Message ID | 20200924192111.GA2528225@coredump.intra.peff.net (mailing list archive) |
---|---|
State | Accepted |
Commit | c578e29ba0791041ad7fabf1166dd6f7e7f26d1f |
Headers | show |
Series | drop unaligned loads | expand |
Am 24.09.20 um 21:21 schrieb Jeff King: > Our put_be32() routine and its variants (get_be32(), put_be64(), etc) > has two implementations: on some platforms we cast memory in place and > use nothl()/htonl(), which can cause unaligned memory access. And on > others, we pick out the individual bytes using bitshifts. > > This introduces extra complexity, and sometimes causes compilers to > generate warnings about type-punning. And it's not clear there's any > performance advantage. > > This split goes back to 660231aa97 (block-sha1: support for > architectures with memory alignment restrictions, 2009-08-12). The > unaligned versions were part of the original block-sha1 code in > d7c208a92e (Add new optimized C 'block-sha1' routines, 2009-08-05), > which says it is: > > 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. > > Back then, Linus provided timings versus the mozilla code which showed a > 27% improvement: > > https://lore.kernel.org/git/alpine.LFD.2.01.0908051545000.3390@localhost.localdomain/ > > However, the unaligned loads were either not the useful part of that > speedup, or perhaps compilers and processors have changed since then. > Here are times for computing the sha1 of 4GB of random data, with and > without -DNO_UNALIGNED_LOADS (and BLK_SHA1=1, of course). This is with > gcc 10, -O2, and the processor is a Core i9-9880H. > > [stock] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] > Range (min … max): 6.550 s … 6.841 s 10 runs > > [-DNO_UNALIGNED_LOADS] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 6.418 s ± 0.015 s [User: 6.058 s, System: 0.360 s] > Range (min … max): 6.394 s … 6.447 s 10 runs > > And here's the same test run on an AMD A8-7600, using gcc 8. > > [stock] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 11.721 s ± 0.113 s [User: 10.761 s, System: 0.951 s] > Range (min … max): 11.509 s … 11.861 s 10 runs > > [-DNO_UNALIGNED_LOADS] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 11.744 s ± 0.066 s [User: 10.807 s, System: 0.928 s] > Range (min … max): 11.637 s … 11.863 s 10 runs Yay, benchmarks! GCC 10.2 with -O2 on an i5-9600K without NO_UNALIGNED_LOADS: Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.547 s ± 0.015 s [User: 6.127 s, System: 0.395 s] Range (min … max): 6.531 s … 6.583 s 10 runs ... and with NO_UNALIGNED_LOADS set: Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.496 s ± 0.011 s [User: 6.135 s, System: 0.360 s] Range (min … max): 6.486 s … 6.519 s 10 runs clang 10 without NO_UNALIGNED_LOADS: Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.697 s ± 0.028 s [User: 6.343 s, System: 0.354 s] Range (min … max): 6.675 s … 6.754 s 10 runs ... and with NO_UNALIGNED_LOADS set: Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.714 s ± 0.049 s [User: 6.320 s, System: 0.375 s] Range (min … max): 6.651 s … 6.791 s 10 runs > +cc René because I know he is going to feed the two of them into > godbolt; I could do that, too, but he will provide much better analysis > on top ;) Weeell, I don't know about that, but I couldn't resist taking a quick look at what some compilers do with the 32-bit functions, which are the ones used in block-sha1: https://www.godbolt.org/z/rhKMTM. Older versions of gcc and clang didn't see through the shifting put_be32() implementation. If you go back further there are also versions that didn't optimize the shifting get_be32(). And the latest icc still can't do that. gcc 10.2 just optimizes all functions to a bswap and a mov. Can't do any better than that, can you? But why do we then see a difference in our benchmark results? Not sure, but https://www.godbolt.org/z/7xh8ao shows that gcc is shuffling some instructions around depending on the implementation. Switch to clang if you want to see more vigorous shuffling. The performance of bigger pieces of code seems to be a matter of luck to some extent. :-/ René
