clang/test/CodeGen/AArch64/ls64.c

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
   2 // RUN: %clang_cc1 -triple aarch64 -target-feature +ls64 -emit-llvm -x c %s -o - | FileCheck --check-prefixes=CHECK-C %s
   3 // RUN: %clang_cc1 -triple aarch64 -target-feature +ls64 -emit-llvm -x c++ %s -o - | FileCheck --check-prefixes=CHECK-CXX %s
   4 // RUN: %clang_cc1 -triple aarch64_be-eabi -target-feature +ls64 -emit-llvm -x c %s -o - | FileCheck  --check-prefixes=CHECK-C %s
   5 // RUN: %clang_cc1 -triple aarch64_be-eabi -target-feature +ls64 -emit-llvm -x c++ %s -o - | FileCheck  --check-prefixes=CHECK-CXX %s
   6
   7 #include <arm_acle.h>
   8
   9 #ifdef __cplusplus
  10 #define EXTERN_C extern "C"
  11 #else
  12 #define EXTERN_C
  13 #endif
  14
  15 data512_t val;
  16 void *addr;
  17 uint64_t status;
  18
  19 // CHECK-C-LABEL: define {{[^@]+}}@test_ld64b(
  20 // CHECK-C-NEXT:  entry:
  21 // CHECK-C-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
  22 // CHECK-C-NEXT:    [[TMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
  23 // CHECK-C-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
  24 // CHECK-C-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META2:![0-9]+]])
  25 // CHECK-C-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8, !noalias !2
  26 // CHECK-C-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8, !noalias !2
  27 // CHECK-C-NEXT:    [[TMP2:%.*]] = call { i64, i64, i64, i64, i64, i64, i64, i64 } @llvm.aarch64.ld64b(ptr [[TMP1]]), !noalias !2
  28 // CHECK-C-NEXT:    [[TMP3:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 0
  29 // CHECK-C-NEXT:    store i64 [[TMP3]], ptr [[TMP]], align 8, !alias.scope !2
  30 // CHECK-C-NEXT:    [[TMP4:%.*]] = getelementptr i64, ptr [[TMP]], i32 1
  31 // CHECK-C-NEXT:    [[TMP5:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 1
  32 // CHECK-C-NEXT:    store i64 [[TMP5]], ptr [[TMP4]], align 8, !alias.scope !2
  33 // CHECK-C-NEXT:    [[TMP6:%.*]] = getelementptr i64, ptr [[TMP]], i32 2
  34 // CHECK-C-NEXT:    [[TMP7:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 2
  35 // CHECK-C-NEXT:    store i64 [[TMP7]], ptr [[TMP6]], align 8, !alias.scope !2
  36 // CHECK-C-NEXT:    [[TMP8:%.*]] = getelementptr i64, ptr [[TMP]], i32 3
  37 // CHECK-C-NEXT:    [[TMP9:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 3
  38 // CHECK-C-NEXT:    store i64 [[TMP9]], ptr [[TMP8]], align 8, !alias.scope !2
  39 // CHECK-C-NEXT:    [[TMP10:%.*]] = getelementptr i64, ptr [[TMP]], i32 4
  40 // CHECK-C-NEXT:    [[TMP11:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 4
  41 // CHECK-C-NEXT:    store i64 [[TMP11]], ptr [[TMP10]], align 8, !alias.scope !2
  42 // CHECK-C-NEXT:    [[TMP12:%.*]] = getelementptr i64, ptr [[TMP]], i32 5
  43 // CHECK-C-NEXT:    [[TMP13:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 5
  44 // CHECK-C-NEXT:    store i64 [[TMP13]], ptr [[TMP12]], align 8, !alias.scope !2
  45 // CHECK-C-NEXT:    [[TMP14:%.*]] = getelementptr i64, ptr [[TMP]], i32 6
  46 // CHECK-C-NEXT:    [[TMP15:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 6
  47 // CHECK-C-NEXT:    store i64 [[TMP15]], ptr [[TMP14]], align 8, !alias.scope !2
  48 // CHECK-C-NEXT:    [[TMP16:%.*]] = getelementptr i64, ptr [[TMP]], i32 7
  49 // CHECK-C-NEXT:    [[TMP17:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 7
  50 // CHECK-C-NEXT:    store i64 [[TMP17]], ptr [[TMP16]], align 8, !alias.scope !2
  51 // CHECK-C-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 @val, ptr align 8 [[TMP]], i64 64, i1 false)
  52 // CHECK-C-NEXT:    ret void
  53 //
  54 // CHECK-CXX-LABEL: define {{[^@]+}}@test_ld64b(
  55 // CHECK-CXX-NEXT:  entry:
  56 // CHECK-CXX-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
  57 // CHECK-CXX-NEXT:    [[REF_TMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
  58 // CHECK-CXX-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
  59 // CHECK-CXX-NEXT:    call void @llvm.experimental.noalias.scope.decl(metadata [[META2:![0-9]+]])
  60 // CHECK-CXX-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8, !noalias !2
  61 // CHECK-CXX-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8, !noalias !2
  62 // CHECK-CXX-NEXT:    [[TMP2:%.*]] = call { i64, i64, i64, i64, i64, i64, i64, i64 } @llvm.aarch64.ld64b(ptr [[TMP1]]), !noalias !2
  63 // CHECK-CXX-NEXT:    [[TMP3:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 0
  64 // CHECK-CXX-NEXT:    store i64 [[TMP3]], ptr [[REF_TMP]], align 8, !alias.scope !2
  65 // CHECK-CXX-NEXT:    [[TMP4:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 1
  66 // CHECK-CXX-NEXT:    [[TMP5:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 1
  67 // CHECK-CXX-NEXT:    store i64 [[TMP5]], ptr [[TMP4]], align 8, !alias.scope !2
  68 // CHECK-CXX-NEXT:    [[TMP6:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 2
