[LoongArch][BF16] Add support for the __bf16 type (#142548) · llvm/llvm-project@0ed5d9a (original) (raw)
``
1
`+
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
`
``
2
`+
; RUN: llc -mtriple=loongarch64 -mattr=+d -target-abi=lp64d < %s | FileCheck --check-prefixes=CHECK,LA64 %s
`
``
3
`+
; RUN: llc -mtriple=loongarch32 -mattr=+d -target-abi=ilp32d < %s | FileCheck --check-prefixes=CHECK,LA32 %s
`
``
4
+
``
5
`+
define void @test_load_store(ptr %p, ptr %q) nounwind {
`
``
6
`+
; CHECK-LABEL: test_load_store:
`
``
7
`+
; CHECK: # %bb.0:
`
``
8
`+
; CHECK-NEXT: ld.h a0,a0, a0,a0, 0
`
``
9
`+
; CHECK-NEXT: st.h a0,a0, a0,a1, 0
`
``
10
`+
; CHECK-NEXT: ret
`
``
11
`+
%a = load bfloat, ptr %p
`
``
12
`+
store bfloat %a, ptr %q
`
``
13
`+
ret void
`
``
14
`+
}
`
``
15
+
``
16
`+
define float @test_fpextend_float(ptr %p) nounwind {
`
``
17
`+
; LA64-LABEL: test_fpextend_float:
`
``
18
`+
; LA64: # %bb.0:
`
``
19
`+
; LA64-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
20
`+
; LA64-NEXT: slli.d a0,a0, a0,a0, 16
`
``
21
`+
; LA64-NEXT: movgr2fr.w fa0,fa0, fa0,a0
`
``
22
`+
; LA64-NEXT: ret
`
``
23
`+
;
`
``
24
`+
; LA32-LABEL: test_fpextend_float:
`
``
25
`+
; LA32: # %bb.0:
`
``
26
`+
; LA32-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
27
`+
; LA32-NEXT: slli.w a0,a0, a0,a0, 16
`
``
28
`+
; LA32-NEXT: movgr2fr.w fa0,fa0, fa0,a0
`
``
29
`+
; LA32-NEXT: ret
`
``
30
`+
%a = load bfloat, ptr %p
`
``
31
`+
%r = fpext bfloat %a to float
`
``
32
`+
ret float %r
`
``
33
`+
}
`
``
34
+
``
35
`+
define double @test_fpextend_double(ptr %p) nounwind {
`
``
36
`+
; LA64-LABEL: test_fpextend_double:
`
``
37
`+
; LA64: # %bb.0:
`
``
38
`+
; LA64-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
39
`+
; LA64-NEXT: slli.d a0,a0, a0,a0, 16
`
``
40
`+
; LA64-NEXT: movgr2fr.w fa0,fa0, fa0,a0
`
``
41
`+
; LA64-NEXT: fcvt.d.s fa0,fa0, fa0,fa0
`
``
42
`+
; LA64-NEXT: ret
`
``
43
`+
;
`
``
44
`+
; LA32-LABEL: test_fpextend_double:
`
``
45
`+
; LA32: # %bb.0:
`
``
46
`+
; LA32-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
47
`+
; LA32-NEXT: slli.w a0,a0, a0,a0, 16
`
``
48
`+
; LA32-NEXT: movgr2fr.w fa0,fa0, fa0,a0
`
``
49
`+
; LA32-NEXT: fcvt.d.s fa0,fa0, fa0,fa0
`
``
50
`+
; LA32-NEXT: ret
`
``
51
`+
%a = load bfloat, ptr %p
`
``
52
`+
%r = fpext bfloat %a to double
`
``
53
`+
ret double %r
`
``
54
`+
}
`
``
55
+
``
56
`+
define void @test_fptrunc_float(float %f, ptr %p) nounwind {
`
``
57
`+
; LA64-LABEL: test_fptrunc_float:
`
``
58
`+
; LA64: # %bb.0:
`
``
59
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, -16
`
``
60
`+
; LA64-NEXT: st.d ra,ra, ra,sp, 8 # 8-byte Folded Spill
`
``
61
`+
; LA64-NEXT: st.d fp,fp, fp,sp, 0 # 8-byte Folded Spill
`
``
