[LoopVectorize] Enable more early exit vectorisation tests

PR #112138 introduced initial support for dispatching to
multiple exit blocks via split middle blocks. This patch
fixes a few issues so that we can enable more tests to use
the new enable-early-exit-vectorization flag. Fixes are:

1. The code to bail out for any loop live-out values happens
too late. This is because collectUsersInExitBlocks ignores
induction variables, which get dealt with in fixupIVUsers.
I've moved the check much earlier in processLoop by looking
for outside users of loop-defined values.
2. We shouldn't yet be interleaving when vectorising loops
with uncountable early exits, since we've not added support
for this yet.
3. Similarly, we also shouldn't be creating vector epilogues.
4. Similarly, we shouldn't enable tail-folding.
5. The existing implementation doesn't yet support loops
that require scalar epilogues, although I plan to add that
as part of PR #88385.
6. The new split middle blocks weren't being added to the
parent loop.

Author: david-arm

llvm/lib/Transforms/Vectorize/LoopVectorize.cpp

@@ -3016,6 +3016,22 @@ void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) {
   PSE.getSE()->forgetLoop(OrigLoop);
   PSE.getSE()->forgetBlockAndLoopDispositions();
 
+  // When dealing with uncountable early exits we create middle.split blocks
+  // between the vector loop region and the exit block. These blocks need
+  // adding to any outer loop.
+  VPRegionBlock *VectorRegion = State.Plan->getVectorLoopRegion();
+  Loop *OuterLoop = OrigLoop->getParentLoop();
+  if (Legal->hasUncountableEarlyExit() && OuterLoop) {
+    VPBasicBlock *MiddleVPBB = State.Plan->getMiddleBlock();
+    VPBlockBase *PredVPBB = MiddleVPBB->getSinglePredecessor();
+    while (PredVPBB && PredVPBB != VectorRegion) {
+      BasicBlock *MiddleSplitBB =
+          State.CFG.VPBB2IRBB[cast<VPBasicBlock>(PredVPBB)];
+      OuterLoop->addBasicBlockToLoop(MiddleSplitBB, *LI);
+      PredVPBB = PredVPBB->getSinglePredecessor();
+    }
+  }
+
   // After vectorization, the exit blocks of the original loop will have
   // additional predecessors. Invalidate SCEVs for the exit phis in case SE
   // looked through single-entry phis.
@@ -3046,7 +3062,6 @@ void InnerLoopVectorizer::fixVectorizedLoop(VPTransformState &State) {
   for (Instruction *PI : PredicatedInstructions)
     sinkScalarOperands(&*PI);
 
-  VPRegionBlock *VectorRegion = State.Plan->getVectorLoopRegion();
   VPBasicBlock *HeaderVPBB = VectorRegion->getEntryBasicBlock();
   BasicBlock *HeaderBB = State.CFG.VPBB2IRBB[HeaderVPBB];
 
@@ -4123,7 +4138,8 @@ LoopVectorizationCostModel::computeMaxVF(ElementCount UserVF, unsigned UserIC) {
   // a bottom-test and a single exiting block. We'd have to handle the fact
   // that not every instruction executes on the last iteration.  This will
   // require a lane mask which varies through the vector loop body.  (TODO)
-  if (TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
+  if (Legal->hasUncountableEarlyExit() ||
+      TheLoop->getExitingBlock() != TheLoop->getLoopLatch()) {
     // If there was a tail-folding hint/switch, but we can't fold the tail by
     // masking, fallback to a vectorization with a scalar epilogue.
     if (ScalarEpilogueStatus == CM_ScalarEpilogueNotNeededUsePredicate) {
@@ -4753,7 +4769,9 @@ bool LoopVectorizationPlanner::isCandidateForEpilogueVectorization(
   // Epilogue vectorization code has not been auditted to ensure it handles
   // non-latch exits properly.  It may be fine, but it needs auditted and
   // tested.
-  if (OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch())
+  // TODO: Add support for loops with an early exit.
+  if (Legal->hasUncountableEarlyExit() ||
+      OrigLoop->getExitingBlock() != OrigLoop->getLoopLatch())
     return false;
 
   return true;
@@ -5001,6 +5019,10 @@ LoopVectorizationCostModel::selectInterleaveCount(ElementCount VF,
   if (!Legal->isSafeForAnyVectorWidth())
     return 1;
 
