1; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tahiti < %s | FileCheck -check-prefix=FAST64 -check-prefix=GCN %s
2; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=bonaire < %s | FileCheck -check-prefix=SLOW64 -check-prefix=GCN %s
3
4
5; lshr (i64 x), c: c > 32 => reg_sequence lshr (i32 hi_32(x)), (c - 32), 0
6; GCN-LABEL: {{^}}lshr_i64_35:
7; GCN-DAG: buffer_load_dword [[VAL:v[0-9]+]]
8; GCN-DAG: v_lshrrev_b32_e32 v[[LO:[0-9]+]], 3, [[VAL]]
9; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
10; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
11define amdgpu_kernel void @lshr_i64_35(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
12  %val = load i64, i64 addrspace(1)* %in
13  %shl = lshr i64 %val, 35
14  store i64 %shl, i64 addrspace(1)* %out
15  ret void
16}
17
18; GCN-LABEL: {{^}}lshr_i64_63:
19; GCN-DAG: buffer_load_dword [[VAL:v[0-9]+]]
20; GCN-DAG: v_lshrrev_b32_e32 v[[LO:[0-9]+]], 31, [[VAL]]
21; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
22; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
23define amdgpu_kernel void @lshr_i64_63(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
24  %val = load i64, i64 addrspace(1)* %in
25  %shl = lshr i64 %val, 63
26  store i64 %shl, i64 addrspace(1)* %out
27  ret void
28}
29
30; GCN-LABEL: {{^}}lshr_i64_33:
31; GCN-DAG: buffer_load_dword [[VAL:v[0-9]+]]
32; GCN-DAG: v_lshrrev_b32_e32 v[[LO:[0-9]+]], 1, [[VAL]]
33; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
34; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
35define amdgpu_kernel void @lshr_i64_33(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
36  %val = load i64, i64 addrspace(1)* %in
37  %shl = lshr i64 %val, 33
38  store i64 %shl, i64 addrspace(1)* %out
39  ret void
40}
41
42; GCN-LABEL: {{^}}lshr_i64_32:
43; GCN-DAG: buffer_load_dword v[[LO:[0-9]+]]
44; GCN-DAG: v_mov_b32_e32 v[[HI:[0-9]+]], 0{{$}}
45; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
46define amdgpu_kernel void @lshr_i64_32(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
47  %val = load i64, i64 addrspace(1)* %in
48  %shl = lshr i64 %val, 32
49  store i64 %shl, i64 addrspace(1)* %out
50  ret void
51}
52
53; Make sure the and of the constant doesn't prevent bfe from forming
54; after 64-bit shift is split.
55
56; GCN-LABEL: {{^}}lshr_and_i64_35:
57; GCN-DAG: v_mov_b32_e32 v[[ZERO:[0-9]+]], 0{{$}}
58; GCN-DAG: buffer_load_dword v[[LO:[0-9]+]]
59; GCN: v_bfe_u32 v[[BFE:[0-9]+]], v[[LO]], 8, 23
60; GCN: buffer_store_dwordx2 v{{\[}}[[BFE]]:[[ZERO]]{{\]}}
61define amdgpu_kernel void @lshr_and_i64_35(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
62  %val = load i64, i64 addrspace(1)* %in
63  %and = and i64 %val, 9223372036854775807 ; 0x7fffffffffffffff
64  %shl = lshr i64 %and, 40
65  store i64 %shl, i64 addrspace(1)* %out
66  ret void
67}
68
69; lshl (i64 x), c: c > 32 => reg_sequence lshl 0, (i32 lo_32(x)), (c - 32)
70
71; GCN-LABEL: {{^}}shl_i64_const_35:
72; GCN: buffer_load_dword [[VAL:v[0-9]+]]
73; GCN: v_lshlrev_b32_e32 v[[HI:[0-9]+]], 3, [[VAL]]
74; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
75; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
