1 // RUN: %clang_cc1 -emit-llvm -o %t %s
2 // RUN: not grep __builtin %t
3 // RUN: %clang_cc1 %s -emit-llvm -o - -triple x86_64-darwin-apple | FileCheck %s
4 
5 int printf(const char *, ...);
6 
7 void p(char *str, int x) {
8   printf("%s: %d\n", str, x);
9 }
10 void q(char *str, double x) {
11   printf("%s: %f\n", str, x);
12 }
13 void r(char *str, void *ptr) {
14   printf("%s: %p\n", str, ptr);
15 }
16 
17 int random(void);
18 
19 int main() {
20   int N = random();
21 #define P(n,args) p(#n #args, __builtin_##n args)
22 #define Q(n,args) q(#n #args, __builtin_##n args)
23 #define R(n,args) r(#n #args, __builtin_##n args)
24 #define V(n,args) p(#n #args, (__builtin_##n args, 0))
25   P(types_compatible_p, (int, float));
26   P(choose_expr, (0, 10, 20));
27   P(constant_p, (sizeof(10)));
28   P(expect, (N == 12, 0));
29   V(prefetch, (&N));
30   V(prefetch, (&N, 1));
31   V(prefetch, (&N, 1, 0));
32 
33   // Numeric Constants
34 
35   Q(huge_val, ());
36   Q(huge_valf, ());
37   Q(huge_vall, ());
38   Q(inf, ());
39   Q(inff, ());
40   Q(infl, ());
41 
42   P(fpclassify, (0, 1, 2, 3, 4, 1.0));
43   P(fpclassify, (0, 1, 2, 3, 4, 1.0f));
44   P(fpclassify, (0, 1, 2, 3, 4, 1.0l));
45 
46   Q(nan, (""));
47   Q(nanf, (""));
48   Q(nanl, (""));
49   Q(nans, (""));
50   Q(nan, ("10"));
51   Q(nanf, ("10"));
52   Q(nanl, ("10"));
53   Q(nans, ("10"));
54 
55   P(isgreater, (1., 2.));
56   P(isgreaterequal, (1., 2.));
57   P(isless, (1., 2.));
58   P(islessequal, (1., 2.));
59   P(islessgreater, (1., 2.));
60   P(isunordered, (1., 2.));
61 
62   P(isinf, (1.));
63   P(isinf_sign, (1.));
64   P(isnan, (1.));
65 
66   // Bitwise & Numeric Functions
67 
68   P(abs, (N));
69 
70   P(clz, (N));
71   P(clzl, (N));
72   P(clzll, (N));
73   P(ctz, (N));
74   P(ctzl, (N));
75   P(ctzll, (N));
76   P(ffs, (N));
77   P(ffsl, (N));
78   P(ffsll, (N));
79   P(parity, (N));
80   P(parityl, (N));
81   P(parityll, (N));
82   P(popcount, (N));
83   P(popcountl, (N));
84   P(popcountll, (N));
85   Q(powi, (1.2f, N));
86   Q(powif, (1.2f, N));
87   Q(powil, (1.2f, N));
88 
89   // Lib functions
90   int a, b, n = random(); // Avoid optimizing out.
91   char s0[10], s1[] = "Hello";
92   V(strcat, (s0, s1));
93   V(strcmp, (s0, s1));
94   V(strncat, (s0, s1, n));
95   V(strchr, (s0, s1[0]));
96   V(strrchr, (s0, s1[0]));
97   V(strcpy, (s0, s1));
98   V(strncpy, (s0, s1, n));
99 
100   // Object size checking
101   V(__memset_chk, (s0, 0, sizeof s0, n));
102   V(__memcpy_chk, (s0, s1, sizeof s0, n));
103   V(__memmove_chk, (s0, s1, sizeof s0, n));
104   V(__mempcpy_chk, (s0, s1, sizeof s0, n));
105   V(__strncpy_chk, (s0, s1, sizeof s0, n));
106   V(__strcpy_chk, (s0, s1, n));
107   s0[0] = 0;
108   V(__strcat_chk, (s0, s1, n));
109   P(object_size, (s0, 0));
110   P(object_size, (s0, 1));
111   P(object_size, (s0, 2));
112   P(object_size, (s0, 3));
113 
114   // Whatever
115 
116   P(bswap16, (N));
117   P(bswap32, (N));
118   P(bswap64, (N));
119 
120   // CHECK: @llvm.bitreverse.i8
121   // CHECK: @llvm.bitreverse.i16
122   // CHECK: @llvm.bitreverse.i32
123   // CHECK: @llvm.bitreverse.i64
124   P(bitreverse8, (N));
125   P(bitreverse16, (N));
126   P(bitreverse32, (N));
127   P(bitreverse64, (N));
128 
129   // FIXME
130   // V(clear_cache, (&N, &N+1));
131   V(trap, ());
132   R(extract_return_addr, (&N));
133   P(signbit, (1.0));
134 
135   return 0;
136 }
137 
138 
139 
140 void foo() {
141  __builtin_strcat(0, 0);
142 }
143 
144 // CHECK-LABEL: define void @bar(
145 void bar() {
146   float f;
147   double d;
148   long double ld;
149 
150   // LLVM's hex representation of float constants is really unfortunate;
151   // basically it does a float-to-double "conversion" and then prints the
152   // hex form of that.  That gives us weird artifacts like exponents
153   // that aren't numerically similar to the original exponent and
154   // significand bit-patterns that are offset by three bits (because
155   // the exponent was expanded from 8 bits to 11).