  69 // CHECK-CXX-NEXT:    [[TMP7:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 2
  70 // CHECK-CXX-NEXT:    store i64 [[TMP7]], ptr [[TMP6]], align 8, !alias.scope !2
  71 // CHECK-CXX-NEXT:    [[TMP8:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 3
  72 // CHECK-CXX-NEXT:    [[TMP9:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 3
  73 // CHECK-CXX-NEXT:    store i64 [[TMP9]], ptr [[TMP8]], align 8, !alias.scope !2
  74 // CHECK-CXX-NEXT:    [[TMP10:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 4
  75 // CHECK-CXX-NEXT:    [[TMP11:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 4
  76 // CHECK-CXX-NEXT:    store i64 [[TMP11]], ptr [[TMP10]], align 8, !alias.scope !2
  77 // CHECK-CXX-NEXT:    [[TMP12:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 5
  78 // CHECK-CXX-NEXT:    [[TMP13:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 5
  79 // CHECK-CXX-NEXT:    store i64 [[TMP13]], ptr [[TMP12]], align 8, !alias.scope !2
  80 // CHECK-CXX-NEXT:    [[TMP14:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 6
  81 // CHECK-CXX-NEXT:    [[TMP15:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 6
  82 // CHECK-CXX-NEXT:    store i64 [[TMP15]], ptr [[TMP14]], align 8, !alias.scope !2
  83 // CHECK-CXX-NEXT:    [[TMP16:%.*]] = getelementptr i64, ptr [[REF_TMP]], i32 7
  84 // CHECK-CXX-NEXT:    [[TMP17:%.*]] = extractvalue { i64, i64, i64, i64, i64, i64, i64, i64 } [[TMP2]], 7
  85 // CHECK-CXX-NEXT:    store i64 [[TMP17]], ptr [[TMP16]], align 8, !alias.scope !2
  86 // CHECK-CXX-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 @val, ptr align 8 [[REF_TMP]], i64 64, i1 false)
  87 // CHECK-CXX-NEXT:    ret void
  88 //
  89 EXTERN_C void test_ld64b(void)
  90 {
  91     val = __arm_ld64b(addr);
  92 }
  93
  94 // CHECK-C-LABEL: define {{[^@]+}}@test_st64b(
  95 // CHECK-C-NEXT:  entry:
  96 // CHECK-C-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
  97 // CHECK-C-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
  98 // CHECK-C-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
  99 // CHECK-C-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 100 // CHECK-C-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[BYVAL_TEMP]], ptr align 8 @val, i64 64, i1 false)
 101 // CHECK-C-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 102 // CHECK-C-NEXT:    store ptr [[BYVAL_TEMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 103 // CHECK-C-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 104 // CHECK-C-NEXT:    [[TMP2:%.*]] = load i64, ptr [[BYVAL_TEMP]], align 8
 105 // CHECK-C-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 1
 106 // CHECK-C-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 107 // CHECK-C-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 2
 108 // CHECK-C-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 109 // CHECK-C-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 3
 110 // CHECK-C-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 111 // CHECK-C-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 4
 112 // CHECK-C-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 113 // CHECK-C-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 5
 114 // CHECK-C-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 115 // CHECK-C-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 6
 116 // CHECK-C-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 117 // CHECK-C-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 7
 118 // CHECK-C-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 119 // CHECK-C-NEXT:    call void @llvm.aarch64.st64b(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 120 // CHECK-C-NEXT:    ret void
 121 //
 122 // CHECK-CXX-LABEL: define {{[^@]+}}@test_st64b(
 123 // CHECK-CXX-NEXT:  entry:
 124 // CHECK-CXX-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
 125 // CHECK-CXX-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
 126 // CHECK-CXX-NEXT:    [[AGG_TMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
 127 // CHECK-CXX-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 128 // CHECK-CXX-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[AGG_TMP]], ptr align 8 @val, i64 64, i1 false)
 129 // CHECK-CXX-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 130 // CHECK-CXX-NEXT:    store ptr [[AGG_TMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 131 // CHECK-CXX-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 132 // CHECK-CXX-NEXT:    [[TMP2:%.*]] = load i64, ptr [[AGG_TMP]], align 8
 133 // CHECK-CXX-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 1
 134 // CHECK-CXX-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 135 // CHECK-CXX-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 2
