1; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt -enable-misched < %s | FileCheck -strict-whitespace -check-prefix=SI %s
2
3; FIXME: We don't get cases where the address was an SGPR because we
4; get a copy to the address register for each one.
5
6@lds = addrspace(3) global [512 x float] undef, align 4
7 @lds.f64 = addrspace(3) global [512 x double] undef, align 8
8
9; SI-LABEL: @simple_read2_f32
10; SI: ds_read2_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:8
11; SI: s_waitcnt lgkmcnt(0)
12; SI: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[HI_VREG]], v[[LO_VREG]]
13; SI: buffer_store_dword [[RESULT]]
14; SI: s_endpgm
15define void @simple_read2_f32(float addrspace(1)* %out) #0 {
16  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
17  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
18  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
19  %add.x = add nsw i32 %x.i, 8
20  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
21  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
22  %sum = fadd float %val0, %val1
23  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
24  store float %sum, float addrspace(1)* %out.gep, align 4
25  ret void
26}
27
28; SI-LABEL: @simple_read2_f32_max_offset
29; SI: ds_read2_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:255
30; SI: s_waitcnt lgkmcnt(0)
31; SI: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[HI_VREG]], v[[LO_VREG]]
32; SI: buffer_store_dword [[RESULT]]
33; SI: s_endpgm
34define void @simple_read2_f32_max_offset(float addrspace(1)* %out) #0 {
35  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
36  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
37  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
38  %add.x = add nsw i32 %x.i, 255
39  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
40  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
41  %sum = fadd float %val0, %val1
42  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
43  store float %sum, float addrspace(1)* %out.gep, align 4
44  ret void
45}
46
47; SI-LABEL: @simple_read2_f32_too_far
48; SI-NOT ds_read2_b32
49; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
50; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:1028
51; SI: s_endpgm
52define void @simple_read2_f32_too_far(float addrspace(1)* %out) #0 {
53  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
54  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
55  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
56  %add.x = add nsw i32 %x.i, 257
57  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
58  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
59  %sum = fadd float %val0, %val1
60  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
61  store float %sum, float addrspace(1)* %out.gep, align 4
62  ret void
63}
64
65; SI-LABEL: @simple_read2_f32_x2
66; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset1:8
67; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
68; SI: s_endpgm
69define void @simple_read2_f32_x2(float addrspace(1)* %out) #0 {
70  %tid.x = tail call i32 @llvm.r600.read.tidig.x() #1
71  %idx.0 = add nsw i32 %tid.x, 0
72  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
73  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
74
75  %idx.1 = add nsw i32 %tid.x, 8
76  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
77  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
78  %sum.0 = fadd float %val0, %val1
79
80  %idx.2 = add nsw i32 %tid.x, 11
81  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
82  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
83
84  %idx.3 = add nsw i32 %tid.x, 27
85  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
86  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
87  %sum.1 = fadd float %val2, %val3
88
89  %sum = fadd float %sum.0, %sum.1
90  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
91  store float %sum, float addrspace(1)* %out.gep, align 4
92  ret void
93}
94
95; Make sure there is an instruction between the two sets of reads.
96; SI-LABEL: @simple_read2_f32_x2_barrier
97; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset1:8
98; SI: s_barrier
99; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
100; SI: s_endpgm
101define void @simple_read2_f32_x2_barrier(float addrspace(1)* %out) #0 {
102  %tid.x = tail call i32 @llvm.r600.read.tidig.x() #1
103  %idx.0 = add nsw i32 %tid.x, 0
104  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
105  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
106
107  %idx.1 = add nsw i32 %tid.x, 8
108  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
109  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
110  %sum.0 = fadd float %val0, %val1
111
112  call void @llvm.AMDGPU.barrier.local() #2
113
114  %idx.2 = add nsw i32 %tid.x, 11
115  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
116  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
117
118  %idx.3 = add nsw i32 %tid.x, 27
119  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
120  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
121  %sum.1 = fadd float %val2, %val3
122
123  %sum = fadd float %sum.0, %sum.1
124  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
125  store float %sum, float addrspace(1)* %out.gep, align 4
126  ret void
127}
128
129; For some reason adding something to the base address for the first
130; element results in only folding the inner pair.
