1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -passes='default<O3>' -enable-matrix -S %s | FileCheck %s
4 target triple = "arm64-apple-ios"
6 define void @matrix_extract_insert_scalar(i32 %i, i32 %k, i32 %j, ptr nonnull align 8 dereferenceable(1800) %A, ptr nonnull align 8 dereferenceable(1800) %B) #0 {
7 ; CHECK-LABEL: @matrix_extract_insert_scalar(
9 ; CHECK-NEXT: [[CONV:%.*]] = zext i32 [[K:%.*]] to i64
10 ; CHECK-NEXT: [[CONV1:%.*]] = zext i32 [[J:%.*]] to i64
11 ; CHECK-NEXT: [[TMP0:%.*]] = mul nuw nsw i64 [[CONV1]], 15
12 ; CHECK-NEXT: [[TMP1:%.*]] = add nuw nsw i64 [[TMP0]], [[CONV]]
13 ; CHECK-NEXT: [[TMP2:%.*]] = icmp ult i64 [[TMP1]], 225
14 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP2]])
15 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds <225 x double>, ptr [[A:%.*]], i64 0, i64 [[TMP1]]
16 ; CHECK-NEXT: [[MATRIXEXT:%.*]] = load double, ptr [[TMP3]], align 8
17 ; CHECK-NEXT: [[CONV2:%.*]] = zext i32 [[I:%.*]] to i64
18 ; CHECK-NEXT: [[TMP4:%.*]] = add nuw nsw i64 [[TMP0]], [[CONV2]]
19 ; CHECK-NEXT: [[TMP5:%.*]] = icmp ult i64 [[TMP4]], 225
20 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP5]])
21 ; CHECK-NEXT: [[TMP6:%.*]] = getelementptr inbounds <225 x double>, ptr [[B:%.*]], i64 0, i64 [[TMP4]]
22 ; CHECK-NEXT: [[MATRIXEXT4:%.*]] = load double, ptr [[TMP6]], align 8
23 ; CHECK-NEXT: [[MUL:%.*]] = fmul double [[MATRIXEXT]], [[MATRIXEXT4]]
24 ; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP1]]
25 ; CHECK-NEXT: [[MATRIXEXT7:%.*]] = load double, ptr [[TMP7]], align 8
26 ; CHECK-NEXT: [[SUB:%.*]] = fsub double [[MATRIXEXT7]], [[MUL]]
27 ; CHECK-NEXT: store double [[SUB]], ptr [[TMP7]], align 8
28 ; CHECK-NEXT: ret void
31 %i.addr = alloca i32, align 4
32 %k.addr = alloca i32, align 4
33 %j.addr = alloca i32, align 4
34 %A.addr = alloca ptr, align 8
35 %B.addr = alloca ptr, align 8
36 store i32 %i, ptr %i.addr, align 4
37 store i32 %k, ptr %k.addr, align 4
38 store i32 %j, ptr %j.addr, align 4
39 store ptr %A, ptr %A.addr, align 8
40 store ptr %B, ptr %B.addr, align 8
41 %0 = load i32, ptr %k.addr, align 4
42 %conv = zext i32 %0 to i64
43 %1 = load i32, ptr %j.addr, align 4
44 %conv1 = zext i32 %1 to i64
45 %2 = mul i64 %conv1, 15
46 %3 = add i64 %2, %conv
47 %4 = icmp ult i64 %3, 225
48 call void @llvm.assume(i1 %4)
49 %5 = load ptr, ptr %A.addr, align 8
50 %6 = load <225 x double>, ptr %5, align 8
51 %matrixext = extractelement <225 x double> %6, i64 %3
52 %7 = load i32, ptr %i.addr, align 4
53 %conv2 = zext i32 %7 to i64
54 %8 = load i32, ptr %j.addr, align 4
55 %conv3 = zext i32 %8 to i64
56 %9 = mul i64 %conv3, 15
57 %10 = add i64 %9, %conv2
58 %11 = icmp ult i64 %10, 225
59 call void @llvm.assume(i1 %11)
60 %12 = load ptr, ptr %B.addr, align 8
61 %13 = load <225 x double>, ptr %12, align 8
62 %matrixext4 = extractelement <225 x double> %13, i64 %10
63 %mul = fmul double %matrixext, %matrixext4
64 %14 = load ptr, ptr %B.addr, align 8
65 %15 = load i32, ptr %k.addr, align 4
66 %conv5 = zext i32 %15 to i64
67 %16 = load i32, ptr %j.addr, align 4
68 %conv6 = zext i32 %16 to i64
69 %17 = mul i64 %conv6, 15
70 %18 = add i64 %17, %conv5
71 %19 = icmp ult i64 %18, 225
