liblzma: Add experimental ARM64 BCJ filter with a temporary Filter ID.

That is, the Filter ID will be changed once the design is final.
The current version will be removed. So files created with the
tempoary Filter ID won't be supported in the future.

CMakeLists.txt

@@ -86,6 +86,7 @@ add_compile_definitions(
     HAVE_DECODERS
     HAVE_DECODER_ARM
     HAVE_DECODER_ARMTHUMB
+    HAVE_DECODER_ARM64
     HAVE_DECODER_DELTA
     HAVE_DECODER_IA64
     HAVE_DECODER_LZMA1
@@ -96,6 +97,7 @@ add_compile_definitions(
     HAVE_ENCODERS
     HAVE_ENCODER_ARM
     HAVE_ENCODER_ARMTHUMB
+    HAVE_ENCODER_ARM64
     HAVE_ENCODER_DELTA
     HAVE_ENCODER_IA64
     HAVE_ENCODER_LZMA1
@@ -331,6 +333,7 @@ add_library(liblzma
     src/liblzma/rangecoder/range_encoder.h
     src/liblzma/simple/arm.c
     src/liblzma/simple/armthumb.c
+    src/liblzma/simple/arm64.c
     src/liblzma/simple/ia64.c
     src/liblzma/simple/powerpc.c
     src/liblzma/simple/simple_coder.c

configure.ac

@@ -79,8 +79,8 @@ fi
 # Filters #
 ###########
 
-m4_define([SUPPORTED_FILTERS], [lzma1,lzma2,delta,x86,powerpc,ia64,arm,armthumb,sparc])dnl
+m4_define([SUPPORTED_FILTERS], [lzma1,lzma2,delta,x86,powerpc,ia64,arm,armthumb,arm64,sparc])dnl
-m4_define([SIMPLE_FILTERS], [x86,powerpc,ia64,arm,armthumb,sparc])
+m4_define([SIMPLE_FILTERS], [x86,powerpc,ia64,arm,armthumb,arm64,sparc])
 m4_define([LZ_FILTERS], [lzma1,lzma2])
 
 m4_foreach([NAME], [SUPPORTED_FILTERS],

src/liblzma/api/lzma/bcj.h

@@ -49,9 +49,16 @@
 	 * Filter for SPARC binaries.
 	 */
 
+#define LZMA_FILTER_ARM64       LZMA_VLI_C(0x3FDB87B33B27000B)
+       /**<
+        * Filter for ARM64 binaries.
+        *
+        * \note         Unlike the older filters above, this doesn't
+        *               support any options (must be NULL).
+        */
 
 /**
- * \brief       Options for BCJ filters
+ * \brief       Options for BCJ filters (except ARM64)
  *
  * The BCJ filters never change the size of the data. Specifying options
  * for them is optional: if pointer to options is NULL, default value is
@@ -88,3 +95,29 @@ typedef struct {
 	uint32_t start_offset;
 
 } lzma_options_bcj;
+
+/**
+ * \brief       Options for the ARM64 filter
+ *
+ * This filter never changes the size of the data.
+ * Specifying options is mandatory.
+ */
+typedef struct {
+	/**
+	 * \brief       How wide range of relative addresses are converted
+	 *
+	 * The ARM64 BL instruction has 26-bit immediate field that encodes
+	 * a relative address as a multiple of four bytes, so the effective
+	 * range is 2^28 bytes (+/-128 MiB).
+	 *
+	 * If width is 28 bits (LZMA_ARM64_WIDTH_MAX), then all BL
+	 * instructions will be converted. This has a downside of some
+	 * false matches that make compression worse. The best value
+	 * depends on the input file and the differences can be significant;
+	 * with large executables the maximum value is sometimes the best.
+	 */
+	uint32_t width;
+#	define LZMA_ARM64_WIDTH_MIN     18
+#	define LZMA_ARM64_WIDTH_MAX     28
+#	define LZMA_ARM64_WIDTH_DEFAULT 26
+} lzma_options_arm64;