131
132; SI-LABEL: @simple_read2_f32_x2_nonzero_base
133; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset0:2 offset1:8
134; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
135; SI: s_endpgm
136define void @simple_read2_f32_x2_nonzero_base(float addrspace(1)* %out) #0 {
137  %tid.x = tail call i32 @llvm.r600.read.tidig.x() #1
138  %idx.0 = add nsw i32 %tid.x, 2
139  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
140  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
141
142  %idx.1 = add nsw i32 %tid.x, 8
143  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
144  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
145  %sum.0 = fadd float %val0, %val1
146
147  %idx.2 = add nsw i32 %tid.x, 11
148  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
149  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
150
151  %idx.3 = add nsw i32 %tid.x, 27
152  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
153  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
154  %sum.1 = fadd float %val2, %val3
155
156  %sum = fadd float %sum.0, %sum.1
157  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
158  store float %sum, float addrspace(1)* %out.gep, align 4
159  ret void
160}
161
162; Be careful of vectors of pointers. We don't know if the 2 pointers
163; in the vectors are really the same base, so this is not safe to
164; merge.
165; Base pointers come from different subregister of same super
166; register. We can't safely merge this.
167
168; SI-LABEL: @read2_ptr_is_subreg_arg_f32
169; SI-NOT: ds_read2_b32
170; SI: ds_read_b32
171; SI: ds_read_b32
172; SI: s_endpgm
173define void @read2_ptr_is_subreg_arg_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
174  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
175  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
176  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
177  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
178  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
179  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
180  %val0 = load float, float addrspace(3)* %gep.0, align 4
181  %val1 = load float, float addrspace(3)* %gep.1, align 4
182  %add.x = add nsw i32 %x.i, 8
183  %sum = fadd float %val0, %val1
184  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
185  store float %sum, float addrspace(1)* %out.gep, align 4
186  ret void
187}
188
189; Apply a constant scalar offset after the pointer vector extract.  We
190; are rejecting merges that have the same, constant 0 offset, so make
191; sure we are really rejecting it because of the different
192; subregisters.
193
194; SI-LABEL: @read2_ptr_is_subreg_arg_offset_f32
195; SI-NOT: ds_read2_b32
196; SI: ds_read_b32
197; SI: ds_read_b32
198; SI: s_endpgm
199define void @read2_ptr_is_subreg_arg_offset_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
200  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
201  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
202  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
203  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
204  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
205  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
206
207  ; Apply an additional offset after the vector that will be more obviously folded.
208  %gep.1.offset = getelementptr float, float addrspace(3)* %gep.1, i32 8
209
210  %val0 = load float, float addrspace(3)* %gep.0, align 4
211  %val1 = load float, float addrspace(3)* %gep.1.offset, align 4
212  %add.x = add nsw i32 %x.i, 8
213  %sum = fadd float %val0, %val1
214  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
215  store float %sum, float addrspace(1)* %out.gep, align 4
216  ret void
217}
218
219; We should be able to merge in this case, but probably not worth the effort.
