183 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 52; RUN: opt -S -passes='dxil-data-scalarization' -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s3 4; Ensure that insertelement instructions have no side effects on each other5; even in the presence of control flow6define void @test_multiple_insert(i32 %c, i32 %i, i32 %j) {7; CHECK-LABEL: define void @test_multiple_insert(8; CHECK-SAME: i32 [[C:%.*]], i32 [[I:%.*]], i32 [[J:%.*]]) {9; CHECK-NEXT: [[V0_ALLOCA:%.*]] = alloca [2 x i32], align 410; CHECK-NEXT: [[V_ALLOCA:%.*]] = alloca [2 x i32], align 411; CHECK-NEXT: [[V0_0:%.*]] = insertelement <2 x i32> poison, i32 0, i32 012; CHECK-NEXT: [[V0:%.*]] = insertelement <2 x i32> [[V0_0]], i32 0, i32 113; CHECK-NEXT: [[V0_EXTRACT0:%.*]] = extractelement <2 x i32> [[V0]], i64 014; CHECK-NEXT: [[V0_INDEX0:%.*]] = getelementptr inbounds [2 x i32], ptr [[V0_ALLOCA]], i32 0, i32 015; CHECK-NEXT: store i32 [[V0_EXTRACT0]], ptr [[V0_INDEX0]], align 416; CHECK-NEXT: [[V0_EXTRACT1:%.*]] = extractelement <2 x i32> [[V0]], i64 117; CHECK-NEXT: [[V0_INDEX1:%.*]] = getelementptr inbounds [2 x i32], ptr [[V0_ALLOCA]], i32 0, i32 118; CHECK-NEXT: store i32 [[V0_EXTRACT1]], ptr [[V0_INDEX1]], align 419; CHECK-NEXT: [[COND:%.*]] = icmp eq i32 [[C]], 120; CHECK-NEXT: br i1 [[COND]], label %[[IF:.*]], label %[[ELSE:.*]]21; CHECK: [[IF]]:22; CHECK-NEXT: [[V1_INDEX:%.*]] = getelementptr inbounds [2 x i32], ptr [[V0_ALLOCA]], i32 0, i32 [[I]]23; CHECK-NEXT: [[V1_LOAD:%.*]] = load i32, ptr [[V1_INDEX]], align 424; CHECK-NEXT: store i32 1, ptr [[V1_INDEX]], align 425; CHECK-NEXT: [[V1_LOAD0:%.*]] = load i32, ptr [[V0_INDEX0]], align 426; CHECK-NEXT: [[V1_INSERT0:%.*]] = insertelement <2 x i32> poison, i32 [[V1_LOAD0]], i32 027; CHECK-NEXT: [[V1_LOAD1:%.*]] = load i32, ptr [[V0_INDEX1]], align 428; CHECK-NEXT: [[V1_INSERT1:%.*]] = insertelement <2 x i32> [[V1_INSERT0]], i32 [[V1_LOAD1]], i32 129; CHECK-NEXT: store i32 [[V1_LOAD]], ptr [[V1_INDEX]], align 430; CHECK-NEXT: br label %[[EXIT:.*]]31; CHECK: [[ELSE]]:32; CHECK-NEXT: [[V2_INDEX:%.*]] = getelementptr inbounds [2 x i32], ptr [[V0_ALLOCA]], i32 0, i32 [[I]]33; CHECK-NEXT: [[V2_LOAD:%.*]] = load i32, ptr [[V2_INDEX]], align 434; CHECK-NEXT: store i32 2, ptr [[V2_INDEX]], align 435; CHECK-NEXT: [[V2_LOAD0:%.*]] = load i32, ptr [[V0_INDEX0]], align 436; CHECK-NEXT: [[V2_INSERT0:%.*]] = insertelement <2 x i32> poison, i32 [[V2_LOAD0]], i32 037; CHECK-NEXT: [[V2_LOAD1:%.*]] = load i32, ptr [[V0_INDEX1]], align 438; CHECK-NEXT: [[V2_INSERT1:%.*]] = insertelement <2 x i32> [[V2_INSERT0]], i32 [[V2_LOAD1]], i32 139; CHECK-NEXT: store i32 [[V2_LOAD]], ptr [[V2_INDEX]], align 440; CHECK-NEXT: br label %[[EXIT]]41; CHECK: [[EXIT]]:42; CHECK-NEXT: [[V:%.