1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-attributes --check-globals all --version 5 2; RUN: opt -S -mtriple=amdgcn-amd-amdhsa -passes=amdgpu-attributor %s | FileCheck %s 3 4; External call to avoid inferring argument attributes. This makes the 5; final attribute groups easier to read 6declare void @dummy() 7 8define void @extern_callee() { 9; CHECK-LABEL: define void @extern_callee( 10; CHECK-SAME: ) #[[ATTR0:[0-9]+]] { 11; CHECK-NEXT: call void @dummy() 12; CHECK-NEXT: ret void 13; 14 call void @dummy() 15 ret void 16} 17 18define internal void @callee_1_2_3() { 19; CHECK-LABEL: define internal void @callee_1_2_3( 20; CHECK-SAME: ) #[[ATTR1:[0-9]+]] { 21; CHECK-NEXT: call void @dummy() 22; CHECK-NEXT: ret void 23; 24 call void @dummy() 25 ret void 26} 27 28define amdgpu_kernel void @kernel_1_2_3() #0 { 29; CHECK-LABEL: define amdgpu_kernel void @kernel_1_2_3( 30; CHECK-SAME: ) #[[ATTR2:[0-9]+]] { 31; CHECK-NEXT: call void @callee_1_2_3() 32; CHECK-NEXT: call void @extern_callee() 33; CHECK-NEXT: call void @dummy() 34; CHECK-NEXT: ret void 35; 36 call void @callee_1_2_3() 37 call void @extern_callee() 38 call void @dummy() 39 ret void 40} 41 42attributes #0 = {"amdgpu-max-num-workgroups"="1,2,3"} 43 44; -> 100,10,99 45define internal void @callee_merge_100_8_32__16_10_99() { 46; CHECK-LABEL: define internal void @callee_merge_100_8_32__16_10_99( 47; CHECK-SAME: ) #[[ATTR3:[0-9]+]] { 48; CHECK-NEXT: call void @dummy() 49; CHECK-NEXT: ret void 50; 51 call void @dummy() 52 ret void 53} 54 55define amdgpu_kernel void @kernel_100_8_32() #1 { 56; CHECK-LABEL: define amdgpu_kernel void @kernel_100_8_32( 57; CHECK-SAME: ) #[[ATTR4:[0-9]+]] { 58; CHECK-NEXT: call void @callee_merge_100_8_32__16_10_99() 59; CHECK-NEXT: ret void 60; 61 call void @callee_merge_100_8_32__16_10_99() 62 ret void 63} 64 65define amdgpu_cs void @amdgpu_cs_100_8_32() #1 { 66; CHECK-LABEL: define amdgpu_cs void @amdgpu_cs_100_8_32( 67; CHECK-SAME: ) #[[ATTR4]] { 68; CHECK-NEXT: call void @callee_merge_100_8_32__16_10_99() 69; CHECK-NEXT: ret void 70; 71 call void @callee_merge_100_8_32__16_10_99() 72 ret void 73} 74 75attributes #1 = {"amdgpu-max-num-workgroups"="100,8,32"} 76 77define amdgpu_kernel void @kernel_16_10_99() #2 { 78; CHECK-LABEL: define amdgpu_kernel void @kernel_16_10_99( 79; CHECK-SAME: ) #[[ATTR5:[0-9]+]] { 80; CHECK-NEXT: call void @callee_merge_100_8_32__16_10_99() 81; CHECK-NEXT: call void @dummy() 82; CHECK-NEXT: ret void 83; 84 call void @callee_merge_100_8_32__16_10_99() 85 call void @dummy() 86 ret void 87} 88 89attributes #2 = {"amdgpu-max-num-workgroups"="16,10,99"} 90 91define internal void @merge_to_worst_case() { 92; CHECK-LABEL: define internal void @merge_to_worst_case( 93; CHECK-SAME: ) #[[ATTR0]] { 94; CHECK-NEXT: call void @dummy() 95; CHECK-NEXT: ret void 96; 97 call void @dummy() 98 ret void 99} 100 101define internal void @callee_x_worst_case() { 102; CHECK-LABEL: define internal void @callee_x_worst_case( 103; CHECK-SAME: ) #[[ATTR0]] { 104; CHECK-NEXT: call void @dummy() 105; CHECK-NEXT: ret void 106; 107 call void @dummy() 108 ret void 109} 110 111define amdgpu_kernel void @kernel_x_maximum() #3 { 112; CHECK-LABEL: define amdgpu_kernel void @kernel_x_maximum( 113; CHECK-SAME: ) #[[ATTR6:[0-9]+]] { 114; CHECK-NEXT: call void @merge_to_worst_case() 115; CHECK-NEXT: call void @callee_x_worst_case() 116; CHECK-NEXT: call void @dummy() 117; CHECK-NEXT: ret void 118; 119 call void @merge_to_worst_case() 120 call void @callee_x_worst_case() 121 call void @dummy() 122 ret void 123} 124 125attributes #3 = {"amdgpu-max-num-workgroups"="4294967295,1,1"} 126 127define amdgpu_kernel void @kernel_y_maximum() #4 { 128; CHECK-LABEL: define amdgpu_kernel void @kernel_y_maximum( 129; CHECK-SAME: ) #[[ATTR7:[0-9]+]] { 130; CHECK-NEXT: call void @merge_to_worst_case() 131; CHECK-NEXT: call void @dummy() 132; CHECK-NEXT: ret void 133; 134 call void @merge_to_worst_case() 135 call void @dummy() 136 ret void 137} 138 139attributes #4 = {"amdgpu-max-num-workgroups"="1,4294967295,1"} 140 141define amdgpu_kernel void @kernel_z_maximum() #5 { 142; CHECK-LABEL: define amdgpu_kernel void @kernel_z_maximum( 143; CHECK-SAME: ) #[[ATTR8:[0-9]+]] { 144; CHECK-NEXT: call void @merge_to_worst_case() 145; CHECK-NEXT: call void @dummy() 146; CHECK-NEXT: ret void 147; 148 call void @merge_to_worst_case() 149 call void @dummy() 150 ret void 151} 152 153attributes #5 = {"amdgpu-max-num-workgroups"="1,1,4294967295"} 154 155; Make sure the attribute isn't lost from the callee. 