Home | History | Annotate | Download | only in R600
      1 ; RUN: llc < %s -march=r600 -mcpu=verde -verify-machineinstrs | FileCheck %s
      2 
      3 ; On Southern Islands GPUs the local address space(3) uses 32-bit pointers and
      4 ; the global address space(1) uses 64-bit pointers.  These tests check to make sure
      5 ; the correct pointer size is used for the local address space.
      6 
      7 ; The e{{32|64}} suffix on the instructions refers to the encoding size and not
      8 ; the size of the operands.  The operand size is denoted in the instruction name.
      9 ; Instructions with B32, U32, and I32 in their name take 32-bit operands, while
     10 ; instructions with B64, U64, and I64 take 64-bit operands.
     11 
     12 ; CHECK-LABEL: @local_address_load
     13 ; CHECK: V_MOV_B32_e{{32|64}} [[PTR:v[0-9]]]
     14 ; CHECK: DS_READ_B32 v{{[0-9]+}}, [[PTR]]
     15 define void @local_address_load(i32 addrspace(1)* %out, i32 addrspace(3)* %in) {
     16 entry:
     17   %0 = load i32 addrspace(3)* %in
     18   store i32 %0, i32 addrspace(1)* %out
     19   ret void
     20 }
     21 
     22 ; CHECK-LABEL: @local_address_gep
     23 ; CHECK: S_ADD_I32 [[SPTR:s[0-9]]]
     24 ; CHECK: V_MOV_B32_e32 [[VPTR:v[0-9]+]], [[SPTR]]
     25 ; CHECK: DS_READ_B32 [[VPTR]]
     26 define void @local_address_gep(i32 addrspace(1)* %out, i32 addrspace(3)* %in, i32 %offset) {
     27 entry:
     28   %0 = getelementptr i32 addrspace(3)* %in, i32 %offset
     29   %1 = load i32 addrspace(3)* %0
     30   store i32 %1, i32 addrspace(1)* %out
     31   ret void
     32 }
     33 
     34 ; CHECK-LABEL: @local_address_gep_const_offset
     35 ; CHECK: V_MOV_B32_e32 [[VPTR:v[0-9]+]], s{{[0-9]+}}
     36 ; CHECK: DS_READ_B32 v{{[0-9]+}}, [[VPTR]], 0x4,
     37 define void @local_address_gep_const_offset(i32 addrspace(1)* %out, i32 addrspace(3)* %in) {
     38 entry:
     39   %0 = getelementptr i32 addrspace(3)* %in, i32 1
     40   %1 = load i32 addrspace(3)* %0
     41   store i32 %1, i32 addrspace(1)* %out
     42   ret void
     43 }
     44 
     45 ; Offset too large, can't fold into 16-bit immediate offset.
     46 ; CHECK-LABEL: @local_address_gep_large_const_offset
     47 ; CHECK: S_ADD_I32 [[SPTR:s[0-9]]], s{{[0-9]+}}, 0x10004
     48 ; CHECK: V_MOV_B32_e32 [[VPTR:v[0-9]+]], [[SPTR]]
     49 ; CHECK: DS_READ_B32 [[VPTR]]
     50 define void @local_address_gep_large_const_offset(i32 addrspace(1)* %out, i32 addrspace(3)* %in) {
     51 entry:
     52   %0 = getelementptr i32 addrspace(3)* %in, i32 16385
     53   %1 = load i32 addrspace(3)* %0
     54   store i32 %1, i32 addrspace(1)* %out
     55   ret void
     56 }
     57 
     58 ; CHECK-LABEL: @null_32bit_lds_ptr:
     59 ; CHECK: V_CMP_NE_I32
     60 ; CHECK-NOT: V_CMP_NE_I32
     61 ; CHECK: V_CNDMASK_B32
     62 define void @null_32bit_lds_ptr(i32 addrspace(1)* %out, i32 addrspace(3)* %lds) nounwind {
     63   %cmp = icmp ne i32 addrspace(3)* %lds, null
     64   %x = select i1 %cmp, i32 123, i32 456
     65   store i32 %x, i32 addrspace(1)* %out
     66   ret void
     67 }
     68 
