https://gcc.gnu.org/bugzilla/show_bug.cgi?id=117282
--- Comment #3 from Feng Xue <fxue at os dot amperecomputing.com> ---
(In reply to Andrew Pinski from comment #2)
> >could we introduce a new option like -mcmodel-dyn{=tiny/small/large} as a
> >supplement to -mcmode
>
> This would cause too many confusion really. especially depending on the
> definition of your tiny/small/large. I rather see other optimizations happen
> in before this kind of hacks.
This code piece(with conversion) exists in SIR17 523.xalanc, 7% improvement if
optimized.
"tiny/small/large" is from "-mcmodel", target specific, or define a middle-end
aware option like: -fdata-object-size="size-limit", default value is hardware
address space. For normal 64-bit "small" application, we could set
"-fdata-object-size=2^32". Actually, this option could also help VRP,
especially pointer-difference operation.