156   //
157   // 0xAE98 == 1010111010011000
158   // 0x15D3 == 1010111010011
159 
160   f = __builtin_huge_valf();     // CHECK: float    0x7FF0000000000000
161   d = __builtin_huge_val();      // CHECK: double   0x7FF0000000000000
162   ld = __builtin_huge_vall();    // CHECK: x86_fp80 0xK7FFF8000000000000000
163   f = __builtin_nanf("");        // CHECK: float    0x7FF8000000000000
164   d = __builtin_nan("");         // CHECK: double   0x7FF8000000000000
165   ld = __builtin_nanl("");       // CHECK: x86_fp80 0xK7FFFC000000000000000
166   f = __builtin_nanf("0xAE98");  // CHECK: float    0x7FF815D300000000
167   d = __builtin_nan("0xAE98");   // CHECK: double   0x7FF800000000AE98
168   ld = __builtin_nanl("0xAE98"); // CHECK: x86_fp80 0xK7FFFC00000000000AE98
169   f = __builtin_nansf("");       // CHECK: float    0x7FF4000000000000
170   d = __builtin_nans("");        // CHECK: double   0x7FF4000000000000
171   ld = __builtin_nansl("");      // CHECK: x86_fp80 0xK7FFFA000000000000000
172   f = __builtin_nansf("0xAE98"); // CHECK: float    0x7FF015D300000000
173   d = __builtin_nans("0xAE98");  // CHECK: double   0x7FF000000000AE98
174   ld = __builtin_nansl("0xAE98");// CHECK: x86_fp80 0xK7FFF800000000000AE98
175 
176 }
177 // CHECK: }
178 
179 // CHECK-LABEL: define void @test_conditional_bzero
180 void test_conditional_bzero() {
181   char dst[20];
182   int _sz = 20, len = 20;
183   return (_sz
184           ? ((_sz >= len)
185               ? __builtin_bzero(dst, len)
186               : foo())
187           : __builtin_bzero(dst, len));
188   // CHECK: call void @llvm.memset
189   // CHECK: call void @llvm.memset
190   // CHECK-NOT: phi
191 }
192 
193 // CHECK-LABEL: define void @test_float_builtins
194 void test_float_builtins(float F, double D, long double LD) {
195   volatile int res;
196   res = __builtin_isinf(F);
197   // CHECK:  call float @llvm.fabs.f32(float
198   // CHECK:  fcmp oeq float {{.*}}, 0x7FF0000000000000
199 
200   res = __builtin_isinf(D);
201   // CHECK:  call double @llvm.fabs.f64(double
202   // CHECK:  fcmp oeq double {{.*}}, 0x7FF0000000000000
203 
204   res = __builtin_isinf(LD);
205   // CHECK:  call x86_fp80 @llvm.fabs.f80(x86_fp80
206   // CHECK:  fcmp oeq x86_fp80 {{.*}}, 0xK7FFF8000000000000000
207 
208   res = __builtin_isinf_sign(F);
209   // CHECK:  %[[ABS:.*]] = call float @llvm.fabs.f32(float %[[ARG:.*]])
210   // CHECK:  %[[ISINF:.*]] = fcmp oeq float %[[ABS]], 0x7FF0000000000000
211   // CHECK:  %[[BITCAST:.*]] = bitcast float %[[ARG]] to i32
212   // CHECK:  %[[ISNEG:.*]] = icmp slt i32 %[[BITCAST]], 0
213   // CHECK:  %[[SIGN:.*]] = select i1 %[[ISNEG]], i32 -1, i32 1
214   // CHECK:  select i1 %[[ISINF]], i32 %[[SIGN]], i32 0
215 
216   res = __builtin_isinf_sign(D);
217   // CHECK:  %[[ABS:.*]] = call double @llvm.fabs.f64(double %[[ARG:.*]])
218   // CHECK:  %[[ISINF:.*]] = fcmp oeq double %[[ABS]], 0x7FF0000000000000
219   // CHECK:  %[[BITCAST:.*]] = bitcast double %[[ARG]] to i64
220   // CHECK:  %[[ISNEG:.*]] = icmp slt i64 %[[BITCAST]], 0
221   // CHECK:  %[[SIGN:.*]] = select i1 %[[ISNEG]], i32 -1, i32 1
222   // CHECK:  select i1 %[[ISINF]], i32 %[[SIGN]], i32 0
223 
224   res = __builtin_isinf_sign(LD);
225   // CHECK:  %[[ABS:.*]] = call x86_fp80 @llvm.fabs.f80(x86_fp80 %[[ARG:.*]])
226   // CHECK:  %[[ISINF:.*]] = fcmp oeq x86_fp80 %[[ABS]], 0xK7FFF8000000000000000
227   // CHECK:  %[[BITCAST:.*]] = bitcast x86_fp80 %[[ARG]] to i80
228   // CHECK:  %[[ISNEG:.*]] = icmp slt i80 %[[BITCAST]], 0
229   // CHECK:  %[[SIGN:.*]] = select i1 %[[ISNEG]], i32 -1, i32 1
230   // CHECK:  select i1 %[[ISINF]], i32 %[[SIGN]], i32 0
231 
232   res = __builtin_isfinite(F);
233   // CHECK: call float @llvm.fabs.f32(float
234   // CHECK: fcmp one float {{.*}}, 0x7FF0000000000000
235 
236   res = finite(D);
237   // CHECK: call double @llvm.fabs.f64(double
238   // CHECK: fcmp one double {{.*}}, 0x7FF0000000000000
239 
240   res = __builtin_isnormal(F);
241   // CHECK: fcmp oeq float
242   // CHECK: call float @llvm.fabs.f32(float
