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-DAG: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
79; HSA-VI: global_load_dwordx2 v{{\[[0-9]+:[0-9]+\]}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}} offset:17
80; HSA-VI: global_load_dword v{{[0-9]+}}, [[ZERO]], s{{\[[0-9]+:[0-9]+\]}} offset:13
81; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
82; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x4
83define amdgpu_kernel void @packed_struct_argument_alignment(<{i32, i64}> %arg0, i8, <{i32, i64}> %arg1) {
84  %val0 = extractvalue <{i32, i64}> %arg0, 0
85  %val1 = extractvalue <{i32, i64}> %arg0, 1
86  %val2 = extractvalue <{i32, i64}> %arg1, 0
87  %val3 = extractvalue <{i32, i64}> %arg1, 1
88  store volatile i32 %val0, i32 addrspace(1)* null
89  store volatile i64 %val1, i64 addrspace(1)* null
90  store volatile i32 %val2, i32 addrspace(1)* null
91  store volatile i64 %val3, i64 addrspace(1)* null
92  ret void
93}
94
95; GCN-LABEL: {{^}}struct_argument_alignment_after:
96; HSA-VI: kernarg_segment_byte_size = 64
97; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
98; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x8
99; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x18
100; HSA-VI: s_load_dwordx2 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x20
101; HSA-VI: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x30
102define amdgpu_kernel void @struct_argument_alignment_after({i32, i64} %arg0, i8, {i32, i64} %arg2, i8, <4 x i32> %arg4) {
103  %val0 = extractvalue {i32, i64} %arg0, 0
104  %val1 = extractvalue {i32, i64} %arg0, 1
105  %val2 = extractvalue {i32, i64} %arg2, 0
106  %val3 = extractvalue {i32, i64} %arg2, 1
107  store volatile i32 %val0, i32 addrspace(1)* null
108  store volatile i64 %val1, i64 addrspace(1)* null
109  store volatile i32 %val2, i32 addrspace(1)* null
110  store volatile i64 %val3, i64 addrspace(1)* null
111  store volatile <4 x i32> %arg4, <4 x i32> addrspace(1)* null
112  ret void
113}
114
115; GCN-LABEL: {{^}}array_3xi32:
116; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
117; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x4
118; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x8
119; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0xc
120define amdgpu_kernel void @array_3xi32(i16 %arg0, [3 x i32] %arg1) {
121  store volatile i16 %arg0, i16 addrspace(1)* undef
122  store volatile [3 x i32] %arg1, [3 x i32] addrspace(1)* undef
123  ret void
124}
125
126; GCN-LABEL: {{^}}array_3xi16:
127; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x0
128; HSA-VI: s_load_dword s{{[0-9]+}}, s[4:5], 0x4
129define amdgpu_kernel void @array_3xi16(i8 %arg0, [3 x i16] %arg1) {
130  store volatile i8 %arg0, i8 addrspace(1)* undef
131  store volatile [3 x i16] %arg1, [3 x i16] addrspace(1)* undef
132  ret void
133}
134
135; GCN-LABEL: {{^}}v2i15_arg:
136; GCN: s_load_dword [[DWORD:s[0-9]+]]
137; GCN-DAG: s_bfe_u32 [[BFE:s[0-9]+]], [[DWORD]], 0x100010{{$}}
138; GCN-DAG: s_and_b32 [[AND:s[0-9]+]], [[DWORD]], 0x7fff{{$}}
139define amdgpu_kernel void @v2i15_arg(<2 x i15> addrspace(1)* nocapture %out, <2 x i15> %in) {
140entry:
141  store <2 x i15> %in, <2 x i15> addrspace(1)* %out, align 4
142  ret void
143}
144
145; GCN-LABEL: {{^}}v3i15_arg:
146; GCN: s_load_dword [[DWORD:s[0-9]+]]
147; GCN: s_lshl_b64
148; GCN: s_and_b32
149; GCN: s_and_b32
150; GCN: s_or_b32
151define amdgpu_kernel void @v3i15_arg(<3 x i15> addrspace(1)* nocapture %out, <3 x i15> %in) {
152entry:
153  store <3 x i15> %in, <3 x i15> addrspace(1)* %out, align 4
154  ret void
155}
156
157; Byref pointers should only be treated as offsets from kernarg
