AST-33: Codegen for Word8/16/32 operations #216

ggreif · 2019-03-05T22:52:40Z

Stories:

https://dfinity.atlassian.net/browse/AST-33
https://dfinity.atlassian.net/browse/AST-63 (Word32 specific)

Probably nothing interesting for Word32. OTOH, Word8 and Word16 are kept in an MSB-aligned format with Vanilla SR. Carrying operations (add, sub, mul) are correctly working by discarding the ripple. Most logical ops are also not infectious, i.e. won't introduce nonzero bits in the lower part. Thus, the frequent operations come for free.
Similarly Char can be represented 8-bit-left-shifted as Vanilla.

This is the theory, and now also practice, as the tests pass :-)

I managed to keep the generating code mostly uniform, due to the additional bookkeeping being in helpers, that nop when nothing is to ne done.

I have also (kind-of) implemented a C++-like type based templating mechanism, see name_of_type.

exotic operators for Word32 (clz, ctz, popcnt)
exotic operators for Word8/16 (clz, ctz, popcnt) ---> done in AST-33: Support popcnt/clz/ctz/shrs for Word8/Word16 #238
exotic binary operator for Word32: shrs
exotic binary operator for Word8/16: shrs (also: revisit implementation, see comment below, AST-33: Word64 and operations #235) ---> done in AST-33: Support popcnt/clz/ctz/shrs for Word8/Word16 #238
exotic binary operator for Word8/16/32/64: btst (AST-33: bit-testing of words #239)
move helpers around (where? module UnboxedSmallWord, probably)
charToWord32 and word32ToChar
turbocharge (**)
intro bounds check on word32ToChar (interpreter and compiler)

nomeata

So, this means that 0xAB : Word8 is stored as 0xAB000000 : i32, right?

This does not fly yet: The (current) specification for Vanilla is that it is a pointer if the last bit is 0, and a scalar otherwise. So it would have to be 0xAB000001 : i32. You can probably trigger a bug this way if you send a message with your new data types.

(In fact, please do extend test/run-dfinity/data-params.as as you implement new data types).

The two solutions are:

Add the tag bit there (and complicate some of the operations), or
Swap the convention: Pointers are tagged and scalars are not. Andreas originally envisioned this, by representing a pointer to location x as x-1. We’d have to adjust where we call the system API (data.externalize); Load and Store can use the offset to fix that.

If you think we should do the latter, then that should happen on a separate branch first, and well tested, I think. Maybe let’s chat on Slack about it.

never seen this before: `_out/pow.wasm:0x___: resource exhaustion: call stack exhausted`

to make life easier for arithmetic on Word8/Word16 (#216) (This is with GC disabled.)

test/run-dfinity/ok/data-params.dvm-run.ok

test/run-dfinity/ok/data-params.run-ir.ok

ggreif · 2019-03-06T23:58:19Z

~~Blocked on #217.~~

we have a polymorpic expander now!

of type-driven templating

also make sure that we don't emit code for clamping/sanitising Word32

ggreif · 2019-03-08T10:46:02Z

@nomeata I'd like to make popcnt type-polymorphic with specialised implementations (à la Haskell typeclass). Restricted for Word{8,16,32,64}.

Is there an example for such a thing? You wrote some show code back then...

this is not semantics changing, but may reduce the recursion depth

no need to sanitise when we know that the bit is clear

24a55ce Revert "another attempt to go green" dc5e8bc Revert "try go green"

test/run-dfinity/ok/actor-reexport.dvm-run.ok

ggreif · 2019-03-10T15:44:23Z

@nomeata I think this is ready for review now. The last bullet point (bounds check on word32ToChar) can wait, but I can also easily press it in, if so desired. The other bullet points I'd like to resolve in another PR.

into new module UnboxedSmallWord

nomeata

Ups, forgot to submit this comment.