243   // CHECK: fcmp ult float {{.*}}, 0x7FF0000000000000
244   // CHECK: fcmp uge float {{.*}}, 0x3810000000000000
245   // CHECK: and i1
246   // CHECK: and i1
247 }
248 
249 // CHECK-LABEL: define void @test_float_builtin_ops
250 void test_float_builtin_ops(float F, double D, long double LD) {
251   volatile float resf;
252   volatile double resd;
253   volatile long double resld;
254 
255   resf = __builtin_fmodf(F,F);
256   // CHECK: frem float
257 
258   resd = __builtin_fmod(D,D);
259   // CHECK: frem double
260 
261   resld = __builtin_fmodl(LD,LD);
262   // CHECK: frem x86_fp80
263 
264   resf = __builtin_fabsf(F);
265   resd = __builtin_fabs(D);
266   resld = __builtin_fabsl(LD);
267   // CHECK: call float @llvm.fabs.f32(float
268   // CHECK: call double @llvm.fabs.f64(double
269   // CHECK: call x86_fp80 @llvm.fabs.f80(x86_fp80
270 
271   resf = __builtin_canonicalizef(F);
272   resd = __builtin_canonicalize(D);
273   resld = __builtin_canonicalizel(LD);
274   // CHECK: call float @llvm.canonicalize.f32(float
275   // CHECK: call double @llvm.canonicalize.f64(double
276   // CHECK: call x86_fp80 @llvm.canonicalize.f80(x86_fp80
277 
278   resf = __builtin_fminf(F, F);
279   // CHECK: call float @llvm.minnum.f32
280 
281   resd = __builtin_fmin(D, D);
282   // CHECK: call double @llvm.minnum.f64
283 
284   resld = __builtin_fminl(LD, LD);
285   // CHECK: call x86_fp80 @llvm.minnum.f80
286 
287   resf = __builtin_fmaxf(F, F);
288   // CHECK: call float @llvm.maxnum.f32
289 
290   resd = __builtin_fmax(D, D);
291   // CHECK: call double @llvm.maxnum.f64
292 
293   resld = __builtin_fmaxl(LD, LD);
294   // CHECK: call x86_fp80 @llvm.maxnum.f80
295 
296   resf = __builtin_fabsf(F);
297   // CHECK: call float @llvm.fabs.f32
298 
299   resd = __builtin_fabs(D);
300   // CHECK: call double @llvm.fabs.f64
301 
302   resld = __builtin_fabsl(LD);
303   // CHECK: call x86_fp80 @llvm.fabs.f80
304 
305   resf = __builtin_copysignf(F, F);
306   // CHECK: call float @llvm.copysign.f32
307 
308   resd = __builtin_copysign(D, D);
309   // CHECK: call double @llvm.copysign.f64
310 
311   resld = __builtin_copysignl(LD, LD);
312   // CHECK: call x86_fp80 @llvm.copysign.f80
313 
314 
315   resf = __builtin_ceilf(F);
316   // CHECK: call float @llvm.ceil.f32
317 
318   resd = __builtin_ceil(D);
319   // CHECK: call double @llvm.ceil.f64
320 
321   resld = __builtin_ceill(LD);
322   // CHECK: call x86_fp80 @llvm.ceil.f80
323 
324   resf = __builtin_floorf(F);
325   // CHECK: call float @llvm.floor.f32
326 
327   resd = __builtin_floor(D);
328   // CHECK: call double @llvm.floor.f64
329 
330   resld = __builtin_floorl(LD);
331   // CHECK: call x86_fp80 @llvm.floor.f80
332 
333   resf = __builtin_sqrtf(F);
334   // CHECK: call float @llvm.sqrt.f32(
335 
336   resd = __builtin_sqrt(D);
337   // CHECK: call double @llvm.sqrt.f64(
338 
339   resld = __builtin_sqrtl(LD);
340   // CHECK: call x86_fp80 @llvm.sqrt.f80
341 
342   resf = __builtin_truncf(F);
343   // CHECK: call float @llvm.trunc.f32
344 
345   resd = __builtin_trunc(D);
346   // CHECK: call double @llvm.trunc.f64
347 
348   resld = __builtin_truncl(LD);
349   // CHECK: call x86_fp80 @llvm.trunc.f80
350 
351   resf = __builtin_rintf(F);
352   // CHECK: call float @llvm.rint.f32
353 
354   resd = __builtin_rint(D);
355   // CHECK: call double @llvm.rint.f64
356 
357   resld = __builtin_rintl(LD);
358   // CHECK: call x86_fp80 @llvm.rint.f80
359 
360   resf = __builtin_nearbyintf(F);
361   // CHECK: call float @llvm.nearbyint.f32
362 
363   resd = __builtin_nearbyint(D);
364   // CHECK: call double @llvm.nearbyint.f64
365 
366   resld = __builtin_nearbyintl(LD);
367   // CHECK: call x86_fp80 @llvm.nearbyint.f80
368 
369   resf = __builtin_roundf(F);
370   // CHECK: call float @llvm.round.f32
371 
372   resd = __builtin_round(D);
373   // CHECK: call double @llvm.round.f64
374 
375   resld = __builtin_roundl(LD);
376   // CHECK: call x86_fp80 @llvm.round.f80
377 
378 }
379 
380 // __builtin_longjmp isn't supported on all platforms, so only test it on X86.
