|
| 1 | +; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 4 |
| 2 | +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck -enable-var-scope -check-prefix=GFX9 %s |
| 3 | +; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck -enable-var-scope -check-prefix=GFX11 %s |
| 4 | + |
| 5 | +; We've separated this file from call-args-inreg.ll since GlobalISel does not support the bfloat type. |
| 6 | +; Ideally, we should merge the two files once that support lands. |
| 7 | + |
| 8 | +declare hidden void @external_void_func_bf16_inreg(bfloat inreg) #0 |
| 9 | +declare hidden void @external_void_func_v2bf16_inreg(<2 x bfloat> inreg) #0 |
| 10 | + |
| 11 | +define void @test_call_external_void_func_bf16_inreg(bfloat inreg %arg) #0 { |
| 12 | +; GFX9-LABEL: test_call_external_void_func_bf16_inreg: |
| 13 | +; GFX9: ; %bb.0: |
| 14 | +; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| 15 | +; GFX9-NEXT: s_mov_b32 s17, s33 |
| 16 | +; GFX9-NEXT: s_mov_b32 s33, s32 |
| 17 | +; GFX9-NEXT: s_or_saveexec_b64 s[18:19], -1 |
| 18 | +; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill |
| 19 | +; GFX9-NEXT: s_mov_b64 exec, s[18:19] |
| 20 | +; GFX9-NEXT: v_writelane_b32 v40, s17, 2 |
| 21 | +; GFX9-NEXT: s_addk_i32 s32, 0x400 |
| 22 | +; GFX9-NEXT: v_writelane_b32 v40, s30, 0 |
| 23 | +; GFX9-NEXT: s_getpc_b64 s[18:19] |
| 24 | +; GFX9-NEXT: s_add_u32 s18, s18, external_void_func_bf16_inreg@rel32@lo+4 |
| 25 | +; GFX9-NEXT: s_addc_u32 s19, s19, external_void_func_bf16_inreg@rel32@hi+12 |
| 26 | +; GFX9-NEXT: s_mov_b32 s0, s16 |
| 27 | +; GFX9-NEXT: v_writelane_b32 v40, s31, 1 |
| 28 | +; GFX9-NEXT: s_swappc_b64 s[30:31], s[18:19] |
| 29 | +; GFX9-NEXT: v_readlane_b32 s31, v40, 1 |
| 30 | +; GFX9-NEXT: v_readlane_b32 s30, v40, 0 |
| 31 | +; GFX9-NEXT: s_mov_b32 s32, s33 |
| 32 | +; GFX9-NEXT: v_readlane_b32 s4, v40, 2 |
| 33 | +; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1 |
| 34 | +; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload |
| 35 | +; GFX9-NEXT: s_mov_b64 exec, s[6:7] |
| 36 | +; GFX9-NEXT: s_mov_b32 s33, s4 |
| 37 | +; GFX9-NEXT: s_waitcnt vmcnt(0) |
| 38 | +; GFX9-NEXT: s_setpc_b64 s[30:31] |
| 39 | +; |
| 40 | +; GFX11-LABEL: test_call_external_void_func_bf16_inreg: |
| 41 | +; GFX11: ; %bb.0: |
| 42 | +; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| 43 | +; GFX11-NEXT: s_mov_b32 s1, s33 |
| 44 | +; GFX11-NEXT: s_mov_b32 s33, s32 |
| 45 | +; GFX11-NEXT: s_or_saveexec_b32 s2, -1 |
| 46 | +; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill |
| 47 | +; GFX11-NEXT: s_mov_b32 exec_lo, s2 |
| 48 | +; GFX11-NEXT: v_writelane_b32 v40, s1, 2 |
| 49 | +; GFX11-NEXT: s_add_i32 s32, s32, 16 |
| 50 | +; GFX11-NEXT: s_getpc_b64 s[2:3] |
| 51 | +; GFX11-NEXT: s_add_u32 s2, s2, external_void_func_bf16_inreg@rel32@lo+4 |
| 52 | +; GFX11-NEXT: s_addc_u32 s3, s3, external_void_func_bf16_inreg@rel32@hi+12 |
| 53 | +; GFX11-NEXT: v_writelane_b32 v40, s30, 0 |
| 54 | +; GFX11-NEXT: v_writelane_b32 v40, s31, 1 |
| 55 | +; GFX11-NEXT: s_swappc_b64 s[30:31], s[2:3] |
| 56 | +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| 57 | +; GFX11-NEXT: v_readlane_b32 s31, v40, 1 |
| 58 | +; GFX11-NEXT: v_readlane_b32 s30, v40, 0 |
| 59 | +; GFX11-NEXT: s_mov_b32 s32, s33 |
| 60 | +; GFX11-NEXT: v_readlane_b32 s0, v40, 2 |
| 61 | +; GFX11-NEXT: s_or_saveexec_b32 s1, -1 |
| 62 | +; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload |
| 63 | +; GFX11-NEXT: s_mov_b32 exec_lo, s1 |
