1; RUN: llc -mtriple=amdgcn--amdhsa -mcpu=gfx900 --amdhsa-code-object-version=2 -amdgpu-ir-lower-kernel-arguments=0 -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=VI,GCN,HSA-VI,FUNC %s
2
3; Repeat of some problematic tests in kernel-args.ll, with the IR
4; argument lowering pass disabled. Struct padding needs to be
5; accounted for, as well as legalization of types changing offsets.
6
7; FUNC-LABEL: {{^}}i1_arg:
8; HSA-VI: kernarg_segment_byte_size = 12
9; HSA-VI: kernarg_segment_alignment = 4
10
11; GCN: s_load_dword s
12; GCN: s_and_b32
13define amdgpu_kernel void @i1_arg(i1 addrspace(1)* %out, i1 %x) nounwind {
14  store i1 %x, i1 addrspace(1)* %out, align 1
15  ret void
16}
17
18; FUNC-LABEL: {{^}}v3i8_arg:
19; HSA-VI: kernarg_segment_byte_size = 12
20; HSA-VI: kernarg_segment_alignment = 4
21; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x0
22; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x8
23define amdgpu_kernel void @v3i8_arg(<3 x i8> addrspace(1)* nocapture %out, <3 x i8> %in) nounwind {
24entry:
25  store <3 x i8> %in, <3 x i8> addrspace(1)* %out, align 4
26  ret void
27}
28
29; FUNC-LABEL: {{^}}i65_arg:
30; HSA-VI: kernarg_segment_byte_size = 24
31; HSA-VI: kernarg_segment_alignment = 4
32; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x0
33; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x8
34define amdgpu_kernel void @i65_arg(i65 addrspace(1)* nocapture %out, i65 %in) nounwind {
35entry:
36  store i65 %in, i65 addrspace(1)* %out, align 4
37  ret void
38}
39
40; FUNC-LABEL: {{^}}empty_struct_arg:
41; HSA-VI: kernarg_segment_byte_size = 0
42define amdgpu_kernel void @empty_struct_arg({} %in) nounwind {
43  ret void
44}
45
46; The correct load offsets for these:
47; load 4 from 0,
48; load 8 from 8
49; load 4 from 24
50; load 8 from 32
51
52; With the SelectionDAG argument lowering, the alignments for the
53; struct members is not properly considered, making these wrong.
54
55; FIXME: Total argument size is computed wrong
56; FUNC-LABEL: {{^}}struct_argument_alignment:
57; HSA-VI: kernarg_segment_byte_size = 40
58; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
59; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x8
60; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x18
61; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x20
62define amdgpu_kernel void @struct_argument_alignment({i32, i64} %arg0, i8, {i32, i64} %arg1) {
63  %val0 = extractvalue {i32, i64} %arg0, 0
64  %val1 = extractvalue {i32, i64} %arg0, 1
65  %val2 = extractvalue {i32, i64} %arg1, 0
66  %val3 = extractvalue {i32, i64} %arg1, 1
67  store volatile i32 %val0, i32 addrspace(1)* null
68  store volatile i64 %val1, i64 addrspace(1)* null
69  store volatile i32 %val2, i32 addrspace(1)* null
70  store volatile i64 %val3, i64 addrspace(1)* null
71  ret void
72}
73
74; No padding between i8 and next struct, but round up at end to 4 byte
75; multiple.
76; FUNC-LABEL: {{^}}packed_struct_argument_alignment:
77; HSA-VI: kernarg_segment_byte_size = 28
78; HSA-VI: global_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:17
79; HSA-VI: global_load_dword v{{[0-9]+}}, v{{\[[0-9]+:[0-9]+\]}}, off offset:13
80; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
81; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x4
82define amdgpu_kernel void @packed_struct_argument_alignment(<{i32, i64}> %arg0, i8, <{i32, i64}> %arg1) {
83  %val0 = extractvalue <{i32, i64}> %arg0, 0
84  %val1 = extractvalue <{i32, i64}> %arg0, 1
85  %val2 = extractvalue <{i32, i64}> %arg1, 0
86  %val3 = extractvalue <{i32, i64}> %arg1, 1
87  store volatile i32 %val0, i32 addrspace(1)* null
88  store volatile i64 %val1, i64 addrspace(1)* null
89  store volatile i32 %val2, i32 addrspace(1)* null
90  store volatile i64 %val3, i64 addrspace(1)* null
91  ret void
92}
93
94; GCN-LABEL: {{^}}struct_argument_alignment_after:
95; HSA-VI: kernarg_segment_byte_size = 64
96; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
97; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x8
98; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x18
99; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x20
100; HSA-VI: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x30
101define amdgpu_kernel void @struct_argument_alignment_after({i32, i64} %arg0, i8, {i32, i64} %arg2, i8, <4 x i32> %arg4) {
