Home | History | Annotate | Download | only in AMDGPU
      1 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs< %s | FileCheck --check-prefix=SI --check-prefix=BOTH %s
      2 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs< %s | FileCheck --check-prefix=CI --check-prefix=BOTH %s
      3 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs< %s | FileCheck --check-prefix=CI --check-prefix=BOTH %s
      4 
      5 ; BOTH-LABEL: {{^}}local_i32_load
      6 ; BOTH: ds_read_b32 [[REG:v[0-9]+]], v{{[0-9]+}} offset:28
      7 ; BOTH: buffer_store_dword [[REG]],
      8 define void @local_i32_load(i32 addrspace(1)* %out, i32 addrspace(3)* %in) nounwind {
      9   %gep = getelementptr i32, i32 addrspace(3)* %in, i32 7
     10   %val = load i32, i32 addrspace(3)* %gep, align 4
     11   store i32 %val, i32 addrspace(1)* %out, align 4
     12   ret void
     13 }
     14 
     15 ; BOTH-LABEL: {{^}}local_i32_load_0_offset
     16 ; BOTH: ds_read_b32 [[REG:v[0-9]+]], v{{[0-9]+}}
     17 ; BOTH: buffer_store_dword [[REG]],
     18 define void @local_i32_load_0_offset(i32 addrspace(1)* %out, i32 addrspace(3)* %in) nounwind {
     19   %val = load i32, i32 addrspace(3)* %in, align 4
     20   store i32 %val, i32 addrspace(1)* %out, align 4
     21   ret void
     22 }
     23 
     24 ; BOTH-LABEL: {{^}}local_i8_load_i16_max_offset:
     25 ; BOTH-NOT: ADD
     26 ; BOTH: ds_read_u8 [[REG:v[0-9]+]], {{v[0-9]+}} offset:65535
     27 ; BOTH: buffer_store_byte [[REG]],
     28 define void @local_i8_load_i16_max_offset(i8 addrspace(1)* %out, i8 addrspace(3)* %in) nounwind {
     29   %gep = getelementptr i8, i8 addrspace(3)* %in, i32 65535
     30   %val = load i8, i8 addrspace(3)* %gep, align 4
     31   store i8 %val, i8 addrspace(1)* %out, align 4
     32   ret void
     33 }
     34 
     35 ; BOTH-LABEL: {{^}}local_i8_load_over_i16_max_offset:
     36 ; The LDS offset will be 65536 bytes, which is larger than the size of LDS on
     37 ; SI, which is why it is being OR'd with the base pointer.
     38 ; SI: s_or_b32 [[ADDR:s[0-9]+]], s{{[0-9]+}}, 0x10000
     39 ; CI: s_add_i32 [[ADDR:s[0-9]+]], s{{[0-9]+}}, 0x10000
     40 ; BOTH: v_mov_b32_e32 [[VREGADDR:v[0-9]+]], [[ADDR]]
     41 ; BOTH: ds_read_u8 [[REG:v[0-9]+]], [[VREGADDR]]
     42 ; BOTH: buffer_store_byte [[REG]],
     43 define void @local_i8_load_over_i16_max_offset(i8 addrspace(1)* %out, i8 addrspace(3)* %in) nounwind {
     44   %gep = getelementptr i8, i8 addrspace(3)* %in, i32 65536
     45   %val = load i8, i8 addrspace(3)* %gep, align 4
     46   store i8 %val, i8 addrspace(1)* %out, align 4
     47   ret void
     48 }
     49 
     50 ; BOTH-LABEL: {{^}}local_i64_load:
     51 ; BOTH-NOT: ADD
     52 ; BOTH: ds_read_b64 [[REG:v[[0-9]+:[0-9]+]]], v{{[0-9]+}} offset:56
     53 ; BOTH: buffer_store_dwordx2 [[REG]],
     54 define void @local_i64_load(i64 addrspace(1)* %out, i64 addrspace(3)* %in) nounwind {
     55   %gep = getelementptr i64, i64 addrspace(3)* %in, i32 7
     56   %val = load i64, i64 addrspace(3)* %gep, align 8
     57   store i64 %val, i64 addrspace(1)* %out, align 8
     58   ret void
     59 }
     60 
     61 ; BOTH-LABEL: {{^}}local_i64_load_0_offset
     62 ; BOTH: ds_read_b64 [[REG:v\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}
     63 ; BOTH: buffer_store_dwordx2 [[REG]],
     64 define void @local_i64_load_0_offset(i64 addrspace(1)* %out, i64 addrspace(3)* %in) nounwind {
     65   %val = load i64, i64 addrspace(3)* %in, align 8
     66   store i64 %val, i64 addrspace(1)* %out, align 8
     67   ret void
     68 }
     69 
     70 ; BOTH-LABEL: {{^}}local_f64_load:
     71 ; BOTH-NOT: ADD
     72 ; BOTH: ds_read_b64 [[REG:v[[0-9]+:[0-9]+]]], v{{[0-9]+}} offset:56
     73 ; BOTH: buffer_store_dwordx2 [[REG]],
     74 define void @local_f64_load(double addrspace(1)* %out, double addrspace(3)* %in) nounwind {
     75   %gep = getelementptr double, double addrspace(3)* %in, i32 7