+  // We don't attempt to perform interleaving for early exit loops.
+  if (Legal->hasUncountableEarlyExit())
+    return 1;
+
   auto BestKnownTC = getSmallBestKnownTC(PSE, TheLoop);
   const bool HasReductions = !Legal->getReductionVars().empty();
 
@@ -7813,11 +7835,14 @@ DenseMap<const SCEV *, Value *> LoopVectorizationPlanner::executePlan(
   // 2.5 When vectorizing the epilogue, fix reduction and induction resume
   // values from the additional bypass block.
   if (VectorizingEpilogue) {
+    assert(!ILV.Legal->hasUncountableEarlyExit() &&
+           "Epilogue vectorisation not yet supported with early exits");
     BasicBlock *BypassBlock = ILV.getAdditionalBypassBlock();
     for (VPRecipeBase &R : *ExitVPBB) {
       fixReductionScalarResumeWhenVectorizingEpilog(
           &R, State, State.CFG.VPBB2IRBB[ExitVPBB], BypassBlock);
     }
+
     BasicBlock *PH = OrigLoop->getLoopPreheader();
     for (const auto &[IVPhi, _] : Legal->getInductionVars()) {
       auto *Inc = cast<PHINode>(IVPhi->getIncomingValueForBlock(PH));
@@ -10177,13 +10202,33 @@ bool LoopVectorizePass::processLoop(Loop *L) {
     return false;
   }
 
-  if (LVL.hasUncountableEarlyExit() && !EnableEarlyExitVectorization) {
-    reportVectorizationFailure("Auto-vectorization of loops with uncountable "
-                               "early exit is not enabled",
-                               "Auto-vectorization of loops with uncountable "
-                               "early exit is not enabled",
-                               "UncountableEarlyExitLoopsDisabled", ORE, L);
-    return false;
+  if (LVL.hasUncountableEarlyExit()) {
+    if (!EnableEarlyExitVectorization) {
+      reportVectorizationFailure("Auto-vectorization of loops with uncountable "
+                                 "early exit is not enabled",
+                                 "Auto-vectorization of loops with uncountable "
+                                 "early exit is not enabled",
+                                 "UncountableEarlyExitLoopsDisabled", ORE, L);
+      return false;
+    }
+
+    // Needed to prevent InnerLoopVectorizer::fixupIVUsers from crashing.
+    for (BasicBlock *BB : L->blocks()) {
+      for (Instruction &I : *BB) {
+        for (User *U : I.users()) {
+          Instruction *UI = cast<Instruction>(U);
+          if (!L->contains(UI)) {
+            reportVectorizationFailure(
+                "Auto-vectorization of loops with uncountable "
+                "early exit and live-outs is not yet supported",
+                "Auto-vectorization of loop with uncountable "
+                "early exit and live-outs is not yet supported",
+                "UncountableEarlyExitLoopLiveOutsUnsupported", ORE, L);
+            return false;
+          }
+        }
+      }
+    }
   }
 
   // Entrance to the VPlan-native vectorization path. Outer loops are processed
@@ -10208,6 +10253,20 @@ bool LoopVectorizePass::processLoop(Loop *L) {
   if (UseInterleaved)
     IAI.analyzeInterleaving(useMaskedInterleavedAccesses(*TTI));
 
