llvm/test/Transforms/LoopVectorize/X86/x86-predication.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
   2 ; RUN: opt < %s -mattr=avx -force-vector-width=2 -force-vector-interleave=1 -passes=loop-vectorize,simplifycfg -simplifycfg-require-and-preserve-domtree=1 -S | FileCheck %s
   3 ; RUN: opt -mcpu=skylake-avx512 -S -force-vector-width=8 -force-vector-interleave=1 -passes=loop-vectorize < %s | FileCheck %s --check-prefix=SINK-GATHER
   4
   5 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
   6 target triple = "x86_64-apple-macosx10.8.0"
   7
   8 ; This test ensures that we don't scalarize the predicated load. Since the load
   9 ; can be vectorized with predication, scalarizing it would cause its pointer
  10 ; operand to become non-uniform.
  11 ;
  12 define i32 @predicated_sdiv_masked_load(ptr %a, ptr %b, i32 %x, i1 %c) {
  13 ; CHECK-LABEL: @predicated_sdiv_masked_load(
  14 ; CHECK-NEXT:  entry:
  15 ; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i1> poison, i1 [[C:%.*]], i64 0
  16 ; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i1> [[BROADCAST_SPLATINSERT]], <2 x i1> poison, <2 x i32> zeroinitializer
  17 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
  18 ; CHECK:       vector.body:
  19 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[PRED_SDIV_CONTINUE2:%.*]] ]
  20 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <2 x i32> [ zeroinitializer, [[ENTRY]] ], [ [[TMP17:%.*]], [[PRED_SDIV_CONTINUE2]] ]
  21 ; CHECK-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
  22 ; CHECK-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP0]]
  23 ; CHECK-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
  24 ; CHECK-NEXT:    [[WIDE_LOAD:%.*]] = load <2 x i32>, ptr [[TMP2]], align 4
  25 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr i32, ptr [[B:%.*]], i64 [[TMP0]]
  26 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr i32, ptr [[TMP3]], i32 0
  27 ; CHECK-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <2 x i32> @llvm.masked.load.v2i32.p0(ptr [[TMP4]], i32 4, <2 x i1> [[BROADCAST_SPLAT]], <2 x i32> poison)
  28 ; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <2 x i1> [[BROADCAST_SPLAT]], i32 0
  29 ; CHECK-NEXT:    br i1 [[TMP5]], label [[PRED_SDIV_IF:%.*]], label [[PRED_SDIV_CONTINUE:%.*]]
  30 ; CHECK:       pred.sdiv.if:
  31 ; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x i32> [[WIDE_MASKED_LOAD]], i32 0
  32 ; CHECK-NEXT:    [[TMP7:%.*]] = sdiv i32 [[TMP6]], [[X:%.*]]
  33 ; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <2 x i32> poison, i32 [[TMP7]], i32 0
  34 ; CHECK-NEXT:    br label [[PRED_SDIV_CONTINUE]]
  35 ; CHECK:       pred.sdiv.continue:
  36 ; CHECK-NEXT:    [[TMP9:%.*]] = phi <2 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP8]], [[PRED_SDIV_IF]] ]
  37 ; CHECK-NEXT:    [[TMP10:%.*]] = extractelement <2 x i1> [[BROADCAST_SPLAT]], i32 1
  38 ; CHECK-NEXT:    br i1 [[TMP10]], label [[PRED_SDIV_IF1:%.*]], label [[PRED_SDIV_CONTINUE2]]
  39 ; CHECK:       pred.sdiv.if1:
  40 ; CHECK-NEXT:    [[TMP11:%.*]] = extractelement <2 x i32> [[WIDE_MASKED_LOAD]], i32 1
  41 ; CHECK-NEXT:    [[TMP12:%.*]] = sdiv i32 [[TMP11]], [[X]]
  42 ; CHECK-NEXT:    [[TMP13:%.*]] = insertelement <2 x i32> [[TMP9]], i32 [[TMP12]], i32 1
  43 ; CHECK-NEXT:    br label [[PRED_SDIV_CONTINUE2]]
  44 ; CHECK:       pred.sdiv.continue2:
  45 ; CHECK-NEXT:    [[TMP14:%.*]] = phi <2 x i32> [ [[TMP9]], [[PRED_SDIV_CONTINUE]] ], [ [[TMP13]], [[PRED_SDIV_IF1]] ]
  46 ; CHECK-NEXT:    [[TMP15:%.*]] = add nsw <2 x i32> [[TMP14]], [[WIDE_LOAD]]
  47 ; CHECK-NEXT:    [[PREDPHI:%.*]] = select <2 x i1> [[BROADCAST_SPLAT]], <2 x i32> [[TMP15]], <2 x i32> [[WIDE_LOAD]]
  48 ; CHECK-NEXT:    [[TMP17]] = add <2 x i32> [[VEC_PHI]], [[PREDPHI]]
  49 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
  50 ; CHECK-NEXT:    [[TMP18:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10000
  51 ; CHECK-NEXT:    br i1 [[TMP18]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
  52 ; CHECK:       middle.block:
  53 ; CHECK-NEXT:    [[TMP19:%.*]] = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> [[TMP17]])
