534378 - eliminate guards that prevent overflows in otherwise dead expressions

Status: RESOLVED WONTFIX

(Core Graveyard :: Nanojit, defect)

Description

[Andreas Gal :gal]

The expression

a + b;

is dead code. If we record it, we will speculate for an integer addition if a and b are both integers, and emit a guard that checks that no overflow occurs. This expression will never be dead, because the guard keeps it alive. However, the guard serves no purpose. Nobody really uses the value.

The attached patch allows to flags guards as weak. I squirreled the weak flag into the GuardRecord for now. This should probably be encoded in the opcode instead (LIR_wxt/LIR_wxf?).

When assembling a weak guard, we check whether any part of the condition expression is already live. If not, we don't emit the guard.

This passes our tests and sunspider in a debug build. We get a recursion error in an OPT build on SS in controlflow-recursive. David .... help :)

Comment 1

[Andreas Gal :gal]

Attachment: patch

Comment 2

[Nicholas Nethercote [inactive]]

Why not make all guards weak in this way?

Nit: s/weather/whether/.

Comment 3

[Andreas Gal :gal]

If a guard checks a shape, we can't just eliminate it. The guarded expression has to be pure.

Comment 4

[Andreas Gal :gal]

Actually the patch didn't do this. This is broken in TM tip.

Comment 5

[Ryan VanderMeulen [:RyanVM]]

Edwin, does Nanojit want this?

Comment 6

[Nicholas Nethercote [inactive]]

Nanojit has been dead for several years. Its Bugzilla component has been moved to the graveyard (bug 984276).

I checked all the open bugs. They're all uninteresting, so I'm WONTFIXing them all. Apologies for the bugspam.