I noticed that LZ4_decompress_generic is sometimes instantiated with
identical set of parameters, or (what's worse) with a subtly different
sets of parameters. For example, LZ4_decompress_fast_withPrefix64k is
instantiated as follows:
return LZ4_decompress_generic(source, dest, 0, originalSize, endOnOutputSize,
full, 0, withPrefix64k, (BYTE*)dest - 64 KB, NULL, 64 KB);
while the equivalent withPrefix64k call in LZ4_decompress_usingDict_generic
passes 0 for the last argument instead of 64 KB. It turns out that there
is no difference in this case: if you change 64 KB to 0 KB in
LZ4_decompress_fast_withPrefix64k, you get the same binary code.
Moreover, because it's been clarified that LZ4_decompress_fast doesn't
check match offsets, it is now obvious that both of these fast/withPrefix64k
instantiations are simply redundant. Exactly because LZ4_decompress_fast
doesn't check offsets, it serves well with any prefixed dictionary.
There's a difference, though, with LZ4_decompress_safe_withPrefix64k.
It also passes 64 KB as the last argument, and if you change that to 0,
as in LZ4_decompress_usingDict_generic, you get a completely different
binary code. It seems that passing 0 enables offset checking:
const int checkOffset = ((safeDecode) && (dictSize < (int)(64 KB)));
However, the resulting code seems to run a bit faster. How come
enabling extra checks can make the code run faster? Curiouser and
curiouser! This needs extra study. Currently I take the view that
the dictSize should be set to non-zero when nothing else will do,
i.e. when passing the external dictionary via dictStart. Otherwise,
lowPrefix betrays just enough information about the dictionary.
* * *
Anyway, with this change, I instantiate all the necessary cases as
functions with distinctive names, which also take fewer arguments and
are therefore less error-prone. I also make the functions non-inline.
(The compiler won't inline the functions because they are used more than
once. Hence I attach LZ4_FORCE_O2_GCC_PPC64LE to the instances while
removing from the callers.) The number of instances is now is reduced
from 18 (safe+fast+partial+4*continue+4*prefix+4*dict+2*prefix64+forceExtDict)
down to 7 (safe+fast+partial+2*prefix+2*dict). The size of the code is
not the only issue here. Separate helper function are much more
amenable to profile-guided optimization: it is enough to profile only
a few basic functions, while the other less-often used functions, such
as LZ4_decompress_*_continue, will benefit automatically.
This is the list of LZ4_decompress* functions in liblz4.so, sorted by size.
Exported functions are marked with a capital T.
$ nm -S lib/liblz4.so |grep -wi T |grep LZ4_decompress |sort -k2
0000000000016260 0000000000000005 T LZ4_decompress_fast_withPrefix64k
0000000000016dc0 0000000000000025 T LZ4_decompress_fast_usingDict
0000000000016d80 0000000000000040 T LZ4_decompress_safe_usingDict
0000000000016d10 000000000000006b T LZ4_decompress_fast_continue
0000000000016c70 000000000000009f T LZ4_decompress_safe_continue
00000000000156c0 000000000000059c T LZ4_decompress_fast
0000000000014a90 00000000000005fa T LZ4_decompress_safe
0000000000015c60 00000000000005fa T LZ4_decompress_safe_withPrefix64k
0000000000002280 00000000000005fa t LZ4_decompress_safe_withSmallPrefix
0000000000015090 000000000000062f T LZ4_decompress_safe_partial
0000000000002880 00000000000008ea t LZ4_decompress_fast_extDict
0000000000016270 0000000000000993 t LZ4_decompress_safe_forceExtDict
It occurred to me that the formula "The last 5 bytes are always
literals", on the list of "assumptions made by the decoder", is
remarkably ambiguous. Suppose the decoder is presented with 5 bytes.
Are they literals? It may seem that the decoder degenerates
to memcpy on short inputs. But of course the answer is no,
so the formula needs some clarification.
Parsing restrictions should be explained as well, otherwise they look
like arbitrary numbers. The 5-byte restriction has been mentioned
recently in connection with the shortcut in LZ4_decompress_generic,
so I add that. The second restriction is left to be explained
by the author.
I also took the liberty to explain that empty inputs "are either
unrepresentable or can be represented with a null byte". This wording
may actually have some merit: it leaves for the implementation,
as opposed to the spec, to decide whether the encoder can compress
empty inputs, and whether the decoder can produce an empty output
(which the implementation should further clarify).
The bug is a read up to 2 bytes past the end of the buffer.
There are three cases for this bug, one for each test case added.
* An empty input causes `token = *ip++` to read one byte too far.
* A one byte input with `(token >> ML_BITS) == RUN_MASK` causes
one extra byte to be read without validation. This could be
combined with the first bug to cause 2 extra bytes to be read.
* The case pointed out in issue #508, where `ip == iend` at the
beginning of the loop after taking the shortcut.
Benchmarks show no regressions on clang or gcc-7 on both my mac
and devserver.
Fixes#508.
The notes about "security guarantee" and "malicious inputs" seemed
a bit non-technical to me, so I took the liberty to tone them down
and instead describe the actual risks in technical terms. Namely,
the function never writes past the end of the output buffer, so
a direct hostile takeover (resulting in arbitrary code execution
soon after the return from the function) is not possible. However,
the application can crash because of reads from unmapped pages.
I also took the liberty to describe what I believe is the only sensible
usage scenario for the function: "This function is only usable if the
originalSize of uncompressed data is known in advance," etc.
The simple change from
`matchIndex+MAX_DISTANCE < current`
towards
`current - matchIndex > MAX_DISTANCE`
is enough to generate a 10% performance drop under clang.
Quite massive.
(I missed as my eyes were concentrated on gcc performance at that time).
The second version is more robust, because it also survives a situation where
`matchIndex > current`
due to overflows.
The first version requires matchIndex to not overflow.
Hence were added `assert()` conditions.
The only case where this can happen is with dictCtx compression,
in the case where the dictionary context is not initialized before loading the dictionary.
So it's enough to always initialize the context while loading the dictionary.
