GH-138245: Perform boolean guards by testing a single bit, rather than a full pointer comparison.

[markshannon]

This reduces the overhead of performing boolean guards in jitted code.

On Aarch64 reduces the size of the stencil from 5 to 2 instructions.

GUARD_IS_FALSE_POP_r10:

  adrp    x8, 0x0   // _Py_FalseStruct
  ldr     x8, [x8]   // _Py_FalseStruct
  orr     x8, x8, #0x1
  cmp     x24, x8
  b.ne    jump_target

_GUARD_BIT_IS_SET_POP_4_r10:

  tbnz    w24, #0x4, next
  b    jump_target
next:
    // 0000000000000004:  R_AARCH64_JUMP26     _JIT_JUMP_TARGET

• Issue: Faster True/False checks using alignment #138245

[Fidget-Spinner]

Please check this once at the context initialization and set it in the context, then fetch the corresponding op to use from the context.

[markshannon]

I thought about that, but decided that having simpler, stateless code was better than a saving a few cycles out of many thousands optimizing a trace.
These functions are only called 2 or 3 times per trace (on average) and are really small and fast.

[Fidget-Spinner]

Same here.

[Fidget-Spinner]

Also do you have performance numbers? Even microbenchmarks are fine.

I see a pretty big difference (around 10%) in Richards.

[markshannon]

Also do you have performance numbers? Even microbenchmarks are fine.

No. It didn't seem worth measuring as I didn't expect the impact to be above the noise.
The generated code is just better: fewer instructions, no memory access, and shorter instructions (for x86).

I see a pretty big difference (around 10%) in Richards.

I'm surprised it makes much difference.
Better or worse?

[Fidget-Spinner]

I'm surprised it makes much difference. Better or worse?

Sorry that was a fluke, I can't reproduce it anymore.