xref: /llvm-project/llvm/test/CodeGen/AMDGPU/legalize-amdgcn.raw.ptr.buffer.load.format.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9) 
1; NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 2
2; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -stop-after=si-fix-sgpr-copies -verify-machineinstrs < %s | FileCheck -check-prefix=GFX908 %s
3
4; Natural mapping
5define float @raw_ptr_buffer_load_format_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset) {
6  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
7  ; GFX908: bb.0 (%ir-block.0):
8  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
9  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
10  ; GFX908-NEXT: {{  $}}
11  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
12  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
13  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
14  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
15  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
16  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
17  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
18  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
19  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
20  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
21  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
22  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
23  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
24  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
25  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
26  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
27  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
28  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
29  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
30  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
31  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
32  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
33  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
34  ; GFX908-NEXT: {{  $}}
35  ; GFX908-NEXT: bb.1:
36  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
37  ; GFX908-NEXT: {{  $}}
38  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
39  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
40  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
41  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
42  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
43  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
44  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
45  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
46  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
47  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
48  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
49  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
50  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
51  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
52  ; GFX908-NEXT: {{  $}}
53  ; GFX908-NEXT: bb.2:
54  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
55  ; GFX908-NEXT: {{  $}}
56  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_X_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_FORMAT_X_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s32) from %ir.rsrc, align 1, addrspace 8)
57  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
58  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
59  ; GFX908-NEXT: {{  $}}
60  ; GFX908-NEXT: bb.3:
61  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
62  ; GFX908-NEXT:   $vgpr0 = COPY [[BUFFER_LOAD_FORMAT_X_OFFEN]]
63  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0
64  %val = call float @llvm.amdgcn.raw.ptr.buffer.load.format.f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
65  ret float %val
66}
67
68define <2 x float> @raw_ptr_buffer_load_format_v2f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset) {
69  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_v2f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
70  ; GFX908: bb.0 (%ir-block.0):
71  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
72  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
73  ; GFX908-NEXT: {{  $}}
74  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
75  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
76  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
77  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
78  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
79  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
80  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
81  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
82  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
83  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
84  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
85  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
86  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
87  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
88  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
89  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
90  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
91  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
92  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
93  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
94  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
95  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
96  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
97  ; GFX908-NEXT: {{  $}}
98  ; GFX908-NEXT: bb.1:
99  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
100  ; GFX908-NEXT: {{  $}}
101  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
102  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
103  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
104  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
105  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
106  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
107  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
108  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
109  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
110  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
111  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
112  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
