1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2; RUN: llc -mtriple=amdgcn--amdpal -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -enable-var-scope --check-prefix=CI %s
3; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs -mattr=+load-store-opt,-unaligned-access-mode < %s | FileCheck -enable-var-scope -check-prefixes=GFX9,GFX9-ALIGNED %s
4; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx900 -verify-machineinstrs -mattr=+load-store-opt,+unaligned-access-mode < %s | FileCheck -enable-var-scope -check-prefixes=GFX9,GFX9-UNALIGNED %s
5
6; FIXME: We don't get cases where the address was an SGPR because we
7; get a copy to the address register for each one.
8
9@lds = addrspace(3) global [512 x float] undef, align 4
10@lds.f64 = addrspace(3) global [512 x double] undef, align 8
11
12define amdgpu_kernel void @simple_read2_f32(float addrspace(1)* %out) #0 {
13; CI-LABEL: simple_read2_f32:
14; CI:       ; %bb.0:
15; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
16; CI-NEXT:    s_mov_b32 m0, -1
17; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset1:8
18; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
19; CI-NEXT:    s_mov_b32 s3, 0xf000
20; CI-NEXT:    s_mov_b32 s2, 0
21; CI-NEXT:    s_waitcnt lgkmcnt(0)
22; CI-NEXT:    v_add_f32_e32 v2, v1, v2
23; CI-NEXT:    v_mov_b32_e32 v1, 0
24; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
25; CI-NEXT:    s_endpgm
26;
27; GFX9-LABEL: simple_read2_f32:
28; GFX9:       ; %bb.0:
29; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
30; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:8
31; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
32; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
33; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
34; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
35; GFX9-NEXT:    s_endpgm
36  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
37  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
38  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
39  %add.x = add nsw i32 %x.i, 8
40  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
41  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
42  %sum = fadd float %val0, %val1
43  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
44  store float %sum, float addrspace(1)* %out.gep, align 4
45  ret void
46}
47
48define amdgpu_kernel void @simple_read2_f32_max_offset(float addrspace(1)* %out) #0 {
49; CI-LABEL: simple_read2_f32_max_offset:
50; CI:       ; %bb.0:
51; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
52; CI-NEXT:    s_mov_b32 m0, -1
53; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset1:255
54; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
55; CI-NEXT:    s_mov_b32 s3, 0xf000
56; CI-NEXT:    s_mov_b32 s2, 0
57; CI-NEXT:    s_waitcnt lgkmcnt(0)
58; CI-NEXT:    v_add_f32_e32 v2, v1, v2
59; CI-NEXT:    v_mov_b32_e32 v1, 0
60; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
61; CI-NEXT:    s_endpgm
62;
63; GFX9-LABEL: simple_read2_f32_max_offset:
64; GFX9:       ; %bb.0:
65; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
66; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:255
67; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
68; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
69; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
70; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
71; GFX9-NEXT:    s_endpgm
72  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
73  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
74  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
75  %add.x = add nsw i32 %x.i, 255
76  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
77  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
78  %sum = fadd float %val0, %val1
79  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
80  store float %sum, float addrspace(1)* %out.gep, align 4
81  ret void
82}
83
84define amdgpu_kernel void @simple_read2_f32_too_far(float addrspace(1)* %out) #0 {
85; CI-LABEL: simple_read2_f32_too_far:
86; CI:       ; %bb.0:
87; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
88; CI-NEXT:    s_mov_b32 m0, -1
89; CI-NEXT:    ds_read_b32 v1, v0
90; CI-NEXT:    ds_read_b32 v2, v0 offset:1028
91; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
92; CI-NEXT:    s_mov_b32 s3, 0xf000
93; CI-NEXT:    s_mov_b32 s2, 0
94; CI-NEXT:    s_waitcnt lgkmcnt(0)
95; CI-NEXT:    v_add_f32_e32 v2, v1, v2
96; CI-NEXT:    v_mov_b32_e32 v1, 0
97; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
98; CI-NEXT:    s_endpgm
99;
100; GFX9-LABEL: simple_read2_f32_too_far:
101; GFX9:       ; %bb.0:
102; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
103; GFX9-NEXT:    ds_read_b32 v1, v0
104; GFX9-NEXT:    ds_read_b32 v2, v0 offset:1028
105; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
106; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
107; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
108; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
109; GFX9-NEXT:    s_endpgm
110  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
111  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
112  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
113  %add.x = add nsw i32 %x.i, 257
114  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
115  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
116  %sum = fadd float %val0, %val1
117  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
118  store float %sum, float addrspace(1)* %out.gep, align 4
119  ret void
120}
121
122define amdgpu_kernel void @simple_read2_f32_x2(float addrspace(1)* %out) #0 {
123; CI-LABEL: simple_read2_f32_x2:
124; CI:       ; %bb.0:
125; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
126; CI-NEXT:    s_mov_b32 m0, -1
127; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset1:8
128; CI-NEXT:    ds_read2_b32 v[3:4], v0 offset0:11 offset1:27
129; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
130; CI-NEXT:    s_mov_b32 s3, 0xf000
131; CI-NEXT:    s_mov_b32 s2, 0
132; CI-NEXT:    s_waitcnt lgkmcnt(0)
133; CI-NEXT:    v_add_f32_e32 v1, v1, v2
134; CI-NEXT:    v_add_f32_e32 v2, v3, v4
135; CI-NEXT:    v_add_f32_e32 v2, v1, v2
136; CI-NEXT:    v_mov_b32_e32 v1, 0
137; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
138; CI-NEXT:    s_endpgm
139;
140; GFX9-LABEL: simple_read2_f32_x2:
141; GFX9:       ; %bb.0:
142; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
143; GFX9-NEXT:    ds_read2_b32 v[0:1], v4 offset1:8
144; GFX9-NEXT:    ds_read2_b32 v[2:3], v4 offset0:11 offset1:27
145; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
146; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
147; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
148; GFX9-NEXT:    v_add_f32_e32 v1, v2, v3
149; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
150; GFX9-NEXT:    global_store_dword v4, v0, s[0:1]
151; GFX9-NEXT:    s_endpgm
152  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
153  %idx.0 = add nsw i32 %tid.x, 0
154  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
155  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
156
157  %idx.1 = add nsw i32 %tid.x, 8
158  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
159  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
160  %sum.0 = fadd float %val0, %val1
161
162  %idx.2 = add nsw i32 %tid.x, 11
163  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
164  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
165
166  %idx.3 = add nsw i32 %tid.x, 27
167  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
168  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
169  %sum.1 = fadd float %val2, %val3
170
171  %sum = fadd float %sum.0, %sum.1
172  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
173  store float %sum, float addrspace(1)* %out.gep, align 4
174  ret void
175}
176
177; Make sure there is an instruction between the two sets of reads.
178define amdgpu_kernel void @simple_read2_f32_x2_barrier(float addrspace(1)* %out) #0 {
179; CI-LABEL: simple_read2_f32_x2_barrier:
180; CI:       ; %bb.0:
181; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
182; CI-NEXT:    s_mov_b32 m0, -1
183; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset1:8
184; CI-NEXT:    s_waitcnt lgkmcnt(0)
185; CI-NEXT:    s_barrier
186; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
187; CI-NEXT:    s_mov_b32 s3, 0xf000
188; CI-NEXT:    s_mov_b32 s2, 0
189; CI-NEXT:    v_add_f32_e32 v3, v1, v2
190; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset0:11 offset1:27
191; CI-NEXT:    s_waitcnt lgkmcnt(0)
192; CI-NEXT:    v_add_f32_e32 v1, v1, v2
193; CI-NEXT:    v_add_f32_e32 v2, v3, v1
194; CI-NEXT:    v_mov_b32_e32 v1, 0
195; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
196; CI-NEXT:    s_endpgm
197;
198; GFX9-LABEL: simple_read2_f32_x2_barrier:
199; GFX9:       ; %bb.0:
200; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
201; GFX9-NEXT:    ds_read2_b32 v[0:1], v4 offset1:8
202; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
203; GFX9-NEXT:    s_barrier
204; GFX9-NEXT:    ds_read2_b32 v[2:3], v4 offset0:11 offset1:27
205; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
206; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
207; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
208; GFX9-NEXT:    v_add_f32_e32 v1, v2, v3
209; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
210; GFX9-NEXT:    global_store_dword v4, v0, s[0:1]
211; GFX9-NEXT:    s_endpgm
212  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
213  %idx.0 = add nsw i32 %tid.x, 0
214  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
215  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
216
217  %idx.1 = add nsw i32 %tid.x, 8
218  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
219  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
220  %sum.0 = fadd float %val0, %val1
221
222  call void @llvm.amdgcn.s.barrier() #2
223
224  %idx.2 = add nsw i32 %tid.x, 11
225  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
226  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
227
228  %idx.3 = add nsw i32 %tid.x, 27
229  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
230  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
231  %sum.1 = fadd float %val2, %val3
232
233  %sum = fadd float %sum.0, %sum.1
234  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
235  store float %sum, float addrspace(1)* %out.gep, align 4
236  ret void
237}
238
239; For some reason adding something to the base address for the first
240; element results in only folding the inner pair.