  54 ; CHECK-NEXT:    ret i32 [[TMP19]]
  55 ;
  56 ; SINK-GATHER-LABEL: @predicated_sdiv_masked_load(
  57 ; SINK-GATHER-NEXT:  entry:
  58 ; SINK-GATHER-NEXT:    br i1 false, label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
  59 ; SINK-GATHER:       vector.ph:
  60 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i1> poison, i1 [[C:%.*]], i64 0
  61 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i1> [[BROADCAST_SPLATINSERT]], <8 x i1> poison, <8 x i32> zeroinitializer
  62 ; SINK-GATHER-NEXT:    br label [[VECTOR_BODY:%.*]]
  63 ; SINK-GATHER:       vector.body:
  64 ; SINK-GATHER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_SDIV_CONTINUE14:%.*]] ]
  65 ; SINK-GATHER-NEXT:    [[VEC_PHI:%.*]] = phi <8 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP47:%.*]], [[PRED_SDIV_CONTINUE14]] ]
  66 ; SINK-GATHER-NEXT:    [[TMP0:%.*]] = add i64 [[INDEX]], 0
  67 ; SINK-GATHER-NEXT:    [[TMP1:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP0]]
  68 ; SINK-GATHER-NEXT:    [[TMP2:%.*]] = getelementptr inbounds i32, ptr [[TMP1]], i32 0
  69 ; SINK-GATHER-NEXT:    [[WIDE_LOAD:%.*]] = load <8 x i32>, ptr [[TMP2]], align 4
  70 ; SINK-GATHER-NEXT:    [[TMP3:%.*]] = getelementptr i32, ptr [[B:%.*]], i64 [[TMP0]]
  71 ; SINK-GATHER-NEXT:    [[TMP4:%.*]] = getelementptr i32, ptr [[TMP3]], i32 0
  72 ; SINK-GATHER-NEXT:    [[WIDE_MASKED_LOAD:%.*]] = call <8 x i32> @llvm.masked.load.v8i32.p0(ptr [[TMP4]], i32 4, <8 x i1> [[BROADCAST_SPLAT]], <8 x i32> poison)
  73 ; SINK-GATHER-NEXT:    [[TMP5:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 0
  74 ; SINK-GATHER-NEXT:    br i1 [[TMP5]], label [[PRED_SDIV_IF:%.*]], label [[PRED_SDIV_CONTINUE:%.*]]
  75 ; SINK-GATHER:       pred.sdiv.if:
  76 ; SINK-GATHER-NEXT:    [[TMP6:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 0
  77 ; SINK-GATHER-NEXT:    [[TMP7:%.*]] = sdiv i32 [[TMP6]], [[X:%.*]]
  78 ; SINK-GATHER-NEXT:    [[TMP8:%.*]] = insertelement <8 x i32> poison, i32 [[TMP7]], i32 0
  79 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE]]
  80 ; SINK-GATHER:       pred.sdiv.continue:
  81 ; SINK-GATHER-NEXT:    [[TMP9:%.*]] = phi <8 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP8]], [[PRED_SDIV_IF]] ]
  82 ; SINK-GATHER-NEXT:    [[TMP10:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 1
  83 ; SINK-GATHER-NEXT:    br i1 [[TMP10]], label [[PRED_SDIV_IF1:%.*]], label [[PRED_SDIV_CONTINUE2:%.*]]
  84 ; SINK-GATHER:       pred.sdiv.if1:
  85 ; SINK-GATHER-NEXT:    [[TMP11:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 1
  86 ; SINK-GATHER-NEXT:    [[TMP12:%.*]] = sdiv i32 [[TMP11]], [[X]]
  87 ; SINK-GATHER-NEXT:    [[TMP13:%.*]] = insertelement <8 x i32> [[TMP9]], i32 [[TMP12]], i32 1
  88 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE2]]
  89 ; SINK-GATHER:       pred.sdiv.continue2:
  90 ; SINK-GATHER-NEXT:    [[TMP14:%.*]] = phi <8 x i32> [ [[TMP9]], [[PRED_SDIV_CONTINUE]] ], [ [[TMP13]], [[PRED_SDIV_IF1]] ]
  91 ; SINK-GATHER-NEXT:    [[TMP15:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 2
  92 ; SINK-GATHER-NEXT:    br i1 [[TMP15]], label [[PRED_SDIV_IF3:%.*]], label [[PRED_SDIV_CONTINUE4:%.*]]
  93 ; SINK-GATHER:       pred.sdiv.if3:
  94 ; SINK-GATHER-NEXT:    [[TMP16:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 2
  95 ; SINK-GATHER-NEXT:    [[TMP17:%.*]] = sdiv i32 [[TMP16]], [[X]]
  96 ; SINK-GATHER-NEXT:    [[TMP18:%.*]] = insertelement <8 x i32> [[TMP14]], i32 [[TMP17]], i32 2
  97 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE4]]
  98 ; SINK-GATHER:       pred.sdiv.continue4:
  99 ; SINK-GATHER-NEXT:    [[TMP19:%.*]] = phi <8 x i32> [ [[TMP14]], [[PRED_SDIV_CONTINUE2]] ], [ [[TMP18]], [[PRED_SDIV_IF3]] ]
 100 ; SINK-GATHER-NEXT:    [[TMP20:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 3
 101 ; SINK-GATHER-NEXT:    br i1 [[TMP20]], label [[PRED_SDIV_IF5:%.*]], label [[PRED_SDIV_CONTINUE6:%.*]]
 102 ; SINK-GATHER:       pred.sdiv.if5:
 103 ; SINK-GATHER-NEXT:    [[TMP21:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 3
 104 ; SINK-GATHER-NEXT:    [[TMP22:%.*]] = sdiv i32 [[TMP21]], [[X]]