113  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
114  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
115  ; GFX908-NEXT: {{  $}}
116  ; GFX908-NEXT: bb.2:
117  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
118  ; GFX908-NEXT: {{  $}}
119  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_XY_OFFEN:%[0-9]+]]:vreg_64 = BUFFER_LOAD_FORMAT_XY_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s64) from %ir.rsrc, align 1, addrspace 8)
120  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
121  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
122  ; GFX908-NEXT: {{  $}}
123  ; GFX908-NEXT: bb.3:
124  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
125  ; GFX908-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XY_OFFEN]].sub0
126  ; GFX908-NEXT:   [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XY_OFFEN]].sub1
127  ; GFX908-NEXT:   $vgpr0 = COPY [[COPY10]]
128  ; GFX908-NEXT:   $vgpr1 = COPY [[COPY11]]
129  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0, implicit $vgpr1
130  %val = call <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v2f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
131  ret <2 x float> %val
132}
133
134define <3 x float> @raw_ptr_buffer_load_format_v3f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset) {
135  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_v3f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
136  ; GFX908: bb.0 (%ir-block.0):
137  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
138  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
139  ; GFX908-NEXT: {{  $}}
140  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
141  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
142  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
143  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
144  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
145  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
146  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
147  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
148  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
149  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
150  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
151  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
152  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
153  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
154  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
155  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
156  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
157  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
158  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
159  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
160  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
161  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
162  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
163  ; GFX908-NEXT: {{  $}}
164  ; GFX908-NEXT: bb.1:
165  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
166  ; GFX908-NEXT: {{  $}}
167  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
168  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
169  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
170  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
171  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
172  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
173  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
174  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
175  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
176  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
177  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
178  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
179  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
180  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
181  ; GFX908-NEXT: {{  $}}
182  ; GFX908-NEXT: bb.2:
183  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
184  ; GFX908-NEXT: {{  $}}
185  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_XYZ_OFFEN:%[0-9]+]]:vreg_96 = BUFFER_LOAD_FORMAT_XYZ_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s96) from %ir.rsrc, align 1, addrspace 8)
186  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
187  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
188  ; GFX908-NEXT: {{  $}}
189  ; GFX908-NEXT: bb.3:
190  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
191  ; GFX908-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZ_OFFEN]].sub0
192  ; GFX908-NEXT:   [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZ_OFFEN]].sub1
193  ; GFX908-NEXT:   [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZ_OFFEN]].sub2
194  ; GFX908-NEXT:   $vgpr0 = COPY [[COPY10]]
195  ; GFX908-NEXT:   $vgpr1 = COPY [[COPY11]]
196  ; GFX908-NEXT:   $vgpr2 = COPY [[COPY12]]
197  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2
198  %val = call <3 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v3f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
199  ret <3 x float> %val
200}
201
202define <4 x float> @raw_ptr_buffer_load_format_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset) {
203  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset
204  ; GFX908: bb.0 (%ir-block.0):
205  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
206  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
207  ; GFX908-NEXT: {{  $}}
208  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
209  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
210  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
211  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
212  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
213  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
214  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
215  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
216  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
217  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
218  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
219  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
220  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
221  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
222  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
