// Part 3 definition-and-launch carrier inputs for loom-lower-forall-to-thread.
// Each sample is a module of `dataflow.thread` definitions (dense domain,
// private visibility, entry block `(args_*, thread_ctrl, coord_*)`) whose
// bodies carry a mapping-free, effect-form, compile-time fixed-domain
// `scf.forall` with no shared_outs, no results and an empty in_parallel
// terminator, paired with the module-scope launch of each definition.

start: {new NTHREADS = random.randint(1, 3); new T = 0} module_body;

module_body: (T < NTHREADS) thread_unit {T += 1} module_body
           | (T == NTHREADS) '';

thread_unit: {new EXTRA = random.randint(0, 2); new RANK = random.randint(0, 2); new TYPES = []; new A = 0; new B = 0; new C = 0; new D = 0; new G = 0; new H = 0; new K = 0}
    'dataflow.thread private @thread_' t_idx ' domain(#dataflow.thread_domain<dense>)(%buf: memref<8xi32>' extra_decls ') ctrl (%thread_ctrl: none)' iv_clause ' {\n'
    thread_body
    '}\n'
    host_func;

t_idx: [str(T)];

// Payload arguments after the leading memref: the `args_*` prefix.
extra_decls: (A < EXTRA) {TYPES.append(random.choice(['i32', 'f32', 'index']))} ', %arg' a_idx ': ' a_ty {A += 1} extra_decls
           | (A == EXTRA) '';

a_idx: [str(A)];
a_ty: [TYPES[A]];

// Dense coordinate suffix: one `index` slot per launch-domain dimension.
iv_clause: (RANK == 0) ''
         | (RANK > 0) ' iv (' iv_decls ')';

iv_decls: (K < RANK) iv_sep '%coord_' k_idx ': index' {K += 1} iv_decls
        | (K == RANK) '';

iv_sep: (K == 0) '' | (K > 0) ', ';
k_idx: [str(K)];

thread_body: '  %value = arith.constant 7 : i32\n'
             forall_op
             yield_op;

// Effect-form forall: no shared_outs, no results, no mapping attribute,
// compile-time fixed zero-based domain, empty in_parallel terminator.
forall_op: '  scf.forall (%lane) in (' extent ') {\n'
           '    memref.store %value, %buf[%lane] : memref<8xi32>\n'
           '  }\n'
         | '';

extent: '2' | '4' | '8';

yield_op: '  dataflow.thread.yield\n'
        | '  dataflow.thread.yield %thread_ctrl : none\n';

host_func: 'func.func @host_' t_idx '() {\n'
           '  %buf = memref.alloc() : memref<8xi32>\n'
           extra_defs
           grid_defs
           '  %token = dataflow.thread.launch @thread_' t_idx '(%buf' extra_uses ')' grid_clause ' : (memref<8xi32>' extra_types ') -> !dataflow.thread_token\n'
           '  return\n'
           '}\n';

extra_defs: (B < EXTRA) '  %h' b_idx ' = ' const_rhs '\n' {B += 1} extra_defs
          | (B == EXTRA) '';

b_idx: [str(B)];

const_rhs: (TYPES[B] == 'i32') 'arith.constant 3 : i32'
         | (TYPES[B] == 'f32') 'arith.constant 1.000000e+00 : f32'
         | (TYPES[B] == 'index') 'arith.constant 2 : index';

extra_uses: (C < EXTRA) ', %h' c_idx {C += 1} extra_uses
          | (C == EXTRA) '';

c_idx: [str(C)];

extra_types: (D < EXTRA) ', ' d_ty {D += 1} extra_types
           | (D == EXTRA) '';

d_ty: [TYPES[D]];

grid_defs: (G < RANK) '  %g' g_idx ' = arith.constant 4 : index\n' {G += 1} grid_defs
         | (G == RANK) '';

g_idx: [str(G)];

grid_clause: (RANK == 0) ''
           | (RANK > 0) ' grid(' grid_uses ')';

grid_uses: (H < RANK) grid_sep '%g' h_idx {H += 1} grid_uses
         | (H == RANK) '';

grid_sep: (H == 0) '' | (H > 0) ', ';
h_idx: [str(H)];