[SR-2045] Optimization of FloatingPoint `.round` for Float/Double/Float80

[stephentyrone]

Previous ID	SR-2045
Radar	None
Original Reporter	@stephentyrone
Type	Improvement
Status	Closed
Resolution	Done

Additional Detail from JIRA

Votes	0
Component/s	Standard Library
Labels	Improvement
Assignee	@stephentyrone
Priority	Medium

md5: 087779091399356b02a215851a43df40

Issue Description:

These functions contain an extra layer of indirection as currently implemented; they call e.g. `_swift_stdlib_ceilf( )`, which turns around and either calls `ceilf( )` or contains a lowering of `ceilf( )` (arm64).

This extra overhead is especially significant for these functions in particular because they have single-instruction implementations for `Float` and `Double` on arm64 and x86. We're eating a function call where we should simply have a single instruction instead.

[stephentyrone]

#3454

[xwu]

Out of curiosity, is the same indirection also worth addressing for fused multiply-add and remainder?

[stephentyrone]

Yes, definitely (especially for FMA, remainder is relatively rare and expensive no matter what). I think there's a bug for that kicking around somewhere, but if there isn't I'll take a look.

[stephentyrone]

@xwu I threw fma and sqrt in with this change. Remainder requires more fiddling, and is much less important.

[xwu]

Thanks! I'd offer to pitch in, but one look at the plumbing in lib/AST/* and it's clear I don't have that skill.

[stephentyrone]

Found a solution for remainder too. No AST plumbing necessary.

[stephentyrone]

Done for Apple platforms. `squareRoot` and `remainder` will need special handling for Linux, but that can be done under a separate bug.