test/CodeGenOpenCL/amdgpu-attrs.cl

   1 // RUN: %clang_cc1 -triple amdgcn-amd-amdhsa -target-cpu tahiti -O0 -emit-llvm -o - %s | FileCheck %s
   2 // RUN: %clang_cc1 -triple amdgcn-- -target-cpu tahiti -O0 -emit-llvm -o - %s | FileCheck %s -check-prefix=NONAMDHSA
   3 // RUN: %clang_cc1 -triple x86_64-unknown-linux-gnu -O0 -emit-llvm -verify -o - %s | FileCheck -check-prefix=X86 %s
   4
   5 __attribute__((amdgpu_flat_work_group_size(0, 0))) // expected-no-diagnostics
   6 kernel void flat_work_group_size_0_0() {}
   7 __attribute__((amdgpu_waves_per_eu(0))) // expected-no-diagnostics
   8 kernel void waves_per_eu_0() {}
   9 __attribute__((amdgpu_waves_per_eu(0, 0))) // expected-no-diagnostics
  10 kernel void waves_per_eu_0_0() {}
  11 __attribute__((amdgpu_num_sgpr(0))) // expected-no-diagnostics
  12 kernel void num_sgpr0() {}
  13 __attribute__((amdgpu_num_vgpr(0))) // expected-no-diagnostics
  14 kernel void num_vgpr0() {}
  15
  16 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0))) // expected-no-diagnostics
  17 kernel void flat_work_group_size_0_0_waves_per_eu_0() {}
  18 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0, 0))) // expected-no-diagnostics
  19 kernel void flat_work_group_size_0_0_waves_per_eu_0_0() {}
  20 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_num_sgpr(0))) // expected-no-diagnostics
  21 kernel void flat_work_group_size_0_0_num_sgpr_0() {}
  22 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  23 kernel void flat_work_group_size_0_0_num_vgpr_0() {}
  24 __attribute__((amdgpu_waves_per_eu(0), amdgpu_num_sgpr(0))) // expected-no-diagnostics
  25 kernel void waves_per_eu_0_num_sgpr_0() {}
  26 __attribute__((amdgpu_waves_per_eu(0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  27 kernel void waves_per_eu_0_num_vgpr_0() {}
  28 __attribute__((amdgpu_waves_per_eu(0, 0), amdgpu_num_sgpr(0))) // expected-no-diagnostics
  29 kernel void waves_per_eu_0_0_num_sgpr_0() {}
  30 __attribute__((amdgpu_waves_per_eu(0, 0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  31 kernel void waves_per_eu_0_0_num_vgpr_0() {}
  32 __attribute__((amdgpu_num_sgpr(0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  33 kernel void num_sgpr_0_num_vgpr_0() {}
  34
  35 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0), amdgpu_num_sgpr(0))) // expected-no-diagnostics
  36 kernel void flat_work_group_size_0_0_waves_per_eu_0_num_sgpr_0() {}
  37 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  38 kernel void flat_work_group_size_0_0_waves_per_eu_0_num_vgpr_0() {}
  39 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0, 0), amdgpu_num_sgpr(0))) // expected-no-diagnostics
  40 kernel void flat_work_group_size_0_0_waves_per_eu_0_0_num_sgpr_0() {}
  41 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0, 0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  42 kernel void flat_work_group_size_0_0_waves_per_eu_0_0_num_vgpr_0() {}
  43
  44 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0), amdgpu_num_sgpr(0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  45 kernel void flat_work_group_size_0_0_waves_per_eu_0_num_sgpr_0_num_vgpr_0() {}
  46 __attribute__((amdgpu_flat_work_group_size(0, 0), amdgpu_waves_per_eu(0, 0), amdgpu_num_sgpr(0), amdgpu_num_vgpr(0))) // expected-no-diagnostics
  47 kernel void flat_work_group_size_0_0_waves_per_eu_0_0_num_sgpr_0_num_vgpr_0() {}
  48
  49 __attribute__((amdgpu_flat_work_group_size(32, 64))) // expected-no-diagnostics
  50 kernel void flat_work_group_size_32_64() {
  51 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64() [[FLAT_WORK_GROUP_SIZE_32_64:#[0-9]+]]
  52 }
  53 __attribute__((amdgpu_waves_per_eu(2))) // expected-no-diagnostics
  54 kernel void waves_per_eu_2() {
  55 // CHECK: define amdgpu_kernel void @waves_per_eu_2() [[WAVES_PER_EU_2:#[0-9]+]]
  56 }
  57 __attribute__((amdgpu_waves_per_eu(2, 4))) // expected-no-diagnostics
  58 kernel void waves_per_eu_2_4() {
  59 // CHECK: define amdgpu_kernel void @waves_per_eu_2_4() [[WAVES_PER_EU_2_4:#[0-9]+]]
  60 }
  61 __attribute__((amdgpu_num_sgpr(32))) // expected-no-diagnostics
  62 kernel void num_sgpr_32() {
  63 // CHECK: define amdgpu_kernel void @num_sgpr_32() [[NUM_SGPR_32:#[0-9]+]]
  64 }
