One BEND_RTC macro for the six __CUDACC_RTC__ tests, and gpu_hash above the lanes

[costamatheus97]

The two pieces of the HIP lane (#958, merged to hip) that do not depend on it, as offered in that thread. One commit on main at e52cda4, bend2/comp.ts only, +23 −14. No behaviour change.

What it changes

• BEND_RTC. Defined when the file is being compiled by the device's runtime compiler (#if defined(__CUDACC_RTC__)). The six __CUDACC_RTC__ tests read it instead. A second RTC lane adds its macro to that one line and leaves the six sites alone.
• gpu_hash. Moved out of the CUDA block to just above the lane blocks, under #if BEND_CUDA. It reads only BEND_SRC and CUBE_LOG, both defined earlier. A lane that keeps a .gpu cache shares it by widening that guard. hip already builds and runs with the function at this position.

With both on main, hip differs from main on these lines by || defined(__HIPCC_RTC__) and || BEND_HIP, which should keep its rebases quiet in shared code.

Checks

I have no CUDA or Metal hardware, so the main check is on the preprocessor. I emitted C for tests/run/gpu_mark.bend from pristine main and from this branch, removed the #embed __FILE__ line from both (the embedded text differs by construction), and compared clang -E -P output:

configuration	result
host, no GPU lane	identical
device, -D__CUDACC_RTC__	identical
device, -D__METAL_VERSION__	identical
host, -DBEND_CUDA=1 (stub cuda.h / nvrtc.h)	identical except gpu_hash appears earlier

• A one-machine copy of gates/test.ts's rules over gpu_*, compile_*, io_hello*, gfx_*: 120 / 122 on this branch and 120 / 122 on pristine main. The two failures are the same on both, gfx_clicks and gfx_window, which need a display.
• comp.ts is 64967 tokens against the 65000 cap. I do not have ttok here; I counted with cl100k_base, which gives exactly your 64941 for main and 68028 for hip.
• Not run: the gates themselves (gates/repo.ts needs ttok), the full test gate, anything on CUDA or Metal hardware.

[VictorTaelin]

Merged into 2.0.25 under your authorship, thank you. We confirmed the preprocessed C is identical on the host, under CUDA RTC and under Metal, and the Metal tests pass here. hip should now differ from main on those lines only by its own || defined(__HIPCC_RTC__) and || BEND_HIP.

Note: this reply was written by an AI after it reported the issue to me and I made the decision. If anything here is wrong, reply and I will review it myself.