GH-113464: Generate a more efficient JIT

[brandtbucher]

This breaks up the JIT into smaller functions, reduces a lot of branching in hot inner loops, and generally makes the C code cleaner (and probably faster).

Currently, we generate declarative structures at build time that we then loop over in order to emit the desired machine code at runtime. For example, the _STORE_FAST stencil looks like this:

Details

staticconstunsigned char_STORE_FAST_code_body[61] ={0x50, 0x48, 0x8b, 0x45, 0xf8, 0x48, 0x83, 0xc5, 0xf8, 0x0f, 0xb7, 0x0d, 0x00, 0x00, 0x00, 0x00, 0x49, 0x8b, 0x7c, 0xcd, 0x48, 0x49, 0x89, 0x44, 0xcd, 0x48, 0x48, 0x85, 0xff, 0x74, 0x0f, 0x48, 0x8b, 0x07, 0x85, 0xc0, 0x78, 0x08, 0x48, 0xff, 0xc8, 0x48, 0x89, 0x07, 0x74, 0x07, 0x58, 0xff, 0x25, 0x00, 0x00, 0x00, 0x00, 0xff, 0x15, 0x00, 0x00, 0x00, 0x00, 0x58, }; staticconstHole_STORE_FAST_code_holes[4] ={{0xc, HoleKind_R_X86_64_GOTPCREL, HoleValue_DATA, NULL, -0x4},{0x31, HoleKind_R_X86_64_GOTPCRELX, HoleValue_DATA, NULL, 0x4},{0x37, HoleKind_R_X86_64_GOTPCRELX, HoleValue_DATA, NULL, 0xc}, }; staticconstunsigned char_STORE_FAST_data_body[25] ={0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, }; staticconstHole_STORE_FAST_data_holes[4] ={{0x0, HoleKind_R_X86_64_64, HoleValue_OPARG, NULL, 0x0},{0x8, HoleKind_R_X86_64_64, HoleValue_CONTINUE, NULL, 0x0},{0x10, HoleKind_R_X86_64_64, HoleValue_ZERO, &_Py_Dealloc, 0x0}, };

This very general approach means that we have a lot of complex logic in our hot inner loop to decode instructions and set up values for patching that may not even be needed. It's also very branchy, since we're essentially "interpreting" the array of holes for each instruction.

With this PR, jit_stencils.h instead contains the following function:

Details

voidemit__STORE_FAST( unsigned char*code, unsigned char*data, _PyExecutorObject*executor, const_PyUOpInstruction*instruction, uintptr_tinstruction_starts[]){constunsigned charcode_body[60] ={0x50, 0x48, 0x8b, 0x45, 0xf8, 0x48, 0x83, 0xc5, 0xf8, 0x0f, 0xb7, 0x0d, 0x00, 0x00, 0x00, 0x00, 0x49, 0x8b, 0x7c, 0xcd, 0x48, 0x49, 0x89, 0x44, 0xcd, 0x48, 0x48, 0x85, 0xff, 0x74, 0x0f, 0x48, 0x8b, 0x07, 0x85, 0xc0, 0x78, 0x08, 0x48, 0xff, 0xc8, 0x48, 0x89, 0x07, 0x74, 0x07, 0x58, 0xff, 0x25, 0x00, 0x00, 0x00, 0x00, 0xff, 0x15, 0x00, 0x00, 0x00, 0x00, 0x58, }; constunsigned chardata_body[24] ={0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, }; memcpy(data, data_body, sizeof(data_body)); patch_64(data+0x0, instruction->oparg); patch_64(data+0x8, (uintptr_t)code+sizeof(code_body)); patch_64(data+0x10, (uintptr_t)&_Py_Dealloc); memcpy(code, code_body, sizeof(code_body)); patch_32r(code+0xc, (uintptr_t)data+-0x4); patch_x86_64_32rx(code+0x31, (uintptr_t)data+0x4); patch_x86_64_32rx(code+0x37, (uintptr_t)data+0xc)}

This function is called directly to emit the machine code for every _STORE_FAST instruction, and hardcodes the logic for all of the necessary copies and patches. The result is one indirect call, no unnecessary branching, and (in my opinion) cleaner code, since a lot of the tricky logic is now hidden away in generated files.

---

I know this is right before feature freeze, but I'd really like to get this in 3.13 since it will make backporting any fixes much easier. It doesn't change the actual jitted code in any way.

Note to reviewers: the diff is a bit messy, so it may make more sense to compare the before-vs-after files side-by-side instead.

• Issue: JIT Compilation #113464

[brandtbucher]

@savannahostrowski, I'd love to get your review of this if you have a few cycles.

[savannahostrowski]

A couple of comments and questions but after sitting and reading through this code a bunch over the last week or two, I'm excited about how much more readable this will get with this change! 💆‍♀️

[savannahostrowski]

Is there a reason that this needs to be sorted?

[brandtbucher]

Nope, I just like it that way (if you couldn't tell by now). ;)

[savannahostrowski]

Thanks for adding in the comment about the naming conventions - I think that helps! Otherwise, this looks pretty solid to me (barring some Windows CI failures). Lots of moving things into function but it's a whole lot more readable! 🎉

[brandtbucher]

Windows JIT CI fixed in GH-118564.