Add 32 bit compressed pointer support. #1271

zherczeg · 2016-08-17T06:19:43Z

This is a WIP patch, which will not be landed before the 1.0 release.

This patch extends compressed pointers to be 32 bit long when "--cpointer_32_bit on" is passed. After custom allocators will be added to Jerry, this will allow to use the full 32 bit address space on 32 bit systems at the cost of increased memory size. I will do some measurements to estimate this trade-of with the current allocator.

zherczeg · 2016-08-22T07:31:55Z

Around 20% memory increase (9% RSS increase), a little perf gain (2%) and a small binary size reduction (200 byte).

Benchmark	Heap (bytes)	RSS (bytes)	Perf (sec)
3d-cube.js	16376 -> 24568 : -50.024%	57344 -> 65536 : -14.286%	0.944 -> 0.929 : +1.550%
3d-raytrace.js	161152 -> 197488 : -22.548%	217088 -> 212992 : +1.887%	1.116 -> 1.157 : -3.704%
access-binary-trees.js	32760 -> 49136 : -49.988%	53248 -> 69632 : -30.769%	0.592 -> 0.549 : +7.289%
access-fannkuch.js	7608 -> 8184 : -7.571%	24576 -> 24576 : 0.000%	2.363 -> 2.330 : +1.409%
access-nbody.js	6688 -> 8176 : -22.249%	28672 -> 28672 : 0.000%	1.077 -> 1.058 : +1.752%
bitops-3bit-bits-in-byte.js	928 -> 1072 : -15.517%	16384 -> 20480 : -25.000%	0.591 -> 0.579 : +1.951%
bitops-bits-in-byte.js	880 -> 1032 : -17.273%	16384 -> 20480 : -25.000%	0.858 -> 0.853 : +0.591%
bitops-bitwise-and.js	568 -> 664 : -16.901%	16384 -> 16384 : 0.000%	0.934 -> 0.883 : +5.476%
bitops-nsieve-bits.js	138568 -> 175248 : -26.471%	155648 -> 192512 : -23.684%	1.788 -> 1.774 : +0.764%
controlflow-recursive.js	912 -> 1048 : -14.912%	94208 -> 98304 : -4.348%	0.399 -> 0.388 : +2.775%
crypto-aes.js	49176 -> 57376 : -16.675%	94208 -> 102400 : -8.696%	1.113 -> 1.096 : +1.502%
crypto-md5.js	139256 -> 172024 : -23.531%	163840 -> 208896 : -27.500%	0.738 -> 0.751 : -1.886%
crypto-sha1.js	90104 -> 106488 : -18.183%	110592 -> 126976 : -14.815%	0.705 -> 0.704 : +0.008%
date-format-tofte.js	16376 -> 16376 : 0.000%	40960 -> 40960 : 0.000%	0.864 -> 0.841 : +2.626%
date-format-xparb.js	16384 -> 24576 : -50.000%	49152 -> 53248 : -8.333%	0.470 -> 0.440 : +6.270%
math-cordic.js	2128 -> 2640 : -24.060%	20480 -> 20480 : 0.000%	1.345 -> 1.302 : +3.180%
math-partial-sums.js	1432 -> 1776 : -24.022%	16384 -> 20480 : -25.000%	0.708 -> 0.695 : +1.804%
math-spectral-norm.js	8184 -> 8176 : +0.098%	24576 -> 24576 : 0.000%	0.599 -> 0.581 : +3.107%
string-base64.js	92032 -> 93472 : -1.565%	143360 -> 139264 : +2.857%	2.054 -> 1.991 : +3.047%
string-fasta.js	16320 -> 16368 : -0.294%	32768 -> 32768 : 0.000%	1.402 -> 1.338 : +4.586%
Geometric mean:	-19.174%	-9.542%	+2.237%

Binary sizes (bytes)
314e74f:156952
5600c4c:156724

LaszloLango · 2016-08-22T07:37:43Z

Enable cpointer_32_bit by default (this will be removed after testing).

IMHO, we should add a new travis job and test this feature on every new commit.

zherczeg · 2016-08-22T07:44:39Z

As you can see travis is failing, since you need a 32 bit system (travis is 64 bit by default). I dont have any idea how to test it.

LaszloLango · 2016-08-31T12:55:31Z

jerry-core/ecma/base/ecma-globals.h

  jmem_cpointer_t next_property_cp; /**< next cpointer */
+#endif /* JERRY_CPOINTER_32_BIT */
 } ecma_property_header_t;


Can you explain this change? Why did you move the next cpointer on 32 bit?

Because ECMA_PROPERTY_VALUE_PTR expects that the type fields are packed in a 4 byte word, which is followed by maximum of four value words. For 16 bit cpointers the next pointer can be squeezed in the last two bytes of the first word. This is not possible with 32 bit cpointers, where that 16 bit is unused.

LaszloLango · 2016-08-31T13:27:31Z

The code looks good, but docs/04.INTERNALS.md should be updated.

zherczeg · 2016-09-01T08:45:27Z

Description added, patch updated.

