[llvm-project.git] / llvm / test / Transforms / LowerMatrixIntrinsics / multiply-fused-lifetime-ends.ll
| blob | bdd0c6f728ae7874b257e8b1ae87ca9ddd2123b2 |
1 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
2 ; RUN: opt -passes=lower-matrix-intrinsics -fuse-matrix-tile-size=2 -matrix-allow-contract -force-fuse-matrix %s -S | FileCheck %s
4 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
6 ; Tests to make sure no loads are introduced after a lifetime.end by multiply
7 ; fusion.
9 define void @lifetime_for_first_arg_before_multiply(ptr noalias %B, ptr noalias %C) {
10 ; CHECK-LABEL: @lifetime_for_first_arg_before_multiply(
11 ; CHECK-NEXT: entry:
12 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
13 ; CHECK-NEXT: call void @init(ptr [[A]])
14 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
15 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
16 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
17 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
18 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
19 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
20 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
21 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
22 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
23 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
24 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
25 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
26 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
27 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
28 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
29 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
30 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
31 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
32 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
33 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
34 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
35 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
36 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
37 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
38 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
39 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
40 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
41 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
42 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
43 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
44 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
45 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
46 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
47 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
48 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
49 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
50 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
51 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
52 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
53 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
54 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
55 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
56 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
57 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
58 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
59 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
60 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
61 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
62 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
63 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
64 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
65 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
66 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
67 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
68 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
69 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
70 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
71 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
72 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C:%.*]], i64 0
73 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
74 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
75 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
76 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[A]])
77 ; CHECK-NEXT: ret void
78 ;
79 entry:
80 %A = alloca <4 x double>
81 call void @init(ptr %A)
82 %a = load <4 x double>, ptr %A, align 8
83 %b = load <4 x double>, ptr %B, align 8
84 call void @llvm.lifetime.end(i64 -1, ptr %A)
85 %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
86 store <4 x double> %c, ptr %C, align 8
87 ret void
88 }
90 define void @lifetime_for_second_arg_before_multiply(ptr noalias %A, ptr noalias %C) {
91 ; CHECK-LABEL: @lifetime_for_second_arg_before_multiply(
92 ; CHECK-NEXT: entry:
93 ; CHECK-NEXT: [[B:%.*]] = alloca <4 x double>, align 32
94 ; CHECK-NEXT: call void @init(ptr [[B]])
95 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A:%.*]], i64 0
96 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
97 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
98 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
99 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B]], i64 0
100 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
101 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
102 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
103 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
104 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
105 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
106 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
107 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
108 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
109 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
110 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
111 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
112 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
113 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
114 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
115 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
116 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
117 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
118 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
119 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
120 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
121 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
122 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
123 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
124 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
125 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
126 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
127 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
128 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
129 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
130 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
131 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
132 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
133 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
134 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
135 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
136 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
137 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
138 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
139 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
140 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
141 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
