[llvm-dev] flow-sensitive alias analysis

[Oliver Braunsdorf via llvm-dev]

> 
> This assumption is very wrong, AFAIK :)
> The best known precise flow-insensitive algorithms easily handle pretty
> much any  scale these days.
> 
> GCC's points-to is field-sensitive precise flow-insensitive, and it
> pretty much never shows up on profiles.
> By contrast, the best sparse known sparse flow-sensitive analysis varies
> (quite wildly, in fact), is usually between 10x and 100x slower.
> This is true even when, for example, parallelized on GPU's (see,
> e.g., http://dl.acm.org/citation.cfm?id=2555296).
> 
> If you only care about certain pointers, you would be better off with a
> demand-driven technique.
> 
> See, e.g., "Demand-Driven pointer analysis with Strong updates using
> value-flow refinement" 
> 

Seems like I have to read much more :D
Thanks for your advice! I will probably come back with some questions
regarding demand-driven analysis.

-Oliver