  65 __attribute__((amdgpu_num_vgpr(64))) // expected-no-diagnostics
  66 kernel void num_vgpr_64() {
  67 // CHECK: define amdgpu_kernel void @num_vgpr_64() [[NUM_VGPR_64:#[0-9]+]]
  68 }
  69
  70 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2))) // expected-no-diagnostics
  71 kernel void flat_work_group_size_32_64_waves_per_eu_2() {
  72 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2:#[0-9]+]]
  73 }
  74 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2, 4))) // expected-no-diagnostics
  75 kernel void flat_work_group_size_32_64_waves_per_eu_2_4() {
  76 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_4() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4:#[0-9]+]]
  77 }
  78 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_num_sgpr(32))) // expected-no-diagnostics
  79 kernel void flat_work_group_size_32_64_num_sgpr_32() {
  80 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_num_sgpr_32() [[FLAT_WORK_GROUP_SIZE_32_64_NUM_SGPR_32:#[0-9]+]]
  81 }
  82 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_num_vgpr(64))) // expected-no-diagnostics
  83 kernel void flat_work_group_size_32_64_num_vgpr_64() {
  84 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_num_vgpr_64() [[FLAT_WORK_GROUP_SIZE_32_64_NUM_VGPR_64:#[0-9]+]]
  85 }
  86 __attribute__((amdgpu_waves_per_eu(2), amdgpu_num_sgpr(32))) // expected-no-diagnostics
  87 kernel void waves_per_eu_2_num_sgpr_32() {
  88 // CHECK: define amdgpu_kernel void @waves_per_eu_2_num_sgpr_32() [[WAVES_PER_EU_2_NUM_SGPR_32:#[0-9]+]]
  89 }
  90 __attribute__((amdgpu_waves_per_eu(2), amdgpu_num_vgpr(64))) // expected-no-diagnostics
  91 kernel void waves_per_eu_2_num_vgpr_64() {
  92 // CHECK: define amdgpu_kernel void @waves_per_eu_2_num_vgpr_64() [[WAVES_PER_EU_2_NUM_VGPR_64:#[0-9]+]]
  93 }
  94 __attribute__((amdgpu_waves_per_eu(2, 4), amdgpu_num_sgpr(32))) // expected-no-diagnostics
  95 kernel void waves_per_eu_2_4_num_sgpr_32() {
  96 // CHECK: define amdgpu_kernel void @waves_per_eu_2_4_num_sgpr_32() [[WAVES_PER_EU_2_4_NUM_SGPR_32:#[0-9]+]]
  97 }
  98 __attribute__((amdgpu_waves_per_eu(2, 4), amdgpu_num_vgpr(64))) // expected-no-diagnostics
  99 kernel void waves_per_eu_2_4_num_vgpr_64() {
 100 // CHECK: define amdgpu_kernel void @waves_per_eu_2_4_num_vgpr_64() [[WAVES_PER_EU_2_4_NUM_VGPR_64:#[0-9]+]]
 101 }
 102 __attribute__((amdgpu_num_sgpr(32), amdgpu_num_vgpr(64))) // expected-no-diagnostics
 103 kernel void num_sgpr_32_num_vgpr_64() {
 104 // CHECK: define amdgpu_kernel void @num_sgpr_32_num_vgpr_64() [[NUM_SGPR_32_NUM_VGPR_64:#[0-9]+]]
 105 }
 106
 107 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2), amdgpu_num_sgpr(32)))
 108 kernel void flat_work_group_size_32_64_waves_per_eu_2_num_sgpr_32() {
 109 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_num_sgpr_32() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_SGPR_32:#[0-9]+]]
 110 }
 111 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2), amdgpu_num_vgpr(64)))
 112 kernel void flat_work_group_size_32_64_waves_per_eu_2_num_vgpr_64() {
 113 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_num_vgpr_64() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_VGPR_64:#[0-9]+]]
 114 }
 115 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2, 4), amdgpu_num_sgpr(32)))
 116 kernel void flat_work_group_size_32_64_waves_per_eu_2_4_num_sgpr_32() {
 117 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_4_num_sgpr_32() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_SGPR_32:#[0-9]+]]
 118 }
 119 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2, 4), amdgpu_num_vgpr(64)))
 120 kernel void flat_work_group_size_32_64_waves_per_eu_2_4_num_vgpr_64() {
 121 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_4_num_vgpr_64() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_VGPR_64:#[0-9]+]]
 122 }
 123
 124 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2), amdgpu_num_sgpr(32), amdgpu_num_vgpr(64))) // expected-no-diagnostics
 125 kernel void flat_work_group_size_32_64_waves_per_eu_2_num_sgpr_32_num_vgpr_64() {
 126 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_num_sgpr_32_num_vgpr_64() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_SGPR_32_NUM_VGPR_64:#[0-9]+]]
 127 }
