In the block SHA-1 code, we have special assembly code for i386 and
amd64 to perform rotations with assembly. This is supposed to help pick
the correct rotation operation depending on which rotation is smaller,
which can help some systems perform slightly better, since any circular
rotation can be specified as either a rotate left or a rotate right.
However, this isn't needed, so we should remove it.
First, SHA-1, like SHA-2, uses fixed constant rotates. Thus, all
rotation amounts are known at compile time and are in fact baked into
the code. Fortunately, peephole optimizers recognize rotations
specified in the normal way and automatically emit the correct code,
including a preference for choosing a rotate left versus a rotate right.
This has been the case for well over a decade, and is a standard example
of the utility of a peephole optimizer.
Moreover, all modern CPUs, with the exception of extremely limited
embedded CPUs such as some Cortex-M processors, provide a barrel
shifter, which lets the CPU perform rotates of any bit amount in
constant time. This is valuable for many cryptographic algorithms to
improve performance, and is required to prevent timing attacks in
algorithms which use data-dependent rotations (which don't include the
hash algorithms we use). As a result, even though the compiler does the
correct optimization, it isn't even needed here and either a left or a
right rotate is equally acceptable.
In fact, the SHA-256 code already takes this into account and just
writes the simple code using an inline function to let the compiler
optimize it for us.
The downside of using this code, however, is that it uses a GCC
extension, which makes the compiler complain when using -pedantic unless
it's prefixed with __extension__. We could fix that, but since it's
not needed, let's just remove it. We haven't noticed this because
almost everyone uses the SHA1DC code instead, but it still shows up for
some people.
Signed-off-by: brian m. carlson <sandals@crustytoothpaste.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
23119ffb4e (block-sha1: put expanded macro parameters in parentheses,
2012-07-22) added a trailing semicolon to the definition of SHA_MIX
without explanation. It doesn't matter with the current code, but make
sure to avoid potential surprises by removing it again.
This allows the macro to be used almost like a function: Users can
combine it with operators of their choice, but still must not pass an
expression with side-effects as a parameter, as it would be evaluated
multiple times.
Signed-off-by: René Scharfe <l.s.r@web.de>
Reviewed-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The block-sha1 implementation takes an "unsigned long" for the length of
a buffer to hash, but our hash algorithm wrappers take a size_t, as do
other implementations we support like openssl or sha1dc. On many
systems, including Linux, these two are equivalent, but they are not on
Windows (where only a "long long" is 64 bits). As a result, passing
large chunks to a single the_hash_algo->update_fn() would produce wrong
answers there.
Note that we don't need to update any other sizes outside of the
function interface. We store the cumulative size in a "long long" (which
we must do since we hash things bigger than 4GB, like packfiles, even on
32-bit platforms). And internally, we break that size_t len down into
64-byte blocks to feed into the guts of the algorithm.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Borrow the bitmap index into packfiles from JGit to speed up
enumeration of objects involved in a commit range without having to
fully traverse the history.
* jk/pack-bitmap: (26 commits)
ewah: unconditionally ntohll ewah data
ewah: support platforms that require aligned reads
read-cache: use get_be32 instead of hand-rolled ntoh_l
block-sha1: factor out get_be and put_be wrappers
do not discard revindex when re-preparing packfiles
pack-bitmap: implement optional name_hash cache
t/perf: add tests for pack bitmaps
t: add basic bitmap functionality tests
count-objects: recognize .bitmap in garbage-checking
repack: consider bitmaps when performing repacks
repack: handle optional files created by pack-objects
repack: turn exts array into array-of-struct
repack: stop using magic number for ARRAY_SIZE(exts)
pack-objects: implement bitmap writing
rev-list: add bitmap mode to speed up object lists
pack-objects: use bitmaps when packing objects
pack-objects: split add_object_entry
pack-bitmap: add support for bitmap indexes
documentation: add documentation for the bitmap format
ewah: compressed bitmap implementation
...
The BLK_SHA1 code has optimized wrappers for doing endian
conversions on memory that may not be aligned. Let's pull
them out so that we can use them elsewhere, especially the
time-tested list of platforms that prefer each strategy.
