183 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 42; RUN: opt -p loop-vectorize -force-vector-width=4 -force-vector-interleave=1 -S %s | FileCheck %s3 4target datalayout = "e-m:e-p:64:64-i64:64-i128:128-n32:64-S128"5 6; Test cases for https://github.com/llvm/llvm-project/issues/87410.7define void @test_not_first_lane_only_constant(ptr %A, ptr noalias %B) {8; CHECK-LABEL: define void @test_not_first_lane_only_constant(9; CHECK-SAME: ptr [[A:%.*]], ptr noalias [[B:%.*]]) {10; CHECK-NEXT: entry:11; CHECK-NEXT: br label [[VECTOR_PH:%.*]]12; CHECK: vector.ph:13; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]14; CHECK: vector.body:15; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]16; CHECK-NEXT: [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i1617; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[A]], i16 [[OFFSET_IDX]]18; CHECK-NEXT: [[TMP13:%.*]] = load i16, ptr [[B]], align 219; CHECK-NEXT: [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <4 x i16> poison, i16 [[TMP13]], i64 020; CHECK-NEXT: [[BROADCAST_SPLAT6:%.*]] = shufflevector <4 x i16> [[BROADCAST_SPLATINSERT5]], <4 x i16> poison, <4 x i32> zeroinitializer21; CHECK-NEXT: store <4 x i16> [[BROADCAST_SPLAT6]], ptr [[TMP1]], align 222; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 423; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], 100024; CHECK-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]25; CHECK: middle.block:26; CHECK-NEXT: br label [[LOOP_LATCH:%.*]]27; CHECK: exit:28; CHECK-NEXT: ret void29;30entry:31 br label %loop.header32 33loop.header:34 %iv = phi i16 [ 0, %entry ], [ %iv.next, %loop.latch ]35 %gep.A = getelementptr inbounds i16, ptr %A, i16 %iv36 br i1 false, label %loop.latch, label %else.137 38else.1:39 br i1 false, label %then.2, label %else.240 41then.2:42 br label %else.243 44else.2:45 br label %loop.latch46 47loop.latch:48 %merge = phi ptr [ %B, %else.2 ], [ poison, %loop.header ]49 %l = load i16, ptr %merge, align 250 %iv.next = add i16 %iv, 151 store i16 %l, ptr %gep.A52 %c.2 = icmp eq i16 %iv.next, 100053 br i1 %c.2, label %exit, label %loop.header54 55exit:56 ret void57}58 59define void @test_not_first_lane_only_wide_compare(ptr %A, ptr noalias %B, i16 %x, i16 %y) {60; CHECK-LABEL: define void @test_not_first_lane_only_wide_compare(61; CHECK-SAME: ptr [[A:%.*]], ptr noalias [[B:%.*]], i16 [[X:%.*]], i16 [[Y:%.*]]) {62; CHECK-NEXT: entry:63; CHECK-NEXT: br label [[VECTOR_PH:%.*]]64; CHECK: vector.ph:65; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]66; CHECK: vector.body:67; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]68; CHECK-NEXT: [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i1669; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[A]], i16 [[OFFSET_IDX]]70; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP1]], align 271; CHECK-NEXT: [[TMP3:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 072; CHECK-NEXT: [[TMP4:%.*]] = icmp ult i16 [[TMP3]], [[X]]73; CHECK-NEXT: [[TMP12:%.*]] = select i1 [[TMP4]], ptr poison, ptr [[B]]74; CHECK-NEXT: [[TMP13:%.*]] = load i16, ptr [[TMP12]], align 275; CHECK-NEXT: [[BROADCAST_SPLATINSERT5:%.*]] = insertelement <4 x i16> poison, i16 [[TMP13]], i64 076; CHECK-NEXT: [[BROADCAST_SPLAT6:%.*]] = shufflevector <4 x i16> [[BROADCAST_SPLATINSERT5]], <4 x i16> poison, <4 x i32> zeroinitializer77; CHECK-NEXT: store <4 x i16> [[BROADCAST_SPLAT6]], ptr [[TMP1]], align 278; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 479; CHECK-NEXT: [[TMP14:%.*]] = icmp eq i32 [[INDEX_NEXT]], 100080; CHECK-NEXT: br i1 [[TMP14]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP3:![0-9]+]]81; CHECK: middle.block:82; CHECK-NEXT: br label [[LOOP_LATCH:%.