src/liblzma/common/filter_common.c

@@ -97,6 +97,15 @@ static const struct {
 		.changes_size = false,
 	},
 #endif
+#if defined(HAVE_ENCODER_ARM64) || defined(HAVE_DECODER_ARM64)
+       {
+               .id = LZMA_FILTER_ARM64,
+		.options_size = sizeof(lzma_options_arm64),
+               .non_last_ok = true,
+               .last_ok = false,
+               .changes_size = false,
+       },
+#endif
 #if defined(HAVE_ENCODER_SPARC) || defined(HAVE_DECODER_SPARC)
 	{
 		.id = LZMA_FILTER_SPARC,

src/liblzma/common/filter_decoder.c

@@ -99,6 +99,14 @@ static const lzma_filter_decoder decoders[] = {
 		.props_decode = &lzma_simple_props_decode,
 	},
 #endif
+#ifdef HAVE_DECODER_ARM64
+       {
+               .id = LZMA_FILTER_ARM64,
+               .init = &lzma_simple_arm64_decoder_init,
+               .memusage = NULL,
+               .props_decode = &lzma_arm64_props_decode,
+       },
+#endif
 #ifdef HAVE_DECODER_SPARC
 	{
 		.id = LZMA_FILTER_SPARC,

src/liblzma/common/filter_encoder.c

@@ -126,6 +126,17 @@ static const lzma_filter_encoder encoders[] = {
 		.props_encode = &lzma_simple_props_encode,
 	},
 #endif
+#ifdef HAVE_ENCODER_ARM64
+       {
+               .id = LZMA_FILTER_ARM64,
+               .init = &lzma_simple_arm64_encoder_init,
+               .memusage = NULL,
+               .block_size = NULL,
+               .props_size_get = NULL,
+               .props_size_fixed = 1,
+               .props_encode = &lzma_arm64_props_encode,
+       },
+#endif
 #ifdef HAVE_ENCODER_SPARC
 	{
 		.id = LZMA_FILTER_SPARC,

src/liblzma/simple/Makefile.inc

@@ -42,6 +42,10 @@ if COND_FILTER_ARMTHUMB
 liblzma_la_SOURCES += simple/armthumb.c
 endif
 
+if COND_FILTER_ARM64
+liblzma_la_SOURCES += simple/arm64.c
+endif
+
 if COND_FILTER_SPARC
 liblzma_la_SOURCES += simple/sparc.c
 endif

