1; RUN: llc -march=amdgcn -mcpu=tahiti  -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=SI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=SICI -check-prefix=SIVIGFX9 %s
2; RUN: llc -march=amdgcn -mcpu=bonaire -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=CI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=SICI %s
3; RUN: llc -march=amdgcn -mcpu=tonga   -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=VI   -check-prefix=GCN -check-prefix=SICIVI -check-prefix=VIGFX9 -check-prefix=SIVIGFX9 %s
4; RUN: llc -march=amdgcn -mcpu=gfx900  -verify-machineinstrs -show-mc-encoding < %s | FileCheck -check-prefix=GFX9 -check-prefix=GCN -check-prefix=VIGFX9 -check-prefix=SIVIGFX9  %s
5
6; SMRD load with an immediate offset.
7; GCN-LABEL: {{^}}smrd0:
8; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x1 ; encoding: [0x01
9; VIGFX9: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4
10define amdgpu_kernel void @smrd0(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
11entry:
12  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 1
13  %tmp1 = load i32, i32 addrspace(4)* %tmp
14  store i32 %tmp1, i32 addrspace(1)* %out
15  ret void
16}
17
18; SMRD load with the largest possible immediate offset.
19; GCN-LABEL: {{^}}smrd1:
20; SICI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff,0x{{[0-9]+[137]}}
21; VIGFX9: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
22define amdgpu_kernel void @smrd1(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
23entry:
24  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 255
25  %tmp1 = load i32, i32 addrspace(4)* %tmp
26  store i32 %tmp1, i32 addrspace(1)* %out
27  ret void
28}
29
30; SMRD load with an offset greater than the largest possible immediate.
31; GCN-LABEL: {{^}}smrd2:
32; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
33; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
34; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
35; VIGFX9: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
36; GCN: s_endpgm
37define amdgpu_kernel void @smrd2(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
38entry:
39  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 256
40  %tmp1 = load i32, i32 addrspace(4)* %tmp
41  store i32 %tmp1, i32 addrspace(1)* %out
42  ret void
43}
44
45; SMRD load with a 64-bit offset
46; GCN-LABEL: {{^}}smrd3:
47; FIXME: There are too many copies here because we don't fold immediates
48;        through REG_SEQUENCE
49; SI: s_load_dwordx2 s[{{[0-9]:[0-9]}}], s[{{[0-9]:[0-9]}}], 0x13 ; encoding: [0x13
50; TODO: Add VI checks
51; GCN: s_endpgm
52define amdgpu_kernel void @smrd3(i32 addrspace(1)* %out, [8 x i32], i32 addrspace(4)* %ptr) #0 {
53entry:
54  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 4294967296
55  %tmp1 = load i32, i32 addrspace(4)* %tmp
56  store i32 %tmp1, i32 addrspace(1)* %out
57  ret void
58}
59
60; SMRD load with the largest possible immediate offset on VI
61; GCN-LABEL: {{^}}smrd4:
62; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
63; SI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
64; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
65; VIGFX9: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
66define amdgpu_kernel void @smrd4(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
67entry:
68  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 262143
69  %tmp1 = load i32, i32 addrspace(4)* %tmp
70  store i32 %tmp1, i32 addrspace(1)* %out
71  ret void
72}
73
74; SMRD load with an offset greater than the largest possible immediate on VI
75; GCN-LABEL: {{^}}smrd5:
76; SIVIGFX9: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
77; SIVIGFX9: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
78; CI: s_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
79; GCN: s_endpgm
80define amdgpu_kernel void @smrd5(i32 addrspace(1)* %out, i32 addrspace(4)* %ptr) #0 {
81entry:
82  %tmp = getelementptr i32, i32 addrspace(4)* %ptr, i64 262144
83  %tmp1 = load i32, i32 addrspace(4)* %tmp
84  store i32 %tmp1, i32 addrspace(1)* %out
85  ret void
86}
87
88; GCN-LABEL: {{^}}smrd_hazard:
89; GCN-DAG: s_mov_b32 s3, 3
90; GCN-DAG: s_mov_b32 s2, 2
91; GCN-DAG: s_mov_b32 s1, 1
92; GCN-DAG: s_mov_b32 s0, 0
93; SI-NEXT: nop 3
94; GCN-NEXT: s_buffer_load_dword s0, s[0:3], 0x0
95define amdgpu_ps float @smrd_hazard(<4 x i32> inreg %desc) #0 {
96main_body:
97  %d0 = insertelement <4 x i32> undef, i32 0, i32 0
98  %d1 = insertelement <4 x i32> %d0, i32 1, i32 1
99  %d2 = insertelement <4 x i32> %d1, i32 2, i32 2
100  %d3 = insertelement <4 x i32> %d2, i32 3, i32 3
101  %r = call float @llvm.SI.load.const.v4i32(<4 x i32> %d3, i32 0)
102  ret float %r
103}
104
105; SMRD load using the load.const.v4i32 intrinsic with an immediate offset
106; GCN-LABEL: {{^}}smrd_load_const0:
107; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x4 ; encoding: [0x04
108; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x10
109define amdgpu_ps void @smrd_load_const0(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
110main_body:
111  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
112  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
113  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 16)
114  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %tmp21, i1 true, i1 true) #0
115  ret void
116}
117
118; SMRD load using the load.const.v4i32 intrinsic with the largest possible immediate
119; offset.