223  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
224  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
225  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
226  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
227  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
228  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
229  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
230  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
231  ; GFX908-NEXT: {{  $}}
232  ; GFX908-NEXT: bb.1:
233  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
234  ; GFX908-NEXT: {{  $}}
235  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
236  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
237  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
238  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
239  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
240  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
241  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
242  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
243  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
244  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
245  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
246  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
247  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
248  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
249  ; GFX908-NEXT: {{  $}}
250  ; GFX908-NEXT: bb.2:
251  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
252  ; GFX908-NEXT: {{  $}}
253  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.rsrc, align 1, addrspace 8)
254  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
255  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
256  ; GFX908-NEXT: {{  $}}
257  ; GFX908-NEXT: bb.3:
258  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
259  ; GFX908-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub0
260  ; GFX908-NEXT:   [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub1
261  ; GFX908-NEXT:   [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub2
262  ; GFX908-NEXT:   [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub3
263  ; GFX908-NEXT:   $vgpr0 = COPY [[COPY10]]
264  ; GFX908-NEXT:   $vgpr1 = COPY [[COPY11]]
265  ; GFX908-NEXT:   $vgpr2 = COPY [[COPY12]]
266  ; GFX908-NEXT:   $vgpr3 = COPY [[COPY13]]
267  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2, implicit $vgpr3
268  %val = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
269  ret <4 x float> %val
270}
271
272; Waterfall for rsrc and soffset, copy for voffset
273define float @raw_ptr_buffer_load_format_f32__vgpr_rsrc__sgpr_voffset__vgpr_soffset(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset) {
274  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_f32__vgpr_rsrc__sgpr_voffset__vgpr_soffset
275  ; GFX908: bb.0 (%ir-block.0):
276  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
277  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
278  ; GFX908-NEXT: {{  $}}
279  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
280  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
281  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
282  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
283  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
284  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
285  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
286  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
287  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
288  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
289  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
290  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
291  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
292  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
293  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
294  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
295  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
296  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
297  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
298  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
299  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
300  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
301  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
302  ; GFX908-NEXT: {{  $}}
303  ; GFX908-NEXT: bb.1:
304  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
305  ; GFX908-NEXT: {{  $}}
306  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
307  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
308  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
309  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
310  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
311  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
312  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
313  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
314  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
315  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
316  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
317  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
318  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
319  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
320  ; GFX908-NEXT: {{  $}}
321  ; GFX908-NEXT: bb.2:
322  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
323  ; GFX908-NEXT: {{  $}}
324  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_X_OFFEN:%[0-9]+]]:vgpr_32 = BUFFER_LOAD_FORMAT_X_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 0, 0, 0, implicit $exec :: (dereferenceable load (s32) from %ir.rsrc, align 1, addrspace 8)
325  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
326  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
327  ; GFX908-NEXT: {{  $}}
328  ; GFX908-NEXT: bb.3:
329  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
330  ; GFX908-NEXT:   $vgpr0 = COPY [[BUFFER_LOAD_FORMAT_X_OFFEN]]
331  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0
332  %val = call float @llvm.amdgcn.raw.ptr.buffer.load.format.f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
333  ret float %val
334}
335
336define <4 x float> @raw_ptr_buffer_load_format_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add_4095(ptr addrspace(8) %rsrc, i32 %voffset.base, i32 %soffset) {