220; SI-NOT: ds_read2_b32
221; SI: ds_read_b32
222; SI: ds_read_b32
223; SI: s_endpgm
224define void @read2_ptr_is_subreg_f32(float addrspace(1)* %out) #0 {
225  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
226  %ptr.0 = insertelement <2 x [512 x float] addrspace(3)*> undef, [512 x float] addrspace(3)* @lds, i32 0
227  %ptr.1 = insertelement <2 x [512 x float] addrspace(3)*> %ptr.0, [512 x float] addrspace(3)* @lds, i32 1
228  %x.i.v.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
229  %x.i.v.1 = insertelement <2 x i32> %x.i.v.0, i32 %x.i, i32 1
230  %idx = add <2 x i32> %x.i.v.1, <i32 0, i32 8>
231  %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
232  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
233  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
234  %val0 = load float, float addrspace(3)* %gep.0, align 4
235  %val1 = load float, float addrspace(3)* %gep.1, align 4
236  %add.x = add nsw i32 %x.i, 8
237  %sum = fadd float %val0, %val1
238  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
239  store float %sum, float addrspace(1)* %out.gep, align 4
240  ret void
241}
242
243; SI-LABEL: @simple_read2_f32_volatile_0
244; SI-NOT ds_read2_b32
245; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
246; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:32
247; SI: s_endpgm
248define void @simple_read2_f32_volatile_0(float addrspace(1)* %out) #0 {
249  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
250  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
251  %val0 = load volatile float, float addrspace(3)* %arrayidx0, align 4
252  %add.x = add nsw i32 %x.i, 8
253  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
254  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
255  %sum = fadd float %val0, %val1
256  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
257  store float %sum, float addrspace(1)* %out.gep, align 4
258  ret void
259}
260
261; SI-LABEL: @simple_read2_f32_volatile_1
262; SI-NOT ds_read2_b32
263; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
264; SI: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:32
265; SI: s_endpgm
266define void @simple_read2_f32_volatile_1(float addrspace(1)* %out) #0 {
267  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
268  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
269  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
270  %add.x = add nsw i32 %x.i, 8
271  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
272  %val1 = load volatile float, float addrspace(3)* %arrayidx1, align 4
273  %sum = fadd float %val0, %val1
274  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
275  store float %sum, float addrspace(1)* %out.gep, align 4
276  ret void
277}
278
279; Can't fold since not correctly aligned.
280; XXX: This isn't really testing anything useful now. I think CI
281; allows unaligned LDS accesses, which would be a problem here.
282; SI-LABEL: @unaligned_read2_f32
283; SI-NOT: ds_read2_b32
284; SI: s_endpgm
285define void @unaligned_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
286  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
287  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
288  %val0 = load float, float addrspace(3)* %arrayidx0, align 1
289  %add.x = add nsw i32 %x.i, 8
290  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
291  %val1 = load float, float addrspace(3)* %arrayidx1, align 1
292  %sum = fadd float %val0, %val1
293  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
294  store float %sum, float addrspace(1)* %out.gep, align 4
295  ret void
296}
297
298; SI-LABEL: @misaligned_2_simple_read2_f32
299; SI-NOT: ds_read2_b32
300; SI: s_endpgm
301define void @misaligned_2_simple_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
302  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
303  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
304  %val0 = load float, float addrspace(3)* %arrayidx0, align 2
305  %add.x = add nsw i32 %x.i, 8
306  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
307  %val1 = load float, float addrspace(3)* %arrayidx1, align 2
308  %sum = fadd float %val0, %val1
309  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
310  store float %sum, float addrspace(1)* %out.gep, align 4
311  ret void
312}
313
314; SI-LABEL: @simple_read2_f64
315; SI: v_lshlrev_b32_e32 [[VPTR:v[0-9]+]], 3, {{v[0-9]+}}
316; SI: ds_read2_b64 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, [[VPTR]] offset1:8
317; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[LO_VREG]]:{{[0-9]+\]}}, v{{\[[0-9]+}}:[[HI_VREG]]{{\]}}
