clang/test/CodeGen/RISCV/riscv64-vararg.c

   1 // NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py UTC_ARGS: --version 2
   2 // RUN: %clang_cc1 -triple riscv64 -emit-llvm %s -o - | FileCheck %s
   3 // RUN: %clang_cc1 -triple riscv64 -target-feature +f -target-abi lp64f -emit-llvm %s -o - \
   4 // RUN:     | FileCheck %s
   5 // RUN: %clang_cc1 -triple riscv64 -target-feature +d -target-feature +f -target-abi lp64d -emit-llvm %s -o - \
   6 // RUN:     | FileCheck %s
   7
   8 #include <stddef.h>
   9 #include <stdint.h>
  10
  11 struct tiny {
  12   uint16_t a, b, c, d;
  13 };
  14
  15 struct small {
  16   int64_t a, *b;
  17 };
  18
  19 struct small_aligned {
  20   __int128_t a;
  21 };
  22
  23 struct large {
  24   int64_t a, b, c, d;
  25 };
  26
  27 // Ensure that ABI lowering happens as expected for vararg calls.
  28 // Specifically, ensure that signext is emitted for varargs that will be
  29 // passed in registers but not on the stack. Ensure this takes into account
  30 // the use of "aligned" register pairs for varargs with 2*xlen alignment.
  31
  32 int f_va_callee(int, ...);
  33
  34 // CHECK-LABEL: define dso_local void @f_va_caller
  35 // CHECK-SAME: () #[[ATTR0:[0-9]+]] {
  36 // CHECK-NEXT:  entry:
  37 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL:%.*]] = alloca [[STRUCT_TINY:%.*]], align 2
  38 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL1:%.*]] = alloca [[STRUCT_SMALL:%.*]], align 8
  39 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL4:%.*]] = alloca [[STRUCT_SMALL_ALIGNED:%.*]], align 16
  40 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL6:%.*]] = alloca [[STRUCT_LARGE:%.*]], align 8
  41 // CHECK-NEXT:    [[BYVAL_TEMP:%.*]] = alloca [[STRUCT_LARGE]], align 8
  42 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL12:%.*]] = alloca [[STRUCT_SMALL_ALIGNED]], align 16
  43 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL16:%.*]] = alloca [[STRUCT_SMALL]], align 8
  44 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL21:%.*]] = alloca [[STRUCT_SMALL_ALIGNED]], align 16
  45 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL25:%.*]] = alloca [[STRUCT_SMALL]], align 8
  46 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL30:%.*]] = alloca [[STRUCT_SMALL_ALIGNED]], align 16
  47 // CHECK-NEXT:    [[DOTCOMPOUNDLITERAL34:%.*]] = alloca [[STRUCT_SMALL]], align 8
  48 // CHECK-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_TINY]], ptr [[DOTCOMPOUNDLITERAL]], i32 0, i32 0
  49 // CHECK-NEXT:    store i16 6, ptr [[A]], align 2
  50 // CHECK-NEXT:    [[B:%.*]] = getelementptr inbounds [[STRUCT_TINY]], ptr [[DOTCOMPOUNDLITERAL]], i32 0, i32 1
  51 // CHECK-NEXT:    store i16 7, ptr [[B]], align 2
  52 // CHECK-NEXT:    [[C:%.*]] = getelementptr inbounds [[STRUCT_TINY]], ptr [[DOTCOMPOUNDLITERAL]], i32 0, i32 2
  53 // CHECK-NEXT:    store i16 8, ptr [[C]], align 2
  54 // CHECK-NEXT:    [[D:%.*]] = getelementptr inbounds [[STRUCT_TINY]], ptr [[DOTCOMPOUNDLITERAL]], i32 0, i32 3
  55 // CHECK-NEXT:    store i16 9, ptr [[D]], align 2
  56 // CHECK-NEXT:    [[A2:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL1]], i32 0, i32 0
  57 // CHECK-NEXT:    store i64 10, ptr [[A2]], align 8
  58 // CHECK-NEXT:    [[B3:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL1]], i32 0, i32 1
  59 // CHECK-NEXT:    store ptr null, ptr [[B3]], align 8
  60 // CHECK-NEXT:    [[A5:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL4]], i32 0, i32 0
  61 // CHECK-NEXT:    store i128 11, ptr [[A5]], align 16
  62 // CHECK-NEXT:    [[A7:%.*]] = getelementptr inbounds [[STRUCT_LARGE]], ptr [[DOTCOMPOUNDLITERAL6]], i32 0, i32 0
  63 // CHECK-NEXT:    store i64 12, ptr [[A7]], align 8
