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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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:    ds_read2_b32 v[3:4], v0 offset0:11 offset1:27
187; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
188; CI-NEXT:    s_mov_b32 s3, 0xf000
189; CI-NEXT:    v_add_f32_e32 v1, v1, v2
190; CI-NEXT:    s_mov_b32 s2, 0
191; CI-NEXT:    s_waitcnt lgkmcnt(0)
192; CI-NEXT:    v_add_f32_e32 v2, v3, v4
193; CI-NEXT:    v_add_f32_e32 v2, v1, v2
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], 0x0
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], 0x0
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], 0x0
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[2:3], s[0:1], 0x2
305; CI-NEXT:    s_mov_b32 m0, -1
306; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
307; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
308; CI-NEXT:    s_waitcnt lgkmcnt(0)
309; CI-NEXT:    v_mov_b32_e32 v1, s2
310; CI-NEXT:    v_mov_b32_e32 v2, s3
311; CI-NEXT:    ds_read_b32 v1, v1 offset:32
312; CI-NEXT:    ds_read_b32 v2, v2
313; CI-NEXT:    s_mov_b32 s3, 0xf000
314; CI-NEXT:    s_mov_b32 s2, 0
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[0:3], 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], 0x8
324; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
325; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
326; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
327; GFX9-NEXT:    v_mov_b32_e32 v1, s2
328; GFX9-NEXT:    v_mov_b32_e32 v2, s3
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[0:1]
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[2:3], s[0:1], 0x2
358; CI-NEXT:    s_mov_b32 m0, -1
359; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
360; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
361; CI-NEXT:    s_waitcnt lgkmcnt(0)
362; CI-NEXT:    v_mov_b32_e32 v1, s2
363; CI-NEXT:    v_mov_b32_e32 v2, s3
364; CI-NEXT:    ds_read_b32 v1, v1 offset:32
365; CI-NEXT:    ds_read_b32 v2, v2 offset:32
366; CI-NEXT:    s_mov_b32 s3, 0xf000
367; CI-NEXT:    s_mov_b32 s2, 0
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[0:3], 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], 0x8
377; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
378; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
379; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
380; GFX9-NEXT:    v_mov_b32_e32 v1, s2
381; GFX9-NEXT:    v_mov_b32_e32 v2, s3
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[0:1]
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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], 0x0
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_dword s2, s[0:1], 0x2
530; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
531; CI-NEXT:    s_mov_b32 m0, -1
532; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
533; CI-NEXT:    s_mov_b32 s3, 0xf000
534; CI-NEXT:    s_waitcnt lgkmcnt(0)
535; CI-NEXT:    v_add_i32_e32 v1, vcc, s2, v0
536; CI-NEXT:    ds_read_u8 v2, v1 offset:34
537; CI-NEXT:    ds_read_u8 v3, v1 offset:32
538; CI-NEXT:    ds_read_u8 v4, v1 offset:3
539; CI-NEXT:    ds_read_u8 v5, v1 offset:2
540; CI-NEXT:    ds_read_u8 v6, v1 offset:1
541; CI-NEXT:    ds_read_u8 v7, v1
542; CI-NEXT:    ds_read_u8 v8, v1 offset:33
543; CI-NEXT:    ds_read_u8 v1, v1 offset:35
544; CI-NEXT:    s_waitcnt lgkmcnt(5)
545; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
546; CI-NEXT:    s_waitcnt lgkmcnt(3)
547; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
548; CI-NEXT:    v_or_b32_e32 v4, v4, v5
549; CI-NEXT:    s_waitcnt lgkmcnt(1)
550; CI-NEXT:    v_lshlrev_b32_e32 v5, 8, v8
551; CI-NEXT:    s_waitcnt lgkmcnt(0)
552; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
553; CI-NEXT:    v_or_b32_e32 v1, v1, v2
554; CI-NEXT:    v_or_b32_e32 v6, v6, v7
555; CI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
556; CI-NEXT:    v_or_b32_e32 v3, v5, v3
557; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
558; CI-NEXT:    v_or_b32_e32 v4, v4, v6
559; CI-NEXT:    v_or_b32_e32 v1, v1, v3
560; CI-NEXT:    v_add_f32_e32 v2, v4, v1
561; CI-NEXT:    s_mov_b32 s2, 0
562; CI-NEXT:    v_mov_b32_e32 v1, 0
563; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
564; CI-NEXT:    s_endpgm
565;
566; GFX9-ALIGNED-LABEL: unaligned_read2_f32:
567; GFX9-ALIGNED:       ; %bb.0:
568; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x8
569; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
570; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
571; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
572; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
573; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1
574; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:1
575; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:2
576; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:3
577; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:32
578; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:33
579; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:34
580; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:35
581; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
582; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
583; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
584; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
585; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
586; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
587; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
588; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
589; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
590; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
591; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
592; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
593; GFX9-ALIGNED-NEXT:    s_endpgm
594;
595; GFX9-UNALIGNED-LABEL: unaligned_read2_f32:
596; GFX9-UNALIGNED:       ; %bb.0:
597; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
598; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
599; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
600; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
601; GFX9-UNALIGNED-NEXT:    v_add_u32_e32 v0, s2, v2
602; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:8
603; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
604; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
605; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[0:1]
606; GFX9-UNALIGNED-NEXT:    s_endpgm
607  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
608  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
609  %val0 = load float, float addrspace(3)* %arrayidx0, align 1
610  %add.x = add nsw i32 %x.i, 8
611  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
612  %val1 = load float, float addrspace(3)* %arrayidx1, align 1
613  %sum = fadd float %val0, %val1
614  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
615  store float %sum, float addrspace(1)* %out.gep, align 4
616  ret void
617}
618
619define amdgpu_kernel void @unaligned_offset_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
620; CI-LABEL: unaligned_offset_read2_f32:
621; CI:       ; %bb.0:
622; CI-NEXT:    s_load_dword s2, s[0:1], 0x2
623; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
624; CI-NEXT:    s_mov_b32 m0, -1
625; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
626; CI-NEXT:    s_mov_b32 s3, 0xf000
627; CI-NEXT:    s_waitcnt lgkmcnt(0)
628; CI-NEXT:    v_add_i32_e32 v1, vcc, s2, v0
629; CI-NEXT:    ds_read_u8 v2, v1 offset:11
630; CI-NEXT:    ds_read_u8 v3, v1 offset:9
631; CI-NEXT:    ds_read_u8 v4, v1 offset:8
632; CI-NEXT:    ds_read_u8 v5, v1 offset:7
633; CI-NEXT:    ds_read_u8 v6, v1 offset:6
634; CI-NEXT:    ds_read_u8 v7, v1 offset:5
635; CI-NEXT:    ds_read_u8 v8, v1 offset:10
636; CI-NEXT:    ds_read_u8 v1, v1 offset:12
637; CI-NEXT:    s_waitcnt lgkmcnt(5)
638; CI-NEXT:    v_lshlrev_b32_e32 v4, 8, v4
639; CI-NEXT:    s_waitcnt lgkmcnt(3)
640; CI-NEXT:    v_lshlrev_b32_e32 v6, 8, v6
641; CI-NEXT:    v_or_b32_e32 v4, v4, v5
642; CI-NEXT:    s_waitcnt lgkmcnt(1)
643; CI-NEXT:    v_lshlrev_b32_e32 v5, 8, v8
644; CI-NEXT:    s_waitcnt lgkmcnt(0)
645; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
646; CI-NEXT:    v_or_b32_e32 v1, v1, v2
647; CI-NEXT:    v_or_b32_e32 v6, v6, v7
648; CI-NEXT:    v_lshlrev_b32_e32 v4, 16, v4
649; CI-NEXT:    v_or_b32_e32 v3, v5, v3
650; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
651; CI-NEXT:    v_or_b32_e32 v4, v4, v6
652; CI-NEXT:    v_or_b32_e32 v1, v1, v3
653; CI-NEXT:    v_add_f32_e32 v2, v4, v1
654; CI-NEXT:    s_mov_b32 s2, 0
655; CI-NEXT:    v_mov_b32_e32 v1, 0
656; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
657; CI-NEXT:    s_endpgm
658;
659; GFX9-ALIGNED-LABEL: unaligned_offset_read2_f32:
660; GFX9-ALIGNED:       ; %bb.0:
661; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x8
662; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
663; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
664; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
665; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
666; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1 offset:5
667; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:6
668; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:7
669; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:8
670; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:9
671; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:10
672; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:11
673; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:12
674; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
675; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
676; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
677; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
678; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
679; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
680; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
681; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
682; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
683; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
684; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
685; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
686; GFX9-ALIGNED-NEXT:    s_endpgm
687;
688; GFX9-UNALIGNED-LABEL: unaligned_offset_read2_f32:
689; GFX9-UNALIGNED:       ; %bb.0:
690; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
691; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
692; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
693; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
694; GFX9-UNALIGNED-NEXT:    v_add_u32_e32 v0, s2, v2
695; GFX9-UNALIGNED-NEXT:    ds_read_b64 v[0:1], v0 offset:5
696; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
697; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
698; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[0:1]
699; GFX9-UNALIGNED-NEXT:    s_endpgm
700  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
701  %base = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
702  %base.i8 = bitcast float addrspace(3)* %base to i8 addrspace(3)*
703  %addr0.i8 = getelementptr inbounds i8, i8 addrspace(3)* %base.i8, i32 5
704  %addr0 = bitcast i8 addrspace(3)* %addr0.i8 to float addrspace(3)*
705  %val0 = load float, float addrspace(3)* %addr0, align 1
706  %addr1.i8 = getelementptr inbounds i8, i8 addrspace(3)* %base.i8, i32 9
707  %addr1 = bitcast i8 addrspace(3)* %addr1.i8 to float addrspace(3)*
708  %val1 = load float, float addrspace(3)* %addr1, align 1
709  %sum = fadd float %val0, %val1
710  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
711  store float %sum, float addrspace(1)* %out.gep, align 4
712  ret void
713}
714
715define amdgpu_kernel void @misaligned_2_simple_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
716; CI-LABEL: misaligned_2_simple_read2_f32:
717; CI:       ; %bb.0:
718; CI-NEXT:    s_load_dword s2, s[0:1], 0x2
719; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
720; CI-NEXT:    s_mov_b32 m0, -1
721; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
722; CI-NEXT:    s_mov_b32 s3, 0xf000
723; CI-NEXT:    s_waitcnt lgkmcnt(0)
724; CI-NEXT:    v_add_i32_e32 v1, vcc, s2, v0
725; CI-NEXT:    ds_read_u16 v2, v1 offset:32
726; CI-NEXT:    ds_read_u16 v3, v1 offset:2
727; CI-NEXT:    ds_read_u16 v4, v1
728; CI-NEXT:    ds_read_u16 v1, v1 offset:34
729; CI-NEXT:    s_mov_b32 s2, 0
730; CI-NEXT:    s_waitcnt lgkmcnt(2)
731; CI-NEXT:    v_lshlrev_b32_e32 v3, 16, v3
732; CI-NEXT:    s_waitcnt lgkmcnt(1)
733; CI-NEXT:    v_or_b32_e32 v3, v3, v4