src/liblzma/simple/arm64.c

@@ -0,0 +1,227 @@
+///////////////////////////////////////////////////////////////////////////////
+//
+/// \file       arm64.c
+/// \brief      Filter for ARM64 binaries
+///
+//  Authors:    Lasse Collin
+//              Jia Tan
+//
+//  This file has been put into the public domain.
+//  You can do whatever you want with this file.
+//
+///////////////////////////////////////////////////////////////////////////////
+
+#include "simple_private.h"
+
+#ifdef HAVE_ENCODER_ARM64
+#	include "simple_encoder.h"
+#endif
+
+#ifdef HAVE_DECODER_ARM64
+#	include "simple_decoder.h"
+#endif
+
+
+// In ARM64, there are two main branch instructions.
+// bl - branch and link: Calls a function and stores the return address.
+// b - branch: Jumps to a location, but does not store a return address.
+//
+// After some benchmarking, it was determined that only the bl instruction
+// is beneficial for compression. A majority of the jumps for the b
+// instruction are very small (+/- 0xFF). These are typical for loops
+// and if-statements. Encoding them to their absolute address reduces
+// redundancy since many of the small relative jump values are repeated,
+// but very few of the absolute addresses are.
+//
+// Thus, only the bl instruction will be encoded and decoded.
+// The bl instruction is 32 bits in size. The highest 6 bits contain
+// the opcode (10 0101 == 0x25) and the remaining 26 bits are
+// the immediate value. The immediate is a signed integer that
+// encodes the target address as a multiple of four bytes so
+// the range is +/-128 MiB.
+
+// The 6-bit op code for the bl instruction in ARM64
+#define ARM64_BL_OPCODE 0x25
+
+// Once the 26-bit immediate is multiple by four, the address is 28 bits
+// with the two lowest bits being zero. This mask is used to clear the
+// unwanted bits.
+#define ADDR28_MASK 0x0FFFFFFCU
+
+
+typedef struct {
+	uint32_t sign_bit;
+	uint32_t sign_mask;
+} lzma_simple_arm64;
+
+
+static size_t
+arm64_code(void *simple_ptr, uint32_t now_pos, bool is_encoder,
+		uint8_t *buffer, size_t size)
+{
+	const lzma_simple_arm64 *simple = simple_ptr;
+	const uint32_t sign_bit = simple->sign_bit;
+	const uint32_t sign_mask = simple->sign_mask;
+
+	size_t i;
+	for (i = 0; i + 4 <= size; i += 4) {
+		if ((buffer[i + 3] >> 2) == ARM64_BL_OPCODE) {
+			// Get the relative 28-bit address from
+			// the 26-bit immediate.
+			uint32_t src = read32le(buffer + i);
+			src <<= 2;
+			src &= ADDR28_MASK;
+
+			if ((src & sign_mask) != 0
+					&& (src & sign_mask) != sign_mask)
+				continue;
+
+			// Some files like static libraries or Linux kernel
+			// modules have the immediate value filled with
+			// zeros. Converting these placeholder values would
+			// make compression worse so don't touch them.
+			if (src == 0)
+				continue;
+
+			const uint32_t pc = now_pos + (uint32_t)(i);
+
+			uint32_t dest;
+			if (is_encoder)
+				dest = pc + src;
+			else
+				dest = src - pc;
+
+			dest &= ADDR28_MASK;
+
+			// Sign-extend negative values or unset sign bits
+			// from positive values.
+			if (dest & sign_bit)
+				dest |= sign_mask;
+			else
+				dest &= ~sign_mask;
+
+			assert((dest & sign_mask) == 0
+					|| (dest & sign_mask) == sign_mask);
+
+			// Since also the decoder will ignore src values
+			// of 0, we must ensure that nothing is ever encoded
+			// to 0. This is achieved by encoding such values
+			// as pc instead. When decoding, pc will be first
+			// converted to 0 which we will catch here and fix.
+			if (dest == 0) {
+				// We cannot get here if pc is zero because
+				// then src would need to be zero too but we
+				// already ensured that src != 0.
+				assert((pc & ADDR28_MASK) != 0);
+				dest = is_encoder ? pc : 0U - pc;
+				dest &= ADDR28_MASK;
+
+				if (dest & sign_bit)
+					dest |= sign_mask;
+				else
+					dest &= ~sign_mask;
+			}
+
+			assert((dest & sign_mask) == 0
+					|| (dest & sign_mask) == sign_mask);
+			assert((dest & ~ADDR28_MASK) == 0);
+
+			// Construct and store the modified 32-bit instruction.
+			dest >>= 2;
+			dest |= (uint32_t)ARM64_BL_OPCODE << 26;
+			write32le(buffer + i, dest);
+		}
+	}
+
+	return i;
+}
+
+
+#ifdef HAVE_ENCODER_ARM64
+extern lzma_ret
+lzma_arm64_props_encode(const void *options, uint8_t *out)
+{
+	const lzma_options_arm64 *const opt = options;
+
+	if (opt->width < LZMA_ARM64_WIDTH_MIN
+			|| opt->width > LZMA_ARM64_WIDTH_MAX)
+		return LZMA_OPTIONS_ERROR;
+
+	out[0] = (uint8_t)(opt->width - LZMA_ARM64_WIDTH_MIN);
+	return LZMA_OK;
+}
+#endif
+
+
+#ifdef HAVE_DECODER_ARM64
+extern lzma_ret
+lzma_arm64_props_decode(void **options, const lzma_allocator *allocator,
+		const uint8_t *props, size_t props_size)
+{
+	if (props_size != 1)
+		return LZMA_OPTIONS_ERROR;
+
+	if (props[0] > LZMA_ARM64_WIDTH_MAX - LZMA_ARM64_WIDTH_MIN)
+		return LZMA_OPTIONS_ERROR;
+
+	lzma_options_arm64 *opt = lzma_alloc(sizeof(lzma_options_arm64),
+			allocator);
+	if (opt == NULL)
+		return LZMA_MEM_ERROR;
+
+	opt->width = props[0] + LZMA_ARM64_WIDTH_MIN;
+	*options = opt;
+	return LZMA_OK;
+
+}
+#endif
+
+
+static lzma_ret
+arm64_coder_init(lzma_next_coder *next, const lzma_allocator *allocator,
+		const lzma_filter_info *filters, bool is_encoder)
+{
+	if (filters[0].options == NULL)
+		return LZMA_PROG_ERROR;
+
+	const lzma_options_arm64 *opt = filters[0].options;
+	if (opt->width < LZMA_ARM64_WIDTH_MIN
+			|| opt->width > LZMA_ARM64_WIDTH_MAX)
+		return LZMA_OPTIONS_ERROR;
+
+	const lzma_ret ret = lzma_simple_coder_init(next, allocator, filters,
+			&arm64_code, sizeof(lzma_simple_arm64), 4, 4,
+			is_encoder, false);
+
+	if (ret == LZMA_OK) {
+		lzma_simple_coder *coder = next->coder;
+		lzma_simple_arm64 *simple = coder->simple;
+
+		simple->sign_bit = UINT32_C(1) << (opt->width - 1);
+		simple->sign_mask = (UINT32_C(1) << 28) - simple->sign_bit;
+	}
+
+	return ret;
+}
+
+
+#ifdef HAVE_ENCODER_ARM64
+extern lzma_ret
+lzma_simple_arm64_encoder_init(lzma_next_coder *next,
+		const lzma_allocator *allocator,
+		const lzma_filter_info *filters)
+{
+	return arm64_coder_init(next, allocator, filters, true);
+}
+#endif
+
+
+#ifdef HAVE_DECODER_ARM64
+extern lzma_ret
+lzma_simple_arm64_decoder_init(lzma_next_coder *next,
+		const lzma_allocator *allocator,
+		const lzma_filter_info *filters)
+{
+	return arm64_coder_init(next, allocator, filters, false);
+}
+#endif