  64 // CHECK-NEXT:    [[B8:%.*]] = getelementptr inbounds [[STRUCT_LARGE]], ptr [[DOTCOMPOUNDLITERAL6]], i32 0, i32 1
  65 // CHECK-NEXT:    store i64 13, ptr [[B8]], align 8
  66 // CHECK-NEXT:    [[C9:%.*]] = getelementptr inbounds [[STRUCT_LARGE]], ptr [[DOTCOMPOUNDLITERAL6]], i32 0, i32 2
  67 // CHECK-NEXT:    store i64 14, ptr [[C9]], align 8
  68 // CHECK-NEXT:    [[D10:%.*]] = getelementptr inbounds [[STRUCT_LARGE]], ptr [[DOTCOMPOUNDLITERAL6]], i32 0, i32 3
  69 // CHECK-NEXT:    store i64 15, ptr [[D10]], align 8
  70 // CHECK-NEXT:    [[TMP0:%.*]] = load i64, ptr [[DOTCOMPOUNDLITERAL]], align 2
  71 // CHECK-NEXT:    [[TMP1:%.*]] = load [2 x i64], ptr [[DOTCOMPOUNDLITERAL1]], align 8
  72 // CHECK-NEXT:    [[COERCE_DIVE:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL4]], i32 0, i32 0
  73 // CHECK-NEXT:    [[TMP2:%.*]] = load i128, ptr [[COERCE_DIVE]], align 16
  74 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[BYVAL_TEMP]], ptr align 8 [[DOTCOMPOUNDLITERAL6]], i64 32, i1 false)
  75 // CHECK-NEXT:    [[CALL:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i64 noundef 3, double noundef 4.000000e+00, double noundef 5.000000e+00, i64 [[TMP0]], [2 x i64] [[TMP1]], i128 [[TMP2]], ptr noundef [[BYVAL_TEMP]])
  76 // CHECK-NEXT:    [[CALL11:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, fp128 noundef 0xL00000000000000004001400000000000, i32 noundef signext 6, i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
  77 // CHECK-NEXT:    [[A13:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL12]], i32 0, i32 0
  78 // CHECK-NEXT:    store i128 5, ptr [[A13]], align 16
  79 // CHECK-NEXT:    [[COERCE_DIVE14:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL12]], i32 0, i32 0
  80 // CHECK-NEXT:    [[TMP3:%.*]] = load i128, ptr [[COERCE_DIVE14]], align 16
  81 // CHECK-NEXT:    [[CALL15:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i128 [[TMP3]], i32 noundef signext 6, i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
  82 // CHECK-NEXT:    [[A17:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL16]], i32 0, i32 0
  83 // CHECK-NEXT:    store i64 5, ptr [[A17]], align 8
  84 // CHECK-NEXT:    [[B18:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL16]], i32 0, i32 1
  85 // CHECK-NEXT:    store ptr null, ptr [[B18]], align 8
  86 // CHECK-NEXT:    [[TMP4:%.*]] = load [2 x i64], ptr [[DOTCOMPOUNDLITERAL16]], align 8
  87 // CHECK-NEXT:    [[CALL19:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, [2 x i64] [[TMP4]], i32 noundef signext 6, i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
  88 // CHECK-NEXT:    [[CALL20:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, fp128 noundef 0xL00000000000000004001800000000000, i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
  89 // CHECK-NEXT:    [[A22:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL21]], i32 0, i32 0
  90 // CHECK-NEXT:    store i128 6, ptr [[A22]], align 16
  91 // CHECK-NEXT:    [[COERCE_DIVE23:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL21]], i32 0, i32 0
  92 // CHECK-NEXT:    [[TMP5:%.*]] = load i128, ptr [[COERCE_DIVE23]], align 16
  93 // CHECK-NEXT:    [[CALL24:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, i128 [[TMP5]], i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
  94 // CHECK-NEXT:    [[A26:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL25]], i32 0, i32 0
  95 // CHECK-NEXT:    store i64 6, ptr [[A26]], align 8
  96 // CHECK-NEXT:    [[B27:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL25]], i32 0, i32 1
  97 // CHECK-NEXT:    store ptr null, ptr [[B27]], align 8
  98 // CHECK-NEXT:    [[TMP6:%.*]] = load [2 x i64], ptr [[DOTCOMPOUNDLITERAL25]], align 8
  99 // CHECK-NEXT:    [[CALL28:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, [2 x i64] [[TMP6]], i32 noundef signext 7, i32 noundef signext 8, i32 noundef signext 9)