120; GCN-LABEL: {{^}}smrd_load_const1:
121; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff ; encoding: [0xff
122; SICI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xff glc ; encoding: [0xff
123; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc
124; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3fc glc
125define amdgpu_ps void @smrd_load_const1(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
126main_body:
127  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
128  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
129  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 1020)
130  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
131  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 1020, i32 1)
132  %s.buffer.float = bitcast i32 %s.buffer to float
133  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
134  ret void
135}
136
137; SMRD load using the load.const.v4i32 intrinsic with an offset greater than the
138; largets possible immediate.
139; immediate offset.
140; GCN-LABEL: {{^}}smrd_load_const2:
141; SI: s_movk_i32 s[[OFFSET:[0-9]]], 0x400
142; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
143; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], s[[OFFSET]] ; encoding: [0x0[[OFFSET]]
144; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
145; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x100
146; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
147; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x400
148define amdgpu_ps void @smrd_load_const2(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
149main_body:
150  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
151  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
152  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 1024)
153  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
154  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 1024, i32 0)
155  %s.buffer.float = bitcast i32 %s.buffer to float
156  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
157  ret void
158}
159
160; SMRD load with the largest possible immediate offset on VI
161; GCN-LABEL: {{^}}smrd_load_const3:
162; SI: s_mov_b32 [[OFFSET:s[0-9]+]], 0xffffc
163; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
164; SI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
165; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
166; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x3ffff
167; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
168; VIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0xffffc
169define amdgpu_ps void @smrd_load_const3(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
170main_body:
171  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
172  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
173  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 1048572)
174  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
175  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 1048572, i32 0)
176  %s.buffer.float = bitcast i32 %s.buffer to float
177  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
178  ret void
179}
180
181; SMRD load with an offset greater than the largest possible immediate on VI
182; GCN-LABEL: {{^}}smrd_load_const4:
183; SIVIGFX9: s_mov_b32 [[OFFSET:s[0-9]+]], 0x100000
184; SIVIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
185; SIVIGFX9: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], [[OFFSET]]
186; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
187; CI: s_buffer_load_dword s{{[0-9]}}, s[{{[0-9]:[0-9]}}], 0x40000
188; GCN: s_endpgm
189define amdgpu_ps void @smrd_load_const4(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
190main_body:
191  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
192  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
193  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 1048576)
194  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
195  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 1048576, i32 0)
196  %s.buffer.float = bitcast i32 %s.buffer to float
197  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
198  ret void
199}
200
201; dwordx2 s.buffer.load
202; GCN-LABEL: {{^}}s_buffer_load_dwordx2:
203; VIGFX9: s_buffer_load_dwordx2 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
204; SICI: s_buffer_load_dwordx2 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
205define amdgpu_ps void @s_buffer_load_dwordx2(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
206main_body:
207  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
208  %s.buffer = call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %tmp22, i32 128, i32 0)
209  %s.buffer.0 = extractelement <2 x i32> %s.buffer, i32 0
210  %s.buffer.0.float = bitcast i32 %s.buffer.0 to float
211  %s.buffer.1 = extractelement <2 x i32> %s.buffer, i32 1
212  %s.buffer.1.float = bitcast i32 %s.buffer.1 to float
213  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0.float, float %s.buffer.1.float, float %s.buffer.0.float, float %s.buffer.1.float, i1 true, i1 true) #0
214  ret void
215}
216
