1; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -enable-var-scope -strict-whitespace -check-prefixes=GCN,CI %s
2; RUN: llc -march=amdgcn -mcpu=gfx900 -verify-machineinstrs -mattr=+load-store-opt,+flat-for-global < %s | FileCheck -enable-var-scope -strict-whitespace -check-prefixes=GCN,GFX9 %s
3
4; FIXME: We don't get cases where the address was an SGPR because we
5; get a copy to the address register for each one.
6
7@lds = addrspace(3) global [512 x float] undef, align 4
8@lds.f64 = addrspace(3) global [512 x double] undef, align 8
9
10; GCN-LABEL: {{^}}simple_read2_f32:
11; CI-DAG: s_mov_b32 m0
12; GFX9-NOT: m0
13
14; GCN: ds_read2_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:8
15; GCN: s_waitcnt lgkmcnt(0)
16; GCN: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[LO_VREG]], v[[HI_VREG]]
17; CI: buffer_store_dword [[RESULT]]
18; GFX9: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]]
19; GCN: s_endpgm
20define amdgpu_kernel void @simple_read2_f32(float addrspace(1)* %out) #0 {
21  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
22  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
23  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
24  %add.x = add nsw i32 %x.i, 8
25  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
26  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
27  %sum = fadd float %val0, %val1
28  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
29  store float %sum, float addrspace(1)* %out.gep, align 4
30  ret void
31}
32
33; GCN-LABEL: {{^}}simple_read2_f32_max_offset:
34; CI-DAG: s_mov_b32 m0
35; GFX9-NOT: m0
36
37; GCN: ds_read2_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:255
38; GCN: s_waitcnt lgkmcnt(0)
39; GCN: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[LO_VREG]], v[[HI_VREG]]
40
41; CI: buffer_store_dword [[RESULT]]
42; GFX9: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]]
43define amdgpu_kernel void @simple_read2_f32_max_offset(float addrspace(1)* %out) #0 {
44  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
45  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
46  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
47  %add.x = add nsw i32 %x.i, 255
48  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
49  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
50  %sum = fadd float %val0, %val1
51  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
52  store float %sum, float addrspace(1)* %out.gep, align 4
53  ret void
54}
55
56; GCN-LABEL: @simple_read2_f32_too_far
57; CI-DAG: s_mov_b32 m0
58; GFX9-NOT: m0
59
60; GCN-NOT: ds_read2_b32
61; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
62; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:1028
63; GCN: s_endpgm
64define amdgpu_kernel void @simple_read2_f32_too_far(float addrspace(1)* %out) #0 {
65  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
66  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
67  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
68  %add.x = add nsw i32 %x.i, 257
69  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
70  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
71  %sum = fadd float %val0, %val1
72  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
73  store float %sum, float addrspace(1)* %out.gep, align 4
74  ret void
75}
76
77; GCN-LABEL: @simple_read2_f32_x2
78; CI-DAG: s_mov_b32 m0
79; GFX9-NOT: m0
80
81; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset1:8
82; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
83; GCN: s_endpgm
84define amdgpu_kernel void @simple_read2_f32_x2(float addrspace(1)* %out) #0 {
85  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
86  %idx.0 = add nsw i32 %tid.x, 0
87  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
88  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
89
90  %idx.1 = add nsw i32 %tid.x, 8
91  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
92  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
93  %sum.0 = fadd float %val0, %val1
94
95  %idx.2 = add nsw i32 %tid.x, 11
96  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
97  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
98
99  %idx.3 = add nsw i32 %tid.x, 27
100  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
101  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
102  %sum.1 = fadd float %val2, %val3
103
104  %sum = fadd float %sum.0, %sum.1
105  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
106  store float %sum, float addrspace(1)* %out.gep, align 4
107  ret void
108}
109
110; Make sure there is an instruction between the two sets of reads.