734; CI-NEXT:    s_waitcnt lgkmcnt(0)
735; CI-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
736; CI-NEXT:    v_or_b32_e32 v1, v1, v2
737; CI-NEXT:    v_add_f32_e32 v2, v3, v1
738; CI-NEXT:    v_mov_b32_e32 v1, 0
739; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
740; CI-NEXT:    s_endpgm
741;
742; GFX9-ALIGNED-LABEL: misaligned_2_simple_read2_f32:
743; GFX9-ALIGNED:       ; %bb.0:
744; GFX9-ALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
745; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
746; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
747; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
748; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s2, v0
749; GFX9-ALIGNED-NEXT:    ds_read_u16 v2, v1
750; GFX9-ALIGNED-NEXT:    ds_read_u16 v3, v1 offset:2
751; GFX9-ALIGNED-NEXT:    ds_read_u16 v4, v1 offset:32
752; GFX9-ALIGNED-NEXT:    ds_read_u16 v1, v1 offset:34
753; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
754; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
755; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
756; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
757; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
758; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[0:1]
759; GFX9-ALIGNED-NEXT:    s_endpgm
760;
761; GFX9-UNALIGNED-LABEL: misaligned_2_simple_read2_f32:
762; GFX9-UNALIGNED:       ; %bb.0:
763; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
764; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
765; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
766; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
767; GFX9-UNALIGNED-NEXT:    v_add_u32_e32 v0, s2, v2
768; GFX9-UNALIGNED-NEXT:    ds_read2_b32 v[0:1], v0 offset1:8
769; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
770; GFX9-UNALIGNED-NEXT:    v_add_f32_e32 v0, v0, v1
771; GFX9-UNALIGNED-NEXT:    global_store_dword v2, v0, s[0:1]
772; GFX9-UNALIGNED-NEXT:    s_endpgm
773  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
774  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
775  %val0 = load float, float addrspace(3)* %arrayidx0, align 2
776  %add.x = add nsw i32 %x.i, 8
777  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
778  %val1 = load float, float addrspace(3)* %arrayidx1, align 2
779  %sum = fadd float %val0, %val1
780  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
781  store float %sum, float addrspace(1)* %out.gep, align 4
782  ret void
783}
784
785define amdgpu_kernel void @simple_read2_f64(double addrspace(1)* %out) #0 {
786; CI-LABEL: simple_read2_f64:
787; CI:       ; %bb.0:
788; CI-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
789; CI-NEXT:    s_mov_b32 m0, -1
790; CI-NEXT:    ds_read2_b64 v[0:3], v4 offset1:8
791; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
792; CI-NEXT:    s_mov_b32 s3, 0xf000
793; CI-NEXT:    s_mov_b32 s2, 0
794; CI-NEXT:    v_mov_b32_e32 v5, 0
795; CI-NEXT:    s_waitcnt lgkmcnt(0)
796; CI-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
797; CI-NEXT:    buffer_store_dwordx2 v[0:1], v[4:5], s[0:3], 0 addr64
798; CI-NEXT:    s_endpgm
799;
800; GFX9-LABEL: simple_read2_f64:
801; GFX9:       ; %bb.0:
802; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
803; GFX9-NEXT:    ds_read2_b64 v[0:3], v4 offset1:8
804; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
805; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
806; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
807; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
808; GFX9-NEXT:    s_endpgm
809  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
810  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
811  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
812  %add.x = add nsw i32 %x.i, 8
813  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
814  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
815  %sum = fadd double %val0, %val1
816  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
817  store double %sum, double addrspace(1)* %out.gep, align 8
818  ret void
819}
820
821define amdgpu_kernel void @simple_read2_f64_max_offset(double addrspace(1)* %out) #0 {
822; CI-LABEL: simple_read2_f64_max_offset:
823; CI:       ; %bb.0:
824; CI-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
825; CI-NEXT:    s_mov_b32 m0, -1
826; CI-NEXT:    ds_read2_b64 v[0:3], v4 offset1:255
827; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
828; CI-NEXT:    s_mov_b32 s3, 0xf000
829; CI-NEXT:    s_mov_b32 s2, 0
830; CI-NEXT:    v_mov_b32_e32 v5, 0
831; CI-NEXT:    s_waitcnt lgkmcnt(0)
832; CI-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
833; CI-NEXT:    buffer_store_dwordx2 v[0:1], v[4:5], s[0:3], 0 addr64
834; CI-NEXT:    s_endpgm
835;
836; GFX9-LABEL: simple_read2_f64_max_offset:
837; GFX9:       ; %bb.0:
838; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
839; GFX9-NEXT:    ds_read2_b64 v[0:3], v4 offset1:255
840; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
841; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
842; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
843; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
844; GFX9-NEXT:    s_endpgm
845  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
846  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
847  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
848  %add.x = add nsw i32 %x.i, 255
849  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
850  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
851  %sum = fadd double %val0, %val1
852  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
853  store double %sum, double addrspace(1)* %out.gep, align 8
854  ret void
855}
856
857define amdgpu_kernel void @simple_read2_f64_too_far(double addrspace(1)* %out) #0 {
858; CI-LABEL: simple_read2_f64_too_far:
859; CI:       ; %bb.0:
860; CI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
861; CI-NEXT:    s_mov_b32 m0, -1
862; CI-NEXT:    ds_read_b64 v[1:2], v0
863; CI-NEXT:    ds_read_b64 v[3:4], v0 offset:2056
864; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
865; CI-NEXT:    s_mov_b32 s3, 0xf000
866; CI-NEXT:    s_mov_b32 s2, 0
867; CI-NEXT:    s_waitcnt lgkmcnt(0)
868; CI-NEXT:    v_add_f64 v[2:3], v[1:2], v[3:4]
869; CI-NEXT:    v_mov_b32_e32 v1, 0
870; CI-NEXT:    buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64