158; GCN-LABEL: {{^}}byref_constant_i8_arg:
159; GCN: kernarg_segment_byte_size = 12
160; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
161; GCN: global_load_ubyte v{{[0-9]+}}, [[ZERO]], s[4:5] offset:8
162define amdgpu_kernel void @byref_constant_i8_arg(i32 addrspace(1)* nocapture %out, i8 addrspace(4)* byref(i8) %in.byref) {
163  %in = load i8, i8 addrspace(4)* %in.byref
164  %ext = zext i8 %in to i32
165  store i32 %ext, i32 addrspace(1)* %out, align 4
166  ret void
167}
168
169; GCN-LABEL: {{^}}byref_constant_i16_arg:
170; GCN: kernarg_segment_byte_size = 12
171; GCN: v_mov_b32_e32 [[ZERO:v[0-9]+]], 0{{$}}
172; GCN: global_load_ushort v{{[0-9]+}}, [[ZERO]], s[4:5] offset:8
173define amdgpu_kernel void @byref_constant_i16_arg(i32 addrspace(1)* nocapture %out, i16 addrspace(4)* byref(i16) %in.byref) {
174  %in = load i16, i16 addrspace(4)* %in.byref
175  %ext = zext i16 %in to i32
176  store i32 %ext, i32 addrspace(1)* %out, align 4
177  ret void
178}
179
180; GCN-LABEL: {{^}}byref_constant_i32_arg:
181; GCN: kernarg_segment_byte_size = 16
182; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
183; GCN: s_load_dword [[OFFSET:s[0-9]+]], s[4:5], 0xc{{$}}
184define amdgpu_kernel void @byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in.byref, i32 %after.offset) {
185  %in = load i32, i32 addrspace(4)* %in.byref
186  store volatile i32 %in, i32 addrspace(1)* %out, align 4
187  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
188  ret void
189}
190
191; GCN-LABEL: {{^}}byref_constant_v4i32_arg:
192; GCN: kernarg_segment_byte_size = 36
193; GCN: s_load_dwordx4 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x10{{$}}
194; GCN: s_load_dword s{{[0-9]+}}, s[4:5], 0x20{{$}}
195define 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) {
196  %in = load <4 x i32>, <4 x i32> addrspace(4)* %in.byref
197  store volatile <4 x i32> %in, <4 x i32> addrspace(1)* %out, align 4
198  %out.cast = bitcast <4 x i32> addrspace(1)* %out to i32 addrspace(1)*
199  store volatile i32 %after.offset, i32 addrspace(1)* %out.cast, align 4
200  ret void
201}
202
203; GCN-LABEL: {{^}}byref_align_constant_i32_arg:
204; GCN: kernarg_segment_byte_size = 264
205; GCN-DAG: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x100{{$}}
206; GCN-DAG: s_load_dword [[AFTER_OFFSET:s[0-9]+]], s[4:5], 0x104{{$}}
207; GCN-DAG: v_mov_b32_e32 [[V_IN:v[0-9]+]], [[IN]]
208; GCN-DAG: v_mov_b32_e32 [[V_AFTER_OFFSET:v[0-9]+]], [[AFTER_OFFSET]]
209; GCN: global_store_dword v{{[0-9]+}}, [[V_IN]], s
210; GCN: global_store_dword v{{[0-9]+}}, [[V_AFTER_OFFSET]], s
211define 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) {
212  %in = load i32, i32 addrspace(4)* %in.byref
213  store volatile i32 %in, i32 addrspace(1)* %out, align 4
214  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
215  ret void
216}
217
218; GCN-LABEL: {{^}}byref_natural_align_constant_v16i32_arg:
219; GCN: kernarg_segment_byte_size = 132
220; GCN-DAG: s_load_dword s{{[0-9]+}}, s[4:5], 0x80
221; GCN-DAG: s_load_dwordx16 s{{\[[0-9]+:[0-9]+\]}}, s[4:5], 0x40{{$}}
222define amdgpu_kernel void @byref_natural_align_constant_v16i32_arg(i32 addrspace(1)* nocapture %out, i8, <16 x i32> addrspace(4)* byref(<16 x i32>) align(64) %in.byref, i32 %after.offset) {
223  %in = load <16 x i32>, <16 x i32> addrspace(4)* %in.byref
224  %cast.out = bitcast i32 addrspace(1)* %out to <16 x i32> addrspace(1)*
225  store volatile <16 x i32> %in, <16 x i32> addrspace(1)* %cast.out, align 4
226  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
227  ret void
228}
229
230; Also accept byref kernel arguments with other global address spaces.
