• file
• revisions
• annotate
• diff
• comparison
• raw

gfx/cairo/libpixman/src/pixman-x64-mmx-emulation.h@b8a032363ba2 (annotated)

gfx/cairo/libpixman/src/pixman-x64-mmx-emulation.h

Thu, 22 Jan 2015 13:21:57 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Thu, 22 Jan 2015 13:21:57 +0100

branch

TOR_BUG_9701

changeset 15

b8a032363ba2

permissions

-rw-r--r--

Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6

michael@0	1	#ifndef MMX_X64_H_INCLUDED
michael@0	2	#define MMX_X64_H_INCLUDED
michael@0	3
michael@0	4	/* Implementation of x64 MMX substitition functions, before
michael@0	5	* pixman is reimplemented not to use __m64 type on Visual C++
michael@0	6	*
michael@0	7	* Copyright (C)2009 by George Yohng
michael@0	8	* Released in public domain.
michael@0	9	*/
michael@0	10
michael@0	11	#include <intrin.h>
michael@0	12
michael@0	13	#define M64C(a) (*(const __m64 *)(&a))
michael@0	14	#define M64U(a) (*(const unsigned long long *)(&a))
michael@0	15
michael@0	16	__inline __m64
michael@0	17	_m_from_int (int a)
michael@0	18	{
michael@0	19	long long i64 = a;
michael@0	20
michael@0	21	return M64C (i64);
michael@0	22	}
michael@0	23
michael@0	24	__inline __m64
michael@0	25	_mm_setzero_si64 ()
michael@0	26	{
michael@0	27	long long i64 = 0;
michael@0	28
michael@0	29	return M64C (i64);
michael@0	30	}
michael@0	31
michael@0	32	__inline __m64
michael@0	33	_mm_set_pi32 (int i1, int i0)
michael@0	34	{
michael@0	35	unsigned long long i64 = ((unsigned)i0) + (((unsigned long long)(unsigned)i1) << 32);
michael@0	36
michael@0	37	return M64C (i64);
michael@0	38	}
michael@0	39
michael@0	40	__inline void
michael@0	41	_m_empty ()
michael@0	42	{
michael@0	43	}
michael@0	44
michael@0	45	__inline __m64
michael@0	46	_mm_set1_pi16 (short w)
michael@0	47	{
michael@0	48	unsigned long long i64 = ((unsigned long long)(unsigned short)(w)) * 0x0001000100010001ULL;
michael@0	49
michael@0	50	return M64C (i64);
michael@0	51	}
michael@0	52
michael@0	53	__inline int
michael@0	54	_m_to_int (__m64 m)
michael@0	55	{
michael@0	56	return m.m64_i32[0];
michael@0	57	}
michael@0	58
michael@0	59	__inline __m64
michael@0	60	_mm_movepi64_pi64 (__m128i a)
michael@0	61	{
michael@0	62	return M64C (a.m128i_i64[0]);
michael@0	63	}
michael@0	64
michael@0	65	__inline __m64
michael@0	66	_m_pand (__m64 a, __m64 b)
michael@0	67	{
michael@0	68	unsigned long long i64 = M64U (a) & M64U (b);
michael@0	69
michael@0	70	return M64C (i64);
michael@0	71	}
michael@0	72
michael@0	73	__inline __m64
michael@0	74	_m_por (__m64 a, __m64 b)
michael@0	75	{
michael@0	76	unsigned long long i64 = M64U (a) | M64U (b);
michael@0	77
michael@0	78	return M64C (i64);
michael@0	79	}
michael@0	80
michael@0	81	__inline __m64
michael@0	82	_m_pxor (__m64 a, __m64 b)
michael@0	83	{
michael@0	84	unsigned long long i64 = M64U (a) ^ M64U (b);
michael@0	85
michael@0	86	return M64C (i64);
michael@0	87	}
michael@0	88
michael@0	89	__inline __m64
michael@0	90	_m_pmulhuw (__m64 a, __m64 b) /* unoptimized */
