Bug 194222 – Reduce the number of iterations in B3ReduceStrength

• Status: NEW
• Component: JavaScriptCore

[Robin Morisset]

B3ReduceStrength repeats a sequence of optimizations to a fixpoint.
One of these optimizations is a peephole optimizer, which currently does not revisit any node it adds or changes.
So if it must be called 3 times to fully reduce an expression, the whole sequence of optimizations will be repeated 3 times (at least), even if the rest of the function was fully optimized in one pass.

I'll try to make the peephole optimizer apply repeatedly whenever it adds/changes nodes, and measure how much it improves FTL compile-time.

[Robin Morisset]

Here are some quick numbers for ARES-6:
1	62
2	351
3	997
4	848
5	108
6	18
Meaning that B3ReduceStrength runs only two iterations in 351 cases, runs 4 iterations in 848 cases, etc..
If nothing is done, I expect this to get worse as B3ReduceStrength is taught more optimizations (e.g. https://bugs.webkit.org/show_bug.cgi?id=194081).
Since the entire pass is repeated on each iteration, reducing this number of iterations should directly improve B3 compile time.

[Robin Morisset]

Created attachment 361360 [details]
WIP

Only supports Mul with the new architecture so far.
Is currently recursive, I may have to rearchitect it into an iterative approach if it causes stack issues.

[Robin Morisset]

Created attachment 361361 [details]
WIP

Simpler, and now most recursive calls are tail-recursive which should alleviate the stack concerns.

[Robin Morisset]

Created attachment 361452 [details]
WIP_Arith

Now supports Add/Sub/Neg/Mul/Div/UDiv/Mod/UMod.
Also now takes a kind as argument instead of an opcode, to correctly deal with chill operations.
Finally I commented out an optimization from https://bugs.webkit.org/show_bug.cgi?id=194250 that seems buggy while I investigate.

[Robin Morisset]

Created attachment 361467 [details]
WIP_Arith+Shifts+FP

Add shifts, rotations, abs, ceil, floor.

[Robin Morisset]

Created attachment 361480 [details]
WIP_More

Adds conversions, select, equal.

[Robin Morisset]

On further investigation, it appears that the peephole optimizations are not the bottleneck on the number of iterations in B3ReduceStrength. So I will make this bug an umbrella for other more precise bugs fixing the actual problems.