318; SI: buffer_store_dwordx2 [[RESULT]]
319; SI: s_endpgm
320define void @simple_read2_f64(double addrspace(1)* %out) #0 {
321  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
322  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
323  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
324  %add.x = add nsw i32 %x.i, 8
325  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
326  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
327  %sum = fadd double %val0, %val1
328  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
329  store double %sum, double addrspace(1)* %out.gep, align 8
330  ret void
331}
332
333; SI-LABEL: @simple_read2_f64_max_offset
334; SI: ds_read2_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset1:255
335; SI: s_endpgm
336define void @simple_read2_f64_max_offset(double addrspace(1)* %out) #0 {
337  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
338  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
339  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
340  %add.x = add nsw i32 %x.i, 255
341  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
342  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
343  %sum = fadd double %val0, %val1
344  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
345  store double %sum, double addrspace(1)* %out.gep, align 8
346  ret void
347}
348
349; SI-LABEL: @simple_read2_f64_too_far
350; SI-NOT ds_read2_b64
351; SI: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}
352; SI: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset:2056
353; SI: s_endpgm
354define void @simple_read2_f64_too_far(double addrspace(1)* %out) #0 {
355  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
356  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
357  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
358  %add.x = add nsw i32 %x.i, 257
359  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
360  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
361  %sum = fadd double %val0, %val1
362  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
363  store double %sum, double addrspace(1)* %out.gep, align 8
364  ret void
365}
366
367; Alignment only 4
368; SI-LABEL: @misaligned_read2_f64
369; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset1:1
370; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset0:14 offset1:15
371; SI: s_endpgm
372define void @misaligned_read2_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
373  %x.i = tail call i32 @llvm.r600.read.tidig.x() #1
374  %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
375  %val0 = load double, double addrspace(3)* %arrayidx0, align 4
376  %add.x = add nsw i32 %x.i, 7
377  %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
378  %val1 = load double, double addrspace(3)* %arrayidx1, align 4
379  %sum = fadd double %val0, %val1
380  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
381  store double %sum, double addrspace(1)* %out.gep, align 4
382  ret void
383}
384
385@foo = addrspace(3) global [4 x i32] undef, align 4
386
387; SI-LABEL: @load_constant_adjacent_offsets
388; SI: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
389; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset1:1
390define void @load_constant_adjacent_offsets(i32 addrspace(1)* %out) {
391  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
392  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
393  %sum = add i32 %val0, %val1
394  store i32 %sum, i32 addrspace(1)* %out, align 4
395  ret void
396}
397
398; SI-LABEL: @load_constant_disjoint_offsets
399; SI: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
400; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset1:2
401define void @load_constant_disjoint_offsets(i32 addrspace(1)* %out) {
402  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
403  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
404  %sum = add i32 %val0, %val1
405  store i32 %sum, i32 addrspace(1)* %out, align 4
406  ret void
407}
408
409@bar = addrspace(3) global [4 x i64] undef, align 4
410
411; SI-LABEL: @load_misaligned64_constant_offsets
412; SI: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
413; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset1:1
414; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]] offset0:2 offset1:3
415define void @load_misaligned64_constant_offsets(i64 addrspace(1)* %out) {
416  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 0), align 4
417  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 1), align 4
418  %sum = add i64 %val0, %val1
