clang/test/Driver/hip-rdc-device-only.hip

   1 // REQUIRES: x86-registered-target
   2 // REQUIRES: amdgpu-registered-target
   3
   4 // RUN: %clang -### --target=x86_64-linux-gnu \
   5 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
   6 // RUN:   -c -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
   7 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
   8 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
   9 // RUN: 2>&1 | FileCheck -check-prefixes=COMMON,EMITBC %s
  10
  11 // With `-emit-llvm`, the output should be the same as the aforementioned line
  12 // as `-fgpu-rdc` in HIP implies `-emit-llvm`.
  13
  14 // RUN: %clang -### --target=x86_64-linux-gnu \
  15 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  16 // RUN:   -c -emit-llvm -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  17 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  18 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
  19 // RUN: 2>&1 | FileCheck -check-prefixes=COMMON,EMITBC %s
  20
  21 // With `-fno-hip-emit-relocatable`, the output should be the same as the aforementioned line
  22 // as `-fgpu-rdc` in HIP implies `-fno-hip-emit-relocatable`.
  23
  24 // RUN: %clang -### --target=x86_64-linux-gnu \
  25 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  26 // RUN:   -c -fno-hip-emit-relocatable -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  27 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  28 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
  29 // RUN: 2>&1 | FileCheck -check-prefixes=COMMON,EMITBC %s
  30
  31 // RUN: %clang -### --target=x86_64-linux-gnu \
  32 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  33 // RUN:   -S -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  34 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  35 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
  36 // RUN: 2>&1 | FileCheck -check-prefixes=COMMON,EMITLL %s
  37
  38 // With `-emit-llvm`, the output should be the same as the aforementioned line
  39 // as `-fgpu-rdc` in HIP implies `-emit-llvm`.
  40
  41 // RUN: %clang -### --target=x86_64-linux-gnu \
  42 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  43 // RUN:   -S -emit-llvm -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  44 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  45 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
  46 // RUN: 2>&1 | FileCheck -check-prefixes=COMMON,EMITLL %s
  47
  48 // With `-save-temps`, commane lines for each steps are dumped. For assembly
  49 // output, there should 3 steps (preprocessor, compile, and backend) per source
  50 // and per target, totally 12 steps.
  51
  52 // RUN: %clang -### -save-temps --target=x86_64-linux-gnu \
  53 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  54 // RUN:   -S -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  55 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu \
  56 // RUN:   %S/Inputs/hip_multiple_inputs/b.hip --gpu-bundle-output \
  57 // RUN: 2>&1 | FileCheck -check-prefix=SAVETEMP %s
  58
  59 // Check output one file without bundling cause error.
  60
  61 // RUN: not %clang -### --target=x86_64-linux-gnu \
  62 // RUN:   -x hip --cuda-gpu-arch=gfx803 --cuda-gpu-arch=gfx900 \
  63 // RUN:   -S -nogpuinc -nogpulib --cuda-device-only -fgpu-rdc \
  64 // RUN:   %S/Inputs/hip_multiple_inputs/a.cu -o %t.s --no-gpu-bundle-output \