+  if (LVL.hasUncountableEarlyExit()) {
+    BasicBlock *LoopLatch = L->getLoopLatch();
+    if (IAI.requiresScalarEpilogue() ||
+        llvm::any_of(LVL.getCountableExitingBlocks(),
+                     [LoopLatch](BasicBlock *BB) { return BB != LoopLatch; })) {
+      reportVectorizationFailure("Auto-vectorization of early exit loops "
+                                 "requiring a scalar epilogue is unsupported",
+                                 "Auto-vectorization of early exit loops "
+                                 "requiring a scalar epilogue is unsupported",
+                                 "UncountableEarlyExitUnsupported", ORE, L);
+      return false;
+    }
+  }
+
   // Check the function attributes and profiles to find out if this function
   // should be optimized for size.
   ScalarEpilogueLowering SEL =

llvm/test/Transforms/LoopVectorize/AArch64/simple_early_exit.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -S < %s -p loop-vectorize | FileCheck %s --check-prefixes=CHECK
+; RUN: opt -S < %s -p loop-vectorize -enable-early-exit-vectorization | FileCheck %s --check-prefixes=CHECK
 
 target triple = "aarch64-unknown-linux-gnu"
 
@@ -272,22 +272,66 @@ define i32 @diff_exit_block_needs_scev_check(i32 %end) {
 ; CHECK-NEXT:    call void @init_mem(ptr [[P1]], i64 1024)
 ; CHECK-NEXT:    call void @init_mem(ptr [[P2]], i64 1024)
 ; CHECK-NEXT:    [[END_CLAMPED:%.*]] = and i32 [[END]], 1023
+; CHECK-NEXT:    [[TMP19:%.*]] = trunc i32 [[END]] to i10
+; CHECK-NEXT:    [[TMP20:%.*]] = zext i10 [[TMP19]] to i64
+; CHECK-NEXT:    [[UMAX1:%.*]] = call i64 @llvm.umax.i64(i64 [[TMP20]], i64 1)
+; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[UMAX1]], 12
+; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_SCEVCHECK:%.*]]
+; CHECK:       vector.scevcheck:
+; CHECK-NEXT:    [[UMAX:%.*]] = call i32 @llvm.umax.i32(i32 [[END_CLAMPED]], i32 1)
+; CHECK-NEXT:    [[TMP2:%.*]] = add nsw i32 [[UMAX]], -1
+; CHECK-NEXT:    [[TMP3:%.*]] = trunc i32 [[TMP2]] to i8
+; CHECK-NEXT:    [[TMP4:%.*]] = add i8 1, [[TMP3]]
+; CHECK-NEXT:    [[TMP5:%.*]] = icmp ult i8 [[TMP4]], 1
+; CHECK-NEXT:    [[TMP6:%.*]] = icmp ugt i32 [[TMP2]], 255
+; CHECK-NEXT:    [[TMP7:%.*]] = or i1 [[TMP5]], [[TMP6]]
+; CHECK-NEXT:    br i1 [[TMP7]], label [[SCALAR_PH]], label [[VECTOR_PH:%.*]]
+; CHECK:       vector.ph:
+; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[UMAX1]], 4
+; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[UMAX1]], [[N_MOD_VF]]
+; CHECK-NEXT:    [[IND_END:%.*]] = trunc i64 [[N_VEC]] to i8
 ; CHECK-NEXT:    br label [[FOR_BODY1:%.*]]
+; CHECK:       vector.body:
+; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[FOR_BODY1]] ]
+; CHECK-NEXT:    [[TMP8:%.*]] = add i64 [[INDEX]], 0
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds i32, ptr [[P1]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds i32, ptr [[TMP9]], i32 0
+; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <4 x i32>, ptr [[TMP10]], align 4
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[P2]], i64 [[TMP8]]
