@ROCStaticLocalArray's default zeroinit costs ~11x: every lane memsets the whole LDS array at kernel entry

@ROCStaticLocalArray (and @ROCDynamicLocalArray) default to zeroinit=true, and the zero-initialization is implemented in a way that costs an order of magnitude on any kernel that uses it.

Measurement (MI300A, ROCm 7.2.4, Julia 1.12.6): a 256-thread LDS tree reduction over 2²⁶ Float32, identical code except the macro argument:

Why it is this expensive. zeroinit_lds! (src/compiler/zeroinit_lds.jl) inserts a memset of the whole LDS global at kernel entry followed by s_barrier. Every work-item executes that memset, so for a 1 KiB array the ISA is 64 × ds_write_b128 per wave, all 64 lanes writing the same address each time (ds_write_b128 v1, v[2:5] offset:N with v1 = 0), which serializes on bank conflicts — and that's repeated by every wave in the workgroup, on every launch. The prologue dwarfs the reduction itself.

Related surfaces

• KernelAbstractions.@localmem (KA.SharedMemory → alloc_special without zeroinit) and src/kernels/mapreduce.jl (@ROCDynamicLocalArray(T, items, false)) already avoid it, so library code is fine; only the user-facing macros pay by default.
• @ROCDynamicLocalArray's zeroinit is a per-work-item serial loop over the whole array (for idx in 1:prod(dims); DA[idx] = zero(T)), same pathology.
• CUDA.jl's CuStaticSharedArray does not zero shared memory; neither does HIP's __shared__.

Suggested fix: default zeroinit=false (matching CUDA.jl/HIP semantics; document that LDS contents are undefined), and if zeroing is requested, distribute it — each work-item clears its strided slice (for i in tid:groupsize:len) before the barrier — instead of every lane clearing everything.

Start with src/compiler/zeroinit_lds.jl and trace how @ROCStaticLocalArray and @ROCDynamicLocalArray request zero-initialization. Compare the existing false paths, then validate the reported reduction benchmark. Done means zero-initialization is opt-in, requested clearing is distributed across work-items, and the undefined-LDS behavior is documented.