156define internal void @annotated_callee_from_unannotated_kernel() #6 { 157; CHECK-LABEL: define internal void @annotated_callee_from_unannotated_kernel( 158; CHECK-SAME: ) #[[ATTR9:[0-9]+]] { 159; CHECK-NEXT: call void @dummy() 160; CHECK-NEXT: ret void 161; 162 call void @dummy() 163 ret void 164} 165 166attributes #6 = {"amdgpu-max-num-workgroups"="42,99,123"} 167 168define amdgpu_kernel void @unannotated_kernel_calls_annotated_callee() { 169; CHECK-LABEL: define amdgpu_kernel void @unannotated_kernel_calls_annotated_callee( 170; CHECK-SAME: ) #[[ATTR10:[0-9]+]] { 171; CHECK-NEXT: call void @annotated_callee_from_unannotated_kernel() 172; CHECK-NEXT: ret void 173; 174 call void @annotated_callee_from_unannotated_kernel() 175 ret void 176} 177 178 179define internal void @annotated_callee_merge_caller() #7 { 180; CHECK-LABEL: define internal void @annotated_callee_merge_caller( 181; CHECK-SAME: ) #[[ATTR11:[0-9]+]] { 182; CHECK-NEXT: call void @dummy() 183; CHECK-NEXT: ret void 184; 185 call void @dummy() 186 ret void 187} 188 189attributes #7 = {"amdgpu-max-num-workgroups"="512,256,1024"} 190 191define amdgpu_kernel void @call_annotated_callee_merge_caller() #8 { 192; CHECK-LABEL: define amdgpu_kernel void @call_annotated_callee_merge_caller( 193; CHECK-SAME: ) #[[ATTR12:[0-9]+]] { 194; CHECK-NEXT: call void @annotated_callee_merge_caller() 195; CHECK-NEXT: ret void 196; 197 call void @annotated_callee_merge_caller() 198 ret void 199} 200 201attributes #8 = {"amdgpu-max-num-workgroups"="256,128,2048"} 202 203define internal void @called_by_explicit_worst_case() { 204; CHECK-LABEL: define internal void @called_by_explicit_worst_case( 205; CHECK-SAME: ) #[[ATTR0]] { 206; CHECK-NEXT: call void @dummy() 207; CHECK-NEXT: ret void 208; 209 call void @dummy() 210 ret void 211} 212 213define amdgpu_kernel void @kernel_explicit_worst_case() #9 { 214; CHECK-LABEL: define amdgpu_kernel void @kernel_explicit_worst_case( 215; CHECK-SAME: ) #[[ATTR13:[0-9]+]] { 216; CHECK-NEXT: call void @called_by_explicit_worst_case() 217; CHECK-NEXT: ret void 218; 219 call void @called_by_explicit_worst_case() 220 ret void 221} 222 223attributes #9 = {"amdgpu-max-num-workgroups"="4294967295,4294967295,4294967295"} 224 225;. 226; CHECK: attributes #[[ATTR0]] = { "uniform-work-group-size"="false" } 227; CHECK: attributes #[[ATTR1]] = { "amdgpu-max-num-workgroups"="1,2,3" "uniform-work-group-size"="false" } 228; CHECK: attributes #[[ATTR2]] = { "amdgpu-max-num-workgroups"="1,2,3" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 229; CHECK: attributes #[[ATTR3]] = { "amdgpu-max-num-workgroups"="100,10,99" "uniform-work-group-size"="false" } 230; CHECK: attributes #[[ATTR4]] = { "amdgpu-max-num-workgroups"="100,8,32" "uniform-work-group-size"="false" } 231; CHECK: attributes #[[ATTR5]] = { "amdgpu-max-num-workgroups"="16,10,99" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 232; CHECK: attributes #[[ATTR6]] = { "amdgpu-max-num-workgroups"="4294967295,1,1" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 233; CHECK: attributes #[[ATTR7]] = { "amdgpu-max-num-workgroups"="1,4294967295,1" "uniform-work-group-size"="false" } 234; CHECK: attributes #[[ATTR8]] = { "amdgpu-max-num-workgroups"="1,1,4294967295" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 235; CHECK: attributes #[[ATTR9]] = { "amdgpu-max-num-workgroups"="42,99,123" "uniform-work-group-size"="false" } 236; CHECK: attributes #[[ATTR10]] = { "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 237; CHECK: attributes #[[ATTR11]] = { "amdgpu-max-num-workgroups"="256,128,1024" "uniform-work-group-size"="false" } 238; CHECK: attributes #[[ATTR12]] = { "amdgpu-max-num-workgroups"="256,128,2048" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 239; CHECK: attributes #[[ATTR13]] = { "amdgpu-max-num-workgroups"="4294967295,4294967295,4294967295" "amdgpu-waves-per-eu"="4,10" "uniform-work-group-size"="false" } 240;. 241