142 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
143 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
144 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
145 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
146 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
147 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
148 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
149 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
150 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
151 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
152 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
153 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C:%.*]], i64 0
154 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
155 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
156 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
157 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[B]])
158 ; CHECK-NEXT: ret void
159 ;
160 entry:
161 %B = alloca <4 x double>
162 call void @init(ptr %B)
163 %a = load <4 x double>, ptr %A, align 8
164 %b = load <4 x double>, ptr %B, align 8
165 call void @llvm.lifetime.end(i64 -1, ptr %B)
166 %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
167 store <4 x double> %c, ptr %C, align 8
168 ret void
169 }
171 define void @lifetime_for_first_arg_before_multiply_load_from_offset(ptr noalias %B, ptr noalias %C) {
172 ; CHECK-LABEL: @lifetime_for_first_arg_before_multiply_load_from_offset(
173 ; CHECK-NEXT: entry:
174 ; CHECK-NEXT: [[A:%.*]] = alloca <8 x double>, align 64
175 ; CHECK-NEXT: call void @init(ptr [[A]])
176 ; CHECK-NEXT: [[GEP_8:%.*]] = getelementptr i8, ptr [[A]], i64 8
177 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[GEP_8]], i64 0
178 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
179 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
180 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
181 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
182 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
183 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
184 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
185 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
186 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
187 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
188 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
189 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
190 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
191 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
192 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
193 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
194 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
195 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
196 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
197 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
198 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
199 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
200 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
201 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
202 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
203 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
204 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
205 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
206 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
207 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
208 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
209 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
210 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
211 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
212 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
213 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
214 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
215 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
216 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
217 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
218 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
219 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
220 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
221 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
222 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
223 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
224 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
225 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
226 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
227 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
228 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
229 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
230 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
231 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
232 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
233 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
234 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
235 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C:%.*]], i64 0
236 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
237 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
238 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
239 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[A]])
240 ; CHECK-NEXT: ret void
241 ;
242 entry:
243 %A = alloca <8 x double>
244 call void @init(ptr %A)
245 %gep.8 = getelementptr i8, ptr %A, i64 8
246 %a = load <4 x double>, ptr %gep.8, align 8
247 %b = load <4 x double>, ptr %B, align 8
248 call void @llvm.lifetime.end(i64 -1, ptr %A)
249 %c = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
250 store <4 x double> %c, ptr %C, align 8
251 ret void
252 }
254 define void @lifetime_for_first_arg_before_multiply_lifetime_does_not_dominate(ptr noalias %B, ptr noalias %C, i1 %c.0) {
255 ; CHECK-LABEL: @lifetime_for_first_arg_before_multiply_lifetime_does_not_dominate(
256 ; CHECK-NEXT: entry:
257 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
258 ; CHECK-NEXT: call void @init(ptr [[A]])
259 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
260 ; CHECK: then:
261 ; CHECK-NEXT: br label [[EXIT]]
262 ; CHECK: exit:
263 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
264 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
265 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
266 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
267 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
268 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
269 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
270 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
271 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
272 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
273 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
274 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
275 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
276 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
277 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
278 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
279 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
280 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
281 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
282 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
283 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
284 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
285 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
286 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
287 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
288 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