On 2020-09-24 at 19:21:11, Jeff King wrote: > However, the unaligned loads were either not the useful part of that > speedup, or perhaps compilers and processors have changed since then. > Here are times for computing the sha1 of 4GB of random data, with and > without -DNO_UNALIGNED_LOADS (and BLK_SHA1=1, of course). This is with > gcc 10, -O2, and the processor is a Core i9-9880H. > > [stock] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] > Range (min … max): 6.550 s … 6.841 s 10 runs > > [-DNO_UNALIGNED_LOADS] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 6.418 s ± 0.015 s [User: 6.058 s, System: 0.360 s] > Range (min … max): 6.394 s … 6.447 s 10 runs > > And here's the same test run on an AMD A8-7600, using gcc 8. > > [stock] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 11.721 s ± 0.113 s [User: 10.761 s, System: 0.951 s] > Range (min … max): 11.509 s … 11.861 s 10 runs > > [-DNO_UNALIGNED_LOADS] > Benchmark #1: t/helper/test-tool sha1 <foo.rand > Time (mean ± σ): 11.744 s ± 0.066 s [User: 10.807 s, System: 0.928 s] > Range (min … max): 11.637 s … 11.863 s 10 runs I think this is a fine and desirable change, both for performance and correctness. It is, as usual, well explained. > So the unaligned loads don't seem to help much, and actually make things > worse. It's possible there are platforms where they provide more > benefit, but: > > - the non-x86 platforms for which we use this code are old and obscure > (powerpc and s390). I cannot speak for s390, since I have never owned one, but my understanding on unaligned access is that typically there is a tiny penalty on x86 (about a cycle) and a more significant penalty on PowerPC, although that may have changed with newer POWER chips. So my gut tells me this is an improvement either way, although I no longer own any such bootable hardware to measure for certain. Anyway, as René found, the latest versions of GCC already use the peephole optimizer to recognize and optimize this on x86, so I expect they'll do so on other architectures as well. Byte swapping is a pretty common operation. > - the main caller that cares about performance is block-sha1. But > these days it is rarely used anyway, in favor of sha1dc (which is > already much slower, and nobody seems to have cared that much). I think block-sha256 uses it as well, but in any case, it's still faster than sha1dc and people who care desperately about performance will use a crypto library instead.
On Fri, Sep 25, 2020 at 12:02:38AM +0200, René Scharfe wrote: > Older versions of gcc and clang didn't see through the shifting > put_be32() implementation. If you go back further there are also > versions that didn't optimize the shifting get_be32(). And the latest > icc still can't do that. > > gcc 10.2 just optimizes all functions to a bswap and a mov. Can't do > any better than that, can you? > > But why do we then see a difference in our benchmark results? Not sure, > but https://www.godbolt.org/z/7xh8ao shows that gcc is shuffling some > instructions around depending on the implementation. Switch to clang if > you want to see more vigorous shuffling. We do redefine ntohl(), etc in compat/bswap.h. Looking at them, I'd think the result would end up as a bswap instruction, though. And indeed, trying to feed that to godbolt results in the same output you got. It does sound like older compilers were benefiting from the unaligned versions. Building with gcc-4.8 (from debian jessie in a docker container on the same machine), I get ~6.25s with the unaligned load versus ~6.6s with the bit-shifting code. So that's the opposite of how the newer compiler behaves. Benchmarking clang-8 (which your results showed doesn't handle the shifting version well). It likewise is just _slightly_ slower after my patch (11.47s versus 11.57s). Given that newer compilers behave the opposite way, and the overall small magnitude of the impact, I'm still comfortable with the change. It's nice to have a better understanding of how the compiler is impacting it (even if I am still confused how anything at all changes on the newer compilers). Thanks for digging. -Peff