419  store i64 %sum, i64 addrspace(1)* %out, align 8
420  ret void
421}
422
423@bar.large = addrspace(3) global [4096 x i64] undef, align 4
424
425; SI-LABEL: @load_misaligned64_constant_large_offsets
426; SI-DAG: v_mov_b32_e32 [[BASE0:v[0-9]+]], 0x7ff8{{$}}
427; SI-DAG: v_mov_b32_e32 [[BASE1:v[0-9]+]], 0x4000
428; SI-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASE0]] offset1:1
429; SI-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASE1]] offset1:1
430; SI: s_endpgm
431define void @load_misaligned64_constant_large_offsets(i64 addrspace(1)* %out) {
432  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 2048), align 4
433  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 4095), align 4
434  %sum = add i64 %val0, %val1
435  store i64 %sum, i64 addrspace(1)* %out, align 8
436  ret void
437}
438
439@sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] undef, align 4
440@sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] undef, align 4
441
442define void @sgemm_inner_loop_read2_sequence(float addrspace(1)* %C, i32 %lda, i32 %ldb) #0 {
443  %x.i = tail call i32 @llvm.r600.read.tgid.x() #1
444  %y.i = tail call i32 @llvm.r600.read.tidig.y() #1
445  %arrayidx44 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %x.i
446  %tmp16 = load float, float addrspace(3)* %arrayidx44, align 4
447  %add47 = add nsw i32 %x.i, 1
448  %arrayidx48 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add47
449  %tmp17 = load float, float addrspace(3)* %arrayidx48, align 4
450  %add51 = add nsw i32 %x.i, 16
451  %arrayidx52 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add51
452  %tmp18 = load float, float addrspace(3)* %arrayidx52, align 4
453  %add55 = add nsw i32 %x.i, 17
454  %arrayidx56 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add55
455  %tmp19 = load float, float addrspace(3)* %arrayidx56, align 4
456  %arrayidx60 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %y.i
457  %tmp20 = load float, float addrspace(3)* %arrayidx60, align 4
458  %add63 = add nsw i32 %y.i, 1
459  %arrayidx64 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add63
460  %tmp21 = load float, float addrspace(3)* %arrayidx64, align 4
461  %add67 = add nsw i32 %y.i, 32
462  %arrayidx68 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add67
463  %tmp22 = load float, float addrspace(3)* %arrayidx68, align 4
464  %add71 = add nsw i32 %y.i, 33
465  %arrayidx72 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add71
466  %tmp23 = load float, float addrspace(3)* %arrayidx72, align 4
467  %add75 = add nsw i32 %y.i, 64
468  %arrayidx76 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add75
469  %tmp24 = load float, float addrspace(3)* %arrayidx76, align 4
470  %add79 = add nsw i32 %y.i, 65
471  %arrayidx80 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add79
472  %tmp25 = load float, float addrspace(3)* %arrayidx80, align 4
473  %sum.0 = fadd float %tmp16, %tmp17
474  %sum.1 = fadd float %sum.0, %tmp18
475  %sum.2 = fadd float %sum.1, %tmp19
476  %sum.3 = fadd float %sum.2, %tmp20
477  %sum.4 = fadd float %sum.3, %tmp21
478  %sum.5 = fadd float %sum.4, %tmp22
479  %sum.6 = fadd float %sum.5, %tmp23
480  %sum.7 = fadd float %sum.6, %tmp24
481  %sum.8 = fadd float %sum.7, %tmp25
482  store float %sum.8, float addrspace(1)* %C, align 4
483  ret void
484}
485
486define void @misaligned_read2_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(3)* %in) #0 {
487  %load = load <2 x i32>, <2 x i32> addrspace(3)* %in, align 4
488  store <2 x i32> %load, <2 x i32> addrspace(1)* %out, align 8
489  ret void
490}
491
492define void @misaligned_read2_i64(i64 addrspace(1)* %out, i64 addrspace(3)* %in) #0 {
493  %load = load i64, i64 addrspace(3)* %in, align 4
494  store i64 %load, i64 addrspace(1)* %out, align 8
495  ret void
496}
497
498; Function Attrs: nounwind readnone
499declare i32 @llvm.r600.read.tgid.x() #1
500
501; Function Attrs: nounwind readnone
502declare i32 @llvm.r600.read.tgid.y() #1
503
504; Function Attrs: nounwind readnone
505declare i32 @llvm.r600.read.tidig.x() #1
506
507; Function Attrs: nounwind readnone
508declare i32 @llvm.r600.read.tidig.y() #1
509
510; Function Attrs: noduplicate nounwind
511declare void @llvm.AMDGPU.barrier.local() #2
512
513attributes #0 = { nounwind "less-precise-fpmad"="false" "no-frame-pointer-elim"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-realign-stack" "stack-protector-buffer-size"="8" "unsafe-fp-math"="false" "use-soft-float"="false" }
514attributes #1 = { nounwind readnone }
515attributes #2 = { noduplicate nounwind }
516