LZX, LZMS: Annotate unaligned memory accesses in x86 filtering

diff --git a/Makefile.am b/Makefile.am
index 612de3ec66aad130368a48340e9cad3cb2e3d01b..2c94bdcb9288d4daad065fdb5cb070b2f02e5784 100644 (file)
--- a/Makefile.am
+++ b/Makefile.am
@@ -126,6 +126,7 @@ libwim_la_SOURCES =         \
        include/wimlib/textfile.h       \
        include/wimlib/timestamp.h      \
        include/wimlib/types.h          \
+       include/wimlib/unaligned.h      \
        include/wimlib/unix_data.h      \
        include/wimlib/util.h           \
        include/wimlib/version.h        \

diff --git a/NEWS b/NEWS
index 7ade11fda37c1adb330ead2748acd7a292ff55c5..159a0ed9a44801072ade561dd1959935794e7ac9 100644 (file)
--- a/NEWS
+++ b/NEWS
@@ -1,6 +1,8 @@
 Version 1.7.3-BETA:
        Fix for very slow export from solid WIM / ESD files.
 
+       Fix for LZX and LZMS algorithms on non-x86 architectures, such as ARM.
+
        New progress message: WIMLIB_PROGRESS_MSG_HANDLE_ERROR.  Applications
        may use this to treat some types of errors as non-fatal.
 

diff --git a/README b/README
index b64d6e88584fc4e0e8bad7a4348a961f091153f7..e8f5523366263f819f94404cf880ee6b70126089 100644 (file)
--- a/README
+++ b/README
@@ -316,9 +316,8 @@ extraction are implemented separately for Windows, UNIX, and UNIX (NTFS-3g
 mode), to ensure a fast and feature-rich implementation of each platform/mode.
 
 wimlib is mainly used on x86 and x86_64 CPUs, but it should also work on a
-number of other GCC-supported 32-bit or 64-bit architectures.  No assumptions
-are made about endianness, but some code assumes that unaligned memory accesses
-are supported and relatively efficient.
+number of other GCC-supported 32-bit or 64-bit architectures.  It has been
+tested on the ARM architecture.
 
 Currently, gcc and clang are the only supported compilers.  A few nonstandard
 extensions are used in the code.

diff --git a/include/wimlib/unaligned.h b/include/wimlib/unaligned.h
new file mode 100644 (file)
index 0000000..d30c9f2
--- /dev/null
+++ b/include/wimlib/unaligned.h
@@ -0,0 +1,38 @@
+/*
+ * unaligned.h
+ *
+ * Inline functions for unaligned memory accesses.
+ *
+ * The author dedicates this file to the public domain.
+ * You can do whatever you want with this file.
+ */
+
+#ifndef _WIMLIB_UNALIGNED_H
+#define _WIMLIB_UNALIGNED_H
+
+#include "compiler.h"
+#include "endianness.h"
+#include "types.h"
+
+#define DEFINE_UNALIGNED_TYPE(type)                            \
+struct type##_unaligned {                                      \
+       type v;                                                 \
+} _packed_attribute;                                           \
+                                                               \
+static inline type                                             \
+load_##type##_unaligned(const void *p)                         \
+{                                                              \
+       return ((const struct type##_unaligned *)p)->v;         \
+}                                                              \
+                                                               \
+static inline void                                             \
+store_##type##_unaligned(type val, void *p)                    \
+{                                                              \
+       ((struct type##_unaligned *)p)->v = val;                \
+}
+
+DEFINE_UNALIGNED_TYPE(le16);
+DEFINE_UNALIGNED_TYPE(le32);
+DEFINE_UNALIGNED_TYPE(le64);
+
+#endif /* _WIMLIB_UNALIGNED_H */

diff --git a/src/lzms-common.c b/src/lzms-common.c
index b5d4fcd5a15d1593bb35f511a9d2763673924615..95682d53a56490fd382d57b7d779b001e78782b7 100644 (file)
--- a/src/lzms-common.c
+++ b/src/lzms-common.c
@@ -28,6 +28,7 @@
 
 #include "wimlib/endianness.h"
 #include "wimlib/lzms.h"
+#include "wimlib/unaligned.h"
 #include "wimlib/util.h"
 
