1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -passes=slp-vectorizer -mtriple=arm64-apple-ios -S %s | FileCheck %s
4 ; Test case where not vectorizing is more profitable because multiple
5 ; fmul/{fadd,fsub} pairs can be lowered to fma instructions.
6 define void @slp_not_profitable_with_fast_fmf(ptr %A, ptr %B) {
7 ; CHECK-LABEL: @slp_not_profitable_with_fast_fmf(
8 ; CHECK-NEXT: [[GEP_B_1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1
9 ; CHECK-NEXT: [[A_0:%.*]] = load float, ptr [[A:%.*]], align 4
10 ; CHECK-NEXT: [[B_0:%.*]] = load float, ptr [[B]], align 4
11 ; CHECK-NEXT: [[GEP_B_2:%.*]] = getelementptr inbounds float, ptr [[B]], i64 2
12 ; CHECK-NEXT: [[B_2:%.*]] = load float, ptr [[GEP_B_2]], align 4
13 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[GEP_B_1]], align 4
14 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[B_0]], i32 0
15 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
16 ; CHECK-NEXT: [[TMP4:%.*]] = fmul fast <2 x float> [[TMP3]], [[TMP1]]
17 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
18 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
19 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
20 ; CHECK-NEXT: [[TMP8:%.*]] = fmul fast <2 x float> [[TMP1]], [[TMP7]]
21 ; CHECK-NEXT: [[TMP9:%.*]] = fsub fast <2 x float> [[TMP8]], [[TMP5]]
22 ; CHECK-NEXT: [[TMP10:%.*]] = fadd fast <2 x float> [[TMP8]], [[TMP5]]
23 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
24 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[A]], align 4
25 ; CHECK-NEXT: store float [[B_2]], ptr [[B]], align 4
26 ; CHECK-NEXT: ret void
28 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
29 %A.0 = load float, ptr %A, align 4
30 %B.1 = load float, ptr %gep.B.1, align 4
31 %mul.0 = fmul fast float %B.1, %A.0
32 %B.0 = load float, ptr %B, align 4
33 %gep.B.2 = getelementptr inbounds float, ptr %B, i64 2
34 %B.2 = load float, ptr %gep.B.2, align 4
35 %mul.1 = fmul fast float %B.2, %B.0
36 %sub = fsub fast float %mul.0, %mul.1
37 %mul.2 = fmul fast float %B.0, %B.1
38 %mul.3 = fmul fast float %B.2, %A.0
39 %add = fadd fast float %mul.3, %mul.2
40 store float %sub, ptr %A, align 4
41 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
42 store float %add, ptr %gep.A.1, align 4
43 store float %B.2, ptr %B, align 4
47 define void @slp_not_profitable_with_reassoc_fmf(ptr %A, ptr %B) {
48 ; CHECK-LABEL: @slp_not_profitable_with_reassoc_fmf(
49 ; CHECK-NEXT: [[GEP_B_1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1
50 ; CHECK-NEXT: [[A_0:%.*]] = load float, ptr [[A:%.*]], align 4
51 ; CHECK-NEXT: [[B_0:%.*]] = load float, ptr [[B]], align 4
52 ; CHECK-NEXT: [[GEP_B_2:%.*]] = getelementptr inbounds float, ptr [[B]], i64 2
53 ; CHECK-NEXT: [[B_2:%.*]] = load float, ptr [[GEP_B_2]], align 4
54 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[GEP_B_1]], align 4
55 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[B_0]], i32 0
56 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
57 ; CHECK-NEXT: [[TMP4:%.*]] = fmul <2 x float> [[TMP3]], [[TMP1]]
58 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
59 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
60 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
61 ; CHECK-NEXT: [[TMP8:%.*]] = fmul reassoc <2 x float> [[TMP1]], [[TMP7]]
62 ; CHECK-NEXT: [[TMP9:%.*]] = fsub reassoc <2 x float> [[TMP8]], [[TMP5]]
63 ; CHECK-NEXT: [[TMP10:%.*]] = fadd reassoc <2 x float> [[TMP8]], [[TMP5]]
64 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
65 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[A]], align 4
66 ; CHECK-NEXT: store float [[B_2]], ptr [[B]], align 4
67 ; CHECK-NEXT: ret void
69 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
70 %A.0 = load float, ptr %A, align 4
71 %B.1 = load float, ptr %gep.B.1, align 4
72 %mul.0 = fmul reassoc float %B.1, %A.0
73 %B.0 = load float, ptr %B, align 4
74 %gep.B.2 = getelementptr inbounds float, ptr %B, i64 2
75 %B.2 = load float, ptr %gep.B.2, align 4
76 %mul.1 = fmul float %B.2, %B.0
77 %sub = fsub reassoc float %mul.0, %mul.1
78 %mul.2 = fmul float %B.0, %B.1
79 %mul.3 = fmul reassoc float %B.2, %A.0
80 %add = fadd reassoc float %mul.3, %mul.2
81 store float %sub, ptr %A, align 4
82 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
83 store float %add, ptr %gep.A.1, align 4
84 store float %B.2, ptr %B, align 4
88 ; FMA cannot be used due to missing fast-math flags, so SLP should kick in.