111; GCN-LABEL: @simple_read2_f32_x2_barrier
112; CI-DAG: s_mov_b32 m0
113; GFX9-NOT: m0
114
115; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset1:8
116; GCN: s_barrier
117; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
118; GCN: s_endpgm
119define amdgpu_kernel void @simple_read2_f32_x2_barrier(float addrspace(1)* %out) #0 {
120  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
121  %idx.0 = add nsw i32 %tid.x, 0
122  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
123  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
124
125  %idx.1 = add nsw i32 %tid.x, 8
126  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
127  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
128  %sum.0 = fadd float %val0, %val1
129
130  call void @llvm.amdgcn.s.barrier() #2
131
132  %idx.2 = add nsw i32 %tid.x, 11
133  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
134  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
135
136  %idx.3 = add nsw i32 %tid.x, 27
137  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
138  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
139  %sum.1 = fadd float %val2, %val3
140
141  %sum = fadd float %sum.0, %sum.1
142  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
143  store float %sum, float addrspace(1)* %out.gep, align 4
144  ret void
145}
146
147; For some reason adding something to the base address for the first
148; element results in only folding the inner pair.
149
150; GCN-LABEL: @simple_read2_f32_x2_nonzero_base
151; CI-DAG: s_mov_b32 m0
152; GFX9-NOT: m0
153
154; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR:v[0-9]+]] offset0:2 offset1:8
155; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[BASEADDR]] offset0:11 offset1:27
156; GCN: s_endpgm
157define amdgpu_kernel void @simple_read2_f32_x2_nonzero_base(float addrspace(1)* %out) #0 {
158  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x() #1
159  %idx.0 = add nsw i32 %tid.x, 2
160  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.0
161  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
162
163  %idx.1 = add nsw i32 %tid.x, 8
164  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.1
165  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
166  %sum.0 = fadd float %val0, %val1
167
168  %idx.2 = add nsw i32 %tid.x, 11
169  %arrayidx2 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.2
170  %val2 = load float, float addrspace(3)* %arrayidx2, align 4
171
172  %idx.3 = add nsw i32 %tid.x, 27
173  %arrayidx3 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %idx.3
174  %val3 = load float, float addrspace(3)* %arrayidx3, align 4
175  %sum.1 = fadd float %val2, %val3
176
177  %sum = fadd float %sum.0, %sum.1
178  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %idx.0
179  store float %sum, float addrspace(1)* %out.gep, align 4
180  ret void
181}
182
183; Be careful of vectors of pointers. We don't know if the 2 pointers
184; in the vectors are really the same base, so this is not safe to
185; merge.
186; Base pointers come from different subregister of same super
187; register. We can't safely merge this.
188
189; GCN-LABEL: @read2_ptr_is_subreg_arg_f32
190; CI-DAG: s_mov_b32 m0
191; GFX9-NOT: m0
192
193; GCN-NOT: ds_read2_b32
194; GCN: ds_read_b32
195; GCN: ds_read_b32
196; GCN: s_endpgm
197define amdgpu_kernel void @read2_ptr_is_subreg_arg_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
198  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
199  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
200  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
201  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
202  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
203  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
204  %val0 = load float, float addrspace(3)* %gep.0, align 4
205  %val1 = load float, float addrspace(3)* %gep.1, align 4
206  %add.x = add nsw i32 %x.i, 8
207  %sum = fadd float %val0, %val1
208  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
209  store float %sum, float addrspace(1)* %out.gep, align 4
210  ret void
211}
212
213; Apply a constant scalar offset after the pointer vector extract.  We
214; are rejecting merges that have the same, constant 0 offset, so make
215; sure we are really rejecting it because of the different
216; subregisters.
217
218; GCN-LABEL: @read2_ptr_is_subreg_arg_offset_f32
219; CI-DAG: s_mov_b32 m0
220; GFX9-NOT: m0
221
222; GCN-NOT: ds_read2_b32
223; GCN: ds_read_b32
224; GCN: ds_read_b32
225; GCN: s_endpgm
226define amdgpu_kernel void @read2_ptr_is_subreg_arg_offset_f32(float addrspace(1)* %out, <2 x float addrspace(3)*> %lds.ptr) #0 {
227  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
228  %index.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
229  %index.1 = insertelement <2 x i32> %index.0, i32 8, i32 0
230  %gep = getelementptr inbounds float, <2 x float addrspace(3)*> %lds.ptr, <2 x i32> %index.1
231  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
232  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
233
234  ; Apply an additional offset after the vector that will be more obviously folded.