 128 __attribute__((amdgpu_flat_work_group_size(32, 64), amdgpu_waves_per_eu(2, 4), amdgpu_num_sgpr(32), amdgpu_num_vgpr(64))) // expected-no-diagnostics
 129 kernel void flat_work_group_size_32_64_waves_per_eu_2_4_num_sgpr_32_num_vgpr_64() {
 130 // CHECK: define amdgpu_kernel void @flat_work_group_size_32_64_waves_per_eu_2_4_num_sgpr_32_num_vgpr_64() [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_SGPR_32_NUM_VGPR_64:#[0-9]+]]
 131 }
 132
 133 __attribute__((reqd_work_group_size(32, 2, 1))) // expected-no-diagnostics
 134 kernel void reqd_work_group_size_32_2_1() {
 135 // CHECK: define amdgpu_kernel void @reqd_work_group_size_32_2_1() [[FLAT_WORK_GROUP_SIZE_64_64:#[0-9]+]]
 136 }
 137 __attribute__((reqd_work_group_size(32, 2, 1), amdgpu_flat_work_group_size(16, 128))) // expected-no-diagnostics
 138 kernel void reqd_work_group_size_32_2_1_flat_work_group_size_16_128() {
 139 // CHECK: define amdgpu_kernel void @reqd_work_group_size_32_2_1_flat_work_group_size_16_128() [[FLAT_WORK_GROUP_SIZE_16_128:#[0-9]+]]
 140 }
 141
 142 void a_function() {
 143 // CHECK: define void @a_function() [[A_FUNCTION:#[0-9]+]]
 144 }
 145
 146
 147 // Make sure this is silently accepted on other targets.
 148 // X86-NOT: "amdgpu-flat-work-group-size"
 149 // X86-NOT: "amdgpu-waves-per-eu"
 150 // X86-NOT: "amdgpu-num-vgpr"
 151 // X86-NOT: "amdgpu-num-sgpr"
 152 // X86-NOT: "amdgpu-implicitarg-num-bytes"
 153 // NONAMDHSA-NOT: "amdgpu-implicitarg-num-bytes"
 154
 155 // CHECK-NOT: "amdgpu-flat-work-group-size"="0,0"
 156 // CHECK-NOT: "amdgpu-waves-per-eu"="0"
 157 // CHECK-NOT: "amdgpu-waves-per-eu"="0,0"
 158 // CHECK-NOT: "amdgpu-num-sgpr"="0"
 159 // CHECK-NOT: "amdgpu-num-vgpr"="0"
 160
 161 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48"
 162 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_64_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="64,64" "amdgpu-implicitarg-num-bytes"="48"
 163 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_16_128]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="16,128" "amdgpu-implicitarg-num-bytes"="48"
 164 // CHECK-DAG: attributes [[WAVES_PER_EU_2]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-waves-per-eu"="2"
 165 // CHECK-DAG: attributes [[WAVES_PER_EU_2_4]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-waves-per-eu"="2,4"
 166 // CHECK-DAG: attributes [[NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32"
 167 // CHECK-DAG: attributes [[NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64"
 168
 169 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-waves-per-eu"="2"
 170 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-waves-per-eu"="2,4"
 171 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32"
 172 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64"
 173 // CHECK-DAG: attributes [[WAVES_PER_EU_2_NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-waves-per-eu"="2"
 174 // CHECK-DAG: attributes [[WAVES_PER_EU_2_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2"
 175 // CHECK-DAG: attributes [[WAVES_PER_EU_2_4_NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-waves-per-eu"="2,4"
 176 // CHECK-DAG: attributes [[WAVES_PER_EU_2_4_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2,4"
 177 // CHECK-DAG: attributes [[NUM_SGPR_32_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-num-vgpr"="64"
 178
 179 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-waves-per-eu"="2"
 180 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2"
 181 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_SGPR_32]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-waves-per-eu"="2,4"
 182 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2,4"
 183
 184 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_NUM_SGPR_32_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2"
 185 // CHECK-DAG: attributes [[FLAT_WORK_GROUP_SIZE_32_64_WAVES_PER_EU_2_4_NUM_SGPR_32_NUM_VGPR_64]] = { convergent noinline nounwind optnone "amdgpu-flat-work-group-size"="32,64" "amdgpu-implicitarg-num-bytes"="48" "amdgpu-num-sgpr"="32" "amdgpu-num-vgpr"="64" "amdgpu-waves-per-eu"="2,4"
 186
 187 // CHECK-DAG: attributes [[A_FUNCTION]] = { convergent noinline nounwind optnone "correctly-rounded-divide-sqrt-fp-math"="false"