     76   %val = load double, double addrspace(3)* %gep, align 8
     77   store double %val, double addrspace(1)* %out, align 8
     78   ret void
     79 }
     80 
     81 ; BOTH-LABEL: {{^}}local_f64_load_0_offset
     82 ; BOTH: ds_read_b64 [[REG:v\[[0-9]+:[0-9]+\]]], v{{[0-9]+}}
     83 ; BOTH: buffer_store_dwordx2 [[REG]],
     84 define void @local_f64_load_0_offset(double addrspace(1)* %out, double addrspace(3)* %in) nounwind {
     85   %val = load double, double addrspace(3)* %in, align 8
     86   store double %val, double addrspace(1)* %out, align 8
     87   ret void
     88 }
     89 
     90 ; BOTH-LABEL: {{^}}local_i64_store:
     91 ; BOTH-NOT: ADD
     92 ; BOTH: ds_write_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}} offset:56
     93 define void @local_i64_store(i64 addrspace(3)* %out) nounwind {
     94   %gep = getelementptr i64, i64 addrspace(3)* %out, i32 7
     95   store i64 5678, i64 addrspace(3)* %gep, align 8
     96   ret void
     97 }
     98 
     99 ; BOTH-LABEL: {{^}}local_i64_store_0_offset:
    100 ; BOTH-NOT: ADD
    101 ; BOTH: ds_write_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}
    102 define void @local_i64_store_0_offset(i64 addrspace(3)* %out) nounwind {
    103   store i64 1234, i64 addrspace(3)* %out, align 8
    104   ret void
    105 }
    106 
    107 ; BOTH-LABEL: {{^}}local_f64_store:
    108 ; BOTH-NOT: ADD
    109 ; BOTH: ds_write_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}} offset:56
    110 define void @local_f64_store(double addrspace(3)* %out) nounwind {
    111   %gep = getelementptr double, double addrspace(3)* %out, i32 7
    112   store double 16.0, double addrspace(3)* %gep, align 8
    113   ret void
    114 }
    115 
    116 ; BOTH-LABEL: {{^}}local_f64_store_0_offset
    117 ; BOTH: ds_write_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}
    118 define void @local_f64_store_0_offset(double addrspace(3)* %out) nounwind {
    119   store double 20.0, double addrspace(3)* %out, align 8
    120   ret void
    121 }
    122 
    123 ; BOTH-LABEL: {{^}}local_v2i64_store:
    124 ; BOTH-NOT: ADD
    125 ; BOTH: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:15 offset1:14
    126 ; BOTH: s_endpgm
    127 define void @local_v2i64_store(<2 x i64> addrspace(3)* %out) nounwind {
    128   %gep = getelementptr <2 x i64>, <2 x i64> addrspace(3)* %out, i32 7
    129   store <2 x i64> <i64 5678, i64 5678>, <2 x i64> addrspace(3)* %gep, align 16
    130   ret void
    131 }
    132 
    133 ; BOTH-LABEL: {{^}}local_v2i64_store_0_offset:
    134 ; BOTH-NOT: ADD
    135 ; BOTH: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:1
    136 ; BOTH: s_endpgm
    137 define void @local_v2i64_store_0_offset(<2 x i64> addrspace(3)* %out) nounwind {
    138   store <2 x i64> <i64 1234, i64 1234>, <2 x i64> addrspace(3)* %out, align 16
    139   ret void
    140 }
    141 
    142 ; BOTH-LABEL: {{^}}local_v4i64_store:
    143 ; BOTH-NOT: ADD
    144 ; BOTH-DAG: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:31 offset1:30
    145 ; BOTH-DAG: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:29 offset1:28
    146 ; BOTH: s_endpgm
    147 define void @local_v4i64_store(<4 x i64> addrspace(3)* %out) nounwind {
    148   %gep = getelementptr <4 x i64>, <4 x i64> addrspace(3)* %out, i32 7
    149   store <4 x i64> <i64 5678, i64 5678, i64 5678, i64 5678>, <4 x i64> addrspace(3)* %gep, align 16
    150   ret void
    151 }
    152 
    153 ; BOTH-LABEL: {{^}}local_v4i64_store_0_offset:
    154 ; BOTH-NOT: ADD
    155 ; BOTH-DAG: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:3 offset1:2
    156 ; BOTH-DAG: ds_write2_b64 v{{[0-9]+}}, {{v\[[0-9]+:[0-9]+\]}}, {{v\[[0-9]+:[0-9]+\]}} offset0:1
    157 ; BOTH: s_endpgm
    158 define void @local_v4i64_store_0_offset(<4 x i64> addrspace(3)* %out) nounwind {
    159   store <4 x i64> <i64 1234, i64 1234, i64 1234, i64 1234>, <4 x i64> addrspace(3)* %out, align 16
    160   ret void
    161 }
    162