Faster True/False checks using alignment

[markshannon]

Currently to check whether a value is True or False we need to compare it either True or False.
In the both the JIT and the interpreter (although this is worse in the JIT) we need to load a full 8 byte value from the instruction stream in order to perform the comparison.
Here's the x86 stencil for _GUARD_IS_TRUE_POP (with TOS caching):

    // 0: 48 b8 00 00 00 00 00 00 00 00 movabsq $0x0, %rax
    // 0000000000000002:  R_X86_64_64  _Py_TrueStruct+0x1
    // a: 4c 39 f8                      cmpq    %r15, %rax
    // d: 0f 85 00 00 00 00             jne     exit

By putting True and False in an aligned array:

struct _booleans {
     PyLongObject False;
     PyLongObject True;
};

alignas(sizeof(struct _booleans)) struct _booleans _PyBooleans = {
   /* Data for False */
   /* Data for True */
};

we can use the alignment to check for True or False by testing a single bit: ref.bits & sizeof(PyLongObject), resulting in this very efficient stencil for _GUARD_IS_TRUE_POP:

    // 0: 41 f6 c7 20                   testb   $0x20, %r15b
    // 4: 0f 84 00 00 00 00             je      exit

Linked PRs

• GH-138245: Perform boolean guards by testing a single bit, rather than a full pointer comparison. #143810

[CarlosEduR]

I find this issue really interesting!! I don't have much experience with CPython internals yet, but I'd like to study this area and hopefully work on similar issues in the future.

[markshannon]

It looks like _Py_FalseStruct and _Py_TrueStruct are part of the stable ABI until 3.13.
So it will be tricky do this without an ABI break.

[markshannon]

It doesn't matter which bit differs, as long as some bit differs (and they must).
We just need to compute the bit at runtime.
This means we can't do this optimization in the interpreter efficiently (yet), but we can still do it in the JIT