src/liblzma/simple/simple_coder.h

@@ -61,6 +61,15 @@ extern lzma_ret lzma_simple_armthumb_decoder_init(lzma_next_coder *next,
 		const lzma_filter_info *filters);
 
 
+extern lzma_ret lzma_simple_arm64_encoder_init(lzma_next_coder *next,
+               const lzma_allocator *allocator,
+               const lzma_filter_info *filters);
+
+extern lzma_ret lzma_simple_arm64_decoder_init(lzma_next_coder *next,
+               const lzma_allocator *allocator,
+               const lzma_filter_info *filters);
+
+
 extern lzma_ret lzma_simple_sparc_encoder_init(lzma_next_coder *next,
 		const lzma_allocator *allocator,
 		const lzma_filter_info *filters);

src/liblzma/simple/simple_decoder.h

@@ -19,4 +19,8 @@ extern lzma_ret lzma_simple_props_decode(
 		void **options, const lzma_allocator *allocator,
 		const uint8_t *props, size_t props_size);
 
+extern lzma_ret lzma_arm64_props_decode(
+		void **options, const lzma_allocator *allocator,
+		const uint8_t *props, size_t props_size);
+
 #endif

src/liblzma/simple/simple_encoder.h

@@ -20,4 +20,6 @@ extern lzma_ret lzma_simple_props_size(uint32_t *size, const void *options);
 
 extern lzma_ret lzma_simple_props_encode(const void *options, uint8_t *out);
 
+extern lzma_ret lzma_arm64_props_encode(const void *options, uint8_t *out);
+
 #endif