  65 // RUN: 2>&1 | FileCheck -check-prefix=FAIL %s
  66
  67 // COMMON: [[CLANG:".*clang.*"]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  68 // COMMON-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
  69 // EMITBC-SAME: "-emit-llvm-bc"
  70 // EMITLL-SAME: "-emit-llvm"
  71 // COMMON-SAME: {{.*}} "-main-file-name" "a.cu"
  72 // COMMON-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
  73 // COMMON-SAME: "-fapply-global-visibility-to-externs"
  74 // COMMON-SAME: "-target-cpu" "gfx803"
  75 // COMMON-SAME: "-fgpu-rdc"
  76 // EMITBC-SAME: {{.*}} "-o" {{".*a.*bc"}} "-x" "hip"
  77 // EMITLL-SAME: {{.*}} "-o" {{".*a.*ll"}} "-x" "hip"
  78 // CHECK-SAME: {{.*}} {{".*a.cu"}}
  79
  80 // COMMON: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  81 // COMMON-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
  82 // EMITBC-SAME: "-emit-llvm-bc"
  83 // EMITLL-SAME: "-emit-llvm"
  84 // COMMON-SAME: {{.*}} "-main-file-name" "a.cu"
  85 // COMMON-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
  86 // COMMON-SAME: "-fapply-global-visibility-to-externs"
  87 // COMMON-SAME: "-target-cpu" "gfx900"
  88 // COMMON-SAME: "-fgpu-rdc"
  89 // EMITBC-SAME: {{.*}} "-o" {{".*a.*bc"}} "-x" "hip"
  90 // EMITLL-SAME: {{.*}} "-o" {{".*a.*ll"}} "-x" "hip"
  91 // COMMON-SAME: {{.*}} {{".*a.cu"}}
  92
  93 // COMMON: "{{.*}}clang-offload-bundler" "-type={{(bc|ll)}}"
  94 // COMMON-SAME: "-targets=hip-amdgcn-amd-amdhsa-gfx803,hip-amdgcn-amd-amdhsa-gfx900"
  95 // COMMON-SAME: "-output=a-hip-amdgcn-amd-amdhsa.{{(bc|ll)}}"
  96
  97 // COMMON: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
  98 // COMMON-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
  99 // EMITBC-SAME: "-emit-llvm-bc"
 100 // EMITLL-SAME: "-emit-llvm"
 101 // COMMON-SAME: {{.*}} "-main-file-name" "b.hip"
 102 // COMMON-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
 103 // COMMON-SAME: "-fapply-global-visibility-to-externs"
 104 // COMMON-SAME: "-target-cpu" "gfx803"
 105 // COMMON-SAME: "-fgpu-rdc"
 106 // EMITBC-SAME: {{.*}} "-o" {{".*b.*bc"}} "-x" "hip"
 107 // EMITLL-SAME: {{.*}} "-o" {{".*b.*ll"}} "-x" "hip"
 108 // COMMON-SAME: {{.*}} {{".*b.hip"}}
 109
 110 // COMMON: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa"
 111 // COMMON-SAME: "-aux-triple" "x86_64-unknown-linux-gnu"
 112 // EMITBC-SAME: "-emit-llvm-bc"
 113 // EMITLL-SAME: "-emit-llvm"
 114 // COMMON-SAME: {{.*}} "-main-file-name" "b.hip"
 115 // COMMON-SAME: "-fcuda-is-device" "-fcuda-allow-variadic-functions" "-fvisibility=hidden"
 116 // COMMON-SAME: "-fapply-global-visibility-to-externs"
 117 // COMMON-SAME: "-target-cpu" "gfx900"
 118 // COMMON-SAME: "-fgpu-rdc"
 119 // EMITBC-SAME: {{.*}} "-o" {{".*b.*bc"}} "-x" "hip"
 120 // EMITLL-SAME: {{.*}} "-o" {{".*b.*ll"}} "-x" "hip"
 121 // COMMON-SAME: {{.*}} {{".*b.hip"}}
 122
 123 // COMMON: "{{.*}}clang-offload-bundler" "-type={{(bc|ll)}}"
 124 // COMMON-SAME: "-targets=hip-amdgcn-amd-amdhsa-gfx803,hip-amdgcn-amd-amdhsa-gfx900"
 125 // COMMON-SAME: "-output=b-hip-amdgcn-amd-amdhsa.{{(bc|ll)}}"
 126
 127 // SAVETEMP: [[CLANG:".*clang.*"]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 128 // SAVETEMP-SAME: "-E"
 129 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx803"
 130 // SAVETEMP-SAME: {{.*}} "-o" [[A_GFX803_CUI:"a.*hipi"]] "-x" "hip" {{".*a.cu"}}