871; CI-NEXT:    s_endpgm
872;
873; GFX9-LABEL: simple_read2_f64_too_far:
874; GFX9:       ; %bb.0:
875; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
876; GFX9-NEXT:    ds_read_b64 v[0:1], v4
877; GFX9-NEXT:    ds_read_b64 v[2:3], v4 offset:2056
878; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
879; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
880; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
881; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
882; GFX9-NEXT:    s_endpgm
883  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
884  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
885  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
886  %add.x = add nsw i32 %x.i, 257
887  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
888  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
889  %sum = fadd double %val0, %val1
890  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
891  store double %sum, double addrspace(1)* %out.gep, align 8
892  ret void
893}
894
895; Alignment only 4
896define amdgpu_kernel void @misaligned_read2_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
897; CI-LABEL: misaligned_read2_f64:
898; CI:       ; %bb.0:
899; CI-NEXT:    s_load_dword s2, s[0:1], 0x2
900; CI-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
901; CI-NEXT:    s_mov_b32 m0, -1
902; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
903; CI-NEXT:    s_mov_b32 s3, 0xf000
904; CI-NEXT:    s_waitcnt lgkmcnt(0)
905; CI-NEXT:    v_add_i32_e32 v3, vcc, s2, v0
906; CI-NEXT:    ds_read2_b32 v[1:2], v3 offset1:1
907; CI-NEXT:    ds_read2_b32 v[3:4], v3 offset0:14 offset1:15
908; CI-NEXT:    s_mov_b32 s2, 0
909; CI-NEXT:    s_waitcnt lgkmcnt(0)
910; CI-NEXT:    v_add_f64 v[2:3], v[1:2], v[3:4]
911; CI-NEXT:    v_mov_b32_e32 v1, 0
912; CI-NEXT:    buffer_store_dwordx2 v[2:3], v[0:1], s[0:3], 0 addr64
913; CI-NEXT:    s_endpgm
914;
915; GFX9-LABEL: misaligned_read2_f64:
916; GFX9:       ; %bb.0:
917; GFX9-NEXT:    s_load_dword s2, s[0:1], 0x8
918; GFX9-NEXT:    v_lshlrev_b32_e32 v4, 3, v0
919; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
920; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
921; GFX9-NEXT:    v_add_u32_e32 v2, s2, v4
922; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
923; GFX9-NEXT:    ds_read2_b32 v[2:3], v2 offset0:14 offset1:15
924; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
925; GFX9-NEXT:    v_add_f64 v[0:1], v[0:1], v[2:3]
926; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
927; GFX9-NEXT:    s_endpgm
928  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
929  %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
930  %val0 = load double, double addrspace(3)* %arrayidx0, align 4
931  %add.x = add nsw i32 %x.i, 7
932  %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
933  %val1 = load double, double addrspace(3)* %arrayidx1, align 4
934  %sum = fadd double %val0, %val1
935  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
936  store double %sum, double addrspace(1)* %out.gep, align 4
937  ret void
938}
939
940@foo = addrspace(3) global [4 x i32] undef, align 4
941
942define amdgpu_kernel void @load_constant_adjacent_offsets(i32 addrspace(1)* %out) {
943; CI-LABEL: load_constant_adjacent_offsets:
944; CI:       ; %bb.0:
945; CI-NEXT:    v_mov_b32_e32 v0, 0
946; CI-NEXT:    s_mov_b32 m0, -1
947; CI-NEXT:    ds_read_b64 v[0:1], v0
948; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
949; CI-NEXT:    s_mov_b32 s3, 0xf000
950; CI-NEXT:    s_mov_b32 s2, -1
951; CI-NEXT:    s_waitcnt lgkmcnt(0)
952; CI-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
953; CI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
954; CI-NEXT:    s_endpgm
955;
956; GFX9-LABEL: load_constant_adjacent_offsets:
957; GFX9:       ; %bb.0:
958; GFX9-NEXT:    v_mov_b32_e32 v2, 0
959; GFX9-NEXT:    ds_read_b64 v[0:1], v2
960; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
961; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
962; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
963; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
964; GFX9-NEXT:    s_endpgm
965  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
966  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
967  %sum = add i32 %val0, %val1
968  store i32 %sum, i32 addrspace(1)* %out, align 4
969  ret void
970}
971
972define amdgpu_kernel void @load_constant_disjoint_offsets(i32 addrspace(1)* %out) {
973; CI-LABEL: load_constant_disjoint_offsets:
974; CI:       ; %bb.0:
975; CI-NEXT:    v_mov_b32_e32 v0, 0
976; CI-NEXT:    s_mov_b32 m0, -1
977; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:2
978; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
979; CI-NEXT:    s_mov_b32 s3, 0xf000
980; CI-NEXT:    s_mov_b32 s2, -1
981; CI-NEXT:    s_waitcnt lgkmcnt(0)
982; CI-NEXT:    v_add_i32_e32 v0, vcc, v1, v0
983; CI-NEXT:    buffer_store_dword v0, off, s[0:3], 0
984; CI-NEXT:    s_endpgm
985;
986; GFX9-LABEL: load_constant_disjoint_offsets:
987; GFX9:       ; %bb.0:
988; GFX9-NEXT:    v_mov_b32_e32 v2, 0
989; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:2
990; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
991; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
992; GFX9-NEXT:    v_add_u32_e32 v0, v0, v1
993; GFX9-NEXT:    global_store_dword v2, v0, s[0:1]
994; GFX9-NEXT:    s_endpgm
995  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
996  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
997  %sum = add i32 %val0, %val1
998  store i32 %sum, i32 addrspace(1)* %out, align 4
999  ret void
1000}
1001
1002@bar = addrspace(3) global [4 x i64] undef, align 4
1003
1004define amdgpu_kernel void @load_misaligned64_constant_offsets(i64 addrspace(1)* %out) {
1005; CI-LABEL: load_misaligned64_constant_offsets:
1006; CI:       ; %bb.0:
1007; CI-NEXT:    v_mov_b32_e32 v0, 0
1008; CI-NEXT:    s_mov_b32 m0, -1
1009; CI-NEXT:    ds_read_b128 v[0:3], v0
1010; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1011; CI-NEXT:    s_mov_b32 s3, 0xf000
1012; CI-NEXT:    s_mov_b32 s2, -1
1013; CI-NEXT:    s_waitcnt lgkmcnt(0)
1014; CI-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
1015; CI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
1016; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1017; CI-NEXT:    s_endpgm
1018;