76define amdgpu_kernel void @shl_i64_const_35(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
77  %val = load i64, i64 addrspace(1)* %in
78  %shl = shl i64 %val, 35
79  store i64 %shl, i64 addrspace(1)* %out
80  ret void
81}
82
83; GCN-LABEL: {{^}}shl_i64_const_32:
84; GCN-DAG: buffer_load_dword v[[HI:[0-9]+]]
85; GCN-DAG: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
86; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
87define amdgpu_kernel void @shl_i64_const_32(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
88  %val = load i64, i64 addrspace(1)* %in
89  %shl = shl i64 %val, 32
90  store i64 %shl, i64 addrspace(1)* %out
91  ret void
92}
93
94; GCN-LABEL: {{^}}shl_i64_const_63:
95; GCN: buffer_load_dword [[VAL:v[0-9]+]]
96; GCN: v_lshlrev_b32_e32 v[[HI:[0-9]+]], 31, [[VAL]]
97; GCN: v_mov_b32_e32 v[[LO:[0-9]+]], 0{{$}}
98; GCN: buffer_store_dwordx2 v{{\[}}[[LO]]:[[HI]]{{\]}}
99define amdgpu_kernel void @shl_i64_const_63(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
100  %val = load i64, i64 addrspace(1)* %in
101  %shl = shl i64 %val, 63
102  store i64 %shl, i64 addrspace(1)* %out
103  ret void
104}
105
106; ashr (i64 x), 63 => (ashr lo(x), 31), lo(x)
107
108; GCN-LABEL: {{^}}ashr_i64_const_32:
109define amdgpu_kernel void @ashr_i64_const_32(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
110  %val = load i64, i64 addrspace(1)* %in
111  %shl = ashr i64 %val, 32
112  store i64 %shl, i64 addrspace(1)* %out
113  ret void
114}
115
116; GCN-LABEL: {{^}}ashr_i64_const_63:
117define amdgpu_kernel void @ashr_i64_const_63(i64 addrspace(1)* %out, i64 addrspace(1)* %in) {
118  %val = load i64, i64 addrspace(1)* %in
119  %shl = ashr i64 %val, 63
120  store i64 %shl, i64 addrspace(1)* %out
121  ret void
122}
123
124; GCN-LABEL: {{^}}trunc_shl_31_i32_i64:
125; GCN: buffer_load_dword [[VAL:v[0-9]+]]
126; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 31, [[VAL]]
127; GCN: buffer_store_dword [[SHL]]
128define amdgpu_kernel void @trunc_shl_31_i32_i64(i32 addrspace(1)* %out, i64 addrspace(1)* %in) {
129  %val = load i64, i64 addrspace(1)* %in
130  %shl = shl i64 %val, 31
131  %trunc = trunc i64 %shl to i32
132  store i32 %trunc, i32 addrspace(1)* %out
133  ret void
134}
135
136; GCN-LABEL: {{^}}trunc_shl_15_i16_i64:
137; GCN: buffer_load_dword [[VAL:v[0-9]+]]
138; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 15, [[VAL]]
139; GCN: buffer_store_short [[SHL]]
140define amdgpu_kernel void @trunc_shl_15_i16_i64(i16 addrspace(1)* %out, i64 addrspace(1)* %in) {
141  %val = load i64, i64 addrspace(1)* %in
142  %shl = shl i64 %val, 15
143  %trunc = trunc i64 %shl to i16
144  store i16 %trunc, i16 addrspace(1)* %out
145  ret void
146}
147
148; GCN-LABEL: {{^}}trunc_shl_15_i16_i32:
149; GCN: buffer_load_dword [[VAL:v[0-9]+]]
150; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 15, [[VAL]]
151; GCN: buffer_store_short [[SHL]]
152define amdgpu_kernel void @trunc_shl_15_i16_i32(i16 addrspace(1)* %out, i32 addrspace(1)* %in) {
153  %val = load i32, i32 addrspace(1)* %in
154  %shl = shl i32 %val, 15
155  %trunc = trunc i32 %shl to i16
156  store i16 %trunc, i16 addrspace(1)* %out
157  ret void
158}
159
160; GCN-LABEL: {{^}}trunc_shl_7_i8_i64:
161; GCN: buffer_load_dword [[VAL:v[0-9]+]]
162; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 7, [[VAL]]
163; GCN: buffer_store_byte [[SHL]]
164define amdgpu_kernel void @trunc_shl_7_i8_i64(i8 addrspace(1)* %out, i64 addrspace(1)* %in) {
165  %val = load i64, i64 addrspace(1)* %in
166  %shl = shl i64 %val, 7
167  %trunc = trunc i64 %shl to i8
168  store i8 %trunc, i8 addrspace(1)* %out
169  ret void
170}
171
172; GCN-LABEL: {{^}}trunc_shl_1_i2_i64:
173; GCN: buffer_load_dword [[VAL:v[0-9]+]]
174; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 1, [[VAL]]
175; GCN: v_and_b32_e32 [[AND:v[0-9]+]], 2, [[SHL]]
176; GCN: buffer_store_byte [[AND]]
177define amdgpu_kernel void @trunc_shl_1_i2_i64(i2 addrspace(1)* %out, i64 addrspace(1)* %in) {
178  %val = load i64, i64 addrspace(1)* %in
179  %shl = shl i64 %val, 1
180  %trunc = trunc i64 %shl to i2
181  store i2 %trunc, i2 addrspace(1)* %out
182  ret void
183}
184
185; GCN-LABEL: {{^}}trunc_shl_1_i32_i64:
186; GCN: buffer_load_dword [[VAL:v[0-9]+]]
187; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 1, [[VAL]]
188; GCN: buffer_store_dword [[SHL]]
189define amdgpu_kernel void @trunc_shl_1_i32_i64(i32 addrspace(1)* %out, i64 addrspace(1)* %in) {
190  %val = load i64, i64 addrspace(1)* %in
191  %shl = shl i64 %val, 1
192  %trunc = trunc i64 %shl to i32
193  store i32 %trunc, i32 addrspace(1)* %out
194  ret void
195}
196
197; GCN-LABEL: {{^}}trunc_shl_16_i32_i64:
198; GCN: buffer_load_dword [[VAL:v[0-9]+]]
199; GCN: v_lshlrev_b32_e32 [[SHL:v[0-9]+]], 16, [[VAL]]
200; GCN: buffer_store_dword [[SHL]]
201define amdgpu_kernel void @trunc_shl_16_i32_i64(i32 addrspace(1)* %out, i64 addrspace(1)* %in) {
202  %val = load i64, i64 addrspace(1)* %in
203  %shl = shl i64 %val, 16
204  %trunc = trunc i64 %shl to i32
205  store i32 %trunc, i32 addrspace(1)* %out
206  ret void
207}
208
209; GCN-LABEL: {{^}}trunc_shl_33_i32_i64:
210; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
211; GCN: buffer_store_dword [[ZERO]]
212define amdgpu_kernel void @trunc_shl_33_i32_i64(i32 addrspace(1)* %out, i64 addrspace(1)* %in) {
213  %val = load i64, i64 addrspace(1)* %in
214  %shl = shl i64 %val, 33
215  %trunc = trunc i64 %shl to i32
216  store i32 %trunc, i32 addrspace(1)* %out
217  ret void
218}
219
220; GCN-LABEL: {{^}}trunc_shl_16_v2i32_v2i64:
221; GCN: buffer_load_dwordx4 v{{\[}}[[LO:[0-9]+]]:[[HI:[0-9]+]]{{\]}}
222; GCN-DAG: v_lshlrev_b32_e32 v[[RESHI:[0-9]+]], 16, v{{[0-9]+}}
223; GCN-DAG: v_lshlrev_b32_e32 v[[RESLO:[0-9]+]], 16, v[[LO]]
224; GCN: buffer_store_dwordx2 v{{\[}}[[RESLO]]:[[RESHI]]{{\]}}
225define amdgpu_kernel void @trunc_shl_16_v2i32_v2i64(<2 x i32> addrspace(1)* %out, <2 x i64> addrspace(1)* %in) {
226  %val = load <2 x i64>, <2 x i64> addrspace(1)* %in
227  %shl = shl <2 x i64> %val, <i64 16, i64 16>
228  %trunc = trunc <2 x i64> %shl to <2 x i32>
229  store <2 x i32> %trunc, <2 x i32> addrspace(1)* %out
230  ret void
231}
232
233; GCN-LABEL: {{^}}trunc_shl_31_i32_i64_multi_use:
234; GCN: buffer_load_dwordx2 [[VAL:v\[[0-9]+:[0-9]+\]]]
235; GCN: v_lshl_b64 v{{\[}}[[RESLO:[0-9]+]]:[[RESHI:[0-9]+]]{{\]}}, [[VAL]], 31
236; GCN: buffer_store_dword v[[RESLO]]
237; GCN: buffer_store_dwordx2 v{{\[}}[[RESLO]]:[[RESHI]]{{\]}}