src/compile.ml

fix serialisation of Tagged.SmallWord this was a carry-over from before the pointer tagging was changed

instead of memcpy_words_skewed this version is probably more efficient (less calls) ``` asm i32.const 1 call $alloc_words tee_local $x' i32.const 11 i32.store offset=1 get_local $x' get_local $x i32.load offset=5 i32.store offset=5 get_local $x' ```

ggreif · 2019-03-11T10:24:49Z

@nomeata I was going to suggest to have helpers for alloc+store_tag, etc. Will use the Tagged.obj helper!

the resulting code is ``` asm i32.const 2 call $alloc_words tee_local $heap_object i32.const 11 i32.store offset=1 get_local $heap_object get_local $x i32.load offset=5 i32.store offset=5 get_local $heap_object ```

src/compile.ml

nomeata

Thanks, good job!

## Changelog for candid: Branch: master Commits: [dfinity/candid@68135b51...8cb12349](dfinity/candid@68135b5...8cb1234) * [`cb117b7b`](dfinity/candid@cb117b7) JS binding for BigInt ([dfinity/candid⁠#215](http://r.duckduckgo.com/l/?uddg=https://github.com/dfinity/candid/issues/215)) * [`8cb12349`](dfinity/candid@8cb1234) release ([dfinity/candid⁠#216](http://r.duckduckgo.com/l/?uddg=https://github.com/dfinity/candid/issues/216))

## Changelog for ic-hs: Branch: master Commits: [dfinity/ic-hs@e727feca...a5f89cf5](dfinity/ic-hs@e727fec...a5f89cf) * [`99efc33f`](dfinity/ic-hs@99efc33) sync node height before making an update call in compliance tests ([dfinity/ic-hs⁠#213](https://togithub.com/dfinity/ic-hs/issues/213)) * [`4bdf6c35`](dfinity/ic-hs@4bdf6c3) sync node height before getStateCert ([dfinity/ic-hs⁠#214](https://togithub.com/dfinity/ic-hs/issues/214)) * [`f718f74d`](dfinity/ic-hs@f718f74) retry on http status of 429 ([dfinity/ic-hs⁠#215](https://togithub.com/dfinity/ic-hs/issues/215)) * [`a5f89cf5`](dfinity/ic-hs@a5f89cf) add note that DFINITY stopped contributing ([dfinity/ic-hs⁠#216](https://togithub.com/dfinity/ic-hs/issues/216))

ggreif added 5 commits March 5, 2019 23:49

AST-33: codegen for Word8/16

03d6936

generate special code for Word8/16 infectious operations

9f9a4dd

fix embarrassing thinko about Xor

3a12d86

oooops, cannot share, type dependent

f3287ec

refactoring

16d81ef

nomeata reviewed Mar 6, 2019

View reviewed changes

ggreif added 5 commits March 6, 2019 14:24

add Word32 test actor

fb5ae0c

Merge remote-tracking branch 'origin/master' into gabor/conversions

489e6f1

use share_code1 and provide dummy PowOp

00777ed

very naive pow for Word32

cae5449

silly me

4d7bb4a

never seen this before: `_out/pow.wasm:0x___: resource exhaustion: call stack exhausted`

nomeata added a commit that referenced this pull request Mar 6, 2019

Tag pointers, not scalars

07e8192

to make life easier for arithmetic on Word8/Word16 (#216) (This is with GC disabled.)

ggreif added 2 commits March 6, 2019 22:14

fix tests (hint: they look a bit strange)

c8e3074

pow16

00f5ab2

ggreif commented Mar 6, 2019

View reviewed changes

test/run-dfinity/ok/data-params.dvm-run.ok Outdated Show resolved Hide resolved

ggreif commented Mar 6, 2019

View reviewed changes

test/run-dfinity/ok/data-params.run-ir.ok Outdated Show resolved Hide resolved

add GC and serialisation support for SmallWord

895ad21

ggreif added 11 commits March 7, 2019 01:21

third one's the charm: Word8

2a2c1e3

we have a polymorpic expander now!