LaszloLango · 2016-09-01T08:55:06Z

LGTM

dbatyai · 2016-09-01T09:31:14Z

jerry-core/config.h

@@ -39,8 +39,6 @@
 */
 #ifndef CONFIG_MEM_HEAP_AREA_SIZE
 # define CONFIG_MEM_HEAP_AREA_SIZE (512 * 1024)
-#elif CONFIG_MEM_HEAP_AREA_SIZE > (512 * 1024)
-# error "Currently, maximum 512 kilobytes heap size is supported"


Removing this may be a bad idea, I think it would be better to add && !JERRY_CPOINTER_32_BIT to the condition instead.

Another assertion is added instead of it.

dbatyai · 2016-09-01T09:32:11Z

LGTM otherwise.

zherczeg · 2016-09-01T09:58:55Z

jerry-core/jmem/jmem-heap.c

-                       two_pow_mem_heap_offset_should_not_be_less_than_mem_heap_size);
+#ifndef JERRY_CPOINTER_32_BIT
+  JERRY_STATIC_ASSERT (((UINT16_MAX + 1) << JMEM_ALIGNMENT_LOG) >= JMEM_HEAP_SIZE,
+                       maximum_heap_size_for_16_bit_compressed_pointers_is_512K);


Oh, I see. Okay.

dbatyai · 2016-09-01T11:06:08Z

LGTM

JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg [email protected]

zherczeg force-pushed the cpointer_32 branch from eec6cdf to 29bc81c Compare August 17, 2016 06:28

LaszloLango added memory management Related to memory management or garbage collection development Feature implementation labels Aug 17, 2016

zherczeg force-pushed the cpointer_32 branch 2 times, most recently from 910da8c to 5600c4c Compare August 22, 2016 06:50

zherczeg force-pushed the cpointer_32 branch from 5600c4c to c7f876f Compare August 22, 2016 13:13

LaszloLango reviewed Aug 31, 2016
View reviewed changes

zherczeg force-pushed the cpointer_32 branch from c7f876f to 3f1c9a1 Compare September 1, 2016 08:45

dbatyai reviewed Sep 1, 2016
View reviewed changes

zherczeg reviewed Sep 1, 2016
View reviewed changes

Add 32 bit compressed pointer support.

da47ded

JerryScript-DCO-1.0-Signed-off-by: Zoltan Herczeg [email protected]

zherczeg force-pushed the cpointer_32 branch from 3f1c9a1 to da47ded Compare September 1, 2016 11:17

zherczeg merged commit da47ded into jerryscript-project:master Sep 1, 2016

zherczeg deleted the cpointer_32 branch September 1, 2016 11:53

LaszloLango mentioned this pull request Sep 5, 2016

I need more memory than 256KB jerryscript-project/iotjs#298

Closed

jiangzidong mentioned this pull request Sep 13, 2016

fix bug in vm call_stack_size calculation #1345

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Add 32 bit compressed pointer support. #1271

Add 32 bit compressed pointer support. #1271

Uh oh!

zherczeg commented Aug 17, 2016

Uh oh!

zherczeg commented Aug 22, 2016

Uh oh!

LaszloLango commented Aug 22, 2016

Uh oh!

zherczeg commented Aug 22, 2016

Uh oh!

LaszloLango Aug 31, 2016

Uh oh!

zherczeg Aug 31, 2016 •

edited

Loading

Uh oh!

LaszloLango commented Aug 31, 2016

Uh oh!

zherczeg commented Sep 1, 2016

Uh oh!

LaszloLango commented Sep 1, 2016

Uh oh!

dbatyai Sep 1, 2016

Uh oh!

zherczeg Sep 1, 2016

Uh oh!

dbatyai commented Sep 1, 2016

Uh oh!

zherczeg Sep 1, 2016

Uh oh!

dbatyai Sep 1, 2016

Uh oh!

dbatyai commented Sep 1, 2016

Uh oh!

Uh oh!

Add 32 bit compressed pointer support. #1271

Add 32 bit compressed pointer support. #1271

Uh oh!

Conversation

zherczeg commented Aug 17, 2016

Uh oh!

zherczeg commented Aug 22, 2016

Uh oh!

LaszloLango commented Aug 22, 2016

Uh oh!

zherczeg commented Aug 22, 2016

Uh oh!

LaszloLango Aug 31, 2016

Choose a reason for hiding this comment

Uh oh!

zherczeg Aug 31, 2016 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

LaszloLango commented Aug 31, 2016

Uh oh!

zherczeg commented Sep 1, 2016

Uh oh!

LaszloLango commented Sep 1, 2016

Uh oh!

dbatyai Sep 1, 2016

Choose a reason for hiding this comment

Uh oh!

zherczeg Sep 1, 2016

Choose a reason for hiding this comment

Uh oh!

dbatyai commented Sep 1, 2016

Uh oh!

zherczeg Sep 1, 2016

Choose a reason for hiding this comment

Uh oh!

dbatyai Sep 1, 2016

Choose a reason for hiding this comment

Uh oh!

dbatyai commented Sep 1, 2016

Uh oh!

Uh oh!

zherczeg Aug 31, 2016 •

edited

Loading