Faster dispatch

[markshannon]

For review purposes. Please don't merge.

This should 🤞 result in MSVC producing a dispatch sequence as efficient as GCC with computed gotos.
This (in theory) works by:

• Creating an exhaustive set of switch targets and limiting the opcode to a single byte, so there is no need for any range checking.
• Generating unique code for each target, so the compiler can't merge them.
• Removing all code between the fetch/decode and dispatch, allowing the compiler to merge the goto with the computed jump.

The dispatch code should end up looking something like:

        movzx   ecx, WORD PTR [edi]
        add     edi, 2
        mov     esi, ecx
        movzx   eax, cl
        shr     esi, 8
        jmp     TABLE[eax*4]

The important thing is that there is a computed jump at the end of the instruction code.

[markshannon]

If this works, we can remove all the PREDICT macros as well.

[gvanrossum]

See email I sent with Disassembly.txt

[gvanrossum]

Here's the disassembly for all of _PyEval_EvalFrameDefault() with some extra stuff before and after it.

Disassembly.txt

[markshannon]

There still seem to a lot of conditional jumps and even a call in the DISPATCH() code.
Do you have LLTRACE or DTRACE enabled?

[markshannon]

Closing this.
I'll make another PR with just the extra opcodes, once https://bugs.python.org/issue43760 is done

[gvanrossum]

(And no, I did not have LLTRACE or DTRACE explicitly enabled. Whether they're enabled by default I have no idea -- if you want me to find out I can though.)