72 call void @llvm.assume(i1 %19)
73 %20 = load <225 x double>, ptr %14, align 8
74 %matrixext7 = extractelement <225 x double> %20, i64 %18
75 %sub = fsub double %matrixext7, %mul
76 %21 = icmp ult i64 %18, 225
77 call void @llvm.assume(i1 %21)
78 %22 = load <225 x double>, ptr %14, align 8
79 %matins = insertelement <225 x double> %22, double %sub, i64 %18
80 store <225 x double> %matins, ptr %14, align 8
83 define void @matrix_extract_insert_loop(i32 %i, ptr nonnull align 8 dereferenceable(1800) %A, ptr nonnull align 8 dereferenceable(1800) %B) {
84 ; CHECK-LABEL: @matrix_extract_insert_loop(
86 ; CHECK-NEXT: [[CMP210_NOT:%.*]] = icmp eq i32 [[I:%.*]], 0
87 ; CHECK-NEXT: [[CONV6:%.*]] = zext i32 [[I]] to i64
88 ; CHECK-NEXT: br i1 [[CMP210_NOT]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_COND1_PREHEADER_US:%.*]]
89 ; CHECK: for.cond1.preheader.us:
90 ; CHECK-NEXT: [[TMP0:%.*]] = icmp ult i32 [[I]], 225
91 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP0]])
92 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr inbounds <225 x double>, ptr [[B:%.*]], i64 0, i64 [[CONV6]]
93 ; CHECK-NEXT: br label [[FOR_BODY4_US:%.*]]
94 ; CHECK: for.body4.us:
95 ; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[FOR_COND1_PREHEADER_US]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY4_US]] ]
96 ; CHECK-NEXT: [[TMP2:%.*]] = icmp ult i64 [[INDVARS_IV]], 225
97 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP2]])
98 ; CHECK-NEXT: [[TMP3:%.*]] = getelementptr inbounds <225 x double>, ptr [[A:%.*]], i64 0, i64 [[INDVARS_IV]]
99 ; CHECK-NEXT: [[MATRIXEXT_US:%.*]] = load double, ptr [[TMP3]], align 8
100 ; CHECK-NEXT: [[MATRIXEXT8_US:%.*]] = load double, ptr [[TMP1]], align 8
101 ; CHECK-NEXT: [[MUL_US:%.*]] = fmul double [[MATRIXEXT_US]], [[MATRIXEXT8_US]]
102 ; CHECK-NEXT: [[TMP4:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[INDVARS_IV]]
103 ; CHECK-NEXT: [[MATRIXEXT11_US:%.*]] = load double, ptr [[TMP4]], align 8
104 ; CHECK-NEXT: [[SUB_US:%.*]] = fsub double [[MATRIXEXT11_US]], [[MUL_US]]
105 ; CHECK-NEXT: store double [[SUB_US]], ptr [[TMP4]], align 8
106 ; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
107 ; CHECK-NEXT: [[EXITCOND_NOT:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[CONV6]]
108 ; CHECK-NEXT: br i1 [[EXITCOND_NOT]], label [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US:%.*]], label [[FOR_BODY4_US]]
109 ; CHECK: for.cond1.for.cond.cleanup3_crit_edge.us:
110 ; CHECK-NEXT: [[TMP5:%.*]] = add nuw nsw i64 [[CONV6]], 15
111 ; CHECK-NEXT: [[TMP6:%.*]] = icmp ult i32 [[I]], 210
112 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP6]])
113 ; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP5]]
114 ; CHECK-NEXT: br label [[FOR_BODY4_US_1:%.*]]
115 ; CHECK: for.body4.us.1:
116 ; CHECK-NEXT: [[INDVARS_IV_1:%.*]] = phi i64 [ 0, [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US]] ], [ [[INDVARS_IV_NEXT_1:%.*]], [[FOR_BODY4_US_1]] ]
117 ; CHECK-NEXT: [[TMP8:%.*]] = add nuw nsw i64 [[INDVARS_IV_1]], 15
118 ; CHECK-NEXT: [[TMP9:%.*]] = icmp ult i64 [[INDVARS_IV_1]], 210
119 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP9]])
120 ; CHECK-NEXT: [[TMP10:%.*]] = getelementptr inbounds <225 x double>, ptr [[A]], i64 0, i64 [[TMP8]]