289 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
290 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
291 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
292 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
293 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
294 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
295 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
296 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
297 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
298 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
299 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
300 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
301 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
302 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
303 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
304 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
305 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
306 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
307 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
308 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
309 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
310 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
311 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
312 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
313 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
314 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
315 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
316 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
317 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
318 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
319 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
320 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
321 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
322 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
323 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
324 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
325 ; CHECK-NEXT: ret void
326 ;
327 entry:
328 %A = alloca <4 x double>
329 call void @init(ptr %A)
330 %a = load <4 x double>, ptr %A, align 8
331 %b = load <4 x double>, ptr %B, align 8
332 br i1 %c.0, label %then, label %exit
334 then:
335 call void @llvm.lifetime.end(i64 -1, ptr %A)
336 br label %exit
338 exit:
339 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
340 store <4 x double> %m, ptr %C, align 8
341 ret void
342 }
344 define void @lifetime_for_second_arg_before_multiply_lifetime_does_not_dominate(ptr noalias %A, ptr noalias %C, i1 %c.0) {
345 ; CHECK-LABEL: @lifetime_for_second_arg_before_multiply_lifetime_does_not_dominate(
346 ; CHECK-NEXT: entry:
347 ; CHECK-NEXT: [[B:%.*]] = alloca <4 x double>, align 32
348 ; CHECK-NEXT: call void @init(ptr [[B]])
349 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
350 ; CHECK: then:
351 ; CHECK-NEXT: br label [[EXIT]]
352 ; CHECK: exit:
353 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A:%.*]], i64 0
354 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
355 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
356 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
357 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B]], i64 0
358 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
359 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
360 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
361 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
362 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
363 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
364 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
365 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
366 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
367 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
368 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
369 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
370 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
371 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
372 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
373 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
374 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
375 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
376 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
377 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
378 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
379 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
380 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
381 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
382 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
383 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
384 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
385 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
386 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
387 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
388 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
389 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
390 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
391 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
392 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
393 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
394 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
395 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
396 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
397 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
398 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
399 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
400 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
401 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
402 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
403 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
404 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
405 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
406 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
407 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
408 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
409 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
410 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
411 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
412 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
413 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
414 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
415 ; CHECK-NEXT: ret void
416 ;
417 entry:
418 %B = alloca <4 x double>
419 call void @init(ptr %B)
420 %a = load <4 x double>, ptr %A, align 8
421 %b = load <4 x double>, ptr %B, align 8
422 br i1 %c.0, label %then, label %exit
424 then:
425 call void @llvm.lifetime.end(i64 -1, ptr %B)
426 br label %exit
428 exit:
429 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
430 store <4 x double> %m, ptr %C, align 8
431 ret void
432 }
434 define void @lifetime_for_ptr_first_arg_before_multiply(ptr noalias %A, ptr noalias %B, ptr noalias %C, i1 %c.0) {
435 ; CHECK-LABEL: @lifetime_for_ptr_first_arg_before_multiply(
436 ; CHECK-NEXT: entry:
437 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
438 ; CHECK: then:
439 ; CHECK-NEXT: br label [[EXIT]]
440 ; CHECK: exit:
441 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A:%.*]], i64 0
442 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
443 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
444 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
445 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
446 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
