xref: /linux/arch/parisc/include/asm/hash.h (revision 95298d63c67673c654c08952672d016212b26054)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _ASM_HASH_H
3 #define _ASM_HASH_H
4 
5 /*
6  * HP-PA only implements integer multiply in the FPU.  However, for
7  * integer multiplies by constant, it has a number of shift-and-add
8  * (but no shift-and-subtract, sigh!) instructions that a compiler
9  * can synthesize a code sequence with.
10  *
11  * Unfortunately, GCC isn't very efficient at using them.  For example
12  * it uses three instructions for "x *= 21" when only two are needed.
13  * But we can find a sequence manually.
14  */
15 
16 #define HAVE_ARCH__HASH_32 1
17 
18 /*
19  * This is a multiply by GOLDEN_RATIO_32 = 0x61C88647 optimized for the
20  * PA7100 pairing rules.  This is an in-order 2-way superscalar processor.
21  * Only one instruction in a pair may be a shift (by more than 3 bits),
22  * but other than that, simple ALU ops (including shift-and-add by up
23  * to 3 bits) may be paired arbitrarily.
24  *
25  * PA8xxx processors also dual-issue ALU instructions, although with
26  * fewer constraints, so this schedule is good for them, too.
27  *
28  * This 6-step sequence was found by Yevgen Voronenko's implementation
29  * of the Hcub algorithm at http://spiral.ece.cmu.edu/mcm/gen.html.
30  */
31 static inline u32 __attribute_const__ __hash_32(u32 x)
32 {
33 	u32 a, b, c;
34 
35 	/*
36 	 * Phase 1: Compute  a = (x << 19) + x,
37 	 * b = (x << 9) + a, c = (x << 23) + b.
38 	 */
39 	a = x << 19;		/* Two shifts can't be paired */
40 	b = x << 9;	a += x;
41 	c = x << 23;	b += a;
42 			c += b;
43 	/* Phase 2: Return (b<<11) + (c<<6) + (a<<3) - c */
44 	b <<= 11;
45 	a += c << 3;	b -= c;
46 	return (a << 3) + b;
47 }
48 
49 #if BITS_PER_LONG == 64
50 
51 #define HAVE_ARCH_HASH_64 1
52 
53 /*
54  * Finding a good shift-and-add chain for GOLDEN_RATIO_64 is tricky,
55  * because available software for the purpose chokes on constants this
56  * large.  (It's mostly designed for compiling FIR filter coefficients
57  * into FPGAs.)
58  *
59  * However, Jason Thong pointed out a work-around.  The Hcub software
60  * (http://spiral.ece.cmu.edu/mcm/gen.html) is designed for *multiple*
61  * constant multiplication, and is good at finding shift-and-add chains
62  * which share common terms.
63  *
64  * Looking at 0x0x61C8864680B583EB in binary:
65  * 0110000111001000100001100100011010000000101101011000001111101011
66  *  \______________/    \__________/       \_______/     \________/
67  *   \____________________________/         \____________________/
68  * you can see the non-zero bits are divided into several well-separated
69  * blocks.  Hcub can find algorithms for those terms separately, which
70  * can then be shifted and added together.
71  *
72  * Dividing the input into 2, 3 or 4 blocks, Hcub can find solutions
73  * with 10, 9 or 8 adds, respectively, making a total of 11 for the
74  * whole number.
75  *
76  * Using just two large blocks, 0xC3910C8D << 31 in the high bits,
77  * and 0xB583EB in the low bits, produces as good an algorithm as any,
78  * and with one more small shift than alternatives.
79  *
80  * The high bits are a larger number and more work to compute, as well
81  * as needing one extra cycle to shift left 31 bits before the final
82  * addition, so they are the critical path for scheduling.  The low bits
83  * can fit into the scheduling slots left over.
84  */
85 
86 
87 /*
88  * This _ASSIGN(dst, src) macro performs "dst = src", but prevents GCC
89  * from inferring anything about the value assigned to "dest".
90  *
91  * This prevents it from mis-optimizing certain sequences.
92  * In particular, gcc is annoyingly eager to combine consecutive shifts.
93  * Given "x <<= 19; y += x; z += x << 1;", GCC will turn this into
94  * "y += x << 19; z += x << 20;" even though the latter sequence needs
95  * an additional instruction and temporary register.
96  *
97  * Because no actual assembly code is generated, this construct is
98  * usefully portable across all GCC platforms, and so can be test-compiled
99  * on non-PA systems.
100  *
101  * In two places, additional unused input dependencies are added.  This
102  * forces GCC's scheduling so it does not rearrange instructions too much.
103  * Because the PA-8xxx is out of order, I'm not sure how much this matters,
104  * but why make it more difficult for the processor than necessary?
105  */
106 #define _ASSIGN(dst, src, ...) asm("" : "=r" (dst) : "0" (src), ##__VA_ARGS__)
107 
108 /*
109  * Multiply by GOLDEN_RATIO_64 = 0x0x61C8864680B583EB using a heavily
110  * optimized shift-and-add sequence.
111  *
112  * Without the final shift, the multiply proper is 19 instructions,
113  * 10 cycles and uses only 4 temporaries.  Whew!
114  *
115  * You are not expected to understand this.
116  */
117 static __always_inline u32 __attribute_const__
118 hash_64(u64 a, unsigned int bits)
119 {
120 	u64 b, c, d;
121 
122 	/*
123 	 * Encourage GCC to move a dynamic shift to %sar early,
124 	 * thereby freeing up an additional temporary register.
125 	 */
126 	if (!__builtin_constant_p(bits))
127 		asm("" : "=q" (bits) : "0" (64 - bits));
128 	else
129 		bits = 64 - bits;
130 
131 	_ASSIGN(b, a*5);	c = a << 13;
132 	b = (b << 2) + a;	_ASSIGN(d, a << 17);
133 	a = b + (a << 1);	c += d;
134 	d = a << 10;		_ASSIGN(a, a << 19);
135 	d = a - d;		_ASSIGN(a, a << 4, "X" (d));
136 	c += b;			a += b;
137 	d -= c;			c += a << 1;
138 	a += c << 3;		_ASSIGN(b, b << (7+31), "X" (c), "X" (d));
139 	a <<= 31;		b += d;
140 	a += b;
141 	return a >> bits;
142 }
143 #undef _ASSIGN	/* We're a widely-used header file, so don't litter! */
144 
145 #endif /* BITS_PER_LONG == 64 */
146 
147 #endif /* _ASM_HASH_H */
148