217; dwordx4 s.buffer.load
218; GCN-LABEL: {{^}}s_buffer_load_dwordx4:
219; VIGFX9: s_buffer_load_dwordx4 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
220; SICI: s_buffer_load_dwordx4 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
221define amdgpu_ps void @s_buffer_load_dwordx4(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
222main_body:
223  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
224  %s.buffer = call <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32> %tmp22, i32 128, i32 0)
225  %s.buffer.0 = extractelement <4 x i32> %s.buffer, i32 0
226  %s.buffer.0.float = bitcast i32 %s.buffer.0 to float
227  %s.buffer.1 = extractelement <4 x i32> %s.buffer, i32 1
228  %s.buffer.1.float = bitcast i32 %s.buffer.1 to float
229  %s.buffer.2 = extractelement <4 x i32> %s.buffer, i32 2
230  %s.buffer.2.float = bitcast i32 %s.buffer.2 to float
231  %s.buffer.3 = extractelement <4 x i32> %s.buffer, i32 3
232  %s.buffer.3.float = bitcast i32 %s.buffer.3 to float
233  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0.float, float %s.buffer.1.float, float %s.buffer.2.float, float %s.buffer.3.float, i1 true, i1 true) #0
234  ret void
235}
236
237; dwordx8 s.buffer.load
238; GCN-LABEL: {{^}}s_buffer_load_dwordx8:
239; VIGFX9: s_buffer_load_dwordx8 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
240; SICI: s_buffer_load_dwordx8 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
241define amdgpu_ps void @s_buffer_load_dwordx8(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
242main_body:
243  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
244  %s.buffer = call <8 x i32> @llvm.amdgcn.s.buffer.load.v8i32(<4 x i32> %tmp22, i32 128, i32 0)
245  %s.buffer.0 = extractelement <8 x i32> %s.buffer, i32 0
246  %s.buffer.0.float = bitcast i32 %s.buffer.0 to float
247  %s.buffer.1 = extractelement <8 x i32> %s.buffer, i32 2
248  %s.buffer.1.float = bitcast i32 %s.buffer.1 to float
249  %s.buffer.2 = extractelement <8 x i32> %s.buffer, i32 5
250  %s.buffer.2.float = bitcast i32 %s.buffer.2 to float
251  %s.buffer.3 = extractelement <8 x i32> %s.buffer, i32 7
252  %s.buffer.3.float = bitcast i32 %s.buffer.3 to float
253  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0.float, float %s.buffer.1.float, float %s.buffer.2.float, float %s.buffer.3.float, i1 true, i1 true) #0
254  ret void
255}
256
257; dwordx16 s.buffer.load
258; GCN-LABEL: {{^}}s_buffer_load_dwordx16:
259; VIGFX9: s_buffer_load_dwordx16 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x80
260; SICI: s_buffer_load_dwordx16 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]:[0-9]}}], 0x20
261define amdgpu_ps void @s_buffer_load_dwordx16(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in) #0 {
262main_body:
263  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
264  %s.buffer = call <16 x i32> @llvm.amdgcn.s.buffer.load.v16i32(<4 x i32> %tmp22, i32 128, i32 0)
265  %s.buffer.0 = extractelement <16 x i32> %s.buffer, i32 0
266  %s.buffer.0.float = bitcast i32 %s.buffer.0 to float
267  %s.buffer.1 = extractelement <16 x i32> %s.buffer, i32 3
268  %s.buffer.1.float = bitcast i32 %s.buffer.1 to float
269  %s.buffer.2 = extractelement <16 x i32> %s.buffer, i32 12
270  %s.buffer.2.float = bitcast i32 %s.buffer.2 to float
271  %s.buffer.3 = extractelement <16 x i32> %s.buffer, i32 15
272  %s.buffer.3.float = bitcast i32 %s.buffer.3 to float
273  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %s.buffer.0.float, float %s.buffer.1.float, float %s.buffer.2.float, float %s.buffer.3.float, i1 true, i1 true) #0
274  ret void
275}
276
277; GCN-LABEL: {{^}}smrd_sgpr_offset:
278; GCN: s_buffer_load_dword s{{[0-9]}}, s[0:3], s4
279define amdgpu_ps float @smrd_sgpr_offset(<4 x i32> inreg %desc, i32 inreg %offset) #0 {
280main_body:
281  %r = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %offset)
282  ret float %r
283}
284
285; GCN-LABEL: {{^}}smrd_vgpr_offset:
286; GCN: buffer_load_dword v{{[0-9]}}, v0, s[0:3], 0 offen ;
287define amdgpu_ps float @smrd_vgpr_offset(<4 x i32> inreg %desc, i32 %offset) #0 {
288main_body:
289  %r = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %offset)
290  ret float %r
291}
292
293; GCN-LABEL: {{^}}smrd_vgpr_offset_imm:
294; GCN-NEXT: %bb.
295; GCN-NEXT: buffer_load_dword v{{[0-9]}}, v0, s[0:3], 0 offen offset:4092 ;
296define amdgpu_ps float @smrd_vgpr_offset_imm(<4 x i32> inreg %desc, i32 %offset) #0 {
297main_body:
298  %off = add i32 %offset, 4092
299  %r = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %off)
300  ret float %r
301}
302
303; GCN-LABEL: {{^}}smrd_vgpr_offset_imm_too_large:
304; GCN-NEXT: %bb.
305; SICI-NEXT: v_add_{{i|u}}32_e32 v0, {{(vcc, )?}}0x1000, v0
306; SICI-NEXT: buffer_load_dword v{{[0-9]}}, v0, s[0:3], 0 offen ;
307; VIGFX9-NEXT: buffer_load_dword v{{[0-9]}}, v0, s[0:3], 4 offen offset:4092 ;
308define amdgpu_ps float @smrd_vgpr_offset_imm_too_large(<4 x i32> inreg %desc, i32 %offset) #0 {
309main_body:
310  %off = add i32 %offset, 4096
311  %r = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %off)
312  ret float %r
313}
314
315; GCN-LABEL: {{^}}smrd_imm_merged:
316; GCN-NEXT: %bb.