 105 ; SINK-GATHER-NEXT:    [[TMP23:%.*]] = insertelement <8 x i32> [[TMP19]], i32 [[TMP22]], i32 3
 106 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE6]]
 107 ; SINK-GATHER:       pred.sdiv.continue6:
 108 ; SINK-GATHER-NEXT:    [[TMP24:%.*]] = phi <8 x i32> [ [[TMP19]], [[PRED_SDIV_CONTINUE4]] ], [ [[TMP23]], [[PRED_SDIV_IF5]] ]
 109 ; SINK-GATHER-NEXT:    [[TMP25:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 4
 110 ; SINK-GATHER-NEXT:    br i1 [[TMP25]], label [[PRED_SDIV_IF7:%.*]], label [[PRED_SDIV_CONTINUE8:%.*]]
 111 ; SINK-GATHER:       pred.sdiv.if7:
 112 ; SINK-GATHER-NEXT:    [[TMP26:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 4
 113 ; SINK-GATHER-NEXT:    [[TMP27:%.*]] = sdiv i32 [[TMP26]], [[X]]
 114 ; SINK-GATHER-NEXT:    [[TMP28:%.*]] = insertelement <8 x i32> [[TMP24]], i32 [[TMP27]], i32 4
 115 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE8]]
 116 ; SINK-GATHER:       pred.sdiv.continue8:
 117 ; SINK-GATHER-NEXT:    [[TMP29:%.*]] = phi <8 x i32> [ [[TMP24]], [[PRED_SDIV_CONTINUE6]] ], [ [[TMP28]], [[PRED_SDIV_IF7]] ]
 118 ; SINK-GATHER-NEXT:    [[TMP30:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 5
 119 ; SINK-GATHER-NEXT:    br i1 [[TMP30]], label [[PRED_SDIV_IF9:%.*]], label [[PRED_SDIV_CONTINUE10:%.*]]
 120 ; SINK-GATHER:       pred.sdiv.if9:
 121 ; SINK-GATHER-NEXT:    [[TMP31:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 5
 122 ; SINK-GATHER-NEXT:    [[TMP32:%.*]] = sdiv i32 [[TMP31]], [[X]]
 123 ; SINK-GATHER-NEXT:    [[TMP33:%.*]] = insertelement <8 x i32> [[TMP29]], i32 [[TMP32]], i32 5
 124 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE10]]
 125 ; SINK-GATHER:       pred.sdiv.continue10:
 126 ; SINK-GATHER-NEXT:    [[TMP34:%.*]] = phi <8 x i32> [ [[TMP29]], [[PRED_SDIV_CONTINUE8]] ], [ [[TMP33]], [[PRED_SDIV_IF9]] ]
 127 ; SINK-GATHER-NEXT:    [[TMP35:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 6
 128 ; SINK-GATHER-NEXT:    br i1 [[TMP35]], label [[PRED_SDIV_IF11:%.*]], label [[PRED_SDIV_CONTINUE12:%.*]]
 129 ; SINK-GATHER:       pred.sdiv.if11:
 130 ; SINK-GATHER-NEXT:    [[TMP36:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 6
 131 ; SINK-GATHER-NEXT:    [[TMP37:%.*]] = sdiv i32 [[TMP36]], [[X]]
 132 ; SINK-GATHER-NEXT:    [[TMP38:%.*]] = insertelement <8 x i32> [[TMP34]], i32 [[TMP37]], i32 6
 133 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE12]]
 134 ; SINK-GATHER:       pred.sdiv.continue12:
 135 ; SINK-GATHER-NEXT:    [[TMP39:%.*]] = phi <8 x i32> [ [[TMP34]], [[PRED_SDIV_CONTINUE10]] ], [ [[TMP38]], [[PRED_SDIV_IF11]] ]
 136 ; SINK-GATHER-NEXT:    [[TMP40:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 7
 137 ; SINK-GATHER-NEXT:    br i1 [[TMP40]], label [[PRED_SDIV_IF13:%.*]], label [[PRED_SDIV_CONTINUE14]]
 138 ; SINK-GATHER:       pred.sdiv.if13:
 139 ; SINK-GATHER-NEXT:    [[TMP41:%.*]] = extractelement <8 x i32> [[WIDE_MASKED_LOAD]], i32 7
 140 ; SINK-GATHER-NEXT:    [[TMP42:%.*]] = sdiv i32 [[TMP41]], [[X]]
 141 ; SINK-GATHER-NEXT:    [[TMP43:%.*]] = insertelement <8 x i32> [[TMP39]], i32 [[TMP42]], i32 7
 142 ; SINK-GATHER-NEXT:    br label [[PRED_SDIV_CONTINUE14]]
 143 ; SINK-GATHER:       pred.sdiv.continue14:
 144 ; SINK-GATHER-NEXT:    [[TMP44:%.*]] = phi <8 x i32> [ [[TMP39]], [[PRED_SDIV_CONTINUE12]] ], [ [[TMP43]], [[PRED_SDIV_IF13]] ]
 145 ; SINK-GATHER-NEXT:    [[TMP45:%.*]] = add nsw <8 x i32> [[TMP44]], [[WIDE_LOAD]]
 146 ; SINK-GATHER-NEXT:    [[PREDPHI:%.*]] = select <8 x i1> [[BROADCAST_SPLAT]], <8 x i32> [[TMP45]], <8 x i32> [[WIDE_LOAD]]
 147 ; SINK-GATHER-NEXT:    [[TMP47]] = add <8 x i32> [[VEC_PHI]], [[PREDPHI]]
 148 ; SINK-GATHER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
 149 ; SINK-GATHER-NEXT:    [[TMP48:%.*]] = icmp eq i64 [[INDEX_NEXT]], 10000
 150 ; SINK-GATHER-NEXT:    br i1 [[TMP48]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
 151 ; SINK-GATHER:       middle.block:
 152 ; SINK-GATHER-NEXT:    [[TMP49:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP47]])
 153 ; SINK-GATHER-NEXT:    br i1 true, label [[FOR_END:%.*]], label [[SCALAR_PH]]
 154 ; SINK-GATHER:       scalar.ph:
 155 ; SINK-GATHER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ 10000, [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 156 ; SINK-GATHER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP49]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
 157 ; SINK-GATHER-NEXT:    br label [[FOR_BODY:%.*]]
 158 ; SINK-GATHER:       for.body:
 159 ; SINK-GATHER-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[FOR_INC:%.*]] ]
 160 ; SINK-GATHER-NEXT:    [[R:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[T7:%.*]], [[FOR_INC]] ]
 161 ; SINK-GATHER-NEXT:    [[T0:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I]]
 162 ; SINK-GATHER-NEXT:    [[T1:%.*]] = load i32, ptr [[T0]], align 4
 163 ; SINK-GATHER-NEXT:    br i1 [[C]], label [[IF_THEN:%.*]], label [[FOR_INC]]
 164 ; SINK-GATHER:       if.then:
 165 ; SINK-GATHER-NEXT:    [[T2:%.*]] = getelementptr inbounds i32, ptr [[B]], i64 [[I]]
 166 ; SINK-GATHER-NEXT:    [[T3:%.*]] = load i32, ptr [[T2]], align 4
 167 ; SINK-GATHER-NEXT:    [[T4:%.*]] = sdiv i32 [[T3]], [[X]]
 168 ; SINK-GATHER-NEXT:    [[T5:%.*]] = add nsw i32 [[T4]], [[T1]]
 169 ; SINK-GATHER-NEXT:    br label [[FOR_INC]]
 170 ; SINK-GATHER:       for.inc:
 171 ; SINK-GATHER-NEXT:    [[T6:%.*]] = phi i32 [ [[T1]], [[FOR_BODY]] ], [ [[T5]], [[IF_THEN]] ]
 172 ; SINK-GATHER-NEXT:    [[T7]] = add i32 [[R]], [[T6]]
 173 ; SINK-GATHER-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
 174 ; SINK-GATHER-NEXT:    [[COND:%.*]] = icmp eq i64 [[I_NEXT]], 10000
 175 ; SINK-GATHER-NEXT:    br i1 [[COND]], label [[FOR_END]], label [[FOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
 176 ; SINK-GATHER:       for.end:
 177 ; SINK-GATHER-NEXT:    [[T8:%.*]] = phi i32 [ [[T7]], [[FOR_INC]] ], [ [[TMP49]], [[MIDDLE_BLOCK]] ]
 178 ; SINK-GATHER-NEXT:    ret i32 [[T8]]
 179 ;
 180 entry:
 181   br label %for.body
 182
 183 for.body:
 184   %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
 185   %r = phi i32 [ 0, %entry ], [ %t7, %for.inc ]
 186   %t0 = getelementptr inbounds i32, ptr %a, i64 %i
 187   %t1 = load i32, ptr %t0, align 4
 188   br i1 %c, label %if.then, label %for.inc
 189
 190 if.then:
 191   %t2 = getelementptr inbounds i32, ptr %b, i64 %i
 192   %t3 = load i32, ptr %t2, align 4
 193   %t4 = sdiv i32 %t3, %x
 194   %t5 = add nsw i32 %t4, %t1
 195   br label %for.inc
 196
 197 for.inc:
 198   %t6 = phi i32 [ %t1, %for.body ], [ %t5, %if.then]
 199   %t7 = add i32 %r, %t6
 200   %i.next = add nuw nsw i64 %i, 1
 201   %cond = icmp eq i64 %i.next, 10000
 202   br i1 %cond, label %for.end, label %for.body
 203
 204 for.end:
 205   %t8 = phi i32 [ %t7, %for.inc ]
 206   ret i32 %t8
 207 }
 208
 209 ; This test ensures that a load, which would have been widened otherwise is
 210 ; instead scalarized if Cost-Model so decided as part of its
 211 ; sink-scalar-operands optimization for predicated instructions.
 212 define i32 @scalarize_and_sink_gather(ptr %a, i1 %c, i32 %x, i64 %n) {
 213 ; CHECK-LABEL: @scalarize_and_sink_gather(
 214 ; CHECK-NEXT:  entry:
 215 ; CHECK-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
 216 ; CHECK-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[SMAX]], 2
 217 ; CHECK-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 218 ; CHECK:       vector.ph:
 219 ; CHECK-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[SMAX]], 2
 220 ; CHECK-NEXT:    [[N_VEC:%.*]] = sub i64 [[SMAX]], [[N_MOD_VF]]
 221 ; CHECK-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <2 x i1> poison, i1 [[C:%.*]], i64 0
 222 ; CHECK-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <2 x i1> [[BROADCAST_SPLATINSERT]], <2 x i1> poison, <2 x i32> zeroinitializer
 223 ; CHECK-NEXT:    [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <2 x i32> poison, i32 [[X:%.*]], i64 0
 224 ; CHECK-NEXT:    [[BROADCAST_SPLAT4:%.*]] = shufflevector <2 x i32> [[BROADCAST_SPLATINSERT3]], <2 x i32> poison, <2 x i32> zeroinitializer
 225 ; CHECK-NEXT:    br label [[VECTOR_BODY:%.*]]
 226 ; CHECK:       vector.body:
 227 ; CHECK-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_UDIV_CONTINUE2:%.*]] ]