michael@0	91	{
michael@0	92	unsigned short d[4] =
michael@0	93	{
michael@0	94	(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0]) >> 16),
michael@0	95	(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]) >> 16),
michael@0	96	(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]) >> 16),
michael@0	97	(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]) >> 16)
michael@0	98	};
michael@0	99
michael@0	100	return M64C (d[0]);
michael@0	101	}
michael@0	102
michael@0	103	__inline __m64
michael@0	104	_m_pmullw2 (__m64 a, __m64 b) /* unoptimized */
michael@0	105	{
michael@0	106	unsigned short d[4] =
michael@0	107	{
michael@0	108	(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0])),
michael@0	109	(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1])),
michael@0	110	(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2])),
michael@0	111	(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))
michael@0	112	};
michael@0	113
michael@0	114	return M64C (d[0]);
michael@0	115	}
michael@0	116
michael@0	117	__inline __m64
michael@0	118	_m_pmullw (__m64 a, __m64 b) /* unoptimized */
michael@0	119	{
michael@0	120	unsigned long long x =
michael@0	121	((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[0]) * b.m64_u16[0]))) +
michael@0	122	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[1]) * b.m64_u16[1]))) << 16) +
michael@0	123	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[2]) * b.m64_u16[2]))) << 32) +
michael@0	124	(((unsigned long long)(unsigned short)((((unsigned)a.m64_u16[3]) * b.m64_u16[3]))) << 48);
michael@0	125
michael@0	126	return M64C (x);
michael@0	127	}
michael@0	128
michael@0	129	__inline __m64
michael@0	130	_m_paddusb (__m64 a, __m64 b) /* unoptimized */
michael@0	131	{
michael@0	132	unsigned long long x = (M64U (a) & 0x00FF00FF00FF00FFULL) +
michael@0	133	(M64U (b) & 0x00FF00FF00FF00FFULL);
michael@0	134
michael@0	135	unsigned long long y = ((M64U (a) >> 8) & 0x00FF00FF00FF00FFULL) +
michael@0	136	((M64U (b) >> 8) & 0x00FF00FF00FF00FFULL);
michael@0	137
michael@0	138	x |= ((x & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
michael@0	139	y |= ((y & 0xFF00FF00FF00FF00ULL) >> 8) * 0xFF;
michael@0	140
michael@0	141	x = (x & 0x00FF00FF00FF00FFULL) | ((y & 0x00FF00FF00FF00FFULL) << 8);
michael@0	142
michael@0	143	return M64C (x);
michael@0	144	}
michael@0	145
michael@0	146	__inline __m64
michael@0	147	_m_paddusw (__m64 a, __m64 b) /* unoptimized */
michael@0	148	{
michael@0	149	unsigned long long x = (M64U (a) & 0x0000FFFF0000FFFFULL) +
michael@0	150	(M64U (b) & 0x0000FFFF0000FFFFULL);
michael@0	151
michael@0	152	unsigned long long y = ((M64U (a) >> 16) & 0x0000FFFF0000FFFFULL) +
michael@0	153	((M64U (b) >> 16) & 0x0000FFFF0000FFFFULL);
michael@0	154
michael@0	155	x |= ((x & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
michael@0	156	y |= ((y & 0xFFFF0000FFFF0000) >> 16) * 0xFFFF;
michael@0	157
michael@0	158	x = (x & 0x0000FFFF0000FFFFULL) | ((y & 0x0000FFFF0000FFFFULL) << 16);