317; SICI-NEXT: s_buffer_load_dwordx4 s[{{[0-9]}}:{{[0-9]}}], s[0:3], 0x1
318; SICI-NEXT: s_buffer_load_dwordx2 s[{{[0-9]}}:{{[0-9]}}], s[0:3], 0x7
319; VIGFX9-NEXT: s_buffer_load_dwordx4 s[{{[0-9]}}:{{[0-9]}}], s[0:3], 0x4
320; VIGFX9-NEXT: s_buffer_load_dwordx2 s[{{[0-9]}}:{{[0-9]}}], s[0:3], 0x1c
321define amdgpu_ps void @smrd_imm_merged(<4 x i32> inreg %desc) #0 {
322main_body:
323  %r1 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 4)
324  %r2 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 8)
325  %r3 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 12)
326  %r4 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 16)
327  %r5 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 28)
328  %r6 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 32)
329  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r1, float %r2, float %r3, float %r4, i1 true, i1 true) #0
330  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r5, float %r6, float undef, float undef, i1 true, i1 true) #0
331  ret void
332}
333
334; GCN-LABEL: {{^}}smrd_imm_merge_m0:
335;
336; GCN: s_buffer_load_dwordx2
337; SICIVI: s_mov_b32 m0
338; SICIVI_DAG: v_interp_p1_f32
339; SICIVI_DAG: v_interp_p1_f32
340; SICIVI_DAG: v_interp_p1_f32
341; SICIVI_DAG: v_interp_p2_f32
342; SICIVI_DAG: v_interp_p2_f32
343; SICIVI_DAG: v_interp_p2_f32
344;
345; extractelement does not result in movrels anymore for vectors gitting 8 dwords
346; SICIVI-NOT: s_mov_b32 m0
347; SICIVI-NOT: v_movrels_b32_e32
348; v_cndmask_b32_e32
349; v_cndmask_b32_e32
350;
351; Merging is still thwarted on GFX9 due to s_set_gpr_idx
352;
353define amdgpu_ps float @smrd_imm_merge_m0(<4 x i32> inreg %desc, i32 inreg %prim, float %u, float %v) #0 {
354main_body:
355  %idx1.f = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 0)
356  %idx1 = bitcast float %idx1.f to i32
357
358  %v0.x1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 0, i32 0, i32 %prim)
359  %v0.x = call nsz float @llvm.amdgcn.interp.p2(float %v0.x1, float %v, i32 0, i32 0, i32 %prim)
360  %v0.y1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 0, i32 1, i32 %prim)
361  %v0.y = call nsz float @llvm.amdgcn.interp.p2(float %v0.y1, float %v, i32 0, i32 1, i32 %prim)
362  %v0.z1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 0, i32 2, i32 %prim)
363  %v0.z = call nsz float @llvm.amdgcn.interp.p2(float %v0.z1, float %v, i32 0, i32 2, i32 %prim)
364  %v0.tmp0 = insertelement <3 x float> undef, float %v0.x, i32 0
365  %v0.tmp1 = insertelement <3 x float> %v0.tmp0, float %v0.y, i32 1
366  %v0 = insertelement <3 x float> %v0.tmp1, float %v0.z, i32 2
367  %a = extractelement <3 x float> %v0, i32 %idx1
368
369  %v1.x1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 1, i32 0, i32 %prim)
370  %v1.x = call nsz float @llvm.amdgcn.interp.p2(float %v1.x1, float %v, i32 1, i32 0, i32 %prim)
371  %v1.y1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 1, i32 1, i32 %prim)
372  %v1.y = call nsz float @llvm.amdgcn.interp.p2(float %v1.y1, float %v, i32 1, i32 1, i32 %prim)
373  %v1.z1 = call nsz float @llvm.amdgcn.interp.p1(float %u, i32 1, i32 2, i32 %prim)
374  %v1.z = call nsz float @llvm.amdgcn.interp.p2(float %v1.z1, float %v, i32 1, i32 2, i32 %prim)
375  %v1.tmp0 = insertelement <3 x float> undef, float %v0.x, i32 0
376  %v1.tmp1 = insertelement <3 x float> %v0.tmp0, float %v0.y, i32 1
377  %v1 = insertelement <3 x float> %v0.tmp1, float %v0.z, i32 2
378
379  %b = extractelement <3 x float> %v1, i32 %idx1
380  %c = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 4)
381
382  %res.tmp = fadd float %a, %b
383  %res = fadd float %res.tmp, %c
384  ret float %res
385}
386
387; GCN-LABEL: {{^}}smrd_vgpr_merged:
388; GCN-NEXT: %bb.
