llvm/test/CodeGen/AMDGPU/lower-work-group-id-intrinsics-pal.ll

   1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
   2 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9 %s
   3 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=-architected-sgprs -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9 %s
   4 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9ARCH-SDAG %s
   5 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 -mattr=+architected-sgprs -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9ARCH-GISEL %s
   6 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12,GFX12-SDAG %s
   7 ; RUN: llc -mtriple=amdgcn-amd-amdpal -mcpu=gfx1200 -global-isel=1 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX12,GFX12-GISEL %s
   8
   9 define amdgpu_cs void @_amdgpu_cs_main() {
  10 ; GFX9-LABEL: _amdgpu_cs_main:
  11 ; GFX9:       ; %bb.0: ; %.entry
  12 ; GFX9-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  13 ; GFX9-NEXT:    s_endpgm
  14 ;
  15 ; GFX9ARCH-SDAG-LABEL: _amdgpu_cs_main:
  16 ; GFX9ARCH-SDAG:       ; %bb.0: ; %.entry
  17 ; GFX9ARCH-SDAG-NEXT:    s_lshr_b32 s0, ttmp7, 16
  18 ; GFX9ARCH-SDAG-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  19 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
  20 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v1, s1
  21 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v2, s0
  22 ; GFX9ARCH-SDAG-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  23 ; GFX9ARCH-SDAG-NEXT:    s_endpgm
  24 ;
  25 ; GFX9ARCH-GISEL-LABEL: _amdgpu_cs_main:
  26 ; GFX9ARCH-GISEL:       ; %bb.0: ; %.entry
  27 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s0, ttmp9
  28 ; GFX9ARCH-GISEL-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  29 ; GFX9ARCH-GISEL-NEXT:    s_lshr_b32 s2, ttmp7, 16
  30 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s0
  31 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v1, s1
  32 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v2, s2
  33 ; GFX9ARCH-GISEL-NEXT:    buffer_store_dwordx3 v[0:2], off, s[0:3], 0
  34 ; GFX9ARCH-GISEL-NEXT:    s_endpgm
  35 ;
  36 ; GFX12-SDAG-LABEL: _amdgpu_cs_main:
  37 ; GFX12-SDAG:       ; %bb.0: ; %.entry
  38 ; GFX12-SDAG-NEXT:    s_and_b32 s0, ttmp7, 0xffff
  39 ; GFX12-SDAG-NEXT:    s_lshr_b32 s1, ttmp7, 16
  40 ; GFX12-SDAG-NEXT:    v_dual_mov_b32 v0, ttmp9 :: v_dual_mov_b32 v1, s0
  41 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v2, s1
  42 ; GFX12-SDAG-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], null
  43 ; GFX12-SDAG-NEXT:    s_nop 0
  44 ; GFX12-SDAG-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  45 ; GFX12-SDAG-NEXT:    s_endpgm
  46 ;
  47 ; GFX12-GISEL-LABEL: _amdgpu_cs_main:
  48 ; GFX12-GISEL:       ; %bb.0: ; %.entry
  49 ; GFX12-GISEL-NEXT:    s_mov_b32 s0, ttmp9
  50 ; GFX12-GISEL-NEXT:    s_and_b32 s1, ttmp7, 0xffff
  51 ; GFX12-GISEL-NEXT:    s_lshr_b32 s2, ttmp7, 16
  52 ; GFX12-GISEL-NEXT:    v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s1
  53 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v2, s2
  54 ; GFX12-GISEL-NEXT:    buffer_store_b96 v[0:2], off, s[0:3], null
  55 ; GFX12-GISEL-NEXT:    s_nop 0
  56 ; GFX12-GISEL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
  57 ; GFX12-GISEL-NEXT:    s_endpgm
  58 .entry:
  59   %idx = call i32 @llvm.amdgcn.workgroup.id.x()