 228 ; CHECK-NEXT:    [[VEC_IND:%.*]] = phi <2 x i64> [ <i64 0, i64 1>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[PRED_UDIV_CONTINUE2]] ]
 229 ; CHECK-NEXT:    [[VEC_PHI:%.*]] = phi <2 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP18:%.*]], [[PRED_UDIV_CONTINUE2]] ]
 230 ; CHECK-NEXT:    [[TMP0:%.*]] = mul <2 x i64> [[VEC_IND]], <i64 777, i64 777>
 231 ; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x i1> [[BROADCAST_SPLAT]], i32 0
 232 ; CHECK-NEXT:    br i1 [[TMP1]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
 233 ; CHECK:       pred.udiv.if:
 234 ; CHECK-NEXT:    [[TMP2:%.*]] = extractelement <2 x i64> [[TMP0]], i32 0
 235 ; CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP2]]
 236 ; CHECK-NEXT:    [[TMP4:%.*]] = load i32, ptr [[TMP3]], align 4
 237 ; CHECK-NEXT:    [[TMP5:%.*]] = udiv i32 [[TMP4]], [[X]]
 238 ; CHECK-NEXT:    [[TMP6:%.*]] = insertelement <2 x i32> poison, i32 [[TMP5]], i32 0
 239 ; CHECK-NEXT:    br label [[PRED_UDIV_CONTINUE]]
 240 ; CHECK:       pred.udiv.continue:
 241 ; CHECK-NEXT:    [[TMP8:%.*]] = phi <2 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_UDIV_IF]] ]
 242 ; CHECK-NEXT:    [[TMP9:%.*]] = extractelement <2 x i1> [[BROADCAST_SPLAT]], i32 1
 243 ; CHECK-NEXT:    br i1 [[TMP9]], label [[PRED_UDIV_IF1:%.*]], label [[PRED_UDIV_CONTINUE2]]
 244 ; CHECK:       pred.udiv.if1:
 245 ; CHECK-NEXT:    [[TMP10:%.*]] = extractelement <2 x i64> [[TMP0]], i32 1
 246 ; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP10]]
 247 ; CHECK-NEXT:    [[TMP12:%.*]] = load i32, ptr [[TMP11]], align 4
 248 ; CHECK-NEXT:    [[TMP13:%.*]] = udiv i32 [[TMP12]], [[X]]
 249 ; CHECK-NEXT:    [[TMP14:%.*]] = insertelement <2 x i32> [[TMP8]], i32 [[TMP13]], i32 1
 250 ; CHECK-NEXT:    br label [[PRED_UDIV_CONTINUE2]]
 251 ; CHECK:       pred.udiv.continue2:
 252 ; CHECK-NEXT:    [[TMP16:%.*]] = phi <2 x i32> [ [[TMP8]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP14]], [[PRED_UDIV_IF1]] ]
 253 ; CHECK-NEXT:    [[PREDPHI:%.*]] = select <2 x i1> [[BROADCAST_SPLAT]], <2 x i32> [[TMP16]], <2 x i32> [[BROADCAST_SPLAT4]]
 254 ; CHECK-NEXT:    [[TMP18]] = add <2 x i32> [[VEC_PHI]], [[PREDPHI]]
 255 ; CHECK-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
 256 ; CHECK-NEXT:    [[VEC_IND_NEXT]] = add <2 x i64> [[VEC_IND]], <i64 2, i64 2>
 257 ; CHECK-NEXT:    [[TMP19:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 258 ; CHECK-NEXT:    br i1 [[TMP19]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]
 259 ; CHECK:       middle.block:
 260 ; CHECK-NEXT:    [[TMP20:%.*]] = call i32 @llvm.vector.reduce.add.v2i32(<2 x i32> [[TMP18]])
 261 ; CHECK-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[SMAX]], [[N_VEC]]
 262 ; CHECK-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
 263 ; CHECK:       scalar.ph:
 264 ; CHECK-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 265 ; CHECK-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP20]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
 266 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 267 ; CHECK:       for.body:
 268 ; CHECK-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[FOR_INC:%.*]] ]
 269 ; CHECK-NEXT:    [[R:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[T6:%.*]], [[FOR_INC]] ]
 270 ; CHECK-NEXT:    [[I7:%.*]] = mul i64 [[I]], 777
 271 ; CHECK-NEXT:    br i1 [[C]], label [[IF_THEN:%.*]], label [[FOR_INC]]
 272 ; CHECK:       if.then:
 273 ; CHECK-NEXT:    [[T0:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I7]]
 274 ; CHECK-NEXT:    [[T2:%.*]] = load i32, ptr [[T0]], align 4
 275 ; CHECK-NEXT:    [[T4:%.*]] = udiv i32 [[T2]], [[X]]
 276 ; CHECK-NEXT:    br label [[FOR_INC]]
 277 ; CHECK:       for.inc:
 278 ; CHECK-NEXT:    [[T5:%.*]] = phi i32 [ [[X]], [[FOR_BODY]] ], [ [[T4]], [[IF_THEN]] ]
 279 ; CHECK-NEXT:    [[T6]] = add i32 [[R]], [[T5]]
 280 ; CHECK-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
 281 ; CHECK-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
 282 ; CHECK-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP4:![0-9]+]]
 283 ; CHECK:       for.end:
 284 ; CHECK-NEXT:    [[T7:%.*]] = phi i32 [ [[T6]], [[FOR_INC]] ], [ [[TMP20]], [[MIDDLE_BLOCK]] ]