389; GCN-NEXT: buffer_load_dwordx4 v[{{[0-9]}}:{{[0-9]}}], v0, s[0:3], 0 offen offset:4
390; GCN-NEXT: buffer_load_dwordx2 v[{{[0-9]}}:{{[0-9]}}], v0, s[0:3], 0 offen offset:28
391define amdgpu_ps void @smrd_vgpr_merged(<4 x i32> inreg %desc, i32 %a) #0 {
392main_body:
393  %a1 = add i32 %a, 4
394  %a2 = add i32 %a, 8
395  %a3 = add i32 %a, 12
396  %a4 = add i32 %a, 16
397  %a5 = add i32 %a, 28
398  %a6 = add i32 %a, 32
399  %r1 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a1)
400  %r2 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a2)
401  %r3 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a3)
402  %r4 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a4)
403  %r5 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a5)
404  %r6 = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %a6)
405  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r1, float %r2, float %r3, float %r4, i1 true, i1 true) #0
406  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r5, float %r6, float undef, float undef, i1 true, i1 true) #0
407  ret void
408}
409
410; GCN-LABEL: {{^}}smrd_sgpr_descriptor_promoted
411; GCN: v_readfirstlane
412define amdgpu_cs void @smrd_sgpr_descriptor_promoted([0 x i8] addrspace(4)* inreg noalias dereferenceable(18446744073709551615), i32) #0 {
413main_body:
414  %descptr = bitcast [0 x i8] addrspace(4)* %0 to <4 x i32> addrspace(4)*, !amdgpu.uniform !0
415  br label %.outer_loop_header
416
417ret_block:                                       ; preds = %.outer, %.label22, %main_body
418  ret void
419
420.outer_loop_header:
421  br label %.inner_loop_header
422
423.inner_loop_header:                                     ; preds = %.inner_loop_body, %.outer_loop_header
424  %loopctr.1 = phi i32 [ 0, %.outer_loop_header ], [ %loopctr.2, %.inner_loop_body ]
425  %loopctr.2 = add i32 %loopctr.1, 1
426  %inner_br1 = icmp slt i32 %loopctr.2, 10
427  br i1 %inner_br1, label %.inner_loop_body, label %ret_block
428
429.inner_loop_body:
430  %descriptor = load <4 x i32>, <4 x i32> addrspace(4)* %descptr, align 16, !invariant.load !0
431  %load1result = call float @llvm.SI.load.const.v4i32(<4 x i32> %descriptor, i32 0)
432  store float %load1result, float addrspace(1)* undef
433  %inner_br2 = icmp uge i32 %1, 10
434  br i1 %inner_br2, label %.inner_loop_header, label %.outer_loop_body
435
436.outer_loop_body:
437  %offset = shl i32 %loopctr.2, 6
438  %load2result = call float @llvm.SI.load.const.v4i32(<4 x i32> %descriptor, i32 %offset)
439  %outer_br = fcmp ueq float %load2result, 0x0
440  br i1 %outer_br, label %.outer_loop_header, label %ret_block
441}
442
443; SMRD load with a non-const offset
444; GCN-LABEL: {{^}}smrd_load_nonconst0:
445; SIVIGFX9: s_buffer_load_dword s{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
446; SIVIGFX9: s_buffer_load_dword s{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
447; CI: s_buffer_load_dword s{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
448; CI: s_buffer_load_dword s{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
449; GCN: s_endpgm
450define amdgpu_ps void @smrd_load_nonconst0(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in, i32 inreg %ncoff) #0 {
451main_body:
452  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
453  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
454  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 %ncoff)
455  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
456  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 %ncoff, i32 0)
457  %s.buffer.float = bitcast i32 %s.buffer to float
458  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
459  ret void
460}
461
462; SMRD load with a non-const non-uniform offset
463; GCN-LABEL: {{^}}smrd_load_nonconst1:
464; SIVIGFX9: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
465; SIVIGFX9: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
466; CI: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
467; CI: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
468; GCN: s_endpgm
469define amdgpu_ps void @smrd_load_nonconst1(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in, i32 %ncoff) #0 {