  60   %idy = call i32 @llvm.amdgcn.workgroup.id.y()
  61   %idz = call i32 @llvm.amdgcn.workgroup.id.z()
  62   %ielemx = insertelement <3 x i32> undef, i32 %idx, i64 0
  63   %ielemy = insertelement <3 x i32> %ielemx, i32 %idy, i64 1
  64   %ielemz = insertelement <3 x i32> %ielemy, i32 %idz, i64 2
  65   call void @llvm.amdgcn.raw.ptr.buffer.store.v3i32(<3 x i32> %ielemz, ptr addrspace(8) undef, i32 0, i32 0, i32 0)
  66   ret void
  67 }
  68
  69 define amdgpu_cs void @caller() {
  70 ; GFX9ARCH-SDAG-LABEL: caller:
  71 ; GFX9ARCH-SDAG:       ; %bb.0:
  72 ; GFX9ARCH-SDAG-NEXT:    s_getpc_b64 s[8:9]
  73 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s8, s0
  74 ; GFX9ARCH-SDAG-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
  75 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s5, callee@abs32@hi
  76 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s4, callee@abs32@lo
  77 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
  78 ; GFX9ARCH-SDAG-NEXT:    s_mov_b32 s32, 0
  79 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt lgkmcnt(0)
  80 ; GFX9ARCH-SDAG-NEXT:    s_add_u32 s8, s8, s0
  81 ; GFX9ARCH-SDAG-NEXT:    s_addc_u32 s9, s9, 0
  82 ; GFX9ARCH-SDAG-NEXT:    s_mov_b64 s[0:1], s[8:9]
  83 ; GFX9ARCH-SDAG-NEXT:    s_mov_b64 s[2:3], s[10:11]
  84 ; GFX9ARCH-SDAG-NEXT:    s_swappc_b64 s[30:31], s[4:5]
  85 ; GFX9ARCH-SDAG-NEXT:    s_endpgm
  86 ;
  87 ; GFX9ARCH-GISEL-LABEL: caller:
  88 ; GFX9ARCH-GISEL:       ; %bb.0:
  89 ; GFX9ARCH-GISEL-NEXT:    s_getpc_b64 s[8:9]
  90 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s8, s0
  91 ; GFX9ARCH-GISEL-NEXT:    s_load_dwordx4 s[8:11], s[8:9], 0x10
  92 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s4, callee@abs32@lo
  93 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s5, callee@abs32@hi
  94 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, ttmp9
  95 ; GFX9ARCH-GISEL-NEXT:    s_mov_b32 s32, 0
  96 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt lgkmcnt(0)
  97 ; GFX9ARCH-GISEL-NEXT:    s_add_u32 s8, s8, s0
  98 ; GFX9ARCH-GISEL-NEXT:    s_addc_u32 s9, s9, 0
  99 ; GFX9ARCH-GISEL-NEXT:    s_mov_b64 s[0:1], s[8:9]
 100 ; GFX9ARCH-GISEL-NEXT:    s_mov_b64 s[2:3], s[10:11]
 101 ; GFX9ARCH-GISEL-NEXT:    s_swappc_b64 s[30:31], s[4:5]
 102 ; GFX9ARCH-GISEL-NEXT:    s_endpgm
 103 ;
 104 ; GFX12-SDAG-LABEL: caller:
 105 ; GFX12-SDAG:       ; %bb.0:
 106 ; GFX12-SDAG-NEXT:    v_mov_b32_e32 v0, ttmp9
 107 ; GFX12-SDAG-NEXT:    s_mov_b32 s1, callee@abs32@hi
 108 ; GFX12-SDAG-NEXT:    s_mov_b32 s0, callee@abs32@lo
 109 ; GFX12-SDAG-NEXT:    s_mov_b32 s32, 0
 110 ; GFX12-SDAG-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 111 ; GFX12-SDAG-NEXT:    s_endpgm
 112 ;
 113 ; GFX12-GISEL-LABEL: caller:
 114 ; GFX12-GISEL:       ; %bb.0:
 115 ; GFX12-GISEL-NEXT:    v_mov_b32_e32 v0, ttmp9
 116 ; GFX12-GISEL-NEXT:    s_mov_b32 s0, callee@abs32@lo
 117 ; GFX12-GISEL-NEXT:    s_mov_b32 s1, callee@abs32@hi
 118 ; GFX12-GISEL-NEXT:    s_mov_b32 s32, 0
 119 ; GFX12-GISEL-NEXT:    s_swappc_b64 s[30:31], s[0:1]
 120 ; GFX12-GISEL-NEXT:    s_endpgm
 121   %idx = call i32 @llvm.amdgcn.workgroup.id.x()
 122   call amdgpu_gfx void @callee(i32 %idx)
 123   ret void
 124 }
 125
 126 declare amdgpu_gfx void @callee(i32)
 127