| 64 | +; GFX11-NEXT: s_mov_b32 s33, s0 |
| 65 | +; GFX11-NEXT: s_waitcnt vmcnt(0) |
| 66 | +; GFX11-NEXT: s_setpc_b64 s[30:31] |
| 67 | + call void @external_void_func_bf16_inreg(bfloat inreg %arg) |
| 68 | + ret void |
| 69 | +} |
| 70 | + |
| 71 | +define void @test_call_external_void_func_v2bf16_inreg(<2 x bfloat> inreg %arg) #0 { |
| 72 | +; GFX9-LABEL: test_call_external_void_func_v2bf16_inreg: |
| 73 | +; GFX9: ; %bb.0: |
| 74 | +; GFX9-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| 75 | +; GFX9-NEXT: s_mov_b32 s17, s33 |
| 76 | +; GFX9-NEXT: s_mov_b32 s33, s32 |
| 77 | +; GFX9-NEXT: s_or_saveexec_b64 s[18:19], -1 |
| 78 | +; GFX9-NEXT: buffer_store_dword v40, off, s[0:3], s33 ; 4-byte Folded Spill |
| 79 | +; GFX9-NEXT: s_mov_b64 exec, s[18:19] |
| 80 | +; GFX9-NEXT: v_writelane_b32 v40, s17, 2 |
| 81 | +; GFX9-NEXT: s_addk_i32 s32, 0x400 |
| 82 | +; GFX9-NEXT: v_writelane_b32 v40, s30, 0 |
| 83 | +; GFX9-NEXT: s_getpc_b64 s[18:19] |
| 84 | +; GFX9-NEXT: s_add_u32 s18, s18, external_void_func_v2bf16_inreg@rel32@lo+4 |
| 85 | +; GFX9-NEXT: s_addc_u32 s19, s19, external_void_func_v2bf16_inreg@rel32@hi+12 |
| 86 | +; GFX9-NEXT: s_mov_b32 s0, s16 |
| 87 | +; GFX9-NEXT: v_writelane_b32 v40, s31, 1 |
| 88 | +; GFX9-NEXT: s_swappc_b64 s[30:31], s[18:19] |
| 89 | +; GFX9-NEXT: v_readlane_b32 s31, v40, 1 |
| 90 | +; GFX9-NEXT: v_readlane_b32 s30, v40, 0 |
| 91 | +; GFX9-NEXT: s_mov_b32 s32, s33 |
| 92 | +; GFX9-NEXT: v_readlane_b32 s4, v40, 2 |
| 93 | +; GFX9-NEXT: s_or_saveexec_b64 s[6:7], -1 |
| 94 | +; GFX9-NEXT: buffer_load_dword v40, off, s[0:3], s33 ; 4-byte Folded Reload |
| 95 | +; GFX9-NEXT: s_mov_b64 exec, s[6:7] |
| 96 | +; GFX9-NEXT: s_mov_b32 s33, s4 |
| 97 | +; GFX9-NEXT: s_waitcnt vmcnt(0) |
| 98 | +; GFX9-NEXT: s_setpc_b64 s[30:31] |
| 99 | +; |
| 100 | +; GFX11-LABEL: test_call_external_void_func_v2bf16_inreg: |
| 101 | +; GFX11: ; %bb.0: |
| 102 | +; GFX11-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0) |
| 103 | +; GFX11-NEXT: s_mov_b32 s1, s33 |
| 104 | +; GFX11-NEXT: s_mov_b32 s33, s32 |
| 105 | +; GFX11-NEXT: s_or_saveexec_b32 s2, -1 |
| 106 | +; GFX11-NEXT: scratch_store_b32 off, v40, s33 ; 4-byte Folded Spill |
| 107 | +; GFX11-NEXT: s_mov_b32 exec_lo, s2 |
| 108 | +; GFX11-NEXT: v_writelane_b32 v40, s1, 2 |
| 109 | +; GFX11-NEXT: s_add_i32 s32, s32, 16 |
| 110 | +; GFX11-NEXT: s_getpc_b64 s[2:3] |
| 111 | +; GFX11-NEXT: s_add_u32 s2, s2, external_void_func_v2bf16_inreg@rel32@lo+4 |
| 112 | +; GFX11-NEXT: s_addc_u32 s3, s3, external_void_func_v2bf16_inreg@rel32@hi+12 |
| 113 | +; GFX11-NEXT: v_writelane_b32 v40, s30, 0 |
| 114 | +; GFX11-NEXT: v_writelane_b32 v40, s31, 1 |
| 115 | +; GFX11-NEXT: s_swappc_b64 s[30:31], s[2:3] |
| 116 | +; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1) |
| 117 | +; GFX11-NEXT: v_readlane_b32 s31, v40, 1 |
| 118 | +; GFX11-NEXT: v_readlane_b32 s30, v40, 0 |
| 119 | +; GFX11-NEXT: s_mov_b32 s32, s33 |
| 120 | +; GFX11-NEXT: v_readlane_b32 s0, v40, 2 |
| 121 | +; GFX11-NEXT: s_or_saveexec_b32 s1, -1 |
| 122 | +; GFX11-NEXT: scratch_load_b32 v40, off, s33 ; 4-byte Folded Reload |
| 123 | +; GFX11-NEXT: s_mov_b32 exec_lo, s1 |
| 124 | +; GFX11-NEXT: s_mov_b32 s33, s0 |
| 125 | +; GFX11-NEXT: s_waitcnt vmcnt(0) |
| 126 | +; GFX11-NEXT: s_setpc_b64 s[30:31] |
| 127 | + call void @external_void_func_v2bf16_inreg(<2 x bfloat> inreg %arg) |
| 128 | + ret void |
| 129 | +} |
| 130 | + |
0 commit comments