447 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
448 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
449 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
450 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
451 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
452 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
453 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
454 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
455 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
456 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
457 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
458 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
459 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
460 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
461 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
462 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
463 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
464 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
465 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
466 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
467 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
468 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
469 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
470 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
471 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
472 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
473 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
474 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
475 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
476 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
477 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
478 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
479 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
480 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
481 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
482 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
483 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
484 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
485 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
486 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
487 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
488 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
489 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
490 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
491 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
492 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
493 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
494 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
495 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
496 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
497 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
498 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
499 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
500 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
501 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
502 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
503 ; CHECK-NEXT: ret void
504 ;
505 entry:
506 %a = load <4 x double>, ptr %A, align 8
507 %b = load <4 x double>, ptr %B, align 8
508 br i1 %c.0, label %then, label %exit
510 then:
511 call void @llvm.lifetime.end(i64 -1, ptr %A)
512 br label %exit
514 exit:
515 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
516 store <4 x double> %m, ptr %C, align 8
517 ret void
518 }
520 define void @lifetime_for_both_ptr_args_before_multiply(ptr noalias %A, ptr noalias %B, ptr noalias %C, i1 %c.0) {
521 ; CHECK-LABEL: @lifetime_for_both_ptr_args_before_multiply(
522 ; CHECK-NEXT: entry:
523 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
524 ; CHECK: then:
525 ; CHECK-NEXT: br label [[EXIT]]
526 ; CHECK: exit:
527 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A:%.*]], i64 0
528 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
529 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
530 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
531 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
532 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
533 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
534 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
535 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
536 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
537 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
538 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
539 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
540 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
541 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
542 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
543 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
544 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
545 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
546 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
547 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
548 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
549 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
550 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
551 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
552 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
553 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
554 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
555 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
556 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
557 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
558 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
559 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
560 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
561 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
562 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
563 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
564 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
565 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
566 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
567 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
568 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
569 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
570 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
571 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
572 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
573 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
574 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
575 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
576 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
577 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
578 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
579 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
580 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
581 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
582 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
583 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
584 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
585 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
586 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
587 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
588 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
589 ; CHECK-NEXT: ret void
590 ;
591 entry:
592 %a = load <4 x double>, ptr %A, align 8
593 %b = load <4 x double>, ptr %B, align 8
594 br i1 %c.0, label %then, label %exit
