Commit 94ec2190f8d8c41d19b668511bf31fae32bcd095 | zig

author	Andrew Kelley <andrew@ziglang.org> 2017-10-23 21:43:18-04:00
committer	Andrew Kelley <andrew@ziglang.org> 2017-11-02 21:54:24-04:00
log	94ec2190f8d8c41d19b668511bf31fae32bcd095
tree	c5ab9b20fbaf8f017661f9a159082d1ecaf9f943
parent	abff1b688420eb30d98145d8bc48e7d08f259885

update to llvm master

30 files changed, 676 insertions(+), 553 deletions(-)

README.md+4-4

...	...	@@ -141,7 +141,7 @@ These libraries must be installed on your system, with the development files
141	141	available. The Zig compiler links against them. You have to use the same
142	142	compiler for these libraries as you do to compile Zig.
143	143
144		* LLVM, Clang, and LLD libraries == 5.x
	144	* LLVM, Clang, and LLD libraries == 6.x
145	145
146	146	### Debug / Development Build
147	147
...	...	@@ -163,11 +163,11 @@ make install
163	163	`ZIG_LIBC_LIB_DIR` and `ZIG_LIBC_STATIC_LIB_DIR` are unused.
164	164
165	165	```
166		brew install llvm@5
167		brew outdated llvm@5 \|\| brew upgrade llvm@5
	166	brew install llvm@6
	167	brew outdated llvm@6 \|\| brew upgrade llvm@6
168	168	mkdir build
169	169	cd build
170		cmake .. -DCMAKE_PREFIX_PATH=/usr/local/opt/llvm@5/ -DCMAKE_INSTALL_PREFIX=$(pwd)
	170	cmake .. -DCMAKE_PREFIX_PATH=/usr/local/opt/llvm@6/ -DCMAKE_INSTALL_PREFIX=$(pwd)
171	171	make install
172	172	./zig build --build-file ../build.zig test
173	173	```

c_headers/__clang_cuda_intrinsics.h+124

...	...	@@ -92,6 +92,130 @@ __MAKE_SHUFFLES(__shfl_xor, __nvvm_shfl_bfly_i32, __nvvm_shfl_bfly_f32, 0x1f);
92	92
93	93	#endif // !defined(__CUDA_ARCH__) \|\| __CUDA_ARCH__ >= 300
94	94
	95	#if CUDA_VERSION >= 9000
	96	#if (!defined(__CUDA_ARCH__) \|\| __CUDA_ARCH__ >= 300)
	97	// __shfl_sync_* variants available in CUDA-9
	98	#pragma push_macro("__MAKE_SYNC_SHUFFLES")
	99	#define __MAKE_SYNC_SHUFFLES(__FnName, __IntIntrinsic, __FloatIntrinsic, \
	100	__Mask) \
	101	inline __device__ int __FnName(unsigned int __mask, int __val, int __offset, \
	102	int __width = warpSize) { \
	103	return __IntIntrinsic(__mask, __val, __offset, \
	104	((warpSize - __width) << 8) \| (__Mask)); \
	105	} \
	106	inline __device__ float __FnName(unsigned int __mask, float __val, \
	107	int __offset, int __width = warpSize) { \
	108	return __FloatIntrinsic(__mask, __val, __offset, \
	109	((warpSize - __width) << 8) \| (__Mask)); \
	110	} \
	111	inline __device__ unsigned int __FnName(unsigned int __mask, \
	112	unsigned int __val, int __offset, \
	113	int __width = warpSize) { \
	114	return static_cast<unsigned int>( \
	115	::__FnName(__mask, static_cast<int>(__val), __offset, __width)); \
	116	} \
	117	inline __device__ long long __FnName(unsigned int __mask, long long __val, \
	118	int __offset, int __width = warpSize) { \
	119	struct __Bits { \
	120	int __a, __b; \
	121	}; \
	122	_Static_assert(sizeof(__val) == sizeof(__Bits)); \
	123	_Static_assert(sizeof(__Bits) == 2 * sizeof(int)); \
	124	__Bits __tmp; \
	125	memcpy(&__val, &__tmp, sizeof(__val)); \
	126	__tmp.__a = ::__FnName(__mask, __tmp.__a, __offset, __width); \
	127	__tmp.__b = ::__FnName(__mask, __tmp.__b, __offset, __width); \
	128	long long __ret; \
	129	memcpy(&__ret, &__tmp, sizeof(__tmp)); \
	130	return __ret; \
	131	} \
	132	inline __device__ unsigned long long __FnName( \
	133	unsigned int __mask, unsigned long long __val, int __offset, \
	134	int __width = warpSize) { \
	135	return static_cast<unsigned long long>(::__FnName( \
	136	__mask, static_cast<unsigned long long>(__val), __offset, __width)); \
	137	} \
	138	inline __device__ double __FnName(unsigned int __mask, double __val, \
	139	int __offset, int __width = warpSize) { \
	140	long long __tmp; \
	141	_Static_assert(sizeof(__tmp) == sizeof(__val)); \
	142	memcpy(&__tmp, &__val, sizeof(__val)); \
	143	__tmp = ::__FnName(__mask, __tmp, __offset, __width); \
	144	double __ret; \
	145	memcpy(&__ret, &__tmp, sizeof(__ret)); \
	146	return __ret; \
	147	}
	148	__MAKE_SYNC_SHUFFLES(__shfl_sync, __nvvm_shfl_sync_idx_i32,
	149	__nvvm_shfl_sync_idx_f32, 0x1f);
	150	// We use 0 rather than 31 as our mask, because shfl.up applies to lanes >=
	151	// maxLane.
	152	__MAKE_SYNC_SHUFFLES(__shfl_up_sync, __nvvm_shfl_sync_up_i32,
	153	__nvvm_shfl_sync_up_f32, 0);
	154	__MAKE_SYNC_SHUFFLES(__shfl_down_sync, __nvvm_shfl_sync_down_i32,
	155	__nvvm_shfl_sync_down_f32, 0x1f);
	156	__MAKE_SYNC_SHUFFLES(__shfl_xor_sync, __nvvm_shfl_sync_bfly_i32,
	157	__nvvm_shfl_sync_bfly_f32, 0x1f);
	158	#pragma pop_macro("__MAKE_SYNC_SHUFFLES")
	159
	160	inline __device__ void __syncwarp(unsigned int mask = 0xffffffff) {
	161	return __nvvm_bar_warp_sync(mask);
	162	}
	163
	164	inline __device__ void __barrier_sync(unsigned int id) {
	165	__nvvm_barrier_sync(id);
	166	}
	167
	168	inline __device__ void __barrier_sync_count(unsigned int id,
	169	unsigned int count) {
	170	__nvvm_barrier_sync_cnt(id, count);
	171	}
	172
	173	inline __device__ int __all_sync(unsigned int mask, int pred) {
	174	return __nvvm_vote_all_sync(mask, pred);
	175	}
	176
	177	inline __device__ int __any_sync(unsigned int mask, int pred) {
	178	return __nvvm_vote_any_sync(mask, pred);
	179	}
	180
	181	inline __device__ int __uni_sync(unsigned int mask, int pred) {
	182	return __nvvm_vote_uni_sync(mask, pred);
	183	}
	184
	185	inline __device__ unsigned int __ballot_sync(unsigned int mask, int pred) {
	186	return __nvvm_vote_ballot_sync(mask, pred);
	187	}
	188
	189	inline __device__ unsigned int __activemask() { return __nvvm_vote_ballot(1); }
	190
	191	#endif // !defined(__CUDA_ARCH__) \|\| __CUDA_ARCH__ >= 300
	192
	193	// Define __match* builtins CUDA-9 headers expect to see.
	194	#if !defined(__CUDA_ARCH__) \|\| __CUDA_ARCH__ >= 700
	195	inline __device__ unsigned int __match32_any_sync(unsigned int mask,
	196	unsigned int value) {
	197	return __nvvm_match_any_sync_i32(mask, value);
	198	}
	199
	200	inline __device__ unsigned long long
	201	__match64_any_sync(unsigned int mask, unsigned long long value) {
	202	return __nvvm_match_any_sync_i64(mask, value);
	203	}
	204
	205	inline __device__ unsigned int
	206	__match32_all_sync(unsigned int mask, unsigned int value, int *pred) {
	207	return __nvvm_match_all_sync_i32p(mask, value, pred);
	208	}
	209
	210	inline __device__ unsigned long long
	211	__match64_all_sync(unsigned int mask, unsigned long long value, int *pred) {
	212	return __nvvm_match_all_sync_i64p(mask, value, pred);
	213	}
	214	#include "crt/sm_70_rt.hpp"
	215
	216	#endif // !defined(__CUDA_ARCH__) \|\| __CUDA_ARCH__ >= 700
	217	#endif // __CUDA_VERSION >= 9000
	218
95	219	// sm_32 intrinsics: __ldg and __funnelshift_{l,lc,r,rc}.
96	220
97	221	// Prevent the vanilla sm_32 intrinsics header from being included.

c_headers/__clang_cuda_runtime_wrapper.h+29-1

...	...	@@ -62,7 +62,7 @@
62	62	#include "cuda.h"
63	63	#if !defined(CUDA_VERSION)
64	64	#error "cuda.h did not define CUDA_VERSION"
65		#elif CUDA_VERSION < 7000 \|\| CUDA_VERSION > 8000
	65	#elif CUDA_VERSION < 7000 \|\| CUDA_VERSION > 9000
66	66	#error "Unsupported CUDA version!"
67	67	#endif
68	68
...	...	@@ -86,7 +86,11 @@
86	86	#define __COMMON_FUNCTIONS_H__
87	87
88	88	#undef __CUDACC__
	89	#if CUDA_VERSION < 9000
89	90	#define __CUDABE__
	91	#else
	92	#define __CUDA_LIBDEVICE__
	93	#endif
90	94	// Disables definitions of device-side runtime support stubs in
91	95	// cuda_device_runtime_api.h
92	96	#include "driver_types.h"
...	...	@@ -94,6 +98,7 @@
94	98	#include "host_defines.h"
95	99
96	100	#undef __CUDABE__
	101	#undef __CUDA_LIBDEVICE__
97	102	#define __CUDACC__
98	103	#include "cuda_runtime.h"
99	104
...	...	@@ -105,7 +110,9 @@
105	110	#define __nvvm_memcpy(s, d, n, a) __builtin_memcpy(s, d, n)
106	111	#define __nvvm_memset(d, c, n, a) __builtin_memset(d, c, n)
107	112
	113	#if CUDA_VERSION < 9000
108	114	#include "crt/device_runtime.h"
	115	#endif
109	116	#include "crt/host_runtime.h"
110	117	// device_runtime.h defines __cxa_* macros that will conflict with
111	118	// cxxabi.h.
...	...	@@ -166,7 +173,18 @@ inline __host__ double __signbitd(double x) {
166	173	// __device__.
167	174	#pragma push_macro("__forceinline__")
168	175	#define __forceinline__ __device__ __inline__ __attribute__((always_inline))
	176
	177	#pragma push_macro("__float2half_rn")
	178	#if CUDA_VERSION >= 9000
	179	// CUDA-9 has conflicting prototypes for __float2half_rn(float f) in
	180	// cuda_fp16.h[pp] and device_functions.hpp. We need to get the one in
	181	// device_functions.hpp out of the way.
	182	#define __float2half_rn __float2half_rn_disabled
	183	#endif
	184
169	185	#include "device_functions.hpp"
	186	#pragma pop_macro("__float2half_rn")
	187
170	188
171	189	// math_function.hpp uses the __USE_FAST_MATH__ macro to determine whether we
172	190	// get the slow-but-accurate or fast-but-inaccurate versions of functions like
...	...	@@ -247,7 +265,17 @@ static inline __device__ void __brkpt(int __c) { __brkpt(); }
247	265	#pragma push_macro("__GNUC__")
248	266	#undef __GNUC__
249	267	#define signbit __ignored_cuda_signbit
	268
	269	// CUDA-9 omits device-side definitions of some math functions if it sees
	270	// include guard from math.h wrapper from libstdc++. We have to undo the header
	271	// guard temporarily to get the definitions we need.
	272	#pragma push_macro("_GLIBCXX_MATH_H")
	273	#if CUDA_VERSION >= 9000
	274	#undef _GLIBCXX_MATH_H
	275	#endif
	276
250	277	#include "math_functions.hpp"
	278	#pragma pop_macro("_GLIBCXX_MATH_H")
251	279	#pragma pop_macro("__GNUC__")
252	280	#pragma pop_macro("signbit")
253	281

c_headers/arm64intr.h created+49

...	...	@@ -0,0 +1,49 @@
	1	/*===---- arm64intr.h - ARM64 Windows intrinsics -------------------------------===
	2	*
	3	* Permission is hereby granted, free of charge, to any person obtaining a copy
	4	* of this software and associated documentation files (the "Software"), to deal
	5	* in the Software without restriction, including without limitation the rights
	6	* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
	7	* copies of the Software, and to permit persons to whom the Software is
	8	* furnished to do so, subject to the following conditions:
	9	*
	10	* The above copyright notice and this permission notice shall be included in
	11	* all copies or substantial portions of the Software.
	12	*
	13	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	14	* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	15	* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
	16	* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
	17	* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
	18	* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
	19	* THE SOFTWARE.
	20	*
	21	*===-----------------------------------------------------------------------===
	22	*/
	23
	24	/* Only include this if we're compiling for the windows platform. */
	25	#ifndef _MSC_VER
	26	#include_next <arm64intr.h>
	27	#else
	28
	29	#ifndef __ARM64INTR_H
	30	#define __ARM64INTR_H
	31
	32	typedef enum
	33	{
	34	_ARM64_BARRIER_SY = 0xF,
	35	_ARM64_BARRIER_ST = 0xE,
	36	_ARM64_BARRIER_LD = 0xD,
	37	_ARM64_BARRIER_ISH = 0xB,
	38	_ARM64_BARRIER_ISHST = 0xA,
	39	_ARM64_BARRIER_ISHLD = 0x9,
	40	_ARM64_BARRIER_NSH = 0x7,
	41	_ARM64_BARRIER_NSHST = 0x6,
	42	_ARM64_BARRIER_NSHLD = 0x5,
	43	_ARM64_BARRIER_OSH = 0x3,
	44	_ARM64_BARRIER_OSHST = 0x2,
	45	_ARM64_BARRIER_OSHLD = 0x1
	46	} _ARM64INTR_BARRIER_TYPE;
	47
	48	#endif /* __ARM64INTR_H */
	49	#endif /* _MSC_VER */

