(Bug 3147) Reimplemented Value-Range Propagation.

[kennytm]

This should have fixed bug 3147 (and bug 6000, bug 5225). The changes are:

• VRP is applied on all integer arithmetic operators, i.e. + - * / % << >> >>> & | ^, and the unary ~ -.
• Range with signed numbers is properly supported by using a 65-bit number internally.
• The algorithm for & | ^ are rewritten to produce a much tighter range than before.
• Several unlisted bugs producing a wrong range regarding << >> >>> are fixed.

[donc]

Awesome!

[andralex]

There was this large discussion about VRP for & and |. Some people came up with pretty compelling algorithms - did you follow that? If not, you may want to compare notes.

[andralex]

One more thing - is VRP effective in switch statements? Consider:

int x;
switch (x & 7) { ... }

With VRP in tow, the compiler should figure that only cases 0 through 7 are needed.

[kennytm]

@andralex: Thanks for reminding. I've checked Adam's algorithm, and his provides a tighter upper bound, but a looser lower bound. So I have combined the two. This should be able to provide a tightest bound in eighty-something percent of all possible cases for |. But & and ^ are still not very good. See https://gist.github.com/1022794 for the result.

As for switch statement, VRP may not produce satisfactory result, e.g.

int x;
final switch (x & 0xf0) {
   ...
}

will require every (241) case in (0, 1, 2, ..., 0xf0) be present, although only 16 cases (0, 0x10, ..., 0xf0) are needed.

[andralex]

@kennytm: Thanks, that's great. The exhaustive checker has its role and merits, but I'd generally hope for bounds that are provably the tightest. (Then the exhaustive checker would serve as a debugger.) Absent the theoretical proof, of course this pull is much better than what we have.

Regarding switch - good point. I think a common pattern is to use switch with a small power-of-two-minus-one mask, so we should make sure we get that covered.

[andralex]

I'll leave it to Walter and Don to pull this.

[WalterBright]

Halp! This can't be auto-merged, what do I do? People didn't like when I did it manually :-(

[CyberShadow]

Do what Linus does, and tell the submitter to do the honors instead!

http://www.youtube.com/watch?v=4XpnKHJAok8#t=35m45s

[WalterBright]

If Linus says so, it must be ok!

[kennytm]

@WalterBright: done :)

[WalterBright]

awesome!