Skip to content

Commit c363a3f

Browse files
david-armdavid-arm
authored
[LV] Ensure getScaledReductions only matches extends inside the loop (#148264)
In getScaledReductions for the case where we try to match a partial reduction of the form: %phi = phi i32 ... ... %add = add i32 %phi, %zext where %zext = i8 %some_val to i32 we should ensure that %zext is actually inside the loop. Fixes llvm/llvm-project#148260
1 parent 69ea174 commit c363a3f

File tree

2 files changed

+262
-1
lines changed

2 files changed

+262
-1
lines changed

‎llvm/lib/Transforms/Vectorize/LoopVectorize.cpp‎

Lines changed: 6 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -8112,14 +8112,19 @@ bool VPRecipeBuilder::getScaledReductions(
81128112
std::optional<unsigned> BinOpc;
81138113
Type *ExtOpTypes[2] = {nullptr};
81148114

8115-
auto CollectExtInfo = [&Exts,
8115+
auto CollectExtInfo = [this, &Exts,
81168116
&ExtOpTypes](SmallVectorImpl<Value *> &Ops) -> bool {
81178117
unsigned I = 0;
81188118
for (Value *OpI : Ops) {
81198119
Value *ExtOp;
81208120
if (!match(OpI, m_ZExtOrSExt(m_Value(ExtOp))))
81218121
return false;
81228122
Exts[I] = cast<Instruction>(OpI);
8123+
8124+
// TODO: We should be able to support live-ins.
8125+
if (!CM.TheLoop->contains(Exts[I]))
8126+
return false;
8127+
81238128
ExtOpTypes[I] = ExtOp->getType();
81248129
I++;
81258130
}

‎llvm/test/Transforms/LoopVectorize/AArch64/partial-reduce.ll‎

Lines changed: 256 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -911,6 +911,262 @@ for.exit: ; preds = %for.body
911911
ret i32 %add
912912
}
913913