381 #ifdef __x86_64__
382 
383 // CHECK-LABEL: define void @test_builtin_longjmp
384 void test_builtin_longjmp(void **buffer) {
385   // CHECK: [[BITCAST:%.*]] = bitcast
386   // CHECK-NEXT: call void @llvm.eh.sjlj.longjmp(i8* [[BITCAST]])
387   __builtin_longjmp(buffer, 1);
388   // CHECK-NEXT: unreachable
389 }
390 
391 #endif
392 
393 // CHECK-LABEL: define i64 @test_builtin_readcyclecounter
394 long long test_builtin_readcyclecounter() {
395   // CHECK: call i64 @llvm.readcyclecounter()
396   return __builtin_readcyclecounter();
397 }
398 
399 // Behavior of __builtin_os_log differs between platforms, so only test on X86
400 #ifdef __x86_64__
401 
402 // CHECK-LABEL: define void @test_builtin_os_log
403 // CHECK: (i8* %[[BUF:.*]], i32 %[[I:.*]], i8* %[[DATA:.*]])
404 void test_builtin_os_log(void *buf, int i, const char *data) {
405   volatile int len;
406   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
407   // CHECK: %[[I_ADDR:.*]] = alloca i32, align 4
408   // CHECK: %[[DATA_ADDR:.*]] = alloca i8*, align 8
409   // CHECK: %[[LEN:.*]] = alloca i32, align 4
410   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
411   // CHECK: store i32 %[[I]], i32* %[[I_ADDR]], align 4
412   // CHECK: store i8* %[[DATA]], i8** %[[DATA_ADDR]], align 8
413 
414   // CHECK: store volatile i32 34, i32* %[[LEN]]
415   len = __builtin_os_log_format_buffer_size("%d %{public}s %{private}.16P", i, data, data);
416 
417   // CHECK: %[[V1:.*]] = load i8*, i8** %[[BUF_ADDR]]
418   // CHECK: %[[V2:.*]] = load i32, i32* %[[I_ADDR]]
419   // CHECK: %[[V3:.*]] = load i8*, i8** %[[DATA_ADDR]]
420   // CHECK: %[[V4:.*]] = ptrtoint i8* %[[V3]] to i64
421   // CHECK: %[[V5:.*]] = load i8*, i8** %[[DATA_ADDR]]
422   // CHECK: %[[V6:.*]] = ptrtoint i8* %[[V5]] to i64
423   // CHECK: call void @__os_log_helper_1_3_4_4_0_8_34_4_17_8_49(i8* %[[V1]], i32 %[[V2]], i64 %[[V4]], i32 16, i64 %[[V6]])
424   __builtin_os_log_format(buf, "%d %{public}s %{private}.16P", i, data, data);
425 }
426 
427 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_3_4_4_0_8_34_4_17_8_49
428 // CHECK: (i8* %[[BUFFER:.*]], i32 %[[ARG0:.*]], i64 %[[ARG1:.*]], i32 %[[ARG2:.*]], i64 %[[ARG3:.*]])
429 
430 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
431 // CHECK: %[[ARG0_ADDR:.*]] = alloca i32, align 4
432 // CHECK: %[[ARG1_ADDR:.*]] = alloca i64, align 8
433 // CHECK: %[[ARG2_ADDR:.*]] = alloca i32, align 4
434 // CHECK: %[[ARG3_ADDR:.*]] = alloca i64, align 8
435 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
436 // CHECK: store i32 %[[ARG0]], i32* %[[ARG0_ADDR]], align 4
437 // CHECK: store i64 %[[ARG1]], i64* %[[ARG1_ADDR]], align 8
438 // CHECK: store i32 %[[ARG2]], i32* %[[ARG2_ADDR]], align 4
439 // CHECK: store i64 %[[ARG3]], i64* %[[ARG3_ADDR]], align 8
440 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
441 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
442 // CHECK: store i8 3, i8* %[[SUMMARY]], align 1
443 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
444 // CHECK: store i8 4, i8* %[[NUMARGS]], align 1
445 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
446 // CHECK: store i8 0, i8* %[[ARGDESCRIPTOR]], align 1
447 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
448 // CHECK: store i8 4, i8* %[[ARGSIZE]], align 1
449 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
450 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i32*
451 // CHECK: %[[V0:.*]] = load i32, i32* %[[ARG0_ADDR]], align 4
452 // CHECK: store i32 %[[V0]], i32* %[[ARGDATACAST]], align 1
453 // CHECK: %[[ARGDESCRIPTOR1:.*]] = getelementptr i8, i8* %[[BUF]], i64 8
454 // CHECK: store i8 34, i8* %[[ARGDESCRIPTOR1]], align 1
455 // CHECK: %[[ARGSIZE2:.*]] = getelementptr i8, i8* %[[BUF]], i64 9
456 // CHECK: store i8 8, i8* %[[ARGSIZE2]], align 1
457 // CHECK: %[[ARGDATA3:.*]] = getelementptr i8, i8* %[[BUF]], i64 10
458 // CHECK: %[[ARGDATACAST4:.*]] = bitcast i8* %[[ARGDATA3]] to i64*
459 // CHECK: %[[V1:.*]] = load i64, i64* %[[ARG1_ADDR]], align 8
460 // CHECK: store i64 %[[V1]], i64* %[[ARGDATACAST4]], align 1
461 // CHECK: %[[ARGDESCRIPTOR5:.*]] = getelementptr i8, i8* %[[BUF]], i64 18
462 // CHECK: store i8 17, i8* %[[ARGDESCRIPTOR5]], align 1
463 // CHECK: %[[ARGSIZE6:.*]] = getelementptr i8, i8* %[[BUF]], i64 19
464 // CHECK: store i8 4, i8* %[[ARGSIZE6]], align 1
465 // CHECK: %[[ARGDATA7:.*]] = getelementptr i8, i8* %[[BUF]], i64 20