470main_body:
471  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
472  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
473  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 %ncoff)
474  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
475  %s.buffer = call i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32> %tmp22, i32 %ncoff, i32 0)
476  %s.buffer.float = bitcast i32 %s.buffer to float
477  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
478  ret void
479}
480
481; SMRD load with a non-const non-uniform offset of > 4 dwords (requires splitting)
482; GCN-LABEL: {{^}}smrd_load_nonconst2:
483; SIVIGFX9: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
484; SIVIGFX9: buffer_load_dwordx4 v[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
485; CI: buffer_load_dword v{{[0-9]+}}, v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
486; CI: buffer_load_dwordx4 v[{{[0-9]+:[0-9]+}}], v{{[0-9]+}}, s[{{[0-9]+:[0-9]+}}], 0 offen
487; GCN: s_endpgm
488define amdgpu_ps void @smrd_load_nonconst2(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in, i32 %ncoff) #0 {
489main_body:
490  %tmp = getelementptr <4 x i32>, <4 x i32> addrspace(4)* %arg, i32 0
491  %tmp20 = load <4 x i32>, <4 x i32> addrspace(4)* %tmp
492  %tmp21 = call float @llvm.SI.load.const.v4i32(<4 x i32> %tmp20, i32 %ncoff)
493  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
494  %s.buffer = call <8 x i32> @llvm.amdgcn.s.buffer.load.v8i32(<4 x i32> %tmp22, i32 %ncoff, i32 0)
495  %s.buffer.elt = extractelement <8 x i32> %s.buffer, i32 1
496  %s.buffer.float = bitcast i32 %s.buffer.elt to float
497  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %tmp21, float %tmp21, float %tmp21, float %s.buffer.float, i1 true, i1 true) #0
498  ret void
499}
500
501; SMRD load with a non-const non-uniform offset of > 4 dwords (requires splitting)
502; GCN-LABEL: {{^}}smrd_load_nonconst3:
503; GCN-DAG: buffer_load_dwordx4 v[0:3], v{{[0-9]+}}, s[0:3], 0 offen ;
504; GCN-DAG: buffer_load_dwordx4 v[4:7], v{{[0-9]+}}, s[0:3], 0 offen offset:16 ;
505; GCN-DAG: buffer_load_dwordx4 v[8:11], v{{[0-9]+}}, s[0:3], 0 offen offset:32 ;
506; GCN-DAG: buffer_load_dwordx4 v[12:15], v{{[0-9]+}}, s[0:3], 0 offen offset:48 ;
507; GCN: ; return to shader part epilog
508define amdgpu_ps <16 x float> @smrd_load_nonconst3(<4 x i32> inreg %rsrc, i32 %off) #0 {
509main_body:
510  %ld = call <16 x i32> @llvm.amdgcn.s.buffer.load.v16i32(<4 x i32> %rsrc, i32 %off, i32 0)
511  %bc = bitcast <16 x i32> %ld to <16 x float>
512  ret <16 x float> %bc
513}
514
515; GCN-LABEL: {{^}}smrd_load_nonconst4:
516; SICI: v_add_i32_e32 v{{[0-9]+}}, vcc, 0xff8, v0 ;
517; SICI-DAG: buffer_load_dwordx4 v[0:3], v{{[0-9]+}}, s[0:3], 0 offen ;
518; SICI-DAG: buffer_load_dwordx4 v[4:7], v{{[0-9]+}}, s[0:3], 0 offen offset:16 ;
519; SICI-DAG: buffer_load_dwordx4 v[8:11], v{{[0-9]+}}, s[0:3], 0 offen offset:32 ;
520; SICI-DAG: buffer_load_dwordx4 v[12:15], v{{[0-9]+}}, s[0:3], 0 offen offset:48 ;
521; VIGFX9-DAG: buffer_load_dwordx4 v[0:3], v{{[0-9]+}}, s[0:3], 56 offen offset:4032 ;
522; VIGFX9-DAG: buffer_load_dwordx4 v[4:7], v{{[0-9]+}}, s[0:3], 56 offen offset:4048 ;
523; VIGFX9-DAG: buffer_load_dwordx4 v[8:11], v{{[0-9]+}}, s[0:3], 56 offen offset:4064 ;
524; VIGFX9-DAG: buffer_load_dwordx4 v[12:15], v{{[0-9]+}}, s[0:3], 56 offen offset:4080 ;
525; GCN: ; return to shader part epilog
526define amdgpu_ps <16 x float> @smrd_load_nonconst4(<4 x i32> inreg %rsrc, i32 %off) #0 {
527main_body:
528  %off.2 = add i32 %off, 4088
529  %ld = call <16 x i32> @llvm.amdgcn.s.buffer.load.v16i32(<4 x i32> %rsrc, i32 %off.2, i32 0)
530  %bc = bitcast <16 x i32> %ld to <16 x float>
531  ret <16 x float> %bc
532}
533
534; GCN-LABEL: {{^}}smrd_load_nonconst5:
535; SICI: v_add_i32_e32 v{{[0-9]+}}, vcc, 0x1004, v0
536; SICI-DAG: buffer_load_dwordx4 v[0:3], v{{[0-9]+}}, s[0:3], 0 offen ;
537; SICI-DAG: buffer_load_dwordx4 v[4:7], v{{[0-9]+}}, s[0:3], 0 offen offset:16 ;
538; SICI-DAG: buffer_load_dwordx4 v[8:11], v{{[0-9]+}}, s[0:3], 0 offen offset:32 ;
539; SICI-DAG: buffer_load_dwordx4 v[12:15], v{{[0-9]+}}, s[0:3], 0 offen offset:48 ;
540; VIGFX9: s_movk_i32 s4, 0xfc0
