Home | History | Annotate | Download | only in AMDGPU
      1 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
      2 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
      3 
      4 ; FUNC-LABEL {{^}}sextload_i1_to_i32_trunc_cmp_eq_0:
      5 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
      6 ; SI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
      7 ; SI: v_cmp_eq_i32_e32 vcc, 0, [[TMP]]{{$}}
      8 ; SI: v_cndmask_b32_e64
      9 ; SI: buffer_store_byte
     10 define void @sextload_i1_to_i32_trunc_cmp_eq_0(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     11   %load = load i1, i1 addrspace(1)* %in
     12   %ext = sext i1 %load to i32
     13   %cmp = icmp eq i32 %ext, 0
     14   store i1 %cmp, i1 addrspace(1)* %out
     15   ret void
     16 }
     17 
     18 ; FIXME: The negate should be inverting the compare.
     19 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_eq_0:
     20 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
     21 ; SI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
     22 ; SI: v_cmp_eq_i32_e32 vcc, 1, [[TMP]]{{$}}
     23 ; SI-NEXT: s_xor_b64 [[NEG:s\[[0-9]+:[0-9]+\]]], vcc, -1
     24 ; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, [[NEG]]
     25 ; SI: buffer_store_byte [[RESULT]]
     26 define void @zextload_i1_to_i32_trunc_cmp_eq_0(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     27   %load = load i1, i1 addrspace(1)* %in
     28   %ext = zext i1 %load to i32
     29   %cmp = icmp eq i32 %ext, 0
     30   store i1 %cmp, i1 addrspace(1)* %out
     31   ret void
     32 }
     33 
     34 ; FUNC-LABEL: {{^}}sextload_i1_to_i32_trunc_cmp_eq_1:
     35 ; SI: v_mov_b32_e32 [[RESULT:v[0-9]+]], 0{{$}}
     36 ; SI: buffer_store_byte [[RESULT]]
     37 define void @sextload_i1_to_i32_trunc_cmp_eq_1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     38   %load = load i1, i1 addrspace(1)* %in
     39   %ext = sext i1 %load to i32
     40   %cmp = icmp eq i32 %ext, 1
     41   store i1 %cmp, i1 addrspace(1)* %out
     42   ret void
     43 }
     44 
     45 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_eq_1:
     46 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
     47 ; SI: v_and_b32_e32 [[RESULT:v[0-9]+]], 1, [[LOAD]]
     48 ; SI: buffer_store_byte [[RESULT]]
     49 define void @zextload_i1_to_i32_trunc_cmp_eq_1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     50   %load = load i1, i1 addrspace(1)* %in
     51   %ext = zext i1 %load to i32
     52   %cmp = icmp eq i32 %ext, 1
     53   store i1 %cmp, i1 addrspace(1)* %out
     54   ret void
     55 }
     56 
     57 ; FUNC-LABEL: {{^}}sextload_i1_to_i32_trunc_cmp_eq_neg1:
     58 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
     59 ; SI: v_and_b32_e32 [[RESULT:v[0-9]+]], 1, [[LOAD]]
     60 ; SI: buffer_store_byte [[RESULT]]
     61 define void @sextload_i1_to_i32_trunc_cmp_eq_neg1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     62   %load = load i1, i1 addrspace(1)* %in
     63   %ext = sext i1 %load to i32
     64   %cmp = icmp eq i32 %ext, -1
     65   store i1 %cmp, i1 addrspace(1)* %out
     66   ret void
     67 }
     68 
     69 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_eq_neg1:
     70 ; SI: v_mov_b32_e32 [[RESULT:v[0-9]+]], 0{{$}}
     71 ; SI: buffer_store_byte [[RESULT]]
     72 define void @zextload_i1_to_i32_trunc_cmp_eq_neg1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     73   %load = load i1, i1 addrspace(1)* %in
     74   %ext = zext i1 %load to i32
     75   %cmp = icmp eq i32 %ext, -1
     76   store i1 %cmp, i1 addrspace(1)* %out
     77   ret void
     78 }
     79 
     80 
     81 ; FUNC-LABEL {{^}}sextload_i1_to_i32_trunc_cmp_ne_0:
     82 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
     83 ; SI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
     84 ; SI: buffer_store_byte [[RESULT]]