121 ; CHECK-NEXT: [[MATRIXEXT_US_1:%.*]] = load double, ptr [[TMP10]], align 8
122 ; CHECK-NEXT: [[MATRIXEXT8_US_1:%.*]] = load double, ptr [[TMP7]], align 8
123 ; CHECK-NEXT: [[MUL_US_1:%.*]] = fmul double [[MATRIXEXT_US_1]], [[MATRIXEXT8_US_1]]
124 ; CHECK-NEXT: [[TMP11:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP8]]
125 ; CHECK-NEXT: [[MATRIXEXT11_US_1:%.*]] = load double, ptr [[TMP11]], align 8
126 ; CHECK-NEXT: [[SUB_US_1:%.*]] = fsub double [[MATRIXEXT11_US_1]], [[MUL_US_1]]
127 ; CHECK-NEXT: store double [[SUB_US_1]], ptr [[TMP11]], align 8
128 ; CHECK-NEXT: [[INDVARS_IV_NEXT_1]] = add nuw nsw i64 [[INDVARS_IV_1]], 1
129 ; CHECK-NEXT: [[EXITCOND_NOT_1:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_1]], [[CONV6]]
130 ; CHECK-NEXT: br i1 [[EXITCOND_NOT_1]], label [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US_1:%.*]], label [[FOR_BODY4_US_1]]
131 ; CHECK: for.cond1.for.cond.cleanup3_crit_edge.us.1:
132 ; CHECK-NEXT: [[TMP12:%.*]] = add nuw nsw i64 [[CONV6]], 30
133 ; CHECK-NEXT: [[TMP13:%.*]] = icmp ult i32 [[I]], 195
134 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP13]])
135 ; CHECK-NEXT: [[TMP14:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP12]]
136 ; CHECK-NEXT: br label [[FOR_BODY4_US_2:%.*]]
137 ; CHECK: for.body4.us.2:
138 ; CHECK-NEXT: [[INDVARS_IV_2:%.*]] = phi i64 [ 0, [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US_1]] ], [ [[INDVARS_IV_NEXT_2:%.*]], [[FOR_BODY4_US_2]] ]
139 ; CHECK-NEXT: [[TMP15:%.*]] = add nuw nsw i64 [[INDVARS_IV_2]], 30
140 ; CHECK-NEXT: [[TMP16:%.*]] = icmp ult i64 [[INDVARS_IV_2]], 195
141 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP16]])
142 ; CHECK-NEXT: [[TMP17:%.*]] = getelementptr inbounds <225 x double>, ptr [[A]], i64 0, i64 [[TMP15]]
143 ; CHECK-NEXT: [[MATRIXEXT_US_2:%.*]] = load double, ptr [[TMP17]], align 8
144 ; CHECK-NEXT: [[MATRIXEXT8_US_2:%.*]] = load double, ptr [[TMP14]], align 8
145 ; CHECK-NEXT: [[MUL_US_2:%.*]] = fmul double [[MATRIXEXT_US_2]], [[MATRIXEXT8_US_2]]
146 ; CHECK-NEXT: [[TMP18:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP15]]
147 ; CHECK-NEXT: [[MATRIXEXT11_US_2:%.*]] = load double, ptr [[TMP18]], align 8
148 ; CHECK-NEXT: [[SUB_US_2:%.*]] = fsub double [[MATRIXEXT11_US_2]], [[MUL_US_2]]
149 ; CHECK-NEXT: store double [[SUB_US_2]], ptr [[TMP18]], align 8
150 ; CHECK-NEXT: [[INDVARS_IV_NEXT_2]] = add nuw nsw i64 [[INDVARS_IV_2]], 1
151 ; CHECK-NEXT: [[EXITCOND_NOT_2:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_2]], [[CONV6]]
152 ; CHECK-NEXT: br i1 [[EXITCOND_NOT_2]], label [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US_2:%.*]], label [[FOR_BODY4_US_2]]
153 ; CHECK: for.cond1.for.cond.cleanup3_crit_edge.us.2:
154 ; CHECK-NEXT: [[TMP19:%.*]] = add nuw nsw i64 [[CONV6]], 45
155 ; CHECK-NEXT: [[TMP20:%.*]] = icmp ult i32 [[I]], 180
156 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP20]])
157 ; CHECK-NEXT: [[TMP21:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP19]]
158 ; CHECK-NEXT: br label [[FOR_BODY4_US_3:%.*]]
159 ; CHECK: for.body4.us.3:
160 ; CHECK-NEXT: [[INDVARS_IV_3:%.*]] = phi i64 [ 0, [[FOR_COND1_FOR_COND_CLEANUP3_CRIT_EDGE_US_2]] ], [ [[INDVARS_IV_NEXT_3:%.*]], [[FOR_BODY4_US_3]] ]
161 ; CHECK-NEXT: [[TMP22:%.*]] = add nuw nsw i64 [[INDVARS_IV_3]], 45
162 ; CHECK-NEXT: [[TMP23:%.*]] = icmp ult i64 [[INDVARS_IV_3]], 180
163 ; CHECK-NEXT: tail call void @llvm.assume(i1 [[TMP23]])
164 ; CHECK-NEXT: [[TMP24:%.*]] = getelementptr inbounds <225 x double>, ptr [[A]], i64 0, i64 [[TMP22]]
165 ; CHECK-NEXT: [[MATRIXEXT_US_3:%.*]] = load double, ptr [[TMP24]], align 8
166 ; CHECK-NEXT: [[MATRIXEXT8_US_3:%.*]] = load double, ptr [[TMP21]], align 8
167 ; CHECK-NEXT: [[MUL_US_3:%.*]] = fmul double [[MATRIXEXT_US_3]], [[MATRIXEXT8_US_3]]
168 ; CHECK-NEXT: [[TMP25:%.*]] = getelementptr inbounds <225 x double>, ptr [[B]], i64 0, i64 [[TMP22]]
169 ; CHECK-NEXT: [[MATRIXEXT11_US_3:%.*]] = load double, ptr [[TMP25]], align 8
170 ; CHECK-NEXT: [[SUB_US_3:%.*]] = fsub double [[MATRIXEXT11_US_3]], [[MUL_US_3]]
171 ; CHECK-NEXT: store double [[SUB_US_3]], ptr [[TMP25]], align 8
172 ; CHECK-NEXT: [[INDVARS_IV_NEXT_3]] = add nuw nsw i64 [[INDVARS_IV_3]], 1
173 ; CHECK-NEXT: [[EXITCOND_NOT_3:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT_3]], [[CONV6]]
174 ; CHECK-NEXT: br i1 [[EXITCOND_NOT_3]], label [[FOR_COND_CLEANUP]], label [[FOR_BODY4_US_3]]
175 ; CHECK: for.cond.cleanup:
176 ; CHECK-NEXT: ret void
179 %i.addr = alloca i32, align 4
180 %A.addr = alloca ptr, align 8
181 %B.addr = alloca ptr, align 8
182 %j = alloca i32, align 4
183 %cleanup.dest.slot = alloca i32, align 4
184 %k = alloca i32, align 4
185 store i32 %i, ptr %i.addr, align 4
186 store ptr %A, ptr %A.addr, align 8
187 store ptr %B, ptr %B.addr, align 8
188 call void @llvm.lifetime.start.p0(i64 4, ptr %j) #3
189 store i32 0, ptr %j, align 4
192 for.cond: ; preds = %for.inc12, %entry
193 %0 = load i32, ptr %j, align 4
194 %cmp = icmp ult i32 %0, 4
195 br i1 %cmp, label %for.body, label %for.cond.cleanup
197 for.cond.cleanup: ; preds = %for.cond
198 store i32 2, ptr %cleanup.dest.slot, align 4
199 call void @llvm.lifetime.end.p0(i64 4, ptr %j) #3
202 for.body: ; preds = %for.cond
203 call void @llvm.lifetime.start.p0(i64 4, ptr %k) #3
204 store i32 0, ptr %k, align 4
207 for.cond1: ; preds = %for.inc, %for.body
208 %1 = load i32, ptr %k, align 4
209 %2 = load i32, ptr %i.addr, align 4
210 %cmp2 = icmp ult i32 %1, %2
211 br i1 %cmp2, label %for.body4, label %for.cond.cleanup3
213 for.cond.cleanup3: ; preds = %for.cond1
214 store i32 5, ptr %cleanup.dest.slot, align 4
215 call void @llvm.lifetime.end.p0(i64 4, ptr %k) #3
218 for.body4: ; preds = %for.cond1
219 %3 = load i32, ptr %k, align 4
220 %conv = zext i32 %3 to i64
221 %4 = load i32, ptr %j, align 4
222 %conv5 = zext i32 %4 to i64
223 %5 = mul i64 %conv5, 15
224 %6 = add i64 %5, %conv
225 %7 = icmp ult i64 %6, 225
226 call void @llvm.assume(i1 %7)
227 %8 = load ptr, ptr %A.addr, align 8
228 %9 = load <225 x double>, ptr %8, align 8
229 %matrixext = extractelement <225 x double> %9, i64 %6
230 %10 = load i32, ptr %i.addr, align 4
231 %conv6 = zext i32 %10 to i64
232 %11 = load i32, ptr %j, align 4
233 %conv7 = zext i32 %11 to i64
234 %12 = mul i64 %conv7, 15
235 %13 = add i64 %12, %conv6
236 %14 = icmp ult i64 %13, 225