On Thu, Sep 24, 2020 at 6:15 PM brian m. carlson <sandals@crustytoothpaste.net> wrote: > On 2020-09-24 at 19:21:11, Jeff King wrote: > > [stock] > > Benchmark #1: t/helper/test-tool sha1 <foo.rand > > Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] > > Range (min … max): 6.550 s … 6.841 s 10 runs slightly offtopic but what generates this nicely formatted output? > I cannot speak for s390, since I have never owned one I happen to be lucky enough to have access to one (RHEL 8.2/z15, gcc 8.3.1) and seems (third consecutive run): stock: user: 7.555s, system: 1.191s -DNO_UNALIGNED_LOADS: user: 7.561s, system: 1.189s Carlo
On Fri, Sep 25, 2020 at 02:05:09AM -0700, Carlo Arenas wrote: > > > [stock] > > > Benchmark #1: t/helper/test-tool sha1 <foo.rand > > > Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] > > > Range (min … max): 6.550 s … 6.841 s 10 runs > > slightly offtopic but what generates this nicely formatted output? It's this: https://github.com/sharkdp/hyperfine It will actually run both versions and compare them, but it's a little more involved to set up (since you have to do a build step in between). > > I cannot speak for s390, since I have never owned one > > I happen to be lucky enough to have access to one (RHEL 8.2/z15, gcc > 8.3.1) and seems (third consecutive run): > > stock: user: 7.555s, system: 1.191s > -DNO_UNALIGNED_LOADS: user: 7.561s, system: 1.189s Thanks. That's not too surprising. gcc 8 seems to be able to optimize both versions to the same thing (though I have no idea if s390 has a bswap instruction). -Peff
On 2020-09-25 05:05, Carlo Arenas wrote: > On Thu, Sep 24, 2020 at 6:15 PM brian m. carlson > <sandals@crustytoothpaste.net> wrote: >> On 2020-09-24 at 19:21:11, Jeff King wrote: >>> [stock] >>> Benchmark #1: t/helper/test-tool sha1 <foo.rand >>> Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] >>> Range (min … max): 6.550 s … 6.841 s 10 runs > > slightly offtopic but what generates this nicely formatted output? It turns out I may have finally reproduced my diff speed regression at home - just today - and I read this :) I may use it to test, nice timing! About my unrelated regression, TL;DR, have anyone benchmarked under a VM? (more context below...) >> I cannot speak for s390, since I have never owned one > > I happen to be lucky enough to have access to one (RHEL 8.2/z15, gcc > 8.3.1) and seems (third consecutive run): > > stock: user: 7.555s, system: 1.191s > -DNO_UNALIGNED_LOADS: user: 7.561s, system: 1.189s So also somewhat offtopic too, but I've been trying lately to reproduce a speed regression with "git diff --no-ext-diff --quiet" (used by git prompt in contrib) on some large repos between 2.9.5 and 2.16.2. Overall the diff was over 3x as slow where I tested, and I measured a couple timings between mmap of the same two large files between the two version, the delta was approx 10x! (no other syscalls so it's all computing and - if my theory proves right - I guess a lot of pagefaults or similar handling control back to the host). I prepped a VM where I will do more testing (I ran one preliminary test and there seemed to be a visible difference, but I didn't have a proper repo to test with yet). The point being it might be worth comparing the two on VM's as well. Regards, Thomas
diff --git a/Makefile b/Makefile index 92d188fd37..d46765aa9e 100644 --- a/Makefile +++ b/Makefile @@ -1218,7 +1218,6 @@ SANITIZERS := $(foreach flag,$(subst $(comma),$(space),$(SANITIZE)),$(flag)) BASIC_CFLAGS += -fsanitize=$(SANITIZE) -fno-sanitize-recover=$(SANITIZE) BASIC_CFLAGS += -fno-omit-frame-pointer ifneq ($(filter undefined,$(SANITIZERS)),) -BASIC_CFLAGS += -DNO_UNALIGNED_LOADS BASIC_CFLAGS += -DSHA1DC_FORCE_ALIGNED_ACCESS endif ifneq ($(filter leak,$(SANITIZERS)),) diff --git a/compat/bswap.h b/compat/bswap.h index e4e25735ce..c0bb744adc 100644 --- a/compat/bswap.h +++ b/compat/bswap.h @@ -145,28 +145,6 @@ static inline uint64_t git_bswap64(uint64_t x) #endif -/* - * Performance might be improved