*]] = phi <2 x i32> [ [[V1_INSERT1]], %[[IF]] ], [ [[V2_INSERT1]], %[[ELSE]] ]43; CHECK-NEXT: [[V_EXTRACT:%.*]] = extractelement <2 x i32> [[V]], i64 044; CHECK-NEXT: [[V_INDEX:%.*]] = getelementptr inbounds [2 x i32], ptr [[V_ALLOCA]], i32 0, i32 045; CHECK-NEXT: store i32 [[V_EXTRACT]], ptr [[V_INDEX]], align 446; CHECK-NEXT: [[V_EXTRACT10:%.*]] = extractelement <2 x i32> [[V]], i64 147; CHECK-NEXT: [[V_INDEX1:%.*]] = getelementptr inbounds [2 x i32], ptr [[V_ALLOCA]], i32 0, i32 148; CHECK-NEXT: store i32 [[V_EXTRACT10]], ptr [[V_INDEX1]], align 449; CHECK-NEXT: [[V3_INDEXJ:%.*]] = getelementptr inbounds [2 x i32], ptr [[V_ALLOCA]], i32 0, i32 [[J]]50; CHECK-NEXT: [[V3_LOAD:%.*]] = load i32, ptr [[V3_INDEXJ]], align 451; CHECK-NEXT: store i32 3, ptr [[V3_INDEXJ]], align 452; CHECK-NEXT: [[V3_LOAD0:%.*]] = load i32, ptr [[V_INDEX]], align 453; CHECK-NEXT: [[V3_INSERT0:%.*]] = insertelement <2 x i32> poison, i32 [[V3_LOAD0]], i32 054; CHECK-NEXT: [[V3_LOAD1:%.*]] = load i32, ptr [[V_INDEX1]], align 455; CHECK-NEXT: [[V3_INSERT1:%.*]] = insertelement <2 x i32> [[V3_INSERT0]], i32 [[V3_LOAD1]], i32 156; CHECK-NEXT: store i32 [[V3_LOAD]], ptr [[V3_INDEXJ]], align 457; CHECK-NEXT: ret void58;59 %v0_0 = insertelement <2 x i32> poison, i32 0, i32 060 %v0 = insertelement <2 x i32> %v0_0, i32 0, i32 161 %cond = icmp eq i32 %c, 162 br i1 %cond, label %if, label %else63if:64 %v1 = insertelement <2 x i32> %v0, i32 1, i32 %i65 br label %exit66else:67 %v2 = insertelement <2 x i32> %v0, i32 2, i32 %i68 br label %exit69exit:70 %v = phi <2 x i32> [ %v1, %if ], [ %v2, %else ]71 %v3 = insertelement <2 x i32> %v, i32 3, i32 %j72 ret void73}74 75; Allocas can be reused across insert/extractelement instructions on the same vector76define void @test_alloca_reuse(<3 x i32> %v, i32 %a, i32 %i) {77; CHECK-LABEL: define void @test_alloca_reuse(78; CHECK-SAME: <3 x i32> [[V:%.*]], i32 [[A:%.*]], i32 [[I:%.*]]) {79; CHECK-NEXT: [[EE1_ALLOCA:%.*]] = alloca [3 x i32], align 480; CHECK-NEXT: [[EE1_EXTRACT:%.*]] = extractelement <3 x i32> [[V]], i64 081; CHECK-NEXT: [[EE1_INDEX:%.*]] = getelementptr inbounds [3 x i32], ptr [[EE1_ALLOCA]], i32 0, i32 082; CHECK-NEXT: store i32 [[EE1_EXTRACT]], ptr [[EE1_INDEX]], align 483; CHECK-NEXT: [[EE1_EXTRACT1:%.*]] = extractelement <3 x i32> [[V]], i64 184; CHECK-NEXT: [[EE1_INDEX2:%.*]] = getelementptr inbounds [3 x i32], ptr [[EE1_ALLOCA]], i32 0, i32 185; CHECK-NEXT: store i32 [[EE1_EXTRACT1]], ptr [[EE1_INDEX2]], align 486; CHECK-NEXT: [[EE1_EXTRACT3:%.*]] = extractelement <3 x i32> [[V]], i64 287; CHECK-NEXT: [[EE1_INDEX4:%.*]] = getelementptr inbounds [3 x i32], ptr [[EE1_ALLOCA]], i32 0, i32 288; CHECK-NEXT: store i32 [[EE1_EXTRACT3]], ptr [[EE1_INDEX4]], align 489; CHECK-NEXT: [[EE1_INDEX5:%.