 136 // CHECK-CXX-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 137 // CHECK-CXX-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 3
 138 // CHECK-CXX-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 139 // CHECK-CXX-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 4
 140 // CHECK-CXX-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 141 // CHECK-CXX-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 5
 142 // CHECK-CXX-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 143 // CHECK-CXX-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 6
 144 // CHECK-CXX-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 145 // CHECK-CXX-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 7
 146 // CHECK-CXX-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 147 // CHECK-CXX-NEXT:    call void @llvm.aarch64.st64b(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 148 // CHECK-CXX-NEXT:    ret void
 149 //
 150 EXTERN_C void test_st64b(void)
 151 {
 152     __arm_st64b(addr, val);
 153 }
 154
 155 // CHECK-C-LABEL: define {{[^@]+}}@test_st64bv(
 156 // CHECK-C-NEXT:  entry:
 157 // CHECK-C-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
 158 // CHECK-C-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
 159 // CHECK-C-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
 160 // CHECK-C-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 161 // CHECK-C-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[BYVAL_TEMP]], ptr align 8 @val, i64 64, i1 false)
 162 // CHECK-C-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 163 // CHECK-C-NEXT:    store ptr [[BYVAL_TEMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 164 // CHECK-C-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 165 // CHECK-C-NEXT:    [[TMP2:%.*]] = load i64, ptr [[BYVAL_TEMP]], align 8
 166 // CHECK-C-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 1
 167 // CHECK-C-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 168 // CHECK-C-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 2
 169 // CHECK-C-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 170 // CHECK-C-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 3
 171 // CHECK-C-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 172 // CHECK-C-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 4
 173 // CHECK-C-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 174 // CHECK-C-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 5
 175 // CHECK-C-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 176 // CHECK-C-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 6
 177 // CHECK-C-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 178 // CHECK-C-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 7
 179 // CHECK-C-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 180 // CHECK-C-NEXT:    [[TMP17:%.*]] = call i64 @llvm.aarch64.st64bv(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 181 // CHECK-C-NEXT:    store i64 [[TMP17]], ptr @status, align 8
 182 // CHECK-C-NEXT:    ret void
 183 //
 184 // CHECK-CXX-LABEL: define {{[^@]+}}@test_st64bv(
 185 // CHECK-CXX-NEXT:  entry:
 186 // CHECK-CXX-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
 187 // CHECK-CXX-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
 188 // CHECK-CXX-NEXT:    [[AGG_TMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
 189 // CHECK-CXX-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 190 // CHECK-CXX-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[AGG_TMP]], ptr align 8 @val, i64 64, i1 false)
 191 // CHECK-CXX-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 192 // CHECK-CXX-NEXT:    store ptr [[AGG_TMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 193 // CHECK-CXX-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 194 // CHECK-CXX-NEXT:    [[TMP2:%.*]] = load i64, ptr [[AGG_TMP]], align 8
 195 // CHECK-CXX-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 1
 196 // CHECK-CXX-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 197 // CHECK-CXX-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 2
 198 // CHECK-CXX-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 199 // CHECK-CXX-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 3
 200 // CHECK-CXX-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 201 // CHECK-CXX-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 4
 202 // CHECK-CXX-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 203 // CHECK-CXX-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 5
 204 // CHECK-CXX-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 205 // CHECK-CXX-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 6
 206 // CHECK-CXX-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 207 // CHECK-CXX-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 7
