xref: /freebsd/contrib/xz/src/liblzma/simple/arm64.c (revision 783d3ff6d7fae619db8a7990b8a6387de0c677b5)
1 // SPDX-License-Identifier: 0BSD
2 
3 ///////////////////////////////////////////////////////////////////////////////
4 //
5 /// \file       arm64.c
6 /// \brief      Filter for ARM64 binaries
7 ///
8 /// This converts ARM64 relative addresses in the BL and ADRP immediates
9 /// to absolute values to increase redundancy of ARM64 code.
10 ///
11 /// Converting B or ADR instructions was also tested but it's not useful.
12 /// A majority of the jumps for the B instruction are very small (+/- 0xFF).
13 /// These are typical for loops and if-statements. Encoding them to their
14 /// absolute address reduces redundancy since many of the small relative
15 /// jump values are repeated, but very few of the absolute addresses are.
16 //
17 //  Authors:    Lasse Collin
18 //              Jia Tan
19 //              Igor Pavlov
20 //
21 ///////////////////////////////////////////////////////////////////////////////
22 
23 #include "simple_private.h"
24 
25 
26 static size_t
27 arm64_code(void *simple lzma_attribute((__unused__)),
28 		uint32_t now_pos, bool is_encoder,
29 		uint8_t *buffer, size_t size)
30 {
31 	size_t i;
32 
33 	// Clang 14.0.6 on x86-64 makes this four times bigger and 40 % slower
34 	// with auto-vectorization that is enabled by default with -O2.
35 	// Such vectorization bloat happens with -O2 when targeting ARM64 too
36 	// but performance hasn't been tested.
37 #ifdef __clang__
38 #	pragma clang loop vectorize(disable)
39 #endif
40 	for (i = 0; i + 4 <= size; i += 4) {
41 		uint32_t pc = (uint32_t)(now_pos + i);
42 		uint32_t instr = read32le(buffer + i);
43 
44 		if ((instr >> 26) == 0x25) {
45 			// BL instruction:
46 			// The full 26-bit immediate is converted.
47 			// The range is +/-128 MiB.
48 			//
49 			// Using the full range is helps quite a lot with
50 			// big executables. Smaller range would reduce false
51 			// positives in non-code sections of the input though
52 			// so this is a compromise that slightly favors big
53 			// files. With the full range only six bits of the 32
54 			// need to match to trigger a conversion.
55 			const uint32_t src = instr;
56 			instr = 0x94000000;
57 
58 			pc >>= 2;
59 			if (!is_encoder)
60 				pc = 0U - pc;
61 
62 			instr |= (src + pc) & 0x03FFFFFF;
63 			write32le(buffer + i, instr);
64 
65 		} else if ((instr & 0x9F000000) == 0x90000000) {
66 			// ADRP instruction:
67 			// Only values in the range +/-512 MiB are converted.
68 			//
69 			// Using less than the full +/-4 GiB range reduces
70 			// false positives on non-code sections of the input
71 			// while being excellent for executables up to 512 MiB.
72 			// The positive effect of ADRP conversion is smaller
73 			// than that of BL but it also doesn't hurt so much in
74 			// non-code sections of input because, with +/-512 MiB
75 			// range, nine bits of 32 need to match to trigger a
76 			// conversion (two 10-bit match choices = 9 bits).
77 			const uint32_t src = ((instr >> 29) & 3)
78 					| ((instr >> 3) & 0x001FFFFC);
79 
80 			// With the addition only one branch is needed to
81 			// check the +/- range. This is usually false when
82 			// processing ARM64 code so branch prediction will
83 			// handle it well in terms of performance.
84 			//
85 			//if ((src & 0x001E0000) != 0
86 			// && (src & 0x001E0000) != 0x001E0000)
87 			if ((src + 0x00020000) & 0x001C0000)
88 				continue;
89 
90 			instr &= 0x9000001F;
91 
92 			pc >>= 12;
93 			if (!is_encoder)
94 				pc = 0U - pc;
95 
96 			const uint32_t dest = src + pc;
97 			instr |= (dest & 3) << 29;
98 			instr |= (dest & 0x0003FFFC) << 3;
99 			instr |= (0U - (dest & 0x00020000)) & 0x00E00000;
100 			write32le(buffer + i, instr);
101 		}
102 	}
103 
104 	return i;
105 }
106 
107 
108 static lzma_ret
109 arm64_coder_init(lzma_next_coder *next, const lzma_allocator *allocator,
110 		const lzma_filter_info *filters, bool is_encoder)
111 {
112 	return lzma_simple_coder_init(next, allocator, filters,
113 			&arm64_code, 0, 4, 4, is_encoder);
114 }
115 
116 
117 #ifdef HAVE_ENCODER_ARM64
118 extern lzma_ret
119 lzma_simple_arm64_encoder_init(lzma_next_coder *next,
120 		const lzma_allocator *allocator,
121 		const lzma_filter_info *filters)
122 {
123 	return arm64_coder_init(next, allocator, filters, true);
124 }
125 #endif
126 
127 
128 #ifdef HAVE_DECODER_ARM64
129 extern lzma_ret
130 lzma_simple_arm64_decoder_init(lzma_next_coder *next,
131 		const lzma_allocator *allocator,
132 		const lzma_filter_info *filters)
133 {
134 	return arm64_coder_init(next, allocator, filters, false);
135 }
136 #endif
137