1 ; RUN: llc -march=amdgcn -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
2 ; RUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
4 ; SI-LABEL: {{^}}load_i8_to_f32:
5 ; SI: buffer_load_ubyte [[LOADREG:v[0-9]+]],
8 ; SI: v_cvt_f32_ubyte0_e32 [[CONV:v[0-9]+]], [[LOADREG]]
9 ; SI: buffer_store_dword [[CONV]],
10 define void @load_i8_to_f32(float addrspace(1)* noalias %out, i8 addrspace(1)* noalias %in) nounwind {
11 %load = load i8, i8 addrspace(1)* %in, align 1
12 %cvt = uitofp i8 %load to float
13 store float %cvt, float addrspace(1)* %out, align 4
17 ; SI-LABEL: {{^}}load_v2i8_to_v2f32:
18 ; SI: buffer_load_ushort [[LOADREG:v[0-9]+]],
22 ; SI-DAG: v_cvt_f32_ubyte1_e32 v[[HIRESULT:[0-9]+]], [[LOADREG]]
23 ; SI-DAG: v_cvt_f32_ubyte0_e32 v[[LORESULT:[0-9]+]], [[LOADREG]]
24 ; SI: buffer_store_dwordx2 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
25 define void @load_v2i8_to_v2f32(<2 x float> addrspace(1)* noalias %out, <2 x i8> addrspace(1)* noalias %in) nounwind {
26 %load = load <2 x i8>, <2 x i8> addrspace(1)* %in, align 2
27 %cvt = uitofp <2 x i8> %load to <2 x float>
28 store <2 x float> %cvt, <2 x float> addrspace(1)* %out, align 16
32 ; SI-LABEL: {{^}}load_v3i8_to_v3f32:
34 ; SI-NOT: v_cvt_f32_ubyte3_e32
35 ; SI-DAG: v_cvt_f32_ubyte2_e32
36 ; SI-DAG: v_cvt_f32_ubyte1_e32
37 ; SI-DAG: v_cvt_f32_ubyte0_e32
38 ; SI: buffer_store_dwordx2 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
39 define void @load_v3i8_to_v3f32(<3 x float> addrspace(1)* noalias %out, <3 x i8> addrspace(1)* noalias %in) nounwind {
40 %load = load <3 x i8>, <3 x i8> addrspace(1)* %in, align 4
41 %cvt = uitofp <3 x i8> %load to <3 x float>
42 store <3 x float> %cvt, <3 x float> addrspace(1)* %out, align 16
46 ; SI-LABEL: {{^}}load_v4i8_to_v4f32:
47 ; SI: buffer_load_dword [[LOADREG:v[0-9]+]]
50 ; SI-DAG: v_cvt_f32_ubyte3_e32 v[[HIRESULT:[0-9]+]], [[LOADREG]]
51 ; SI-DAG: v_cvt_f32_ubyte2_e32 v{{[0-9]+}}, [[LOADREG]]
52 ; SI-DAG: v_cvt_f32_ubyte1_e32 v{{[0-9]+}}, [[LOADREG]]
53 ; SI-DAG: v_cvt_f32_ubyte0_e32 v[[LORESULT:[0-9]+]], [[LOADREG]]
54 ; SI: buffer_store_dwordx4 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
55 define void @load_v4i8_to_v4f32(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
56 %load = load <4 x i8>, <4 x i8> addrspace(1)* %in, align 4
57 %cvt = uitofp <4 x i8> %load to <4 x float>
58 store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
62 ; This should not be adding instructions to shift into the correct
63 ; position in the word for the component.
65 ; SI-LABEL: {{^}}load_v4i8_to_v4f32_unaligned:
66 ; SI: buffer_load_ubyte [[LOADREG3:v[0-9]+]]
67 ; SI: buffer_load_ubyte [[LOADREG2:v[0-9]+]]
68 ; SI: buffer_load_ubyte [[LOADREG1:v[0-9]+]]
69 ; SI: buffer_load_ubyte [[LOADREG0:v[0-9]+]]
70 ; SI-NOT: v_lshlrev_b32
73 ; SI-DAG: v_cvt_f32_ubyte0_e32 v[[LORESULT:[0-9]+]], [[LOADREG0]]
74 ; SI-DAG: v_cvt_f32_ubyte0_e32 v{{[0-9]+}}, [[LOADREG1]]
75 ; SI-DAG: v_cvt_f32_ubyte0_e32 v{{[0-9]+}}, [[LOADREG2]]
76 ; SI-DAG: v_cvt_f32_ubyte0_e32 v[[HIRESULT:[0-9]+]], [[LOADREG3]]
78 ; SI: buffer_store_dwordx4 v{{\[}}[[LORESULT]]:[[HIRESULT]]{{\]}},
79 define void @load_v4i8_to_v4f32_unaligned(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
80 %load = load <4 x i8>, <4 x i8> addrspace(1)* %in, align 1
81 %cvt = uitofp <4 x i8> %load to <4 x float>
82 store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
86 ; XXX - This should really still be able to use the v_cvt_f32_ubyte0
87 ; for each component, but computeKnownBits doesn't handle vectors very
90 ; SI-LABEL: {{^}}load_v4i8_to_v4f32_2_uses:
91 ; SI: buffer_load_ubyte
92 ; SI: buffer_load_ubyte
93 ; SI: buffer_load_ubyte
94 ; SI: buffer_load_ubyte
95 ; SI: v_cvt_f32_ubyte0_e32
96 ; SI: v_cvt_f32_ubyte0_e32
97 ; SI: v_cvt_f32_ubyte0_e32
98 ; SI: v_cvt_f32_ubyte0_e32
100 ; XXX - replace with this when v4i8 loads aren't scalarized anymore.
