1; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -stress-regalloc=8 < %s | FileCheck -check-prefixes=GCN,MUBUF %s
2; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 -enable-misched=0 -post-RA-scheduler=0 -stress-regalloc=8 -amdgpu-enable-flat-scratch < %s | FileCheck -check-prefixes=GCN,FLATSCR %s
3
4; Test that the VGPR spiller correctly switches to SGPR offsets when the
5; instruction offset field would overflow, and that it accounts for memory
6; swizzling.
7
8; GCN-LABEL: test_inst_offset_kernel
9define amdgpu_kernel void @test_inst_offset_kernel() {
10entry:
11  ; Occupy 4092 bytes of scratch, so the offset of the spill of %a just fits in
12  ; the instruction offset field.
13  %alloca = alloca i8, i32 4088, align 4, addrspace(5)
14  %buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
15
16  %aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
17  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4092 ; 4-byte Folded Spill
18  ; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s{{[0-9]+}} ; 4-byte Folded Spill
19  %a = load volatile i32, i32 addrspace(5)* %aptr
20
21  ; Force %a to spill.
22  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
23
24  %outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
25  store volatile i32 %a, i32 addrspace(5)* %outptr
26
27  ret void
28}
29
30; GCN-LABEL: test_sgpr_offset_kernel
31define amdgpu_kernel void @test_sgpr_offset_kernel() {
32entry:
33  ; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
34  ; fit in the instruction, and has to live in the SGPR offset.
35  %alloca = alloca i8, i32 4092, align 4, addrspace(5)
36  %buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
37
38  %aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
39  ; 0x40000 / 64 = 4096 (for wave64)
40  ; MUBUF:   s_mov_b32 s4, 0x40000
41  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
42  ; FLATSCR: s_movk_i32 s2, 0x1000
43  ; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s2 ; 4-byte Folded Spill
44  %a = load volatile i32, i32 addrspace(5)* %aptr
45
46  ; Force %a to spill
47  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
48
49  %outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
50  store volatile i32 %a, i32 addrspace(5)* %outptr
51
52  ret void
53}
54
55; FIXME: If we fail to scavenge an SGPR in a kernel we don't have a stack
56; pointer to temporarily update, so we just crash.
57
58; GCN-LABEL: test_sgpr_offset_function_scavenge_fail
59define void @test_sgpr_offset_function_scavenge_fail() #2 {
60entry:
61  ; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
62  ; fit in the instruction, and has to live in the SGPR offset.
63  %alloca = alloca i8, i32 4096, align 4, addrspace(5)
64  %buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
65
66  %aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
67
68  %asm.0 = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
69  %asm0.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 0
70  %asm1.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 1
71  %asm2.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 2
72  %asm3.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 3
73  %asm4.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 4
74  %asm5.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 5
75  %asm6.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 6
76  %asm7.0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm.0, 7
77
78  ; 0x40000 / 64 = 4096 (for wave64)
79  %a = load volatile i32, i32 addrspace(5)* %aptr
80
81  ; MUBUF:   s_add_i32 s32, s32, 0x40100
82  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s32 ; 4-byte Folded Spill
83  ; MUBUF:   s_add_i32 s32, s32, 0xfffbff00
84  ; FLATSCR: s_add_i32 [[SOFF:s[0-9]+]], s32, 0x1004
85  ; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, [[SOFF]] ; 4-byte Folded Spill
86  call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0.0, i32 %asm1.0, i32 %asm2.0, i32 %asm3.0, i32 %asm4.0, i32 %asm5.0, i32 %asm6.0, i32 %asm7.0, i32 %a)
87
88  %asm = call { i32, i32, i32, i32, i32, i32, i32, i32 } asm sideeffect "", "=s,=s,=s,=s,=s,=s,=s,=s"()
89  %asm0 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 0
90  %asm1 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 1
91  %asm2 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 2
92  %asm3 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 3
93  %asm4 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 4
94  %asm5 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 5
95  %asm6 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 6
96  %asm7 = extractvalue { i32, i32, i32, i32, i32, i32, i32, i32 } %asm, 7
97
98  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}"() #0
99
100  ; MUBUF:   s_add_i32 s32, s32, 0x40100
101  ; MUBUF:   buffer_load_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s32 ; 4-byte Folded Reload
102  ; MUBUF:   s_add_i32 s32, s32, 0xfffbff00
103  ; FLATSCR: s_add_i32 [[SOFF:s[0-9]+]], s32, 0x1004
104  ; FLATSCR: scratch_load_dword v{{[0-9]+}}, off, [[SOFF]] ; 4-byte Folded Reload
105
106   ; Force %a to spill with no free SGPRs
107  call void asm sideeffect "", "s,s,s,s,s,s,s,s,v"(i32 %asm0, i32 %asm1, i32 %asm2, i32 %asm3, i32 %asm4, i32 %asm5, i32 %asm6, i32 %asm7, i32 %a)
108  ret void
109}
110
111; GCN-LABEL: test_sgpr_offset_subregs_kernel
112define amdgpu_kernel void @test_sgpr_offset_subregs_kernel() {
113entry:
114  ; Occupy 4088 bytes of scratch, so that the spill of the last subreg of %a
115  ; still fits below offset 4096 (4088 + 8 - 4 = 4092), and can be placed in
116  ; the instruction offset field.