89 define void @slp_profitable_missing_fmf_on_fadd_fsub(ptr %A, ptr %B) {
90 ; CHECK-LABEL: @slp_profitable_missing_fmf_on_fadd_fsub(
91 ; CHECK-NEXT: [[GEP_B_1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1
92 ; CHECK-NEXT: [[A_0:%.*]] = load float, ptr [[A:%.*]], align 4
93 ; CHECK-NEXT: [[B_0:%.*]] = load float, ptr [[B]], align 4
94 ; CHECK-NEXT: [[GEP_B_2:%.*]] = getelementptr inbounds float, ptr [[B]], i64 2
95 ; CHECK-NEXT: [[B_2:%.*]] = load float, ptr [[GEP_B_2]], align 4
96 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[GEP_B_1]], align 4
97 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[B_0]], i32 0
98 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
99 ; CHECK-NEXT: [[TMP4:%.*]] = fmul fast <2 x float> [[TMP3]], [[TMP1]]
100 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
101 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
102 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
103 ; CHECK-NEXT: [[TMP8:%.*]] = fmul fast <2 x float> [[TMP1]], [[TMP7]]
104 ; CHECK-NEXT: [[TMP9:%.*]] = fsub <2 x float> [[TMP8]], [[TMP5]]
105 ; CHECK-NEXT: [[TMP10:%.*]] = fadd <2 x float> [[TMP8]], [[TMP5]]
106 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
107 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[A]], align 4
108 ; CHECK-NEXT: store float [[B_2]], ptr [[B]], align 4
109 ; CHECK-NEXT: ret void
111 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
112 %A.0 = load float, ptr %A, align 4
113 %B.1 = load float, ptr %gep.B.1, align 4
114 %mul.0 = fmul fast float %B.1, %A.0
115 %B.0 = load float, ptr %B, align 4
116 %gep.B.2 = getelementptr inbounds float, ptr %B, i64 2
117 %B.2 = load float, ptr %gep.B.2, align 4
118 %mul.1 = fmul fast float %B.2, %B.0
119 %sub = fsub float %mul.0, %mul.1
120 %mul.2 = fmul fast float %B.0, %B.1
121 %mul.3 = fmul fast float %B.2, %A.0
122 %add = fadd float %mul.3, %mul.2
123 store float %sub, ptr %A, align 4
124 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
125 store float %add, ptr %gep.A.1, align 4
126 store float %B.2, ptr %B, align 4
130 ; FMA cannot be used due to missing fast-math flags, so SLP should kick in.
131 define void @slp_profitable_missing_fmf_on_fmul_fadd_fsub(ptr %A, ptr %B) {
132 ; CHECK-LABEL: @slp_profitable_missing_fmf_on_fmul_fadd_fsub(
133 ; CHECK-NEXT: [[GEP_B_1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1
134 ; CHECK-NEXT: [[A_0:%.*]] = load float, ptr [[A:%.*]], align 4
135 ; CHECK-NEXT: [[B_0:%.*]] = load float, ptr [[B]], align 4
136 ; CHECK-NEXT: [[GEP_B_2:%.*]] = getelementptr inbounds float, ptr [[B]], i64 2
137 ; CHECK-NEXT: [[B_2:%.*]] = load float, ptr [[GEP_B_2]], align 4
138 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[GEP_B_1]], align 4
139 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[B_0]], i32 0
140 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
141 ; CHECK-NEXT: [[TMP4:%.*]] = fmul <2 x float> [[TMP3]], [[TMP1]]
142 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
143 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
144 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
145 ; CHECK-NEXT: [[TMP8:%.*]] = fmul <2 x float> [[TMP1]], [[TMP7]]
146 ; CHECK-NEXT: [[TMP9:%.*]] = fsub <2 x float> [[TMP8]], [[TMP5]]
147 ; CHECK-NEXT: [[TMP10:%.*]] = fadd <2 x float> [[TMP8]], [[TMP5]]
148 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
149 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[A]], align 4
150 ; CHECK-NEXT: store float [[B_2]], ptr [[B]], align 4
151 ; CHECK-NEXT: ret void
153 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
154 %A.0 = load float, ptr %A, align 4
155 %B.1 = load float, ptr %gep.B.1, align 4
156 %mul.0 = fmul float %B.1, %A.0
157 %B.0 = load float, ptr %B, align 4
158 %gep.B.2 = getelementptr inbounds float, ptr %B, i64 2
159 %B.2 = load float, ptr %gep.B.2, align 4
160 %mul.1 = fmul float %B.2, %B.0
161 %sub = fsub float %mul.0, %mul.1
162 %mul.2 = fmul float %B.0, %B.1
163 %mul.3 = fmul float %B.2, %A.0
164 %add = fadd float %mul.3, %mul.2
165 store float %sub, ptr %A, align 4
166 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
167 store float %add, ptr %gep.A.1, align 4
168 store float %B.2, ptr %B, align 4
172 ; FMA cannot be used due to missing fast-math flags, so SLP should kick in.