914+
915+
define i32 @add_of_zext_outside_loop(i32 %a, ptr noalias %b, i8 %c, i32 %d) #0 {
916+
; CHECK-INTERLEAVE1-LABEL: define i32 @add_of_zext_outside_loop(
917+
; CHECK-INTERLEAVE1-SAME: i32 [[A:%.*]], ptr noalias [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
918+
; CHECK-INTERLEAVE1-NEXT: entry:
919+
; CHECK-INTERLEAVE1-NEXT: [[CONV1:%.*]] = zext i8 [[C]] to i32
920+
; CHECK-INTERLEAVE1-NEXT: [[TMP0:%.*]] = sub i32 1024, [[D]]
921+
; CHECK-INTERLEAVE1-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP0]], 16
922+
; CHECK-INTERLEAVE1-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
923+
; CHECK-INTERLEAVE1: vector.ph:
924+
; CHECK-INTERLEAVE1-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 16
925+
; CHECK-INTERLEAVE1-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[N_MOD_VF]]
926+
; CHECK-INTERLEAVE1-NEXT: [[TMP1:%.*]] = add i32 [[D]], [[N_VEC]]
927+
; CHECK-INTERLEAVE1-NEXT: [[TMP2:%.*]] = insertelement <16 x i32> zeroinitializer, i32 [[A]], i32 0
928+
; CHECK-INTERLEAVE1-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i32> poison, i32 [[CONV1]], i64 0
929+
; CHECK-INTERLEAVE1-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i32> [[BROADCAST_SPLATINSERT]], <16 x i32> poison, <16 x i32> zeroinitializer
930+
; CHECK-INTERLEAVE1-NEXT: br label [[FOR_BODY:%.*]]
931+
; CHECK-INTERLEAVE1: vector.body:
932+
; CHECK-INTERLEAVE1-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[FOR_BODY]] ]
933+
; CHECK-INTERLEAVE1-NEXT: [[VEC_PHI:%.*]] = phi <16 x i32> [ [[TMP2]], [[VECTOR_PH]] ], [ [[TMP5:%.*]], [[FOR_BODY]] ]
934+
; CHECK-INTERLEAVE1-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[INDEX]]
935+
; CHECK-INTERLEAVE1-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
936+
; CHECK-INTERLEAVE1-NEXT: [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[TMP3]], i32 0
937+
; CHECK-INTERLEAVE1-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP4]], align 1
938+
; CHECK-INTERLEAVE1-NEXT: [[TMP5]] = add <16 x i32> [[VEC_PHI]], [[BROADCAST_SPLAT]]
939+
; CHECK-INTERLEAVE1-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
940+
; CHECK-INTERLEAVE1-NEXT: [[TMP6:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
941+
; CHECK-INTERLEAVE1-NEXT: br i1 [[TMP6]], label [[MIDDLE_BLOCK:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
942+
; CHECK-INTERLEAVE1: middle.block:
943+
; CHECK-INTERLEAVE1-NEXT: [[TMP7:%.*]] = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> [[TMP5]])
944+
; CHECK-INTERLEAVE1-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP0]], [[N_VEC]]
945+
; CHECK-INTERLEAVE1-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
946+
; CHECK-INTERLEAVE1: scalar.ph:
947+
;
948+
; CHECK-INTERLEAVED-LABEL: define i32 @add_of_zext_outside_loop(
949+
; CHECK-INTERLEAVED-SAME: i32 [[A:%.*]], ptr noalias [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
950+
; CHECK-INTERLEAVED-NEXT: entry:
951+
; CHECK-INTERLEAVED-NEXT: [[CONV1:%.*]] = zext i8 [[C]] to i32
952+
; CHECK-INTERLEAVED-NEXT: [[TMP2:%.*]] = sub i32 1024, [[D]]
953+
; CHECK-INTERLEAVED-NEXT: [[TMP11:%.*]] = icmp ult i32 [[TMP2]], 32
954+
; CHECK-INTERLEAVED-NEXT: br i1 [[TMP11]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
