Question for the chip design crowd here: Would it be feasible for Intel/AMD to design a fast ISA from the ground up and add it as a runtime option to their x86_64 chips? Like, on context switch to userland, the kernel puts the chip into whatever ISA is in the respective process's binary. And the syscall instruction and interrupts put the chip back into the kernel ISA. So you would have backwards compatiblity, but new stuff could use the faster ISA. I guess my main question is if there would be enough die space to support two separate instruction decoders.
The short answer is no. Basically, you are asking them to really have two sets of cores on a single die and switch back and forth.
I think the more realistic point of view is that x86 has likely run its course, and now it is time to move on.
The long answer is look up Transmeta and do some reading. Basically they designed a RISC CPU core with an external microcode engine, that was able to run x86 instructions. It was a valiant effort, but it didn't work as well as hoped.
Here is a fun history read that mentions Transmeta as well as P.A.Semi. P.A.Semi is the core team that designed the A1 (and following Apple cores) after being bought by Apple in 2008. Prior to that, the team had designed the DEC Alpha (awesome processor) and StrongARM cores. Later, they developed PowerPC cores. A team with really deep CPU talent, and in particular, prior ARM ISA experience, when they were acquired in 2008.
You say no, but the PowerPC 615 is an example of a core which did that. You just need two instruction decoders (and get to keep one of them off at a time, which is awesome for dark silicon reasons).
This is what Intel attempted with Itanium. The hardware implementation was too slow so they pulled it.
More recently is Nvidia's Denver project, which has a private ISA and dynamically translates ARM instructions. It's an open secret they had it working for x86 too, but it was scuttled due to patent threats.
I suspect they never shipped the x86 version because it wasn't competitive. Nvidia's CPUs couldn't keep up with contemporary Arm Cortex parts (probably a good part of the reason they bought them), let alone Intel and AMD stuff.