596 then:
597 call void @llvm.lifetime.end(i64 -1, ptr %B)
598 call void @llvm.lifetime.end(i64 -1, ptr %A)
599 br label %exit
601 exit:
602 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
603 store <4 x double> %m, ptr %C, align 8
604 ret void
605 }
607 define void @multiple_unrelated_lifetimes(ptr noalias %A, ptr noalias %B, ptr noalias %C, i1 %c.0) {
608 ; CHECK-LABEL: @multiple_unrelated_lifetimes(
609 ; CHECK-NEXT: entry:
610 ; CHECK-NEXT: [[ALLOC_1:%.*]] = alloca i32, align 4
611 ; CHECK-NEXT: [[ALLOC_2:%.*]] = alloca i32, align 4
612 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
613 ; CHECK: then:
614 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[ALLOC_1]])
615 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[ALLOC_2]])
616 ; CHECK-NEXT: br label [[EXIT]]
617 ; CHECK: exit:
618 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A:%.*]], i64 0
619 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
620 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
621 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
622 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
623 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
624 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
625 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
626 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
627 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
628 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
629 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
630 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
631 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
632 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
633 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
634 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
635 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
636 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
637 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
638 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
639 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
640 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
641 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
642 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
643 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
644 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
645 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
646 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
647 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
648 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
649 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
650 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
651 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
652 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
653 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
654 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
655 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
656 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
657 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
658 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
659 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
660 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
661 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
662 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
663 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
664 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
665 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
666 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
667 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
668 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
669 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
670 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
671 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
672 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
673 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
674 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
675 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
676 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
677 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
678 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
679 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
680 ; CHECK-NEXT: ret void
681 ;
682 entry:
683 %alloc.1 = alloca i32
684 %alloc.2 = alloca i32
685 %a = load <4 x double>, ptr %A, align 8
686 %b = load <4 x double>, ptr %B, align 8
687 br i1 %c.0, label %then, label %exit
689 then:
690 call void @llvm.lifetime.end(i64 -1, ptr %B)
691 call void @llvm.lifetime.end(i64 -1, ptr %alloc.1)
692 call void @llvm.lifetime.end(i64 -1, ptr %A)
693 call void @llvm.lifetime.end(i64 -1, ptr %alloc.2)
694 br label %exit
696 exit:
697 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
698 store <4 x double> %m, ptr %C, align 8
699 ret void
700 }
702 define void @lifetime_for_ptr_select_before_multiply(ptr noalias %A, ptr noalias %B, ptr noalias %C, i1 %c.0, i1 %c.1) {
703 ; CHECK-LABEL: @lifetime_for_ptr_select_before_multiply(
704 ; CHECK-NEXT: entry:
705 ; CHECK-NEXT: [[P:%.*]] = select i1 [[C_0:%.*]], ptr [[A:%.*]], ptr [[B:%.*]]
706 ; CHECK-NEXT: br i1 [[C_1:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
707 ; CHECK: then:
708 ; CHECK-NEXT: br label [[EXIT]]
709 ; CHECK: exit:
710 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[P]], i64 0
711 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
712 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
713 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
714 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B]], i64 0
715 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
716 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
717 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
718 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
719 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
720 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
721 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
722 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
723 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
724 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
725 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
726 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
727 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
728 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
729 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
730 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
731 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
732 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
733 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
734 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
735 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
736 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
737 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
738 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
739 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
740 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
741 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
742 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
743 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
744 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
745 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
746 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
747 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
748 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