if the CPU architecture is OK with - * unaligned 32-bit loads and a fast ntohl() is available. - * Otherwise fall back to byte loads and shifts which is portable, - * and is faster on architectures with memory alignment issues. - */ - -#if !defined(NO_UNALIGNED_LOADS) && ( \ - defined(__i386__) || defined(__x86_64__) || \ - defined(_M_IX86) || defined(_M_X64) || \ - defined(__ppc__) || defined(__ppc64__) || \ - defined(__powerpc__) || defined(__powerpc64__) || \ - defined(__s390__) || defined(__s390x__)) - -#define get_be16(p) ntohs(*(unsigned short *)(p)) -#define get_be32(p) ntohl(*(unsigned int *)(p)) -#define get_be64(p) ntohll(*(uint64_t *)(p)) -#define put_be32(p, v) do { *(unsigned int *)(p) = htonl(v); } while (0) -#define put_be64(p, v) do { *(uint64_t *)(p) = htonll(v); } while (0) - -#else - static inline uint16_t get_be16(const void *ptr) { const unsigned char *p = ptr; @@ -212,6 +190,4 @@ static inline void put_be64(void *ptr, uint64_t value) p[7] = value >> 0; } -#endif - #endif /* COMPAT_BSWAP_H */
Our put_be32() routine and its variants (get_be32(), put_be64(), etc) has two implementations: on some platforms we cast memory in place and use nothl()/htonl(), which can cause unaligned memory access. And on others, we pick out the individual bytes using bitshifts. This introduces extra complexity, and sometimes causes compilers to generate warnings about type-punning. And it's not clear there's any performance advantage. This split goes back to 660231aa97 (block-sha1: support for architectures with memory alignment restrictions, 2009-08-12). The unaligned versions were part of the original block-sha1 code in d7c208a92e (Add new optimized C 'block-sha1' routines, 2009-08-05), which says it is: 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. Back then, Linus provided timings versus the mozilla code which showed a 27% improvement: https://lore.kernel.org/git/alpine.LFD.2.01.0908051545000.3390@localhost.localdomain/ However, the unaligned loads were either not the useful part of that speedup, or perhaps compilers and processors have changed since then. Here are times for computing the sha1 of 4GB of random data, with and without -DNO_UNALIGNED_LOADS (and BLK_SHA1=1, of course). This is with gcc 10, -O2, and the processor is a Core i9-9880H. [stock] Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.638 s ± 0.081 s [User: 6.269 s, System: 0.368 s] Range (min … max): 6.550 s … 6.841 s 10 runs [-DNO_UNALIGNED_LOADS] Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 6.418 s ± 0.015 s [User: 6.058 s, System: 0.360 s] Range (min … max): 6.394 s … 6.447 s 10 runs And here's the same test run on an AMD A8-7600, using gcc 8. [stock] Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 11.721 s ± 0.113 s [User: 10.761 s, System: 0.951 s] Range (min … max): 11.509 s … 11.861 s 10 runs [-DNO_UNALIGNED_LOADS] Benchmark #1: t/helper/test-tool sha1 <foo.rand Time (mean ± σ): 11.744 s ± 0.066 s [User: 10.807 s, System: 0.928 s] Range (min … max): 11.637 s … 11.863 s 10 runs So the unaligned loads don't seem to help much, and actually make things worse. It's possible there are platforms where they provide more benefit, but: - the non-x86 platforms for which we use this code are old and obscure (powerpc and s390). - the main caller that cares about performance is block-sha1. But these days it is rarely used anyway, in favor of sha1dc (which is already much slower, and nobody seems to have cared that much). Let's just drop unaligned versions entirely in the name of simplicity. Signed-off-by: Jeff King <peff@peff.net> --- +cc Linus for any thoughts or performance wisdom, and in particular if there might be other platforms worth benchmarking (though unless the improvement is dramatic on some system, avoiding the complexity and pointer aliasing seems worthwhile to me). +cc René because I know he is going to feed the two of them into godbolt; I could do that, too, but he will provide much better analysis on top ;) Makefile | 1 - compat/bswap.h | 24 ------------------------ 2 files changed, 25 deletions(-)