241define amdgpu_kernel void @simple_read2_f32_x2_nonzero_base(float addrspace(1)* %out) #0 {
242; CI-LABEL: simple_read2_f32_x2_nonzero_base:
243; CI:       ; %bb.0:
244; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
245; CI-NEXT:    s_mov_b32 m0, -1
246; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset0:2 offset1:8
247; CI-NEXT:    ds_read2_b32 v[3:4], v0 offset0:11 offset1:27
248; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
249; CI-NEXT:    s_mov_b32 s3, 0xf000
250; CI-NEXT:    s_mov_b32 s2, 0
251; CI-NEXT:    s_waitcnt lgkmcnt(0)
252; CI-NEXT:    v_add_f32_e32 v1, v1, v2
253; CI-NEXT:    v_add_f32_e32 v2, v3, v4
254; CI-NEXT:    v_add_f32_e32 v2, v1, v2
255; CI-NEXT:    v_mov_b32_e32 v1, 0
256; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64 offset:8
257; CI-NEXT:    s_endpgm
258;
259; GFX9-LABEL: simple_read2_f32_x2_nonzero_base:
260; GFX9:       ; %bb.0:
261; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 2, v0
262; GFX9-NEXT:    ds_read2_b32 v[0:1], v4 offset0:2 offset1:8
263; GFX9-NEXT:    ds_read2_b32 v[2:3], v4 offset0:11 offset1:27
264; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
265; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
266; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
267; GFX9-NEXT:    v_add_f32_e32 v1, v2, v3
268; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
269; GFX9-NEXT:    global_store_dword v4, v0, s[0:1] offset:8
270; GFX9-NEXT:    s_endpgm
271  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
272  %idx.0 = add nsw i32 %tid.x, 2
273  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
274  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
275
276  %idx.1 = add nsw i32 %tid.x, 8
277  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
278  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
279  %sum.0 = fadd float %val0, %val1
280
281  %idx.2 = add nsw i32 %tid.x, 11
282  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
283  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
284
285  %idx.3 = add nsw i32 %tid.x, 27
286  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
287  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
288  %sum.1 = fadd float %val2, %val3
289
290  %sum = fadd float %sum.0, %sum.1
291  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
292  store float %sum, float addrspace(1)* %out.gep, align 4
293  ret void
294}
295
296; Be careful of vectors of pointers. We don't know if the 2 pointers
297; in the vectors are really the same base, so this is not safe to
298; merge.
299; Base pointers come from different subregister of same super
300; register. We can't safely merge this.
301define amdgpu_kernel void @read2_ptr_is_subreg_arg_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
302; CI-LABEL: read2_ptr_is_subreg_arg_f32:
303; CI:       ; %bb.0:
304; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
305; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
306; CI-NEXT:    s_mov_b32 m0, -1
307; CI-NEXT:    s_mov_b32 s7, 0xf000
308; CI-NEXT:    s_mov_b32 s6, 0
309; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
310; CI-NEXT:    s_waitcnt lgkmcnt(0)
311; CI-NEXT:    v_mov_b32_e32 v1, s0
312; CI-NEXT:    v_mov_b32_e32 v2, s1
313; CI-NEXT:    ds_read_b32 v1, v1 offset:32
314; CI-NEXT:    ds_read_b32 v2, v2
315; CI-NEXT:    s_waitcnt lgkmcnt(0)
316; CI-NEXT:    v_add_f32_e32 v2, v1, v2
317; CI-NEXT:    v_mov_b32_e32 v1, 0
318; CI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
319; CI-NEXT:    s_endpgm
320;
321; GFX9-LABEL: read2_ptr_is_subreg_arg_f32:
322; GFX9:       ; %bb.0:
323; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
324; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x2c
325; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
326; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
327; GFX9-NEXT:    v_mov_b32_e32 v1, s4
328; GFX9-NEXT:    v_mov_b32_e32 v2, s5
329; GFX9-NEXT:    ds_read_b32 v1, v1 offset:32
330; GFX9-NEXT:    ds_read_b32 v2, v2
331; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
332; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
333; GFX9-NEXT:    global_store_dword v0, v1, s[2:3]
334; GFX9-NEXT:    s_endpgm
335  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
336  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
337  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
338  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
339  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
340  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
341  %val0 = load float, float addrspace(3)* %gep.0, align 4
342  %val1 = load float, float addrspace(3)* %gep.1, align 4
343  %add.x = add nsw i32 %x.i, 8
344  %sum = fadd float %val0, %val1
345  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
346  store float %sum, float addrspace(1)* %out.gep, align 4
347  ret void
348}
349
350; Apply a constant scalar offset after the pointer vector extract.  We
351; are rejecting merges that have the same, constant 0 offset, so make
352; sure we are really rejecting it because of the different
353; subregisters.
354define amdgpu_kernel void @read2_ptr_is_subreg_arg_offset_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
355; CI-LABEL: read2_ptr_is_subreg_arg_offset_f32:
356; CI:       ; %bb.0:
357; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
358; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xb
359; CI-NEXT:    s_mov_b32 m0, -1
360; CI-NEXT:    s_mov_b32 s7, 0xf000
361; CI-NEXT:    s_mov_b32 s6, 0
362; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
363; CI-NEXT:    s_waitcnt lgkmcnt(0)
364; CI-NEXT:    v_mov_b32_e32 v1, s0
365; CI-NEXT:    v_mov_b32_e32 v2, s1
366; CI-NEXT:    ds_read_b32 v1, v1 offset:32
367; CI-NEXT:    ds_read_b32 v2, v2 offset:32
368; CI-NEXT:    s_waitcnt lgkmcnt(0)
369; CI-NEXT:    v_add_f32_e32 v2, v1, v2
370; CI-NEXT:    v_mov_b32_e32 v1, 0
371; CI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
372; CI-NEXT:    s_endpgm
373;
374; GFX9-LABEL: read2_ptr_is_subreg_arg_offset_f32:
375; GFX9:       ; %bb.0:
376; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
377; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x2c
378; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
379; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
380; GFX9-NEXT:    v_mov_b32_e32 v1, s4
381; GFX9-NEXT:    v_mov_b32_e32 v2, s5
382; GFX9-NEXT:    ds_read_b32 v1, v1 offset:32
383; GFX9-NEXT:    ds_read_b32 v2, v2 offset:32
384; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
385; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
386; GFX9-NEXT:    global_store_dword v0, v1, s[2:3]
387; GFX9-NEXT:    s_endpgm
388  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
389  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
390  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
391  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
392  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
393  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
394
395  ; Apply an additional offset after the vector that will be more obviously folded.