 100 // CHECK-NEXT:    [[CALL29:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, i32 noundef signext 6, fp128 noundef 0xL00000000000000004001C00000000000, i32 noundef signext 8, i32 noundef signext 9)
 101 // CHECK-NEXT:    [[A31:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL30]], i32 0, i32 0
 102 // CHECK-NEXT:    store i128 7, ptr [[A31]], align 16
 103 // CHECK-NEXT:    [[COERCE_DIVE32:%.*]] = getelementptr inbounds [[STRUCT_SMALL_ALIGNED]], ptr [[DOTCOMPOUNDLITERAL30]], i32 0, i32 0
 104 // CHECK-NEXT:    [[TMP7:%.*]] = load i128, ptr [[COERCE_DIVE32]], align 16
 105 // CHECK-NEXT:    [[CALL33:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, i32 noundef signext 6, i128 [[TMP7]], i32 noundef signext 8, i32 noundef signext 9)
 106 // CHECK-NEXT:    [[A35:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL34]], i32 0, i32 0
 107 // CHECK-NEXT:    store i64 7, ptr [[A35]], align 8
 108 // CHECK-NEXT:    [[B36:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[DOTCOMPOUNDLITERAL34]], i32 0, i32 1
 109 // CHECK-NEXT:    store ptr null, ptr [[B36]], align 8
 110 // CHECK-NEXT:    [[TMP8:%.*]] = load [2 x i64], ptr [[DOTCOMPOUNDLITERAL34]], align 8
 111 // CHECK-NEXT:    [[CALL37:%.*]] = call signext i32 (i32, ...) @f_va_callee(i32 noundef signext 1, i32 noundef signext 2, i32 noundef signext 3, i32 noundef signext 4, i32 noundef signext 5, i32 noundef signext 6, [2 x i64] [[TMP8]], i32 noundef signext 8, i32 noundef signext 9)
 112 // CHECK-NEXT:    ret void
 113 //
 114 void f_va_caller(void) {
 115   f_va_callee(1, 2, 3LL, 4.0f, 5.0, (struct tiny){6, 7, 8, 9},
 116               (struct small){10, NULL}, (struct small_aligned){11},
 117               (struct large){12, 13, 14, 15});
 118   f_va_callee(1, 2, 3, 4, 5.0L, 6, 7, 8, 9);
 119   f_va_callee(1, 2, 3, 4, (struct small_aligned){5}, 6, 7, 8, 9);
 120   f_va_callee(1, 2, 3, 4, (struct small){5, NULL}, 6, 7, 8, 9);
 121   f_va_callee(1, 2, 3, 4, 5, 6.0L, 7, 8, 9);
 122   f_va_callee(1, 2, 3, 4, 5, (struct small_aligned){6}, 7, 8, 9);
 123   f_va_callee(1, 2, 3, 4, 5, (struct small){6, NULL}, 7, 8, 9);
 124   f_va_callee(1, 2, 3, 4, 5, 6, 7.0L, 8, 9);
 125   f_va_callee(1, 2, 3, 4, 5, 6, (struct small_aligned){7}, 8, 9);
 126   f_va_callee(1, 2, 3, 4, 5, 6, (struct small){7, NULL}, 8, 9);
 127 }
 128
 129 // CHECK-LABEL: define dso_local signext i32 @f_va_1
 130 // CHECK-SAME: (ptr noundef [[FMT:%.*]], ...) #[[ATTR0]] {
 131 // CHECK-NEXT:  entry:
 132 // CHECK-NEXT:    [[FMT_ADDR:%.*]] = alloca ptr, align 8
 133 // CHECK-NEXT:    [[VA:%.*]] = alloca ptr, align 8
 134 // CHECK-NEXT:    [[V:%.*]] = alloca i32, align 4
 135 // CHECK-NEXT:    store ptr [[FMT]], ptr [[FMT_ADDR]], align 8
 136 // CHECK-NEXT:    call void @llvm.va_start(ptr [[VA]])
 137 // CHECK-NEXT:    [[ARGP_CUR:%.*]] = load ptr, ptr [[VA]], align 8
 138 // CHECK-NEXT:    [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i64 8
 139 // CHECK-NEXT:    store ptr [[ARGP_NEXT]], ptr [[VA]], align 8
 140 // CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARGP_CUR]], align 8
 141 // CHECK-NEXT:    store i32 [[TMP0]], ptr [[V]], align 4
 142 // CHECK-NEXT:    call void @llvm.va_end(ptr [[VA]])
 143 // CHECK-NEXT:    [[TMP1:%.*]] = load i32, ptr [[V]], align 4
 144 // CHECK-NEXT:    ret i32 [[TMP1]]
 145 //
 146 int f_va_1(char *fmt, ...) {
 147   __builtin_va_list va;
 148
 149   __builtin_va_start(va, fmt);
 150   int v = __builtin_va_arg(va, int);
 151   __builtin_va_end(va);
 152
 153   return v;
 154 }
 155
 156 // An "aligned" register pair (where the first register is even-numbered) is
 157 // used to pass varargs with 2x xlen alignment and 2x xlen size. Ensure the
 158 // correct offsets are used.