c_headers/avx2intrin.h+10-2

...	...	@@ -145,13 +145,21 @@ _mm256_andnot_si256(__m256i __a, __m256i __b)
145	145	static __inline__ __m256i __DEFAULT_FN_ATTRS
146	146	_mm256_avg_epu8(__m256i __a, __m256i __b)
147	147	{
148		return (__m256i)__builtin_ia32_pavgb256((__v32qi)__a, (__v32qi)__b);
	148	typedef unsigned short __v32hu __attribute__((__vector_size__(64)));
	149	return (__m256i)__builtin_convertvector(
	150	((__builtin_convertvector((__v32qu)__a, __v32hu) +
	151	__builtin_convertvector((__v32qu)__b, __v32hu)) + 1)
	152	>> 1, __v32qu);
149	153	}
150	154
151	155	static __inline__ __m256i __DEFAULT_FN_ATTRS
152	156	_mm256_avg_epu16(__m256i __a, __m256i __b)
153	157	{
154		return (__m256i)__builtin_ia32_pavgw256((__v16hi)__a, (__v16hi)__b);
	158	typedef unsigned int __v16su __attribute__((__vector_size__(64)));
	159	return (__m256i)__builtin_convertvector(
	160	((__builtin_convertvector((__v16hu)__a, __v16su) +
	161	__builtin_convertvector((__v16hu)__b, __v16su)) + 1)
	162	>> 1, __v16hu);
155	163	}
156	164
157	165	static __inline__ __m256i __DEFAULT_FN_ATTRS

c_headers/avx512bwintrin.h+34-37

...	...	@@ -706,57 +706,55 @@ _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
706	706	static __inline__ __m512i __DEFAULT_FN_ATTRS
707	707	_mm512_avg_epu8 (__m512i __A, __m512i __B)
708	708	{
709		return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
710		(__v64qi) __B,
711		(__v64qi) _mm512_setzero_qi(),
712		(__mmask64) -1);
	709	typedef unsigned short __v64hu __attribute__((__vector_size__(128)));
	710	return (__m512i)__builtin_convertvector(
	711	((__builtin_convertvector((__v64qu) __A, __v64hu) +
	712	__builtin_convertvector((__v64qu) __B, __v64hu)) + 1)
	713	>> 1, __v64qu);
713	714	}
714	715
715	716	static __inline__ __m512i __DEFAULT_FN_ATTRS
716	717	_mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A,
717	718	__m512i __B)
718	719	{
719		return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
720		(__v64qi) __B,
721		(__v64qi) __W,
722		(__mmask64) __U);
	720	return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
	721	(__v64qi)_mm512_avg_epu8(__A, __B),
	722	(__v64qi)__W);
723	723	}
724	724
725	725	static __inline__ __m512i __DEFAULT_FN_ATTRS
726	726	_mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B)
727	727	{
728		return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
729		(__v64qi) __B,
730		(__v64qi) _mm512_setzero_qi(),
731		(__mmask64) __U);
	728	return (__m512i)__builtin_ia32_selectb_512((__mmask64)__U,
	729	(__v64qi)_mm512_avg_epu8(__A, __B),
	730	(__v64qi)_mm512_setzero_qi());
732	731	}
733	732
734	733	static __inline__ __m512i __DEFAULT_FN_ATTRS
735	734	_mm512_avg_epu16 (__m512i __A, __m512i __B)
736	735	{
737		return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
738		(__v32hi) __B,
739		(__v32hi) _mm512_setzero_hi(),
740		(__mmask32) -1);
	736	typedef unsigned int __v32su __attribute__((__vector_size__(128)));
	737	return (__m512i)__builtin_convertvector(
	738	((__builtin_convertvector((__v32hu) __A, __v32su) +
	739	__builtin_convertvector((__v32hu) __B, __v32su)) + 1)
	740	>> 1, __v32hu);
741	741	}
742	742
743	743	static __inline__ __m512i __DEFAULT_FN_ATTRS
744	744	_mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A,
745	745	__m512i __B)
746	746	{
747		return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
748		(__v32hi) __B,
749		(__v32hi) __W,
750		(__mmask32) __U);
	747	return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U,
	748	(__v32hi)_mm512_avg_epu16(__A, __B),
	749	(__v32hi)__W);
751	750	}
752	751
753	752	static __inline__ __m512i __DEFAULT_FN_ATTRS
754	753	_mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
755	754	{
756		return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
757		(__v32hi) __B,
758		(__v32hi) _mm512_setzero_hi(),
759		(__mmask32) __U);
	755	return (__m512i)__builtin_ia32_selectw_512((__mmask32)__U,
	756	(__v32hi)_mm512_avg_epu16(__A, __B),
	757	(__v32hi) _mm512_setzero_hi());
760	758	}
761	759
762	760	static __inline__ __m512i __DEFAULT_FN_ATTRS
...	...	@@ -2028,18 +2026,17 @@ _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A)
2028	2026	static __inline__ __m512i __DEFAULT_FN_ATTRS
2029	2027	_mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A)
2030	2028	{
2031		return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A,
2032		(__v64qi) __O,
2033		__M);
	2029	return (__m512i) __builtin_ia32_selectb_512(__M,
	2030	(__v64qi)_mm512_set1_epi8(__A),
	2031	(__v64qi) __O);
2034	2032	}
2035	2033
2036	2034	static __inline__ __m512i __DEFAULT_FN_ATTRS
2037	2035	_mm512_maskz_set1_epi8 (__mmask64 __M, char __A)
2038	2036	{
2039		return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A,
2040		(__v64qi)
2041		_mm512_setzero_qi(),
2042		__M);
	2037	return (__m512i) __builtin_ia32_selectb_512(__M,
	2038	(__v64qi) _mm512_set1_epi8(__A),
	2039	(__v64qi) _mm512_setzero_si512());
2043	2040	}
2044	2041
2045	2042	static __inline__ __mmask64 __DEFAULT_FN_ATTRS
...	...	@@ -2219,17 +2216,17 @@ _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A)
2219	2216	static __inline__ __m512i __DEFAULT_FN_ATTRS
2220	2217	_mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A)
2221	2218	{
2222		return (__m512i) __builtin_ia32_pbroadcastw512_gpr_mask (__A,
2223		(__v32hi) __O,
2224		__M);
	2219	return (__m512i) __builtin_ia32_selectw_512(__M,
	2220	(__v32hi) _mm512_set1_epi16(__A),
	2221	(__v32hi) __O);
2225	2222	}
2226	2223
2227	2224	static __inline__ __m512i __DEFAULT_FN_ATTRS
2228	2225	_mm512_maskz_set1_epi16 (__mmask32 __M, short __A)
2229	2226	{
2230		return (__m512i) __builtin_ia32_pbroadcastw512_gpr_mask (__A,
2231		(__v32hi) _mm512_setzero_hi(),
2232		__M);
	2227	return (__m512i) __builtin_ia32_selectw_512(__M,
	2228	(__v32hi) _mm512_set1_epi16(__A),
	2229	(__v32hi) _mm512_setzero_si512());
2233	2230	}
2234	2231
2235	2232	static __inline__ __m512i __DEFAULT_FN_ATTRS

c_headers/avx512dqintrin.h+20-18

...	...	@@ -973,25 +973,26 @@ _mm512_movepi64_mask (__m512i __A)
973	973	static __inline__ __m512 __DEFAULT_FN_ATTRS
974	974	_mm512_broadcast_f32x2 (__m128 __A)
975	975	{
976		return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A,
977		(__v16sf)_mm512_undefined_ps(),
978		(__mmask16) -1);
	976	return (__m512)__builtin_shufflevector((__v4sf)__A,
	977	(__v4sf)_mm_undefined_ps(),
	978	0, 1, 0, 1, 0, 1, 0, 1,
	979	0, 1, 0, 1, 0, 1, 0, 1);
979	980	}
980	981
981	982	static __inline__ __m512 __DEFAULT_FN_ATTRS
982	983	_mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A)
983	984	{
984		return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A,
985		(__v16sf)
986		__O, __M);
	985	return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
	986	(__v16sf)_mm512_broadcast_f32x2(__A),
	987	(__v16sf)__O);
987	988	}
988	989
989	990	static __inline__ __m512 __DEFAULT_FN_ATTRS
990	991	_mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A)
991	992	{
992		return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A,
993		(__v16sf)_mm512_setzero_ps (),
994		__M);
	993	return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
	994	(__v16sf)_mm512_broadcast_f32x2(__A),
	995	(__v16sf)_mm512_setzero_ps());
995	996	}
996	997
997	998	static __inline__ __m512 __DEFAULT_FN_ATTRS
...	...	@@ -1044,25 +1045,26 @@ _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A)
1044	1045	static __inline__ __m512i __DEFAULT_FN_ATTRS
1045	1046	_mm512_broadcast_i32x2 (__m128i __A)
1046	1047	{
1047		return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A,
1048		(__v16si)_mm512_setzero_si512(),
1049		(__mmask16) -1);
	1048	return (__m512i)__builtin_shufflevector((__v4si)__A,
	1049	(__v4si)_mm_undefined_si128(),
	1050	0, 1, 0, 1, 0, 1, 0, 1,
	1051	0, 1, 0, 1, 0, 1, 0, 1);
1050	1052	}
1051	1053
1052	1054	static __inline__ __m512i __DEFAULT_FN_ATTRS
1053	1055	_mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A)
1054	1056	{
1055		return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A,
1056		(__v16si)
1057		__O, __M);
	1057	return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
	1058	(__v16si)_mm512_broadcast_i32x2(__A),
	1059	(__v16si)__O);
1058	1060	}
1059	1061
1060	1062	static __inline__ __m512i __DEFAULT_FN_ATTRS
1061	1063	_mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A)
1062	1064	{
1063		return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A,
1064		(__v16si)_mm512_setzero_si512 (),
1065		__M);
	1065	return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
	1066	(__v16si)_mm512_broadcast_i32x2(__A),
	1067	(__v16si)_mm512_setzero_si512());
1066	1068	}
1067	1069
1068	1070	static __inline__ __m512i __DEFAULT_FN_ATTRS

c_headers/avx512fintrin.h+25-29

...	...	@@ -258,30 +258,6 @@ _mm512_maskz_broadcastq_epi64 (__mmask8 __M, __m128i __A)
258	258	(__v8di) _mm512_setzero_si512());
259	259	}
260	260
261		static __inline __m512i __DEFAULT_FN_ATTRS
262		_mm512_maskz_set1_epi32(__mmask16 __M, int __A)
263		{
264		return (__m512i) __builtin_ia32_pbroadcastd512_gpr_mask (__A,
265		(__v16si)
266		_mm512_setzero_si512 (),
267		__M);
268		}
269
270		static __inline __m512i __DEFAULT_FN_ATTRS
271		_mm512_maskz_set1_epi64(__mmask8 __M, long long __A)
272		{
273		#ifdef __x86_64__
274		return (__m512i) __builtin_ia32_pbroadcastq512_gpr_mask (__A,
275		(__v8di)
276		_mm512_setzero_si512 (),
277		__M);
278		#else
279		return (__m512i) __builtin_ia32_pbroadcastq512_mem_mask (__A,
280		(__v8di)
281		_mm512_setzero_si512 (),
282		__M);
283		#endif
284		}
285	261
286	262	static __inline __m512 __DEFAULT_FN_ATTRS
287	263	_mm512_setzero_ps(void)
...	...	@@ -340,12 +316,30 @@ _mm512_set1_epi32(int __s)
340	316	__s, __s, __s, __s, __s, __s, __s, __s };
341	317	}
342	318
	319	static __inline __m512i __DEFAULT_FN_ATTRS
	320	_mm512_maskz_set1_epi32(__mmask16 __M, int __A)
	321	{
	322	return (__m512i)__builtin_ia32_selectd_512(__M,
	323	(__v16si)_mm512_set1_epi32(__A),
	324	(__v16si)_mm512_setzero_si512());
	325	}
	326
343	327	static __inline __m512i __DEFAULT_FN_ATTRS
344	328	_mm512_set1_epi64(long long __d)
345	329	{
346	330	return (__m512i)(__v8di){ __d, __d, __d, __d, __d, __d, __d, __d };
347	331	}
348	332
	333	#ifdef __x86_64__
	334	static __inline __m512i __DEFAULT_FN_ATTRS
	335	_mm512_maskz_set1_epi64(__mmask8 __M, long long __A)
	336	{
	337	return (__m512i)__builtin_ia32_selectq_512(__M,
	338	(__v8di)_mm512_set1_epi64(__A),
	339	(__v8di)_mm512_setzero_si512());
	340	}
	341	#endif
	342
349	343	static __inline__ __m512 __DEFAULT_FN_ATTRS
350	344	_mm512_broadcastss_ps(__m128 __A)
351	345	{
...	...	@@ -9040,7 +9034,7 @@ _mm512_stream_si512 (__m512i * __P, __m512i __A)
9040	9034	}
9041	9035
9042	9036	static __inline__ __m512i __DEFAULT_FN_ATTRS
9043		_mm512_stream_load_si512 (void *__P)
	9037	_mm512_stream_load_si512 (void const *__P)
9044	9038	{
9045	9039	typedef __v8di __v8di_aligned __attribute__((aligned(64)));
9046	9040	return (__m512i) __builtin_nontemporal_load((const __v8di_aligned *)__P);
...	...	@@ -9742,16 +9736,18 @@ _mm_cvtu64_ss (__m128 __A, unsigned long long __B)
9742	9736	static __inline__ __m512i __DEFAULT_FN_ATTRS
9743	9737	_mm512_mask_set1_epi32 (__m512i __O, __mmask16 __M, int __A)
9744	9738	{
9745		return (__m512i) __builtin_ia32_pbroadcastd512_gpr_mask (__A, (__v16si) __O,
9746		__M);
	9739	return (__m512i) __builtin_ia32_selectd_512(__M,
	9740	(__v16si) _mm512_set1_epi32(__A),
	9741	(__v16si) __O);
9747	9742	}
9748	9743
9749	9744	#ifdef __x86_64__
9750	9745	static __inline__ __m512i __DEFAULT_FN_ATTRS
9751	9746	_mm512_mask_set1_epi64 (__m512i __O, __mmask8 __M, long long __A)
9752	9747	{
9753		return (__m512i) __builtin_ia32_pbroadcastq512_gpr_mask (__A, (__v8di) __O,
9754		__M);
	9748	return (__m512i) __builtin_ia32_selectq_512(__M,
	9749	(__v8di) _mm512_set1_epi64(__A),
	9750	(__v8di) __O);
9755	9751	}
9756	9752	#endif
9757	9753

c_headers/avx512vlbwintrin.h+24-26

...	...	@@ -2660,35 +2660,33 @@ _mm256_maskz_mov_epi8 (__mmask32 __U, __m256i __A)
2660	2660	static __inline__ __m128i __DEFAULT_FN_ATTRS
2661	2661	_mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A)
2662	2662	{
2663		return (__m128i) __builtin_ia32_pbroadcastb128_gpr_mask (__A,
2664		(__v16qi) __O,
2665		__M);
	2663	return (__m128i) __builtin_ia32_selectb_128(__M,
	2664	(__v16qi) _mm_set1_epi8(__A),
	2665	(__v16qi) __O);
2666	2666	}
2667	2667
2668	2668	static __inline__ __m128i __DEFAULT_FN_ATTRS
2669	2669	_mm_maskz_set1_epi8 (__mmask16 __M, char __A)
2670	2670	{
2671		return (__m128i) __builtin_ia32_pbroadcastb128_gpr_mask (__A,
2672		(__v16qi)
2673		_mm_setzero_si128 (),
2674		__M);
	2671	return (__m128i) __builtin_ia32_selectb_128(__M,
	2672	(__v16qi) _mm_set1_epi8(__A),
	2673	(__v16qi) _mm_setzero_si128());
2675	2674	}
2676	2675
2677	2676	static __inline__ __m256i __DEFAULT_FN_ATTRS
2678	2677	_mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A)
2679	2678	{
2680		return (__m256i) __builtin_ia32_pbroadcastb256_gpr_mask (__A,
2681		(__v32qi) __O,
2682		__M);
	2679	return (__m256i) __builtin_ia32_selectb_256(__M,
	2680	(__v32qi) _mm256_set1_epi8(__A),
	2681	(__v32qi) __O);
2683	2682	}
2684	2683
2685	2684	static __inline__ __m256i __DEFAULT_FN_ATTRS
2686	2685	_mm256_maskz_set1_epi8 (__mmask32 __M, char __A)
2687	2686	{
2688		return (__m256i) __builtin_ia32_pbroadcastb256_gpr_mask (__A,
2689		(__v32qi)
2690		_mm256_setzero_si256 (),
2691		__M);
	2687	return (__m256i) __builtin_ia32_selectb_256(__M,
	2688	(__v32qi) _mm256_set1_epi8(__A),
	2689	(__v32qi) _mm256_setzero_si256());
2692	2690	}
2693	2691
2694	2692	static __inline__ __m128i __DEFAULT_FN_ATTRS
...	...	@@ -3025,33 +3023,33 @@ _mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A)
3025	3023	static __inline__ __m256i __DEFAULT_FN_ATTRS
3026	3024	_mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A)
3027	3025	{
3028		return (__m256i) __builtin_ia32_pbroadcastw256_gpr_mask (__A,
3029		(__v16hi) __O,
3030		__M);
	3026	return (__m256i) __builtin_ia32_selectw_256 (__M,
	3027	(__v16hi) _mm256_set1_epi16(__A),
	3028	(__v16hi) __O);
3031	3029	}
3032	3030
3033	3031	static __inline__ __m256i __DEFAULT_FN_ATTRS
3034	3032	_mm256_maskz_set1_epi16 (__mmask16 __M, short __A)
3035	3033	{
3036		return (__m256i) __builtin_ia32_pbroadcastw256_gpr_mask (__A,
3037		(__v16hi) _mm256_setzero_si256 (),
3038		__M);
	3034	return (__m256i) __builtin_ia32_selectw_256(__M,
	3035	(__v16hi)_mm256_set1_epi16(__A),
	3036	(__v16hi) _mm256_setzero_si256());
3039	3037	}
3040	3038
3041	3039	static __inline__ __m128i __DEFAULT_FN_ATTRS
3042	3040	_mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A)
3043	3041	{
3044		return (__m128i) __builtin_ia32_pbroadcastw128_gpr_mask (__A,
3045		(__v8hi) __O,
3046		__M);
	3042	return (__m128i) __builtin_ia32_selectw_128(__M,
	3043	(__v8hi) _mm_set1_epi16(__A),
	3044	(__v8hi) __O);
3047	3045	}
3048	3046
3049	3047	static __inline__ __m128i __DEFAULT_FN_ATTRS
3050	3048	_mm_maskz_set1_epi16 (__mmask8 __M, short __A)
3051	3049	{
3052		return (__m128i) __builtin_ia32_pbroadcastw128_gpr_mask (__A,
3053		(__v8hi) _mm_setzero_si128 (),
3054		__M);
	3050	return (__m128i) __builtin_ia32_selectw_128(__M,
	3051	(__v8hi) _mm_set1_epi16(__A),
	3052	(__v8hi) _mm_setzero_si128());
3055	3053	}
3056	3054
3057	3055	static __inline__ __m128i __DEFAULT_FN_ATTRS

c_headers/avx512vldqintrin.h+27-27

...	...	@@ -978,25 +978,25 @@ _mm256_movepi64_mask (__m256i __A)
978	978	static __inline__ __m256 __DEFAULT_FN_ATTRS
979	979	_mm256_broadcast_f32x2 (__m128 __A)
980	980	{
981		return (__m256) __builtin_ia32_broadcastf32x2_256_mask ((__v4sf) __A,
982		(__v8sf)_mm256_undefined_ps(),
983		(__mmask8) -1);
	981	return (__m256)__builtin_shufflevector((__v4sf)__A,
	982	(__v4sf)_mm_undefined_ps(),
	983	0, 1, 0, 1, 0, 1, 0, 1);
984	984	}
985	985
986	986	static __inline__ __m256 __DEFAULT_FN_ATTRS
987	987	_mm256_mask_broadcast_f32x2 (__m256 __O, __mmask8 __M, __m128 __A)
988	988	{
989		return (__m256) __builtin_ia32_broadcastf32x2_256_mask ((__v4sf) __A,
990		(__v8sf) __O,
991		__M);
	989	return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
	990	(__v8sf)_mm256_broadcast_f32x2(__A),
	991	(__v8sf)__O);
992	992	}
993	993
994	994	static __inline__ __m256 __DEFAULT_FN_ATTRS
995	995	_mm256_maskz_broadcast_f32x2 (__mmask8 __M, __m128 __A)
996	996	{
997		return (__m256) __builtin_ia32_broadcastf32x2_256_mask ((__v4sf) __A,
998		(__v8sf) _mm256_setzero_ps (),
999		__M);
	997	return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
	998	(__v8sf)_mm256_broadcast_f32x2(__A),
	999	(__v8sf)_mm256_setzero_ps());
1000	1000	}
1001	1001
1002	1002	static __inline__ __m256d __DEFAULT_FN_ATTRS
...	...	@@ -1025,49 +1025,49 @@ _mm256_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A)
1025	1025	static __inline__ __m128i __DEFAULT_FN_ATTRS
1026	1026	_mm_broadcast_i32x2 (__m128i __A)
1027	1027	{
1028		return (__m128i) __builtin_ia32_broadcasti32x2_128_mask ((__v4si) __A,
1029		(__v4si)_mm_undefined_si128(),
1030		(__mmask8) -1);
	1028	return (__m128i)__builtin_shufflevector((__v4si)__A,
	1029	(__v4si)_mm_undefined_si128(),
	1030	0, 1, 0, 1);
1031	1031	}
1032	1032
1033	1033	static __inline__ __m128i __DEFAULT_FN_ATTRS
1034	1034	_mm_mask_broadcast_i32x2 (__m128i __O, __mmask8 __M, __m128i __A)
1035	1035	{
1036		return (__m128i) __builtin_ia32_broadcasti32x2_128_mask ((__v4si) __A,
1037		(__v4si) __O,
1038		__M);
	1036	return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
	1037	(__v4si)_mm_broadcast_i32x2(__A),
	1038	(__v4si)__O);
1039	1039	}
1040	1040
1041	1041	static __inline__ __m128i __DEFAULT_FN_ATTRS
1042	1042	_mm_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A)
1043	1043	{
1044		return (__m128i) __builtin_ia32_broadcasti32x2_128_mask ((__v4si) __A,
1045		(__v4si) _mm_setzero_si128 (),
1046		__M);
	1044	return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
	1045	(__v4si)_mm_broadcast_i32x2(__A),
	1046	(__v4si)_mm_setzero_si128());
1047	1047	}
1048	1048
1049	1049	static __inline__ __m256i __DEFAULT_FN_ATTRS
1050	1050	_mm256_broadcast_i32x2 (__m128i __A)
1051	1051	{
1052		return (__m256i) __builtin_ia32_broadcasti32x2_256_mask ((__v4si) __A,
1053		(__v8si)_mm256_undefined_si256(),
1054		(__mmask8) -1);
	1052	return (__m256i)__builtin_shufflevector((__v4si)__A,
	1053	(__v4si)_mm_undefined_si128(),
	1054	0, 1, 0, 1, 0, 1, 0, 1);
1055	1055	}
1056	1056
1057	1057	static __inline__ __m256i __DEFAULT_FN_ATTRS
1058	1058	_mm256_mask_broadcast_i32x2 (__m256i __O, __mmask8 __M, __m128i __A)
1059	1059	{
1060		return (__m256i) __builtin_ia32_broadcasti32x2_256_mask ((__v4si) __A,
1061		(__v8si) __O,
1062		__M);
	1060	return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
	1061	(__v8si)_mm256_broadcast_i32x2(__A),
	1062	(__v8si)__O);
1063	1063	}
1064	1064
1065	1065	static __inline__ __m256i __DEFAULT_FN_ATTRS
1066	1066	_mm256_maskz_broadcast_i32x2 (__mmask8 __M, __m128i __A)
1067	1067	{
1068		return (__m256i) __builtin_ia32_broadcasti32x2_256_mask ((__v4si) __A,
1069		(__v8si) _mm256_setzero_si256 (),
1070		__M);
	1068	return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
	1069	(__v8si)_mm256_broadcast_i32x2(__A),
	1070	(__v8si)_mm256_setzero_si256());
1071	1071	}
1072	1072
1073	1073	static __inline__ __m256i __DEFAULT_FN_ATTRS