117  %alloca = alloca i8, i32 4084, align 4, addrspace(5)
118  %bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
119  %bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
120
121  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4088 ; 4-byte Folded Spill
122  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], 0 offset:4092 ; 4-byte Folded Spill
123  ; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0xff8
124  ; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], [[SOFF]]          ; 8-byte Folded Spill
125  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
126  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
127
128  ; Force %a to spill.
129  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
130
131  ; Ensure the alloca sticks around.
132  %bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
133  %b = load volatile i32, i32 addrspace(5)* %bptr
134
135  ; Ensure the spill is of the full super-reg.
136  call void asm sideeffect "; $0", "r"(<2 x i32> %a)
137
138  ret void
139}
140
141; GCN-LABEL: test_inst_offset_subregs_kernel
142define amdgpu_kernel void @test_inst_offset_subregs_kernel() {
143entry:
144  ; Occupy 4092 bytes of scratch, so that the spill of the last subreg of %a
145  ; does not fit below offset 4096 (4092 + 8 - 4 = 4096), and has to live
146  ; in the SGPR offset.
147  %alloca = alloca i8, i32 4088, align 4, addrspace(5)
148  %bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
149  %bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
150
151  ; 0x3ff00 / 64 = 4092 (for wave64)
152  ; MUBUF:   s_mov_b32 s4, 0x3ff00
153  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
154  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 offset:4 ; 4-byte Folded Spill
155  ; FLATSCR: s_movk_i32 [[SOFF:s[0-9]+]], 0xffc
156  ; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], [[SOFF]]          ; 8-byte Folded Spill
157  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
158  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
159
160  ; Force %a to spill.
161  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
162
163  ; Ensure the alloca sticks around.
164  %bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
165  %b = load volatile i32, i32 addrspace(5)* %bptr
166
167  ; Ensure the spill is of the full super-reg.
168  call void asm sideeffect "; $0", "r"(<2 x i32> %a)
169
170  ret void
171}
172
173; GCN-LABEL: test_inst_offset_function
174define void @test_inst_offset_function() {
175entry:
176  ; Occupy enough bytes of scratch, so the offset of the spill of %a
177  ; just fits in the instruction offset field when the emergency stack
178  ; slot is added. It's hard to hit the actual limit since we're also
179  ; going to insert the emergency stack slot for large frames.
180  %alloca = alloca i8, i32 4088, align 4, addrspace(5)
181  %buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
182
183  %aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
184  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
185  ; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
186  %a = load volatile i32, i32 addrspace(5)* %aptr
187
188  ; Force %a to spill.
189  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
190
191  %outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
192  store volatile i32 %a, i32 addrspace(5)* %outptr
193
194  ret void
195}
196
197; GCN-LABEL: test_sgpr_offset_function
198define void @test_sgpr_offset_function() {
199entry:
200  ; Occupy 4096 bytes of scratch, so the offset of the spill of %a does not
201  ; fit in the instruction, and has to live in the SGPR offset.
