reference, declarationdefinition
definition → references, declarations, derived classes, virtual overrides
reference to multiple definitions → definitions
unreferenced
    1
    2
    3
    4
    5
    6
    7
    8
    9
   10
   11
   12
   13
   14
   15
   16
   17
   18
   19
   20
   21
   22
   23
   24
   25
   26
   27
   28
   29
   30
   31
   32
   33
   34
   35
   36
   37
   38
   39
   40
   41
   42
   43
   44
   45
   46
   47
   48
   49
   50
   51
   52
   53
   54
   55
   56
   57
   58
   59
   60
   61
   62
   63
   64
   65
   66
   67
   68
   69
   70
   71
   72
   73
   74
   75
   76
   77
   78
   79
   80
   81
   82
   83
   84
   85
   86
   87
   88
   89
   90
   91
   92
   93
   94
   95
   96
   97
   98
   99
  100
  101
  102
  103
  104
  105
  106
  107
  108
  109
  110
  111
  112
  113
  114
  115
  116
  117
  118
  119
  120
  121
  122
  123
  124
  125
  126
  127
  128
  129
  130
  131
  132
  133
  134
  135
  136
  137
  138
  139
  140
  141
  142
  143
  144
  145
  146
  147
  148
  149
  150
  151
  152
  153
  154
  155
  156
  157
  158
  159
  160
  161
  162
  163
  164
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=kaveri -verify-machineinstrs < %s | FileCheck -check-prefix=CI -check-prefix=GCN -check-prefix=CIVI %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=VI -check-prefix=GFX89 -check-prefix=GCN -check-prefix=CIVI %s
; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX89 -check-prefix=GFX9 -check-prefix=GCN %s

; GCN-LABEL: {{^}}fneg_fabs_fadd_f16:
; CI-DAG: v_cvt_f32_f16_e32
; CI-DAG: v_cvt_f32_f16_e64 [[CVT_ABS_X:v[0-9]+]], |s{{[0-9]+}}|
; CI: v_sub_f32_e32 v{{[0-9]+}}, v{{[0-9]+}}, [[CVT_ABS_X]]

; GFX89-NOT: _and
; GFX89: v_sub_f16_e64 {{v[0-9]+}}, {{s[0-9]+}}, |{{v[0-9]+}}|
define amdgpu_kernel void @fneg_fabs_fadd_f16(half addrspace(1)* %out, half %x, half %y) {
  %fabs = call half @llvm.fabs.f16(half %x)
  %fsub = fsub half -0.0, %fabs
  %fadd = fadd half %y, %fsub
  store half %fadd, half addrspace(1)* %out, align 2
  ret void
}

; GCN-LABEL: {{^}}fneg_fabs_fmul_f16:
; CI-DAG: v_cvt_f32_f16_e32
; CI-DAG: v_cvt_f32_f16_e64 [[CVT_NEG_ABS_X:v[0-9]+]], -|{{s[0-9]+}}|
; CI: v_mul_f32_e32 {{v[0-9]+}},  {{v[0-9]+}}, [[CVT_NEG_ABS_X]]
; CI: v_cvt_f16_f32_e32

; GFX89-NOT: _and
; GFX89: v_mul_f16_e64 [[MUL:v[0-9]+]], {{s[0-9]+}}, -|{{v[0-9]+}}|
; GFX89-NOT: [[MUL]]
; GFX89: {{flat|global}}_store_short v{{\[[0-9]+:[0-9]+\]}}, [[MUL]]
define amdgpu_kernel void @fneg_fabs_fmul_f16(half addrspace(1)* %out, half %x, half %y) {
  %fabs = call half @llvm.fabs.f16(half %x)
  %fsub = fsub half -0.0, %fabs
  %fmul = fmul half %y, %fsub
  store half %fmul, half addrspace(1)* %out, align 2
  ret void
}

; DAGCombiner will transform:
; (fabs (f16 bitcast (i16 a))) => (f16 bitcast (and (i16 a), 0x7FFFFFFF))
; unless isFabsFree returns true

; GCN-LABEL: {{^}}fneg_fabs_free_f16:
; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, 0x8000
define amdgpu_kernel void @fneg_fabs_free_f16(half addrspace(1)* %out, i16 %in) {
  %bc = bitcast i16 %in to half
  %fabs = call half @llvm.fabs.f16(half %bc)
  %fsub = fsub half -0.0, %fabs
  store half %fsub, half addrspace(1)* %out
  ret void
}

; GCN-LABEL: {{^}}fneg_fabs_f16:
; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, 0x8000
define amdgpu_kernel void @fneg_fabs_f16(half addrspace(1)* %out, half %in) {
  %fabs = call half @llvm.fabs.f16(half %in)
  %fsub = fsub half -0.0, %fabs
  store half %fsub, half addrspace(1)* %out, align 2
  ret void
}

; GCN-LABEL: {{^}}v_fneg_fabs_f16:
; GCN: v_or_b32_e32 v{{[0-9]+}}, 0x8000, v{{[0-9]+}}
define amdgpu_kernel void @v_fneg_fabs_f16(half addrspace(1)* %out, half addrspace(1)* %in) {
  %val = load half, half addrspace(1)* %in, align 2
  %fabs = call half @llvm.fabs.f16(half %val)
  %fsub = fsub half -0.0, %fabs
  store half %fsub, half addrspace(1)* %out, align 2
  ret void
}