c_headers/avx512vlintrin.h+41-28

...	...	@@ -5723,59 +5723,72 @@ _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A)
5723	5723	(__v4df)_mm256_setzero_pd());
5724	5724	}
5725	5725
	5726	static __inline__ __m128i __DEFAULT_FN_ATTRS
	5727	_mm_mask_set1_epi32(__m128i __O, __mmask8 __M, int __A)
	5728	{
	5729	return (__m128i)__builtin_ia32_selectd_128(__M,
	5730	(__v4si) _mm_set1_epi32(__A),
	5731	(__v4si)__O);
	5732	}
5726	5733
5727		#define _mm_mask_set1_epi32(O, M, A) __extension__ ({ \
5728		(__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \
5729		(__v4si)(__m128i)(O), \
5730		(__mmask8)(M)); })
	5734	static __inline__ __m128i __DEFAULT_FN_ATTRS
	5735	_mm_maskz_set1_epi32( __mmask8 __M, int __A)
	5736	{
	5737	return (__m128i)__builtin_ia32_selectd_128(__M,
	5738	(__v4si) _mm_set1_epi32(__A),
	5739	(__v4si)_mm_setzero_si128());
	5740	}
5731	5741
5732		#define _mm_maskz_set1_epi32(M, A) __extension__ ({ \
5733		(__m128i)__builtin_ia32_pbroadcastd128_gpr_mask((int)(A), \
5734		(__v4si)_mm_setzero_si128(), \
5735		(__mmask8)(M)); })
	5742	static __inline__ __m256i __DEFAULT_FN_ATTRS
	5743	_mm256_mask_set1_epi32(__m256i __O, __mmask8 __M, int __A)
	5744	{
	5745	return (__m256i)__builtin_ia32_selectd_256(__M,
	5746	(__v8si) _mm256_set1_epi32(__A),
	5747	(__v8si)__O);
	5748	}
5736	5749
5737		#define _mm256_mask_set1_epi32(O, M, A) __extension__ ({ \
5738		(__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \
5739		(__v8si)(__m256i)(O), \
5740		(__mmask8)(M)); })
	5750	static __inline__ __m256i __DEFAULT_FN_ATTRS
	5751	_mm256_maskz_set1_epi32( __mmask8 __M, int __A)
	5752	{
	5753	return (__m256i)__builtin_ia32_selectd_256(__M,
	5754	(__v8si) _mm256_set1_epi32(__A),
	5755	(__v8si)_mm256_setzero_si256());
	5756	}
5741	5757
5742		#define _mm256_maskz_set1_epi32(M, A) __extension__ ({ \
5743		(__m256i)__builtin_ia32_pbroadcastd256_gpr_mask((int)(A), \
5744		(__v8si)_mm256_setzero_si256(), \
5745		(__mmask8)(M)); })
5746	5758
5747	5759	#ifdef __x86_64__
5748	5760	static __inline__ __m128i __DEFAULT_FN_ATTRS
5749	5761	_mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
5750	5762	{
5751		return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A, (__v2di) __O,
5752		__M);
	5763	return (__m128i) __builtin_ia32_selectq_128(__M,
	5764	(__v2di) _mm_set1_epi64x(__A),
	5765	(__v2di) __O);
5753	5766	}
5754	5767
5755	5768	static __inline__ __m128i __DEFAULT_FN_ATTRS
5756	5769	_mm_maskz_set1_epi64 (__mmask8 __M, long long __A)
5757	5770	{
5758		return (__m128i) __builtin_ia32_pbroadcastq128_gpr_mask (__A,
5759		(__v2di)
5760		_mm_setzero_si128 (),
5761		__M);
	5771	return (__m128i) __builtin_ia32_selectq_128(__M,
	5772	(__v2di) _mm_set1_epi64x(__A),
	5773	(__v2di) _mm_setzero_si128());
5762	5774	}
5763	5775
5764	5776	static __inline__ __m256i __DEFAULT_FN_ATTRS
5765	5777	_mm256_mask_set1_epi64 (__m256i __O, __mmask8 __M, long long __A)
5766	5778	{
5767		return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A, (__v4di) __O,
5768		__M);
	5779	return (__m256i) __builtin_ia32_selectq_256(__M,
	5780	(__v4di) _mm256_set1_epi64x(__A),
	5781	(__v4di) __O) ;
5769	5782	}
5770	5783
5771	5784	static __inline__ __m256i __DEFAULT_FN_ATTRS
5772	5785	_mm256_maskz_set1_epi64 (__mmask8 __M, long long __A)
5773	5786	{
5774		return (__m256i) __builtin_ia32_pbroadcastq256_gpr_mask (__A,
5775		(__v4di)
5776		_mm256_setzero_si256 (),
5777		__M);
	5787	return (__m256i) __builtin_ia32_selectq_256(__M,
	5788	(__v4di) _mm256_set1_epi64x(__A),
	5789	(__v4di) _mm256_setzero_si256());
5778	5790	}
	5791
5779	5792	#endif
5780	5793
5781	5794	#define _mm_fixupimm_pd(A, B, C, imm) __extension__ ({ \

c_headers/clflushoptintrin.h+1-1

...	...	@@ -32,7 +32,7 @@
32	32	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("clflushopt")))
33	33
34	34	static __inline__ void __DEFAULT_FN_ATTRS
35		_mm_clflushopt(char * __m) {
	35	_mm_clflushopt(void const * __m) {
36	36	__builtin_ia32_clflushopt(__m);
37	37	}
38	38

c_headers/clwbintrin.h created+52

...	...	@@ -0,0 +1,52 @@
	1	/*===---- clwbintrin.h - CLWB intrinsic ------------------------------------===
	2	*
	3	* Permission is hereby granted, free of charge, to any person obtaining a copy
	4	* of this software and associated documentation files (the "Software"), to deal
	5	* in the Software without restriction, including without limitation the rights
	6	* to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
	7	* copies of the Software, and to permit persons to whom the Software is
	8	* furnished to do so, subject to the following conditions:
	9	*
	10	* The above copyright notice and this permission notice shall be included in
	11	* all copies or substantial portions of the Software.
	12	*
	13	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	14	* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	15	* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
	16	* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
	17	* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
	18	* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
	19	* THE SOFTWARE.
	20	*
	21	*===-----------------------------------------------------------------------===
	22	*/
	23
	24	#ifndef __IMMINTRIN_H
	25	#error "Never use <clwbintrin.h> directly; include <immintrin.h> instead."
	26	#endif
	27
	28	#ifndef __CLWBINTRIN_H
	29	#define __CLWBINTRIN_H
	30
	31	/* Define the default attributes for the functions in this file. */
	32	#define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("clwb")))
	33
	34	/// \brief Writes back to memory the cache line (if modified) that contains the
	35	/// linear address specified in \a __p from any level of the cache hierarchy in
	36	/// the cache coherence domain
	37	///
	38	/// \headerfile <immintrin.h>
	39	///
	40	/// This intrinsic corresponds to the <c> CLWB </c> instruction.
	41	///
	42	/// \param __p
	43	/// A pointer to the memory location used to identify the cache line to be
	44	/// written back.
	45	static __inline__ void __DEFAULT_FN_ATTRS
	46	_mm_clwb(void const *__p) {
	47	__builtin_ia32_clwb(__p);
	48	}
	49
	50	#undef __DEFAULT_FN_ATTRS
	51
	52	#endif

c_headers/cuda_wrappers/new+50-1

...	...	@@ -26,7 +26,6 @@
26	26
27	27	#include_next <new>
28	28
29		// Device overrides for placement new and delete.
30	29	#pragma push_macro("CUDA_NOEXCEPT")
31	30	#if __cplusplus >= 201103L
32	31	#define CUDA_NOEXCEPT noexcept
...	...	@@ -34,6 +33,55 @@
34	33	#define CUDA_NOEXCEPT
35	34	#endif
36	35
	36	// Device overrides for non-placement new and delete.
	37	__device__ inline void *operator new(__SIZE_TYPE__ size) {
	38	if (size == 0) {
	39	size = 1;
	40	}
	41	return ::malloc(size);
	42	}
	43	__device__ inline void *operator new(__SIZE_TYPE__ size,
	44	const std::nothrow_t &) CUDA_NOEXCEPT {
	45	return ::operator new(size);
	46	}
	47
	48	__device__ inline void *operator new[](__SIZE_TYPE__ size) {
	49	return ::operator new(size);
	50	}
	51	__device__ inline void *operator new[](__SIZE_TYPE__ size,
	52	const std::nothrow_t &) {
	53	return ::operator new(size);
	54	}
	55
	56	__device__ inline void operator delete(void* ptr) CUDA_NOEXCEPT {
	57	if (ptr) {
	58	::free(ptr);
	59	}
	60	}
	61	__device__ inline void operator delete(void *ptr,
	62	const std::nothrow_t &) CUDA_NOEXCEPT {
	63	::operator delete(ptr);
	64	}
	65
	66	__device__ inline void operator delete[](void* ptr) CUDA_NOEXCEPT {
	67	::operator delete(ptr);
	68	}
	69	__device__ inline void operator delete[](void *ptr,
	70	const std::nothrow_t &) CUDA_NOEXCEPT {
	71	::operator delete(ptr);
	72	}
	73
	74	// Sized delete, C++14 only.
	75	#if __cplusplus >= 201402L
	76	__device__ void operator delete(void *ptr, __SIZE_TYPE__ size) CUDA_NOEXCEPT {
	77	::operator delete(ptr);
	78	}
	79	__device__ void operator delete[](void *ptr, __SIZE_TYPE__ size) CUDA_NOEXCEPT {
	80	::operator delete(ptr);
	81	}
	82	#endif
	83
	84	// Device overrides for placement new and delete.
37	85	__device__ inline void operator new(__SIZE_TYPE__, void __ptr) CUDA_NOEXCEPT {
38	86	return __ptr;
39	87	}
...	...	@@ -42,6 +90,7 @@ __device__ inline void operator new[](__SIZE_TYPE__, void __ptr) CUDA_NOEXCEPT
42	90	}
43	91	__device__ inline void operator delete(void , void ) CUDA_NOEXCEPT {}
44	92	__device__ inline void operator delete[](void , void ) CUDA_NOEXCEPT {}
	93
45	94	#pragma pop_macro("CUDA_NOEXCEPT")
46	95
47	96	#endif // include guard

c_headers/emmintrin.h+10-2

...	...	@@ -2258,7 +2258,11 @@ _mm_adds_epu16(__m128i __a, __m128i __b)
2258	2258	static __inline__ __m128i __DEFAULT_FN_ATTRS
2259	2259	_mm_avg_epu8(__m128i __a, __m128i __b)
2260	2260	{
2261		return (__m128i)__builtin_ia32_pavgb128((__v16qi)__a, (__v16qi)__b);
	2261	typedef unsigned short __v16hu __attribute__ ((__vector_size__ (32)));
	2262	return (__m128i)__builtin_convertvector(
	2263	((__builtin_convertvector((__v16qu)__a, __v16hu) +
	2264	__builtin_convertvector((__v16qu)__b, __v16hu)) + 1)
	2265	>> 1, __v16qu);
2262	2266	}
2263	2267
2264	2268	/// \brief Computes the rounded avarages of corresponding elements of two
...	...	@@ -2278,7 +2282,11 @@ _mm_avg_epu8(__m128i __a, __m128i __b)
2278	2282	static __inline__ __m128i __DEFAULT_FN_ATTRS
2279	2283	_mm_avg_epu16(__m128i __a, __m128i __b)
2280	2284	{
2281		return (__m128i)__builtin_ia32_pavgw128((__v8hi)__a, (__v8hi)__b);
	2285	typedef unsigned int __v8su __attribute__ ((__vector_size__ (32)));
	2286	return (__m128i)__builtin_convertvector(
	2287	((__builtin_convertvector((__v8hu)__a, __v8su) +
	2288	__builtin_convertvector((__v8hu)__b, __v8su)) + 1)
	2289	>> 1, __v8hu);
2282	2290	}
2283	2291
2284	2292	/// \brief Multiplies the corresponding elements of two 128-bit signed [8 x i16]

c_headers/float.h+14

...	...	@@ -143,4 +143,18 @@
143	143	# define LDBL_DECIMAL_DIG __LDBL_DECIMAL_DIG__
144	144	#endif
145	145
	146	#ifdef __STDC_WANT_IEC_60559_TYPES_EXT__
	147	# define FLT16_MANT_DIG __FLT16_MANT_DIG__
	148	# define FLT16_DECIMAL_DIG __FLT16_DECIMAL_DIG__
	149	# define FLT16_DIG __FLT16_DIG__
	150	# define FLT16_MIN_EXP __FLT16_MIN_EXP__
	151	# define FLT16_MIN_10_EXP __FLT16_MIN_10_EXP__
	152	# define FLT16_MAX_EXP __FLT16_MAX_EXP__
	153	# define FLT16_MAX_10_EXP __FLT16_MAX_10_EXP__
	154	# define FLT16_MAX __FLT16_MAX__
	155	# define FLT16_EPSILON __FLT16_EPSILON__
	156	# define FLT16_MIN __FLT16_MIN__
	157	# define FLT16_TRUE_MIN __FLT16_TRUE_MIN__
	158	#endif /* __STDC_WANT_IEC_60559_TYPES_EXT__ */
	159
146	160	#endif /* __FLOAT_H */

c_headers/immintrin.h+4

...	...	@@ -58,6 +58,10 @@
58	58	#include <clflushoptintrin.h>
59	59	#endif
60	60
	61	#if !defined(_MSC_VER) \|\| __has_feature(modules) \|\| defined(__CLWB__)
	62	#include <clwbintrin.h>
	63	#endif
	64
61	65	#if !defined(_MSC_VER) \|\| __has_feature(modules) \|\| defined(__AVX__)
62	66	#include <avxintrin.h>
63	67	#endif

c_headers/intrin.h+5-1

...	...	@@ -38,6 +38,10 @@
38	38	#include <armintr.h>
39	39	#endif
40	40
	41	#if defined(_M_ARM64)
	42	#include <arm64intr.h>
	43	#endif
	44
41	45	/* For the definition of jmp_buf. */
42	46	#if __STDC_HOSTED__
43	47	#include <setjmp.h>
...	...	@@ -828,7 +832,7 @@ _InterlockedCompareExchange_nf(long volatile *_Destination,
828	832	__ATOMIC_SEQ_CST, __ATOMIC_RELAXED);
829	833	return _Comparand;
830	834	}
831		static __inline__ short __DEFAULT_FN_ATTRS
	835	static __inline__ long __DEFAULT_FN_ATTRS
832	836	_InterlockedCompareExchange_rel(long volatile *_Destination,
833	837	long _Exchange, long _Comparand) {
834	838	__atomic_compare_exchange(_Destination, &_Comparand, &_Exchange, 0,

