Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
On 17.07.2008, at 16:20, Matt Wilmas wrote: Hi all, Replying to this OLD message about the ZEND_SIGNED_MULTIPLY_LONG() change... It's just a small thing, but it's been on the TODO for a while, so I attempted to correct the issue with my original patch that used long long type, which wasn't available on the older MSVC versions, etc. I created a zend_long64 type (when possible, #if logic copied from snprintf.h's WIDE_INT definition), hopefully it's mostly correct/OK, and maybe it can be used elsewhere too. Since the new memory manager, ..._MULTIPLY_LONG() isn't used for safe_emalloc() anymore, so I made another safe_address() variation (overflow safe, I hope). I noticed ~0.6% slowdown on bench.php, though it seems this zend_ulong64 version should use fewer cycles than the doubles one for sure, if someone wants to analyze it more. :-) Finally, is it even possible to also have a Windows assembly version of these 2 things, where I left the comments? There is for zend_mm_[high|low]_bit(), but they're very simple... http://realplain.com/php/multiply_long.diff http://realplain.com/php/multiply_long_5_3.diff any takers for this patch? otherwise it will probably not make the cut by the 24th regards, Lukas Kahwe Smith [EMAIL PROTECTED] -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Hi all, Replying to this OLD message about the ZEND_SIGNED_MULTIPLY_LONG() change... It's just a small thing, but it's been on the TODO for a while, so I attempted to correct the issue with my original patch that used long long type, which wasn't available on the older MSVC versions, etc. I created a zend_long64 type (when possible, #if logic copied from snprintf.h's WIDE_INT definition), hopefully it's mostly correct/OK, and maybe it can be used elsewhere too. Since the new memory manager, ..._MULTIPLY_LONG() isn't used for safe_emalloc() anymore, so I made another safe_address() variation (overflow safe, I hope). I noticed ~0.6% slowdown on bench.php, though it seems this zend_ulong64 version should use fewer cycles than the doubles one for sure, if someone wants to analyze it more. :-) Finally, is it even possible to also have a Windows assembly version of these 2 things, where I left the comments? There is for zend_mm_[high|low]_bit(), but they're very simple... http://realplain.com/php/multiply_long.diff http://realplain.com/php/multiply_long_5_3.diff - Matt - Original Message - From: Matt Wilmas Sent: Wednesday, November 08, 2006 Hi, Here's an additional ZEND_SIGNED_MULTIPLY_LONG() for platforms with 32-bit longs that don't use the assembly version (so all Windows systems at least?). On my Windows system, mul_function() is 40% faster with this version (no overflow), which makes PHP's * operator 20% faster; with overflow mul_function() is 20% faster (though I don't see any difference with the * operator). The macro is also used in safe_emalloc()... Patch just against 5_2 as the file versions are the only difference. Matt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Did it not work for you as-is on Windows...? It did for me, since it seems there's already a typedef for __int64 - long long. I'd wondered the same, I think some VC versions on windows don't know long long but do know __int64, but I don't remember from the top of my head which version, but I think maybe VC6 might have a problem with that. Otherwise looks fine, so if adjusted for VC6 it's OK. I forgot to mention that I had seen this in main/config.w32.h: /* MSVC.6/NET don't allow 'long long' or know 'intmax_t' */ #define SIZEOF_LONG_LONG_INT 0 #define SIZEOF_LONG_LONG 8 /* defined as __int64 */ SIZEOF_LONG_LONG was changed from 0 to 8 in v1.86 and the comment about __int64 added. So, I also assumed long long could be used as-is because of that. Why have SIZEOF as 8 if it's unusable...? It was changed to fix Bug #34052, I know, but I wouldn't think it was just a hack to make it non-zero -- especially with the __int64 comment. yes, it is just an hack. (/me hides). But anyway, I'll leave it to the experts if the patch needs tweaking. ;-) Me too, as I don't have karma on that module. Nuno -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Great patch! From my quick tests I got about 35% performance increase. Ah, it needs a little tweak to work on windows, as 'long long' is defined as __int64 (probably we could do a typedef once for all, as there are already many #ifdefs in the code because of this). Nuno - Original Message - Nice patch! Looks good to me... Any other thoughts? -Original Message- From: Matt Wilmas [mailto:[EMAIL PROTECTED] Sent: Wednesday, November 08, 2006 8:47 PM To: internals@lists.php.net Subject: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems Hi, Here's an additional ZEND_SIGNED_MULTIPLY_LONG() for platforms with 32-bit longs that don't use the assembly version (so all Windows systems at least?). On my Windows system, mul_function() is 40% faster with this version (no overflow), which makes PHP's * operator 20% faster; with overflow mul_function() is 20% faster (though I don't see any difference with the * operator). The macro is also used in safe_emalloc()... Patch just against 5_2 as the file versions are the only difference. Matt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Hi Nuno, - Original Message - From: Nuno Lopes Sent: Thursday, November 09, 2006 Great patch! From my quick tests I got about 35% performance increase. Ah, it needs a little tweak to work on windows, as 'long long' is defined as __int64 (probably we could do a typedef once for all, as there are already many #ifdefs in the code because of this). Did it not work for you as-is on Windows...? It did for me, since it seems there's already a typedef for __int64 - long long. I'd wondered the same, but before I added the (double) cast in case of overflow, I got the Conversion from *__int64* to double: possible loss of precision warning. :-) Glad you and Andi like the patch though! I didn't know if there'd be any