955+
; CHECK-INTERLEAVED: vector.ph:
956+
; CHECK-INTERLEAVED-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP2]], 32
957+
; CHECK-INTERLEAVED-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP2]], [[N_MOD_VF]]
958+
; CHECK-INTERLEAVED-NEXT: [[TMP1:%.*]] = add i32 [[D]], [[N_VEC]]
959+
; CHECK-INTERLEAVED-NEXT: [[TMP10:%.*]] = insertelement <16 x i32> zeroinitializer, i32 [[A]], i32 0
960+
; CHECK-INTERLEAVED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i32> poison, i32 [[CONV1]], i64 0
961+
; CHECK-INTERLEAVED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i32> [[BROADCAST_SPLATINSERT]], <16 x i32> poison, <16 x i32> zeroinitializer
962+
; CHECK-INTERLEAVED-NEXT: br label [[VECTOR_BODY:%.*]]
963+
; CHECK-INTERLEAVED: vector.body:
964+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP21:%.*]], [[VECTOR_BODY]] ]
965+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI:%.*]] = phi <16 x i32> [ [[TMP10]], [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[VECTOR_BODY]] ]
966+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI2:%.*]] = phi <16 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
967+
; CHECK-INTERLEAVED-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[VEC_PHI1]]
968+
; CHECK-INTERLEAVED-NEXT: [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
969+
; CHECK-INTERLEAVED-NEXT: [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[TMP3]], i32 0
970+
; CHECK-INTERLEAVED-NEXT: [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP3]], i32 16
971+
; CHECK-INTERLEAVED-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP4]], align 1
972+
; CHECK-INTERLEAVED-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP5]], align 1
973+
; CHECK-INTERLEAVED-NEXT: [[TMP6]] = add <16 x i32> [[VEC_PHI]], [[BROADCAST_SPLAT]]
974+
; CHECK-INTERLEAVED-NEXT: [[TMP7]] = add <16 x i32> [[VEC_PHI2]], [[BROADCAST_SPLAT]]
975+
; CHECK-INTERLEAVED-NEXT: [[TMP21]] = add nuw i32 [[VEC_PHI1]], 32
976+
; CHECK-INTERLEAVED-NEXT: [[TMP8:%.*]] = icmp eq i32 [[TMP21]], [[N_VEC]]
977+
; CHECK-INTERLEAVED-NEXT: br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
978+
; CHECK-INTERLEAVED: middle.block:
979+
; CHECK-INTERLEAVED-NEXT: [[BIN_RDX:%.*]] = add <16 x i32> [[TMP7]], [[TMP6]]
980+
; CHECK-INTERLEAVED-NEXT: [[TMP9:%.*]] = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> [[BIN_RDX]])
981+
; CHECK-INTERLEAVED-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP2]], [[N_VEC]]
982+
; CHECK-INTERLEAVED-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
983+
; CHECK-INTERLEAVED: scalar.ph:
984+
;
985+
; CHECK-MAXBW-LABEL: define i32 @add_of_zext_outside_loop(
986+
; CHECK-MAXBW-SAME: i32 [[A:%.*]], ptr noalias [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
987+
; CHECK-MAXBW-NEXT: entry:
988+
; CHECK-MAXBW-NEXT: [[CONV1:%.*]] = zext i8 [[C]] to i32
989+
; CHECK-MAXBW-NEXT: [[TMP0:%.*]] = sub i32 1024, [[D]]
990+
; CHECK-MAXBW-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()
991+
; CHECK-MAXBW-NEXT: [[TMP2:%.*]] = mul nuw i32 [[TMP1]], 16
992+
; CHECK-MAXBW-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP0]], [[TMP2]]
993+