c_headers/opencl-c.h+23-336

...	...	@@ -11381,6 +11381,8 @@ half16 __ovld __cnfn bitselect(half16 a, half16 b, half16 c);
11381	11381	* For each component of a vector type,
11382	11382	* result[i] = if MSB of c[i] is set ? b[i] : a[i].
11383	11383	* For a scalar type, result = c ? b : a.
	11384	* b and a must have the same type.
	11385	* c must have the same number of elements and bits as a.
11384	11386	*/
11385	11387	char __ovld __cnfn select(char a, char b, char c);
11386	11388	uchar __ovld __cnfn select(uchar a, uchar b, char c);
...	...	@@ -11394,60 +11396,7 @@ char8 __ovld __cnfn select(char8 a, char8 b, char8 c);
11394	11396	uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, char8 c);
11395	11397	char16 __ovld __cnfn select(char16 a, char16 b, char16 c);
11396	11398	uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, char16 c);
11397		short __ovld __cnfn select(short a, short b, char c);
11398		ushort __ovld __cnfn select(ushort a, ushort b, char c);
11399		short2 __ovld __cnfn select(short2 a, short2 b, char2 c);
11400		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, char2 c);
11401		short3 __ovld __cnfn select(short3 a, short3 b, char3 c);
11402		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, char3 c);
11403		short4 __ovld __cnfn select(short4 a, short4 b, char4 c);
11404		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, char4 c);
11405		short8 __ovld __cnfn select(short8 a, short8 b, char8 c);
11406		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, char8 c);
11407		short16 __ovld __cnfn select(short16 a, short16 b, char16 c);
11408		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, char16 c);
11409		int __ovld __cnfn select(int a, int b, char c);
11410		uint __ovld __cnfn select(uint a, uint b, char c);
11411		int2 __ovld __cnfn select(int2 a, int2 b, char2 c);
11412		uint2 __ovld __cnfn select(uint2 a, uint2 b, char2 c);
11413		int3 __ovld __cnfn select(int3 a, int3 b, char3 c);
11414		uint3 __ovld __cnfn select(uint3 a, uint3 b, char3 c);
11415		int4 __ovld __cnfn select(int4 a, int4 b, char4 c);
11416		uint4 __ovld __cnfn select(uint4 a, uint4 b, char4 c);
11417		int8 __ovld __cnfn select(int8 a, int8 b, char8 c);
11418		uint8 __ovld __cnfn select(uint8 a, uint8 b, char8 c);
11419		int16 __ovld __cnfn select(int16 a, int16 b, char16 c);
11420		uint16 __ovld __cnfn select(uint16 a, uint16 b, char16 c);
11421		long __ovld __cnfn select(long a, long b, char c);
11422		ulong __ovld __cnfn select(ulong a, ulong b, char c);
11423		long2 __ovld __cnfn select(long2 a, long2 b, char2 c);
11424		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, char2 c);
11425		long3 __ovld __cnfn select(long3 a, long3 b, char3 c);
11426		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, char3 c);
11427		long4 __ovld __cnfn select(long4 a, long4 b, char4 c);
11428		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, char4 c);
11429		long8 __ovld __cnfn select(long8 a, long8 b, char8 c);
11430		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, char8 c);
11431		long16 __ovld __cnfn select(long16 a, long16 b, char16 c);
11432		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, char16 c);
11433		float __ovld __cnfn select(float a, float b, char c);
11434		float2 __ovld __cnfn select(float2 a, float2 b, char2 c);
11435		float3 __ovld __cnfn select(float3 a, float3 b, char3 c);
11436		float4 __ovld __cnfn select(float4 a, float4 b, char4 c);
11437		float8 __ovld __cnfn select(float8 a, float8 b, char8 c);
11438		float16 __ovld __cnfn select(float16 a, float16 b, char16 c);
11439		char __ovld __cnfn select(char a, char b, short c);
11440		uchar __ovld __cnfn select(uchar a, uchar b, short c);
11441		char2 __ovld __cnfn select(char2 a, char2 b, short2 c);
11442		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, short2 c);
11443		char3 __ovld __cnfn select(char3 a, char3 b, short3 c);
11444		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, short3 c);
11445		char4 __ovld __cnfn select(char4 a, char4 b, short4 c);
11446		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, short4 c);
11447		char8 __ovld __cnfn select(char8 a, char8 b, short8 c);
11448		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, short8 c);
11449		char16 __ovld __cnfn select(char16 a, char16 b, short16 c);
11450		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, short16 c);
	11399
11451	11400	short __ovld __cnfn select(short a, short b, short c);
11452	11401	ushort __ovld __cnfn select(ushort a, ushort b, short c);
11453	11402	short2 __ovld __cnfn select(short2 a, short2 b, short2 c);
...	...	@@ -11460,60 +11409,7 @@ short8 __ovld __cnfn select(short8 a, short8 b, short8 c);
11460	11409	ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, short8 c);
11461	11410	short16 __ovld __cnfn select(short16 a, short16 b, short16 c);
11462	11411	ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, short16 c);
11463		int __ovld __cnfn select(int a, int b, short c);
11464		uint __ovld __cnfn select(uint a, uint b, short c);
11465		int2 __ovld __cnfn select(int2 a, int2 b, short2 c);
11466		uint2 __ovld __cnfn select(uint2 a, uint2 b, short2 c);
11467		int3 __ovld __cnfn select(int3 a, int3 b, short3 c);
11468		uint3 __ovld __cnfn select(uint3 a, uint3 b, short3 c);
11469		int4 __ovld __cnfn select(int4 a, int4 b, short4 c);
11470		uint4 __ovld __cnfn select(uint4 a, uint4 b, short4 c);
11471		int8 __ovld __cnfn select(int8 a, int8 b, short8 c);
11472		uint8 __ovld __cnfn select(uint8 a, uint8 b, short8 c);
11473		int16 __ovld __cnfn select(int16 a, int16 b, short16 c);
11474		uint16 __ovld __cnfn select(uint16 a, uint16 b, short16 c);
11475		long __ovld __cnfn select(long a, long b, short c);
11476		ulong __ovld __cnfn select(ulong a, ulong b, short c);
11477		long2 __ovld __cnfn select(long2 a, long2 b, short2 c);
11478		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, short2 c);
11479		long3 __ovld __cnfn select(long3 a, long3 b, short3 c);
11480		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, short3 c);
11481		long4 __ovld __cnfn select(long4 a, long4 b, short4 c);
11482		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, short4 c);
11483		long8 __ovld __cnfn select(long8 a, long8 b, short8 c);
11484		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, short8 c);
11485		long16 __ovld __cnfn select(long16 a, long16 b, short16 c);
11486		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, short16 c);
11487		float __ovld __cnfn select(float a, float b, short c);
11488		float2 __ovld __cnfn select(float2 a, float2 b, short2 c);
11489		float3 __ovld __cnfn select(float3 a, float3 b, short3 c);
11490		float4 __ovld __cnfn select(float4 a, float4 b, short4 c);
11491		float8 __ovld __cnfn select(float8 a, float8 b, short8 c);
11492		float16 __ovld __cnfn select(float16 a, float16 b, short16 c);
11493		char __ovld __cnfn select(char a, char b, int c);
11494		uchar __ovld __cnfn select(uchar a, uchar b, int c);
11495		char2 __ovld __cnfn select(char2 a, char2 b, int2 c);
11496		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, int2 c);
11497		char3 __ovld __cnfn select(char3 a, char3 b, int3 c);
11498		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, int3 c);
11499		char4 __ovld __cnfn select(char4 a, char4 b, int4 c);
11500		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, int4 c);
11501		char8 __ovld __cnfn select(char8 a, char8 b, int8 c);
11502		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, int8 c);
11503		char16 __ovld __cnfn select(char16 a, char16 b, int16 c);
11504		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, int16 c);
11505		short __ovld __cnfn select(short a, short b, int c);
11506		ushort __ovld __cnfn select(ushort a, ushort b, int c);
11507		short2 __ovld __cnfn select(short2 a, short2 b, int2 c);
11508		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, int2 c);
11509		short3 __ovld __cnfn select(short3 a, short3 b, int3 c);
11510		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, int3 c);
11511		short4 __ovld __cnfn select(short4 a, short4 b, int4 c);
11512		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, int4 c);
11513		short8 __ovld __cnfn select(short8 a, short8 b, int8 c);
11514		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, int8 c);
11515		short16 __ovld __cnfn select(short16 a, short16 b, int16 c);
11516		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, int16 c);
	11412
11517	11413	int __ovld __cnfn select(int a, int b, int c);
11518	11414	uint __ovld __cnfn select(uint a, uint b, int c);
11519	11415	int2 __ovld __cnfn select(int2 a, int2 b, int2 c);
...	...	@@ -11526,60 +11422,13 @@ int8 __ovld __cnfn select(int8 a, int8 b, int8 c);
11526	11422	uint8 __ovld __cnfn select(uint8 a, uint8 b, int8 c);
11527	11423	int16 __ovld __cnfn select(int16 a, int16 b, int16 c);
11528	11424	uint16 __ovld __cnfn select(uint16 a, uint16 b, int16 c);
11529		long __ovld __cnfn select(long a, long b, int c);
11530		ulong __ovld __cnfn select(ulong a, ulong b, int c);
11531		long2 __ovld __cnfn select(long2 a, long2 b, int2 c);
11532		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, int2 c);
11533		long3 __ovld __cnfn select(long3 a, long3 b, int3 c);
11534		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, int3 c);
11535		long4 __ovld __cnfn select(long4 a, long4 b, int4 c);
11536		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, int4 c);
11537		long8 __ovld __cnfn select(long8 a, long8 b, int8 c);
11538		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, int8 c);
11539		long16 __ovld __cnfn select(long16 a, long16 b, int16 c);
11540		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, int16 c);
11541	11425	float __ovld __cnfn select(float a, float b, int c);
11542	11426	float2 __ovld __cnfn select(float2 a, float2 b, int2 c);
11543	11427	float3 __ovld __cnfn select(float3 a, float3 b, int3 c);
11544	11428	float4 __ovld __cnfn select(float4 a, float4 b, int4 c);
11545	11429	float8 __ovld __cnfn select(float8 a, float8 b, int8 c);
11546	11430	float16 __ovld __cnfn select(float16 a, float16 b, int16 c);
11547		char __ovld __cnfn select(char a, char b, long c);
11548		uchar __ovld __cnfn select(uchar a, uchar b, long c);
11549		char2 __ovld __cnfn select(char2 a, char2 b, long2 c);
11550		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, long2 c);
11551		char3 __ovld __cnfn select(char3 a, char3 b, long3 c);
11552		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, long3 c);
11553		char4 __ovld __cnfn select(char4 a, char4 b, long4 c);
11554		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, long4 c);
11555		char8 __ovld __cnfn select(char8 a, char8 b, long8 c);
11556		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, long8 c);
11557		char16 __ovld __cnfn select(char16 a, char16 b, long16 c);
11558		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, long16 c);
11559		short __ovld __cnfn select(short a, short b, long c);
11560		ushort __ovld __cnfn select(ushort a, ushort b, long c);
11561		short2 __ovld __cnfn select(short2 a, short2 b, long2 c);
11562		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, long2 c);
11563		short3 __ovld __cnfn select(short3 a, short3 b, long3 c);
11564		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, long3 c);
11565		short4 __ovld __cnfn select(short4 a, short4 b, long4 c);
11566		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, long4 c);
11567		short8 __ovld __cnfn select(short8 a, short8 b, long8 c);
11568		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, long8 c);
11569		short16 __ovld __cnfn select(short16 a, short16 b, long16 c);
11570		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, long16 c);
11571		int __ovld __cnfn select(int a, int b, long c);
11572		uint __ovld __cnfn select(uint a, uint b, long c);
11573		int2 __ovld __cnfn select(int2 a, int2 b, long2 c);
11574		uint2 __ovld __cnfn select(uint2 a, uint2 b, long2 c);
11575		int3 __ovld __cnfn select(int3 a, int3 b, long3 c);
11576		uint3 __ovld __cnfn select(uint3 a, uint3 b, long3 c);
11577		int4 __ovld __cnfn select(int4 a, int4 b, long4 c);
11578		uint4 __ovld __cnfn select(uint4 a, uint4 b, long4 c);
11579		int8 __ovld __cnfn select(int8 a, int8 b, long8 c);
11580		uint8 __ovld __cnfn select(uint8 a, uint8 b, long8 c);
11581		int16 __ovld __cnfn select(int16 a, int16 b, long16 c);
11582		uint16 __ovld __cnfn select(uint16 a, uint16 b, long16 c);
	11431
11583	11432	long __ovld __cnfn select(long a, long b, long c);
11584	11433	ulong __ovld __cnfn select(ulong a, ulong b, long c);
11585	11434	long2 __ovld __cnfn select(long2 a, long2 b, long2 c);
...	...	@@ -11592,12 +11441,7 @@ long8 __ovld __cnfn select(long8 a, long8 b, long8 c);
11592	11441	ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, long8 c);
11593	11442	long16 __ovld __cnfn select(long16 a, long16 b, long16 c);
11594	11443	ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, long16 c);
11595		float __ovld __cnfn select(float a, float b, long c);
11596		float2 __ovld __cnfn select(float2 a, float2 b, long2 c);
11597		float3 __ovld __cnfn select(float3 a, float3 b, long3 c);
11598		float4 __ovld __cnfn select(float4 a, float4 b, long4 c);
11599		float8 __ovld __cnfn select(float8 a, float8 b, long8 c);
11600		float16 __ovld __cnfn select(float16 a, float16 b, long16 c);
	11444
11601	11445	char __ovld __cnfn select(char a, char b, uchar c);
11602	11446	uchar __ovld __cnfn select(uchar a, uchar b, uchar c);
11603	11447	char2 __ovld __cnfn select(char2 a, char2 b, uchar2 c);
...	...	@@ -11610,60 +11454,7 @@ char8 __ovld __cnfn select(char8 a, char8 b, uchar8 c);
11610	11454	uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, uchar8 c);
11611	11455	char16 __ovld __cnfn select(char16 a, char16 b, uchar16 c);
11612	11456	uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, uchar16 c);
11613		short __ovld __cnfn select(short a, short b, uchar c);
11614		ushort __ovld __cnfn select(ushort a, ushort b, uchar c);
11615		short2 __ovld __cnfn select(short2 a, short2 b, uchar2 c);
11616		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, uchar2 c);
11617		short3 __ovld __cnfn select(short3 a, short3 b, uchar3 c);