337  ; GFX908-LABEL: name: raw_ptr_buffer_load_format_v4f32__sgpr_rsrc__vgpr_voffset__sgpr_soffset_voffset_add_4095
338  ; GFX908: bb.0 (%ir-block.0):
339  ; GFX908-NEXT:   successors: %bb.1(0x80000000)
340  ; GFX908-NEXT:   liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5
341  ; GFX908-NEXT: {{  $}}
342  ; GFX908-NEXT:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr5
343  ; GFX908-NEXT:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr4
344  ; GFX908-NEXT:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr3
345  ; GFX908-NEXT:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr2
346  ; GFX908-NEXT:   [[COPY4:%[0-9]+]]:vgpr_32 = COPY $vgpr1
347  ; GFX908-NEXT:   [[COPY5:%[0-9]+]]:vgpr_32 = COPY $vgpr0
348  ; GFX908-NEXT:   [[DEF:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
349  ; GFX908-NEXT:   [[DEF1:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
350  ; GFX908-NEXT:   [[REG_SEQUENCE:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY3]], %subreg.sub0, [[COPY2]], %subreg.sub1
351  ; GFX908-NEXT:   [[COPY6:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub1
352  ; GFX908-NEXT:   [[COPY7:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE]].sub0
353  ; GFX908-NEXT:   [[DEF2:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
354  ; GFX908-NEXT:   [[DEF3:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
355  ; GFX908-NEXT:   [[REG_SEQUENCE1:%[0-9]+]]:vreg_64 = REG_SEQUENCE [[COPY5]], %subreg.sub0, [[COPY4]], %subreg.sub1
356  ; GFX908-NEXT:   [[COPY8:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub1
357  ; GFX908-NEXT:   [[COPY9:%[0-9]+]]:vgpr_32 = COPY [[REG_SEQUENCE1]].sub0
358  ; GFX908-NEXT:   [[DEF4:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
359  ; GFX908-NEXT:   [[DEF5:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
360  ; GFX908-NEXT:   [[DEF6:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
361  ; GFX908-NEXT:   [[DEF7:%[0-9]+]]:sgpr_32 = IMPLICIT_DEF
362  ; GFX908-NEXT:   [[REG_SEQUENCE2:%[0-9]+]]:vreg_128 = REG_SEQUENCE [[COPY9]], %subreg.sub0, [[COPY8]], %subreg.sub1, [[COPY7]], %subreg.sub2, [[COPY6]], %subreg.sub3
363  ; GFX908-NEXT:   [[DEF8:%[0-9]+]]:sreg_32 = IMPLICIT_DEF
364  ; GFX908-NEXT:   [[S_MOV_B64_:%[0-9]+]]:sreg_64_xexec = S_MOV_B64 $exec
365  ; GFX908-NEXT: {{  $}}
366  ; GFX908-NEXT: bb.1:
367  ; GFX908-NEXT:   successors: %bb.2(0x80000000)
368  ; GFX908-NEXT: {{  $}}
369  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub0, implicit $exec
370  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_1:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub1, implicit $exec
371  ; GFX908-NEXT:   [[REG_SEQUENCE3:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1
372  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE3]], [[REG_SEQUENCE2]].sub0_sub1, implicit $exec
373  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_2:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub2, implicit $exec
374  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_3:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[REG_SEQUENCE2]].sub3, implicit $exec
375  ; GFX908-NEXT:   [[REG_SEQUENCE4:%[0-9]+]]:sgpr_64 = REG_SEQUENCE [[V_READFIRSTLANE_B32_2]], %subreg.sub0, [[V_READFIRSTLANE_B32_3]], %subreg.sub1
376  ; GFX908-NEXT:   [[V_CMP_EQ_U64_e64_1:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U64_e64 [[REG_SEQUENCE4]], [[REG_SEQUENCE2]].sub2_sub3, implicit $exec
377  ; GFX908-NEXT:   [[S_AND_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[V_CMP_EQ_U64_e64_]], [[V_CMP_EQ_U64_e64_1]], implicit-def $scc
378  ; GFX908-NEXT:   [[REG_SEQUENCE5:%[0-9]+]]:sgpr_128 = REG_SEQUENCE [[V_READFIRSTLANE_B32_]], %subreg.sub0, [[V_READFIRSTLANE_B32_1]], %subreg.sub1, [[V_READFIRSTLANE_B32_2]], %subreg.sub2, [[V_READFIRSTLANE_B32_3]], %subreg.sub3
379  ; GFX908-NEXT:   [[V_READFIRSTLANE_B32_4:%[0-9]+]]:sgpr_32 = V_READFIRSTLANE_B32 [[COPY]], implicit $exec
380  ; GFX908-NEXT:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64_xexec = V_CMP_EQ_U32_e64 [[V_READFIRSTLANE_B32_4]], [[COPY]], implicit $exec
381  ; GFX908-NEXT:   [[S_AND_B64_1:%[0-9]+]]:sreg_64_xexec = S_AND_B64 [[S_AND_B64_]], [[V_CMP_EQ_U32_e64_]], implicit-def $scc
382  ; GFX908-NEXT:   [[S_AND_SAVEEXEC_B64_:%[0-9]+]]:sreg_64_xexec = S_AND_SAVEEXEC_B64 killed [[S_AND_B64_1]], implicit-def $exec, implicit-def $scc, implicit $exec
383  ; GFX908-NEXT: {{  $}}
384  ; GFX908-NEXT: bb.2:
385  ; GFX908-NEXT:   successors: %bb.1(0x40000000), %bb.3(0x40000000)
386  ; GFX908-NEXT: {{  $}}
387  ; GFX908-NEXT:   [[BUFFER_LOAD_FORMAT_XYZW_OFFEN:%[0-9]+]]:vreg_128 = BUFFER_LOAD_FORMAT_XYZW_OFFEN [[COPY1]], killed [[REG_SEQUENCE5]], killed [[V_READFIRSTLANE_B32_4]], 4095, 0, 0, implicit $exec :: (dereferenceable load (s128) from %ir.rsrc, align 1, addrspace 8)
388  ; GFX908-NEXT:   $exec = S_XOR_B64_term $exec, [[S_AND_SAVEEXEC_B64_]], implicit-def $scc
389  ; GFX908-NEXT:   SI_WATERFALL_LOOP %bb.1, implicit $exec
390  ; GFX908-NEXT: {{  $}}
391  ; GFX908-NEXT: bb.3:
392  ; GFX908-NEXT:   $exec = S_MOV_B64 [[S_MOV_B64_]]
393  ; GFX908-NEXT:   [[COPY10:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub0
394  ; GFX908-NEXT:   [[COPY11:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub1
395  ; GFX908-NEXT:   [[COPY12:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub2
396  ; GFX908-NEXT:   [[COPY13:%[0-9]+]]:vgpr_32 = COPY [[BUFFER_LOAD_FORMAT_XYZW_OFFEN]].sub3
397  ; GFX908-NEXT:   $vgpr0 = COPY [[COPY10]]
398  ; GFX908-NEXT:   $vgpr1 = COPY [[COPY11]]
399  ; GFX908-NEXT:   $vgpr2 = COPY [[COPY12]]
400  ; GFX908-NEXT:   $vgpr3 = COPY [[COPY13]]
401  ; GFX908-NEXT:   SI_RETURN implicit $vgpr0, implicit $vgpr1, implicit $vgpr2, implicit $vgpr3
402  %voffset = add i32 %voffset.base, 4095
403  %val = call <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8) %rsrc, i32 %voffset, i32 %soffset, i32 0)
404  ret <4 x float> %val
405}
406
407declare float @llvm.amdgcn.raw.ptr.buffer.load.format.f32(ptr addrspace(8), i32, i32, i32 immarg) #0
408declare <2 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v2f32(ptr addrspace(8), i32, i32, i32 immarg) #0
409declare <3 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v3f32(ptr addrspace(8), i32, i32, i32 immarg) #0
410declare <4 x float> @llvm.amdgcn.raw.ptr.buffer.load.format.v4f32(ptr addrspace(8), i32, i32, i32 immarg) #0
411
412attributes #0 = { nounwind readonly }
413