Hi,
On Thu, January 1, 2015 13:30, Dmitry Stogov wrote:
> Thanks for analizing.
> did you say "8.5 more" instructions? in my test I saw "less".
>
> It's really an intermedie step, and we probably will check if other
> optimizations are reaaly benefitable before committing this. But yes, this
> patch doesn't make any degradation, except for slight increase of memory
> consumption.
>
yeah, i'm afraid that's a typo, less instruction retired. Here's the full
report I have, rerun the profiling a couple of times again to be sure. The
numbers on the right are from the patch and on the left from the
mainstream.
Elapsed Time: 2.428s - 2.088s = 0.339s
Clockticks: 5,891,879,618 - 5,800,714,085 = 91,165,533
Instructions Retired: 10,148,429,042 - 10,121,040,536 = 27,388,506
CPI Rate: 0.581 - 0.573 = 0.007
MUX Reliability: 0.985 - 0.977 = 0.008
Paused Time: 0.263s - 0s = 0.263s
Filled Pipeline Slots:
Retiring: 0.483 - 0.488 = -0.006
Assists: 0.000 - 0.000 = 0.000
Bad Speculation: 1.000 - 0.890 = 0.110
Branch Mispredict: 1.000 - 1.000 = 0.000
Machine Clears: 0.000 - 0.000 = 0.000
Unfilled Pipeline Slots (Stalls):
Back-end Bound: 0.049 - 0.105 = -0.056
DIV Active: 0.000 - 0.000 = 0.000
Flags Merge Stalls: 0.006 - 0.000 = 0.006
Slow LEA Stalls: 0.006 - 0.000 = 0.006
Memory Latency:
LLC Miss: 0.000 - 0.000 = 0.000
LLC Hit: 0.000 - 0.000 = 0.000
DTLB Overhead: 0.000 - 0.001 = -0.001
Contested Accesses: 0.000 - 0.000 = 0.000
Data Sharing: 0.000 - 0.000 = 0.000
Memory Replacements:
L1D Replacement Percentage: 0.000 - 0.000 = 0.000
L2 Replacement Percentage: 1.000 - 1.000 = 0.000
LLC Replacement Percentage: 0.000 - 0.000 = 0.000
Memory Reissues:
Loads Blocked by Store Forwarding: 0.000 - 0.000 = 0.000
Split Loads: 0.000 - 0.000 = 0.000
Split Stores: 0.000 - 0.000 = 0.000
4K Aliasing: 0.011 - 0.013 = -0.003
Front-end Bound: 0.137 - 0.100 = 0.037
ICache Misses: 0.000 - 0.000 = 0.000
ITLB Overhead: 0.000 - 0.000 = 0.000
DSB Switches: 0.000 - 0.000 = 0.000
As one sees, the instruction count vary, but that's most likely because of
the measurement uncertainty.
Regards
Anatol
--
PHP Internals - PHP Runtime Development Mailing List
To unsubscribe, visit: http://www.php.net/unsub.php