173 define void @slp_profitable_missing_fmf_nnans_only(ptr %A, ptr %B) {
174 ; CHECK-LABEL: @slp_profitable_missing_fmf_nnans_only(
175 ; CHECK-NEXT: [[GEP_B_1:%.*]] = getelementptr inbounds float, ptr [[B:%.*]], i64 1
176 ; CHECK-NEXT: [[A_0:%.*]] = load float, ptr [[A:%.*]], align 4
177 ; CHECK-NEXT: [[B_0:%.*]] = load float, ptr [[B]], align 4
178 ; CHECK-NEXT: [[GEP_B_2:%.*]] = getelementptr inbounds float, ptr [[B]], i64 2
179 ; CHECK-NEXT: [[B_2:%.*]] = load float, ptr [[GEP_B_2]], align 4
180 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[GEP_B_1]], align 4
181 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[B_0]], i32 0
182 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
183 ; CHECK-NEXT: [[TMP4:%.*]] = fmul nnan <2 x float> [[TMP3]], [[TMP1]]
184 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
185 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[A_0]], i32 0
186 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
187 ; CHECK-NEXT: [[TMP8:%.*]] = fmul nnan <2 x float> [[TMP1]], [[TMP7]]
188 ; CHECK-NEXT: [[TMP9:%.*]] = fsub nnan <2 x float> [[TMP8]], [[TMP5]]
189 ; CHECK-NEXT: [[TMP10:%.*]] = fadd nnan <2 x float> [[TMP8]], [[TMP5]]
190 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
191 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[A]], align 4
192 ; CHECK-NEXT: store float [[B_2]], ptr [[B]], align 4
193 ; CHECK-NEXT: ret void
195 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
196 %A.0 = load float, ptr %A, align 4
197 %B.1 = load float, ptr %gep.B.1, align 4
198 %mul.0 = fmul nnan float %B.1, %A.0
199 %B.0 = load float, ptr %B, align 4
200 %gep.B.2 = getelementptr inbounds float, ptr %B, i64 2
201 %B.2 = load float, ptr %gep.B.2, align 4
202 %mul.1 = fmul nnan float %B.2, %B.0
203 %sub = fsub nnan float %mul.0, %mul.1
204 %mul.2 = fmul nnan float %B.0, %B.1
205 %mul.3 = fmul nnan float %B.2, %A.0
206 %add = fadd nnan float %mul.3, %mul.2
207 store float %sub, ptr %A, align 4
208 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
209 store float %add, ptr %gep.A.1, align 4
210 store float %B.2, ptr %B, align 4
214 ; Test case where not vectorizing is more profitable because multiple
215 ; fmul/{fadd,fsub} pairs can be lowered to fma instructions.
216 define float @slp_not_profitable_in_loop(float %x, ptr %A) {
217 ; CHECK-LABEL: @slp_not_profitable_in_loop(
219 ; CHECK-NEXT: [[GEP_A_2:%.*]] = getelementptr inbounds float, ptr [[A:%.*]], i64 2
220 ; CHECK-NEXT: [[L_1:%.*]] = load float, ptr [[GEP_A_2]], align 4
221 ; CHECK-NEXT: [[TMP0:%.*]] = load <2 x float>, ptr [[A]], align 4
222 ; CHECK-NEXT: [[L_3:%.*]] = load float, ptr [[A]], align 4
223 ; CHECK-NEXT: [[TMP1:%.*]] = insertelement <2 x float> <float poison, float 3.000000e+00>, float [[X:%.*]], i32 0
224 ; CHECK-NEXT: br label [[LOOP:%.*]]
226 ; CHECK-NEXT: [[IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[IV_NEXT:%.*]], [[LOOP]] ]
227 ; CHECK-NEXT: [[RED:%.*]] = phi float [ 0.000000e+00, [[ENTRY]] ], [ [[RED_NEXT:%.*]], [[LOOP]] ]
228 ; CHECK-NEXT: [[TMP2:%.*]] = fmul fast <2 x float> [[TMP1]], [[TMP0]]
229 ; CHECK-NEXT: [[MUL12:%.*]] = fmul fast float 3.000000e+00, [[L_1]]
230 ; CHECK-NEXT: [[MUL16:%.*]] = fmul fast float 3.000000e+00, [[L_3]]
231 ; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x float> [[TMP2]], i32 1