237 call void @llvm.assume(i1 %14)
238 %15 = load ptr, ptr %B.addr, align 8
239 %16 = load <225 x double>, ptr %15, align 8
240 %matrixext8 = extractelement <225 x double> %16, i64 %13
241 %mul = fmul double %matrixext, %matrixext8
242 %17 = load ptr, ptr %B.addr, align 8
243 %18 = load i32, ptr %k, align 4
244 %conv9 = zext i32 %18 to i64
245 %19 = load i32, ptr %j, align 4
246 %conv10 = zext i32 %19 to i64
247 %20 = mul i64 %conv10, 15
248 %21 = add i64 %20, %conv9
249 %22 = icmp ult i64 %21, 225
250 call void @llvm.assume(i1 %22)
251 %23 = load <225 x double>, ptr %17, align 8
252 %matrixext11 = extractelement <225 x double> %23, i64 %21
253 %sub = fsub double %matrixext11, %mul
254 %24 = icmp ult i64 %21, 225
255 call void @llvm.assume(i1 %24)
256 %25 = load <225 x double>, ptr %17, align 8
257 %matins = insertelement <225 x double> %25, double %sub, i64 %21
258 store <225 x double> %matins, ptr %17, align 8
261 for.inc: ; preds = %for.body4
262 %26 = load i32, ptr %k, align 4
263 %inc = add i32 %26, 1
264 store i32 %inc, ptr %k, align 4
267 for.end: ; preds = %for.cond.cleanup3
270 for.inc12: ; preds = %for.end
271 %27 = load i32, ptr %j, align 4
272 %inc13 = add i32 %27, 1
273 store i32 %inc13, ptr %j, align 4
276 for.end14: ; preds = %for.cond.cleanup
280 ; Function Attrs: argmemonly nofree nosync nounwind willreturn
281 declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #1
283 ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn
284 declare void @llvm.assume(i1 noundef) #2
286 ; Function Attrs: argmemonly nofree nosync nounwind willreturn
287 declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #1
289 ; Function Attrs: nounwind ssp uwtable mustprogress
291 define <4 x float> @reverse_hadd_v4f32(<4 x float> %a, <4 x float> %b) {
292 ; CHECK-LABEL: @reverse_hadd_v4f32(
293 ; CHECK-NEXT: [[TMP1:%.*]] = shufflevector <4 x float> [[B:%.*]], <4 x float> [[A:%.*]], <4 x i32> <i32 2, i32 0, i32 6, i32 4>
294 ; CHECK-NEXT: [[TMP2:%.*]] = shufflevector <4 x float> [[B]], <4 x float> [[A]], <4 x i32> <i32 3, i32 1, i32 7, i32 5>
295 ; CHECK-NEXT: [[TMP3:%.*]] = fadd <4 x float> [[TMP1]], [[TMP2]]
296 ; CHECK-NEXT: ret <4 x float> [[TMP3]]
298 %vecext = extractelement <4 x float> %a, i32 0
299 %vecext1 = extractelement <4 x float> %a, i32 1
300 %add = fadd float %vecext, %vecext1
301 %vecinit = insertelement <4 x float> undef, float %add, i32 0
302 %vecext2 = extractelement <4 x float> %a, i32 2
303 %vecext3 = extractelement <4 x float> %a, i32 3
304 %add4 = fadd float %vecext2, %vecext3
305 %vecinit5 = insertelement <4 x float> %vecinit, float %add4, i32 1
306 %vecext6 = extractelement <4 x float> %b, i32 0
307 %vecext7 = extractelement <4 x float> %b, i32 1
308 %add8 = fadd float %vecext6, %vecext7
309 %vecinit9 = insertelement <4 x float> %vecinit5, float %add8, i32 2
310 %vecext10 = extractelement <4 x float> %b, i32 2
311 %vecext11 = extractelement <4 x float> %b, i32 3
312 %add12 = fadd float %vecext10, %vecext11
313 %vecinit13 = insertelement <4 x float> %vecinit9, float %add12, i32 3
314 %shuffle = shufflevector <4 x float> %vecinit13, <4 x float> %a, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
315 ret <4 x float> %shuffle