| blob | aba9056c78cdaa0920fcd14f63a972f6ae14859c |
1 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
2 ; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=1 | FileCheck %s --check-prefixes=NO_FOLDING
3 ; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=2 | FileCheck %s --check-prefixes=NO_FOLDING
4 ; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING,ZVFH
5 ; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfhmin,+f,+d -verify-machineinstrs %s -o - --riscv-lower-ext-max-web-size=3 | FileCheck %s --check-prefixes=FOLDING,ZVFHMIN
6 ; Check that the default value enables the web folding and
7 ; that it is bigger than 3.
8 ; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh,+f,+d -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=FOLDING
10 define void @vfwmul_v2f116_multiple_users(ptr %x, ptr %y, ptr %z, <2 x half> %a, <2 x half> %b, <2 x half> %b2) {
11 ; NO_FOLDING-LABEL: vfwmul_v2f116_multiple_users:
12 ; NO_FOLDING: # %bb.0:
13 ; NO_FOLDING-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
14 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v11, v8
15 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
16 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
17 ; NO_FOLDING-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
18 ; NO_FOLDING-NEXT: vfmul.vv v10, v11, v8
19 ; NO_FOLDING-NEXT: vfadd.vv v11, v11, v9
20 ; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
21 ; NO_FOLDING-NEXT: vse32.v v10, (a0)
22 ; NO_FOLDING-NEXT: vse32.v v11, (a1)
23 ; NO_FOLDING-NEXT: vse32.v v8, (a2)
24 ; NO_FOLDING-NEXT: ret
25 ;
26 ; ZVFH-LABEL: vfwmul_v2f116_multiple_users:
27 ; ZVFH: # %bb.0:
28 ; ZVFH-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
29 ; ZVFH-NEXT: vfwmul.vv v11, v8, v9
30 ; ZVFH-NEXT: vfwadd.vv v12, v8, v10
31 ; ZVFH-NEXT: vfwsub.vv v8, v9, v10
32 ; ZVFH-NEXT: vse32.v v11, (a0)
33 ; ZVFH-NEXT: vse32.v v12, (a1)
34 ; ZVFH-NEXT: vse32.v v8, (a2)
35 ; ZVFH-NEXT: ret
36 ;
37 ; ZVFHMIN-LABEL: vfwmul_v2f116_multiple_users:
38 ; ZVFHMIN: # %bb.0:
39 ; ZVFHMIN-NEXT: vsetivli zero, 2, e16, mf4, ta, ma
40 ; ZVFHMIN-NEXT: vfwcvt.f.f.v v11, v8
41 ; ZVFHMIN-NEXT: vfwcvt.f.f.v v8, v9
42 ; ZVFHMIN-NEXT: vfwcvt.f.f.v v9, v10
43 ; ZVFHMIN-NEXT: vsetvli zero, zero, e32, mf2, ta, ma
44 ; ZVFHMIN-NEXT: vfmul.vv v10, v11, v8
45 ; ZVFHMIN-NEXT: vfadd.vv v11, v11, v9
46 ; ZVFHMIN-NEXT: vfsub.vv v8, v8, v9
47 ; ZVFHMIN-NEXT: vse32.v v10, (a0)
48 ; ZVFHMIN-NEXT: vse32.v v11, (a1)
49 ; ZVFHMIN-NEXT: vse32.v v8, (a2)
50 ; ZVFHMIN-NEXT: ret
51 %c = fpext <2 x half> %a to <2 x float>
52 %d = fpext <2 x half> %b to <2 x float>
53 %d2 = fpext <2 x half> %b2 to <2 x float>
54 %e = fmul <2 x float> %c, %d
55 %f = fadd <2 x float> %c, %d2
56 %g = fsub <2 x float> %d, %d2
57 store <2 x float> %e, ptr %x
58 store <2 x float> %f, ptr %y
59 store <2 x float> %g, ptr %z
60 ret void
61 }
63 define void @vfwmul_v2f32_multiple_users(ptr %x, ptr %y, ptr %z, <2 x float> %a, <2 x float> %b, <2 x float> %b2) {
64 ; NO_FOLDING-LABEL: vfwmul_v2f32_multiple_users:
65 ; NO_FOLDING: # %bb.0:
66 ; NO_FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
67 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v11, v8
68 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
69 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
70 ; NO_FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
71 ; NO_FOLDING-NEXT: vfmul.vv v10, v11, v8
72 ; NO_FOLDING-NEXT: vfadd.vv v11, v11, v9
73 ; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
74 ; NO_FOLDING-NEXT: vse64.v v10, (a0)
75 ; NO_FOLDING-NEXT: vse64.v v11, (a1)
76 ; NO_FOLDING-NEXT: vse64.v v8, (a2)
77 ; NO_FOLDING-NEXT: ret
78 ;
79 ; FOLDING-LABEL: vfwmul_v2f32_multiple_users:
80 ; FOLDING: # %bb.0:
81 ; FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
82 ; FOLDING-NEXT: vfwmul.vv v11, v8, v9
83 ; FOLDING-NEXT: vfwadd.vv v12, v8, v10
84 ; FOLDING-NEXT: vfwsub.vv v8, v9, v10
85 ; FOLDING-NEXT: vse64.v v11, (a0)
86 ; FOLDING-NEXT: vse64.v v12, (a1)
87 ; FOLDING-NEXT: vse64.v v8, (a2)
88 ; FOLDING-NEXT: ret
89 %c = fpext <2 x float> %a to <2 x double>
90 %d = fpext <2 x float> %b to <2 x double>
91 %d2 = fpext <2 x float> %b2 to <2 x double>
92 %e = fmul <2 x double> %c, %d
93 %f = fadd <2 x double> %c, %d2
94 %g = fsub <2 x double> %d, %d2
95 store <2 x double> %e, ptr %x
96 store <2 x double> %f, ptr %y
97 store <2 x double> %g, ptr %z
98 ret void
99 }
101 define void @vfwmacc_v2f32_multiple_users(ptr %x, ptr %y, ptr %z, <2 x float> %a, <2 x float> %b, <2 x float> %b2, <2 x double> %w) {
102 ; NO_FOLDING-LABEL: vfwmacc_v2f32_multiple_users:
103 ; NO_FOLDING: # %bb.0:
104 ; NO_FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
105 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v12, v8
106 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
107 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
108 ; NO_FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
109 ; NO_FOLDING-NEXT: vfmul.vv v10, v12, v8
110 ; NO_FOLDING-NEXT: vfmadd.vv v12, v9, v11
111 ; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
112 ; NO_FOLDING-NEXT: vse64.v v10, (a0)
113 ; NO_FOLDING-NEXT: vse64.v v12, (a1)
114 ; NO_FOLDING-NEXT: vse64.v v8, (a2)
115 ; NO_FOLDING-NEXT: ret
116 ;
117 ; FOLDING-LABEL: vfwmacc_v2f32_multiple_users:
118 ; FOLDING: # %bb.0:
119 ; FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
120 ; FOLDING-NEXT: vfwmul.vv v12, v8, v9
121 ; FOLDING-NEXT: vfwsub.vv v13, v9, v10
122 ; FOLDING-NEXT: vfwmacc.vv v11, v8, v10
123 ; FOLDING-NEXT: vse64.v v12, (a0)
124 ; FOLDING-NEXT: vse64.v v11, (a1)
125 ; FOLDING-NEXT: vse64.v v13, (a2)
126 ; FOLDING-NEXT: ret
127 %c = fpext <2 x float> %a to <2 x double>
128 %d = fpext <2 x float> %b to <2 x double>
129 %d2 = fpext <2 x float> %b2 to <2 x double>
130 %e = fmul <2 x double> %c, %d
131 %f = call <2 x double> @llvm.fma(<2 x double> %c, <2 x double> %d2, <2 x double> %w)
132 %g = fsub <2 x double> %d, %d2
133 store <2 x double> %e, ptr %x
134 store <2 x double> %f, ptr %y
135 store <2 x double> %g, ptr %z
136 ret void
137 }
139 ; Negative test. We can't fold because the FMA addend is a user.
