478211 - eliminate null pointer checks in CodegenLIR independently from the verifier

Status: VERIFIED WONTFIX

(Tamarin Graveyard :: Baseline JIT (CodegenLIR), defect)

Description

[Edwin Smith]

It is possible to avoid using FrameState.notNull and instead eliminate null checks in CodegenLIR with a fairly simple Nanojit filter.

rough sketch:
* the (p == 0) expression can already be matched by CseFilter
* a new filter can maintain a set of LIns* for previously seen pointers that are not null.  (if you see jt(p==0), then from that point until the next label, p is in the set).  
* if a second jt(p==0) is seen, it can be dropped since it's always false.
* similar logic can be used for handling the check for Atom pointers being == nullAtom or undefinedAtom (the relation needed is p <= undefinedAtom)

Since this works on the LIns* for p, we don't need the loop that searches for copies of p in FrameState that is in Verifier::emitSetNull().

With this change, FrameState.notNull will be write-only and can be eliminated.  notNull is not captured by OP_newclass or OP_newfunction so it's only used as an optimization within the verifier pipeline, and only with the jit enabled.

Note: there is no going back.  removing notNull will relax the verifier's type-meet rules, after which point all future vm's must support the more relaxed rules.  luckily, this is what we want.  

This bug depends on removing MIR so we don't have to do all this twice.

Comment 1

[Edwin Smith]

Attachment: first draft, not ready for review but here for safe keeping

Comment 2

[Edwin Smith]

Attachment: show consts, labels, and branches, in live instruction list

Also makes live expression list optional (too much output most of the time).

Comment 3

[Edwin Smith]

Comment on attachment 365433 [details] [diff] [review]
show consts, labels, and branches, in live instruction list

pushed http://hg.mozilla.org/tamarin-redux/rev/6701ace052c6

Comment 4

[Edwin Smith]

Attachment: merged with latest verifier changes

A few more functions added to the not-null list, updated to work with either pointer size.

Comment 5

[Jeff Dyer]

As an experiment I did the following:

- stop managing notNull in frame state
- still call coder->writeCheckNull from the verifier
- change assert_nz to call is_nz to check the nz-ness of a value

The hypothesis is that this will uncover missing nz copies in the NullCheckFilter. 

When there are no more asserts we can move the call to writeCheckNull out of the verifier and into CodegenLIR.

Comment 6

[William Maddox]

Null check elimination has now been decoupled from the verifier, in VarTracker.
Is this bug still relevant?

Comment 7

[Edwin Smith]

obsoleted by verifier work in spring 2010

Comment 8

[Trevor Baker]

bulk verifying resolved !fixed issues