1864209 - Experimentation with a MIR-level inliner for wasm

Status: RESOLVED WONTFIX

(Core :: JavaScript: WebAssembly, enhancement, P2)

Description

[Julian Seward [:jseward]]

Let's do some experimentation to assess the difficulties in
building a MIR-level inliner that can be used for wasm code.

Comment 1

[Julian Seward [:jseward]]

Attachment: bug1864209-inliner.diff-5

Initial code for supporting serialize/deserialize classes. Also, top level
coordination of compilation, enough to make testing of the inlining feasible.

Comment 2

[Julian Seward [:jseward]]

Attachment: bug1864209-inliner.diff-6

With classes Serializer/Deserializer tidied up.

Comment 3

[Julian Seward [:jseward]]

Attachment: bug1864209-inliner.diff-22

With redesigned serialization library ({Se,Des}erializer2), and further
integration with MIR. WIP backup, unusable for anything as-is.

Comment 4

[Julian Seward [:jseward]]

Attachment: bug1864209-inliner.diff-28

Is able to round-trip (serialize and deserialize) some simple
MIR graphs resulting from wasm inputs. Hence it covers most
of the heavy lifting for MIR generic framework stuff
(MIRGraph, MBasicBlock, MPhi, MUse)
and for Vector, FixedList, InlineList, InlineForwardsList
as well as a handful of specific MInstructions (parameter,
call, compare, test, goto, add/sub/mul/div/shl/shr, return)
needed for the test cases shown in comment 5.

Comment 5

[Julian Seward [:jseward]]

Attachment: badness124.js

Test case for the patch in comment 4. Run with
IONFLAGS=dump-mir-expr ./dist/bin/js --no-threads --no-ion /path/to/badness124.js
See comment 6 for resulting MIR.

Comment 6

[Julian Seward [:jseward]]

Attachment: MIR for the test case in comment 5

Round-trippable MIR for the test case in comment 5
as processed using the patch in comment 4.

Comment 7

[Julian Seward [:jseward]]

Attachment: bug1864209-inliner.diff-29

Marginally cleaned up version of comment 4 patch.
Applies against m-c 684483:5d6699b34edc.

Comment 8

[Ryan Hunt [:rhunt]]

Julian, do you want this still open?

Comment 9

[Julian Seward [:jseward]]

Closing as WONTFIX, since it seems like this isn't a road we plan to travel
along any time soon. The scheme under development in bug 1868521 seems more
promising.