466 // CHECK: %[[ARGDATACAST8:.*]] = bitcast i8* %[[ARGDATA7]] to i32*
467 // CHECK: %[[V2:.*]] = load i32, i32* %[[ARG2_ADDR]], align 4
468 // CHECK: store i32 %[[V2]], i32* %[[ARGDATACAST8]], align 1
469 // CHECK: %[[ARGDESCRIPTOR9:.*]] = getelementptr i8, i8* %[[BUF]], i64 24
470 // CHECK: store i8 49, i8* %[[ARGDESCRIPTOR9]], align 1
471 // CHECK: %[[ARGSIZE10:.*]] = getelementptr i8, i8* %[[BUF]], i64 25
472 // CHECK: store i8 8, i8* %[[ARGSIZE10]], align 1
473 // CHECK: %[[ARGDATA11:.*]] = getelementptr i8, i8* %[[BUF]], i64 26
474 // CHECK: %[[ARGDATACAST12:.*]] = bitcast i8* %[[ARGDATA11]] to i64*
475 // CHECK: %[[V3:.*]] = load i64, i64* %[[ARG3_ADDR]], align 8
476 // CHECK: store i64 %[[V3]], i64* %[[ARGDATACAST12]], align 1
477 
478 // CHECK-LABEL: define void @test_builtin_os_log_wide
479 // CHECK: (i8* %[[BUF:.*]], i8* %[[DATA:.*]], i32* %[[STR:.*]])
480 typedef int wchar_t;
481 void test_builtin_os_log_wide(void *buf, const char *data, wchar_t *str) {
482   volatile int len;
483 
484   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
485   // CHECK: %[[DATA_ADDR:.*]] = alloca i8*, align 8
486   // CHECK: %[[STR_ADDR:.*]] = alloca i32*, align 8
487   // CHECK: %[[LEN:.*]] = alloca i32, align 4
488   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
489   // CHECK: store i8* %[[DATA]], i8** %[[DATA_ADDR]], align 8
490   // CHECK: store i32* %[[STR]], i32** %[[STR_ADDR]], align 8
491 
492   // CHECK: store volatile i32 12, i32* %[[LEN]], align 4
493   len = __builtin_os_log_format_buffer_size("%S", str);
494 
495   // CHECK: %[[V1:.*]] = load i8*, i8** %[[BUF_ADDR]], align 8
496   // CHECK: %[[V2:.*]] = load i32*, i32** %[[STR_ADDR]], align 8
497   // CHECK: %[[V3:.*]] = ptrtoint i32* %[[V2]] to i64
498   // CHECK: call void @__os_log_helper_1_2_1_8_80(i8* %[[V1]], i64 %[[V3]])
499 
500   __builtin_os_log_format(buf, "%S", str);
501 }
502 
503 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_2_1_8_80
504 // CHECK: (i8* %[[BUFFER:.*]], i64 %[[ARG0:.*]])
505 
506 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
507 // CHECK: %[[ARG0_ADDR:.*]] = alloca i64, align 8
508 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
509 // CHECK: store i64 %[[ARG0]], i64* %[[ARG0_ADDR]], align 8
510 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
511 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
512 // CHECK: store i8 2, i8* %[[SUMMARY]], align 1
513 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
514 // CHECK: store i8 1, i8* %[[NUMARGS]], align 1
515 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
516 // CHECK: store i8 80, i8* %[[ARGDESCRIPTOR]], align 1
517 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
518 // CHECK: store i8 8, i8* %[[ARGSIZE]], align 1
519 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
520 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i64*
521 // CHECK: %[[V0:.*]] = load i64, i64* %[[ARG0_ADDR]], align 8
522 // CHECK: store i64 %[[V0]], i64* %[[ARGDATACAST]], align 1
523 
524 // CHECK-LABEL: define void @test_builtin_os_log_precision_width
525 // CHECK: (i8* %[[BUF:.*]], i8* %[[DATA:.*]], i32 %[[PRECISION:.*]], i32 %[[WIDTH:.*]])
526 void test_builtin_os_log_precision_width(void *buf, const char *data,
527                                          int precision, int width) {
528   volatile int len;
529   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
530   // CHECK: %[[DATA_ADDR:.*]] = alloca i8*, align 8
531   // CHECK: %[[PRECISION_ADDR:.*]] = alloca i32, align 4
532   // CHECK: %[[WIDTH_ADDR:.*]] = alloca i32, align 4
533   // CHECK: %[[LEN:.*]] = alloca i32, align 4
534   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
535   // CHECK: store i8* %[[DATA]], i8** %[[DATA_ADDR]], align 8
536   // CHECK: store i32 %[[PRECISION]], i32* %[[PRECISION_ADDR]], align 4
537   // CHECK: store i32 %[[WIDTH]], i32* %[[WIDTH_ADDR]], align 4
538 
539   // CHECK: store volatile i32 24, i32* %[[LEN]], align 4
540   len = __builtin_os_log_format_buffer_size("Hello %*.*s World", precision, width, data);
541 
542   // CHECK: %[[V1:.*]] = load i8*, i8** %[[BUF_ADDR]], align 8
543   // CHECK: %[[V2:.*]] = load i32, i32* %[[PRECISION_ADDR]], align 4
544   // CHECK: %[[V3:.*]] = load i32, i32* %[[WIDTH_ADDR]], align 4
545   // CHECK: %[[V4:.*]] = load i8*, i8** %[[DATA_ADDR]], align 8
