1 ; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -mattr=+load-store-opt < %s | FileCheck -check-prefix=SI %s
3 @lds = addrspace(3) global [512 x float] undef, align 4
4 @lds.f64 = addrspace(3) global [512 x double] undef, align 8
7 ; SI-LABEL: @simple_read2st64_f32_0_1
8 ; SI: ds_read2st64_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:1
9 ; SI: s_waitcnt lgkmcnt(0)
10 ; SI: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[LO_VREG]], v[[HI_VREG]]
11 ; SI: buffer_store_dword [[RESULT]]
13 define amdgpu_kernel void @simple_read2st64_f32_0_1(float addrspace(1)* %out) #0 {
14 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
15 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
16 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
17 %add.x = add nsw i32 %x.i, 64
18 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
19 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
20 %sum = fadd float %val0, %val1
21 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
22 store float %sum, float addrspace(1)* %out.gep, align 4
26 ; SI-LABEL: @simple_read2st64_f32_1_2
27 ; SI: ds_read2st64_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1 offset1:2
28 ; SI: s_waitcnt lgkmcnt(0)
29 ; SI: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[LO_VREG]], v[[HI_VREG]]
30 ; SI: buffer_store_dword [[RESULT]]
32 define amdgpu_kernel void @simple_read2st64_f32_1_2(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
33 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
34 %add.x.0 = add nsw i32 %x.i, 64
35 %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.0
36 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
37 %add.x.1 = add nsw i32 %x.i, 128
38 %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.1
39 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
40 %sum = fadd float %val0, %val1
41 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
42 store float %sum, float addrspace(1)* %out.gep, align 4
46 ; SI-LABEL: @simple_read2st64_f32_max_offset
47 ; SI: ds_read2st64_b32 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1 offset1:255
48 ; SI: s_waitcnt lgkmcnt(0)
49 ; SI: v_add_f32_e32 [[RESULT:v[0-9]+]], v[[LO_VREG]], v[[HI_VREG]]
50 ; SI: buffer_store_dword [[RESULT]]
52 define amdgpu_kernel void @simple_read2st64_f32_max_offset(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
53 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
54 %add.x.0 = add nsw i32 %x.i, 64
55 %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.0
56 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
57 %add.x.1 = add nsw i32 %x.i, 16320
58 %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.1
59 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
60 %sum = fadd float %val0, %val1
61 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
62 store float %sum, float addrspace(1)* %out.gep, align 4
66 ; SI-LABEL: @simple_read2st64_f32_over_max_offset
67 ; SI-NOT: ds_read2st64_b32
68 ; SI-DAG: v_add_i32_e32 [[BIGADD:v[0-9]+]], vcc, 0x10000, {{v[0-9]+}}
69 ; SI-DAG: ds_read_b32 {{v[0-9]+}}, {{v[0-9]+}} offset:256
70 ; SI-DAG: ds_read_b32 {{v[0-9]+}}, [[BIGADD]]{{$}}
72 define amdgpu_kernel void @simple_read2st64_f32_over_max_offset(float addrspace(1)* %out, float addrspace(3)* %lds) #0 {
73 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
74 %add.x.0 = add nsw i32 %x.i, 64
75 %arrayidx0 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.0
76 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
77 %add.x.1 = add nsw i32 %x.i, 16384
78 %arrayidx1 = getelementptr inbounds float, float addrspace(3)* %lds, i32 %add.x.1
79 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
80 %sum = fadd float %val0, %val1
81 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
82 store float %sum, float addrspace(1)* %out.gep, align 4
86 ; SI-LABEL: @odd_invalid_read2st64_f32_0
87 ; SI-NOT: ds_read2st64_b32
89 define amdgpu_kernel void @odd_invalid_read2st64_f32_0(float addrspace(1)* %out) #0 {
90 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
91 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %x.i
92 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
93 %add.x = add nsw i32 %x.i, 63
94 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x
95 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
96 %sum = fadd float %val0, %val1
97 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
98 store float %sum, float addrspace(1)* %out.gep, align 4
102 ; SI-LABEL: @odd_invalid_read2st64_f32_1
103 ; SI-NOT: ds_read2st64_b32
105 define amdgpu_kernel void @odd_invalid_read2st64_f32_1(float addrspace(1)* %out) #0 {
106 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
107 %add.x.0 = add nsw i32 %x.i, 64
108 %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x.0
109 %val0 = load float, float addrspace(3)* %arrayidx0, align 4
110 %add.x.1 = add nsw i32 %x.i, 127
111 %arrayidx1 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds, i32 0, i32 %add.x.1
112 %val1 = load float, float addrspace(3)* %arrayidx1, align 4
113 %sum = fadd float %val0, %val1
114 %out.gep = getelementptr inbounds float, float addrspace(1)* %out, i32 %x.i
115 store float %sum, float addrspace(1)* %out.gep, align 4
119 ; SI-LABEL: @simple_read2st64_f64_0_1
120 ; SI: ds_read2st64_b64 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset1:1
121 ; SI: s_waitcnt lgkmcnt(0)
122 ; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[LO_VREG]]:{{[0-9]+\]}}, v{{\[[0-9]+}}:[[HI_VREG]]{{\]}}
123 ; SI: buffer_store_dwordx2 [[RESULT]]
125 define amdgpu_kernel void @simple_read2st64_f64_0_1(double addrspace(1)* %out) #0 {
126 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
127 %arrayidx0 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %x.i
128 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
129 %add.x = add nsw i32 %x.i, 64
130 %arrayidx1 = getelementptr inbounds [512 x double], [512 x double] addrspace(3)* @lds.f64, i32 0, i32 %add.x