 159
 160 // CHECK-LABEL: define dso_local fp128 @f_va_2
 161 // CHECK-SAME: (ptr noundef [[FMT:%.*]], ...) #[[ATTR0]] {
 162 // CHECK-NEXT:  entry:
 163 // CHECK-NEXT:    [[FMT_ADDR:%.*]] = alloca ptr, align 8
 164 // CHECK-NEXT:    [[VA:%.*]] = alloca ptr, align 8
 165 // CHECK-NEXT:    [[V:%.*]] = alloca fp128, align 16
 166 // CHECK-NEXT:    store ptr [[FMT]], ptr [[FMT_ADDR]], align 8
 167 // CHECK-NEXT:    call void @llvm.va_start(ptr [[VA]])
 168 // CHECK-NEXT:    [[ARGP_CUR:%.*]] = load ptr, ptr [[VA]], align 8
 169 // CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i32 15
 170 // CHECK-NEXT:    [[ARGP_CUR_ALIGNED:%.*]] = call ptr @llvm.ptrmask.p0.i64(ptr [[TMP0]], i64 -16)
 171 // CHECK-NEXT:    [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR_ALIGNED]], i64 16
 172 // CHECK-NEXT:    store ptr [[ARGP_NEXT]], ptr [[VA]], align 8
 173 // CHECK-NEXT:    [[TMP1:%.*]] = load fp128, ptr [[ARGP_CUR_ALIGNED]], align 16
 174 // CHECK-NEXT:    store fp128 [[TMP1]], ptr [[V]], align 16
 175 // CHECK-NEXT:    call void @llvm.va_end(ptr [[VA]])
 176 // CHECK-NEXT:    [[TMP2:%.*]] = load fp128, ptr [[V]], align 16
 177 // CHECK-NEXT:    ret fp128 [[TMP2]]
 178 //
 179 long double f_va_2(char *fmt, ...) {
 180   __builtin_va_list va;
 181
 182   __builtin_va_start(va, fmt);
 183   long double v = __builtin_va_arg(va, long double);
 184   __builtin_va_end(va);
 185
 186   return v;
 187 }
 188
 189 // Two "aligned" register pairs.