 285 ; CHECK-NEXT:    ret i32 [[T7]]
 286 ;
 287 ; SINK-GATHER-LABEL: @scalarize_and_sink_gather(
 288 ; SINK-GATHER-NEXT:  entry:
 289 ; SINK-GATHER-NEXT:    [[SMAX:%.*]] = call i64 @llvm.smax.i64(i64 [[N:%.*]], i64 1)
 290 ; SINK-GATHER-NEXT:    [[MIN_ITERS_CHECK:%.*]] = icmp ult i64 [[SMAX]], 8
 291 ; SINK-GATHER-NEXT:    br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
 292 ; SINK-GATHER:       vector.ph:
 293 ; SINK-GATHER-NEXT:    [[N_MOD_VF:%.*]] = urem i64 [[SMAX]], 8
 294 ; SINK-GATHER-NEXT:    [[N_VEC:%.*]] = sub i64 [[SMAX]], [[N_MOD_VF]]
 295 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLATINSERT:%.*]] = insertelement <8 x i1> poison, i1 [[C:%.*]], i64 0
 296 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLAT:%.*]] = shufflevector <8 x i1> [[BROADCAST_SPLATINSERT]], <8 x i1> poison, <8 x i32> zeroinitializer
 297 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLATINSERT15:%.*]] = insertelement <8 x i32> poison, i32 [[X:%.*]], i64 0
 298 ; SINK-GATHER-NEXT:    [[BROADCAST_SPLAT16:%.*]] = shufflevector <8 x i32> [[BROADCAST_SPLATINSERT15]], <8 x i32> poison, <8 x i32> zeroinitializer
 299 ; SINK-GATHER-NEXT:    br label [[VECTOR_BODY:%.*]]
 300 ; SINK-GATHER:       vector.body:
 301 ; SINK-GATHER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[PRED_UDIV_CONTINUE14:%.*]] ]
 302 ; SINK-GATHER-NEXT:    [[VEC_IND:%.*]] = phi <8 x i64> [ <i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6, i64 7>, [[VECTOR_PH]] ], [ [[VEC_IND_NEXT:%.*]], [[PRED_UDIV_CONTINUE14]] ]
 303 ; SINK-GATHER-NEXT:    [[VEC_PHI:%.*]] = phi <8 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP66:%.*]], [[PRED_UDIV_CONTINUE14]] ]
 304 ; SINK-GATHER-NEXT:    [[TMP0:%.*]] = mul <8 x i64> [[VEC_IND]], <i64 777, i64 777, i64 777, i64 777, i64 777, i64 777, i64 777, i64 777>
 305 ; SINK-GATHER-NEXT:    [[TMP1:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 0
 306 ; SINK-GATHER-NEXT:    br i1 [[TMP1]], label [[PRED_UDIV_IF:%.*]], label [[PRED_UDIV_CONTINUE:%.*]]
 307 ; SINK-GATHER:       pred.udiv.if:
 308 ; SINK-GATHER-NEXT:    [[TMP2:%.*]] = extractelement <8 x i64> [[TMP0]], i32 0
 309 ; SINK-GATHER-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP2]]
 310 ; SINK-GATHER-NEXT:    [[TMP4:%.*]] = load i32, ptr [[TMP3]], align 4
 311 ; SINK-GATHER-NEXT:    [[TMP5:%.*]] = udiv i32 [[TMP4]], [[X]]
 312 ; SINK-GATHER-NEXT:    [[TMP6:%.*]] = insertelement <8 x i32> poison, i32 [[TMP5]], i32 0
 313 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE]]
 314 ; SINK-GATHER:       pred.udiv.continue:
 315 ; SINK-GATHER-NEXT:    [[TMP8:%.*]] = phi <8 x i32> [ poison, [[VECTOR_BODY]] ], [ [[TMP6]], [[PRED_UDIV_IF]] ]
 316 ; SINK-GATHER-NEXT:    [[TMP9:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 1
 317 ; SINK-GATHER-NEXT:    br i1 [[TMP9]], label [[PRED_UDIV_IF1:%.*]], label [[PRED_UDIV_CONTINUE2:%.*]]
 318 ; SINK-GATHER:       pred.udiv.if1:
 319 ; SINK-GATHER-NEXT:    [[TMP10:%.*]] = extractelement <8 x i64> [[TMP0]], i32 1
 320 ; SINK-GATHER-NEXT:    [[TMP11:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP10]]
 321 ; SINK-GATHER-NEXT:    [[TMP12:%.*]] = load i32, ptr [[TMP11]], align 4
 322 ; SINK-GATHER-NEXT:    [[TMP13:%.*]] = udiv i32 [[TMP12]], [[X]]
 323 ; SINK-GATHER-NEXT:    [[TMP14:%.*]] = insertelement <8 x i32> [[TMP8]], i32 [[TMP13]], i32 1
 324 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE2]]
 325 ; SINK-GATHER:       pred.udiv.continue2:
 326 ; SINK-GATHER-NEXT:    [[TMP16:%.*]] = phi <8 x i32> [ [[TMP8]], [[PRED_UDIV_CONTINUE]] ], [ [[TMP14]], [[PRED_UDIV_IF1]] ]
 327 ; SINK-GATHER-NEXT:    [[TMP17:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 2
 328 ; SINK-GATHER-NEXT:    br i1 [[TMP17]], label [[PRED_UDIV_IF3:%.*]], label [[PRED_UDIV_CONTINUE4:%.*]]
 329 ; SINK-GATHER:       pred.udiv.if3:
 330 ; SINK-GATHER-NEXT:    [[TMP18:%.*]] = extractelement <8 x i64> [[TMP0]], i32 2
 331 ; SINK-GATHER-NEXT:    [[TMP19:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP18]]
 332 ; SINK-GATHER-NEXT:    [[TMP20:%.*]] = load i32, ptr [[TMP19]], align 4
 333 ; SINK-GATHER-NEXT:    [[TMP21:%.*]] = udiv i32 [[TMP20]], [[X]]
