187 lines · plain
1// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py2// REQUIRES: amdgpu-registered-target3 4// RUN: %clang_cc1 -include __clang_hip_runtime_wrapper.h \5// RUN: -internal-isystem %S/../../lib/Headers/cuda_wrappers \6// RUN: -internal-isystem %S/Inputs/include \7// RUN: -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown \8// RUN: -target-cpu gfx906 -emit-llvm %s -fcuda-is-device -O1 -o - \9// RUN: -D__HIPCC_RTC__ | FileCheck -check-prefixes=DEFAULT,CORRECT-DIV-SQRT %s10 11// Check that we end up with fast math flags set on intrinsic calls12// RUN: %clang_cc1 -include __clang_hip_runtime_wrapper.h \13// RUN: -internal-isystem %S/../../lib/Headers/cuda_wrappers \14// RUN: -internal-isystem %S/Inputs/include \15// RUN: -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown \16// RUN: -target-cpu gfx906 -emit-llvm %s -fcuda-is-device -O1 -menable-no-infs \17// RUN: -menable-no-nans -o - \18// RUN: -D__HIPCC_RTC__ | FileCheck -check-prefix=FINITEONLY %s19 20// Check that we end up with fpmath metadata set on sqrt calls21// RUN: %clang_cc1 -include __clang_hip_runtime_wrapper.h \22// RUN: -internal-isystem %S/../../lib/Headers/cuda_wrappers \23// RUN: -internal-isystem %S/Inputs/include \24// RUN: -triple amdgcn-amd-amdhsa -aux-triple x86_64-unknown-unknown \25// RUN: -target-cpu gfx906 -emit-llvm %s -fcuda-is-device -O1 \26// RUN: -fno-hip-fp32-correctly-rounded-divide-sqrt -o - \27// RUN: -D__HIPCC_RTC__ | FileCheck -check-prefixes=DEFAULT,NO-CORRECT-DIV-SQRT %s28 29// DEFAULT-LABEL: @test_fma_f16(30// DEFAULT-NEXT: entry:31// DEFAULT-NEXT: [[TMP0:%.*]] = tail call contract noundef half @llvm.fma.f16(half [[X:%.*]], half [[Y:%.*]], half [[Z:%.*]])32// DEFAULT-NEXT: ret half [[TMP0]]33//34// FINITEONLY-LABEL: @test_fma_f16(35// FINITEONLY-NEXT: entry:36// FINITEONLY-NEXT: [[TMP0:%.*]] = tail call nnan ninf contract noundef half @llvm.fma.f16(half nofpclass(nan inf) [[X:%.*]], half nofpclass(nan inf) [[Y:%.*]], half nofpclass(nan inf) [[Z:%.*]])37// FINITEONLY-NEXT: ret half [[TMP0]]38//39extern "C" __device__ _Float16 test_fma_f16(_Float16 x, _Float16 y,40 _Float16 z) {41 return fma(x, y, z);42}43 44// DEFAULT-LABEL: @test_pow_f16(45// DEFAULT-NEXT: entry:46// DEFAULT-NEXT: [[CALL_I:%.*]] = tail call contract noundef half @__ocml_pown_f16(half noundef [[X:%.*]], i32 noundef [[Y:%.*]]) #[[ATTR9:[0-9]+]]47// DEFAULT-NEXT: ret half [[CALL_I]]48//49// FINITEONLY-LABEL: @test_pow_f16(50// FINITEONLY-NEXT: entry:51// FINITEONLY-NEXT: [[CALL_I:%.*]] = tail call nnan ninf contract noundef nofpclass(nan inf) half @__ocml_pown_f16(half noundef nofpclass(nan inf) [[X:%.*]], i32 noundef [[Y:%.*]]) #[[ATTR9:[0-9]+]]52// FINITEONLY-NEXT: ret half [[CALL_I]]53//54extern "C" __device__ _Float16 test_pow_f16(_Float16 x, int y) {55 return pow(x, y);56}57 58// DEFAULT-LABEL: @test_fabs_f32(59// DEFAULT-NEXT: entry:60// DEFAULT-NEXT: [[TMP0:%.*]] = tail call contract noundef float @llvm.fabs.f32(float [[X:%.