[llvm] [VPlan] Introduce child regions as VPlan transform. (PR #129402)

[Florian Hahn via llvm-commits]

================
@@ -14,26 +14,93 @@
 #include "LoopVectorizationPlanner.h"
 #include "VPlan.h"
 #include "VPlanCFG.h"
+#include "VPlanDominatorTree.h"
 #include "VPlanTransforms.h"
 #include "llvm/Analysis/LoopInfo.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 
 using namespace llvm;
 
-void VPlanTransforms::introduceTopLevelVectorLoopRegion(
-    VPlan &Plan, Type *InductionTy, PredicatedScalarEvolution &PSE,
-    bool RequiresScalarEpilogueCheck, bool TailFolded, Loop *TheLoop) {
-  // TODO: Generalize to introduce all loop regions.
-  auto *HeaderVPBB = cast<VPBasicBlock>(Plan.getEntry()->getSingleSuccessor());
-  VPBlockUtils::disconnectBlocks(Plan.getEntry(), HeaderVPBB);
+/// Checks if \p HeaderVPB is a loop header block in the plain CFG; that is, it
+/// has exactly 2 predecessors (preheader and latch), where the block
+/// dominates the latch and the preheader dominates the block. If it is a
+/// header block, returns a pair with the corresponding preheader and latch
+/// blocks. Otherwise return std::nullopt.
+static std::optional<std::pair<VPBasicBlock *, VPBasicBlock *>>
+getPreheaderAndLatch(VPBlockBase *HeaderVPB, const VPDominatorTree &VPDT) {
----------------
fhahn wrote:

This means effectively going back to the orignial verison, which enforced the specific order? The current version moves the canonicalization of order outside the CFG construction, and leaves the plain CFG a 1-1 mapping of the incomig CFG, with both phi recipes and VPBB predecessor order matching the order in the original CFG.

https://github.com/llvm/llvm-project/pull/129402