235  %gep.1.offset = getelementptr float, float addrspace(3)* %gep.1, i32 8
236
237  %val0 = load float, float addrspace(3)* %gep.0, align 4
238  %val1 = load float, float addrspace(3)* %gep.1.offset, align 4
239  %add.x = add nsw i32 %x.i, 8
240  %sum = fadd float %val0, %val1
241  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
242  store float %sum, float addrspace(1)* %out.gep, align 4
243  ret void
244}
245
246; GCN-LABEL: {{^}}read2_ptr_is_subreg_f32:
247; CI-DAG: s_mov_b32 m0
248; GFX9-NOT: m0
249
250; GCN: ds_read2_b32 {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset1:8{{$}}
251; GCN: s_endpgm
252define amdgpu_kernel void @read2_ptr_is_subreg_f32(float addrspace(1)* %out) #0 {
253  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
254  %ptr.0 = insertelement <2 x [512 x float] addrspace(3)*> undef, [512 x float] addrspace(3)* @lds, i32 0
255  %ptr.1 = insertelement <2 x [512 x float] addrspace(3)*> %ptr.0, [512 x float] addrspace(3)* @lds, i32 1
256  %x.i.v.0 = insertelement <2 x i32> undef, i32 %x.i, i32 0
257  %x.i.v.1 = insertelement <2 x i32> %x.i.v.0, i32 %x.i, i32 1
258  %idx = add <2 x i32> %x.i.v.1, <i32 0, i32 8>
259  %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
260  %gep.0 = extractelement <2 x float addrspace(3)*> %gep, i32 0
261  %gep.1 = extractelement <2 x float addrspace(3)*> %gep, i32 1
262  %val0 = load float, float addrspace(3)* %gep.0, align 4
263  %val1 = load float, float addrspace(3)* %gep.1, align 4
264  %add.x = add nsw i32 %x.i, 8
265  %sum = fadd float %val0, %val1
266  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
267  store float %sum, float addrspace(1)* %out.gep, align 4
268  ret void
269}
270
271; GCN-LABEL: @simple_read2_f32_volatile_0
272; CI-DAG: s_mov_b32 m0
273; GFX9-NOT: m0
274
275; GCN-NOT: ds_read2_b32
276; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
277; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:32
278; GCN: s_endpgm
279define amdgpu_kernel void @simple_read2_f32_volatile_0(float addrspace(1)* %out) #0 {
280  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
281  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
282  %val0 = load volatile float, float addrspace(3)* %arrayidx0, align 4
283  %add.x = add nsw i32 %x.i, 8
284  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
285  %val1 = load float, float addrspace(3)* %arrayidx1, align 4
286  %sum = fadd float %val0, %val1
287  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
288  store float %sum, float addrspace(1)* %out.gep, align 4
289  ret void
290}
291
292; GCN-LABEL: @simple_read2_f32_volatile_1
293; CI-DAG: s_mov_b32 m0
294; GFX9-NOT: m0
295
296; GCN-NOT: ds_read2_b32
297; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}}
298; GCN: ds_read_b32 v{{[0-9]+}}, v{{[0-9]+}} offset:32
299; GCN: s_endpgm
300define amdgpu_kernel void @simple_read2_f32_volatile_1(float addrspace(1)* %out) #0 {
301  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
302  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
303  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
304  %add.x = add nsw i32 %x.i, 8
305  %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
306  %val1 = load volatile float, float addrspace(3)* %arrayidx1, align 4
307  %sum = fadd float %val0, %val1
308  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
309  store float %sum, float addrspace(1)* %out.gep, align 4
310  ret void
311}
312
313; Can't fold since not correctly aligned.
314; XXX: This isn't really testing anything useful now. I think CI
315; allows unaligned LDS accesses, which would be a problem here.