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds i32, ptr [[TMP11]], i32 0
+; CHECK-NEXT:    [[WIDE_LOAD3:%.*]] = load <4 x i32>, ptr [[TMP12]], align 4
+; CHECK-NEXT:    [[TMP13:%.*]] = icmp eq <4 x i32> [[WIDE_LOAD]], [[WIDE_LOAD3]]
+; CHECK-NEXT:    [[TMP14:%.*]] = xor <4 x i1> [[TMP13]], splat (i1 true)
+; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 4
+; CHECK-NEXT:    [[TMP15:%.*]] = xor <4 x i1> [[TMP14]], splat (i1 true)
+; CHECK-NEXT:    [[TMP16:%.*]] = call i1 @llvm.vector.reduce.or.v4i1(<4 x i1> [[TMP15]])
+; CHECK-NEXT:    [[TMP17:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
+; CHECK-NEXT:    [[TMP18:%.*]] = or i1 [[TMP16]], [[TMP17]]
+; CHECK-NEXT:    br i1 [[TMP18]], label [[MIDDLE_SPLIT:%.*]], label [[FOR_BODY1]], !llvm.loop [[LOOP0:![0-9]+]]
+; CHECK:       middle.split:
+; CHECK-NEXT:    br i1 [[TMP16]], label [[FOUND:%.*]], label [[MIDDLE_BLOCK:%.*]]
+; CHECK:       middle.block:
+; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[UMAX1]], [[N_VEC]]
+; CHECK-NEXT:    br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
+; CHECK:       scalar.ph:
+; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i8 [ [[IND_END]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ], [ 0, [[VECTOR_SCEVCHECK]] ]
+; CHECK-NEXT:    [[BC_RESUME_VAL2:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ], [ 0, [[VECTOR_SCEVCHECK]] ]
+; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
-; CHECK-NEXT:    [[IND:%.*]] = phi i8 [ [[IND_NEXT:%.*]], [[FOR_INC:%.*]] ], [ 0, [[ENTRY:%.*]] ]
-; CHECK-NEXT:    [[GEP_IND:%.*]] = phi i64 [ [[GEP_IND_NEXT:%.*]], [[FOR_INC]] ], [ 0, [[ENTRY]] ]
+; CHECK-NEXT:    [[IND:%.*]] = phi i8 [ [[IND_NEXT:%.*]], [[FOR_INC:%.*]] ], [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ]
+; CHECK-NEXT:    [[GEP_IND:%.*]] = phi i64 [ [[GEP_IND_NEXT:%.*]], [[FOR_INC]] ], [ [[BC_RESUME_VAL2]], [[SCALAR_PH]] ]
 ; CHECK-NEXT:    [[ARRAYIDX1:%.*]] = getelementptr inbounds i32, ptr [[P1]], i64 [[GEP_IND]]
 ; CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARRAYIDX1]], align 4
 ; CHECK-NEXT:    [[ARRAYIDX2:%.*]] = getelementptr inbounds i32, ptr [[P2]], i64 [[GEP_IND]]
 ; CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[ARRAYIDX2]], align 4
 ; CHECK-NEXT:    [[CMP_EARLY:%.*]] = icmp eq i32 [[TMP0]], [[TMP1]]
-; CHECK-NEXT:    br i1 [[CMP_EARLY]], label [[FOUND:%.*]], label [[FOR_INC]]
+; CHECK-NEXT:    br i1 [[CMP_EARLY]], label [[FOUND]], label [[FOR_INC]]
 ; CHECK:       for.inc:
 ; CHECK-NEXT:    [[IND_NEXT]] = add i8 [[IND]], 1
 ; CHECK-NEXT:    [[CONV:%.*]] = zext i8 [[IND_NEXT]] to i32
 ; CHECK-NEXT:    [[GEP_IND_NEXT]] = add i64 [[GEP_IND]], 1
 ; CHECK-NEXT:    [[CMP:%.*]] = icmp ult i32 [[CONV]], [[END_CLAMPED]]
-; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY1]], label [[EXIT:%.*]]
+; CHECK-NEXT:    br i1 [[CMP]], label [[FOR_BODY]], label [[EXIT]], !llvm.loop [[LOOP3:![0-9]+]]
 ; CHECK:       found:
 ; CHECK-NEXT:    ret i32 1
 ; CHECK:       exit:
@@ -325,9 +369,39 @@ exit:
   ret i32 0
 }
 
