Host-side prefetch fails to compile for targets with the hvx feature #9496

Generated with GitHub Copilot CLI

Prefetches in host code fail to compile when the target has the hvx feature (e.g. host-hvx), even if nothing is offloaded with .hexagon().

Reproducer:

#include "Halide.h"
using namespace Halide;
int main() {
    Var x, y;
    ImageParam input(Int(32), 2, "input");
    Func f("f");
    f(x, y) = input(x, y) * 2;
    f.prefetch(input, y, y, 2);
    Buffer<int> in(4, 4);
    in.fill(3);
    input.set(in);
    Buffer<int> r = f.realize({4, 4});
    printf("ok %d\n", r(0, 0));
}

With HL_JIT_TARGET=host this prints ok 6. With HL_JIT_TARGET=host-hvx it fails with:

Error: Only prefetch of 1 cache line is supported.

reduce_prefetch_dimension (Prefetch.cpp) chooses Hexagon's 2D prefetch form whenever t.has_feature(Target::HVX). It runs in lower() before inject_hexagon_rpc splits off the .hexagon() loops, so prefetches that stay in host code also get the 2D form, which CodeGen_LLVM rejects. It should probably only use the 2D form when t.arch == Target::Hexagon or the prefetch is inside a DeviceAPI::Hexagon loop.