Real-world scenario of x>y? abs(x-y+1):0 and its optimization?

I’ve been attending Open Source Promotion Plan 2023 for the last 6 weeks under the instruction of @Allen and nearly finished my project about improving the InstCombine optimization based on ⚙ D122013 [InstCombine] Fold abs of known negative operand when source is sub.

In this project, I added some pattern matching schemes for x>y? abs(x-y+1): 0 and optimize out the abs() using the isImpliedByDomCondition() API. After the optimization, the ternary expression is replaced with x>y? x-y+1: 0 and can generate only one AArch64 instruction csinc. I also try to match similar patterns such as x<y? abs(x-y-1): 0, x>y? abs(2*(x-y)): 0 and optimize them based on arithmetic relationship.

This optimization has already been done in GCC, click here for a demo.

While this optimization seems non-trivial, I can not figure out its real-world application. I try to search for the code pattern needs to be optimized in some large code bases such as Linux kernel using weggli, but there’s no matching result.

weggli --cpp "{$x>$y ? abs($x-$y+1): 0}" .

Can anybody tell me about the similar code pattern you’ve met? Or how to find the optimization scenario the right way? Thanks!

Hey, I once applied for this project too. And congratulations.
From my perspective, there is no much need for real-world motivation to fold that. If IR become “better” after the fold, it’s worthy to do.
Besides, there may be other equivalent forms like

  z = abs(expr)

where expr may be a complex expression folded into x-y+1 after optimization, which would be indirect real-world application.

Thanks. I have finished this project two weeks before, and it turns out that this project really improves the IR by optimizing out the intrinsic abs function, see PR and test code for more details.

As for the equivalent form you have provided, this project can not handle it now because what have been implemented is a peephole optimization and only considers the basic value-flow inside ternary expression. In order for optimizing such case, we need a longer value-flow analysis, do you have any advice on that?