[slyfox1186]

Problem and result

A CUDA build prints nvcc warning #128-D "loop is not reachable" 208 times (reported in #1585, which was closed in the
three-day sweep and still reproduces on ffa80d73). Nine kernels end an if constexpr branch with return or
continue and run their plain path after it. In every instantiation that takes the branch, the plain path is dead
code, and nvcc warns once per instantiation:

File	Kernel	Branch
native_bf16.cu	bf16_f32_mmvf_rows_kernel	TS
native_mmvq.cu	native_mmvq_multi_kernel	TS
fused_gr.cu	gr_down_v3_kernel	RB && R2V > 1
iq_kernels.cu	s26_gu_l_kernel, s26_down_l_kernel, s27_gu_kernel, s27_down_kernel	TS
iq_kernels.cu	native_gu_lds_kernel	TI
dflash2.cu	gemm8_kernel	!COMPUTE

After this change the same build prints no #128-D, and the generated code is unchanged.

Related discussion

#1585 (closed by the stale sweep, not fixed). This is a focused fix with no change in behavior.

Approach

Each plain path moves into the branch's else, so it is no longer instantiated where it cannot run. The return or
continue that ended the branch goes away, because the else now reaches the same end of the function or loop
body. gemm8_kernel tested its template parameter with a plain if (!COMPUTE); that becomes if constexpr.

The bodies keep their indentation, the way native_bf16.cu already wraps its plain loop at lines 182-196, so the diff
is only the branch structure: 19 lines in 5 files, identical with git diff -w.

Validation

Linux x86-64, RTX 4090 (sm_89), CUDA 13.4.92, host GCC 15.2.0, the engine build's own nvcc command lines
(-O3 --use_fast_math --generate-code=arch=compute_89,code=[compute_89,sm_89]) on ffa80d73.

Passed:

• Warning count before: replaying each of the build's 83 CUDA compile commands on ffa80d73 gives 208 --gpus with mixed architectures fails: engine not rebuilt when adding a GPU with a different compute capability #128-D
  warnings, all at these nine places (fused_gr 16, native_mmvq 147, native_bf16 32, iq_kernels 11, dflash2 2).
• Warning count after: compiling the five changed files with the same command lines exits 0 and prints no --gpus with mixed architectures fails: engine not rebuilt when adding a GPU with a different compute capability #128-D.
  The other warnings in those files are unchanged (generate: residency-table uploads wait for their own copy before the non-blocking streams read the table #550 dflash2.cu:775, #20199 iq_kernels.cu:3840, #20058
  iq_kernels.cu:5583).
• Generated code: cuobjdump -sass of each object before and after (anonymous-namespace names normalized, since
  they hash the source path) is byte-identical for all 1917 kernels: native_bf16 128, native_mmvq 1005,
  fused_gr 211, dflash2 83, iq_kernels 490.

Failed:
none observed

Not run:

• HIP, SYCL and Metal builds. The change is plain C++17 control flow inside device templates, with no
  backend-specific code, but I have no AMD or Intel GPU toolchain here to compile it.
• Hardware execution. Identical SASS means the sm_89 kernels are the same instructions as before.

Limits

Only #128-D is addressed. The build's other warnings (#63 shift count in qsa_select.cu:2501, #20280 in
s2_expert_grouped.cu, #611, #550, #20199, #20058) are separate problems.