; GCN-LABEL: {{^}}s_fneg_fabs_v2f16_non_bc_src:
; GFX9-DAG: s_load_dword [[VAL:s[0-9]+]]
; GFX9-DAG: v_mov_b32_e32 [[K:v[0-9]+]], 0x40003c00
; GFX9: v_pk_add_f16 [[ADD:v[0-9]+]], [[VAL]], [[K]]
; GFX9: v_or_b32_e32 [[RESULT:v[0-9]+]], 0x80008000, [[ADD]]

; VI: v_or_b32_e32 v{{[0-9]+}}, 0x80008000, v{{[0-9]+}}
define amdgpu_kernel void @s_fneg_fabs_v2f16_non_bc_src(<2 x half> addrspace(1)* %out, <2 x half> %in) {
  %add = fadd <2 x half> %in, <half 1.0, half 2.0>
  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %add)
  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
  store <2 x half> %fneg.fabs, <2 x half> addrspace(1)* %out
  ret void
}

; FIXME: single bit op

; Combine turns this into integer op when bitcast source (from load)

; GCN-LABEL: {{^}}s_fneg_fabs_v2f16_bc_src:
; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, 0x80008000
define amdgpu_kernel void @s_fneg_fabs_v2f16_bc_src(<2 x half> addrspace(1)* %out, <2 x half> %in) {
  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
  store <2 x half> %fneg.fabs, <2 x half> addrspace(1)* %out
  ret void
}

; GCN-LABEL: {{^}}fneg_fabs_v4f16:
; GCN: s_mov_b32 [[MASK:s[0-9]+]], 0x80008000
; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, [[MASK]]
; GCN: s_or_b32 s{{[0-9]+}}, s{{[0-9]+}}, [[MASK]]
; GCN: {{flat|global}}_store_dwordx2
define amdgpu_kernel void @fneg_fabs_v4f16(<4 x half> addrspace(1)* %out, <4 x half> %in) {
  %fabs = call <4 x half> @llvm.fabs.v4f16(<4 x half> %in)
  %fsub = fsub <4 x half> <half -0.0, half -0.0, half -0.0, half -0.0>, %fabs
  store <4 x half> %fsub, <4 x half> addrspace(1)* %out
  ret void
}

; GCN-LABEL: {{^}}fold_user_fneg_fabs_v2f16:
; CI: s_load_dword [[IN:s[0-9]+]]
; CI: s_lshr_b32
; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|
; CI: v_cvt_f32_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|
; CI: v_mul_f32_e32 v{{[0-9]+}}, -4.0, v{{[0-9]+}}
; CI: v_mul_f32_e32 v{{[0-9]+}}, -4.0, v{{[0-9]+}}

; VI: v_mul_f16_e64 v{{[0-9]+}}, |s{{[0-9]+}}|, -4.0
; VI: v_mul_f16_sdwa v{{[0-9]+}}, |v{{[0-9]+}}|, v{{[0-9]+}} dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD

; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
; GFX9: v_pk_mul_f16 v{{[0-9]+}}, [[ABS]], -4.0 op_sel_hi:[1,0]
define amdgpu_kernel void @fold_user_fneg_fabs_v2f16(<2 x half> addrspace(1)* %out, <2 x half> %in) #0 {
  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
  %fneg.fabs = fsub <2 x half> <half -0.0, half -0.0>, %fabs
  %mul = fmul <2 x half> %fneg.fabs, <half 4.0, half 4.0>
  store <2 x half> %mul, <2 x half> addrspace(1)* %out
  ret void
}

; GCN-LABEL: {{^}}s_fneg_multi_use_fabs_v2f16:
; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
; GFX9: v_mov_b32_e32 [[V_ABS:v[0-9]+]], [[ABS]]
; GFX9: s_xor_b32 [[NEG:s[0-9]+]], [[ABS]], 0x80008000
; GFX9-DAG: v_mov_b32_e32 [[V_NEG:v[0-9]+]], [[NEG]]
; GFX9-DAG: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_ABS]]
; GFX9: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_NEG]]
define amdgpu_kernel void @s_fneg_multi_use_fabs_v2f16(<2 x half> addrspace(1)* %out0, <2 x half> addrspace(1)* %out1, <2 x half> %in) {
  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
  %fneg = fsub <2 x half> <half -0.0, half -0.0>, %fabs
  store <2 x half> %fabs, <2 x half> addrspace(1)* %out0
  store <2 x half> %fneg, <2 x half> addrspace(1)* %out1
  ret void
}

; GCN-LABEL: {{^}}s_fneg_multi_use_fabs_foldable_neg_v2f16:
; GFX9: s_and_b32 [[ABS:s[0-9]+]], s{{[0-9]+}}, 0x7fff7fff
; GFX9: v_pk_mul_f16 v{{[0-9]+}}, [[ABS]], -4.0 op_sel_hi:[1,0]
define amdgpu_kernel void @s_fneg_multi_use_fabs_foldable_neg_v2f16(<2 x half> addrspace(1)* %out0, <2 x half> addrspace(1)* %out1, <2 x half> %in) {
  %fabs = call <2 x half> @llvm.fabs.v2f16(<2 x half> %in)
  %fneg = fsub <2 x half> <half -0.0, half -0.0>, %fabs
  %mul = fmul <2 x half> %fneg, <half 4.0, half 4.0>
  store <2 x half> %fabs, <2 x half> addrspace(1)* %out0
  store <2 x half> %mul, <2 x half> addrspace(1)* %out1
  ret void
}

declare half @llvm.fabs.f16(half) #1
declare <2 x half> @llvm.fabs.v2f16(<2 x half>) #1
declare <4 x half> @llvm.fabs.v4f16(<4 x half>) #1

attributes #0 = { nounwind }
attributes #1 = { nounwind readnone }