 131 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 132 // SAVETEMP-SAME: "-emit-llvm-bc"
 133 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx803"
 134 // SAVETEMP-SAME: {{.*}} "-o" [[A_GFX803_TMP_BC:"a.*tmp.bc"]] "-x" "hip-cpp-output" [[A_GFX803_CUI]]
 135 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 136 // SAVETEMP-SAME: "-emit-llvm"
 137 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx803"
 138 // SAVETEMP-SAME: {{.*}} "-o" {{"a.*.ll"}} "-x" "ir" [[A_GFX803_TMP_BC]]
 139
 140 // SAVETEMP: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 141 // SAVETEMP-SAME: "-E"
 142 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx900"
 143 // SAVETEMP-SAME: {{.*}} "-o" [[A_GFX900_CUI:"a.*hipi"]] "-x" "hip" {{".*a.cu"}}
 144 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 145 // SAVETEMP-SAME: "-emit-llvm-bc"
 146 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx900"
 147 // SAVETEMP-SAME: {{.*}} "-o" [[A_GFX900_TMP_BC:"a.*tmp.bc"]] "-x" "hip-cpp-output" [[A_GFX900_CUI]]
 148 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 149 // SAVETEMP-SAME: "-emit-llvm"
 150 // SAVETEMP-SAME: {{.*}} "-main-file-name" "a.cu" {{.*}} "-target-cpu" "gfx900"
 151 // SAVETEMP-SAME: {{.*}} "-o" {{"a.*.ll"}} "-x" "ir" [[A_GFX900_TMP_BC]]
 152
 153 // SAVETEMP: "{{.*}}clang-offload-bundler" "-type=ll"
 154 // SAVETEMP-SAME: "-targets=hip-amdgcn-amd-amdhsa-gfx803,hip-amdgcn-amd-amdhsa-gfx900"
 155 // SAVETEMP-SAME: "-output=a-hip-amdgcn-amd-amdhsa.ll"
 156
 157 // SAVETEMP: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 158 // SAVETEMP-SAME: "-E"
 159 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx803"
 160 // SAVETEMP-SAME: {{.*}} "-o" [[B_GFX803_CUI:"b.*hipi"]] "-x" "hip" {{".*b.hip"}}
 161 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 162 // SAVETEMP-SAME: "-emit-llvm-bc"
 163 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx803"
 164 // SAVETEMP-SAME: {{.*}} "-o" [[B_GFX803_TMP_BC:"b.*tmp.bc"]] "-x" "hip-cpp-output" [[B_GFX803_CUI]]
 165 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 166 // SAVETEMP-SAME: "-emit-llvm"
 167 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx803"
 168 // SAVETEMP-SAME: {{.*}} "-o" {{"b.*.ll"}} "-x" "ir" [[B_GFX803_TMP_BC]]
 169
 170 // SAVETEMP: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 171 // SAVETEMP-SAME: "-E"
 172 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx900"
 173 // SAVETEMP-SAME: {{.*}} "-o" [[B_GFX900_CUI:"b.*hipi"]] "-x" "hip" {{".*b.hip"}}
 174 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 175 // SAVETEMP-SAME: "-emit-llvm-bc"
 176 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx900"
 177 // SAVETEMP-SAME: {{.*}} "-o" [[B_GFX900_TMP_BC:"b.*tmp.bc"]] "-x" "hip-cpp-output" [[B_GFX900_CUI]]
 178 // SAVETEMP-NEXT: [[CLANG]] "-cc1" "-triple" "amdgcn-amd-amdhsa" "-aux-triple" "x86_64-unknown-linux-gnu"
 179 // SAVETEMP-SAME: "-emit-llvm"
 180 // SAVETEMP-SAME: {{.*}} "-main-file-name" "b.hip" {{.*}} "-target-cpu" "gfx900"
 181 // SAVETEMP-SAME: {{.*}} "-o" {{"b.*.ll"}} "-x" "ir" [[B_GFX900_TMP_BC]]
 182
 183 // SAVETEMP: "{{.*}}clang-offload-bundler" "-type=ll"
 184 // SAVETEMP-SAME: "-targets=hip-amdgcn-amd-amdhsa-gfx803,hip-amdgcn-amd-amdhsa-gfx900"
 185 // SAVETEMP-SAME: "-output=b-hip-amdgcn-amd-amdhsa.ll"
 186
 187 // FAIL: error: cannot specify -o when generating multiple output files