1 | // Test for linking with CUDA's libdevice as outlined in |
2 | // http://llvm.org/docs/NVPTXUsage.html#linking-with-libdevice |
3 | // |
4 | // REQUIRES: nvptx-registered-target |
5 | // |
6 | // Prepare bitcode file to link with |
7 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -emit-llvm-bc \ |
8 | // RUN: -disable-llvm-passes -o %t.bc %S/Inputs/device-code.ll |
9 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -emit-llvm-bc \ |
10 | // RUN: -disable-llvm-passes -o %t-2.bc %S/Inputs/device-code-2.ll |
11 | // |
12 | // Make sure function in device-code gets linked in and internalized. |
13 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -fcuda-is-device \ |
14 | // RUN: -mlink-builtin-bitcode %t.bc -emit-llvm \ |
15 | // RUN: -disable-llvm-passes -o - %s \ |
16 | // RUN: | FileCheck %s -check-prefix CHECK-IR |
17 | |
18 | // Make sure legacy flag name works |
19 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -fcuda-is-device \ |
20 | // RUN: -mlink-cuda-bitcode %t.bc -emit-llvm \ |
21 | // RUN: -disable-llvm-passes -o - %s \ |
22 | // RUN: | FileCheck %s -check-prefix CHECK-IR |
23 | // |
24 | // Make sure we can link two bitcode files. |
25 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -fcuda-is-device \ |
26 | // RUN: -mlink-builtin-bitcode %t.bc -mlink-builtin-bitcode %t-2.bc \ |
27 | // RUN: -emit-llvm -disable-llvm-passes -o - %s \ |
28 | // RUN: | FileCheck %s -check-prefix CHECK-IR -check-prefix CHECK-IR-2 |
29 | // |
30 | // Make sure function in device-code gets linked but is not internalized |
31 | // without -fcuda-uses-libdevice |
32 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -fcuda-is-device \ |
33 | // RUN: -mlink-bitcode-file %t.bc -emit-llvm \ |
34 | // RUN: -disable-llvm-passes -o - %s \ |
35 | // RUN: | FileCheck %s -check-prefix CHECK-IR-NLD |
36 | // |
37 | // Make sure NVVMReflect pass is enabled in NVPTX back-end. |
38 | // RUN: %clang_cc1 -triple nvptx-unknown-cuda -fcuda-is-device \ |
39 | // RUN: -mlink-builtin-bitcode %t.bc -S -o /dev/null %s \ |
40 | // RUN: -mllvm -debug-pass=Structure 2>&1 \ |
41 | // RUN: | FileCheck %s -check-prefix CHECK-REFLECT |
42 | |
43 | #include "Inputs/cuda.h" |
44 | |
45 | __device__ float device_mul_or_add(float a, float b); |
46 | extern "C" __device__ double __nv_sin(double x); |
47 | extern "C" __device__ double __nv_exp(double x); |
48 | |
49 | // CHECK-IR-LABEL: define void @_Z26should_not_be_internalizedPf( |
50 | // CHECK-PTX-LABEL: .visible .func _Z26should_not_be_internalizedPf( |
51 | __device__ void should_not_be_internalized(float *data) {} |
52 | |
53 | // Make sure kernel call has not been internalized. |
54 | // CHECK-IR-LABEL: define void @_Z6kernelPfS_ |
55 | // CHECK-PTX-LABEL: .visible .entry _Z6kernelPfS_( |
56 | __global__ __attribute__((used)) void kernel(float *out, float *in) { |
57 | *out = device_mul_or_add(in[0], in[1]); |
58 | *out += __nv_exp(__nv_sin(*out)); |
59 | should_not_be_internalized(out); |
60 | } |
61 | |
62 | // Make sure device_mul_or_add() is present in IR, is internal and |
63 | // calls __nvvm_reflect(). |
64 | // CHECK-IR-LABEL: define internal float @_Z17device_mul_or_addff( |
65 | // CHECK-IR-NLD-LABEL: define float @_Z17device_mul_or_addff( |
66 | // CHECK-IR: call i32 @__nvvm_reflect |
67 | // CHECK-IR: ret float |
68 | |
69 | // Make sure we've linked in and internalized only needed functions |
70 | // from the second bitcode file. |
71 | // CHECK-IR-2-LABEL: define internal double @__nv_sin |
72 | // CHECK-IR-2-LABEL: define internal double @__nv_exp |
73 | // CHECK-IR-2-NOT: double @__unused |
74 | |
75 | // Verify that NVVMReflect pass is among the passes run by NVPTX back-end. |
76 | // CHECK-REFLECT: Replace occurrences of __nvvm_reflect() calls with 0/1 |
77 | |