541; VIGFX9-DAG: buffer_load_dwordx4 v[0:3], v{{[0-9]+}}, s[0:3], s4 offen offset:68 ;
542; VIGFX9-DAG: buffer_load_dwordx4 v[4:7], v{{[0-9]+}}, s[0:3], s4 offen offset:84 ;
543; VIGFX9-DAG: buffer_load_dwordx4 v[8:11], v{{[0-9]+}}, s[0:3], s4 offen offset:100 ;
544; VIGFX9-DAG: buffer_load_dwordx4 v[12:15], v{{[0-9]+}}, s[0:3], s4 offen offset:116 ;
545; GCN: ; return to shader part epilog
546define amdgpu_ps <16 x float> @smrd_load_nonconst5(<4 x i32> inreg %rsrc, i32 %off) #0 {
547main_body:
548  %off.2 = add i32 %off, 4100
549  %ld = call <16 x i32> @llvm.amdgcn.s.buffer.load.v16i32(<4 x i32> %rsrc, i32 %off.2, i32 0)
550  %bc = bitcast <16 x i32> %ld to <16 x float>
551  ret <16 x float> %bc
552}
553
554; SMRD load dwordx2
555; GCN-LABEL: {{^}}smrd_load_dwordx2:
556; SIVIGFX9: s_buffer_load_dwordx2 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
557; CI: s_buffer_load_dwordx2 s[{{[0-9]+:[0-9]+}}], s[{{[0-9]+:[0-9]+}}], s{{[0-9]+}}
558; GCN: s_endpgm
559define amdgpu_ps void @smrd_load_dwordx2(<4 x i32> addrspace(4)* inreg %arg, <4 x i32> addrspace(4)* inreg %arg1, <32 x i8> addrspace(4)* inreg %arg2, i32 inreg %arg3, <2 x i32> %arg4, <2 x i32> %arg5, <2 x i32> %arg6, <3 x i32> %arg7, <2 x i32> %arg8, <2 x i32> %arg9, <2 x i32> %arg10, float %arg11, float %arg12, float %arg13, float %arg14, float %arg15, float %arg16, float %arg17, float %arg18, float %arg19, <4 x i32> addrspace(4)* inreg %in, i32 inreg %ncoff) #0 {
560main_body:
561  %tmp22 = load <4 x i32>, <4 x i32> addrspace(4)* %in
562  %s.buffer = call <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32> %tmp22, i32 %ncoff, i32 0)
563  %s.buffer.float = bitcast <2 x i32> %s.buffer to <2 x float>
564  %r.1 = extractelement <2 x float> %s.buffer.float, i32 0
565  %r.2 = extractelement <2 x float> %s.buffer.float, i32 1
566  call void @llvm.amdgcn.exp.f32(i32 0, i32 15, float %r.1, float %r.1, float %r.1, float %r.2, i1 true, i1 true) #0
567  ret void
568}
569
570; GCN-LABEL: {{^}}smrd_uniform_loop:
571;
572; TODO: we should keep the loop counter in an SGPR
573;
574; GCN: v_readfirstlane_b32
575; GCN: s_buffer_load_dword
576define amdgpu_ps float @smrd_uniform_loop(<4 x i32> inreg %desc, i32 %bound) #0 {
577main_body:
578  br label %loop
579
580loop:
581  %counter = phi i32 [ 0, %main_body ], [ %counter.next, %loop ]
582  %sum = phi float [ 0.0, %main_body ], [ %sum.next, %loop ]
583  %offset = shl i32 %counter, 2
584  %v = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %offset)
585  %sum.next = fadd float %sum, %v
586  %counter.next = add i32 %counter, 1
587  %cc = icmp uge i32 %counter.next, %bound
588  br i1 %cc, label %exit, label %loop
589
590exit:
591  ret float %sum.next
592}
593
594
595; GCN-LABEL: {{^}}smrd_uniform_loop2:
596; (this test differs from smrd_uniform_loop by the more complex structure of phis,
597; which used to confuse the DivergenceAnalysis after structurization)
598;
599; TODO: we should keep the loop counter in an SGPR and use an S_BUFFER_LOAD
600;
601; GCN: buffer_load_dword
602define amdgpu_ps float @smrd_uniform_loop2(<4 x i32> inreg %desc, i32 %bound, i32 %bound.a) #0 {
603main_body:
604  br label %loop
605
606loop:
607  %counter = phi i32 [ 0, %main_body ], [ %counter.next, %loop.a ], [ %counter.next, %loop.b ]
608  %sum = phi float [ 0.0, %main_body ], [ %sum.next, %loop.a ], [ %sum.next.b, %loop.b ]
609  %offset = shl i32 %counter, 2
610  %v = call float @llvm.SI.load.const.v4i32(<4 x i32> %desc, i32 %offset)
611  %sum.next = fadd float %sum, %v
612  %counter.next = add i32 %counter, 1
613  %cc = icmp uge i32 %counter.next, %bound
614  br i1 %cc, label %exit, label %loop.a
615
616loop.a:
617  %cc.a = icmp uge i32 %counter.next, %bound.a
618  br i1 %cc, label %loop, label %loop.b
619
620loop.b:
621  %sum.next.b = fadd float %sum.next, 1.0
622  br label %loop
623
624exit:
625  ret float %sum.next
626}
627
628; This test checks that the load after some control flow with an offset based
629; on a divergent shader input is correctly recognized as divergent. This was
630; reduced from an actual regression. Yes, the %unused argument matters, as
631; well as the fact that %arg4 is a vector.