 208 // CHECK-CXX-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 209 // CHECK-CXX-NEXT:    [[TMP17:%.*]] = call noundef i64 @llvm.aarch64.st64bv(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 210 // CHECK-CXX-NEXT:    store i64 [[TMP17]], ptr @status, align 8
 211 // CHECK-CXX-NEXT:    ret void
 212 //
 213 EXTERN_C void test_st64bv(void)
 214 {
 215     status = __arm_st64bv(addr, val);
 216 }
 217
 218 // CHECK-C-LABEL: define {{[^@]+}}@test_st64bv0(
 219 // CHECK-C-NEXT:  entry:
 220 // CHECK-C-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
 221 // CHECK-C-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
 222 // CHECK-C-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
 223 // CHECK-C-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 224 // CHECK-C-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[BYVAL_TEMP]], ptr align 8 @val, i64 64, i1 false)
 225 // CHECK-C-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 226 // CHECK-C-NEXT:    store ptr [[BYVAL_TEMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 227 // CHECK-C-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 228 // CHECK-C-NEXT:    [[TMP2:%.*]] = load i64, ptr [[BYVAL_TEMP]], align 8
 229 // CHECK-C-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 1
 230 // CHECK-C-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 231 // CHECK-C-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 2
 232 // CHECK-C-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 233 // CHECK-C-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 3
 234 // CHECK-C-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 235 // CHECK-C-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 4
 236 // CHECK-C-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 237 // CHECK-C-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 5
 238 // CHECK-C-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 239 // CHECK-C-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 6
 240 // CHECK-C-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 241 // CHECK-C-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[BYVAL_TEMP]], i32 7
 242 // CHECK-C-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 243 // CHECK-C-NEXT:    [[TMP17:%.*]] = call i64 @llvm.aarch64.st64bv0(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 244 // CHECK-C-NEXT:    store i64 [[TMP17]], ptr @status, align 8
 245 // CHECK-C-NEXT:    ret void
 246 //
 247 // CHECK-CXX-LABEL: define {{[^@]+}}@test_st64bv0(
 248 // CHECK-CXX-NEXT:  entry:
 249 // CHECK-CXX-NEXT:    [[__ADDR_ADDR_I:%.*]] = alloca ptr, align 8
 250 // CHECK-CXX-NEXT:    [[VALUE_INDIRECT_ADDR:%.*]] = alloca ptr, align 8
 251 // CHECK-CXX-NEXT:    [[AGG_TMP:%.*]] = alloca [[STRUCT_DATA512_T:%.*]], align 8
 252 // CHECK-CXX-NEXT:    [[TMP0:%.*]] = load ptr, ptr @addr, align 8
 253 // CHECK-CXX-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[AGG_TMP]], ptr align 8 @val, i64 64, i1 false)
 254 // CHECK-CXX-NEXT:    store ptr [[TMP0]], ptr [[__ADDR_ADDR_I]], align 8
 255 // CHECK-CXX-NEXT:    store ptr [[AGG_TMP]], ptr [[VALUE_INDIRECT_ADDR]], align 8
 256 // CHECK-CXX-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[__ADDR_ADDR_I]], align 8
 257 // CHECK-CXX-NEXT:    [[TMP2:%.*]] = load i64, ptr [[AGG_TMP]], align 8
 258 // CHECK-CXX-NEXT:    [[TMP3:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 1
 259 // CHECK-CXX-NEXT:    [[TMP4:%.*]] = load i64, ptr [[TMP3]], align 8
 260 // CHECK-CXX-NEXT:    [[TMP5:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 2
 261 // CHECK-CXX-NEXT:    [[TMP6:%.*]] = load i64, ptr [[TMP5]], align 8
 262 // CHECK-CXX-NEXT:    [[TMP7:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 3
 263 // CHECK-CXX-NEXT:    [[TMP8:%.*]] = load i64, ptr [[TMP7]], align 8
 264 // CHECK-CXX-NEXT:    [[TMP9:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 4
 265 // CHECK-CXX-NEXT:    [[TMP10:%.*]] = load i64, ptr [[TMP9]], align 8
 266 // CHECK-CXX-NEXT:    [[TMP11:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 5
 267 // CHECK-CXX-NEXT:    [[TMP12:%.*]] = load i64, ptr [[TMP11]], align 8
 268 // CHECK-CXX-NEXT:    [[TMP13:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 6
 269 // CHECK-CXX-NEXT:    [[TMP14:%.*]] = load i64, ptr [[TMP13]], align 8
 270 // CHECK-CXX-NEXT:    [[TMP15:%.*]] = getelementptr i64, ptr [[AGG_TMP]], i32 7
 271 // CHECK-CXX-NEXT:    [[TMP16:%.*]] = load i64, ptr [[TMP15]], align 8
 272 // CHECK-CXX-NEXT:    [[TMP17:%.*]] = call noundef i64 @llvm.aarch64.st64bv0(ptr [[TMP1]], i64 [[TMP2]], i64 [[TMP4]], i64 [[TMP6]], i64 [[TMP8]], i64 [[TMP10]], i64 [[TMP12]], i64 [[TMP14]], i64 [[TMP16]])
 273 // CHECK-CXX-NEXT:    store i64 [[TMP17]], ptr @status, align 8
 274 // CHECK-CXX-NEXT:    ret void
 275 //
 276 EXTERN_C void test_st64bv0(void)
 277 {
 278     status = __arm_st64bv0(addr, val);
 279 }