interest, LOL. I'll respond to Andi's message later... Nuno Matt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Matt Wilmas wrote: Hi Nuno, - Original Message - From: Nuno Lopes Sent: Thursday, November 09, 2006 Great patch! From my quick tests I got about 35% performance increase. Ah, it needs a little tweak to work on windows, as 'long long' is defined as __int64 (probably we could do a typedef once for all, as there are already many #ifdefs in the code because of this). Did it not work for you as-is on Windows...? It did for me, since it seems there's already a typedef for __int64 - long long. I'd wondered the same, but before I added the (double) cast in case of overflow, I got the Conversion from *__int64* to double: possible loss of precision warning. :-) Glad you and Andi like the patch though! I didn't know if there'd be any interest, LOL. I'll respond to Andi's message later... Could be compiler version. VC6 doesn't support long long, but I've heard that the new versions supported it. Edin -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Great patch! From my quick tests I got about 35% performance increase. Ah, it needs a little tweak to work on windows, as 'long long' is defined as __int64 (probably we could do a typedef once for all, as there are already many #ifdefs in the code because of this). Did it not work for you as-is on Windows...? It did for me, since it seems there's already a typedef for __int64 - long long. I'd wondered the same, but before I added the (double) cast in case of overflow, I got the Conversion from *__int64* to double: possible loss of precision warning. :-) I haven't tried the patch on windows, I was just guessing. But Edin's mail explains why it worked for you (we use VC 6 for official builds). Nuno -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Matt, Looks like a good patch to me. On 8-Nov-06, at 11:46 PM, Matt Wilmas wrote: Hi, Here's an additional ZEND_SIGNED_MULTIPLY_LONG() for platforms with 32-bit longs that don't use the assembly version (so all Windows systems at least?). On my Windows system, mul_function() is 40% faster with this version (no overflow), which makes PHP's * operator 20% faster; with overflow mul_function() is 20% faster (though I don't see any difference with the * operator). The macro is also used in safe_emalloc()... Patch just against 5_2 as the file versions are the only difference. Matt multiply_long.txt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php Ilia Alshanetsky -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Did it not work for you as-is on Windows...? It did for me, since it seems there's already a typedef for __int64 - long long. I'd wondered the same, I think some VC versions on windows don't know long long but do know __int64, but I don't remember from the top of my head which version, but I think maybe VC6 might have a problem with that. Otherwise looks fine, so if adjusted for VC6 it's OK. -- Stanislav Malyshev, Zend Products Engineer [EMAIL PROTECTED] http://www.zend.com/ -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
Re: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Hi Andi, - Original Message - From: Andi Gutmans Sent: Wednesday, November 08, 2006 Nice patch! Looks good to me... Any other thoughts? Since I don't have a *nix system to test on, I was wondering how my version compares to the assembly one. But I guess the latter is still fastest where supported, at least in case of no overflow. I thought I was going to mention something else I noticed when benchmarking the * operator, which is why I didn't reply sooner, but it's not multiplication-specific, so nevermind. :-) I'll send a new e-mail after looking into it more... Matt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
[PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Hi,
Here's an additional ZEND_SIGNED_MULTIPLY_LONG() for platforms with 32-bit
longs that don't use the assembly version (so all Windows systems at
least?). On my Windows system, mul_function() is 40% faster with this
version (no overflow), which makes PHP's * operator 20% faster; with
overflow mul_function() is 20% faster (though I don't see any difference
with the * operator). The macro is also used in safe_emalloc()...
Patch just against 5_2 as the file versions are the only difference.
Matt
Index: zend_multiply.h
===
RCS file: /repository/ZendEngine2/zend_multiply.h,v
retrieving revision 1.10.2.1
diff -u -r1.10.2.1 zend_multiply.h
--- zend_multiply.h 4 Jan 2006 23:53:04 - 1.10.2.1
+++ zend_multiply.h 9 Nov 2006 03:54:01 -
@@ -31,6 +31,19 @@
else (lval) = __tmpvar;
\
} while (0)
+#elif SIZEOF_LONG_LONG SIZEOF_LONG
+
+#define ZEND_SIGNED_MULTIPLY_LONG(a, b, lval, dval, usedval) do { \
+ long long __result = (long long) (a) * (long long) (b);
\
+ if (__result LONG_MAX || __result LONG_MIN) {
\
+ (dval) = (double) __result;
\
+ (usedval) = 1;
\
+ } else {
\
+ (lval) = (long) __result;
\
+ (usedval) = 0;
\
+ }
\
+} while (0)
+
#else
#define ZEND_SIGNED_MULTIPLY_LONG(a, b, lval, dval, usedval) do {
\
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
RE: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems
Nice patch! Looks good to me... Any other thoughts? -Original Message- From: Matt Wilmas [mailto:[EMAIL PROTECTED] Sent: Wednesday, November 08, 2006 8:47 PM To: internals@lists.php.net Subject: [PHP-DEV] Optimization for ..._MULTIPLY_LONG on more systems Hi, Here's an additional ZEND_SIGNED_MULTIPLY_LONG() for platforms with 32-bit longs that don't use the assembly version (so all Windows systems at least?). On my Windows system, mul_function() is 40% faster with this version (no overflow), which makes PHP's * operator 20% faster; with overflow mul_function() is 20% faster (though I don't see any difference with the * operator). The macro is also used in safe_emalloc()... Patch just against 5_2 as the file versions are the only difference. Matt -- PHP Internals - PHP Runtime Development Mailing List To unsubscribe, visit: http://www.php.net/unsub.php