202  %alloca = alloca i8, i32 4096, align 4, addrspace(5)
203  %buf = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
204
205  %aptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
206  ; 0x40000 / 64 = 4096 (for wave64)
207  ; MUBUF:   s_add_i32 s4, s32, 0x40100
208  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
209  ; FLATSCR: s_add_i32 s0, s32, 0x1004
210  ; FLATSCR: scratch_store_dword off, v{{[0-9]+}}, s0 ; 4-byte Folded Spill
211  %a = load volatile i32, i32 addrspace(5)* %aptr
212
213  ; Force %a to spill
214  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
215
216  %outptr = getelementptr i32, i32 addrspace(5)* %buf, i32 1
217  store volatile i32 %a, i32 addrspace(5)* %outptr
218
219  ret void
220}
221
222; GCN-LABEL: test_sgpr_offset_subregs_function
223define void @test_sgpr_offset_subregs_function() {
224entry:
225  ; We want to test the spill of the last subreg of %a is the highest
226  ; valid value for the immediate offset. We enable the emergency
227  ; stack slot for large frames, so it's hard to get the frame layout
228  ; exactly as we want to test it.
229  ;
230  ; Occupy 4084 bytes of scratch, so that the spill of the last subreg of %a
231  ; still fits below offset 4096 (4084 + 8 - 4 = 4092), and can be placed in
232  ; the instruction offset field.
233  %alloca = alloca i8, i32 4084, align 4, addrspace(5)
234  %bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
235  %bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
236
237  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4084 ; 4-byte Folded Spill
238  ; MUBUF:   buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}} offset:4088 ; 4-byte Folded Spill
239  ; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], s32 offset:4084 ; 8-byte Folded Spill
240  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
241  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
242
243  ; Force %a to spill.
244  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
245
246  ; Ensure the alloca sticks around.
247  %bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
248  %b = load volatile i32, i32 addrspace(5)* %bptr
249
250  ; Ensure the spill is of the full super-reg.
251  call void asm sideeffect "; $0", "r"(<2 x i32> %a)
252
253  ret void
254}
255
256; GCN-LABEL: test_inst_offset_subregs_function
257define void @test_inst_offset_subregs_function() {
258entry:
259  ; Occupy 4088 bytes of scratch, so that the spill of the last subreg of %a
260  ; does not fit below offset 4096 (408 + 4 + 8 - 4 = 4096), and has to live
261  ; in the SGPR offset.
262  %alloca = alloca i8, i32 4088, align 4, addrspace(5)
263  %bufv1 = bitcast i8 addrspace(5)* %alloca to i32 addrspace(5)*
264  %bufv2 = bitcast i8 addrspace(5)* %alloca to <2 x i32> addrspace(5)*
265
266  ; 0x3ff0000 / 64 = 4092 (for wave64)
267  ; MUBUF: s_add_i32 s4, s32, 0x3ff00
268  ; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 ; 4-byte Folded Spill
269  ; MUBUF: buffer_store_dword v{{[0-9]+}}, off, s[{{[0-9]+:[0-9]+}}], s4 offset:4 ; 4-byte Folded Spill
270  ; FLATSCR: scratch_store_dwordx2 off, v[{{[0-9:]+}}], s32 offset:4092 ; 8-byte Folded Spill
271  %aptr = getelementptr <2 x i32>, <2 x i32> addrspace(5)* %bufv2, i32 1
272  %a = load volatile <2 x i32>, <2 x i32> addrspace(5)* %aptr
273
274  ; Force %a to spill.
275  call void asm sideeffect "", "~{v0},~{v1},~{v2},~{v3},~{v4},~{v5},~{v6},~{v7}" ()
276
277  ; Ensure the alloca sticks around.
278  %bptr = getelementptr i32, i32 addrspace(5)* %bufv1, i32 1
279  %b = load volatile i32, i32 addrspace(5)* %bptr
280
281  ; Ensure the spill is of the full super-reg.
282  call void asm sideeffect "; $0", "r"(<2 x i32> %a)
283
284  ret void
285}
286
287attributes #0 = { nounwind }
288attributes #1 = { nounwind "amdgpu-num-sgpr"="17" "amdgpu-num-vgpr"="8" }
289attributes #2 = { nounwind "amdgpu-num-sgpr"="14" "amdgpu-num-vgpr"="8" }
290