316; GCN-LABEL: @unaligned_read2_f32
317; CI-DAG: s_mov_b32 m0
318; GFX9-NOT: m0
319
320; GCN-NOT: ds_read2_b32
321; GCN: s_endpgm
322define amdgpu_kernel void @unaligned_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
323  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
324  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
325  %val0 = load float, float addrspace(3)* %arrayidx0, align 1
326  %add.x = add nsw i32 %x.i, 8
327  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
328  %val1 = load float, float addrspace(3)* %arrayidx1, align 1
329  %sum = fadd float %val0, %val1
330  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
331  store float %sum, float addrspace(1)* %out.gep, align 4
332  ret void
333}
334
335; GCN-LABEL: @misaligned_2_simple_read2_f32
336; CI-DAG: s_mov_b32 m0
337; GFX9-NOT: m0
338
339; GCN-NOT: ds_read2_b32
340; GCN: s_endpgm
341define amdgpu_kernel void @misaligned_2_simple_read2_f32(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
342  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
343  %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %x.i
344  %val0 = load float, float addrspace(3)* %arrayidx0, align 2
345  %add.x = add nsw i32 %x.i, 8
346  %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x
347  %val1 = load float, float addrspace(3)* %arrayidx1, align 2
348  %sum = fadd float %val0, %val1
349  %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
350  store float %sum, float addrspace(1)* %out.gep, align 4
351  ret void
352}
353
354; GCN-LABEL: @simple_read2_f64
355; CI-DAG: s_mov_b32 m0
356; GFX9-NOT: m0
357
358; GCN-DAG: v_lshlrev_b32_e32 [[VOFS:v[0-9]+]], 3, {{v[0-9]+}}
359; GCN-DAG: v_add_{{[iu]}}32_e32 [[VPTR:v[0-9]+]], {{(vcc, )?}}lds.f64@abs32@lo, [[VOFS]]
360; GCN: ds_read2_b64 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, [[VPTR]] offset1:8
361; GCN: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[LO_VREG]]:{{[0-9]+\]}}, v{{\[[0-9]+}}:[[HI_VREG]]{{\]}}
362
363; CI: buffer_store_dwordx2 [[RESULT]]
364; GFX9: global_store_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[RESULT]]
365define amdgpu_kernel void @simple_read2_f64(double addrspace(1)* %out) #0 {
366  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
367  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
368  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
369  %add.x = add nsw i32 %x.i, 8
370  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
371  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
372  %sum = fadd double %val0, %val1
373  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
374  store double %sum, double addrspace(1)* %out.gep, align 8
375  ret void
376}
377
378; GCN-LABEL: @simple_read2_f64_max_offset
379; CI-DAG: s_mov_b32 m0
380; GFX9-NOT: m0
381
382; GCN: ds_read2_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset1:255
383; GCN: s_endpgm
384define amdgpu_kernel void @simple_read2_f64_max_offset(double addrspace(1)* %out) #0 {
385  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
386  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
387  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
388  %add.x = add nsw i32 %x.i, 255
389  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
390  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
391  %sum = fadd double %val0, %val1
392  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
393  store double %sum, double addrspace(1)* %out.gep, align 8
394  ret void
395}
396
397; GCN-LABEL: @simple_read2_f64_too_far
398; CI-DAG: s_mov_b32 m0
399; GFX9-NOT: m0
400
401; GCN-NOT: ds_read2_b64
402; GCN: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}
403; GCN: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset:2056
404; GCN: s_endpgm
405define amdgpu_kernel void @simple_read2_f64_too_far(double addrspace(1)* %out) #0 {
406  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
407  %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