     69 ; CHECK-LABEL: @mul_32bit_ptr:
     70 ; CHECK: V_MUL_LO_I32
     71 ; CHECK-NEXT: V_ADD_I32_e32
     72 ; CHECK-NEXT: DS_READ_B32
     73 define void @mul_32bit_ptr(float addrspace(1)* %out, [3 x float] addrspace(3)* %lds, i32 %tid) {
     74   %ptr = getelementptr [3 x float] addrspace(3)* %lds, i32 %tid, i32 0
     75   %val = load float addrspace(3)* %ptr
     76   store float %val, float addrspace(1)* %out
     77   ret void
     78 }
     79 
     80 @g_lds = addrspace(3) global float zeroinitializer, align 4
     81 
     82 ; CHECK-LABEL: @infer_ptr_alignment_global_offset:
     83 ; CHECK: V_MOV_B32_e32 [[REG:v[0-9]+]], 0
     84 ; CHECK: DS_READ_B32 v{{[0-9]+}}, [[REG]]
     85 define void @infer_ptr_alignment_global_offset(float addrspace(1)* %out, i32 %tid) {
     86   %val = load float addrspace(3)* @g_lds
     87   store float %val, float addrspace(1)* %out
     88   ret void
     89 }
     90 
     91 
     92 @ptr = addrspace(3) global i32 addrspace(3)* null
     93 @dst = addrspace(3) global [16384 x i32] zeroinitializer
     94 
     95 ; CHECK-LABEL: @global_ptr:
     96 ; CHECK: DS_WRITE_B32
     97 define void @global_ptr() nounwind {
     98   store i32 addrspace(3)* getelementptr ([16384 x i32] addrspace(3)* @dst, i32 0, i32 16), i32 addrspace(3)* addrspace(3)* @ptr
     99   ret void
    100 }
    101 
    102 ; CHECK-LABEL: @local_address_store
    103 ; CHECK: DS_WRITE_B32
    104 define void @local_address_store(i32 addrspace(3)* %out, i32 %val) {
    105   store i32 %val, i32 addrspace(3)* %out
    106   ret void
    107 }
    108 
    109 ; CHECK-LABEL: @local_address_gep_store
    110 ; CHECK: S_ADD_I32 [[SADDR:s[0-9]+]],
    111 ; CHECK: V_MOV_B32_e32 [[ADDR:v[0-9]+]], [[SADDR]]
    112 ; CHECK: DS_WRITE_B32 [[ADDR]], v{{[0-9]+}},
    113 define void @local_address_gep_store(i32 addrspace(3)* %out, i32, i32 %val, i32 %offset) {
    114   %gep = getelementptr i32 addrspace(3)* %out, i32 %offset
    115   store i32 %val, i32 addrspace(3)* %gep, align 4
    116   ret void
    117 }
    118 
    119 ; CHECK-LABEL: @local_address_gep_const_offset_store
    120 ; CHECK: V_MOV_B32_e32 [[VPTR:v[0-9]+]], s{{[0-9]+}}
    121 ; CHECK: V_MOV_B32_e32 [[VAL:v[0-9]+]], s{{[0-9]+}}
    122 ; CHECK: DS_WRITE_B32 [[VPTR]], [[VAL]], 0x4
    123 define void @local_address_gep_const_offset_store(i32 addrspace(3)* %out, i32 %val) {
    124   %gep = getelementptr i32 addrspace(3)* %out, i32 1
    125   store i32 %val, i32 addrspace(3)* %gep, align 4
    126   ret void
    127 }
    128 
    129 ; Offset too large, can't fold into 16-bit immediate offset.
    130 ; CHECK-LABEL: @local_address_gep_large_const_offset_store
    131 ; CHECK: S_ADD_I32 [[SPTR:s[0-9]]], s{{[0-9]+}}, 0x10004
    132 ; CHECK: V_MOV_B32_e32 [[VPTR:v[0-9]+]], [[SPTR]]
    133 ; CHECK: DS_WRITE_B32 [[VPTR]], v{{[0-9]+}}, 0
    134 define void @local_address_gep_large_const_offset_store(i32 addrspace(3)* %out, i32 %val) {
    135   %gep = getelementptr i32 addrspace(3)* %out, i32 16385
    136   store i32 %val, i32 addrspace(3)* %gep, align 4
    137   ret void
    138 }
    139