1019; GFX9-LABEL: load_misaligned64_constant_offsets:
1020; GFX9:       ; %bb.0:
1021; GFX9-NEXT:    v_mov_b32_e32 v4, 0
1022; GFX9-NEXT:    ds_read_b128 v[0:3], v4
1023; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1024; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1025; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
1026; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
1027; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
1028; GFX9-NEXT:    s_endpgm
1029  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 0), align 4
1030  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 1), align 4
1031  %sum = add i64 %val0, %val1
1032  store i64 %sum, i64 addrspace(1)* %out, align 8
1033  ret void
1034}
1035
1036@bar.large = addrspace(3) global [4096 x i64] undef, align 4
1037
1038define amdgpu_kernel void @load_misaligned64_constant_large_offsets(i64 addrspace(1)* %out) {
1039; CI-LABEL: load_misaligned64_constant_large_offsets:
1040; CI:       ; %bb.0:
1041; CI-NEXT:    v_mov_b32_e32 v2, 0
1042; CI-NEXT:    s_mov_b32 m0, -1
1043; CI-NEXT:    ds_read_b64 v[0:1], v2 offset:16384
1044; CI-NEXT:    ds_read_b64 v[2:3], v2 offset:32760
1045; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1046; CI-NEXT:    s_mov_b32 s3, 0xf000
1047; CI-NEXT:    s_mov_b32 s2, -1
1048; CI-NEXT:    s_waitcnt lgkmcnt(0)
1049; CI-NEXT:    v_add_i32_e32 v0, vcc, v0, v2
1050; CI-NEXT:    v_addc_u32_e32 v1, vcc, v1, v3, vcc
1051; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1052; CI-NEXT:    s_endpgm
1053;
1054; GFX9-LABEL: load_misaligned64_constant_large_offsets:
1055; GFX9:       ; %bb.0:
1056; GFX9-NEXT:    v_mov_b32_e32 v4, 0
1057; GFX9-NEXT:    ds_read_b64 v[0:1], v4 offset:16384
1058; GFX9-NEXT:    ds_read_b64 v[2:3], v4 offset:32760
1059; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1060; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1061; GFX9-NEXT:    v_add_co_u32_e32 v0, vcc, v0, v2
1062; GFX9-NEXT:    v_addc_co_u32_e32 v1, vcc, v1, v3, vcc
1063; GFX9-NEXT:    global_store_dwordx2 v4, v[0:1], s[0:1]
1064; GFX9-NEXT:    s_endpgm
1065  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 2048), align 4
1066  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 4095), align 4
1067  %sum = add i64 %val0, %val1
1068  store i64 %sum, i64 addrspace(1)* %out, align 8
1069  ret void
1070}
1071
1072@sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] undef, align 4
1073@sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] undef, align 4
1074
1075define amdgpu_kernel void @sgemm_inner_loop_read2_sequence(float addrspace(1)* %C, i32 %lda, i32 %ldb) #0 {
1076; CI-LABEL: sgemm_inner_loop_read2_sequence:
1077; CI:       ; %bb.0:
1078; CI-NEXT:    s_load_dwordx2 s[4:5], s[0:1], 0x0
1079; CI-NEXT:    s_lshl_b32 s0, s2, 2
1080; CI-NEXT:    s_add_i32 s1, s0, 0xc20
1081; CI-NEXT:    s_addk_i32 s0, 0xc60
1082; CI-NEXT:    v_mov_b32_e32 v0, s1
1083; CI-NEXT:    v_mov_b32_e32 v2, s0
1084; CI-NEXT:    v_lshlrev_b32_e32 v8, 2, v1
1085; CI-NEXT:    s_mov_b32 m0, -1
1086; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1087; CI-NEXT:    ds_read2_b32 v[2:3], v2 offset1:1
1088; CI-NEXT:    ds_read2_b32 v[4:5], v8 offset1:1
1089; CI-NEXT:    ds_read2_b32 v[6:7], v8 offset0:32 offset1:33
1090; CI-NEXT:    ds_read2_b32 v[8:9], v8 offset0:64 offset1:65
1091; CI-NEXT:    s_waitcnt lgkmcnt(0)
1092; CI-NEXT:    v_add_f32_e32 v0, v0, v1
1093; CI-NEXT:    v_add_f32_e32 v0, v0, v2
1094; CI-NEXT:    v_add_f32_e32 v0, v0, v3
1095; CI-NEXT:    v_add_f32_e32 v0, v0, v4
1096; CI-NEXT:    v_add_f32_e32 v0, v0, v5
1097; CI-NEXT:    v_add_f32_e32 v0, v0, v6
1098; CI-NEXT:    v_add_f32_e32 v0, v0, v7
1099; CI-NEXT:    v_add_f32_e32 v0, v0, v8
1100; CI-NEXT:    s_mov_b32 s7, 0xf000
1101; CI-NEXT:    s_mov_b32 s6, -1
1102; CI-NEXT:    v_add_f32_e32 v0, v0, v9
1103; CI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
1104; CI-NEXT:    s_endpgm
1105;
1106; GFX9-LABEL: sgemm_inner_loop_read2_sequence:
1107; GFX9:       ; %bb.0:
1108; GFX9-NEXT:    s_lshl_b32 s2, s2, 2
1109; GFX9-NEXT:    s_add_i32 s3, s2, 0xc20
1110; GFX9-NEXT:    s_addk_i32 s2, 0xc60
1111; GFX9-NEXT:    v_mov_b32_e32 v0, s3
1112; GFX9-NEXT:    v_mov_b32_e32 v2, s2
1113; GFX9-NEXT:    v_lshlrev_b32_e32 v8, 2, v1
1114; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1115; GFX9-NEXT:    ds_read2_b32 v[2:3], v2 offset1:1
1116; GFX9-NEXT:    ds_read2_b32 v[4:5], v8 offset1:1
1117; GFX9-NEXT:    ds_read2_b32 v[6:7], v8 offset0:32 offset1:33
1118; GFX9-NEXT:    ds_read2_b32 v[8:9], v8 offset0:64 offset1:65
1119; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
1120; GFX9-NEXT:    v_add_f32_e32 v0, v0, v1
1121; GFX9-NEXT:    s_waitcnt lgkmcnt(3)
1122; GFX9-NEXT:    v_add_f32_e32 v0, v0, v2
1123; GFX9-NEXT:    v_add_f32_e32 v0, v0, v3
1124; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
1125; GFX9-NEXT:    v_add_f32_e32 v0, v0, v4
1126; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1127; GFX9-NEXT:    v_add_f32_e32 v0, v0, v5
1128; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1129; GFX9-NEXT:    v_add_f32_e32 v0, v0, v6
1130; GFX9-NEXT:    v_add_f32_e32 v0, v0, v7
1131; GFX9-NEXT:    v_add_f32_e32 v0, v0, v8
1132; GFX9-NEXT:    v_mov_b32_e32 v10, 0
1133; GFX9-NEXT:    v_add_f32_e32 v0, v0, v9
1134; GFX9-NEXT:    global_store_dword v10, v0, s[0:1]
1135; GFX9-NEXT:    s_endpgm
1136  %x.i = tail call i32 @llvm.amdgcn.workgroup.id.x() #1
1137  %y.i = tail call i32 @llvm.amdgcn.workitem.id.y() #1
1138  %arrayidx44 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %x.i
1139  %tmp16 = load float, float addrspace(3)* %arrayidx44, align 4
1140  %add47 = add nsw i32 %x.i, 1
1141  %arrayidx48 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add47
1142  %tmp17 = load float, float addrspace(3)* %arrayidx48, align 4
1143  %add51 = add nsw i32 %x.i, 16
1144  %arrayidx52 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add51
1145  %tmp18 = load float, float addrspace(3)* %arrayidx52, align 4
1146  %add55 = add nsw i32 %x.i, 17
1147  %arrayidx56 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add55
1148  %tmp19 = load float, float addrspace(3)* %arrayidx56, align 4
1149  %arrayidx60 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %y.i
1150  %tmp20 = load float, float addrspace(3)* %arrayidx60, align 4
1151  %add63 = add nsw i32 %y.i, 1