396  %gep.1.offset = getelementptr float, float addrspace(3)* %gep.1, i32 8
397
398  %val0 = load float, float addrspace(3)* %gep.0, align 4
399  %val1 = load float, float addrspace(3)* %gep.1.offset, align 4
400  %add.x = add nsw i32 %x.i, 8
401  %sum = fadd float %val0, %val1
402  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
403  store float %sum, float addrspace(1)* %out.gep, align 4
404  ret void
405}
406
407define amdgpu_kernel void @read2_ptr_is_subreg_f32(float addrspace(1)* %out) #0 {
408; CI-LABEL: read2_ptr_is_subreg_f32:
409; CI:       ; %bb.0:
410; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
411; CI-NEXT:    s_mov_b32 m0, -1
412; CI-NEXT:    ds_read2_b32 v[1:2], v0 offset1:8
413; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
414; CI-NEXT:    s_mov_b32 s3, 0xf000
415; CI-NEXT:    s_mov_b32 s2, 0
416; CI-NEXT:    s_waitcnt lgkmcnt(0)
417; CI-NEXT:    v_add_f32_e32 v2, v1, v2
418; CI-NEXT:    v_mov_b32_e32 v1, 0
419; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
420; CI-NEXT:    s_endpgm
421;
422; GFX9-LABEL: read2_ptr_is_subreg_f32:
423; GFX9:       ; %bb.0:
424; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
425; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:8
426; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
427; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
428; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
429; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
430; GFX9-NEXT:    s_endpgm
431  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
432  %ptr.0 = insertelement <2 x [512 x float] addrspace(3)*> undef, [512 x float] addrspace(3)* @lds, i32 0
433  %ptr.1 = insertelement <2 x [512 x float] addrspace(3)*> %ptr.0, [512 x float] addrspace(3)* @lds, i32 1
434  %x.i.v.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
435  %x.i.v.1 = insertelement <2 x i32> %x.i.v.0, i32 %x.i, i32 1
436  %idx = add <2 x i32> %x.i.v.1, <i32 0, i32 8>
437  %gep = getelementptr inbounds [512 x float], <2 x [512 x float] addrspace(3)*> %ptr.1, <2 x i32> <i32 0, i32 0>, <2 x i32> %idx
438  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
439  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
440  %val0 = load float, float addrspace(3)* %gep.0, align 4
441  %val1 = load float, float addrspace(3)* %gep.1, align 4
442  %add.x = add nsw i32 %x.i, 8
443  %sum = fadd float %val0, %val1
444  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
445  store float %sum, float addrspace(1)* %out.gep, align 4
446  ret void
447}
448
449define amdgpu_kernel void @simple_read2_f32_volatile_0(float addrspace(1)* %out) #0 {
450; CI-LABEL: simple_read2_f32_volatile_0:
451; CI:       ; %bb.0:
452; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
453; CI-NEXT:    s_mov_b32 m0, -1
454; CI-NEXT:    ds_read_b32 v1, v0
455; CI-NEXT:    ds_read_b32 v2, v0 offset:32
456; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
457; CI-NEXT:    s_mov_b32 s3, 0xf000
458; CI-NEXT:    s_mov_b32 s2, 0
459; CI-NEXT:    s_waitcnt lgkmcnt(0)
460; CI-NEXT:    v_add_f32_e32 v2, v1, v2
461; CI-NEXT:    v_mov_b32_e32 v1, 0
462; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
463; CI-NEXT:    s_endpgm
464;
465; GFX9-LABEL: simple_read2_f32_volatile_0:
466; GFX9:       ; %bb.0:
467; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
468; GFX9-NEXT:    ds_read_b32 v1, v0
469; GFX9-NEXT:    ds_read_b32 v2, v0 offset:32
470; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
471; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
472; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
473; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
474; GFX9-NEXT:    s_endpgm
475  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
476  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
477  %val0 = load volatile float, float addrspace(3)* %arrayidx0, align 4
478  %add.x = add nsw i32 %x.i, 8
479  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
480  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
481  %sum = fadd float %val0, %val1
482  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
483  store float %sum, float addrspace(1)* %out.gep, align 4
484  ret void
485}
486
487define amdgpu_kernel void @simple_read2_f32_volatile_1(float addrspace(1)* %out) #0 {
488; CI-LABEL: simple_read2_f32_volatile_1:
489; CI:       ; %bb.0:
490; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
491; CI-NEXT:    s_mov_b32 m0, -1
492; CI-NEXT:    ds_read_b32 v1, v0
493; CI-NEXT:    ds_read_b32 v2, v0 offset:32
494; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
495; CI-NEXT:    s_mov_b32 s3, 0xf000
496; CI-NEXT:    s_mov_b32 s2, 0
497; CI-NEXT:    s_waitcnt lgkmcnt(0)
498; CI-NEXT:    v_add_f32_e32 v2, v1, v2
499; CI-NEXT:    v_mov_b32_e32 v1, 0
500; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
501; CI-NEXT:    s_endpgm
502;
503; GFX9-LABEL: simple_read2_f32_volatile_1:
504; GFX9:       ; %bb.0:
505; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
506; GFX9-NEXT:    ds_read_b32 v1, v0
507; GFX9-NEXT:    ds_read_b32 v2, v0 offset:32
508; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
509; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
510; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
511; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
512; GFX9-NEXT:    s_endpgm
513  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
514  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
515  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
516  %add.x = add nsw i32 %x.i, 8
517  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
518  %val1 = load volatile float, float addrspace(3)* %arrayidx1, align 4
519  %sum = fadd float %val0, %val1
520  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
521  store float %sum, float addrspace(1)* %out.gep, align 4
522  ret void
523}
524
525; Can't fold since not correctly aligned.
526define amdgpu_kernel void @unaligned_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
527; CI-LABEL: unaligned_read2_f32:
528; CI:       ; %bb.0:
529; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
530; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
531; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
532; CI-NEXT:    s_mov_b32 m0, -1
533; CI-NEXT:    s_mov_b32 s7, 0xf000
534; CI-NEXT:    s_mov_b32 s6, 0
535; CI-NEXT:    s_waitcnt lgkmcnt(0)
536; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
537; CI-NEXT:    ds_read_u8 v2, v1 offset:35
538; CI-NEXT:    ds_read_u8 v3, v1 offset:34
539; CI-NEXT:    ds_read_u8 v4, v1 offset:33
540; CI-NEXT:    ds_read_u8 v5, v1 offset:32
541; CI-NEXT:    ds_read_u8 v6, v1 offset:3
542; CI-NEXT:    ds_read_u8 v7, v1 offset:2
543; CI-NEXT:    ds_read_u8 v8, v1 offset:1
544; CI-NEXT:    ds_read_u8 v1, v1
545; CI-NEXT:    s_waitcnt lgkmcnt(7)
546; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
547; CI-NEXT:    s_waitcnt lgkmcnt(3)
548; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
549; CI-NEXT:    s_waitcnt lgkmcnt(2)
550; CI-NEXT:    v_or_b32_e32 v6, v6, v7
551; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
552; CI-NEXT:    v_or_b32_e32 v2, v2, v3
553; CI-NEXT:    s_waitcnt lgkmcnt(1)
554; CI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
555; CI-NEXT:    s_waitcnt lgkmcnt(0)
556; CI-NEXT:    v_or_b32_e32 v1, v8, v1
557; CI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
558; CI-NEXT:    v_or_b32_e32 v4, v4, v5
559; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
560; CI-NEXT:    v_or_b32_e32 v1, v6, v1
561; CI-NEXT:    v_or_b32_e32 v2, v2, v4
562; CI-NEXT:    v_add_f32_e32 v2, v1, v2
563; CI-NEXT:    v_mov_b32_e32 v1, 0
564; CI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
565; CI-NEXT:    s_endpgm
566;
567; GFX9-ALIGNED-LABEL: unaligned_read2_f32:
568; GFX9-ALIGNED:       ; %bb.0:
569; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
570; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
571; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
572; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
573; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
574; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1
575; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:1
576; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:2
577; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:3
578; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:32
579; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:33
580; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:34
581; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:35
582; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
583; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
584; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
585; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
586; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
587; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
588; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
589; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
590; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
591; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
592; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
593; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
594; GFX9-ALIGNED-NEXT:    s_endpgm
595;
596; GFX9-UNALIGNED-LABEL: unaligned_read2_f32:
597; GFX9-UNALIGNED:       ; %bb.0:
598; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
599; GFX9-UNALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
600; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
601; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
602; GFX9-UNALIGNED-NEXT:    v_add_u32_e32 v0, s4, v2
603; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:8
604; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
605; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
606; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[2:3]
607; GFX9-UNALIGNED-NEXT:    s_endpgm
608  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
609  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
610  %val0 = load float, float addrspace(3)* %arrayidx0, align 1
611  %add.x = add nsw i32 %x.i, 8
612  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
613  %val1 = load float, float addrspace(3)* %arrayidx1, align 1
614  %sum = fadd float %val0, %val1
615  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
616  store float %sum, float addrspace(1)* %out.gep, align 4
617  ret void
618}
619
620define amdgpu_kernel void @unaligned_offset_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
621; CI-LABEL: unaligned_offset_read2_f32:
622; CI:       ; %bb.0:
623; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
624; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
625; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
626; CI-NEXT:    s_mov_b32 m0, -1
627; CI-NEXT:    s_mov_b32 s7, 0xf000
628; CI-NEXT:    s_mov_b32 s6, 0
629; CI-NEXT:    s_waitcnt lgkmcnt(0)
630; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
631; CI-NEXT:    ds_read_u8 v2, v1 offset:12
632; CI-NEXT:    ds_read_u8 v3, v1 offset:11
633; CI-NEXT:    ds_read_u8 v4, v1 offset:10
634; CI-NEXT:    ds_read_u8 v5, v1 offset:9
635; CI-NEXT:    ds_read_u8 v6, v1 offset:8
636; CI-NEXT:    ds_read_u8 v7, v1 offset:7
637; CI-NEXT:    ds_read_u8 v8, v1 offset:6
638; CI-NEXT:    ds_read_u8 v1, v1 offset:5
639; CI-NEXT:    s_waitcnt lgkmcnt(7)
640; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
641; CI-NEXT:    s_waitcnt lgkmcnt(3)
642; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
643; CI-NEXT:    s_waitcnt lgkmcnt(2)
644; CI-NEXT:    v_or_b32_e32 v6, v6, v7
645; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
646; CI-NEXT:    v_or_b32_e32 v2, v2, v3
647; CI-NEXT:    s_waitcnt lgkmcnt(1)
648; CI-NEXT:    v_lshlrev_b32_e32 v8, 8, v8
649; CI-NEXT:    s_waitcnt lgkmcnt(0)
650; CI-NEXT:    v_or_b32_e32 v1, v8, v1
651; CI-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
652; CI-NEXT:    v_or_b32_e32 v4, v4, v5
653; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
654; CI-NEXT:    v_or_b32_e32 v1, v6, v1
655; CI-NEXT:    v_or_b32_e32 v2, v2, v4
656; CI-NEXT:    v_add_f32_e32 v2, v1, v2
657; CI-NEXT:    v_mov_b32_e32 v1, 0
658; CI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
659; CI-NEXT:    s_endpgm
660;
661; GFX9-ALIGNED-LABEL: unaligned_offset_read2_f32:
662; GFX9-ALIGNED:       ; %bb.0:
663; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
664; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
665; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
666; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
667; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
668; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1 offset:5
669; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:6
670; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:7
671; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:8
672; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:9
673; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:10
674; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:11
675; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:12
676; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
677; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
678; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
679; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
680; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
681; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
682; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
683; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
684; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
685; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
686; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
687; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
688; GFX9-ALIGNED-NEXT:    s_endpgm
689;
690; GFX9-UNALIGNED-LABEL: unaligned_offset_read2_f32:
691; GFX9-UNALIGNED:       ; %bb.0:
692; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
693; GFX9-UNALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
694; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
695; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
696; GFX9-UNALIGNED-NEXT:    v_add3_u32 v0, s4, v2, 5
697; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
698; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
699; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
700; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[2:3]
701; GFX9-UNALIGNED-NEXT:    s_endpgm
702  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
703  %base = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
704  %base.i8 = bitcast float addrspace(3)* %base to i8 addrspace(3)*
705  %addr0.i8 = getelementptr inbounds i8, i8 addrspace(3)* %base.i8, i32 5
706  %addr0 = bitcast i8 addrspace(3)* %addr0.i8 to float addrspace(3)*
707  %val0 = load float, float addrspace(3)* %addr0, align 1
708  %addr1.i8 = getelementptr inbounds i8, i8 addrspace(3)* %base.i8, i32 9
709  %addr1 = bitcast i8 addrspace(3)* %addr1.i8 to float addrspace(3)*
710  %val1 = load float, float addrspace(3)* %addr1, align 1
711  %sum = fadd float %val0, %val1
712  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
713  store float %sum, float addrspace(1)* %out.gep, align 4
714  ret void
715}
716
717define amdgpu_kernel void @misaligned_2_simple_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
718; CI-LABEL: misaligned_2_simple_read2_f32:
719; CI:       ; %bb.0:
720; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
721; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
722; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
723; CI-NEXT:    s_mov_b32 m0, -1
724; CI-NEXT:    s_mov_b32 s7, 0xf000
725; CI-NEXT:    s_mov_b32 s6, 0
726; CI-NEXT:    s_waitcnt lgkmcnt(0)
727; CI-NEXT:    v_add_i32_e32 v1, vcc, s0, v0
728; CI-NEXT:    ds_read_u16 v2, v1 offset:34
729; CI-NEXT:    ds_read_u16 v3, v1 offset:32
730; CI-NEXT:    ds_read_u16 v4, v1 offset:2
731; CI-NEXT:    ds_read_u16 v1, v1
732; CI-NEXT:    s_waitcnt lgkmcnt(3)
733; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
734; CI-NEXT:    s_waitcnt lgkmcnt(2)
735; CI-NEXT:    v_or_b32_e32 v2, v2, v3
736; CI-NEXT:    s_waitcnt lgkmcnt(1)
737; CI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
738; CI-NEXT:    s_waitcnt lgkmcnt(0)
739; CI-NEXT:    v_or_b32_e32 v1, v4, v1
740; CI-NEXT:    v_add_f32_e32 v2, v1, v2
741; CI-NEXT:    v_mov_b32_e32 v1, 0
742; CI-NEXT:    buffer_store_dword v2, v[0:1], s[4:7], 0 addr64
743; CI-NEXT:    s_endpgm
744;
745; GFX9-ALIGNED-LABEL: misaligned_2_simple_read2_f32:
746; GFX9-ALIGNED:       ; %bb.0:
747; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
748; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
749; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
750; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
751; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
752; GFX9-ALIGNED-NEXT:    ds_read_u16 v2, v1
753; GFX9-ALIGNED-NEXT:    ds_read_u16 v3, v1 offset:2
754; GFX9-ALIGNED-NEXT:    ds_read_u16 v4, v1 offset:32
755; GFX9-ALIGNED-NEXT:    ds_read_u16 v1, v1 offset:34
756; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
757; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
758; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
759; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
760; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
761; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
762; GFX9-ALIGNED-NEXT:    s_endpgm
763;
764; GFX9-UNALIGNED-LABEL: misaligned_2_simple_read2_f32:
765; GFX9-UNALIGNED:       ; %bb.0:
766; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
767; GFX9-UNALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x2c
768; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
769; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
770; GFX9-UNALIGNED-NEXT:    v_add_u32_e32 v0, s4, v2
771; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:8
772; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
773; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
774; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[2:3]
775; GFX9-UNALIGNED-NEXT:    s_endpgm
776  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
777  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
778  %val0 = load float, float addrspace(3)* %arrayidx0, align 2
779  %add.x = add nsw i32 %x.i, 8
780  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
781  %val1 = load float, float addrspace(3)* %arrayidx1, align 2
782  %sum = fadd float %val0, %val1
783  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
784  store float %sum, float addrspace(1)* %out.gep, align 4
785  ret void
786}
787
788define amdgpu_kernel void @simple_read2_f64(double addrspace(1)* %out) #0 {
789; CI-LABEL: simple_read2_f64:
790; CI:       ; %bb.0:
791; CI-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
792; CI-NEXT:    s_mov_b32 m0, -1
793; CI-NEXT:    ds_read2_b64 v[0:3], v4 offset1:8
794; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
795; CI-NEXT:    s_mov_b32 s3, 0xf000
796; CI-NEXT:    s_mov_b32 s2, 0
797; CI-NEXT:    v_mov_b32_e32 v5, 0
798; CI-NEXT:    s_waitcnt lgkmcnt(0)
799; CI-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
800; CI-NEXT:    buffer_store_dwordx2 v[0:1], v[4:5], s[0:3], 0 addr64
801; CI-NEXT:    s_endpgm
802;
803; GFX9-LABEL: simple_read2_f64:
804; GFX9:       ; %bb.0:
805; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
806; GFX9-NEXT:    ds_read2_b64 v[0:3], v4 offset1:8
807; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
808; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
809; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
810; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
811; GFX9-NEXT:    s_endpgm
812  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
813  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
814  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
815  %add.x = add nsw i32 %x.i, 8
816  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
817  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
818  %sum = fadd double %val0, %val1
819  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
820  store double %sum, double addrspace(1)* %out.gep, align 8
821  ret void
822}
823
824define amdgpu_kernel void @simple_read2_f64_max_offset(double addrspace(1)* %out) #0 {
825; CI-LABEL: simple_read2_f64_max_offset:
826; CI:       ; %bb.0:
827; CI-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
828; CI-NEXT:    s_mov_b32 m0, -1
829; CI-NEXT:    ds_read2_b64 v[0:3], v4 offset1:255
830; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
831; CI-NEXT:    s_mov_b32 s3, 0xf000
832; CI-NEXT:    s_mov_b32 s2, 0
833; CI-NEXT:    v_mov_b32_e32 v5, 0
834; CI-NEXT:    s_waitcnt lgkmcnt(0)
835; CI-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
836; CI-NEXT:    buffer_store_dwordx2 v[0:1], v[4:5], s[0:3], 0 addr64
837; CI-NEXT:    s_endpgm
838;
839; GFX9-LABEL: simple_read2_f64_max_offset:
840; GFX9:       ; %bb.0:
841; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
842; GFX9-NEXT:    ds_read2_b64 v[0:3], v4 offset1:255
843; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
844; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
845; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
846; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
847; GFX9-NEXT:    s_endpgm
848  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
849  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
850  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
851  %add.x = add nsw i32 %x.i, 255
852  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
853  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
854  %sum = fadd double %val0, %val1
855  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
856  store double %sum, double addrspace(1)* %out.gep, align 8
857  ret void
858}
859
860define amdgpu_kernel void @simple_read2_f64_too_far(double addrspace(1)* %out) #0 {
861; CI-LABEL: simple_read2_f64_too_far:
862; CI:       ; %bb.0:
863; CI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
864; CI-NEXT:    s_mov_b32 m0, -1
865; CI-NEXT:    ds_read_b64 v[1:2], v0
866; CI-NEXT:    ds_read_b64 v[3:4], v0 offset:2056
867; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
868; CI-NEXT:    s_mov_b32 s3, 0xf000
869; CI-NEXT:    s_mov_b32 s2, 0
870; CI-NEXT:    s_waitcnt lgkmcnt(0)
871; CI-NEXT:    v_add_f64 v[2:3], v[1:2], v[3:4]
872; CI-NEXT:    v_mov_b32_e32 v1, 0
873; CI-NEXT:    buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64
874; CI-NEXT:    s_endpgm
875;
876; GFX9-LABEL: simple_read2_f64_too_far:
877; GFX9:       ; %bb.0:
878; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
879; GFX9-NEXT:    ds_read_b64 v[0:1], v4
880; GFX9-NEXT:    ds_read_b64 v[2:3], v4 offset:2056
881; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
882; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
883; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
884; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
885; GFX9-NEXT:    s_endpgm
886  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
887  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
888  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
889  %add.x = add nsw i32 %x.i, 257
890  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
891  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
892  %sum = fadd double %val0, %val1
893  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
894  store double %sum, double addrspace(1)* %out.gep, align 8
895  ret void
896}
897
898; Alignment only 4
899define amdgpu_kernel void @misaligned_read2_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
900; CI-LABEL: misaligned_read2_f64:
901; CI:       ; %bb.0:
902; CI-NEXT:    s_load_dword s2, s[0:1], 0xb
903; CI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
904; CI-NEXT:    s_mov_b32 m0, -1
905; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
906; CI-NEXT:    s_mov_b32 s3, 0xf000
907; CI-NEXT:    s_waitcnt lgkmcnt(0)
908; CI-NEXT:    v_add_i32_e32 v3, vcc, s2, v0
909; CI-NEXT:    ds_read2_b32 v[1:2], v3 offset1:1
910; CI-NEXT:    ds_read2_b32 v[3:4], v3 offset0:14 offset1:15
911; CI-NEXT:    s_mov_b32 s2, 0
912; CI-NEXT:    s_waitcnt lgkmcnt(0)
913; CI-NEXT:    v_add_f64 v[2:3], v[1:2], v[3:4]
914; CI-NEXT:    v_mov_b32_e32 v1, 0
915; CI-NEXT:    buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64
916; CI-NEXT:    s_endpgm
917;
918; GFX9-LABEL: misaligned_read2_f64:
919; GFX9:       ; %bb.0:
920; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x2c
921; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
922; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
923; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
924; GFX9-NEXT:    v_add_u32_e32 v2, s2, v4
925; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
926; GFX9-NEXT:    ds_read2_b32 v[2:3], v2 offset0:14 offset1:15
927; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
928; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
929; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
930; GFX9-NEXT:    s_endpgm
931  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
932  %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
933  %val0 = load double, double addrspace(3)* %arrayidx0, align 4
934  %add.x = add nsw i32 %x.i, 7
935  %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
936  %val1 = load double, double addrspace(3)* %arrayidx1, align 4
937  %sum = fadd double %val0, %val1
938  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
939  store double %sum, double addrspace(1)* %out.gep, align 4
940  ret void
941}
942
943@foo = addrspace(3) global [4 x i32] undef, align 4
944
945define amdgpu_kernel void @load_constant_adjacent_offsets(i32 addrspace(1)* %out) {
946; CI-LABEL: load_constant_adjacent_offsets:
947; CI:       ; %bb.0:
948; CI-NEXT:    v_mov_b32_e32 v0, 0
949; CI-NEXT:    s_mov_b32 m0, -1
950; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
951; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
952; CI-NEXT:    s_mov_b32 s3, 0xf000
953; CI-NEXT:    s_mov_b32 s2, -1
954; CI-NEXT:    s_waitcnt lgkmcnt(0)
955; CI-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
956; CI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
957; CI-NEXT:    s_endpgm
958;
959; GFX9-LABEL: load_constant_adjacent_offsets:
960; GFX9:       ; %bb.0:
961; GFX9-NEXT:    v_mov_b32_e32 v2, 0
962; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
963; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
964; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
965; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
966; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
967; GFX9-NEXT:    s_endpgm
968  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
969  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
970  %sum = add i32 %val0, %val1
971  store i32 %sum, i32 addrspace(1)* %out, align 4
972  ret void
973}
974
975define amdgpu_kernel void @load_constant_disjoint_offsets(i32 addrspace(1)* %out) {
976; CI-LABEL: load_constant_disjoint_offsets:
977; CI:       ; %bb.0:
978; CI-NEXT:    v_mov_b32_e32 v0, 0