238define amdgpu_kernel void @trunc_shl_31_i32_i64_multi_use(i32 addrspace(1)* %out, i64 addrspace(1)* %in) {
239  %val = load i64, i64 addrspace(1)* %in
240  %shl = shl i64 %val, 31
241  %trunc = trunc i64 %shl to i32
242  store volatile i32 %trunc, i32 addrspace(1)* %out
243  store volatile i64 %shl, i64 addrspace(1)* %in
244  ret void
245}
246
247; GCN-LABEL: {{^}}trunc_shl_and31:
248; GCN:     s_and_b32 s[[AMT:[0-9]+]], s{{[0-9]+}}, 31
249; GCN:     v_lshlrev_b32_e32 v{{[0-9]+}}, s[[AMT]], v{{[0-9]+}}
250; GCN-NOT: v_lshl_b64
251; GCN-NOT: v_lshlrev_b64
252define amdgpu_kernel void @trunc_shl_and31(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1, i32 %arg2) {
253bb:
254  %tmp = load i64, i64 addrspace(1)* %arg, align 8
255  %tmp3 = and i32 %arg2, 31
256  %tmp4 = zext i32 %tmp3 to i64
257  %tmp5 = shl i64 %tmp, %tmp4
258  %tmp6 = trunc i64 %tmp5 to i32
259  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
260  ret void
261}
262
263; GCN-LABEL: {{^}}trunc_shl_and30:
264; GCN:     s_and_b32 s[[AMT:[0-9]+]], s{{[0-9]+}}, 30
265; GCN:     v_lshlrev_b32_e32 v{{[0-9]+}}, s[[AMT]], v{{[0-9]+}}
266; GCN-NOT: v_lshl_b64
267; GCN-NOT: v_lshlrev_b64
268define amdgpu_kernel void @trunc_shl_and30(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1, i32 %arg2) {
269bb:
270  %tmp = load i64, i64 addrspace(1)* %arg, align 8
271  %tmp3 = and i32 %arg2, 30
272  %tmp4 = zext i32 %tmp3 to i64
273  %tmp5 = shl i64 %tmp, %tmp4
274  %tmp6 = trunc i64 %tmp5 to i32
275  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
276  ret void
277}
278
279; GCN-LABEL: {{^}}trunc_shl_wrong_and63:
280; Negative test, wrong constant
281; GCN: v_lshl_b64
282define amdgpu_kernel void @trunc_shl_wrong_and63(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1, i32 %arg2) {
283bb:
284  %tmp = load i64, i64 addrspace(1)* %arg, align 8
285  %tmp3 = and i32 %arg2, 63
286  %tmp4 = zext i32 %tmp3 to i64
287  %tmp5 = shl i64 %tmp, %tmp4
288  %tmp6 = trunc i64 %tmp5 to i32
289  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
290  ret void
291}
292
293; GCN-LABEL: {{^}}trunc_shl_no_and:
294; Negative test, shift can be full 64 bit
295; GCN: v_lshl_b64
296define amdgpu_kernel void @trunc_shl_no_and(i64 addrspace(1)* nocapture readonly %arg, i32 addrspace(1)* nocapture %arg1, i32 %arg2) {
297bb:
298  %tmp = load i64, i64 addrspace(1)* %arg, align 8
299  %tmp4 = zext i32 %arg2 to i64
300  %tmp5 = shl i64 %tmp, %tmp4
301  %tmp6 = trunc i64 %tmp5 to i32
302  store i32 %tmp6, i32 addrspace(1)* %arg1, align 4
303  ret void
304}
305
306; GCN-LABEL: {{^}}trunc_shl_vec_vec:
307; GCN-DAG: v_lshl_b64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], 3
308; GCN-DAG: v_lshl_b64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], 4
309; GCN-DAG: v_lshl_b64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], 5
310; GCN-DAG: v_lshl_b64 v[{{[0-9:]+}}], v[{{[0-9:]+}}], 6
311define amdgpu_kernel void @trunc_shl_vec_vec(<4 x i64> addrspace(1)* %arg) {
312bb:
313  %v = load <4 x i64>, <4 x i64> addrspace(1)* %arg, align 32
314  %shl = shl <4 x i64> %v, <i64 3, i64 4, i64 5, i64 6>
315  store <4 x i64> %shl, <4 x i64> addrspace(1)* %arg, align 32
316  ret void
317}
318