408  %val0 = load double, double addrspace(3)* %arrayidx0, align 8
409  %add.x = add nsw i32 %x.i, 257
410  %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
411  %val1 = load double, double addrspace(3)* %arrayidx1, align 8
412  %sum = fadd double %val0, %val1
413  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
414  store double %sum, double addrspace(1)* %out.gep, align 8
415  ret void
416}
417
418; Alignment only 4
419; GCN-LABEL: @misaligned_read2_f64
420; CI-DAG: s_mov_b32 m0
421; GFX9-NOT: m0
422
423; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset1:1
424; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset0:14 offset1:15
425; GCN: s_endpgm
426define amdgpu_kernel void @misaligned_read2_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
427  %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
428  %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
429  %val0 = load double, double addrspace(3)* %arrayidx0, align 4
430  %add.x = add nsw i32 %x.i, 7
431  %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
432  %val1 = load double, double addrspace(3)* %arrayidx1, align 4
433  %sum = fadd double %val0, %val1
434  %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
435  store double %sum, double addrspace(1)* %out.gep, align 4
436  ret void
437}
438
439@foo = addrspace(3) global [4 x i32] undef, align 4
440
441; GCN-LABEL: @load_constant_adjacent_offsets
442; CI-DAG: s_mov_b32 m0
443; GFX9-NOT: m0
444
445; GCN-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], foo@abs32@lo{{$}}
446; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[PTR]] offset1:1
447define amdgpu_kernel void @load_constant_adjacent_offsets(i32 addrspace(1)* %out) {
448  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
449  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 1), align 4
450  %sum = add i32 %val0, %val1
451  store i32 %sum, i32 addrspace(1)* %out, align 4
452  ret void
453}
454
455; GCN-LABEL: @load_constant_disjoint_offsets
456; CI-DAG: s_mov_b32 m0
457; GFX9-NOT: m0
458
459; GCN-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], foo@abs32@lo{{$}}
460; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[PTR]] offset1:2
461define amdgpu_kernel void @load_constant_disjoint_offsets(i32 addrspace(1)* %out) {
462  %val0 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 0), align 4
463  %val1 = load i32, i32 addrspace(3)* getelementptr inbounds ([4 x i32], [4 x i32] addrspace(3)* @foo, i32 0, i32 2), align 4
464  %sum = add i32 %val0, %val1
465  store i32 %sum, i32 addrspace(1)* %out, align 4
466  ret void
467}
468
469@bar = addrspace(3) global [4 x i64] undef, align 4
470
471; GCN-LABEL: @load_misaligned64_constant_offsets
472; CI-DAG: s_mov_b32 m0
473; GFX9-NOT: m0
474
475; GCN-DAG: v_mov_b32_e32 [[PTR:v[0-9]+]], bar@abs32@lo{{$}}
476; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[PTR]] offset0:2 offset1:3
477; GCN: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[PTR]] offset1:1
478define amdgpu_kernel void @load_misaligned64_constant_offsets(i64 addrspace(1)* %out) {
479  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 0), align 4
480  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4 x i64], [4 x i64] addrspace(3)* @bar, i32 0, i32 1), align 4
481  %sum = add i64 %val0, %val1
482  store i64 %sum, i64 addrspace(1)* %out, align 8
483  ret void
484}
485
486@bar.large = addrspace(3) global [4096 x i64] undef, align 4
487
488; GCN-LABEL: @load_misaligned64_constant_large_offsets
489; CI-DAG: s_mov_b32 m0
490; GFX9-NOT: m0
491
492; GCN-DAG: s_mov_b32 [[SBASE0:s[0-9]+]], bar.large@abs32@lo
493; GCN-DAG: s_add_i32 [[SBASE1:s[0-9]+]], [[SBASE0]], 0x4000{{$}}
494; GCN-DAG: s_addk_i32 [[SBASE0]], 0x7ff8{{$}}
495; GCN-DAG: v_mov_b32_e32 [[VBASE0:v[0-9]+]], [[SBASE0]]
496; GCN-DAG: v_mov_b32_e32 [[VBASE1:v[0-9]+]], [[SBASE1]]
497; GCN-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[VBASE0]] offset1:1
498; GCN-DAG: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, [[VBASE1]] offset1:1