michael@0	159
michael@0	160	return M64C (x);
michael@0	161	}
michael@0	162
michael@0	163	__inline __m64
michael@0	164	_m_pshufw (__m64 a, int n) /* unoptimized */
michael@0	165	{
michael@0	166	unsigned short d[4] =
michael@0	167	{
michael@0	168	a.m64_u16[n & 3],
michael@0	169	a.m64_u16[(n >> 2) & 3],
michael@0	170	a.m64_u16[(n >> 4) & 3],
michael@0	171	a.m64_u16[(n >> 6) & 3]
michael@0	172	};
michael@0	173
michael@0	174	return M64C (d[0]);
michael@0	175	}
michael@0	176
michael@0	177	__inline unsigned char
michael@0	178	sat16 (unsigned short d)
michael@0	179	{
michael@0	180	if (d > 0xFF) return 0xFF;
michael@0	181	else return d & 0xFF;
michael@0	182	}
michael@0	183
michael@0	184	__inline __m64
michael@0	185	_m_packuswb (__m64 m1, __m64 m2) /* unoptimized */
michael@0	186	{
michael@0	187	unsigned char d[8] =
michael@0	188	{
michael@0	189	sat16 (m1.m64_u16[0]),
michael@0	190	sat16 (m1.m64_u16[1]),
michael@0	191	sat16 (m1.m64_u16[2]),
michael@0	192	sat16 (m1.m64_u16[3]),
michael@0	193	sat16 (m2.m64_u16[0]),
michael@0	194	sat16 (m2.m64_u16[1]),
michael@0	195	sat16 (m2.m64_u16[2]),
michael@0	196	sat16 (m2.m64_u16[3])
michael@0	197	};
michael@0	198
michael@0	199	return M64C (d[0]);
michael@0	200	}
michael@0	201
michael@0	202	__inline __m64 _m_punpcklbw (__m64 m1, __m64 m2) /* unoptimized */
michael@0	203	{
michael@0	204	unsigned char d[8] =
michael@0	205	{
michael@0	206	m1.m64_u8[0],
michael@0	207	m2.m64_u8[0],
michael@0	208	m1.m64_u8[1],
michael@0	209	m2.m64_u8[1],
michael@0	210	m1.m64_u8[2],
michael@0	211	m2.m64_u8[2],
michael@0	212	m1.m64_u8[3],
michael@0	213	m2.m64_u8[3],
michael@0	214	};
michael@0	215
michael@0	216	return M64C (d[0]);
michael@0	217	}
michael@0	218
michael@0	219	__inline __m64 _m_punpckhbw (__m64 m1, __m64 m2) /* unoptimized */
michael@0	220	{
michael@0	221	unsigned char d[8] =
michael@0	222	{
michael@0	223	m1.m64_u8[4],
michael@0	224	m2.m64_u8[4],
michael@0	225	m1.m64_u8[5],
michael@0	226	m2.m64_u8[5],
michael@0	227	m1.m64_u8[6],
michael@0	228	m2.m64_u8[6],
michael@0	229	m1.m64_u8[7],
michael@0	230	m2.m64_u8[7],
michael@0	231	};
michael@0	232
michael@0	233	return M64C (d[0]);
michael@0	234	}
michael@0	235
michael@0	236	__inline __m64 _m_psrlwi (__m64 a, int n) /* unoptimized */
michael@0	237	{
michael@0	238	unsigned short d[4] =
michael@0	239	{
michael@0	240	a.m64_u16[0] >> n,
michael@0	241	a.m64_u16[1] >> n,
michael@0	242	a.m64_u16[2] >> n,
michael@0	243	a.m64_u16[3] >> n
michael@0	244	};
michael@0	245
michael@0	246	return M64C (d[0]);
michael@0	247	}
michael@0	248
michael@0	249	__inline __m64 _m_psrlqi (__m64 m, int n)
michael@0	250	{
michael@0	251	unsigned long long x = M64U (m) >> n;
michael@0	252
michael@0	253	return M64C (x);
michael@0	254	}
michael@0	255
michael@0	256	__inline __m64 _m_psllqi (__m64 m, int n)
michael@0	257	{
michael@0	258	unsigned long long x = M64U (m) << n;
michael@0	259
michael@0	260	return M64C (x);
michael@0	261	}
michael@0	262
michael@0	263	#endif /* MMX_X64_H_INCLUDED */