Support for allocation of executable memory

[Maurizio Cimadamore]

On 22/11/2021 16:41, Paul Sandoz wrote:
> As shown in the thread it's possible with Panama’s FFM API to do something similar, but the degree of optimization compared to C2 intrinsics will be limited right now. The code snippets prototype was more optimally integrated with C2, more intrinsic-like than native-like. Perhaps over time it may be possible to close the gap for certain kinds of native functionality (i.e. like that internal JNI fast path the name of which escapes me right now), which depends on the properties of that functionality.

I think it would be ultimately great to land in the place you describe. 
One caveat: to get there, having executable memory segments is just the 
tip of the iceberg. That is, of course to enable that at some point we 
will need to create a memory segment which points to a piece of 
executable memory - but that's a low-level detail. The high-order bit 
missing is that we need a way to describe the computation that is 
associated with that intrinsic. Directly acccepting assembly is too 
low-level - and specifying behavior in Java is too high-level. As in the 
case of the Vector API, it seems like something "in between" is needed.

In other words, the Foreign API gives us the basic capability to create 
a memory segment and (through linker API, as Glavo noted) change its 
permission to executable. We can then even link the address of that 
segment into a method handle, using the linker API, and have it behave 
as a regular native function. But to get to C2 intrinsification nirvana, 
we need a way to describe what's inside the computation bubble.

Maurizio