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 %s
3# RUN: llc -march=amdgcn -mcpu=gfx1010 -verify-machineinstrs -run-pass si-insert-waitcnts -o - %s | FileCheck -check-prefixes=GFX10 %s
4
5--- |
6    define amdgpu_kernel void @max-counter-lgkmcnt() #0 { ret void }
7    define amdgpu_kernel void @max-counter-vmcnt() #0 { ret void }
8    define amdgpu_kernel void @max-counter-expcnt() #0 { ret void }
9
10    attributes #0 = { "amdgpu-flat-work-group-size"="1,256" }
11...
12
13# Check that we handle cases where a counter has overflowed.
14
15# Overflows lgkmcnt with gfx9 but not with gfx10.
16---
17name:            max-counter-lgkmcnt
18body:             |
19  bb.0:
20    liveins: $vgpr99
21
22    ; GFX9-LABEL: name: max-counter-lgkmcnt
23    ; GFX9: S_WAITCNT 0
24    ; GFX9: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
25    ; GFX9: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
26    ; GFX9: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
27    ; GFX9: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
28    ; GFX9: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
29    ; GFX9: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
30    ; GFX9: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
31    ; GFX9: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
32    ; GFX9: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
33    ; GFX9: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
34    ; GFX9: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
35    ; GFX9: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
36    ; GFX9: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
37    ; GFX9: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
38    ; GFX9: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
39    ; GFX9: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
40    ; GFX9: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
41    ; GFX9: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
42    ; GFX9: S_WAITCNT 52863
43    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
44    ; GFX9: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
45    ; GFX9: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
46    ; GFX9: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
47    ; GFX9: S_ENDPGM 0
48    ; GFX10-LABEL: name: max-counter-lgkmcnt
49    ; GFX10: S_WAITCNT 0
50    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
51    ; GFX10: $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
52    ; GFX10: $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
53    ; GFX10: $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
54    ; GFX10: $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
55    ; GFX10: $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
56    ; GFX10: $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
57    ; GFX10: $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
58    ; GFX10: $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
59    ; GFX10: $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
60    ; GFX10: $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
61    ; GFX10: $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
62    ; GFX10: $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
63    ; GFX10: $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
64    ; GFX10: $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
65    ; GFX10: $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
66    ; GFX10: $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
67    ; GFX10: $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
68    ; GFX10: $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
69    ; GFX10: S_WAITCNT 53631
70    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
71    ; GFX10: S_WAITCNT 53375
72    ; GFX10: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
73    ; GFX10: S_WAITCNT 53119
74    ; GFX10: $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
75    ; GFX10: S_WAITCNT 52863
76    ; GFX10: $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
77    ; GFX10: S_ENDPGM 0
78    $vgpr0_vgpr1 = DS_READ2_B32_gfx9 renamable $vgpr99, 0, 1, 0, implicit $exec
79    $vgpr2_vgpr3 = DS_READ2_B32_gfx9 renamable $vgpr99, 2, 3, 0, implicit $exec
80    $vgpr4_vgpr5 = DS_READ2_B32_gfx9 renamable $vgpr99, 4, 5, 0, implicit $exec
81    $vgpr6_vgpr7 = DS_READ2_B32_gfx9 renamable $vgpr99, 6, 7, 0, implicit $exec
82    $vgpr8_vgpr9 = DS_READ2_B32_gfx9 renamable $vgpr99, 8, 9, 0, implicit $exec
83    $vgpr10_vgpr11 = DS_READ2_B32_gfx9 renamable $vgpr99, 10, 11, 0, implicit $exec
84    $vgpr12_vgpr13 = DS_READ2_B32_gfx9 renamable $vgpr99, 12, 13, 0, implicit $exec
85    $vgpr14_vgpr15 = DS_READ2_B32_gfx9 renamable $vgpr99, 14, 15, 0, implicit $exec
86    $vgpr16_vgpr17 = DS_READ2_B32_gfx9 renamable $vgpr99, 16, 17, 0, implicit $exec
87    $vgpr18_vgpr19 = DS_READ2_B32_gfx9 renamable $vgpr99, 18, 19, 0, implicit $exec
88    $vgpr20_vgpr21 = DS_READ2_B32_gfx9 renamable $vgpr99, 20, 21, 0, implicit $exec
89    $vgpr22_vgpr23 = DS_READ2_B32_gfx9 renamable $vgpr99, 22, 23, 0, implicit $exec
90    $vgpr24_vgpr25 = DS_READ2_B32_gfx9 renamable $vgpr99, 24, 25, 0, implicit $exec
91    $vgpr26_vgpr27 = DS_READ2_B32_gfx9 renamable $vgpr99, 26, 27, 0, implicit $exec
92    $vgpr28_vgpr29 = DS_READ2_B32_gfx9 renamable $vgpr99, 28, 29, 0, implicit $exec
93    $vgpr30_vgpr31 = DS_READ2_B32_gfx9 renamable $vgpr99, 30, 31, 0, implicit $exec
94    $vgpr32_vgpr33 = DS_READ2_B32_gfx9 renamable $vgpr99, 32, 33, 0, implicit $exec
95    $vgpr34_vgpr35 = DS_READ2_B32_gfx9 renamable $vgpr99, 34, 35, 0, implicit $exec
96    $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
97    $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
98    $vgpr4 = V_MAC_F32_e32 0, $vgpr5, $vgpr4, implicit $mode, implicit $exec
99    $vgpr6 = V_MAC_F32_e32 0, $vgpr7, $vgpr6, implicit $mode, implicit $exec
100    S_ENDPGM 0
101...
102
103# Overflows vmcnt with gfx9 and gfx10.
104---
105name:            max-counter-vmcnt
106body:             |
107  bb.0:
108    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4
109
110    ; GFX9-LABEL: name: max-counter-vmcnt
111    ; GFX9: S_WAITCNT 0
112    ; GFX9: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
113    ; GFX9: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