 334 ; SINK-GATHER-NEXT:    [[TMP22:%.*]] = insertelement <8 x i32> [[TMP16]], i32 [[TMP21]], i32 2
 335 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE4]]
 336 ; SINK-GATHER:       pred.udiv.continue4:
 337 ; SINK-GATHER-NEXT:    [[TMP24:%.*]] = phi <8 x i32> [ [[TMP16]], [[PRED_UDIV_CONTINUE2]] ], [ [[TMP22]], [[PRED_UDIV_IF3]] ]
 338 ; SINK-GATHER-NEXT:    [[TMP25:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 3
 339 ; SINK-GATHER-NEXT:    br i1 [[TMP25]], label [[PRED_UDIV_IF5:%.*]], label [[PRED_UDIV_CONTINUE6:%.*]]
 340 ; SINK-GATHER:       pred.udiv.if5:
 341 ; SINK-GATHER-NEXT:    [[TMP26:%.*]] = extractelement <8 x i64> [[TMP0]], i32 3
 342 ; SINK-GATHER-NEXT:    [[TMP27:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP26]]
 343 ; SINK-GATHER-NEXT:    [[TMP28:%.*]] = load i32, ptr [[TMP27]], align 4
 344 ; SINK-GATHER-NEXT:    [[TMP29:%.*]] = udiv i32 [[TMP28]], [[X]]
 345 ; SINK-GATHER-NEXT:    [[TMP30:%.*]] = insertelement <8 x i32> [[TMP24]], i32 [[TMP29]], i32 3
 346 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE6]]
 347 ; SINK-GATHER:       pred.udiv.continue6:
 348 ; SINK-GATHER-NEXT:    [[TMP32:%.*]] = phi <8 x i32> [ [[TMP24]], [[PRED_UDIV_CONTINUE4]] ], [ [[TMP30]], [[PRED_UDIV_IF5]] ]
 349 ; SINK-GATHER-NEXT:    [[TMP33:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 4
 350 ; SINK-GATHER-NEXT:    br i1 [[TMP33]], label [[PRED_UDIV_IF7:%.*]], label [[PRED_UDIV_CONTINUE8:%.*]]
 351 ; SINK-GATHER:       pred.udiv.if7:
 352 ; SINK-GATHER-NEXT:    [[TMP34:%.*]] = extractelement <8 x i64> [[TMP0]], i32 4
 353 ; SINK-GATHER-NEXT:    [[TMP35:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP34]]
 354 ; SINK-GATHER-NEXT:    [[TMP36:%.*]] = load i32, ptr [[TMP35]], align 4
 355 ; SINK-GATHER-NEXT:    [[TMP37:%.*]] = udiv i32 [[TMP36]], [[X]]
 356 ; SINK-GATHER-NEXT:    [[TMP38:%.*]] = insertelement <8 x i32> [[TMP32]], i32 [[TMP37]], i32 4
 357 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE8]]
 358 ; SINK-GATHER:       pred.udiv.continue8:
 359 ; SINK-GATHER-NEXT:    [[TMP40:%.*]] = phi <8 x i32> [ [[TMP32]], [[PRED_UDIV_CONTINUE6]] ], [ [[TMP38]], [[PRED_UDIV_IF7]] ]
 360 ; SINK-GATHER-NEXT:    [[TMP41:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 5
 361 ; SINK-GATHER-NEXT:    br i1 [[TMP41]], label [[PRED_UDIV_IF9:%.*]], label [[PRED_UDIV_CONTINUE10:%.*]]
 362 ; SINK-GATHER:       pred.udiv.if9:
 363 ; SINK-GATHER-NEXT:    [[TMP42:%.*]] = extractelement <8 x i64> [[TMP0]], i32 5
 364 ; SINK-GATHER-NEXT:    [[TMP43:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP42]]
 365 ; SINK-GATHER-NEXT:    [[TMP44:%.*]] = load i32, ptr [[TMP43]], align 4
 366 ; SINK-GATHER-NEXT:    [[TMP45:%.*]] = udiv i32 [[TMP44]], [[X]]
 367 ; SINK-GATHER-NEXT:    [[TMP46:%.*]] = insertelement <8 x i32> [[TMP40]], i32 [[TMP45]], i32 5
 368 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE10]]
 369 ; SINK-GATHER:       pred.udiv.continue10:
 370 ; SINK-GATHER-NEXT:    [[TMP48:%.*]] = phi <8 x i32> [ [[TMP40]], [[PRED_UDIV_CONTINUE8]] ], [ [[TMP46]], [[PRED_UDIV_IF9]] ]
 371 ; SINK-GATHER-NEXT:    [[TMP49:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 6
 372 ; SINK-GATHER-NEXT:    br i1 [[TMP49]], label [[PRED_UDIV_IF11:%.*]], label [[PRED_UDIV_CONTINUE12:%.*]]
 373 ; SINK-GATHER:       pred.udiv.if11:
 374 ; SINK-GATHER-NEXT:    [[TMP50:%.*]] = extractelement <8 x i64> [[TMP0]], i32 6
 375 ; SINK-GATHER-NEXT:    [[TMP51:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP50]]
 376 ; SINK-GATHER-NEXT:    [[TMP52:%.*]] = load i32, ptr [[TMP51]], align 4
 377 ; SINK-GATHER-NEXT:    [[TMP53:%.*]] = udiv i32 [[TMP52]], [[X]]
 378 ; SINK-GATHER-NEXT:    [[TMP54:%.*]] = insertelement <8 x i32> [[TMP48]], i32 [[TMP53]], i32 6
 379 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE12]]
 380 ; SINK-GATHER:       pred.udiv.continue12:
 381 ; SINK-GATHER-NEXT:    [[TMP56:%.*]] = phi <8 x i32> [ [[TMP48]], [[PRED_UDIV_CONTINUE10]] ], [ [[TMP54]], [[PRED_UDIV_IF11]] ]
 382 ; SINK-GATHER-NEXT:    [[TMP57:%.*]] = extractelement <8 x i1> [[BROADCAST_SPLAT]], i32 7