1152  %arrayidx64 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add63
1153  %tmp21 = load float, float addrspace(3)* %arrayidx64, align 4
1154  %add67 = add nsw i32 %y.i, 32
1155  %arrayidx68 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add67
1156  %tmp22 = load float, float addrspace(3)* %arrayidx68, align 4
1157  %add71 = add nsw i32 %y.i, 33
1158  %arrayidx72 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add71
1159  %tmp23 = load float, float addrspace(3)* %arrayidx72, align 4
1160  %add75 = add nsw i32 %y.i, 64
1161  %arrayidx76 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add75
1162  %tmp24 = load float, float addrspace(3)* %arrayidx76, align 4
1163  %add79 = add nsw i32 %y.i, 65
1164  %arrayidx80 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add79
1165  %tmp25 = load float, float addrspace(3)* %arrayidx80, align 4
1166  %sum.0 = fadd float %tmp16, %tmp17
1167  %sum.1 = fadd float %sum.0, %tmp18
1168  %sum.2 = fadd float %sum.1, %tmp19
1169  %sum.3 = fadd float %sum.2, %tmp20
1170  %sum.4 = fadd float %sum.3, %tmp21
1171  %sum.5 = fadd float %sum.4, %tmp22
1172  %sum.6 = fadd float %sum.5, %tmp23
1173  %sum.7 = fadd float %sum.6, %tmp24
1174  %sum.8 = fadd float %sum.7, %tmp25
1175  store float %sum.8, float addrspace(1)* %C, align 4
1176  ret void
1177}
1178
1179define amdgpu_kernel void @misaligned_read2_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(3)* %in) #0 {
1180; CI-LABEL: misaligned_read2_v2i32:
1181; CI:       ; %bb.0:
1182; CI-NEXT:    s_load_dword s2, s[0:1], 0x2
1183; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1184; CI-NEXT:    s_mov_b32 m0, -1
1185; CI-NEXT:    s_mov_b32 s3, 0xf000
1186; CI-NEXT:    s_waitcnt lgkmcnt(0)
1187; CI-NEXT:    v_mov_b32_e32 v0, s2
1188; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1189; CI-NEXT:    s_mov_b32 s2, -1
1190; CI-NEXT:    s_waitcnt lgkmcnt(0)
1191; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1192; CI-NEXT:    s_endpgm
1193;
1194; GFX9-LABEL: misaligned_read2_v2i32:
1195; GFX9:       ; %bb.0:
1196; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x8
1197; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
1198; GFX9-NEXT:    v_mov_b32_e32 v2, 0
1199; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1200; GFX9-NEXT:    v_mov_b32_e32 v0, s4
1201; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1202; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1203; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
1204; GFX9-NEXT:    s_endpgm
1205  %load = load <2 x i32>, <2 x i32> addrspace(3)* %in, align 4
1206  store <2 x i32> %load, <2 x i32> addrspace(1)* %out, align 8
1207  ret void
1208}
1209
1210define amdgpu_kernel void @misaligned_read2_i64(i64 addrspace(1)* %out, i64 addrspace(3)* %in) #0 {
1211; CI-LABEL: misaligned_read2_i64:
1212; CI:       ; %bb.0:
1213; CI-NEXT:    s_load_dword s2, s[0:1], 0x2
1214; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1215; CI-NEXT:    s_mov_b32 m0, -1
1216; CI-NEXT:    s_mov_b32 s3, 0xf000
1217; CI-NEXT:    s_waitcnt lgkmcnt(0)
1218; CI-NEXT:    v_mov_b32_e32 v0, s2
1219; CI-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1220; CI-NEXT:    s_mov_b32 s2, -1
1221; CI-NEXT:    s_waitcnt lgkmcnt(0)
1222; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1223; CI-NEXT:    s_endpgm
1224;
1225; GFX9-LABEL: misaligned_read2_i64:
1226; GFX9:       ; %bb.0:
1227; GFX9-NEXT:    s_load_dword s4, s[0:1], 0x8
1228; GFX9-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
1229; GFX9-NEXT:    v_mov_b32_e32 v2, 0
1230; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1231; GFX9-NEXT:    v_mov_b32_e32 v0, s4
1232; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:1
1233; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1234; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[2:3]
1235; GFX9-NEXT:    s_endpgm
1236  %load = load i64, i64 addrspace(3)* %in, align 4
1237  store i64 %load, i64 addrspace(1)* %out, align 8
1238  ret void
1239}
1240
1241define amdgpu_kernel void @ds_read_diff_base_interleaving(
1242; CI-LABEL: ds_read_diff_base_interleaving:
1243; CI:       ; %bb.0: ; %bb
1244; CI-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x2
1245; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1246; CI-NEXT:    v_lshlrev_b32_e32 v1, 4, v1
1247; CI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
1248; CI-NEXT:    s_mov_b32 m0, -1
1249; CI-NEXT:    s_waitcnt lgkmcnt(0)
1250; CI-NEXT:    v_add_i32_e32 v2, vcc, s4, v1
1251; CI-NEXT:    v_add_i32_e32 v3, vcc, s5, v0
1252; CI-NEXT:    v_add_i32_e32 v4, vcc, s6, v1
1253; CI-NEXT:    v_add_i32_e32 v6, vcc, s7, v0
1254; CI-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
1255; CI-NEXT:    ds_read2_b32 v[2:3], v3 offset1:4
1256; CI-NEXT:    ds_read2_b32 v[4:5], v4 offset1:1
1257; CI-NEXT:    ds_read2_b32 v[6:7], v6 offset1:4
1258; CI-NEXT:    s_mov_b32 s3, 0xf000
1259; CI-NEXT:    s_mov_b32 s2, -1
1260; CI-NEXT:    s_waitcnt lgkmcnt(2)
1261; CI-NEXT:    v_mul_f32_e32 v0, v0, v2
1262; CI-NEXT:    v_add_f32_e32 v0, 2.0, v0
1263; CI-NEXT:    s_waitcnt lgkmcnt(0)
1264; CI-NEXT:    v_mul_f32_e32 v2, v4, v6
1265; CI-NEXT:    v_sub_f32_e32 v0, v0, v2
1266; CI-NEXT:    v_mul_f32_e32 v1, v1, v3
1267; CI-NEXT:    v_sub_f32_e32 v0, v0, v1
1268; CI-NEXT:    v_mul_f32_e32 v1, v5, v7
1269; CI-NEXT:    v_sub_f32_e32 v0, v0, v1
1270; CI-NEXT:    buffer_store_dword v0, off, s[0:3], 0 offset:40
1271; CI-NEXT:    s_endpgm
1272;
1273; GFX9-LABEL: ds_read_diff_base_interleaving:
1274; GFX9:       ; %bb.0: ; %bb
1275; GFX9-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x8
1276; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 4, v1
1277; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
1278; GFX9-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1279; GFX9-NEXT:    v_mov_b32_e32 v8, 0
1280; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1281; GFX9-NEXT:    v_add_u32_e32 v2, s4, v1
1282; GFX9-NEXT:    v_add_u32_e32 v3, s5, v0
1283; GFX9-NEXT:    v_add_u32_e32 v4, s6, v1
1284; GFX9-NEXT:    v_add_u32_e32 v6, s7, v0
1285; GFX9-NEXT:    ds_read2_b32 v[0:1], v2 offset1:1
1286; GFX9-NEXT:    ds_read2_b32 v[2:3], v3 offset1:4
1287; GFX9-NEXT:    ds_read2_b32 v[4:5], v4 offset1:1
1288; GFX9-NEXT:    ds_read2_b32 v[6:7], v6 offset1:4
1289; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
1290; GFX9-NEXT:    v_mul_f32_e32 v0, v0, v2
