xref: /llvm-project/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.raw.ptr.tbuffer.load.ll (revision 9e9907f1cfa424366fba58d9520f9305b537cec9) 
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2;RUN: llc < %s -mtriple=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck -check-prefix=PREGFX10 %s
3;RUN: llc < %s -mtriple=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck -check-prefix=PREGFX10 %s
4;RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1010 -verify-machineinstrs | FileCheck -check-prefix=GFX10 %s
5;RUN: llc < %s -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs | FileCheck -check-prefix=GFX11 %s
6
7define amdgpu_vs {<4 x float>, <4 x float>, <4 x float>, <4 x float>} @tbuffer_load(ptr addrspace(8) inreg) {
8; PREGFX10-LABEL: tbuffer_load:
9; PREGFX10:       ; %bb.0: ; %main_body
10; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT]
11; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] glc
12; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] slc
13; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[12:15], off, s[0:3], 0 format:[BUF_DATA_FORMAT_10_11_11,BUF_NUM_FORMAT_SNORM] glc
14; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
15; PREGFX10-NEXT:    ; return to shader part epilog
16;
17; GFX10-LABEL: tbuffer_load:
18; GFX10:       ; %bb.0: ; %main_body
19; GFX10-NEXT:    s_clause 0x3
20; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:78
21; GFX10-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_FMT_32_32_SINT] glc
22; GFX10-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT] slc
23; GFX10-NEXT:    tbuffer_load_format_xyzw v[12:15], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT] glc dlc
24; GFX10-NEXT:    s_waitcnt vmcnt(0)
25; GFX10-NEXT:    ; return to shader part epilog
26;
27; GFX11-LABEL: tbuffer_load:
28; GFX11:       ; %bb.0: ; %main_body
29; GFX11-NEXT:    s_clause 0x3
30; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:78
31; GFX11-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] glc
32; GFX11-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT] slc
33; GFX11-NEXT:    tbuffer_load_format_xyzw v[12:15], off, s[0:3], 0 format:[BUF_FMT_32_FLOAT] glc dlc
34; GFX11-NEXT:    s_waitcnt vmcnt(0)
35; GFX11-NEXT:    ; return to shader part epilog
36main_body:
37    %vdata     = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 0, i32 0, i32 78, i32 0)
38    %vdata_glc = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 0, i32 0, i32 63, i32 1)
39    %vdata_slc = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 0, i32 0, i32 22, i32 2)
40    %vdata_f32 = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 0, i32 0, i32 22, i32 5)
41    %vdata.f     = bitcast <4 x i32> %vdata to <4 x float>
42    %vdata_glc.f = bitcast <4 x i32> %vdata_glc to <4 x float>
43    %vdata_slc.f = bitcast <4 x i32> %vdata_slc to <4 x float>
44    %r0 = insertvalue {<4 x float>, <4 x float>, <4 x float>, <4 x float>} undef, <4 x float> %vdata.f, 0
45    %r1 = insertvalue {<4 x float>, <4 x float>, <4 x float>, <4 x float>} %r0, <4 x float> %vdata_glc.f, 1
46    %r2 = insertvalue {<4 x float>, <4 x float>, <4 x float>, <4 x float>} %r1, <4 x float> %vdata_slc.f, 2
47    %r3 = insertvalue {<4 x float>, <4 x float>, <4 x float>, <4 x float>} %r2, <4 x float> %vdata_f32, 3
48    ret {<4 x float>, <4 x float>, <4 x float>, <4 x float>} %r3
49}
50
51define amdgpu_vs <4 x float> @tbuffer_load_immoffs(ptr addrspace(8) inreg) {
52; PREGFX10-LABEL: tbuffer_load_immoffs:
53; PREGFX10:       ; %bb.0: ; %main_body
54; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT] offset:42
55; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
56; PREGFX10-NEXT:    ; return to shader part epilog
57;
58; GFX10-LABEL: tbuffer_load_immoffs:
59; GFX10:       ; %bb.0: ; %main_body
60; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:78 offset:42
61; GFX10-NEXT:    s_waitcnt vmcnt(0)
62; GFX10-NEXT:    ; return to shader part epilog
63;
64; GFX11-LABEL: tbuffer_load_immoffs:
65; GFX11:       ; %bb.0: ; %main_body
66; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:78 offset:42
67; GFX11-NEXT:    s_waitcnt vmcnt(0)
68; GFX11-NEXT:    ; return to shader part epilog
69main_body:
70    %vdata   = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 42, i32 0, i32 78, i32 0)
71    %vdata.f = bitcast <4 x i32> %vdata to <4 x float>
72    ret <4 x float> %vdata.f
73}
74
75define amdgpu_ps <4 x float> @buffer_load_voffset_large_12bit(ptr addrspace(8) inreg) {
76; PREGFX10-LABEL: buffer_load_voffset_large_12bit:
77; PREGFX10:       ; %bb.0: ; %main_body
78; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offset:4092
79; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
80; PREGFX10-NEXT:    ; return to shader part epilog
81;
82; GFX10-LABEL: buffer_load_voffset_large_12bit:
83; GFX10:       ; %bb.0: ; %main_body
84; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offset:4092
85; GFX10-NEXT:    s_waitcnt vmcnt(0)
86; GFX10-NEXT:    ; return to shader part epilog
87;
88; GFX11-LABEL: buffer_load_voffset_large_12bit:
89; GFX11:       ; %bb.0: ; %main_body
90; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offset:4092
91; GFX11-NEXT:    s_waitcnt vmcnt(0)
92; GFX11-NEXT:    ; return to shader part epilog
93main_body:
94  %data = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 4092, i32 0, i32 63, i32 0)
95  ret <4 x float> %data
96}
97
98define amdgpu_ps <4 x float> @tbuffer_load_voffset_large_13bit(ptr addrspace(8) inreg) {
99; PREGFX10-LABEL: tbuffer_load_voffset_large_13bit:
100; PREGFX10:       ; %bb.0: ; %main_body
101; PREGFX10-NEXT:    v_mov_b32_e32 v0, 0x1000
102; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092
103; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
104; PREGFX10-NEXT:    ; return to shader part epilog
105;
106; GFX10-LABEL: tbuffer_load_voffset_large_13bit:
107; GFX10:       ; %bb.0: ; %main_body
108; GFX10-NEXT:    v_mov_b32_e32 v0, 0x1000
109; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092
110; GFX10-NEXT:    s_waitcnt vmcnt(0)
111; GFX10-NEXT:    ; return to shader part epilog
112;
113; GFX11-LABEL: tbuffer_load_voffset_large_13bit:
114; GFX11:       ; %bb.0: ; %main_body
115; GFX11-NEXT:    v_mov_b32_e32 v0, 0x1000
116; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092
117; GFX11-NEXT:    s_waitcnt vmcnt(0)
118; GFX11-NEXT:    ; return to shader part epilog
119main_body:
120  %data = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 8188, i32 0, i32 63, i32 0)
121  ret <4 x float> %data
122}
123
124define amdgpu_ps <4 x float> @tbuffer_load_voffset_large_16bit(ptr addrspace(8) inreg) {
125; PREGFX10-LABEL: tbuffer_load_voffset_large_16bit:
126; PREGFX10:       ; %bb.0: ; %main_body
127; PREGFX10-NEXT:    v_mov_b32_e32 v0, 0xf000
128; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092
129; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
130; PREGFX10-NEXT:    ; return to shader part epilog
131;