131 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
132 %sum = fadd double %val0, %val1
133 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
134 store double %sum, double addrspace(1)* %out.gep, align 8
138 ; SI-LABEL: @simple_read2st64_f64_1_2
139 ; SI: ds_read2st64_b64 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:1 offset1:2
140 ; SI: s_waitcnt lgkmcnt(0)
141 ; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[LO_VREG]]:{{[0-9]+\]}}, v{{\[[0-9]+}}:[[HI_VREG]]{{\]}}
142 ; SI: buffer_store_dwordx2 [[RESULT]]
144 define amdgpu_kernel void @simple_read2st64_f64_1_2(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
145 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
146 %add.x.0 = add nsw i32 %x.i, 64
147 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.0
148 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
149 %add.x.1 = add nsw i32 %x.i, 128
150 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.1
151 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
152 %sum = fadd double %val0, %val1
153 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
154 store double %sum, double addrspace(1)* %out.gep, align 8
160 ; SI-LABEL: @misaligned_read2st64_f64
161 ; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset1:1
162 ; SI: ds_read2_b32 v{{\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset0:128 offset1:129
164 define amdgpu_kernel void @misaligned_read2st64_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
165 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
166 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
167 %val0 = load double, double addrspace(3)* %arrayidx0, align 4
168 %add.x = add nsw i32 %x.i, 64
169 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
170 %val1 = load double, double addrspace(3)* %arrayidx1, align 4
171 %sum = fadd double %val0, %val1
172 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
173 store double %sum, double addrspace(1)* %out.gep, align 4
177 ; The maximum is not the usual 0xff because 0xff * 8 * 64 > 0xffff
178 ; SI-LABEL: @simple_read2st64_f64_max_offset
179 ; SI: ds_read2st64_b64 v{{\[}}[[LO_VREG:[0-9]+]]:[[HI_VREG:[0-9]+]]{{\]}}, v{{[0-9]+}} offset0:4 offset1:127
180 ; SI: s_waitcnt lgkmcnt(0)
181 ; SI: v_add_f64 [[RESULT:v\[[0-9]+:[0-9]+\]]], v{{\[}}[[LO_VREG]]:{{[0-9]+\]}}, v{{\[[0-9]+}}:[[HI_VREG]]{{\]}}
182 ; SI: buffer_store_dwordx2 [[RESULT]]
184 define amdgpu_kernel void @simple_read2st64_f64_max_offset(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
185 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
186 %add.x.0 = add nsw i32 %x.i, 256
187 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.0
188 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
189 %add.x.1 = add nsw i32 %x.i, 8128
190 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.1
191 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
192 %sum = fadd double %val0, %val1
193 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
194 store double %sum, double addrspace(1)* %out.gep, align 8
198 ; SI-LABEL: @simple_read2st64_f64_over_max_offset
199 ; SI-NOT: ds_read2st64_b64
200 ; SI-DAG: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, {{v[0-9]+}} offset:512
201 ; SI-DAG: v_add_i32_e32 [[BIGADD:v[0-9]+]], vcc, 0x10000, {{v[0-9]+}}
202 ; SI: ds_read_b64 {{v\[[0-9]+:[0-9]+\]}}, [[BIGADD]]
204 define amdgpu_kernel void @simple_read2st64_f64_over_max_offset(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
205 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
206 %add.x.0 = add nsw i32 %x.i, 64
207 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.0
208 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
209 %add.x.1 = add nsw i32 %x.i, 8192
210 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.1
211 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
212 %sum = fadd double %val0, %val1
213 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
214 store double %sum, double addrspace(1)* %out.gep, align 8
218 ; SI-LABEL: @invalid_read2st64_f64_odd_offset
219 ; SI-NOT: ds_read2st64_b64
221 define amdgpu_kernel void @invalid_read2st64_f64_odd_offset(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
222 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
223 %add.x.0 = add nsw i32 %x.i, 64
224 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.0
225 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
226 %add.x.1 = add nsw i32 %x.i, 8129
227 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x.1
228 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
229 %sum = fadd double %val0, %val1
230 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
231 store double %sum, double addrspace(1)* %out.gep, align 8
235 ; The stride of 8 elements is 8 * 8 bytes. We need to make sure the
236 ; stride in elements, not bytes, is a multiple of 64.
238 ; SI-LABEL: @byte_size_only_divisible_64_read2_f64
239 ; SI-NOT: ds_read2st_b64
240 ; SI: ds_read2_b64 v{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}} offset1:8
242 define amdgpu_kernel void @byte_size_only_divisible_64_read2_f64(double addrspace(1)* %out, double addrspace(3)* %lds) #0 {
243 %x.i = tail call i32 @llvm.amdgcn.workitem.id.x() #1
244 %arrayidx0 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %x.i
245 %val0 = load double, double addrspace(3)* %arrayidx0, align 8
246 %add.x = add nsw i32 %x.i, 8
247 %arrayidx1 = getelementptr inbounds double, double addrspace(3)* %lds, i32 %add.x
248 %val1 = load double, double addrspace(3)* %arrayidx1, align 8
249 %sum = fadd double %val0, %val1
250 %out.gep = getelementptr inbounds double, double addrspace(1)* %out, i32 %x.i
251 store double %sum, double addrspace(1)* %out.gep, align 4
255 ; Function Attrs: nounwind readnone
256 declare i32 @llvm.amdgcn.workitem.id.x() #1
258 ; Function Attrs: nounwind readnone
259 declare i32 @llvm.amdgcn.workitem.id.y() #1
261 attributes #0 = { nounwind }
262 attributes #1 = { nounwind readnone }