11618		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, uchar3 c);
11619		short4 __ovld __cnfn select(short4 a, short4 b, uchar4 c);
11620		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, uchar4 c);
11621		short8 __ovld __cnfn select(short8 a, short8 b, uchar8 c);
11622		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, uchar8 c);
11623		short16 __ovld __cnfn select(short16 a, short16 b, uchar16 c);
11624		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, uchar16 c);
11625		int __ovld __cnfn select(int a, int b, uchar c);
11626		uint __ovld __cnfn select(uint a, uint b, uchar c);
11627		int2 __ovld __cnfn select(int2 a, int2 b, uchar2 c);
11628		uint2 __ovld __cnfn select(uint2 a, uint2 b, uchar2 c);
11629		int3 __ovld __cnfn select(int3 a, int3 b, uchar3 c);
11630		uint3 __ovld __cnfn select(uint3 a, uint3 b, uchar3 c);
11631		int4 __ovld __cnfn select(int4 a, int4 b, uchar4 c);
11632		uint4 __ovld __cnfn select(uint4 a, uint4 b, uchar4 c);
11633		int8 __ovld __cnfn select(int8 a, int8 b, uchar8 c);
11634		uint8 __ovld __cnfn select(uint8 a, uint8 b, uchar8 c);
11635		int16 __ovld __cnfn select(int16 a, int16 b, uchar16 c);
11636		uint16 __ovld __cnfn select(uint16 a, uint16 b, uchar16 c);
11637		long __ovld __cnfn select(long a, long b, uchar c);
11638		ulong __ovld __cnfn select(ulong a, ulong b, uchar c);
11639		long2 __ovld __cnfn select(long2 a, long2 b, uchar2 c);
11640		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, uchar2 c);
11641		long3 __ovld __cnfn select(long3 a, long3 b, uchar3 c);
11642		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, uchar3 c);
11643		long4 __ovld __cnfn select(long4 a, long4 b, uchar4 c);
11644		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, uchar4 c);
11645		long8 __ovld __cnfn select(long8 a, long8 b, uchar8 c);
11646		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, uchar8 c);
11647		long16 __ovld __cnfn select(long16 a, long16 b, uchar16 c);
11648		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, uchar16 c);
11649		float __ovld __cnfn select(float a, float b, uchar c);
11650		float2 __ovld __cnfn select(float2 a, float2 b, uchar2 c);
11651		float3 __ovld __cnfn select(float3 a, float3 b, uchar3 c);
11652		float4 __ovld __cnfn select(float4 a, float4 b, uchar4 c);
11653		float8 __ovld __cnfn select(float8 a, float8 b, uchar8 c);
11654		float16 __ovld __cnfn select(float16 a, float16 b, uchar16 c);
11655		char __ovld __cnfn select(char a, char b, ushort c);
11656		uchar __ovld __cnfn select(uchar a, uchar b, ushort c);
11657		char2 __ovld __cnfn select(char2 a, char2 b, ushort2 c);
11658		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, ushort2 c);
11659		char3 __ovld __cnfn select(char3 a, char3 b, ushort3 c);
11660		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, ushort3 c);
11661		char4 __ovld __cnfn select(char4 a, char4 b, ushort4 c);
11662		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, ushort4 c);
11663		char8 __ovld __cnfn select(char8 a, char8 b, ushort8 c);
11664		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, ushort8 c);
11665		char16 __ovld __cnfn select(char16 a, char16 b, ushort16 c);
11666		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, ushort16 c);
	11457
11667	11458	short __ovld __cnfn select(short a, short b, ushort c);
11668	11459	ushort __ovld __cnfn select(ushort a, ushort b, ushort c);
11669	11460	short2 __ovld __cnfn select(short2 a, short2 b, ushort2 c);
...	...	@@ -11676,60 +11467,7 @@ short8 __ovld __cnfn select(short8 a, short8 b, ushort8 c);
11676	11467	ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, ushort8 c);
11677	11468	short16 __ovld __cnfn select(short16 a, short16 b, ushort16 c);
11678	11469	ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, ushort16 c);
11679		int __ovld __cnfn select(int a, int b, ushort c);
11680		uint __ovld __cnfn select(uint a, uint b, ushort c);
11681		int2 __ovld __cnfn select(int2 a, int2 b, ushort2 c);
11682		uint2 __ovld __cnfn select(uint2 a, uint2 b, ushort2 c);
11683		int3 __ovld __cnfn select(int3 a, int3 b, ushort3 c);
11684		uint3 __ovld __cnfn select(uint3 a, uint3 b, ushort3 c);
11685		int4 __ovld __cnfn select(int4 a, int4 b, ushort4 c);
11686		uint4 __ovld __cnfn select(uint4 a, uint4 b, ushort4 c);
11687		int8 __ovld __cnfn select(int8 a, int8 b, ushort8 c);
11688		uint8 __ovld __cnfn select(uint8 a, uint8 b, ushort8 c);
11689		int16 __ovld __cnfn select(int16 a, int16 b, ushort16 c);
11690		uint16 __ovld __cnfn select(uint16 a, uint16 b, ushort16 c);
11691		long __ovld __cnfn select(long a, long b, ushort c);
11692		ulong __ovld __cnfn select(ulong a, ulong b, ushort c);
11693		long2 __ovld __cnfn select(long2 a, long2 b, ushort2 c);
11694		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, ushort2 c);
11695		long3 __ovld __cnfn select(long3 a, long3 b, ushort3 c);
11696		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, ushort3 c);
11697		long4 __ovld __cnfn select(long4 a, long4 b, ushort4 c);
11698		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, ushort4 c);
11699		long8 __ovld __cnfn select(long8 a, long8 b, ushort8 c);
11700		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, ushort8 c);
11701		long16 __ovld __cnfn select(long16 a, long16 b, ushort16 c);
11702		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, ushort16 c);
11703		float __ovld __cnfn select(float a, float b, ushort c);
11704		float2 __ovld __cnfn select(float2 a, float2 b, ushort2 c);
11705		float3 __ovld __cnfn select(float3 a, float3 b, ushort3 c);
11706		float4 __ovld __cnfn select(float4 a, float4 b, ushort4 c);
11707		float8 __ovld __cnfn select(float8 a, float8 b, ushort8 c);
11708		float16 __ovld __cnfn select(float16 a, float16 b, ushort16 c);
11709		char __ovld __cnfn select(char a, char b, uint c);
11710		uchar __ovld __cnfn select(uchar a, uchar b, uint c);
11711		char2 __ovld __cnfn select(char2 a, char2 b, uint2 c);
11712		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, uint2 c);
11713		char3 __ovld __cnfn select(char3 a, char3 b, uint3 c);
11714		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, uint3 c);
11715		char4 __ovld __cnfn select(char4 a, char4 b, uint4 c);
11716		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, uint4 c);
11717		char8 __ovld __cnfn select(char8 a, char8 b, uint8 c);
11718		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, uint8 c);
11719		char16 __ovld __cnfn select(char16 a, char16 b, uint16 c);
11720		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, uint16 c);
11721		short __ovld __cnfn select(short a, short b, uint c);
11722		ushort __ovld __cnfn select(ushort a, ushort b, uint c);
11723		short2 __ovld __cnfn select(short2 a, short2 b, uint2 c);
11724		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, uint2 c);
11725		short3 __ovld __cnfn select(short3 a, short3 b, uint3 c);
11726		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, uint3 c);
11727		short4 __ovld __cnfn select(short4 a, short4 b, uint4 c);
11728		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, uint4 c);
11729		short8 __ovld __cnfn select(short8 a, short8 b, uint8 c);
11730		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, uint8 c);
11731		short16 __ovld __cnfn select(short16 a, short16 b, uint16 c);
11732		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, uint16 c);
	11470
11733	11471	int __ovld __cnfn select(int a, int b, uint c);
11734	11472	uint __ovld __cnfn select(uint a, uint b, uint c);
11735	11473	int2 __ovld __cnfn select(int2 a, int2 b, uint2 c);
...	...	@@ -11742,60 +11480,13 @@ int8 __ovld __cnfn select(int8 a, int8 b, uint8 c);
11742	11480	uint8 __ovld __cnfn select(uint8 a, uint8 b, uint8 c);
11743	11481	int16 __ovld __cnfn select(int16 a, int16 b, uint16 c);
11744	11482	uint16 __ovld __cnfn select(uint16 a, uint16 b, uint16 c);
11745		long __ovld __cnfn select(long a, long b, uint c);
11746		ulong __ovld __cnfn select(ulong a, ulong b, uint c);
11747		long2 __ovld __cnfn select(long2 a, long2 b, uint2 c);
11748		ulong2 __ovld __cnfn select(ulong2 a, ulong2 b, uint2 c);
11749		long3 __ovld __cnfn select(long3 a, long3 b, uint3 c);
11750		ulong3 __ovld __cnfn select(ulong3 a, ulong3 b, uint3 c);
11751		long4 __ovld __cnfn select(long4 a, long4 b, uint4 c);
11752		ulong4 __ovld __cnfn select(ulong4 a, ulong4 b, uint4 c);
11753		long8 __ovld __cnfn select(long8 a, long8 b, uint8 c);
11754		ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, uint8 c);
11755		long16 __ovld __cnfn select(long16 a, long16 b, uint16 c);
11756		ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, uint16 c);
11757	11483	float __ovld __cnfn select(float a, float b, uint c);
11758	11484	float2 __ovld __cnfn select(float2 a, float2 b, uint2 c);
11759	11485	float3 __ovld __cnfn select(float3 a, float3 b, uint3 c);
11760	11486	float4 __ovld __cnfn select(float4 a, float4 b, uint4 c);
11761	11487	float8 __ovld __cnfn select(float8 a, float8 b, uint8 c);
11762	11488	float16 __ovld __cnfn select(float16 a, float16 b, uint16 c);
11763		char __ovld __cnfn select(char a, char b, ulong c);
11764		uchar __ovld __cnfn select(uchar a, uchar b, ulong c);
11765		char2 __ovld __cnfn select(char2 a, char2 b, ulong2 c);
11766		uchar2 __ovld __cnfn select(uchar2 a, uchar2 b, ulong2 c);
11767		char3 __ovld __cnfn select(char3 a, char3 b, ulong3 c);
11768		uchar3 __ovld __cnfn select(uchar3 a, uchar3 b, ulong3 c);
11769		char4 __ovld __cnfn select(char4 a, char4 b, ulong4 c);
11770		uchar4 __ovld __cnfn select(uchar4 a, uchar4 b, ulong4 c);
11771		char8 __ovld __cnfn select(char8 a, char8 b, ulong8 c);
11772		uchar8 __ovld __cnfn select(uchar8 a, uchar8 b, ulong8 c);
11773		char16 __ovld __cnfn select(char16 a, char16 b, ulong16 c);
11774		uchar16 __ovld __cnfn select(uchar16 a, uchar16 b, ulong16 c);
11775		short __ovld __cnfn select(short a, short b, ulong c);
11776		ushort __ovld __cnfn select(ushort a, ushort b, ulong c);
11777		short2 __ovld __cnfn select(short2 a, short2 b, ulong2 c);
11778		ushort2 __ovld __cnfn select(ushort2 a, ushort2 b, ulong2 c);
11779		short3 __ovld __cnfn select(short3 a, short3 b, ulong3 c);
11780		ushort3 __ovld __cnfn select(ushort3 a, ushort3 b, ulong3 c);
11781		short4 __ovld __cnfn select(short4 a, short4 b, ulong4 c);
11782		ushort4 __ovld __cnfn select(ushort4 a, ushort4 b, ulong4 c);
11783		short8 __ovld __cnfn select(short8 a, short8 b, ulong8 c);
11784		ushort8 __ovld __cnfn select(ushort8 a, ushort8 b, ulong8 c);
11785		short16 __ovld __cnfn select(short16 a, short16 b, ulong16 c);
11786		ushort16 __ovld __cnfn select(ushort16 a, ushort16 b, ulong16 c);
11787		int __ovld __cnfn select(int a, int b, ulong c);
11788		uint __ovld __cnfn select(uint a, uint b, ulong c);
11789		int2 __ovld __cnfn select(int2 a, int2 b, ulong2 c);
11790		uint2 __ovld __cnfn select(uint2 a, uint2 b, ulong2 c);
11791		int3 __ovld __cnfn select(int3 a, int3 b, ulong3 c);
11792		uint3 __ovld __cnfn select(uint3 a, uint3 b, ulong3 c);
11793		int4 __ovld __cnfn select(int4 a, int4 b, ulong4 c);
11794		uint4 __ovld __cnfn select(uint4 a, uint4 b, ulong4 c);
11795		int8 __ovld __cnfn select(int8 a, int8 b, ulong8 c);
11796		uint8 __ovld __cnfn select(uint8 a, uint8 b, ulong8 c);
11797		int16 __ovld __cnfn select(int16 a, int16 b, ulong16 c);
11798		uint16 __ovld __cnfn select(uint16 a, uint16 b, ulong16 c);
	11489
11799	11490	long __ovld __cnfn select(long a, long b, ulong c);
11800	11491	ulong __ovld __cnfn select(ulong a, ulong b, ulong c);
11801	11492	long2 __ovld __cnfn select(long2 a, long2 b, ulong2 c);
...	...	@@ -11808,12 +11499,7 @@ long8 __ovld __cnfn select(long8 a, long8 b, ulong8 c);
11808	11499	ulong8 __ovld __cnfn select(ulong8 a, ulong8 b, ulong8 c);
11809	11500	long16 __ovld __cnfn select(long16 a, long16 b, ulong16 c);
11810	11501	ulong16 __ovld __cnfn select(ulong16 a, ulong16 b, ulong16 c);
11811		float __ovld __cnfn select(float a, float b, ulong c);
11812		float2 __ovld __cnfn select(float2 a, float2 b, ulong2 c);
11813		float3 __ovld __cnfn select(float3 a, float3 b, ulong3 c);
11814		float4 __ovld __cnfn select(float4 a, float4 b, ulong4 c);
11815		float8 __ovld __cnfn select(float8 a, float8 b, ulong8 c);
11816		float16 __ovld __cnfn select(float16 a, float16 b, ulong16 c);
	11502
11817	11503	#ifdef cl_khr_fp64
11818	11504	double __ovld __cnfn select(double a, double b, long c);
11819	11505	double2 __ovld __cnfn select(double2 a, double2 b, long2 c);
...	...	@@ -13141,13 +12827,14 @@ void __ovld __conv barrier(cl_mem_fence_flags flags);
13141	12827
13142	12828	#if __OPENCL_C_VERSION__ >= CL_VERSION_2_0
13143	12829
13144		typedef enum memory_scope
13145		{
13146		memory_scope_work_item,
13147		memory_scope_work_group,
13148		memory_scope_device,
13149		memory_scope_all_svm_devices,
13150		memory_scope_sub_group
	12830	typedef enum memory_scope {
	12831	memory_scope_work_item = __OPENCL_MEMORY_SCOPE_WORK_ITEM,
	12832	memory_scope_work_group = __OPENCL_MEMORY_SCOPE_WORK_GROUP,
	12833	memory_scope_device = __OPENCL_MEMORY_SCOPE_DEVICE,
	12834	memory_scope_all_svm_devices = __OPENCL_MEMORY_SCOPE_ALL_SVM_DEVICES,
	12835	#if defined(cl_intel_subgroups) \|\| defined(cl_khr_subgroups)
	12836	memory_scope_sub_group = __OPENCL_MEMORY_SCOPE_SUB_GROUP
	12837	#endif
13151	12838	} memory_scope;
13152	12839
13153	12840	void __ovld __conv work_group_barrier(cl_mem_fence_flags flags, memory_scope scope);
...	...	@@ -13952,11 +13639,11 @@ unsigned long __ovld atom_xor(volatile __local unsigned long *p, unsigned long v
13952	13639	// enum values aligned with what clang uses in EmitAtomicExpr()
13953	13640	typedef enum memory_order
13954	13641	{
13955		memory_order_relaxed,
13956		memory_order_acquire,
13957		memory_order_release,
13958		memory_order_acq_rel,
13959		memory_order_seq_cst
	13642	memory_order_relaxed = __ATOMIC_RELAXED,
	13643	memory_order_acquire = __ATOMIC_ACQUIRE,
	13644	memory_order_release = __ATOMIC_RELEASE,
	13645	memory_order_acq_rel = __ATOMIC_ACQ_REL,
	13646	memory_order_seq_cst = __ATOMIC_SEQ_CST
13960	13647	} memory_order;
13961	13648
13962	13649	// double atomics support requires extensions cl_khr_int64_base_atomics and cl_khr_int64_extended_atomics