101 ; XSI: buffer_load_dword
102 ; XSI: v_cvt_f32_u32_e32
103 ; XSI: v_cvt_f32_u32_e32
104 ; XSI: v_cvt_f32_u32_e32
105 ; XSI: v_cvt_f32_u32_e32
107 define void @load_v4i8_to_v4f32_2_uses(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %out2, <4 x i8> addrspace(1)* noalias %in) nounwind {
108 %load = load <4 x i8>, <4 x i8> addrspace(1)* %in, align 4
109 %cvt = uitofp <4 x i8> %load to <4 x float>
110 store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16
111 %add = add <4 x i8> %load, <i8 9, i8 9, i8 9, i8 9> ; Second use of %load
112 store <4 x i8> %add, <4 x i8> addrspace(1)* %out2, align 4
116 ; Make sure this doesn't crash.
117 ; SI-LABEL: {{^}}load_v7i8_to_v7f32:
119 define void @load_v7i8_to_v7f32(<7 x float> addrspace(1)* noalias %out, <7 x i8> addrspace(1)* noalias %in) nounwind {
120 %load = load <7 x i8>, <7 x i8> addrspace(1)* %in, align 1
121 %cvt = uitofp <7 x i8> %load to <7 x float>
122 store <7 x float> %cvt, <7 x float> addrspace(1)* %out, align 16
126 ; SI-LABEL: {{^}}load_v8i8_to_v8f32:
127 ; SI: buffer_load_dwordx2 v{{\[}}[[LOLOAD:[0-9]+]]:[[HILOAD:[0-9]+]]{{\]}},
130 ; SI-DAG: v_cvt_f32_ubyte3_e32 v{{[0-9]+}}, v[[LOLOAD]]
131 ; SI-DAG: v_cvt_f32_ubyte2_e32 v{{[0-9]+}}, v[[LOLOAD]]
132 ; SI-DAG: v_cvt_f32_ubyte1_e32 v{{[0-9]+}}, v[[LOLOAD]]
133 ; SI-DAG: v_cvt_f32_ubyte0_e32 v{{[0-9]+}}, v[[LOLOAD]]
134 ; SI-DAG: v_cvt_f32_ubyte3_e32 v{{[0-9]+}}, v[[HILOAD]]
135 ; SI-DAG: v_cvt_f32_ubyte2_e32 v{{[0-9]+}}, v[[HILOAD]]
136 ; SI-DAG: v_cvt_f32_ubyte1_e32 v{{[0-9]+}}, v[[HILOAD]]
137 ; SI-DAG: v_cvt_f32_ubyte0_e32 v{{[0-9]+}}, v[[HILOAD]]
140 ; SI: buffer_store_dword
141 ; SI: buffer_store_dword
142 ; SI: buffer_store_dword
143 ; SI: buffer_store_dword
144 ; SI: buffer_store_dword
145 ; SI: buffer_store_dword
146 ; SI: buffer_store_dword
147 ; SI: buffer_store_dword
148 define void @load_v8i8_to_v8f32(<8 x float> addrspace(1)* noalias %out, <8 x i8> addrspace(1)* noalias %in) nounwind {
149 %load = load <8 x i8>, <8 x i8> addrspace(1)* %in, align 8
150 %cvt = uitofp <8 x i8> %load to <8 x float>
151 store <8 x float> %cvt, <8 x float> addrspace(1)* %out, align 16
155 ; SI-LABEL: {{^}}i8_zext_inreg_i32_to_f32:
156 ; SI: buffer_load_dword [[LOADREG:v[0-9]+]],
157 ; SI: v_add_i32_e32 [[ADD:v[0-9]+]], vcc, 2, [[LOADREG]]
158 ; SI-NEXT: v_cvt_f32_ubyte0_e32 [[CONV:v[0-9]+]], [[ADD]]
159 ; SI: buffer_store_dword [[CONV]],
160 define void @i8_zext_inreg_i32_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
161 %load = load i32, i32 addrspace(1)* %in, align 4
162 %add = add i32 %load, 2
163 %inreg = and i32 %add, 255
164 %cvt = uitofp i32 %inreg to float
165 store float %cvt, float addrspace(1)* %out, align 4
169 ; SI-LABEL: {{^}}i8_zext_inreg_hi1_to_f32:
170 define void @i8_zext_inreg_hi1_to_f32(float addrspace(1)* noalias %out, i32 addrspace(1)* noalias %in) nounwind {
171 %load = load i32, i32 addrspace(1)* %in, align 4
172 %inreg = and i32 %load, 65280
173 %shr = lshr i32 %inreg, 8
174 %cvt = uitofp i32 %shr to float
175 store float %cvt, float addrspace(1)* %out, align 4
180 ; We don't get these ones because of the zext, but instcombine removes
181 ; them so it shouldn't really matter.
182 define void @i8_zext_i32_to_f32(float addrspace(1)* noalias %out, i8 addrspace(1)* noalias %in) nounwind {
183 %load = load i8, i8 addrspace(1)* %in, align 1
184 %ext = zext i8 %load to i32
185 %cvt = uitofp i32 %ext to float
186 store float %cvt, float addrspace(1)* %out, align 4
190 define void @v4i8_zext_v4i32_to_v4f32(<4 x float> addrspace(1)* noalias %out, <4 x i8> addrspace(1)* noalias %in) nounwind {
191 %load = load <4 x i8>, <4 x i8> addrspace(1)* %in, align 1
192 %ext = zext <4 x i8> %load to <4 x i32>
193 %cvt = uitofp <4 x i32> %ext to <4 x float>
194 store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16