1291; GFX9-NEXT:    v_add_f32_e32 v0, 2.0, v0
1292; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1293; GFX9-NEXT:    v_mul_f32_e32 v2, v4, v6
1294; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v2
1295; GFX9-NEXT:    v_mul_f32_e32 v1, v1, v3
1296; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v1
1297; GFX9-NEXT:    v_mul_f32_e32 v1, v5, v7
1298; GFX9-NEXT:    v_sub_f32_e32 v0, v0, v1
1299; GFX9-NEXT:    global_store_dword v8, v0, s[0:1] offset:40
1300; GFX9-NEXT:    s_endpgm
1301  float addrspace(1)* nocapture %arg,
1302  [4 x [4 x float]] addrspace(3)* %arg1,
1303  [4 x [4 x float]] addrspace(3)* %arg2,
1304  [4 x [4 x float]] addrspace(3)* %arg3,
1305  [4 x [4 x float]] addrspace(3)* %arg4) #1 {
1306bb:
1307  %tmp = getelementptr float, float addrspace(1)* %arg, i64 10
1308  %tmp5 = tail call i32 @llvm.amdgcn.workitem.id.x() #2
1309  %tmp6 = tail call i32 @llvm.amdgcn.workitem.id.y() #2
1310  %tmp7 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 0
1311  %tmp8 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 0, i32 %tmp5
1312  %tmp9 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 0
1313  %tmp10 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 0, i32 %tmp5
1314  %tmp11 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 1
1315  %tmp12 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 1, i32 %tmp5
1316  %tmp13 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 1
1317  %tmp14 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 1, i32 %tmp5
1318  %tmp15 = load float, float addrspace(3)* %tmp7
1319  %tmp16 = load float, float addrspace(3)* %tmp8
1320  %tmp17 = fmul float %tmp15, %tmp16
1321  %tmp18 = fadd float 2.000000e+00, %tmp17
1322  %tmp19 = load float, float addrspace(3)* %tmp9
1323  %tmp20 = load float, float addrspace(3)* %tmp10
1324  %tmp21 = fmul float %tmp19, %tmp20
1325  %tmp22 = fsub float %tmp18, %tmp21
1326  %tmp23 = load float, float addrspace(3)* %tmp11
1327  %tmp24 = load float, float addrspace(3)* %tmp12
1328  %tmp25 = fmul float %tmp23, %tmp24
1329  %tmp26 = fsub float %tmp22, %tmp25
1330  %tmp27 = load float, float addrspace(3)* %tmp13
1331  %tmp28 = load float, float addrspace(3)* %tmp14
1332  %tmp29 = fmul float %tmp27, %tmp28
1333  %tmp30 = fsub float %tmp26, %tmp29
1334  store float %tmp30, float addrspace(1)* %tmp
1335  ret void
1336}
1337
1338define amdgpu_kernel void @ds_read_call_read(i32 addrspace(1)* %out, i32 addrspace(3)* %arg) {
1339; CI-LABEL: ds_read_call_read:
1340; CI:       ; %bb.0:
1341; CI-NEXT:    s_getpc_b64 s[40:41]
1342; CI-NEXT:    s_mov_b32 s40, s0
1343; CI-NEXT:    s_load_dwordx4 s[40:43], s[40:41], 0x0
1344; CI-NEXT:    s_mov_b32 s14, s10
1345; CI-NEXT:    s_mov_b32 s12, s8
1346; CI-NEXT:    s_mov_b32 s13, s9
1347; CI-NEXT:    v_lshlrev_b32_e32 v3, 2, v0
1348; CI-NEXT:    s_waitcnt lgkmcnt(0)
1349; CI-NEXT:    s_add_u32 s40, s40, s11
1350; CI-NEXT:    s_mov_b64 s[10:11], s[6:7]
1351; CI-NEXT:    s_load_dwordx2 s[36:37], s[4:5], 0x0
1352; CI-NEXT:    s_load_dword s6, s[4:5], 0x2
1353; CI-NEXT:    s_addc_u32 s41, s41, 0
1354; CI-NEXT:    s_add_u32 s8, s4, 12
1355; CI-NEXT:    s_addc_u32 s9, s5, 0
1356; CI-NEXT:    s_getpc_b64 s[4:5]
1357; CI-NEXT:    s_add_u32 s4, s4, void_func_void@gotpcrel32@lo+4
1358; CI-NEXT:    s_addc_u32 s5, s5, void_func_void@gotpcrel32@hi+12
1359; CI-NEXT:    s_waitcnt lgkmcnt(0)
1360; CI-NEXT:    v_add_i32_e32 v40, vcc, s6, v3
1361; CI-NEXT:    s_mov_b32 m0, -1
1362; CI-NEXT:    s_load_dwordx2 s[16:17], s[4:5], 0x0
1363; CI-NEXT:    ds_read_b32 v41, v40
1364; CI-NEXT:    v_lshlrev_b32_e32 v1, 10, v1
1365; CI-NEXT:    v_lshlrev_b32_e32 v2, 20, v2
1366; CI-NEXT:    v_or_b32_e32 v0, v0, v1
1367; CI-NEXT:    s_mov_b64 s[4:5], s[0:1]
1368; CI-NEXT:    s_mov_b64 s[6:7], s[2:3]
1369; CI-NEXT:    s_mov_b64 s[0:1], s[40:41]
1370; CI-NEXT:    v_or_b32_e32 v31, v0, v2
1371; CI-NEXT:    s_mov_b64 s[2:3], s[42:43]
1372; CI-NEXT:    s_mov_b32 s32, 0
1373; CI-NEXT:    s_mov_b32 s39, 0xf000
1374; CI-NEXT:    s_mov_b32 s38, -1
1375; CI-NEXT:    s_waitcnt lgkmcnt(0)
1376; CI-NEXT:    s_swappc_b64 s[30:31], s[16:17]
1377; CI-NEXT:    ds_read_b32 v0, v40 offset:4
1378; CI-NEXT:    s_waitcnt lgkmcnt(0)
1379; CI-NEXT:    v_add_i32_e32 v0, vcc, v41, v0
1380; CI-NEXT:    buffer_store_dword v0, off, s[36:39], 0
1381; CI-NEXT:    s_endpgm
1382;
1383; GFX9-LABEL: ds_read_call_read:
1384; GFX9:       ; %bb.0:
1385; GFX9-NEXT:    s_getpc_b64 s[36:37]
1386; GFX9-NEXT:    s_mov_b32 s36, s0
1387; GFX9-NEXT:    s_load_dwordx4 s[36:39], s[36:37], 0x0
1388; GFX9-NEXT:    s_mov_b32 s14, s10
1389; GFX9-NEXT:    s_mov_b32 s12, s8
1390; GFX9-NEXT:    s_mov_b32 s13, s9
1391; GFX9-NEXT:    v_lshlrev_b32_e32 v2, 20, v2
1392; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1393; GFX9-NEXT:    s_add_u32 s36, s36, s11
1394; GFX9-NEXT:    s_addc_u32 s37, s37, 0
1395; GFX9-NEXT:    s_mov_b64 s[10:11], s[6:7]
1396; GFX9-NEXT:    s_load_dword s6, s[4:5], 0x8
1397; GFX9-NEXT:    s_load_dwordx2 s[34:35], s[4:5], 0x0
1398; GFX9-NEXT:    s_add_u32 s8, s4, 12
1399; GFX9-NEXT:    s_addc_u32 s9, s5, 0
1400; GFX9-NEXT:    s_getpc_b64 s[4:5]
1401; GFX9-NEXT:    s_add_u32 s4, s4, void_func_void@gotpcrel32@lo+4
1402; GFX9-NEXT:    s_addc_u32 s5, s5, void_func_void@gotpcrel32@hi+12
1403; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1404; GFX9-NEXT:    v_lshl_add_u32 v41, v0, 2, s6
1405; GFX9-NEXT:    s_load_dwordx2 s[16:17], s[4:5], 0x0
1406; GFX9-NEXT:    ds_read_b32 v42, v41
1407; GFX9-NEXT:    v_lshlrev_b32_e32 v1, 10, v1
1408; GFX9-NEXT:    s_mov_b64 s[4:5], s[0:1]
1409; GFX9-NEXT:    s_mov_b64 s[6:7], s[2:3]
1410; GFX9-NEXT:    s_mov_b64 s[0:1], s[36:37]
1411; GFX9-NEXT:    v_or3_b32 v31, v0, v1, v2
1412; GFX9-NEXT:    s_mov_b64 s[2:3], s[38:39]
1413; GFX9-NEXT:    s_mov_b32 s32, 0
1414; GFX9-NEXT:    v_mov_b32_e32 v40, 0
1415; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1416; GFX9-NEXT:    s_swappc_b64 s[30:31], s[16:17]
1417; GFX9-NEXT:    ds_read_b32 v0, v41 offset:4
1418; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1419; GFX9-NEXT:    v_add_u32_e32 v0, v42, v0
1420; GFX9-NEXT:    global_store_dword v40, v0, s[34:35]
1421; GFX9-NEXT:    s_endpgm
1422  %x = call i32 @llvm.amdgcn.workitem.id.x()
1423  %arrayidx0 = getelementptr i32, i32 addrspace(3)* %arg, i32 %x
1424  %arrayidx1 = getelementptr i32, i32 addrspace(3)* %arrayidx0, i32 1
