1# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py 2# RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX9,GFX9_10 %s 3# RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX10,GFX9_10 %s 4 5# Check that we handle cases where a counter has overflowed. 6 7# Overflows lgkmcnt with gfx9 but not with gfx10. 8--- 9name: max-counter-lgkmcnt 10body: | 11 bb.0: 12 liveins: $vgpr99 13 14 ; GFX9-LABEL: name: max-counter-lgkmcnt 15 ; GFX9: S_WAITCNT 0 16 ; GFX9: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec 17 ; GFX9-NOT: S_WAITCNT 53119 18 ; GFX9-NEXT: S_WAITCNT 52863 19 ; GFX9-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 20 ; GFX9-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec 21 ; GFX9-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec 22 ; GFX9-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec 23 ; GFX9-NEXT: S_ENDPGM 0 24 ; GFX10-LABEL: name: max-counter-lgkmcnt 25 ; GFX10: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec 26 ; GFX10-NEXT: S_WAITCNT 53631 27 ; GFX10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 28 ; GFX10-NEXT: S_WAITCNT 53375 29 ; GFX10-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec 30 ; GFX10-NEXT: S_WAITCNT 53119 31 ; GFX10-NEXT: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec 32 ; GFX10-NEXT: S_WAITCNT 52863 33 ; GFX10-NEXT: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec 34 ; GFX10-NEXT: S_ENDPGM 0 35 $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec 36 $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec 37 $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec 38 $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec 39 $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec 40 $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec 41 $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec 42 $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec 43 $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec 44 $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec 45 $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec 46 $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec 47 $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec 48 $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec 49 $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec 50 $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec 51 $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec 52 $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec 53 $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 54 $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec 55 $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec 56 $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec 57 S_ENDPGM 0 58... 59 60# Overflows vmcnt with gfx9 and gfx10. 61--- 62name: max-counter-vmcnt 63body: | 64 bb.0: 65 liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4 66 67 ; GFX9_10-LABEL: name: max-counter-vmcnt 68 ; GFX9_10: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, 0, 0, implicit $exec 69 ; GFX9-NOT: S_WAITCNT 53119 70 ; GFX10-NOT: S_WAITCNT 65407 71 ; GFX9-NEXT: S_WAITCNT 53118 72 ; GFX10-NEXT: S_WAITCNT 65406 73 ; GFX9_10-NEXT: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 74 ; GFX9_10-NEXT: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec 75 ; GFX9_10-NEXT: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec 76 ; GFX9_10-NEXT: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec 77 ; GFX9_10-NEXT: S_ENDPGM 0 78 $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, 0, 0, implicit $exec 79 $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, 0, 0, implicit $exec 80 $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, 0, 0, implicit $exec 81 $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, 0, 0, implicit $exec 82 $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, 0, 0, implicit $exec 83 $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, 0, 0, implicit $exec 84 $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, 0, 0, implicit $exec 85 $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, 0, 0, implicit $exec 86 $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, 0, 0, implicit $exec 87 $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, 0, 0, implicit $exec 88 $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, 0, 0, implicit $exec 89 $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, 0, 0, implicit $exec 90 $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, 0, 0, implicit $exec 91 $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, 0, 0, implicit $exec 92 $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, 0, 0, implicit $exec 93 $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, 0, 0, implicit $exec 94 $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, 0, 0, implicit $exec 95 $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, 0, 0, implicit $exec 96 $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, 0, 0, implicit $exec 97 $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, 0, 0, implicit $exec 98 $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, 0, 0, implicit $exec 99 $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, 0, 0, implicit $exec 100 $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, 0, 0, implicit $exec 101 $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, 0, 0, implicit $exec 102 $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, 0, 0, implicit $exec 103 $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, 0, 0, implicit $exec 104 $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, 0, 0, implicit $exec 105 $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, 0, 0, implicit $exec 106 $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, 0, 0, implicit $exec 107 $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, 0, 0, implicit $exec 108 $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, 0, 0, implicit $exec 109 $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, 0, 0, implicit $exec 110 $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, 0, 0, implicit $exec 111 $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, 0, 0, implicit $exec 112 $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, 0, 0, implicit $exec 113 $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, 0, 0, implicit $exec 114 $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, 0, 0, implicit $exec 115 $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, 0, 0, implicit $exec 116 $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, 0, 0, implicit $exec 117 $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, 0, 0, implicit $exec 118 $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, 0, 0, implicit $exec 119 $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, 0, 0, implicit $exec 120 $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, 0, 0, implicit $exec 121 $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, 0, 0, implicit $exec 122 $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, 0, 0, implicit $exec 123 $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, 0, 0, implicit $exec 124 $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, 0, 0, implicit $exec 125 $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, 0, 0, implicit $exec 126 $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, 0, 0, implicit $exec 127 $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, 0, 0, implicit $exec 128 $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, 0, 0, implicit $exec 129 $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, 0, 0, implicit $exec 130 $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, 0, 0, implicit $exec 131 $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, 0, 0, implicit $exec 132 $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, 0, 0, implicit $exec 133 $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, 0, 0, implicit $exec 134 $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, 0, 0, implicit $exec 135 $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, 0, 0, implicit $exec 136 $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, 0, 0, implicit $exec 137 $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, 0, 0, implicit $exec 138 $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, 0, 0, implicit $exec 139 $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, 0, 0, implicit $exec 140 $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, 0, 0, implicit $exec 141 $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, 0, 0, implicit $exec 142 $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, 0, 0, implicit $exec 143 $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, 0, 0, implicit $exec 144 $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, 0, 0, implicit $exec 145 $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 146 $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec 147 $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec 148 $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec 149 S_ENDPGM 0 150... 151 152# Should be no waitcnt if expcnt overflows. 153--- 154name: max-counter-expcnt 155body: | 156 bb.0: 157 liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0, $vgpr1 158 159 ; GFX9_10-LABEL: name: max-counter-expcnt 160 ; GFX9_10: EXP 161 ; GFX9_10-NOT: S_WAITCNT 162 EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec 163 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 164 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 165 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 166 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 167 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 168 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 169 EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec 170 $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec 171 S_ENDPGM 0 172... 173