546   // CHECK: %[[V5:.*]] = ptrtoint i8* %[[V4]] to i64
547   // CHECK: call void @__os_log_helper_1_2_3_4_0_4_16_8_32(i8* %[[V1]], i32 %[[V2]], i32 %[[V3]], i64 %[[V5]])
548   __builtin_os_log_format(buf, "Hello %*.*s World", precision, width, data);
549 }
550 
551 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_2_3_4_0_4_16_8_32
552 // CHECK: (i8* %[[BUFFER:.*]], i32 %[[ARG0:.*]], i32 %[[ARG1:.*]], i64 %[[ARG2:.*]])
553 
554 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
555 // CHECK: %[[ARG0_ADDR:.*]] = alloca i32, align 4
556 // CHECK: %[[ARG1_ADDR:.*]] = alloca i32, align 4
557 // CHECK: %[[ARG2_ADDR:.*]] = alloca i64, align 8
558 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
559 // CHECK: store i32 %[[ARG0]], i32* %[[ARG0_ADDR]], align 4
560 // CHECK: store i32 %[[ARG1]], i32* %[[ARG1_ADDR]], align 4
561 // CHECK: store i64 %[[ARG2]], i64* %[[ARG2_ADDR]], align 8
562 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
563 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
564 // CHECK: store i8 2, i8* %[[SUMMARY]], align 1
565 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
566 // CHECK: store i8 3, i8* %[[NUMARGS]], align 1
567 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
568 // CHECK: store i8 0, i8* %[[ARGDESCRIPTOR]], align 1
569 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
570 // CHECK: store i8 4, i8* %[[ARGSIZE]], align 1
571 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
572 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i32*
573 // CHECK: %[[V0:.*]] = load i32, i32* %[[ARG0_ADDR]], align 4
574 // CHECK: store i32 %[[V0]], i32* %[[ARGDATACAST]], align 1
575 // CHECK: %[[ARGDESCRIPTOR1:.*]] = getelementptr i8, i8* %[[BUF]], i64 8
576 // CHECK: store i8 16, i8* %[[ARGDESCRIPTOR1]], align 1
577 // CHECK: %[[ARGSIZE2:.*]] = getelementptr i8, i8* %[[BUF]], i64 9
578 // CHECK: store i8 4, i8* %[[ARGSIZE2]], align 1
579 // CHECK: %[[ARGDATA3:.*]] = getelementptr i8, i8* %[[BUF]], i64 10
580 // CHECK: %[[ARGDATACAST4:.*]] = bitcast i8* %[[ARGDATA3]] to i32*
581 // CHECK: %[[V1:.*]] = load i32, i32* %[[ARG1_ADDR]], align 4
582 // CHECK: store i32 %[[V1]], i32* %[[ARGDATACAST4]], align 1
583 // CHECK: %[[ARGDESCRIPTOR5:.*]] = getelementptr i8, i8* %[[BUF]], i64 14
584 // CHECK: store i8 32, i8* %[[ARGDESCRIPTOR5]], align 1
585 // CHECK: %[[ARGSIZE6:.*]] = getelementptr i8, i8* %[[BUF]], i64 15
586 // CHECK: store i8 8, i8* %[[ARGSIZE6]], align 1
587 // CHECK: %[[ARGDATA7:.*]] = getelementptr i8, i8* %[[BUF]], i64 16
588 // CHECK: %[[ARGDATACAST8:.*]] = bitcast i8* %[[ARGDATA7]] to i64*
589 // CHECK: %[[V2:.*]] = load i64, i64* %[[ARG2_ADDR]], align 8
590 // CHECK: store i64 %[[V2]], i64* %[[ARGDATACAST8]], align 1
591 
592 // CHECK-LABEL: define void @test_builtin_os_log_invalid
593 // CHECK: (i8* %[[BUF:.*]], i32 %[[DATA:.*]])
594 void test_builtin_os_log_invalid(void *buf, int data) {
595   volatile int len;
596   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
597   // CHECK: %[[DATA_ADDR:.*]] = alloca i32, align 4
598   // CHECK: %[[LEN:.*]] = alloca i32, align 4
599   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
600   // CHECK: store i32 %[[DATA]], i32* %[[DATA_ADDR]], align 4
601 
602   // CHECK: store volatile i32 8, i32* %[[LEN]], align 4
603   len = __builtin_os_log_format_buffer_size("invalid specifier %: %d even a trailing one%", data);
604 
605   // CHECK: %[[V1:.*]] = load i8*, i8** %[[BUF_ADDR]], align 8
606   // CHECK: %[[V2:.*]] = load i32, i32* %[[DATA_ADDR]], align 4
607   // CHECK: call void @__os_log_helper_1_0_1_4_0(i8* %[[V1]], i32 %[[V2]])
608 
609   __builtin_os_log_format(buf, "invalid specifier %: %d even a trailing one%", data);
610 }
611 
612 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_0_1_4_0
613 // CHECK: (i8* %[[BUFFER:.*]], i32 %[[ARG0:.*]])
614 
615 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
616 // CHECK: %[[ARG0_ADDR:.*]] = alloca i32, align 4
617 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
618 // CHECK: store i32 %[[ARG0]], i32* %[[ARG0_ADDR]], align 4
619 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
620 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
621 // CHECK: store i8 0, i8* %[[SUMMARY]], align 1
622 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
623 // CHECK: store i8 1, i8* %[[NUMARGS]], align 1
624 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