c_headers/unwind.h+59-21

...	...	@@ -76,7 +76,13 @@ typedef intptr_t _sleb128_t;
76	76	typedef uintptr_t _uleb128_t;
77	77
78	78	struct _Unwind_Context;
	79	#if defined(__arm__) && !(defined(__USING_SJLJ_EXCEPTIONS__) \|\| defined(__ARM_DWARF_EH__))
	80	struct _Unwind_Control_Block;
	81	typedef struct _Unwind_Control_Block _Unwind_Exception; /* Alias */
	82	#else
79	83	struct _Unwind_Exception;
	84	typedef struct _Unwind_Exception _Unwind_Exception;
	85	#endif
80	86	typedef enum {
81	87	_URC_NO_REASON = 0,
82	88	#if defined(__arm__) && !defined(__USING_SJLJ_EXCEPTIONS__) && \
...	...	@@ -109,8 +115,42 @@ typedef enum {
109	115	} _Unwind_Action;
110	116
111	117	typedef void (*_Unwind_Exception_Cleanup_Fn)(_Unwind_Reason_Code,
112		struct _Unwind_Exception *);
113
	118	_Unwind_Exception *);
	119
	120	#if defined(__arm__) && !(defined(__USING_SJLJ_EXCEPTIONS__) \|\| defined(__ARM_DWARF_EH__))
	121	typedef struct _Unwind_Control_Block _Unwind_Control_Block;
	122	typedef uint32_t _Unwind_EHT_Header;
	123
	124	struct _Unwind_Control_Block {
	125	uint64_t exception_class;
	126	void (exception_cleanup)(_Unwind_Reason_Code, _Unwind_Control_Block );
	127	/* unwinder cache (private fields for the unwinder's use) */
	128	struct {
	129	uint32_t reserved1; /* forced unwind stop function, 0 if not forced */
	130	uint32_t reserved2; /* personality routine */
	131	uint32_t reserved3; /* callsite */
	132	uint32_t reserved4; /* forced unwind stop argument */
	133	uint32_t reserved5;
	134	} unwinder_cache;
	135	/* propagation barrier cache (valid after phase 1) */
	136	struct {
	137	uint32_t sp;
	138	uint32_t bitpattern[5];
	139	} barrier_cache;
	140	/* cleanup cache (preserved over cleanup) */
	141	struct {
	142	uint32_t bitpattern[4];
	143	} cleanup_cache;
	144	/* personality cache (for personality's benefit) */
	145	struct {
	146	uint32_t fnstart; /* function start address */
	147	_Unwind_EHT_Header ehtp; / pointer to EHT entry header word */
	148	uint32_t additional; /* additional data */
	149	uint32_t reserved1;
	150	} pr_cache;
	151	long long int : 0; /* force alignment of next item to 8-byte boundary */
	152	} __attribute__((__aligned__(8)));
	153	#else
114	154	struct _Unwind_Exception {
115	155	_Unwind_Exception_Class exception_class;
116	156	_Unwind_Exception_Cleanup_Fn exception_cleanup;
...	...	@@ -120,23 +160,24 @@ struct _Unwind_Exception {
120	160	* aligned". GCC has interpreted this to mean "use the maximum useful
121	161	* alignment for the target"; so do we. */
122	162	} __attribute__((__aligned__));
	163	#endif
123	164
124	165	typedef _Unwind_Reason_Code (*_Unwind_Stop_Fn)(int, _Unwind_Action,
125	166	_Unwind_Exception_Class,
126		struct _Unwind_Exception *,
	167	_Unwind_Exception *,
127	168	struct _Unwind_Context *,
128	169	void *);
129	170
130		typedef _Unwind_Reason_Code (*_Unwind_Personality_Fn)(
131		int, _Unwind_Action, _Unwind_Exception_Class, struct _Unwind_Exception *,
132		struct _Unwind_Context *);
	171	typedef _Unwind_Reason_Code (*_Unwind_Personality_Fn)(int, _Unwind_Action,
	172	_Unwind_Exception_Class,
	173	_Unwind_Exception *,
	174	struct _Unwind_Context *);
133	175	typedef _Unwind_Personality_Fn __personality_routine;
134	176
135	177	typedef _Unwind_Reason_Code (_Unwind_Trace_Fn)(struct _Unwind_Context ,
136	178	void *);
137	179
138		#if defined(__arm__) && !defined(__APPLE__)
139
	180	#if defined(__arm__) && !(defined(__USING_SJLJ_EXCEPTIONS__) \|\| defined(__ARM_DWARF_EH__))
140	181	typedef enum {
141	182	_UVRSC_CORE = 0, /* integer register */
142	183	_UVRSC_VFP = 1, /* vfp */
...	...	@@ -158,14 +199,12 @@ typedef enum {
158	199	_UVRSR_FAILED = 2
159	200	} _Unwind_VRS_Result;
160	201
161		#if !defined(__USING_SJLJ_EXCEPTIONS__) && !defined(__ARM_DWARF_EH__)
162	202	typedef uint32_t _Unwind_State;
163	203	#define _US_VIRTUAL_UNWIND_FRAME ((_Unwind_State)0)
164	204	#define _US_UNWIND_FRAME_STARTING ((_Unwind_State)1)
165	205	#define _US_UNWIND_FRAME_RESUME ((_Unwind_State)2)
166	206	#define _US_ACTION_MASK ((_Unwind_State)3)
167	207	#define _US_FORCE_UNWIND ((_Unwind_State)8)
168		#endif
169	208
170	209	_Unwind_VRS_Result _Unwind_VRS_Get(struct _Unwind_Context *__context,
171	210	_Unwind_VRS_RegClass __regclass,
...	...	@@ -224,13 +263,12 @@ _Unwind_Ptr _Unwind_GetRegionStart(struct _Unwind_Context *);
224	263
225	264	/* DWARF EH functions; currently not available on Darwin/ARM */
226	265	#if !defined(__APPLE__) \|\| !defined(__arm__)
227
228		_Unwind_Reason_Code _Unwind_RaiseException(struct _Unwind_Exception *);
229		_Unwind_Reason_Code _Unwind_ForcedUnwind(struct _Unwind_Exception *,
230		_Unwind_Stop_Fn, void *);
231		void _Unwind_DeleteException(struct _Unwind_Exception *);
232		void _Unwind_Resume(struct _Unwind_Exception *);
233		_Unwind_Reason_Code _Unwind_Resume_or_Rethrow(struct _Unwind_Exception *);
	266	_Unwind_Reason_Code _Unwind_RaiseException(_Unwind_Exception *);
	267	_Unwind_Reason_Code _Unwind_ForcedUnwind(_Unwind_Exception *, _Unwind_Stop_Fn,
	268	void *);
	269	void _Unwind_DeleteException(_Unwind_Exception *);
	270	void _Unwind_Resume(_Unwind_Exception *);
	271	_Unwind_Reason_Code _Unwind_Resume_or_Rethrow(_Unwind_Exception *);
234	272
235	273	#endif
236	274
...	...	@@ -241,11 +279,11 @@ typedef struct SjLj_Function_Context *_Unwind_FunctionContext_t;
241	279
242	280	void _Unwind_SjLj_Register(_Unwind_FunctionContext_t);
243	281	void _Unwind_SjLj_Unregister(_Unwind_FunctionContext_t);
244		_Unwind_Reason_Code _Unwind_SjLj_RaiseException(struct _Unwind_Exception *);
245		_Unwind_Reason_Code _Unwind_SjLj_ForcedUnwind(struct _Unwind_Exception *,
	282	_Unwind_Reason_Code _Unwind_SjLj_RaiseException(_Unwind_Exception *);
	283	_Unwind_Reason_Code _Unwind_SjLj_ForcedUnwind(_Unwind_Exception *,
246	284	_Unwind_Stop_Fn, void *);
247		void _Unwind_SjLj_Resume(struct _Unwind_Exception *);
248		_Unwind_Reason_Code _Unwind_SjLj_Resume_or_Rethrow(struct _Unwind_Exception *);
	285	void _Unwind_SjLj_Resume(_Unwind_Exception *);
	286	_Unwind_Reason_Code _Unwind_SjLj_Resume_or_Rethrow(_Unwind_Exception *);
249	287
250	288	void _Unwind_FindEnclosingFunction(void );
251	289

ci/appveyor/build_script.bat+2-2

...	...	@@ -7,13 +7,13 @@ SET "PATH=C:\msys64\mingw64\bin;C:\msys64\usr\bin;%PATH%"
7	7	SET "MSYSTEM=MINGW64"
8	8	SET "APPVEYOR_CACHE_ENTRY_ZIP_ARGS=-m0=Copy"
9	9
10		bash -lc "cd ${APPVEYOR_BUILD_FOLDER} && if [ -s ""llvm+clang-5.0.0-win64-msvc-release.tar.xz"" ]; then echo 'skipping LLVM download'; else wget 'https://s3.amazonaws.com/superjoe/temp/llvm%%2bclang-5.0.0-win64-msvc-release.tar.xz'; fi && tar xf llvm+clang-5.0.0-win64-msvc-release.tar.xz" \|\| exit /b
	10	bash -lc "cd ${APPVEYOR_BUILD_FOLDER} && if [ -s ""llvm+clang-6.0.0-win64-msvc-release.tar.xz"" ]; then echo 'skipping LLVM download'; else wget 'https://s3.amazonaws.com/superjoe/temp/llvm%%2bclang-6.0.0-win64-msvc-release.tar.xz'; fi && tar xf llvm+clang-6.0.0-win64-msvc-release.tar.xz" \|\| exit /b
11	11
12	12
13	13	SET "PATH=%PREVPATH%"
14	14	SET "MSYSTEM=%PREVMSYSTEM%"
15	15	SET "ZIGBUILDDIR=%APPVEYOR_BUILD_FOLDER%\build-msvc-release"
16		SET "ZIGPREFIXPATH=%APPVEYOR_BUILD_FOLDER%\llvm+clang-5.0.0-win64-msvc-release"
	16	SET "ZIGPREFIXPATH=%APPVEYOR_BUILD_FOLDER%\llvm+clang-6.0.0-win64-msvc-release"
17	17
18	18	mkdir %ZIGBUILDDIR%
19	19	cd %ZIGBUILDDIR%

ci/travis_linux_install+1-1

...	...	@@ -4,4 +4,4 @@ set -x
4	4
5	5	sudo apt-get remove -y llvm-*
6	6	sudo rm -rf /usr/local/*
7		sudo apt-get install -y clang-5.0 libclang-5.0 libclang-5.0-dev llvm-5.0 llvm-5.0-dev liblld-5.0 liblld-5.0-dev cmake wine1.6-amd64
	7	sudo apt-get install -y clang-6.0 libclang-6.0 libclang-6.0-dev llvm-6.0 llvm-6.0-dev liblld-6.0 liblld-6.0-dev cmake wine1.6-amd64