; CHECK-MAXBW-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
994+
; CHECK-MAXBW: vector.ph:
995+
; CHECK-MAXBW-NEXT: [[TMP3:%.*]] = call i32 @llvm.vscale.i32()
996+
; CHECK-MAXBW-NEXT: [[TMP4:%.*]] = mul nuw i32 [[TMP3]], 16
997+
; CHECK-MAXBW-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], [[TMP4]]
998+
; CHECK-MAXBW-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[N_MOD_VF]]
999+
; CHECK-MAXBW-NEXT: [[TMP5:%.*]] = call i32 @llvm.vscale.i32()
1000+
; CHECK-MAXBW-NEXT: [[TMP6:%.*]] = mul nuw i32 [[TMP5]], 16
1001+
; CHECK-MAXBW-NEXT: [[TMP7:%.*]] = add i32 [[D]], [[N_VEC]]
1002+
; CHECK-MAXBW-NEXT: [[TMP8:%.*]] = insertelement <vscale x 16 x i32> zeroinitializer, i32 [[A]], i32 0
1003+
; CHECK-MAXBW-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i32> poison, i32 [[CONV1]], i64 0
1004+
; CHECK-MAXBW-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i32> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i32> poison, <vscale x 16 x i32> zeroinitializer
1005+
; CHECK-MAXBW-NEXT: br label [[FOR_BODY:%.*]]
1006+
; CHECK-MAXBW: vector.body:
1007+
; CHECK-MAXBW-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[FOR_BODY]] ]
1008+
; CHECK-MAXBW-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 16 x i32> [ [[TMP8]], [[VECTOR_PH]] ], [ [[TMP11:%.*]], [[FOR_BODY]] ]
1009+
; CHECK-MAXBW-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[INDEX]]
1010+
; CHECK-MAXBW-NEXT: [[TMP9:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
1011+
; CHECK-MAXBW-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[TMP9]], i32 0
1012+
; CHECK-MAXBW-NEXT: store <vscale x 16 x i8> zeroinitializer, ptr [[TMP10]], align 1
1013+
; CHECK-MAXBW-NEXT: [[TMP11]] = add <vscale x 16 x i32> [[VEC_PHI]], [[BROADCAST_SPLAT]]
1014+
; CHECK-MAXBW-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP6]]
1015+
; CHECK-MAXBW-NEXT: [[TMP12:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
1016+
; CHECK-MAXBW-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP18:![0-9]+]]
1017+
; CHECK-MAXBW: middle.block:
1018+
; CHECK-MAXBW-NEXT: [[TMP13:%.*]] = call i32 @llvm.vector.reduce.add.nxv16i32(<vscale x 16 x i32> [[TMP11]])
1019+
; CHECK-MAXBW-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP0]], [[N_VEC]]
1020+
; CHECK-MAXBW-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
1021+
; CHECK-MAXBW: scalar.ph:
1022+
;
1023+
entry:
1024+
%conv1 = zext i8 %c to i32
1025+
br label %for.body
1026+
1027+
for.body:
1028+
%iv = phi i32 [ %d, %entry ], [ %iv.next, %for.body ]
1029+
%rdx = phi i32 [ %a, %entry ], [ %rdx.next, %for.body ]
1030+
%arrayidx = getelementptr inbounds i8, ptr %b, i32 %iv
1031+
store i8 0, ptr %arrayidx, align 1
1032+
%rdx.next = add nsw i32 %rdx, %conv1
1033+
%iv.next = add nuw nsw i32 %iv, 1
1034+
%cmp = icmp eq i32 %iv.next, 1024
1035+
br i1 %cmp, label %exit, label %for.body
1036+
1037+
exit:
1038+
%add.lcssa = phi i32 [ %rdx.next, %for.body ]
1039+
ret i32 %add.lcssa
1040+
}
1041+
1042+
define i32 @add_of_loop_invariant_zext(i32 %a, ptr %b, i8 %c, i32 %d) #0 {
1043+
; CHECK-INTERLEAVE1-LABEL: define i32 @add_of_loop_invariant_zext(
1044+
; CHECK-INTERLEAVE1-SAME: i32 [[A:%.*]], ptr [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
