariant`, which may not be able to identify all invariant cases.

This way we also don't rely on an implementation detail of `isDereferenceableAndAlignedInLoop` (the check for the pointer being an AddRec, i.e. there's nothing fundamental preventing it to look through things like pointer selects).

https://github.com/llvm/llvm-project/pull/145663