*]] = getelementptr inbounds [3 x i32], ptr [[EE1_ALLOCA]], i32 0, i32 [[I]]90; CHECK-NEXT: [[EE1_LOAD:%.*]] = load i32, ptr [[EE1_INDEX5]], align 491; CHECK-NEXT: [[IE1_DYNINDEX:%.*]] = getelementptr inbounds [3 x i32], ptr [[EE1_ALLOCA]], i32 0, i32 [[I]]92; CHECK-NEXT: [[IE1_LOAD1:%.*]] = load i32, ptr [[IE1_DYNINDEX]], align 493; CHECK-NEXT: store i32 [[A]], ptr [[IE1_DYNINDEX]], align 494; CHECK-NEXT: [[IE1_LOAD:%.*]] = load i32, ptr [[EE1_INDEX]], align 495; CHECK-NEXT: [[IE1_INSERT:%.*]] = insertelement <3 x i32> poison, i32 [[IE1_LOAD]], i32 096; CHECK-NEXT: [[IE1_LOAD6:%.*]] = load i32, ptr [[EE1_INDEX2]], align 497; CHECK-NEXT: [[IE1_INSERT7:%.*]] = insertelement <3 x i32> [[IE1_INSERT]], i32 [[IE1_LOAD6]], i32 198; CHECK-NEXT: [[IE1_LOAD8:%.*]] = load i32, ptr [[EE1_INDEX4]], align 499; CHECK-NEXT: [[IE1_INSERT9:%.*]] = insertelement <3 x i32> [[IE1_INSERT7]], i32 [[IE1_LOAD8]], i32 2100; CHECK-NEXT: store i32 [[IE1_LOAD1]], ptr [[IE1_DYNINDEX]], align 4101; CHECK-NEXT: ret void102;103 %ee1 = extractelement <3 x i32> %v, i32 %i104 %ie1 = insertelement <3 x i32> %v, i32 %a, i32 %i105 ret void106}107 108define float @extract_float_vec_dynamic(<4 x float> %v, i32 %i) {109; CHECK-LABEL: define float @extract_float_vec_dynamic(110; CHECK-SAME: <4 x float> [[V:%.*]], i32 [[I:%.*]]) {111; CHECK-NEXT: [[EE_ALLOCA:%.*]] = alloca [4 x float], align 4112; CHECK-NEXT: [[EE_EXTRACT:%.*]] = extractelement <4 x float> [[V]], i64 0113; CHECK-NEXT: [[EE_INDEX:%.*]] = getelementptr inbounds [4 x float], ptr [[EE_ALLOCA]], i32 0, i32 0114; CHECK-NEXT: store float [[EE_EXTRACT]], ptr [[EE_INDEX]], align 4115; CHECK-NEXT: [[EE_EXTRACT1:%.*]] = extractelement <4 x float> [[V]], i64 1116; CHECK-NEXT: [[EE_INDEX2:%.*]] = getelementptr inbounds [4 x float], ptr [[EE_ALLOCA]], i32 0, i32 1117; CHECK-NEXT: store float [[EE_EXTRACT1]], ptr [[EE_INDEX2]], align 4118; CHECK-NEXT: [[EE_EXTRACT3:%.*]] = extractelement <4 x float> [[V]], i64 2119; CHECK-NEXT: [[EE_INDEX4:%.*]] = getelementptr inbounds [4 x float], ptr [[EE_ALLOCA]], i32 0, i32 2120; CHECK-NEXT: store float [[EE_EXTRACT3]], ptr [[EE_INDEX4]], align 4121; CHECK-NEXT: [[EE_EXTRACT5:%.*]] = extractelement <4 x float> [[V]], i64 3122; CHECK-NEXT: [[EE_INDEX6:%.*]] = getelementptr inbounds [4 x float], ptr [[EE_ALLOCA]], i32 0, i32 3123; CHECK-NEXT: store float [[EE_EXTRACT5]], ptr [[EE_INDEX6]], align 4124; CHECK-NEXT: [[EE_INDEX7:%.*]] = getelementptr inbounds [4 x float], ptr [[EE_ALLOCA]], i32 0, i32 [[I]]125; CHECK-NEXT: [[EE_LOAD:%.*]] = load float, ptr [[EE_INDEX7]], align 4126; CHECK-NEXT: ret float [[EE_LOAD]]127;128 %ee = extractelement <4 x float> %v, i32 %i129 ret float %ee130}131 132define <3 x i32> @insert_i32_vec_dynamic(<3 x i32> %v, i32 %a, i32 %i) {133; CHECK-LABEL: define <3 x i32> @insert_i32_vec_dynamic(134; CHECK-SAME: <3 x i32> [[V:%.*]], i32 [[A:%.*]], i32 [[I:%.*]]) {135; CHECK-NEXT: [[IE_ALLOCA:%.*]] = alloca [3 x i32], align 4136; CHECK-NEXT: [[IE_EXTRACT:%.*]] = extractelement <3 x i32> [[V]], i64 0137; CHECK-NEXT: [[IE_INDEX:%.*]] = getelementptr inbounds [3 x i32], ptr [[IE_ALLOCA]], i32 0, i32 0138; CHECK-NEXT: store i32 [[IE_EXTRACT]], ptr [[IE_INDEX]], align 4139; CHECK-NEXT: [[IE_EXTRACT1:%.*]] = extractelement <3 x i32> [[V]], i64 1140; CHECK-NEXT: [[IE_INDEX2:%.*]] = getelementptr inbounds [3 x i32], ptr [[IE_ALLOCA]], i32 0, i32 1141; CHECK-NEXT: store i32 [[IE_EXTRACT1]], ptr [[IE_INDEX2]], align 4142; CHECK-NEXT: [[IE_EXTRACT3:%.*]] = extractelement <3 x i32> [[V]], i64 2143; CHECK-NEXT: [[IE_INDEX4:%.*]] = getelementptr inbounds [3 x i32], ptr [[IE_ALLOCA]], i32 0, i32 2144; CHECK-NEXT: store i32 [[IE_EXTRACT3]], ptr [[IE_INDEX4]], align 4145; CHECK-NEXT: [[IE_DYNINDEX:%.*]] = getelementptr inbounds [3 x i32], ptr [[IE_ALLOCA]], i32 0, i32 [[I]]146; CHECK-NEXT: [[IE_LOAD1:%.*]] = load i32, ptr [[IE_DYNINDEX]], align 4147; CHECK-NEXT: store i32 [[A]], ptr [[IE_DYNINDEX]], align 4148; CHECK-NEXT: [[IE_LOAD:%.*]] = load i32, ptr [[IE_INDEX]], align 4149; CHECK-NEXT: [[IE_INSERT:%.*]] = insertelement <3 x i32> poison, i32 [[IE_LOAD]], i32 0150; CHECK-NEXT: [[IE_LOAD5:%.*]] = load i32, ptr [[IE_INDEX2]], align 4151; CHECK-NEXT: [[IE_INSERT6:%.*]] = insertelement <3 x i32> [[IE_INSERT]], i32 [[IE_LOAD5]], i32 1152; CHECK-NEXT: [[IE_LOAD7:%.*]] = load i32, ptr [[IE_INDEX4]], align 4153; CHECK-NEXT: [[IE_INSERT8:%.*]] = insertelement <3 x i32> [[IE_INSERT6]], i32 [[IE_LOAD7]], i32 2154; CHECK-NEXT: store i32 [[IE_LOAD1]], ptr [[IE_DYNINDEX]], align 4155; CHECK-NEXT: ret <3 x i32> [[IE_INSERT8]]156;157 %ie = insertelement <3 x i32> %v, i32 %a, i32 %i158 ret <3 x i32> %ie159}160 161; An extractelement with a constant index should not be converted to array form162define i16 @extract_i16_vec_constant(<4 x i16> %v) {163; CHECK-LABEL: define i16 @extract_i16_vec_constant(164; CHECK-SAME: <4 x i16> [[V:%.*]]) {165; CHECK-NEXT: [[EE:%.*]] = extractelement <4 x i16> [[V]], i32 1166; CHECK-NEXT: ret i16 [[EE]]167;168 %ee = extractelement <4 x i16> %v, i32 1169 ret i16 %ee170}171 172; An insertelement with a constant index should not be converted to array form173define <2 x half> @insert_half_vec_constant(<2 x half> %v, half %a) {174; CHECK-LABEL: define <2 x half> @insert_half_vec_constant(175; CHECK-SAME: <2 x half> [[V:%.*]], half [[A:%.*]]) {176; CHECK-NEXT: [[IE:%.*]] = insertelement <2 x half> [[V]], half [[A]], i32 1177; CHECK-NEXT: ret <2 x half> [[IE]]178;179 %ie = insertelement <2 x half> %v, half %a, i32 1180 ret <2 x half> %ie181}182 183