One more missed bounds check optimization

[bugadani]

https://godbolt.org/z/3TGbK9

It looks like bounds checks are not eliminated when more than one restriction is applied to the index variable (i.e. when it's range checked from two sides). Late edit: It's also interesting that the issue disappears for certain lower limits. For example if idx > 2 && idx < 8 is optimized, while if idx > 5 && idx < 8 is not.

Using assumptions, the bounds checks are removed as expected: https://godbolt.org/z/Ksn8dW
However, if any of the assumptions is a GE or LE instead of a GT or LT, the bounds checks are back: https://godbolt.org/z/YhK1rT

[tesuji]

@rustbot modify labels: A-LLVM A-mir-opt C-enhancement I-slow T-compiler

[tesuji]

This could be improved by optimizing MIR. But the problem rely on LLVM too.
C version: https://godbolt.org/z/M3Ynq6
Clang trunk has more optimizations than v10.0.1.
GCC is still ahead in these optimizations.

[the8472]

Same as #73396?

[davidbolvansky]

Posted on llvm bugzilla:
https://bugs.llvm.org/show_bug.cgi?id=47172

[bugadani]

Thanks for that report!

I've played some more with my examples and these three cases are particularly interesting for me: https://godbolt.org/z/z6o3Pc
I'm curious if the improvements shown by Clang trunk are caused by clang or llvm, and if llvm, will it fix these as well. I guess time will tell this :)

[davidbolvansky]

Since Clang does no optimizations, this is all about about LLVM:)

Also worth to see:
https://reviews.llvm.org/D84547