*]]83; CHECK: exit:84; CHECK-NEXT: ret void85;86entry:87 br label %loop.header88 89loop.header:90 %iv = phi i16 [ 0, %entry ], [ %iv.next, %loop.latch ]91 %gep.A = getelementptr inbounds i16, ptr %A, i16 %iv92 %l.0 = load i16, ptr %gep.A93 %c.0 = icmp ult i16 %l.0, %x94 br i1 %c.0, label %loop.latch, label %else.195 96else.1:97 %c.1 = icmp ult i16 %l.0, %y98 br i1 %c.1, label %then.2, label %else.299 100then.2:101 br label %else.2102 103else.2:104 br label %loop.latch105 106loop.latch:107 %merge = phi ptr [ %B, %else.2 ], [ poison, %loop.header ]108 %l = load i16, ptr %merge, align 2109 %iv.next = add i16 %iv, 1110 store i16 %l, ptr %gep.A111 %c.2 = icmp eq i16 %iv.next, 1000112 br i1 %c.2, label %exit, label %loop.header113 114exit:115 ret void116}117 118define void @test_not_first_lane_only_wide_compare_incoming_order_swapped(ptr %A, ptr noalias %B, i16 %x, i16 %y) {119; CHECK-LABEL: define void @test_not_first_lane_only_wide_compare_incoming_order_swapped(120; CHECK-SAME: ptr [[A:%.*]], ptr noalias [[B:%.*]], i16 [[X:%.*]], i16 [[Y:%.*]]) {121; CHECK-NEXT: entry:122; CHECK-NEXT: br label [[VECTOR_PH:%.*]]123; CHECK: vector.ph:124; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]125; CHECK: vector.body:126; CHECK-NEXT: [[INDEX:%.*]] = phi i32 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]127; CHECK-NEXT: [[OFFSET_IDX:%.*]] = trunc i32 [[INDEX]] to i16128; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds i16, ptr [[A]], i16 [[OFFSET_IDX]]129; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <4 x i16>, ptr [[TMP1]], align 2130; CHECK-NEXT: [[TMP3:%.*]] = extractelement <4 x i16> [[WIDE_LOAD]], i32 0131; CHECK-NEXT: [[TMP4:%.*]] = icmp ult i16 [[TMP3]], [[X]]132; CHECK-NEXT: [[PREDPHI:%.*]] = select i1 [[TMP4]], ptr poison, ptr [[B]]133; CHECK-NEXT: [[TMP12:%.*]] = load i16, ptr [[PREDPHI]], align 2134; CHECK-NEXT: [[BROADCAST_SPLATINSERT3:%.*]] = insertelement <4 x i16> poison, i16 [[TMP12]], i64 0135; CHECK-NEXT: [[BROADCAST_SPLAT4:%.*]] = shufflevector <4 x i16> [[BROADCAST_SPLATINSERT3]], <4 x i16> poison, <4 x i32> zeroinitializer136; CHECK-NEXT: store <4 x i16> [[BROADCAST_SPLAT4]], ptr [[TMP1]], align 2137; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i32 [[INDEX]], 4138; CHECK-NEXT: [[TMP13:%.*]] = icmp eq i32 [[INDEX_NEXT]], 1000139; CHECK-NEXT: br i1 [[TMP13]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP4:![0-9]+]]140; CHECK: middle.block:141; CHECK-NEXT: br label [[LOOP_LATCH:%.*]]142; CHECK: exit:143; CHECK-NEXT: ret void144;145entry:146 br label %loop.header147 148loop.header:149 %iv = phi i16 [ 0, %entry ], [ %iv.next, %loop.latch ]150 %gep.A = getelementptr inbounds i16, ptr %A, i16 %iv151 %l.0 = load i16, ptr %gep.A152 %c.0 = icmp ult i16 %l.0, %x153 br i1 %c.0, label %loop.latch, label %else.1154 155else.1:156 %c.1 = icmp ult i16 %l.0, %y157 br i1 %c.1, label %then.2, label %else.2158 159then.2:160 br label %else.2161 162else.2:163 br label %loop.latch164 165loop.latch:166 %merge = phi ptr [ poison, %loop.header ], [ %B, %else.2 ]167 %l = load i16, ptr %merge, align 2168 %iv.next = add i16 %iv, 1169 store i16 %l, ptr %gep.A170 %c.2 = icmp eq i16 %iv.next, 1000171 br i1 %c.2, label %exit, label %loop.header172 173exit:174 ret void175}176;.177; CHECK: [[LOOP0]] = distinct !{[[LOOP0]], [[META1:![0-9]+]], [[META2:![0-9]+]]}178; CHECK: [[META1]] = !{!"llvm.loop.isvectorized", i32 1}179; CHECK: [[META2]] = !{!"llvm.loop.unroll.runtime.disable"}180; CHECK: [[LOOP3]] = distinct !{[[LOOP3]], [[META1]], [[META2]]}181; CHECK: [[LOOP4]] = distinct !{[[LOOP4]], [[META1]], [[META2]]}182;.183