1425  %v0 = load i32, i32 addrspace(3)* %arrayidx0, align 4
1426  call void @void_func_void()
1427  %v1 = load i32, i32 addrspace(3)* %arrayidx1, align 4
1428  %r = add i32 %v0, %v1
1429  store i32 %r, i32 addrspace(1)* %out, align 4
1430  ret void
1431}
1432
1433define amdgpu_ps <2 x float> @ds_read_interp_read(i32 inreg %prims, float addrspace(3)* %inptr) {
1434; CI-LABEL: ds_read_interp_read:
1435; CI:       ; %bb.0:
1436; CI-NEXT:    s_mov_b32 m0, -1
1437; CI-NEXT:    ds_read_b32 v2, v0
1438; CI-NEXT:    s_mov_b32 m0, s0
1439; CI-NEXT:    v_interp_mov_f32 v1, p10, attr0.x
1440; CI-NEXT:    s_mov_b32 m0, -1
1441; CI-NEXT:    ds_read_b32 v0, v0 offset:16
1442; CI-NEXT:    s_waitcnt lgkmcnt(0)
1443; CI-NEXT:    v_add_f32_e32 v1, v0, v1
1444; CI-NEXT:    v_mov_b32_e32 v0, v2
1445; CI-NEXT:    ; return to shader part epilog
1446;
1447; GFX9-LABEL: ds_read_interp_read:
1448; GFX9:       ; %bb.0:
1449; GFX9-NEXT:    ds_read2_b32 v[0:1], v0 offset1:4
1450; GFX9-NEXT:    s_mov_b32 m0, s0
1451; GFX9-NEXT:    s_nop 0
1452; GFX9-NEXT:    v_interp_mov_f32_e32 v2, p10, attr0.x
1453; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
1454; GFX9-NEXT:    v_add_f32_e32 v1, v1, v2
1455; GFX9-NEXT:    ; return to shader part epilog
1456  %v0 = load float, float addrspace(3)* %inptr, align 4
1457  %intrp = call float @llvm.amdgcn.interp.mov(i32 0, i32 0, i32 0, i32 %prims)
1458  %ptr1 = getelementptr float, float addrspace(3)* %inptr, i32 4
1459  %v1 = load float, float addrspace(3)* %ptr1, align 4
1460  %v1b = fadd float %v1, %intrp
1461  %r0 = insertelement <2 x float> undef, float %v0, i32 0
1462  %r1 = insertelement <2 x float> %r0, float %v1b, i32 1
1463  ret <2 x float> %r1
1464}
1465
1466@v2i32_align1 = internal addrspace(3) global [100 x <2 x i32>] undef, align 1
1467
1468define amdgpu_kernel void @read2_v2i32_align1_odd_offset(<2 x i32> addrspace(1)* %out) {
1469; CI-LABEL: read2_v2i32_align1_odd_offset:
1470; CI:       ; %bb.0: ; %entry
1471; CI-NEXT:    v_mov_b32_e32 v0, 0
1472; CI-NEXT:    s_mov_b32 m0, -1
1473; CI-NEXT:    ds_read_u8 v1, v0 offset:70
1474; CI-NEXT:    ds_read_u8 v2, v0 offset:72
1475; CI-NEXT:    ds_read_u8 v3, v0 offset:71
1476; CI-NEXT:    ds_read_u8 v4, v0 offset:69
1477; CI-NEXT:    ds_read_u8 v5, v0 offset:68
1478; CI-NEXT:    s_waitcnt lgkmcnt(4)
1479; CI-NEXT:    v_lshlrev_b32_e32 v1, 8, v1
1480; CI-NEXT:    s_waitcnt lgkmcnt(3)
1481; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v2
1482; CI-NEXT:    s_waitcnt lgkmcnt(2)
1483; CI-NEXT:    v_or_b32_e32 v2, v2, v3
1484; CI-NEXT:    s_waitcnt lgkmcnt(1)
1485; CI-NEXT:    v_or_b32_e32 v1, v1, v4
1486; CI-NEXT:    ds_read_u8 v4, v0 offset:67
1487; CI-NEXT:    ds_read_u8 v6, v0 offset:66
1488; CI-NEXT:    ds_read_u8 v0, v0 offset:65
1489; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
1490; CI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1491; CI-NEXT:    v_or_b32_e32 v1, v2, v1
1492; CI-NEXT:    s_waitcnt lgkmcnt(0)
1493; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v6
1494; CI-NEXT:    v_or_b32_e32 v0, v2, v0
1495; CI-NEXT:    v_lshlrev_b32_e32 v2, 8, v5
1496; CI-NEXT:    v_or_b32_e32 v2, v2, v4
1497; CI-NEXT:    v_lshlrev_b32_e32 v2, 16, v2
1498; CI-NEXT:    s_mov_b32 s3, 0xf000
1499; CI-NEXT:    s_mov_b32 s2, -1
1500; CI-NEXT:    v_or_b32_e32 v0, v2, v0
1501; CI-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
1502; CI-NEXT:    s_endpgm
1503;
1504; GFX9-ALIGNED-LABEL: read2_v2i32_align1_odd_offset:
1505; GFX9-ALIGNED:       ; %bb.0: ; %entry
1506; GFX9-ALIGNED-NEXT:    v_mov_b32_e32 v2, 0
1507; GFX9-ALIGNED-NEXT:    ds_read_u8 v0, v2 offset:70
1508; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v2 offset:65
1509; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v2 offset:66
1510; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v2 offset:67
1511; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v2 offset:68
1512; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v2 offset:69
1513; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v2 offset:72
1514; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v2 offset:71
1515; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(7)
1516; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
1517; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1518; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1519; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
1520; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v1, 8, v7
1521; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v1, v1, v8 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
1522; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v1, v1, v0
1523; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 8, v4
1524; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v0, v3
1525; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v3, 8, v6
1526; GFX9-ALIGNED-NEXT:    v_or_b32_sdwa v3, v3, v5 dst_sel:WORD_1 dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:DWORD
1527; GFX9-ALIGNED-NEXT:    v_or_b32_e32 v0, v3, v0
1528; GFX9-ALIGNED-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
1529; GFX9-ALIGNED-NEXT:    s_endpgm
1530;
1531; GFX9-UNALIGNED-LABEL: read2_v2i32_align1_odd_offset:
1532; GFX9-UNALIGNED:       ; %bb.0: ; %entry
1533; GFX9-UNALIGNED-NEXT:    v_mov_b32_e32 v2, 0
1534; GFX9-UNALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
1535; GFX9-UNALIGNED-NEXT:    ds_read_b64 v[0:1], v2 offset:65
1536; GFX9-UNALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
1537; GFX9-UNALIGNED-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
1538; GFX9-UNALIGNED-NEXT:    s_endpgm
1539entry:
1540  %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
1541  store <2 x i32> %load, <2 x i32> addrspace(1)* %out
1542  ret void
1543}
1544
1545declare void @void_func_void() #3
1546
1547declare i32 @llvm.amdgcn.workgroup.id.x() #1
1548declare i32 @llvm.amdgcn.workgroup.id.y() #1
1549declare i32 @llvm.amdgcn.workitem.id.x() #1
1550declare i32 @llvm.amdgcn.workitem.id.y() #1
1551
1552declare float @llvm.amdgcn.interp.mov(i32, i32, i32, i32) nounwind readnone
1553
1554declare void @llvm.amdgcn.s.barrier() #2
1555
1556attributes #0 = { nounwind }
1557attributes #1 = { nounwind readnone speculatable }
1558attributes #2 = { convergent nounwind }
1559attributes #3 = { nounwind noinline }
1560