132; GFX10-LABEL: tbuffer_load_voffset_large_16bit:
133; GFX10:       ; %bb.0: ; %main_body
134; GFX10-NEXT:    v_mov_b32_e32 v0, 0xf000
135; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092
136; GFX10-NEXT:    s_waitcnt vmcnt(0)
137; GFX10-NEXT:    ; return to shader part epilog
138;
139; GFX11-LABEL: tbuffer_load_voffset_large_16bit:
140; GFX11:       ; %bb.0: ; %main_body
141; GFX11-NEXT:    v_mov_b32_e32 v0, 0xf000
142; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092
143; GFX11-NEXT:    s_waitcnt vmcnt(0)
144; GFX11-NEXT:    ; return to shader part epilog
145main_body:
146  %data = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 65532, i32 0, i32 63, i32 0)
147  ret <4 x float> %data
148}
149
150define amdgpu_ps <4 x float> @tbuffer_load_voffset_large_23bit(ptr addrspace(8) inreg) {
151; PREGFX10-LABEL: tbuffer_load_voffset_large_23bit:
152; PREGFX10:       ; %bb.0: ; %main_body
153; PREGFX10-NEXT:    v_mov_b32_e32 v0, 0x7ff000
154; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092
155; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
156; PREGFX10-NEXT:    ; return to shader part epilog
157;
158; GFX10-LABEL: tbuffer_load_voffset_large_23bit:
159; GFX10:       ; %bb.0: ; %main_body
160; GFX10-NEXT:    v_mov_b32_e32 v0, 0x7ff000
161; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092
162; GFX10-NEXT:    s_waitcnt vmcnt(0)
163; GFX10-NEXT:    ; return to shader part epilog
164;
165; GFX11-LABEL: tbuffer_load_voffset_large_23bit:
166; GFX11:       ; %bb.0: ; %main_body
167; GFX11-NEXT:    v_mov_b32_e32 v0, 0x7ff000
168; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092
169; GFX11-NEXT:    s_waitcnt vmcnt(0)
170; GFX11-NEXT:    ; return to shader part epilog
171main_body:
172  %data = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 8388604, i32 0, i32 63, i32 0)
173  ret <4 x float> %data
174}
175
176define amdgpu_ps <4 x float> @tbuffer_load_voffset_large_24bit(ptr addrspace(8) inreg) {
177; PREGFX10-LABEL: tbuffer_load_voffset_large_24bit:
178; PREGFX10:       ; %bb.0: ; %main_body
179; PREGFX10-NEXT:    v_mov_b32_e32 v0, 0xfff000
180; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_SSCALED] offen offset:4092
181; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
182; PREGFX10-NEXT:    ; return to shader part epilog
183;
184; GFX10-LABEL: tbuffer_load_voffset_large_24bit:
185; GFX10:       ; %bb.0: ; %main_body
186; GFX10-NEXT:    v_mov_b32_e32 v0, 0xfff000
187; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_SINT] offen offset:4092
188; GFX10-NEXT:    s_waitcnt vmcnt(0)
189; GFX10-NEXT:    ; return to shader part epilog
190;
191; GFX11-LABEL: tbuffer_load_voffset_large_24bit:
192; GFX11:       ; %bb.0: ; %main_body
193; GFX11-NEXT:    v_mov_b32_e32 v0, 0xfff000
194; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT] offen offset:4092
195; GFX11-NEXT:    s_waitcnt vmcnt(0)
196; GFX11-NEXT:    ; return to shader part epilog
197main_body:
198  %data = call <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8) %0, i32 16777212, i32 0, i32 63, i32 0)
199  ret <4 x float> %data
200}
201
202define amdgpu_vs {<4 x float>, <4 x float>, <4 x float>} @tbuffer_load_immoffs_large(ptr addrspace(8) inreg, i32 inreg %soffs) {
203; PREGFX10-LABEL: tbuffer_load_immoffs_large:
204; PREGFX10:       ; %bb.0:
205; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 61 format:[BUF_DATA_FORMAT_RESERVED_15,BUF_NUM_FORMAT_USCALED] offset:4095
206; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], s4 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_SSCALED] offset:73
207; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], s4 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT] offset:1
208; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
209; PREGFX10-NEXT:    ; return to shader part epilog
210;
211; GFX10-LABEL: tbuffer_load_immoffs_large:
212; GFX10:       ; %bb.0:
213; GFX10-NEXT:    s_clause 0x2
214; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 61 format:[BUF_FMT_10_10_10_2_SSCALED] offset:4095
215; GFX10-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], s4 format:[BUF_FMT_32_32_UINT] offset:73
216; GFX10-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], s4 format:[BUF_FMT_32_32_32_32_FLOAT] offset:1
217; GFX10-NEXT:    s_waitcnt vmcnt(0)
218; GFX10-NEXT:    ; return to shader part epilog
219;
220; GFX11-LABEL: tbuffer_load_immoffs_large:
221; GFX11:       ; %bb.0:
222; GFX11-NEXT:    s_clause 0x2
223; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], off, s[0:3], 61 format:[BUF_FMT_8_8_8_8_SINT] offset:4095
224; GFX11-NEXT:    tbuffer_load_format_xyzw v[4:7], off, s[0:3], s4 format:[BUF_FMT_32_32_32_32_SINT] offset:73
225; GFX11-NEXT:    tbuffer_load_format_xyzw v[8:11], off, s[0:3], s4 format:77 offset:1
226; GFX11-NEXT:    s_waitcnt vmcnt(0)
227; GFX11-NEXT:    ; return to shader part epilog
228    %vdata     = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 4095, i32 61, i32 47, i32 0)
229    %vdata_glc = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 73, i32 %soffs, i32 62, i32 0)
230    %vdata_slc = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 1, i32 %soffs, i32 77, i32 0)
231    %vdata.f     = bitcast <4 x i32> %vdata to <4 x float>
232    %vdata_glc.f = bitcast <4 x i32> %vdata_glc to <4 x float>
233    %vdata_slc.f = bitcast <4 x i32> %vdata_slc to <4 x float>
234    %r0 = insertvalue {<4 x float>, <4 x float>, <4 x float>} undef, <4 x float> %vdata.f, 0
235    %r1 = insertvalue {<4 x float>, <4 x float>, <4 x float>} %r0, <4 x float> %vdata_glc.f, 1
236    %r2 = insertvalue {<4 x float>, <4 x float>, <4 x float>} %r1, <4 x float> %vdata_slc.f, 2
237    ret {<4 x float>, <4 x float>, <4 x float>} %r2
238}
239
240define amdgpu_vs <4 x float> @tbuffer_load_ofs(ptr addrspace(8) inreg, i32 %voffs) {
241; PREGFX10-LABEL: tbuffer_load_ofs:
242; PREGFX10:       ; %bb.0: ; %main_body
243; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT] offen
244; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
245; PREGFX10-NEXT:    ; return to shader part epilog
246;
247; GFX10-LABEL: tbuffer_load_ofs:
248; GFX10:       ; %bb.0: ; %main_body
249; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:78 offen
250; GFX10-NEXT:    s_waitcnt vmcnt(0)
251; GFX10-NEXT:    ; return to shader part epilog
252;
253; GFX11-LABEL: tbuffer_load_ofs:
254; GFX11:       ; %bb.0: ; %main_body
255; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:78 offen
256; GFX11-NEXT:    s_waitcnt vmcnt(0)
257; GFX11-NEXT:    ; return to shader part epilog
258main_body:
259    %vdata   = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 %voffs, i32 0, i32 78, i32 0)
260    %vdata.f = bitcast <4 x i32> %vdata to <4 x float>
261    ret <4 x float> %vdata.f
262}
263
264define amdgpu_vs <4 x float> @tbuffer_load_ofs_imm(ptr addrspace(8) inreg, i32 %voffs) {
265; PREGFX10-LABEL: tbuffer_load_ofs_imm:
266; PREGFX10:       ; %bb.0: ; %main_body
267; PREGFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32_32,BUF_NUM_FORMAT_UINT] offen offset:52
268; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
269; PREGFX10-NEXT:    ; return to shader part epilog
270;
271; GFX10-LABEL: tbuffer_load_ofs_imm:
272; GFX10:       ; %bb.0: ; %main_body
273; GFX10-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:78 offen offset:52
274; GFX10-NEXT:    s_waitcnt vmcnt(0)
275; GFX10-NEXT:    ; return to shader part epilog
276;
277; GFX11-LABEL: tbuffer_load_ofs_imm:
278; GFX11:       ; %bb.0: ; %main_body
279; GFX11-NEXT:    tbuffer_load_format_xyzw v[0:3], v0, s[0:3], 0 format:78 offen offset:52
280; GFX11-NEXT:    s_waitcnt vmcnt(0)
281; GFX11-NEXT:    ; return to shader part epilog
282main_body:
283    %ofs = add i32 %voffs, 52
284    %vdata   = call <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8) %0, i32 %ofs, i32 0, i32 78, i32 0)
285    %vdata.f = bitcast <4 x i32> %vdata to <4 x float>
286    ret <4 x float> %vdata.f
287}
288
289define amdgpu_vs <2 x float> @buffer_load_xy(ptr addrspace(8) inreg %rsrc) {
290; PREGFX10-LABEL: buffer_load_xy:
291; PREGFX10:       ; %bb.0:
292; PREGFX10-NEXT:    tbuffer_load_format_xy v[0:1], off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT]
293; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
294; PREGFX10-NEXT:    ; return to shader part epilog
295;
296; GFX10-LABEL: buffer_load_xy:
297; GFX10:       ; %bb.0:
298; GFX10-NEXT:    tbuffer_load_format_xy v[0:1], off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT]
299; GFX10-NEXT:    s_waitcnt vmcnt(0)
300; GFX10-NEXT:    ; return to shader part epilog
301;
302; GFX11-LABEL: buffer_load_xy:
303; GFX11:       ; %bb.0:
304; GFX11-NEXT:    tbuffer_load_format_xy v[0:1], off, s[0:3], 0 format:77
305; GFX11-NEXT:    s_waitcnt vmcnt(0)
306; GFX11-NEXT:    ; return to shader part epilog
307    %vdata = call <2 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v2i32(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 77, i32 0)
308    %vdata.f = bitcast <2 x i32> %vdata to <2 x float>
309    ret <2 x float> %vdata.f
310}
311
312define amdgpu_vs float @buffer_load_x(ptr addrspace(8) inreg %rsrc) {
313; PREGFX10-LABEL: buffer_load_x:
314; PREGFX10:       ; %bb.0:
315; PREGFX10-NEXT:    tbuffer_load_format_x v0, off, s[0:3], 0 format:[BUF_DATA_FORMAT_32_32_32,BUF_NUM_FORMAT_UINT]
316; PREGFX10-NEXT:    s_waitcnt vmcnt(0)
317; PREGFX10-NEXT:    ; return to shader part epilog
318;
319; GFX10-LABEL: buffer_load_x:
320; GFX10:       ; %bb.0:
321; GFX10-NEXT:    tbuffer_load_format_x v0, off, s[0:3], 0 format:[BUF_FMT_32_32_32_32_FLOAT]
322; GFX10-NEXT:    s_waitcnt vmcnt(0)
323; GFX10-NEXT:    ; return to shader part epilog
324;
325; GFX11-LABEL: buffer_load_x:
326; GFX11:       ; %bb.0:
327; GFX11-NEXT:    tbuffer_load_format_x v0, off, s[0:3], 0 format:77
328; GFX11-NEXT:    s_waitcnt vmcnt(0)
329; GFX11-NEXT:    ; return to shader part epilog
330    %vdata = call i32 @llvm.amdgcn.raw.ptr.tbuffer.load.i32(ptr addrspace(8) %rsrc, i32 0, i32 0, i32 77, i32 0)
331    %vdata.f = bitcast i32 %vdata to float
332    ret float %vdata.f
333}
334
335declare i32 @llvm.amdgcn.raw.ptr.tbuffer.load.i32(ptr addrspace(8), i32, i32, i32, i32)
336declare <2 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v2i32(ptr addrspace(8), i32, i32, i32, i32)
337declare <4 x i32> @llvm.amdgcn.raw.ptr.tbuffer.load.v4i32(ptr addrspace(8), i32, i32, i32, i32)
338declare <4 x float> @llvm.amdgcn.raw.ptr.tbuffer.load.v4f32(ptr addrspace(8), i32, i32, i32, i32)
339