Signed-off-by: Jeff King <peff@peff.net>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The code to load a word one-byte-at-a-time was optimized into a
word-wide load instruction even when the pointer was not aligned,
which caused issues on architectures that do not like unaligned
access.
* jn/block-sha1:
Makefile: BLK_SHA1 does not require fast htonl() and unaligned loads
block-sha1: put expanded macro parameters in parentheses
block-sha1: avoid pointer conversion that violates alignment constraints
't' is currently always a numeric constant, but it can't hurt to
prepare for the day that it becomes useful for a caller to pass in a
more complex expression.
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With 660231aa (block-sha1: support for architectures with memory
alignment restrictions, 2009-08-12), blk_SHA1_Update was modified to
access 32-bit chunks of memory one byte at a time on arches that
prefer that:
#define get_be32(p) ( \
(*((unsigned char *)(p) + 0) << 24) | \
(*((unsigned char *)(p) + 1) << 16) | \
(*((unsigned char *)(p) + 2) << 8) | \
(*((unsigned char *)(p) + 3) << 0) )
The code previously accessed these values by just using htonl(*p).
Unfortunately, Michael noticed on an Alpha machine that git was using
plain 32-bit reads anyway. As soon as we convert a pointer to int *,
the compiler can assume that the object pointed to is correctly
aligned as an int (C99 section 6.3.2.3 "pointer conversions"
paragraph 7), and gcc takes full advantage by using a single 32-bit
load, resulting in a whole bunch of unaligned access traps.
So we need to obey the alignment constraints even when only dealing
with pointers instead of actual values. Do so by changing the type
of 'data' to void *. This patch renames 'data' to 'block' at the same
time to make sure all references are updated to reflect the new type.
Reported-tested-and-explained-by: Michael Cree <mcree@orcon.net.nz>
Signed-off-by: Jonathan Nieder <jrnieder@gmail.com>
Acked-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
On Intel machines, the msvc compiler defines the CPU architecture
macros _M_IX86 and _M_X64 (equivalent to __i386__ and __x86_64__
respectively). Use these macros in the pre-processor expression
to select the "fast" definition of the {get,put}_be32() macros.
Signed-off-by: Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Acked-by: Jonathan Nieder <jrnieder@gmail.com>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In particular, using the normal (or production) compiler
warning level (-W3), msvc complains as follows:
.../sha1.c(244) : warning C4018: '<' : signed/unsigned mismatch
.../sha1.c(270) : warning C4244: 'function' : conversion from \
'unsigned __int64' to 'unsigned long', possible loss of data
.../sha1.c(271) : warning C4244: 'function' : conversion from \
'unsigned __int64' to 'unsigned long', possible loss of data
Note that gcc issues a similar complaint about line 244 when
compiling with -Wextra.
Signed-off-by: Ramsay Jones <ramsay@ramsay1.demon.co.uk>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
They are both slower than the new BLK_SHA1 implementation, so it is
pointless to keep them around.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
With this, the code should now be portable to any C compiler.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
We rely on ntohl() and htonl() to perform byte swapping in many places.
However, some platforms have libraries providing really poor
implementations of those which might cause significant performance
issues, especially with the block-sha1 code.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Some compilers produce errors when arithmetic is attempted on pointers to
void. We want computations done on byte addresses, so cast them to char *
to work them around.
Signed-off-by: Brandon Casey <casey@nrlssc.navy.mil>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
In addition to X86, PowerPC and S390 are capable of unaligned memory
accesses.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is needed on architectures with poor or non-existent unaligned memory
support and/or no fast byte swap instruction (such as ARM) by using byte
accesses to memory and shifting the result together.
This also makes the code portable, therefore the byte access methods are
the defaults. Any architecture that properly supports unaligned word
accesses in hardware simply has to enable the alternative methods.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This is to make it easier for them to be selected individually depending
on the architecture instead of the other way around i.e. having each
architecture select a list of hacks up front. That makes for clearer
documentation as well.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Move the code around so specific architecture hacks are defined first.
Also make one line comments actually one line. No code change.
Signed-off-by: Nicolas Pitre <nico@cam.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
For x86 performance (especially in 32-bit mode) I added that hack to write
the SHA1 internal temporary hash using a volatile pointer, in order to get
gcc to not try to cache the array contents. Because gcc will do all the
wrong things, and then spill things in insane random ways.