140 define void @vfwmacc_v2f32_multiple_users_addend_user(ptr %x, ptr %y, ptr %z, <2 x float> %a, <2 x float> %b, <2 x float> %b2) {
141 ; NO_FOLDING-LABEL: vfwmacc_v2f32_multiple_users_addend_user:
142 ; NO_FOLDING: # %bb.0:
143 ; NO_FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
144 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v11, v8
145 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
146 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v9, v10
147 ; NO_FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
148 ; NO_FOLDING-NEXT: vfmul.vv v10, v11, v8
149 ; NO_FOLDING-NEXT: vfmadd.vv v11, v9, v8
150 ; NO_FOLDING-NEXT: vfsub.vv v8, v8, v9
151 ; NO_FOLDING-NEXT: vse64.v v10, (a0)
152 ; NO_FOLDING-NEXT: vse64.v v11, (a1)
153 ; NO_FOLDING-NEXT: vse64.v v8, (a2)
154 ; NO_FOLDING-NEXT: ret
155 ;
156 ; FOLDING-LABEL: vfwmacc_v2f32_multiple_users_addend_user:
157 ; FOLDING: # %bb.0:
158 ; FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
159 ; FOLDING-NEXT: vfwcvt.f.f.v v11, v8
160 ; FOLDING-NEXT: vfwcvt.f.f.v v8, v9
161 ; FOLDING-NEXT: vfwcvt.f.f.v v9, v10
162 ; FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
163 ; FOLDING-NEXT: vfmul.vv v10, v11, v8
164 ; FOLDING-NEXT: vfmadd.vv v11, v9, v8
165 ; FOLDING-NEXT: vfsub.vv v8, v8, v9
166 ; FOLDING-NEXT: vse64.v v10, (a0)
167 ; FOLDING-NEXT: vse64.v v11, (a1)
168 ; FOLDING-NEXT: vse64.v v8, (a2)
169 ; FOLDING-NEXT: ret
170 %c = fpext <2 x float> %a to <2 x double>
171 %d = fpext <2 x float> %b to <2 x double>
172 %d2 = fpext <2 x float> %b2 to <2 x double>
173 %e = fmul <2 x double> %c, %d
174 %f = call <2 x double> @llvm.fma(<2 x double> %c, <2 x double> %d2, <2 x double> %d)
175 %g = fsub <2 x double> %d, %d2
176 store <2 x double> %e, ptr %x
177 store <2 x double> %f, ptr %y
178 store <2 x double> %g, ptr %z
179 ret void
180 }
182 ; Negative test. We can't fold because the FMA addend is a user.
183 define void @vfwmacc_v2f32_addend_user(ptr %x, <2 x float> %a, <2 x float> %b) {
184 ; NO_FOLDING-LABEL: vfwmacc_v2f32_addend_user:
185 ; NO_FOLDING: # %bb.0:
186 ; NO_FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
187 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v10, v8
188 ; NO_FOLDING-NEXT: vfwcvt.f.f.v v8, v9
189 ; NO_FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
190 ; NO_FOLDING-NEXT: vfmadd.vv v8, v10, v8
191 ; NO_FOLDING-NEXT: vse64.v v8, (a0)
192 ; NO_FOLDING-NEXT: ret
193 ;
194 ; FOLDING-LABEL: vfwmacc_v2f32_addend_user:
195 ; FOLDING: # %bb.0:
196 ; FOLDING-NEXT: vsetivli zero, 2, e32, mf2, ta, ma
197 ; FOLDING-NEXT: vfwcvt.f.f.v v10, v8
198 ; FOLDING-NEXT: vfwcvt.f.f.v v8, v9
199 ; FOLDING-NEXT: vsetvli zero, zero, e64, m1, ta, ma
200 ; FOLDING-NEXT: vfmadd.vv v8, v10, v8
201 ; FOLDING-NEXT: vse64.v v8, (a0)
202 ; FOLDING-NEXT: ret
203 %c = fpext <2 x float> %a to <2 x double>
204 %d = fpext <2 x float> %b to <2 x double>
205 %f = call <2 x double> @llvm.fma(<2 x double> %c, <2 x double> %d, <2 x double> %d)
206 store <2 x double> %f, ptr %x
207 ret void
208 }