625 // CHECK: store i8 0, i8* %[[ARGDESCRIPTOR]], align 1
626 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
627 // CHECK: store i8 4, i8* %[[ARGSIZE]], align 1
628 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
629 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i32*
630 // CHECK: %[[V0:.*]] = load i32, i32* %[[ARG0_ADDR]], align 4
631 // CHECK: store i32 %[[V0]], i32* %[[ARGDATACAST]], align 1
632 
633 // CHECK-LABEL: define void @test_builtin_os_log_percent
634 // CHECK: (i8* %[[BUF:.*]], i8* %[[DATA1:.*]], i8* %[[DATA2:.*]])
635 // Check that the %% which does not consume any argument is correctly handled
636 void test_builtin_os_log_percent(void *buf, const char *data1, const char *data2) {
637   volatile int len;
638   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
639   // CHECK: %[[DATA1_ADDR:.*]] = alloca i8*, align 8
640   // CHECK: %[[DATA2_ADDR:.*]] = alloca i8*, align 8
641   // CHECK: %[[LEN:.*]] = alloca i32, align 4
642   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
643   // CHECK: store i8* %[[DATA1]], i8** %[[DATA1_ADDR]], align 8
644   // CHECK: store i8* %[[DATA2]], i8** %[[DATA2_ADDR]], align 8
645   // CHECK: store volatile i32 22, i32* %[[LEN]], align 4
646 
647   len = __builtin_os_log_format_buffer_size("%s %% %s", data1, data2);
648 
649   // CHECK: %[[V1:.*]] = load i8*, i8** %[[BUF_ADDR]], align 8
650   // CHECK: %[[V2:.*]] = load i8*, i8** %[[DATA1_ADDR]], align 8
651   // CHECK: %[[V3:.*]] = ptrtoint i8* %[[V2]] to i64
652   // CHECK: %[[V4:.*]] = load i8*, i8** %[[DATA2_ADDR]], align 8
653   // CHECK: %[[V5:.*]] = ptrtoint i8* %[[V4]] to i64
654   // CHECK: call void @__os_log_helper_1_2_2_8_32_8_32(i8* %[[V1]], i64 %[[V3]], i64 %[[V5]])
655 
656   __builtin_os_log_format(buf, "%s %% %s", data1, data2);
657 }
658 
659 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_2_2_8_32_8_32
660 // CHECK: (i8* %[[BUFFER:.*]], i64 %[[ARG0:.*]], i64 %[[ARG1:.*]])
661 
662 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
663 // CHECK: %[[ARG0_ADDR:.*]] = alloca i64, align 8
664 // CHECK: %[[ARG1_ADDR:.*]] = alloca i64, align 8
665 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
666 // CHECK: store i64 %[[ARG0]], i64* %[[ARG0_ADDR]], align 8
667 // CHECK: store i64 %[[ARG1]], i64* %[[ARG1_ADDR]], align 8
668 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
669 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
670 // CHECK: store i8 2, i8* %[[SUMMARY]], align 1
671 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
672 // CHECK: store i8 2, i8* %[[NUMARGS]], align 1
673 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
674 // CHECK: store i8 32, i8* %[[ARGDESCRIPTOR]], align 1
675 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
676 // CHECK: store i8 8, i8* %[[ARGSIZE]], align 1
677 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
678 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i64*
679 // CHECK: %[[V0:.*]] = load i64, i64* %[[ARG0_ADDR]], align 8
680 // CHECK: store i64 %[[V0]], i64* %[[ARGDATACAST]], align 1
681 // CHECK: %[[ARGDESCRIPTOR1:.*]] = getelementptr i8, i8* %[[BUF]], i64 12
682 // CHECK: store i8 32, i8* %[[ARGDESCRIPTOR1]], align 1
683 // CHECK: %[[ARGSIZE2:.*]] = getelementptr i8, i8* %[[BUF]], i64 13
684 // CHECK: store i8 8, i8* %[[ARGSIZE2]], align 1
685 // CHECK: %[[ARGDATA3:.*]] = getelementptr i8, i8* %[[BUF]], i64 14
686 // CHECK: %[[ARGDATACAST4:.*]] = bitcast i8* %[[ARGDATA3]] to i64*
687 // CHECK: %[[V1:.*]] = load i64, i64* %[[ARG1_ADDR]], align 8
688 // CHECK: store i64 %[[V1]], i64* %[[ARGDATACAST4]], align 1
689 
690 // Check that the following two functions call the same helper function.
691 
692 // CHECK-LABEL: define void @test_builtin_os_log_merge_helper0
693 // CHECK: call void @__os_log_helper_1_0_2_4_0_8_0(
694 void test_builtin_os_log_merge_helper0(void *buf, int i, double d) {
695   __builtin_os_log_format(buf, "%d %f", i, d);
696 }
697 
698 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_0_2_4_0_8_0(
699 
700 // CHECK-LABEL: define void @test_builtin_os_log_merge_helper1
701 // CHECK: call void @__os_log_helper_1_0_2_4_0_8_0(
702 void test_builtin_os_log_merge_helper1(void *buf, unsigned u, long long ll) {
703   __builtin_os_log_format(buf, "%u %lld", u, ll);
704 }
705 
706 // Check that this function doesn't write past the end of array 'buf'.
