Slice: use reflect.Append instead of pre-allocating whole slice.

[gagliardetto]

No description provided.

[gagliardetto]

@terorie pls take a look next week

[riptl]

We can save a lot of bounds checking here by

• doing a bounds check for the entire array
• allocating the entire slice
• then filling it

Same for all the other primitive arrays and slices

reflect.Append is worse than pre-allocation when the slice size is known because the end result is the same (in terms of memory usage), but the former is less efficient

[riptl]

Two more optimization ideas:

• On little endian machines without alignment requirements, no conversion is necessary for integer primitive types like uint16`. We could simply unsafely copy into the backing memory buffer of a slice.
• In all simple copy cases, allocating a slice isn't even necessary – The decoder could unsafely construct a slice that points to a place in the serialized buffer

[gagliardetto]

Thanks!

Before:

Benchmark_uintSlices_Decode-8     	    6078	    179633 ns/op	   38688 B/op	    1297 allocs/op
Benchmark_uintSlices_append-8     	    3704	    302197 ns/op	   64144 B/op	    2829 allocs/op
Benchmark_uintSlices_allocate-8   	   10000	    103067 ns/op	   16408 B/op	    1026 allocs/op

After:

Benchmark_uintSlices_Decode-8        	   95216	     12386 ns/op	    4256 B/op	       6 allocs/op
Benchmark_uintSlices_append-8        	   13005	     92601 ns/op	   16432 B/op	    1027 allocs/op
Benchmark_uintSlices_preallocate-8   	   12608	     92314 ns/op	   24624 B/op	    1028 allocs/op

(NOTE: the name poorly reflects what happens inside in the AFTER)

[gagliardetto]

Okay, so now for slices of uints:

• checking if remaining bytes are enough for the desired slice size.
• slice of that specific type is pre-allocated.
• no reflect is involved in the creation of the slice or the writing to it.
• when the slice is ready, it's set upstream using reflect.

[gagliardetto]

Ooops, wrong type in the benchmark.

[gagliardetto]

Final benchmark:

BEFORE optimizing uint slices (underlying use of reflect.Append)

goos: linux
goarch: amd64
pkg: github.com/gagliardetto/binary
cpu: Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
Benchmark_uintSlice64_Decode_noMake-8         	    3636	    326426 ns/op	  131864 B/op	    5137 allocs/op
Benchmark_uintSlice64_Decode_make-8           	    3724	    308653 ns/op	  131864 B/op	    5137 allocs/op
Benchmark_uintSlice64_Decode_field_noMake-8   	    3643	    302873 ns/op	  132008 B/op	    5143 allocs/op
Benchmark_uintSlice64_Decode_field_make-8     	    3644	    303124 ns/op	  132008 B/op	    5143 allocs/op
Benchmark_uintSlice64_reflect_noMake-8        	   24592	     50456 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice64_reflect_make-8          	   24235	     49264 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice32_Decode_noMake-8         	    4402	    246662 ns/op	  107296 B/op	    4113 allocs/op
Benchmark_uintSlice32_Decode_make-8           	    4700	    251511 ns/op	  107296 B/op	    4113 allocs/op
Benchmark_uintSlice32_Decode_field_noMake-8   	    4592	    253552 ns/op	  107440 B/op	    4119 allocs/op
Benchmark_uintSlice32_Decode_field_make-8     	    4332	    256575 ns/op	  107440 B/op	    4119 allocs/op
Benchmark_uintSlice32_reflect_noMake-8        	  173953	      6631 ns/op	    4192 B/op	       4 allocs/op
Benchmark_uintSlice32_reflect_make-8          	  165073	      6599 ns/op	    4192 B/op	       4 allocs/op
PASS
ok  	github.com/gagliardetto/binary	15.178s

BEFORE optimizing uint slices (underlying use of reflect.MakeSlice + rv.Index(i).Set)

goos: linux
goarch: amd64
pkg: github.com/gagliardetto/binary
cpu: Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
Benchmark_uintSlice64_Decode_noMake-8         	    5041	    244716 ns/op	   90272 B/op	    4102 allocs/op
Benchmark_uintSlice64_Decode_make-8           	    5509	    216908 ns/op	   90272 B/op	    4102 allocs/op
Benchmark_uintSlice64_Decode_field_noMake-8   	    5554	    217488 ns/op	   90416 B/op	    4108 allocs/op
Benchmark_uintSlice64_Decode_field_make-8     	    4992	    227863 ns/op	   90416 B/op	    4108 allocs/op
Benchmark_uintSlice64_reflect_noMake-8        	   22930	     53717 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice64_reflect_make-8          	   23380	     49189 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice32_Decode_noMake-8         	    6961	    166826 ns/op	   73888 B/op	    3078 allocs/op
Benchmark_uintSlice32_Decode_make-8           	    6466	    169598 ns/op	   73888 B/op	    3078 allocs/op
Benchmark_uintSlice32_Decode_field_noMake-8   	    6379	    166451 ns/op	   74032 B/op	    3084 allocs/op
Benchmark_uintSlice32_Decode_field_make-8     	    6475	    170186 ns/op	   74032 B/op	    3084 allocs/op
Benchmark_uintSlice32_reflect_noMake-8        	  179046	      6314 ns/op	    4192 B/op	       4 allocs/op
Benchmark_uintSlice32_reflect_make-8          	  169556	      6537 ns/op	    4192 B/op	       4 allocs/op
PASS
ok  	github.com/gagliardetto/binary	18.773s

