1 // RUN: %clang_cc1 -emit-llvm %s -o - -fcuda-is-device -triple nvptx-unknown-unknown | FileCheck %s 2 3 // Verifies Clang emits correct address spaces and addrspacecast instructions 4 // for CUDA code. 5 6 #include "Inputs/cuda.h" 7 8 // CHECK: @i = addrspace(1) externally_initialized global 9 __device__ int i; 10 11 // CHECK: @j = addrspace(4) externally_initialized global 12 __constant__ int j; 13 14 // CHECK: @k = addrspace(3) global 15 __shared__ int k; 16 17 struct MyStruct { 18 int data1; 19 int data2; 20 }; 21 22 // CHECK: @_ZZ5func0vE1a = internal addrspace(3) global %struct.MyStruct zeroinitializer 23 // CHECK: @_ZZ5func1vE1a = internal addrspace(3) global float 0.000000e+00 24 // CHECK: @_ZZ5func2vE1a = internal addrspace(3) global [256 x float] zeroinitializer 25 // CHECK: @_ZZ5func3vE1a = internal addrspace(3) global float 0.000000e+00 26 // CHECK: @_ZZ5func4vE1a = internal addrspace(3) global float 0.000000e+00 27 // CHECK: @b = addrspace(3) global float undef 28 29 __device__ void foo() { 30 // CHECK: load i32, i32* addrspacecast (i32 addrspace(1)* @i to i32*) 31 i++; 32 33 // CHECK: load i32, i32* addrspacecast (i32 addrspace(4)* @j to i32*) 34 j++; 35 36 // CHECK: load i32, i32* addrspacecast (i32 addrspace(3)* @k to i32*) 37 k++; 38 39 __shared__ int lk; 40 // CHECK: load i32, i32* addrspacecast (i32 addrspace(3)* @_ZZ3foovE2lk to i32*) 41 lk++; 42 } 43 44 __device__ void func0() { 45 __shared__ MyStruct a; 46 MyStruct *ap = &a; // composite type 47 ap->data1 = 1; 48 ap->data2 = 2; 49 } 50 // CHECK: define void @_Z5func0v() 51 // CHECK: store %struct.MyStruct* addrspacecast (%struct.MyStruct addrspace(3)* @_ZZ5func0vE1a to %struct.MyStruct*), %struct.MyStruct** %ap 52 53 __device__ void callee(float *ap) { 54 *ap = 1.0f; 55 } 56 57 __device__ void func1() { 58 __shared__ float a; 59 callee(&a); // implicit cast from parameters 60 } 61 // CHECK: define void @_Z5func1v() 62 // CHECK: call void @_Z6calleePf(float* addrspacecast (float addrspace(3)* @_ZZ5func1vE1a to float*)) 63 64 __device__ void func2() { 65 __shared__ float a[256]; 66 float *ap = &a[128]; // implicit cast from a decayed array 67 *ap = 1.0f; 68 } 69 // CHECK: define void @_Z5func2v() 70 // CHECK: store float* getelementptr inbounds ([256 x float], [256 x float]* addrspacecast ([256 x float] addrspace(3)* @_ZZ5func2vE1a to [256 x float]*), i32 0, i32 128), float** %ap 71 72 __device__ void func3() { 73 __shared__ float a; 74 float *ap = reinterpret_cast<float *>(&a); // explicit cast 75 *ap = 1.0f; 76 } 77 // CHECK: define void @_Z5func3v() 78 // CHECK: store float* addrspacecast (float addrspace(3)* @_ZZ5func3vE1a to float*), float** %ap 79 80 __device__ void func4() { 81 __shared__ float a; 82 float *ap = (float *)&a; // explicit c-style cast 83 *ap = 1.0f; 84 } 85 // CHECK: define void @_Z5func4v() 86 // CHECK: store float* addrspacecast (float addrspace(3)* @_ZZ5func4vE1a to float*), float** %ap 87 88 __shared__ float b; 89 90 __device__ float *func5() { 91 return &b; // implicit cast from a return value 92 } 93 // CHECK: define float* @_Z5func5v() 94 // CHECK: ret float* addrspacecast (float addrspace(3)* @b to float*) 95
