[llvm] [LV] Optimise latch exit induction users for some early exit loops (PR #128880)

[Benjamin Maxwell via llvm-commits]

================
@@ -730,67 +730,74 @@ static VPWidenInductionRecipe *getOptimizableIVOf(VPValue *VPV) {
   return IsWideIVInc() ? WideIV : nullptr;
 }
 
-void VPlanTransforms::optimizeInductionExitUsers(
-    VPlan &Plan, DenseMap<VPValue *, VPValue *> &EndValues) {
+static VPValue *
+optimizeLatchExitInductionUser(VPlan &Plan, VPTypeAnalysis &TypeInfo,
+                               VPBlockBase *PredVPBB, VPValue *Op,
+                               DenseMap<VPValue *, VPValue *> &EndValues) {
   using namespace VPlanPatternMatch;
-  SmallVector<VPIRBasicBlock *> ExitVPBBs(Plan.getExitBlocks());
-  if (ExitVPBBs.size() != 1)
-    return;
 
-  VPIRBasicBlock *ExitVPBB = ExitVPBBs[0];
-  VPBlockBase *PredVPBB = ExitVPBB->getSinglePredecessor();
-  if (!PredVPBB)
-    return;
-  assert(PredVPBB == Plan.getMiddleBlock() &&
-         "predecessor must be the middle block");
-
-  VPTypeAnalysis TypeInfo(Plan.getCanonicalIV()->getScalarType());
-  VPBuilder B(Plan.getMiddleBlock()->getTerminator());
-  for (VPRecipeBase &R : *ExitVPBB) {
-    auto *ExitIRI = cast<VPIRInstruction>(&R);
-    if (!isa<PHINode>(ExitIRI->getInstruction()))
-      break;
+  VPValue *Incoming;
+  if (!match(Op, m_VPInstruction<VPInstruction::ExtractFromEnd>(
+                     m_VPValue(Incoming), m_SpecificInt(1))))
+    return nullptr;
 
-    VPValue *Incoming;
-    if (!match(ExitIRI->getOperand(0),
-               m_VPInstruction<VPInstruction::ExtractFromEnd>(
-                   m_VPValue(Incoming), m_SpecificInt(1))))
-      continue;
+  auto *WideIV = getOptimizableIVOf(Incoming);
+  if (!WideIV)
+    return nullptr;
 
-    auto *WideIV = getOptimizableIVOf(Incoming);
-    if (!WideIV)
-      continue;
-    VPValue *EndValue = EndValues.lookup(WideIV);
-    assert(EndValue && "end value must have been pre-computed");
+  VPValue *EndValue = EndValues.lookup(WideIV);
----------------
MacDue wrote:

I see the end values are populated in `addScalarResumePhis()`, am I right in thinking this map goes from `IV` -> exit IV value from scalar loop? Could this be non-constant? (in the changed tests it is always a constant). 

https://github.com/llvm/llvm-project/pull/128880