But on architectures like PPC, where you have 32 registers, it's actually
perfectly reasonable to put the whole temporary array[] into the register
set, and gcc can do so.
So make the 'volatile unsigned int *' cast be dependent on a
SMALL_REGISTER_SET preprocessor symbol, and enable it (currently) on just
x86 and x86-64. With that, the routine is fairly reasonable even when
compared to the hand-scheduled PPC version. Ben Herrenschmidt reports on
a G5:
* Paulus asm version: about 3.67s
* Yours with no change: about 5.74s
* Yours without "volatile": about 3.78s
so with this the C version is within about 3% of the asm one.
And add a lot of commentary on what the heck is going on.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
I think I have found a way to avoid the gcc crazyness.
Lookie here:
# TIME[s] SPEED[MB/s]
rfc3174 5.094 119.8
rfc3174 5.098 119.7
linus 1.462 417.5
linusas 2.008 304
linusas2 1.878 325
mozilla 5.566 109.6
mozillaas 5.866 104.1
openssl 1.609 379.3
spelvin 1.675 364.5
spelvina 1.601 381.3
nettle 1.591 383.6
notice? I outperform all the hand-tuned asm on 32-bit too. By quite a
margin, in fact.
Now, I didn't try a P4, and it's possible that it won't do that there, but
the 32-bit code generation sure looks impressive on my Nehalem box. The
magic? I force the stores to the 512-bit hash bucket to be done in order.
That seems to help a lot.
The diff is trivial (on top of the "rename registers with cpp" patch), as
appended. And it does seem to fix the P4 issues too, although I can
obviously (once again) only test Prescott, and only in 64-bit mode:
# TIME[s] SPEED[MB/s]
rfc3174 1.662 36.73
rfc3174 1.64 37.22
linus 0.2523 241.9
linusas 0.4367 139.8
linusas2 0.4487 136
mozilla 0.9704 62.9
mozillaas 0.9399 64.94
that's some really impressive improvement. All from just saying "do the
stores in the order I told you to, dammit!" to the compiler.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Instead of letting the compiler to figure out the optimal way to rotate
register usage, explicitly rotate the register names with cpp.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
.. and simplify the ctx->size logic.
We now count the size in bytes, which means that 'lenW' was always just
the low 6 bits of the total size, so we don't carry it around separately
any more. And we do the 'size in bits' shift at the end.
Suggested by Nicolas Pitre and linux@horizon.com.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
It's an equivalent expression, but the '+' gives us some freedom in
instruction selection (for example, we can use 'lea' rather than 'add'),
and associates with the other additions around it to give some minor
scheduling freedom.
Suggested-by: linux@horizon.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Avoid repeating the shared parts of the different rounds by adding a
macro layer or two. It was already more cpp than C.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
The mozilla-SHA1 code did this 80-word array for the 80 iterations. But
the SHA1 state is really just 512 bits, and you can actually keep it in
a kind of "circular queue" of just 16 words instead.
This requires us to do the xor updates as we go along (rather than as a
pre-phase), but that's really what we want to do anyway.
This gets me really close to the OpenSSL performance on my Nehalem.
Look ma, all C code (ok, there's the rol/ror hack, but that one doesn't
strictly even matter on my Nehalem, it's just a local optimization).
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
This helps a teeny bit. But what I -really- want to do is to avoid the
whole 80-array loop, and do the xor updates as I go along..
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Bert Wesarg noticed non-x86 version of SHA_ROT() had a typo.
Also spell in-line assembly as __asm__(), otherwise I seem to get
error: implicit declaration of function 'asm' from my compiler.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Use the one with the smaller constant. It _can_ generate slightly
smaller code (a constant of 1 is special), but perhaps more importantly
it's possibly faster on any uarch that does a rotate with a loop.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Undo the change I picked up from the mailing list discussion suggested
by Nico, not because it is wrong, but it will be done at the end of the
follow-up series.
Signed-off-by: Junio C Hamano <gitster@pobox.com>
Based on the mozilla SHA1 routine, but doing the input data accesses a
word at a time and with 'htonl()' instead of loading bytes and shifting.
It requires an architecture that is ok with unaligned 32-bit loads and a
fast htonl().
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Junio C Hamano <gitster@pobox.com>