102  %val0 = extractvalue {i32, i64} %arg0, 0
103  %val1 = extractvalue {i32, i64} %arg0, 1
104  %val2 = extractvalue {i32, i64} %arg2, 0
105  %val3 = extractvalue {i32, i64} %arg2, 1
106  store volatile i32 %val0, i32 addrspace(1)* null
107  store volatile i64 %val1, i64 addrspace(1)* null
108  store volatile i32 %val2, i32 addrspace(1)* null
109  store volatile i64 %val3, i64 addrspace(1)* null
110  store volatile <4 x i32> %arg4, <4 x i32> addrspace(1)* null
111  ret void
112}
113
114; GCN-LABEL: {{^}}array_3xi32:
115; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
116; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x4
117; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x8
118; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0xc
119define amdgpu_kernel void @array_3xi32(i16 %arg0, [3 x i32] %arg1) {
120  store volatile i16 %arg0, i16 addrspace(1)* undef
121  store volatile [3 x i32] %arg1, [3 x i32] addrspace(1)* undef
122  ret void
123}
124
125; GCN-LABEL: {{^}}array_3xi16:
126; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
127; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x4
128define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
129  store volatile i8 %arg0, i8 addrspace(1)* undef
130  store volatile [3 x i16] %arg1, [3 x i16] addrspace(1)* undef
131  ret void
132}
133
134; GCN-LABEL: {{^}}v2i15_arg:
135; GCN: s_load_dword [[DWORD:s[0-9]+]]
136; GCN-DAG: s_bfe_u32 [[BFE:s[0-9]+]], [[DWORD]], 0x100010{{$}}
137; GCN-DAG: s_and_b32 [[AND:s[0-9]+]], [[DWORD]], 0x7fff{{$}}
138define amdgpu_kernel void @v2i15_arg(<2 x i15> addrspace(1)* nocapture %out, <2 x i15> %in) {
139entry:
140  store <2 x i15> %in, <2 x i15> addrspace(1)* %out, align 4
141  ret void
142}
143
144; GCN-LABEL: {{^}}v3i15_arg:
145; GCN: s_load_dword [[DWORD:s[0-9]+]]
146; GCN: s_lshl_b64
147; GCN: s_and_b32
148; GCN: s_and_b32
149; GCN: s_or_b32
150define amdgpu_kernel void @v3i15_arg(<3 x i15> addrspace(1)* nocapture %out, <3 x i15> %in) {
151entry:
152  store <3 x i15> %in, <3 x i15> addrspace(1)* %out, align 4
153  ret void
154}
155
156; Byref pointers should only be treated as offsets from kernarg
157; GCN-LABEL: {{^}}byref_constant_i8_arg:
158; GCN: kernarg_segment_byte_size = 12
159; GCN: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s4
160; GCN: v_mov_b32_e32 v[[VPTR_HI:[0-9]+]], s5
161; GCN: global_load_ubyte v{{[0-9]+}}, v{{\[}}[[VPTR_LO]]:[[VPTR_HI]]{{\]}}, off offset:8
162define amdgpu_kernel void @byref_constant_i8_arg(i32 addrspace(1)* nocapture %out, i8 addrspace(4)* byref(i8) %in.byref) {
163  %in = load i8, i8 addrspace(4)* %in.byref
164  %ext = zext i8 %in to i32
165  store i32 %ext, i32 addrspace(1)* %out, align 4
166  ret void
167}
168
169; GCN-LABEL: {{^}}byref_constant_i16_arg:
170; GCN: kernarg_segment_byte_size = 12
171; GCN: v_mov_b32_e32 v[[VPTR_LO:[0-9]+]], s4
172; GCN: v_mov_b32_e32 v[[VPTR_HI:[0-9]+]], s5
173; GCN: global_load_ushort v{{[0-9]+}}, v{{\[}}[[VPTR_LO]]:[[VPTR_HI]]{{\]}}, off offset:8
174define amdgpu_kernel void @byref_constant_i16_arg(i32 addrspace(1)* nocapture %out, i16 addrspace(4)* byref(i16) %in.byref) {
175  %in = load i16, i16 addrspace(4)* %in.byref
176  %ext = zext i16 %in to i32
177  store i32 %ext, i32 addrspace(1)* %out, align 4
178  ret void
179}
180
181; GCN-LABEL: {{^}}byref_constant_i32_arg:
182; GCN: kernarg_segment_byte_size = 16
183; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
184; GCN: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0xc{{$}}
185define amdgpu_kernel void @byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in.byref, i32 %after.offset) {
186  %in = load i32, i32 addrspace(4)* %in.byref
187  store volatile i32 %in, i32 addrspace(1)* %out, align 4
188  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
189  ret void
190}
191
192; GCN-LABEL: {{^}}byref_constant_v4i32_arg:
193; GCN: kernarg_segment_byte_size = 36
194; GCN: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x10{{$}}
195; GCN: s_load_dword s{{[0-9]+}}, s[4:5], 0x20{{$}}
196define amdgpu_kernel void @byref_constant_v4i32_arg(<4 x i32> addrspace(1)* nocapture %out, <4 x i32> addrspace(4)* byref(<4 x i32>) %in.byref, i32 %after.offset) {
197  %in = load <4 x i32>, <4 x i32> addrspace(4)* %in.byref
198  store volatile <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
199  %out.cast = bitcast <4 x i32> addrspace(1)* %out to i32 addrspace(1)*
200  store volatile i32 %after.offset, i32 addrspace(1)* %out.cast, align 4
201  ret void
202}
203
204; GCN-LABEL: {{^}}byref_align_constant_i32_arg:
205; GCN: kernarg_segment_byte_size = 264
206; GCN-DAG: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x100{{$}}
207; GCN-DAG: s_load_dword [[AFTER_OFFSET:s[0-9]+]], s[4:5], 0x104{{$}}
208; GCN-DAG: v_mov_b32_e32 [[V_IN:v[0-9]+]], [[IN]]
209; GCN-DAG: v_mov_b32_e32 [[V_AFTER_OFFSET:v[0-9]+]], [[AFTER_OFFSET]]
210; GCN: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_IN]]
211; GCN: global_store_dword v{{\[[0-9]+:[0-9]+\]}}, [[V_AFTER_OFFSET]]
212define amdgpu_kernel void @byref_align_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) align(256) %in.byref, i32 %after.offset) {
213  %in = load i32, i32 addrspace(4)* %in.byref
214  store volatile i32 %in, i32 addrspace(1)* %out, align 4
215  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
216  ret void
217}
218
219; GCN-LABEL: {{^}}byref_natural_align_constant_v16i32_arg:
220; GCN: kernarg_segment_byte_size = 132
221; GCN-DAG: s_load_dword s{{[0-9]+}}, s[4:5], 0x80
222; GCN-DAG: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x40{{$}}
223define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) align(64) %in.byref, i32 %after.offset) {
224  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
225  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
226  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
227  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
228  ret void
229}
230
231; Also accept byref kernel arguments with other global address spaces.
232; GCN-LABEL: {{^}}byref_global_i32_arg:
233; GCN: kernarg_segment_byte_size = 12
234; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
235define amdgpu_kernel void @byref_global_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* byref(i32) %in.byref) {
236  %in = load i32, i32 addrspace(1)* %in.byref
237  store i32 %in, i32 addrspace(1)* %out, align 4
238  ret void
239}
240
241; GCN-LABEL: {{^}}byref_flat_i32_arg:
242; GCN: flat_load_dword [[IN:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}} offset:8{{$}}
243define amdgpu_kernel void @byref_flat_i32_arg(i32 addrspace(1)* nocapture %out, i32* byref(i32) %in.byref) {
244  %in = load i32, i32* %in.byref
245  store i32 %in, i32 addrspace(1)* %out, align 4
246  ret void
247}
248
249; GCN-LABEL: {{^}}byref_constant_32bit_i32_arg:
250; GCN: s_add_i32 s[[PTR_LO:[0-9]+]], s4, 8
251; GCN: s_mov_b32 s[[PTR_HI:[0-9]+]], 0{{$}}
252; GCN: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}}
253define amdgpu_kernel void @byref_constant_32bit_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(6)* byref(i32) %in.byref) {
254  %in = load i32, i32 addrspace(6)* %in.byref
255  store i32 %in, i32 addrspace(1)* %out, align 4
256  ret void
257}
258
259; define amdgpu_kernel void @byref_unknown_as_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(999)* byref %in.byref) {
260;   %in = load i32, i32 addrspace(999)* %in.byref
261;   store i32 %in, i32 addrspace(1)* %out, align 4
262;   ret void
263; }
264
265; GCN-LABEL: {{^}}multi_byref_constant_i32_arg:
266; GCN: kernarg_segment_byte_size = 20
267; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x8
268; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0xc
269; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x10
270define amdgpu_kernel void @multi_byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in0.byref, i32 addrspace(4)* byref(i32) %in1.byref, i32 %after.offset) {
271  %in0 = load i32, i32 addrspace(4)* %in0.byref
272  %in1 = load i32, i32 addrspace(4)* %in1.byref
273  store volatile i32 %in0, i32 addrspace(1)* %out, align 4
274  store volatile i32 %in1, i32 addrspace(1)* %out, align 4
275  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
276  ret void
277}
278
279; GCN-LABEL: {{^}}byref_constant_i32_arg_offset0:
280; GCN: kernarg_segment_byte_size = 4
281; GCN-NOT: s4
282; GCN-NOT: s5
283; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x0{{$}}
284define amdgpu_kernel void @byref_constant_i32_arg_offset0(i32 addrspace(4)* byref(i32) %in.byref) {
285  %in = load i32, i32 addrspace(4)* %in.byref
286  store i32 %in, i32 addrspace(1)* undef, align 4
287  ret void
288}
289