1 // REQUIRES: x86-registered-target 2 // REQUIRES: amdgpu-registered-target 3 4 // RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -fcuda-is-device -emit-llvm -x hip %s -o - | FileCheck --check-prefixes=COMMON,CHECK %s 5 // RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -fcuda-is-device -emit-llvm -x hip %s -disable-O0-optnone -o - | opt -S -O2 | FileCheck %s --check-prefixes=COMMON,OPT 6 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -emit-llvm -x hip %s -o - | FileCheck -check-prefix=HOST %s 7 8 #include "Inputs/cuda.h" 9 10 // Coerced struct from `struct S` without all generic pointers lowered into 11 // global ones. 12 // COMMON: %struct.S.coerce = type { i32 addrspace(1)*, float addrspace(1)* } 13 // COMMON: %struct.T.coerce = type { [2 x float addrspace(1)*] } 14 15 // On the host-side compilation, generic pointer won't be coerced. 16 // HOST-NOT: %struct.S.coerce 17 // HOST-NOT: %struct.T.coerce 18 19 // HOST: define void @_Z22__device_stub__kernel1Pi(i32* %x) 20 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel1Pi(i32 addrspace(1)*{{.*}} %x.coerce) 21 // CHECK: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 22 // CHECK-NOT: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 23 // OPT: [[VAL:%.*]] = load i32, i32 addrspace(1)* %x.coerce, align 4 24 // OPT: [[INC:%.*]] = add nsw i32 [[VAL]], 1 25 // OPT: store i32 [[INC]], i32 addrspace(1)* %x.coerce, align 4 26 // OPT: ret void 27 __global__ void kernel1(int *x) { 28 x[0]++; 29 } 30 31 // HOST: define void @_Z22__device_stub__kernel2Ri(i32* nonnull align 4 dereferenceable(4) %x) 32 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel2Ri(i32 addrspace(1)*{{.*}} nonnull align 4 dereferenceable(4) %x.coerce) 33 // CHECK: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 34 // CHECK-NOT: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 35 // OPT: [[VAL:%.*]] = load i32, i32 addrspace(1)* %x.coerce, align 4 36 // OPT: [[INC:%.*]] = add nsw i32 [[VAL]], 1 37 // OPT: store i32 [[INC]], i32 addrspace(1)* %x.coerce, align 4 38 // OPT: ret void 39 __global__ void kernel2(int &x) { 40 x++; 41 } 42 43 // HOST: define void @_Z22__device_stub__kernel3PU3AS2iPU3AS1i(i32 addrspace(2)* %x, i32 addrspace(1)* %y) 44 // CHECK-LABEL: define amdgpu_kernel void @_Z7kernel3PU3AS2iPU3AS1i(i32 addrspace(2)*{{.*}} %x, i32 addrspace(1)*{{.*}} %y) 45 // CHECK-NOT: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 46 __global__ void kernel3(__attribute__((address_space(2))) int *x, 47 __attribute__((address_space(1))) int *y) { 48 y[0] = x[0]; 49 } 50 51 // COMMON-LABEL: define void @_Z4funcPi(i32*{{.*}} %x) 52 // CHECK-NOT: = addrspacecast [[TYPE:.*]] addrspace(1)* %{{.*}} to [[TYPE]]* 53 __device__ void func(int *x) { 54 x[0]++; 55 } 56 57 struct S { 58 int *x; 59 float *y; 60 }; 61 // `by-val` struct will be coerced into a similar struct with all generic 62 // pointers lowerd into global ones. 63 // HOST: define void @_Z22__device_stub__kernel41S(i32* %s.coerce0, float* %s.coerce1) 64 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel41S(%struct.S.coerce %s.coerce) 65 // OPT: [[P0:%.*]] = extractvalue %struct.S.coerce %s.coerce, 0 66 // OPT: [[P1:%.*]] = extractvalue %struct.S.coerce %s.coerce, 1 67 // OPT: [[V0:%.*]] = load i32, i32 addrspace(1)* [[P0]], align 4 68 // OPT: [[INC:%.*]] = add nsw i32 [[V0]], 1 69 // OPT: store i32 [[INC]], i32 addrspace(1)* [[P0]], align 4 70 // OPT: [[V1:%.*]] = load float, float addrspace(1)* [[P1]], align 4 71 // OPT: [[ADD:%.*]] = fadd contract float [[V1]], 1.000000e+00 72 // OPT: store float [[ADD]], float addrspace(1)* [[P1]], align 4 73 // OPT: ret void 74 __global__ void kernel4(struct S s) { 75 s.x[0]++; 76 s.y[0] += 1.f; 77 } 78 79 // If a pointer to struct is passed, only the pointer itself is coerced into the global one. 80 // HOST: define void @_Z22__device_stub__kernel5P1S(%struct.S* %s) 81 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel5P1S(%struct.S addrspace(1)*{{.*}} %s.coerce) 82 __global__ void kernel5(struct S *s) { 83 s->x[0]++; 84 s->y[0] += 1.f; 85 } 86 87 struct T { 88 float *x[2]; 89 }; 90 // `by-val` array is also coerced. 91 // HOST: define void @_Z22__device_stub__kernel61T(float* %t.coerce0, float* %t.coerce1) 92 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel61T(%struct.T.coerce %t.coerce) 93 // OPT: [[ARR:%.*]] = extractvalue %struct.T.coerce %t.coerce, 0 94 // OPT: [[P0:%.*]] = extractvalue [2 x float addrspace(1)*] [[ARR]], 0 95 // OPT: [[P1:%.*]] = extractvalue [2 x float addrspace(1)*] [[ARR]], 1 96 // OPT: [[V0:%.*]] = load float, float addrspace(1)* [[P0]], align 4 97 // OPT: [[ADD0:%.*]] = fadd contract float [[V0]], 1.000000e+00 98 // OPT: store float [[ADD0]], float addrspace(1)* [[P0]], align 4 99 // OPT: [[V1:%.*]] = load float, float addrspace(1)* [[P1]], align 4 100 // OPT: [[ADD1:%.*]] = fadd contract float [[V1]], 2.000000e+00 101 // OPT: store float [[ADD1]], float addrspace(1)* [[P1]], align 4 102 // OPT: ret void 103 __global__ void kernel6(struct T t) { 104 t.x[0][0] += 1.f; 105 t.x[1][0] += 2.f; 106 } 107 108 // Check that coerced pointers retain the noalias attribute when qualified with __restrict. 109 // HOST: define void @_Z22__device_stub__kernel7Pi(i32* noalias %x) 110 // COMMON-LABEL: define amdgpu_kernel void @_Z7kernel7Pi(i32 addrspace(1)* noalias{{.*}} %x.coerce) 111 __global__ void kernel7(int *__restrict x) { 112 x[0]++; 113 } 114