[LV] Fix recording of BranchTakenCount for FoldTail
authorAyal Zaks <ayal.zaks@intel.com>
Sat, 25 Apr 2020 00:44:38 +0000 (03:44 +0300)
committerAyal Zaks <ayal.zaks@intel.com>
Sun, 26 Apr 2020 17:13:10 +0000 (20:13 +0300)
When folding tail, branch taken count is computed during initial VPlan execution
and recorded to be used by the compare computing the loop's mask. This recording
should directly set the State, instead of reusing Value2VPValue mapping which
serves original Values present prior to vectorization.
The branch taken count may be a constant Value, which may be used elsewhere in
the loop; trying to employ Value2VPValue for both leads to the issue reported in
https://reviews.llvm.org/D76992#inline-721028

Differential Revision: https://reviews.llvm.org/D78847

llvm/lib/Transforms/Vectorize/VPlan.cpp
llvm/lib/Transforms/Vectorize/VPlan.h
llvm/test/Transforms/LoopVectorize/X86/x86-interleaved-accesses-masked-group.ll
llvm/test/Transforms/LoopVectorize/tail-folding-counting-down.ll

index b094a6f..1a38090 100644 (file)
@@ -442,7 +442,9 @@ void VPlan::execute(VPTransformState *State) {
     IRBuilder<> Builder(State->CFG.PrevBB->getTerminator());
     auto *TCMO = Builder.CreateSub(TC, ConstantInt::get(TC->getType(), 1),
                                    "trip.count.minus.1");
-    Value2VPValue[TCMO] = BackedgeTakenCount;
+    Value *VTCMO = Builder.CreateVectorSplat(State->VF, TCMO, "broadcast");
+    for (unsigned Part = 0, UF = State->UF; Part < UF; ++Part)
+      State->set(BackedgeTakenCount, VTCMO, Part);
   }
 
   // 0. Set the reverse mapping from VPValues to Values for code generation.
index 25afb09..1a112d3 100644 (file)
@@ -1544,10 +1544,9 @@ public:
     if (Entry)
       VPBlockBase::deleteCFG(Entry);
     for (auto &MapEntry : Value2VPValue)
-      if (MapEntry.second != BackedgeTakenCount)
-        delete MapEntry.second;
+      delete MapEntry.second;
     if (BackedgeTakenCount)
-      delete BackedgeTakenCount; // Delete once, if in Value2VPValue or not.
+      delete BackedgeTakenCount;
     for (VPValue *Def : VPExternalDefs)
       delete Def;
     for (VPValue *CBV : VPCBVs)
index 27250dd..ca468b7 100644 (file)
@@ -202,10 +202,10 @@ for.end:
 ; ENABLED_MASKED_STRIDED-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 7
 ; ENABLED_MASKED_STRIDED-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -8
 ; ENABLED_MASKED_STRIDED-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = add i32 [[N]], -1
-; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> undef, i32 [[CONV]], i32 0
-; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <8 x i32> undef, i32 [[TRIP_COUNT_MINUS_1]], i32 0
 ; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT1]], <8 x i32> undef, <8 x i32> zeroinitializer
+; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> undef, i32 [[CONV]], i32 0
+; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; ENABLED_MASKED_STRIDED-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; ENABLED_MASKED_STRIDED:       vector.body:
 ; ENABLED_MASKED_STRIDED-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
@@ -300,10 +300,10 @@ for.end:
 ; ENABLED_MASKED_STRIDED-NEXT:    [[N_RND_UP:%.*]] = add i32 [[N]], 7
 ; ENABLED_MASKED_STRIDED-NEXT:    [[N_VEC:%.*]] = and i32 [[N_RND_UP]], -8
 ; ENABLED_MASKED_STRIDED-NEXT:    [[TRIP_COUNT_MINUS_1:%.*]] = add i32 [[N]], -1
-; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> undef, i32 [[CONV]], i32 0
-; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT1:%.*]] = insertelement <8 x i32> undef, i32 [[TRIP_COUNT_MINUS_1]], i32 0
 ; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT2:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT1]], <8 x i32> undef, <8 x i32> zeroinitializer
+; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i32> undef, i32 [[CONV]], i32 0
+; ENABLED_MASKED_STRIDED-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT]], <8 x i32> undef, <8 x i32> zeroinitializer
 ; ENABLED_MASKED_STRIDED-NEXT:    br label [[VECTOR_BODY:%.*]]
 ; ENABLED_MASKED_STRIDED:       vector.body:
 ; ENABLED_MASKED_STRIDED-NEXT:    [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
index b51cb56..a090b15 100644 (file)
@@ -39,3 +39,36 @@ while.end.loopexit:
 while.end:
   ret void
 }
+
+; Make sure a loop is successfully vectorized with fold-tail when the backedge
+; taken count is constant and used inside the loop. Issue revealed by D76992.
+;
+define void @reuse_const_btc(i8* %A) optsize {
+; CHECK-LABEL: @reuse_const_btc
+; CHECK: {{%.*}} = icmp ule <4 x i32> {{%.*}}, <i32 13, i32 13, i32 13, i32 13>
+; CHECK: {{%.*}} = select <4 x i1> {{%.*}}, <4 x i32> <i32 12, i32 12, i32 12, i32 12>, <4 x i32> <i32 13, i32 13, i32 13, i32 13>
+;
+entry:
+  br label %loop
+
+loop:
+  %riv = phi i32 [ 13, %entry ], [ %rivMinus1, %merge ]
+  %sub = sub nuw nsw i32 20, %riv
+  %arrayidx = getelementptr inbounds i8, i8* %A, i32 %sub
+  %cond0 = icmp eq i32 %riv, 7
+  br i1 %cond0, label %then, label %else
+then:
+  br label %merge
+else:
+  br label %merge
+merge:
+  %blend = phi i32 [ 13, %then ], [ 12, %else ]
+  %trunc = trunc i32 %blend to i8
+  store i8 %trunc, i8* %arrayidx, align 1
+  %rivMinus1 = add nuw nsw i32 %riv, -1
+  %cond = icmp eq i32 %riv, 0
+  br i1 %cond, label %exit, label %loop
+
+exit:
+  ret void
+}