749 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
750 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
751 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
752 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
753 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
754 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
755 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
756 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
757 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
758 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
759 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
760 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
761 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
762 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
763 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
764 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
765 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
766 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
767 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
768 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C:%.*]], i64 0
769 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
770 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
771 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
772 ; CHECK-NEXT: ret void
773 ;
774 entry:
775 %P = select i1 %c.0, ptr %A, ptr %B
776 %a = load <4 x double>, ptr %P, align 8
777 %b = load <4 x double>, ptr %B, align 8
778 br i1 %c.1, label %then, label %exit
780 then:
781 call void @llvm.lifetime.end(i64 -1, ptr %P)
782 br label %exit
784 exit:
785 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
786 store <4 x double> %m, ptr %C, align 8
787 ret void
788 }
790 define void @lifetimes_for_args_in_different_blocks(ptr noalias %B, ptr noalias %C, i1 %c.0) {
791 ; CHECK-LABEL: @lifetimes_for_args_in_different_blocks(
792 ; CHECK-NEXT: entry:
793 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
794 ; CHECK-NEXT: call void @init(ptr [[A]])
795 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
796 ; CHECK: then:
797 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
798 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
799 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
800 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
801 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
802 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
803 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
804 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
805 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
806 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
807 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
808 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
809 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
810 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
811 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
812 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
813 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
814 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
815 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
816 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
817 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
818 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
819 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
820 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
821 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
822 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
823 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
824 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
825 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
826 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
827 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
828 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
829 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
830 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
831 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
832 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
833 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
834 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
835 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
836 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
837 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
838 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
839 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
840 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
841 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
842 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
843 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
844 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
845 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
846 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
847 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
848 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
849 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
850 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
851 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
852 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
853 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
854 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
855 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
856 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
857 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
858 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
859 ; CHECK-NEXT: br label [[EXIT]]
860 ; CHECK: exit:
861 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[A]])
862 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[B]])
863 ; CHECK-NEXT: ret void
864 ;
865 entry:
866 %A = alloca <4 x double>
867 call void @init(ptr %A)
868 br i1 %c.0, label %then, label %exit
870 then:
871 %a = load <4 x double>, ptr %A, align 8
872 %b = load <4 x double>, ptr %B, align 8
873 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
874 store <4 x double> %m, ptr %C, align 8
875 br label %exit
877 exit:
878 call void @llvm.lifetime.end(i64 -1, ptr %A)
879 call void @llvm.lifetime.end(i64 -1, ptr %B)
880 ret void
881 }
883 define void @lifetimes_for_args_in_different_blocks2(ptr noalias %B, ptr noalias %C, i1 %c.0) {
884 ; CHECK-LABEL: @lifetimes_for_args_in_different_blocks2(
885 ; CHECK-NEXT: entry:
886 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
887 ; CHECK-NEXT: call void @init(ptr [[A]])
888 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
889 ; CHECK: then:
890 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[A]])
891 ; CHECK-NEXT: call void @llvm.lifetime.end.p0(i64 -1, ptr [[B:%.*]])
892 ; CHECK-NEXT: br label [[EXIT]]
893 ; CHECK: exit:
894 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
895 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
896 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
897 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
898 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B]], i64 0
899 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
900 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
901 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
902 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
903 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
904 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
905 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
906 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
907 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