*]])61// DEFAULT-NEXT: ret float [[TMP0]]62//63// FINITEONLY-LABEL: @test_fabs_f32(64// FINITEONLY-NEXT: entry:65// FINITEONLY-NEXT: [[TMP0:%.*]] = tail call nnan ninf contract noundef float @llvm.fabs.f32(float nofpclass(nan inf) [[X:%.*]])66// FINITEONLY-NEXT: ret float [[TMP0]]67//68extern "C" __device__ float test_fabs_f32(float x) {69 return fabs(x);70}71 72// DEFAULT-LABEL: @test_sin_f32(73// DEFAULT-NEXT: entry:74// DEFAULT-NEXT: [[CALL_I1:%.*]] = tail call contract noundef float @__ocml_sin_f32(float noundef [[X:%.*]]) #[[ATTR10:[0-9]+]]75// DEFAULT-NEXT: ret float [[CALL_I1]]76//77// FINITEONLY-LABEL: @test_sin_f32(78// FINITEONLY-NEXT: entry:79// FINITEONLY-NEXT: [[CALL_I1:%.*]] = tail call nnan ninf contract noundef nofpclass(nan inf) float @__ocml_sin_f32(float noundef nofpclass(nan inf) [[X:%.*]]) #[[ATTR10:[0-9]+]]80// FINITEONLY-NEXT: ret float [[CALL_I1]]81//82extern "C" __device__ float test_sin_f32(float x) {83 return sin(x);84}85 86// DEFAULT-LABEL: @test_cos_f32(87// DEFAULT-NEXT: entry:88// DEFAULT-NEXT: [[CALL_I1:%.*]] = tail call contract noundef float @__ocml_cos_f32(float noundef [[X:%.*]]) #[[ATTR10]]89// DEFAULT-NEXT: ret float [[CALL_I1]]90//91// FINITEONLY-LABEL: @test_cos_f32(92// FINITEONLY-NEXT: entry:93// FINITEONLY-NEXT: [[CALL_I1:%.*]] = tail call nnan ninf contract noundef nofpclass(nan inf) float @__ocml_cos_f32(float noundef nofpclass(nan inf) [[X:%.*]]) #[[ATTR10]]94// FINITEONLY-NEXT: ret float [[CALL_I1]]95//96extern "C" __device__ float test_cos_f32(float x) {97 return cos(x);98}99 100// Check user defined type which can be converted to float and double but not101// specializes __numeric_type will not cause ambiguity diagnostics.102struct user_bfloat16 {103 __host__ __device__ user_bfloat16(float);104 operator float();105 operator double();106};107 108namespace user_namespace {109// DEFAULT-LABEL: @_ZN14user_namespace3fmaE13user_bfloat16S0_S0_(110// DEFAULT-NEXT: entry:111// DEFAULT-NEXT: ret void112//113// FINITEONLY-LABEL: @_ZN14user_namespace3fmaE13user_bfloat16S0_S0_(114// FINITEONLY-NEXT: entry:115// FINITEONLY-NEXT: ret void116//117 __device__ user_bfloat16 fma(const user_bfloat16 a, const user_bfloat16 b, const user_bfloat16 c) {118 return a;119 }120 121// DEFAULT-LABEL: @_ZN14user_namespace8test_fmaEv(122// DEFAULT-NEXT: entry:123// DEFAULT-NEXT: [[A:%.*]] = alloca [[STRUCT_USER_BFLOAT16:%.*]], align 1, addrspace(5)124// DEFAULT-NEXT: [[B:%.*]] = alloca [[STRUCT_USER_BFLOAT16]], align 1, addrspace(5)125// DEFAULT-NEXT: [[A_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A]] to ptr126// DEFAULT-NEXT: [[B_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[B]] to ptr127// DEFAULT-NEXT: call void @llvm.lifetime.start.p5(ptr addrspace(5) [[A]]) #[[ATTR11:[0-9]+]]128// DEFAULT-NEXT: call void @_ZN13user_bfloat16C1Ef(ptr noundef nonnull align 1 dereferenceable(1) [[A_ASCAST]], float noundef 1.000000e+00) #[[ATTR10]]129// DEFAULT-NEXT: call void @llvm.lifetime.start.p5(ptr addrspace(5) [[B]]) #[[ATTR11]]130// DEFAULT-NEXT: call void @_ZN13user_bfloat16C1Ef(ptr noundef nonnull align 1 dereferenceable(1) [[B_ASCAST]], float noundef 2.000000e+00) #[[ATTR10]]131// DEFAULT-NEXT: call void @llvm.lifetime.end.p5(ptr addrspace(5) [[B]]) #[[ATTR11]]132// DEFAULT-NEXT: call void @llvm.lifetime.end.p5(ptr addrspace(5) [[A]]) #[[ATTR11]]133// DEFAULT-NEXT: ret void134//135// FINITEONLY-LABEL: @_ZN14user_namespace8test_fmaEv(136// FINITEONLY-NEXT: entry:137// FINITEONLY-NEXT: [[A:%.