Some minor optimizations:

1. The table base address is currently a 10-byte MOV instruction for each opcode. We should use a RIP-relative LEA here (7 bytes, the difference adds up to a few hundred bytes and it's what C++ compilers do). More importantly, on ARM64 we currently use LDR where we can similarly use ADR (because we know the interpreter code is much smaller than 1 MB).

2. We could order the list of bytecode ops in BaselineCodeGen.h on measured frequency in the browser. This should be a bit more cache friendly.

3. The toggled call for the debugger is unfortunately implemented pretty inefficiently on ARM64 (sync stack pointer, LDR, NOP). We should get this down to one instruction, probably with toggledJump.

This affects the following platforms:

• x64: use a RIP-relative LEA instead of an immediate MOV. This saves a few
  hundred bytes total and seems to be a little bit faster on interpreter
  micro-benchmarks.

• arm64: use ADR instead of LDR.

Depends on D43398

We now use real NOPs on all platforms. On x86/x64 this used to be a CMP
instruction and on ARM64 this involved an unconditional LDR with some
other instructions.

Depends on D43413

(In reply to Jan de Mooij [:jandem] from comment #0)

2. We could order the list of bytecode ops in BaselineCodeGen.h on measured frequency in the browser. This should be a bit more cache friendly.

I'm leaving this one for now, we can always try it later.

https://hg.mozilla.org/mozilla-central/rev/dae1e9839adc
https://hg.mozilla.org/mozilla-central/rev/feec09fd96eb
https://hg.mozilla.org/mozilla-central/rev/e72770318826