979; CI-NEXT:    s_mov_b32 m0, -1
980; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:2
981; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
982; CI-NEXT:    s_mov_b32 s3, 0xf000
983; CI-NEXT:    s_mov_b32 s2, -1
984; CI-NEXT:    s_waitcnt lgkmcnt(0)
985; CI-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
986; CI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
987; CI-NEXT:    s_endpgm
988;
989; GFX9-LABEL: load_constant_disjoint_offsets:
990; GFX9:       ; %bb.0:
991; GFX9-NEXT:    v_mov_b32_e32 v2, 0
992; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:2
993; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
994; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
995; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
996; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
997; GFX9-NEXT:    s_endpgm
998  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
999  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
1000  %sum = add i32 %val0, %val1
1001  store i32 %sum, i32 addrspace(1)* %out, align 4
1002  ret void
1003}
1004
1005@bar = addrspace(3) global [4 x i64] undef, align 4
1006
1007define amdgpu_kernel void @load_misaligned64_constant_offsets(i64 addrspace(1)* %out) {
1008; CI-LABEL: load_misaligned64_constant_offsets:
1009; CI:       ; %bb.0:
1010; CI-NEXT:    v_mov_b32_e32 v2, 0
1011; CI-NEXT:    s_mov_b32 m0, -1
1012; CI-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
1013; CI-NEXT:    ds_read2_b32 v[2:3], v2 offset0:2 offset1:3
1014; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
1015; CI-NEXT:    s_mov_b32 s3, 0xf000
1016; CI-NEXT:    s_mov_b32 s2, -1
1017; CI-NEXT:    s_waitcnt lgkmcnt(0)
1018; CI-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
1019; CI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
1020; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1021; CI-NEXT:    s_endpgm
1022;
1023; GFX9-ALIGNED-LABEL: load_misaligned64_constant_offsets:
1024; GFX9-ALIGNED:       ; %bb.0:
1025; GFX9-ALIGNED-NEXT:    v_mov_b32_e32 v4, 0
1026; GFX9-ALIGNED-NEXT:    ds_read2_b32 v[0:1], v4 offset1:1
1027; GFX9-ALIGNED-NEXT:    ds_read2_b32 v[2:3], v4 offset0:2 offset1:3
1028; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1029; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1030; GFX9-ALIGNED-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
1031; GFX9-ALIGNED-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
1032; GFX9-ALIGNED-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
1033; GFX9-ALIGNED-NEXT:    s_endpgm
1034;
1035; GFX9-UNALIGNED-LABEL: load_misaligned64_constant_offsets:
1036; GFX9-UNALIGNED:       ; %bb.0:
1037; GFX9-UNALIGNED-NEXT:    v_mov_b32_e32 v4, 0
1038; GFX9-UNALIGNED-NEXT:    ds_read_b128 v[0:3], v4
1039; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1040; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1041; GFX9-UNALIGNED-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
1042; GFX9-UNALIGNED-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
1043; GFX9-UNALIGNED-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
1044; GFX9-UNALIGNED-NEXT:    s_endpgm
1045  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 0), align 4
1046  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 1), align 4
1047  %sum = add i64 %val0, %val1
1048  store i64 %sum, i64 addrspace(1)* %out, align 8
1049  ret void
1050}
1051
1052@bar.large = addrspace(3) global [4096 x i64] undef, align 4
1053
1054define amdgpu_kernel void @load_misaligned64_constant_large_offsets(i64 addrspace(1)* %out) {
1055; CI-LABEL: load_misaligned64_constant_large_offsets:
1056; CI:       ; %bb.0:
1057; CI-NEXT:    v_mov_b32_e32 v0, 0x4000
1058; CI-NEXT:    v_mov_b32_e32 v2, 0x7ff8
1059; CI-NEXT:    s_mov_b32 m0, -1
1060; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1061; CI-NEXT:    ds_read2_b32 v[2:3], v2 offset1:1
1062; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
1063; CI-NEXT:    s_mov_b32 s3, 0xf000
1064; CI-NEXT:    s_mov_b32 s2, -1
1065; CI-NEXT:    s_waitcnt lgkmcnt(0)
1066; CI-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
1067; CI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
1068; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1069; CI-NEXT:    s_endpgm
1070;
1071; GFX9-LABEL: load_misaligned64_constant_large_offsets:
1072; GFX9:       ; %bb.0:
1073; GFX9-NEXT:    v_mov_b32_e32 v0, 0x4000
1074; GFX9-NEXT:    v_mov_b32_e32 v2, 0x7ff8
1075; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1076; GFX9-NEXT:    ds_read2_b32 v[2:3], v2 offset1:1
1077; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1078; GFX9-NEXT:    v_mov_b32_e32 v4, 0
1079; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1080; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
1081; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
1082; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
1083; GFX9-NEXT:    s_endpgm
1084  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 2048), align 4
1085  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 4095), align 4
1086  %sum = add i64 %val0, %val1
1087  store i64 %sum, i64 addrspace(1)* %out, align 8
1088  ret void
1089}
1090
1091@sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] undef, align 4
1092@sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] undef, align 4
1093
1094define amdgpu_kernel void @sgemm_inner_loop_read2_sequence(float addrspace(1)* %C, i32 %lda, i32 %ldb) #0 {
1095; CI-LABEL: sgemm_inner_loop_read2_sequence:
1096; CI:       ; %bb.0:
1097; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
1098; CI-NEXT:    s_lshl_b32 s0, s2, 2
1099; CI-NEXT:    s_add_i32 s1, s0, 0xc20
1100; CI-NEXT:    s_addk_i32 s0, 0xc60
1101; CI-NEXT:    v_mov_b32_e32 v0, s1
1102; CI-NEXT:    v_mov_b32_e32 v4, s0
1103; CI-NEXT:    s_mov_b32 m0, -1
1104; CI-NEXT:    ds_read2_b32 v[2:3], v0 offset1:1
1105; CI-NEXT:    ds_read2_b32 v[4:5], v4 offset1:1
1106; CI-NEXT:    v_lshlrev_b32_e32 v8, 2, v1
1107; CI-NEXT:    ds_read2_b32 v[0:1], v8 offset1:1
1108; CI-NEXT:    ds_read2_b32 v[6:7], v8 offset0:32 offset1:33
1109; CI-NEXT:    ds_read2_b32 v[8:9], v8 offset0:64 offset1:65
1110; CI-NEXT:    s_mov_b32 s7, 0xf000
1111; CI-NEXT:    s_waitcnt lgkmcnt(0)
1112; CI-NEXT:    v_add_f32_e32 v2, v2, v3
1113; CI-NEXT:    v_add_f32_e32 v2, v2, v4
1114; CI-NEXT:    v_add_f32_e32 v2, v2, v5
1115; CI-NEXT:    v_add_f32_e32 v0, v2, v0
1116; CI-NEXT:    v_add_f32_e32 v0, v0, v1
1117; CI-NEXT:    v_add_f32_e32 v0, v0, v6
1118; CI-NEXT:    v_add_f32_e32 v0, v0, v7
1119; CI-NEXT:    v_add_f32_e32 v0, v0, v8
1120; CI-NEXT:    s_mov_b32 s6, -1
1121; CI-NEXT:    v_add_f32_e32 v0, v0, v9
1122; CI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
1123; CI-NEXT:    s_endpgm
1124;
1125; GFX9-LABEL: sgemm_inner_loop_read2_sequence:
1126; GFX9:       ; %bb.0:
1127; GFX9-NEXT:    s_lshl_b32 s2, s2, 2
1128; GFX9-NEXT:    s_add_i32 s3, s2, 0xc20
1129; GFX9-NEXT:    s_addk_i32 s2, 0xc60
1130; GFX9-NEXT:    v_mov_b32_e32 v0, s3
1131; GFX9-NEXT:    v_mov_b32_e32 v4, s2
1132; GFX9-NEXT:    ds_read2_b32 v[2:3], v0 offset1:1
1133; GFX9-NEXT:    ds_read2_b32 v[4:5], v4 offset1:1
1134; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 2, v1
1135; GFX9-NEXT:    ds_read2_b32 v[0:1], v8 offset1:1
1136; GFX9-NEXT:    ds_read2_b32 v[6:7], v8 offset0:32 offset1:33
1137; GFX9-NEXT:    ds_read2_b32 v[8:9], v8 offset0:64 offset1:65
1138; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1139; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1140; GFX9-NEXT:    v_add_f32_e32 v2, v2, v3
1141; GFX9-NEXT:    v_add_f32_e32 v2, v2, v4
1142; GFX9-NEXT:    v_add_f32_e32 v2, v2, v5
1143; GFX9-NEXT:    v_add_f32_e32 v0, v2, v0
1144; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
1145; GFX9-NEXT:    v_add_f32_e32 v0, v0, v6
1146; GFX9-NEXT:    v_add_f32_e32 v0, v0, v7
1147; GFX9-NEXT:    v_add_f32_e32 v0, v0, v8
1148; GFX9-NEXT:    v_mov_b32_e32 v10, 0
1149; GFX9-NEXT:    v_add_f32_e32 v0, v0, v9
1150; GFX9-NEXT:    global_store_dword v10, v0, s[0:1]
1151; GFX9-NEXT:    s_endpgm
1152  %x.i = tail call i32 @llvm.amdgcn.workgroup.id.x() #1
1153  %y.i = tail call i32 @llvm.amdgcn.workitem.id.y() #1
1154  %arrayidx44 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %x.i
1155  %tmp16 = load float, float addrspace(3)* %arrayidx44, align 4
1156  %add47 = add nsw i32 %x.i, 1
1157  %arrayidx48 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add47
1158  %tmp17 = load float, float addrspace(3)* %arrayidx48, align 4
1159  %add51 = add nsw i32 %x.i, 16
1160  %arrayidx52 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add51
1161  %tmp18 = load float, float addrspace(3)* %arrayidx52, align 4
1162  %add55 = add nsw i32 %x.i, 17
1163  %arrayidx56 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add55
1164  %tmp19 = load float, float addrspace(3)* %arrayidx56, align 4
1165  %arrayidx60 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %y.i
1166  %tmp20 = load float, float addrspace(3)* %arrayidx60, align 4
1167  %add63 = add nsw i32 %y.i, 1
1168  %arrayidx64 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add63
