Home | History | Annotate | Download | only in AMDGPU 
      1 ; RUN: opt -mtriple=amdgcn-unknown-amdhsa -S -amdgpu-annotate-kernel-features < %s | FileCheck -check-prefix=HSA -check-prefix=ALL %s
      2 ; RUN: opt -S -amdgpu-annotate-kernel-features < %s | FileCheck -check-prefix=NOHSA -check-prefix=ALL %s
      3 
      4 declare i32 @llvm.r600.read.tgid.x() #0
      5 declare i32 @llvm.r600.read.tgid.y() #0
      6 declare i32 @llvm.r600.read.tgid.z() #0
      7 
      8 declare i32 @llvm.r600.read.tidig.x() #0
      9 declare i32 @llvm.r600.read.tidig.y() #0
     10 declare i32 @llvm.r600.read.tidig.z() #0
     11 
     12 declare i32 @llvm.r600.read.local.size.x() #0
     13 declare i32 @llvm.r600.read.local.size.y() #0
     14 declare i32 @llvm.r600.read.local.size.z() #0
     15 
     16 declare i32 @llvm.r600.read.global.size.x() #0
     17 declare i32 @llvm.r600.read.global.size.y() #0
     18 declare i32 @llvm.r600.read.global.size.z() #0
     19 
     20 
     21 ; ALL: define void @use_tgid_x(i32 addrspace(1)* %ptr) #1 {
     22 define void @use_tgid_x(i32 addrspace(1)* %ptr) #1 {
     23   %val = call i32 @llvm.r600.read.tgid.x()
     24   store i32 %val, i32 addrspace(1)* %ptr
     25   ret void
     26 }
     27 
     28 ; ALL: define void @use_tgid_y(i32 addrspace(1)* %ptr) #2 {
     29 define void @use_tgid_y(i32 addrspace(1)* %ptr) #1 {
     30   %val = call i32 @llvm.r600.read.tgid.y()
     31   store i32 %val, i32 addrspace(1)* %ptr
     32   ret void
     33 }
     34 
     35 ; ALL: define void @multi_use_tgid_y(i32 addrspace(1)* %ptr) #2 {
     36 define void @multi_use_tgid_y(i32 addrspace(1)* %ptr) #1 {
     37   %val0 = call i32 @llvm.r600.read.tgid.y()
     38   store volatile i32 %val0, i32 addrspace(1)* %ptr
     39   %val1 = call i32 @llvm.r600.read.tgid.y()
     40   store volatile i32 %val1, i32 addrspace(1)* %ptr
     41   ret void
     42 }
     43 
     44 ; ALL: define void @use_tgid_x_y(i32 addrspace(1)* %ptr) #2 {
     45 define void @use_tgid_x_y(i32 addrspace(1)* %ptr) #1 {
     46   %val0 = call i32 @llvm.r600.read.tgid.x()
     47   %val1 = call i32 @llvm.r600.read.tgid.y()
     48   store volatile i32 %val0, i32 addrspace(1)* %ptr
     49   store volatile i32 %val1, i32 addrspace(1)* %ptr
     50   ret void
     51 }
     52 
     53 ; ALL: define void @use_tgid_z(i32 addrspace(1)* %ptr) #3 {
     54 define void @use_tgid_z(i32 addrspace(1)* %ptr) #1 {
     55   %val = call i32 @llvm.r600.read.tgid.z()
     56   store i32 %val, i32 addrspace(1)* %ptr
     57   ret void
     58 }
     59 
     60 ; ALL: define void @use_tgid_x_z(i32 addrspace(1)* %ptr) #3 {
     61 define void @use_tgid_x_z(i32 addrspace(1)* %ptr) #1 {
     62   %val0 = call i32 @llvm.r600.read.tgid.x()
     63   %val1 = call i32 @llvm.r600.read.tgid.z()
     64   store volatile i32 %val0, i32 addrspace(1)* %ptr
     65   store volatile i32 %val1, i32 addrspace(1)* %ptr
     66   ret void
     67 }
     68 