1045+
; CHECK-INTERLEAVE1-NEXT: entry:
1046+
; CHECK-INTERLEAVE1-NEXT: [[TMP0:%.*]] = sub i32 1024, [[D]]
1047+
; CHECK-INTERLEAVE1-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP0]], 16
1048+
; CHECK-INTERLEAVE1-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1049+
; CHECK-INTERLEAVE1: vector.ph:
1050+
; CHECK-INTERLEAVE1-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], 16
1051+
; CHECK-INTERLEAVE1-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[N_MOD_VF]]
1052+
; CHECK-INTERLEAVE1-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i8> poison, i8 [[C]], i64 0
1053+
; CHECK-INTERLEAVE1-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i8> [[BROADCAST_SPLATINSERT]], <16 x i8> poison, <16 x i32> zeroinitializer
1054+
; CHECK-INTERLEAVE1-NEXT: [[TMP1:%.*]] = add i32 [[D]], [[N_VEC]]
1055+
; CHECK-INTERLEAVE1-NEXT: [[TMP2:%.*]] = insertelement <16 x i32> zeroinitializer, i32 [[A]], i32 0
1056+
; CHECK-INTERLEAVE1-NEXT: [[TMP3:%.*]] = zext <16 x i8> [[BROADCAST_SPLAT]] to <16 x i32>
1057+
; CHECK-INTERLEAVE1-NEXT: br label [[FOR_BODY:%.*]]
1058+
; CHECK-INTERLEAVE1: vector.body:
1059+
; CHECK-INTERLEAVE1-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[FOR_BODY]] ]
1060+
; CHECK-INTERLEAVE1-NEXT: [[VEC_PHI:%.*]] = phi <16 x i32> [ [[TMP2]], [[VECTOR_PH]] ], [ [[TMP6:%.*]], [[FOR_BODY]] ]
1061+
; CHECK-INTERLEAVE1-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[INDEX]]
1062+
; CHECK-INTERLEAVE1-NEXT: [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
1063+
; CHECK-INTERLEAVE1-NEXT: [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 0
1064+
; CHECK-INTERLEAVE1-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP5]], align 1
1065+
; CHECK-INTERLEAVE1-NEXT: [[TMP6]] = add <16 x i32> [[VEC_PHI]], [[TMP3]]
1066+
; CHECK-INTERLEAVE1-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 16
1067+
; CHECK-INTERLEAVE1-NEXT: [[TMP7:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
1068+
; CHECK-INTERLEAVE1-NEXT: br i1 [[TMP7]], label [[MIDDLE_BLOCK:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
1069+
; CHECK-INTERLEAVE1: middle.block:
1070+
; CHECK-INTERLEAVE1-NEXT: [[TMP8:%.*]] = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> [[TMP6]])
1071+
; CHECK-INTERLEAVE1-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP0]], [[N_VEC]]
1072+
; CHECK-INTERLEAVE1-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
1073+
; CHECK-INTERLEAVE1: scalar.ph:
1074+
;
1075+
; CHECK-INTERLEAVED-LABEL: define i32 @add_of_loop_invariant_zext(
1076+
; CHECK-INTERLEAVED-SAME: i32 [[A:%.*]], ptr [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
1077+
; CHECK-INTERLEAVED-NEXT: entry:
1078+
; CHECK-INTERLEAVED-NEXT: [[TMP2:%.*]] = sub i32 1024, [[D]]
1079+
; CHECK-INTERLEAVED-NEXT: [[TMP11:%.*]] = icmp ult i32 [[TMP2]], 32
1080+
; CHECK-INTERLEAVED-NEXT: br i1 [[TMP11]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1081+
; CHECK-INTERLEAVED: vector.ph:
1082+
; CHECK-INTERLEAVED-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP2]], 32
1083+
; CHECK-INTERLEAVED-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP2]], [[N_MOD_VF]]
1084+