1169  %tmp21 = load float, float addrspace(3)* %arrayidx64, align 4
1170  %add67 = add nsw i32 %y.i, 32
1171  %arrayidx68 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add67
1172  %tmp22 = load float, float addrspace(3)* %arrayidx68, align 4
1173  %add71 = add nsw i32 %y.i, 33
1174  %arrayidx72 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add71
1175  %tmp23 = load float, float addrspace(3)* %arrayidx72, align 4
1176  %add75 = add nsw i32 %y.i, 64
1177  %arrayidx76 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add75
1178  %tmp24 = load float, float addrspace(3)* %arrayidx76, align 4
1179  %add79 = add nsw i32 %y.i, 65
1180  %arrayidx80 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add79
1181  %tmp25 = load float, float addrspace(3)* %arrayidx80, align 4
1182  %sum.0 = fadd float %tmp16, %tmp17
1183  %sum.1 = fadd float %sum.0, %tmp18
1184  %sum.2 = fadd float %sum.1, %tmp19
1185  %sum.3 = fadd float %sum.2, %tmp20
1186  %sum.4 = fadd float %sum.3, %tmp21
1187  %sum.5 = fadd float %sum.4, %tmp22
1188  %sum.6 = fadd float %sum.5, %tmp23
1189  %sum.7 = fadd float %sum.6, %tmp24
1190  %sum.8 = fadd float %sum.7, %tmp25
1191  store float %sum.8, float addrspace(1)* %C, align 4
1192  ret void
1193}
1194
1195define amdgpu_kernel void @misaligned_read2_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(3)* %in) #0 {
1196; CI-LABEL: misaligned_read2_v2i32:
1197; CI:       ; %bb.0:
1198; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
1199; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
1200; CI-NEXT:    s_mov_b32 m0, -1
1201; CI-NEXT:    s_mov_b32 s7, 0xf000
1202; CI-NEXT:    s_mov_b32 s6, -1
1203; CI-NEXT:    s_waitcnt lgkmcnt(0)
1204; CI-NEXT:    v_mov_b32_e32 v0, s0
1205; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1206; CI-NEXT:    s_waitcnt lgkmcnt(0)
1207; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
1208; CI-NEXT:    s_endpgm
1209;
1210; GFX9-LABEL: misaligned_read2_v2i32:
1211; GFX9:       ; %bb.0:
1212; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
1213; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
1214; GFX9-NEXT:    v_mov_b32_e32 v2, 0
1215; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1216; GFX9-NEXT:    v_mov_b32_e32 v0, s4
1217; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1218; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1219; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
1220; GFX9-NEXT:    s_endpgm
1221  %load = load <2 x i32>, <2 x i32> addrspace(3)* %in, align 4
1222  store <2 x i32> %load, <2 x i32> addrspace(1)* %out, align 8
1223  ret void
1224}
1225
1226define amdgpu_kernel void @misaligned_read2_i64(i64 addrspace(1)* %out, i64 addrspace(3)* %in) #0 {
1227; CI-LABEL: misaligned_read2_i64:
1228; CI:       ; %bb.0:
1229; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
1230; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
1231; CI-NEXT:    s_mov_b32 m0, -1
1232; CI-NEXT:    s_mov_b32 s7, 0xf000
1233; CI-NEXT:    s_mov_b32 s6, -1
1234; CI-NEXT:    s_waitcnt lgkmcnt(0)
1235; CI-NEXT:    v_mov_b32_e32 v0, s0
1236; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1237; CI-NEXT:    s_waitcnt lgkmcnt(0)
1238; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[4:7], 0
1239; CI-NEXT:    s_endpgm
1240;
1241; GFX9-LABEL: misaligned_read2_i64:
1242; GFX9:       ; %bb.0:
1243; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
1244; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x2c
1245; GFX9-NEXT:    v_mov_b32_e32 v2, 0
1246; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1247; GFX9-NEXT:    v_mov_b32_e32 v0, s4
1248; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1249; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1250; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
1251; GFX9-NEXT:    s_endpgm
1252  %load = load i64, i64 addrspace(3)* %in, align 4
1253  store i64 %load, i64 addrspace(1)* %out, align 8
1254  ret void
1255}
1256
1257define amdgpu_kernel void @ds_read_diff_base_interleaving(
1258; CI-LABEL: ds_read_diff_base_interleaving:
1259; CI:       ; %bb.0: ; %bb
1260; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x9
1261; CI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0xb
1262; CI-NEXT:    v_lshlrev_b32_e32 v1, 4, v1
1263; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
1264; CI-NEXT:    s_mov_b32 m0, -1
1265; CI-NEXT:    s_mov_b32 s7, 0xf000
1266; CI-NEXT:    s_waitcnt lgkmcnt(0)
1267; CI-NEXT:    v_add_i32_e32 v2, vcc, s0, v1
1268; CI-NEXT:    v_add_i32_e32 v4, vcc, s1, v0
1269; CI-NEXT:    v_add_i32_e32 v3, vcc, s2, v1
1270; CI-NEXT:    v_add_i32_e32 v6, vcc, s3, v0
1271; CI-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
1272; CI-NEXT:    ds_read2_b32 v[2:3], v3 offset1:1
1273; CI-NEXT:    ds_read2_b32 v[4:5], v4 offset1:4
1274; CI-NEXT:    s_mov_b32 s6, -1
1275; CI-NEXT:    s_waitcnt lgkmcnt(0)
1276; CI-NEXT:    v_mul_f32_e32 v0, v0, v4
1277; CI-NEXT:    v_add_f32_e32 v4, 2.0, v0
1278; CI-NEXT:    v_mul_f32_e32 v5, v1, v5
1279; CI-NEXT:    ds_read2_b32 v[0:1], v6 offset1:4
1280; CI-NEXT:    s_waitcnt lgkmcnt(0)
1281; CI-NEXT:    v_mul_f32_e32 v0, v2, v0
1282; CI-NEXT:    v_sub_f32_e32 v0, v4, v0
1283; CI-NEXT:    v_sub_f32_e32 v0, v0, v5
1284; CI-NEXT:    v_mul_f32_e32 v1, v3, v1
1285; CI-NEXT:    v_sub_f32_e32 v0, v0, v1
1286; CI-NEXT:    buffer_store_dword v0, off, s[4:7], 0 offset:40
1287; CI-NEXT:    s_endpgm
1288;
1289; GFX9-LABEL: ds_read_diff_base_interleaving:
1290; GFX9:       ; %bb.0: ; %bb
1291; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
1292; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x2c
1293; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 4, v1
1294; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
1295; GFX9-NEXT:    v_mov_b32_e32 v8, 0
1296; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1297; GFX9-NEXT:    v_add_u32_e32 v2, s4, v1
1298; GFX9-NEXT:    v_add_u32_e32 v3, s5, v0
1299; GFX9-NEXT:    v_add_u32_e32 v4, s6, v1
1300; GFX9-NEXT:    v_add_u32_e32 v6, s7, v0
1301; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
1302; GFX9-NEXT:    ds_read2_b32 v[2:3], v3 offset1:4
1303; GFX9-NEXT:    ds_read2_b32 v[4:5], v4 offset1:1
1304; GFX9-NEXT:    ds_read2_b32 v[6:7], v6 offset1:4
1305; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
1306; GFX9-NEXT:    v_mul_f32_e32 v0, v0, v2
1307; GFX9-NEXT:    v_add_f32_e32 v0, 2.0, v0
1308; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1309; GFX9-NEXT:    v_mul_f32_e32 v2, v4, v6
1310; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v2
1311; GFX9-NEXT:    v_mul_f32_e32 v1, v1, v3
1312; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v1
1313; GFX9-NEXT:    v_mul_f32_e32 v1, v5, v7
1314; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v1
1315; GFX9-NEXT:    global_store_dword v8, v0, s[2:3] offset:40
1316; GFX9-NEXT:    s_endpgm
1317  float addrspace(1)* nocapture %arg,
1318  [4 x [4 x float]] addrspace(3)* %arg1,
1319  [4 x [4 x float]] addrspace(3)* %arg2,
1320  [4 x [4 x float]] addrspace(3)* %arg3,
1321  [4 x [4 x float]] addrspace(3)* %arg4) #1 {
1322bb:
1323  %tmp = getelementptr float, float addrspace(1)* %arg, i64 10
1324  %tmp5 = tail call i32 @llvm.amdgcn.workitem.id.x() #2
1325  %tmp6 = tail call i32 @llvm.amdgcn.workitem.id.y() #2
1326  %tmp7 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 0
1327  %tmp8 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 0, i32 %tmp5
1328  %tmp9 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 0
1329  %tmp10 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 0, i32 %tmp5
1330  %tmp11 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 1
1331  %tmp12 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 1, i32 %tmp5
1332  %tmp13 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 1
1333  %tmp14 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 1, i32 %tmp5
1334  %tmp15 = load float, float addrspace(3)* %tmp7
1335  %tmp16 = load float, float addrspace(3)* %tmp8
1336  %tmp17 = fmul float %tmp15, %tmp16
1337  %tmp18 = fadd float 2.000000e+00, %tmp17
1338  %tmp19 = load float, float addrspace(3)* %tmp9
1339  %tmp20 = load float, float addrspace(3)* %tmp10
1340  %tmp21 = fmul float %tmp19, %tmp20
1341  %tmp22 = fsub float %tmp18, %tmp21
1342  %tmp23 = load float, float addrspace(3)* %tmp11
1343  %tmp24 = load float, float addrspace(3)* %tmp12
1344  %tmp25 = fmul float %tmp23, %tmp24
1345  %tmp26 = fsub float %tmp22, %tmp25
1346  %tmp27 = load float, float addrspace(3)* %tmp13
1347  %tmp28 = load float, float addrspace(3)* %tmp14
1348  %tmp29 = fmul float %tmp27, %tmp28
1349  %tmp30 = fsub float %tmp26, %tmp29
1350  store float %tmp30, float addrspace(1)* %tmp
1351  ret void
1352}
1353
1354define amdgpu_kernel void @ds_read_call_read(i32 addrspace(1)* %out, i32 addrspace(3)* %arg) {
1355; CI-LABEL: ds_read_call_read:
1356; CI:       ; %bb.0:
1357; CI-NEXT:    s_getpc_b64 s[40:41]
1358; CI-NEXT:    s_mov_b32 s40, s0
1359; CI-NEXT:    s_load_dwordx4 s[40:43], s[40:41], 0x0
1360; CI-NEXT:    s_load_dwordx2 s[36:37], s[0:1], 0x9
1361; CI-NEXT:    s_load_dword s0, s[0:1], 0xb
1362; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