908 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
909 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
910 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
911 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
912 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
913 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
914 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
915 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
916 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
917 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
918 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
919 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
920 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
921 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
922 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
923 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
924 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
925 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
926 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
927 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
928 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
929 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
930 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
931 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
932 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
933 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
934 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
935 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
936 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
937 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
938 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
939 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
940 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
941 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
942 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
943 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
944 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
945 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
946 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
947 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
948 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
949 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
950 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
951 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
952 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
953 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
954 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
955 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
956 ; CHECK-NEXT: ret void
957 ;
958 entry:
959 %A = alloca <4 x double>
960 call void @init(ptr %A)
961 br i1 %c.0, label %then, label %exit
963 then:
964 call void @llvm.lifetime.end(i64 -1, ptr %A)
965 call void @llvm.lifetime.end(i64 -1, ptr %B)
966 br label %exit
968 exit:
969 %a = load <4 x double>, ptr %A, align 8
970 %b = load <4 x double>, ptr %B, align 8
971 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
972 store <4 x double> %m, ptr %C, align 8
973 ret void
974 }
976 define void @lifetimes_for_args_load0_in_different_block(ptr noalias %B, ptr noalias %C, i1 %c.0) {
977 ; CHECK-LABEL: @lifetimes_for_args_load0_in_different_block(
978 ; CHECK-NEXT: entry:
979 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
980 ; CHECK-NEXT: call void @init(ptr [[A]])
981 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
982 ; CHECK: then:
983 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
984 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
985 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
986 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
987 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
988 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
989 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
990 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
991 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
992 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
993 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
994 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
995 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
996 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
997 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
998 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
999 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
1000 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
1001 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1002 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
1003 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
1004 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
1005 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
1006 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
1007 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
1008 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
1009 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
1010 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
1011 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
1012 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
1013 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
1014 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1015 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
1016 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
1017 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
1018 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
1019 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
1020 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
1021 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
1022 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
1023 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
1024 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
1025 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
1026 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1027 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
1028 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
1029 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
1030 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
1031 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
1032 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
1033 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
1034 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
1035 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
1036 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
1037 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
1038 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
1039 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1040 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
1041 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
1042 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
1043 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
1044 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
1045 ; CHECK-NEXT: br label [[EXIT]]
1046 ; CHECK: exit:
1047 ; CHECK-NEXT: ret void
1048 ;
1049 entry:
1050 %A = alloca <4 x double>
1051 call void @init(ptr %A)
1052 %a = load <4 x double>, ptr %A, align 8
1053 call void @llvm.lifetime.end(i64 -1, ptr %A)
1054 br i1 %c.0, label %then, label %exit
1056 then:
1057 %b = load <4 x double>, ptr %B, align 8
1058 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
1059 store <4 x double> %m, ptr %C, align 8