632;
633; GCN-LABEL: {{^}}arg_divergence:
634; GCN: buffer_load_dword v0, v0,
635; GCN-NEXT: s_waitcnt
636; GCN-NEXT: ; return to shader part epilog
637define amdgpu_cs float @arg_divergence(i32 inreg %unused, <3 x i32> %arg4) #0 {
638main_body:
639  br i1 undef, label %if1, label %endif1
640
641if1:                                              ; preds = %main_body
642  store i32 0, i32 addrspace(3)* undef, align 4
643  br label %endif1
644
645endif1:                                           ; preds = %if1, %main_body
646  %tmp13 = extractelement <3 x i32> %arg4, i32 0
647  %tmp97 = call float @llvm.SI.load.const.v4i32(<4 x i32> undef, i32 %tmp13)
648  ret float %tmp97
649}
650
651; GCN-LABEL: {{^}}s_buffer_load_f32:
652; GCN: s_buffer_load_dword s0, s[0:3], s4
653define amdgpu_ps void @s_buffer_load_f32(<4 x i32> inreg %rsrc, i32 inreg %offset) {
654  %sgpr = call float @llvm.amdgcn.s.buffer.load.f32(<4 x i32> %rsrc, i32 %offset, i32 0)
655  call void asm sideeffect "; use $0", "s"(float %sgpr)
656  ret void
657}
658
659; GCN-LABEL: {{^}}s_buffer_load_v2f32:
660; GCN: s_buffer_load_dwordx2 s[0:1], s[0:3], s4
661define amdgpu_ps void @s_buffer_load_v2f32(<4 x i32> inreg %rsrc, i32 inreg %offset) {
662  %sgpr = call <2 x float> @llvm.amdgcn.s.buffer.load.v2f32(<4 x i32> %rsrc, i32 %offset, i32 0)
663  call void asm sideeffect "; use $0", "s"(<2 x float> %sgpr)
664  ret void
665}
666
667; GCN-LABEL: {{^}}s_buffer_load_v4f32:
668; GCN: s_buffer_load_dwordx4 s[0:3], s[0:3], s4
669define amdgpu_ps void @s_buffer_load_v4f32(<4 x i32> inreg %rsrc, i32 inreg %offset) {
670  %sgpr = call <4 x float> @llvm.amdgcn.s.buffer.load.v4f32(<4 x i32> %rsrc, i32 %offset, i32 0)
671  call void asm sideeffect "; use $0", "s"(<4 x float> %sgpr)
672  ret void
673}
674
675; GCN-LABEL: {{^}}s_buffer_load_v8f32:
676; GCN: s_buffer_load_dwordx8 s[0:7], s[0:3], s4
677define amdgpu_ps void @s_buffer_load_v8f32(<4 x i32> inreg %rsrc, i32 inreg %offset) {
678  %sgpr = call <8 x float> @llvm.amdgcn.s.buffer.load.v8f32(<4 x i32> %rsrc, i32 %offset, i32 0)
679  call void asm sideeffect "; use $0", "s"(<8 x float> %sgpr)
680  ret void
681}
682
683; GCN-LABEL: {{^}}s_buffer_load_v16f32:
684; GCN: s_buffer_load_dwordx16 s[0:15], s[0:3], s4
685define amdgpu_ps void @s_buffer_load_v16f32(<4 x i32> inreg %rsrc, i32 inreg %offset) {
686  %sgpr = call <16 x float> @llvm.amdgcn.s.buffer.load.v16f32(<4 x i32> %rsrc, i32 %offset, i32 0)
687  call void asm sideeffect "; use $0", "s"(<16 x float> %sgpr)
688  ret void
689}
690
691declare void @llvm.amdgcn.exp.f32(i32, i32, float, float, float, float, i1, i1) #0
692declare float @llvm.SI.load.const.v4i32(<4 x i32>, i32) #1
693declare float @llvm.amdgcn.interp.p1(float, i32, i32, i32) #2
694declare float @llvm.amdgcn.interp.p2(float, float, i32, i32, i32) #2
695declare i32 @llvm.amdgcn.s.buffer.load.i32(<4 x i32>, i32, i32)
696declare <2 x i32> @llvm.amdgcn.s.buffer.load.v2i32(<4 x i32>, i32, i32)
697declare <4 x i32> @llvm.amdgcn.s.buffer.load.v4i32(<4 x i32>, i32, i32)
698declare <8 x i32> @llvm.amdgcn.s.buffer.load.v8i32(<4 x i32>, i32, i32)
699declare <16 x i32> @llvm.amdgcn.s.buffer.load.v16i32(<4 x i32>, i32, i32)
700
701declare float @llvm.amdgcn.s.buffer.load.f32(<4 x i32>, i32, i32)
702declare <2 x float> @llvm.amdgcn.s.buffer.load.v2f32(<4 x i32>, i32, i32)
703declare <4 x float> @llvm.amdgcn.s.buffer.load.v4f32(<4 x i32>, i32, i32)
704declare <8 x float> @llvm.amdgcn.s.buffer.load.v8f32(<4 x i32>, i32, i32)
705declare <16 x float> @llvm.amdgcn.s.buffer.load.v16f32(<4 x i32>, i32, i32)
706
707attributes #0 = { nounwind }
708attributes #1 = { nounwind readnone }
709attributes #2 = { nounwind readnone speculatable }
710
711!0 = !{}
712