114    ; GFX9: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
115    ; GFX9: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
116    ; GFX9: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
117    ; GFX9: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
118    ; GFX9: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
119    ; GFX9: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
120    ; GFX9: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
121    ; GFX9: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
122    ; GFX9: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
123    ; GFX9: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
124    ; GFX9: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
125    ; GFX9: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
126    ; GFX9: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
127    ; GFX9: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
128    ; GFX9: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
129    ; GFX9: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
130    ; GFX9: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
131    ; GFX9: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
132    ; GFX9: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
133    ; GFX9: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
134    ; GFX9: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
135    ; GFX9: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
136    ; GFX9: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
137    ; GFX9: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
138    ; GFX9: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
139    ; GFX9: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
140    ; GFX9: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
141    ; GFX9: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
142    ; GFX9: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
143    ; GFX9: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
144    ; GFX9: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
145    ; GFX9: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
146    ; GFX9: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
147    ; GFX9: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
148    ; GFX9: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
149    ; GFX9: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
150    ; GFX9: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
151    ; GFX9: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
152    ; GFX9: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
153    ; GFX9: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
154    ; GFX9: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
155    ; GFX9: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
156    ; GFX9: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
157    ; GFX9: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
158    ; GFX9: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
159    ; GFX9: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
160    ; GFX9: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
161    ; GFX9: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
162    ; GFX9: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
163    ; GFX9: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
164    ; GFX9: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
165    ; GFX9: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
166    ; GFX9: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
167    ; GFX9: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
168    ; GFX9: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
169    ; GFX9: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
170    ; GFX9: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
171    ; GFX9: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
172    ; GFX9: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
173    ; GFX9: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
174    ; GFX9: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
175    ; GFX9: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
176    ; GFX9: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
177    ; GFX9: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
178    ; GFX9: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
179    ; GFX9: S_WAITCNT 53118
180    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
181    ; GFX9: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
182    ; GFX9: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
183    ; GFX9: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
184    ; GFX9: S_ENDPGM 0
185    ; GFX10-LABEL: name: max-counter-vmcnt
186    ; GFX10: S_WAITCNT 0
187    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
188    ; GFX10: $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
189    ; GFX10: $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
190    ; GFX10: $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
191    ; GFX10: $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
192    ; GFX10: $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
193    ; GFX10: $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
194    ; GFX10: $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
195    ; GFX10: $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
196    ; GFX10: $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
197    ; GFX10: $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
198    ; GFX10: $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
199    ; GFX10: $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
200    ; GFX10: $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
201    ; GFX10: $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
202    ; GFX10: $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
203    ; GFX10: $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
204    ; GFX10: $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
205    ; GFX10: $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
206    ; GFX10: $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
207    ; GFX10: $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
208    ; GFX10: $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