231; GCN-LABEL: {{^}}byref_global_i32_arg:
232; GCN: kernarg_segment_byte_size = 12
233; GCN: s_load_dword [[IN:s[0-9]+]], s[4:5], 0x8{{$}}
234define amdgpu_kernel void @byref_global_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(1)* byref(i32) %in.byref) {
235  %in = load i32, i32 addrspace(1)* %in.byref
236  store i32 %in, i32 addrspace(1)* %out, align 4
237  ret void
238}
239
240; GCN-LABEL: {{^}}byref_flat_i32_arg:
241; GCN: flat_load_dword [[IN:v[0-9]+]], v{{\[[0-9]+:[0-9]+\]}} offset:8{{$}}
242define amdgpu_kernel void @byref_flat_i32_arg(i32 addrspace(1)* nocapture %out, i32* byref(i32) %in.byref) {
243  %in = load i32, i32* %in.byref
244  store i32 %in, i32 addrspace(1)* %out, align 4
245  ret void
246}
247
248; GCN-LABEL: {{^}}byref_constant_32bit_i32_arg:
249; GCN: s_add_i32 s[[PTR_LO:[0-9]+]], s4, 8
250; GCN: s_mov_b32 s[[PTR_HI:[0-9]+]], 0{{$}}
251; GCN: s_load_dword s{{[0-9]+}}, s{{\[}}[[PTR_LO]]:[[PTR_HI]]{{\]}}, 0x0{{$}}
252define amdgpu_kernel void @byref_constant_32bit_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(6)* byref(i32) %in.byref) {
253  %in = load i32, i32 addrspace(6)* %in.byref
254  store i32 %in, i32 addrspace(1)* %out, align 4
255  ret void
256}
257
258; define amdgpu_kernel void @byref_unknown_as_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(999)* byref %in.byref) {
259;   %in = load i32, i32 addrspace(999)* %in.byref
260;   store i32 %in, i32 addrspace(1)* %out, align 4
261;   ret void
262; }
263
264; GCN-LABEL: {{^}}multi_byref_constant_i32_arg:
265; GCN: kernarg_segment_byte_size = 20
266; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x8
267; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0xc
268; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x10
269define amdgpu_kernel void @multi_byref_constant_i32_arg(i32 addrspace(1)* nocapture %out, i32 addrspace(4)* byref(i32) %in0.byref, i32 addrspace(4)* byref(i32) %in1.byref, i32 %after.offset) {
270  %in0 = load i32, i32 addrspace(4)* %in0.byref
271  %in1 = load i32, i32 addrspace(4)* %in1.byref
272  store volatile i32 %in0, i32 addrspace(1)* %out, align 4
273  store volatile i32 %in1, i32 addrspace(1)* %out, align 4
274  store volatile i32 %after.offset, i32 addrspace(1)* %out, align 4
275  ret void
276}
277
278; GCN-LABEL: {{^}}byref_constant_i32_arg_offset0:
279; GCN: kernarg_segment_byte_size = 4
280; GCN-NOT: s4
281; GCN-NOT: s5
282; GCN: s_load_dword {{s[0-9]+}}, s[4:5], 0x0{{$}}
283define amdgpu_kernel void @byref_constant_i32_arg_offset0(i32 addrspace(4)* byref(i32) %in.byref) {
284  %in = load i32, i32 addrspace(4)* %in.byref
285  store i32 %in, i32 addrspace(1)* undef, align 4
286  ret void
287}
288