Adding Bitand, Neg, lowbit tests, and lowbit optimizations
#714
Conversation
FTRobbin
changed the title
Bitand, Neg, and lowbit testsBitand, Neg, lowbit tests, and lowbit optimizations
|
Will merge after successful nightly! Being careful right now because it seems like eggcc might have crashed the nightly last night... |
| ;; PopcountIterations guarantees termination for non-zero values | ||
| ;; lowbit(0) is undefined behavior | ||
|
|
||
| (rule ( |
There was a problem hiding this comment.
This technically violates the weak linearity because we have not provided an equivalent value for other values of the then branch?
There was a problem hiding this comment.
My understanding is that if you are computing something else in your loop, you will not be able to extract without the loop.
There was a problem hiding this comment.
Yihong is right- we need to provide a loop that doesn't have the state edge in it as an alternative. See the existing state edge passthrough file for an example of how it's done for if statements.
If you are confused why we have to do this, hopefully the paper clears it up! Good chance to read that section
There was a problem hiding this comment.
Also, move this rule to that file and ruleset
There was a problem hiding this comment.
I have fixed this issue, but I feel it makes more sense for the rule to stay in this file.
| ;; PopcountIterations guarantees termination for non-zero values | ||
| ;; lowbit(0) is undefined behavior | ||
|
|
||
| (rule ( |
There was a problem hiding this comment.
Yihong is right- we need to provide a loop that doesn't have the state edge in it as an alternative. See the existing state edge passthrough file for an example of how it's done for if statements.
If you are confused why we have to do this, hopefully the paper clears it up! Good chance to read that section
| ;; PopcountIterations guarantees termination for non-zero values | ||
| ;; lowbit(0) is undefined behavior | ||
|
|
||
| (rule ( |
There was a problem hiding this comment.
Also, move this rule to that file and ruleset
| (let newlpinputs (TupleRemoveAt lpinputs j)) | ||
| (let newpred_outputs (TupleRemoveAt pred_outputs (+ j 1))) | ||
|
|
||
| (let newlpctx (DummyLoopContext newlpinputs newpred_outputs pred_outputs)) |
There was a problem hiding this comment.
These days we've decided to use TmpCtx instead, then make sure to delete it at the end of the rule. It's safer in case you forget to include some important information in the dummy context
|
Nightly looks great! |
This PR adds two new operators
BitandandNeg, some tests, and analysis and optimization in Egglog that rewrite the naive loop implementation of the lowbit function into the Hacker's Delight one-liner.This now outperforms llvm-O3-O3 by an order of magnitude on my machine because the closed form is an O(log n) speedup (n = 2 * 10^5).
The RVSDGs before and after the optimization:
