124 lines · plain
1// RUN: %clang_cc1 -finclude-default-header -triple dxil-pc-shadermodel6.3-library -x hlsl -emit-llvm -disable-llvm-passes -o - %s | FileCheck %s2 3// vector flat cast from array4// CHECK-LABEL: define void {{.*}}call25// CHECK: [[A:%.*]] = alloca [2 x [1 x i32]], align 46// CHECK-NEXT: [[B:%.*]] = alloca <2 x i32>, align 87// CHECK-NEXT: [[Tmp:%.*]] = alloca [2 x [1 x i32]], align 48// CHECK-NEXT: [[Tmp2:%.*]] = alloca <2 x i32>, align 89// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[A]], ptr align 4 {{.*}}, i32 8, i1 false)10// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[Tmp]], ptr align 4 [[A]], i32 8, i1 false)11// CHECK-NEXT: [[G1:%.*]] = getelementptr inbounds [2 x [1 x i32]], ptr [[Tmp]], i32 0, i32 0, i32 012// CHECK-NEXT: [[G2:%.*]] = getelementptr inbounds [2 x [1 x i32]], ptr [[Tmp]], i32 0, i32 1, i32 013// CHECK-NEXT: [[C:%.*]] = load <2 x i32>, ptr [[Tmp2]], align 814// CHECK-NEXT: [[L:%.*]] = load i32, ptr [[G1]], align 415// CHECK-NEXT: [[D:%.*]] = insertelement <2 x i32> [[C]], i32 [[L]], i64 016// CHECK-NEXT: [[L2:%.*]] = load i32, ptr [[G2]], align 417// CHECK-NEXT: [[E:%.*]] = insertelement <2 x i32> [[D]], i32 [[L2]], i64 118// CHECK-NEXT: store <2 x i32> [[E]], ptr [[B]], align 819export void call2() {20 int A[2][1] = {{1},{2}};21 int2 B = (int2)A;22}23 24struct S {25 int X;26 float Y;27};28 29// vector flat cast from struct30// CHECK-LABEL: define void {{.*}}call331// CHECK: [[s:%.*]] = alloca %struct.S, align 132// CHECK-NEXT: [[A:%.*]] = alloca <2 x i32>, align 833// CHECK-NEXT: [[Tmp:%.*]] = alloca %struct.S, align 134// CHECK-NEXT: [[Tmp2:%.*]] = alloca <2 x i32>, align 835// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 1 [[s]], ptr align 1 {{.*}}, i32 8, i1 false)36// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 1 [[Tmp]], ptr align 1 [[s]], i32 8, i1 false)37// CHECK-NEXT: [[G1:%.*]] = getelementptr inbounds %struct.S, ptr [[Tmp]], i32 0, i32 038// CHECK-NEXT: [[G2:%.*]] = getelementptr inbounds %struct.S, ptr [[Tmp]], i32 0, i32 139// CHECK-NEXT: [[B:%.*]] = load <2 x i32>, ptr [[Tmp2]], align 840// CHECK-NEXT: [[L:%.*]] = load i32, ptr [[G1]], align 441// CHECK-NEXT: [[C:%.*]] = insertelement <2 x i32> [[B]], i32 [[L]], i64 042// CHECK-NEXT: [[L2:%.*]] = load float, ptr [[G2]], align 443// CHECK-NEXT: [[D:%.*]] = fptosi float [[L2]] to i3244// CHECK-NEXT: [[E:%.*]] = insertelement <2 x i32> [[C]], i32 [[D]], i64 145// CHECK-NEXT: store <2 x i32> [[E]], ptr [[A]], align 846export void call3() {47 S s = {1, 2.0};48 int2 A = (int2)s;49}50 51// truncate array to scalar52// CHECK-LABEL: define void {{.*}}call453// CHECK: [[A:%.*]] = alloca [2 x i32], align 454// CHECK-NEXT: [[B:%.*]] = alloca i32, align 455// CHECK-NEXT: [[Tmp:%.*]] = alloca [2 x i32], align 456// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[A]], ptr align 4 {{.*}}, i32 8, i1 false)57// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 4 [[Tmp]], ptr align 4 [[A]], i32 8, i1 false)58// CHECK-NEXT: [[G1:%.*]] = getelementptr inbounds [2 x i32], ptr [[Tmp]], i32 0, i32 059// CHECK-NEXT: [[G2:%.*]] = getelementptr inbounds [2 x i32], ptr [[Tmp]], i32 0, i32 160// CHECK-NEXT: [[L:%.*]] = load i32, ptr [[G1]], align 461// CHECK-NEXT: store i32 [[L]], ptr [[B]], align 462export void call4() {63 int A[2] = {1,2};64 int B = (int)A;65}66 67// truncate struct to scalar68// CHECK-LABEL: define void {{.*}}call569// CHECK: [[s:%.*]] = alloca %struct.S, align 170// CHECK-NEXT: [[A:%.*]] = alloca i32, align 471// CHECK-NEXT: [[Tmp:%.*]] = alloca %struct.S, align 172// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 1 [[s]], ptr align 1 {{.*}}, i32 8, i1 false)73// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 1 [[Tmp]], ptr align 1 [[s]], i32 8, i1 false)74// CHECK-NEXT: [[G1:%.*]] = getelementptr inbounds %struct.S, ptr [[Tmp]], i32 0, i32 075// CHECK-NEXT: [[G2:%.*]] = getelementptr inbounds %struct.S, ptr [[Tmp]], i32 0, i32 176// CHECK-NEXT: [[L:%.*]] = load i32, ptr [[G1]], align 477// CHECK-NEXT: store i32 [[L]], ptr [[A]], align 478export void call5() {79 S s = {1, 2.0};80 int A = (int)s;81}82 83struct BFields {84 double D;85 int E: 15;86 int : 8;87 float F;88};89 90struct Derived : BFields {91 int G;92};93 94// vector flat cast from derived struct with bitfield95// CHECK-LABEL: call696// CHECK: [[A:%.*]] = alloca <4 x i32>, align 1697// CHECK-NEXT: [[Tmp:%.*]] = alloca %struct.Derived, align 198// CHECK-NEXT: [[FlatTmp:%.*]] = alloca <4 x i32>, align 1699// CHECK-NEXT: call void @llvm.memcpy.p0.p0.i32(ptr align 1 [[Tmp]], ptr align 1 %D, i32 19, i1 false)100// CHECK-NEXT: [[Gep:%.*]] = getelementptr inbounds %struct.Derived, ptr [[Tmp]], i32 0, i32 0101// CHECK-NEXT: [[E:%.*]] = getelementptr inbounds nuw %struct.BFields, ptr [[Gep]], i32 0, i32 1102// CHECK-NEXT: [[Gep1:%.*]] = getelementptr inbounds %struct.Derived, ptr [[Tmp]], i32 0, i32 0, i32 0103// CHECK-NEXT: [[Gep2:%.*]] = getelementptr inbounds %struct.Derived, ptr [[Tmp]], i32 0, i32 0, i32 2104// CHECK-NEXT: [[Gep3:%.*]] = getelementptr inbounds %struct.Derived, ptr [[Tmp]], i32 0, i32 1105// CHECK-NEXT: [[Z:%.*]] = load <4 x i32>, ptr [[FlatTmp]], align 16106// CHECK-NEXT: [[Y:%.*]] = load double, ptr [[Gep1]], align 8107// CHECK-NEXT: [[C:%.*]] = fptosi double [[Y]] to i32108// CHECK-NEXT: [[X:%.*]] = insertelement <4 x i32> [[Z]], i32 [[C]], i64 0109// CHECK-NEXT: [[BFL:%.*]] = load i24, ptr [[E]], align 1110// CHECK-NEXT: [[BFShl:%.*]] = shl i24 [[BFL]], 9111// CHECK-NEXT: [[BFAshr:%.*]] = ashr i24 [[BFShl]], 9112// CHECK-NEXT: [[BFC:%.*]] = sext i24 [[BFAshr]] to i32113// CHECK-NEXT: [[W:%.*]] = insertelement <4 x i32> [[X]], i32 [[BFC]], i64 1114// CHECK-NEXT: [[V:%.*]] = load float, ptr [[Gep2]], align 4115// CHECK-NEXT: [[C4:%.*]] = fptosi float [[V]] to i32116// CHECK-NEXT: [[U:%.*]] = insertelement <4 x i32> [[W]], i32 [[C4]], i64 2117// CHECK-NEXT: [[T:%.*]] = load i32, ptr [[Gep3]], align 4118// CHECK-NEXT: [[S:%.*]] = insertelement <4 x i32> [[U]], i32 [[T]], i64 3119// CHECK-NEXT: store <4 x i32> [[S]], ptr [[A]], align 16120// CHECK-NEXT: ret void121export void call6(Derived D) {122 int4 A = (int4)D;123}124