+%my.struct = type { i8, i8 }
+
+define i64 @same_exit_block_requires_interleaving() {
+entry:
+  %p1 = alloca [128 x %my.struct]
+  call void @init_mem(ptr %p1, i64 256)
+  br label %loop
+
+loop:
+  %index = phi i64 [ %index.next, %loop.latch ], [ 3, %entry ]
+  %arrayidx = getelementptr inbounds [128 x %my.struct], ptr %p1, i64 0, i64 %index
+  %ld1 = load i8, ptr %arrayidx, align 1
+  %cmp3 = icmp eq i8 %ld1, 3
+  br i1 %cmp3, label %loop.latch, label %loop.end
+
+loop.latch:
+  %index.next = add i64 %index, 1
+  %exitcond = icmp ne i64 %index.next, 69
+  br i1 %exitcond, label %loop, label %loop.end
+
+loop.end:
+  %retval = phi i64 [ 0, %loop.latch ], [ 1, %loop ]
+  ret i64 %retval
+}
 
 declare i32 @foo(i32) readonly
 declare <vscale x 4 x i32> @foo_vec(<vscale x 4 x i32>)
 
 attributes #0 = { "vector-function-abi-variant"="_ZGVsNxv_foo(foo_vec)" }
 attributes #1 = { "target-features"="+sve" vscale_range(1,16) }
+;.
+; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}
+; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}
+; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}
+; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]]}
+;.

llvm/test/Transforms/LoopVectorize/early_exit_legality.ll

@@ -49,7 +49,7 @@ define i64 @same_exit_block_pre_inc_use1() {
 ; CHECK-LABEL: LV: Checking a loop in 'same_exit_block_pre_inc_use1'
 ; CHECK:       LV: Found an early exit loop with symbolic max backedge taken count: 63
 ; CHECK-NEXT:  LV: We can vectorize this loop!
-; CHECK-NOT:   LV: Not vectorizing
+; CHECK-NOT:   LV: Auto-vectorization of loops with uncountable early exit and live-outs is not yet supported.
 entry:
   %p1 = alloca [1024 x i8]
   %p2 = alloca [1024 x i8]
@@ -141,7 +141,7 @@ define i64 @loop_contains_load_after_early_exit(ptr dereferenceable(1024) align(
 ; CHECK-LABEL: LV: Checking a loop in 'loop_contains_load_after_early_exit'
 ; CHECK:       LV: Found an early exit loop with symbolic max backedge taken count: 63
 ; CHECK-NEXT:  LV: We can vectorize this loop!
-; CHECK:       LV: Not vectorizing: Some exit values in loop with uncountable exit not supported yet.
+; CHECK:       LV: Not vectorizing: Auto-vectorization of loops with uncountable early exit and live-outs is not yet supported.
 entry:
   %p1 = alloca [1024 x i8]
   call void @init_mem(ptr %p1, i64 1024)

llvm/test/Transforms/LoopVectorize/multi_early_exit.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -S < %s -p loop-vectorize | FileCheck %s
+; RUN: opt -S < %s -p loop-vectorize -enable-early-exit-vectorization | FileCheck %s
 
 declare void @init_mem(ptr, i64);
 

llvm/test/Transforms/LoopVectorize/multi_early_exit_live_outs.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -S < %s -p loop-vectorize | FileCheck %s
+; RUN: opt -S < %s -p loop-vectorize -enable-early-exit-vectorization | FileCheck %s
 
 declare void @init_mem(ptr, i64);
 

llvm/test/Transforms/LoopVectorize/single_early_exit_live_outs.ll

@@ -1,5 +1,5 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 4
-; RUN: opt -S < %s -p loop-vectorize | FileCheck %s
+; RUN: opt -S < %s -p loop-vectorize -enable-early-exit-vectorization | FileCheck %s
 
 declare void @init_mem(ptr, i64);