Lines Matching refs:MUBUF

2 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx906 < %s | FileCheck -check-prefixes=GCN,GFX9,MUBUF
20 ; MUBUF-LABEL: local_stack_offset_uses_sp:
21 ; MUBUF: ; %bb.0: ; %entry
22 ; MUBUF-NEXT: s_load_dwordx2 s[4:5], s[4:5], 0x0
23 ; MUBUF-NEXT: s_add_u32 flat_scratch_lo, s6, s9
24 ; MUBUF-NEXT: s_addc_u32 flat_scratch_hi, s7, 0
25 ; MUBUF-NEXT: s_add_u32 s0, s0, s9
26 ; MUBUF-NEXT: v_mov_b32_e32 v1, 0x3000
27 ; MUBUF-NEXT: s_addc_u32 s1, s1, 0
28 ; MUBUF-NEXT: v_add_u32_e32 v0, 64, v1
29 ; MUBUF-NEXT: v_mov_b32_e32 v2, 0
30 ; MUBUF-NEXT: v_mov_b32_e32 v3, 0x2000
31 ; MUBUF-NEXT: s_mov_b32 s6, 0
32 ; MUBUF-NEXT: buffer_store_dword v2, v3, s[0:3], 0 offen
33 ; MUBUF-NEXT: BB0_1: ; %loadstoreloop
34 ; MUBUF-NEXT: ; =>This Inner Loop Header: Depth=1
35 ; MUBUF-NEXT: v_add_u32_e32 v3, s6, v1
36 ; MUBUF-NEXT: s_add_i32 s6, s6, 1
37 ; MUBUF-NEXT: s_cmpk_lt_u32 s6, 0x2120
38 ; MUBUF-NEXT: buffer_store_byte v2, v3, s[0:3], 0 offen
39 ; MUBUF-NEXT: s_cbranch_scc1 BB0_1
40 ; MUBUF-NEXT: ; %bb.2: ; %split
41 ; MUBUF-NEXT: v_mov_b32_e32 v1, 0x3000
42 ; MUBUF-NEXT: v_add_u32_e32 v1, 0x20d0, v1
43 ; MUBUF-NEXT: buffer_load_dword v2, v1, s[0:3], 0 offen
44 ; MUBUF-NEXT: buffer_load_dword v1, v1, s[0:3], 0 offen offset:4
45 ; MUBUF-NEXT: buffer_load_dword v3, v0, s[0:3], 0 offen
46 ; MUBUF-NEXT: buffer_load_dword v4, v0, s[0:3], 0 offen offset:4
47 ; MUBUF-NEXT: s_waitcnt vmcnt(1)
48 ; MUBUF-NEXT: v_add_co_u32_e32 v0, vcc, v2, v3
49 ; MUBUF-NEXT: s_waitcnt vmcnt(0)
50 ; MUBUF-NEXT: v_addc_co_u32_e32 v1, vcc, v1, v4, vcc
51 ; MUBUF-NEXT: v_mov_b32_e32 v2, 0
52 ; MUBUF-NEXT: s_waitcnt lgkmcnt(0)
53 ; MUBUF-NEXT: global_store_dwordx2 v2, v[0:1], s[4:5]
54 ; MUBUF-NEXT: s_endpgm
101 ; MUBUF-LABEL: func_local_stack_offset_uses_sp:
102 ; MUBUF: ; %bb.0: ; %entry
103 ; MUBUF-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
104 ; MUBUF-NEXT: s_add_u32 s4, s32, 0x7ffc0
105 ; MUBUF-NEXT: s_mov_b32 s5, s33
106 ; MUBUF-NEXT: s_and_b32 s33, s4, 0xfff80000
107 ; MUBUF-NEXT: v_lshrrev_b32_e64 v3, 6, s33
108 ; MUBUF-NEXT: v_add_u32_e32 v3, 0x1000, v3
109 ; MUBUF-NEXT: v_mov_b32_e32 v4, 0
110 ; MUBUF-NEXT: v_add_u32_e32 v2, 64, v3
111 ; MUBUF-NEXT: s_mov_b32 s4, 0
112 ; MUBUF-NEXT: s_add_u32 s32, s32, 0x180000
113 ; MUBUF-NEXT: buffer_store_dword v4, off, s[0:3], s33
114 ; MUBUF-NEXT: BB1_1: ; %loadstoreloop
115 ; MUBUF-NEXT: ; =>This Inner Loop Header: Depth=1
116 ; MUBUF-NEXT: v_add_u32_e32 v5, s4, v3
117 ; MUBUF-NEXT: s_add_i32 s4, s4, 1
118 ; MUBUF-NEXT: s_cmpk_lt_u32 s4, 0x2120
119 ; MUBUF-NEXT: buffer_store_byte v4, v5, s[0:3], 0 offen
120 ; MUBUF-NEXT: s_cbranch_scc1 BB1_1
121 ; MUBUF-NEXT: ; %bb.2: ; %split
122 ; MUBUF-NEXT: v_lshrrev_b32_e64 v3, 6, s33
123 ; MUBUF-NEXT: v_add_u32_e32 v3, 0x1000, v3
124 ; MUBUF-NEXT: v_add_u32_e32 v3, 0x20d0, v3
125 ; MUBUF-NEXT: buffer_load_dword v4, v3, s[0:3], 0 offen
126 ; MUBUF-NEXT: buffer_load_dword v3, v3, s[0:3], 0 offen offset:4
127 ; MUBUF-NEXT: buffer_load_dword v5, v2, s[0:3], 0 offen
128 ; MUBUF-NEXT: buffer_load_dword v6, v2, s[0:3], 0 offen offset:4
129 ; MUBUF-NEXT: s_sub_u32 s32, s32, 0x180000
130 ; MUBUF-NEXT: s_mov_b32 s33, s5
131 ; MUBUF-NEXT: s_waitcnt vmcnt(1)
132 ; MUBUF-NEXT: v_add_co_u32_e32 v2, vcc, v4, v5
133 ; MUBUF-NEXT: s_waitcnt vmcnt(0)
134 ; MUBUF-NEXT: v_addc_co_u32_e32 v3, vcc, v3, v6, vcc
135 ; MUBUF-NEXT: global_store_dwordx2 v[0:1], v[2:3], off
136 ; MUBUF-NEXT: s_waitcnt vmcnt(0)
137 ; MUBUF-NEXT: s_setpc_b64 s[30:31]