232 ; CHECK-NEXT: [[ADD:%.*]] = fadd fast float [[MUL12]], [[TMP3]]
233 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x float> [[TMP2]], i32 0
234 ; CHECK-NEXT: [[ADD13:%.*]] = fadd fast float [[ADD]], [[TMP4]]
235 ; CHECK-NEXT: [[RED_NEXT]] = fadd fast float [[ADD13]], [[MUL16]]
236 ; CHECK-NEXT: [[IV_NEXT]] = add nuw nsw i64 [[IV]], 1
237 ; CHECK-NEXT: [[CMP:%.*]] = icmp eq i64 [[IV]], 10
238 ; CHECK-NEXT: br i1 [[CMP]], label [[EXIT:%.*]], label [[LOOP]]
240 ; CHECK-NEXT: ret float [[RED_NEXT]]
243 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
244 %l.0 = load float, ptr %gep.A.1, align 4
245 %gep.A.2 = getelementptr inbounds float, ptr %A, i64 2
246 %l.1 = load float, ptr %gep.A.2, align 4
247 %l.2 = load float, ptr %A, align 4
248 %l.3 = load float, ptr %A, align 4
252 %iv = phi i64 [ 0, %entry ], [ %iv.next, %loop ]
253 %red = phi float [ 0.000000e+00, %entry ], [ %red.next, %loop ]
254 %mul11 = fmul fast float 3.000000e+00, %l.0
255 %mul12 = fmul fast float 3.000000e+00, %l.1
256 %mul14 = fmul fast float %x, %l.2
257 %mul16 = fmul fast float 3.000000e+00, %l.3
258 %add = fadd fast float %mul12, %mul11
259 %add13 = fadd fast float %add, %mul14
260 %red.next = fadd fast float %add13, %mul16
261 %iv.next = add nuw nsw i64 %iv, 1
262 %cmp = icmp eq i64 %iv, 10
263 br i1 %cmp, label %exit, label %loop
269 define void @slp_profitable(ptr %A, ptr %B, float %0) {
270 ; CHECK-LABEL: @slp_profitable(
272 ; CHECK-NEXT: [[SUB_I1096:%.*]] = fsub fast float 1.000000e+00, [[TMP0:%.*]]
273 ; CHECK-NEXT: [[TMP1:%.*]] = load <2 x float>, ptr [[A:%.*]], align 4
274 ; CHECK-NEXT: [[TMP2:%.*]] = insertelement <2 x float> poison, float [[TMP0]], i32 0
275 ; CHECK-NEXT: [[TMP3:%.*]] = shufflevector <2 x float> [[TMP2]], <2 x float> poison, <2 x i32> zeroinitializer
276 ; CHECK-NEXT: [[TMP4:%.*]] = fmul fast <2 x float> [[TMP1]], [[TMP3]]
277 ; CHECK-NEXT: [[TMP5:%.*]] = shufflevector <2 x float> [[TMP4]], <2 x float> poison, <2 x i32> <i32 1, i32 0>
278 ; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x float> poison, float [[SUB_I1096]], i32 0
279 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x float> [[TMP6]], <2 x float> poison, <2 x i32> zeroinitializer
280 ; CHECK-NEXT: [[TMP8:%.*]] = fmul fast <2 x float> [[TMP1]], [[TMP7]]
281 ; CHECK-NEXT: [[TMP9:%.*]] = fadd fast <2 x float> [[TMP5]], [[TMP8]]
282 ; CHECK-NEXT: [[TMP10:%.*]] = fsub fast <2 x float> [[TMP5]], [[TMP8]]
283 ; CHECK-NEXT: [[TMP11:%.*]] = shufflevector <2 x float> [[TMP9]], <2 x float> [[TMP10]], <2 x i32> <i32 0, i32 3>
284 ; CHECK-NEXT: store <2 x float> [[TMP11]], ptr [[B:%.*]], align 4
285 ; CHECK-NEXT: ret void
288 %gep.A.1 = getelementptr inbounds float, ptr %A, i64 1
289 %sub.i1096 = fsub fast float 1.000000e+00, %0
290 %1 = load float, ptr %A, align 4
291 %mul.i1100 = fmul fast float %1, %sub.i1096
292 %2 = load float, ptr %gep.A.1, align 4
293 %mul7.i1101 = fmul fast float %2, %0
294 %add.i1102 = fadd fast float %mul7.i1101, %mul.i1100
295 %mul14.i = fmul fast float %1, %0
296 %3 = fmul fast float %2, %sub.i1096
297 %add15.i = fsub fast float %mul14.i, %3
298 store float %add.i1102, ptr %B, align 4
299 %gep.B.1 = getelementptr inbounds float, ptr %B, i64 1
300 store float %add15.i, ptr %gep.B.1, align 4