     69 ; ALL: define void @use_tgid_y_z(i32 addrspace(1)* %ptr) #4 {
     70 define void @use_tgid_y_z(i32 addrspace(1)* %ptr) #1 {
     71   %val0 = call i32 @llvm.r600.read.tgid.y()
     72   %val1 = call i32 @llvm.r600.read.tgid.z()
     73   store volatile i32 %val0, i32 addrspace(1)* %ptr
     74   store volatile i32 %val1, i32 addrspace(1)* %ptr
     75   ret void
     76 }
     77 
     78 ; ALL: define void @use_tgid_x_y_z(i32 addrspace(1)* %ptr) #4 {
     79 define void @use_tgid_x_y_z(i32 addrspace(1)* %ptr) #1 {
     80   %val0 = call i32 @llvm.r600.read.tgid.x()
     81   %val1 = call i32 @llvm.r600.read.tgid.y()
     82   %val2 = call i32 @llvm.r600.read.tgid.z()
     83   store volatile i32 %val0, i32 addrspace(1)* %ptr
     84   store volatile i32 %val1, i32 addrspace(1)* %ptr
     85   store volatile i32 %val2, i32 addrspace(1)* %ptr
     86   ret void
     87 }
     88 
     89 ; ALL: define void @use_tidig_x(i32 addrspace(1)* %ptr) #1 {
     90 define void @use_tidig_x(i32 addrspace(1)* %ptr) #1 {
     91   %val = call i32 @llvm.r600.read.tidig.x()
     92   store i32 %val, i32 addrspace(1)* %ptr
     93   ret void
     94 }
     95 
     96 ; ALL: define void @use_tidig_y(i32 addrspace(1)* %ptr) #5 {
     97 define void @use_tidig_y(i32 addrspace(1)* %ptr) #1 {
     98   %val = call i32 @llvm.r600.read.tidig.y()
     99   store i32 %val, i32 addrspace(1)* %ptr
    100   ret void
    101 }
    102 
    103 ; ALL: define void @use_tidig_z(i32 addrspace(1)* %ptr) #6 {
    104 define void @use_tidig_z(i32 addrspace(1)* %ptr) #1 {
    105   %val = call i32 @llvm.r600.read.tidig.z()
    106   store i32 %val, i32 addrspace(1)* %ptr
    107   ret void
    108 }
    109 
    110 ; ALL: define void @use_tidig_x_tgid_x(i32 addrspace(1)* %ptr) #1 {
    111 define void @use_tidig_x_tgid_x(i32 addrspace(1)* %ptr) #1 {
    112   %val0 = call i32 @llvm.r600.read.tidig.x()
    113   %val1 = call i32 @llvm.r600.read.tgid.x()
    114   store volatile i32 %val0, i32 addrspace(1)* %ptr
    115   store volatile i32 %val1, i32 addrspace(1)* %ptr
    116   ret void
    117 }
    118 
    119 ; ALL: define void @use_tidig_y_tgid_y(i32 addrspace(1)* %ptr) #7 {
    120 define void @use_tidig_y_tgid_y(i32 addrspace(1)* %ptr) #1 {
    121   %val0 = call i32 @llvm.r600.read.tidig.y()
    122   %val1 = call i32 @llvm.r600.read.tgid.y()
    123   store volatile i32 %val0, i32 addrspace(1)* %ptr
    124   store volatile i32 %val1, i32 addrspace(1)* %ptr
    125   ret void
    126 }
    127 
    128 ; ALL: define void @use_tidig_x_y_z(i32 addrspace(1)* %ptr) #8 {
    129 define void @use_tidig_x_y_z(i32 addrspace(1)* %ptr) #1 {
    130   %val0 = call i32 @llvm.r600.read.tidig.x()
    131   %val1 = call i32 @llvm.r600.read.tidig.y()
    132   %val2 = call i32 @llvm.r600.read.tidig.z()
    133   store volatile i32 %val0, i32 addrspace(1)* %ptr
    134   store volatile i32 %val1, i32 addrspace(1)* %ptr
    135   store volatile i32 %val2, i32 addrspace(1)* %ptr