 190
 191 // CHECK-LABEL: define dso_local fp128 @f_va_3
 192 // CHECK-SAME: (ptr noundef [[FMT:%.*]], ...) #[[ATTR0]] {
 193 // CHECK-NEXT:  entry:
 194 // CHECK-NEXT:    [[FMT_ADDR:%.*]] = alloca ptr, align 8
 195 // CHECK-NEXT:    [[VA:%.*]] = alloca ptr, align 8
 196 // CHECK-NEXT:    [[V:%.*]] = alloca fp128, align 16
 197 // CHECK-NEXT:    [[W:%.*]] = alloca i32, align 4
 198 // CHECK-NEXT:    [[X:%.*]] = alloca fp128, align 16
 199 // CHECK-NEXT:    store ptr [[FMT]], ptr [[FMT_ADDR]], align 8
 200 // CHECK-NEXT:    call void @llvm.va_start(ptr [[VA]])
 201 // CHECK-NEXT:    [[ARGP_CUR:%.*]] = load ptr, ptr [[VA]], align 8
 202 // CHECK-NEXT:    [[TMP0:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i32 15
 203 // CHECK-NEXT:    [[ARGP_CUR_ALIGNED:%.*]] = call ptr @llvm.ptrmask.p0.i64(ptr [[TMP0]], i64 -16)
 204 // CHECK-NEXT:    [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR_ALIGNED]], i64 16
 205 // CHECK-NEXT:    store ptr [[ARGP_NEXT]], ptr [[VA]], align 8
 206 // CHECK-NEXT:    [[TMP1:%.*]] = load fp128, ptr [[ARGP_CUR_ALIGNED]], align 16
 207 // CHECK-NEXT:    store fp128 [[TMP1]], ptr [[V]], align 16
 208 // CHECK-NEXT:    [[ARGP_CUR1:%.*]] = load ptr, ptr [[VA]], align 8
 209 // CHECK-NEXT:    [[ARGP_NEXT2:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR1]], i64 8
 210 // CHECK-NEXT:    store ptr [[ARGP_NEXT2]], ptr [[VA]], align 8
 211 // CHECK-NEXT:    [[TMP2:%.*]] = load i32, ptr [[ARGP_CUR1]], align 8
 212 // CHECK-NEXT:    store i32 [[TMP2]], ptr [[W]], align 4
 213 // CHECK-NEXT:    [[ARGP_CUR3:%.*]] = load ptr, ptr [[VA]], align 8
 214 // CHECK-NEXT:    [[TMP3:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR3]], i32 15
 215 // CHECK-NEXT:    [[ARGP_CUR3_ALIGNED:%.*]] = call ptr @llvm.ptrmask.p0.i64(ptr [[TMP3]], i64 -16)
 216 // CHECK-NEXT:    [[ARGP_NEXT4:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR3_ALIGNED]], i64 16
 217 // CHECK-NEXT:    store ptr [[ARGP_NEXT4]], ptr [[VA]], align 8
 218 // CHECK-NEXT:    [[TMP4:%.*]] = load fp128, ptr [[ARGP_CUR3_ALIGNED]], align 16
 219 // CHECK-NEXT:    store fp128 [[TMP4]], ptr [[X]], align 16
 220 // CHECK-NEXT:    call void @llvm.va_end(ptr [[VA]])
 221 // CHECK-NEXT:    [[TMP5:%.*]] = load fp128, ptr [[V]], align 16
 222 // CHECK-NEXT:    [[TMP6:%.*]] = load fp128, ptr [[X]], align 16
 223 // CHECK-NEXT:    [[ADD:%.*]] = fadd fp128 [[TMP5]], [[TMP6]]
 224 // CHECK-NEXT:    ret fp128 [[ADD]]
 225 //
 226 long double f_va_3(char *fmt, ...) {
 227   __builtin_va_list va;
 228
 229   __builtin_va_start(va, fmt);
 230   long double v = __builtin_va_arg(va, long double);
 231   int w = __builtin_va_arg(va, int);
 232   long double x = __builtin_va_arg(va, long double);
 233   __builtin_va_end(va);
 234
 235   return v + x;
 236 }
 237
 238 // CHECK-LABEL: define dso_local signext i32 @f_va_4
 239 // CHECK-SAME: (ptr noundef [[FMT:%.*]], ...) #[[ATTR0]] {
 240 // CHECK-NEXT:  entry:
 241 // CHECK-NEXT:    [[FMT_ADDR:%.*]] = alloca ptr, align 8
 242 // CHECK-NEXT:    [[VA:%.*]] = alloca ptr, align 8
 243 // CHECK-NEXT:    [[V:%.*]] = alloca i32, align 4
 244 // CHECK-NEXT:    [[TS:%.*]] = alloca [[STRUCT_TINY:%.*]], align 2
 245 // CHECK-NEXT:    [[SS:%.*]] = alloca [[STRUCT_SMALL:%.*]], align 8