499; GCN: s_endpgm
500define amdgpu_kernel void @load_misaligned64_constant_large_offsets(i64 addrspace(1)* %out) {
501  %val0 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 2048), align 4
502  %val1 = load i64, i64 addrspace(3)* getelementptr inbounds ([4096 x i64], [4096 x i64] addrspace(3)* @bar.large, i32 0, i32 4095), align 4
503  %sum = add i64 %val0, %val1
504  store i64 %sum, i64 addrspace(1)* %out, align 8
505  ret void
506}
507
508@sgemm.lA = internal unnamed_addr addrspace(3) global [264 x float] undef, align 4
509@sgemm.lB = internal unnamed_addr addrspace(3) global [776 x float] undef, align 4
510
511; GCN-LABEL: {{^}}sgemm_inner_loop_read2_sequence:
512; CI-DAG: s_mov_b32 m0
513; GFX9-NOT: m0
514
515define amdgpu_kernel void @sgemm_inner_loop_read2_sequence(float addrspace(1)* %C, i32 %lda, i32 %ldb) #0 {
516  %x.i = tail call i32 @llvm.amdgcn.workgroup.id.x() #1
517  %y.i = tail call i32 @llvm.amdgcn.workitem.id.y() #1
518  %arrayidx44 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %x.i
519  %tmp16 = load float, float addrspace(3)* %arrayidx44, align 4
520  %add47 = add nsw i32 %x.i, 1
521  %arrayidx48 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add47
522  %tmp17 = load float, float addrspace(3)* %arrayidx48, align 4
523  %add51 = add nsw i32 %x.i, 16
524  %arrayidx52 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add51
525  %tmp18 = load float, float addrspace(3)* %arrayidx52, align 4
526  %add55 = add nsw i32 %x.i, 17
527  %arrayidx56 = getelementptr inbounds [264 x float], [264 x float] addrspace(3)* @sgemm.lA, i32 0, i32 %add55
528  %tmp19 = load float, float addrspace(3)* %arrayidx56, align 4
529  %arrayidx60 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %y.i
530  %tmp20 = load float, float addrspace(3)* %arrayidx60, align 4
531  %add63 = add nsw i32 %y.i, 1
532  %arrayidx64 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add63
533  %tmp21 = load float, float addrspace(3)* %arrayidx64, align 4
534  %add67 = add nsw i32 %y.i, 32
535  %arrayidx68 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add67
536  %tmp22 = load float, float addrspace(3)* %arrayidx68, align 4
537  %add71 = add nsw i32 %y.i, 33
538  %arrayidx72 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add71
539  %tmp23 = load float, float addrspace(3)* %arrayidx72, align 4
540  %add75 = add nsw i32 %y.i, 64
541  %arrayidx76 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add75
542  %tmp24 = load float, float addrspace(3)* %arrayidx76, align 4
543  %add79 = add nsw i32 %y.i, 65
544  %arrayidx80 = getelementptr inbounds [776 x float], [776 x float] addrspace(3)* @sgemm.lB, i32 0, i32 %add79
545  %tmp25 = load float, float addrspace(3)* %arrayidx80, align 4
546  %sum.0 = fadd float %tmp16, %tmp17
547  %sum.1 = fadd float %sum.0, %tmp18
548  %sum.2 = fadd float %sum.1, %tmp19
549  %sum.3 = fadd float %sum.2, %tmp20
550  %sum.4 = fadd float %sum.3, %tmp21
551  %sum.5 = fadd float %sum.4, %tmp22
552  %sum.6 = fadd float %sum.5, %tmp23
553  %sum.7 = fadd float %sum.6, %tmp24
554  %sum.8 = fadd float %sum.7, %tmp25
555  store float %sum.8, float addrspace(1)* %C, align 4
556  ret void
557}
558
559; GCN-LABEL: {{^}}misaligned_read2_v2i32:
560; CI-DAG: s_mov_b32 m0
561; GFX9-NOT: m0
562define amdgpu_kernel void @misaligned_read2_v2i32(<2 x i32> addrspace(1)* %out, <2 x i32> addrspace(3)* %in) #0 {
563  %load = load <2 x i32>, <2 x i32> addrspace(3)* %in, align 4
564  store <2 x i32> %load, <2 x i32> addrspace(1)* %out, align 8
565  ret void
566}
567
568; GCN-LABEL: {{^}}misaligned_read2_i64:
569; CI-DAG: s_mov_b32 m0
570; GFX9-NOT: m0
571define amdgpu_kernel void @misaligned_read2_i64(i64 addrspace(1)* %out, i64 addrspace(3)* %in) #0 {
572  %load = load i64, i64 addrspace(3)* %in, align 4
573  store i64 %load, i64 addrspace(1)* %out, align 8
574  ret void
575}
576
577; GCN-LABEL: ds_read_diff_base_interleaving
578; CI-DAG: s_mov_b32 m0
579; GFX9-NOT: m0
580
581; GCN-NOT: ds_read_b32
