Consider using new LLVM intrinsic for i32::saturating_add etc.

[hanna-kruppe]

As featured in LLVM Weekly, https://reviews.llvm.org/rL344629 introduces @llvm.sadd.sat for saturating signed integer addition. Not sure whether it's as transparent to the optimizer yet as open-coding it based on checked_add (as we currently do), but we might want to try it when we update to an LLVM version that has that intrinsic available.

[nikic]

Right now add.sat is completely opaque. I've submitted https://reviews.llvm.org/D54237 for ConstantFolding and InstCombine support, which should be enough to properly handle #52203.

[nikic]

All the optimizations have landed, so we can give this a try after the next LLVM update.

[scottmcm]

Looks like a sufficiently-new LLVM has landed in nightly, and the folding logic is working: https://play.rust-lang.org/?version=nightly&mode=release&edition=2018&gist=75e0b8364fc0d815edc1d091840fca47

[nikic]

Work on saturating add/sub intrinsics is still ongoing (turns out that getting new intrinsics to optimize and codegen well is a pretty big rabbit hole), but I believe at this point they're good enough that we should give them a try in Rust. I'll submit a PR later this week, if no one beats me to it.