xor_template.c

  1// SPDX-License-Identifier: GPL-2.0-or-later
  2/*
  3 * Copyright (C) 2023 WANG Xuerui <git@xen0n.name>
  4 *
  5 * Template for XOR operations, instantiated in xor_simd.c.
  6 *
  7 * Expected preprocessor definitions:
  8 *
  9 * - LINE_WIDTH
 10 * - XOR_FUNC_NAME(nr)
 11 * - LD_INOUT_LINE(buf)
 12 * - LD_AND_XOR_LINE(buf)
 13 * - ST_LINE(buf)
 14 */
 15
 16void XOR_FUNC_NAME(2)(unsigned long bytes,
 17		      unsigned long * __restrict v1,
 18		      const unsigned long * __restrict v2)
 19{
 20	unsigned long lines = bytes / LINE_WIDTH;
 21
 22	do {
 23		__asm__ __volatile__ (
 24			LD_INOUT_LINE(v1)
 25			LD_AND_XOR_LINE(v2)
 26			ST_LINE(v1)
 27		: : [v1] "r"(v1), [v2] "r"(v2) : "memory"
 28		);
 29
 30		v1 += LINE_WIDTH / sizeof(unsigned long);
 31		v2 += LINE_WIDTH / sizeof(unsigned long);
 32	} while (--lines > 0);
 33}
 34
 35void XOR_FUNC_NAME(3)(unsigned long bytes,
 36		      unsigned long * __restrict v1,
 37		      const unsigned long * __restrict v2,
 38		      const unsigned long * __restrict v3)
 39{
 40	unsigned long lines = bytes / LINE_WIDTH;
 41
 42	do {
 43		__asm__ __volatile__ (
 44			LD_INOUT_LINE(v1)
 45			LD_AND_XOR_LINE(v2)
 46			LD_AND_XOR_LINE(v3)
 47			ST_LINE(v1)
 48		: : [v1] "r"(v1), [v2] "r"(v2), [v3] "r"(v3) : "memory"
 49		);
 50
 51		v1 += LINE_WIDTH / sizeof(unsigned long);
 52		v2 += LINE_WIDTH / sizeof(unsigned long);
 53		v3 += LINE_WIDTH / sizeof(unsigned long);
 54	} while (--lines > 0);
 55}
 56
 57void XOR_FUNC_NAME(4)(unsigned long bytes,
 58		      unsigned long * __restrict v1,
 59		      const unsigned long * __restrict v2,
 60		      const unsigned long * __restrict v3,
 61		      const unsigned long * __restrict v4)
 62{
 63	unsigned long lines = bytes / LINE_WIDTH;
 64
 65	do {
 66		__asm__ __volatile__ (
 67			LD_INOUT_LINE(v1)
 68			LD_AND_XOR_LINE(v2)
 69			LD_AND_XOR_LINE(v3)
 70			LD_AND_XOR_LINE(v4)
 71			ST_LINE(v1)
 72		: : [v1] "r"(v1), [v2] "r"(v2), [v3] "r"(v3), [v4] "r"(v4)
 73		: "memory"
 74		);
 75
 76		v1 += LINE_WIDTH / sizeof(unsigned long);
 77		v2 += LINE_WIDTH / sizeof(unsigned long);
 78		v3 += LINE_WIDTH / sizeof(unsigned long);
 79		v4 += LINE_WIDTH / sizeof(unsigned long);
 80	} while (--lines > 0);
 81}
 82
 83void XOR_FUNC_NAME(5)(unsigned long bytes,
 84		      unsigned long * __restrict v1,
 85		      const unsigned long * __restrict v2,
 86		      const unsigned long * __restrict v3,
 87		      const unsigned long * __restrict v4,
 88		      const unsigned long * __restrict v5)
 89{
 90	unsigned long lines = bytes / LINE_WIDTH;
 91
 92	do {
 93		__asm__ __volatile__ (
 94			LD_INOUT_LINE(v1)
 95			LD_AND_XOR_LINE(v2)
 96			LD_AND_XOR_LINE(v3)
 97			LD_AND_XOR_LINE(v4)
 98			LD_AND_XOR_LINE(v5)
 99			ST_LINE(v1)
100		: : [v1] "r"(v1), [v2] "r"(v2), [v3] "r"(v3), [v4] "r"(v4),
101		    [v5] "r"(v5) : "memory"
102		);
103
104		v1 += LINE_WIDTH / sizeof(unsigned long);
105		v2 += LINE_WIDTH / sizeof(unsigned long);
106		v3 += LINE_WIDTH / sizeof(unsigned long);
107		v4 += LINE_WIDTH / sizeof(unsigned long);
108		v5 += LINE_WIDTH / sizeof(unsigned long);
109	} while (--lines > 0);
110}