62
`+
; LA64-NEXT: move fp,fp, fp,a0
`
``
63
`+
; LA64-NEXT: pcaddu18i $ra, %call36(__truncsfbf2)
`
``
64
`+
; LA64-NEXT: jirl ra,ra, ra,ra, 0
`
``
65
`+
; LA64-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
66
`+
; LA64-NEXT: st.h a0,a0, a0,fp, 0
`
``
67
`+
; LA64-NEXT: ld.d fp,fp, fp,sp, 0 # 8-byte Folded Reload
`
``
68
`+
; LA64-NEXT: ld.d ra,ra, ra,sp, 8 # 8-byte Folded Reload
`
``
69
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, 16
`
``
70
`+
; LA64-NEXT: ret
`
``
71
`+
;
`
``
72
`+
; LA32-LABEL: test_fptrunc_float:
`
``
73
`+
; LA32: # %bb.0:
`
``
74
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, -16
`
``
75
`+
; LA32-NEXT: st.w ra,ra, ra,sp, 12 # 4-byte Folded Spill
`
``
76
`+
; LA32-NEXT: st.w fp,fp, fp,sp, 8 # 4-byte Folded Spill
`
``
77
`+
; LA32-NEXT: move fp,fp, fp,a0
`
``
78
`+
; LA32-NEXT: bl __truncsfbf2
`
``
79
`+
; LA32-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
80
`+
; LA32-NEXT: st.h a0,a0, a0,fp, 0
`
``
81
`+
; LA32-NEXT: ld.w fp,fp, fp,sp, 8 # 4-byte Folded Reload
`
``
82
`+
; LA32-NEXT: ld.w ra,ra, ra,sp, 12 # 4-byte Folded Reload
`
``
83
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, 16
`
``
84
`+
; LA32-NEXT: ret
`
``
85
`+
%a = fptrunc float %f to bfloat
`
``
86
`+
store bfloat %a, ptr %p
`
``
87
`+
ret void
`
``
88
`+
}
`
``
89
+
``
90
`+
define void @test_fptrunc_double(double %d, ptr %p) nounwind {
`
``
91
`+
; LA64-LABEL: test_fptrunc_double:
`
``
92
`+
; LA64: # %bb.0:
`
``
93
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, -16
`
``
94
`+
; LA64-NEXT: st.d ra,ra, ra,sp, 8 # 8-byte Folded Spill
`
``
95
`+
; LA64-NEXT: st.d fp,fp, fp,sp, 0 # 8-byte Folded Spill
`
``
96
`+
; LA64-NEXT: move fp,fp, fp,a0
`
``
97
`+
; LA64-NEXT: pcaddu18i $ra, %call36(__truncdfbf2)
`
``
98
`+
; LA64-NEXT: jirl ra,ra, ra,ra, 0
`
``
99
`+
; LA64-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
100
`+
; LA64-NEXT: st.h a0,a0, a0,fp, 0
`
``
101
`+
; LA64-NEXT: ld.d fp,fp, fp,sp, 0 # 8-byte Folded Reload
`
``
102
`+
; LA64-NEXT: ld.d ra,ra, ra,sp, 8 # 8-byte Folded Reload
`
``
103
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, 16
`
``
104
`+
; LA64-NEXT: ret
`
``
105
`+
;
`
``
106
`+
; LA32-LABEL: test_fptrunc_double:
`
``
107
`+
; LA32: # %bb.0:
`
``
108
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, -16
`
``
109
`+
; LA32-NEXT: st.w ra,ra, ra,sp, 12 # 4-byte Folded Spill
`
``
110
`+
; LA32-NEXT: st.w fp,fp, fp,sp, 8 # 4-byte Folded Spill
`
``
111
`+
; LA32-NEXT: move fp,fp, fp,a0
`
``
112
`+
; LA32-NEXT: bl __truncdfbf2
`
``
113
`+
; LA32-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
114
`+
; LA32-NEXT: st.h a0,a0, a0,fp, 0
`
``
115
`+
; LA32-NEXT: ld.w fp,fp, fp,sp, 8 # 4-byte Folded Reload