AFTER optimizing for uint slices

$ go test  -bench "uintSlice" -benchmem

goos: linux
goarch: amd64
pkg: github.com/gagliardetto/binary
cpu: Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
Benchmark_uintSlice64_Decode_noMake-8         	   22210	     52099 ns/op	   16544 B/op	    1030 allocs/op
Benchmark_uintSlice64_Decode_make-8           	   22062	     54300 ns/op	   16544 B/op	    1030 allocs/op
Benchmark_uintSlice64_Decode_field_noMake-8   	   21668	     54316 ns/op	   16688 B/op	    1036 allocs/op
Benchmark_uintSlice64_Decode_field_make-8     	   22428	     54018 ns/op	   16688 B/op	    1036 allocs/op
Benchmark_uintSlice64_reflect_noMake-8        	   22785	     53440 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice64_reflect_make-8          	   22976	     52906 ns/op	   16480 B/op	    1028 allocs/op
Benchmark_uintSlice32_Decode_noMake-8         	  149244	      7155 ns/op	    4256 B/op	       6 allocs/op
Benchmark_uintSlice32_Decode_make-8           	  160652	      7149 ns/op	    4256 B/op	       6 allocs/op
Benchmark_uintSlice32_Decode_field_noMake-8   	  151970	      7551 ns/op	    4400 B/op	      12 allocs/op
Benchmark_uintSlice32_Decode_field_make-8     	  151761	      7708 ns/op	    4400 B/op	      12 allocs/op
Benchmark_uintSlice32_reflect_noMake-8        	  160108	      7006 ns/op	    4192 B/op	       4 allocs/op
Benchmark_uintSlice32_reflect_make-8          	  154717	      7048 ns/op	    4192 B/op	       4 allocs/op
PASS
ok  	github.com/gagliardetto/binary	17.686s

[gagliardetto]

If this looks good, I'll proceed tomorrow to merge. @terorie

[gagliardetto]

After simplifying readNBytes:

goos: linux
goarch: amd64
pkg: github.com/gagliardetto/binary
cpu: Intel(R) Core(TM) i7-7700HQ CPU @ 2.80GHz
Benchmark_uintSlice64_Decode_noMake-8                    	  109407	     10240 ns/op	    8352 B/op	       6 allocs/op
Benchmark_uintSlice64_Decode_make-8                      	  104725	     10590 ns/op	    8352 B/op	       6 allocs/op
Benchmark_uintSlice64_Decode_field_noMake-8              	  106695	     10673 ns/op	    8496 B/op	      12 allocs/op
Benchmark_uintSlice64_Decode_field_make-8                	  109975	     11053 ns/op	    8496 B/op	      12 allocs/op
Benchmark_uintSlice64_readArray_noMake-8                 	  109645	     10126 ns/op	    8288 B/op	       4 allocs/op
Benchmark_uintSlice64_readArray_make-8                   	  104791	     10162 ns/op	    8288 B/op	       4 allocs/op
Benchmark_uintSlice64_Decode_field_withCustomDecoder-8   	  110866	     10326 ns/op	    8240 B/op	       2 allocs/op
Benchmark_uintSlice32_Decode_noMake-8                    	  162374	      7324 ns/op	    4256 B/op	       6 allocs/op
Benchmark_uintSlice32_Decode_make-8                      	  140947	      7607 ns/op	    4256 B/op	       6 allocs/op
Benchmark_uintSlice32_Decode_field_noMake-8              	  130076	      7981 ns/op	    4400 B/op	      12 allocs/op
Benchmark_uintSlice32_Decode_field_make-8                	  127812	      8376 ns/op	    4400 B/op	      12 allocs/op
Benchmark_uintSlice32_readArray_noMake-8                 	  150466	      7251 ns/op	    4192 B/op	       4 allocs/op
Benchmark_uintSlice32_readArray_make-8                   	  160896	      6962 ns/op	    4192 B/op	       4 allocs/op
Benchmark_uintSlice32_Decode_field_withCustomDecoder-8   	  151675	      7046 ns/op	    4144 B/op	       2 allocs/op