209    ; GFX10: $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
210    ; GFX10: $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
211    ; GFX10: $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
212    ; GFX10: $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
213    ; GFX10: $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
214    ; GFX10: $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
215    ; GFX10: $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
216    ; GFX10: $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
217    ; GFX10: $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
218    ; GFX10: $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
219    ; GFX10: $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
220    ; GFX10: $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
221    ; GFX10: $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
222    ; GFX10: $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
223    ; GFX10: $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
224    ; GFX10: $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
225    ; GFX10: $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
226    ; GFX10: $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
227    ; GFX10: $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
228    ; GFX10: $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
229    ; GFX10: $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
230    ; GFX10: $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
231    ; GFX10: $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
232    ; GFX10: $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
233    ; GFX10: $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
234    ; GFX10: $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
235    ; GFX10: $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
236    ; GFX10: $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
237    ; GFX10: $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
238    ; GFX10: $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
239    ; GFX10: $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
240    ; GFX10: $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
241    ; GFX10: $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
242    ; GFX10: $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
243    ; GFX10: $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
244    ; GFX10: $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
245    ; GFX10: $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
246    ; GFX10: $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
247    ; GFX10: $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
248    ; GFX10: $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
249    ; GFX10: $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
250    ; GFX10: $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
251    ; GFX10: $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
252    ; GFX10: $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
253    ; GFX10: $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
254    ; GFX10: $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
255    ; GFX10: S_WAITCNT 65406
256    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
257    ; GFX10: $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
258    ; GFX10: $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
259    ; GFX10: $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
260    ; GFX10: S_ENDPGM 0
261    $vgpr0 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 0, 0, 0, 0, implicit $exec
262    $vgpr1 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 4, 0, 0, 0, implicit $exec
263    $vgpr2 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 8, 0, 0, 0, implicit $exec
264    $vgpr3 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 12, 0, 0, 0, implicit $exec
265    $vgpr4 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 16, 0, 0, 0, implicit $exec
266    $vgpr5 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 20, 0, 0, 0, implicit $exec
267    $vgpr6 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 24, 0, 0, 0, implicit $exec
268    $vgpr7 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 28, 0, 0, 0, implicit $exec
269    $vgpr8 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 32, 0, 0, 0, implicit $exec
270    $vgpr9 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 36, 0, 0, 0, implicit $exec
271    $vgpr10 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 40, 0, 0, 0, implicit $exec
272    $vgpr11 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 44, 0, 0, 0, implicit $exec
273    $vgpr12 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 48, 0, 0, 0, implicit $exec
274    $vgpr13 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 52, 0, 0, 0, implicit $exec
275    $vgpr14 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 56, 0, 0, 0, implicit $exec
276    $vgpr15 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 60, 0, 0, 0, implicit $exec
277    $vgpr16 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 64, 0, 0, 0, implicit $exec
278    $vgpr17 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 68, 0, 0, 0, implicit $exec
279    $vgpr18 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 72, 0, 0, 0, implicit $exec
280    $vgpr19 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 76, 0, 0, 0, implicit $exec
281    $vgpr20 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 80, 0, 0, 0, implicit $exec
282    $vgpr21 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 84, 0, 0, 0, implicit $exec
283    $vgpr22 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 88, 0, 0, 0, implicit $exec
284    $vgpr23 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 92, 0, 0, 0, implicit $exec
285    $vgpr24 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 96, 0, 0, 0, implicit $exec
286    $vgpr25 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 100, 0, 0, 0, implicit $exec
287    $vgpr26 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 104, 0, 0, 0, implicit $exec
288    $vgpr27 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 108, 0, 0, 0, implicit $exec