`
``
116
`+
; LA32-NEXT: ld.w ra,ra, ra,sp, 12 # 4-byte Folded Reload
`
``
117
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, 16
`
``
118
`+
; LA32-NEXT: ret
`
``
119
`+
%a = fptrunc double %d to bfloat
`
``
120
`+
store bfloat %a, ptr %p
`
``
121
`+
ret void
`
``
122
`+
}
`
``
123
+
``
124
`+
define void @test_fadd(ptr %p, ptr %q) nounwind {
`
``
125
`+
; LA64-LABEL: test_fadd:
`
``
126
`+
; LA64: # %bb.0:
`
``
127
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, -16
`
``
128
`+
; LA64-NEXT: st.d ra,ra, ra,sp, 8 # 8-byte Folded Spill
`
``
129
`+
; LA64-NEXT: st.d fp,fp, fp,sp, 0 # 8-byte Folded Spill
`
``
130
`+
; LA64-NEXT: ld.hu a1,a1, a1,a1, 0
`
``
131
`+
; LA64-NEXT: move fp,fp, fp,a0
`
``
132
`+
; LA64-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
133
`+
; LA64-NEXT: slli.d a1,a1, a1,a1, 16
`
``
134
`+
; LA64-NEXT: movgr2fr.w fa0,fa0, fa0,a1
`
``
135
`+
; LA64-NEXT: slli.d a0,a0, a0,a0, 16
`
``
136
`+
; LA64-NEXT: movgr2fr.w fa1,fa1, fa1,a0
`
``
137
`+
; LA64-NEXT: fadd.s fa0,fa0, fa0,fa1, $fa0
`
``
138
`+
; LA64-NEXT: pcaddu18i $ra, %call36(__truncsfbf2)
`
``
139
`+
; LA64-NEXT: jirl ra,ra, ra,ra, 0
`
``
140
`+
; LA64-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
141
`+
; LA64-NEXT: st.h a0,a0, a0,fp, 0
`
``
142
`+
; LA64-NEXT: ld.d fp,fp, fp,sp, 0 # 8-byte Folded Reload
`
``
143
`+
; LA64-NEXT: ld.d ra,ra, ra,sp, 8 # 8-byte Folded Reload
`
``
144
`+
; LA64-NEXT: addi.d sp,sp, sp,sp, 16
`
``
145
`+
; LA64-NEXT: ret
`
``
146
`+
;
`
``
147
`+
; LA32-LABEL: test_fadd:
`
``
148
`+
; LA32: # %bb.0:
`
``
149
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, -16
`
``
150
`+
; LA32-NEXT: st.w ra,ra, ra,sp, 12 # 4-byte Folded Spill
`
``
151
`+
; LA32-NEXT: st.w fp,fp, fp,sp, 8 # 4-byte Folded Spill
`
``
152
`+
; LA32-NEXT: ld.hu a1,a1, a1,a1, 0
`
``
153
`+
; LA32-NEXT: move fp,fp, fp,a0
`
``
154
`+
; LA32-NEXT: ld.hu a0,a0, a0,a0, 0
`
``
155
`+
; LA32-NEXT: slli.w a1,a1, a1,a1, 16
`
``
156
`+
; LA32-NEXT: movgr2fr.w fa0,fa0, fa0,a1
`
``
157
`+
; LA32-NEXT: slli.w a0,a0, a0,a0, 16
`
``
158
`+
; LA32-NEXT: movgr2fr.w fa1,fa1, fa1,a0
`
``
159
`+
; LA32-NEXT: fadd.s fa0,fa0, fa0,fa1, $fa0
`
``
160
`+
; LA32-NEXT: bl __truncsfbf2
`
``
161
`+
; LA32-NEXT: movfr2gr.s a0,a0, a0,fa0
`
``
162
`+
; LA32-NEXT: st.h a0,a0, a0,fp, 0
`
``
163
`+
; LA32-NEXT: ld.w fp,fp, fp,sp, 8 # 4-byte Folded Reload
`
``
164
`+
; LA32-NEXT: ld.w ra,ra, ra,sp, 12 # 4-byte Folded Reload
`
``
165
`+
; LA32-NEXT: addi.w sp,sp, sp,sp, 16
`
``
166
`+
; LA32-NEXT: ret
`
``
167
`+
%a = load bfloat, ptr %p
`
``
168
`+
%b = load bfloat, ptr %q
`
``
169
`+
%r = fadd bfloat %a, %b
`
``
170
`+
store bfloat %r, ptr %p
`
``
171
`+
ret void
`
``
172
`+
}
`