*]] = alloca [[STRUCT_USER_BFLOAT16:%.*]], align 1, addrspace(5)138// FINITEONLY-NEXT: [[B:%.*]] = alloca [[STRUCT_USER_BFLOAT16]], align 1, addrspace(5)139// FINITEONLY-NEXT: [[A_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[A]] to ptr140// FINITEONLY-NEXT: [[B_ASCAST:%.*]] = addrspacecast ptr addrspace(5) [[B]] to ptr141// FINITEONLY-NEXT: call void @llvm.lifetime.start.p5(ptr addrspace(5) [[A]]) #[[ATTR11:[0-9]+]]142// FINITEONLY-NEXT: call void @_ZN13user_bfloat16C1Ef(ptr noundef nonnull align 1 dereferenceable(1) [[A_ASCAST]], float noundef nofpclass(nan inf) 1.000000e+00) #[[ATTR10]]143// FINITEONLY-NEXT: call void @llvm.lifetime.start.p5(ptr addrspace(5) [[B]]) #[[ATTR11]]144// FINITEONLY-NEXT: call void @_ZN13user_bfloat16C1Ef(ptr noundef nonnull align 1 dereferenceable(1) [[B_ASCAST]], float noundef nofpclass(nan inf) 2.000000e+00) #[[ATTR10]]145// FINITEONLY-NEXT: call void @llvm.lifetime.end.p5(ptr addrspace(5) [[B]]) #[[ATTR11]]146// FINITEONLY-NEXT: call void @llvm.lifetime.end.p5(ptr addrspace(5) [[A]]) #[[ATTR11]]147// FINITEONLY-NEXT: ret void148//149 __global__ void test_fma() {150 user_bfloat16 a = 1.0f, b = 2.0f;151 fma(a, b, b);152 }153}154 155// CORRECT-DIV-SQRT-LABEL: @test_sqrt_f32(156// CORRECT-DIV-SQRT-NEXT: entry:157// CORRECT-DIV-SQRT-NEXT: [[TMP0:%.*]] = tail call contract noundef float @llvm.sqrt.f32(float [[X:%.*]])158// CORRECT-DIV-SQRT-NEXT: ret float [[TMP0]]159//160// FINITEONLY-LABEL: @test_sqrt_f32(161// FINITEONLY-NEXT: entry:162// FINITEONLY-NEXT: [[TMP0:%.*]] = tail call nnan ninf contract noundef float @llvm.sqrt.f32(float nofpclass(nan inf) [[X:%.*]])163// FINITEONLY-NEXT: ret float [[TMP0]]164//165// NO-CORRECT-DIV-SQRT-LABEL: @test_sqrt_f32(166// NO-CORRECT-DIV-SQRT-NEXT: entry:167// NO-CORRECT-DIV-SQRT-NEXT: [[TMP0:%.*]] = tail call contract noundef float @llvm.sqrt.f32(float [[X:%.*]]), !fpmath [[META4:![0-9]+]]168// NO-CORRECT-DIV-SQRT-NEXT: ret float [[TMP0]]169//170extern "C" __device__ float test_sqrt_f32(float x) {171 return sqrt(x);172}173 174// DEFAULT-LABEL: @test_sqrt_f64(175// DEFAULT-NEXT: entry:176// DEFAULT-NEXT: [[TMP0:%.*]] = tail call contract noundef double @llvm.sqrt.f64(double [[X:%.*]])177// DEFAULT-NEXT: ret double [[TMP0]]178//179// FINITEONLY-LABEL: @test_sqrt_f64(180// FINITEONLY-NEXT: entry:181// FINITEONLY-NEXT: [[TMP0:%.*]] = tail call nnan ninf contract noundef double @llvm.sqrt.f64(double nofpclass(nan inf) [[X:%.*]])182// FINITEONLY-NEXT: ret double [[TMP0]]183//184extern "C" __device__ double test_sqrt_f64(double x) {185 return sqrt(x);186}187