707 
708 // CHECK-LABEL: define void @test_builtin_os_log_errno
709 void test_builtin_os_log_errno() {
710   // CHECK: %[[VLA:.*]] = alloca i8, i64 4, align 16
711   // CHECK: call void @__os_log_helper_16_2_1_0_96(i8* %[[VLA]])
712 
713   char buf[__builtin_os_log_format_buffer_size("%m")];
714   __builtin_os_log_format(buf, "%m");
715 }
716 
717 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_16_2_1_0_96
718 // CHECK: (i8* %[[BUFFER:.*]])
719 
720 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
721 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
722 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
723 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
724 // CHECK: store i8 2, i8* %[[SUMMARY]], align 16
725 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
726 // CHECK: store i8 1, i8* %[[NUMARGS]], align 1
727 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
728 // CHECK: store i8 96, i8* %[[ARGDESCRIPTOR]], align 2
729 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
730 // CHECK: store i8 0, i8* %[[ARGSIZE]], align 1
731 // CHECK-NEXT: ret void
732 
733 // CHECK-LABEL: define void @test_builtin_os_log_long_double
734 // CHECK: (i8* %[[BUF:.*]], x86_fp80 %[[LD:.*]])
735 void test_builtin_os_log_long_double(void *buf, long double ld) {
736   // CHECK: %[[BUF_ADDR:.*]] = alloca i8*, align 8
737   // CHECK: %[[LD_ADDR:.*]] = alloca x86_fp80, align 16
738   // CHECK: %[[COERCE:.*]] = alloca i128, align 16
739   // CHECK: store i8* %[[BUF]], i8** %[[BUF_ADDR]], align 8
740   // CHECK: store x86_fp80 %[[LD]], x86_fp80* %[[LD_ADDR]], align 16
741   // CHECK: %[[V0:.*]] = load i8*, i8** %[[BUF_ADDR]], align 8
742   // CHECK: %[[V1:.*]] = load x86_fp80, x86_fp80* %[[LD_ADDR]], align 16
743   // CHECK: %[[V2:.*]] = bitcast x86_fp80 %[[V1]] to i80
744   // CHECK: %[[V3:.*]] = zext i80 %[[V2]] to i128
745   // CHECK: store i128 %[[V3]], i128* %[[COERCE]], align 16
746   // CHECK: %[[V4:.*]] = bitcast i128* %[[COERCE]] to { i64, i64 }*
747   // CHECK: %[[V5:.*]] = getelementptr inbounds { i64, i64 }, { i64, i64 }* %[[V4]], i32 0, i32 0
748   // CHECK: %[[V6:.*]] = load i64, i64* %[[V5]], align 16
749   // CHECK: %[[V7:.*]] = getelementptr inbounds { i64, i64 }, { i64, i64 }* %[[V4]], i32 0, i32 1
750   // CHECK: %[[V8:.*]] = load i64, i64* %[[V7]], align 8
751   // CHECK: call void @__os_log_helper_1_0_1_16_0(i8* %[[V0]], i64 %[[V6]], i64 %[[V8]])
752 
753   __builtin_os_log_format(buf, "%Lf", ld);
754 }
755 
756 // CHECK-LABEL: define linkonce_odr hidden void @__os_log_helper_1_0_1_16_0
757 // CHECK: (i8* %[[BUFFER:.*]], i64 %[[ARG0_COERCE0:.*]], i64 %[[ARG0_COERCE1:.*]])
758 
759 // CHECK: %[[ARG0:.*]] = alloca i128, align 16
760 // CHECK: %[[BUFFER_ADDR:.*]] = alloca i8*, align 8
761 // CHECK: %[[ARG0_ADDR:.*]] = alloca i128, align 16
762 // CHECK: %[[V0:.*]] = bitcast i128* %[[ARG0]] to { i64, i64 }*
763 // CHECK: %[[V1:.*]] = getelementptr inbounds { i64, i64 }, { i64, i64 }* %[[V0]], i32 0, i32 0
764 // CHECK: store i64 %[[ARG0_COERCE0]], i64* %[[V1]], align 16
765 // CHECK: %[[V2:.*]] = getelementptr inbounds { i64, i64 }, { i64, i64 }* %[[V0]], i32 0, i32 1
766 // CHECK: store i64 %[[ARG0_COERCE1]], i64* %[[V2]], align 8
767 // CHECK: %[[ARG01:.*]] = load i128, i128* %[[ARG0]], align 16
768 // CHECK: store i8* %[[BUFFER]], i8** %[[BUFFER_ADDR]], align 8
769 // CHECK: store i128 %[[ARG01]], i128* %[[ARG0_ADDR]], align 16
770 // CHECK: %[[BUF:.*]] = load i8*, i8** %[[BUFFER_ADDR]], align 8
771 // CHECK: %[[SUMMARY:.*]] = getelementptr i8, i8* %[[BUF]], i64 0
772 // CHECK: store i8 0, i8* %[[SUMMARY]], align 1
773 // CHECK: %[[NUMARGS:.*]] = getelementptr i8, i8* %[[BUF]], i64 1
774 // CHECK: store i8 1, i8* %[[NUMARGS]], align 1
775 // CHECK: %[[ARGDESCRIPTOR:.*]] = getelementptr i8, i8* %[[BUF]], i64 2
776 // CHECK: store i8 0, i8* %[[ARGDESCRIPTOR]], align 1
777 // CHECK: %[[ARGSIZE:.*]] = getelementptr i8, i8* %[[BUF]], i64 3
778 // CHECK: store i8 16, i8* %[[ARGSIZE]], align 1
779 // CHECK: %[[ARGDATA:.*]] = getelementptr i8, i8* %[[BUF]], i64 4
780 // CHECK: %[[ARGDATACAST:.*]] = bitcast i8* %[[ARGDATA]] to i128*
781 // CHECK: %[[V3:.*]] = load i128, i128* %[[ARG0_ADDR]], align 16
782 // CHECK: store i128 %[[V3]], i128* %[[ARGDATACAST]], align 1
783 
784 #endif
785