289    $vgpr28 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 112, 0, 0, 0, implicit $exec
290    $vgpr29 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 116, 0, 0, 0, implicit $exec
291    $vgpr30 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 120, 0, 0, 0, implicit $exec
292    $vgpr31 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 124, 0, 0, 0, implicit $exec
293    $vgpr32 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 128, 0, 0, 0, implicit $exec
294    $vgpr33 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 132, 0, 0, 0, implicit $exec
295    $vgpr34 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 136, 0, 0, 0, implicit $exec
296    $vgpr35 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 140, 0, 0, 0, implicit $exec
297    $vgpr36 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 144, 0, 0, 0, implicit $exec
298    $vgpr37 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 148, 0, 0, 0, implicit $exec
299    $vgpr38 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 152, 0, 0, 0, implicit $exec
300    $vgpr39 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 156, 0, 0, 0, implicit $exec
301    $vgpr40 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 160, 0, 0, 0, implicit $exec
302    $vgpr41 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 164, 0, 0, 0, implicit $exec
303    $vgpr42 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 168, 0, 0, 0, implicit $exec
304    $vgpr43 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 172, 0, 0, 0, implicit $exec
305    $vgpr44 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 176, 0, 0, 0, implicit $exec
306    $vgpr45 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 180, 0, 0, 0, implicit $exec
307    $vgpr46 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 184, 0, 0, 0, implicit $exec
308    $vgpr47 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 188, 0, 0, 0, implicit $exec
309    $vgpr48 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 192, 0, 0, 0, implicit $exec
310    $vgpr49 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 196, 0, 0, 0, implicit $exec
311    $vgpr50 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 200, 0, 0, 0, implicit $exec
312    $vgpr51 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 204, 0, 0, 0, implicit $exec
313    $vgpr52 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 208, 0, 0, 0, implicit $exec
314    $vgpr53 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 212, 0, 0, 0, implicit $exec
315    $vgpr54 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 216, 0, 0, 0, implicit $exec
316    $vgpr55 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 220, 0, 0, 0, implicit $exec
317    $vgpr56 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 224, 0, 0, 0, implicit $exec
318    $vgpr57 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 228, 0, 0, 0, implicit $exec
319    $vgpr58 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 232, 0, 0, 0, implicit $exec
320    $vgpr59 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 236, 0, 0, 0, implicit $exec
321    $vgpr60 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 240, 0, 0, 0, implicit $exec
322    $vgpr61 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 244, 0, 0, 0, implicit $exec
323    $vgpr62 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 248, 0, 0, 0, implicit $exec
324    $vgpr63 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 252, 0, 0, 0, implicit $exec
325    $vgpr64 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 256, 0, 0, 0, implicit $exec
326    $vgpr65 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 260, 0, 0, 0, implicit $exec
327    $vgpr66 = BUFFER_LOAD_DWORD_OFFSET $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, 264, 0, 0, 0, implicit $exec
328    $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
329    $vgpr1 = V_MAC_F32_e32 0, $vgpr2, $vgpr1, implicit $mode, implicit $exec
330    $vgpr2 = V_MAC_F32_e32 0, $vgpr3, $vgpr2, implicit $mode, implicit $exec
331    $vgpr3 = V_MAC_F32_e32 0, $vgpr4, $vgpr3, implicit $mode, implicit $exec
332    S_ENDPGM 0
333...
334
335# Should be no waitcnt if expcnt overflows.
336---
337name:            max-counter-expcnt
338body:             |
339  bb.0:
340    liveins: $sgpr0_sgpr1_sgpr2_sgpr3, $sgpr4, $vgpr0, $vgpr1
341
342    ; GFX9-LABEL: name: max-counter-expcnt
343    ; GFX9: S_WAITCNT 0
344    ; GFX9: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
345    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
346    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
347    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
348    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
349    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
350    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
351    ; GFX9: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
352    ; GFX9: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
353    ; GFX9: S_ENDPGM 0
354    ; GFX10-LABEL: name: max-counter-expcnt
355    ; GFX10: S_WAITCNT 0
356    ; GFX10: S_WAITCNT_VSCNT undef $sgpr_null, 0
357    ; GFX10: EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
358    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
359    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
360    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
361    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
362    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
363    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
364    ; GFX10: EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
365    ; GFX10: $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
366    ; GFX10: S_ENDPGM 0
367    EXP 0, $vgpr0, $vgpr0, $vgpr0, $vgpr0, -1, -1, 15, implicit $exec
368    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
369    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
370    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
371    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
372    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
373    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
374    EXP 0, $vgpr1, $vgpr1, $vgpr1, $vgpr1, -1, -1, 15, implicit $exec
375    $vgpr0 = V_MAC_F32_e32 0, $vgpr1, $vgpr0, implicit $mode, implicit $exec
376    S_ENDPGM 0
377...
378