 246 // CHECK-NEXT:    [[LS:%.*]] = alloca [[STRUCT_LARGE:%.*]], align 8
 247 // CHECK-NEXT:    [[RET:%.*]] = alloca i32, align 4
 248 // CHECK-NEXT:    store ptr [[FMT]], ptr [[FMT_ADDR]], align 8
 249 // CHECK-NEXT:    call void @llvm.va_start(ptr [[VA]])
 250 // CHECK-NEXT:    [[ARGP_CUR:%.*]] = load ptr, ptr [[VA]], align 8
 251 // CHECK-NEXT:    [[ARGP_NEXT:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR]], i64 8
 252 // CHECK-NEXT:    store ptr [[ARGP_NEXT]], ptr [[VA]], align 8
 253 // CHECK-NEXT:    [[TMP0:%.*]] = load i32, ptr [[ARGP_CUR]], align 8
 254 // CHECK-NEXT:    store i32 [[TMP0]], ptr [[V]], align 4
 255 // CHECK-NEXT:    [[ARGP_CUR1:%.*]] = load ptr, ptr [[VA]], align 8
 256 // CHECK-NEXT:    [[ARGP_NEXT2:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR1]], i64 8
 257 // CHECK-NEXT:    store ptr [[ARGP_NEXT2]], ptr [[VA]], align 8
 258 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 2 [[TS]], ptr align 8 [[ARGP_CUR1]], i64 8, i1 false)
 259 // CHECK-NEXT:    [[ARGP_CUR3:%.*]] = load ptr, ptr [[VA]], align 8
 260 // CHECK-NEXT:    [[ARGP_NEXT4:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR3]], i64 16
 261 // CHECK-NEXT:    store ptr [[ARGP_NEXT4]], ptr [[VA]], align 8
 262 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[SS]], ptr align 8 [[ARGP_CUR3]], i64 16, i1 false)
 263 // CHECK-NEXT:    [[ARGP_CUR5:%.*]] = load ptr, ptr [[VA]], align 8
 264 // CHECK-NEXT:    [[ARGP_NEXT6:%.*]] = getelementptr inbounds i8, ptr [[ARGP_CUR5]], i64 8
 265 // CHECK-NEXT:    store ptr [[ARGP_NEXT6]], ptr [[VA]], align 8
 266 // CHECK-NEXT:    [[TMP1:%.*]] = load ptr, ptr [[ARGP_CUR5]], align 8
 267 // CHECK-NEXT:    call void @llvm.memcpy.p0.p0.i64(ptr align 8 [[LS]], ptr align 8 [[TMP1]], i64 32, i1 false)
 268 // CHECK-NEXT:    call void @llvm.va_end(ptr [[VA]])
 269 // CHECK-NEXT:    [[A:%.*]] = getelementptr inbounds [[STRUCT_TINY]], ptr [[TS]], i32 0, i32 0
 270 // CHECK-NEXT:    [[TMP2:%.*]] = load i16, ptr [[A]], align 2
 271 // CHECK-NEXT:    [[CONV:%.*]] = zext i16 [[TMP2]] to i64
 272 // CHECK-NEXT:    [[A7:%.*]] = getelementptr inbounds [[STRUCT_SMALL]], ptr [[SS]], i32 0, i32 0
 273 // CHECK-NEXT:    [[TMP3:%.*]] = load i64, ptr [[A7]], align 8
 274 // CHECK-NEXT:    [[ADD:%.*]] = add nsw i64 [[CONV]], [[TMP3]]
 275 // CHECK-NEXT:    [[C:%.*]] = getelementptr inbounds [[STRUCT_LARGE]], ptr [[LS]], i32 0, i32 2
 276 // CHECK-NEXT:    [[TMP4:%.*]] = load i64, ptr [[C]], align 8
 277 // CHECK-NEXT:    [[ADD8:%.*]] = add nsw i64 [[ADD]], [[TMP4]]
 278 // CHECK-NEXT:    [[CONV9:%.*]] = trunc i64 [[ADD8]] to i32
 279 // CHECK-NEXT:    store i32 [[CONV9]], ptr [[RET]], align 4
 280 // CHECK-NEXT:    [[TMP5:%.*]] = load i32, ptr [[RET]], align 4
 281 // CHECK-NEXT:    ret i32 [[TMP5]]
 282 //
 283 int f_va_4(char *fmt, ...) {
 284   __builtin_va_list va;
 285
 286   __builtin_va_start(va, fmt);
 287   int v = __builtin_va_arg(va, int);
 288   struct tiny ts = __builtin_va_arg(va, struct tiny);
 289   struct small ss = __builtin_va_arg(va, struct small);
 290   struct large ls = __builtin_va_arg(va, struct large);
 291   __builtin_va_end(va);
 292
 293   int ret = ts.a + ss.a + ls.c;
 294
 295   return ret;
 296 }