finish up demo

11f9646

of type-driven templating

implement codegen for Word8/16 Rotl/Rotr

0bb9d8e

Merge branch 'master' into gabor/conversions

da9f84e

also test Word16 in actors

ea6caf7

clamp_shift_amount

395b402

Merge branch 'master' into gabor/conversions

72b83be

bug fixed

94e73d0

refactor

4271874

also make sure that we don't emit code for clamping/sanitising Word32

a bit more of refactoring

31459b1

bit-fiddling doesn't need an environment

40c4e87

ggreif and others added 9 commits March 10, 2019 12:20

undo bad change

6cef59a

refacor pow generator

57bc97e

sanitise left shift

4781ab6

this is not semantics changing, but may reduce the recursion depth

minor tweak

269c571

no need to sanitise when we know that the bit is clear

simplify

40732d6

another attempt to go green

6d70731

ok files

a82204b

no idea where this comes from

32c105c

262858d Revert "ok files"

e6a48b1

24a55ce Revert "another attempt to go green" dc5e8bc Revert "try go green"

ggreif commented Mar 10, 2019

View reviewed changes

test/run-dfinity/ok/actor-reexport.dvm-run.ok Outdated Show resolved Hide resolved

move helper functions

25cdb78

into new module UnboxedSmallWord

nomeata reviewed Mar 10, 2019

View reviewed changes

src/compile.ml Show resolved Hide resolved

nomeata reviewed Mar 10, 2019

View reviewed changes

src/compile.ml Outdated Show resolved Hide resolved

src/compile.ml Show resolved Hide resolved

ggreif self-assigned this Mar 11, 2019

ggreif added 2 commits March 11, 2019 09:22

review feedback

322db4c

fix serialisation of Tagged.SmallWord this was a carry-over from before the pointer tagging was changed

use load_field/store_field

aa8c8da

instead of memcpy_words_skewed this version is probably more efficient (less calls) ``` asm i32.const 1 call $alloc_words tee_local $x' i32.const 11 i32.store offset=1 get_local $x' get_local $x i32.load offset=5 i32.store offset=5 get_local $x' ```

use Tagged.obj

4dcc7ec

the resulting code is ``` asm i32.const 2 call $alloc_words tee_local $heap_object i32.const 11 i32.store offset=1 get_local $heap_object get_local $x i32.load offset=5 i32.store offset=5 get_local $heap_object ```

ggreif commented Mar 11, 2019

View reviewed changes

src/compile.ml Outdated Show resolved Hide resolved

Merge branch 'master' into gabor/conversions

3d3cbb9

ggreif requested a review from nomeata March 11, 2019 16:48

nomeata approved these changes Mar 12, 2019

View reviewed changes

ggreif merged commit 9563d54 into master Mar 12, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

AST-33: Codegen for Word8/16/32 operations #216

AST-33: Codegen for Word8/16/32 operations #216

ggreif commented Mar 5, 2019 •

edited

Loading

nomeata left a comment •

edited by ggreif

Loading

ggreif commented Mar 6, 2019 •

edited

Loading

ggreif commented Mar 8, 2019

ggreif commented Mar 10, 2019

nomeata left a comment

ggreif commented Mar 11, 2019 •

edited

Loading

nomeata left a comment

AST-33: Codegen for Word8/16/32 operations #216

AST-33: Codegen for Word8/16/32 operations #216

Conversation

ggreif commented Mar 5, 2019 • edited Loading

nomeata left a comment • edited by ggreif Loading

Choose a reason for hiding this comment

ggreif commented Mar 6, 2019 • edited Loading

ggreif commented Mar 8, 2019

ggreif commented Mar 10, 2019

nomeata left a comment

Choose a reason for hiding this comment

ggreif commented Mar 11, 2019 • edited Loading

nomeata left a comment

Choose a reason for hiding this comment

ggreif commented Mar 5, 2019 •

edited

Loading

nomeata left a comment •

edited by ggreif

Loading

ggreif commented Mar 6, 2019 •

edited

Loading

ggreif commented Mar 11, 2019 •

edited

Loading