     85 define void @sextload_i1_to_i32_trunc_cmp_ne_0(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     86   %load = load i1, i1 addrspace(1)* %in
     87   %ext = sext i1 %load to i32
     88   %cmp = icmp ne i32 %ext, 0
     89   store i1 %cmp, i1 addrspace(1)* %out
     90   ret void
     91 }
     92 
     93 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_ne_0:
     94 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
     95 ; SI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
     96 ; SI: buffer_store_byte [[RESULT]]
     97 define void @zextload_i1_to_i32_trunc_cmp_ne_0(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
     98   %load = load i1, i1 addrspace(1)* %in
     99   %ext = zext i1 %load to i32
    100   %cmp = icmp ne i32 %ext, 0
    101   store i1 %cmp, i1 addrspace(1)* %out
    102   ret void
    103 }
    104 
    105 ; FUNC-LABEL: {{^}}sextload_i1_to_i32_trunc_cmp_ne_1:
    106 ; SI: v_mov_b32_e32 [[RESULT:v[0-9]+]], 1{{$}}
    107 ; SI: buffer_store_byte [[RESULT]]
    108 define void @sextload_i1_to_i32_trunc_cmp_ne_1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
    109   %load = load i1, i1 addrspace(1)* %in
    110   %ext = sext i1 %load to i32
    111   %cmp = icmp ne i32 %ext, 1
    112   store i1 %cmp, i1 addrspace(1)* %out
    113   ret void
    114 }
    115 
    116 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_ne_1:
    117 ; SI: buffer_load_ubyte [[LOAD:v[0-9]+]]
    118 ; SI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
    119 ; SI: v_cmp_eq_i32_e32 vcc, 1, [[TMP]]{{$}}
    120 ; SI-NEXT: s_xor_b64 [[NEG:s\[[0-9]+:[0-9]+\]]], vcc, -1
    121 ; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, [[NEG]]
    122 ; SI: buffer_store_byte [[RESULT]]
    123 define void @zextload_i1_to_i32_trunc_cmp_ne_1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
    124   %load = load i1, i1 addrspace(1)* %in
    125   %ext = zext i1 %load to i32
    126   %cmp = icmp ne i32 %ext, 1
    127   store i1 %cmp, i1 addrspace(1)* %out
    128   ret void
    129 }
    130 
    131 ; FIXME: This should be one compare.
    132 ; FUNC-LABEL: {{^}}sextload_i1_to_i32_trunc_cmp_ne_neg1:
    133 ; XSI: buffer_load_ubyte [[LOAD:v[0-9]+]]
    134 ; XSI: v_and_b32_e32 [[TMP:v[0-9]+]], 1, [[LOAD]]
    135 ; XSI: v_cmp_eq_i32_e64 [[CMP0:s\[[0-9]+:[0-9]+\]]], [[TMP]], 0{{$}}
    136 ; XSI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, 1, [[CMP0]]
    137 ; XSI-NEXT: buffer_store_byte [[RESULT]]
    138 define void @sextload_i1_to_i32_trunc_cmp_ne_neg1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
    139   %load = load i1, i1 addrspace(1)* %in
    140   %ext = sext i1 %load to i32
    141   %cmp = icmp ne i32 %ext, -1
    142   store i1 %cmp, i1 addrspace(1)* %out
    143   ret void
    144 }
    145 
    146 ; FUNC-LABEL: {{^}}zextload_i1_to_i32_trunc_cmp_ne_neg1:
    147 ; SI: v_mov_b32_e32 [[RESULT:v[0-9]+]], 1{{$}}
    148 ; SI: buffer_store_byte [[RESULT]]
    149 define void @zextload_i1_to_i32_trunc_cmp_ne_neg1(i1 addrspace(1)* %out, i1 addrspace(1)* %in) nounwind {
    150   %load = load i1, i1 addrspace(1)* %in
    151   %ext = zext i1 %load to i32
    152   %cmp = icmp ne i32 %ext, -1
    153   store i1 %cmp, i1 addrspace(1)* %out
    154   ret void
    155 }
    156 
    157 ; FUNC-LABEL: {{^}}masked_load_i1_to_i32_trunc_cmp_ne_neg1:
    158 ; SI: buffer_load_sbyte [[LOAD:v[0-9]+]]
    159 ; SI: v_cmp_ne_i32_e32 vcc, -1, [[LOAD]]{{$}}
    160 ; SI-NEXT: v_cndmask_b32_e64
    161 ; SI: buffer_store_byte
    162 define void @masked_load_i1_to_i32_trunc_cmp_ne_neg1(i1 addrspace(1)* %out, i8 addrspace(1)* %in) nounwind {
    163   %load = load i8, i8 addrspace(1)* %in
    164   %masked = and i8 %load, 255
    165   %ext = sext i8 %masked to i32
    166   %cmp = icmp ne i32 %ext, -1
    167   store i1 %cmp, i1 addrspace(1)* %out
    168   ret void
    169 }
    170