; CHECK-INTERLEAVED-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <16 x i8> poison, i8 [[C]], i64 0
1085+
; CHECK-INTERLEAVED-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <16 x i8> [[BROADCAST_SPLATINSERT]], <16 x i8> poison, <16 x i32> zeroinitializer
1086+
; CHECK-INTERLEAVED-NEXT: [[TMP1:%.*]] = add i32 [[D]], [[N_VEC]]
1087+
; CHECK-INTERLEAVED-NEXT: [[TMP12:%.*]] = insertelement <16 x i32> zeroinitializer, i32 [[A]], i32 0
1088+
; CHECK-INTERLEAVED-NEXT: [[TMP3:%.*]] = zext <16 x i8> [[BROADCAST_SPLAT]] to <16 x i32>
1089+
; CHECK-INTERLEAVED-NEXT: br label [[VECTOR_BODY:%.*]]
1090+
; CHECK-INTERLEAVED: vector.body:
1091+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI1:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[TMP22:%.*]], [[VECTOR_BODY]] ]
1092+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI:%.*]] = phi <16 x i32> [ [[TMP12]], [[VECTOR_PH]] ], [ [[TMP7:%.*]], [[VECTOR_BODY]] ]
1093+
; CHECK-INTERLEAVED-NEXT: [[VEC_PHI2:%.*]] = phi <16 x i32> [ zeroinitializer, [[VECTOR_PH]] ], [ [[TMP8:%.*]], [[VECTOR_BODY]] ]
1094+
; CHECK-INTERLEAVED-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[VEC_PHI1]]
1095+
; CHECK-INTERLEAVED-NEXT: [[TMP4:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
1096+
; CHECK-INTERLEAVED-NEXT: [[TMP5:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 0
1097+
; CHECK-INTERLEAVED-NEXT: [[TMP6:%.*]] = getelementptr inbounds i8, ptr [[TMP4]], i32 16
1098+
; CHECK-INTERLEAVED-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP5]], align 1
1099+
; CHECK-INTERLEAVED-NEXT: store <16 x i8> zeroinitializer, ptr [[TMP6]], align 1
1100+
; CHECK-INTERLEAVED-NEXT: [[TMP7]] = add <16 x i32> [[VEC_PHI]], [[TMP3]]
1101+
; CHECK-INTERLEAVED-NEXT: [[TMP8]] = add <16 x i32> [[VEC_PHI2]], [[TMP3]]
1102+
; CHECK-INTERLEAVED-NEXT: [[TMP22]] = add nuw i32 [[VEC_PHI1]], 32
1103+
; CHECK-INTERLEAVED-NEXT: [[TMP9:%.*]] = icmp eq i32 [[TMP22]], [[N_VEC]]
1104+
; CHECK-INTERLEAVED-NEXT: br i1 [[TMP9]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
1105+
; CHECK-INTERLEAVED: middle.block:
1106+
; CHECK-INTERLEAVED-NEXT: [[BIN_RDX:%.*]] = add <16 x i32> [[TMP8]], [[TMP7]]
1107+
; CHECK-INTERLEAVED-NEXT: [[TMP10:%.*]] = call i32 @llvm.vector.reduce.add.v16i32(<16 x i32> [[BIN_RDX]])
1108+
; CHECK-INTERLEAVED-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP2]], [[N_VEC]]
1109+
; CHECK-INTERLEAVED-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
1110+
; CHECK-INTERLEAVED: scalar.ph:
1111+
;
1112+
; CHECK-MAXBW-LABEL: define i32 @add_of_loop_invariant_zext(
1113+
; CHECK-MAXBW-SAME: i32 [[A:%.*]], ptr [[B:%.*]], i8 [[C:%.*]], i32 [[D:%.*]]) #[[ATTR0]] {
1114+
; CHECK-MAXBW-NEXT: entry:
1115+
; CHECK-MAXBW-NEXT: [[TMP0:%.*]] = sub i32 1024, [[D]]
1116+
; CHECK-MAXBW-NEXT: [[TMP1:%.*]] = call i32 @llvm.vscale.i32()
1117+
; CHECK-MAXBW-NEXT: [[TMP2:%.*]] = mul nuw i32 [[TMP1]], 16
1118+
; CHECK-MAXBW-NEXT: [[MIN_ITERS_CHECK:%.*]] = icmp ult i32 [[TMP0]], [[TMP2]]
1119+
; CHECK-MAXBW-NEXT: br i1 [[MIN_ITERS_CHECK]], label [[SCALAR_PH:%.*]], label [[VECTOR_PH:%.*]]
1120+
; CHECK-MAXBW: vector.ph:
1121+
; CHECK-MAXBW-NEXT: [[TMP3:%.*]] = call i32 @llvm.vscale.i32()
1122+
; CHECK-MAXBW-NEXT: [[TMP4:%.*]] = mul nuw i32 [[TMP3]], 16