1060 br label %exit
1062 exit:
1063 call void @llvm.lifetime.end(i64 -1, ptr %B)
1064 ret void
1065 }
1067 define void @lifetimes_for_args_load1_in_different_block(ptr noalias %B, ptr noalias %C, i1 %c.0) {
1068 ; CHECK-LABEL: @lifetimes_for_args_load1_in_different_block(
1069 ; CHECK-NEXT: entry:
1070 ; CHECK-NEXT: [[A:%.*]] = alloca <4 x double>, align 32
1071 ; CHECK-NEXT: call void @init(ptr [[A]])
1072 ; CHECK-NEXT: br i1 [[C:%.*]], label [[THEN:%.*]], label [[EXIT:%.*]]
1073 ; CHECK: then:
1074 ; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[A]], i64 0
1075 ; CHECK-NEXT: [[COL_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
1076 ; CHECK-NEXT: [[VEC_GEP:%.*]] = getelementptr double, ptr [[TMP0]], i64 2
1077 ; CHECK-NEXT: [[COL_LOAD1:%.*]] = load <2 x double>, ptr [[VEC_GEP]], align 8
1078 ; CHECK-NEXT: [[TMP1:%.*]] = getelementptr double, ptr [[B:%.*]], i64 0
1079 ; CHECK-NEXT: [[COL_LOAD2:%.*]] = load <2 x double>, ptr [[TMP1]], align 8
1080 ; CHECK-NEXT: [[VEC_GEP3:%.*]] = getelementptr double, ptr [[TMP1]], i64 2
1081 ; CHECK-NEXT: [[COL_LOAD4:%.*]] = load <2 x double>, ptr [[VEC_GEP3]], align 8
1082 ; CHECK-NEXT: [[BLOCK:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
1083 ; CHECK-NEXT: [[TMP2:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
1084 ; CHECK-NEXT: [[SPLAT_SPLATINSERT:%.*]] = insertelement <1 x double> poison, double [[TMP2]], i64 0
1085 ; CHECK-NEXT: [[SPLAT_SPLAT:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT]], <1 x double> poison, <1 x i32> zeroinitializer
1086 ; CHECK-NEXT: [[TMP3:%.*]] = fmul contract <1 x double> [[BLOCK]], [[SPLAT_SPLAT]]
1087 ; CHECK-NEXT: [[BLOCK5:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
1088 ; CHECK-NEXT: [[TMP4:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
1089 ; CHECK-NEXT: [[SPLAT_SPLATINSERT6:%.*]] = insertelement <1 x double> poison, double [[TMP4]], i64 0
1090 ; CHECK-NEXT: [[SPLAT_SPLAT7:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT6]], <1 x double> poison, <1 x i32> zeroinitializer
1091 ; CHECK-NEXT: [[TMP5:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK5]], <1 x double> [[SPLAT_SPLAT7]], <1 x double> [[TMP3]])
1092 ; CHECK-NEXT: [[TMP6:%.*]] = shufflevector <1 x double> [[TMP5]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1093 ; CHECK-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP6]], <2 x i32> <i32 2, i32 1>
1094 ; CHECK-NEXT: [[BLOCK8:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> poison, <1 x i32> <i32 1>
1095 ; CHECK-NEXT: [[BLOCK9:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
1096 ; CHECK-NEXT: [[TMP8:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 0
1097 ; CHECK-NEXT: [[SPLAT_SPLATINSERT10:%.*]] = insertelement <1 x double> poison, double [[TMP8]], i64 0
1098 ; CHECK-NEXT: [[SPLAT_SPLAT11:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT10]], <1 x double> poison, <1 x i32> zeroinitializer
1099 ; CHECK-NEXT: [[TMP9:%.*]] = fmul contract <1 x double> [[BLOCK9]], [[SPLAT_SPLAT11]]
1100 ; CHECK-NEXT: [[BLOCK12:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
1101 ; CHECK-NEXT: [[TMP10:%.*]] = extractelement <2 x double> [[COL_LOAD2]], i64 1
1102 ; CHECK-NEXT: [[SPLAT_SPLATINSERT13:%.*]] = insertelement <1 x double> poison, double [[TMP10]], i64 0
1103 ; CHECK-NEXT: [[SPLAT_SPLAT14:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT13]], <1 x double> poison, <1 x i32> zeroinitializer
1104 ; CHECK-NEXT: [[TMP11:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK12]], <1 x double> [[SPLAT_SPLAT14]], <1 x double> [[TMP9]])
1105 ; CHECK-NEXT: [[TMP12:%.*]] = shufflevector <1 x double> [[TMP11]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1106 ; CHECK-NEXT: [[TMP13:%.*]] = shufflevector <2 x double> [[TMP7]], <2 x double> [[TMP12]], <2 x i32> <i32 0, i32 2>
1107 ; CHECK-NEXT: [[BLOCK15:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> zeroinitializer
1108 ; CHECK-NEXT: [[TMP14:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
1109 ; CHECK-NEXT: [[SPLAT_SPLATINSERT16:%.*]] = insertelement <1 x double> poison, double [[TMP14]], i64 0
1110 ; CHECK-NEXT: [[SPLAT_SPLAT17:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT16]], <1 x double> poison, <1 x i32> zeroinitializer
1111 ; CHECK-NEXT: [[TMP15:%.*]] = fmul contract <1 x double> [[BLOCK15]], [[SPLAT_SPLAT17]]
1112 ; CHECK-NEXT: [[BLOCK18:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> zeroinitializer
1113 ; CHECK-NEXT: [[TMP16:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
1114 ; CHECK-NEXT: [[SPLAT_SPLATINSERT19:%.*]] = insertelement <1 x double> poison, double [[TMP16]], i64 0
1115 ; CHECK-NEXT: [[SPLAT_SPLAT20:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT19]], <1 x double> poison, <1 x i32> zeroinitializer
1116 ; CHECK-NEXT: [[TMP17:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK18]], <1 x double> [[SPLAT_SPLAT20]], <1 x double> [[TMP15]])
1117 ; CHECK-NEXT: [[TMP18:%.*]] = shufflevector <1 x double> [[TMP17]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1118 ; CHECK-NEXT: [[TMP19:%.*]] = shufflevector <2 x double> zeroinitializer, <2 x double> [[TMP18]], <2 x i32> <i32 2, i32 1>
1119 ; CHECK-NEXT: [[BLOCK21:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> poison, <1 x i32> <i32 1>
1120 ; CHECK-NEXT: [[BLOCK22:%.*]] = shufflevector <2 x double> [[COL_LOAD]], <2 x double> poison, <1 x i32> <i32 1>
1121 ; CHECK-NEXT: [[TMP20:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 0
1122 ; CHECK-NEXT: [[SPLAT_SPLATINSERT23:%.*]] = insertelement <1 x double> poison, double [[TMP20]], i64 0
1123 ; CHECK-NEXT: [[SPLAT_SPLAT24:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT23]], <1 x double> poison, <1 x i32> zeroinitializer
1124 ; CHECK-NEXT: [[TMP21:%.*]] = fmul contract <1 x double> [[BLOCK22]], [[SPLAT_SPLAT24]]
1125 ; CHECK-NEXT: [[BLOCK25:%.*]] = shufflevector <2 x double> [[COL_LOAD1]], <2 x double> poison, <1 x i32> <i32 1>
1126 ; CHECK-NEXT: [[TMP22:%.*]] = extractelement <2 x double> [[COL_LOAD4]], i64 1
1127 ; CHECK-NEXT: [[SPLAT_SPLATINSERT26:%.*]] = insertelement <1 x double> poison, double [[TMP22]], i64 0
1128 ; CHECK-NEXT: [[SPLAT_SPLAT27:%.*]] = shufflevector <1 x double> [[SPLAT_SPLATINSERT26]], <1 x double> poison, <1 x i32> zeroinitializer
1129 ; CHECK-NEXT: [[TMP23:%.*]] = call contract <1 x double> @llvm.fmuladd.v1f64(<1 x double> [[BLOCK25]], <1 x double> [[SPLAT_SPLAT27]], <1 x double> [[TMP21]])
1130 ; CHECK-NEXT: [[TMP24:%.*]] = shufflevector <1 x double> [[TMP23]], <1 x double> poison, <2 x i32> <i32 0, i32 poison>
1131 ; CHECK-NEXT: [[TMP25:%.*]] = shufflevector <2 x double> [[TMP19]], <2 x double> [[TMP24]], <2 x i32> <i32 0, i32 2>
1132 ; CHECK-NEXT: [[TMP26:%.*]] = getelementptr double, ptr [[C1:%.*]], i64 0
1133 ; CHECK-NEXT: store <2 x double> [[TMP13]], ptr [[TMP26]], align 8
1134 ; CHECK-NEXT: [[VEC_GEP28:%.*]] = getelementptr double, ptr [[TMP26]], i64 2
1135 ; CHECK-NEXT: store <2 x double> [[TMP25]], ptr [[VEC_GEP28]], align 8
1136 ; CHECK-NEXT: br label [[EXIT]]
1137 ; CHECK: exit:
1138 ; CHECK-NEXT: ret void
1139 ;
1140 entry:
1141 %A = alloca <4 x double>
1142 call void @init(ptr %A)
1143 %b = load <4 x double>, ptr %B, align 8
1144 call void @llvm.lifetime.end(i64 -1, ptr %B)
1145 br i1 %c.0, label %then, label %exit
1147 then:
1148 %a = load <4 x double>, ptr %A, align 8
1149 %m = call <4 x double> @llvm.matrix.multiply(<4 x double> %a, <4 x double> %b, i32 2, i32 2, i32 2)
1150 store <4 x double> %m, ptr %C, align 8
1151 br label %exit
1153 exit:
1154 call void @llvm.lifetime.end(i64 -1, ptr %A)
1155 ret void
1156 }
1158 declare void @init(ptr)
1159 declare void @llvm.lifetime.end(i64, ptr)
1161 declare <4 x double> @llvm.matrix.multiply(<4 x double>, <4 x double>, i32, i32, i32)