1363; CI-NEXT:    s_mov_b32 m0, -1
1364; CI-NEXT:    s_mov_b32 s32, 0
1365; CI-NEXT:    s_waitcnt lgkmcnt(0)
1366; CI-NEXT:    s_add_u32 s40, s40, s3
1367; CI-NEXT:    s_addc_u32 s41, s41, 0
1368; CI-NEXT:    v_add_i32_e32 v40, vcc, s0, v0
1369; CI-NEXT:    s_getpc_b64 s[0:1]
1370; CI-NEXT:    s_add_u32 s0, s0, void_func_void@gotpcrel32@lo+4
1371; CI-NEXT:    s_addc_u32 s1, s1, void_func_void@gotpcrel32@hi+12
1372; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
1373; CI-NEXT:    ds_read_b32 v41, v40
1374; CI-NEXT:    s_mov_b64 s[0:1], s[40:41]
1375; CI-NEXT:    s_mov_b64 s[2:3], s[42:43]
1376; CI-NEXT:    s_mov_b32 s39, 0xf000
1377; CI-NEXT:    s_mov_b32 s38, -1
1378; CI-NEXT:    s_waitcnt lgkmcnt(0)
1379; CI-NEXT:    s_swappc_b64 s[30:31], s[4:5]
1380; CI-NEXT:    ds_read_b32 v0, v40 offset:4
1381; CI-NEXT:    s_waitcnt lgkmcnt(0)
1382; CI-NEXT:    v_add_i32_e32 v0, vcc, v41, v0
1383; CI-NEXT:    buffer_store_dword v0, off, s[36:39], 0
1384; CI-NEXT:    s_endpgm
1385;
1386; GFX9-LABEL: ds_read_call_read:
1387; GFX9:       ; %bb.0:
1388; GFX9-NEXT:    s_getpc_b64 s[36:37]
1389; GFX9-NEXT:    s_mov_b32 s36, s0
1390; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[36:37], 0x0
1391; GFX9-NEXT:    s_nop 0
1392; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[0:1], 0x24
1393; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x2c
1394; GFX9-NEXT:    s_mov_b32 s32, 0
1395; GFX9-NEXT:    v_mov_b32_e32 v40, 0
1396; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1397; GFX9-NEXT:    s_add_u32 s36, s36, s3
1398; GFX9-NEXT:    s_addc_u32 s37, s37, 0
1399; GFX9-NEXT:    s_getpc_b64 s[0:1]
1400; GFX9-NEXT:    s_add_u32 s0, s0, void_func_void@gotpcrel32@lo+4
1401; GFX9-NEXT:    s_addc_u32 s1, s1, void_func_void@gotpcrel32@hi+12
1402; GFX9-NEXT:    v_lshl_add_u32 v41, v0, 2, s2
1403; GFX9-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
1404; GFX9-NEXT:    ds_read_b32 v42, v41
1405; GFX9-NEXT:    s_mov_b64 s[0:1], s[36:37]
1406; GFX9-NEXT:    s_mov_b64 s[2:3], s[38:39]
1407; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1408; GFX9-NEXT:    s_swappc_b64 s[30:31], s[4:5]
1409; GFX9-NEXT:    ds_read_b32 v0, v41 offset:4
1410; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1411; GFX9-NEXT:    v_add_u32_e32 v0, v42, v0
1412; GFX9-NEXT:    global_store_dword v40, v0, s[34:35]
1413; GFX9-NEXT:    s_endpgm
1414  %x = call i32 @llvm.amdgcn.workitem.id.x()
1415  %arrayidx0 = getelementptr i32, i32 addrspace(3)* %arg, i32 %x
1416  %arrayidx1 = getelementptr i32, i32 addrspace(3)* %arrayidx0, i32 1
1417  %v0 = load i32, i32 addrspace(3)* %arrayidx0, align 4
1418  call void @void_func_void()
1419  %v1 = load i32, i32 addrspace(3)* %arrayidx1, align 4
1420  %r = add i32 %v0, %v1
1421  store i32 %r, i32 addrspace(1)* %out, align 4
1422  ret void
1423}
1424
1425define amdgpu_ps <2 x float> @ds_read_interp_read(i32 inreg %prims, float addrspace(3)* %inptr) {
1426; CI-LABEL: ds_read_interp_read:
1427; CI:       ; %bb.0:
1428; CI-NEXT:    s_mov_b32 m0, -1
1429; CI-NEXT:    ds_read_b32 v2, v0
1430; CI-NEXT:    s_mov_b32 m0, s0
1431; CI-NEXT:    v_interp_mov_f32 v1, p10, attr0.x
1432; CI-NEXT:    s_mov_b32 m0, -1
1433; CI-NEXT:    ds_read_b32 v0, v0 offset:16
1434; CI-NEXT:    s_waitcnt lgkmcnt(0)
1435; CI-NEXT:    v_add_f32_e32 v1, v0, v1
1436; CI-NEXT:    v_mov_b32_e32 v0, v2
1437; CI-NEXT:    ; return to shader part epilog
1438;
1439; GFX9-LABEL: ds_read_interp_read:
1440; GFX9:       ; %bb.0:
1441; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:4
1442; GFX9-NEXT:    s_mov_b32 m0, s0
1443; GFX9-NEXT:    s_nop 0
1444; GFX9-NEXT:    v_interp_mov_f32_e32 v2, p10, attr0.x
1445; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1446; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
1447; GFX9-NEXT:    ; return to shader part epilog
1448  %v0 = load float, float addrspace(3)* %inptr, align 4
1449  %intrp = call float @llvm.amdgcn.interp.mov(i32 0, i32 0, i32 0, i32 %prims)
1450  %ptr1 = getelementptr float, float addrspace(3)* %inptr, i32 4
1451  %v1 = load float, float addrspace(3)* %ptr1, align 4
1452  %v1b = fadd float %v1, %intrp
1453  %r0 = insertelement <2 x float> undef, float %v0, i32 0
1454  %r1 = insertelement <2 x float> %r0, float %v1b, i32 1
1455  ret <2 x float> %r1
1456}
1457
1458@v2i32_align1 = internal addrspace(3) global [100 x <2 x i32>] undef, align 1
1459
1460define amdgpu_kernel void @read2_v2i32_align1_odd_offset(<2 x i32> addrspace(1)* %out) {
1461; CI-LABEL: read2_v2i32_align1_odd_offset:
1462; CI:       ; %bb.0: ; %entry
1463; CI-NEXT:    v_mov_b32_e32 v0, 0
1464; CI-NEXT:    s_mov_b32 m0, -1
1465; CI-NEXT:    ds_read_u8 v1, v0 offset:72
1466; CI-NEXT:    ds_read_u8 v2, v0 offset:71
1467; CI-NEXT:    ds_read_u8 v3, v0 offset:70
1468; CI-NEXT:    ds_read_u8 v4, v0 offset:69
1469; CI-NEXT:    ds_read_u8 v5, v0 offset:68
1470; CI-NEXT:    s_waitcnt lgkmcnt(4)
1471; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
1472; CI-NEXT:    s_waitcnt lgkmcnt(3)
1473; CI-NEXT:    v_or_b32_e32 v1, v1, v2
1474; CI-NEXT:    s_waitcnt lgkmcnt(2)
1475; CI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
1476; CI-NEXT:    s_waitcnt lgkmcnt(1)
1477; CI-NEXT:    v_or_b32_e32 v3, v3, v4
1478; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
1479; CI-NEXT:    v_or_b32_e32 v1, v1, v3
1480; CI-NEXT:    ds_read_u8 v2, v0 offset:67
1481; CI-NEXT:    ds_read_u8 v3, v0 offset:66
1482; CI-NEXT:    ds_read_u8 v0, v0 offset:65
1483; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
1484; CI-NEXT:    s_mov_b32 s3, 0xf000
1485; CI-NEXT:    s_mov_b32 s2, -1
1486; CI-NEXT:    s_waitcnt lgkmcnt(0)
1487; CI-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
1488; CI-NEXT:    v_or_b32_e32 v0, v3, v0
1489; CI-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
1490; CI-NEXT:    v_or_b32_e32 v2, v3, v2
1491; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
1492; CI-NEXT:    v_or_b32_e32 v0, v2, v0
1493; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1494; CI-NEXT:    s_endpgm
1495;
1496; GFX9-ALIGNED-LABEL: read2_v2i32_align1_odd_offset:
1497; GFX9-ALIGNED:       ; %bb.0: ; %entry
1498; GFX9-ALIGNED-NEXT:    v_mov_b32_e32 v2, 0
1499; GFX9-ALIGNED-NEXT:    ds_read_u8 v0, v2 offset:65
1500; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v2 offset:66
1501; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v2 offset:67
1502; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v2 offset:68
1503; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v2 offset:69
1504; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v2 offset:70
1505; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v2 offset:71
1506; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v2 offset:72
1507; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1508; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1509; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v3
1510; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
1511; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v6, v1
1512; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v6, 8, v8
1513; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v3, v0
1514; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v5
1515; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v6, v6, v7 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
1516; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v3, v3, v4 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
1517; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v6, v1
1518; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v3, v0
1519; GFX9-ALIGNED-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
1520; GFX9-ALIGNED-NEXT:    s_endpgm
1521;
1522; GFX9-UNALIGNED-LABEL: read2_v2i32_align1_odd_offset:
1523; GFX9-UNALIGNED:       ; %bb.0: ; %entry
1524; GFX9-UNALIGNED-NEXT:    v_mov_b32_e32 v0, 0x41
1525; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
1526; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1527; GFX9-UNALIGNED-NEXT:    v_mov_b32_e32 v2, 0
1528; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1529; GFX9-UNALIGNED-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
1530; GFX9-UNALIGNED-NEXT:    s_endpgm
1531entry:
1532  %load = load <2 x i32>, <2 x i32> addrspace(3)* bitcast (i8 addrspace(3)* getelementptr (i8, i8 addrspace(3)* bitcast ([100 x <2 x i32>] addrspace(3)* @v2i32_align1 to i8 addrspace(3)*), i32 65) to <2 x i32> addrspace(3)*), align 1
1533  store <2 x i32> %load, <2 x i32> addrspace(1)* %out
1534  ret void
1535}
1536
1537declare void @void_func_void() #3
1538
1539declare i32 @llvm.amdgcn.workgroup.id.x() #1
1540declare i32 @llvm.amdgcn.workgroup.id.y() #1
1541declare i32 @llvm.amdgcn.workitem.id.x() #1
1542declare i32 @llvm.amdgcn.workitem.id.y() #1
1543
1544declare float @llvm.amdgcn.interp.mov(i32, i32, i32, i32) nounwind readnone
1545
1546declare void @llvm.amdgcn.s.barrier() #2
1547
1548attributes #0 = { nounwind }
1549attributes #1 = { nounwind readnone speculatable }
1550attributes #2 = { convergent nounwind }
1551attributes #3 = { nounwind noinline }
1552