 128 define amdgpu_gfx void @workgroup_ids_gfx(ptr addrspace(1) %outx, ptr addrspace(1) %outy, ptr addrspace(1) %outz) {
 129 ; GFX9-LABEL: workgroup_ids_gfx:
 130 ; GFX9:       ; %bb.0:
 131 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 132 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 133 ;
 134 ; GFX9ARCH-SDAG-LABEL: workgroup_ids_gfx:
 135 ; GFX9ARCH-SDAG:       ; %bb.0:
 136 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 137 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v6, ttmp9
 138 ; GFX9ARCH-SDAG-NEXT:    s_and_b32 s34, ttmp7, 0xffff
 139 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[0:1], v6, off
 140 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 141 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, s34
 142 ; GFX9ARCH-SDAG-NEXT:    s_lshr_b32 s34, ttmp7, 16
 143 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[2:3], v0, off
 144 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 145 ; GFX9ARCH-SDAG-NEXT:    v_mov_b32_e32 v0, s34
 146 ; GFX9ARCH-SDAG-NEXT:    global_store_dword v[4:5], v0, off
 147 ; GFX9ARCH-SDAG-NEXT:    s_waitcnt vmcnt(0)
 148 ; GFX9ARCH-SDAG-NEXT:    s_setpc_b64 s[30:31]
 149 ;
 150 ; GFX9ARCH-GISEL-LABEL: workgroup_ids_gfx:
 151 ; GFX9ARCH-GISEL:       ; %bb.0:
 152 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 153 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v6, ttmp9
 154 ; GFX9ARCH-GISEL-NEXT:    s_and_b32 s34, ttmp7, 0xffff
 155 ; GFX9ARCH-GISEL-NEXT:    s_lshr_b32 s35, ttmp7, 16
 156 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[0:1], v6, off
 157 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 158 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s34
 159 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[2:3], v0, off
 160 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 161 ; GFX9ARCH-GISEL-NEXT:    v_mov_b32_e32 v0, s35
 162 ; GFX9ARCH-GISEL-NEXT:    global_store_dword v[4:5], v0, off
 163 ; GFX9ARCH-GISEL-NEXT:    s_waitcnt vmcnt(0)
 164 ; GFX9ARCH-GISEL-NEXT:    s_setpc_b64 s[30:31]
 165 ;
 166 ; GFX12-LABEL: workgroup_ids_gfx:
 167 ; GFX12:       ; %bb.0:
 168 ; GFX12-NEXT:    s_wait_loadcnt_dscnt 0x0
 169 ; GFX12-NEXT:    s_wait_expcnt 0x0
 170 ; GFX12-NEXT:    s_wait_samplecnt 0x0
 171 ; GFX12-NEXT:    s_wait_bvhcnt 0x0
 172 ; GFX12-NEXT:    s_wait_kmcnt 0x0
 173 ; GFX12-NEXT:    s_and_b32 s0, ttmp7, 0xffff
 174 ; GFX12-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
 175 ; GFX12-NEXT:    v_dual_mov_b32 v6, ttmp9 :: v_dual_mov_b32 v7, s0
 176 ; GFX12-NEXT:    s_lshr_b32 s1, ttmp7, 16
 177 ; GFX12-NEXT:    v_mov_b32_e32 v8, s1
 178 ; GFX12-NEXT:    s_wait_storecnt 0x0
 179 ; GFX12-NEXT:    global_store_b32 v[0:1], v6, off scope:SCOPE_SYS
 180 ; GFX12-NEXT:    s_wait_storecnt 0x0
 181 ; GFX12-NEXT:    global_store_b32 v[2:3], v7, off scope:SCOPE_SYS
 182 ; GFX12-NEXT:    s_wait_storecnt 0x0
 183 ; GFX12-NEXT:    global_store_b32 v[4:5], v8, off scope:SCOPE_SYS
 184 ; GFX12-NEXT:    s_wait_storecnt 0x0
 185 ; GFX12-NEXT:    s_setpc_b64 s[30:31]
 186   %id.x = call i32 @llvm.amdgcn.workgroup.id.x()
 187   %id.y = call i32 @llvm.amdgcn.workgroup.id.y()
 188   %id.z = call i32 @llvm.amdgcn.workgroup.id.z()
 189   store volatile i32 %id.x, ptr addrspace(1) %outx
 190   store volatile i32 %id.y, ptr addrspace(1) %outy
 191   store volatile i32 %id.z, ptr addrspace(1) %outz
 192   ret void
 193 }