    136   ret void
    137 }
    138 
    139 ; ALL: define void @use_all_workitems(i32 addrspace(1)* %ptr) #9 {
    140 define void @use_all_workitems(i32 addrspace(1)* %ptr) #1 {
    141   %val0 = call i32 @llvm.r600.read.tidig.x()
    142   %val1 = call i32 @llvm.r600.read.tidig.y()
    143   %val2 = call i32 @llvm.r600.read.tidig.z()
    144   %val3 = call i32 @llvm.r600.read.tgid.x()
    145   %val4 = call i32 @llvm.r600.read.tgid.y()
    146   %val5 = call i32 @llvm.r600.read.tgid.z()
    147   store volatile i32 %val0, i32 addrspace(1)* %ptr
    148   store volatile i32 %val1, i32 addrspace(1)* %ptr
    149   store volatile i32 %val2, i32 addrspace(1)* %ptr
    150   store volatile i32 %val3, i32 addrspace(1)* %ptr
    151   store volatile i32 %val4, i32 addrspace(1)* %ptr
    152   store volatile i32 %val5, i32 addrspace(1)* %ptr
    153   ret void
    154 }
    155 
    156 ; HSA: define void @use_get_local_size_x(i32 addrspace(1)* %ptr) #10 {
    157 ; NOHSA: define void @use_get_local_size_x(i32 addrspace(1)* %ptr) #1 {
    158 define void @use_get_local_size_x(i32 addrspace(1)* %ptr) #1 {
    159   %val = call i32 @llvm.r600.read.local.size.x()
    160   store i32 %val, i32 addrspace(1)* %ptr
    161   ret void
    162 }
    163 
    164 ; HSA: define void @use_get_local_size_y(i32 addrspace(1)* %ptr) #10 {
    165 ; NOHSA: define void @use_get_local_size_y(i32 addrspace(1)* %ptr) #1 {
    166 define void @use_get_local_size_y(i32 addrspace(1)* %ptr) #1 {
    167   %val = call i32 @llvm.r600.read.local.size.y()
    168   store i32 %val, i32 addrspace(1)* %ptr
    169   ret void
    170 }
    171 
    172 ; HSA: define void @use_get_local_size_z(i32 addrspace(1)* %ptr) #10 {
    173 ; NOHSA: define void @use_get_local_size_z(i32 addrspace(1)* %ptr) #1 {
    174 define void @use_get_local_size_z(i32 addrspace(1)* %ptr) #1 {
    175   %val = call i32 @llvm.r600.read.local.size.z()
    176   store i32 %val, i32 addrspace(1)* %ptr
    177   ret void
    178 }
    179 
    180 attributes #0 = { nounwind readnone }
    181 attributes #1 = { nounwind }
    182 
    183 ; HSA: attributes #0 = { nounwind readnone }
    184 ; HSA: attributes #1 = { nounwind }
    185 ; HSA: attributes #2 = { nounwind "amdgpu-work-group-id-y" }
    186 ; HSA: attributes #3 = { nounwind "amdgpu-work-group-id-z" }
    187 ; HSA: attributes #4 = { nounwind "amdgpu-work-group-id-y" "amdgpu-work-group-id-z" }
    188 ; HSA: attributes #5 = { nounwind "amdgpu-work-item-id-y" }
    189 ; HSA: attributes #6 = { nounwind "amdgpu-work-item-id-z" }
    190 ; HSA: attributes #7 = { nounwind "amdgpu-work-group-id-y" "amdgpu-work-item-id-y" }
    191 ; HSA: attributes #8 = { nounwind "amdgpu-work-item-id-y" "amdgpu-work-item-id-z" }
    192 ; HSA: attributes #9 = { nounwind "amdgpu-work-group-id-y" "amdgpu-work-group-id-z" "amdgpu-work-item-id-y" "amdgpu-work-item-id-z" }
    193 ; HSA: attributes #10 = { nounwind "amdgpu-dispatch-ptr" }
    194