582define amdgpu_kernel void @ds_read_diff_base_interleaving(
583  float addrspace(1)* nocapture %arg,
584  [4 x [4 x float]] addrspace(3)* %arg1,
585  [4 x [4 x float]] addrspace(3)* %arg2,
586  [4 x [4 x float]] addrspace(3)* %arg3,
587  [4 x [4 x float]] addrspace(3)* %arg4) #1 {
588bb:
589  %tmp = getelementptr float, float addrspace(1)* %arg, i64 10
590  %tmp5 = tail call i32 @llvm.amdgcn.workitem.id.x() #2
591  %tmp6 = tail call i32 @llvm.amdgcn.workitem.id.y() #2
592  %tmp7 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 0
593  %tmp8 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 0, i32 %tmp5
594  %tmp9 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 0
595  %tmp10 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 0, i32 %tmp5
596  %tmp11 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg1, i32 0, i32 %tmp6, i32 1
597  %tmp12 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg2, i32 0, i32 1, i32 %tmp5
598  %tmp13 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg3, i32 0, i32 %tmp6, i32 1
599  %tmp14 = getelementptr [4 x [4 x float]], [4 x [4 x float]] addrspace(3)* %arg4, i32 0, i32 1, i32 %tmp5
600  %tmp15 = load float, float addrspace(3)* %tmp7
601  %tmp16 = load float, float addrspace(3)* %tmp8
602  %tmp17 = fmul float %tmp15, %tmp16
603  %tmp18 = fadd float 2.000000e+00, %tmp17
604  %tmp19 = load float, float addrspace(3)* %tmp9
605  %tmp20 = load float, float addrspace(3)* %tmp10
606  %tmp21 = fmul float %tmp19, %tmp20
607  %tmp22 = fsub float %tmp18, %tmp21
608  %tmp23 = load float, float addrspace(3)* %tmp11
609  %tmp24 = load float, float addrspace(3)* %tmp12
610  %tmp25 = fmul float %tmp23, %tmp24
611  %tmp26 = fsub float %tmp22, %tmp25
612  %tmp27 = load float, float addrspace(3)* %tmp13
613  %tmp28 = load float, float addrspace(3)* %tmp14
614  %tmp29 = fmul float %tmp27, %tmp28
615  %tmp30 = fsub float %tmp26, %tmp29
616  store float %tmp30, float addrspace(1)* %tmp
617  ret void
618}
619
620; GCN-LABEL: ds_read_call_read:
621; GCN: ds_read_b32
622; GCN: s_swappc_b64
623; GCN: ds_read_b32
624define amdgpu_kernel void @ds_read_call_read(i32 addrspace(1)* %out, i32 addrspace(3)* %arg) {
625  %x = call i32 @llvm.amdgcn.workitem.id.x()
626  %arrayidx0 = getelementptr i32, i32 addrspace(3)* %arg, i32 %x
627  %arrayidx1 = getelementptr i32, i32 addrspace(3)* %arrayidx0, i32 1
628  %v0 = load i32, i32 addrspace(3)* %arrayidx0, align 4
629  call void @void_func_void()
630  %v1 = load i32, i32 addrspace(3)* %arrayidx1, align 4
631  %r = add i32 %v0, %v1
632  store i32 %r, i32 addrspace(1)* %out, align 4
633  ret void
634}
635
636; GCN-LABEL: {{^}}ds_read_interp_read:
637; CI: s_mov_b32 m0, -1
638; CI: ds_read_b32
639; CI: s_mov_b32 m0, s0
640; CI: v_interp_mov_f32
641; CI: s_mov_b32 m0, -1
642; CI: ds_read_b32
643; GFX9: ds_read2_b32 v[0:1], v0 offset1:4
644; GFX9: s_mov_b32 m0, s0
645; GFX9: v_interp_mov_f32
646define amdgpu_ps <2 x float> @ds_read_interp_read(i32 inreg %prims, float addrspace(3)* %inptr) {
647  %v0 = load float, float addrspace(3)* %inptr, align 4
648  %intrp = call float @llvm.amdgcn.interp.mov(i32 0, i32 0, i32 0, i32 %prims)
649  %ptr1 = getelementptr float, float addrspace(3)* %inptr, i32 4
650  %v1 = load float, float addrspace(3)* %ptr1, align 4
651  %v1b = fadd float %v1, %intrp
652  %r0 = insertelement <2 x float> undef, float %v0, i32 0
653  %r1 = insertelement <2 x float> %r0, float %v1b, i32 1
654  ret <2 x float> %r1
655}
656
657declare void @void_func_void() #3
658
659declare i32 @llvm.amdgcn.workgroup.id.x() #1
660declare i32 @llvm.amdgcn.workgroup.id.y() #1
661declare i32 @llvm.amdgcn.workitem.id.x() #1
662declare i32 @llvm.amdgcn.workitem.id.y() #1
663
664declare float @llvm.amdgcn.interp.mov(i32, i32, i32, i32) nounwind readnone
665
666declare void @llvm.amdgcn.s.barrier() #2
667
668attributes #0 = { nounwind }
669attributes #1 = { nounwind readnone speculatable }
670attributes #2 = { convergent nounwind }
671attributes #3 = { nounwind noinline }
672