cmake/Findclang.cmake+4-4

...	...	@@ -26,16 +26,16 @@ if(MSVC)
26	26	else()
27	27	find_path(CLANG_INCLUDE_DIRS NAMES clang/Frontend/ASTUnit.h
28	28	PATHS
29		/usr/lib/llvm/5/include
30		/usr/lib/llvm-5.0/include
	29	/usr/lib/llvm/6/include
	30	/usr/lib/llvm-6.0/include
31	31	/mingw64/include)
32	32
33	33	macro(FIND_AND_ADD_CLANG_LIB _libname_)
34	34	string(TOUPPER ${_libname_} _prettylibname_)
35	35	find_library(CLANG_${_prettylibname_}_LIB NAMES ${_libname_}
36	36	PATHS
37		/usr/lib/llvm/5/lib
38		/usr/lib/llvm-5.0/lib
	37	/usr/lib/llvm/6/lib
	38	/usr/lib/llvm-6.0/lib
39	39	/mingw64/lib
40	40	/c/msys64/mingw64/lib
41	41	c:\\msys64\\mingw64\\lib)

cmake/Findlld.cmake+6-5

...	...	@@ -6,12 +6,12 @@
6	6	# LLD_INCLUDE_DIRS
7	7	# LLD_LIBRARIES
8	8
9		find_path(LLD_INCLUDE_DIRS NAMES lld/Driver/Driver.h
	9	find_path(LLD_INCLUDE_DIRS NAMES lld/Common/Driver.h
10	10	PATHS
11		/usr/lib/llvm-5.0/include
	11	/usr/lib/llvm-6.0/include
12	12	/mingw64/include)
13	13
14		find_library(LLD_LIBRARY NAMES lld-5.0 lld PATHS /usr/lib/llvm-5.0/lib)
	14	find_library(LLD_LIBRARY NAMES lld-6.0 lld PATHS /usr/lib/llvm-6.0/lib)
15	15	if(EXISTS ${LLD_LIBRARY})
16	16	set(LLD_LIBRARIES ${LLD_LIBRARY})
17	17	else()
...	...	@@ -19,7 +19,7 @@ else()
19	19	string(TOUPPER ${_libname_} _prettylibname_)
20	20	find_library(LLD_${_prettylibname_}_LIB NAMES ${_libname_}
21	21	PATHS
22		/usr/lib/llvm-5.0/lib
	22	/usr/lib/llvm-6.0/lib
23	23	/mingw64/lib
24	24	/c/msys64/mingw64/lib
25	25	c:/msys64/mingw64/lib)
...	...	@@ -29,13 +29,14 @@ else()
29	29	endmacro(FIND_AND_ADD_LLD_LIB)
30	30
31	31	FIND_AND_ADD_LLD_LIB(lldDriver)
	32	FIND_AND_ADD_LLD_LIB(lldMinGW)
32	33	FIND_AND_ADD_LLD_LIB(lldELF)
33	34	FIND_AND_ADD_LLD_LIB(lldCOFF)
34	35	FIND_AND_ADD_LLD_LIB(lldMachO)
35	36	FIND_AND_ADD_LLD_LIB(lldReaderWriter)
36	37	FIND_AND_ADD_LLD_LIB(lldCore)
37	38	FIND_AND_ADD_LLD_LIB(lldYAML)
38		FIND_AND_ADD_LLD_LIB(lldConfig)
	39	FIND_AND_ADD_LLD_LIB(lldCommon)
39	40	endif()
40	41
41	42	include(FindPackageHandleStandardArgs)

cmake/Findllvm.cmake+3-3

...	...	@@ -8,12 +8,12 @@
8	8	# LLVM_LIBDIRS
9	9
10	10	find_program(LLVM_CONFIG_EXE
11		NAMES llvm-config-5.0 llvm-config
	11	NAMES llvm-config-6.0 llvm-config
12	12	PATHS
13	13	"/mingw64/bin"
14	14	"/c/msys64/mingw64/bin"
15	15	"c:/msys64/mingw64/bin"
16		"C:/Libraries/llvm-5.0.0/bin")
	16	"C:/Libraries/llvm-6.0.0/bin")
17	17
18	18	if(NOT(CMAKE_BUILD_TYPE STREQUAL "Debug"))
19	19	execute_process(
...	...	@@ -62,7 +62,7 @@ execute_process(
62	62	set(LLVM_LIBRARIES ${LLVM_LIBRARIES} ${LLVM_SYSTEM_LIBS})
63	63
64	64	if(NOT LLVM_LIBRARIES)
65		find_library(LLVM_LIBRARIES NAMES LLVM LLVM-5.0 LLVM-5)
	65	find_library(LLVM_LIBRARIES NAMES LLVM LLVM-6.0 LLVM-6)
66	66	endif()
67	67
68	68	link_directories("${CMAKE_PREFIX_PATH}/lib")

src/parsec.cpp+3

...	...	@@ -570,6 +570,8 @@ static AstNode trans_type(Context c, const Type *ty, const SourceLocation &sou
570	570	return trans_create_node_symbol_str(c, "f64");
571	571	case BuiltinType::Float128:
572	572	return trans_create_node_symbol_str(c, "f128");
	573	case BuiltinType::Float16:
	574	return trans_create_node_symbol_str(c, "f16");
573	575	case BuiltinType::LongDouble:
574	576	return trans_create_node_symbol_str(c, "c_longdouble");
575	577	case BuiltinType::WChar_U:
...	...	@@ -856,6 +858,7 @@ static AstNode trans_type(Context c, const Type *ty, const SourceLocation &sou
856	858	case Type::Pipe:
857	859	case Type::ObjCTypeParam:
858	860	case Type::DeducedTemplateSpecialization:
	861	case Type::DependentAddressSpace:
859	862	emit_warning(c, source_loc, "unsupported type: '%s'", ty->getTypeClassName());
860	863	return nullptr;
861	864	}

src/target.cpp+25-1

...	...	@@ -13,6 +13,7 @@
13	13	#include <stdio.h>
14	14
15	15	static const ArchType arch_list[] = {
	16	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v8_3a},
16	17	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v8_2a},
17	18	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v8_1a},
18	19	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v8},
...	...	@@ -33,9 +34,30 @@ static const ArchType arch_list[] = {
33	34	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v5te},
34	35	{ZigLLVM_arm, ZigLLVM_ARMSubArch_v4t},
35	36
36		{ZigLLVM_armeb, ZigLLVM_NoSubArch},
	37	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8_3a},
	38	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8_2a},
	39	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8_1a},
	40	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8},
	41	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8r},
	42	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8m_baseline},
	43	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v8m_mainline},
	44	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7},
	45	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7em},
	46	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7m},
	47	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7s},
	48	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7k},
	49	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v7ve},
	50	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v6},
	51	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v6m},
	52	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v6k},
	53	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v6t2},
	54	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v5},
	55	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v5te},
	56	{ZigLLVM_armeb, ZigLLVM_ARMSubArch_v4t},
	57
37	58	{ZigLLVM_aarch64, ZigLLVM_NoSubArch},
38	59	{ZigLLVM_aarch64_be, ZigLLVM_NoSubArch},
	60	{ZigLLVM_arc, ZigLLVM_NoSubArch},
39	61	{ZigLLVM_avr, ZigLLVM_NoSubArch},
40	62	{ZigLLVM_bpfel, ZigLLVM_NoSubArch},
41	63	{ZigLLVM_bpfeb, ZigLLVM_NoSubArch},
...	...	@@ -345,6 +367,7 @@ void resolve_target_object_format(ZigTarget *target) {
345	367	case ZigLLVM_amdil:
346	368	case ZigLLVM_amdil64:
347	369	case ZigLLVM_armeb:
	370	case ZigLLVM_arc:
348	371	case ZigLLVM_avr:
349	372	case ZigLLVM_bpfeb:
350	373	case ZigLLVM_bpfel:
...	...	@@ -407,6 +430,7 @@ static int get_arch_pointer_bit_width(ZigLLVM_ArchType arch) {
407	430	case ZigLLVM_msp430:
408	431	return 16;
409	432
	433	case ZigLLVM_arc:
410	434	case ZigLLVM_arm:
411	435	case ZigLLVM_armeb:
412	436	case ZigLLVM_hexagon:

src/zig_llvm.cpp+6-2

...	...	@@ -37,7 +37,7 @@
37	37	#include <llvm/Transforms/IPO/AlwaysInliner.h>
38	38	#include <llvm/Transforms/Scalar.h>
39	39
40		#include <lld/Driver/Driver.h>
	40	#include <lld/Common/Driver.h>
41	41
42	42	using namespace llvm;
43	43
...	...	@@ -605,11 +605,13 @@ static_assert((Triple::ArchType)ZigLLVM_LastArchType == Triple::LastArchType, ""
605	605	static_assert((Triple::VendorType)ZigLLVM_LastVendorType == Triple::LastVendorType, "");
606	606	static_assert((Triple::OSType)ZigLLVM_LastOSType == Triple::LastOSType, "");
607	607	static_assert((Triple::EnvironmentType)ZigLLVM_LastEnvironmentType == Triple::LastEnvironmentType, "");
	608	static_assert((Triple::SubArchType)ZigLLVM_KalimbaSubArch_v5 == Triple::KalimbaSubArch_v5, "");
608	609
609	610	static_assert((Triple::ObjectFormatType)ZigLLVM_UnknownObjectFormat == Triple::UnknownObjectFormat, "");
610	611	static_assert((Triple::ObjectFormatType)ZigLLVM_COFF == Triple::COFF, "");
611	612	static_assert((Triple::ObjectFormatType)ZigLLVM_ELF == Triple::ELF, "");
612	613	static_assert((Triple::ObjectFormatType)ZigLLVM_MachO == Triple::MachO, "");
	614	static_assert((Triple::ObjectFormatType)ZigLLVM_Wasm == Triple::Wasm, "");
613	615
614	616	const char *ZigLLVMGetArchTypeName(ZigLLVM_ArchType arch) {
615	617	return (const char*)Triple::getArchTypeName((Triple::ArchType)arch).bytes_begin();
...	...	@@ -648,6 +650,8 @@ const char *ZigLLVMGetSubArchTypeName(ZigLLVM_SubArchType sub_arch) {
648	650	switch (sub_arch) {
649	651	case ZigLLVM_NoSubArch:
650	652	return "(none)";
	653	case ZigLLVM_ARMSubArch_v8_3a:
	654	return "v8_3a";
651	655	case ZigLLVM_ARMSubArch_v8_2a:
652	656	return "v8_2a";
653	657	case ZigLLVM_ARMSubArch_v8_1a:
...	...	@@ -775,7 +779,7 @@ bool ZigLLDLink(ZigLLVM_ObjectFormatType oformat, const char **args, size_t arg_
775	779	zig_unreachable();
776	780
777	781	case ZigLLVM_COFF:
778		return lld::coff::link(array_ref_args);
	782	return lld::coff::link(array_ref_args, false);
779	783
780	784	case ZigLLVM_ELF:
781	785	return lld::elf::link(array_ref_args, false, diag);

src/zig_llvm.hpp+3

...	...	@@ -177,6 +177,7 @@ enum ZigLLVM_ArchType {
177	177	ZigLLVM_armeb, // ARM (big endian): armeb
178	178	ZigLLVM_aarch64, // AArch64 (little endian): aarch64
179	179	ZigLLVM_aarch64_be, // AArch64 (big endian): aarch64_be
	180	ZigLLVM_arc, // ARC: Synopsys ARC
180	181	ZigLLVM_avr, // AVR: Atmel AVR microcontroller
181	182	ZigLLVM_bpfel, // eBPF or extended BPF or 64-bit BPF (little endian)
182	183	ZigLLVM_bpfeb, // eBPF or extended BPF or 64-bit BPF (big endian)
...	...	@@ -229,6 +230,7 @@ enum ZigLLVM_ArchType {
229	230	enum ZigLLVM_SubArchType {
230	231	ZigLLVM_NoSubArch,
231	232
	233	ZigLLVM_ARMSubArch_v8_3a,
232	234	ZigLLVM_ARMSubArch_v8_2a,
233	235	ZigLLVM_ARMSubArch_v8_1a,
234	236	ZigLLVM_ARMSubArch_v8,
...	...	@@ -318,6 +320,7 @@ enum ZigLLVM_EnvironmentType {
318	320	ZigLLVM_UnknownEnvironment,
319	321
320	322	ZigLLVM_GNU,
	323	ZigLLVM_GNUABIN32,
321	324	ZigLLVM_GNUABI64,
322	325	ZigLLVM_GNUEABI,
323	326	ZigLLVM_GNUEABIHF,

std/special/compiler_rt/index.zig+18-1

...	...	@@ -93,7 +93,24 @@ fn isArmArch() -> bool {
93	93	builtin.Arch.armv5,
94	94	builtin.Arch.armv5te,
95	95	builtin.Arch.armv4t,
96		builtin.Arch.armeb => true,
	96	builtin.Arch.armebv8_2a,
	97	builtin.Arch.armebv8_1a,
	98	builtin.Arch.armebv8,
	99	builtin.Arch.armebv8r,
	100	builtin.Arch.armebv8m_baseline,
	101	builtin.Arch.armebv8m_mainline,
	102	builtin.Arch.armebv7,
	103	builtin.Arch.armebv7em,
	104	builtin.Arch.armebv7m,
	105	builtin.Arch.armebv7s,
	106	builtin.Arch.armebv7k,
	107	builtin.Arch.armebv6,
	108	builtin.Arch.armebv6m,
	109	builtin.Arch.armebv6k,
	110	builtin.Arch.armebv6t2,
	111	builtin.Arch.armebv5,
	112	builtin.Arch.armebv5te,
	113	builtin.Arch.armebv4t => true,
97	114	else => false,
98	115	};
99	116	}