 #include <pthread.h>
@@ -157,20 +158,20 @@ lzms_maybe_do_x86_translation(u8 data[restrict], s32 i, s32 num_op_bytes,
                if (i - *closest_target_usage_p <= max_trans_offset) {
                        LZMS_DEBUG("Undid x86 translation at position %d "
                                   "(opcode 0x%02x)", i, data[i]);
-                       le32 *p32 = (le32*)&data[i + num_op_bytes];
-                       u32 n = le32_to_cpu(*p32);
-                       *p32 = cpu_to_le32(n - i);
+                       void *p32 = &data[i + num_op_bytes];
+                       u32 n = le32_to_cpu(load_le32_unaligned(p32));
+                       store_le32_unaligned(cpu_to_le32(n - i), p32);
                }
-               pos = i + le16_to_cpu(*(const le16*)&data[i + num_op_bytes]);
+               pos = i + le16_to_cpu(load_le16_unaligned(&data[i + num_op_bytes]));
        } else {
-               pos = i + le16_to_cpu(*(const le16*)&data[i + num_op_bytes]);
+               pos = i + le16_to_cpu(load_le16_unaligned(&data[i + num_op_bytes]));
 
                if (i - *closest_target_usage_p <= max_trans_offset) {
                        LZMS_DEBUG("Did x86 translation at position %d "
                                   "(opcode 0x%02x)", i, data[i]);
-                       le32 *p32 = (le32*)&data[i + num_op_bytes];
-                       u32 n = le32_to_cpu(*p32);
-                       *p32 = cpu_to_le32(n + i);
+                       void *p32 = &data[i + num_op_bytes];
+                       u32 n = le32_to_cpu(load_le32_unaligned(p32));
+                       store_le32_unaligned(cpu_to_le32(n + i), p32);
                }
        }
 

diff --git a/src/lzx-common.c b/src/lzx-common.c
index dac43badd9c8d50ea15c2ee877c11e50ea05e3d8..9f55f171f220689bb61dc9e50006507a210ad66c 100644 (file)
--- a/src/lzx-common.c
+++ b/src/lzx-common.c
@@ -25,6 +25,7 @@
 
 #include "wimlib/endianness.h"
 #include "wimlib/lzx.h"
+#include "wimlib/unaligned.h"
 #include "wimlib/util.h"
 
 #ifdef __SSE2__
@@ -120,12 +121,11 @@ lzx_get_num_main_syms(unsigned window_order)
 }
 
 static void
-do_translate_target(sle32 *target, s32 input_pos)
+do_translate_target(void *target, s32 input_pos)
 {
        s32 abs_offset, rel_offset;
 
-       /* XXX: This assumes unaligned memory accesses are okay.  */
-       rel_offset = le32_to_cpu(*target);
+       rel_offset = le32_to_cpu(load_le32_unaligned(target));
        if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
                if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
                        /* "good translation" */
@@ -134,30 +134,29 @@ do_translate_target(sle32 *target, s32 input_pos)
                        /* "compensating translation" */
                        abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
                }
-               *target = cpu_to_le32(abs_offset);
+               store_le32_unaligned(cpu_to_le32(abs_offset), target);
        }
 }
 
 static void
-undo_translate_target(sle32 *target, s32 input_pos)
+undo_translate_target(void *target, s32 input_pos)
 {
        s32 abs_offset, rel_offset;
 
-       /* XXX: This assumes unaligned memory accesses are okay.  */
-       abs_offset = le32_to_cpu(*target);
+       abs_offset = le32_to_cpu(load_le32_unaligned(target));
        if (abs_offset >= 0) {
                if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
                        /* "good translation" */
                        rel_offset = abs_offset - input_pos;
 
-                       *target = cpu_to_le32(rel_offset);
+                       store_le32_unaligned(cpu_to_le32(rel_offset), target);
                }
        } else {
                if (abs_offset >= -input_pos) {
                        /* "compensating translation" */
                        rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
 
-                       *target = cpu_to_le32(rel_offset);
+                       store_le32_unaligned(cpu_to_le32(rel_offset), target);
                }
        }
 }
@@ -194,7 +193,7 @@ inline  /* Although inlining the 'process_target' function still speeds up the
           SSE2 case, it bloats the binary more.  */
 #endif
 void
-lzx_e8_filter(u8 *data, u32 size, void (*process_target)(sle32 *, s32))
+lzx_e8_filter(u8 *data, u32 size, void (*process_target)(void *, s32))
 {
 #ifdef __SSE2__
        /* SSE2 vectorized implementation for x86_64.  This speeds up LZX
@@ -249,7 +248,7 @@ lzx_e8_filter(u8 *data, u32 size, void (*process_target)(sle32 *, s32))
 
                                        /* Do (or undo) the e8 translation.  */
                                        u8 *p8 = (u8 *)p128 + bit;
-                                       (*process_target)((sle32 *)(p8 + 1),
+                                       (*process_target)(p8 + 1,
                                                          p8 - data);
 
                                        /* Don't start an e8 translation in the
@@ -279,7 +278,7 @@ lzx_e8_filter(u8 *data, u32 size, void (*process_target)(sle32 *, s32))
                u8 *p8_end = data + size - 10;
                do {
                        if (*p8 == 0xe8) {
-                               (*process_target)((sle32 *)(p8 + 1), p8 - data);
+                               (*process_target)(p8 + 1, p8 - data);
                                p8 += 5;
                        } else {
                                p8++;