1123+
; CHECK-MAXBW-NEXT: [[N_MOD_VF:%.*]] = urem i32 [[TMP0]], [[TMP4]]
1124+
; CHECK-MAXBW-NEXT: [[N_VEC:%.*]] = sub i32 [[TMP0]], [[N_MOD_VF]]
1125+
; CHECK-MAXBW-NEXT: [[TMP5:%.*]] = call i32 @llvm.vscale.i32()
1126+
; CHECK-MAXBW-NEXT: [[TMP6:%.*]] = mul nuw i32 [[TMP5]], 16
1127+
; CHECK-MAXBW-NEXT: [[BROADCAST_SPLATINSERT:%.*]] = insertelement <vscale x 16 x i8> poison, i8 [[C]], i64 0
1128+
; CHECK-MAXBW-NEXT: [[BROADCAST_SPLAT:%.*]] = shufflevector <vscale x 16 x i8> [[BROADCAST_SPLATINSERT]], <vscale x 16 x i8> poison, <vscale x 16 x i32> zeroinitializer
1129+
; CHECK-MAXBW-NEXT: [[TMP7:%.*]] = add i32 [[D]], [[N_VEC]]
1130+
; CHECK-MAXBW-NEXT: [[TMP8:%.*]] = insertelement <vscale x 4 x i32> zeroinitializer, i32 [[A]], i32 0
1131+
; CHECK-MAXBW-NEXT: [[TMP9:%.*]] = zext <vscale x 16 x i8> [[BROADCAST_SPLAT]] to <vscale x 16 x i32>
1132+
; CHECK-MAXBW-NEXT: br label [[FOR_BODY:%.*]]
1133+
; CHECK-MAXBW: vector.body:
1134+
; CHECK-MAXBW-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[FOR_BODY]] ]
1135+
; CHECK-MAXBW-NEXT: [[VEC_PHI:%.*]] = phi <vscale x 4 x i32> [ [[TMP8]], [[VECTOR_PH]] ], [ [[PARTIAL_REDUCE:%.*]], [[FOR_BODY]] ]
1136+
; CHECK-MAXBW-NEXT: [[OFFSET_IDX:%.*]] = add i32 [[D]], [[INDEX]]
1137+
; CHECK-MAXBW-NEXT: [[TMP10:%.*]] = getelementptr inbounds i8, ptr [[B]], i32 [[OFFSET_IDX]]
1138+
; CHECK-MAXBW-NEXT: [[TMP11:%.*]] = getelementptr inbounds i8, ptr [[TMP10]], i32 0
1139+
; CHECK-MAXBW-NEXT: store <vscale x 16 x i8> zeroinitializer, ptr [[TMP11]], align 1
1140+
; CHECK-MAXBW-NEXT: [[PARTIAL_REDUCE]] = call <vscale x 4 x i32> @llvm.experimental.vector.partial.reduce.add.nxv4i32.nxv16i32(<vscale x 4 x i32> [[VEC_PHI]], <vscale x 16 x i32> [[TMP9]])
1141+
; CHECK-MAXBW-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], [[TMP6]]
1142+
; CHECK-MAXBW-NEXT: [[TMP12:%.*]] = icmp eq i32 [[INDEX_NEXT]], [[N_VEC]]
1143+
; CHECK-MAXBW-NEXT: br i1 [[TMP12]], label [[MIDDLE_BLOCK:%.*]], label [[FOR_BODY]], !llvm.loop [[LOOP20:![0-9]+]]
1144+
; CHECK-MAXBW: middle.block:
1145+
; CHECK-MAXBW-NEXT: [[TMP13:%.*]] = call i32 @llvm.vector.reduce.add.nxv4i32(<vscale x 4 x i32> [[PARTIAL_REDUCE]])
1146+
; CHECK-MAXBW-NEXT: [[CMP_N:%.*]] = icmp eq i32 [[TMP0]], [[N_VEC]]
1147+
; CHECK-MAXBW-NEXT: br i1 [[CMP_N]], label [[EXIT:%.*]], label [[SCALAR_PH]]
1148+
; CHECK-MAXBW: scalar.ph:
1149+
;
1150+
entry:
1151+
br label %for.body
1152+
1153+
for.body:
1154+
%iv = phi i32 [ %d, %entry ], [ %iv.next, %for.body ]
1155+
%rdx = phi i32 [ %a, %entry ], [ %rdx.next, %for.body ]
1156+
%arrayidx = getelementptr inbounds i8, ptr %b, i32 %iv
1157+
store i8 0, ptr %arrayidx, align 1
1158+
%conv1 = zext i8 %c to i32
1159+
%rdx.next = add nsw i32 %rdx, %conv1
1160+
%iv.next = add nuw nsw i32 %iv, 1
1161+
%cmp = icmp eq i32 %iv.next, 1024
1162+
br i1 %cmp, label %exit, label %for.body
1163+
1164+
exit:
1165+
%add.lcssa = phi i32 [ %rdx.next, %for.body ]
1166+
ret i32 %add.lcssa
1167+
}
1168+
1169+
9141170
!0 = distinct !{!0, !1}
9151171
!1 = !{!"llvm.loop.vectorize.predicate.enable", i1 true}
9161172
attributes #0 = { vscale_range(1,16) "target-features"="+sve" }

0 commit comments

Comments
 (0)