 383 ; SINK-GATHER-NEXT:    br i1 [[TMP57]], label [[PRED_UDIV_IF13:%.*]], label [[PRED_UDIV_CONTINUE14]]
 384 ; SINK-GATHER:       pred.udiv.if13:
 385 ; SINK-GATHER-NEXT:    [[TMP58:%.*]] = extractelement <8 x i64> [[TMP0]], i32 7
 386 ; SINK-GATHER-NEXT:    [[TMP59:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[TMP58]]
 387 ; SINK-GATHER-NEXT:    [[TMP60:%.*]] = load i32, ptr [[TMP59]], align 4
 388 ; SINK-GATHER-NEXT:    [[TMP61:%.*]] = udiv i32 [[TMP60]], [[X]]
 389 ; SINK-GATHER-NEXT:    [[TMP62:%.*]] = insertelement <8 x i32> [[TMP56]], i32 [[TMP61]], i32 7
 390 ; SINK-GATHER-NEXT:    br label [[PRED_UDIV_CONTINUE14]]
 391 ; SINK-GATHER:       pred.udiv.continue14:
 392 ; SINK-GATHER-NEXT:    [[TMP64:%.*]] = phi <8 x i32> [ [[TMP56]], [[PRED_UDIV_CONTINUE12]] ], [ [[TMP62]], [[PRED_UDIV_IF13]] ]
 393 ; SINK-GATHER-NEXT:    [[PREDPHI:%.*]] = select <8 x i1> [[BROADCAST_SPLAT]], <8 x i32> [[TMP64]], <8 x i32> [[BROADCAST_SPLAT16]]
 394 ; SINK-GATHER-NEXT:    [[TMP66]] = add <8 x i32> [[VEC_PHI]], [[PREDPHI]]
 395 ; SINK-GATHER-NEXT:    [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 8
 396 ; SINK-GATHER-NEXT:    [[VEC_IND_NEXT]] = add <8 x i64> [[VEC_IND]], <i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8, i64 8>
 397 ; SINK-GATHER-NEXT:    [[TMP67:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
 398 ; SINK-GATHER-NEXT:    br i1 [[TMP67]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]
 399 ; SINK-GATHER:       middle.block:
 400 ; SINK-GATHER-NEXT:    [[TMP68:%.*]] = call i32 @llvm.vector.reduce.add.v8i32(<8 x i32> [[TMP66]])
 401 ; SINK-GATHER-NEXT:    [[CMP_N:%.*]] = icmp eq i64 [[SMAX]], [[N_VEC]]
 402 ; SINK-GATHER-NEXT:    br i1 [[CMP_N]], label [[FOR_END:%.*]], label [[SCALAR_PH]]
 403 ; SINK-GATHER:       scalar.ph:
 404 ; SINK-GATHER-NEXT:    [[BC_RESUME_VAL:%.*]] = phi i64 [ [[N_VEC]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY:%.*]] ]
 405 ; SINK-GATHER-NEXT:    [[BC_MERGE_RDX:%.*]] = phi i32 [ [[TMP68]], [[MIDDLE_BLOCK]] ], [ 0, [[ENTRY]] ]
 406 ; SINK-GATHER-NEXT:    br label [[FOR_BODY:%.*]]
 407 ; SINK-GATHER:       for.body:
 408 ; SINK-GATHER-NEXT:    [[I:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[I_NEXT:%.*]], [[FOR_INC:%.*]] ]
 409 ; SINK-GATHER-NEXT:    [[R:%.*]] = phi i32 [ [[BC_MERGE_RDX]], [[SCALAR_PH]] ], [ [[T6:%.*]], [[FOR_INC]] ]
 410 ; SINK-GATHER-NEXT:    [[I7:%.*]] = mul i64 [[I]], 777
 411 ; SINK-GATHER-NEXT:    br i1 [[C]], label [[IF_THEN:%.*]], label [[FOR_INC]]
 412 ; SINK-GATHER:       if.then:
 413 ; SINK-GATHER-NEXT:    [[T0:%.*]] = getelementptr inbounds i32, ptr [[A]], i64 [[I7]]
 414 ; SINK-GATHER-NEXT:    [[T2:%.*]] = load i32, ptr [[T0]], align 4
 415 ; SINK-GATHER-NEXT:    [[T4:%.*]] = udiv i32 [[T2]], [[X]]
 416 ; SINK-GATHER-NEXT:    br label [[FOR_INC]]
 417 ; SINK-GATHER:       for.inc:
 418 ; SINK-GATHER-NEXT:    [[T5:%.*]] = phi i32 [ [[X]], [[FOR_BODY]] ], [ [[T4]], [[IF_THEN]] ]
 419 ; SINK-GATHER-NEXT:    [[T6]] = add i32 [[R]], [[T5]]
 420 ; SINK-GATHER-NEXT:    [[I_NEXT]] = add nuw nsw i64 [[I]], 1
 421 ; SINK-GATHER-NEXT:    [[COND:%.*]] = icmp slt i64 [[I_NEXT]], [[N]]
 422 ; SINK-GATHER-NEXT:    br i1 [[COND]], label [[FOR_BODY]], label [[FOR_END]], !llvm.loop [[LOOP5:![0-9]+]]
 423 ; SINK-GATHER:       for.end:
 424 ; SINK-GATHER-NEXT:    [[T7:%.*]] = phi i32 [ [[T6]], [[FOR_INC]] ], [ [[TMP68]], [[MIDDLE_BLOCK]] ]
 425 ; SINK-GATHER-NEXT:    ret i32 [[T7]]
 426 ;
 427 entry:
 428   br label %for.body
 429
 430 for.body:
 431   %i = phi i64 [ 0, %entry ], [ %i.next, %for.inc ]
 432   %r = phi i32 [ 0, %entry ], [ %t6, %for.inc ]
 433   %i7 = mul i64 %i, 777
 434   br i1 %c, label %if.then, label %for.inc
 435
 436 if.then:
 437   %t0 = getelementptr inbounds i32, ptr %a, i64 %i7
 438   %t2 = load i32, ptr %t0, align 4
 439   %t4 = udiv i32 %t2, %x
 440   br label %for.inc
 441
 442 for.inc:
 443   %t5 = phi i32 [ %x, %for.body ], [ %t4, %if.then]
 444   %t6 = add i32 %r, %t5
 445   %i.next = add nuw nsw i64 %i, 1
 446   %cond = icmp slt i64 %i.next, %n
 447   br i1 %cond, label %for.body, label %for.end
 448
 449 for.end:
 450   %t7 = phi i32 [ %t6, %for.inc ]
 451   ret i32 %t7
 452 }